test(macos): cover cached image selection

fix(rebase): correct test_cached_lookup key and drop duplicate test
Two issues introduced by the rebase conflict resolution: - test_cached_lookup_requires_ready_marker used _dockerfile_hash as the cache-dir key, but cached_agent_rootfs_dir now uses _rootfs_digest (which also folds in the guest init). Updated the test to match. - test_pyproject_toml_change_bumps_version appeared twice in test_infra_artifact.py (once from main, once from the PR commit that added pyproject.toml support). Removed the duplicate. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-07-21 05:54:42 +00:00 · 2026-07-21 01:50:45 -04:00 · 2026-07-21 01:50:45 -04:00 · 2026-07-21 01:50:45 -04:00 · 2026-07-21 01:50:45 -04:00 · 2026-07-21 01:50:45 -04:00
169 changed files with 13492 additions and 2083 deletions
@@ -22,10 +22,7 @@ jobs:
      - name: Checkout
        uses: actions/checkout@v4

-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.12"
-
+      # No actions/setup-python: canaries are stdlib unittest on the image's
+      # system Python 3.12 (older act_runner mishandles setup-python's PATH).
      - name: Run canaries
        run: python3 -m unittest discover -t . -s tests/canaries -v
@@ -13,20 +13,30 @@ jobs:
    steps:
      - uses: actions/checkout@v3

-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.12"
-
+      # No actions/setup-python: the runner image already ships Python 3.12,
+      # and older act_runner engines mishandle setup-python's PATH. Install
+      # into the ephemeral job container's system Python — the pylint/pyright
+      # console scripts land on /usr/local/bin (on PATH) so the steps below
+      # still resolve. --break-system-packages is safe: the container is
+      # disposable.
      - name: Install dev dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-dev.txt
+        run: python3 -m pip install --break-system-packages -r requirements-dev.txt

      - name: Run pylint
        run: |
-          # Run pylint on all Python files in the repo
-          find . -name '*.py' -not -path './.venv/*' -not -path './.git/*' | xargs pylint --fail-under=8.0
+          # Pylint's normal exit code is nonzero for any emitted finding,
+          # regardless of --fail-under. Preserve the full report but enforce
+          # the aggregate score this workflow promises.
+          set +e
+          find . -name '*.py' -not -path './.venv/*' -not -path './.git/*' \
+            | xargs pylint --fail-under=8.0 \
+            | tee /tmp/pylint-output.txt
+          set -e
+          SCORE=$(sed -n \
+            's/^Your code has been rated at \([-0-9.]*\)\/10.*/\1/p' \
+            /tmp/pylint-output.txt | tail -1)
+          test -n "$SCORE"
+          awk -v score="$SCORE" 'BEGIN { exit !(score >= 8.0) }'

      - name: Run pyright
        run: |
@@ -37,11 +37,8 @@ jobs:
          fetch-depth: 0
          token: ${{ secrets.GITHUB_TOKEN }}

-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.12"
-
+      # No actions/setup-python: the inline script is stdlib-only on the
+      # image's system Python 3.12 (older act_runner mishandles its PATH).
      - name: Configure git
        run: |
          git config user.name "github-actions[bot]"
@@ -4,16 +4,15 @@
 # dependencies are required to execute it. Tests are split by directory:
 #
 #   tests/unit/         — pure unit tests; always run
-#   tests/integration/  — need a reachable Docker daemon; skip cleanly
-#                         (via tests/_docker.py:skip_unless_docker) when
-#                         Docker isn't available on the runner
+#   tests/integration/  — need a reachable backend; skip cleanly when
+#                         the backend isn't available on the runner
 #   tests/canaries/     — upstream regression canaries; run on a separate
 #                         schedule (see canaries.yml), not here
 #
-# This workflow assumes the Gitea Actions runner exposes the host Docker
-# socket to the job container so `docker` commands inside the job can
-# reach the daemon. If that's not yet configured on the runner the
-# integration tests will skip rather than fail.
+# Integration tests run once per backend in separate jobs. Each job sets
+# BOT_BOTTLE_BACKEND explicitly so the test suite uses the right backend.
+# Backends that aren't available on the runner fail the preflight step
+# rather than silently skipping inside the test output.

 name: test

@@ -23,24 +22,84 @@ on:
      - main
    paths:
      - '**.py'
+      - '.gitea/workflows/**.yml'
+      - 'scripts/**'
+      - 'README.md'
+      # Dockerfiles and pyproject.toml are baked into the infra rootfs; a
+      # change here alters what the integration/coverage jobs build locally.
+      - 'Dockerfile*'
+      - 'pyproject.toml'
  pull_request:
    paths:
      - '**.py'
+      - '.gitea/workflows/**.yml'
+      - 'scripts/**'
+      - 'README.md'
+      - 'Dockerfile*'
+      - 'pyproject.toml'
+  workflow_dispatch:

 jobs:
+  stage-firecracker-inputs:
+    runs-on: [self-hosted, kvm]
+    # Same guard as the other KVM-runner jobs: don't spin the privileged
+    # runner for fork PRs (this only copies a non-secret static binary, but
+    # keep the posture consistent — build-infra/integration/coverage all
+    # depend on it, so gating here gates the whole Firecracker chain).
+    if: >-
+      github.event_name == 'push' ||
+      github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'pull_request' &&
+       github.event.pull_request.head.repo.full_name == github.repository)
+    steps:
+      - name: Stage the provisioned static dropbear
+        run: |
+          mkdir -p firecracker-inputs
+          cp /var/cache/bot-bottle-fc/dropbear firecracker-inputs/dropbear
+
+      - name: Upload Firecracker build inputs
+        uses: actions/upload-artifact@v3
+        with:
+          name: firecracker-inputs
+          path: firecracker-inputs/
+
+  build-infra:
+    needs: stage-firecracker-inputs
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Download Firecracker build inputs
+        uses: actions/download-artifact@v3
+        with:
+          name: firecracker-inputs
+          path: firecracker-inputs
+
+      - name: Build infra candidate from this checkout
+        env:
+          BOT_BOTTLE_FC_DROPBEAR: ${{ github.workspace }}/firecracker-inputs/dropbear
+        run: python3 -m bot_bottle.backend.firecracker.publish_infra --output infra-candidate
+
+      - name: Upload infra candidate
+        uses: actions/upload-artifact@v3
+        with:
+          name: infra-candidate
+          path: infra-candidate/
+
  unit:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout
        uses: actions/checkout@v4

-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.12"
-
+      # No actions/setup-python: the runner image already ships Python 3.12,
+      # and older act_runner engines mishandle setup-python's PATH (coverage
+      # lands in one interpreter, `python3` resolves to another). Install
+      # straight into the ephemeral job container's system Python —
+      # --break-system-packages is safe because the container is disposable.
      - name: Install dev requirements
-        run: python3 -m pip install -r requirements-dev.txt
+        run: python3 -m pip install --break-system-packages -r requirements-dev.txt

      - name: Run unit tests
        run: python3 -m coverage run -m unittest discover -t . -s tests/unit -v
@@ -48,17 +107,14 @@ jobs:
      - name: Report unit coverage
        run: python3 -m coverage report -m

-  integration:
+  integration-docker:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout
        uses: actions/checkout@v4

-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.12"
-
+      # No actions/setup-python (see the note in the `unit` job); the
+      # container's system Python 3.12 runs the stdlib test suite directly.
      - name: Show environment
        run: |
          python3 --version
@@ -68,33 +124,150 @@ jobs:
            echo "docker not on PATH — integration tests will skip"
          fi

-      - name: Run integration tests
+      - name: Run integration tests (docker)
+        env:
+          BOT_BOTTLE_BACKEND: docker
        run: python3 -m unittest discover -t . -s tests/integration -v

-  # Combined unit+integration coverage report (informational). See
-  # docs/decisions/0004-coverage-policy.md.
+  # Integration tests against the Firecracker backend. Runs on a self-hosted
+  # KVM runner (label `kvm`) where /dev/kvm and the TAP/nft pool are available.
  #
-  # The hard diff-coverage gate (changed lines >= 90%) is DEFERRED: the
-  # Firecracker backend's VM/SSH orchestration is covered by the integration
-  # suite, which needs /dev/kvm + the provisioned TAP/nft pool — a
-  # container-based runner skips it and those lines read uncovered, so the
-  # gate can't pass here. Re-enabling it on a self-hosted KVM runner is
-  # tracked separately (see PRD 0069 / #348 and the ci-runner branch).
+  # Restricted to same-repo PRs, push to main, and workflow_dispatch — fork
+  # PRs don't execute untrusted code on the privileged runner.
+  #
+  # Runner prerequisites (provision once; see README "Firecracker on Linux"):
+  # `firecracker` on PATH, `/dev/kvm` accessible, cached kernel +
+  # static dropbear, and the pool as a persistent systemd unit.
+  integration-firecracker:
+    needs: build-infra
+    runs-on: [self-hosted, kvm]
+    if: >-
+      github.event_name == 'push' ||
+      github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'pull_request' &&
+       github.event.pull_request.head.repo.full_name == github.repository)
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Preflight — Firecracker host is ready
+        run: |
+          command -v firecracker >/dev/null || {
+            echo "firecracker not on PATH — provision the runner (README: Firecracker on Linux)"; exit 1; }
+          test -e /dev/kvm || { echo "/dev/kvm missing — KVM not available on this runner"; exit 1; }
+          # `backend status` exits non-zero unless the TAP pool is up + no
+          # range overlap; it prints the exact `backend setup` fix.
+          python3 cli.py backend status --backend=firecracker
+
+      - name: Download the candidate built from this checkout
+        uses: actions/download-artifact@v3
+        with:
+          name: infra-candidate
+          path: infra-candidate
+
+      - name: Replace the persistent infra VM with the candidate
+        run: python3 -c 'from bot_bottle.backend.firecracker import infra_vm; infra_vm.stop()'
+
+      # No dev-requirements install: the integration suite runs on stdlib
+      # `unittest` (pylint/pyright are lint.yml's concern, not this job's),
+      # and the self-hosted runner's Nix python env has no `pip` module
+      # (`python3 -m pip` → "No module named pip"). Nothing to install.
+      - name: Run integration tests (firecracker)
+        env:
+          BOT_BOTTLE_BACKEND: firecracker
+          BOT_BOTTLE_INFRA_ARTIFACT_DIR: ${{ github.workspace }}/infra-candidate
+        run: python3 -m unittest discover -t . -s tests/integration -v
+
+  # Combined unit+integration coverage + the diff-coverage gate (the hard
+  # gate: new/changed lines >= 90%). See docs/decisions/0004-coverage-policy.md.
+  #
+  # This runs on a self-hosted KVM runner (label `kvm`), NOT ubuntu-latest,
+  # because the Firecracker backend's subprocess/VM orchestration
+  # (launch/boot/SSH/isolation-probe) is covered by the integration suite,
+  # and that suite needs `/dev/kvm` + the provisioned TAP/nft pool — which a
+  # container-based runner doesn't have. On such a runner the firecracker
+  # integration test skips and its ~230 orchestration lines read as
+  # uncovered, so the gate can't pass there.
+  #
+  # Restricted to the same events as integration-firecracker (same-repo PRs,
+  # push, workflow_dispatch) for the same security reason.
+  #
+  # See #414 for the planned follow-up: artifact-based coverage combination
+  # (run tests once in their respective jobs, combine .coverage files here).
+  #
+  # build-infra creates one candidate from the checkout. This job boots that
+  # same candidate after integration-firecracker has exercised it; the main
+  # push path publishes the identical bytes only after every required job.
  coverage:
-    runs-on: ubuntu-latest
+    needs: [build-infra, integration-firecracker]
+    timeout-minutes: 15
+    runs-on: [self-hosted, kvm]
+    if: >-
+      github.event_name == 'push' ||
+      github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'pull_request' &&
+       github.event.pull_request.head.repo.full_name == github.repository)
    steps:
      - name: Checkout
        uses: actions/checkout@v4
        with:
          fetch-depth: 0

-      - name: Set up Python
-        uses: actions/setup-python@v5
+      - name: Preflight — Firecracker host is ready
+        run: |
+          command -v firecracker >/dev/null || {
+            echo "firecracker not on PATH — provision the runner (README: Firecracker on Linux)"; exit 1; }
+          test -e /dev/kvm || { echo "/dev/kvm missing — KVM not available on this runner"; exit 1; }
+          # `backend status` exits non-zero unless the TAP pool is up + no
+          # range overlap; it prints the exact `backend setup` fix.
+          python3 cli.py backend status --backend=firecracker
+
+      - name: Download the candidate already exercised by integration
+        uses: actions/download-artifact@v3
        with:
-          python-version: "3.12"
+          name: infra-candidate
+          path: infra-candidate

-      - name: Install dev requirements
-        run: python3 -m pip install -r requirements-dev.txt
-
-      - name: Combined coverage report (unit + integration)
+      # No dev-requirements install: `coverage` is already provided by the
+      # self-hosted runner's Nix python env, and that env has no `pip`
+      # module to install into anyway. `scripts/coverage.sh` +
+      # `diff_coverage.py` need only `coverage` (not pylint/pyright).
+      - name: Combined coverage (unit + integration, incl. firecracker)
+        env:
+          BOT_BOTTLE_CI_INFRA_ARTIFACT_DIR: ${{ github.workspace }}/infra-candidate
        run: PYTHON=python3 bash scripts/coverage.sh critical
+
+      - name: Diff-coverage gate (changed lines >= 90%)
+        run: |
+          git fetch --no-tags origin main:refs/remotes/origin/main
+          python3 scripts/diff_coverage.py --base origin/main --min 90
+
+  publish-infra:
+    needs: [stage-firecracker-inputs, build-infra, unit, integration-docker, integration-firecracker, coverage]
+    runs-on: ubuntu-latest
+    if: github.event_name == 'push' && github.ref == 'refs/heads/main'
+    steps:
+      - name: Checkout the tested revision
+        uses: actions/checkout@v4
+
+      - name: Download the tested candidate
+        uses: actions/download-artifact@v3
+        with:
+          name: infra-candidate
+          path: infra-candidate
+
+      # publish_infra re-derives the version from the checkout to confirm the
+      # bundle matches before uploading, and the version hashes the dropbear
+      # bytes. Stage the SAME dropbear build-infra used, or the recheck
+      # computes a "<missing>"-dropbear version and rejects the candidate.
+      - name: Download the staged dropbear (matches build-infra's version)
+        uses: actions/download-artifact@v3
+        with:
+          name: firecracker-inputs
+          path: firecracker-inputs
+
+      - name: Publish the tested candidate
+        env:
+          BOT_BOTTLE_INFRA_ARTIFACT_TOKEN: ${{ secrets.BOT_BOTTLE_INFRA_ARTIFACT_TOKEN }}
+          BOT_BOTTLE_FC_DROPBEAR: ${{ github.workspace }}/firecracker-inputs/dropbear
+        run: python3 -m bot_bottle.backend.firecracker.publish_infra --publish-dir infra-candidate
@@ -0,0 +1,17 @@
+name: tracker-policy-issues
+
+on:
+  issues:
+    types: [opened, unlabeled]
+
+jobs:
+  label-issue:
+    runs-on: ubuntu-latest
+    permissions:
+      issues: write
+    steps:
+      - uses: actions/checkout@v4
+      - name: Ensure the issue has a label
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: python3 scripts/tracker_policy.py label-issue
@@ -0,0 +1,18 @@
+name: tracker-policy-pr
+
+on:
+  pull_request:
+    types: [opened, edited, reopened, synchronize, labeled, unlabeled]
+
+jobs:
+  check-pr:
+    runs-on: ubuntu-latest
+    permissions:
+      issues: read
+      pull-requests: read
+    steps:
+      - uses: actions/checkout@v4
+      - name: Require an unlabeled PR linked to an issue
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: python3 scripts/tracker_policy.py check-pr
@@ -20,21 +20,18 @@ jobs:
          fetch-depth: 0
          token: ${{ secrets.GITHUB_TOKEN }}

-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.12'
-
+      # No actions/setup-python: the runner image ships Python 3.12 and older
+      # act_runner engines mishandle setup-python's PATH. Install into the
+      # ephemeral job container's system Python (--break-system-packages is
+      # safe because the container is disposable).
      - name: Install dev dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-dev.txt
+        run: python3 -m pip install --break-system-packages -r requirements-dev.txt

      - name: Run coverage and extract percentage
        id: coverage
        run: |
-          python -m coverage run -m unittest discover -t . -s tests/unit > /dev/null 2>&1 || true
-          PERCENT=$(python -m coverage report 2>/dev/null | grep '^TOTAL' | grep -oP '\d+(?=%)' | tail -1)
+          python3 -m coverage run -m unittest discover -t . -s tests/unit > /dev/null 2>&1 || true
+          PERCENT=$(python3 -m coverage report 2>/dev/null | grep '^TOTAL' | grep -oP '\d+(?=%)' | tail -1)
          echo "percent=$PERCENT" >> $GITHUB_OUTPUT
          echo "Coverage: $PERCENT%"

@@ -45,7 +42,7 @@ jobs:
          # the single source of truth in scripts/critical-modules.txt; every
          # core module is unit-tested, so the unit-only run is accurate for it.
          INCLUDE=$(grep -vE '^[[:space:]]*(#|$)' scripts/critical-modules.txt | paste -sd, -)
-          PERCENT=$(python -m coverage report --include="$INCLUDE" 2>/dev/null | grep '^TOTAL' | grep -oP '\d+(?=%)' | tail -1)
+          PERCENT=$(python3 -m coverage report --include="$INCLUDE" 2>/dev/null | grep '^TOTAL' | grep -oP '\d+(?=%)' | tail -1)
          echo "percent=$PERCENT" >> $GITHUB_OUTPUT
          echo "Core coverage: $PERCENT%"

@@ -16,10 +16,11 @@
 # Layout:
 #
 #   /usr/bin/gitleaks                    gitleaks binary
-#   /app/egress_addon.py + siblings      mitmproxy addon (egress)
+#   /app/egress_addon.py                 mitmproxy addon entry point
 #   /app/egress-entrypoint.sh            mitmdump launcher
-#   /app/supervise_server.py + .py       supervise MCP server
-#   /app/gateway_init.py                 PID 1 supervisor
+#   /usr/local/lib/python*/bot_bottle/   installed package (all daemons + shared modules)
+#   /app/egress_addon.py                 one-line shim: re-exports addons from package
+#                                        (mitmdump -s requires a file path, not a module)
 #   /etc/egress/routes.yaml              bind-mounted at run time
 #   /etc/git-gate/pre-receive            docker-cp'd at start time
 #   /git-gate-entrypoint.sh              docker-cp'd at start time
@@ -34,57 +35,73 @@
 #   9420  git-gate smart HTTP (VM-backend agent-facing transport)
 #   9100  supervise (MCP HTTP)

-# Stage 1: gitleaks binary. The upstream gitleaks image is alpine
-# with the binary at /usr/bin/gitleaks. Pinned by digest in lockstep
-# with Dockerfile.git-gate's prior base (now deleted at chunk 3).
-FROM zricethezav/gitleaks@sha256:c00b6bd0aeb3071cbcb79009cb16a60dd9e0a7c60e2be9ab65d25e6bc8abbb7f AS gitleaks-src
-
-# Stage 2: assembly. mitmproxy/mitmproxy is debian-slim-based with
-# Python + mitmdump pre-installed — heavier than the others, so
-# this stage starts there and pulls the standalone binaries in.
-FROM mitmproxy/mitmproxy:11.1.3
-
-# Run as root inside the bundle. The bundle is the isolation
-# boundary; per-daemon user separation inside it is not load-bearing
-# and complicates the supervisor's spawn path.
-USER root
+# Based on `python:3.12-slim` (Debian trixie) rather than the
+# `mitmproxy/mitmproxy` image (Debian bookworm) so the whole stack —
+# gateway here, and the firecracker infra image that builds FROM this —
+# lands on trixie, whose buildah (1.39) can build agent Dockerfiles that
+# use heredocs. mitmproxy is pip-installed to the same effect as the
+# upstream image. (bookworm's buildah is 1.28, which can't parse
+# `RUN ... <<EOF`; see the infra image + PR discussion.)
+FROM python:3.12-slim

 # Runtime system deps:
 #   git supplies the `git daemon` subcommand (no separate package)
 #     plus the core `git` binary the pre-receive hook invokes.
 #   openssh-client supplies the upstream SSH transport the
 #     pre-receive hook uses to forward accepted refs.
-#   ca-certificates is needed for mitmdump upstream TLS (the
-#     base image already has it; listed for explicitness).
+#   ca-certificates is needed for mitmdump upstream TLS.
 RUN apt-get update \
 && apt-get install -y --no-install-recommends \
      git openssh-client ca-certificates \
 && rm -rf /var/lib/apt/lists/*

-# Pull the standalone binaries into the final image.
-COPY --from=gitleaks-src /usr/bin/gitleaks /usr/bin/gitleaks
+# mitmdump (the egress data plane). The upstream mitmproxy image baked
+# this in; on the plain python base we pip-install the same pinned
+# version. Its CA dir is set explicitly via `--set confdir=` in
+# egress-entrypoint.sh, so it doesn't depend on a `mitmproxy` home user.
+RUN pip install --no-cache-dir mitmproxy==11.1.3

-# Project Python: addon + server modules + the init supervisor.
-# Kept flat under /app/ so mitmdump's loader resolves them as
-# top-level siblings (absolute imports), matching the prior
-# Dockerfile.egress / Dockerfile.supervise layout.
-COPY bot_bottle/egress_addon_core.py /app/egress_addon_core.py
-COPY bot_bottle/egress_dlp_config.py /app/egress_dlp_config.py
-COPY bot_bottle/egress_addon.py      /app/egress_addon.py
-COPY bot_bottle/policy_resolver.py   /app/policy_resolver.py
-COPY bot_bottle/dlp_detectors.py     /app/dlp_detectors.py
-COPY bot_bottle/yaml_subset.py       /app/yaml_subset.py
-COPY bot_bottle/paths.py             /app/paths.py
-COPY bot_bottle/migrations.py        /app/migrations.py
-COPY bot_bottle/db_store.py          /app/db_store.py
-COPY bot_bottle/supervise_types.py   /app/supervise_types.py
-COPY bot_bottle/queue_store.py       /app/queue_store.py
-COPY bot_bottle/audit_store.py       /app/audit_store.py
-COPY bot_bottle/store_manager.py     /app/store_manager.py
-COPY bot_bottle/supervise.py         /app/supervise.py
-COPY bot_bottle/supervise_server.py  /app/supervise_server.py
-COPY bot_bottle/gateway_init.py      /app/gateway_init.py
-COPY bot_bottle/git_http_backend.py  /app/git_http_backend.py
+# gitleaks (the pre-receive hook's secret scanner). Installed from its
+# official release, pinned by version + SHA256 and verified — rather than
+# using a third-party image as a build stage (supply-chain surface, and it
+# would pin us to that image's cadence). python (already present) does the
+# download so we add no curl/wget. trixie apt also ships gitleaks, but an
+# older 8.16; the pinned download keeps the verified 8.30.1.
+#
+# Arch-aware: the asset + SHA are picked from the build's target
+# architecture so an arm64 host (Apple Silicon) gets the arm64 binary
+# rather than an x86_64 one that dies with "Exec format error" the first
+# time the pre-receive hook runs it. TARGETARCH is auto-populated by
+# BuildKit; the dpkg fallback keeps it correct under a legacy builder.
+ARG GITLEAKS_VERSION=8.30.1
+ARG GITLEAKS_SHA256_AMD64=551f6fc83ea457d62a0d98237cbad105af8d557003051f41f3e7ca7b3f2470eb
+ARG GITLEAKS_SHA256_ARM64=e4a487ee7ccd7d3a7f7ec08657610aa3606637dab924210b3aee62570fb4b080
+ARG TARGETARCH
+RUN arch="${TARGETARCH:-$(dpkg --print-architecture)}" \
+ && case "$arch" in \
+      amd64) asset="linux_x64";   sha="${GITLEAKS_SHA256_AMD64}" ;; \
+      arm64) asset="linux_arm64"; sha="${GITLEAKS_SHA256_ARM64}" ;; \
+      *) echo "unsupported gitleaks target arch: $arch" >&2; exit 1 ;; \
+    esac \
+ && url="https://github.com/gitleaks/gitleaks/releases/download/v${GITLEAKS_VERSION}/gitleaks_${GITLEAKS_VERSION}_${asset}.tar.gz" \
+ && python3 -c "import sys,urllib.request; urllib.request.urlretrieve(sys.argv[1], '/tmp/gitleaks.tar.gz')" "$url" \
+ && echo "${sha}  /tmp/gitleaks.tar.gz" | sha256sum -c - \
+ && tar -xzf /tmp/gitleaks.tar.gz -C /usr/bin gitleaks \
+ && rm /tmp/gitleaks.tar.gz
+
+# Install bot_bottle as a proper package so entry-point scripts can use
+# `from bot_bottle.X import Y` absolute imports. A rename or a missing
+# module is caught at pip-install time — not at container runtime.
+COPY pyproject.toml /src/
+COPY bot_bottle/    /src/bot_bottle/
+RUN pip install --no-cache-dir /src/
+
+# mitmdump -s requires a file path, not a module. Write a one-line shim that
+# re-exports `addons` from the installed package; mitmdump finds it there.
+# WORKDIR here also creates /app so the shim + COPYs below can write into it
+# (nothing created /app before this point).
+WORKDIR /app
+RUN printf 'from bot_bottle.egress_addon import addons\n' > /app/egress_addon.py
 COPY bot_bottle/egress_entrypoint.sh /app/egress-entrypoint.sh
 RUN chmod +x /app/egress-entrypoint.sh

@@ -103,10 +120,6 @@ RUN mkdir -p \
 # subset the bottle uses.
 EXPOSE 8888 9099 9418 9420 9100

-# WORKDIR matches Dockerfile.supervise's prior layout so the
-# in-app same-dir import in supervise_server.py stays deterministic.
-WORKDIR /app
-
 # PID 1 is the supervisor. It owns signal handling and exit-code
 # propagation; no `exec` chain in the entrypoint itself.
-ENTRYPOINT ["python3", "/app/gateway_init.py"]
+ENTRYPOINT ["python3", "-m", "bot_bottle.gateway_init"]
@@ -0,0 +1,22 @@
+# Shared infra image: gateway data plane + orchestrator control plane.
+#
+# Used directly by the Docker backend (run as one `bot-bottle-infra`
+# container, replacing the prior two-container split). The Firecracker
+# backend extends this via Dockerfile.infra.fc, adding buildah/crun/
+# netavark for in-VM agent-image building.
+#
+# Dockerfile.orchestrator is the single definition of the orchestrator
+# content (the lean `bot_bottle` package on python:3.12-slim). Both this
+# image and Dockerfile.infra.fc pull it in via `COPY --from`.
+#
+# multi-`FROM` can't union two bases (that's multi-stage, not multiple
+# inheritance), so the orchestrator content is pulled in via `COPY --from`
+# rather than a second base. Both images share the trixie `python:3.12-slim`
+# base, so the copy is clean (same python; future installed deps copy too).
+FROM bot-bottle-gateway:latest
+
+# The orchestrator content, from its single definition. The gateway image
+# already has the flat daemon modules under /app; this adds the full
+# `bot_bottle` package so `python3 -m bot_bottle.orchestrator` resolves —
+# used by gateway_init when BOT_BOTTLE_GATEWAY_DAEMONS includes `orchestrator`.
+COPY --from=bot-bottle-orchestrator:latest /app/bot_bottle /app/bot_bottle
@@ -0,0 +1,23 @@
+# Firecracker infra VM image (PRD 0070 Stage B).
+#
+# Extends the shared infra base (Dockerfile.infra: gateway + orchestrator
+# control plane) with the in-VM agent-image builder. The Firecracker backend
+# builds users' agent Dockerfiles *inside this VM* with buildah (rootless,
+# daemonless) instead of on the host — no host Docker daemon, no
+# root-equivalent `docker` group.
+#
+# Requires the trixie base from bot-bottle-gateway (buildah 1.39: bookworm's
+# 1.28 can't parse Dockerfile heredocs that agent images use).
+#
+# `crun` is the OCI runtime; `netavark` + `aardvark-dns` are the network
+# backend for `FROM` pulls + `RUN` egress. `vfs` + `chroot`: buildah works
+# as root in the bare microVM (no fuse-overlayfs / overlay module / subuid
+# maps). Matches image_builder.
+FROM bot-bottle-infra:latest
+
+RUN apt-get update \
+ && apt-get install -y --no-install-recommends \
+      buildah crun netavark aardvark-dns \
+ && rm -rf /var/lib/apt/lists/*
+ENV STORAGE_DRIVER=vfs \
+    BUILDAH_ISOLATION=chroot
@@ -1,27 +1,36 @@
 # Orchestrator control-plane image (PRD 0070, #384).
 #
-# The per-host orchestrator runs `python3 -m bot_bottle.orchestrator`.
-# The `bot_bottle` package is **stdlib-only** by design, so the control
-# plane needs nothing but a Python runtime — none of the gateway's
-# mitmproxy / git / gitleaks payload (that is the separate
-# `bot-bottle-gateway` image, Dockerfile.gateway). Splitting them keeps
-# the secret-dense control plane (it concentrates every bottle's egress
-# tokens — see PRD 0070's "secret concentration") on a minimal
-# dependency surface.
+# This is the **single definition of the orchestrator's content** — the
+# `bot_bottle` package baked onto a Python runtime — referenced by BOTH:
+#   * the docker backend, which runs this image directly as the lean
+#     control-plane container; and
+#   * the firecracker infra image (Dockerfile.infra), which `COPY --from`s
+#     this image's `/app/bot_bottle` so the single infra VM runs the same
+#     control plane. Keeping it in one place means future orchestrator deps
+#     (e.g. iroh) are added here once, not duplicated per backend.
 #
-# The repo is bind-mounted read-only into the container at run time (see
-# `orchestrator/lifecycle.py`), so the source is NOT copied in here: the
-# image is just the runtime. `ensure_running` recreates the container
-# only when the bind-mounted source hash changes (#381), which is why
-# the code stays a mount rather than a baked layer.
+# It stays deliberately lean: the control plane is **stdlib-only** today, so
+# no third-party payload — none of the gateway's mitmproxy/git/gitleaks
+# (that's Dockerfile.gateway) and no buildah (that's the firecracker
+# builder, and lives only in Dockerfile.infra). Keeping the secret-dense
+# control plane on a minimal dependency surface is the point (PRD 0070's
+# "secret concentration").
+#
+# Shares the trixie `python:3.12-slim` base with the gateway image, so when
+# the orchestrator grows real deps they can be `COPY --from`'d into the
+# infra image cleanly (same base/python — installed packages copy safely).

 FROM python:3.12-slim

-# No third-party deps to install — stdlib only. Kept as an explicit,
-# self-documenting stage so a future confinement step (baking the
-# package, dropping the bind mount) has an obvious home.
 WORKDIR /app

+# The orchestrator content. Baked so the image is self-contained (runs from
+# a built image, no runtime bind-mount); the docker backend may still
+# bind-mount /app for dev live-reload, which simply overlays this copy.
+# `.dockerignore` keeps .git/docs/*.md out of the context. (Future deps like
+# iroh go here too — a shared requirements installed on this same base.)
+COPY bot_bottle /app/bot_bottle
+
 # Documentation only; lifecycle.py overrides the entrypoint to
 # `python3 -m bot_bottle.orchestrator` with the runtime flags.
 ENTRYPOINT ["python3", "-m", "bot_bottle.orchestrator"]
@@ -5,8 +5,8 @@
 # bot-bottle

 [![test](https://gitea.dideric.is/didericis/bot-bottle/actions/workflows/test.yml/badge.svg?branch=main)](https://gitea.dideric.is/didericis/bot-bottle/actions?workflow=test.yml)
-[![coverage](https://img.shields.io/badge/coverage-84%25-brightgreen)](https://coverage.readthedocs.io/)
-[![core coverage](https://img.shields.io/badge/core%20coverage-95%25-brightgreen)](https://gitea.dideric.is/didericis/bot-bottle/src/branch/main/docs/decisions/0004-coverage-policy.md)
+[![coverage](https://img.shields.io/badge/coverage-81%25-brightgreen)](https://coverage.readthedocs.io/)
+[![core coverage](https://img.shields.io/badge/core%20coverage-94%25-brightgreen)](https://gitea.dideric.is/didericis/bot-bottle/src/branch/main/docs/decisions/0004-coverage-policy.md)

 **Problem:** Developer wants to run a coding agent without supervision, but they don't want a prompt injected or misbehaving agent wrecking their environment or exfiltrating sensitive data.

@@ -90,6 +90,8 @@ BOT_BOTTLE_BACKEND=firecracker ./cli.py start <agent>

 > **NixOS:** enable `virtualisation.docker`, ensure the KVM module is loaded (`boot.kernelModules = [ "kvm-intel" ];` or `kvm-amd`), and add your user to the `kvm` and `docker` groups. For the network pool, consume the flake module — `imports = [ inputs.bot-bottle.nixosModules.firecracker-netpool ]; services.bot-bottle-firecracker = { enable = true; owner = "you"; };` — then `nixos-rebuild switch` (imperative nft/TAP rules don't survive a rebuild; channel users can `imports = [ <bot-bottle>/nix/firecracker-netpool.nix ]`). `firecracker` isn't in nixpkgs by default as a user binary — install the release binary (pin the version) and put it on `PATH`.

+> **CI:** the coverage gate (`.gitea/workflows/test.yml` → `coverage` job) runs on a self-hosted runner labelled `kvm`, because the Firecracker backend's VM/SSH orchestration is exercised only by the integration suite, which needs `/dev/kvm` + the provisioned pool (a container runner would skip it and read as uncovered). Provision that runner exactly like a normal Firecracker host — `firecracker` on `PATH`, `/dev/kvm`, the cached guest kernel + static dropbear, and the pool installed as the persistent systemd unit — then register it with the `kvm` label. A Docker-capable hosted job builds the candidate once; KVM tests boot those exact bytes, and a successful main run publishes them. The unit/lint jobs still run on `ubuntu-latest`.
+
 ```sh
 ./cli.py start <agent>   # builds the image on first run, drops you into claude
 ```
@@ -171,6 +173,15 @@ When an outbound DLP detector matches a token, the route's `dlp.outbound_on_matc

 More examples in `examples/`. Full design lives under `docs/prds/`; the trust-boundary rationale is in `docs/prds/0011-per-file-md-manifest.md`.

+## Tracker policy
+
+Issues are the canonical work items and own all tracker labels; every issue
+must have at least one. Pull requests stay unlabeled and deliberately reference
+an issue with `Closes #…`, `Part of #…`, or another form defined in
+[`ADR 0005`](docs/decisions/0005-issues-own-tracker-metadata.md). Gitea Actions
+enforces the convention for new work from 2026-07-18 onward. Earlier closed
+PRs are grandfathered rather than given artificial retrospective issues.
+
 ## Trademarks

 bot-bottle is an independent project and is not affiliated with, endorsed by, or sponsored by Anthropic, PBC. "Claude" and "Claude Code" are trademarks of Anthropic, PBC; the project name uses "claude" descriptively to indicate that the tool runs Claude Code inside a sandbox.
@@ -61,6 +61,13 @@ class AgentProviderRuntime:
    prompt_mode: PromptMode
    bypass_args: tuple[str, ...]
    resume_args: tuple[str, ...]
+    # argv run inside a throwaway container of a freshly built agent
+    # image, right after `build_image()`, to catch a build that
+    # exited 0 but produced a broken CLI (e.g. an npm
+    # optionalDependencies fetch for a platform-native binary that
+    # silently no-ops on a transient failure). Empty tuple skips the
+    # check — not every provider has opted in yet.
+    smoke_test: tuple[str, ...] = ()


@dataclass(frozen=True)
@@ -259,6 +266,7 @@ class AgentProvider(ABC):
            gate_scheme = getattr(plan, "git_gate_insteadof_scheme", "git")
            content = git_gate_render_gitconfig(
                manifest_bottle.git, gate_host, scheme=gate_scheme,
+                identity_token=getattr(plan, "identity_token", ""),
            )
            guest_gitconfig = f"{plan.guest_home}/.gitconfig"
            with tempfile.NamedTemporaryFile(
@@ -42,7 +42,7 @@ class AuditStore(DbStore):
        super().__init__(db_path or host_db_path(), migrations)

    def write_audit_entry(self, entry: AuditEntry) -> Path:
-        with self._connect() as conn:
+        with self._connection() as conn:
            conn.execute(
                """
                INSERT INTO supervise_audit_entries (
@@ -66,7 +66,7 @@ class AuditStore(DbStore):
    def read_audit_entries(self, component: str, slug: str) -> list[AuditEntry]:
        if not self.db_path.is_file():
            return []
-        with self._connect() as conn:
+        with self._connection() as conn:
            rows = conn.execute(
                """
                SELECT * FROM supervise_audit_entries
@@ -37,15 +37,16 @@ import os
 import shlex
 import sys
 from abc import ABC, abstractmethod
-from contextlib import AbstractContextManager
+from contextlib import AbstractContextManager, contextmanager
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Any, Generic, Sequence, TypeVar
+from typing import TYPE_CHECKING, Any, Generator, Generic, Sequence, TypeVar

 from ..agent_provider import AgentProvisionPlan, get_provider, build_agent_provision_plan
 from ..egress import EgressPlan
 from ..git_gate import GitGatePlan
-from ..log import die, info
+from ..log import die, info, warn
+from ..util import read_tty_line
 from ..manifest import Manifest, ManifestIndex
 from ..supervise import SupervisePlan
 from ..util import expand_tilde
@@ -54,6 +55,9 @@ from ..workspace import WorkspacePlan, workspace_plan
 from .print_util import print_multi, visible_agent_env_names
 from .util import host_skill_dir

+if TYPE_CHECKING:
+    from .freeze import CommitCancelled, Freezer, get_freezer
+

@dataclass(frozen=True)
 class BottleSpec:
@@ -79,6 +83,9 @@ class BottleSpec:
    # True when launched via --headless (no TTY, no interactive prompts).
    # The git-gate host-key preflight uses this to error rather than prompt.
    headless: bool = False
+    # Image startup policy. "fresh" preserves the normal build path;
+    # "cached" reuses the current local image/artifact without rebuilding.
+    image_policy: str = "fresh"


@dataclass(frozen=True)
@@ -274,6 +281,18 @@ PlanT = TypeVar("PlanT", bound=BottlePlan)
 CleanupT = TypeVar("CleanupT", bound=BottleCleanupPlan)


+@dataclass(frozen=True)
+class BottleImages:
+    """Resolved image references (or artifact paths) for a bottle launch.
+
+    For Docker/macOS-container backends, `agent` and `sidecar` are string
+    image refs. For the smolmachines backend they are Path objects pointing
+    to pre-built `.smolmachine` artifacts."""
+
+    agent: str | Path
+    sidecar: str | Path = ""
+
+
 class BottleBackend(ABC, Generic[PlanT, CleanupT]):
    """Abstract base for selectable bottle backends. Concrete subclasses
    (e.g. DockerBottleBackend) own their own prepare/launch impls.
@@ -433,9 +452,27 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
        prompt file, Dockerfile path, and guest home all live on
        `agent_provision_plan` — the source of truth."""

+    def prelaunch_checks(self, plan: PlanT) -> None:
+        """Raise StaleImageError if any cached image used by this plan is stale.
+        No-op default; backends override to call the shared check_stale*
+        helpers on their image/artifact timestamps. Called by the CLI before
+        launch so the operator can be prompted outside the launch context."""
+
+    @contextmanager
+    def launch(self, plan: PlanT) -> Generator[Bottle, None, None]:
+        """Template: build or load images, then delegate to _launch_impl."""
+        images = self._build_or_load_images(plan)
+        with self._launch_impl(plan, images) as bottle:
+            yield bottle
+
    @abstractmethod
-    def launch(self, plan: PlanT) -> AbstractContextManager[Bottle]:
-        """Build/run the bottle and yield a handle; tear down on exit."""
+    def _build_or_load_images(self, plan: PlanT) -> BottleImages:
+        """Return the agent and sidecar image references (or artifact paths)
+        for this plan, building fresh images when the policy requires it."""
+
+    @abstractmethod
+    def _launch_impl(self, plan: PlanT, images: BottleImages) -> AbstractContextManager[Bottle]:
+        """Bring up the bottle using pre-resolved images; yield a handle; tear down on exit."""

    def provision(self, plan: PlanT, bottle: "Bottle") -> str | None:
        """Copy host-side files (CA cert, prompt, skills, .git) into
@@ -511,6 +548,18 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
        del plan
        return ""

+    def ensure_orchestrator(self) -> str:
+        """Bring up this backend's per-host orchestrator + shared gateway
+        (idempotent) and return the host-reachable control-plane URL.
+
+        This is the backend-agnostic bring-up entry point: `launch` calls
+        it as part of starting a bottle, and operator tools (`supervise`)
+        call it to start the control plane on demand when none is running
+        yet. Docker starts the orchestrator + gateway containers;
+        firecracker boots the infra VM. Backends with no orchestrator
+        (macos-container) die with a pointer — the default here."""
+        die(f"backend {self.name!r} has no orchestrator control plane")
+
    @abstractmethod
    def prepare_cleanup(self) -> CleanupT:
        """Enumerate orphaned resources from previous bottles. No side
@@ -572,68 +621,170 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
        Not called by the launch path or the test suite."""


-# Import concrete backend classes AFTER the base types are defined, so
-# each backend module can pull BottleSpec / BottlePlan / BottleBackend
-# via `from . import ...` without hitting a partially-initialized module.
-from .docker import DockerBottleBackend  # noqa: E402  # pylint: disable=wrong-import-position
-from .firecracker import FirecrackerBottleBackend  # noqa: E402  # pylint: disable=wrong-import-position
-from .macos_container import MacosContainerBottleBackend  # noqa: E402  # pylint: disable=wrong-import-position
-
-# Freezer is imported after the backend classes for the same reason:
-# Freezer.commit_slug constructs ActiveAgent, which must be fully
-# defined first.
-from .freeze import CommitCancelled, Freezer, get_freezer  # noqa: E402  # pylint: disable=wrong-import-position
+# _backends is None until the first call to _get_backends(), at which
+# point all three concrete backend classes are imported and instantiated.
+# Keeping the imports out of module scope means that importing any
+# backend sub-module (e.g. `backend.docker.util`) no longer drags the
+# firecracker and macos-container implementations into memory.
+#
+# Tests may replace _backends with a {name: fake} dict via patch.object;
+# _get_backends() returns the current module-level value as-is when it
+# is not None, so test fakes take effect without triggering real imports.
+_backends: dict[str, BottleBackend[Any, Any]] | None = None


-# The dict is heterogeneous: each value is a BottleBackend specialized
-# over its own plan type. Concrete plan types are erased here because
-# the registry is selected at runtime and the CLI only needs the
-# unparameterized methods (prepare → plan → launch(plan), cleanup, etc.).
-_BACKENDS: dict[str, BottleBackend[Any, Any]] = {
-    "docker": DockerBottleBackend(),
-    "firecracker": FirecrackerBottleBackend(),
-    "macos-container": MacosContainerBottleBackend(),
-}
+def _get_backends() -> dict[str, BottleBackend[Any, Any]]:
+    """Return the registry of all backend instances, loading lazily on first call."""
+    global _backends  # pylint: disable=global-statement
+    if _backends is None:
+        from .docker import DockerBottleBackend
+        from .firecracker import FirecrackerBottleBackend
+        from .macos_container import MacosContainerBottleBackend
+        _backends = {
+            "docker": DockerBottleBackend(),
+            "firecracker": FirecrackerBottleBackend(),
+            "macos-container": MacosContainerBottleBackend(),
+        }
+    return _backends
+
+
+def __getattr__(name: str) -> Any:
+    """Lazily surface concrete backend classes and freeze symbols at the
+    package level so existing `from bot_bottle.backend import X` and
+    `patch.object(backend_mod, X, ...)` call-sites keep working without
+    forcing an import of every backend at module-init time."""
+    if name == "DockerBottleBackend":
+        from .docker import DockerBottleBackend
+        globals()[name] = DockerBottleBackend
+        return DockerBottleBackend
+    if name == "FirecrackerBottleBackend":
+        from .firecracker import FirecrackerBottleBackend
+        globals()[name] = FirecrackerBottleBackend
+        return FirecrackerBottleBackend
+    if name == "MacosContainerBottleBackend":
+        from .macos_container import MacosContainerBottleBackend
+        globals()[name] = MacosContainerBottleBackend
+        return MacosContainerBottleBackend
+    if name == "CommitCancelled":
+        from .freeze import CommitCancelled
+        globals()[name] = CommitCancelled
+        return CommitCancelled
+    if name == "Freezer":
+        from .freeze import Freezer
+        globals()[name] = Freezer
+        return Freezer
+    if name == "get_freezer":
+        from .freeze import get_freezer
+        globals()[name] = get_freezer
+        return get_freezer
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


 def get_bottle_backend(
    name: str | None = None,
+    *,
+    prompt: bool = True,
 ) -> BottleBackend[Any, Any]:
    """Resolve the bottle backend.

    `name` precedence:
-      1. explicit arg (CLI `--backend=<name>` passes through here)
+      1. explicit arg (e.g. resume passes the recorded backend name)
      2. BOT_BOTTLE_BACKEND env var
-      3. `macos-container` on compatible macOS hosts
-      4. `firecracker` on KVM-capable Linux hosts
-      5. default `docker`
+      3. auto-selection: VM backend first, docker fallback with prompt
+
+    `prompt` controls whether auto-selection may block on an interactive
+    [i/d/q] prompt when falling back to docker. Pass `prompt=False` in
+    non-interactive contexts (headless launches, CI) so the call dies
+    with an actionable message instead of hanging.

    Dies with a pointer at the known backends if the chosen name
    isn't implemented."""
-    resolved = name or os.environ.get("BOT_BOTTLE_BACKEND") or _default_backend_name()
-    if resolved not in _BACKENDS:
-        known = ", ".join(sorted(_BACKENDS))
+    resolved = name or os.environ.get("BOT_BOTTLE_BACKEND")
+    if resolved is None:
+        resolved = _auto_select_backend(prompt=prompt)
+    backends = _get_backends()
+    if resolved not in backends:
+        known = ", ".join(sorted(backends))
        die(f"unknown backend {resolved!r}; known backends: {known}")
-    return _BACKENDS[resolved]
+    return backends[resolved]


-def _default_backend_name() -> str:
+def _platform_vm_suggestion() -> str:
+    """Platform-appropriate VM backend name for install suggestions."""
+    return "macos-container" if sys.platform == "darwin" else "firecracker"
+
+
+def _print_vm_install_instructions() -> None:
+    """Print platform-appropriate VM backend install instructions to stderr."""
+    vm = _platform_vm_suggestion()
+    if vm == "macos-container":
+        info("Install Apple Container: https://github.com/apple/container/releases")
+        info("Then start the service:  container system start")
+    else:
+        info("Install Firecracker: https://github.com/firecracker-microvm/firecracker/releases")
+        info("Configure the host:  ./cli.py backend setup")
+
+
+def _auto_select_backend(prompt: bool = True) -> str:
+    """Tier-1 / tier-2 backend auto-selection.
+
+    Tier 1: VM backend — macos-container on macOS when Apple Container is
+    installed; firecracker on KVM-capable Linux even before the binary is
+    present (its preflight prints an install pointer).
+
+    Tier 2: docker, with a security warning and an interactive prompt.
+    When `prompt=False` (headless / CI), dies with an actionable message
+    instead of blocking on a TTY read. When docker is also absent, prints
+    VM install instructions and exits.
+    """
+    # --- Tier 1: VM backend -----------------------------------------
    if has_backend("macos-container"):
        return "macos-container"
    # A KVM-capable Linux host defaults to firecracker even when the
    # `firecracker` binary isn't installed yet: selecting it here routes
    # start through firecracker's preflight, which prints an install
    # pointer, instead of silently falling back to docker.
+    from .firecracker import FirecrackerBottleBackend
    if FirecrackerBottleBackend.is_host_capable():
        return "firecracker"
-    return "docker"
+
+    # --- Tier 2: docker fallback ------------------------------------
+    if not has_backend("docker"):
+        info("No backend available on this host.")
+        _print_vm_install_instructions()
+        die("no backend available; install a VM backend and re-run")
+
+    vm = _platform_vm_suggestion()
+    warn(
+        "docker is less secure than VM backends — "
+        "containers share the host kernel."
+    )
+    if not prompt:
+        die(
+            f"no VM backend available; set BOT_BOTTLE_BACKEND=docker to proceed "
+            f"with docker, or install the {vm!r} backend."
+        )
+    sys.stderr.write(
+        f"bot-bottle: For better isolation, install the {vm!r} backend.\n"
+        f"  [i] show {vm} install instructions and exit\n"
+        "  [d] use docker anyway\n"
+        "  [q] quit\n"
+        "bot-bottle: choice [i/d/q]: "
+    )
+    sys.stderr.flush()
+    reply = read_tty_line().strip().lower()
+    if reply == "d":
+        return "docker"
+    if reply == "i":
+        _print_vm_install_instructions()
+    die("not proceeding with docker; install a VM backend or set BOT_BOTTLE_BACKEND=docker")


 def known_backend_names() -> tuple[str, ...]:
-    """Sorted tuple of all backend keys in `_BACKENDS`. Used by
+    """Sorted tuple of all backend keys in `_get_backends()`. Used by
    argparse (`--backend` choices) and the dashboard's backend
    picker."""
-    return tuple(sorted(_BACKENDS))
+    return tuple(sorted(_get_backends()))


 def has_backend(name: str) -> bool:
@@ -645,9 +796,10 @@ def has_backend(name: str) -> bool:

    Returns False for unknown names so callers can pass
    arbitrary input without separate validation."""
-    if name not in _BACKENDS:
+    backends = _get_backends()
+    if name not in backends:
        return False
-    return _BACKENDS[name].is_available()
+    return backends[name].is_available()


 def enumerate_active_agents() -> list[ActiveAgent]:
@@ -663,10 +815,11 @@ def enumerate_active_agents() -> list[ActiveAgent]:
    deterministic tiebreaker. Agents with missing metadata
    (`started_at == ""`) sort first."""
    out: list[ActiveAgent] = []
-    for name in known_backend_names():
-        if not has_backend(name):
+    backends = _get_backends()
+    for name in sorted(backends):
+        if not backends[name].is_available():
            continue
-        out.extend(_BACKENDS[name].enumerate_active())
+        out.extend(backends[name].enumerate_active())
    out.sort(key=lambda a: (a.started_at, a.slug))
    return out

@@ -0,0 +1,60 @@
+"""Shared helpers for the consolidated launch sequence (PRD 0070).
+
+Logic that was duplicated across the docker, macos_container, and
+firecracker consolidated_launch modules — extracted so each backend
+imports it rather than re-implementing it.
+"""
+
+from __future__ import annotations
+
+from ..egress import EgressPlan
+from ..git_gate import GitGatePlan
+from ..orchestrator.client import OrchestratorClient
+from ..orchestrator.registration import registration_inputs
+from .docker.gateway_provision import GatewayTransport, deprovision_git_gate, provision_git_gate
+
+
+def provision_bottle(
+    client: OrchestratorClient,
+    source_ip: str,
+    egress_plan: EgressPlan,
+    git_gate_plan: GitGatePlan,
+    transport: GatewayTransport,
+    *,
+    image_ref: str = "",
+    tokens: dict[str, str] | None = None,
+):
+    """Register the bottle and provision its git-gate state. Rolls back the
+    registration if provisioning fails so no orphan is left. Returns the
+    `RegisteredBottle` from the orchestrator."""
+    inputs = registration_inputs(egress_plan)
+    reg = client.register_bottle(
+        source_ip, image_ref=image_ref, policy=inputs.policy,
+        metadata=inputs.metadata, tokens=tokens,
+    )
+    try:
+        provision_git_gate(transport, reg.bottle_id, git_gate_plan)
+    except Exception:
+        client.teardown_bottle(reg.bottle_id)
+        raise
+    return reg
+
+
+def teardown_consolidated(
+    bottle_id: str,
+    transport: GatewayTransport,
+    *,
+    orchestrator_url: str,
+    timeout: float | None = None,
+) -> None:
+    """Deregister the bottle and remove its git-gate state. Both steps are
+    idempotent so this is safe from a cleanup trap."""
+    from ..orchestrator.config_store import DEFAULT_TEARDOWN_TIMEOUT_SECONDS
+    OrchestratorClient(
+        orchestrator_url,
+        timeout=timeout if timeout is not None else DEFAULT_TEARDOWN_TIMEOUT_SECONDS,
+    ).teardown_bottle(bottle_id)
+    deprovision_git_gate(transport, bottle_id)
+
+
+__all__ = ["provision_bottle", "teardown_consolidated"]
@@ -31,7 +31,7 @@ from ...env import ResolvedEnv
 from ...git_gate import GitGatePlan
 from ...supervise import SupervisePlan
 from ...manifest import Manifest
-from .. import ActiveAgent, BottleBackend, BottleSpec
+from .. import ActiveAgent, BottleBackend, BottleImages, BottleSpec
 from . import cleanup as _cleanup
 from . import enumerate as _enumerate
 from . import launch as _launch
@@ -100,11 +100,21 @@ class DockerBottleBackend(BottleBackend["DockerBottlePlan", "DockerBottleCleanup
            stage_dir=stage_dir,
        )

+    def prelaunch_checks(self, plan: DockerBottlePlan) -> None:
+        _launch.stale_checks(plan)
+
+    def _build_or_load_images(self, plan: DockerBottlePlan) -> BottleImages:
+        return _launch.build_or_load_images(plan)
+
    @contextmanager
-    def launch(self, plan: DockerBottlePlan) -> Generator[DockerBottle, None, None]:
-        with _launch.launch(plan, provision=self.provision) as bottle:
+    def _launch_impl(self, plan: DockerBottlePlan, images: BottleImages) -> Generator[DockerBottle, None, None]:
+        with _launch.launch(plan, images, provision=self.provision) as bottle:
            yield bottle

+    def ensure_orchestrator(self) -> str:
+        from ...orchestrator.lifecycle import OrchestratorService
+        return OrchestratorService().ensure_running()
+
    def supervise_mcp_url(self, plan: DockerBottlePlan) -> str:
        """Docker bottles reach the supervise daemon via the
        compose-network alias `supervise:9100`. No per-bottle URL
@@ -35,6 +35,10 @@ class DockerBottlePlan(BottlePlan):
    # Likewise the supervise MCP endpoint at the gateway (`http://<gw>:9100/`);
    # empty → the single-tenant `supervise` alias.
    agent_supervise_url: str = ""
+    # Per-bottle identity token the agent presents on every attributed request
+    # (egress proxy credentials, git-gate/supervise headers); set by launch
+    # from the orchestrator registration. Empty pre-registration.
+    identity_token: str = ""

    @property
    def container_name(self) -> str:
@@ -31,7 +31,13 @@ def consolidated_agent_compose(
 ) -> dict[str, Any]:
    """A compose spec with only the agent service, on the external gateway
    network at `source_ip`, proxying egress through `gateway_ip`."""
-    proxy_url = f"http://{gateway_ip}:{EGRESS_PORT}"
+    # Deliver the identity token as egress proxy credentials — the gateway
+    # reads Proxy-Authorization, validates the (source_ip, token) pair, and
+    # strips it before upstream. git-http/supervise get it via their own
+    # headers (git config extraHeader / MCP header).
+    token = getattr(plan, "identity_token", "")
+    cred = f"bottle:{token}@" if token else ""
+    proxy_url = f"http://{cred}{gateway_ip}:{EGRESS_PORT}"
    # git-http + supervise live on the gateway too and must NOT go through the
    # egress proxy — the agent reaches them directly by the gateway address.
    no_proxy = f"localhost,127.0.0.1,{gateway_ip}"
@@ -1,19 +1,13 @@
 """Consolidated bottle launch sequence for the docker backend (PRD 0070).

-Composes the orchestrator primitives into the register/teardown sequence that
-replaces the per-bottle gateway:
+Composes the orchestrator primitives into the register/teardown sequence:

-  1. ensure the orchestrator control plane + shared gateway are up;
-  2. allocate the bottle a pinned source IP on the gateway network (the
-     attribution key), skipping the gateway's own address + live bottles;
-  3. register it (egress policy blob + slug metadata) → bottle id + identity
-     token;
-  4. provision its git-gate repos/creds into the running gateway.
+  1. ensure the single infra container (control plane + gateway) is up;
+  2. allocate the bottle a pinned source IP on the gateway network;
+  3. register it and provision its git-gate repos/creds into the gateway.

-It returns a `LaunchContext` with everything the agent container needs to
-attach — network, pinned IP, the gateway's address (its proxy target), the
-orchestrator URL, and the identity token. The agent `docker run` itself is
-the backend's job (it owns provider provisioning); this owns the
+Returns a `LaunchContext` with everything the agent container needs to
+attach. The agent `docker run` itself is the backend's job; this owns the
 orchestrator-facing wiring so that sequence stays testable in isolation.
 """

@@ -25,11 +19,12 @@ from ...docker_cmd import run_docker
 from ...egress import EgressPlan
 from ...git_gate import GitGatePlan
 from ...orchestrator.client import OrchestratorClient
-from ...orchestrator.gateway import GATEWAY_NAME, GATEWAY_NETWORK
-from ...orchestrator.lifecycle import OrchestratorService
-from ...orchestrator.registration import registration_inputs
+from ...orchestrator.gateway import GATEWAY_NETWORK
+from ...orchestrator.lifecycle import INFRA_NAME, OrchestratorService
+from ..consolidated_util import provision_bottle
+from ..consolidated_util import teardown_consolidated as _teardown_util
+from .gateway_provision import DockerGatewayTransport
 from .gateway_net import next_free_ip
-from .gateway_provision import deprovision_git_gate, provision_git_gate


 class ConsolidatedLaunchError(RuntimeError):
@@ -71,24 +66,21 @@ def _container_ip(name: str, network: str) -> str:
    ip = proc.stdout.strip()
    if proc.returncode != 0 or not ip:
        raise ConsolidatedLaunchError(
-            f"gateway {name} has no address on {network}: {proc.stderr.strip()}"
+            f"container {name} has no address on {network}: {proc.stderr.strip()}"
        )
    return ip


 def _network_container_ips(network: str) -> list[str]:
    """Every address currently assigned on the gateway network — the ground
-    truth for "in use": the gateway + orchestrator infrastructure containers
-    and every live agent. Read from the network so a new bottle can't collide
-    with anything actually attached (a registry-only view would miss the
-    orchestrator/gateway containers)."""
+    truth for "in use": the infra container and every live agent. Read from
+    the network so a new bottle can't collide with anything actually attached."""
    proc = run_docker([
        "docker", "network", "inspect", "--format",
        "{{range .Containers}}{{.IPv4Address}} {{end}}", network,
    ])
    ips: list[str] = []
    for entry in proc.stdout.split():
-        # entries look like "172.20.0.2/16" — keep the address.
        ips.append(entry.split("/", 1)[0])
    return ips

@@ -100,32 +92,24 @@ def launch_consolidated(
    image_ref: str = "",
    tokens: dict[str, str] | None = None,
    service: OrchestratorService | None = None,
-    gateway_name: str = GATEWAY_NAME,
+    infra_name: str = INFRA_NAME,
    network: str = GATEWAY_NETWORK,
 ) -> LaunchContext:
-    """Ensure the orchestrator + gateway are up, allocate + register the
-    bottle, and provision its git-gate state. Returns the agent's attach
-    context. Raises `ConsolidatedLaunchError` (or the primitives' own errors)
-    if any step fails — the caller tears down on failure."""
+    """Ensure the infra container is up, allocate + register the bottle, and
+    provision its git-gate state. Returns the agent's attach context."""
    service = service or OrchestratorService()
    url = service.ensure_running()
    client = OrchestratorClient(url)

    cidr = _network_cidr(network)
-    gateway_ip = _container_ip(gateway_name, network)
+    gateway_ip = _container_ip(infra_name, network)
    source_ip = next_free_ip(cidr, _network_container_ips(network))

-    inputs = registration_inputs(egress_plan)
-    reg = client.register_bottle(
-        source_ip, image_ref=image_ref, policy=inputs.policy,
-        metadata=inputs.metadata, tokens=tokens,
+    transport = DockerGatewayTransport(infra_name)
+    reg = provision_bottle(
+        client, source_ip, egress_plan, git_gate_plan, transport,
+        image_ref=image_ref, tokens=tokens,
    )
-    try:
-        provision_git_gate(gateway_name, reg.bottle_id, git_gate_plan)
-    except Exception:
-        # Roll the registration back so a provisioning failure leaves no orphan.
-        client.teardown_bottle(reg.bottle_id)
-        raise
    return LaunchContext(
        bottle_id=reg.bottle_id,
        identity_token=reg.identity_token,
@@ -137,12 +121,12 @@ def launch_consolidated(


 def teardown_consolidated(
-    bottle_id: str, *, orchestrator_url: str, gateway_name: str = GATEWAY_NAME,
+    bottle_id: str, *, orchestrator_url: str, infra_name: str = INFRA_NAME,
+    timeout: float | None = None,
 ) -> None:
-    """Deregister the bottle and remove its git-gate state from the gateway.
-    Both steps are idempotent so this is safe from a cleanup trap."""
-    OrchestratorClient(orchestrator_url).teardown_bottle(bottle_id)
-    deprovision_git_gate(gateway_name, bottle_id)
+    """Deregister the bottle and remove its git-gate state. Idempotent."""
+    _teardown_util(bottle_id, DockerGatewayTransport(infra_name),
+                   orchestrator_url=orchestrator_url, timeout=timeout)


 __all__ = [
@@ -15,14 +15,15 @@ bottle's push credentials out of another's repos on the shared gateway.
 from __future__ import annotations

 import re
+from typing import Protocol

 from ...docker_cmd import run_docker
 from ...git_gate import GitGatePlan, git_gate_render_provision

 # bottle ids index the gateway's per-bottle repo + creds dirs; they land in
-# `docker cp`/`rm` path arguments, so validate before any path is built (a
-# traversal id like "../etc" must never reach the container). Registry ids are
-# token_hex — this is defense in depth at the docker boundary.
+# exec/cp path arguments, so validate before any path is built (a traversal
+# id like "../etc" must never reach the gateway). Registry ids are token_hex —
+# this is defense in depth at the transport boundary.
 _SAFE_BOTTLE_ID = re.compile(r"[A-Za-z0-9_-]+")


@@ -30,6 +31,42 @@ class GatewayProvisionError(RuntimeError):
    """A git-gate provisioning step against the running gateway failed."""


+class GatewayTransport(Protocol):
+    """How the launcher stages files + runs commands in the running gateway.
+    Backend-neutral so the same provisioning logic serves the docker gateway
+    (exec/cp over the docker socket) and the firecracker gateway VM (over
+    SSH)."""
+
+    def exec(self, argv: list[str]) -> None:
+        """Run `argv` in the gateway, raising `GatewayProvisionError` on
+        failure."""
+
+    def cp_into(self, src: str, dest: str) -> None:
+        """Copy host file `src` to `dest` in the gateway, raising on
+        failure."""
+
+
+class DockerGatewayTransport:
+    """`GatewayTransport` for the docker gateway container (exec/cp)."""
+
+    def __init__(self, gateway: str) -> None:
+        self.gateway = gateway
+
+    def exec(self, argv: list[str]) -> None:
+        proc = run_docker(["docker", "exec", self.gateway, *argv])
+        if proc.returncode != 0:
+            raise GatewayProvisionError(
+                f"gateway exec {argv!r} failed: {proc.stderr.strip()}"
+            )
+
+    def cp_into(self, src: str, dest: str) -> None:
+        proc = run_docker(["docker", "cp", src, f"{self.gateway}:{dest}"])
+        if proc.returncode != 0:
+            raise GatewayProvisionError(
+                f"gateway cp {src} -> {dest} failed: {proc.stderr.strip()}"
+            )
+
+
 def _require_safe(bottle_id: str) -> None:
    if not _SAFE_BOTTLE_ID.fullmatch(bottle_id):
        raise GatewayProvisionError(f"unsafe bottle id {bottle_id!r}")
@@ -39,27 +76,11 @@ def _creds_dir(bottle_id: str) -> str:
    return f"/git-gate/creds/{bottle_id}"


-def _exec(gateway: str, argv: list[str]) -> None:
-    """`docker exec` a command in the gateway, raising on non-zero exit."""
-    proc = run_docker(["docker", "exec", gateway, *argv])
-    if proc.returncode != 0:
-        raise GatewayProvisionError(
-            f"gateway exec {argv!r} failed: {proc.stderr.strip()}"
-        )
-
-
-def _cp_into(gateway: str, src: str, dest: str) -> None:
-    """`docker cp` a host file into the gateway, raising on non-zero exit."""
-    proc = run_docker(["docker", "cp", src, f"{gateway}:{dest}"])
-    if proc.returncode != 0:
-        raise GatewayProvisionError(
-            f"gateway cp {src} -> {dest} failed: {proc.stderr.strip()}"
-        )
-
-
-def provision_git_gate(gateway: str, bottle_id: str, plan: GitGatePlan) -> None:
-    """Place `bottle_id`'s git-gate credentials into the running `gateway`
-    container and init its bare repos under `/git/<bottle_id>/`.
+def provision_git_gate(
+    transport: GatewayTransport, bottle_id: str, plan: GitGatePlan,
+) -> None:
+    """Place `bottle_id`'s git-gate credentials into the running gateway and
+    init its bare repos under `/git/<bottle_id>/`.

    Copies each upstream's identity key (and known_hosts, when present) into
    `/git-gate/creds/<bottle_id>/`, then runs the namespaced provisioning
@@ -70,31 +91,42 @@ def provision_git_gate(gateway: str, bottle_id: str, plan: GitGatePlan) -> None:
    # The pre-receive + access hooks are bottle-agnostic and shared by every
    # bottle's repos; install them into the gateway (idempotent — same content
    # each time). The per-bottle model cp'd these into each bundle at start.
-    _exec(gateway, ["mkdir", "-p", "/etc/git-gate"])
-    _cp_into(gateway, str(plan.hook_script), "/etc/git-gate/pre-receive")
-    _cp_into(gateway, str(plan.access_hook_script), "/etc/git-gate/access-hook")
+    transport.exec(["mkdir", "-p", "/etc/git-gate"])
+    transport.cp_into(str(plan.hook_script), "/etc/git-gate/pre-receive")
+    transport.cp_into(str(plan.access_hook_script), "/etc/git-gate/access-hook")
+    # The access-hook is exec'd directly (not via `sh`), so it needs the x bit.
+    # Set it here rather than trusting the copy to carry the staged 0o700:
+    # `docker cp` preserves source mode, but the Apple `container cp` does not,
+    # landing the hook 0o644 → EACCES when the git-http handler tries to exec it.
+    # chmod on the gateway side is backend-neutral and fixes every transport.
+    transport.exec(["chmod", "+x", "/etc/git-gate/access-hook"])
    creds = _creds_dir(bottle_id)
-    _exec(gateway, ["mkdir", "-p", creds])
+    transport.exec(["mkdir", "-p", creds])
    for u in plan.upstreams:
        if u.identity_file:
-            _cp_into(gateway, u.identity_file, f"{creds}/{u.name}-key")
+            transport.cp_into(u.identity_file, f"{creds}/{u.name}-key")
        known_hosts = str(u.known_hosts_file)
        if known_hosts and known_hosts != ".":
-            _cp_into(gateway, known_hosts, f"{creds}/{u.name}-known_hosts")
+            transport.cp_into(known_hosts, f"{creds}/{u.name}-known_hosts")
    # Init the bare repos + per-repo credential config for this namespace.
    script = git_gate_render_provision(bottle_id, plan.upstreams)
-    _exec(gateway, ["sh", "-c", script])
+    transport.exec(["sh", "-c", script])


-def deprovision_git_gate(gateway: str, bottle_id: str) -> None:
+def deprovision_git_gate(transport: GatewayTransport, bottle_id: str) -> None:
    """Remove a bottle's repos + creds from the gateway on teardown. Idempotent
    — an already-absent namespace is a clean no-op (best effort; a stray dir
    can't leak, since attribution is by source IP and the bottle is gone)."""
    _require_safe(bottle_id)
-    run_docker([
-        "docker", "exec", gateway, "rm", "-rf",
-        f"/git/{bottle_id}", _creds_dir(bottle_id),
-    ])
+    try:
+        transport.exec([
+            "rm", "-rf", f"/git/{bottle_id}", _creds_dir(bottle_id),
+        ])
+    except GatewayProvisionError:
+        pass  # best-effort teardown; absent namespace is success


-__all__ = ["provision_git_gate", "deprovision_git_gate", "GatewayProvisionError"]
+__all__ = [
+    "provision_git_gate", "deprovision_git_gate",
+    "GatewayProvisionError", "GatewayTransport", "DockerGatewayTransport",
+]
@@ -36,12 +36,15 @@ from contextlib import ExitStack, contextmanager
 from pathlib import Path
 from typing import Callable, Generator

+from ...agent_provider import runtime_for
 from ...egress import egress_resolve_token_values
 from ...git_gate import (
    provision_git_gate_dynamic_keys,
    revoke_git_gate_provisioned_keys,
 )
-from ...log import info, warn
+from ...image_cache import check_stale
+from ...log import die, info, warn
+from .. import BottleImages
 from . import util as docker_mod
 from .bottle import DockerBottle
 from .bottle_plan import DockerBottlePlan
@@ -61,6 +64,7 @@ from .compose import (
    write_compose_file,
 )
 from .consolidated_compose import consolidated_agent_compose
+from ...orchestrator.config_store import resolve_teardown_timeout
 from .consolidated_launch import launch_consolidated, teardown_consolidated
 from ...orchestrator.gateway import DockerGateway

@@ -69,16 +73,47 @@ from ...orchestrator.gateway import DockerGateway
 _REPO_DIR = str(Path(__file__).resolve().parent.parent.parent.parent)


+def build_or_load_images(plan: DockerBottlePlan) -> BottleImages:
+    """Resolve the agent image ref for this plan.
+
+    Returns the committed snapshot if one exists, the cached image when the
+    policy is 'cached', or builds a fresh image and returns that."""
+    committed = read_committed_image(plan.slug)
+    if committed and docker_mod.image_exists(committed):
+        info(f"using committed image {committed!r}")
+        return BottleImages(agent=committed)
+    if plan.spec.image_policy == "cached":
+        if not docker_mod.image_exists(plan.image):
+            die(
+                f"cached agent image {plan.image!r} not found; "
+                "run without --cached-images to build it"
+            )
+        info(f"using cached agent image {plan.image!r}")
+        return BottleImages(agent=plan.image)
+    docker_mod.build_image(plan.image, _REPO_DIR, dockerfile=plan.dockerfile_path)
+    docker_mod.verify_agent_image(
+        plan.image, runtime_for(plan.agent_provider_template).smoke_test,
+    )
+    return BottleImages(agent=plan.image)
+
+
@contextmanager
 def launch(
    plan: DockerBottlePlan,
+    images: BottleImages,
    *,
    provision: Callable[[DockerBottlePlan, "DockerBottle"], str | None],
 ) -> Generator[DockerBottle, None, None]:
-    """Build, launch, and provision a Docker bottle via compose.
-    Teardown on exit."""
+    """Launch and provision a Docker bottle via compose. Teardown on exit."""
    stack = ExitStack()

+    # Stamp the resolved agent image ref into the plan so compose rendering
+    # picks up the right image (may be a committed snapshot or cached ref).
+    plan = dataclasses.replace(
+        plan,
+        agent_provision=dataclasses.replace(plan.agent_provision, image=str(images.agent)),
+    )
+
    _bottle_for_revoke = plan.manifest.bottle
    _git_gate_dir_for_revoke = git_gate_state_dir(plan.slug)

@@ -95,22 +130,6 @@ def launch(
        )

    try:
-        # Step 1: agent image. Use a committed snapshot when one exists
-        # and is present in the local daemon; otherwise build from the
-        # Dockerfile. (The gateway image is built by the orchestrator.)
-        committed = read_committed_image(plan.slug)
-        if committed and docker_mod.image_exists(committed):
-            info(f"using committed image {committed!r}")
-            plan = dataclasses.replace(
-                plan,
-                agent_provision=dataclasses.replace(plan.agent_provision, image=committed),
-            )
-        else:
-            docker_mod.build_image(
-                plan.image, _REPO_DIR,
-                dockerfile=plan.dockerfile_path,
-            )
-
        # Step 2: mint the git-gate dynamic (gitea) deploy keys, if any, before
        # provisioning the bottle's repos into the shared gateway.
        git_gate_plan = plan.git_gate_plan
@@ -129,11 +148,14 @@ def launch(
        token_values = egress_resolve_token_values(
            plan.egress_plan.token_env_map, effective_env,
        )
+        teardown_timeout = resolve_teardown_timeout()
        ctx = launch_consolidated(
            plan.egress_plan, git_gate_plan, image_ref=plan.image, tokens=token_values,
        )
        stack.callback(
-            teardown_consolidated, ctx.bottle_id, orchestrator_url=ctx.orchestrator_url,
+            teardown_consolidated, ctx.bottle_id,
+            orchestrator_url=ctx.orchestrator_url,
+            timeout=teardown_timeout,
        )

        # Step 4: install the SHARED gateway CA into the agent (replaces the
@@ -163,6 +185,7 @@ def launch(
            egress_plan=egress_plan,
            agent_git_gate_url=git_gate_url,
            agent_supervise_url=supervise_url,
+            identity_token=ctx.identity_token,
        )

        # Step 5: render + up the agent-only compose, pinned on the shared
@@ -202,3 +225,21 @@ def launch(
        yield bottle
    finally:
        teardown()
+
+
+def stale_checks(plan: DockerBottlePlan) -> None:
+    """Raise StaleImageError if a cached image is older than the configured
+    threshold. Only runs when image_policy is 'cached'. Called by the backend
+    class's _image_stale_checks before _launch_impl starts any resources."""
+    if plan.spec.image_policy != "cached":
+        return
+    committed = read_committed_image(plan.slug)
+    if committed and docker_mod.image_exists(committed):
+        ts = docker_mod.image_created_at(committed)
+        if ts is not None:
+            check_stale(f"agent image {committed!r}", ts)
+        return
+    if docker_mod.image_exists(plan.image):
+        ts = docker_mod.image_created_at(plan.image)
+        if ts is not None:
+            check_stale(f"agent image {plan.image!r}", ts)
@@ -27,10 +27,14 @@ def _docker_on_path() -> bool:
 def _daemon_reachable() -> bool:
    if not _docker_on_path():
        return False
-    return subprocess.run(
-        ["docker", "info"],
-        stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, check=False,
-    ).returncode == 0
+    try:
+        return subprocess.run(
+            ["docker", "info"],
+            stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+            check=False, timeout=5,
+        ).returncode == 0
+    except subprocess.TimeoutExpired:
+        return False


 def _print_install_pointer() -> None:
@@ -4,11 +4,14 @@ existence, and building images."""

 from __future__ import annotations

+import os
+from datetime import datetime, timezone
 import re
 import shutil
 import subprocess
-from typing import Iterable, Iterator
+from typing import Iterator

+from ...docker_cmd import run_docker
 from ...log import die, info
 # from ...workspace import WorkspacePlan

@@ -30,12 +33,7 @@ def container_name_candidates(base: str) -> Iterator[str]:
 def runsc_available() -> bool:
    """Return True if the Docker daemon has the gVisor (`runsc`) runtime
    registered. Called once per prepare; the result lives on the plan."""
-    r = subprocess.run(
-        ["docker", "info", "--format", "{{json .Runtimes}}"],
-        capture_output=True,
-        text=True,
-        check=False,
-    )
+    r = run_docker(["docker", "info", "--format", "{{json .Runtimes}}"])
    return r.returncode == 0 and "runsc" in r.stdout


@@ -49,20 +47,15 @@ def require_docker() -> None:


 def image_exists(ref: str) -> bool:
-    return _silent_run(["docker", "image", "inspect", ref]) == 0
+    return run_docker(["docker", "image", "inspect", ref]).returncode == 0


 def container_exists(name: str) -> bool:
    """Returns True if a container (running or stopped) with the given
    name exists. Uses `docker ps -a -q -f name=^<name>$` so substring
    matches don't false-positive."""
-    result = subprocess.run(
-        ["docker", "ps", "-a", "-q", "-f", f"name=^{name}$"],
-        capture_output=True,
-        text=True,
-        check=True,
-    )
-    return bool(result.stdout.strip())
+    result = run_docker(["docker", "ps", "-a", "-q", "-f", f"name=^{name}$"])
+    return result.returncode == 0 and bool(result.stdout.strip())


 def force_remove_container(name: str) -> None:
@@ -70,12 +63,7 @@ def force_remove_container(name: str) -> None:
    doesn't — and the rm itself is best-effort (errors swallowed) so
    this is safe to register as a teardown callback."""
    if container_exists(name):
-        subprocess.run(
-            ["docker", "rm", "-f", name],
-            stdout=subprocess.DEVNULL,
-            stderr=subprocess.DEVNULL,
-            check=False,
-        )
+        run_docker(["docker", "rm", "-f", name])


 def docker_exec_root(container: str, argv: list[str]) -> None:
@@ -88,6 +76,29 @@ def docker_exec_root(container: str, argv: list[str]) -> None:
    )


+def docker_exec(container: str, argv: list[str], *, user: str = "") -> None:
+    """Run `docker exec` in the named container, dying with the
+    command's own stderr on failure. Pass `user=\"0\"` to run as root."""
+    cmd = ["docker", "exec"]
+    if user:
+        cmd += ["-u", user]
+    cmd += [container, *argv]
+    result = run_docker(cmd)
+    if result.returncode != 0:
+        die(
+            f"docker exec in {container} failed: "
+            f"{(result.stderr or '').strip() or '<no stderr>'}"
+        )
+
+
+def docker_cp(src: str, dest: str) -> None:
+    """Run `docker cp`, dying with the command's own stderr on failure."""
+    result = run_docker(["docker", "cp", src, dest])
+    if result.returncode != 0:
+        die(f"docker cp {src} -> {dest} failed: "
+            f"{(result.stderr or '').strip() or '<no stderr>'}")
+
+
 _SLUG_RE = re.compile(r"[^a-z0-9]+")


@@ -108,15 +119,40 @@ def build_image(ref: str, context: str, *, dockerfile: str = "") -> None:

    `dockerfile` is an optional path (relative to `context`, or
    absolute) for callers that need to build from a non-default
-    Dockerfile in the same context — e.g. `Dockerfile.git-gate`."""
+    Dockerfile in the same context — e.g. `Dockerfile.git-gate`.
+
+    Set `BOT_BOTTLE_NO_CACHE=1` (the `start --no-cache` flag) to force
+    `--no-cache`. The npm/curl installers some provider Dockerfiles
+    shell out to can silently no-op on a transient network failure —
+    e.g. an `optionalDependencies` fetch for a platform-native binary —
+    and Docker will then cache that broken layer indefinitely."""
    info(f"building image {ref} from {context} (layer cache keeps repeat builds fast)")
    args = ["docker", "build", "-t", ref]
+    if os.environ.get("BOT_BOTTLE_NO_CACHE") == "1":
+        args.append("--no-cache")
    if dockerfile:
        args.extend(["-f", dockerfile])
    args.append(context)
    subprocess.run(args, check=True)


+def verify_agent_image(image: str, argv: tuple[str, ...]) -> None:
+    """Run `argv` inside a throwaway container of a freshly built agent
+    image and die loudly if it fails, instead of shipping an image
+    whose CLI only breaks at first real use. No-op when the provider
+    hasn't declared a smoke test (`AgentProviderRuntime.smoke_test`)."""
+    if not argv:
+        return
+    result = run_docker(["docker", "run", "--rm", "--entrypoint", argv[0], image, *argv[1:]])
+    if result.returncode != 0:
+        detail = (result.stderr or result.stdout or "").strip()
+        die(
+            f"agent image {image!r} failed its post-build smoke test "
+            f"({' '.join(argv)}): {detail}\n"
+            f"Try rebuilding from scratch: bot-bottle start --no-cache"
+        )
+
+
 # def build_image_with_cwd(
 #     derived: str,
 #     base: str,
@@ -155,10 +191,7 @@ def build_image(ref: str, context: str, *, dockerfile: str = "") -> None:
 def commit_container(container_name: str, image_tag: str) -> None:
    """Run `docker commit <container_name> <image_tag>` to snapshot the
    running container's filesystem state as a local Docker image."""
-    result = subprocess.run(
-        ["docker", "commit", container_name, image_tag],
-        capture_output=True, text=True, check=False,
-    )
+    result = run_docker(["docker", "commit", container_name, image_tag])
    if result.returncode != 0:
        die(
            f"docker commit {container_name!r} → {image_tag!r} failed: "
@@ -167,10 +200,44 @@ def commit_container(container_name: str, image_tag: str) -> None:
    info(f"committed {container_name!r} → {image_tag!r}")


-def _silent_run(cmd: Iterable[str]) -> int:
-    return subprocess.run(
-        list(cmd),
-        stdout=subprocess.DEVNULL,
-        stderr=subprocess.DEVNULL,
+def image_created_at(ref: str) -> datetime | None:
+    """Return Docker's image Created timestamp as an aware UTC datetime, or
+    None when the field is absent or unparseable. Callers should skip the
+    stale check when None is returned."""
+    r = subprocess.run(
+        ["docker", "image", "inspect", "--format", "{{.Created}}", ref],
+        capture_output=True,
+        text=True,
        check=False,
-    ).returncode
+    )
+    if r.returncode != 0:
+        die(
+            f"docker image inspect for {ref!r} failed: "
+            f"{(r.stderr or '').strip() or '<no stderr>'}"
+        )
+    raw = r.stdout.strip()
+    if not raw:
+        return None
+    try:
+        return _parse_docker_timestamp(raw)
+    except ValueError:
+        return None
+
+
+def _parse_docker_timestamp(raw: str) -> datetime:
+    text = raw.strip()
+    if text.endswith("Z"):
+        text = text[:-1] + "+00:00"
+    dot = text.find(".")
+    if dot != -1:
+        tz_plus = text.find("+", dot)
+        tz_minus = text.find("-", dot)
+        tz_candidates = [pos for pos in (tz_plus, tz_minus) if pos != -1]
+        if tz_candidates:
+            tz_pos = min(tz_candidates)
+            frac = text[dot + 1:tz_pos]
+            text = text[:dot + 1] + frac[:6].ljust(6, "0") + text[tz_pos:]
+    dt = datetime.fromisoformat(text)
+    if dt.tzinfo is None:
+        dt = dt.replace(tzinfo=timezone.utc)
+    return dt.astimezone(timezone.utc)
@@ -18,7 +18,7 @@ from ...env import ResolvedEnv
 from ...git_gate import GitGatePlan
 from ...manifest import Manifest
 from ...supervise import SupervisePlan
-from .. import ActiveAgent, BottleBackend, BottleSpec
+from .. import ActiveAgent, BottleBackend, BottleImages, BottleSpec
 from . import cleanup as _cleanup
 from . import enumerate as _enumerate
 from . import launch as _launch
@@ -92,11 +92,18 @@ class FirecrackerBottleBackend(
            stage_dir=stage_dir,
        )

+    def _build_or_load_images(self, plan: FirecrackerBottlePlan) -> BottleImages:
+        return BottleImages(agent=_launch.build_or_load_agent_base(plan))
+
+    def prelaunch_checks(self, plan: FirecrackerBottlePlan) -> None:
+        _launch.stale_checks(plan)
+
    @contextmanager
-    def launch(
-        self, plan: FirecrackerBottlePlan
+    def _launch_impl(
+        self, plan: FirecrackerBottlePlan, images: BottleImages,
    ) -> Generator[FirecrackerBottle, None, None]:
-        with _launch.launch(plan, provision=self.provision) as bottle:
+        assert isinstance(images.agent, Path)
+        with _launch.launch(plan, images.agent, provision=self.provision) as bottle:
            yield bottle

    def prepare_cleanup(self) -> FirecrackerBottleCleanupPlan:
@@ -110,3 +117,7 @@ class FirecrackerBottleBackend(

    def supervise_mcp_url(self, plan: FirecrackerBottlePlan) -> str:
        return plan.agent_supervise_url
+
+    def ensure_orchestrator(self) -> str:
+        from . import infra_vm
+        return infra_vm.ensure_running().control_plane_url
@@ -105,10 +105,9 @@ class FirecrackerBottle(Bottle):
        # root-owned and unreadable by node, which breaks Node's
        # process.cwd(), the shell-snapshot machinery, and `/doctor`.
        # Use `env --chdir` rather than a `sh -c 'cd … && exec "$@"'`
-        # wrapper: ssh space-joins everything after the host into one
-        # string for the guest shell, so a quoted script + $@ would be
-        # re-split and mangled (exec'ing the $0 placeholder). All-simple
-        # words survive that join.
+        # wrapper: it keeps the guest command a flat argv that `agent_argv`
+        # can quote token-by-token for the ssh→guest-shell round trip,
+        # avoiding a fragile nested-quoting `"$@"` script.
        workdir = self.agent_workdir or _HOME_FOR["node"]
        remote = ["runuser", "-u", "node", "--",
                  "env", f"--chdir={workdir}",
@@ -117,7 +116,15 @@ class FirecrackerBottle(Bottle):
        return remote

    def agent_argv(self, argv: list[str], *, tty: bool = True) -> list[str]:
-        return [*self._ssh(tty=tty), "--", *self._agent_remote_argv(argv)]
+        # ssh space-joins everything after the host into one line the guest
+        # shell re-parses, so pre-quote each remote token for that shell.
+        # Simple words are unchanged (existing behaviour); an arg containing
+        # spaces — e.g. codex's `read_prompt_file` positional "Read and follow
+        # the instructions in <path>." — is quoted so it survives as ONE
+        # argument instead of being re-split (which made codex parse "and" as
+        # a subcommand).
+        remote = self._agent_remote_argv(argv)
+        return [*self._ssh(tty=tty), "--", *(shlex.quote(t) for t in remote)]

    def exec_agent(self, argv: list[str], *, tty: bool = True) -> int:
        agent_argv = self.agent_argv(argv, tty=tty)
@@ -18,6 +18,10 @@ class FirecrackerBottlePlan(BottlePlan):
    agent_proxy_url: str = ""
    agent_git_gate_url: str = ""
    agent_supervise_url: str = ""
+    # Per-bottle identity token the agent presents on every attributed request
+    # (egress proxy credentials, git-gate/supervise headers); set by launch
+    # from the orchestrator registration. Empty pre-registration.
+    identity_token: str = ""

    @property
    def container_name(self) -> str:
@@ -0,0 +1,103 @@
+"""Consolidated bottle launch sequence for the Firecracker backend
+(PRD 0070, Stage B).
+
+The shared gateway + orchestrator control plane run in a single persistent
+per-host **infra VM** (`infra_vm.py`), not Docker containers. Agent VMs reach
+the gateway's egress / supervise / git-http ports at the infra VM via a
+PREROUTING DNAT on their own host-side TAP IP (see
+`scripts/firecracker-netpool.sh`), and the host CLI reaches the control plane
+over HTTP at the infra VM's guest IP.
+
+Attribution is by the agent VM's guest IP, unspoofable by construction: the
+/31 point-to-point TAP + the `bot_bottle_fc` nft table ensure only the
+expected VM can source-IP that address.
+
+Sequence:
+  1. ensure the infra VM (control plane + gateway) is up (a singleton — a
+     prior launcher may already have booted it);
+  2. register the bottle by its guest IP (attribution key) → bottle id +
+     identity token;
+  3. provision its git-gate repos/creds into the gateway VM (over SSH);
+  4. fetch the shared gateway CA for the provisioner to install in the rootfs.
+
+The TAP slot allocation, rootfs build, and VM boot are the caller's job.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from ...egress import EgressPlan
+from ...git_gate import GitGatePlan
+from ...orchestrator.client import OrchestratorClient
+from ...orchestrator.lifecycle import (
+    OrchestratorStartError,  # re-exported so callers can catch it
+)
+from ..consolidated_util import provision_bottle, teardown_consolidated as _teardown_util
+from . import infra_vm
+
+
+class ConsolidatedLaunchError(RuntimeError):
+    """The consolidated register/provision sequence could not complete."""
+
+
+@dataclass(frozen=True)
+class LaunchContext:
+    """What the Firecracker launch needs from the consolidated sequence."""
+
+    bottle_id: str
+    identity_token: str
+    source_ip: str          # the VM's guest IP — the attribution key
+    gateway_ca_pem: str     # the shared gateway CA the provisioner installs
+    orchestrator_url: str
+
+
+def launch_consolidated(
+    egress_plan: EgressPlan,
+    git_gate_plan: GitGatePlan,
+    *,
+    guest_ip: str,
+    image_ref: str = "",
+    tokens: dict[str, str] | None = None,
+) -> LaunchContext:
+    """Ensure the infra VM is up, register the bottle by its guest IP, and
+    provision its git-gate state into the gateway VM. Returns the context the
+    agent-VM launch needs. Raises on failure — the caller tears down."""
+    infra = infra_vm.ensure_running()
+    url = infra.control_plane_url
+    client = OrchestratorClient(url)
+
+    transport = infra_vm.gateway_transport()
+    reg = provision_bottle(
+        client, guest_ip, egress_plan, git_gate_plan, transport,
+        image_ref=image_ref, tokens=tokens,
+    )
+    # The shared gateway CA every agent on this host trusts for TLS
+    # interception — fetched from the infra VM over SSH.
+    return LaunchContext(
+        bottle_id=reg.bottle_id,
+        identity_token=reg.identity_token,
+        source_ip=guest_ip,
+        gateway_ca_pem=infra.gateway_ca_pem(),
+        orchestrator_url=url,
+    )
+
+
+def teardown_consolidated(
+    bottle_id: str, *, orchestrator_url: str, timeout: float | None = None,
+) -> None:
+    """Deregister the bottle and remove its git-gate state from the gateway
+    VM. Both steps are idempotent so this is safe from a cleanup trap. Does
+    NOT stop the infra VM — it's a persistent per-host singleton shared by
+    every bottle."""
+    _teardown_util(bottle_id, infra_vm.gateway_transport(),
+                   orchestrator_url=orchestrator_url, timeout=timeout)
+
+
+__all__ = [
+    "LaunchContext",
+    "launch_consolidated",
+    "teardown_consolidated",
+    "ConsolidatedLaunchError",
+    "OrchestratorStartError",
+]
@@ -78,20 +78,32 @@ def _config(
    vcpus: int,
    mem_mib: int,
    guest_mac: str,
+    data_drive: Path | None = None,
 ) -> dict[str, object]:
+    drives: list[dict[str, object]] = [
+        {
+            "drive_id": "rootfs",
+            "path_on_host": str(rootfs),
+            "is_root_device": True,
+            "is_read_only": False,
+        }
+    ]
+    # A second virtio-block device (guest /dev/vdb) — the infra VM's
+    # persistent registry "volume", a host-side ext4 file that outlives the
+    # ephemeral rootfs across VM restarts.
+    if data_drive is not None:
+        drives.append({
+            "drive_id": "data",
+            "path_on_host": str(data_drive),
+            "is_root_device": False,
+            "is_read_only": False,
+        })
    return {
        "boot-source": {
            "kernel_image_path": str(util.kernel_path()),
            "boot_args": _boot_args(guest_ip, host_ip, pubkey),
        },
-        "drives": [
-            {
-                "drive_id": "rootfs",
-                "path_on_host": str(rootfs),
-                "is_root_device": True,
-                "is_read_only": False,
-            }
-        ],
+        "drives": drives,
        "network-interfaces": [
            {
                "iface_id": "eth0",
@@ -118,9 +130,16 @@ def boot(
    vcpus: int = 2,
    mem_mib: int = 2048,
    guest_mac: str = "06:00:AC:10:00:02",
+    detached: bool = False,
+    data_drive: Path | None = None,
 ) -> VmHandle:
    """Write the config and launch the VMM. Returns once the process is
-    spawned; callers wait for SSH readiness separately."""
+    spawned; callers wait for SSH readiness separately.
+
+    `detached` starts the VMM in its own session (`start_new_session`) so it
+    survives the launcher exiting — used for the persistent per-host infra
+    VM, which must outlive the short-lived `start` process (agent VMs stay
+    attached and are torn down with the launcher)."""
    run_dir.mkdir(parents=True, exist_ok=True)
    config_path = run_dir / "config.json"
    console_log = run_dir / "console.log"
@@ -128,6 +147,7 @@ def boot(
        _config(
            rootfs=rootfs, tap=tap, guest_ip=guest_ip, host_ip=host_ip,
            pubkey=pubkey, vcpus=vcpus, mem_mib=mem_mib, guest_mac=guest_mac,
+            data_drive=data_drive,
        ),
        indent=2,
    ))
@@ -137,6 +157,7 @@ def boot(
    process = subprocess.Popen(
        ["firecracker", "--no-api", "--config-file", str(config_path)],
        stdout=log_fh, stderr=subprocess.STDOUT, stdin=subprocess.DEVNULL,
+        start_new_session=detached,
    )
    return VmHandle(process=process, guest_ip=guest_ip, console_log=console_log)

@@ -1,9 +1,12 @@
-"""FirecrackerFreezer — snapshot a running microVM to a Docker image.
+"""FirecrackerFreezer — snapshot a running microVM to a rootfs tar.

 The VM is live and can't be block-copied safely, so — like the macOS
 backend — we stream the guest root filesystem out over the control
-channel (SSH here) and rebuild an image from it. The bottle keeps
-running after the snapshot.
+channel (SSH here). Unlike the other backends this needs no Docker: the
+tar *is* the resumable artifact. `resume` extracts it and rebuilds a
+fresh per-bottle ext4 with `mke2fs -d` (see `util.build_committed_rootfs_dir`
+and `launch.build_or_load_agent_base`). The bottle keeps running after the
+snapshot.
 """

 from __future__ import annotations
@@ -11,9 +14,9 @@ from __future__ import annotations
 import json
 import os
 import subprocess
-import tempfile
 from pathlib import Path

+from ...bottle_state import committed_rootfs_path
 from ...log import die, info
 from .. import ActiveAgent
 from ..freeze import Freezer
@@ -30,14 +33,13 @@ class FirecrackerFreezer(Freezer):
        if not private_key.is_file() or not guest_ip:
            die(f"cannot freeze {agent.slug}: run dir {run_dir} is missing the "
                f"SSH key or VM config (is the bottle still running?)")
-        image_tag = f"bot-bottle-committed-{agent.slug}:latest"
-        _commit_via_ssh(private_key, guest_ip, image_tag)
-        info(f"committed {agent.slug} -> {image_tag!r}")
-        return image_tag
+        tar_path = committed_rootfs_path(agent.slug)
+        _commit_rootfs_via_ssh(private_key, guest_ip, tar_path)
+        info(f"committed {agent.slug} -> {tar_path}")
+        return str(tar_path)

    def _export_hint(self, slug: str, image_ref: str) -> None:
-        info(f"to export for migration:      docker image save {image_ref} "
-             f"-o {slug}.tar")
+        info(f"to export for migration:      cp {image_ref} {slug}.tar")


 def _guest_ip_from_config(config_path: Path) -> str:
@@ -53,24 +55,36 @@ def _guest_ip_from_config(config_path: Path) -> str:
    return ""


-def _commit_via_ssh(private_key: Path, guest_ip: str, image_tag: str) -> None:
-    with tempfile.TemporaryDirectory(prefix="bot-bottle-fc-commit.") as tmp:
-        rootfs_tar = os.path.join(tmp, "rootfs.tar")
-        ssh = util.ssh_base_argv(private_key, guest_ip)
-        with open(rootfs_tar, "wb") as tar_out:
-            result = subprocess.run(
-                [*ssh, "--", "tar", "--create", "--one-file-system",
-                 "--exclude=./proc", "--exclude=./sys", "--exclude=./dev",
-                 "--exclude=./run", "--file=-", "--directory=/", "."],
-                stdout=tar_out, stderr=subprocess.PIPE, check=False,
-            )
-        if result.returncode != 0:
-            die(f"ssh tar for {guest_ip} failed: "
-                f"{(result.stderr or b'').decode().strip() or '<no stderr>'}")
-        with open(os.path.join(tmp, "Dockerfile"), "w", encoding="utf-8") as f:
-            f.write("FROM scratch\nADD rootfs.tar /\nUSER node\nWORKDIR /home/node\n")
-        build = subprocess.run(
-            ["docker", "build", "-t", image_tag, tmp], check=False,
+def _commit_rootfs_via_ssh(private_key: Path, guest_ip: str, tar_path: Path) -> None:
+    """Stream the guest rootfs out over SSH into `tar_path`. Excludes the
+    virtual/live mounts (proc/sys/dev/run) — resume recreates those empty
+    mount points. Written to a `.partial` sibling and renamed on success so
+    a failed freeze never leaves a truncated artifact in its place."""
+    tar_path.parent.mkdir(parents=True, exist_ok=True)
+    partial = tar_path.with_name(tar_path.name + ".partial")
+    ssh = util.ssh_base_argv(private_key, guest_ip)
+    # The snapshot can contain the bottle's private workspace, so keep it
+    # owner-only (0600) for the whole stream. The `os.open` mode only applies
+    # on *creation*, so unlink any leftover partial (a prior interrupted run
+    # could have left it world-readable, or something could swap in a symlink
+    # at this predictable name) and exclusively recreate it — O_EXCL|O_NOFOLLOW
+    # — then fchmod immediately so umask can't loosen it. Re-assert after the
+    # rename too (os.replace carries the source mode, but be explicit).
+    partial.unlink(missing_ok=True)
+    fd = os.open(
+        partial, os.O_WRONLY | os.O_CREAT | os.O_EXCL | os.O_NOFOLLOW, 0o600
+    )
+    os.fchmod(fd, 0o600)
+    with os.fdopen(fd, "wb") as tar_out:
+        result = subprocess.run(
+            [*ssh, "--", "tar", "--create", "--one-file-system",
+             "--exclude=./proc", "--exclude=./sys", "--exclude=./dev",
+             "--exclude=./run", "--file=-", "--directory=/", "."],
+            stdout=tar_out, stderr=subprocess.PIPE, check=False,
        )
-        if build.returncode != 0:
-            die(f"docker build for {image_tag!r} failed")
+    if result.returncode != 0:
+        partial.unlink(missing_ok=True)
+        die(f"ssh tar for {guest_ip} failed: "
+            f"{(result.stderr or b'').decode().strip() or '<no stderr>'}")
+    os.replace(partial, tar_path)
+    os.chmod(tar_path, 0o600)
@@ -0,0 +1,244 @@
+"""Docker-free agent-image builds for the Firecracker backend (PRD 0069 Stage 3).
+
+Agent Dockerfiles build **inside the persistent per-host infra VM**
+(`infra_vm.py`), which carries buildah (rootless, daemonless): no host Docker
+daemon, no root-equivalent `docker` group. The build runs over SSH against the
+infra VM and its rootfs streams back to the host, where the existing
+`mke2fs -d` path (`util.build_rootfs_ext4`) turns it into a bootable ext4.
+
+Building in the infra VM — rather than a throwaway builder VM — means there is
+one buildah image (`bot-bottle-infra`) and no contention for the orchestrator
+TAP. Tradeoff: an untrusted Dockerfile's `RUN` steps share the VM with the
+control plane + gateway (buildah `--isolation chroot` isn't a hard boundary) —
+the accepted single-VM blast-radius tradeoff, re-splittable into a disposable
+builder (booted from this same image on its own TAP) later.
+"""
+
+from __future__ import annotations
+
+import fcntl
+import hashlib
+import os
+import shutil
+import subprocess
+from contextlib import contextmanager
+from pathlib import Path
+from typing import Generator
+
+from ...log import die, info
+from . import infra_vm, util
+
+# vfs + chroot: buildah works as root in the microVM (no fuse-overlayfs /
+# overlay module / subuid maps). `--isolation` is a build/run-only flag;
+# `from`/`mount` take just the store.
+_BUILD_FLAGS = "--isolation chroot --storage-driver vfs"
+_STORE_FLAG = "--storage-driver vfs"
+
+_BUILD_TIMEOUT_SECONDS = 900.0
+
+
+def _dockerfile_hash(dockerfile: Path) -> str:
+    """The Dockerfile's content hash. The shipped agent Dockerfiles COPY
+    nothing from the build context (see .dockerignore), so their content fully
+    determines the built image; a Dockerfile that adds COPY will want the
+    context folded in here too."""
+    return hashlib.sha256(dockerfile.read_bytes()).hexdigest()[:16]
+
+
+def _rootfs_digest(dockerfile: Path) -> str:
+    """Cache key for the built AND boot-injected agent rootfs. Two inputs
+    determine the on-disk rootfs: the Dockerfile (the image) and the guest init
+    injected into it (`util._GUEST_INIT`). Folding the init in means a fix to
+    it — e.g. making /tmp world-writable — busts the cache instead of silently
+    reusing a stale rootfs built with the old init."""
+    h = hashlib.sha256()
+    h.update(_dockerfile_hash(dockerfile).encode())
+    h.update(b"\0")
+    h.update(util._GUEST_INIT.encode())
+    return h.hexdigest()[:16]
+
+
+def cached_agent_rootfs_dir(dockerfile: Path) -> Path | None:
+    """Return the ready cached rootfs for ``dockerfile``, if one exists."""
+    base = util.cache_dir() / "rootfs" / f"agent-{_rootfs_digest(dockerfile)}"
+    return base if (base / ".bb-ready").is_file() else None
+
+
+def build_agent_rootfs_dir(
+    dockerfile: Path, *, image_tag: str, smoke_test: tuple[str, ...] = (),
+) -> Path:
+    """Build `dockerfile` in the infra VM (buildah, no host docker), export its
+    rootfs, inject the guest boot bits, and return the cached base dir — the
+    same shape `util.build_rootfs_ext4` consumes. Cached by Dockerfile content
+    + injected guest init, so a repeat launch skips the rebuild but an init or
+    Dockerfile change rebuilds.
+
+    `smoke_test` (the provider's declared argv, e.g. `("claude","--version")`)
+    is run in the freshly built image before export, catching an npm
+    silent-failure image at build time rather than at first agent use."""
+    digest = _rootfs_digest(dockerfile)
+    base = util.cache_dir() / "rootfs" / f"agent-{digest}"
+    if cached_agent_rootfs_dir(dockerfile) is not None:
+        info(f"using cached agent rootfs {base.name}")
+        return base
+
+    # Serialize builds: the infra VM's buildah store + this cache dir are
+    # shared, so concurrent `start`s must not build into them at once. The
+    # lock covers the cache lookup + build + atomic publish; the ready
+    # fast-path above takes no lock.
+    with _build_lock():
+        if (base / ".bb-ready").is_file():  # another build finished while we waited
+            info(f"using cached agent rootfs {base.name}")
+            return base
+        # Build into a temp dir and publish by atomic rename, so a partial
+        # build is never visible as `agent-<digest>`.
+        staging = util.cache_dir() / "rootfs" / f".building-{digest}"
+        shutil.rmtree(staging, ignore_errors=True)
+        staging.mkdir(parents=True)
+        info(f"building agent image {image_tag!r} in the infra VM")
+        _build_in_infra(dockerfile, staging, smoke_test, digest)
+        util.inject_guest_boot(staging)
+        (staging / ".bb-ready").write_text("ok\n")
+        shutil.rmtree(base, ignore_errors=True)
+        os.rename(staging, base)
+    return base
+
+
+@contextmanager
+def _build_lock() -> Generator[None, None, None]:
+    """Host-level exclusive lock serializing agent-image builds (shared infra
+    buildah store + cache dir). flock auto-releases on a crash."""
+    lock_path = util.cache_dir() / "rootfs" / ".build.lock"
+    lock_path.parent.mkdir(parents=True, exist_ok=True)
+    handle = open(lock_path, "w", encoding="utf-8")
+    try:
+        fcntl.flock(handle, fcntl.LOCK_EX)
+        yield
+    finally:
+        handle.close()
+
+
+def _build_in_infra(
+    dockerfile: Path, base: Path, smoke_test: tuple[str, ...], digest: str,
+) -> None:
+    """Ensure the infra VM is up, `buildah build` the Dockerfile in it, smoke
+    test the image, and stream its rootfs into `base`. The infra VM persists;
+    only the per-build container/image/context are cleaned up."""
+    infra = infra_vm.ensure_running()
+    key, ip = infra.private_key, infra.guest_ip
+    tag = f"bot-bottle-agent-build-{digest}"
+    ctx = f"/tmp/agent-build-{digest}"
+    smoke_ctr, export_ctr = f"{tag}-smoke", f"{tag}-export"
+
+    def _cleanup() -> None:
+        # Remove only THIS build's working containers/image/context — never
+        # `buildah rm -a`, which would nuke a concurrent build's container.
+        _ssh(key, ip,
+             f"buildah rm {smoke_ctr} {export_ctr} >/dev/null 2>&1; "
+             f"buildah rmi {_STORE_FLAG} {tag} >/dev/null 2>&1; rm -rf {ctx}",
+             timeout=60)
+
+    _cleanup()  # clear leftovers from a crashed prior build of this digest
+    try:
+        prep = _ssh(key, ip, f"mkdir -p {ctx}/ctx")
+        if prep.returncode != 0:
+            die(f"preparing build dir in the infra VM failed: {prep.stderr.strip()}")
+        _send_dockerfile(key, ip, dockerfile, ctx)
+        _buildah_build(key, ip, ctx, tag)
+        _smoke_test(key, ip, tag, smoke_ctr, smoke_test)
+        _stream_rootfs(key, ip, tag, export_ctr, base)
+    finally:
+        _cleanup()
+
+
+def _ssh(private_key: Path, guest_ip: str, script: str,
+         *, timeout: float = 60.0) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        util.ssh_base_argv(private_key, guest_ip) + [script],
+        capture_output=True, text=True, timeout=timeout, check=False,
+    )
+
+
+def _ssh_streamed(private_key: Path, guest_ip: str, script: str,
+                  *, timeout: float) -> int:
+    """Run an SSH command letting the remote's stdout/stderr flow straight to
+    ours (no capture), for long chatty steps where live progress beats a
+    silent wait. Returns the exit code."""
+    proc = subprocess.run(
+        util.ssh_base_argv(private_key, guest_ip) + [script],
+        timeout=timeout, check=False,
+    )
+    return proc.returncode
+
+
+def _send_dockerfile(private_key: Path, guest_ip: str, dockerfile: Path, ctx: str) -> None:
+    proc = subprocess.run(
+        util.ssh_base_argv(private_key, guest_ip) + [f"cat > {ctx}/Dockerfile"],
+        input=dockerfile.read_bytes(), capture_output=True, timeout=30, check=False,
+    )
+    if proc.returncode != 0:
+        die(f"sending Dockerfile to the infra VM failed: "
+            f"{proc.stderr.decode(errors='replace').strip()}")
+
+
+def _buildah_build(private_key: Path, guest_ip: str, ctx: str, tag: str) -> None:
+    # Stream buildah's step-by-step output straight to our stderr (like the
+    # docker backend's `docker build`), so a long first build (base pull +
+    # apt/npm installs) shows live progress instead of a silent wait. The
+    # remote stderr is where buildah writes its `STEP i/n` lines.
+    info(f"buildah build {tag} in the infra VM (streaming output)")
+    rc = _ssh_streamed(
+        private_key, guest_ip,
+        f"buildah build {_BUILD_FLAGS} -t {tag} -f {ctx}/Dockerfile {ctx}/ctx",
+        timeout=_BUILD_TIMEOUT_SECONDS,
+    )
+    if rc != 0:
+        die(f"buildah build in the infra VM failed (exit {rc}); "
+            "see the build output above.")
+
+
+def _smoke_test(private_key: Path, guest_ip: str, tag: str, ctr: str,
+                argv: tuple[str, ...]) -> None:
+    """Run the provider's smoke argv inside the freshly built image
+    (`buildah run`, which uses the image's own PATH), failing the build
+    loudly if the CLI is a broken stub. No-op without a declared test. Uses a
+    named working container (`ctr`) so cleanup is scoped to this build."""
+    if not argv:
+        return
+    cmd = (
+        f"set -e; buildah from {_STORE_FLAG} --name {ctr} {tag} >/dev/null; "
+        f"buildah run {_BUILD_FLAGS} {ctr} -- {' '.join(argv)}; rc=$?; "
+        f"buildah rm {ctr} >/dev/null 2>&1 || true; exit $rc"
+    )
+    result = _ssh(private_key, guest_ip, cmd, timeout=120)
+    if result.returncode != 0:
+        detail = (result.stdout + result.stderr).strip().splitlines()[-10:]
+        die(f"agent image failed its post-build smoke test "
+            f"({' '.join(argv)}):\n" + "\n".join(detail))
+
+
+def _stream_rootfs(private_key: Path, guest_ip: str, tag: str, ctr: str, base: Path) -> None:
+    """`buildah mount` the built image in the infra VM and pipe its rootfs tar
+    straight into `base` on the host (extracted as the non-root host user, so
+    uid 0 isn't preserved — the guest init restores /root ownership). Uses a
+    named working container so cleanup is scoped to this build."""
+    export = (
+        f"set -e; buildah from {_STORE_FLAG} --name {ctr} {tag} >/dev/null; "
+        f"mnt=$(buildah mount {_STORE_FLAG} {ctr}); "
+        f"tar -C \"$mnt\" -cf - ."
+    )
+    ssh_proc = subprocess.Popen(
+        util.ssh_base_argv(private_key, guest_ip) + [export],
+        stdout=subprocess.PIPE, stderr=subprocess.PIPE,
+    )
+    assert ssh_proc.stdout is not None
+    untar = subprocess.run(
+        ["tar", "-x", "-C", str(base)], stdin=ssh_proc.stdout, check=False,
+    )
+    ssh_proc.stdout.close()
+    ssh_err = (ssh_proc.stderr.read().decode(errors="replace")
+               if ssh_proc.stderr else "")
+    rc = ssh_proc.wait()
+    if rc != 0 or untar.returncode != 0:
+        die(f"exporting the built rootfs from the infra VM failed: "
+            f"{ssh_err.strip() or '<no stderr>'}")
@@ -0,0 +1,232 @@
+"""Prebuilt infra-VM rootfs, pulled as an artifact (PRD 0069 Stage 2).
+
+The Firecracker infra VM boots a fixed rootfs (orchestrator control plane +
+gateway + buildah, control-plane init as PID 1) that does not vary per launch —
+the per-boot bits (authorized_keys, guest IP) ride the kernel cmdline, so one
+rootfs boots on any host. Instead of building that rootfs on the launch host
+with Docker, we build it **off-host** and publish it as a versioned, ready-to-
+boot ext4 (gzip-compressed) to a Gitea **generic package**; the launch host
+downloads + verifies + boots it. No Docker, no image tooling on the launch
+host — just an HTTP fetch and gunzip.
+
+  publish (off-host, see publish_infra.py):
+    docker build -> rootfs dir -> mke2fs -> gzip -> PUT generic package
+  pull (this module, launch host):
+    GET .../rootfs.ext4.gz (+ .sha256) -> verify -> gunzip -> boot
+
+The artifact **version** is a content hash of everything baked into the rootfs
+(the shipped bot_bottle package, the three Dockerfiles, and the init), so a
+launch host always pulls the artifact matching its code and a content change
+can't silently boot a stale rootfs. A checksum mismatch fails closed.
+
+Set `BOT_BOTTLE_INFRA_BUILD=local` to skip the pull and build the rootfs
+locally with Docker (dev iteration on the Dockerfiles) — see `infra_vm`.
+"""
+
+from __future__ import annotations
+
+import gzip
+import hashlib
+import os
+import shutil
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+from ...log import die, info
+from . import util
+
+# Bump if the on-disk artifact *format* changes (compression, layout) so a new
+# scheme can't collide with a cached/published artifact of the old one.
+_ARTIFACT_FORMAT = "1"
+
+_REPO_ROOT = Path(__file__).resolve().parents[3]
+_DOCKERFILES = ("Dockerfile.orchestrator", "Dockerfile.gateway", "Dockerfile.infra", "Dockerfile.infra.fc")
+
+_DEFAULT_BASE = "https://gitea.dideric.is"
+_DEFAULT_OWNER = "didericis"
+_PACKAGE = "bot-bottle-firecracker-infra"
+
+# Streaming copy chunk for the (hundreds-of-MB) download.
+_CHUNK = 1 << 20
+
+
+def local_build_requested() -> bool:
+    """True when the operator opted into the dev Docker-build path instead of
+    pulling the published artifact (`BOT_BOTTLE_INFRA_BUILD=local`)."""
+    return os.environ.get("BOT_BOTTLE_INFRA_BUILD", "").strip().lower() == "local"
+
+
+def infra_artifact_version(init_script: str, *, repo_root: Path = _REPO_ROOT) -> str:
+    """Content hash (16 hex) of everything baked into the infra rootfs: the
+    whole shipped `bot_bottle` package, the three fixed Dockerfiles, and the
+    guest init. Deterministic across the publish host and the launch host when
+    both run the same checkout, so the tag the launch host pulls is exactly the
+    tag publish produced.
+
+    The package is `COPY bot_bottle /app/bot_bottle`'d wholesale into the image,
+    so hash *every* regular file under it — not just `*.py`. Non-Python inputs
+    (e.g. `egress_entrypoint.sh`, `netpool.defaults.env`) are baked in too, and
+    a change to one must bump the version or a launch host could boot a stale
+    rootfs whose code differs from its checkout. `__pycache__`/`.pyc` are the
+    only exclusions — build artifacts, never copied."""
+    h = hashlib.sha256()
+    h.update(f"format={_ARTIFACT_FORMAT}\n".encode())
+    pkg = repo_root / "bot_bottle"
+    for path in sorted(pkg.rglob("*")):
+        if not path.is_file():
+            continue
+        if "__pycache__" in path.parts or path.suffix == ".pyc":
+            continue
+        h.update(str(path.relative_to(repo_root)).encode())
+        h.update(b"\0")
+        h.update(path.read_bytes())
+    for name in _DOCKERFILES:
+        h.update(name.encode())
+        h.update(b"\0")
+        h.update((repo_root / name).read_bytes())
+    h.update(b"pyproject.toml\0")
+    h.update((repo_root / "pyproject.toml").read_bytes())
+    h.update(b"dropbear\0")
+    dropbear = util.dropbear_path()
+    h.update(dropbear.read_bytes() if dropbear.is_file() else b"<missing>")
+    h.update(b"init\0")
+    h.update(init_script.encode())
+    return h.hexdigest()[:16]
+
+
+def _config() -> tuple[str, str, str]:
+    """(base_url, owner, token) for the generic-package endpoint. Base + owner
+    are overridable for other deployments / mirrors; the token comes solely from
+    `BOT_BOTTLE_INFRA_ARTIFACT_TOKEN` (a dedicated package-scoped token, kept
+    separate from the general-purpose Gitea token) and is optional — a public
+    package needs none to pull."""
+    base = os.environ.get("BOT_BOTTLE_INFRA_ARTIFACT_BASE", _DEFAULT_BASE).rstrip("/")
+    owner = os.environ.get("BOT_BOTTLE_INFRA_ARTIFACT_OWNER", _DEFAULT_OWNER)
+    token = os.environ.get("BOT_BOTTLE_INFRA_ARTIFACT_TOKEN", "")
+    return base, owner, token
+
+
+def artifact_url(version: str, filename: str) -> str:
+    """The generic-package download URL for one file of this version's
+    artifact (`rootfs.ext4.gz` / `rootfs.ext4.gz.sha256`)."""
+    base, owner, _ = _config()
+    return f"{base}/api/packages/{owner}/generic/{_PACKAGE}/{version}/{filename}"
+
+
+_GZ_NAME = "rootfs.ext4.gz"
+_SHA_NAME = "rootfs.ext4.gz.sha256"
+_CANDIDATE_DIR_ENV = "BOT_BOTTLE_INFRA_ARTIFACT_DIR"
+
+
+def _cache_root(version: str) -> Path:
+    return util.cache_dir() / "infra-artifact" / version
+
+
+def _open(url: str) -> urllib.request.Request:
+    _, _, token = _config()
+    req = urllib.request.Request(url)
+    if token:
+        req.add_header("Authorization", f"token {token}")
+    return req
+
+
+def _download(url: str, dest: Path) -> None:
+    """Stream `url` to `dest` (atomic via a `.part` sibling)."""
+    tmp = dest.with_suffix(dest.suffix + ".part")
+    try:
+        with urllib.request.urlopen(_open(url)) as resp, open(tmp, "wb") as out:
+            shutil.copyfileobj(resp, out, _CHUNK)
+    except urllib.error.HTTPError as e:
+        tmp.unlink(missing_ok=True)
+        if e.code == 404:
+            die(
+                f"infra artifact not published for this code version.\n"
+                f"    missing: {url}\n"
+                f"    publish it from a build host (Docker):\n"
+                f"        python3 -m bot_bottle.backend.firecracker.publish_infra\n"
+                f"    or build the rootfs locally: BOT_BOTTLE_INFRA_BUILD=local"
+            )
+        die(f"downloading infra artifact failed (HTTP {e.code}): {url}")
+    except urllib.error.URLError as e:
+        tmp.unlink(missing_ok=True)
+        die(f"infra artifact registry unreachable: {url} ({e.reason})")
+    tmp.replace(dest)
+
+
+def _sha256_file(path: Path) -> str:
+    h = hashlib.sha256()
+    with open(path, "rb") as fh:
+        for chunk in iter(lambda: fh.read(_CHUNK), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+def ensure_artifact_gz(version: str) -> Path:
+    """The verified, cached `rootfs.ext4.gz` for `version` — downloading it (and
+    its `.sha256`) once, then reusing it. Fail-closed on a checksum mismatch:
+    the partial is removed and we die rather than boot an unverified rootfs."""
+    candidate_dir = os.environ.get(_CANDIDATE_DIR_ENV, "").strip()
+    if candidate_dir:
+        root = Path(candidate_dir)
+        version_file = root / "version.txt"
+        # Guard the read so a missing version.txt is a clean error, not a raw
+        # FileNotFoundError.
+        if not version_file.is_file():
+            die(f"infra candidate bundle is incomplete: {root}")
+        declared = version_file.read_text(encoding="utf-8").strip()
+        if declared != version:
+            die(
+                f"infra candidate version mismatch: expected {version}, "
+                f"bundle contains {declared or '<empty>'}"
+            )
+        gz = root / _GZ_NAME
+        sha = root / _SHA_NAME
+        if not gz.is_file() or not sha.is_file():
+            die(f"infra candidate bundle is incomplete: {root}")
+        expected = sha.read_text().split()[0].strip().lower()
+        actual = _sha256_file(gz)
+        if actual != expected:
+            die(
+                f"infra candidate checksum mismatch for {version}:\n"
+                f"    expected {expected}\n    actual   {actual}"
+            )
+        return gz
+
+    root = _cache_root(version)
+    root.mkdir(parents=True, exist_ok=True)
+    gz = root / _GZ_NAME
+    ok = root / ".verified"
+    if gz.is_file() and ok.is_file():
+        return gz
+
+    info(f"pulling infra rootfs artifact {_PACKAGE}/{version}")
+    _download(artifact_url(version, _GZ_NAME), gz)
+    sha = root / _SHA_NAME
+    _download(artifact_url(version, _SHA_NAME), sha)
+
+    expected = sha.read_text().split()[0].strip().lower()
+    actual = _sha256_file(gz)
+    if actual != expected:
+        gz.unlink(missing_ok=True)
+        sha.unlink(missing_ok=True)
+        die(
+            f"infra artifact checksum mismatch for {version}:\n"
+            f"    expected {expected}\n"
+            f"    actual   {actual}\n"
+            f"    refusing to boot an unverified rootfs."
+        )
+    ok.write_text("ok\n")
+    return gz
+
+
+def materialize_ext4(version: str, dest: Path) -> None:
+    """Ensure the verified artifact is cached, then gunzip it to `dest` — a
+    fresh, writable per-boot rootfs (the VM mutates it; the cached `.gz` stays
+    pristine). Atomic via a `.part` sibling."""
+    gz = ensure_artifact_gz(version)
+    tmp = dest.with_suffix(dest.suffix + ".part")
+    info(f"expanding infra rootfs -> {dest}")
+    with gzip.open(gz, "rb") as src, open(tmp, "wb") as out:
+        shutil.copyfileobj(src, out, _CHUNK)
+    tmp.replace(dest)
@@ -0,0 +1,504 @@
+"""The per-host infra VM for the Firecracker backend (PRD 0070 Stage B).
+
+A single persistent microVM that runs the orchestrator **control plane** (and,
+in a following step, the gateway **data plane**) — the trusted per-host service
+the docker backend runs as containers. It boots on the NAT'd orchestrator link
+(`netpool.orch_slot()`): the host CLI reaches its control plane over HTTP at the
+guest IP, and agent VMs reach its gateway ports over VM-to-VM routing.
+
+Build-from-source (the default while the design churns): the rootfs is exported
+from the locally built orchestrator image, which bakes the stdlib-only
+control-plane source. A pull-from-registry mode (Gitea's OCI registry) becomes
+the default later.
+
+SSH is left enabled for debugging; the control plane is the load-bearing
+surface.
+"""
+
+from __future__ import annotations
+
+import fcntl
+import hashlib
+import os
+import shlex
+import signal
+import stat
+import subprocess
+import time
+import urllib.error
+import urllib.request
+from contextlib import contextmanager
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Generator
+
+from ...log import die, info
+from .. import util as backend_util
+from ..docker import util as docker_mod
+from ..docker.gateway_provision import GatewayProvisionError
+from . import firecracker_vm, infra_artifact, netpool, util
+
+# The single infra-VM image: gateway data plane + baked control-plane source
+# (Dockerfile.infra FROM the gateway image). Built from source by default;
+# a pull-from-registry mode lands later.
+_INFRA_IMAGE = "bot-bottle-infra:latest"
+_GATEWAY_IMAGE = "bot-bottle-gateway:latest"
+_ORCHESTRATOR_IMAGE = "bot-bottle-orchestrator:latest"
+_REPO_ROOT = Path(__file__).resolve().parents[3]
+
+CONTROL_PLANE_PORT = 8099
+# Gateway data-plane ports (agent-facing): egress proxy, supervise MCP,
+# git-http. Reached by agent VMs over VM-to-VM routing (added next).
+EGRESS_PORT = 9099
+SUPERVISE_PORT = 9100
+GIT_HTTP_PORT = 9420
+# mitmproxy writes its CA here a beat after start; agents install it to trust
+# the gateway's TLS interception.
+_GATEWAY_CA_PATH = "/home/mitmproxy/.mitmproxy/mitmproxy-ca-cert.pem"
+
+# The infra VM makes direct upstream connections (gateway egress, and buildah
+# during builds), and the kernel `ip=` cmdline sets no resolver. Public for
+# now; routing DNS through a filtered path is a later refinement.
+_INFRA_RESOLVER = "1.1.1.1"
+
+_HEALTH_TIMEOUT_SECONDS = 45.0
+_HEALTH_POLL_SECONDS = 0.5
+_CA_TIMEOUT_SECONDS = 30.0
+
+
+@dataclass
+class InfraVm:
+    """A handle to the per-host infra VM: its guest IP and the stable SSH key
+    used to fetch the gateway CA / provision git-gate. `vm` is the live VMM
+    handle when this process booted it, and None when adopting a singleton a
+    prior launcher started (teardown then goes through the PID file)."""
+
+    guest_ip: str
+    private_key: Path
+    vm: firecracker_vm.VmHandle | None = None
+
+    @property
+    def control_plane_url(self) -> str:
+        return f"http://{self.guest_ip}:{CONTROL_PLANE_PORT}"
+
+    def terminate(self) -> None:
+        """Stop the infra VM — via the live handle if we booted it, else the
+        PID file (adopting-process case)."""
+        if self.vm is not None:
+            self.vm.terminate()
+        else:
+            _kill_pidfile()
+        _pid_file().unlink(missing_ok=True)
+
+    def gateway_ca_pem(self, *, timeout: float = _CA_TIMEOUT_SECONDS) -> str:
+        """The gateway's mitmproxy CA (PEM) that agents install to trust its
+        TLS interception. Generated a moment after boot, so this polls over
+        SSH until it appears (mirrors DockerGateway.ca_cert_pem)."""
+        def _fetch() -> str | None:
+            proc = subprocess.run(
+                util.ssh_base_argv(self.private_key, self.guest_ip)
+                + [f"cat {_GATEWAY_CA_PATH}"],
+                capture_output=True, text=True, timeout=15, check=False,
+            )
+            ok = proc.returncode == 0 and "BEGIN CERTIFICATE" in proc.stdout
+            return proc.stdout if ok else None
+        try:
+            return backend_util.poll_ca_cert(_fetch, timeout=timeout)
+        except TimeoutError as exc:
+            die(str(exc))
+
+
+def ensure_built() -> None:
+    """Ensure the infra rootfs is available before boot.
+
+    Default (docker-free, PRD 0069 Stage 2): download + verify the prebuilt
+    rootfs artifact matching this code version (see `infra_artifact`); the
+    launch host needs no Docker. `BOT_BOTTLE_INFRA_BUILD=local` instead builds
+    the three fixed images from source with host Docker — the infra image
+    `COPY --from`s the orchestrator image and is `FROM` the gateway image, so
+    both must exist first — for iterating on the Dockerfiles."""
+    if infra_artifact.local_build_requested():
+        build_infra_images_with_docker()
+        return
+    infra_artifact.ensure_artifact_gz(
+        infra_artifact.infra_artifact_version(_infra_init()))
+
+
+def build_infra_images_with_docker() -> None:
+    """Build the four fixed images from source with host Docker: orchestrator,
+    gateway, the shared infra base (Dockerfile.infra), then the Firecracker
+    infra image (Dockerfile.infra.fc: FROM infra + buildah). The launch host
+    uses this only in `BOT_BOTTLE_INFRA_BUILD=local` mode; `publish_infra`
+    uses it off-host to produce the published artifact."""
+    docker_mod.build_image(
+        _ORCHESTRATOR_IMAGE, str(_REPO_ROOT), dockerfile="Dockerfile.orchestrator")
+    docker_mod.build_image(
+        _GATEWAY_IMAGE, str(_REPO_ROOT), dockerfile="Dockerfile.gateway")
+    docker_mod.build_image(
+        "bot-bottle-infra:latest", str(_REPO_ROOT), dockerfile="Dockerfile.infra")
+    docker_mod.build_image(
+        _INFRA_IMAGE, str(_REPO_ROOT), dockerfile="Dockerfile.infra.fc")
+
+
+def build_infra_rootfs_dir() -> Path:
+    """The infra VM's base rootfs: the infra image prepared with the
+    control-plane + gateway init as PID 1. The init's content is folded into
+    the cache key so an init change rebuilds the rootfs (the base image digest
+    alone wouldn't catch it)."""
+    init = _infra_init()
+    tag = hashlib.sha256(init.encode()).hexdigest()[:8]
+    return util.build_base_rootfs_dir(
+        _INFRA_IMAGE, variant=f"-infra-{tag}", init_script=init,
+    )
+
+
+def ensure_running() -> InfraVm:
+    """Idempotent per-host singleton. Adopt the infra VM if its control plane
+    is already healthy (a prior launcher booted it — it outlives short-lived
+    `start` processes); otherwise clear any stale VM and boot a fresh one.
+    Returns a handle usable for CA fetch / git-gate provisioning.
+
+    Concurrency-safe: the cold stop/build/boot path is serialized by a host
+    flock, so two simultaneous first launches don't both boot on the same
+    rootfs/PID. The healthy fast-path takes no lock."""
+    slot = netpool.orch_slot()
+    url = f"http://{slot.guest_ip}:{CONTROL_PLANE_PORT}"
+    key = _infra_dir() / "id_ed25519"
+    want = _expected_version()
+    if _adoptable(key, url, want):
+        info(f"adopting running infra VM at {url}")
+        return InfraVm(guest_ip=slot.guest_ip, private_key=key)
+
+    with _singleton_lock():
+        # Re-check under the lock: another launcher may have booted it while
+        # we waited for the lock (double-checked, so we adopt not re-boot).
+        if _adoptable(key, url, want):
+            info(f"adopting running infra VM at {url}")
+            return InfraVm(guest_ip=slot.guest_ip, private_key=key)
+        # Clear a stale/hung/OUTDATED VM holding the link before booting fresh.
+        stop()
+        ensure_built()
+        infra = boot()
+        wait_for_health(infra)
+        _record_booted_version(want)
+        return infra
+
+
+@contextmanager
+def _singleton_lock() -> Generator[None, None, None]:
+    """Host-level exclusive lock serializing the infra VM's cold create path
+    (`stop`/`ensure_built`/`boot`). flock auto-releases if the launcher
+    crashes, so the lock is never leaked."""
+    lock_path = _infra_dir() / "singleton.lock"
+    handle = open(lock_path, "w", encoding="utf-8")
+    try:
+        fcntl.flock(handle, fcntl.LOCK_EX)
+        yield
+    finally:
+        handle.close()
+
+
+def stop() -> None:
+    """Stop the infra VM singleton (idempotent — absent is success). Reaps the
+    recorded VMM AND any orphaned firecracker still bound to the infra config —
+    the PID file drifts after crashes / out-of-band kills, and a survivor would
+    hold the orchestrator TAP so the next boot dies with "tap … Resource busy".
+    Drops the version marker so a stopped VM is never treated as adoptable."""
+    _kill_pidfile()
+    _kill_infra_firecrackers()
+    _pid_file().unlink(missing_ok=True)
+    _version_file().unlink(missing_ok=True)
+
+
+def boot() -> InfraVm:
+    """Boot the infra VM (detached, so it outlives the launcher) on the
+    orchestrator link, recording its PID. Prefer `ensure_running`."""
+    slot = netpool.orch_slot()
+    if not netpool.tap_present(slot.iface):
+        die(f"orchestrator link {slot.iface} not present.\n"
+            f"    ./cli.py backend setup --backend=firecracker")
+
+    run_dir = _infra_dir()
+    rootfs = run_dir / "rootfs.ext4"
+    if infra_artifact.local_build_requested():
+        util.build_rootfs_ext4(build_infra_rootfs_dir(), rootfs, slack_mib=8192)
+    else:
+        # Prebuilt artifact already carries the buildah build slack; expand it
+        # to a fresh writable rootfs for this boot.
+        infra_artifact.materialize_ext4(
+            infra_artifact.infra_artifact_version(_infra_init()), rootfs)
+    private_key, pubkey = _stable_keypair()
+
+    info(f"booting infra VM on {slot.iface} (guest {slot.guest_ip})")
+    vm = firecracker_vm.boot(
+        name="bot-bottle-infra", rootfs=rootfs, tap=slot.iface,
+        guest_ip=slot.guest_ip, host_ip=slot.host_ip, pubkey=pubkey,
+        run_dir=run_dir, mem_mib=4096, detached=True,
+        data_drive=_ensure_registry_volume(),
+    )
+    _pid_file().write_text(str(vm.process.pid))
+    return InfraVm(guest_ip=slot.guest_ip, private_key=private_key, vm=vm)
+
+
+def _infra_dir() -> Path:
+    d = util.cache_dir() / "infra"
+    d.mkdir(parents=True, exist_ok=True)
+    return d
+
+
+def _pid_file() -> Path:
+    return _infra_dir() / "vm.pid"
+
+
+def _version_file() -> Path:
+    """Records the infra-artifact version the *running* VM booted from, so a
+    later launcher can tell whether the singleton it found is the current code.
+    Without it, a healthy VM built from an older image gets adopted forever and
+    the new code never boots — every infra change would need an out-of-band
+    kill to dislodge the stale VM (and races whatever launched next)."""
+    return _infra_dir() / "booted-version"
+
+
+def _expected_version() -> str:
+    return infra_artifact.infra_artifact_version(_infra_init())
+
+
+def _adoptable(key: Path, url: str, want: str) -> bool:
+    """Adopt a running infra VM only if it booted from the CURRENT version and
+    its control plane is healthy. A missing/mismatched marker means a prior
+    launcher booted an older infra image — reboot rather than reuse stale code."""
+    if not key.exists():
+        return False
+    try:
+        booted = _version_file().read_text(encoding="utf-8").strip()
+    except OSError:
+        return False
+    return booted == want and _health_ok(url)
+
+
+def _record_booted_version(version: str) -> None:
+    _version_file().write_text(version + "\n", encoding="utf-8")
+
+
+# The registry "volume": a host-side ext4 file attached to the infra VM as a
+# second virtio-block device (guest /dev/vdb), mounted at the control plane's
+# DB dir. It outlives the ephemeral rootfs, so the bottle registry survives an
+# infra-VM restart — the firecracker analogue of a docker volume. It is a
+# plain ext4 file: `sudo mount -o loop <path>` on the host (with the VM
+# stopped) to inspect bot-bottle.db directly.
+_REGISTRY_SIZE = "512M"
+
+
+def registry_volume_path() -> Path:
+    return _infra_dir() / "registry.ext4"
+
+
+def _ensure_registry_volume() -> Path:
+    """Create the empty ext4 registry volume on first use; reuse it after."""
+    vol = registry_volume_path()
+    if vol.exists():
+        return vol
+    info(f"creating infra registry volume {vol} ({_REGISTRY_SIZE})")
+    proc = subprocess.run(
+        ["mke2fs", "-q", "-t", "ext4", "-F", str(vol), _REGISTRY_SIZE],
+        capture_output=True, text=True, check=False,
+    )
+    if proc.returncode != 0:
+        vol.unlink(missing_ok=True)
+        die(f"creating registry volume failed: {proc.stderr.strip()}")
+    return vol
+
+
+def _stable_keypair() -> tuple[Path, str]:
+    """The infra VM's SSH keypair — generated once and reused, so any later
+    launcher can SSH in (fetch CA / provision) even though a different process
+    booted the VM. The pubkey is re-injected on every boot via the cmdline."""
+    d = _infra_dir()
+    key, pub = d / "id_ed25519", d / "id_ed25519.pub"
+    if key.exists() and pub.exists():
+        return key, pub.read_text().strip()
+    key.unlink(missing_ok=True)
+    pub.unlink(missing_ok=True)
+    subprocess.run(
+        ["ssh-keygen", "-t", "ed25519", "-N", "", "-q", "-f", str(key),
+         "-C", "bot-bottle-infra"],
+        check=True,
+    )
+    return key, pub.read_text().strip()
+
+
+def _kill_pidfile() -> None:
+    """SIGTERM (then SIGKILL) the recorded infra VMM, if it's still ours.
+    Guards against a recycled PID by checking the process is firecracker."""
+    try:
+        pid = int(_pid_file().read_text().strip())
+    except (OSError, ValueError):
+        return
+    try:
+        comm = Path(f"/proc/{pid}/comm").read_text().strip()
+    except OSError:
+        return  # already gone
+    if comm != "firecracker":
+        return  # PID recycled by an unrelated process
+    try:
+        os.kill(pid, signal.SIGTERM)
+        for _ in range(50):
+            if not Path(f"/proc/{pid}").exists():
+                return
+            time.sleep(0.1)
+        os.kill(pid, signal.SIGKILL)
+    except OSError:
+        pass
+
+
+def _kill_infra_firecrackers(proc_root: Path = Path("/proc")) -> None:
+    """SIGKILL any firecracker VMM whose `--config-file` is this host's infra
+    config, independent of the PID file — reaps orphans it lost track of so the
+    orchestrator TAP is free to rebind. Scoped to the infra config path, so the
+    interactive pool's agent/infra VMs (other config paths) are untouched."""
+    cfg = str(_infra_dir() / "config.json")
+    for entry in proc_root.iterdir():
+        if not entry.name.isdigit():
+            continue
+        try:
+            if (entry / "comm").read_text().strip() != "firecracker":
+                continue
+            args = (entry / "cmdline").read_bytes().split(b"\0")
+        except OSError:
+            continue  # process vanished / not ours
+        if any(a.decode("utf-8", "replace") == cfg for a in args):
+            try:
+                os.kill(int(entry.name), signal.SIGKILL)
+            except (OSError, ValueError):
+                pass
+
+
+def _health_ok(url: str) -> bool:
+    try:
+        with urllib.request.urlopen(f"{url}/health", timeout=1.0) as resp:
+            return resp.status == 200
+    except (urllib.error.URLError, TimeoutError, OSError):
+        return False
+
+
+class SshGatewayTransport:
+    """`GatewayTransport` for the gateway running in the infra VM — the docker
+    exec/cp equivalents over SSH (dropbear + the stable infra key)."""
+
+    def __init__(self, private_key: Path, guest_ip: str) -> None:
+        self._key = private_key
+        self._ip = guest_ip
+
+    def exec(self, argv: list[str]) -> None:
+        proc = subprocess.run(
+            util.ssh_base_argv(self._key, self._ip) + [shlex.join(argv)],
+            capture_output=True, text=True, timeout=60, check=False,
+        )
+        if proc.returncode != 0:
+            raise GatewayProvisionError(
+                f"infra gateway exec {argv!r} failed: {proc.stderr.strip()}")
+
+    def cp_into(self, src: str, dest: str) -> None:
+        # Preserve the source mode (docker cp does): the access-hook is staged
+        # 0700 and git-http execs it directly — a plain `cat >` would land it
+        # 0644 and the exec fails with EACCES; keys stay 0600.
+        mode = stat.S_IMODE(os.stat(src).st_mode)
+        q = shlex.quote(dest)
+        proc = subprocess.run(
+            util.ssh_base_argv(self._key, self._ip)
+            + [f"cat > {q} && chmod {mode:o} {q}"],
+            input=Path(src).read_bytes(), capture_output=True, timeout=30, check=False,
+        )
+        if proc.returncode != 0:
+            raise GatewayProvisionError(
+                f"infra gateway cp {src} -> {dest} failed: "
+                f"{proc.stderr.decode(errors='replace').strip()}")
+
+
+def gateway_transport() -> SshGatewayTransport:
+    """git-gate provisioning transport for the gateway in the infra VM, built
+    from the stable key + the orchestrator link's guest IP. Needs no live VM
+    handle, so teardown can use it too."""
+    return SshGatewayTransport(
+        _infra_dir() / "id_ed25519", netpool.orch_slot().guest_ip)
+
+
+def wait_for_health(
+    infra: InfraVm, *, timeout: float = _HEALTH_TIMEOUT_SECONDS,
+) -> None:
+    """Poll the control plane's /health until it answers 200 or the deadline
+    passes. Dies (with the console tail) if the VMM exits early."""
+    url = f"{infra.control_plane_url}/health"
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        if infra.vm is not None and not infra.vm.is_alive():
+            die(f"infra VM exited during boot (rc={infra.vm.process.returncode}).\n"
+                f"{firecracker_vm._console_tail(infra.vm.console_log)}")
+        try:
+            with urllib.request.urlopen(url, timeout=1.0) as resp:
+                if resp.status == 200:
+                    info(f"infra control plane healthy at {infra.control_plane_url}")
+                    return
+        except (urllib.error.URLError, TimeoutError, OSError):
+            pass
+        time.sleep(_HEALTH_POLL_SECONDS)
+    tail = (firecracker_vm._console_tail(infra.vm.console_log)
+            if infra.vm is not None else "")
+    die(f"infra control plane at {url} did not become healthy within "
+        f"{timeout:.0f}s.\n{tail}")
+
+
+def _infra_init() -> str:
+    """PID-1 init for the infra VM: mount the pseudo-filesystems, wire a
+    resolver, start dropbear (debug SSH), then launch the control plane and
+    the gateway data plane (multi-tenant against the local control plane)."""
+    return f"""#!/bin/sh
+# bot-bottle Firecracker infra VM init (PID 1).
+mount -t proc proc /proc 2>/dev/null
+mount -t sysfs sys /sys 2>/dev/null
+mount -t devtmpfs dev /dev 2>/dev/null
+mkdir -p /dev/pts && mount -t devpts devpts /dev/pts 2>/dev/null
+mount -o remount,rw / 2>/dev/null
+
+# Export a real PATH: a bare-init shell resolves its own execs via a
+# built-in default path, but that isn't in the *environment*, so
+# gateway_init's subprocess daemons (spawned as `python3 ...`) would
+# inherit no PATH and fail to find python3. Export it for all children.
+export PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+
+# Direct upstream resolver (control-plane / gateway egress + buildah).
+printf 'nameserver {_INFRA_RESOLVER}\\n' > /etc/resolv.conf 2>/dev/null
+
+# Debug SSH: install the per-boot pubkey from the kernel cmdline.
+KEY=$(sed -n 's/.*bb_pubkey=\\([^ ]*\\).*/\\1/p' /proc/cmdline | base64 -d 2>/dev/null)
+if [ -n "$KEY" ]; then
+    mkdir -p /root/.ssh
+    printf '%s\\n' "$KEY" > /root/.ssh/authorized_keys
+    chmod 700 /root/.ssh && chmod 600 /root/.ssh/authorized_keys
+fi
+chown -R 0:0 /root 2>/dev/null || true
+mkdir -p /etc/dropbear /run /var/lib/bot-bottle
+
+# Persistent registry volume (second virtio-block device, /dev/vdb) mounted
+# at the control plane's DB dir, so bot-bottle.db survives infra-VM restarts.
+mount -t ext4 /dev/vdb /var/lib/bot-bottle 2>/dev/null || true
+
+/bb-dropbear -R -E -p 22 &
+
+# Control plane. Source is baked at /app; the package is stdlib-only.
+cd /app
+BOT_BOTTLE_ROOT=/var/lib/bot-bottle python3 -m bot_bottle.orchestrator \\
+    --host 0.0.0.0 --port {CONTROL_PLANE_PORT} --broker stub &
+
+# Gateway data plane, multi-tenant: each request resolves source-IP ->
+# policy against the local control plane. The VM backend reaches git over
+# git-http (9420), so the git:// daemon (git-gate, needs a per-bottle
+# entrypoint the consolidated model doesn't use) is left out.
+BOT_BOTTLE_GATEWAY_DAEMONS=egress,git-http,supervise \\
+BOT_BOTTLE_ORCHESTRATOR_URL=http://127.0.0.1:{CONTROL_PLANE_PORT} \\
+SUPERVISE_DB_PATH=/var/lib/bot-bottle/db/bot-bottle.db \\
+    python3 -m bot_bottle.gateway_init &
+
+# Reap as PID 1; children are backgrounded, so `wait` blocks.
+while : ; do wait ; done
+"""
@@ -1,39 +1,284 @@
-"""Launch flow for the Firecracker backend — temporarily disabled (#385).
+"""Launch flow for the Firecracker backend (PRD 0070, consolidated).

-The firecracker backend launched a per-bottle companion container (the
-egress / git-gate / supervise data plane) alongside each microVM. That
-per-bottle-companion architecture was removed in the companion-container removal;
-firecracker's replacement — the consolidated per-host gateway — lands in
-its own cutover (#354).
+Per bottle:
+  1. build the agent rootfs in a builder VM (buildah, no host docker), or
+     resume a frozen bottle from its committed rootfs tar; cache the ext4;
+  2. ensure the per-host orchestrator + shared gateway are up;
+  3. claim a free TAP pool slot (rootless flock);
+  4. register the bottle on the orchestrator by the VM's guest IP (the
+     attribution key) and provision its git-gate state into the gateway;
+  5. boot the microVM on that TAP; wait for SSH;
+  6. provision (shared gateway CA, prompt, skills, workspace, git, supervise)
+     over SSH.

-Until that lands, launching a firecracker bottle fails closed rather than
-silently running the removed path. `prepare` / `status` / cleanup still
-work, so `backend status --backend=firecracker` and orphan cleanup are
-unaffected.
+The per-bottle Docker sidecar bundle is gone. The shared gateway handles
+egress / git-gate / supervise for every VM; Docker's PREROUTING DNAT routes
+the VMs' traffic to it, and the nft table's `ct status dnat accept` rule
+in the forward chain lets it pass. The VM still sends to `host_tap_ip:PORT`
+— the address its world is, by nft design, limited to.
+
+Isolation is enforced by the operator-provisioned nft table (checked
+fail-closed in preflight): a VM reaches only the sidecar (DNAT'd from
+the host TAP IP) and nothing else.
 """

 from __future__ import annotations

-from contextlib import contextmanager
+import dataclasses
+import os
+from contextlib import ExitStack, contextmanager
+from pathlib import Path
 from typing import Callable, Generator

-from ...log import die
+from ...agent_provider import runtime_for
+from ...bottle_state import (
+    committed_rootfs_path,
+    egress_state_dir,
+    git_gate_state_dir,
+    read_committed_image,
+)
+from ...egress import (
+    egress_agent_env_entries,
+    egress_resolve_token_values,
+)
+from ...git_gate import (
+    provision_git_gate_dynamic_keys,
+    revoke_git_gate_provisioned_keys,
+)
+from ...image_cache import check_stale_path
+from ...log import die, info, warn
+from ...supervise import SUPERVISE_PORT
+from ..docker.egress import EGRESS_PORT
+from ..util import AGENT_CA_BUNDLE, AGENT_CA_PATH
+from . import firecracker_vm, image_builder, isolation_probe, netpool, util
 from .bottle import FirecrackerBottle
 from .bottle_plan import FirecrackerBottlePlan
+from ...orchestrator.config_store import resolve_teardown_timeout
+from .consolidated_launch import (
+    launch_consolidated,
+    teardown_consolidated,
+)
+
+
+_GIT_HTTP_PORT = 9420


@contextmanager
 def launch(
    plan: FirecrackerBottlePlan,
+    agent_base: Path,
    *,
    provision: Callable[[FirecrackerBottlePlan, "FirecrackerBottle"], str | None],
 ) -> Generator[FirecrackerBottle, None, None]:
-    """Fail closed: the firecracker backend is disabled while its
-    consolidated (gateway-backed) launch is built in #354."""
-    del plan, provision
-    die(
-        "the firecracker backend is temporarily disabled during the "
-        "companion-container removal (#385); its consolidated relaunch "
-        "lands in #354. Use --backend=docker for now."
+    """Build, launch, and provision a Firecracker bottle via the consolidated
+    orchestrator. Teardown on exit."""
+    stack = ExitStack()
+    bottle_for_revoke = plan.manifest.bottle
+    git_gate_dir_for_revoke = git_gate_state_dir(plan.slug)
+
+    def teardown() -> None:
+        teardown_exc: BaseException | None = None
+        try:
+            stack.close()
+        except BaseException as exc:  # noqa: W0718 - teardown must continue
+            teardown_exc = exc
+            warn(f"firecracker teardown failed: {exc!r}")
+        revoke_git_gate_provisioned_keys(bottle_for_revoke, git_gate_dir_for_revoke)
+        if teardown_exc is not None:
+            raise teardown_exc
+
+    try:
+        # Step 1 (rootfs resolution/build) runs in BottleBackend.launch before
+        # this context starts resources. ``agent_base`` is the selected cache,
+        # fresh build, or committed snapshot.
+        # Step 2: mint the git-gate dynamic (gitea) deploy keys, if any.
+        git_gate_plan = plan.git_gate_plan
+        if git_gate_plan.upstreams:
+            git_gate_plan = provision_git_gate_dynamic_keys(
+                plan.manifest.bottle, git_gate_plan, git_gate_state_dir(plan.slug),
+            )
+
+        # Step 3: claim a TAP slot; the flock is held until teardown.
+        slot, lock = netpool.allocate(plan.slug)
+        stack.callback(lock.close)
+        info(f"firecracker slot {slot.iface}: host={slot.host_ip} "
+             f"guest={slot.guest_ip}")
+
+        # Step 4: register on the orchestrator + provision this bottle's
+        # git-gate state into the shared gateway. The per-bottle egress tokens
+        # are resolved from the host env now and handed to the orchestrator
+        # (in memory) for the gateway to inject — the agent never sees them.
+        # Attribution is by the VM's guest IP (unspoofable via /31 TAP + nft).
+        effective_env = {**os.environ, **plan.agent_provision.provisioned_env}
+        token_values = egress_resolve_token_values(
+            plan.egress_plan.token_env_map, effective_env,
+        )
+        teardown_timeout = resolve_teardown_timeout()
+        ctx = launch_consolidated(
+            plan.egress_plan, git_gate_plan,
+            guest_ip=slot.guest_ip,
+            image_ref=plan.image,
+            tokens=token_values,
+        )
+        stack.callback(
+            teardown_consolidated, ctx.bottle_id,
+            orchestrator_url=ctx.orchestrator_url,
+            timeout=teardown_timeout,
+        )
+
+        # Step 5: install the SHARED gateway CA (replaces the per-bottle CA).
+        # Write it to a stable host path so the provisioner can copy it over SSH.
+        ca_dir = egress_state_dir(plan.slug) / "gateway-ca"
+        ca_dir.mkdir(parents=True, exist_ok=True)
+        ca_file = ca_dir / "gateway-ca.pem"
+        ca_file.write_text(ctx.gateway_ca_pem)
+        egress_plan = dataclasses.replace(
+            plan.egress_plan,
+            mitmproxy_ca_host_path=ca_file,
+            mitmproxy_ca_cert_only_host_path=ca_file,
+        )
+        # Point the agent's git-gate insteadOf rewrites and supervise MCP URL
+        # at the shared gateway (reached at the slot's host TAP IP — the VM
+        # sends there and Docker DNAT routes to the gateway container).
+        git_gate_url = (
+            f"http://{slot.host_ip}:{_GIT_HTTP_PORT}" if git_gate_plan.upstreams else ""
+        )
+        supervise_url = (
+            f"http://{slot.host_ip}:{SUPERVISE_PORT}/"
+            if plan.supervise_plan is not None else ""
+        )
+        plan = dataclasses.replace(
+            plan,
+            git_gate_plan=git_gate_plan,
+            egress_plan=egress_plan,
+            identity_token=ctx.identity_token,
+            # Deliver the identity token as egress proxy credentials — clients
+            # honor `HTTPS_PROXY=http://id:token@gw` without app changes; the
+            # gateway reads Proxy-Authorization, validates the (source_ip,
+            # token) pair, and strips it before upstream.
+            agent_proxy_url=(
+                f"http://bottle:{ctx.identity_token}"
+                f"@{slot.host_ip}:{EGRESS_PORT}"
+            ),
+            agent_git_gate_url=git_gate_url,
+            agent_supervise_url=supervise_url,
+        )
+
+        # Step 6: build the per-bottle rootfs + SSH key, then boot.
+        run_dir = util.cache_dir() / "run" / plan.slug
+        run_dir.mkdir(parents=True, exist_ok=True)
+        rootfs = run_dir / "rootfs.ext4"
+        util.build_rootfs_ext4(agent_base, rootfs)
+        private_key, pubkey = util.generate_keypair(run_dir)
+
+        vm = firecracker_vm.boot(
+            name=plan.container_name,
+            rootfs=rootfs,
+            tap=slot.iface,
+            guest_ip=slot.guest_ip,
+            host_ip=slot.host_ip,
+            pubkey=pubkey,
+            run_dir=run_dir,
+        )
+        stack.callback(vm.terminate)
+        firecracker_vm.wait_for_ssh(vm, private_key)
+
+        # Authoritative fail-closed egress-boundary check, before the agent
+        # runs: prove the VM cannot reach the host directly.
+        isolation_probe.verify_isolation(private_key, slot.guest_ip)
+
+        bottle = FirecrackerBottle(
+            plan.container_name,
+            private_key=private_key,
+            guest_ip=slot.guest_ip,
+            guest_env=_agent_guest_env(plan, slot.host_ip),
+            agent_command=plan.agent_command,
+            agent_prompt_mode=plan.agent_prompt_mode,
+            agent_provider_template=plan.agent_provider_template,
+            terminal_title=(
+                f"{plan.spec.label} ({plan.spec.agent_name})"
+                if plan.spec.label else plan.spec.agent_name
+            ),
+            terminal_color=plan.spec.color,
+            agent_workdir=plan.workspace_plan.workdir,
+        )
+        bottle.prompt_path = provision(plan, bottle)
+
+        yield bottle
+    finally:
+        teardown()
+
+
+def build_or_load_agent_base(plan: FirecrackerBottlePlan) -> Path:
+    """Produce the agent's base rootfs dir. Primary path: build the Dockerfile
+    inside a Firecracker builder VM (buildah, no host docker), smoke-testing
+    the image before export. A committed snapshot (freeze/migrate) is resumed
+    directly from the rootfs tar the freezer wrote — no host docker either."""
+    committed = read_committed_image(plan.slug)
+    committed_tar = committed_rootfs_path(plan.slug)
+    if committed and committed_tar.is_file():
+        info(f"resuming from committed rootfs {committed_tar}")
+        return util.build_committed_rootfs_dir(committed_tar)
+    dockerfile = Path(plan.dockerfile_path)
+    if plan.spec.image_policy == "cached":
+        cached = image_builder.cached_agent_rootfs_dir(dockerfile)
+        if cached is None:
+            die(
+                f"cached agent rootfs for {plan.image!r} not found; "
+                "run without --cached-images to build it"
+            )
+        info(f"using cached agent rootfs {cached.name}")
+        return cached
+    return image_builder.build_agent_rootfs_dir(
+        dockerfile,
+        image_tag=plan.image,
+        smoke_test=runtime_for(plan.agent_provider_template).smoke_test,
    )
-    yield  # unreachable — `die` raises; keeps this a generator/contextmanager
+
+
+def stale_checks(plan: FirecrackerBottlePlan) -> None:
+    """Raise when the cached rootfs selected by this plan is stale."""
+    if plan.spec.image_policy != "cached":
+        return
+    committed = read_committed_image(plan.slug)
+    committed_tar = committed_rootfs_path(plan.slug)
+    if committed and committed_tar.is_file():
+        check_stale_path(f"agent rootfs {committed_tar}", committed_tar)
+        return
+    cached = image_builder.cached_agent_rootfs_dir(Path(plan.dockerfile_path))
+    if cached is not None:
+        check_stale_path(f"agent rootfs {cached}", cached / ".bb-ready")
+
+
+# --- agent guest env -------------------------------------------------
+
+def _agent_guest_env(plan: FirecrackerBottlePlan, host_ip: str) -> dict[str, str]:
+    """Env injected into every agent/exec call over SSH. The VM has no
+    baked process env (it just runs init), so the proxy/CA/git/supervise
+    wiring is applied per-invocation."""
+    # Carries the identity token as proxy credentials (set in `launch`).
+    proxy_url = plan.agent_proxy_url or f"http://{host_ip}:{EGRESS_PORT}"
+    no_proxy = f"localhost,127.0.0.1,{host_ip}"
+    env: dict[str, str] = {
+        "HTTPS_PROXY": proxy_url, "HTTP_PROXY": proxy_url,
+        "https_proxy": proxy_url, "http_proxy": proxy_url,
+        "NO_PROXY": no_proxy, "no_proxy": no_proxy,
+        "NODE_EXTRA_CA_CERTS": AGENT_CA_PATH,
+        "SSL_CERT_FILE": AGENT_CA_BUNDLE,
+        "REQUESTS_CA_BUNDLE": AGENT_CA_BUNDLE,
+    }
+    if plan.agent_git_gate_url:
+        env["GIT_GATE_URL"] = plan.agent_git_gate_url
+    if plan.agent_supervise_url:
+        env["MCP_SUPERVISE_URL"] = plan.agent_supervise_url
+    for entry in egress_agent_env_entries(plan.egress_plan):
+        key, _, value = entry.partition("=")
+        env[key] = value
+    env.update(plan.agent_provision.guest_env)
+    # Forwarded (bare-name) env: resolve host values now, since the VM
+    # can't inherit them from a `docker run --env NAME`.
+    for name in plan.forwarded_env:
+        value = os.environ.get(name)
+        if value is not None:
+            env[name] = value
+    return env
@@ -15,3 +15,11 @@ BOT_BOTTLE_FC_POOL_SIZE=8
 BOT_BOTTLE_FC_IP_BASE=10.243.0.0
 BOT_BOTTLE_FC_IFACE_PREFIX=bbfc
 BOT_BOTTLE_FC_NFT_TABLE=bot_bottle_fc
+# The orchestrator/gateway VM's own TAP — a dedicated link OUTSIDE the
+# bbfc* agent pool. Unlike agent VMs (which reach only their gateway),
+# the orchestrator is trusted infra that needs real NAT'd internet
+# egress: to FROM-pull + apt/npm during in-VM agent-image builds
+# (buildah) and to forward agent egress upstream (Stage B gateway). Its
+# /31 is the top of the IP_BASE /16 (host x.y.255.0, guest x.y.255.1),
+# clear of the pool near the bottom of the block.
+BOT_BOTTLE_FC_ORCH_IFACE=bborch0
@@ -79,6 +79,12 @@ def _cfg(key: str) -> str:
 IFACE_PREFIX = _cfg("BOT_BOTTLE_FC_IFACE_PREFIX")
 NFT_TABLE = _cfg("BOT_BOTTLE_FC_NFT_TABLE")

+# The orchestrator/gateway VM's dedicated TAP — outside the bbfc* agent
+# pool and, unlike it, NAT'd to the internet (see `orch_slot`). The
+# orchestrator is trusted infra: it builds agent images in-VM (buildah
+# needs to FROM-pull + apt/npm) and forwards agent egress upstream.
+ORCH_IFACE = _cfg("BOT_BOTTLE_FC_ORCH_IFACE")
+

 def pool_size() -> int:
    return int(_cfg("BOT_BOTTLE_FC_POOL_SIZE"))
@@ -123,6 +129,25 @@ def all_slots() -> list[Slot]:
    return [slot(i) for i in range(pool_size())]


+def orch_slot() -> Slot:
+    """The orchestrator/gateway VM's dedicated link — its own TAP
+    (`ORCH_IFACE`) on a /31 at the TOP of the IP_BASE /16 (host
+    x.y.255.0, guest x.y.255.1), well clear of the agent pool near the
+    bottom of the block. Unlike a pool `Slot`, this link is NAT'd out to
+    the internet by the setup (the orchestrator is trusted infra), so it
+    is deliberately *not* one of the isolated `bbfc*` slots.
+
+    `index` is -1 (sentinel: not a pool index)."""
+    base16 = int(ipaddress.IPv4Address(ip_base())) & 0xFFFF0000
+    host = base16 + 0xFF00
+    return Slot(
+        index=-1,
+        iface=ORCH_IFACE,
+        host_ip=str(ipaddress.IPv4Address(host)),
+        guest_ip=str(ipaddress.IPv4Address(host + 1)),
+    )
+
+
 # --- fail-closed verification ---------------------------------------

 def _run_ok(argv: list[str]) -> bool:
@@ -0,0 +1,212 @@
+"""Build the infra rootfs and publish it as a Gitea generic package.
+
+The off-host (build / CI) half of PRD 0069 Stage 2: this DOES use Docker, but
+never on the launch host. It runs the same pipeline the launch host used to run
+locally — `docker build` the three fixed images, export to a rootfs dir, inject
+the guest boot, `mke2fs` to an ext4 with the buildah build slack — then gzips
+the ext4 and PUTs it (plus a `.sha256`) to
+`…/api/packages/<owner>/generic/bot-bottle-firecracker-infra/<version>/`.
+
+The `<version>` is `infra_artifact.infra_artifact_version(...)`, the content
+hash of the rootfs inputs, so a launch host at the same code checkout resolves
+the exact artifact this produced.
+
+  python3 -m bot_bottle.backend.firecracker.publish_infra --output DIR
+  python3 -m bot_bottle.backend.firecracker.publish_infra --publish-dir DIR
+
+Auth: a token with `write:package` on the target owner, from
+`BOT_BOTTLE_INFRA_ARTIFACT_TOKEN`.
+"""
+
+from __future__ import annotations
+
+import argparse
+import gzip
+import hashlib
+import shutil
+import sys
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+from . import infra_artifact, infra_vm, util
+
+_CHUNK = 1 << 20
+
+# A human-readable description shipped alongside the artifact — generic packages
+# have no description field, so this file *is* the description on the package
+# page. Uploaded on every publish so it never goes stale.
+_ABOUT_NAME = "about.txt"
+_ABOUT_TEXT = (
+    "bot-bottle infra rootfs for the Firecracker backend (PRD 0069 Stage 2, "
+    "#348): the per-host infra VM (orchestrator control plane + gateway + "
+    "buildah). Prebuilt off-host, gzip ext4; the launch host downloads + "
+    "sha256-verifies + boots it, no host Docker. The version tag is a content "
+    "hash of the rootfs inputs. Files: rootfs.ext4.gz + rootfs.ext4.gz.sha256.\n"
+)
+
+
+def _gzip(src: Path, dest: Path) -> None:
+    with open(src, "rb") as fh, gzip.open(dest, "wb") as out:
+        shutil.copyfileobj(fh, out, _CHUNK)
+
+
+def _sha256(path: Path) -> str:
+    h = hashlib.sha256()
+    with open(path, "rb") as fh:
+        for chunk in iter(lambda: fh.read(_CHUNK), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+def _put(url: str, body: "bytes | Path", token: str) -> None:
+    """PUT `body` (raw bytes, or a Path streamed from disk) to `url`. The rootfs
+    is hundreds of MB, so it is passed as a Path and streamed — `urlopen` reads
+    the open file in blocks rather than materializing it in memory (with an
+    explicit Content-Length, which Gitea requires and which also stops urllib
+    from `len()`-ing a non-bytes body)."""
+    handle = None
+    if isinstance(body, Path):
+        length = body.stat().st_size
+        handle = open(body, "rb")
+        data: object = handle
+    else:
+        length = len(body)
+        data = body
+    req = urllib.request.Request(url, data=data, method="PUT")  # type: ignore[arg-type]
+    req.add_header("Content-Length", str(length))
+    if token:
+        req.add_header("Authorization", f"token {token}")
+    req.add_header("Content-Type", "application/octet-stream")
+    try:
+        with urllib.request.urlopen(req) as resp:
+            print(f"  uploaded {url} (HTTP {resp.status})")
+    except urllib.error.HTTPError as e:
+        if e.code == 409:
+            raise SystemExit(
+                f"artifact already published at {url} (HTTP 409); "
+                f"bump the code version or pass --force to overwrite"
+            )
+        raise SystemExit(f"upload failed (HTTP {e.code}): {url}\n{e.read().decode(errors='replace')}")
+    except urllib.error.URLError as e:
+        raise SystemExit(f"registry unreachable: {url} ({e.reason})")
+    finally:
+        if handle is not None:
+            handle.close()
+
+
+def _delete(url: str, token: str) -> None:
+    req = urllib.request.Request(url, method="DELETE")
+    if token:
+        req.add_header("Authorization", f"token {token}")
+    try:
+        with urllib.request.urlopen(req):
+            pass
+    except urllib.error.HTTPError as e:
+        if e.code != 404:
+            raise SystemExit(f"could not overwrite existing artifact (HTTP {e.code}): {url}")
+    except urllib.error.URLError as e:
+        raise SystemExit(f"registry unreachable: {url} ({e.reason})")
+
+
+def build_artifact(out_dir: Path) -> tuple[str, Path, Path]:
+    """Build the infra rootfs ext4, gzip it, and write the checksum. Returns
+    `(version, gz_path, sha_path)`. Uses host Docker (off-host / CI)."""
+    version = infra_artifact.infra_artifact_version(infra_vm._infra_init())
+    print(f"building infra rootfs artifact {version} (docker)")
+    infra_vm.build_infra_images_with_docker()
+    base = infra_vm.build_infra_rootfs_dir()
+
+    ext4 = out_dir / "rootfs.ext4"
+    util.build_rootfs_ext4(base, ext4, slack_mib=8192)
+    gz = out_dir / "rootfs.ext4.gz"
+    print("compressing rootfs")
+    _gzip(ext4, gz)
+    ext4.unlink(missing_ok=True)
+
+    sha = out_dir / "rootfs.ext4.gz.sha256"
+    digest = _sha256(gz)
+    sha.write_text(f"{digest}  rootfs.ext4.gz\n")
+    print(f"  {gz.name}: {gz.stat().st_size / 1e6:.0f} MB  sha256={digest}")
+    return version, gz, sha
+
+
+def _publish_bundle(root: Path, token: str) -> str:
+    version_file = root / "version.txt"
+    # Guard the read so a missing version.txt is a clean error, not a raw
+    # FileNotFoundError.
+    if not version_file.is_file():
+        raise SystemExit(f"incomplete artifact bundle: {root}")
+    version = version_file.read_text(encoding="utf-8").strip()
+    expected = infra_artifact.infra_artifact_version(infra_vm._infra_init())
+    if version != expected:
+        raise SystemExit(
+            f"artifact bundle version {version!r} does not match checkout {expected!r}"
+        )
+    gz = root / "rootfs.ext4.gz"
+    sha = root / "rootfs.ext4.gz.sha256"
+    if not gz.is_file() or not sha.is_file():
+        raise SystemExit(f"incomplete artifact bundle: {root}")
+    expected_sha = sha.read_text().split()[0].strip().lower()
+    if _sha256(gz) != expected_sha:
+        raise SystemExit("artifact bundle checksum mismatch")
+
+    gz_url = infra_artifact.artifact_url(version, gz.name)
+    sha_url = infra_artifact.artifact_url(version, sha.name)
+    about_url = infra_artifact.artifact_url(version, _ABOUT_NAME)
+
+    # Publishing is idempotent. If this exact complete artifact is already
+    # present, a test-only main commit is a no-op. Otherwise clear any partial
+    # upload left by an interrupted prior attempt and upload the complete set.
+    try:
+        with urllib.request.urlopen(infra_artifact._open(sha_url)) as resp:
+            remote_sha = resp.read().decode("utf-8").split()[0].strip().lower()
+    except urllib.error.HTTPError as e:
+        if e.code != 404:
+            raise SystemExit(f"checking existing artifact failed (HTTP {e.code})")
+        remote_sha = ""
+    except urllib.error.URLError as e:
+        raise SystemExit(f"registry unreachable: {sha_url} ({e.reason})")
+    if remote_sha == expected_sha:
+        print(f"infra rootfs {version} already published")
+        return version
+
+    for url in (gz_url, sha_url, about_url):
+        _delete(url, token)
+    _put(gz_url, gz, token)
+    _put(sha_url, sha.read_bytes(), token)
+    _put(about_url, _ABOUT_TEXT.encode(), token)
+    return version
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(
+        prog="publish_infra", description="Build + publish the infra rootfs artifact.")
+    mode = parser.add_mutually_exclusive_group(required=True)
+    mode.add_argument("--output", type=Path,
+                      help="build a candidate bundle in DIR without publishing")
+    mode.add_argument("--publish-dir", type=Path,
+                      help="publish an already-built and tested candidate bundle")
+    args = parser.parse_args(argv)
+
+    _, _, token = infra_artifact._config()
+    if args.publish_dir is not None and not token:
+        raise SystemExit(
+            "no publish token: set BOT_BOTTLE_INFRA_ARTIFACT_TOKEN to a token "
+            "with write:package")
+
+    if args.output is not None:
+        args.output.mkdir(parents=True, exist_ok=True)
+        version, _gz, _sha = build_artifact(args.output)
+        (args.output / "version.txt").write_text(version + "\n", encoding="utf-8")
+        print(f"built infra rootfs candidate {version}")
+        return 0
+
+    assert args.publish_dir is not None
+    version = _publish_bundle(args.publish_dir, token)
+    print(f"published infra rootfs {version}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -14,6 +14,7 @@ and `./cli.py backend setup --backend=firecracker`.

 from __future__ import annotations

+import hashlib
 import os
 import platform
 import shutil
@@ -87,7 +88,7 @@ def require_firecracker() -> None:
    booting a VM without it."""
    if not is_linux():
        die("firecracker backend is only supported on Linux (KVM). "
-            "On macOS use --backend=macos-container.")
+            "On macOS use the macos-container backend.")
    if shutil.which("firecracker") is None:
        info("Firecracker is required but was not found on PATH.")
        info("Install: https://github.com/firecracker-microvm/firecracker/releases")
@@ -159,15 +160,22 @@ def docker_image_id(ref: str) -> str:
    return result.stdout.strip().replace("sha256:", "")[:16]


-def build_base_rootfs_dir(image_ref: str) -> Path:
-    """Export the agent image's filesystem and inject the guest init +
-    static dropbear. Cached by image digest — the per-bottle bits
+def build_base_rootfs_dir(
+    image_ref: str, *, variant: str = "", init_script: str | None = None,
+) -> Path:
+    """Export the image's filesystem and inject the guest init + static
+    dropbear. Cached by image digest — the per-bottle bits
    (authorized_keys, IP) are passed at boot via the kernel cmdline, so
    this tree carries nothing bottle-specific and is safely shared.

+    `variant` suffixes the cache key so the same image can be prepared
+    with a different `init_script` (e.g. the infra VM boots the same
+    orchestrator image as the builder but runs the control plane as
+    PID 1, not the SSH-only agent init) without a cache collision.
+
    Returns the prepared directory (read as the `mke2fs -d` source)."""
    digest = docker_image_id(image_ref)
-    base = cache_dir() / "rootfs" / digest
+    base = cache_dir() / "rootfs" / f"{digest}{variant}"
    ready = base / ".bb-ready"
    if ready.is_file():
        return base
@@ -200,18 +208,85 @@ def build_base_rootfs_dir(image_ref: str) -> Path:
            stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, check=False,
        )

-    _inject_guest_boot(base)
+    inject_guest_boot(base, init_script=init_script)
    ready.write_text("ok\n")
    return base


-def _inject_guest_boot(rootfs: Path) -> None:
-    """Drop the static dropbear and the PID-1 init into the rootfs."""
-    shutil.copy2(dropbear_path(), rootfs / "bb-dropbear")
-    os.chmod(rootfs / "bb-dropbear", 0o755)
-    init = rootfs / "bb-init"
-    init.write_text(_GUEST_INIT)
-    os.chmod(init, 0o755)
+def build_committed_rootfs_dir(tar_path: Path) -> Path:
+    """Prepare a base rootfs dir from a frozen-bottle snapshot tar (the
+    freeze/resume path — no Docker). Extracts the snapshot, recreates the
+    virtual mount points the freezer excluded, and injects the guest init +
+    static dropbear, mirroring `build_base_rootfs_dir` but sourced from a tar
+    we control rather than a Docker image.
+
+    Cached under the rootfs cache, keyed by the tar's size+mtime so a
+    re-freeze re-extracts but repeated resumes of the same snapshot don't.
+    Returns the prepared directory (read as the `mke2fs -d` source)."""
+    st = tar_path.stat()
+    fingerprint = hashlib.sha256(
+        f"{tar_path}:{st.st_size}:{st.st_mtime_ns}".encode()
+    ).hexdigest()[:16]
+    base = cache_dir() / "rootfs" / f"committed-{fingerprint}"
+    ready = base / ".bb-ready"
+    if ready.is_file():
+        return base
+
+    if base.exists():
+        shutil.rmtree(base, ignore_errors=True)
+    base.mkdir(parents=True)
+
+    info(f"extracting committed rootfs {tar_path} -> {base}")
+    result = subprocess.run(
+        ["tar", "-x", "-f", str(tar_path), "-C", str(base)],
+        capture_output=True, text=True, check=False,
+    )
+    if result.returncode != 0:
+        die(f"extracting committed rootfs {tar_path} failed: "
+            f"{result.stderr.strip() or '<no stderr>'}")
+
+    # The freezer excludes the live/virtual filesystems from the snapshot;
+    # recreate them as empty mount points so the guest init can mount
+    # proc/sys/dev and dropbear has a writable /run.
+    for mount_point in ("proc", "sys", "dev", "run"):
+        (base / mount_point).mkdir(mode=0o755, exist_ok=True)
+
+    inject_guest_boot(base)
+    ready.write_text("ok\n")
+    return base
+
+
+def inject_guest_boot(rootfs: Path, init_script: str | None = None) -> None:
+    """Drop the static dropbear and the PID-1 init into the rootfs.
+    `init_script` defaults to the SSH-only agent init; the infra VM
+    passes its own (control plane + gateway) init.
+
+    A committed snapshot is guest-controlled, so `bb-dropbear`/`bb-init`
+    may already exist as symlinks aimed at a host file (e.g. bb-init ->
+    ~/.bashrc). Replace whatever is there and create the files with
+    O_EXCL|O_NOFOLLOW so the write always lands a fresh regular file in
+    the staging tree and never follows a planted symlink out of it."""
+    _write_staged_file(rootfs / "bb-dropbear", dropbear_path().read_bytes())
+    _write_staged_file(rootfs / "bb-init", (init_script or _GUEST_INIT).encode())
+
+
+def _write_staged_file(path: Path, data: bytes) -> None:
+    """Write `data` to `path` (mode 0755) as a fresh regular file inside a
+    staging rootfs, replacing any pre-existing entry without following a
+    symlink at `path`. Fails closed on anything unexpected there."""
+    if path.is_symlink() or path.exists():
+        if path.is_dir() and not path.is_symlink():
+            shutil.rmtree(path)
+        else:
+            path.unlink()
+    fd = os.open(
+        path, os.O_WRONLY | os.O_CREAT | os.O_EXCL | os.O_NOFOLLOW, 0o755
+    )
+    try:
+        os.write(fd, data)
+    finally:
+        os.close(fd)
+    os.chmod(path, 0o755)


 def build_rootfs_ext4(base_dir: Path, out_path: Path, *, slack_mib: int = 1024) -> None:
@@ -293,6 +368,11 @@ mount -t devtmpfs dev /dev 2>/dev/null
 mkdir -p /dev/pts && mount -t devpts devpts /dev/pts 2>/dev/null
 mount -o remount,rw / 2>/dev/null

+# /tmp must be world-writable + sticky. The rootless rootfs build can land
+# it 0755/root-owned, leaving the agent (uid 1000 node) unable to create
+# scratch dirs there — git worktrees, build temp, `git init /tmp/...`, etc.
+mkdir -p /tmp && chmod 1777 /tmp
+
 # Install the per-bottle SSH pubkey from the kernel cmdline.
 KEY=$(sed -n 's/.*bb_pubkey=\([^ ]*\).*/\1/p' /proc/cmdline | base64 -d 2>/dev/null)
 if [ -n "$KEY" ]; then
@@ -12,7 +12,7 @@ from ...env import ResolvedEnv
 from ...git_gate import GitGatePlan
 from ...supervise import SupervisePlan
 from ...manifest import Manifest
-from .. import ActiveAgent, BottleBackend, BottleSpec
+from .. import ActiveAgent, BottleBackend, BottleImages, BottleSpec
 from . import cleanup as _cleanup
 from . import enumerate as _enumerate
 from . import launch as _launch
@@ -82,13 +82,27 @@ class MacosContainerBottleBackend(
            stage_dir=stage_dir,
        )

+    def prelaunch_checks(self, plan: MacosContainerBottlePlan) -> None:
+        _launch.stale_checks(plan)
+
+    def _build_or_load_images(self, plan: MacosContainerBottlePlan) -> BottleImages:
+        return _launch.build_or_load_images(plan)
+
    @contextmanager
-    def launch(
-        self, plan: MacosContainerBottlePlan
+    def _launch_impl(
+        self, plan: MacosContainerBottlePlan, images: BottleImages
    ) -> Generator[MacosContainerBottle, None, None]:
-        with _launch.launch(plan, provision=self.provision) as bottle:
+        with _launch.launch(plan, images, provision=self.provision) as bottle:
            yield bottle

+    def ensure_orchestrator(self) -> str:
+        """Bring up the per-host infra container (control plane + gateway) and
+        return its control-plane URL — the on-demand entry point operator tools
+        (`supervise`) call when no control plane is running yet. Mirrors
+        firecracker's infra-VM bring-up."""
+        from .infra import MacosInfraService
+        return MacosInfraService().ensure_running().control_plane_url
+
    def prepare_cleanup(self) -> MacosContainerBottleCleanupPlan:
        return _cleanup.prepare_cleanup()

@@ -52,6 +52,7 @@ class MacosContainerBottle(Bottle):
        terminal_title: str = "",
        terminal_color: str = "",
        agent_workdir: str = "/home/node",
+        exec_env: dict[str, str] | None = None,
    ):
        self.name = container
        self._teardown = teardown
@@ -62,6 +63,20 @@ class MacosContainerBottle(Bottle):
        self.terminal_color = terminal_color
        self.agent_provider_template = agent_provider_template
        self.agent_workdir = agent_workdir
+        # Env applied to the agent process at `container exec` time, on top of
+        # what the container was run with. This is how the identity token
+        # reaches the agent (PRD 0070): registration mints it *after* the
+        # container exists — its source IP is the registration key and Apple
+        # Container assigns that by DHCP — so it cannot be in the run-time env
+        # the way docker's compose spec does it.
+        #
+        # `container exec --env` does NOT override a run-time value — it
+        # appends, leaving duplicate entries in the agent's `environ` whose
+        # resolution is runtime-specific (Node last-wins, Rust first-wins). So
+        # nothing here may rely on superseding: the proxy vars are supplied
+        # *only* at exec time and are deliberately absent from the run-time
+        # env. See `launch._agent_env_entries`.
+        self._exec_env = dict(exec_env or {})
        self._closed = False

    def agent_argv(self, argv: list[str], *, tty: bool = True) -> list[str]:
@@ -74,6 +89,12 @@ class MacosContainerBottle(Bottle):
            )
        )
        container_exec = ["container", "exec"]
+        # Bare env names, same rule as the terminal hints below: the value
+        # stays in the child env `exec_agent` builds and never reaches argv —
+        # the proxy URL here carries the identity token, which `ps` would
+        # otherwise expose to every process on the host.
+        for name in sorted(self._exec_env):
+            container_exec.extend(["--env", name])
        if tty:
            container_exec.extend(["--interactive", "--tty"])
            # Forward terminal capability hints so TUIs can enable modified-key
@@ -94,21 +115,33 @@ class MacosContainerBottle(Bottle):

    def exec_agent(self, argv: list[str], *, tty: bool = True) -> int:
        agent_argv = self.agent_argv(argv, tty=tty)
+        # The values behind the bare `--env` names in `agent_argv`. `sh -lc`
+        # below is in this process tree, so the child env reaches `container
+        # exec` either way.
+        env = {**os.environ, **self._exec_env} if self._exec_env else None
        script = (
            exec_shell_script(agent_argv, self.terminal_title, self.terminal_color)
            if tty else None
        )
        if script is None:
-            return subprocess.run(agent_argv, check=False).returncode
-        return subprocess.run(["sh", "-lc", script], check=False).returncode
+            return subprocess.run(agent_argv, env=env, check=False).returncode
+        return subprocess.run(["sh", "-lc", script], env=env, check=False).returncode

    def exec(self, script: str, *, user: str = "node") -> ExecResult:
+        # Carry the same exec env the agent gets: provisioning steps run
+        # through here, and a provider whose provision step fetches anything
+        # would egress without the identity token and be denied by /resolve.
+        # Bare `--env NAME` again, so the token stays off argv.
+        argv = ["container", "exec", "--user", user, "--interactive"]
+        for name in sorted(self._exec_env):
+            argv.extend(["--env", name])
+        argv.extend([self.name, "sh", "-s"])
        result = subprocess.run(
-            ["container", "exec", "--user", user, "--interactive",
-             self.name, "sh", "-s"],
+            argv,
            input=script,
            capture_output=True,
            text=True,
+            env={**os.environ, **self._exec_env} if self._exec_env else None,
            check=False,
        )
        return ExecResult(
@@ -13,9 +13,13 @@ from .. import BottlePlan
 class MacosContainerBottlePlan(BottlePlan):
    slug: str
    forwarded_env: dict[str, str] = field(repr=False)
-    agent_proxy_url: str = ""
    agent_git_gate_url: str = ""
    agent_supervise_url: str = ""
+    # Read by provision-time consumers (git extraHeader, supervise MCP header)
+    # via getattr(plan, "identity_token", ""); stamped in launch after the
+    # bottle is registered. See launch.py's stamp for why it lives here and not
+    # only in the exec-time proxy env.
+    identity_token: str = ""

    @property
    def container_name(self) -> str:
@@ -0,0 +1,178 @@
+"""Consolidated bottle launch sequence for the macOS backend (PRD 0070).
+
+The docker backend allocates a free address, pins the agent to it with
+`--ip`, registers it, *then* starts the agent — registration precedes launch
+because the pinned address is known up front.
+
+**Apple Container 1.0.0 has no `--ip`.** The `--network` flag takes only
+`<name>[,mac=…][,mtu=…]`; the address is assigned by vmnet's DHCP and is
+knowable only once the container is running. So the macOS order inverts:
+
+    ensure_gateway()  ->  caller starts the agent  ->  register_agent(source_ip)
+
+That is why this module exposes two functions where docker has one — the
+caller has to start the agent in between. `ensure_gateway` runs first because
+the agent's proxy env needs the gateway's address at `container run` time; the
+agent's *own* address (the attribution key) only exists afterwards.
+
+The control plane and the gateway are one **infra container** here (see
+`infra`), so `gateway_ip` and the control-plane host are the same address.
+
+The consequence for the identity token: it is minted by registration, i.e.
+*after* the agent container exists, so it cannot be baked into the run-time
+env the way docker's compose spec does. It is delivered at `container exec`
+time instead — see `bottle.MacosContainerBottle`.
+
+That delivery is load-bearing, not a nicety: `/resolve` requires a matching
+`(source_ip, identity_token)` pair and fail-closes with no source-IP-only
+fallback (#366). So egress that does not carry the token is denied — which is
+the safe direction, and is why the agent's init process is a bare `sleep` and
+every real command arrives through `container exec`.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from ...egress import EgressPlan
+from ...git_gate import GitGatePlan
+from ...log import info
+from ...orchestrator.client import OrchestratorClient, OrchestratorClientError
+from ..consolidated_util import provision_bottle, teardown_consolidated as _teardown_util
+from . import util as container_mod
+from .enumerate import CONTAINER_NAME_PREFIX, EnumerationError, enumerate_active
+from .gateway import GATEWAY_NETWORK
+from .gateway_provision import AppleGatewayTransport
+from .infra import MacosInfraService, OrchestratorStartError
+
+
+class ConsolidatedLaunchError(RuntimeError):
+    """The consolidated register/provision sequence could not complete."""
+
+
+@dataclass(frozen=True)
+class GatewayEndpoint:
+    """What the agent `container run` needs to reach the shared gateway (the
+    infra container). `gateway_ip` is that container's host-only address, the
+    same host the control-plane URL points at."""
+
+    orchestrator_url: str
+    gateway_ip: str         # the gateway's address — the agent's proxy target
+    gateway_ca_pem: str     # the shared CA the provisioner installs
+    network: str            # the shared host-only network to attach to
+
+
+@dataclass(frozen=True)
+class LaunchContext:
+    """What the running agent needs once it has been registered."""
+
+    bottle_id: str
+    identity_token: str
+    source_ip: str          # the agent's DHCP-assigned address (attribution key)
+    gateway_ip: str
+    network: str
+    orchestrator_url: str
+
+
+def ensure_gateway(
+    *, service: MacosInfraService | None = None,
+) -> GatewayEndpoint:
+    """Ensure the per-host infra container (control plane + gateway) is up and
+    report how to reach it. Idempotent — one singleton, so N bottle launches
+    share it. Call before starting the agent container: the agent's proxy env
+    needs `gateway_ip` at run time."""
+    service = service or MacosInfraService()
+    infra = service.ensure_running()
+    return GatewayEndpoint(
+        orchestrator_url=infra.control_plane_url,
+        gateway_ip=infra.gateway_ip,
+        gateway_ca_pem=service.ca_cert_pem(),
+        network=service.network,
+    )
+
+
+def live_source_ips(network: str) -> list[str]:
+    """Every running agent container's address on `network`.
+
+    The reconciliation input: the orchestrator lives inside the infra
+    container and cannot enumerate the host's containers, so the host has to
+    tell it which bottles are actually up. Containers that have not been
+    assigned an address yet contribute nothing — the reap's grace window, not
+    this list, is what protects an in-flight launch.
+
+    Raises `EnumerationError` when the live set cannot be determined
+    authoritatively: either the container listing fails or any individual
+    inspect fails. Callers must skip reconciliation in that case to avoid
+    unregistering healthy bottles."""
+    ips: list[str] = []
+    for agent in enumerate_active():
+        name = f"{CONTAINER_NAME_PREFIX}{agent.slug}"
+        ip = container_mod.inspect_container_network_ip(name, network)
+        if ip is None:
+            raise EnumerationError(
+                f"container inspect {name!r} failed; live set is not authoritative"
+            )
+        if ip:
+            ips.append(ip)
+    return ips
+
+
+def register_agent(
+    egress_plan: EgressPlan,
+    git_gate_plan: GitGatePlan,
+    *,
+    source_ip: str,
+    endpoint: GatewayEndpoint,
+    image_ref: str = "",
+    tokens: dict[str, str] | None = None,
+) -> LaunchContext:
+    """Register the (already running) agent by its address and provision its
+    git-gate state into the gateway. `source_ip` must be read from the live
+    container — it is the attribution key the gateway resolves policy by.
+    Raises on failure; the caller tears down."""
+    client = OrchestratorClient(endpoint.orchestrator_url)
+    # Self-heal before registering: a launcher that died hard (SIGKILL, closed
+    # terminal, host sleep) never ran its teardown callback, leaving an active
+    # row with no container. vmnet recycles addresses, so such a row can
+    # collide with this bottle's — and `by_source_ip` fail-closes on ambiguity,
+    # which would resolve no policy at all and deny every host. Best-effort: a
+    # reconciliation failure must not block an otherwise-fine launch.
+    try:
+        client.reconcile(live_source_ips(endpoint.network))
+    except (OrchestratorClientError, EnumerationError) as e:
+        info(f"registry reconciliation skipped: {e}")
+    reg = provision_bottle(
+        client, source_ip, egress_plan, git_gate_plan, AppleGatewayTransport(),
+        image_ref=image_ref, tokens=tokens,
+    )
+    return LaunchContext(
+        bottle_id=reg.bottle_id,
+        identity_token=reg.identity_token,
+        source_ip=source_ip,
+        gateway_ip=endpoint.gateway_ip,
+        network=endpoint.network,
+        orchestrator_url=endpoint.orchestrator_url,
+    )
+
+
+def teardown_consolidated(
+    bottle_id: str, *, orchestrator_url: str, timeout: float | None = None,
+) -> None:
+    """Deregister the bottle and remove its git-gate state from the gateway.
+    Both steps are idempotent so this is safe from a cleanup trap. Does NOT
+    stop the gateway — it's a persistent per-host singleton."""
+    _teardown_util(bottle_id, AppleGatewayTransport(),
+                   orchestrator_url=orchestrator_url, timeout=timeout)
+
+
+__all__ = [
+    "GatewayEndpoint",
+    "LaunchContext",
+    "ensure_gateway",
+    "live_source_ips",
+    "register_agent",
+    "teardown_consolidated",
+    "ConsolidatedLaunchError",
+    "OrchestratorStartError",
+    "GATEWAY_NETWORK",
+]
@@ -1,9 +1,11 @@
 """Host-side egress route-apply for the macos-container backend.

 The per-bottle companion container this used to signal (`container kill
--signal HUP <container>`) was removed in the companion-container removal (#385),
-along with the disabled macOS launch path. Fails closed until the macOS
-backend grows the consolidated gateway.
+--signal HUP <container>`) was removed in the companion-container removal
+(#385). In the consolidated model the shared gateway resolves egress policy
+per-request against the orchestrator rather than reloading a per-bottle routes
+file, so the live per-bottle reload is not supported here and fails closed
+until the gateway-side apply lands — same posture as the docker backend.
 """

 from __future__ import annotations
@@ -16,8 +18,8 @@ class MacOSContainerEgressApplicator(EgressApplicator):
        del slug
        raise EgressApplyError(
            "live egress route-apply was removed with the per-bottle "
-            "companion container (#385); the macos-container backend is "
-            "disabled until it uses the consolidated gateway."
+            "companion container (#385); route changes will flow through "
+            "the consolidated gateway in a follow-up."
        )


@@ -1,14 +1,53 @@
-"""Active-agent enumeration for the macOS Apple Container backend.
-
-The backend is disabled during the companion-container removal (#385) — it can't
-launch bottles, so there are none to enumerate. Enumeration returns when
-the backend grows the consolidated gateway.
-"""
+"""Active-agent enumeration for the macOS Apple Container backend."""

 from __future__ import annotations

+import subprocess
+
+from ...bottle_state import read_metadata
 from .. import ActiveAgent
+from .infra import INFRA_NAME
+
+# The name every agent container carries: `bot-bottle-<slug>`. Exported
+# because callers that act on a running bottle (gateway-host rewrites,
+# registry reconciliation) have to map an enumerated slug back to a
+# container name.
+CONTAINER_NAME_PREFIX = "bot-bottle-"
+# The shared per-host infra container carries the same prefix as agent
+# containers but is infrastructure, not a bottle — one control plane + gateway
+# serves every agent, so listing it as an agent would invent one per host.
+_INFRA_NAMES = frozenset({INFRA_NAME})
+
+
+class EnumerationError(RuntimeError):
+    """container list failed; the resulting live set is not authoritative."""


 def enumerate_active() -> list[ActiveAgent]:
-    return []
+    result = subprocess.run(
+        ["container", "list", "--quiet"],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        raise EnumerationError(
+            f"container list failed: "
+            f"{(result.stderr or '').strip() or '<no stderr>'}"
+        )
+    out: list[ActiveAgent] = []
+    for name in sorted(line.strip() for line in result.stdout.splitlines()):
+        if not name.startswith(CONTAINER_NAME_PREFIX) or name in _INFRA_NAMES:
+            continue
+        slug = name[len(CONTAINER_NAME_PREFIX):]
+        metadata = read_metadata(slug)
+        out.append(ActiveAgent(
+            backend_name="macos-container",
+            slug=slug,
+            agent_name=metadata.agent_name if metadata else "?",
+            started_at=metadata.started_at if metadata else "",
+            services=(),
+            label=metadata.label if metadata else "",
+            color=metadata.color if metadata else "",
+        ))
+    return out
@@ -0,0 +1,46 @@
+"""Shared network/image constants for the macOS consolidated infra container.
+
+The gateway data plane no longer runs as its own Apple container — it shares a
+single per-host **infra container** with the control plane (see `infra`),
+because two Apple-Container guests writing one `bot-bottle.db` over virtiofs
+would race incoherent `fcntl` locks. This module holds the pieces both the
+infra service and the launch/provision glue need: the network names, the
+gateway image, and the network-creation helper.
+"""
+
+from __future__ import annotations
+
+import os
+
+from ...orchestrator.gateway import GatewayError
+from . import util as container_mod
+
+# The shared host-only network the infra container and every agent bottle sit
+# on. The agent's address here is the attribution key. Distinct from the docker
+# names so both backends can coexist on one host.
+GATEWAY_NETWORK = "bot-bottle-mac-gateway"
+# The NAT network that gives the infra container (and only it) a route out.
+GATEWAY_EGRESS_NETWORK = "bot-bottle-mac-egress"
+
+GATEWAY_IMAGE = os.environ.get("BOT_BOTTLE_GATEWAY_IMAGE", "bot-bottle-gateway:latest")
+
+DEFAULT_CA_TIMEOUT_SECONDS = 30.0
+
+
+def ensure_networks(
+    network: str = GATEWAY_NETWORK, egress_network: str = GATEWAY_EGRESS_NETWORK,
+) -> None:
+    """Create the shared host-only network + the NAT egress network. Idempotent
+    — `create_network` tolerates 'already exists'."""
+    container_mod.create_network(egress_network)
+    container_mod.create_network(network, internal=True)
+
+
+__all__ = [
+    "GATEWAY_NETWORK",
+    "GATEWAY_EGRESS_NETWORK",
+    "GATEWAY_IMAGE",
+    "GatewayError",
+    "DEFAULT_CA_TIMEOUT_SECONDS",
+    "ensure_networks",
+]
@@ -0,0 +1,96 @@
+"""Stable gateway name for macOS agents, via each bottle's `/etc/hosts`.
+
+The shared gateway's address is assigned by vmnet's DHCP and changes whenever
+the infra container is recreated — a source-hash bump, an image upgrade, a
+crash. Every agent-facing URL (egress proxy, git-http, supervise) embeds that
+address, and the proxy URL reaches the agent as **process environment** at
+`container exec` time. A running process's `environ` cannot be rewritten from
+outside, so a moved gateway used to strand every running bottle permanently:
+not degraded, unreachable, until the bottle was relaunched and its session
+thrown away.
+
+So the agent never learns the address. It is given a stable *name*
+(`GATEWAY_HOSTNAME`) in every URL, resolved through its own `/etc/hosts`.
+Unlike `environ`, that is a file — it can be rewritten inside a container that
+is already running, so a gateway that comes back at a new address is picked up
+by live bottles instead of orphaning them.
+
+Apple Container 1.0 offers no container-name DNS on a user network (the only
+nameserver an agent sees is vmnet's, which does not know container names) and
+`container run` has no `--add-host`, so the entry is written by exec after the
+container starts.
+
+Writing it needs root, and the agent runs as `node`: the agent therefore
+cannot repoint its own gateway name, while the host (which drives `container
+exec --user root`) can. That asymmetry is deliberate — keep it.
+"""
+
+from __future__ import annotations
+
+from ...log import warn
+from . import util as container_mod
+from .enumerate import CONTAINER_NAME_PREFIX, enumerate_active
+
+# The name every agent-facing gateway URL uses. Must not collide with a real
+# DNS name the agent might resolve; it is bottle-local by construction.
+GATEWAY_HOSTNAME = "bot-bottle-gateway"
+
+# Marker so the rewrite is idempotent and only ever touches our own line —
+# the rest of /etc/hosts (localhost, the container's own name) is preserved.
+_MARKER = "# bot-bottle gateway"
+
+
+def _rewrite_script(gateway_ip: str) -> str:
+    """A shell one-liner that replaces our managed line in `/etc/hosts`.
+
+    Rewrites in place via a temp file + `cat` rather than `mv`, so the file
+    keeps its original inode, ownership, and mode — a bind-mounted or
+    pre-created `/etc/hosts` must not be replaced by a root-owned 0644 copy
+    that the runtime then refuses to update.
+    """
+    return (
+        "set -e; "
+        f"grep -v '{_MARKER}' /etc/hosts > /tmp/.bb-hosts || true; "
+        f"printf '%s %s %s\\n' '{gateway_ip}' '{GATEWAY_HOSTNAME}' "
+        f"'{_MARKER}' >> /tmp/.bb-hosts; "
+        "cat /tmp/.bb-hosts > /etc/hosts; "
+        "rm -f /tmp/.bb-hosts"
+    )
+
+
+def set_gateway_host(container_name: str, gateway_ip: str) -> None:
+    """Point `GATEWAY_HOSTNAME` at `gateway_ip` inside one running container.
+
+    Must run before the agent is exec'd: the agent's proxy URL names the
+    gateway, so the entry has to exist for its first connection. Idempotent —
+    re-running with the same address is a no-op in effect.
+    """
+    container_mod.exec_container_as_root(
+        container_name, ["sh", "-c", _rewrite_script(gateway_ip)],
+    )
+
+
+def refresh_gateway_host(gateway_ip: str) -> list[str]:
+    """Re-point every running bottle at the current gateway address.
+
+    Called once the shared gateway is known to be up, so a bottle stranded by
+    an earlier gateway restart re-attaches instead of needing a relaunch.
+    Returns the containers updated.
+
+    Best-effort per bottle: one container that refuses the write (already
+    exiting, say) must not stop the others from being repaired, and must not
+    fail the launch that triggered the sweep.
+    """
+    updated: list[str] = []
+    for agent in enumerate_active():
+        name = f"{CONTAINER_NAME_PREFIX}{agent.slug}"
+        try:
+            set_gateway_host(name, gateway_ip)
+            updated.append(name)
+        # One bad bottle must not stop the sweep, so this is deliberately broad.
+        except Exception as e:  # noqa: BLE001  # pylint: disable=broad-exception-caught
+            warn(f"could not re-point {name} at the gateway: {e}")
+    return updated
+
+
+__all__ = ["GATEWAY_HOSTNAME", "set_gateway_host", "refresh_gateway_host"]
@@ -0,0 +1,44 @@
+"""`GatewayTransport` for the Apple infra container (PRD 0070).
+
+The provisioning *logic* (per-bottle creds dirs, namespaced repo init) is
+backend-neutral and lives in `backend.docker.gateway_provision`; this is only
+the transport — how files and commands reach the running gateway. Docker uses
+`docker exec`/`docker cp` and Firecracker uses SSH; Apple uses the `container`
+CLI's equivalents against the infra container that hosts the gateway daemons.
+"""
+
+from __future__ import annotations
+
+from ..docker.gateway_provision import GatewayProvisionError
+from . import util as container_mod
+from .infra import INFRA_NAME
+
+
+class AppleGatewayTransport:
+    """`GatewayTransport` for the gateway daemons in the Apple infra container."""
+
+    def __init__(self, gateway: str = INFRA_NAME) -> None:
+        self.gateway = gateway
+
+    def exec(self, argv: list[str]) -> None:
+        result = container_mod.run_container_argv(
+            ["container", "exec", self.gateway, *argv]
+        )
+        if result.returncode != 0:
+            raise GatewayProvisionError(
+                f"gateway exec {argv!r} failed: "
+                f"{(result.stderr or '').strip() or '<no stderr>'}"
+            )
+
+    def cp_into(self, src: str, dest: str) -> None:
+        result = container_mod.run_container_argv(
+            ["container", "cp", src, f"{self.gateway}:{dest}"]
+        )
+        if result.returncode != 0:
+            raise GatewayProvisionError(
+                f"gateway cp {src} -> {dest} failed: "
+                f"{(result.stderr or '').strip() or '<no stderr>'}"
+            )
+
+
+__all__ = ["AppleGatewayTransport", "GatewayProvisionError"]
@@ -0,0 +1,304 @@
+"""The per-host infra container for the macOS backend (PRD 0070).
+
+A single persistent Apple container that runs BOTH the orchestrator control
+plane and the gateway data plane — the macOS analogue of the Firecracker infra
+VM (`backend/firecracker/infra_vm.py`), not the docker backend's two separate
+containers.
+
+Why one container, not two: Apple Containers are lightweight VMs, each with its
+own kernel. The docker backend runs the orchestrator and gateway as two
+containers safely because they share the host kernel, so their concurrent
+writes to the one `bot-bottle.db` (the orchestrator's registry + the gateway
+supervise daemon's queue) are serialized by coherent `fcntl` locks. Across two
+*guest* kernels sharing a virtiofs-mounted DB those locks are not coherent, and
+concurrent writers can corrupt the file. Firecracker solved this by putting
+both services in one guest with the DB on a device only that guest mounts; this
+does the same with Apple primitives.
+
+Two consequences fall out of the single container, both simplifications:
+
+- **No DNS dance.** The control plane and the gateway daemons reach each other
+  over `127.0.0.1`, so nothing depends on Apple's (absent) container DNS and
+  there is no orchestrator-before-gateway ordering to get right.
+- **The DB is never host-shared.** It lives on a container-only volume, so no
+  host process opens the live file. The host CLI reaches registry + supervise
+  state through the control-plane HTTP surface (`cli/supervise.py` already uses
+  `OrchestratorClient`), exactly as it does for firecracker.
+
+The control-plane source is bind-mounted (like the docker orchestrator), so a
+code change takes effect on the next launch without an image rebuild; the
+gateway daemons are baked in the gateway image and rebuild through its own
+digest check.
+"""
+
+from __future__ import annotations
+
+import os
+import time
+import urllib.error
+import urllib.request
+from dataclasses import dataclass
+from pathlib import Path
+
+from ... import log
+from ...orchestrator.gateway import GATEWAY_CA_CERT
+from ...orchestrator.lifecycle import (
+    DEFAULT_PORT,
+    DEFAULT_STARTUP_TIMEOUT_SECONDS,
+    OrchestratorStartError,
+    source_hash,
+)
+from ...paths import (
+    CONTROL_PLANE_TOKEN_ENV,
+    HOST_DB_FILENAME,
+    host_control_plane_token,
+)
+from .. import util as backend_util
+from . import util as container_mod
+from .gateway import (
+    DEFAULT_CA_TIMEOUT_SECONDS,
+    GATEWAY_EGRESS_NETWORK,
+    GATEWAY_IMAGE,
+    GATEWAY_NETWORK,
+    GatewayError,
+    ensure_networks,
+)
+
+# The one per-host infra container: control plane + gateway data plane.
+INFRA_NAME = "bot-bottle-mac-infra"
+INFRA_LABEL = "bot-bottle-mac-infra=1"
+# Container-only volume holding bot-bottle.db. No host bind-mount, so the DB is
+# written by exactly one kernel (this container's). Survives recreation.
+INFRA_DB_VOLUME = "bot-bottle-mac-db"
+
+# BOT_BOTTLE_ROOT inside the container; host_db_path() resolves the DB to
+# <root>/db/<filename> and the supervise daemon writes the same file.
+_DB_ROOT_IN_CONTAINER = "/var/lib/bot-bottle"
+_DB_PATH_IN_CONTAINER = f"{_DB_ROOT_IN_CONTAINER}/db/{HOST_DB_FILENAME}"
+_SRC_IN_CONTAINER = "/bot-bottle-src"
+
+_REPO_ROOT = Path(__file__).resolve().parents[3]
+
+_HEALTH_POLL_SECONDS = 0.25
+_HEALTH_REQUEST_TIMEOUT_SECONDS = 1.0
+_CA_POLL_SECONDS = 0.5
+
+# The gateway subset the consolidated model runs (no per-bottle git:// daemon).
+_GATEWAY_DAEMONS = "egress,git-http,supervise"
+
+
+def _init_script(port: int) -> str:
+    """PID-1 init: start the control plane and the gateway daemons, both in
+    this container, reaching each other over loopback. Backgrounded so `wait`
+    reaps as PID 1. No `set -e` — a transient daemon failure must not kill the
+    whole container (gateway_init applies the same 'stay up' policy)."""
+    return (
+        "export PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin\n"
+        f"mkdir -p $(dirname {_DB_PATH_IN_CONTAINER})\n"
+        # Control plane, from the bind-mounted source (stdlib-only package).
+        f"( cd {_SRC_IN_CONTAINER} && BOT_BOTTLE_ROOT={_DB_ROOT_IN_CONTAINER} "
+        f"python3 -m bot_bottle.orchestrator --host 0.0.0.0 --port {port} "
+        "--broker stub ) &\n"
+        # Gateway data plane, multi-tenant against the local control plane.
+        f"( cd /app && BOT_BOTTLE_GATEWAY_DAEMONS={_GATEWAY_DAEMONS} "
+        f"BOT_BOTTLE_ORCHESTRATOR_URL=http://127.0.0.1:{port} "
+        f"SUPERVISE_DB_PATH={_DB_PATH_IN_CONTAINER} python3 -m bot_bottle.gateway_init ) &\n"
+        "while : ; do wait ; done\n"
+    )
+
+
+@dataclass(frozen=True)
+class InfraEndpoint:
+    """How to reach the running infra container. The control plane and the
+    gateway are the same container, so one address serves both."""
+
+    control_plane_url: str   # http://<infra ip>:8099 — host CLI + registration
+    gateway_ip: str          # same container; agents' proxy / git-http / MCP target
+
+
+class MacosInfraService:
+    """Manages the single per-host infra container. Callers use
+    `ensure_running()` (returns the endpoint) and `ca_cert_pem()`."""
+
+    def __init__(
+        self,
+        *,
+        port: int = DEFAULT_PORT,
+        network: str = GATEWAY_NETWORK,
+        egress_network: str = GATEWAY_EGRESS_NETWORK,
+        image: str = GATEWAY_IMAGE,
+        repo_root: Path = _REPO_ROOT,
+        name: str = INFRA_NAME,
+        db_volume: str = INFRA_DB_VOLUME,
+    ) -> None:
+        self.port = port
+        self.network = network
+        self.egress_network = egress_network
+        self.image = image
+        self._repo_root = repo_root
+        self._name = name
+        self._db_volume = db_volume
+
+    def _resolve_url(self) -> str:
+        """The control-plane URL, or "" while the container has no address."""
+        ip = container_mod.try_container_ipv4_on_network(self._name, self.network)
+        return f"http://{ip}:{self.port}" if ip else ""
+
+    def is_healthy(
+        self, url: str, *, timeout: float = _HEALTH_REQUEST_TIMEOUT_SECONDS,
+    ) -> bool:
+        if not url:
+            return False
+        try:
+            with urllib.request.urlopen(f"{url}/health", timeout=timeout) as resp:
+                return resp.status == 200
+        except (urllib.error.URLError, TimeoutError, OSError):
+            return False
+
+    def _source_current(self, current_hash: str) -> bool:
+        """True iff the running infra container was created from the current
+        bind-mounted control-plane source. The control-plane process loads that
+        code at startup and won't reload it, so a stale container keeps serving
+        OLD code."""
+        if not container_mod.container_is_running(self._name):
+            return False
+        env = container_mod.container_env(self._name)
+        if not env:
+            return True  # can't compare → don't churn a working container
+        return env.get("BOT_BOTTLE_SOURCE_HASH") == current_hash
+
+    def _running_healthy_endpoint(self, current_hash: str) -> InfraEndpoint | None:
+        """The endpoint if the running container is BOTH source-current and
+        answering /health, else None (→ recreate). Health, not just the source
+        label, is what lets a wedged-but-current container self-heal instead of
+        being polled to death forever."""
+        if not self._source_current(current_hash):
+            return None
+        url = self._resolve_url()
+        if url and self.is_healthy(url):
+            return InfraEndpoint(control_plane_url=url, gateway_ip=_ip_of(url))
+        return None
+
+    def ensure_built(self) -> None:
+        """Ensure the gateway data-plane image exists. The control-plane source
+        is bind-mounted, not baked, so only the gateway image needs building."""
+        container_mod.build_image(
+            self.image, str(self._repo_root), dockerfile="Dockerfile.gateway",
+        )
+
+    def ensure_running(
+        self, *, startup_timeout: float = DEFAULT_STARTUP_TIMEOUT_SECONDS,
+    ) -> InfraEndpoint:
+        """Ensure the single infra container is up; return how to reach it.
+        Idempotent per-host singleton — a healthy container on current source
+        is left untouched, so N launches share the one control plane + gateway.
+        Raises `OrchestratorStartError` on startup timeout."""
+        current_hash = source_hash(self._repo_root)
+        endpoint = self._running_healthy_endpoint(current_hash)
+        if endpoint is not None:
+            return endpoint
+        self.ensure_built()
+        log.info("starting infra container", context={"name": self._name})
+        self._run_container(current_hash)
+        return self._wait_healthy(startup_timeout)
+
+    def _run_container(self, current_hash: str) -> None:
+        ensure_networks(self.network, self.egress_network)
+        container_mod.force_remove_container(self._name)
+        argv = [
+            "container", "run", "--detach",
+            "--name", self._name,
+            "--label", "bot-bottle.backend=macos-container",
+            "--label", INFRA_LABEL,
+            # NAT network FIRST so the gateway's egress has a default route;
+            # the host-only network is where agents (and the host CLI) reach it.
+            "--network", self.egress_network,
+            "--network", self.network,
+            "--dns", container_mod.dns_server(),
+            # Container-only DB volume: one kernel writes bot-bottle.db, never
+            # shared with the host or another guest.
+            "--volume", f"{self._db_volume}:{_DB_ROOT_IN_CONTAINER}",
+            # Bind-mount the control-plane source (read-only); a code change
+            # takes effect on relaunch with no image rebuild.
+            "--mount",
+            container_mod.bind_mount_spec(
+                str(self._repo_root), _SRC_IN_CONTAINER, readonly=True),
+            # Baked onto the container so `_source_current` can detect a real
+            # control-plane code change and recreate.
+            "--env", f"BOT_BOTTLE_SOURCE_HASH={current_hash}",
+            # The control-plane secret, for BOTH the control plane (to require
+            # it) and the gateway's PolicyResolver (to present it) — they share
+            # this one container. Bare `--env NAME` inherits the value from the
+            # run process below, so the secret never lands on argv or in
+            # `container inspect`'s command line. The agent runs in a SEPARATE
+            # container that is never given this var, which is the whole point.
+            "--env", CONTROL_PLANE_TOKEN_ENV,
+            "--entrypoint", "sh",
+            self.image,
+            "-c", _init_script(self.port),
+        ]
+        run_env = {**os.environ, CONTROL_PLANE_TOKEN_ENV: host_control_plane_token()}
+        result = container_mod.run_container_argv(argv, env=run_env)
+        if result.returncode != 0:
+            raise OrchestratorStartError(
+                f"infra container failed to start: "
+                f"{(result.stderr or '').strip() or '<no stderr>'}"
+            )
+
+    def _wait_healthy(self, startup_timeout: float) -> InfraEndpoint:
+        deadline = time.monotonic() + startup_timeout
+        while True:
+            url = self._resolve_url()
+            if url and self.is_healthy(url):
+                log.info("infra container healthy", context={"url": url})
+                return InfraEndpoint(control_plane_url=url, gateway_ip=_ip_of(url))
+            if time.monotonic() >= deadline:
+                raise OrchestratorStartError(
+                    f"infra container did not become healthy within "
+                    f"{startup_timeout:g}s"
+                )
+            time.sleep(_HEALTH_POLL_SECONDS)
+
+    def ca_cert_pem(self, *, timeout: float = DEFAULT_CA_TIMEOUT_SECONDS) -> str:
+        """The gateway's mitmproxy CA (PEM) agents install to trust its TLS
+        interception. Read out of the container (the CA lives on a
+        container-internal path, not a host mount); polls because mitmproxy
+        writes it a beat after start."""
+        def _fetch() -> str | None:
+            result = container_mod.run_container_argv(
+                ["container", "exec", self._name, "cat", GATEWAY_CA_CERT])
+            return result.stdout if result.returncode == 0 and result.stdout.strip() else None
+        try:
+            return backend_util.poll_ca_cert(_fetch, timeout=timeout)
+        except TimeoutError as exc:
+            raise GatewayError(
+                f"gateway CA not available in {self._name} after {timeout:g}s"
+            ) from exc
+
+    def stop(self) -> None:
+        """Remove the infra container (idempotent). The DB volume persists."""
+        container_mod.force_remove_container(self._name)
+
+
+def _ip_of(url: str) -> str:
+    """The host from an http://host:port URL."""
+    return url.split("://", 1)[-1].rsplit(":", 1)[0]
+
+
+def probe_control_plane_url(port: int = DEFAULT_PORT) -> str:
+    """The running infra container's control-plane URL, or "" if it isn't up.
+    Used by host-side control-plane discovery (`discover_orchestrator_url`);
+    safe to call on any host — returns "" when the container or the `container`
+    CLI isn't present."""
+    ip = container_mod.try_container_ipv4_on_network(INFRA_NAME, GATEWAY_NETWORK)
+    return f"http://{ip}:{port}" if ip else ""
+
+
+__all__ = [
+    "MacosInfraService",
+    "InfraEndpoint",
+    "OrchestratorStartError",
+    "GatewayError",
+    "INFRA_NAME",
+    "INFRA_DB_VOLUME",
+]
@@ -1,39 +1,419 @@
-"""Launch flow for the macOS Apple Container backend — disabled (#385).
+"""Launch flow for the macOS Apple Container backend (PRD 0070).

-This backend launched a per-bottle companion container (the egress /
-git-gate / supervise data plane) alongside the agent container, with the
-agent's proxy env pointed at the companion's host-only IP. That
-per-bottle-companion architecture was removed in the companion-container removal;
-the macOS backend will be re-enabled once it grows the consolidated
-per-host gateway the docker backend already uses.
+The agent container attaches to the **shared host-only gateway network** and
+proxies egress through the one per-host gateway, replacing the per-bottle
+companion container removed in #385.

-Until then, launching a macOS bottle fails closed. `prepare` / `status`
-/ cleanup still work.
+The order differs from docker's, forced by Apple Container 1.0.0 having no
+`--ip` (see `consolidated_launch`): the agent is started *before* it is
+registered, because its DHCP-assigned address — the attribution key — does not
+exist until then.
+
+    gateway up -> run agent -> read its IP -> register it -> provision
+
+Two things follow from that inversion:
+
+- The **identity token** is minted by registration and so cannot be in the
+  agent's run-time env; it rides the proxy URL applied at `container exec`
+  time (`bottle.MacosContainerBottle`). `/resolve` requires it (#366), so
+  egress without it is denied — hence the bare `sleep` init: every real agent
+  command goes through exec and therefore carries the token.
+- The agent is run with `--cap-drop CAP_NET_RAW`. Apple Container grants
+  NET_RAW by default, which would let an agent open a raw socket and forge a
+  neighbour's source address on the shared segment. NET_ADMIN is already
+  absent (the agent cannot change its own address or route), so dropping
+  NET_RAW is what closes the source-address half of PRD 0070's invariant:
+  "a packet's source address, as seen by the orchestrator, provably identifies
+  the originating bottle." The identity token is the other half — an attacker
+  would need to forge the address *and* steal the token — but the invariant is
+  a stated precondition of consolidation, so it is enforced on its own terms
+  rather than left to the token.
 """

 from __future__ import annotations

-from contextlib import contextmanager
+import dataclasses
+import os
+import subprocess
+from contextlib import ExitStack, contextmanager
+from pathlib import Path
 from typing import Callable, Generator

-from ...log import die
+from ...bottle_state import (
+    egress_state_dir,
+    git_gate_state_dir,
+    read_committed_image,
+)
+from ...egress import (
+    egress_agent_env_entries,
+    egress_resolve_token_values,
+)
+from ...git_gate import (
+    provision_git_gate_dynamic_keys,
+    revoke_git_gate_provisioned_keys,
+)
+from ...git_http_backend import DEFAULT_PORT as _GIT_HTTP_PORT
+from ...image_cache import check_stale
+from ...log import die, info, warn
+from .. import BottleImages
+from ...supervise import SUPERVISE_PORT
+from ..docker.egress import EGRESS_PORT
+from ..util import AGENT_CA_BUNDLE, AGENT_CA_PATH
+from . import util as container_mod
 from .bottle import MacosContainerBottle
+from .gateway_hosts import (
+    GATEWAY_HOSTNAME,
+    refresh_gateway_host,
+    set_gateway_host,
+)
 from .bottle_plan import MacosContainerBottlePlan
+from ...orchestrator.config_store import resolve_teardown_timeout
+from .consolidated_launch import (
+    GatewayEndpoint,
+    ensure_gateway,
+    register_agent,
+    teardown_consolidated,
+)
+
+_REPO_DIR = str(Path(__file__).resolve().parent.parent.parent.parent)
+_AGENT_SLEEP_SECONDS = "2147483647"
+
+
+def build_or_load_images(plan: MacosContainerBottlePlan) -> BottleImages:
+    """Resolve the agent image ref for this plan. The gateway's own image is
+    built by `ensure_gateway` — it belongs to the shared singleton."""
+    committed = read_committed_image(plan.slug)
+    if committed and container_mod.image_exists(committed):
+        info(f"using committed image {committed!r}")
+        return BottleImages(agent=committed)
+    if plan.spec.image_policy == "cached":
+        if not container_mod.image_exists(plan.image):
+            die(
+                f"cached agent image {plan.image!r} not found; "
+                "run without --cached-images to build it"
+            )
+        info(f"using cached agent image {plan.image!r}")
+        return BottleImages(agent=plan.image)
+    container_mod.build_image(plan.image, _REPO_DIR, dockerfile=plan.dockerfile_path)
+    return BottleImages(agent=plan.image)


@contextmanager
 def launch(
    plan: MacosContainerBottlePlan,
+    images: BottleImages,
    *,
    provision: Callable[[MacosContainerBottlePlan, "MacosContainerBottle"], str | None],
 ) -> Generator[MacosContainerBottle, None, None]:
-    """Fail closed: the macOS backend is disabled until it grows the
-    consolidated per-host gateway (the companion-container path it used
-    was removed in #385)."""
-    del plan, provision
-    die(
-        "the macos-container backend is temporarily disabled during the "
-        "companion-container removal (#385); it will return once it uses "
-        "the consolidated gateway. Use --backend=docker for now."
+    """Run, register, provision, and yield an Apple Container bottle on the
+    shared per-host gateway."""
+    stack = ExitStack()
+    bottle_for_revoke = plan.manifest.bottle
+    git_gate_dir_for_revoke = git_gate_state_dir(plan.slug)
+
+    plan = dataclasses.replace(
+        plan,
+        agent_provision=dataclasses.replace(plan.agent_provision, image=str(images.agent)),
    )
-    yield  # unreachable — `die` raises; keeps this a generator/contextmanager
+
+    def teardown() -> None:
+        teardown_exc: BaseException | None = None
+        try:
+            stack.close()
+        except BaseException as exc:  # noqa: W0718 - teardown must continue
+            teardown_exc = exc
+            warn(f"macos-container teardown failed: {exc!r}")
+        revoke_git_gate_provisioned_keys(bottle_for_revoke, git_gate_dir_for_revoke)
+        if teardown_exc is not None:
+            raise teardown_exc
+
+    try:
+        # Step 1: the per-host singletons. Must precede the agent run — its
+        # proxy env needs the gateway's address at `container run` time.
+        endpoint = ensure_gateway()
+        # The gateway's address may have changed since these bottles launched
+        # (any infra recreate re-runs DHCP). They name the gateway rather than
+        # address it, so re-pointing /etc/hosts re-attaches them in place
+        # instead of leaving them stranded until relaunch.
+        refresh_gateway_host(endpoint.gateway_ip)
+
+        # Step 2: mint this bottle's deploy keys, then point it at the SHARED
+        # gateway's CA + git-http/supervise ports.
+        plan = _provision_git_gate_keys(plan)
+        plan = _install_gateway_ca(plan, endpoint)
+        plan = _stamp_agent_urls(plan, endpoint)
+
+        # Step 3: run the agent. It has no identity token yet — registration
+        # needs the address this run assigns.
+        container_mod.force_remove_container(plan.container_name)
+        _start_agent(plan, endpoint)
+        stack.callback(container_mod.force_remove_container, plan.container_name)
+
+        # Step 4: read the assigned address and register by it. This is the
+        # attribution key; `--cap-drop CAP_NET_RAW` at run is what makes it
+        # unforgeable. Poll: `container run --detach` can return before vmnet's
+        # DHCP has assigned the address.
+        # Resolve the gateway name before anything execs: every agent-facing
+        # URL uses it, so the entry must exist for the first connection.
+        set_gateway_host(plan.container_name, endpoint.gateway_ip)
+        source_ip = container_mod.wait_container_ipv4_on_network(
+            plan.container_name, endpoint.network,
+        )
+        if not source_ip:
+            die(
+                f"agent {plan.container_name} never got an address on "
+                f"{endpoint.network}"
+            )
+        effective_env = {**os.environ, **plan.agent_provision.provisioned_env}
+        token_values = egress_resolve_token_values(
+            plan.egress_plan.token_env_map, effective_env,
+        )
+        teardown_timeout = resolve_teardown_timeout()
+        ctx = register_agent(
+            plan.egress_plan,
+            plan.git_gate_plan,
+            source_ip=source_ip,
+            endpoint=endpoint,
+            image_ref=plan.image,
+            tokens=token_values,
+        )
+        stack.callback(
+            teardown_consolidated, ctx.bottle_id,
+            orchestrator_url=ctx.orchestrator_url,
+            timeout=teardown_timeout,
+        )
+        info(
+            f"agent {plan.container_name} registered "
+            f"(gateway {endpoint.gateway_ip}, ip {source_ip})"
+        )
+
+        # Stamp the token onto the plan so provision-time consumers can read it,
+        # not only the exec-time egress proxy. git-gate's gitconfig extraHeader
+        # and the supervise MCP --header both reach the gateway on NO_PROXY (they
+        # bypass the egress proxy that carries the token), so without this the
+        # gateway's /resolve fail-closes and every git fetch/push and supervise
+        # call from the bottle is denied. Registration already produced the
+        # token above, so — unlike the run-time env — the plan CAN carry it.
+        plan = dataclasses.replace(plan, identity_token=ctx.identity_token)
+
+        bottle = MacosContainerBottle(
+            plan.container_name,
+            teardown,
+            None,
+            agent_command=plan.agent_command,
+            agent_prompt_mode=plan.agent_prompt_mode,
+            agent_provider_template=plan.agent_provider_template,
+            terminal_title=(
+                f"{plan.spec.label} ({plan.spec.agent_name})"
+                if plan.spec.label else plan.spec.agent_name
+            ),
+            terminal_color=plan.spec.color,
+            agent_workdir=plan.workspace_plan.workdir,
+            exec_env=_identity_proxy_env(endpoint, ctx.identity_token),
+        )
+        bottle.prompt_path = provision(plan, bottle)
+
+        yield bottle
+    finally:
+        teardown()
+
+
+
+def stale_checks(plan: MacosContainerBottlePlan) -> None:
+    """Raise StaleImageError if a cached image is older than the configured
+    threshold. Only runs when image_policy is 'cached'. Called by the backend
+    class's _image_stale_checks before _launch_impl starts any resources."""
+    if plan.spec.image_policy != "cached":
+        return
+    committed = read_committed_image(plan.slug)
+    if committed and container_mod.image_exists(committed):
+        ts = container_mod.image_created_at(committed)
+        if ts is not None:
+            check_stale(f"agent image {committed!r}", ts)
+        return
+    if container_mod.image_exists(plan.image):
+        ts = container_mod.image_created_at(plan.image)
+        if ts is not None:
+            check_stale(f"agent image {plan.image!r}", ts)
+
+
+def _provision_git_gate_keys(
+    plan: MacosContainerBottlePlan,
+) -> MacosContainerBottlePlan:
+    if not plan.git_gate_plan.upstreams:
+        return plan
+    git_gate_plan = provision_git_gate_dynamic_keys(
+        plan.manifest.bottle,
+        plan.git_gate_plan,
+        git_gate_state_dir(plan.slug),
+    )
+    return dataclasses.replace(plan, git_gate_plan=git_gate_plan)
+
+
+def _install_gateway_ca(
+    plan: MacosContainerBottlePlan, endpoint: GatewayEndpoint,
+) -> MacosContainerBottlePlan:
+    """Stage the SHARED gateway's CA for the provisioner to install, replacing
+    the per-bottle CA the companion container used to mint. Every bottle on
+    this host trusts this one CA."""
+    ca_dir = egress_state_dir(plan.slug) / "gateway-ca"
+    ca_dir.mkdir(parents=True, exist_ok=True)
+    ca_file = ca_dir / "gateway-ca.pem"
+    ca_file.write_text(endpoint.gateway_ca_pem)
+    egress_plan = dataclasses.replace(
+        plan.egress_plan,
+        mitmproxy_ca_host_path=ca_file,
+        mitmproxy_ca_cert_only_host_path=ca_file,
+    )
+    return dataclasses.replace(plan, egress_plan=egress_plan)
+
+
+def _stamp_agent_urls(
+    plan: MacosContainerBottlePlan, endpoint: GatewayEndpoint,
+) -> MacosContainerBottlePlan:
+    """Point the agent's git-gate insteadOf rewrites + supervise MCP at the
+    shared gateway's ports. Both bypass the egress proxy (NO_PROXY covers the
+    gateway name).
+
+    Addressed by `GATEWAY_HOSTNAME`, never by IP: these URLs are baked into
+    the agent's gitconfig and MCP config at provision time, so an address here
+    would strand the bottle the moment the gateway moved. The name is resolved
+    per connection through `/etc/hosts`, which stays rewritable while the
+    bottle runs."""
+    del endpoint  # addressed by name; the address reaches the bottle via /etc/hosts
+    git_gate_url = (
+        f"http://{GATEWAY_HOSTNAME}:{_GIT_HTTP_PORT}"
+        if plan.git_gate_plan.upstreams else ""
+    )
+    supervise_url = (
+        f"http://{GATEWAY_HOSTNAME}:{SUPERVISE_PORT}/"
+        if plan.supervise_plan is not None else ""
+    )
+    return dataclasses.replace(
+        plan,
+        agent_git_gate_url=git_gate_url,
+        agent_supervise_url=supervise_url,
+    )
+
+
+def _proxy_url(identity_token: str = "") -> str:
+    """The agent's egress proxy URL. The identity token rides as proxy
+    credentials — the gateway reads Proxy-Authorization, resolves the
+    (source_ip, token) pair against the control plane, and strips it before
+    upstream. Without a valid pair `/resolve` denies the request (#366).
+
+    Names the gateway rather than addressing it: this URL reaches the agent as
+    process environment, which cannot be rewritten once the agent is running,
+    so an address baked here is unfixable if the gateway moves."""
+    cred = f"bottle:{identity_token}@" if identity_token else ""
+    return f"http://{cred}{GATEWAY_HOSTNAME}:{EGRESS_PORT}"
+
+
+def _no_proxy() -> str:
+    # git-http + supervise live on the gateway and must NOT go through the
+    # egress proxy — the agent reaches them directly by name. Deliberately
+    # address-free: NO_PROXY is baked into the run-time env and is therefore
+    # just as unfixable as the proxy URL if the gateway moves.
+    return f"localhost,127.0.0.1,{GATEWAY_HOSTNAME}"
+
+
+def _identity_proxy_env(
+    endpoint: GatewayEndpoint, identity_token: str,
+) -> dict[str, str]:
+    """The token-bearing proxy env applied at `container exec` — the only way
+    to get the token in, since it does not exist until after the container
+    runs (registration keys on the DHCP-assigned address).
+
+    This is the *sole* source of `*_PROXY` for the agent. It deliberately does
+    not rely on overriding a run-time value: `container exec --env` appends
+    rather than replaces, so a run-time `HTTPS_PROXY` would survive alongside
+    this one and first-wins runtimes would read the wrong entry. See
+    `_agent_env_entries`."""
+    if not identity_token:
+        return {}
+    del endpoint  # the gateway is named, not addressed
+    url = _proxy_url(identity_token)
+    return {
+        "HTTPS_PROXY": url, "HTTP_PROXY": url,
+        "https_proxy": url, "http_proxy": url,
+    }
+
+
+def _start_agent(plan: MacosContainerBottlePlan, endpoint: GatewayEndpoint) -> None:
+    argv = _agent_run_argv(plan, endpoint)
+    env = {**os.environ, **plan.forwarded_env}
+    info(f"container run agent {plan.container_name}")
+    result = subprocess.run(
+        argv, capture_output=True, text=True, env=env, check=False,
+    )
+    if result.returncode != 0:
+        die(
+            f"container run for agent {plan.container_name} failed: "
+            f"{(result.stderr or '').strip() or '<no stderr>'}"
+        )
+
+
+def _agent_run_argv(
+    plan: MacosContainerBottlePlan, endpoint: GatewayEndpoint,
+) -> list[str]:
+    argv = [
+        "container", "run",
+        "--name", plan.container_name,
+        "--detach",
+        "--label", "bot-bottle.backend=macos-container",
+        "--network", endpoint.network,
+        # The attribution invariant: without NET_RAW the agent cannot open a
+        # raw socket, so it cannot source-IP-spoof its neighbours on the shared
+        # segment. NET_ADMIN is not granted by default, so its address and
+        # route are already fixed. See the module docstring.
+        "--cap-drop", "CAP_NET_RAW",
+    ]
+    for entry in _agent_env_entries(plan, endpoint):
+        argv += ["--env", entry]
+    # The init process is a no-op: every agent command arrives via
+    # `container exec`, which is also how the identity token gets in.
+    argv += [plan.image, "sleep", _AGENT_SLEEP_SECONDS]
+    return argv
+
+
+def _agent_env_entries(
+    plan: MacosContainerBottlePlan, endpoint: GatewayEndpoint,
+) -> tuple[str, ...]:
+    # No `*_PROXY` here on purpose. The token-bearing URL is applied at
+    # `container exec` (`_identity_proxy_env`), and Apple's `container exec
+    # --env` **appends** to the run-time environment rather than replacing it:
+    # setting a token-less value here leaves two `HTTPS_PROXY` entries in the
+    # agent's `environ`, token-less first. Which one a runtime reads is then
+    # pure luck — Node takes the last (and worked), Rust's `std::env::var`
+    # takes the first, so Codex proxied without its identity token and
+    # `/resolve` fail-closed on every request.
+    #
+    # A token-less proxy URL has no legitimate consumer anyway: the init
+    # process is `sleep` and everything that egresses arrives via exec. Its
+    # only value was a tidy 403 for unattributed callers, which is not worth
+    # silently dropping attribution for. Without it a process that egresses
+    # before the exec-time env still fails closed — the agent network is
+    # host-only, so there is no route off it except the gateway.
+    no_proxy = _no_proxy()
+    env = [
+        f"NO_PROXY={no_proxy}",
+        f"no_proxy={no_proxy}",
+        f"NODE_EXTRA_CA_CERTS={AGENT_CA_PATH}",
+        f"SSL_CERT_FILE={AGENT_CA_BUNDLE}",
+        f"REQUESTS_CA_BUNDLE={AGENT_CA_BUNDLE}",
+    ]
+    if plan.agent_git_gate_url:
+        env.append(f"GIT_GATE_URL={plan.agent_git_gate_url}")
+    if plan.agent_supervise_url:
+        env.append(f"MCP_SUPERVISE_URL={plan.agent_supervise_url}")
+    for name, value in sorted(plan.agent_provision.guest_env.items()):
+        env.append(f"{name}={value}")
+    # Forwarded vars: bare name → inherits from the `container run` process env
+    # so the secret value never lands on argv.
+    for name in sorted(plan.forwarded_env.keys()):
+        env.append(name)
+    env.extend(egress_agent_env_entries(plan.egress_plan))
+    return tuple(env)
+
+
+__all__ = ["launch"]
@@ -10,6 +10,7 @@ import shutil
 import subprocess
 import tempfile
 import time
+from datetime import datetime, timezone
 from typing import Iterable

 from ...log import die, info
@@ -60,13 +61,21 @@ def dns_server() -> str:


 def build_image(ref: str, context: str, *, dockerfile: str = "") -> None:
-    """Build an OCI image with Apple's BuildKit-backed `container build`."""
+    """Build an OCI image with Apple's BuildKit-backed `container build`.
+
+    Set `BOT_BOTTLE_NO_CACHE=1` (the `start --no-cache` flag) to force
+    `--no-cache`. The npm/curl installers some provider Dockerfiles
+    shell out to can silently no-op on a transient network failure —
+    e.g. an `optionalDependencies` fetch for a platform-native binary —
+    and the builder will then cache that broken layer indefinitely."""
    info(
        f"building image {ref} from {context} with Apple Container "
        "(layer cache keeps repeat builds fast)"
    )
    _ensure_builder_dns()
    args = [_CONTAINER, "build", "-t", ref, "--dns", dns_server()]
+    if os.environ.get("BOT_BOTTLE_NO_CACHE") == "1":
+        args.append("--no-cache")
    if dockerfile:
        # `container build` resolves -f relative to the current working
        # directory, not the build context. Anchor a relative Dockerfile to
@@ -78,6 +87,28 @@ def build_image(ref: str, context: str, *, dockerfile: str = "") -> None:
    subprocess.run(args, check=True)


+def verify_agent_image(image: str, argv: tuple[str, ...]) -> None:
+    """Run `argv` inside a throwaway container of a freshly built agent
+    image and die loudly if it fails, instead of shipping an image
+    whose CLI only breaks at first real use. No-op when the provider
+    hasn't declared a smoke test (`AgentProviderRuntime.smoke_test`)."""
+    if not argv:
+        return
+    result = subprocess.run(
+        [_CONTAINER, "run", "--rm", "--entrypoint", argv[0], image, *argv[1:]],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        detail = (result.stderr or result.stdout or "").strip()
+        die(
+            f"agent image {image!r} failed its post-build smoke test "
+            f"({' '.join(argv)}): {detail}\n"
+            f"Try rebuilding from scratch: bot-bottle start --no-cache"
+        )
+
+
 def commit_container(container_name: str, image_tag: str) -> None:
    """Snapshot a running Apple Container as a local image.

@@ -330,6 +361,21 @@ def exec_container(name: str, argv: list[str]) -> None:
        )


+def exec_container_as_root(name: str, argv: list[str]) -> None:
+    """`exec_container`, but as uid 0 inside the container.
+
+    For host-driven maintenance the agent itself must not be able to perform —
+    rewriting `/etc/hosts` to point the gateway name at an address. The agent
+    runs as `node`, so it cannot repoint its own gateway; the host can.
+    """
+    result = _run_container_op([_CONTAINER, "exec", "--user", "root", name, *argv])
+    if result.returncode != 0:
+        die(
+            f"container exec (root) in {name} failed: "
+            f"{(result.stderr or '').strip() or '<no stderr>'}"
+        )
+
+
 def _run_container_op(cmd: list[str]) -> subprocess.CompletedProcess[str]:
    result = subprocess.run(
        cmd,
@@ -407,22 +453,180 @@ def inspect_container(name: str) -> dict[str, object]:


 def container_ipv4_on_network(name: str, network: str) -> str:
-    data = inspect_container(name)
+    """The container's IPv4 address on `network`. Fatal if absent — callers
+    that can tolerate "not yet" want `try_container_ipv4_on_network`."""
+    ip = try_container_ipv4_on_network(name, network)
+    if not ip:
+        die(f"container {name} has no IPv4 address on {network}")
+    return ip
+
+
+def run_container_argv(
+    argv: list[str], *, env: dict[str, str] | None = None,
+) -> subprocess.CompletedProcess[str]:
+    """Run a `container` command, returning the result for the caller to
+    interpret. Unlike the `die`-on-failure helpers above, this lets callers
+    that raise their own typed errors (the gateway / orchestrator lifecycle)
+    keep control of the failure path.
+
+    `env` sets the child process environment — used to hand a secret to a bare
+    `--env NAME` flag (Apple's "just key → inherit from host" form) so the
+    value is inherited from this process, never written onto argv or into
+    `container inspect`'s recorded command line."""
+    return subprocess.run(
+        argv, capture_output=True, text=True, check=False, env=env)
+
+
+def bind_mount_spec(source: str, target: str, *, readonly: bool = False) -> str:
+    """A `container run --mount` bind spec. One definition so the gateway and
+    orchestrator emit an identical string — a divergence here would silently
+    break one backend's mounts while the other kept working."""
+    spec = f"type=bind,source={source},target={target}"
+    if readonly:
+        spec += ",readonly"
+    return spec
+
+
+def _normalize_digest(value: str) -> str:
+    return value.split(":", 1)[1] if ":" in value else value
+
+
+def _inspect_first(argv: list[str]) -> dict[str, object]:
+    """Run an inspect command and return its first JSON object, or {} on any
+    failure (non-zero exit, malformed JSON, unexpected shape). {} is the shared
+    'don't know' signal all the non-fatal inspect readers below build on — a
+    caller comparing against it treats it as 'leave the working container
+    alone', never as a mismatch."""
+    result = run_container_argv(argv)
+    if result.returncode != 0:
+        return {}
+    try:
+        data = json.loads(result.stdout or "[]")
+    except json.JSONDecodeError:
+        return {}
+    if isinstance(data, list):
+        data = data[0] if data else {}
+    return data if isinstance(data, dict) else {}
+
+
+def _descriptor_digest(node: object) -> str:
+    """The normalized digest under a `{... "descriptor": {"digest": ...}}`
+    node, or "". Both the image and container inspect shapes nest the image's
+    identity this way, so the digest readers stay symmetric — a difference
+    between them is what would spuriously recreate a container."""
+    if not isinstance(node, dict):
+        return ""
+    descriptor = node.get("descriptor")
+    if isinstance(descriptor, dict) and descriptor.get("digest"):
+        return _normalize_digest(str(descriptor["digest"]))
+    return ""
+
+
+def image_digest(ref: str) -> str:
+    """The digest of image `ref`, or "" if it can't be read. Reads exactly the
+    field `container_image_digest` reads (`configuration.descriptor.digest`) so
+    the two are comparable; "" means 'don't know' → callers don't churn."""
+    data = _inspect_first([_CONTAINER, "image", "inspect", ref])
+    return _descriptor_digest(data.get("configuration"))
+
+
+def container_image_digest(name: str) -> str:
+    """The digest of the image container `name` was created from, or "" if it
+    can't be read. Compare with `image_digest(ref)` to tell whether a running
+    container predates an image rebuild."""
+    config = _inspect_first([_CONTAINER, "inspect", name]).get("configuration")
+    image = config.get("image") if isinstance(config, dict) else None
+    return _descriptor_digest(image)
+
+
+def container_env(name: str) -> dict[str, str]:
+    """The env container `name` was started with, or {} if unreadable. Lets a
+    caller tell whether a running container's baked-in configuration still
+    matches what it would pass today."""
+    config = _inspect_first([_CONTAINER, "inspect", name]).get("configuration")
+    init = config.get("initProcess") if isinstance(config, dict) else None
+    entries = init.get("environment") if isinstance(init, dict) else None
+    if not isinstance(entries, list):
+        return {}
+    env: dict[str, str] = {}
+    for entry in entries:
+        if isinstance(entry, str) and "=" in entry:
+            key, value = entry.split("=", 1)
+            env[key] = value
+    return env
+
+
+def try_container_ipv4_on_network(name: str, network: str) -> str:
+    """`container_ipv4_on_network` without the fatal exit: "" when the address
+    isn't readable yet. For pollers — a container is created before it has an
+    address, so "not yet" is an expected state there, not an error."""
+    status = _inspect_first([_CONTAINER, "inspect", name]).get("status")
+    networks = status.get("networks") if isinstance(status, dict) else None
+    if not isinstance(networks, list):
+        return ""
+    for entry in networks:
+        if not isinstance(entry, dict) or entry.get("network") != network:
+            continue
+        raw = entry.get("ipv4Address")
+        if isinstance(raw, str) and raw:
+            return raw.split("/", 1)[0]
+    return ""
+
+
+def inspect_container_network_ip(name: str, network: str) -> str | None:
+    """IP of `name` on `network`, distinguishing inspect failure from "not yet".
+
+    Returns:
+      - the IP string when the container has one on `network`
+      - "" when inspect succeeds but no address is assigned yet (in-flight DHCP)
+      - None when the inspect command itself fails (authoritative list impossible)
+    """
+    result = subprocess.run(
+        [_CONTAINER, "inspect", name],
+        capture_output=True, text=True, check=False,
+    )
+    if result.returncode != 0:
+        return None
+    try:
+        data = json.loads(result.stdout or "[]")
+    except json.JSONDecodeError:
+        return None
+    if isinstance(data, list):
+        data = data[0] if data else {}
+    if not isinstance(data, dict):
+        return None
    status = data.get("status")
    networks = status.get("networks") if isinstance(status, dict) else None
    if not isinstance(networks, list):
-        die(f"container inspect {name} did not include status.networks")
+        return ""
    for entry in networks:
-        if not isinstance(entry, dict):
-            continue
-        if entry.get("network") != network:
+        if not isinstance(entry, dict) or entry.get("network") != network:
            continue
        raw = entry.get("ipv4Address")
-        if not isinstance(raw, str) or not raw:
-            die(f"container {name} has no IPv4 address on {network}")
-        return raw.split("/", 1)[0]
-    die(f"container {name} is not attached to network {network}")
-    raise AssertionError("unreachable")
+        if isinstance(raw, str) and raw:
+            return raw.split("/", 1)[0]
+    return ""
+
+
+def wait_container_ipv4_on_network(
+    name: str, network: str, *, timeout: float = 15.0, poll: float = 0.25,
+) -> str:
+    """Poll for the container's DHCP-assigned address on `network`, returning
+    it once available or "" on timeout.
+
+    Apple Container has no `--ip`: `container run --detach` can return before
+    vmnet's DHCP has populated `status.networks[].ipv4Address`, so a bare read
+    right after start races the assignment. Callers that need the address (the
+    attribution key, the gateway's proxy target) poll through here instead of
+    the fatal `container_ipv4_on_network`."""
+    deadline = time.monotonic() + timeout
+    while True:
+        ip = try_container_ipv4_on_network(name, network)
+        if ip:
+            return ip
+        if time.monotonic() >= deadline:
+            return ""
+        time.sleep(poll)


 def image_id(ref: str) -> str:
@@ -458,6 +662,39 @@ def image_id(ref: str) -> str:
    raise AssertionError("unreachable")


+def image_created_at(ref: str) -> datetime | None:
+    """Return the image creation timestamp as an aware UTC datetime, or None
+    when the field is absent or unparseable (e.g. FROM-scratch images, images
+    pulled from registries that omit the field). Callers should skip the stale
+    check when None is returned rather than treating it as an error."""
+    result = subprocess.run(
+        [_CONTAINER, "image", "inspect", ref],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        die(
+            f"container image inspect for {ref!r} failed: "
+            f"{(result.stderr or '').strip() or '<no stderr>'}"
+        )
+    try:
+        data = json.loads(result.stdout or "{}")
+    except json.JSONDecodeError as exc:
+        die(f"container image inspect for {ref!r} returned malformed JSON: {exc}")
+    if isinstance(data, list) and data:
+        data = data[0]
+    if isinstance(data, dict):
+        value = data.get("created") or data.get("Created")
+        if isinstance(value, str) and value:
+            try:
+                ts = value.rstrip("Z")
+                return datetime.fromisoformat(ts).replace(tzinfo=timezone.utc)
+            except ValueError:
+                pass
+    return None
+
+
 def save(ref: str, output: str) -> None:
    subprocess.run([_CONTAINER, "image", "save", ref, "-o", output], check=True)

@@ -7,6 +7,8 @@ from __future__ import annotations
 import hashlib
 import os
 import ssl
+import time
+from collections.abc import Callable
 from pathlib import Path
 from typing import TYPE_CHECKING

@@ -15,6 +17,24 @@ from ..log import die, info
 if TYPE_CHECKING:
    from ..egress import EgressPlan

+_CA_POLL_INTERVAL = 0.5
+
+
+def poll_ca_cert(fetch: Callable[[], str | None], *, timeout: float) -> str:
+    """Poll `fetch` until it returns a non-empty PEM string or `timeout` expires.
+
+    `fetch` should return the PEM on success and `None` (or empty string) when
+    the cert is not yet available. Raises `TimeoutError` if the cert never
+    appears within `timeout` seconds."""
+    deadline = time.monotonic() + timeout
+    while True:
+        result = fetch()
+        if result:
+            return result
+        if time.monotonic() >= deadline:
+            raise TimeoutError(f"CA cert not available after {timeout:g}s")
+        time.sleep(_CA_POLL_INTERVAL)
+

 # Debian-family CA layout, shared by every backend (all guest images
 # are Debian-family). AGENT_CA_PATH is the source path that
@@ -44,6 +44,7 @@ from .paths import bot_bottle_root
 _STATE_SUBDIR = "state"
 _PER_BOTTLE_DOCKERFILE_NAME = "Dockerfile"
 _COMMITTED_IMAGE_NAME = "committed-image"
+_COMMITTED_ROOTFS_NAME = "committed-rootfs.tar"
 _TRANSCRIPT_SUBDIR = "transcript"
 # Per-daemon scratch subdirs. PRD 0018 chunk 2: bind-mount sources
 # live here so chunk 3's `docker compose up` can find them at stable
@@ -200,6 +201,15 @@ def committed_image_path(identity: str) -> Path:
    return bottle_state_dir(identity) / _COMMITTED_IMAGE_NAME


+def committed_rootfs_path(identity: str) -> Path:
+    """Where the Firecracker freezer stores a snapshot of the bottle's
+    guest rootfs (a plain tar). This is the resumable/migratable artifact
+    the Firecracker backend boots from — no Docker image involved. The
+    matching `committed-image` state file records that a snapshot exists
+    (and its path); `resume` boots from this tar when both are present."""
+    return bottle_state_dir(identity) / _COMMITTED_ROOTFS_NAME
+
+
 def write_committed_image(identity: str, image_tag: str) -> Path:
    """Persist the committed image tag for `identity`. The next
    `cli.py resume <identity>` will boot from this image instead of
@@ -354,6 +364,7 @@ __all__ = [
    "cleanup_state",
    "clear_preserve_marker",
    "committed_image_path",
+    "committed_rootfs_path",
    "egress_state_dir",
    "git_gate_state_dir",
    "is_preserved",
@@ -38,6 +38,13 @@ COMMANDS = {
    "supervise": cmd_supervise,
 }

+# Commands that manage host prerequisites (or are otherwise store-free) and
+# must run before — or without — a migrated DB. `backend` provisions/probes
+# the host (TAP pool, /dev/kvm, firecracker) and never opens the store, so
+# gating it on the schema breaks preflight on a fresh CI runner where stdin
+# isn't a TTY and the migration prompt can't be answered.
+NO_MIGRATION_COMMANDS = frozenset({"backend"})
+

 def usage() -> None:
    sys.stderr.write(f"usage: {PROG} <command> [args...]\n\n")
@@ -80,7 +87,7 @@ def main(argv: list[str] | None = None) -> int:
        usage()
        die(f"unknown command: {command}")
    mgr = StoreManager.instance()
-    if not mgr.is_migrated():
+    if command not in NO_MIGRATION_COMMANDS and not mgr.is_migrated():
        sys.stderr.write("bot-bottle: database schema is out of date\n")
        sys.stderr.write("Migrate now? [y/N] ")
        sys.stderr.flush()
@@ -3,18 +3,10 @@
 from __future__ import annotations

 import os
-import sys
 from pathlib import Path

+from ..util import read_tty_line as read_tty_line
+
 PROG = "cli.py"
 USER_CWD = os.getcwd()
 REPO_DIR = str(Path(__file__).resolve().parent.parent.parent)
-
-
-def read_tty_line() -> str:
-    """Mirror `IFS= read -r REPLY </dev/tty`. Falls back to stdin."""
-    try:
-        with open("/dev/tty", "r", encoding="utf-8") as tty:
-            return tty.readline().rstrip("\n")
-    except OSError:
-        return sys.stdin.readline().rstrip("\n")
@@ -21,16 +21,20 @@ from __future__ import annotations

 import sys

-from ..backend import get_bottle_backend, known_backend_names
+from ..backend import get_bottle_backend, has_backend, known_backend_names
 from ..log import info
 from ._common import read_tty_line


 def cmd_cleanup(_argv: list[str]) -> int:
    # Order: stable backend iteration so the y/N output is
-    # deterministic across runs.
+    # deterministic across runs. Skip backends whose runtime
+    # isn't available on this host so e.g. macos-container
+    # doesn't error on Linux.
    plans = [
-        (name, get_bottle_backend(name)) for name in known_backend_names()
+        (name, get_bottle_backend(name))
+        for name in known_backend_names()
+        if has_backend(name)
    ]
    prepared = [(name, b, b.prepare_cleanup()) for name, b in plans]

@@ -27,7 +27,6 @@ from ..backend import (
    BottleSpec,
    enumerate_active_agents,
    get_bottle_backend,
-    known_backend_names,
 )
 from ..backend.docker import util as docker_mod
 from ..backend.docker.bottle_plan import DockerBottlePlan
@@ -36,6 +35,7 @@ from ..bottle_state import (
    is_preserved,
    mark_preserved,
 )
+from ..image_cache import StaleImageError
 from ..log import info, die
 from ..manifest import Manifest, ManifestIndex
 from ._common import PROG, USER_CWD, read_tty_line
@@ -47,12 +47,14 @@ def cmd_start(argv: list[str]) -> int:
    parser.add_argument("--dry-run", action="store_true")
    parser.add_argument("--cwd", action="store_true", help="copy host cwd into the running bottle")
    parser.add_argument(
-        "--backend",
-        choices=known_backend_names(),
-        default=None,
+        "--no-cache",
+        action="store_true",
        help=(
-            "backend to launch the bottle on (default: $BOT_BOTTLE_BACKEND "
-            "or host auto-selection). Overrides the env var when set."
+            "rebuild agent/sidecar images from scratch, bypassing the "
+            "build layer cache. Use when an image looks broken after a "
+            "dependency bump — e.g. an installer's optionalDependencies "
+            "fetch silently no-op'd on a transient failure and got baked "
+            "into a cached layer."
        ),
    )
    parser.add_argument(
@@ -63,6 +65,14 @@ def cmd_start(argv: list[str]) -> int:
            "skip all prompts. For orchestrators, CI, and webhooks."
        ),
    )
+    parser.add_argument(
+        "--cached-images",
+        action="store_true",
+        help=(
+            "quickstart with existing local agent and sidecar images; "
+            "only valid with --headless"
+        ),
+    )
    parser.add_argument(
        "--bottle",
        action="append",
@@ -95,15 +105,21 @@ def cmd_start(argv: list[str]) -> int:
        help="agent name defined in bot-bottle.json (omit to pick interactively)",
    )
    args = parser.parse_args(argv)
+    if args.cached_images and not args.headless:
+        die("--cached-images is only supported with --headless")

    dry_run = args.dry_run or os.environ.get("BOT_BOTTLE_DRY_RUN") == "1"
+    if args.no_cache or os.environ.get("BOT_BOTTLE_NO_CACHE") == "1":
+        # Read by build_image() in each backend's util module — set here
+        # so both the interactive and --headless paths pick it up without
+        # threading a no_cache field through every backend's plan dataclass.
+        os.environ["BOT_BOTTLE_NO_CACHE"] = "1"

    manifest = ManifestIndex.resolve(USER_CWD)
-    backend_name: str | None = args.backend

    if args.headless:
        return _start_headless(
-            manifest, args, dry_run=dry_run, backend_name=backend_name
+            manifest, args, dry_run=dry_run
        )

    agent_name: str | None = args.name
@@ -142,6 +158,10 @@ def cmd_start(argv: list[str]) -> int:
    label, color = tui.name_color_modal(default_label=agent_name)
    label, color = _resolve_unique_label(label, color)

+    image_policy = _select_image_policy()
+    if image_policy is None:
+        return 0
+
    spec = BottleSpec(
        manifest=manifest,
        agent_name=agent_name,
@@ -150,11 +170,11 @@ def cmd_start(argv: list[str]) -> int:
        label=label,
        color=color,
        bottle_names=bottle_names,
+        image_policy=image_policy,
    )
    return _launch_bottle(
        spec,
        dry_run=dry_run,
-        backend_name=backend_name,
    )


@@ -166,7 +186,6 @@ def _start_headless(
    args: argparse.Namespace,
    *,
    dry_run: bool,
-    backend_name: str | None,
 ) -> int:
    """Non-interactive launch path for orchestrators / CI / webhooks.

@@ -210,11 +229,11 @@ def _start_headless(
        color=args.color or "",
        bottle_names=bottle_names,
        headless=True,
+        image_policy="cached" if args.cached_images else "fresh",
    )
    return _launch_bottle(
        spec,
        dry_run=dry_run,
-        backend_name=backend_name,
        assume_yes=True,
        headless_prompt_text=prompt,
    )
@@ -252,15 +271,18 @@ def prepare_with_preflight(
    injected callable, prompt y/N via the injected callable.

    `backend_name` selects which backend prepares the plan
-    (`None` → `$BOT_BOTTLE_BACKEND` → host auto-selection). The CLI
-    passes whatever `--backend` resolved to.
+    (`None` → `$BOT_BOTTLE_BACKEND` → host auto-selection).
+
+    When `spec.headless` is True the docker-fallback prompt is suppressed:
+    auto-selection dies with an actionable message rather than blocking
+    on a TTY read (which would hang CI, webhook dispatch, and orchestrators).

    Returns `(plan, identity)`. `plan` is None on dry-run or
    operator-N, but `identity` is set as soon as `backend.prepare`
    returns so callers can reap the prepare-time state dir via
    `settle_state(identity)` in their finally — exactly the existing
    semantics."""
-    backend = get_bottle_backend(backend_name)
+    backend = get_bottle_backend(backend_name, prompt=not spec.headless)
    plan = backend.prepare(spec, stage_dir=stage_dir)
    identity = _identity_from_plan(plan)

@@ -390,6 +412,13 @@ def _text_prompt_yes() -> bool:
    return reply in ("y", "Y", "yes", "YES")


+def _select_image_policy() -> str | None:
+    return tui.filter_select(
+        ["fresh", "cached"],
+        title="Select image startup mode",
+    )
+
+
 def _text_render_preflight():
    def _render(plan: DockerBottlePlan, backend_name: str) -> None:
        print(file=sys.stderr)
@@ -532,6 +561,15 @@ def _launch_bottle(
            return 0

        backend = get_bottle_backend(backend_name)
+        try:
+            backend.prelaunch_checks(plan)
+        except StaleImageError as exc:
+            if assume_yes:
+                die(str(exc))
+            sys.stderr.write(f"bot-bottle: {exc}\nLaunch anyway? [y/N] ")
+            sys.stderr.flush()
+            if read_tty_line() not in ("y", "Y", "yes", "YES"):
+                return 0
        with backend.launch(plan) as bottle:
            agent_provider_template = getattr(plan, "agent_provider_template", "claude")
            extra_args: tuple[str, ...] = ()
@@ -550,10 +588,6 @@ def _launch_bottle(
                f"session ended (exit {exit_code}); "
                f"container {bottle.name} will be removed"
            )
-            # While the container is still alive: always snapshot the
-            # transcript and — if the agent exited non-zero — mark
-            # the state for preservation. This picks up crashes /
-            # Ctrl-Cs / OOM kills before cleanup removes the state dir.
            if agent_provider_template == "claude":
                capture_claude_session_state(identity, exit_code)
        return 0
@@ -20,32 +20,19 @@ from datetime import datetime, timezone
 from pathlib import Path

 from ..paths import bot_bottle_root
-from ..bottle_state import read_metadata
-from ..backend.docker.egress_apply import (
-    EgressApplyError,
-    applicator as _docker_applicator,
-)
-from ..backend.macos_container.egress_apply import (
-    applicator as _macos_applicator,
-)
 from ..log import Die, error, info
+from ..orchestrator.client import (
+    OrchestratorClient,
+    OrchestratorClientError,
+    discover_orchestrator_url,
+)

 from ..supervise import (
-    COMPONENT_FOR_TOOL,
-    AuditEntry,
    Proposal,
-    Response,
-    STATUS_APPROVED,
-    STATUS_MODIFIED,
-    STATUS_REJECTED,
    TOOL_EGRESS_ALLOW,
    TOOL_EGRESS_BLOCK,
    TOOL_GITLEAKS_ALLOW,
    TOOL_EGRESS_TOKEN_ALLOW,
-    list_all_pending_proposals,
-    render_diff,
-    write_audit_entry,
-    write_response,
 )
 from ._common import PROG

@@ -60,30 +47,61 @@ _REPORT_ONLY_TOOLS: tuple[str, ...] = (TOOL_GITLEAKS_ALLOW, TOOL_EGRESS_TOKEN_AL

@dataclass(frozen=True)
 class QueuedProposal:
-    """A pending proposal from the supervise queue."""
+    """A pending proposal from the supervise queue.
+
+    `label` is the operator-facing bottle name (the human slug the
+    orchestrator resolved from the registry); `proposal.bottle_slug` is the
+    opaque bottle_id every operator action is keyed by. Display uses `label`;
+    respond calls use `proposal.bottle_slug`."""

    proposal: Proposal
+    label: str = ""


-# Errors any remediation engine may raise. Caught by the TUI key
-# handlers and surfaced in the status line so a failed apply keeps
-# the proposal pending rather than crashing curses.
-ApplyError = (EgressApplyError,)
+# A failed operator action (orchestrator unreachable, bottle torn down,
+# 409) is caught by the TUI key handlers and surfaced in the status line so
+# the proposal stays pending rather than crashing curses.
+ApplyError = (OrchestratorClientError,)


-def apply_routes_change(slug: str, content: str) -> tuple[str, str]:
-    meta = read_metadata(slug)
-    backend = meta.backend if meta is not None else ""
-    if backend == "macos-container":
-        return _macos_applicator.apply_routes_change(slug, content)
-    return _docker_applicator.apply_routes_change(slug, content)
+# The one per-host orchestrator, discovered lazily on first use. Every
+# operator action — list, approve, reject — goes through its HTTP control
+# plane (the orchestrator owns the single DB + live policy); there is no
+# direct-DB path and no backend branching here.
+_client_instance: OrchestratorClient | None = None
+
+
+def _resolve_orchestrator_url() -> str:
+    """URL of the running orchestrator control plane, starting one on demand.
+
+    Supervise is often the first thing an operator runs — before any bottle
+    has booted the control plane. So when discovery finds nothing, bring up
+    the selected backend's orchestrator + gateway (idempotent) rather than
+    failing with "launch a bottle first"."""
+    try:
+        return discover_orchestrator_url()
+    except OrchestratorClientError:
+        from ..backend import get_bottle_backend
+        backend = get_bottle_backend()
+        info(f"no orchestrator control plane running; starting one ({backend.name})…")
+        return backend.ensure_orchestrator()
+
+
+def _client() -> OrchestratorClient:
+    global _client_instance  # noqa: PLW0603 — CLI-session singleton
+    if _client_instance is None:
+        _client_instance = OrchestratorClient(_resolve_orchestrator_url())
+    return _client_instance


 def discover_pending() -> list[QueuedProposal]:
-    """Collect pending proposals across bottles."""
+    """Collect pending proposals across bottles from the orchestrator."""
    out = [
-        QueuedProposal(proposal=proposal)
-        for proposal in list_all_pending_proposals()
+        QueuedProposal(
+            proposal=Proposal.from_dict(d),
+            label=str(d.get("bottle_label") or d.get("bottle_slug") or ""),
+        )
+        for d in _client().supervise_pending()
    ]
    out.sort(key=lambda q: q.proposal.arrival_timestamp)
    return out
@@ -91,8 +109,8 @@ def discover_pending() -> list[QueuedProposal]:

 def _approval_status(qp: QueuedProposal, verb: str) -> str:
    """Status-line text after a successful approval."""
-    base = f"{verb} {qp.proposal.tool} for [{qp.proposal.bottle_slug}]"
-    return f"{base}; resume: ./cli.py resume {qp.proposal.bottle_slug}"
+    base = f"{verb} {qp.proposal.tool} for [{qp.label}]"
+    return f"{base}; resume: ./cli.py resume {qp.label}"


 def _detail_lines(
@@ -103,7 +121,7 @@ def _detail_lines(
    """Return the detail-view body as (text, curses-attr) tuples."""
    p = qp.proposal
    out: list[tuple[str, int]] = [
-        (f"bottle: {p.bottle_slug}", 0),
+        (f"bottle: {qp.label}", 0),
        (f"tool: {p.tool}", 0),
        (f"id: {p.id}", 0),
        (f"arrived: {p.arrival_timestamp}", 0),
@@ -136,39 +154,27 @@ def approve(
    notes: str = "",
    final_file: str | None = None,
 ) -> None:
-    """Apply the proposal, write the waiting response, and audit it."""
-    status = STATUS_MODIFIED if final_file is not None else STATUS_APPROVED
-    file_to_apply = final_file if final_file is not None else qp.proposal.proposed_file
-
-    diff_before, diff_after = "", ""
-    if qp.proposal.tool in (TOOL_EGRESS_ALLOW, TOOL_EGRESS_BLOCK):
-        diff_before, diff_after = apply_routes_change(
-            qp.proposal.bottle_slug,
-            file_to_apply,
-        )
-
-    response = Response(
-        proposal_id=qp.proposal.id,
-        status=status,
+    """Approve (or, with `final_file`, modify-then-approve) via the
+    orchestrator: it applies the route change to the bottle's live policy,
+    writes the response that unblocks the agent, and audits it — one atomic
+    server-side op. Raises `OrchestratorClientError` on failure."""
+    _client().supervise_respond(
+        qp.proposal.id,
+        bottle_slug=qp.proposal.bottle_slug,
+        decision="modify" if final_file is not None else "approve",
        notes=notes,
        final_file=final_file,
    )
-    write_response(qp.proposal.bottle_slug, response)
-    _write_audit(
-        qp, action=status, notes=notes,
-        diff_before=diff_before, diff_after=diff_after,
-    )
+

 def reject(qp: QueuedProposal, *, reason: str) -> None:
-    """Write a rejection response and an audit entry."""
-    response = Response(
-        proposal_id=qp.proposal.id,
-        status=STATUS_REJECTED,
+    """Reject via the orchestrator (writes the response + audit)."""
+    _client().supervise_respond(
+        qp.proposal.id,
+        bottle_slug=qp.proposal.bottle_slug,
+        decision="reject",
        notes=reason,
-        final_file=None,
    )
-    write_response(qp.proposal.bottle_slug, response)
-    _write_audit(qp, action=STATUS_REJECTED, notes=reason, diff_before="", diff_after="")


 def _approve_from_tui(
@@ -188,29 +194,6 @@ def _approve_from_tui(
    return _approval_status(qp, verb)


-def _write_audit(
-    qp: QueuedProposal,
-    *,
-    action: str,
-    notes: str,
-    diff_before: str,
-    diff_after: str,
-) -> None:
-    """Audit log for egress tool."""
-    component = COMPONENT_FOR_TOOL.get(qp.proposal.tool)
-    if component is None:
-        return
-    write_audit_entry(AuditEntry(
-        timestamp=datetime.now(timezone.utc).isoformat(),
-        bottle_slug=qp.proposal.bottle_slug,
-        component=component,
-        operator_action=action,
-        operator_notes=notes,
-        justification=qp.proposal.justification,
-        diff=render_diff(diff_before, diff_after, label=component),
-    ))
-
-
 # --- $EDITOR integration --------------------------------------------------


@@ -245,6 +228,20 @@ def cmd_supervise(argv: list[str]) -> int:
    )
    args = parser.parse_args(argv)

+    # Establish the orchestrator connection up front so a missing control
+    # plane is a clean one-line error, not a curses crash mid-loop. This also
+    # starts the orchestrator on demand when none is running (see `_client`).
+    try:
+        _client()
+    except OrchestratorClientError as e:
+        error(str(e))
+        return 1
+    except Die as e:
+        # Backend has no orchestrator to start (e.g. macos-container).
+        if e.message:
+            error(e.message)
+        return e.code if isinstance(e.code, int) else 1
+
    if args.once:
        return _list_once()
    try:
@@ -299,7 +296,7 @@ def _list_once() -> int:
    for qp in pending:
        sys.stdout.write(
            f"{qp.proposal.arrival_timestamp}  "
-            f"[{qp.proposal.bottle_slug}]  "
+            f"[{qp.label}]  "
            f"{qp.proposal.tool}  "
            f"{qp.proposal.id}\n"
        )
@@ -396,7 +393,7 @@ def _main_loop(stdscr: "curses._CursesWindow") -> None:  # type: ignore  # pragm
            reason = _prompt(stdscr, "reject reason: ")
            if reason:
                reject(qp, reason=reason)
-                status_line = f"rejected {qp.proposal.tool} for [{qp.proposal.bottle_slug}]"
+                status_line = f"rejected {qp.proposal.tool} for [{qp.label}]"
            else:
                status_line = "reject aborted (empty reason)"

@@ -435,7 +432,7 @@ def _render(
            cursor = "> " if i == selected else "  "
            line = (
                f"{cursor}{ts_short}  "
-                f"[{p.bottle_slug}]  {p.tool:<18}  {p.id[:8]}"
+                f"[{qp.label}]  {p.tool:<18}  {p.id[:8]}"
            )
            attr = curses.A_REVERSE if i == selected else curses.A_NORMAL
            stdscr.addnstr(row, 0, line, w - 1, attr)
@@ -0,0 +1,71 @@
+"""SQLite-backed bot-bottle configuration store."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+try:
+    from .db_store import DbStore
+    from .migrations import TableMigrations
+    from .paths import host_db_path
+except ImportError:
+    from db_store import DbStore  # type: ignore[import-not-found]  # pylint: disable=import-error,no-name-in-module
+    from migrations import TableMigrations  # type: ignore[import-not-found]  # pylint: disable=import-error,no-name-in-module
+    from paths import host_db_path  # type: ignore[import-not-found]  # pylint: disable=import-error,no-name-in-module
+
+
+DEFAULT_CACHED_IMAGE_STALE_WARNING_DAYS = 1
+
+
+class ConfigStore(DbStore):
+    """SQLite configuration for host-side bot-bottle settings."""
+
+    def __init__(self, db_path: Path | None = None) -> None:
+        migrations = TableMigrations("config_store", [
+            # v1 — host-side bot-bottle settings
+            """
+            CREATE TABLE IF NOT EXISTS bot_bottle_config (
+                id INTEGER PRIMARY KEY CHECK (id = 1),
+                cached_image_stale_warning_days INTEGER NOT NULL DEFAULT 1
+            )
+            """,
+        ])
+        super().__init__(db_path or host_db_path(), migrations)
+
+    def cached_image_stale_warning_days(self) -> int:
+        if not self.db_path.is_file():
+            return DEFAULT_CACHED_IMAGE_STALE_WARNING_DAYS
+        with self._connect() as conn:
+            row = conn.execute(
+                """
+                SELECT cached_image_stale_warning_days
+                FROM bot_bottle_config
+                WHERE id = 1
+                """,
+            ).fetchone()
+        if row is None:
+            return DEFAULT_CACHED_IMAGE_STALE_WARNING_DAYS
+        try:
+            return int(row["cached_image_stale_warning_days"])
+        except (TypeError, ValueError):
+            return DEFAULT_CACHED_IMAGE_STALE_WARNING_DAYS
+
+    def set_cached_image_stale_warning_days(self, days: int) -> Path:
+        with self._connect() as conn:
+            conn.execute(
+                """
+                INSERT INTO bot_bottle_config (id, cached_image_stale_warning_days)
+                VALUES (1, ?)
+                ON CONFLICT(id) DO UPDATE SET
+                    cached_image_stale_warning_days = excluded.cached_image_stale_warning_days
+                """,
+                (days,),
+            )
+        self._chmod()
+        return self.db_path
+
+
+__all__ = [
+    "DEFAULT_CACHED_IMAGE_STALE_WARNING_DAYS",
+    "ConfigStore",
+]
@@ -0,0 +1,17 @@
+"""Shared wire-protocol constants for gateway-bundled modules.
+
+Single source of truth for values that appear across the egress addon,
+git-http backend, supervise server, and git-gate renderer. Importing
+from this module instead of duplicating the literals means a rename is
+a one-line change and is caught by the type checker at the import site."""
+
+# App-layer identity token header. Delivered as proxy credentials
+# (HTTPS_PROXY=http://<bottle_id>:<token>@gw) by launch; the egress
+# addon reads and strips it, the supervise server and git-http backend
+# read it for attribution, and none of them forward it upstream.
+IDENTITY_HEADER = "x-bot-bottle-identity"
+
+# Shared timeout (seconds) for all git-gate subprocess and CGI calls:
+# git daemon (--timeout/--init-timeout), the access-hook subprocess in
+# git_http_backend, and the git http-backend CGI subprocess.
+GIT_GATE_TIMEOUT_SECS = 15
@@ -32,6 +32,8 @@ if TYPE_CHECKING:


 _SUPERVISE_MCP_NAME = "supervise"
+# App-layer identity token header (mirrors egress_addon / git_http_backend).
+_IDENTITY_HEADER = "x-bot-bottle-identity"


 def _skills_dir(guest_home: str) -> str:
@@ -91,6 +93,7 @@ _RUNTIME = AgentProviderRuntime(
    prompt_mode="append_file",
    bypass_args=("--dangerously-skip-permissions",),
    resume_args=("--continue",),
+    smoke_test=("claude", "--version"),
 )


@@ -300,9 +303,15 @@ class ClaudeAgentProvider(AgentProvider):
        if plan.supervise_plan is None:
            return
        info(f"registering supervise MCP server in agent claude config → {supervise_url}")
+        # Deliver the identity token as an MCP request header — the supervise
+        # daemon requires it (mandatory (source_ip, token) attribution).
+        token = getattr(plan, "identity_token", "")
+        header = (
+            f" --header {shlex.quote(f'{_IDENTITY_HEADER}: {token}')}" if token else ""
+        )
        r = bottle.exec(
            f"claude mcp add --scope user --transport http "
-            f"{_SUPERVISE_MCP_NAME} {supervise_url}",
+            f"{_SUPERVISE_MCP_NAME} {supervise_url}{header}",
            user="node",
        )
        if r.returncode != 0:
@@ -9,6 +9,7 @@ invocation that registers the supervise daemon in Codex's

 from __future__ import annotations

+import base64
 import os
 import shlex
 from pathlib import Path
@@ -26,7 +27,7 @@ from ...agent_provider import (
 )
 from .codex_auth import codex_host_access_token, write_codex_dummy_auth_file
 from ...egress import CODEX_HOST_CREDENTIAL_TOKEN_REF, EgressRoute
-from ...log import die, info, warn
+from ...log import die, info


 if TYPE_CHECKING:
@@ -34,6 +35,8 @@ if TYPE_CHECKING:


 _SUPERVISE_MCP_NAME = "supervise"
+# App-layer identity token header (mirrors egress_addon / git_http_backend).
+_IDENTITY_HEADER = "x-bot-bottle-identity"
 _CODEX_CLI = "/home/node/.codex/packages/standalone/current/bin/codex"
 _CODEX_CLI_PATH = (
    "/home/node/.local/bin:"
@@ -42,6 +45,41 @@ _CODEX_CLI_PATH = (
 )


+def _toml_basic_string(value: str) -> str:
+    """Quote `value` as a TOML basic (double-quoted) string."""
+    escaped = (
+        value.replace("\\", "\\\\")
+        .replace('"', '\\"')
+        .replace("\n", "\\n")
+        .replace("\t", "\\t")
+    )
+    return f'"{escaped}"'
+
+
+def _supervise_mcp_config_toml(supervise_url: str, token: str) -> str:
+    """Render the `[mcp_servers.supervise]` streamable-HTTP entry for
+    Codex's `config.toml`.
+
+    The Codex CLI has no `mcp add --header` flag; a static request
+    header on an HTTP MCP server is only expressible via the
+    `http_headers` config key (see `RawMcpServerConfig` /
+    `McpServerTransportConfig::StreamableHttp`). We deliver the
+    mandatory identity token (source_ip, token attribution) that way.
+    Only Codex-supported streamable-HTTP keys (`url`, `http_headers`)
+    are emitted."""
+    lines = [
+        "",
+        f"[mcp_servers.{_SUPERVISE_MCP_NAME}]",
+        f"url = {_toml_basic_string(supervise_url)}",
+    ]
+    if token:
+        key = _toml_basic_string(_IDENTITY_HEADER)
+        val = _toml_basic_string(token)
+        lines.append(f"http_headers = {{ {key} = {val} }}")
+    lines.append("")
+    return "\n".join(lines)
+
+
 def _skills_dir(guest_home: str) -> str:
    # Codex agents still read skills from the claude-code convention
    # (~/.claude/skills/) — the bot-bottle-codex image follows the
@@ -61,6 +99,7 @@ _RUNTIME = AgentProviderRuntime(
    prompt_mode="read_prompt_file",
    bypass_args=("--dangerously-bypass-approvals-and-sandbox",),
    resume_args=("resume", "--last"),
+    smoke_test=(_CODEX_CLI, "--version"),
 )


@@ -265,25 +304,39 @@ class CodexAgentProvider(AgentProvider):
        bottle: "Bottle",
        supervise_url: str,
    ) -> None:
-        """Run `codex mcp add` inside the agent guest to register the
-        supervise daemon in Codex's user config (~/.codex/config.toml).
+        """Register the supervise daemon as a streamable-HTTP MCP
+        server in Codex's user config (`~/.codex/config.toml`).

-        Mirrors the Claude provider's `claude mcp add` flow — failure
-        is logged but not fatal."""
+        We write the `[mcp_servers.supervise]` entry directly rather
+        than shelling out to `codex mcp add`: the CLI's `add` has no
+        way to attach a static request header, and the identity token
+        (mandatory (source_ip, token) attribution) MUST ride on the
+        MCP request as `http_headers`. Failure is FATAL when supervise
+        is enabled — a silently-unregistered server leaves the agent
+        with no supervise access and, under mandatory attribution, no
+        way to recover from inside the bottle."""
        if plan.supervise_plan is None:
            return
        info(f"registering supervise MCP server in agent codex config → {supervise_url}")
-        r = bottle.exec(
-            f"{shlex.quote(_CODEX_CLI)} mcp add {_SUPERVISE_MCP_NAME} --url "
-            f"{shlex.quote(supervise_url)}",
-            user="node",
+        token = getattr(plan, "identity_token", "")
+        block = _supervise_mcp_config_toml(supervise_url, token)
+        auth_dir = plan.agent_provision.guest_env.get("CODEX_HOME") \
+            or f"{plan.guest_home}/.codex"
+        config_path = f"{auth_dir}/config.toml"
+        # Append via base64 so the TOML payload never has to survive a
+        # shell-quoting round trip. node owns the config file, so append
+        # as node to preserve ownership/mode.
+        payload = base64.b64encode(block.encode()).decode()
+        script = (
+            f"printf %s {shlex.quote(payload)} | base64 -d "
+            f">> {shlex.quote(config_path)}"
        )
+        r = bottle.exec(script, user="node")
        if r.returncode != 0:
-            warn(
-                f"`codex mcp add supervise` failed (exit {r.returncode}): "
-                f"{(r.stderr or r.stdout or '').strip()}. Inside the bottle, "
-                f"register manually with: "
-                f"codex mcp add supervise --url {shlex.quote(supervise_url)}"
+            die(
+                "agent provider provisioning: could not register supervise "
+                f"MCP server in {config_path}: "
+                f"{(r.stderr or r.stdout or '').strip()}"
            )

    def headless_prompt(self, prompt: str) -> list[str]:
@@ -3,6 +3,7 @@
 from __future__ import annotations

 import sqlite3
+from contextlib import contextmanager
 from pathlib import Path

 try:
@@ -28,12 +29,21 @@ class DbStore:
        conn.row_factory = sqlite3.Row
        return conn

+    @contextmanager
+    def _connection(self):
+        conn = self._connect()
+        try:
+            with conn:
+                yield conn
+        finally:
+            conn.close()
+
    def is_migrated(self) -> bool:
        """Return True if the DB is fully up-to-date, False if migration is needed."""
        if not self.db_path.exists():
            return False
        try:
-            with self._connect() as conn:
+            with self._connection() as conn:
                row = conn.execute(
                    "SELECT version FROM schema_versions WHERE module = ?",
                    (self._migrations.schema_key,),
@@ -45,7 +55,7 @@ class DbStore:

    def migrate(self) -> None:
        """Apply any pending migrations and set permissions on the DB file."""
-        with self._connect() as conn:
+        with self._connection() as conn:
            self._migrations.apply(conn)
        self._chmod()

@@ -3,9 +3,8 @@
 Pure Python, no mitmproxy dependency. Each detector is a module-level
 function returning `ScanResult | None`.

-Ships flat into the gateway image alongside
-`egress_addon_core.py` — both this file and the package source use
-the same try/except import shim pattern.
+Available in the gateway via the installed `bot_bottle` package
+(see `Dockerfile.gateway`).
 """

 from __future__ import annotations
@@ -20,10 +19,7 @@ from math import log2
 from collections import Counter
 from urllib.parse import quote as url_quote

-try:
-    from egress_addon_core import ScanResult  # type: ignore[import-not-found]
-except ImportError:  # pragma: no cover - host-side path
-    from .egress_addon_core import ScanResult
+from .egress_addon_core import ScanResult


 # ---------------------------------------------------------------------------
@@ -14,13 +14,20 @@ from __future__ import annotations
 import subprocess


-def run_docker(argv: list[str]) -> subprocess.CompletedProcess[str]:
+def run_docker(
+    argv: list[str], *, env: dict[str, str] | None = None,
+) -> subprocess.CompletedProcess[str]:
    """Run a `docker` command, capturing stdout/stderr as text. Never raises
    on a non-zero exit — callers inspect `returncode` / `stderr` so they can
    stay fail-closed or tolerate idempotent no-ops (e.g. removing an
-    already-absent container)."""
+    already-absent container).
+
+    `env` sets the child process environment — used to hand a secret to a bare
+    `--env NAME` flag (docker inherits its value from this process) so the
+    value never lands on argv or in `docker inspect`'s recorded command line."""
    return subprocess.run(
-        argv, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False,
+        argv, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True,
+        check=False, env=env,
    )


@@ -6,16 +6,18 @@ egress container."""
 from __future__ import annotations

 import asyncio
+import base64
+import binascii
 import json
 import os
-import signal
 import sys
 import typing
-from pathlib import Path

 from mitmproxy import http  # type: ignore[import-not-found]  # pylint: disable=import-error

-from egress_addon_core import (  # type: ignore[import-not-found]  # pylint: disable=import-error
+from bot_bottle.constants import IDENTITY_HEADER
+from bot_bottle.dlp_detectors import redact_tokens, strip_crlf
+from bot_bottle.egress_addon_core import (
    LOG_BLOCKS,
    LOG_FULL,
    DEFAULT_OUTBOUND_ON_MATCH,
@@ -31,7 +33,6 @@ from egress_addon_core import (  # type: ignore[import-not-found]  # pylint: dis
    decide_git_fetch,
    is_git_fetch_request,
    is_git_push_request,
-    load_config,
    match_route,
    resolve_client_context,
    outbound_scan_headers,
@@ -39,39 +40,40 @@ from egress_addon_core import (  # type: ignore[import-not-found]  # pylint: dis
    scan_inbound,
    scan_outbound,
 )
+from bot_bottle import supervise as _sv
+from bot_bottle.policy_resolver import PolicyResolver

-try:
-    from dlp_detectors import redact_tokens, strip_crlf  # type: ignore[import-not-found]
-except ImportError:  # pragma: no cover - host-side path
-    from bot_bottle.dlp_detectors import (  # type: ignore[import-not-found]
-        redact_tokens,
-        strip_crlf,
-    )
-
-try:
-    import supervise as _sv  # type: ignore[import-not-found]
-except ImportError:  # pragma: no cover - host-side path
-    from bot_bottle import supervise as _sv  # type: ignore[import-not-found]
-
-try:
-    from policy_resolver import PolicyResolver  # type: ignore[import-not-found]
-except ImportError:  # pragma: no cover - host-side path
-    from bot_bottle.policy_resolver import PolicyResolver
-
-
-DEFAULT_ROUTES_PATH = "/etc/egress/routes.yaml"

 INTROSPECT_HOST = "_egress.local"

-# Consolidated (multi-tenant) mode: when this points at the per-host
-# orchestrator's control plane, the addon resolves each client's Config by
-# source IP per request instead of using a single static routes file. Unset
-# → legacy per-bottle single-tenant mode (unchanged).
+# The per-host orchestrator control plane the addon resolves every request's
+# Config against, by source IP (PRD 0070). Mandatory: the consolidated gateway
+# is the only topology now — there is no static per-bottle routes file to fall
+# back to — so an unset value is a fatal misconfiguration (see __init__).
 ORCHESTRATOR_URL_ENV = "BOT_BOTTLE_ORCHESTRATOR_URL"

-# App-layer identity token (defense-in-depth over the source-IP invariant);
-# the agent injects it, the addon strips it so it never leaks upstream.
-IDENTITY_HEADER = "x-bot-bottle-identity"
+# Per-flow key under which `request()` stashes the resolved (Config, supervise
+# slug, env) so the later `response()` and `websocket_message()` hooks scan
+# against the *calling bottle's* policy — the same one the request was decided
+# on — without a second `/resolve` per response or per WebSocket frame. A hook
+# on a flow that never resolved (no stash) fails closed to deny-all, so it's a
+# safe no-op rather than an unscanned pass.
+_FLOW_CTX_KEY = "bot_bottle_egress_ctx"
+
+
+def _token_from_proxy_auth(header: str) -> str:
+    """Extract the identity token (the password) from a `Proxy-Authorization:
+    Basic base64(<bottle_id>:<token>)` header. Empty on any malformed value —
+    the mandatory `/resolve` then fail-closes on the empty token."""
+    scheme, _, encoded = header.partition(" ")
+    if scheme.lower() != "basic" or not encoded:
+        return ""
+    try:
+        decoded = base64.b64decode(encoded, validate=True).decode("utf-8")
+    except (binascii.Error, ValueError, UnicodeDecodeError):
+        return ""
+    _, _, password = decoded.partition(":")
+    return password

 # Seconds the egress proxy holds a token-blocked request open waiting for the
 # operator's supervisor decision (PRD 0062), overridable via env.
@@ -89,33 +91,47 @@ _TOKEN_ALLOW_JUSTIFICATION = (


 class EgressAddon:
-    # Class default so addons built via __new__ (e.g. in tests) default to
-    # single-tenant; __init__ sets the instance attribute for real runs.
-    _resolver: "PolicyResolver | None" = None
+    # Bare annotations (no class value): __init__ sets a live PolicyResolver for
+    # real runs, and every host-side test builds an addon via __new__ and sets a
+    # fake resolver. Egress is resolver-only now — the per-request policy always
+    # comes from the orchestrator's /resolve (PRD 0070); there is no static
+    # per-bottle routes file, SIGHUP reload, or single-tenant fallback.
+    _resolver: "PolicyResolver"
+    # Class default so __new__-built addons have it (real runs get a fresh
+    # per-instance dict in __init__; only http_connect mutates it, which the
+    # request-flow tests don't exercise).
+    _conn_tokens: "dict[str, str]" = {}

    def __init__(self) -> None:
-        self.routes_path = os.environ.get("EGRESS_ROUTES", DEFAULT_ROUTES_PATH)
-        self.config: Config = Config(routes=())
-        # Consolidated mode: resolve per-client Config from the orchestrator.
-        # Absent → single-tenant (static routes file); behaviour unchanged.
+        # Resolver-only: the gateway is always multi-tenant, resolving each
+        # request's policy by source IP against the orchestrator control plane
+        # (PRD 0070). The URL is mandatory — without a policy source the gateway
+        # must not come up (fail-closed), rather than silently allowing nothing.
        orch_url = os.environ.get(ORCHESTRATOR_URL_ENV, "").strip()
-        self._resolver = PolicyResolver(orch_url) if orch_url else None
+        if not orch_url:
+            raise RuntimeError(
+                f"{ORCHESTRATOR_URL_ENV} is required: the egress gateway "
+                "resolves every request's policy from the orchestrator and has "
+                "no static routes file to fall back to."
+            )
+        self._resolver = PolicyResolver(orch_url)
        # Tokens the operator has approved this session (PRD 0062), keyed by
        # bottle so the shared gateway keeps each bottle's safelist separate —
        # a global set would let bottle A's approved secret pass bottle B's DLP
        # scan. In-memory only (a restart re-prompts); mutated only from the
        # asyncio loop that runs the addon hooks, so no lock is needed.
        self._safe_tokens: dict[str, set[str]] = {}
-        self._supervise_slug = os.environ.get("SUPERVISE_BOTTLE_SLUG", "").strip()
+        # Per-client-connection identity token captured from the CONNECT's
+        # `Proxy-Authorization` (HTTPS tunnels don't repeat it on the bumped
+        # inner requests). Keyed by client_conn.id; cleared on disconnect.
+        self._conn_tokens: dict[str, str] = {}
        self._token_allow_timeout = _token_allow_timeout_from_env(os.environ)
-        self._reload(initial=True)
-        self._install_sighup()

    @staticmethod
    def _supervise_available(slug: str) -> bool:
        """Supervise is reachable for this request iff we resolved a bottle to
-        attribute its proposals to (single-tenant env slug, or a source-IP
-        -attributed bottle id). Empty → fail closed (no queue to write to)."""
+        attribute its proposals to (the source-IP-attributed bottle id). Empty
+        → fail closed (no queue to write to)."""
        return bool(slug)

    def _safe_tokens_for(self, slug: str) -> set[str]:
@@ -124,40 +140,15 @@ class EgressAddon:
        bottle's approved token into another's scan."""
        return self._safe_tokens.setdefault(slug, set())

-    def _reload(self, *, initial: bool = False) -> None:
-        try:
-            text = Path(self.routes_path).read_text(encoding="utf-8")
-            new_config = load_config(text)
-        except (OSError, ValueError) as e:
-            tag = "boot" if initial else "SIGHUP"
-            sys.stderr.write(
-                f"egress: {tag} load failed: {e}\n"
-            )
-            if initial:
-                self.config = Config(routes=())
-            return
-        self.config = new_config
-        log_label = ("off", "blocks", "full")[self.config.log]
-        sys.stderr.write(
-            f"egress: loaded {len(self.config.routes)} route(s): "
-            f"{', '.join(r.host for r in self.config.routes)}"
-            f" [log={log_label}]\n"
-        )
-
-    def _install_sighup(self) -> None:
-        if not hasattr(signal, "SIGHUP"):
-            return
-
-        def handler(signum: int, frame: object) -> None:
-            del signum, frame
-            self._reload()
-
-        signal.signal(signal.SIGHUP, handler)
-
-    def _serve_introspection(self, flow: http.HTTPFlow, path: str) -> None:
+    def _serve_introspection(
+        self, flow: http.HTTPFlow, path: str, config: Config,
+    ) -> None:
+        """Serve the calling bottle's own allowlist. `config` is this flow's
+        resolved policy (the same one every hook uses), so the agent sees the
+        routes that actually apply to it."""
        if path == "/allowlist":
            payload = json.dumps(
-                {"routes": [route_to_yaml_dict(r) for r in self.config.routes]},
+                {"routes": [route_to_yaml_dict(r) for r in config.routes]},
                indent=2,
            ).encode("utf-8")
            flow.response = http.Response.make(
@@ -171,11 +162,21 @@ class EgressAddon:
            {"Content-Type": "text/plain; charset=utf-8"},
        )

+    def _flow_log(self, flow: http.HTTPFlow) -> int:
+        """This flow's log level, from the policy `request()` resolved and
+        stashed. The block/redact log gates were a single global in the static-
+        config world; they are per bottle now, so they read it from the flow."""
+        return self._flow_ctx(flow)[0].log
+
    def _req_ctx(self, flow: http.HTTPFlow) -> dict[str, object]:
+        # Redact with this flow's resolved env overlay (process env + the
+        # bottle's /resolve tokens), so the ctx scrubs the calling bottle's
+        # provisioned secrets, not just os.environ's.
+        env = self._flow_ctx(flow)[2]
        return {
-            "host": redact_tokens(flow.request.pretty_host, env=os.environ),
+            "host": redact_tokens(flow.request.pretty_host, env=env),
            "method": flow.request.method,
-            "path": redact_tokens(flow.request.path, env=os.environ),
+            "path": redact_tokens(flow.request.path, env=env),
        }

    def _block(
@@ -184,7 +185,7 @@ class EgressAddon:
        reason: str,
        ctx: dict[str, object] | None = None,
    ) -> None:
-        if self.config.log >= LOG_BLOCKS:
+        if self._flow_log(flow) >= LOG_BLOCKS:
            entry: dict[str, object] = {"event": "egress_block", "reason": reason}
            if ctx:
                entry.update(ctx)
@@ -195,31 +196,39 @@ class EgressAddon:
            {"Content-Type": "text/plain; charset=utf-8"},
        )

-    def _log_request(self, flow: http.HTTPFlow) -> None:
+    def _log_request(
+        self, flow: http.HTTPFlow, env: "typing.Mapping[str, str]",
+    ) -> None:
+        # `env` is the per-flow resolved overlay (process env + this bottle's
+        # /resolve tokens), so the log redaction scrubs the calling bottle's
+        # provisioned secrets — not just the process-level ones in os.environ.
        headers = {
-            k: redact_tokens(v, env=os.environ)
+            k: redact_tokens(v, env=env)
            for k, v in flow.request.headers.items()
            if k.lower() != "authorization"
        }
-        body = redact_tokens(flow.request.get_text(strict=False) or "", env=os.environ)
+        body = redact_tokens(flow.request.get_text(strict=False) or "", env=env)
        sys.stderr.write(
            json.dumps({
                "event": "egress_request",
-                "host": redact_tokens(flow.request.pretty_host, env=os.environ),
+                "host": redact_tokens(flow.request.pretty_host, env=env),
                "method": flow.request.method,
-                "path": redact_tokens(flow.request.path, env=os.environ),
+                "path": redact_tokens(flow.request.path, env=env),
                "headers": headers,
                "body": body,
            })
            + "\n"
        )

-    def _log_response(self, flow: http.HTTPFlow) -> None:
+    def _log_response(
+        self, flow: http.HTTPFlow, env: "typing.Mapping[str, str]",
+    ) -> None:
+        # Per-flow env overlay (see _log_request): redact this bottle's tokens.
        headers = {
-            k: redact_tokens(v, env=os.environ)
+            k: redact_tokens(v, env=env)
            for k, v in flow.response.headers.items()
        }
-        body = redact_tokens(flow.response.get_text(strict=False) or "", env=os.environ)
+        body = redact_tokens(flow.response.get_text(strict=False) or "", env=env)
        sys.stderr.write(
            json.dumps({
                "event": "egress_response",
@@ -234,33 +243,94 @@ class EgressAddon:
    def _resolve_flow(
        self, flow: http.HTTPFlow,
    ) -> "tuple[Config, str, typing.Mapping[str, str]]":
-        """The `(Config, supervise slug, env)` to apply to this request.
-        Single-tenant → the static `self.config`, the env slug, and the process
-        env. Consolidated → the calling bottle's Config + bottle id + auth
-        tokens, resolved by source IP in one round-trip (fail-closed to deny-all
-        + empty slug if unattributed); `env` is the process env overlaid with
-        the bottle's tokens, so upstream-auth injection (and DLP) use *this*
-        bottle's credentials — exactly what the per-bottle gateway daemon's env did.
-        The identity token, if the agent injected one, is read then stripped so
-        it never leaks upstream."""
-        if self._resolver is None:
-            return self.config, self._supervise_slug, os.environ
+        """The calling bottle's `(Config, supervise slug, env)`, resolved by
+        source IP in one round-trip against the orchestrator — fail-closed to
+        deny-all + empty slug if unattributed. `env` is the process env overlaid
+        with the bottle's `/resolve` tokens, so upstream-auth injection (and DLP)
+        use *this* bottle's credentials. The identity token, if the agent
+        injected one, is read then stripped so it never leaks upstream."""
        conn = flow.client_conn
        client_ip = conn.peername[0] if conn and conn.peername else ""
-        token = flow.request.headers.get(IDENTITY_HEADER, "")
-        flow.request.headers.pop(IDENTITY_HEADER, None)
+        token = self._request_token(flow)
        config, slug, tokens = resolve_client_context(self._resolver, client_ip, token)
        env = {**os.environ, **tokens} if tokens else os.environ
        return config, slug, env

+    def _stash_flow_ctx(
+        self,
+        flow: http.HTTPFlow,
+        config: Config,
+        slug: str,
+        env: "typing.Mapping[str, str]",
+    ) -> None:
+        """Remember the per-flow context `request()` resolved, so the later
+        `response()` / `websocket_message()` hooks reuse it — scanning against
+        the same bottle's policy the request was decided on, with one `/resolve`
+        per flow rather than one per frame."""
+        meta = getattr(flow, "metadata", None)
+        if isinstance(meta, dict):
+            meta[_FLOW_CTX_KEY] = (config, slug, env)
+
+    def _flow_ctx(
+        self, flow: http.HTTPFlow,
+    ) -> "tuple[Config, str, typing.Mapping[str, str]]":
+        """The `(Config, supervise slug, env)` `request()` resolved for this
+        flow, so a later hook scans against the calling bottle's policy. Falls
+        back to deny-all (empty routes, empty slug) for a flow that never passed
+        through `request()` (or a flow object without metadata) — fail-closed, so
+        a DLP hook on such a flow is a safe no-op rather than an unscanned pass."""
+        meta = getattr(flow, "metadata", None)
+        if isinstance(meta, dict):
+            ctx = meta.get(_FLOW_CTX_KEY)
+            if ctx is not None:
+                return ctx
+        return Config(routes=()), "", os.environ
+
+    def _request_token(self, flow: http.HTTPFlow) -> str:
+        """The per-bottle identity token for this request, from the proxy
+        credentials — the delivery mechanism (`HTTPS_PROXY=http://id:token@gw`)
+        that clients honor without app changes. Plain-HTTP requests carry
+        `Proxy-Authorization` directly; HTTPS bumped requests inherit the token
+        captured from their tunnel's CONNECT. Read then stripped so it never
+        leaks upstream (also strips the legacy header, if present)."""
+        token = _token_from_proxy_auth(
+            flow.request.headers.get("Proxy-Authorization", ""))
+        flow.request.headers.pop("Proxy-Authorization", None)
+        flow.request.headers.pop(IDENTITY_HEADER, None)
+        conn = flow.client_conn
+        if not token and conn is not None:
+            token = self._conn_tokens.get(getattr(conn, "id", ""), "")
+        return token
+
+    def http_connect(self, flow: http.HTTPFlow) -> None:
+        """Capture the identity token from an HTTPS tunnel's CONNECT (the inner
+        bumped requests won't carry `Proxy-Authorization`), keyed by client
+        connection, and strip it so it never reaches upstream."""
+        token = _token_from_proxy_auth(
+            flow.request.headers.get("Proxy-Authorization", ""))
+        flow.request.headers.pop("Proxy-Authorization", None)
+        conn = flow.client_conn
+        if conn is not None and getattr(conn, "id", ""):
+            self._conn_tokens[conn.id] = token
+
+    def client_disconnected(self, client: typing.Any) -> None:
+        """Drop the per-connection token when the client goes away."""
+        self._conn_tokens.pop(getattr(client, "id", ""), None)
+
    async def request(self, flow: http.HTTPFlow) -> None:
        request_path, _, query = flow.request.path.partition("?")

-        if flow.request.pretty_host == INTROSPECT_HOST:
-            self._serve_introspection(flow, request_path)
-            return
-
        config, slug, env = self._resolve_flow(flow)
+        # Stash for the response / websocket hooks so their DLP scans reuse this
+        # bottle's resolved policy (one /resolve per flow — see _flow_ctx).
+        self._stash_flow_ctx(flow, config, slug, env)
+
+        # Introspection ("_egress.local/allowlist") reports the calling bottle's
+        # own resolved routes — served after resolution so it reflects this
+        # bottle's policy, not a stale global.
+        if flow.request.pretty_host == INTROSPECT_HOST:
+            self._serve_introspection(flow, request_path, config)
+            return

        # DLP outbound scan BEFORE stripping auth — catches tokens the
        # agent tried to smuggle in any header, path, query param, or body.
@@ -309,6 +379,7 @@ class EgressAddon:
            env,
            request_method=flow.request.method,
            request_headers=req_headers,
+            deny_reason=config.deny_reason,
        )

        if decision.action == "block":
@@ -319,7 +390,7 @@ class EgressAddon:
            flow.request.headers["authorization"] = decision.inject_authorization

        if config.log >= LOG_FULL:
-            self._log_request(flow)
+            self._log_request(flow, env)

    def _block_dlp(self, flow: http.HTTPFlow, result: ScanResult) -> None:
        ctx = self._req_ctx(flow)
@@ -366,7 +437,7 @@ class EgressAddon:
            # forwards; it fails closed only if a match survives the scrub.
            if policy == ON_MATCH_REDACT:
                if self._redact_outbound(flow, route, env):
-                    if self.config.log >= LOG_BLOCKS:
+                    if self._flow_log(flow) >= LOG_BLOCKS:
                        sys.stderr.write(json.dumps({
                            "event": "egress_redacted",
                            "reason": f"egress DLP: {result.reason}",
@@ -488,7 +559,7 @@ class EgressAddon:
            _sv.STATUS_APPROVED, _sv.STATUS_MODIFIED,
        ):
            self._safe_tokens_for(slug).add(result.matched)
-            if self.config.log >= LOG_BLOCKS:
+            if self._flow_log(flow) >= LOG_BLOCKS:
                sys.stderr.write(json.dumps({
                    "event": "egress_token_allowed",
                    "reason": f"egress DLP: {result.reason}",
@@ -528,14 +599,16 @@ class EgressAddon:
            await asyncio.sleep(TOKEN_ALLOW_POLL_INTERVAL_SECONDS)

    def response(self, flow: http.HTTPFlow) -> None:
-        """DLP inbound scan on response headers and body."""
-        route = match_route(self.config.routes, flow.request.pretty_host)
+        """DLP inbound scan on response headers and body, against the calling
+        bottle's resolved config (`request()` stashed it — see `_flow_ctx`)."""
+        config, _slug, env = self._flow_ctx(flow)
+        route = match_route(config.routes, flow.request.pretty_host)
        if route is None:
            return
        if flow.response is None:
            return
-        if self.config.log >= LOG_FULL:
-            self._log_response(flow)
+        if config.log >= LOG_FULL:
+            self._log_response(flow, env)
        resp_headers = {k.lower(): v for k, v in flow.response.headers.items()}
        body = flow.response.get_text(strict=False) or ""
        scan_text = build_inbound_scan_text(resp_headers, body)
@@ -552,7 +625,7 @@ class EgressAddon:
            resp_ctx = {**resp_ctx, "context": result.context}
        if result.severity == "block":
            self._block(flow, f"egress DLP: {result.reason}", ctx=resp_ctx)
-        elif result.severity == "warn" and self.config.log >= LOG_BLOCKS:
+        elif result.severity == "warn" and config.log >= LOG_BLOCKS:
            sys.stderr.write(
                json.dumps({
                    "event": "egress_warn",
@@ -563,7 +636,9 @@ class EgressAddon:
            )

    def websocket_message(self, flow: http.HTTPFlow) -> None:
-        """DLP scan on WebSocket frames.
+        """DLP scan on WebSocket frames, against the calling bottle's resolved
+        config (see `_flow_ctx`). `request()` resolves and stashes the per-flow
+        (config, slug, env) at the upgrade, and every frame reuses it.

        Outbound frames (from_client) are scanned for credential leakage;
        inbound frames are scanned for prompt injection.  On a block the
@@ -572,10 +647,8 @@ class EgressAddon:
        """
        if flow.websocket is None:  # type: ignore[union-attr]
            return
-        # WebSocket DLP runs against the static config only (single-tenant); in
-        # the consolidated gateway self.config has no routes, so this is inert
-        # until websocket routing is made source-IP-aware (a separate slice).
-        route = match_route(self.config.routes, flow.request.pretty_host)
+        config, slug, env = self._flow_ctx(flow)
+        route = match_route(config.routes, flow.request.pretty_host)
        if route is None:
            return
        message = flow.websocket.messages[-1]  # type: ignore[union-attr]
@@ -584,8 +657,8 @@ class EgressAddon:
            # A WebSocket data frame is not an HTTP request line, so CRLF is
            # not an injection vector here — scan only for credential leakage.
            result = scan_outbound(
-                route, content, os.environ,
-                safe_tokens=self._safe_tokens_for(self._supervise_slug), crlf_text="",
+                route, content, env,
+                safe_tokens=self._safe_tokens_for(slug), crlf_text="",
            )
            if result is not None and result.severity == "block":
                sys.stderr.write(f"egress DLP: {result.reason}\n")
@@ -6,9 +6,9 @@ exercise the parse + decision functions without depending on the
 `mitmproxy.http.HTTPFlow` API and is loaded inside the gateway
 container.

-Imports: stdlib + `yaml_subset` (which is itself stdlib-only and
-ships flat into the gateway image alongside this file —
-see `Dockerfile.gateway`)."""
+Imports: stdlib + sibling package modules (`yaml_subset`,
+`egress_dlp_config`). Available in the gateway via the installed
+`bot_bottle` package (see `Dockerfile.gateway`)."""

 from __future__ import annotations

@@ -16,36 +16,20 @@ import re
 import typing
 from dataclasses import dataclass

-try:
-    from yaml_subset import YamlSubsetError, parse_yaml_subset  # type: ignore[import-not-found]
-except ImportError:  # pragma: no cover - host-side path
-    from .yaml_subset import YamlSubsetError, parse_yaml_subset
+from .yaml_subset import YamlSubsetError, parse_yaml_subset

-# DLP detector-config parsing lives in a sibling module (also flat-bundled
-# into the gateway — see Dockerfile.gateway). Re-exported below so existing
-# `from egress_addon_core import ON_MATCH_*` callers keep working.
-try:
-    from egress_dlp_config import (  # type: ignore[import-not-found]
-        DEFAULT_OUTBOUND_ON_MATCH,
-        INBOUND_DETECTOR_NAMES,
-        ON_MATCH_BLOCK,
-        ON_MATCH_REDACT,
-        ON_MATCH_SUPERVISE,
-        OUTBOUND_DETECTOR_NAMES,
-        OUTBOUND_ON_MATCH_VALUES,
-        parse_dlp_block,
-    )
-except ImportError:  # pragma: no cover - host-side path
-    from .egress_dlp_config import (
-        DEFAULT_OUTBOUND_ON_MATCH,
-        INBOUND_DETECTOR_NAMES,
-        ON_MATCH_BLOCK,
-        ON_MATCH_REDACT,
-        ON_MATCH_SUPERVISE,
-        OUTBOUND_DETECTOR_NAMES,
-        OUTBOUND_ON_MATCH_VALUES,
-        parse_dlp_block,
-    )
+# DLP detector-config parsing lives in a sibling module. Re-exported below
+# so existing `from egress_addon_core import ON_MATCH_*` callers keep working.
+from .egress_dlp_config import (
+    DEFAULT_OUTBOUND_ON_MATCH,
+    INBOUND_DETECTOR_NAMES,
+    ON_MATCH_BLOCK,
+    ON_MATCH_REDACT,
+    ON_MATCH_SUPERVISE,
+    OUTBOUND_DETECTOR_NAMES,
+    OUTBOUND_ON_MATCH_VALUES,
+    parse_dlp_block,
+)


 # ---------------------------------------------------------------------------
@@ -105,6 +89,14 @@ LOG_FULL = 2    # log block/warn events + full request and response bodies
 class Config:
    routes: tuple[Route, ...]
    log: int = LOG_OFF
+    # Why this Config is a deny-all, when it is one for a reason *other* than
+    # the bottle's own policy genuinely not listing the host. A deny-all is
+    # indistinguishable from "policy loaded, host not allowed" at the decision
+    # point — both are simply "no matching route" — so without this the
+    # operator sees `host X is not in the allowlist` and goes hunting for a
+    # missing route that was never the problem. Empty for a normally-parsed
+    # policy; `decide` prefers it over the allowlist wording when set.
+    deny_reason: str = ""


@dataclass(frozen=True)
@@ -421,16 +413,40 @@ class PolicyResolverLike(typing.Protocol):
        ...


+# Deny-all explanations. Each names the *actual* failure so an operator isn't
+# sent looking for a missing egress route when the bottle never had a policy
+# to begin with — the failure mode that made a bricked registration read like
+# a misconfigured allowlist.
+DENY_UNATTRIBUTED = (
+    "egress: this request was not attributed to any bottle, so no egress "
+    "policy applies and every host is denied. Either the bottle's registry "
+    "row is missing/ambiguous (torn down, or another bottle claimed its "
+    "source IP), or the request carried no matching identity token — check "
+    "that the caller's proxy URL includes it. This is not an allowlist problem."
+)
+DENY_UNPARSEABLE = (
+    "egress: this bottle's egress policy could not be parsed, so it is being "
+    "treated as deny-all. Fix the bottle's egress.routes; every host is denied "
+    "until it loads."
+)
+DENY_RESOLVER_ERROR = (
+    "egress: the orchestrator could not be reached to resolve this bottle's "
+    "egress policy, so every host is denied (fail-closed). Check that the "
+    "control plane is up; this is not an allowlist problem."
+)
+
+
 def _config_from_policy(policy: "str | None") -> "Config":
    """Parse a resolved policy blob into a Config, fail-closed: None / empty /
    unparseable all become a deny-all Config (no routes → every request
-    blocked)."""
+    blocked). Each deny-all carries the reason it is one, so the block message
+    names the real fault instead of blaming the allowlist."""
    if not policy:
-        return Config(routes=())  # unattributed or empty → deny-all
+        return Config(routes=(), deny_reason=DENY_UNATTRIBUTED)
    try:
        return load_config(policy)
    except ValueError:
-        return Config(routes=())  # unparseable policy → deny
+        return Config(routes=(), deny_reason=DENY_UNPARSEABLE)


 def resolve_client_config(
@@ -444,7 +460,7 @@ def resolve_client_config(
    try:
        policy = resolver.resolve(client_ip, identity_token)
    except Exception:  # noqa: BLE001  # pylint: disable=broad-exception-caught
-        return Config(routes=())  # orchestrator unreachable/errored → deny
+        return Config(routes=(), deny_reason=DENY_RESOLVER_ERROR)
    return _config_from_policy(policy)


@@ -473,7 +489,7 @@ def resolve_client_context(
            client_ip, identity_token,
        )
    except Exception:  # noqa: BLE001  # pylint: disable=broad-exception-caught
-        return Config(routes=()), "", {}  # orchestrator unreachable/errored → deny
+        return Config(routes=(), deny_reason=DENY_RESOLVER_ERROR), "", {}
    return _config_from_policy(policy), (bottle_id or ""), tokens


@@ -588,12 +604,16 @@ def decide(
    *,
    request_method: str = "GET",
    request_headers: typing.Mapping[str, str] | None = None,
+    deny_reason: str = "",
 ) -> Decision:
+    """`deny_reason` is `Config.deny_reason`: when the deny-all came from a
+    missing/unparseable policy rather than the bottle's own allowlist, report
+    that instead of implying a route is merely absent."""
    route = match_route(routes, request_host)
    if route is None:
        return Decision(
            action="block",
-            reason=(
+            reason=deny_reason or (
                f"egress: host {request_host!r} is not in the "
                f"bottle's egress.routes allowlist. Declare a "
                f"route for it or remove the request."
@@ -868,6 +888,9 @@ __all__ = [
    "is_git_push_request",
    "is_git_fetch_request",
    "load_config",
+    "DENY_UNATTRIBUTED",
+    "DENY_UNPARSEABLE",
+    "DENY_RESOLVER_ERROR",
    "resolve_client_config",
    "resolve_client_context",
    "PolicyResolverLike",
@@ -15,11 +15,23 @@
 #     mitmproxy at it. The option REPLACES mitmproxy's default
 #     trust store, so passing the upstream CA alone would break
 #     non-chained hosts.
-#   * `-s /app/egress_addon.py` loads the addon that reads
-#     /etc/egress/routes.yaml.
+#   * `-s /app/egress_addon.py` loads the addon that resolves each
+#     request's policy from the orchestrator control plane by source
+#     IP (PRD 0070). There is no static routes file.

 set -e

+# Fail closed on a missing policy source. The addon itself raises at
+# load when BOT_BOTTLE_ORCHESTRATOR_URL is unset (so mitmdump exits via
+# its errorcheck addon), but that leaves the fail-closed guarantee at the
+# mercy of a mitmproxy version keeping that behavior. Refuse here too, so
+# a misconfigured gateway can never come up as a bare TLS-bumping open
+# proxy with no policy — independent of mitmproxy's startup-error handling.
+if [ -z "$BOT_BOTTLE_ORCHESTRATOR_URL" ]; then
+  echo "egress: BOT_BOTTLE_ORCHESTRATOR_URL is required (no static routes fallback)" >&2
+  exit 1
+fi
+
 # Pin mitmproxy's config dir to the bind-mount location of its CA
 # regardless of which user mitmdump runs as. In the legacy
 # four-daemon setup (Dockerfile.egress, USER mitmproxy) this
@@ -61,6 +61,11 @@ class _DaemonSpec:
 _EGRESS_ONLY_ENV_PREFIXES: tuple[str, ...] = ("EGRESS_TOKEN_",)
 _READY_GATED_DAEMONS: tuple[str, ...] = ("git-gate", "git-http")

+# Daemons that must be requested explicitly via BOT_BOTTLE_GATEWAY_DAEMONS
+# and are NOT started in the default (env-var-unset) case. The orchestrator
+# only runs in the combined infra container, never in a standalone gateway.
+_OPT_IN_DAEMONS: frozenset[str] = frozenset({"orchestrator"})
+

 def _env_for_daemon(name: str, base_env: dict[str, str]) -> dict[str, str]:
    """Egress sees the full bundle env. Everyone else gets a copy
@@ -75,11 +80,18 @@ def _env_for_daemon(name: str, base_env: dict[str, str]) -> dict[str, str]:
    }


+# The orchestrator is listed first so it starts before the gateway daemons,
+# giving the control plane a head start to accept /resolve calls. The gateway
+# daemons tolerate early /resolve failures and retry per-request.
 _DAEMONS: tuple[_DaemonSpec, ...] = (
+    _DaemonSpec("orchestrator", (
+        "python3", "-m", "bot_bottle.orchestrator",
+        "--host", "0.0.0.0", "--port", "8099", "--broker", "stub",
+    )),
    _DaemonSpec("egress", ("/bin/sh", "/app/egress-entrypoint.sh")),
    _DaemonSpec("git-gate", ("/bin/sh", "/git-gate-entrypoint.sh")),
-    _DaemonSpec("git-http", ("python3", "/app/git_http_backend.py")),
-    _DaemonSpec("supervise", ("python3", "/app/supervise_server.py")),
+    _DaemonSpec("git-http", ("python3", "-m", "bot_bottle.git_http_backend")),
+    _DaemonSpec("supervise", ("python3", "-m", "bot_bottle.supervise_server")),
 )


@@ -103,18 +115,20 @@ def _selected_daemons(
    env: dict[str, str],
    all_daemons: Sequence[_DaemonSpec] | None = None,
 ) -> tuple[_DaemonSpec, ...]:
-    """Filter the daemon set by the BOT_BOTTLE_GATEWAY_DAEMONS env
-    var. Unknown names in the list are ignored — the caller is the
-    source of truth for which daemons are wired.
+    """Filter the daemon set by the BOT_BOTTLE_GATEWAY_DAEMONS env var.

-    `all_daemons` defaults to `_DAEMONS` resolved at call time (not
-    at definition time), so tests can monkey-patch the module-level
-    `_DAEMONS` and have the new value take effect."""
+    When the var is unset/empty, return all non-opt-in daemons (the
+    standard gateway subset). Opt-in daemons (e.g. `orchestrator`) only
+    run when explicitly named — they never start in a plain gateway
+    container that doesn't set the env var. Unknown names are ignored.
+
+    `all_daemons` defaults to `_DAEMONS` resolved at call time (not at
+    definition time), so tests can pass a custom list."""
    if all_daemons is None:
        all_daemons = _DAEMONS
    raw = env.get("BOT_BOTTLE_GATEWAY_DAEMONS", "").strip()
    if not raw:
-        return tuple(all_daemons)
+        return tuple(d for d in all_daemons if d.name not in _OPT_IN_DAEMONS)
    wanted = {n.strip() for n in raw.split(",") if n.strip()}
    return tuple(d for d in all_daemons if d.name in wanted)

@@ -136,7 +150,7 @@ def _pump(name: str, stream: IO[bytes]) -> None:

 def _spawn(spec: _DaemonSpec) -> subprocess.Popen[bytes]:
    env = _env_for_daemon(spec.name, dict(os.environ))
-    proc = subprocess.Popen(
+    proc = subprocess.Popen(  # pylint: disable=consider-using-with
        _argv_for_daemon(spec.name, spec.argv, env),
        stdout=subprocess.PIPE,
        stderr=subprocess.STDOUT,
@@ -183,6 +197,14 @@ class _Supervisor:
                except ProcessLookupError:
                    pass

+    def _sigkill_all(self) -> None:
+        for _, p in self.procs:
+            if p.poll() is None:
+                try:
+                    p.kill()
+                except ProcessLookupError:
+                    pass
+
    def request_restart(self, daemon_name: str) -> bool:
        """Queue a daemon restart for the main loop to process.

@@ -235,12 +257,7 @@ class _Supervisor:
                        f"grace ({_GRACE_SECONDS:.0f}s) elapsed; SIGKILL on "
                        f"{', '.join(still_running)}"
                    )
-                    for _, p in self.procs:
-                        if p.poll() is None:
-                            try:
-                                p.kill()
-                            except ProcessLookupError:
-                                pass
+                    self._sigkill_all()

        done = all(p.poll() is not None for _, p in self.procs)
        if done:
@@ -361,7 +378,10 @@ def main(argv: Sequence[str] | None = None) -> int:
    # --signal HUP <bundle>` after writing routes.yaml. The kernel
    # delivers SIGHUP to PID 1 (this supervisor); forward it to
    # mitmdump so it reloads its addon.
-    signal.signal(signal.SIGHUP, lambda *_: sup.forward_signal(signal.SIGHUP, "egress"))  # type: ignore
+    signal.signal(
+        signal.SIGHUP,
+        lambda *_: sup.forward_signal(signal.SIGHUP, "egress"),  # type: ignore[misc]
+    )

    while not sup.tick():
        time.sleep(_POLL_INTERVAL)
@@ -112,8 +112,10 @@ class GitGate(ABC):
        access_hook = stage_dir / "git_gate_access_hook.sh"
        access_hook.write_text(git_gate_render_access_hook())
        # 0o700 (not 0o600): git daemon execs --access-hook directly,
-        # not via `sh`, so the script needs the x bit. docker cp
-        # preserves source mode into the container.
+        # not via `sh`, so the script needs the x bit. The gateway copy
+        # does not necessarily preserve this mode (`docker cp` does, the
+        # Apple `container cp` does not), so provision_git_gate re-applies
+        # +x on the gateway side — see backend/docker/gateway_provision.py.
        access_hook.chmod(0o700)
        upstreams_with_files: list[GitGateUpstream] = []
        for u in upstreams:
@@ -14,15 +14,12 @@ import shlex
 from dataclasses import dataclass
 from pathlib import Path

+from .constants import GIT_GATE_TIMEOUT_SECS, IDENTITY_HEADER
 from .manifest import ManifestBottle, ManifestGitEntry

 # Short network alias for git-gate inside the gateway. The
 # agent's `.gitconfig` insteadOf rewrites resolve through this name.
 GIT_GATE_HOSTNAME = "git-gate"
-# Shared timeout (seconds) for all git-gate subprocess and CGI calls:
-# git daemon (--timeout/--init-timeout), the access-hook subprocess in
-# git_http_backend, and the git http-backend CGI subprocess.
-GIT_GATE_TIMEOUT_SECS = 15


@dataclass(frozen=True)
@@ -75,6 +72,7 @@ def _gitconfig_validate_value(field: str, value: str) -> None:

 def git_gate_render_gitconfig(
    entries: tuple[ManifestGitEntry, ...], gate_host: str, *, scheme: str = "git",
+    identity_token: str = "",
 ) -> str:
    """Render the agent's ~/.gitconfig content for git-gate
    `insteadOf` rewrites. Pure host-side, no docker / VM;
@@ -96,6 +94,15 @@ def git_gate_render_gitconfig(
        "# the upstream bidirectionally (gitleaks-scanned push;\n",
        "# fetch-from-upstream-before-every-upload-pack via access-hook).\n",
    ]
+    # Over the smart-HTTP transport (VM backends), attach the per-bottle
+    # identity token as a request header on requests to the gate, scoped to
+    # its URL so it never goes to any other remote. git-http requires it (the
+    # gateway's mandatory (source_ip, token) attribution). git:// (single-tenant
+    # docker) carries no header — attribution there is the network alias.
+    if identity_token and scheme == "http":
+        _gitconfig_validate_value("identity_token", identity_token)
+        out.append(f'[http "http://{gate_host}/"]\n')
+        out.append(f"\textraHeader = {IDENTITY_HEADER}: {identity_token}\n")
    for entry in entries:
        _gitconfig_validate_value(f"repos[{entry.Name!r}].url", entry.Upstream)
        out.append(f'[url "{scheme}://{gate_host}/{entry.Name}.git"]\n')
@@ -412,18 +419,24 @@ PY
 while IFS=' ' read -r old new ref; do
  [ -z "$ref" ] && continue
  [ "$new" = "$zero" ] && continue
-  if [ "$old" = "$zero" ]; then
-    # New ref: scan only the commits this push introduces — those
-    # reachable from $new but not from any ref the gate already has.
-    # Everything already on the gate arrived via upstream mirror-fetch
-    # or a previously gitleaks-scanned push, so it's already-upstream
-    # or already-scanned; re-scanning it (the old `$new` full-ancestry
-    # range) only resurfaces historical findings and blocks every new
-    # branch. See PRD 0028 / issue #106.
-    log_opts="$new --not --all"
-  else
-    log_opts="$old..$new"
-  fi
+  # Scan only the commits this push introduces — those reachable from
+  # $new but not from any ref the gate already has. Everything already
+  # on the gate arrived via upstream mirror-fetch or a previously
+  # gitleaks-scanned push, so it's already-upstream or already-scanned;
+  # re-scanning it only resurfaces historical fixture findings.
+  #
+  # Applies to both new refs and updates. The old existing-branch range
+  # `$old..$new` walks commits reachable from the new tip but not the
+  # *old branch tip*: on a rebase/force-push onto a freshly-advanced
+  # main that pulls in all of main's new history (incl. the deliberate
+  # sandbox-escape gitleaks fixtures), blocking the push. `--not --all`
+  # excludes anything already on the gate regardless of ancestry, so it
+  # is also correct for non-fast-forward pushes (a rebase can skip
+  # commits off the direct path). Security-equivalent per PRD 0028's
+  # analysis: the bare repo's refs come only from trusted upstream
+  # mirror-fetch or gitleaks-gated pushes.
+  # See PRD 0028 (open question) / issues #106, #346.
+  log_opts="$new --not --all"
  echo "git-gate: gitleaks scanning $ref ($log_opts)" >&2
  if ! gitleaks git --log-opts="$log_opts" --no-banner --redact 1>&2; then
    echo "git-gate: gitleaks rejected push to $ref" >&2
@@ -7,14 +7,13 @@ wrapper serves the same `/git/*.git` bare repos through
 `git http-backend`, so pre-receive and upstream forwarding remain the
 git-gate enforcement point.

-Consolidated (PRD 0070): when `BOT_BOTTLE_ORCHESTRATOR_URL` is set, one
-shared gateway serves every bottle, and each request is served from the
-calling bottle's repo namespace (`<root>/<bottle_id>`), attributed from
-the unspoofable source IP via the orchestrator. Per-repo credentials +
+One shared gateway serves every bottle (PRD 0070): each request is served
+from the calling bottle's repo namespace (`<root>/<bottle_id>`), attributed
+from the unspoofable source IP via the orchestrator. Per-repo credentials +
 hooks scope by repo directory, so isolating the *root* per bottle isolates
-its creds too. Unattributed clients fail closed (404). Unset → the legacy
-per-bottle single-tenant flat root, unchanged — a transitional path that
-gets stripped out once every backend runs the consolidated gateway.
+its creds too. Unattributed clients — and a missing/unreachable orchestrator
+— fail closed (404). `BOT_BOTTLE_ORCHESTRATOR_URL` is mandatory: there is no
+single-tenant flat-root fallback.
 """

 from __future__ import annotations
@@ -27,36 +26,19 @@ from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 from pathlib import Path
 from urllib.parse import urlsplit

-# policy_resolver ships flat alongside this file in the gateway
-# image (see Dockerfile.gateway); the bot_bottle.* fallback is the
-# host-side / test path. Mirrors egress_addon's import shape.
-try:
-    from policy_resolver import (  # type: ignore[import-not-found]
-        PolicyResolveError,
-        PolicyResolver,
-    )
-except ImportError:  # pragma: no cover - host-side path
-    from bot_bottle.policy_resolver import PolicyResolveError, PolicyResolver
+from bot_bottle.constants import GIT_GATE_TIMEOUT_SECS, IDENTITY_HEADER
+from bot_bottle.policy_resolver import PolicyResolveError, PolicyResolver


 DEFAULT_PORT = 9420

-# Consolidated (multi-tenant) mode: when this points at the per-host
-# orchestrator's control plane, the backend serves each request from the
-# *calling* bottle's repo namespace, selected by source IP, instead of a
-# single flat repo root. Unset → legacy per-bottle single-tenant mode
-# (unchanged). Same env the egress addon reads, so one orchestrator setting
-# flips the whole shared gateway multi-tenant.
+# The per-host orchestrator control plane the backend attributes each request
+# to, serving from the *calling* bottle's repo namespace selected by source IP.
+# Mandatory — the same env the egress addon requires; there is no single flat
+# repo-root fallback.
 ORCHESTRATOR_URL_ENV = "BOT_BOTTLE_ORCHESTRATOR_URL"

-# App-layer identity token (defense-in-depth over the source-IP invariant);
-# the agent injects it, the backend reads it for attribution and never
-# forwards it to `git http-backend`. Mirrors egress_addon.IDENTITY_HEADER
-# (duplicated, not imported: egress_addon pulls in mitmproxy).
-IDENTITY_HEADER = "x-bot-bottle-identity"
-
-# Default flat repo root (single-tenant, and the base under which
-# consolidated mode nests each sandbox's namespace).
+# The base under which each bottle's `<bottle_id>` repo namespace is nested.
 DEFAULT_REPO_ROOT = "/git"


@@ -71,25 +53,16 @@ class ResolverLike(typing.Protocol):


 def resolve_sandbox_root(
-    resolver: "ResolverLike | None",
+    resolver: "ResolverLike",
    base_root: Path,
    source_ip: str,
    identity_token: str = "",
 ) -> Path | None:
-    """The per-sandbox repo root to serve this request from, or None to
-    deny (404).
-
-    Single-tenant (`resolver is None`): the flat `base_root`, unchanged.
-    NOTE: this legacy per-bottle single-tenant path is transitional — it
-    will be stripped out once every backend runs the consolidated gateway
-    (PRD 0070), leaving only the source-IP-attributed path below.
-
-    Consolidated: `base_root/<bottle_id>`, where the sandbox is attributed
-    from the source IP via the orchestrator. Fail-closed — an unattributed
-    client, a resolver error, or a namespace that would escape `base_root`
-    all deny, so one sandbox can never reach another's repos."""
-    if resolver is None:
-        return base_root
+    """The per-sandbox repo root to serve this request from — `base_root/
+    <bottle_id>`, where the sandbox is attributed from the source IP via the
+    orchestrator — or None to deny (404). Fail-closed: an unattributed client, a
+    resolver error, or a namespace that would escape `base_root` all deny, so one
+    sandbox can never reach another's repos."""
    try:
        bottle_id = resolver.resolve_bottle_id(source_ip, identity_token)
    except PolicyResolveError:
@@ -102,13 +75,6 @@ def resolve_sandbox_root(
        return None  # bottle_id tried to escape the root → deny
    return namespace

-# Mirrors git_gate_render.GIT_GATE_TIMEOUT_SECS. Duplicated rather than
-# imported: this module ships as a flat top-level sibling in the gateway
-# bundle image (see Dockerfile.gateway), not as part of the bot_bottle
-# package, so `bot_bottle.git_gate` and its dependency chain aren't
-# available at runtime.
-GIT_GATE_TIMEOUT_SECS = 15
-
 # Bound memory use while still allowing ordinary git push packfiles.
 MAX_BODY_BYTES = 100 * 1024 * 1024

@@ -123,12 +89,13 @@ class GitHttpHandler(BaseHTTPRequestHandler):
        self._run_backend()

    def _sandbox_root(self) -> Path | None:
-        """This request's per-sandbox repo root, or None to deny. Single-tenant
-        unless the server was started with a resolver (consolidated mode), in
-        which case the root is the calling sandbox's source-IP-selected
-        namespace. `GIT_PROJECT_ROOT` keeps git's own env-var name."""
+        """This request's per-sandbox repo root (the calling bottle's source-IP-
+        selected `<base>/<bottle_id>` namespace), or None to deny. `GIT_PROJECT_
+        ROOT` keeps git's own env-var name."""
        base = Path(os.environ.get("GIT_PROJECT_ROOT", DEFAULT_REPO_ROOT))
        resolver = getattr(self.server, "policy_resolver", None)
+        if resolver is None:
+            return None  # server started without a resolver (misconfig) → deny
        token = self.headers.get(IDENTITY_HEADER, "")
        return resolve_sandbox_root(resolver, base, self.client_address[0], token)

@@ -148,12 +115,24 @@ class GitHttpHandler(BaseHTTPRequestHandler):
                "GIT_GATE_ACCESS_HOOK", "/etc/git-gate/access-hook",
            )
            peer = self.client_address[0]
-            hook = subprocess.run(
-                [hook_path, "upload-pack", str(repo_dir), peer, peer],
-                capture_output=True,
-                check=False,
-                timeout=GIT_GATE_TIMEOUT_SECS,
-            )
+            try:
+                hook = subprocess.run(
+                    [hook_path, "upload-pack", str(repo_dir), peer, peer],
+                    capture_output=True,
+                    check=False,
+                    timeout=GIT_GATE_TIMEOUT_SECS,
+                )
+            except (OSError, subprocess.SubprocessError) as exc:
+                # The access-hook couldn't be run (missing, not executable,
+                # timed out, …). Fail closed with a real HTTP error rather
+                # than letting the exception kill the handler thread — an
+                # unhandled exception closes the socket with no response, which
+                # the client sees as an opaque "empty reply from server".
+                self.log_message(
+                    "access-hook could not run for %s: %s", parsed.path, exc,
+                )
+                self.send_error(503, "git-gate access-hook unavailable")
+                return
            if hook.returncode != 0:
                detail = (hook.stderr or hook.stdout).decode(
                    "utf-8", errors="replace",
@@ -188,14 +167,11 @@ class GitHttpHandler(BaseHTTPRequestHandler):
            "SERVER_PORT": str(self.server.server_port),  # type: ignore
            "SERVER_PROTOCOL": self.request_version,
        })
-        # Consolidated mode: attribute the gitleaks-allow supervise proposal
-        # (written by receive-pack's pre-receive hook, a child of the CGI we
-        # spawn below) to the calling bottle. The namespaced root is
-        # `<base>/<bottle_id>`, so its final component is the bottle id — the
-        # same per-bottle key egress uses. Single-tenant leaves the hook's
-        # container-stamped SUPERVISE_BOTTLE_SLUG untouched.
-        if getattr(self.server, "policy_resolver", None) is not None:
-            env["SUPERVISE_BOTTLE_SLUG"] = sandbox_root.name
+        # Attribute the gitleaks-allow supervise proposal (written by
+        # receive-pack's pre-receive hook, a child of the CGI we spawn below) to
+        # the calling bottle. The namespaced root is `<base>/<bottle_id>`, so its
+        # final component is the bottle id — the same per-bottle key egress uses.
+        env["SUPERVISE_BOTTLE_SLUG"] = sandbox_root.name
        for header, variable in (
            ("accept", "HTTP_ACCEPT"),
            ("content-encoding", "HTTP_CONTENT_ENCODING"),
@@ -285,14 +261,20 @@ class GitHttpHandler(BaseHTTPRequestHandler):

 def main() -> int:
    port = int(os.environ.get("GIT_HTTP_PORT", str(DEFAULT_PORT)))
-    server = ThreadingHTTPServer(("0.0.0.0", port), GitHttpHandler)
    orch_url = os.environ.get(ORCHESTRATOR_URL_ENV, "").strip()
-    # Consolidated mode: resolve each request's sandbox namespace by source
-    # IP. Absent → single-tenant (flat repo root); behaviour unchanged.
-    resolver = PolicyResolver(orch_url) if orch_url else None
-    server.policy_resolver = resolver  # type: ignore[attr-defined]
-    mode = "multi-tenant" if orch_url else "single-tenant"
-    sys.stdout.write(f"git-http listening on 0.0.0.0:{port} ({mode})\n")
+    if not orch_url:
+        # Resolver-only: without an orchestrator the backend can't attribute a
+        # request to a bottle namespace, so it must not serve (fail-closed).
+        sys.stderr.write(
+            f"git-http: {ORCHESTRATOR_URL_ENV} is required "
+            "(no single-tenant flat-root fallback)\n"
+        )
+        return 1
+    server = ThreadingHTTPServer(("0.0.0.0", port), GitHttpHandler)
+    # Resolve each request's sandbox namespace by source IP against the
+    # orchestrator control plane.
+    server.policy_resolver = PolicyResolver(orch_url)  # type: ignore[attr-defined]
+    sys.stdout.write(f"git-http listening on 0.0.0.0:{port} (multi-tenant)\n")
    sys.stdout.flush()
    server.serve_forever()
    return 0
@@ -0,0 +1,42 @@
+"""Shared helpers for cached-image quickstart stale checks."""
+
+from __future__ import annotations
+
+from datetime import datetime, timezone
+from pathlib import Path
+
+try:
+    from .config_store import ConfigStore
+except ImportError:
+    from config_store import ConfigStore  # type: ignore[import-not-found]  # pylint: disable=import-error,no-name-in-module
+
+
+class StaleImageError(Exception):
+    """Raised when a cached image or artifact exceeds the configured staleness
+    threshold. Callers can catch this to prompt interactively; headless paths
+    let it propagate as a fatal error."""
+
+
+def check_stale(label: str, created_at: datetime) -> None:
+    """Raise StaleImageError if `created_at` is older than the configured
+    stale-warning threshold. Negative threshold disables the check."""
+    threshold_days = ConfigStore().cached_image_stale_warning_days()
+    if threshold_days < 0:
+        return
+    now = datetime.now(timezone.utc)
+    created = created_at.astimezone(timezone.utc)
+    age = now - created
+    if age.total_seconds() <= threshold_days * 86400:
+        return
+    raise StaleImageError(
+        f"cached {label} is {age.days} day(s) old; "
+        "quickstart does not verify it matches the current Dockerfile/context"
+    )
+
+
+def check_stale_path(label: str, path: Path) -> None:
+    """Raise StaleImageError if `path`'s mtime exceeds the staleness threshold."""
+    check_stale(label, datetime.fromtimestamp(path.stat().st_mtime, tz=timezone.utc))
+
+
+__all__ = ["StaleImageError", "check_stale", "check_stale_path"]
@@ -16,6 +16,7 @@ import secrets
 from pathlib import Path

 from .. import log
+from ..store_manager import StoreManager
 from .broker import LaunchBroker, StubBroker
 from .control_plane import make_server
 from .docker_broker import DockerBroker
@@ -45,6 +46,11 @@ def main(argv: list[str] | None = None) -> int:

    registry = RegistryStore(args.db)
    registry.migrate()
+    # One DB per host: the supervise queue + audit tables live in the SAME
+    # SQLite file the registry owns, so the control plane is the single
+    # source of truth. The in-VM supervise daemon writes here; the host
+    # operator reaches it over HTTP (never a second, disconnected DB).
+    StoreManager(registry.db_path).migrate()

    # An ephemeral signing secret ties the orchestrator (signer) to its
    # broker (verifier). 'stub' records launches instead of starting
@@ -15,11 +15,25 @@ from __future__ import annotations
 import json
 import urllib.error
 import urllib.request
+from collections.abc import Iterable
 from dataclasses import dataclass

+from ..paths import host_control_plane_token
+from .control_plane import CONTROL_AUTH_HEADER
+
 DEFAULT_TIMEOUT_SECONDS = 5.0


+def _host_auth_token() -> str:
+    """The per-host control-plane secret, or "" if it can't be read. "" means
+    'send no auth header' — correct against an open (unconfigured) control
+    plane, and harmlessly rejected by a secured one."""
+    try:
+        return host_control_plane_token()
+    except OSError:
+        return ""
+
+
 class OrchestratorClientError(RuntimeError):
    """A control-plane call failed (unreachable, or an unexpected status)."""

@@ -34,11 +48,24 @@ class RegisteredBottle:


 class OrchestratorClient:
-    """Trusted host-side client for the orchestrator control plane."""
+    """Trusted host-side client for the orchestrator control plane.

-    def __init__(self, base_url: str, *, timeout: float = DEFAULT_TIMEOUT_SECONDS) -> None:
+    Presents the per-host control-plane secret on every call (the header the
+    control plane requires on all routes but `/health`). The secret is read
+    from the host file — this client only ever runs host-side (CLI, launcher,
+    discovery), so it can read what an agent can't. `auth_token` is overridable
+    for tests; the default reads the host file, minting it on first use."""
+
+    def __init__(
+        self,
+        base_url: str,
+        *,
+        timeout: float = DEFAULT_TIMEOUT_SECONDS,
+        auth_token: str | None = None,
+    ) -> None:
        self._base = base_url.rstrip("/")
        self._timeout = timeout
+        self._auth_token = auth_token if auth_token is not None else _host_auth_token()

    def _request(
        self, method: str, path: str, body: dict[str, object] | None = None,
@@ -49,6 +76,8 @@ class OrchestratorClient:
        callers can treat 404 as a meaningful "no such bottle"."""
        data = json.dumps(body).encode() if body is not None else None
        headers = {"Content-Type": "application/json"} if data is not None else {}
+        if self._auth_token:
+            headers[CONTROL_AUTH_HEADER] = self._auth_token
        req = urllib.request.Request(
            f"{self._base}{path}", data=data, method=method, headers=headers,
        )
@@ -119,6 +148,20 @@ class OrchestratorClient:
            raise OrchestratorClientError(f"teardown {bottle_id}: HTTP {status}")
        return True

+    def reconcile(
+        self, live_source_ips: Iterable[str], *, grace_seconds: float | None = None,
+    ) -> list[str]:
+        """Drop registry rows for bottles that are no longer running
+        (`POST /reconcile`), returning the reaped bottle ids. `live_source_ips`
+        is the caller's enumeration of its live bottles — the orchestrator
+        can't see the backend from inside the infra container."""
+        body: dict[str, object] = {"live_source_ips": list(live_source_ips)}
+        if grace_seconds is not None:
+            body["grace_seconds"] = grace_seconds
+        payload = self._ok("POST", "/reconcile", body)
+        reaped = payload.get("reaped")
+        return [r for r in reaped if isinstance(r, str)] if isinstance(reaped, list) else []
+
    def set_policy(self, bottle_id: str, policy: str) -> bool:
        """Live-reload a bottle's policy (`PUT /bottles/<id>/policy`). False on
        404 (unknown bottle)."""
@@ -135,10 +178,78 @@ class OrchestratorClient:
        bottles = payload.get("bottles")
        return bottles if isinstance(bottles, list) else []

+    # --- supervise queue (operator TUI) ------------------------------------
+
+    def supervise_pending(self) -> list[dict[str, object]]:
+        """Pending supervise proposals across all bottles
+        (`GET /supervise/proposals`)."""
+        payload = self._ok("GET", "/supervise/proposals")
+        proposals = payload.get("proposals")
+        return proposals if isinstance(proposals, list) else []
+
+    def supervise_respond(
+        self,
+        proposal_id: str,
+        *,
+        bottle_slug: str,
+        decision: str,
+        notes: str = "",
+        final_file: str | None = None,
+    ) -> None:
+        """Record an operator decision (`POST /supervise/respond`). `decision`
+        is approve/modify/reject. Raises `OrchestratorClientError` if the
+        proposal is gone or the bottle can no longer be applied to (409)."""
+        body: dict[str, object] = {
+            "proposal_id": proposal_id,
+            "bottle_slug": bottle_slug,
+            "decision": decision,
+            "notes": notes,
+        }
+        if final_file is not None:
+            body["final_file"] = final_file
+        self._ok("POST", "/supervise/respond", body)
+
+
+def discover_orchestrator_url(*, timeout: float = 2.0) -> str:
+    """The URL of the one running per-host orchestrator control plane, probing
+    the backends' well-known control-plane addresses (both on port 8099):
+    docker publishes it on loopback; the firecracker infra VM serves it on the
+    orchestrator TAP. Returns the first that answers `/health`; raises if none
+    do (no orchestrator up — launch a bottle first)."""
+    candidates: list[str] = []
+    try:  # docker: loopback-published control plane
+        from .lifecycle import DEFAULT_PORT as _DOCKER_PORT
+        candidates.append(f"http://127.0.0.1:{_DOCKER_PORT}")
+    except Exception:  # noqa: BLE001 — backend optional
+        candidates.append("http://127.0.0.1:8099")
+    try:  # firecracker: infra VM control plane on the orchestrator TAP
+        from ..backend.firecracker import netpool
+        from ..backend.firecracker.infra_vm import CONTROL_PLANE_PORT
+        candidates.append(
+            f"http://{netpool.orch_slot().guest_ip}:{CONTROL_PLANE_PORT}")
+    except Exception:  # noqa: BLE001 — backend optional / not firecracker
+        pass
+    try:  # macOS: infra container control plane on its host-only address
+        from ..backend.macos_container.infra import probe_control_plane_url
+        url = probe_control_plane_url()
+        if url:
+            candidates.append(url)
+    except Exception:  # noqa: BLE001 — backend optional / not macOS
+        pass
+    for url in candidates:
+        if OrchestratorClient(url, timeout=timeout).health():
+            return url
+    raise OrchestratorClientError(
+        "no running orchestrator control plane found (tried "
+        + ", ".join(candidates)
+        + "); launch a bottle first"
+    )
+

 __all__ = [
    "OrchestratorClient",
    "OrchestratorClientError",
    "RegisteredBottle",
    "DEFAULT_TIMEOUT_SECONDS",
+    "discover_orchestrator_url",
 ]
@@ -0,0 +1,107 @@
+"""Per-host orchestrator configuration store (settings in bot-bottle.db).
+
+Co-tenants the shared `bot-bottle.db` via the `DbStore` framework.  Settings
+are readable by the host launch path directly (no HTTP round-trip to the
+orchestrator), so they take effect even before the orchestrator is reachable.
+"""
+
+from __future__ import annotations
+
+import os
+import sqlite3
+from pathlib import Path
+
+from ..db_store import DbStore
+from ..migrations import TableMigrations
+from ..paths import host_db_path
+
+TEARDOWN_TIMEOUT_ENV = "BOT_BOTTLE_ORCHESTRATOR_TEARDOWN_TIMEOUT_SECONDS"
+DEFAULT_TEARDOWN_TIMEOUT_SECONDS = 30.0
+
+_MIGRATIONS = TableMigrations(
+    "orchestrator_config",
+    [
+        """
+        CREATE TABLE IF NOT EXISTS orchestrator_config (
+            id                       INTEGER PRIMARY KEY CHECK (id = 1),
+            teardown_timeout_seconds REAL
+        )
+        """,
+    ],
+)
+
+
+class OrchestratorConfigStore(DbStore):
+    """Orchestrator settings in the shared host DB."""
+
+    def __init__(self, db_path: Path | None = None) -> None:
+        super().__init__(db_path or host_db_path(), _MIGRATIONS)
+
+    def _connect(self) -> sqlite3.Connection:
+        conn = super()._connect()
+        conn.execute("PRAGMA busy_timeout=5000")
+        return conn
+
+    def get_teardown_timeout_seconds(self) -> float | None:
+        """Return the configured teardown timeout, or None if not set."""
+        try:
+            with self._connection() as conn:
+                row = conn.execute(
+                    "SELECT teardown_timeout_seconds FROM orchestrator_config WHERE id = 1"
+                ).fetchone()
+        except sqlite3.OperationalError:
+            return None
+        return row["teardown_timeout_seconds"] if row else None
+
+    def set_teardown_timeout_seconds(self, value: float) -> None:
+        """Persist the teardown timeout."""
+        with self._connection() as conn:
+            conn.execute(
+                "INSERT OR REPLACE INTO orchestrator_config"
+                " (id, teardown_timeout_seconds) VALUES (1, ?)",
+                (value,),
+            )
+        self._chmod()
+
+    def delete_teardown_timeout_seconds(self) -> bool:
+        """Clear the stored teardown timeout. Returns True if a value existed."""
+        with self._connection() as conn:
+            cur = conn.execute(
+                "UPDATE orchestrator_config SET teardown_timeout_seconds = NULL"
+                " WHERE id = 1 AND teardown_timeout_seconds IS NOT NULL"
+            )
+        return cur.rowcount > 0
+
+
+def resolve_teardown_timeout(db_path: Path | None = None) -> float:
+    """Return the teardown timeout to use, in priority order:
+
+    1. ``BOT_BOTTLE_ORCHESTRATOR_TEARDOWN_TIMEOUT_SECONDS`` env var
+    2. ``teardown_timeout_seconds`` in the orchestrator config DB
+    3. ``DEFAULT_TEARDOWN_TIMEOUT_SECONDS`` (30 s)
+    """
+    raw = os.environ.get(TEARDOWN_TIMEOUT_ENV, "").strip()
+    if raw:
+        try:
+            value = float(raw)
+            if value > 0:
+                return value
+        except ValueError:
+            pass
+
+    store = OrchestratorConfigStore(db_path)
+    if not store.is_migrated():
+        store.migrate()
+    db_value = store.get_teardown_timeout_seconds()
+    if db_value is not None and db_value > 0:
+        return db_value
+
+    return DEFAULT_TEARDOWN_TIMEOUT_SECONDS
+
+
+__all__ = [
+    "OrchestratorConfigStore",
+    "resolve_teardown_timeout",
+    "TEARDOWN_TIMEOUT_ENV",
+    "DEFAULT_TEARDOWN_TIMEOUT_SECONDS",
+]
@@ -13,9 +13,16 @@ vsock / unix-socket portability caveats):
    PUT    /bottles/<bottle_id>/policy -> 200 {"updated": true} | 404  (live reload)
                                         body: {"policy"}
    DELETE /bottles/<bottle_id>        -> 200 {"torn_down": true} | 404  (teardown)
+    POST   /reconcile                 -> 200 {"reaped": [bottle_id, ...]}
+                                         body: {"live_source_ips": [...],
+                                                ["grace_seconds"]}
    POST   /attribute                 -> 200 {"bottle_id"} | 403
    POST   /resolve                   -> 200 {"bottle_id","policy"} | 403
                                         body: {"source_ip","identity_token"}
+    GET    /supervise/proposals       -> 200 {"proposals": [ <proposal>, ...]}
+    POST   /supervise/respond         -> 200 {"responded": true} | 409  (operator)
+                                         body: {"proposal_id","bottle_slug",
+                                                "decision", ["notes"],["final_file"]}

 `POST /bottles` / `DELETE` drive the full launch lifecycle: they mint (or
 tear down) the bottle in the registry AND broker the backend-native launch
@@ -30,6 +37,7 @@ returned only once, to the caller that launches the bottle.

 from __future__ import annotations

+import hmac
 import http.server
 import json
 import os
@@ -38,11 +46,20 @@ import sys
 import typing
 from urllib.parse import urlsplit

+from ..paths import CONTROL_PLANE_TOKEN_ENV
 from .service import Orchestrator

 # JSON body payload type (parsed request / rendered response).
 Json = dict[str, object]

+# The request header carrying the per-host control-plane secret. Every route
+# except `GET /health` requires it (see `dispatch`). The trusted callers hold
+# the secret (the gateway's PolicyResolver, the host CLI's OrchestratorClient);
+# an agent that can merely *reach* the port cannot present it, so it can't
+# enumerate bottles, rewrite policy, read injected upstream tokens, or approve
+# its own supervise proposals.
+CONTROL_AUTH_HEADER = "x-bot-bottle-control-auth"
+

 def _parse_json_object(body: bytes) -> Json:
    """Parse a JSON object body. Raises ValueError for non-objects / bad JSON."""
@@ -55,15 +72,29 @@ def _parse_json_object(body: bytes) -> Json:


 def dispatch(  # pylint: disable=too-many-return-statements,too-many-branches
-    orch: Orchestrator, method: str, path: str, body: bytes
+    orch: Orchestrator, method: str, path: str, body: bytes, *, authorized: bool = True,
 ) -> tuple[int, Json]:
    """Route one control-plane request to a (status, payload) pair. Pure —
-    no I/O beyond the orchestrator — so it is fully testable without a socket."""
+    no I/O beyond the orchestrator — so it is fully testable without a socket.
+
+    `authorized` is whether the request presented the control-plane secret (or
+    no secret is configured — see `ControlPlaneServer`). Every route except
+    `GET /health` requires it: the source-IP + identity-token checks inside
+    `/resolve` and `/attribute` authenticate the *bottle* a request is about,
+    not the *caller*, so without this gate any agent that can reach the port
+    could rewrite another bottle's policy, read the injected upstream tokens,
+    or approve its own supervise proposals. Defaults True so unit tests of the
+    routing logic don't have to thread it through."""
    route = urlsplit(path).path.rstrip("/") or "/"

    if method == "GET" and route == "/health":
        return 200, {"status": "ok"}

+    if not authorized:
+        # Everything below is a trusted-caller operation. Deny before touching
+        # the registry / broker / supervise store.
+        return 401, {"error": "control-plane authentication required"}
+
    if method == "GET" and route == "/gateway":
        return 200, orch.gateway_status()

@@ -113,6 +144,27 @@ def dispatch(  # pylint: disable=too-many-return-statements,too-many-branches
            return 200, {"torn_down": True}
        return 404, {"error": "no such bottle"}

+    if method == "POST" and route == "/reconcile":
+        # Host-driven self-heal: the caller enumerates its live bottles (only
+        # the host can see the backend) and the orchestrator drops rows for
+        # every other active bottle. Trusted-caller only — an agent that could
+        # reach this would be able to unregister its neighbours.
+        try:
+            data = _parse_json_object(body)
+        except ValueError as e:
+            return 400, {"error": f"invalid JSON: {e}"}
+        raw_ips = data.get("live_source_ips")
+        if not isinstance(raw_ips, list):
+            return 400, {"error": "live_source_ips (list of strings) is required"}
+        live = [ip for ip in raw_ips if isinstance(ip, str) and ip]
+        grace = data.get("grace_seconds")
+        kwargs = (
+            {"grace_seconds": float(grace)}
+            if isinstance(grace, (int, float)) and not isinstance(grace, bool)
+            else {}
+        )
+        return 200, {"reaped": orch.reconcile(live, **kwargs)}
+
    if method == "POST" and route == "/attribute":
        try:
            data = _parse_json_object(body)
@@ -127,10 +179,44 @@ def dispatch(  # pylint: disable=too-many-return-statements,too-many-branches
            return 403, {"error": "unattributed"}
        return 200, {"bottle_id": rec.bottle_id}

+    if method == "GET" and route == "/supervise/proposals":
+        # Operator TUI: pending supervise proposals across all bottles.
+        return 200, {"proposals": orch.supervise_pending()}
+
+    if method == "POST" and route == "/supervise/respond":
+        # Operator decision: apply (approve/modify rewrites egress policy),
+        # write the queued response, audit — all server-side on the one DB.
+        try:
+            data = _parse_json_object(body)
+        except ValueError as e:
+            return 400, {"error": f"invalid JSON: {e}"}
+        proposal_id = data.get("proposal_id")
+        bottle_slug = data.get("bottle_slug")
+        decision = data.get("decision")
+        if not (isinstance(proposal_id, str) and proposal_id):
+            return 400, {"error": "proposal_id (string) is required"}
+        if not (isinstance(bottle_slug, str) and bottle_slug):
+            return 400, {"error": "bottle_slug (string) is required"}
+        if not (isinstance(decision, str) and decision):
+            return 400, {"error": "decision (string) is required"}
+        notes = data.get("notes")
+        final_file = data.get("final_file")
+        ok, err = orch.supervise_respond(
+            proposal_id,
+            bottle_slug=bottle_slug,
+            decision=decision,
+            notes=notes if isinstance(notes, str) else "",
+            final_file=final_file if isinstance(final_file, str) else None,
+        )
+        if ok:
+            return 200, {"responded": True}
+        return 409, {"error": err}
+
    if method == "POST" and route == "/resolve":
        # The per-request lookup the multi-tenant gateway makes: returns the
-        # bottle's policy. identity_token is OPTIONAL — absent means resolve
-        # by source IP alone (network-layer attribution).
+        # bottle's policy. Requires a matching (source_ip, identity_token)
+        # pair — a missing/empty/mismatched token fail-closes (403), no
+        # source-IP-only fallback.
        try:
            data = _parse_json_object(body)
        except ValueError as e:
@@ -171,8 +257,10 @@ class Handler(http.server.BaseHTTPRequestHandler):
        assert isinstance(server, ControlPlaneServer)
        length = int(self.headers.get("Content-Length") or 0)
        body = self.rfile.read(length) if length > 0 else b""
+        authorized = server.is_authorized(self.headers.get(CONTROL_AUTH_HEADER, ""))
        try:
-            status, payload = dispatch(server.orchestrator, method, self.path, body)
+            status, payload = dispatch(
+                server.orchestrator, method, self.path, body, authorized=authorized)
        except Exception as e:  # noqa: BLE001 — the control plane must stay up
            sys.stderr.write(f"orchestrator: {method} {self.path} failed: {e!r}\n")
            sys.stderr.flush()
@@ -198,15 +286,40 @@ class Handler(http.server.BaseHTTPRequestHandler):


 class ControlPlaneServer(socketserver.ThreadingMixIn, http.server.HTTPServer):
-    """Threading HTTP server that carries the orchestrator for its handlers."""
+    """Threading HTTP server that carries the orchestrator for its handlers.
+
+    Holds the per-host control-plane secret (from `$BOT_BOTTLE_CONTROL_PLANE_TOKEN`,
+    injected by the launcher into this container only). When a secret is set,
+    every route but `/health` requires it; when it is unset the server runs
+    **open** and says so loudly at startup — a fail-visible fallback for tests
+    and any backend that hasn't wired the secret yet (e.g. Firecracker, whose
+    nft boundary already blocks agents from the control-plane port)."""

    daemon_threads = True
    allow_reuse_address = True

    def __init__(self, address: tuple[str, int], orchestrator: Orchestrator) -> None:
        self.orchestrator = orchestrator
+        self._auth_token = os.environ.get(CONTROL_PLANE_TOKEN_ENV, "").strip()
+        if not self._auth_token:
+            sys.stderr.write(
+                "orchestrator: WARNING — no control-plane secret "
+                f"(${CONTROL_PLANE_TOKEN_ENV}); running WITHOUT caller "
+                "authentication. Any client that can reach this port can drive "
+                "it. Backends that put the control plane on an agent-reachable "
+                "network MUST set this.\n"
+            )
+            sys.stderr.flush()
        super().__init__(address, Handler)

+    def is_authorized(self, presented: str) -> bool:
+        """True iff the request may proceed past `/health`: either no secret is
+        configured (open mode) or the presented header matches it. Constant-time
+        compare so a wrong token leaks nothing timing-wise."""
+        if not self._auth_token:
+            return True
+        return hmac.compare_digest(presented, self._auth_token)
+

 def make_server(
    orchestrator: Orchestrator, host: str = "127.0.0.1", port: int = 0
@@ -216,4 +329,7 @@ def make_server(
    return ControlPlaneServer((host, port), orchestrator)


-__all__ = ["dispatch", "Handler", "ControlPlaneServer", "make_server", "Json"]
+__all__ = [
+    "dispatch", "Handler", "ControlPlaneServer", "make_server", "Json",
+    "CONTROL_AUTH_HEADER",
+]
@@ -23,6 +23,17 @@ import time
 from pathlib import Path

 from ..docker_cmd import run_docker
+from ..paths import (
+    CONTROL_PLANE_TOKEN_ENV,
+    host_control_plane_token,
+    host_db_path,
+)
+from ..supervise import DB_PATH_IN_CONTAINER
+
+# The host DB dir is bind-mounted here so the gateway's supervise daemon
+# writes its queued proposals into the ONE host DB (the same file the
+# orchestrator container opens and the operator reaches over HTTP).
+_SUPERVISE_DB_DIR_IN_CONTAINER = os.path.dirname(DB_PATH_IN_CONTAINER)

 # The gateway's mitmproxy writes its CA a beat after the container starts, so
 # reads poll for it rather than assuming it's there on a fresh launch.
@@ -54,6 +65,14 @@ GATEWAY_DOCKERFILE = "Dockerfile.gateway"
 _REPO_ROOT = Path(__file__).resolve().parents[2]


+def _host_db_dir() -> str:
+    """The host DB directory (created if missing), for the gateway's
+    supervise-DB bind-mount."""
+    db_dir = host_db_path().parent
+    db_dir.mkdir(parents=True, exist_ok=True)
+    return str(db_dir)
+
+
 class GatewayError(Exception):
    """The shared gateway failed to build/start/stop (non-zero `docker` exit)."""

@@ -100,16 +119,24 @@ class DockerGateway(Gateway):
        orchestrator_url: str = "",
        build_context: Path | None = None,
        dockerfile: str | None = GATEWAY_DOCKERFILE,
+        host_port_bindings: tuple[int, ...] = (),
    ) -> None:
        self.image_ref = image_ref
        self.name = name
        self.network = network
        # The control-plane URL the gateway's data plane resolves per bottle
        # against — reached by container name over docker DNS on the shared
-        # network (container↔container, no host firewall). Empty → single-tenant.
+        # network (container↔container, no host firewall). Mandatory to *run*
+        # the gateway (see `ensure_running`); empty is tolerated only for the
+        # construct-then-read-CA path (`ca_cert_pem` on an already-running
+        # container), which never launches a container.
        self._orchestrator_url = orchestrator_url
        self._build_context = build_context or _REPO_ROOT
        self._dockerfile = dockerfile
+        # Ports published on the host (0.0.0.0). Used by the Firecracker
+        # backend's dev-harness gateway so VMs can reach it via their TAP link;
+        # Docker's DNAT + the nft `ct status dnat accept` rule handle the rest.
+        self._host_port_bindings = host_port_bindings

    def image_exists(self) -> bool:
        return run_docker(["docker", "image", "inspect", self.image_ref]).returncode == 0
@@ -169,6 +196,16 @@ class DockerGateway(Gateway):
            )

    def ensure_running(self) -> None:
+        # Fail closed on a missing policy source. The data-plane daemons are
+        # resolver-only now (PRD 0070) — without an orchestrator URL egress
+        # raises, git-http exits 1, and supervise exits 2 — so launching a
+        # gateway without one would only crash-loop its daemons. Refuse here so
+        # the misconfiguration surfaces as a clear error, not a broken container.
+        if not self._orchestrator_url:
+            raise GatewayError(
+                "gateway requires an orchestrator URL to run "
+                "(resolver-only data plane; no single-tenant fallback)"
+            )
        # Recreate when the running container's image is stale (a rebuild),
        # so source changes to the gateway's flat daemons take effect — not
        # just when the container is absent.
@@ -187,13 +224,26 @@ class DockerGateway(Gateway):
            # Persist the self-generated CA so it survives restarts (agents
            # trust it) — see GATEWAY_CA_VOLUME.
            "--volume", f"{GATEWAY_CA_VOLUME}:{MITMPROXY_HOME}",
+            # Share the one host DB: the supervise daemon queues proposals
+            # into the same file the orchestrator (and the operator, over
+            # HTTP) reads — no second, disconnected DB in the container.
+            "--volume", f"{_host_db_dir()}:{_SUPERVISE_DB_DIR_IN_CONTAINER}",
+            "--env", f"SUPERVISE_DB_PATH={DB_PATH_IN_CONTAINER}",
        ]
-        if self._orchestrator_url:
-            # Makes the gateway's egress / git / supervise daemons multi-tenant:
-            # each request resolves source-IP -> policy against the control plane.
-            argv += ["--env", f"BOT_BOTTLE_ORCHESTRATOR_URL={self._orchestrator_url}"]
+        for port in self._host_port_bindings:
+            argv += ["--publish", f"0.0.0.0:{port}:{port}"]
+        run_env = dict(os.environ)
+        # The gateway's egress / git / supervise daemons resolve source-IP ->
+        # policy against the control plane per request (guaranteed non-empty by
+        # the check above).
+        argv += ["--env", f"BOT_BOTTLE_ORCHESTRATOR_URL={self._orchestrator_url}"]
+        # ...and present the control-plane secret on those /resolve calls (the
+        # control plane requires it). Bare `--env NAME` keeps the value off argv
+        # / `docker inspect`; only the gateway (not the agent) is given it.
+        argv += ["--env", CONTROL_PLANE_TOKEN_ENV]
+        run_env[CONTROL_PLANE_TOKEN_ENV] = host_control_plane_token()
        argv.append(self.image_ref)
-        proc = run_docker(argv)
+        proc = run_docker(argv, env=run_env)
        if proc.returncode != 0:
            raise GatewayError(f"gateway failed to start: {proc.stderr.strip()}")

@@ -1,17 +1,15 @@
 """Orchestrator + gateway lifecycle (PRD 0070, docker slice).

-Runs the orchestrator control plane **as a container** on the shared gateway
-network, alongside the gateway container. This is the PRD's "virtualize the
-orchestrator": container↔container between the gateway and the orchestrator
-avoids the host firewall (which drops container→host traffic), and the gateway
-reaches the control plane by container name over docker DNS. The host CLI
-reaches it via a published loopback port.
+Runs both the orchestrator control plane and the gateway data plane inside
+a single `bot-bottle-infra` container on the shared gateway network —
+matching the structure already used by the macOS and Firecracker backends.
+`gateway_init` is PID 1 and supervises both; the infra container is an
+idempotent per-host singleton.

-The orchestrator runs with the **register-only broker** — the *backend*
-launches agent containers (compose), so the orchestrator needs no docker
-socket. That keeps this control-plane container unprivileged; the host manages
-both containers. `ensure_running` is an idempotent singleton (fixed container
-names + the published port).
+The combined container replaces the prior two-container split
+(bot-bottle-orchestrator + bot-bottle-orch-gateway). The host CLI reaches
+the control plane via a published loopback port; gateway daemons reach it
+over 127.0.0.1 (same container).
 """

 from __future__ import annotations
@@ -25,50 +23,71 @@ from pathlib import Path

 from .. import log
 from ..docker_cmd import run_docker
-from ..paths import bot_bottle_root
-from .gateway import GATEWAY_IMAGE, GATEWAY_NETWORK, DockerGateway, GatewayError
+from ..paths import CONTROL_PLANE_TOKEN_ENV, bot_bottle_root, host_control_plane_token
+from ..supervise import DB_PATH_IN_CONTAINER
+from .gateway import (
+    GATEWAY_CA_VOLUME,
+    GATEWAY_DOCKERFILE,
+    GATEWAY_IMAGE,
+    GATEWAY_NETWORK,
+    GatewayError,
+    MITMPROXY_HOME,
+    _host_db_dir,
+)

 DEFAULT_PORT = 8099
-ORCHESTRATOR_NAME = "bot-bottle-orchestrator"
-ORCHESTRATOR_LABEL = "bot-bottle-orchestrator=1"
-# The control-plane's own runtime image — lean (python + the stdlib-only
-# `bot_bottle` package, bind-mounted at run time), distinct from the heavy
-# gateway data-plane image it used to borrow (#384). Env override for
-# operators pinning a published build.
+DEFAULT_STARTUP_TIMEOUT_SECONDS = 45.0
+
+INFRA_NAME = "bot-bottle-infra"
+INFRA_LABEL = "bot-bottle-infra=1"
+# The combined infra image: gateway data plane + orchestrator content.
+# Built from Dockerfile.infra (FROM gateway + COPY --from orchestrator).
+INFRA_IMAGE = os.environ.get("BOT_BOTTLE_INFRA_IMAGE", "bot-bottle-infra:latest")
+INFRA_DOCKERFILE = "Dockerfile.infra"
+# Baked as a container label so `ensure_running` can detect whether the
+# running container is executing the current bind-mounted source.
+INFRA_SOURCE_HASH_LABEL = "bot-bottle-infra-source-hash"
+
+# Orchestrator image: the single canonical definition of the control-plane
+# content (lean: python:3.12-slim + bot_bottle package, no mitmproxy/git).
+# Used as a build intermediate: `Dockerfile.infra` COPY --from this image.
 ORCHESTRATOR_IMAGE = os.environ.get(
    "BOT_BOTTLE_ORCHESTRATOR_IMAGE", "bot-bottle-orchestrator:latest"
 )
 ORCHESTRATOR_DOCKERFILE = "Dockerfile.orchestrator"
-# Baked onto the container as a label so `ensure_running` can tell whether the
-# running process is executing the *current* bind-mounted source — see
-# `_source_hash`.
-ORCHESTRATOR_SOURCE_HASH_LABEL = "bot-bottle-orchestrator-source-hash"

-# The repo root is bind-mounted into the control-plane container so
-# `python -m bot_bottle.orchestrator` resolves the package (the orchestrator
-# is stdlib-only, so the lean orchestrator image's python is enough).
-_REPO_ROOT = Path(__file__).resolve().parents[2]
-_APP_DIR = "/app"
+# The gateway daemons + orchestrator the infra container runs.
+# BOT_BOTTLE_GATEWAY_DAEMONS listing `orchestrator` opts it in to
+# gateway_init's supervise tree (see gateway_init._OPT_IN_DAEMONS).
+_INFRA_DAEMONS = "egress,git-http,supervise,orchestrator"
+
+# The bind-mount path for the live control-plane source inside the
+# container. Separate from /app so the gateway's baked scripts
+# (egress_addon.py, egress-entrypoint.sh) are not overlaid.
+_SRC_IN_CONTAINER = "/bot-bottle-src"
+# Bot-bottle host-root bind-mount inside the container (DB + state).
 _ROOT_IN_CONTAINER = "/bot-bottle-root"

+# The supervise daemon writes proposals into the host DB directory.
+_SUPERVISE_DB_DIR_IN_CONTAINER = os.path.dirname(DB_PATH_IN_CONTAINER)
+
 _HEALTH_POLL_SECONDS = 0.25
-DEFAULT_STARTUP_TIMEOUT_SECONDS = 45.0
 _HEALTH_REQUEST_TIMEOUT_SECONDS = 1.0

+_REPO_ROOT = Path(__file__).resolve().parents[2]
+

 class OrchestratorStartError(RuntimeError):
-    """The orchestrator container did not become healthy within the timeout."""
+    """The infra container did not become healthy within the timeout."""


-def _source_hash(repo_root: Path) -> str:
+def source_hash(repo_root: Path) -> str:
    """Content hash of the orchestrator's bind-mounted Python source (the
-    `bot_bottle` package the control-plane process imports). This only
-    changes when the code that would actually run inside the container
-    changes — `ensure_running` recreates the container on a mismatch and
-    otherwise leaves a healthy one alone, so a bottle launch that isn't
-    accompanied by a code change doesn't restart the process and drop every
-    *other* active bottle's in-memory egress tokens (`Orchestrator._tokens`
-    in `service.py`, never persisted to disk by design)."""
+    `bot_bottle` package the control-plane process imports). Changes only
+    when the code that would actually run changes — `ensure_running`
+    recreates the container on a mismatch so a code change takes effect,
+    but leaves a healthy up-to-date container alone to preserve in-memory
+    egress tokens."""
    h = hashlib.sha256()
    for path in sorted((repo_root / "bot_bottle").rglob("*.py")):
        h.update(str(path.relative_to(repo_root)).encode())
@@ -77,42 +96,37 @@ def _source_hash(repo_root: Path) -> str:


 class OrchestratorService:
-    """Manages the orchestrator control-plane container + the shared gateway.
-    Callers only need `ensure_running()` + `url`."""
+    """Manages the single per-host infra container (control plane + gateway).
+    Callers only need `ensure_running()` + `url`.
+
+    `infra_name` / `infra_label` let backends run independent infra containers
+    on the same host without name collisions (e.g. isolated integration tests
+    that can't share the production INFRA_NAME singleton)."""

    def __init__(
        self,
        *,
        port: int = DEFAULT_PORT,
        network: str = GATEWAY_NETWORK,
-        image: str = ORCHESTRATOR_IMAGE,
-        gateway_image: str = GATEWAY_IMAGE,
+        image: str = INFRA_IMAGE,
        repo_root: Path = _REPO_ROOT,
        host_root: Path | None = None,
+        infra_name: str = INFRA_NAME,
+        infra_label: str = INFRA_LABEL,
    ) -> None:
        self.port = port
        self.network = network
-        # Two distinct images (#384): `image` is the lean control-plane
-        # runtime this container runs; `_gateway_image` is the heavy egress /
-        # git-gate / supervise data plane the gateway container runs. They
-        # were one conflated image before the split.
        self.image = image
-        self._gateway_image = gateway_image
        self._repo_root = repo_root
        self._host_root = host_root or bot_bottle_root()
+        self._infra_name = infra_name
+        self._infra_label = infra_label

    @property
    def url(self) -> str:
        """Host-side control-plane URL (published loopback port)."""
        return f"http://127.0.0.1:{self.port}"

-    @property
-    def internal_url(self) -> str:
-        """Control-plane URL as the gateway container reaches it — by name over
-        docker DNS on the shared network. This is the gateway's
-        BOT_BOTTLE_ORCHESTRATOR_URL."""
-        return f"http://{ORCHESTRATOR_NAME}:{self.port}"
-
    def is_healthy(self, *, timeout: float = _HEALTH_REQUEST_TIMEOUT_SECONDS) -> bool:
        try:
            with urllib.request.urlopen(f"{self.url}/health", timeout=timeout) as resp:
@@ -124,125 +138,129 @@ class OrchestratorService:
        proc = run_docker(["docker", "ps", "--filter", f"name=^/{name}$", "--format", "{{.Names}}"])
        return name in proc.stdout.split()

-    def _run_orchestrator_container(self, source_hash: str) -> None:
-        """Start the control-plane container (idempotent: clears a stale
-        fixed-name container first). Register-only broker → no docker socket.
-        Labels the container with `source_hash` so a later `ensure_running`
-        can detect a real code change (see `_source_hash`)."""
-        run_docker(["docker", "rm", "--force", ORCHESTRATOR_NAME])
-        proc = run_docker([
-            "docker", "run", "--detach",
-            "--name", ORCHESTRATOR_NAME,
-            "--label", ORCHESTRATOR_LABEL,
-            "--label", f"{ORCHESTRATOR_SOURCE_HASH_LABEL}={source_hash}",
-            "--network", self.network,
-            # Host CLI reaches the control plane here; bound to loopback so it
-            # is not exposed on the host's external interfaces.
-            "--publish", f"127.0.0.1:{self.port}:{self.port}",
-            "--volume", f"{self._repo_root}:{_APP_DIR}:ro",
-            "--workdir", _APP_DIR,
-            # Persist the registry DB on the host (sole-owner: only the
-            # orchestrator opens bot-bottle.db).
-            "--volume", f"{self._host_root}:{_ROOT_IN_CONTAINER}",
-            "--env", f"BOT_BOTTLE_ROOT={_ROOT_IN_CONTAINER}",
-            "--entrypoint", "python3",
-            self.image,
-            "-m", "bot_bottle.orchestrator",
-            "--host", "0.0.0.0", "--port", str(self.port), "--broker", "stub",
-        ])
-        if proc.returncode != 0:
-            raise OrchestratorStartError(
-                f"orchestrator container failed to start: {proc.stderr.strip()}"
-            )
-
-    def _gateway(self) -> DockerGateway:
-        return DockerGateway(
-            self._gateway_image, network=self.network, orchestrator_url=self.internal_url
-        )
-
-    def _ensure_orchestrator_image(self) -> None:
-        """Build the lean control-plane image from `Dockerfile.orchestrator`
-        when it's missing (#384). Cheap — a `FROM python:*-slim` base with no
-        deps to install, so the layer cache makes rebuilds a no-op. Unlike the
-        gateway image this is build-if-missing, not build-every-time: the
-        control plane bind-mounts its source, so a code change is caught by the
-        source-hash recreate (below), not by an image rebuild."""
-        if run_docker(["docker", "image", "inspect", self.image]).returncode == 0:
-            return
-        argv = ["docker", "build", "-t", self.image,
-                "-f", str(self._repo_root / ORCHESTRATOR_DOCKERFILE),
-                str(self._repo_root)]
-        if os.environ.get("BOT_BOTTLE_NO_CACHE"):
-            argv.insert(2, "--no-cache")
-        proc = run_docker(argv)
-        if proc.returncode != 0:
-            raise GatewayError(
-                f"orchestrator image build failed: {proc.stderr.strip()}"
-            )
-
-    def _orchestrator_source_current(self, current_hash: str) -> bool:
-        """True iff the running orchestrator container was created from the
-        *current* bind-mounted source. Mirrors `DockerGateway`'s
-        image-staleness check, but by content hash rather than image id since
-        the orchestrator runs bind-mounted source, not a built image."""
-        if not self._container_running(ORCHESTRATOR_NAME):
+    def _infra_source_current(self, current_hash: str) -> bool:
+        """True iff the running infra container was started from the current
+        bind-mounted source. Mirrors the macOS backend's `_source_current`."""
+        if not self._container_running(self._infra_name):
            return False
        proc = run_docker([
            "docker", "inspect", "--format",
-            "{{ index .Config.Labels \"" + ORCHESTRATOR_SOURCE_HASH_LABEL + "\" }}",
-            ORCHESTRATOR_NAME,
+            "{{ index .Config.Labels \"" + INFRA_SOURCE_HASH_LABEL + "\" }}",
+            self._infra_name,
        ])
        if proc.returncode != 0:
-            return True  # can't compare -> don't churn a working container
+            return True  # can't compare → don't churn a working container
        return proc.stdout.strip() == current_hash

+    def _ensure_network(self) -> None:
+        if run_docker(["docker", "network", "inspect", self.network]).returncode == 0:
+            return
+        proc = run_docker(["docker", "network", "create", self.network])
+        if proc.returncode != 0 and "already exists" not in proc.stderr:
+            raise GatewayError(
+                f"gateway network {self.network} failed to create: {proc.stderr.strip()}"
+            )
+
+    def _build_images(self) -> None:
+        """Build the gateway base, the orchestrator intermediate, then the
+        infra image. All are cache-aware: a no-op when nothing changed."""
+        for tag, dockerfile in (
+            (GATEWAY_IMAGE, GATEWAY_DOCKERFILE),
+            (ORCHESTRATOR_IMAGE, ORCHESTRATOR_DOCKERFILE),
+            (self.image, INFRA_DOCKERFILE),
+        ):
+            argv = ["docker", "build", "-t", tag,
+                    "-f", str(self._repo_root / dockerfile),
+                    str(self._repo_root)]
+            if os.environ.get("BOT_BOTTLE_NO_CACHE"):
+                argv.insert(2, "--no-cache")
+            proc = run_docker(argv)
+            if proc.returncode != 0:
+                raise GatewayError(f"{dockerfile} build failed: {proc.stderr.strip()}")
+
+    def _run_infra_container(self, current_hash: str) -> None:
+        """Start the combined infra container (idempotent: clears a stale
+        fixed-name container first). Labels the container with `current_hash`
+        so a later `ensure_running` can detect a real code change."""
+        self._ensure_network()
+        run_docker(["docker", "rm", "--force", self._infra_name])
+        proc = run_docker([
+            "docker", "run", "--detach",
+            "--name", self._infra_name,
+            "--label", self._infra_label,
+            "--label", f"{INFRA_SOURCE_HASH_LABEL}={current_hash}",
+            "--network", self.network,
+            # Host CLI reaches the control plane here (loopback only).
+            # gateway_init always starts the orchestrator on DEFAULT_PORT (8099)
+            # inside the container; self.port is the host-side published port.
+            "--publish", f"127.0.0.1:{self.port}:{DEFAULT_PORT}",
+            # Persist the mitmproxy CA so it survives container recreation.
+            "--volume", f"{GATEWAY_CA_VOLUME}:{MITMPROXY_HOME}",
+            # Shared supervise DB (same file the operator reads over HTTP).
+            "--volume", f"{_host_db_dir()}:{_SUPERVISE_DB_DIR_IN_CONTAINER}",
+            "--env", f"SUPERVISE_DB_PATH={DB_PATH_IN_CONTAINER}",
+            # Live control-plane source, mounted to a path that does not
+            # overlay the gateway's baked /app scripts.
+            "--volume", f"{self._repo_root}:{_SRC_IN_CONTAINER}:ro",
+            # PYTHONPATH lets the orchestrator (and other Python daemons)
+            # import the live source ahead of the installed package.
+            "--env", f"PYTHONPATH={_SRC_IN_CONTAINER}",
+            # Orchestrator registry DB on the host (sole writer: control plane).
+            "--volume", f"{self._host_root}:{_ROOT_IN_CONTAINER}",
+            "--env", f"BOT_BOTTLE_ROOT={_ROOT_IN_CONTAINER}",
+            # Control-plane secret: required by the orchestrator (to enforce)
+            # and by the gateway daemons (to present on /resolve calls).
+            "--env", CONTROL_PLANE_TOKEN_ENV,
+            # Gateway daemons reach the orchestrator over loopback at its
+            # fixed internal port (DEFAULT_PORT), independent of self.port.
+            "--env", f"BOT_BOTTLE_ORCHESTRATOR_URL=http://127.0.0.1:{DEFAULT_PORT}",
+            # Opt the orchestrator into gateway_init's supervise tree.
+            "--env", f"BOT_BOTTLE_GATEWAY_DAEMONS={_INFRA_DAEMONS}",
+            self.image,
+        ], env={**os.environ, CONTROL_PLANE_TOKEN_ENV: host_control_plane_token()})
+        if proc.returncode != 0:
+            raise OrchestratorStartError(
+                f"infra container failed to start: {proc.stderr.strip()}"
+            )
+
    def ensure_running(
        self, *, startup_timeout: float = DEFAULT_STARTUP_TIMEOUT_SECONDS,
    ) -> str:
-        """Ensure the control plane + shared gateway are up; return the host
-        control-plane URL. Idempotent — a healthy control plane running
-        current code and a running gateway are left untouched. Raises
-        `OrchestratorStartError` on timeout."""
-        gateway = self._gateway()
-        gateway.ensure_built()  # rebuild the bundle image on a source change
-        gateway.ensure_running()  # creates the shared network + (re)starts gateway
+        """Ensure the infra container (control plane + gateway) is up; return
+        the host control-plane URL. Idempotent — a healthy container on current
+        source is left untouched. Raises `OrchestratorStartError` on timeout."""
+        self._build_images()

-        # Recreate the orchestrator container only when its bind-mounted
-        # source has actually changed since it started — its Python process
-        # loaded that code at startup and won't reload, so a stale container
-        # would keep running OLD control-plane code. Recreating on *every*
-        # launch (the prior behaviour) would drop every other active
-        # bottle's in-memory egress tokens each time a new bottle starts,
-        # since the orchestrator process holds them only in memory (#381).
-        current_hash = _source_hash(self._repo_root)
-        if self.is_healthy() and self._orchestrator_source_current(current_hash):
+        current_hash = source_hash(self._repo_root)
+        if self.is_healthy() and self._infra_source_current(current_hash):
            return self.url

-        self._ensure_orchestrator_image()
-        log.info("starting orchestrator container", context={"name": ORCHESTRATOR_NAME})
-        self._run_orchestrator_container(current_hash)
+        log.info("starting infra container", context={"name": self._infra_name})
+        self._run_infra_container(current_hash)

        deadline = time.monotonic() + startup_timeout
        while time.monotonic() < deadline:
            if self.is_healthy():
-                log.info("orchestrator healthy", context={"url": self.url})
+                log.info("infra container healthy", context={"url": self.url})
                return self.url
            time.sleep(_HEALTH_POLL_SECONDS)
        raise OrchestratorStartError(
-            f"orchestrator at {self.url} did not become healthy within {startup_timeout:g}s"
+            f"infra container at {self.url} did not become healthy within {startup_timeout:g}s"
        )

    def stop(self) -> None:
-        """Remove the orchestrator + gateway containers (idempotent)."""
-        run_docker(["docker", "rm", "--force", ORCHESTRATOR_NAME])
-        self._gateway().stop()
+        """Remove the infra container (idempotent)."""
+        run_docker(["docker", "rm", "--force", self._infra_name])


 __all__ = [
    "OrchestratorService",
    "OrchestratorStartError",
-    "ORCHESTRATOR_NAME",
+    "INFRA_NAME",
+    "INFRA_IMAGE",
+    "INFRA_SOURCE_HASH_LABEL",
    "ORCHESTRATOR_IMAGE",
    "DEFAULT_PORT",
    "DEFAULT_STARTUP_TIMEOUT_SECONDS",
+    "source_hash",
 ]
@@ -32,6 +32,7 @@ import hmac
 import secrets
 import sqlite3
 import time
+from collections.abc import Iterable
 from dataclasses import dataclass
 from pathlib import Path

@@ -42,6 +43,12 @@ from ..paths import host_db_path
 # 256 bits of urandom, URL-safe — unguessable per-bottle identity token.
 IDENTITY_TOKEN_BYTES = 32

+# How recently a row must have been registered to be exempt from
+# `reap_absent`. Covers the window between `container run` and the address
+# becoming visible to another launch's enumeration, so reconciliation never
+# reaps a bottle that is still coming up.
+DEFAULT_REAP_GRACE_SECONDS = 120.0
+

 def new_identity_token() -> str:
    """A fresh per-bottle identity token (PRD 0070 attribution defence)."""
@@ -167,7 +174,7 @@ class RegistryStore(DbStore):
            metadata=metadata,
            policy=policy,
        )
-        with self._connect() as conn:
+        with self._connection() as conn:
            conn.execute(
                "DELETE FROM orchestrator_bottles "
                "WHERE source_ip = ? AND state = 'active' AND bottle_id != ?",
@@ -193,7 +200,7 @@ class RegistryStore(DbStore):
    def set_policy(self, bottle_id: str, policy: str) -> bool:
        """Update a bottle's policy in place (live reload). Returns True if
        the bottle exists."""
-        with self._connect() as conn:
+        with self._connection() as conn:
            cur = conn.execute(
                "UPDATE orchestrator_bottles SET policy = ? WHERE bottle_id = ?",
                (policy, bottle_id),
@@ -203,7 +210,7 @@ class RegistryStore(DbStore):

    def deregister(self, bottle_id: str) -> bool:
        """Remove a bottle. Returns True if a row was deleted."""
-        with self._connect() as conn:
+        with self._connection() as conn:
            cur = conn.execute(
                "DELETE FROM orchestrator_bottles WHERE bottle_id = ?", (bottle_id,)
            )
@@ -211,7 +218,7 @@ class RegistryStore(DbStore):

    def get(self, bottle_id: str) -> BottleRecord | None:
        """Return the bottle by id, or None if absent."""
-        with self._connect() as conn:
+        with self._connection() as conn:
            row = conn.execute(
                "SELECT * FROM orchestrator_bottles WHERE bottle_id = ?", (bottle_id,)
            ).fetchone()
@@ -219,12 +226,76 @@ class RegistryStore(DbStore):

    def all(self) -> list[BottleRecord]:
        """Every registered bottle, oldest first."""
-        with self._connect() as conn:
+        with self._connection() as conn:
            rows = conn.execute(
                "SELECT * FROM orchestrator_bottles ORDER BY created_at"
            ).fetchall()
        return [_row_to_record(r) for r in rows]

+    def reap_absent(
+        self,
+        live_source_ips: Iterable[str],
+        *,
+        grace_seconds: float = DEFAULT_REAP_GRACE_SECONDS,
+        now: float | None = None,
+    ) -> list[BottleRecord]:
+        """Delete active rows whose source IP is not held by a live bottle.
+
+        A row only ever leaves the registry two ways: an explicit
+        `teardown_bottle` (the launcher's cleanup callback) or the supersede
+        sweep in `register`. Neither runs when the launching CLI dies hard —
+        SIGKILL, a closed terminal, a host sleep/crash — so the row outlives
+        its container. That orphan is not inert: source IPs are recycled by
+        the backend's DHCP, and `by_source_ip` fail-closes on ambiguity, so a
+        leftover row at a reused address can brick the *next* bottle that
+        lands on it (no policy resolved -> every host denied, reported to the
+        agent as "not in the allowlist"). Reconciling against the live set at
+        launch keeps the registry from accumulating those landmines.
+
+        Restores the invariant the data plane needs: **at most one active row
+        per live address, and none at all for a dead one.** Two cases, because
+        a dead bottle's address may already have been handed to a live one:
+
+          * no live bottle holds the address — every row there is an orphan;
+          * a live bottle holds it but several rows claim it — the newest
+            registration is authoritative and the rest are orphans, the same
+            rule `register`'s same-IP supersede sweep applies. Without this
+            second case a recycled address stays ambiguous, which is exactly
+            the state that resolves no policy.
+
+        `grace_seconds` protects an in-flight launch: registration happens
+        moments after `container run`, and a concurrent launch's address may
+        not be visible to the caller's enumeration yet. Rows younger than the
+        grace window are never reaped, so reconciliation can't race a bottle
+        that is still coming up. Returns the deleted records."""
+        live = {ip for ip in live_source_ips if ip}
+        cutoff = (time.time() if now is None else now) - grace_seconds
+        with self._connection() as conn:
+            rows = conn.execute(
+                "SELECT * FROM orchestrator_bottles WHERE state = 'active'",
+            ).fetchall()
+            by_ip: dict[str, list[BottleRecord]] = {}
+            for row in rows:
+                rec = _row_to_record(row)
+                by_ip.setdefault(rec.source_ip, []).append(rec)
+            candidates: list[BottleRecord] = []
+            for ip, recs in by_ip.items():
+                if ip not in live:
+                    candidates.extend(recs)
+                    continue
+                # Keep the newest claim on a live address; supersede the rest.
+                recs.sort(key=lambda r: r.created_at)
+                candidates.extend(recs[:-1])
+            doomed = [r for r in candidates if r.created_at <= cutoff]
+            for rec in doomed:
+                conn.execute(
+                    "DELETE FROM orchestrator_bottles WHERE bottle_id = ?",
+                    (rec.bottle_id,),
+                )
+        if doomed:
+            self._chmod()
+        return doomed
+
    def by_source_ip(self, source_ip: str) -> BottleRecord | None:
        """Network-layer attribution: the single active bottle at this source
        IP, or None if unknown or ambiguous (more than one — a
@@ -232,7 +303,7 @@ class RegistryStore(DbStore):
        source IP is unspoofable (Firecracker `/31` + nft) and the control
        plane is reachable only by the trusted gateway; pair with the
        identity token (`attribute`) elsewhere."""
-        with self._connect() as conn:
+        with self._connection() as conn:
            rows = conn.execute(
                "SELECT * FROM orchestrator_bottles "
                "WHERE source_ip = ? AND state = 'active'",
@@ -262,4 +333,5 @@ __all__ = [
    "new_identity_token",
    "default_db_path",
    "IDENTITY_TOKEN_BYTES",
+    "DEFAULT_REAP_GRACE_SECONDS",
 ]
@@ -13,13 +13,46 @@ Launch lifecycle:
    and returns the record. If the broker rejects/fails, the registry entry
    is rolled back so a failed launch leaves no orphan.
  * `teardown_bottle` sends a signed teardown request, then deregisters.
+  * `reconcile` sweeps rows whose bottle is no longer running — the
+    self-heal for the teardown paths that never got to run (a hard-killed
+    launcher), since an orphan row at a recycled source IP bricks the next
+    bottle that lands on it.
 """

 from __future__ import annotations

+import json
+from collections.abc import Iterable
+from datetime import datetime, timezone
+
 from .broker import LaunchBroker, LaunchRequest, sign_request
-from .registry import BottleRecord, RegistryStore
+from .registry import DEFAULT_REAP_GRACE_SECONDS, BottleRecord, RegistryStore
 from .gateway import Gateway
+from ..supervise import (
+    AuditEntry,
+    COMPONENT_FOR_TOOL,
+    Response,
+    STATUS_APPROVED,
+    STATUS_MODIFIED,
+    STATUS_REJECTED,
+    TOOL_EGRESS_ALLOW,
+    TOOL_EGRESS_BLOCK,
+    list_all_pending_proposals,
+    read_proposal,
+    render_diff,
+    write_audit_entry,
+    write_response,
+)
+
+
+# Operator decision → Response.status. The apply half (egress tools) runs
+# for approve/modify only.
+_RESPOND_STATUS = {
+    "approve": STATUS_APPROVED,
+    "modify": STATUS_MODIFIED,
+    "reject": STATUS_REJECTED,
+}
+_APPLY_TOOLS = (TOOL_EGRESS_ALLOW, TOOL_EGRESS_BLOCK)


 class Orchestrator:
@@ -89,6 +122,30 @@ class Orchestrator:
        self._tokens.pop(bottle_id, None)
        return True

+    def reconcile(
+        self,
+        live_source_ips: Iterable[str],
+        *,
+        grace_seconds: float = DEFAULT_REAP_GRACE_SECONDS,
+    ) -> list[str]:
+        """Drop registry rows for bottles that are no longer running, and
+        forget their in-memory egress tokens. Returns the reaped bottle ids.
+
+        The caller supplies the live set because only the host can enumerate
+        its own containers — the orchestrator runs *inside* the infra
+        container and has no view of the backend. Deliberately does not
+        broker a teardown: the container is already gone, so there is nothing
+        to stop, and a broker error must not stop the sweep from clearing
+        the row that would otherwise brick the next bottle at that address.
+
+        See `RegistryStore.reap_absent` for why orphans accumulate and why
+        they are harmful rather than merely untidy."""
+        reaped = self.registry.reap_absent(
+            live_source_ips, grace_seconds=grace_seconds)
+        for rec in reaped:
+            self._tokens.pop(rec.bottle_id, None)
+        return [rec.bottle_id for rec in reaped]
+
    def tokens_for(self, bottle_id: str) -> dict[str, str]:
        """The bottle's in-memory egress auth tokens (env_name -> value), or
        empty. The gateway injects these per request; they are never
@@ -99,22 +156,134 @@ class Orchestrator:
        """Fail-closed attribution (delegates to the registry)."""
        return self.registry.attribute(source_ip, identity_token)

-    def resolve(self, source_ip: str, identity_token: str = "") -> BottleRecord | None:
-        """Resolve the bottle behind a request — the source-IP-keyed lookup
-        the multi-tenant gateway makes per request; the returned record
-        carries its `policy`. With a token, full attribution (source IP +
-        token); without, network-layer attribution by source IP alone
-        (valid where the IP is unspoofable and the control plane is
-        gateway-only)."""
-        if identity_token:
-            return self.registry.attribute(source_ip, identity_token)
-        return self.registry.by_source_ip(source_ip)
+    def resolve(self, source_ip: str, identity_token: str) -> BottleRecord | None:
+        """Resolve the bottle behind a request — the per-request lookup the
+        multi-tenant gateway makes; the returned record carries its `policy`.
+
+        **Mandatory pair**: requires a matching `(source_ip, identity_token)`
+        (constant-time). There is no source-IP-only fallback — the app-layer
+        token is delivered on every attributed data plane (egress proxy
+        credentials, git-gate/supervise headers), so a missing or mismatched
+        token fail-closes. This keeps a spoofed source IP (which the /31 TAP
+        alone does not prevent) from selecting another bottle's policy/tokens
+        without also holding that bottle's unguessable token."""
+        return self.registry.attribute(source_ip, identity_token)

    def set_policy(self, bottle_id: str, policy: str) -> bool:
        """Update a bottle's gateway policy in place (live reload). False if
        the bottle is unknown."""
        return self.registry.set_policy(bottle_id, policy)

+    # --- supervise queue (operator approvals) ------------------------------
+    #
+    # The orchestrator owns the single DB *and* the live policy, so operator
+    # decisions are applied here, server-side, and reached over HTTP by the
+    # host TUI (no direct-DB access, one path for every backend).
+
+    def supervise_pending(self) -> list[dict[str, object]]:
+        """All pending proposals across bottles, FIFO, as JSON dicts
+        (`Proposal.to_dict`, round-trippable via `Proposal.from_dict`).
+
+        Each dict carries an extra `bottle_label`: the bottle's human slug
+        resolved from the registry (the proposal itself is keyed by the
+        orchestrator-assigned bottle_id, which is opaque to an operator). The
+        CLI renders the label but still responds against `bottle_slug`."""
+        out: list[dict[str, object]] = []
+        for p in list_all_pending_proposals():
+            d = p.to_dict()
+            d["bottle_label"] = self._label_for(p.bottle_slug)
+            out.append(d)
+        return out
+
+    def _label_for(self, bottle_slug: str) -> str:
+        """The human slug recorded in registry metadata for a proposal's
+        bottle, or the bottle_slug unchanged when the bottle is gone or has no
+        recorded slug — so the label is always non-empty."""
+        rec = self.registry.get(bottle_slug)
+        if rec is None:
+            return bottle_slug
+        try:
+            meta = json.loads(rec.metadata) if rec.metadata else {}
+        except ValueError:
+            meta = {}
+        slug = meta.get("slug") if isinstance(meta, dict) else None
+        return slug if isinstance(slug, str) and slug else bottle_slug
+
+    def _record_for_slug(self, slug: str) -> BottleRecord | None:
+        """The live registry record for a proposal's bottle, or None (e.g. the
+        bottle was torn down before the operator responded).
+
+        In consolidated mode the supervise server attributes each proposal to
+        the orchestrator-assigned bottle_id and stores that as the proposal's
+        `bottle_slug` (see supervise_server `_attributed_config`), so the fast
+        path is a direct bottle_id lookup. The metadata-slug scan is the
+        fallback for legacy single-tenant proposals keyed by the human slug."""
+        rec = self.registry.get(slug)
+        if rec is not None:
+            return rec
+        for rec in self.registry.all():
+            try:
+                meta = json.loads(rec.metadata) if rec.metadata else {}
+            except ValueError:
+                meta = {}
+            if isinstance(meta, dict) and meta.get("slug") == slug:
+                return rec
+        return None
+
+    def supervise_respond(
+        self,
+        proposal_id: str,
+        *,
+        bottle_slug: str,
+        decision: str,
+        notes: str = "",
+        final_file: str | None = None,
+    ) -> tuple[bool, str]:
+        """Record an operator decision on a queued proposal, applying it
+        server-side. `decision` is approve/modify/reject.
+
+        Approve/modify on an egress tool rewrites the bottle's policy so the
+        gateway serves the new routes on its next `/resolve` (the live apply);
+        then the queued Response is written (unblocking the agent's MCP call)
+        and an audit entry recorded — all against the one DB. Returns
+        (ok, error): ok=False with a message when the proposal or decision is
+        unknown, or the bottle is gone so an approval can't be applied."""
+        status = _RESPOND_STATUS.get(decision)
+        if status is None:
+            return False, f"unknown decision {decision!r}"
+        try:
+            proposal = read_proposal(bottle_slug, proposal_id)
+        except FileNotFoundError:
+            return False, "no such proposal"
+
+        diff_before, diff_after = "", ""
+        if status in (STATUS_APPROVED, STATUS_MODIFIED) and proposal.tool in _APPLY_TOOLS:
+            new_policy = final_file if final_file is not None else proposal.proposed_file
+            rec = self._record_for_slug(bottle_slug)
+            if rec is None:
+                return False, (
+                    f"bottle {bottle_slug!r} is no longer registered; "
+                    "cannot apply the route change"
+                )
+            diff_before, diff_after = rec.policy, new_policy
+            self.set_policy(rec.bottle_id, new_policy)
+
+        write_response(bottle_slug, Response(
+            proposal_id=proposal_id, status=status, notes=notes, final_file=final_file,
+        ))
+        component = COMPONENT_FOR_TOOL.get(proposal.tool)
+        if component is not None:
+            write_audit_entry(AuditEntry(
+                timestamp=datetime.now(timezone.utc).isoformat(),
+                bottle_slug=bottle_slug,
+                component=component,
+                operator_action=status,
+                operator_notes=notes,
+                justification=proposal.justification,
+                diff=render_diff(diff_before, diff_after, label=component),
+            ))
+        return True, ""
+
    # --- consolidated gateway ----------------------------------------------

    def ensure_gateway(self) -> None:
@@ -16,6 +16,8 @@ layer (and to COPY flat into the gateway).
 from __future__ import annotations

 import os
+import secrets
+import stat
 from pathlib import Path

 # The single shared host state DB. All bot-bottle SQLite stores (supervise
@@ -23,6 +25,14 @@ from pathlib import Path
 # TableMigrations schema_key namespaces each store's tables.
 HOST_DB_FILENAME = "bot-bottle.db"

+# The per-host control-plane secret file, and the env var the launchers inject
+# its value into. The control plane requires this secret on every mutating /
+# reading route (see orchestrator/control_plane.py); it is held only by the
+# trusted callers (control plane, gateway, host CLI) and never handed to an
+# agent, so an agent that can reach the control-plane port still can't drive it.
+CONTROL_PLANE_TOKEN_FILENAME = "control-plane-token"
+CONTROL_PLANE_TOKEN_ENV = "BOT_BOTTLE_CONTROL_PLANE_TOKEN"
+

 def bot_bottle_root() -> Path:
    """The app data root — `$BOT_BOTTLE_ROOT` if set, else `~/.bot-bottle`."""
@@ -40,4 +50,52 @@ def host_db_path() -> Path:
    return bot_bottle_root() / "db" / HOST_DB_FILENAME


-__all__ = ["HOST_DB_FILENAME", "bot_bottle_root", "host_db_path"]
+def host_db_dir() -> Path:
+    """The directory holding the shared host state DB, created if missing.
+    Backends bind-mount this into their gateway so the supervise daemon writes
+    to the one DB the orchestrator (and the operator over HTTP) reads."""
+    db_dir = host_db_path().parent
+    db_dir.mkdir(parents=True, exist_ok=True)
+    return db_dir
+
+
+def host_control_plane_token() -> str:
+    """The per-host control-plane secret, minted (256-bit, url-safe) and
+    persisted 0600 on first use, then reused.
+
+    This is the shared secret the launchers inject into the control-plane and
+    gateway containers and that the host CLI presents on every call. It is a
+    *host* artifact — the file lives under the root the agent never mounts, and
+    the env var is set only on the trusted containers — so reading it here is
+    safe on the host launch path but the value never reaches a bottle."""
+    path = bot_bottle_root() / CONTROL_PLANE_TOKEN_FILENAME
+    try:
+        existing = path.read_text().strip()
+        if existing:
+            return existing
+    except OSError:
+        pass
+    path.parent.mkdir(parents=True, exist_ok=True)
+    token = secrets.token_urlsafe(32)
+    # Create 0600 up front (O_EXCL loses a concurrent race harmlessly — we
+    # re-read the winner's token below) so the secret is never briefly world-
+    # readable between write and chmod.
+    try:
+        fd = os.open(path, os.O_WRONLY | os.O_CREAT | os.O_EXCL, 0o600)
+    except FileExistsError:
+        return path.read_text().strip()
+    with os.fdopen(fd, "w") as f:
+        f.write(token)
+    os.chmod(path, stat.S_IRUSR | stat.S_IWUSR)
+    return token
+
+
+__all__ = [
+    "HOST_DB_FILENAME",
+    "CONTROL_PLANE_TOKEN_FILENAME",
+    "CONTROL_PLANE_TOKEN_ENV",
+    "bot_bottle_root",
+    "host_db_path",
+    "host_db_dir",
+    "host_control_plane_token",
+]
@@ -22,18 +22,35 @@ closed too rather than silently serving stale or empty policy.

 The resolved value is the policy blob the orchestrator stores verbatim; the
 consumer parses it (e.g. the egress addon's `load_config`). This module is
-stdlib-only and free of bot-bottle imports so it can be COPYed flat into
-the gateway.
+stdlib-only and free of bot-bottle imports.
 """

 from __future__ import annotations

 import json
+import os
 import urllib.error
 import urllib.request

 DEFAULT_TIMEOUT_SECONDS = 2.0

+# The control-plane secret this gateway presents on every /resolve call, read
+# from the env the launcher injects into the gateway container. The control
+# plane requires it (orchestrator/control_plane.py). Constant + env-var name are
+# duplicated here rather than imported because this module is COPYed flat into
+# the gateway image, free of bot-bottle imports — same rationale as
+# IDENTITY_HEADER in egress_addon / git_http_backend.
+CONTROL_AUTH_HEADER = "x-bot-bottle-control-auth"
+CONTROL_PLANE_TOKEN_ENV = "BOT_BOTTLE_CONTROL_PLANE_TOKEN"
+
+
+def _control_auth_headers() -> dict[str, str]:
+    """The auth header to send, or {} when no secret is configured (an open
+    control plane, e.g. Firecracker behind its nft boundary — sending nothing
+    is correct there and harmlessly ignored)."""
+    token = os.environ.get(CONTROL_PLANE_TOKEN_ENV, "").strip()
+    return {CONTROL_AUTH_HEADER: token} if token else {}
+

 class PolicyResolveError(RuntimeError):
    """The orchestrator was unreachable or returned an unexpected status —
@@ -57,7 +74,7 @@ class PolicyResolver:
        ).encode()
        req = urllib.request.Request(
            f"{self._base}/resolve", data=body, method="POST",
-            headers={"Content-Type": "application/json"},
+            headers={"Content-Type": "application/json", **_control_auth_headers()},
        )
        try:
            with urllib.request.urlopen(req, timeout=self._timeout) as resp:
@@ -66,7 +66,7 @@ class QueueStore(DbStore):
        super().__init__(resolved, migrations)

    def write_proposal(self, proposal: Proposal) -> Path:
-        with self._connect() as conn:
+        with self._connection() as conn:
            conn.execute(
                """
                INSERT OR REPLACE INTO supervise_proposals (
@@ -89,7 +89,7 @@ class QueueStore(DbStore):
        return self.db_path

    def read_proposal(self, proposal_id: str) -> Proposal:
-        with self._connect() as conn:
+        with self._connection() as conn:
            row = conn.execute(
                """
                SELECT * FROM supervise_proposals
@@ -104,7 +104,7 @@ class QueueStore(DbStore):
    def list_pending_proposals(self) -> list[Proposal]:
        if not self.db_path.is_file():
            return []
-        with self._connect() as conn:
+        with self._connection() as conn:
            rows = conn.execute(
                """
                SELECT p.* FROM supervise_proposals p
@@ -125,7 +125,7 @@ class QueueStore(DbStore):
    def list_all_pending_proposals(self) -> list[Proposal]:
        if not self.db_path.is_file():
            return []
-        with self._connect() as conn:
+        with self._connection() as conn:
            rows = conn.execute(
                """
                SELECT p.* FROM supervise_proposals p
@@ -142,7 +142,7 @@ class QueueStore(DbStore):
        return [self._row_to_proposal(row) for row in rows]

    def write_response(self, response: Response) -> Path:
-        with self._connect() as conn:
+        with self._connection() as conn:
            conn.execute(
                """
                INSERT OR REPLACE INTO supervise_responses (
@@ -161,7 +161,7 @@ class QueueStore(DbStore):
        return self.db_path

    def read_response(self, proposal_id: str) -> Response:
-        with self._connect() as conn:
+        with self._connection() as conn:
            row = conn.execute(
                """
                SELECT * FROM supervise_responses
@@ -176,7 +176,7 @@ class QueueStore(DbStore):
    def archive_proposal(self, proposal_id: str) -> None:
        if not self.db_path.is_file():
            return
-        with self._connect() as conn:
+        with self._connection() as conn:
            conn.execute(
                """
                UPDATE supervise_proposals SET archived = 1
@@ -6,9 +6,11 @@ from pathlib import Path

 try:
    from .audit_store import AuditStore
+    from .config_store import ConfigStore
    from .queue_store import QueueStore
 except ImportError:
    from audit_store import AuditStore  # type: ignore[import-not-found]  # pylint: disable=import-error,no-name-in-module
+    from config_store import ConfigStore  # type: ignore[import-not-found]  # pylint: disable=import-error,no-name-in-module
    from queue_store import QueueStore  # type: ignore[import-not-found]  # pylint: disable=import-error,no-name-in-module

 _instance: StoreManager | None = None
@@ -47,11 +49,13 @@ class StoreManager:
        return (
            QueueStore("", self.db_path).is_migrated()
            and AuditStore(self.db_path).is_migrated()
+            and ConfigStore(self.db_path).is_migrated()
        )

    def migrate(self) -> None:
        QueueStore("", self.db_path).migrate()
        AuditStore(self.db_path).migrate()
+        ConfigStore(self.db_path).migrate()


 __all__ = ["StoreManager"]
@@ -37,40 +37,23 @@ from abc import ABC
 from dataclasses import dataclass
 from pathlib import Path

-try:
-    from .supervise_types import (
-        ACTION_OPERATOR_EDIT,
-        AuditEntry,
-        Proposal,
-        Response,
-        STATUSES,
-        STATUS_APPROVED,
-        STATUS_MODIFIED,
-        STATUS_REJECTED,
-        TOOLS,
-        TOOL_EGRESS_ALLOW,
-        TOOL_EGRESS_BLOCK,
-        TOOL_EGRESS_TOKEN_ALLOW,
-        TOOL_GITLEAKS_ALLOW,
-        TOOL_LIST_EGRESS_ROUTES,
-    )
-except ImportError:
-    from supervise_types import (  # type: ignore[import-not-found,no-redef]  # pylint: disable=import-error,no-name-in-module
-        ACTION_OPERATOR_EDIT,
-        AuditEntry,
-        Proposal,
-        Response,
-        STATUSES,
-        STATUS_APPROVED,
-        STATUS_MODIFIED,
-        STATUS_REJECTED,
-        TOOLS,
-        TOOL_EGRESS_ALLOW,
-        TOOL_EGRESS_BLOCK,
-        TOOL_EGRESS_TOKEN_ALLOW,
-        TOOL_GITLEAKS_ALLOW,
-        TOOL_LIST_EGRESS_ROUTES,
-    )
+from .supervise_types import (
+    ACTION_OPERATOR_EDIT,
+    AuditEntry,
+    Proposal,
+    Response,
+    STATUSES,
+    STATUS_APPROVED,
+    STATUS_MODIFIED,
+    STATUS_REJECTED,
+    TOOLS,
+    TOOL_CHECK_PROPOSAL,
+    TOOL_EGRESS_ALLOW,
+    TOOL_EGRESS_BLOCK,
+    TOOL_EGRESS_TOKEN_ALLOW,
+    TOOL_GITLEAKS_ALLOW,
+    TOOL_LIST_EGRESS_ROUTES,
+)


 try:
@@ -281,6 +264,7 @@ __all__ = [
    "TOOLS",
    "EGRESS_FORWARD_PROXY",
    "EGRESS_INTROSPECT_URL",
+    "TOOL_CHECK_PROPOSAL",
    "TOOL_EGRESS_ALLOW",
    "TOOL_EGRESS_BLOCK",
    "TOOL_GITLEAKS_ALLOW",
@@ -2,37 +2,44 @@

 Per-bottle MCP server exposing tools the agent calls to propose egress
 config changes when stuck. The tools are `egress-allow`,
-`egress-block`, and `list-egress-routes`.
+`egress-block`, `list-egress-routes`, and `check-proposal`.

-Each queued tool call:
+Each queued proposal tool call:

  1. Validates the proposed file syntactically.
  2. Writes a Proposal to the host SQLite database.
-  3. Blocks polling for a matching Response row.
-  4. Returns the operator's `{status, notes}` to the agent.
+  3. Blocks polling for a matching Response row, up to a short grace
+     window (`SUPERVISE_RESPONSE_TIMEOUT_SECONDS`, default 30s).
+  4. On a decision within the window, returns the operator's
+     `{status, notes}`. On timeout, returns `status: pending` **with the
+     proposal id** and leaves the proposal queued — the flow is
+     non-blocking past the grace window (PRD prd-new / issue #412).

-The bottle slug arrives via SUPERVISE_BOTTLE_SLUG env (stamped at
-container creation by the backend's start step). SUPERVISE_DB_PATH
+`check-proposal` is the non-blocking companion: given a `proposal_id`
+returned by a `pending` response, it reports the current decision
+(`pending` | `approved` | `modified` | `rejected`) without re-proposing,
+so an approval made out-of-band (e.g. a web review console) can be resumed
+without holding an HTTP request open.
+
+One shared server fronts every bottle (PRD 0070) and attributes each
+proposal to the calling bottle by source IP, resolved from the orchestrator
+— an unattributed or unreachable source fails closed. BOT_BOTTLE_ORCHESTRATOR_URL
+is mandatory: there is no fixed-slug single-tenant fallback. SUPERVISE_DB_PATH
 points at the bind-mounted host database.

-Consolidated (PRD 0070): when BOT_BOTTLE_ORCHESTRATOR_URL is set, one
-shared server fronts every bottle and attributes each proposal to the
-calling bottle by source IP (resolved from the orchestrator) instead of a
-fixed slug — an unattributed source fails closed. Unset → the legacy
-per-bottle single-tenant server, unchanged.
-
 Speaks MCP over HTTP+JSON-RPC. Methods handled:

  * `initialize`           — handshake; returns server info + caps.
  * `notifications/initialized` — ack-only.
  * `tools/list`           — returns the tool definitions.
-  * `tools/call`           — validates, queues, blocks, returns.
+  * `tools/call`           — validates, queues, waits out the grace
+                             window, returns (pending past it); or, for
+                             `check-proposal`, a non-blocking status poll.

 Everything else returns JSON-RPC error -32601 (method not found).

-Stdlib-only. The Dockerfile copies this file + bot_bottle/supervise.py
-into the image; the server imports `supervise` for the queue / Proposal
-plumbing.
+The Dockerfile copies this script to /app/supervise_server.py and installs
+the bot_bottle package so its `from bot_bottle.*` imports resolve.
 """

 from __future__ import annotations
@@ -44,21 +51,14 @@ import socketserver
 import sys
 import time
 import typing
-import urllib.error
-import urllib.request
 from dataclasses import dataclass, replace

-try:
-    # Same-directory imports inside the bundle container; these files are
-    # COPYed flat under /app by Dockerfile.gateway.
-    from egress_addon_core import LOG_OFF, load_config
-    from policy_resolver import PolicyResolveError, PolicyResolver
-    import supervise as _sv
-except ModuleNotFoundError:
-    # Package imports for host-side tests and tooling.
-    from .egress_addon_core import LOG_OFF, load_config
-    from .policy_resolver import PolicyResolveError, PolicyResolver
-    from . import supervise as _sv
+from bot_bottle.constants import IDENTITY_HEADER
+from bot_bottle.egress_addon_core import (
+    LOG_OFF, load_config, resolve_client_context, route_to_yaml_dict,
+)
+from bot_bottle.policy_resolver import PolicyResolveError, PolicyResolver
+from bot_bottle import supervise as _sv


 # --- JSON-RPC / MCP plumbing ----------------------------------------------
@@ -79,12 +79,10 @@ ERR_INTERNAL = -32603

 DEFAULT_RESPONSE_TIMEOUT_SECONDS = 30.0
 MIN_RESPONSE_POLL_INTERVAL_SECONDS = 0.05
-EGRESS_LIST_TIMEOUT_SECONDS = 5.0

-# Consolidated (multi-tenant) mode: when set, one shared supervise server
-# fronts every bottle and attributes each proposal to the calling bottle by
-# source IP (resolved from the orchestrator), instead of a single
-# SUPERVISE_BOTTLE_SLUG env. Unset → legacy per-bottle single-tenant.
+# The per-host orchestrator control plane the shared supervise server attributes
+# each proposal to, by source IP. Mandatory — there is no single-tenant
+# SUPERVISE_BOTTLE_SLUG fallback.
 ORCHESTRATOR_URL_ENV = "BOT_BOTTLE_ORCHESTRATOR_URL"


@@ -246,6 +244,31 @@ TOOL_DEFINITIONS: list[dict[str, object]] = [
        ),
        "inputSchema": _proposal_input_schema(),
    },
+    {
+        "name": _sv.TOOL_CHECK_PROPOSAL,
+        "description": (
+            "Poll a previously queued proposal for the operator's decision "
+            "WITHOUT blocking or re-proposing. Pass the `proposal_id` you "
+            "got back when an `egress-allow`/`egress-block` call returned "
+            "`status: pending`. Returns the current status: `pending` (no "
+            "decision yet — poll again later), `approved`, `modified`, "
+            "`rejected`, or `unknown` (no such queued proposal — wrong id, "
+            "or it was already resolved and read)."
+        ),
+        "inputSchema": {
+            "type": "object",
+            "properties": {
+                "proposal_id": {
+                    "type": "string",
+                    "description": (
+                        "The proposal id from a `pending` response."
+                    ),
+                },
+            },
+            "required": ["proposal_id"],
+            "additionalProperties": False,
+        },
+    },
 ]


@@ -304,42 +327,6 @@ def handle_tools_list(_params: dict[str, object]) -> dict[str, object]:
    return {"tools": TOOL_DEFINITIONS}


-def handle_list_egress_routes(
-    _params: dict[str, object],
-    _config: ServerConfig,
-) -> dict[str, object]:
-    """Fetch the live egress route table via its
-    `_egress.local/allowlist` introspection endpoint. The
-    request goes through egress as a forward proxy; the
-    addon recognises the magic host and synthesizes a response —
-    no real upstream connection, no allowlist enforcement
-    against the magic host. Returns the JSON payload as the
-    tool's text content."""
-    proxy_handler = urllib.request.ProxyHandler({
-        "http": _sv.EGRESS_FORWARD_PROXY,
-    })
-    opener = urllib.request.build_opener(proxy_handler)
-    try:
-        with opener.open(_sv.EGRESS_INTROSPECT_URL, timeout=EGRESS_LIST_TIMEOUT_SECONDS) as resp:
-            body = resp.read().decode("utf-8")
-    except (urllib.error.URLError, OSError) as e:
-        return {
-            "content": [{
-                "type": "text",
-                "text": (
-                    f"list-egress-routes: could not reach "
-                    f"{_sv.EGRESS_INTROSPECT_URL!r} via "
-                    f"{_sv.EGRESS_FORWARD_PROXY!r}: {e}"
-                ),
-            }],
-            "isError": True,
-        }
-    return {
-        "content": [{"type": "text", "text": body}],
-        "isError": False,
-    }
-
-
 def handle_tools_call(
    params: dict[str, object],
    config: ServerConfig,
@@ -347,14 +334,13 @@ def handle_tools_call(
    """Validates the proposal, writes it to the queue, blocks waiting
    for a Response, returns the result wrapped in MCP `content`.

-    Side-effect-free `list-*` tools short-circuit before the queue/
-    blocking machinery — they're read-only introspection that
-    doesn't need operator approval."""
+    `list-egress-routes` never reaches here — the handler answers it from
+    the calling bottle's resolved policy before dispatching (see
+    `MCPHandler._dispatch`); this path is the queued, operator-approved
+    `egress-allow` / `egress-block` tools."""
    name = params.get("name")
    if not isinstance(name, str):
        raise _RpcClientError(ERR_INVALID_PARAMS, "tools/call missing 'name'")
-    if name == _sv.TOOL_LIST_EGRESS_ROUTES:
-        return handle_list_egress_routes(typing.cast(dict[str, object], params.get("arguments", {})), config)

    args_raw = params.get("arguments", {})
    if not isinstance(args_raw, dict):
@@ -404,7 +390,7 @@ def handle_tools_call(
            deadline=deadline,
        )
    except TimeoutError:
-        text = format_pending_response_text(config.response_timeout_seconds)
+        text = format_pending_response_text(proposal.id, config.response_timeout_seconds)
        return {
            "content": [{"type": "text", "text": text}],
            "isError": False,
@@ -421,6 +407,54 @@ def handle_tools_call(
    }


+def handle_check_proposal(
+    params: dict[str, object],
+    config: ServerConfig,
+) -> dict[str, object]:
+    """Non-blocking poll of a queued proposal's decision, by id.
+
+    Never creates a Proposal (so `check-proposal` isn't in `TOOLS`); it only
+    reads the queue. Resolution order mirrors the synchronous path's terminal
+    step — a decided proposal is archived here exactly as `handle_tools_call`
+    archives it after `wait_for_response`, so `pending` proposals stay visible
+    to the operator until they're both decided *and* polled."""
+    args_raw = params.get("arguments", {})
+    if not isinstance(args_raw, dict):
+        raise _RpcClientError(ERR_INVALID_PARAMS, "tools/call 'arguments' must be an object")
+    proposal_id = args_raw.get("proposal_id")
+    if not isinstance(proposal_id, str) or not proposal_id.strip():
+        raise _RpcClientError(
+            ERR_INVALID_PARAMS,
+            "check-proposal: 'proposal_id' is required and must be a non-empty string",
+        )
+    proposal_id = proposal_id.strip()
+
+    try:
+        response = _sv.read_response(config.bottle_slug, proposal_id)
+    except FileNotFoundError:
+        # No decision yet — distinguish "still queued" from "unknown id".
+        try:
+            _sv.read_proposal(config.bottle_slug, proposal_id)
+        except FileNotFoundError:
+            return {
+                "content": [{"type": "text", "text": format_unknown_proposal_text(proposal_id)}],
+                "isError": True,
+            }
+        return {
+            "content": [{"type": "text", "text": format_still_pending_text(proposal_id)}],
+            "isError": False,
+        }
+
+    try:
+        _sv.archive_proposal(config.bottle_slug, proposal_id)
+    except OSError as e:
+        raise _RpcInternalError(f"failed to archive proposal: {e}") from e
+    return {
+        "content": [{"type": "text", "text": format_response_text(response)}],
+        "isError": response.status == _sv.STATUS_REJECTED,
+    }
+
+
 def format_response_text(response: "_sv.Response") -> str:
    """Pretty-print a Response for the tool's text content. The agent
    reads the text and decides whether to retry / give up / surface."""
@@ -433,12 +467,35 @@ def format_response_text(response: "_sv.Response") -> str:
    return "\n".join(lines)


-def format_pending_response_text(timeout_seconds: float) -> str:
+def format_pending_response_text(proposal_id: str, timeout_seconds: float) -> str:
+    """Grace-window timeout: the proposal stays queued, and the agent is
+    told the id so it can `check-proposal` instead of re-proposing."""
    return "\n".join([
        "status: pending",
+        f"proposal_id: {proposal_id}",
        (
-            "notes: operator response timed out after "
-            f"{timeout_seconds:g}s; proposal remains queued"
+            f"notes: no operator decision within {timeout_seconds:g}s; the "
+            "proposal remains queued. Poll it (do not re-propose) by calling "
+            f"`check-proposal` with proposal_id={proposal_id!r}."
+        ),
+    ])
+
+
+def format_still_pending_text(proposal_id: str) -> str:
+    return "\n".join([
+        "status: pending",
+        f"proposal_id: {proposal_id}",
+        "notes: still queued; no operator decision yet. Call `check-proposal` again later.",
+    ])
+
+
+def format_unknown_proposal_text(proposal_id: str) -> str:
+    return "\n".join([
+        "status: unknown",
+        f"proposal_id: {proposal_id}",
+        (
+            "notes: no queued proposal with this id for this bottle — the id "
+            "may be wrong, or the proposal was already resolved and read."
        ),
    ])

@@ -525,24 +582,63 @@ class MCPHandler(http.server.BaseHTTPRequestHandler):
        if method == "tools/list":
            return handle_tools_list(req.params)
        if method == "tools/call":
-            # Attribute the proposal to the calling bottle. Single-tenant → the
-            # env slug on `config`; consolidated → the source-IP-resolved
-            # bottle id, so one shared server queues each bottle's proposal
-            # under its own slug.
+            # `list-egress-routes` is read-only introspection. The shared gateway
+            # has no static route table (routes are resolved per request by
+            # source IP), so answer it from the calling bottle's resolved policy.
+            # Otherwise the agent sees an empty allowlist and composes an egress
+            # proposal that *replaces* the live routes instead of extending them
+            # — silently dropping base routes like api.anthropic.com on approval.
+            if req.params.get("name") == _sv.TOOL_LIST_EGRESS_ROUTES:
+                return self._resolved_routes_payload()
+            # `check-proposal` is a non-blocking read of the calling bottle's
+            # own queue — attributed by source IP like a proposal, but it
+            # never queues or blocks.
+            if req.params.get("name") == _sv.TOOL_CHECK_PROPOSAL:
+                return handle_check_proposal(req.params, self._attributed_config(config))
+            # Attribute the proposal to the source-IP-resolved bottle, so the one
+            # shared server queues each bottle's proposal under its own slug.
            return handle_tools_call(req.params, self._attributed_config(config))
        raise _RpcClientError(ERR_METHOD_NOT_FOUND, f"method not found: {method}")

-    def _attributed_config(self, config: ServerConfig) -> ServerConfig:
-        """The ServerConfig with `bottle_slug` bound to *this request's* bottle.
-        Single-tenant (no resolver): unchanged. Consolidated: the bottle id
-        attributed from the source IP — **fail-closed**, an unattributed or
-        unreachable source raises so no proposal is queued under the wrong (or
-        empty) slug."""
+    def _resolver_or_fail(self) -> "PolicyResolver":
+        """This server's policy resolver. A server started without one is a
+        misconfiguration, not a tenancy mode — fail closed rather than
+        attribute (or list) anything."""
        resolver = getattr(self.server, "policy_resolver", None)
        if resolver is None:
-            return config
+            raise _RpcInternalError("supervise server has no policy resolver")
+        return resolver
+
+    def _resolved_routes_payload(self) -> dict[str, object]:
+        """The calling bottle's live egress routes as the `list-egress-routes`
+        JSON payload, resolved by (source_ip, identity token). Fail-closed like
+        `_attributed_config`: an unattributed source or an unreachable
+        orchestrator yields an empty route list (never another bottle's),
+        courtesy of `resolve_client_context`."""
+        resolver = self._resolver_or_fail()
+        headers = getattr(self, "headers", None)
+        token = headers.get(IDENTITY_HEADER, "") if headers is not None else ""
+        conf, _slug, _tokens = resolve_client_context(
+            resolver, self.client_address[0], token,
+        )
+        body = json.dumps(
+            {"routes": [route_to_yaml_dict(r) for r in conf.routes]}, indent=2,
+        )
+        return {"content": [{"type": "text", "text": body}], "isError": False}
+
+    def _attributed_config(self, config: ServerConfig) -> ServerConfig:
+        """The ServerConfig with `bottle_slug` bound to *this request's* bottle:
+        the bottle id attributed from the source IP — **fail-closed**, an
+        unattributed or unreachable source raises so no proposal is queued under
+        the wrong (or empty) slug."""
+        resolver = self._resolver_or_fail()
+        # The agent's MCP client sends the identity token as a request header
+        # (provisioned via `mcp add --header`); the orchestrator requires the
+        # (source_ip, token) pair, so a missing/wrong token fail-closes below.
+        headers = getattr(self, "headers", None)
+        token = headers.get(IDENTITY_HEADER, "") if headers is not None else ""
        try:
-            bottle_id = resolver.resolve_bottle_id(self.client_address[0])
+            bottle_id = resolver.resolve_bottle_id(self.client_address[0], token)
        except PolicyResolveError as e:
            raise _RpcInternalError(f"orchestrator unreachable, cannot attribute: {e}") from e
        if not bottle_id:
@@ -572,8 +668,9 @@ class MCPServer(socketserver.ThreadingMixIn, http.server.HTTPServer):
    allow_reuse_address = True
    daemon_threads = True
    config: ServerConfig = ServerConfig(bottle_slug="")
-    # None → single-tenant (proposals use config.bottle_slug); set → consolidated
-    # (each proposal attributed to the source-IP-resolved bottle).
+    # Set by `serve`; every proposal is attributed to the source-IP-resolved
+    # bottle. The class default is a placeholder — a server without a resolver
+    # fails closed per request (see `_resolver_or_fail`).
    policy_resolver: "PolicyResolver | None" = None


@@ -582,21 +679,21 @@ class MCPServer(socketserver.ThreadingMixIn, http.server.HTTPServer):

 def serve(
    *,
-    bottle_slug: str,
+    resolver: "PolicyResolver",
    port: int = _sv.SUPERVISE_PORT,
    bind: str = "0.0.0.0",
    response_timeout_seconds: float = DEFAULT_RESPONSE_TIMEOUT_SECONDS,
-    resolver: "PolicyResolver | None" = None,
 ) -> typing.NoReturn:
    server = MCPServer((bind, port), MCPHandler)
+    # bottle_slug is a placeholder: every request's proposal is attributed to
+    # the source-IP-resolved bottle (see MCPHandler._attributed_config).
    server.config = ServerConfig(
-        bottle_slug=bottle_slug,
+        bottle_slug="",
        response_timeout_seconds=response_timeout_seconds,
    )
    server.policy_resolver = resolver
-    mode = "multi-tenant" if resolver else f"slug={bottle_slug!r}"
    sys.stderr.write(
-        f"supervise listening on {bind}:{port}; {mode}; "
+        f"supervise listening on {bind}:{port}; multi-tenant; "
        f"tools: {', '.join(t['name'] for t in TOOL_DEFINITIONS)}\n"  # type: ignore[arg-type]
    )
    sys.stderr.flush()
@@ -612,12 +709,13 @@ def serve(
 def main(argv: list[str]) -> int:
    del argv  # config is env-only, no CLI flags
    orch_url = os.environ.get(ORCHESTRATOR_URL_ENV, "").strip()
-    resolver = PolicyResolver(orch_url) if orch_url else None
-    bottle_slug = os.environ.get("SUPERVISE_BOTTLE_SLUG", "")
-    # Consolidated mode resolves the slug per request, so the env slug is
-    # optional there; single-tenant still requires it.
-    if not bottle_slug and resolver is None:
-        sys.stderr.write("supervise: SUPERVISE_BOTTLE_SLUG env is unset\n")
+    if not orch_url:
+        # Resolver-only: without an orchestrator the server can't attribute a
+        # proposal to a bottle, so it must not serve (fail-closed).
+        sys.stderr.write(
+            f"supervise: {ORCHESTRATOR_URL_ENV} is required "
+            "(no single-tenant SUPERVISE_BOTTLE_SLUG fallback)\n"
+        )
        return 2
    port = int(os.environ.get("SUPERVISE_PORT", str(_sv.SUPERVISE_PORT)))
    bind = os.environ.get("SUPERVISE_BIND", "0.0.0.0")
@@ -627,11 +725,10 @@ def main(argv: list[str]) -> int:
        sys.stderr.write(f"supervise: {e}\n")
        return 2
    serve(
-        bottle_slug=bottle_slug,
+        resolver=PolicyResolver(orch_url),
        port=port,
        bind=bind,
        response_timeout_seconds=response_timeout_seconds,
-        resolver=resolver,
    )
    return 0  # serve() does not return

@@ -20,6 +20,10 @@ TOOL_EGRESS_ALLOW = "egress-allow"
 TOOL_GITLEAKS_ALLOW = "gitleaks-allow"
 TOOL_EGRESS_TOKEN_ALLOW = "egress-token-allow"
 TOOL_LIST_EGRESS_ROUTES = "list-egress-routes"
+# Read-only agent tool: poll a queued proposal for the operator's decision
+# without blocking or re-proposing. It never becomes a `Proposal.tool` (no
+# queue record is created for it), so it is intentionally NOT in `TOOLS`.
+TOOL_CHECK_PROPOSAL = "check-proposal"
 TOOLS: tuple[str, ...] = (
    TOOL_EGRESS_ALLOW,
    TOOL_EGRESS_BLOCK,
@@ -156,6 +160,7 @@ __all__ = [
    "TOOLS",
    "TOOL_EGRESS_ALLOW",
    "TOOL_EGRESS_BLOCK",
+    "TOOL_CHECK_PROPOSAL",
    "TOOL_EGRESS_TOKEN_ALLOW",
    "TOOL_GITLEAKS_ALLOW",
    "TOOL_LIST_EGRESS_ROUTES",
@@ -7,6 +7,7 @@ from __future__ import annotations

 import ipaddress
 import os
+import sys


 def is_ip_literal(value: str) -> bool:
@@ -17,6 +18,15 @@ def is_ip_literal(value: str) -> bool:
    return True


+def read_tty_line() -> str:
+    """Mirror `IFS= read -r REPLY </dev/tty`. Falls back to stdin."""
+    try:
+        with open("/dev/tty", "r", encoding="utf-8") as tty:
+            return tty.readline().rstrip("\n")
+    except OSError:
+        return sys.stdin.readline().rstrip("\n")
+
+
 def expand_tilde(path: str) -> str:
    """Expand a leading '~' to $HOME. Leaves paths without a leading
    tilde unchanged. Falls back to the empty string if $HOME is unset
@@ -0,0 +1,57 @@
+# ADR 0005: Keep tracker metadata on issues
+
+- **Status:** Accepted
+- **Date:** 2026-07-18
+- **Deciders:** didericis
+
+## Context
+
+Gitea exposes labels on both issues and pull requests. Applying the same labels
+to both copies planning metadata, creates a synchronization obligation, and
+makes disagreements between the two records possible. At the same time,
+unlabelled objects look accidental unless the repository states which object
+owns the metadata.
+
+The repository already uses issues as work items and PRs as implementations of
+those work items. At this decision's cutoff, all open PRs reference issues, but
+121 of 219 historically merged PRs do not. Manufacturing retrospective issues
+for that history would create records that never participated in planning and
+would make the issue history less truthful.
+
+## Decision
+
+Issues are the canonical tracker records and own labels. Every issue has at
+least one label. An issue opened or left without labels receives
+`Status/Needs Triage` automatically until it is classified.
+
+Pull requests carry no labels. Every new PR deliberately references at least
+one existing issue in its title or description with one of these forms:
+
+- `Closes #123`, `Fixes #123`, or `Resolves #123` when merging completes it.
+- `Part of #123`, `Related to #123`, `Refs #123`, or `References #123` when it
+  contributes without completing it.
+
+Gitea Actions enforces both PR rules as a status check and repairs the empty
+issue-label state. Branch protection makes the PR policy check required.
+
+The policy applies from 2026-07-18 onward. Existing issues may be labelled as
+they are encountered, but closed PRs are grandfathered: no retrospective
+issues or PR labels are created solely to make history conform.
+
+## Consequences
+
+- Classification, priority, and workflow metadata have one source of truth.
+- A PR's issue link is the navigation path to its planning metadata.
+- Multi-PR issues do not require copied or synchronized labels.
+- `Status/Needs Triage` is an intentional fallback, not a final
+  classification.
+- Direct issue creation remains convenient; automation repairs a missing label
+  immediately after creation because Gitea has no native required-label rule.
+- The required check must be configured in branch protection after this
+  workflow lands.
+
+## Links
+
+- Issue #405.
+- `.gitea/workflows/tracker-policy.yml`.
+- `scripts/tracker_policy.py`.
@@ -1,6 +1,6 @@
 # PRD 0023: smolmachines bottle backend

-> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/landscape-containerized-claude.md`.
+> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/agent-sandbox-landscape.md`.

 - **Status:** Superseded (2026-07-11) — was Active
 - **Author:** didericis
@@ -1,6 +1,6 @@
 # PRD 0032: Decompose smolmachines launch and harden bringup sequencing

-> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/landscape-containerized-claude.md`.
+> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/agent-sandbox-landscape.md`.

 - **Status:** Superseded (2026-07-11) — was Active
 - **Author:** didericis-claude
@@ -1,6 +1,6 @@
 # PRD 0038: smolmachines Env Contract and Secret-Safe Injection

-> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/landscape-containerized-claude.md`.
+> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/agent-sandbox-landscape.md`.

 - **Status:** Superseded (2026-07-11) — was Active
 - **Author:** didericis-codex
@@ -1,6 +1,6 @@
 # PRD 0039: smolmachines Capability-Block Remediation

-> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/landscape-containerized-claude.md`.
+> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/agent-sandbox-landscape.md`.

 - **Status:** Superseded (2026-07-11) — was Active
 - **Author:** didericis-codex
@@ -1,6 +1,6 @@
 # PRD 0042: smolmachines Cross-Backend Parity Tests

-> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/landscape-containerized-claude.md`.
+> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/agent-sandbox-landscape.md`.

 - **Status:** Superseded (2026-07-11) — was Active
 - **Author:** didericis-codex
@@ -1,6 +1,6 @@
 # PRD 0057: Promote smolmachines to default backend; convert Docker to example-only

-> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/landscape-containerized-claude.md`.
+> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/agent-sandbox-landscape.md`.

 - **Status:** Superseded (2026-07-11) — was Active
 - **Author:** didericis
@@ -1,6 +1,6 @@
 # PRD 0068: smolmachines backend on Linux

-> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/landscape-containerized-claude.md`.
+> **Superseded (2026-07-11).** The smolmachines backend was removed — Linux now uses the Firecracker backend, macOS uses macos-container. Kept as a historical record; see the removal commit `c07ebca` and `docs/research/agent-sandbox-landscape.md`.

 - **Status:** Superseded (2026-07-11) — was Active
 - **Author:** Claude
@@ -8,16 +8,20 @@
 > **Superseded in part by [PRD 0070](0070-per-host-orchestrator.md) (#351):**
 > the sidecar-consolidation framing here (Stage 1, per-host sidecar; Stage 4,
 > sidecar-as-VM) is taken over by 0070's per-host orchestrator. This PRD still
-> owns the docker-free **image-building** work — Stage 2 (nix-built fixed
-> images, a dependency of 0070) and Stage 3 (in-VM Dockerfile builder).
+> owns the docker-free **image-provisioning** work — Stage 2 (pull the fixed
+> images from an OCI registry instead of building them with host Docker, a
+> dependency of 0070) and Stage 3 (in-VM Dockerfile builder).

 ## Summary

 Make the Firecracker backend depend on **firecracker + KVM only**, removing
 Docker from the host. Two moves get us there: run the **sidecar bundle as a
 persistent, per-host service** (eventually a Firecracker VM) instead of a
-per-bottle container, and **build agent rootfs images without a host Docker
-daemon** (nix for the fixed images; an in-VM builder for user Dockerfiles).
+per-bottle container, and **provision rootfs images without a host Docker
+daemon** — pull the fixed images (orchestrator/gateway/infra) from an OCI
+registry and unpack them daemonlessly, and build user Dockerfiles in an in-VM
+builder. The images are still *built* with Docker, but off the launch host
+(CI / a publish step) and pushed to the registry; the launch host only pulls.

 ## Motivation

@@ -93,13 +97,51 @@ torn down at exit."
 Can ship as a container first (quick resource/ops win) and become a VM in
 Stage 4.

-### Stage 2 — Fixed images built with nix (no Docker)
+### Stage 2 — Fixed rootfs prebuilt + pulled as an artifact (no host Docker)

-The images bot-bottle *ships* — the sidecar, the agent base, and the builder
-(Stage 3) — are built declaratively with nix (`nixos-generators` /
-`make-ext4-fs` / `pkgs.dockerTools` for the rootfs), producing an ext4 or
-tar with correct ownership. Removes Docker for everything we own and gives
-the rootless-rootfs correctness (#347) for free on these images.
+The one fixed image the Firecracker backend needs at launch — the combined
+**infra** rootfs the infra VM boots (orchestrator control plane + gateway +
+buildah, with the control-plane init as PID 1) — is **prebuilt end-to-end off
+the launch host and published as a versioned, ready-to-boot ext4 artifact**.
+The launch host **downloads the `.ext4` and boots it directly** — no
+`docker build`, no `docker export`, no `mke2fs`, no image tooling at all.
+
+This is possible because the infra rootfs is already **host- and
+bottle-agnostic**: the per-boot bits (authorized_keys, guest IP) arrive on the
+**kernel cmdline**, not in the rootfs (see `build_base_rootfs_dir`). So one
+published ext4 boots on any launch host.
+
+- **Artifact.** `rootfs.ext4` + a `rootfs.ext4.sha256`, published as a Gitea
+  **generic package** (`bot-bottle-firecracker-infra/<tag>`) — generic packages take
+  arbitrary large binaries (no attachment size cap / file-type allowlist that
+  release attachments impose). The matching `vmlinux` kernel can ship the same
+  way, so the whole VM is fetchable.
+- **Pull.** The launch host `GET`s
+  `…/api/packages/<owner>/generic/bot-bottle-firecracker-infra/<tag>/rootfs.ext4` (+
+  `.sha256`) for its pinned tag, verifies the checksum, caches it under the
+  tag, and attaches it as the infra VM's root disk. Host prerequisite is an
+  HTTP client — nothing else. Public packages need no auth to pull; a token
+  with `read:package` covers a private instance.
+- **Registry.** The artifact base URL + owner are configurable, defaulting to
+  this deployment's Gitea (`https://gitea.dideric.is` / `didericis`);
+  overridable via env for other deployments / air-gapped mirrors.
+- **Versioning.** A pinned tag bumped when the infra rootfs contents change
+  (bot_bottle's shipped files, the base deps, or the init), so a launch host
+  pulls the artifact matching its code and a content change can't silently
+  boot a stale rootfs. A checksum mismatch fails closed.
+- **Publish.** A `publish` step (CLI subcommand / CI job) runs the full
+  pipeline **on a build/CI host** — `docker build` the three Dockerfiles →
+  export → inject guest boot → `mke2fs` → upload the `.ext4` + `.sha256`.
+  Building still uses Docker, but never on the launch/runner host, which is
+  the one #348 needs unprivileged.
+- **Dev escape hatch.** An explicit opt-in still builds the rootfs locally
+  with Docker (for iterating on the Dockerfiles without a publish
+  round-trip); it is never the default path.
+
+Removes Docker from the launch host entirely for the fixed image, and the
+launch host needs no OCI/rootfs tooling — just fetch + boot. The build-time
+cache / build-time-egress open problems a from-scratch build would face don't
+arise: the launch host never builds, it downloads a finished disk.

 ### Stage 3 — User Dockerfiles built in a builder VM (the unlock)

@@ -0,0 +1,156 @@
+# PRD prd-new: Consolidate infra backend for Docker
+
+- **Status:** Active
+- **Author:** Claude
+- **Created:** 2026-07-20
+- **Issue:** #431
+
+## Summary
+
+The Docker backend runs two containers — `bot-bottle-orch-gateway` (gateway
+data plane) and `bot-bottle-orchestrator` (control plane) — where the
+macOS and Firecracker backends already run a single combined infra
+unit. This PRD collapses Docker to the same model: one `bot-bottle-infra`
+container running both processes under the `gateway_init` supervise tree, a
+restructured `Dockerfile.infra` as the shared gateway+orchestrator base,
+and a handful of extracted shared utilities (CA cert polling, teardown
+sequence, launch skeleton) that are currently duplicated across all three
+`consolidated_launch.py` files.
+
+## Goals / success criteria
+
+- Docker backend starts exactly one infra container instead of two.
+- `Dockerfile.infra` is the shared base image (gateway + orchestrator, no
+  buildah); the Firecracker image layers buildah on top of it.
+- The orchestrator process runs under the `gateway_init` supervise tree
+  inside the combined container (one PID-1, one restart/health surface).
+- CA cert polling, the teardown sequence, and the shared launch skeleton
+  (ensure-infra → register → provision → return context) live in a single
+  shared module; all three backends import from it.
+- No functional change to macOS or Firecracker launch paths.
+
+## Non-goals
+
+- Changing per-bottle isolation — agents stay one-VM/container-each.
+- Consolidating transport implementations (`DockerGatewayTransport`,
+  `AppleGatewayTransport`, `SshGatewayTransport`) — these are already the
+  right abstraction boundary.
+- macOS DHCP-inversion of registration order — irreducible backend
+  difference, stays as-is.
+- Any changes to the orchestrator RPC protocol or the attribution model.
+
+## Design
+
+### Dockerfile restructuring
+
+**Current shape:**
+
+- `Dockerfile.gateway` — data plane (mitmproxy, gitleaks, git, openssh,
+  supervise daemons)
+- `Dockerfile.orchestrator` — control plane (python:3.12-slim + bot_bottle
+  package; stdlib-only, no third-party deps)
+- `Dockerfile.infra` — Firecracker only: `FROM bot-bottle-gateway` +
+  buildah + `COPY --from bot-bottle-orchestrator`
+
+**New shape:**
+
+- `Dockerfile.gateway` — unchanged
+- `Dockerfile.orchestrator` — unchanged (single definition of orchestrator
+  content; both Docker infra and Firecracker infra `COPY --from` it)
+- `Dockerfile.infra` — **shared base**: `FROM bot-bottle-gateway` + `COPY
+  --from bot-bottle-orchestrator` (no buildah — Docker infra image)
+- `Dockerfile.infra.fc` — Firecracker only: `FROM bot-bottle-infra` +
+  buildah/crun/netavark/aardvark-dns (layered on the shared base, same net
+  result as today)
+
+The comment in `Dockerfile.infra` that says "the docker backend keeps
+orchestrator + gateway as separate images; this combined image exists only
+for the Firecracker single-VM cut" is removed.
+
+### Orchestrator in the supervise tree
+
+`gateway_init` already supervises the data-plane daemons (egress, git-http,
+supervise-MCP). The orchestrator control plane is added as another supervised
+process: `python3 -m bot_bottle.orchestrator --host 0.0.0.0 --port <port>
+--broker stub`.
+
+The orchestrator source is bind-mounted (`/app` → repo root, as today) so
+dev live-reload still works. `source_hash`-based container recreation in
+`OrchestratorService.ensure_running` continues to apply — a code change
+recreates the combined infra container, which bounces both gateway and
+orchestrator. This is acceptable: the docker backend is a dev/legacy target
+where in-flight egress connections across a code deploy are not a hard
+requirement.
+
+### `OrchestratorService` changes
+
+`OrchestratorService` currently starts two containers in sequence: gateway
+first (`DockerGateway.ensure_running`), then orchestrator. After this PRD:
+
+- Single `docker run` of `bot-bottle-infra:latest`
+- Container name: `bot-bottle-infra` (replaces `bot-bottle-orch-gateway` +
+  `bot-bottle-orchestrator`)
+- Published ports: `127.0.0.1:{host_port}:8099` for the control plane
+  (`gateway_init` listens on a fixed internal port 8099; the caller-chosen
+  host port maps to it)
+- Bind mounts: repo root + host root (same as today)
+- `DockerGateway` becomes an implementation detail of `OrchestratorService`
+  rather than a separately started container; the gateway image name
+  (`GATEWAY_IMAGE`) is no longer referenced at runtime, only at build time
+  for the `Dockerfile.infra` base
+
+The `_gateway()` / `ensure_running` two-step in `OrchestratorService` is
+replaced by a single `_run_infra_container()`.
+
+### Shared backend utilities
+
+Three items are duplicated across
+`backend/docker/consolidated_launch.py`,
+`backend/macos_container/consolidated_launch.py`, and
+`backend/firecracker/consolidated_launch.py`:
+
+1. **CA cert polling loop** — `deadline = time.monotonic() + timeout; while
+   ...: try fetch CA; sleep` — extracted to
+   `backend/consolidated_util.py:poll_ca_cert(transport, *, timeout)`.
+
+2. **Teardown sequence** — `OrchestratorClient(url).teardown_bottle(id)` +
+   `deprovision_git_gate(transport, id)` — extracted to
+   `backend/consolidated_util.py:teardown_consolidated(url, transport,
+   bottle_id)`.
+
+3. **Launch skeleton** — all three follow: ensure-infra → allocate/register
+   → provision git-gate → fetch CA cert → return launch context. The macOS
+   inversion (agent starts before registration, source IP from DHCP) is the
+   only deviation. Extract a shared `_provision_bottle(transport, bottle_id,
+   plan, orchestrator_url)` helper covering the register → provision →
+   return-token steps; the backends keep their own `launch_consolidated`
+   wrappers for the before/after (infra-ensure + agent-start + IP
+   allocation), calling the shared helper.
+
+The new `backend/consolidated_util.py` module holds only backend-neutral,
+transport-agnostic logic. All three backends import from it.
+
+## Implementation chunks
+
+1. **(this PR)** Dockerfile restructuring: rename current `Dockerfile.infra`
+   content to `Dockerfile.infra.fc`; write new `Dockerfile.infra` as
+   gateway+orchestrator base. Update Firecracker image-build references from
+   `Dockerfile.infra` → `Dockerfile.infra.fc`.
+
+2. Add orchestrator process to `gateway_init` supervise tree.
+
+3. Collapse `OrchestratorService` to a single-container start; rename
+   container from `bot-bottle-orch-gateway`/`bot-bottle-orchestrator` →
+   `bot-bottle-infra`; update image name constant.
+
+4. Extract `backend/consolidated_util.py` with `poll_ca_cert`,
+   `teardown_consolidated`, and `_provision_bottle`; update all three
+   `consolidated_launch.py` files to import from it.
+
+5. Update tests that reference the old container names or two-container
+   startup sequence.
+
+## Open questions
+
+None — the supervise-tree approach and shared Dockerfile layering were
+confirmed in issue #431.
@@ -0,0 +1,125 @@
+# PRD prd-new: Non-blocking supervise (async approval + proposal polling)
+
+- **Status:** Draft
+- **Author:** didericis
+- **Created:** 2026-07-18
+- **Issue:** #412
+
+## Summary
+
+The per-bottle supervise MCP server (`bot_bottle/supervise_server.py`)
+answers `tools/call` **synchronously**: it queues the agent's proposal and
+blocks the tool call polling for the operator's decision. On timeout it
+returns `status: pending` and leaves the proposal queued — but it hands the
+agent **no proposal id** and offers **no way to poll a specific pending
+proposal**, so the only way to learn the outcome is to re-propose (a
+duplicate).
+
+This PRD makes the MCP flow non-blocking and pollable, so an approval can
+happen out-of-band (a human taking minutes-to-hours in a review console)
+without holding an HTTP request open or wedging the agent:
+
+1. Include the `proposal_id` in the `pending` response.
+2. Add a `check-proposal` MCP tool: a non-blocking status lookup by
+   proposal id.
+3. Keep the short synchronous grace window for the common "operator is
+   right there" fast path.
+
+## Problem
+
+`handle_tools_call` → `_sv.wait_for_response(...)` blocks up to
+`SUPERVISE_RESPONSE_TIMEOUT_SECONDS` (default 30s). Two problems follow:
+
+- **Human latency ≠ tool-call latency.** A real review — rendered diff,
+  RBAC routing to an approver, someone tapping approve on their phone — is
+  minutes-to-hours. Holding the MCP request open that long is fragile
+  (proxy/keepalive timeouts, the mitmproxy egress hop, and the agent
+  harness's own tool-call timeout, which a long block can trip and stall
+  the whole turn).
+- **No resume path.** The pending fallback already exists, but without a
+  proposal id and a poll tool the agent can't reconnect to that specific
+  decision — it re-proposes, duplicating the queue entry.
+
+This is also the precondition for the planned web-console human-review
+flow (RBAC, audit retention, mobile) — see issue #412.
+
+**Safety note:** the MCP tools only *propose* policy changes; enforcement
+stays at the egress proxy and the git-gate. Returning early on `pending`
+therefore opens no hole — the agent still cannot egress or push anything
+unapproved.
+
+## Goals / success criteria
+
+- A `pending` MCP response carries the `proposal_id`.
+- An agent can call `check-proposal(proposal_id)` and get the current
+  state (`pending` | `approved` | `modified` | `rejected`) **without
+  blocking** and **without creating a new proposal**.
+- The synchronous fast path (operator approves within the grace window) is
+  unchanged: the first `tools/call` still returns the decision directly.
+- No change to enforcement, attribution (source-IP → bottle), or the
+  operator-side queue/response schema.
+
+## Non-goals
+
+- The git-gate `pre-receive` path (it is synchronous by nature and cannot
+  poll — its async variant is reject-fast + re-push; tracked as a
+  follow-up).
+- Backpressure / in-flight-proposal caps.
+- MCP server→client notifications (event-driven resume).
+- Any web-console UI (this PRD is the protocol groundwork it needs).
+
+## Design
+
+### `pending` response carries the id
+
+`handle_tools_call`'s timeout branch formats the pending text with the
+`proposal.id` and a pointer to `check-proposal`, so the agent knows what to
+poll.
+
+### `check-proposal` tool
+
+A new read-only MCP tool (`TOOL_CHECK_PROPOSAL = "check-proposal"`),
+attributed to the calling bottle by source IP exactly like the proposal
+tools. Input: `{ "proposal_id": string }`. Behavior:
+
+1. `read_response(slug, id)` →
+   - **found**: archive the proposal (same terminal step the synchronous
+     path takes) and return the decision via `format_response_text`;
+     `isError` iff rejected.
+2. **not found** → `read_proposal(slug, id)` →
+   - **found**: still queued → return `status: pending`.
+   - **not found**: unknown id, or already resolved-and-archived (e.g. a
+     second poll) → return `status: unknown`, `isError: true`.
+
+Both lookups already raise `FileNotFoundError` when absent
+(`queue_store.py`), so the handler needs no new store methods. `check-`
+`proposal` is the only path (besides the synchronous response) that
+archives, so a proposal that times out to `pending` stays visible to the
+operator until it is decided and then polled.
+
+### Grace window
+
+Left at the existing 30s default (`SUPERVISE_RESPONSE_TIMEOUT_SECONDS`),
+which doubles as the instant-approve fast path. Tuning it down is an
+operator setting, not a code change; noted for the console rollout.
+
+## Implementation chunks
+
+1. **(this PR)** `TOOL_CHECK_PROPOSAL` constant; `check-proposal` tool
+   definition + `handle_check_proposal`; dispatch wiring; `proposal_id` in
+   the pending text; unit tests. Files: `bot_bottle/supervise_types.py`,
+   `bot_bottle/supervise.py` (re-export), `bot_bottle/supervise_server.py`,
+   `tests/unit/test_supervise_server.py`.
+2. **(follow-up)** git-gate `pre-receive` reject-fast + re-push.
+3. **(follow-up)** per-bottle in-flight-proposal backpressure cap.
+4. **(follow-up)** MCP notifications for event-driven resume; web-console
+   review flow (RBAC, audit retention) on top.
+
+## Open questions
+
+- Should a resolved-but-unpolled proposal auto-archive after some TTL, or
+  only on poll? (Leaning: only on poll, so a decision is never lost to a
+  reaper before the agent sees it.)
+- Does the agent harness need an explicit "you have a pending proposal"
+  nudge, or is returning `pending` from the original call enough? (Deferred
+  to the notifications chunk.)
@@ -1,32 +1,81 @@
 # Landscape: AI-agent sandbox tools

-A broader survey than [`landscape-containerized-claude.md`](landscape-containerized-claude.md),
-which focused on Claude-Code-specific containerizers. This one covers
-general AI-agent sandbox / containment projects — some Claude-specific,
-some agent-agnostic, some hosted SaaS — and contrasts them with
-bot-bottle's design.
+Survey of AI-agent sandbox and containment projects — including local
+coding-agent wrappers, agent-agnostic runtimes, hosted platforms, and
+governance layers — contrasted with bot-bottle's design. The original
+Claude-Code-specific containerizer survey was folded into this note on
+2026-07-20 so there is one landscape and one positioning verdict.

-Research conducted 2026-05-11.
+Research conducted 2026-05-11. CubeSandbox added 2026-07-18 (see its
+per-project note and the addendum at the end). Also updated 2026-07-18:
+bot-bottle no longer uses **pipelock** — outbound DLP is now bot-bottle's
+own (deliberately simple) egress scanner (a mitmproxy addon with custom
+detectors, PRD 0017 / 0052), and git-push secret scanning is handled by
+**gitleaks** in the git-gate. "pipelock" below has been replaced with the
+current mechanism; it survives only in older PRDs as history.
+
+Updated again 2026-07-18: six additional tools added (Cleanroom,
+container-use, Docker sbx, Anthropic srt, Microsoft AGT, Open Agent
+Passport); an **Agent-tailored policy** row added to the comparison table;
+a separate Governance layers section added for AGT and OAP. See the
+second addendum at the end.
+
+Updated 2026-07-20: the borrowable-ideas status was reconciled with the
+current implementation. In-flight credential injection and the microVM
+backends have shipped, while per-use SSH confirmation was superseded by
+keeping git credentials out of the agent entirely.
+
+Also updated 2026-07-20: **E2B and Daytona added as first-class entries.**
+Earlier revisions mentioned E2B only as the API and lifecycle model that
+CubeSandbox implements, and omitted Daytona entirely. That was a survey gap,
+not a principled scope exclusion: both are major hosted sandbox platforms and
+belong in this landscape even though they target platform builders rather than
+bot-bottle's local single-operator workflow.

 ## Summary

-Eight projects surveyed. None duplicate bot-bottle's combination of
-local Docker, declarative JSON manifest, per-agent egress allowlist via
-pipelock, and bottle/agent split. Two clusters stand out:
+The main table compares bot-bottle against fifteen isolation/sandbox tools.
+Governance/pre-action authorization and credential-only layers are covered
+separately because they don't provide VM or container isolation. None
+duplicate bot-bottle's combination of local
+VM-per-bottle isolation, a declarative per-role manifest, per-agent
+egress allowlist + outbound-content DLP, bottle/agent split, and the
+composable `extends:` policy model. Three clusters stand out:

 - **Closest neighbours** — agent-safehouse and litterbox: local,
  single-user, thin wrappers over an existing OS primitive
  (`sandbox-exec`, Podman + Landlock).
- **Different category** — tilde.run (hosted SaaS), boxlite and
-  microsandbox (microVM libraries for platform builders), endo-familiar
+- **Different category (isolation)** — tilde.run (hosted SaaS), boxlite
+  and microsandbox (microVM libraries for platform builders), E2B and Daytona
+  (hosted sandbox platforms), CubeSandbox (self-hosted multi-tenant microVM
+  service), endo-familiar
  (capability-security paradigm, no OS isolation).
+- **New: governance/pre-action layers** — Microsoft AGT and Open Agent
+  Passport (OAP): framework-embedded tool-call interceptors with
+  per-agent declarative policy. Closest competitors on agent-tailored
+  policy, but operate at the tool-call level rather than providing
+  network/filesystem isolation; they complement rather than substitute.

-The microVM cluster (matchlock, smolmachines, boxlite, microsandbox) is
-the most relevant for the v2 isolation discussion in
+The microVM cluster (matchlock, smolmachines, boxlite, microsandbox,
+CubeSandbox) is the most relevant for the v2 isolation discussion in
 [`stronger-isolation-alternatives.md`](stronger-isolation-alternatives.md):
 libkrun and Apple's Virtualization.framework have made local microVMs
-ergonomic enough that a `"runtime": "microvm"` option on a bottle is now
-plausible without a heavy stack.
+ergonomic enough that microVMs are **now bot-bottle's default backend**
+(Firecracker on KVM Linux, Apple Container on macOS), with Docker kept
+only as a legacy fallback for CI / hosts without KVM or Apple Container.
+That discussion has since shipped, not just been theorized.
+
+**The one that matters most for positioning is CubeSandbox** — it ships
+bot-bottle's bundle of default-deny egress allowlisting, full audit logs, and
+in-flight credential custody *combined with* per-sandbox microVM isolation,
+open-source under Apache 2.0, with Tencent Cloud behind it and 10.4k
+stars. It's a self-hosted multi-tenant service for platform builders, not
+a single-user declarative tool, so it doesn't collide head-on — but it
+narrows the "nobody else bundles egress custody + credential injection"
+claim that the monetization positioning leans on. Daytona now also offers
+domain/CIDR firewall policy plus in-flight header credential substitution and
+response scrubbing, although its higher tiers are not default-deny and its
+production platform is proprietary. See the addendum.

 ## Per-project notes

@@ -63,7 +112,8 @@ plausible without a heavy stack.

 ### agent-safehouse
 - **Source**: https://agent-safehouse.dev/ ; https://github.com/eugene1g/agent-safehouse
- **License**: Apache 2.0 (~1,400 stars)
+- **HN launch**: [#47301085](https://news.ycombinator.com/item?id=47301085) (March 12 2026) — 823 points
+- **License**: Apache 2.0 (~1,781 stars at launch)
 - **Isolation**: macOS `sandbox-exec` (Seatbelt) profiles — kernel-level
  syscall interception, no container.
 - **Locality**: Local, macOS only.
@@ -73,6 +123,16 @@ plausible without a heavy stack.
 - **Config**: Shell functions or custom `sandbox-exec` profile files;
  LLM-assisted profile generation supported.
 - **Network policy**: Not addressed.
+- **Notable from HN thread**: Creator acknowledged the project is "just a
+  policy-generator for `sandbox-exec` — no dependencies, no daemons, no
+  subscription; I did put in many hours to identify the minimum required
+  permissions for agents to continue working." Simon Willison noted that
+  evaluating whether a sandboxing tool actually works as intended is hard.
+  Top community sentiment: *"I honestly think that sandboxing is currently
+  THE major challenge that needs to be solved for the tech to fully realise
+  its potential."* The macOS Docker gap (Docker for Mac runs inside a Linux
+  VM, so `sandbox-exec` is the only native primitive for bare-metal macOS
+  processes) was the stated motivation.
 - **Maturity**: Active through March 2026.

 ### matchlock
@@ -155,73 +215,474 @@ plausible without a heavy stack.
  also supported.
 - **Maturity**: Active through April 2026.

+### E2B *(added 2026-07-20)*
+
+- **Source**: https://github.com/e2b-dev/e2b ; https://e2b.dev/docs
+- **License**: Apache 2.0 (~12.4k stars); commercial hosted service with
+  self-hosting/BYOC support.
+- **Isolation**: Firecracker microVM per sandbox.
+- **Locality**: Cloud-hosted by default; self-hosting uses Terraform on AWS or
+  GCP (with other targets documented as works in progress).
+- **Agent integration**: LLM-agnostic Python and JavaScript/TypeScript SDKs;
+  code-interpreter and desktop-sandbox products. Platform primitive rather
+  than a coding-agent wrapper.
+- **Config**: Programmatic SDK/API plus templates. Network configuration
+  supports internet on/off, outbound allow/deny rules, and a custom egress
+  proxy.
+- **Network policy**: Configurable per sandbox, but not documented as
+  default-deny and no built-in outbound-content DLP is documented.
+- **Credentials**: Environment variables passed to the sandbox are explicitly
+  not private at the OS level. No built-in in-flight application-credential
+  injection is documented.
+- **Persistence**: Full memory + filesystem pause/resume, snapshots, and
+  auto-resume. Continuous runtime is tier-limited, while paused sandboxes are
+  retained indefinitely.
+- **Maturity**: Established hosted platform and the API compatibility target
+  used by CubeSandbox.
+
+### Daytona *(added 2026-07-20)*
+
+- **Source**: https://github.com/daytonaio/daytona ;
+  https://www.daytona.io/docs/
+- **License**: Current production platform is proprietary. The former AGPL
+  repository remains public but is no longer maintained after Daytona moved
+  production development closed-source in June 2026.
+- **Isolation**: Hosted container sandboxes by default, with separate Linux
+  and Windows VM sandbox classes for dedicated-OS workloads. Each sandbox has
+  its own filesystem and network stack; VM-only features include memory
+  pause/resume and forking.
+- **Locality**: Hosted multi-tenant service, with dedicated/custom regions and
+  customer runners available.
+- **Agent integration**: LLM/framework-agnostic SDKs (Python, TypeScript, Go,
+  Ruby, Java), API, and CLI; official agent-framework guides. Platform
+  primitive rather than a local coding-agent wrapper.
+- **Config**: Programmatic per-sandbox image/snapshot, resources, lifecycle,
+  firewall, and secrets.
+- **Network policy**: Per-sandbox IPv4/domain allowlists and block-all mode,
+  subordinate to organization/tier policy. Full internet access is the
+  default on higher tiers, so it is configurable rather than uniformly
+  default-deny.
+- **Credentials**: First-class secret manager with the same phantom-token
+  pattern as bot-bottle: the sandbox environment gets an opaque placeholder,
+  an HTTPS proxy substitutes the real secret in headers only for allowed
+  hosts, and responses are scrubbed back to the placeholder.
+- **Persistence**: Persistent filesystem for stopped container sandboxes;
+  memory + filesystem pause/resume for VM sandboxes; snapshots and configurable
+  auto-stop.
+- **Maturity**: Production commercial platform. Notable April 2026 credential
+  exposure was patched; the June 2026 closed-source transition materially
+  changes its transparency/self-hosting posture.
+
+### Other hosted runtimes carried forward from the earlier survey
+
+- **Northflank Sandboxes** — hosted or customer-cloud, microVM-backed
+  containers with SDK-managed lifecycle, optional persistent volumes, and
+  sub-second claimed boot. This is a platform primitive for untrusted code and
+  agents, not a local agent wrapper or role-policy layer.
+- **Cloudflare Sandbox SDK** — Workers/Durable Objects API over VM-isolated
+  Linux containers for command, file, process, and service execution. It is a
+  hosted TypeScript platform primitive; application authentication,
+  authorization, and credential-proxy patterns remain the integrator's job.
+
+Both belong to the same “build your agent platform on this runtime” category as
+E2B and Daytona. They were named but not analyzed in depth by the original
+Claude-specific note, so they remain outside the main comparison table rather
+than being presented with false precision.
+
+### CubeSandbox *(added 2026-07-18)*
+- **Source**: https://github.com/TencentCloud/CubeSandbox ;
+  HN launch https://news.ycombinator.com/item?id=47863430
+- **License**: Apache 2.0 (~10.4k stars). By Tencent Cloud; described as
+  "battle-tested, production-ready" infra already running in Tencent
+  Cloud. Rust / Go / C.
+- **Isolation**: MicroVMs via RustVMM + KVM — "each sandbox gets its own
+  Guest OS kernel, no Docker shared-kernel escapes." Hardware-level
+  isolation, dedicated kernel per instance.
+- **Locality**: Self-hosted, but **server/cluster-oriented**, not a
+  single-user local CLI. Deploy guides target PVM cloud VMs, bare metal,
+  and dev. A single 96-vCPU host is claimed to run 2,000+ concurrent
+  sandboxes.
+- **Agent integration**: **Drop-in E2B SDK replacement** (single env-var
+  change) — the headline compatibility claim. OpenClaw assistant
+  integration; general LLM-code execution. Aimed at platform builders,
+  not one developer's laptop.
+- **Config**: Programmatic via the E2B-compatible SDK. No declarative
+  manifest.
+- **Network policy**: This is the striking part — **domain allowlists,
+  instant block on unauthorized egress, full audit logs, per-sandbox
+  traffic tokens, policy-routing egress**, enforced by an eBPF-based
+  virtual switch giving kernel-level network isolation. Closest match yet
+  to bot-bottle's own default-deny + per-bottle allowlist egress model.
+- **Credentials**: **Credential vault** — agents call external APIs / LLMs
+  while "keys never enter the sandbox, model context, or logs." Same
+  in-flight-injection idea as matchlock, but productized as a vault.
+- **Performance**: <60ms cold start (claimed 2.5–50× faster than
+  alternatives), <5MB memory per instance; millisecond snapshot rollback
+  is upcoming.
+- **Maturity**: Open-sourced July 2026 off production Tencent Cloud use;
+  most-starred project in this set (~10.4k).
+
+### Cleanroom *(added 2026-07-18)*
+- **Source**: https://github.com/buildkite/cleanroom
+- **License**: Apache 2.0
+- **Isolation**: MicroVM — Firecracker on Linux, Virtualization.framework
+  on macOS. Digest-pinned OCI images.
+- **Locality**: Self-hosted server (CI-oriented).
+- **Agent integration**: Generic process sandbox; CI-first, not a
+  Claude/agent wrapper.
+- **Config**: `cleanroom.yaml` in the repo being sandboxed defines egress
+  rules, resources, and network policy. Cleanroom resolves this from the
+  commit being run.
+- **Network policy**: Default-deny + per-repo hostname allowlist (resolved
+  from DNS answers + destination IP:port). Co-hosted services on the same
+  IP:port are not distinguished. OIDC-backed auth for remote servers.
+- **Credentials**: Host-side only; not injected in-flight but not present
+  in the VM.
+- **Notable**: Policy lives in the *repo being sandboxed*, not in an
+  agent-role definition — closer to per-repo scoping than per-role.
+  Supports Docker-inside-sandbox (`services.docker.required: true`), OIDC
+  authorization, suspend/resume lifecycle.
+- **Maturity**: Active Buildkite product.
+
+### container-use *(added 2026-07-18)*
+- **Source**: https://github.com/dagger/container-use
+- **License**: Apache 2.0
+- **Isolation**: Docker container per agent + git worktree per agent.
+  Containers share the host kernel; stronger than bare host but weaker
+  than microVM.
+- **Locality**: Local.
+- **Agent integration**: MCP stdio server — Claude Code, Cursor, Windsurf.
+  `claude mcp add container-use -- container-use stdio`.
+- **Config**: None for security policy. Environments are provisioned on
+  demand; no allowlist or credential config.
+- **Network policy**: Not addressed.
+- **Notable**: Per-agent git branches (`container-use/<env_name>`);
+  parallel agents without filesystem conflict; real-time log visibility
+  and terminal attach for intervention; git-based review workflow.
+  Oriented toward parallel development safety, not security containment.
+- **Maturity**: Early development, active.
+
+### Docker sbx *(added 2026-07-18)*
+- **Source**: Docker proprietary (`sbx` CLI, separate from `docker`).
+- **License**: Proprietary.
+- **Isolation**: MicroVM (Docker's own implementation) — each session gets
+  its own kernel, Docker daemon inside the VM, and filesystem.
+- **Locality**: Local (macOS and Windows; does not require Docker Desktop).
+- **Agent integration**: Explicit wrapper — Claude Code, Codex, Gemini
+  CLI, Copilot CLI, Kiro. Launches agent inside the VM with
+  `--dangerously-skip-permissions` by default.
+- **Config**: Open / Balanced / Locked Down network presets at launch. No
+  per-role manifest.
+- **Network policy**: Default-deny; preset levels control strictness. TUI
+  dashboard shows a live log of every outbound connection (allowed and
+  blocked) with point-and-click allow/block for hosts.
+- **Credentials**: OS keychain + host-side proxy injection — API keys
+  never enter the VM.
+- **Notable**: Best DX among microVM tools (one command, works like native
+  yolo Claude but inside a VM); branch mode creates a git worktree in
+  `.sbx/`. Network policy is preset-based, not role-declarative.
+- **Maturity**: GA 2026.
+
+### Anthropic srt *(added 2026-07-18)*
+- **Source**: https://github.com/anthropic-experimental/sandbox-runtime
+  (`@anthropic-ai/sandbox-runtime` on npm, `sandbox-runtime` on PyPI)
+- **License**: Apache 2.0 (experimental).
+- **Isolation**: OS-level only — Seatbelt (`sandbox-exec`) on macOS,
+  bubblewrap on Linux, WFP (Windows Filtering Platform) account-fenced on
+  Windows. **No container or VM.** Lowest overhead in the set.
+- **Locality**: Local.
+- **Agent integration**: Claude Code's sandboxed bash tool uses this
+  internally. Can wrap any arbitrary process (`srt <command>`). Cloud
+  Claude Code sessions use full microVMs instead.
+- **Config**: Programmatic per-invocation — allow/deny path lists for
+  filesystem; allow/denylist for network (HTTP proxy + SOCKS5).
+- **Network policy**: Proxy-based filtering (HTTP + SOCKS5); domain
+  allowlist/denylist enforced at proxy layer. Custom proxy supported
+  (e.g. mitmproxy for inspection + audit). Processes that ignore proxy
+  env vars may bypass filtering on some platforms.
+- **Notable**: Cross-platform (macOS/Linux/Windows); wraps any process,
+  not just agents; no role/manifest concept. Annotated as a research
+  preview — APIs may change.
+- **Maturity**: Early research preview.
+
+## Claude-specific wrappers and developer environments
+
+These projects were the focus of the original containerized-Claude survey.
+They remain useful comparisons for local developer experience, but most are
+templates or wrappers rather than policy-bearing sandbox platforms, so they
+are grouped here instead of widening the main table further.
+
+### claudebox
+
+- **Source**: https://github.com/RchGrav/claudebox
+- **Isolation**: Docker, with per-project images, authentication state, and
+  configuration.
+- **Agent integration**: Claude Code wrapper with 15+ preconfigured language
+  and task profiles.
+- **Network policy**: Per-project firewall allowlists.
+- **Closest overlap**: local one-command developer workflow and project-scoped
+  network policy.
+- **Difference**: profiles describe development toolchains, not named agent
+  roles. There is no bottle/agent split, composable role manifest, provider
+  plugin layer, or outbound-content DLP.
+
+### Spritz / claude-code-sandbox
+
+- **Source**: https://github.com/textcortex/claude-code-sandbox (archived;
+  points to its successor, Spritz).
+- **Isolation**: The original project ran Claude Code in local Docker with
+  bypass permissions; Spritz moved toward Kubernetes-native multi-agent
+  infrastructure.
+- **Difference**: the successor targets cluster orchestration rather than a
+  low-dependency local launcher. It is architecturally closer to hosted or
+  Kubernetes platform runtimes than to bot-bottle's single-operator CLI.
+
+### Trail of Bits claude-code-devcontainer
+
+- **Source**: https://github.com/trailofbits/claude-code-devcontainer
+- **Isolation**: A Docker devcontainer that exposes only project files and is
+  designed to run Claude Code with `bypassPermissions` for security audits and
+  untrusted-code review.
+- **Difference**: a hardened, reusable environment definition rather than an
+  agent launcher or fleet. It has no named-role manifest, per-role credential
+  custody, supervision plane, or multi-backend abstraction.
+
+### Smaller wrappers and official templates
+
+Projects such as `arezi/claude-sandbox`, `nkrefman/claude-sandbox`, and
+`VishalJ99/claude-docker`, plus Docker/Anthropic devcontainer templates, prove
+there is steady demand for “Claude in a container.” They are deliberately
+small launch/build configurations. They compete on setup simplicity, not on
+role-aware policy, credential custody, persistent supervision, or a fleet
+model, and are better treated as a product category than as individual rows.
+
+### SuperHQ
+
+- **Source**: https://superhq.ai/
+- **Isolation**: Apple-Silicon desktop application using local microVMs via
+  Virtualization.framework/libkrun-era components.
+- **Agent integration**: Claude Code, Codex, and Pi in a GUI, with mobile
+  remote access.
+- **Credentials and review**: host-side auth gateway injects credentials on
+  the wire; a temporary overlay stages writes for diff-and-accept review.
+- **Closest overlap**: local microVM isolation, multi-provider launching, and
+  credential custody for security-minded individual developers.
+- **Difference**: GUI desktop product on Apple Silicon rather than a
+  cross-platform declarative CLI/fleet layer. The July 2026 snapshot in the
+  original survey recorded a user request for per-run tool-call and network
+  audit logging; treat that as point-in-time rather than a permanent gap.
+
+## Credential gateway without isolation
+
+### OneCLI
+
+[OneCLI](https://onecli.sh/) is a framework-agnostic identity gateway rather
+than a sandbox. Its phantom-token design gives the agent a placeholder and
+substitutes the encrypted real credential at the network layer. It therefore
+matches bot-bottle closely on secret custody, and is more portable because it
+can sit in front of agents launched by anything, but it supplies no container
+or VM boundary, filesystem isolation, role manifest, or egress-content DLP.
+
+The positioning consequence from the earlier survey still holds: secret
+custody alone is not unique. bot-bottle's relevant combination is local
+isolation + default-deny egress + payload DLP + declarative roles + credential
+custody. OneCLI's managed tier also places custody with a third party, whereas
+bot-bottle keeps it within operator-controlled infrastructure. See
+[`agent-credential-proxy-landscape.md`](agent-credential-proxy-landscape.md)
+for the detailed build-versus-adopt analysis.
+
+## Governance / pre-action authorization layers
+
+These two tools don't provide VM or filesystem isolation; they intercept
+tool calls before execution and evaluate them against a per-agent
+declarative policy. They are the closest competitors on **agent-tailored
+policy** and complement isolation sandboxes rather than substituting for
+them.
+
+### Microsoft Agent Governance Toolkit (AGT) *(added 2026-07-18)*
+- **Source**: https://github.com/microsoft/agent-governance-toolkit
+- **License**: MIT (~3.3k stars, open-sourced April 2, 2026).
+- **Isolation**: None (OS/VM). Execution rings (0–3, inspired by CPU
+  privilege levels) control what an agent can do at the framework layer.
+  MCP security gateway treats MCP traffic as an untrusted boundary.
+- **Locality**: Embedded in the agent framework (Python, TypeScript, .NET,
+  Rust, Go; 20+ framework adapters).
+- **Agent integration**: Framework-agnostic. Plugs into Semantic Kernel,
+  AutoGen, and others as a middleware layer.
+- **Config**: YAML policy per agent — tools can be `allowed`, `denied`,
+  `sandboxed`, or routed through an `approval` step. Every action passes
+  through a governance gate checking: agent DID, trust score, risk tier,
+  requested tool, action type, and policy rules.
+- **Network policy**: Not directly — operates at tool-call level.
+- **Credentials**: Per-agent DID (Ed25519 decentralized identifier); agent
+  does not borrow a human's credentials.
+- **Notable**: Dynamic trust score (0–1,000, behavioral decay) —
+  privilege follows observed behaviour, not just provisioning. Covers all
+  10 OWASP Agentic Top 10 risks. Kill switch + SLO monitoring. Sub-ms
+  policy enforcement.
+- **Maturity**: MIT, ~3.3k ⭐, v3.7.0 May 2026.
+
+### Open Agent Passport (OAP) *(added 2026-07-18)*
+- **Source**: https://github.com/aporthq/aport-spec ; spec at
+  https://api.aport.io/spec/spec/oap/oap-spec.md/ ; arXiv 2603.20953
+- **License**: Open specification.
+- **Isolation**: None. Pre-action hook only — intercepts tool calls
+  synchronously before execution, evaluates against a cloud-registry
+  declarative policy, fails closed.
+- **Locality**: Local hook + cloud policy registry.
+- **Agent integration**: Framework-agnostic; hook pattern.
+- **Config**: Declarative policy rules in a cloud registry (evaluated in
+  order; first failing rule denies). Ed25519-signed, hash-chained audit
+  records per decision.
+- **Network policy**: Not directly.
+- **Notable**: 53ms median authorization decision (N=1,000). In an
+  adversarial testbed ($5,000 bounty, 1,151 sessions), social engineering
+  succeeded 74.6% of the time under a permissive policy; under a
+  restrictive OAP policy, 0% success across 879 attempts. Assumes
+  framework runtime is not compromised.
+- **Maturity**: Specification + reference implementation, 2026.
+
 ## Comparison table

-| Axis | bot-bottle | endo-familiar | litterbox | agent-safehouse | matchlock | tilde.run | boxlite | microsandbox | smolmachines |
-|---|---|---|---|---|---|---|---|---|---|
-| Isolation | Docker + internal net + pipelock; gVisor if present | Object-capability (no OS isolation) | Podman + opt. Landlock | macOS `sandbox-exec` | MicroVM (Firecracker / Virt.fw) | Hosted container (unverified) | MicroVM (KVM / Hypervisor.fw) | MicroVM (libkrun) | MicroVM (libkrun / KVM) |
-| Local vs hosted | Local | Local | Local (Linux) | Local (macOS) | Local | Hosted SaaS | Local | Local | Local |
-| Open source | Apache 2.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 | MIT | No | Apache 2.0 | Apache 2.0 | Apache 2.0 |
-| Agent target | Claude Code | Generic (demo) | Generic | Multi-agent wrapper | Generic (+ Claude/OpenAI SDKs) | Claude focus | Generic | Claude + Cursor (MCP/Skills) | Generic (AGENTS.md) |
-| Network policy | Default-deny via pipelock + per-bottle allowlist + DLP | Capability model only | Limited | Not addressed | Default-deny + allowlist + secret-injecting proxy | Default-deny + logging | Per-VM net (unverified) | Not documented | Off by default + allowlist |
-| Parallel agents | Yes (one bottle per agent) | n/a | Not addressed | One at a time | Multiple VMs | Yes (dashboard) | SDK-level | SDK-level | Architectural |
-| Config | JSON manifest (bottles + agents) | Programmatic refs | CLI wizard | Profile files / shell fns | CLI / SDK | DSL + CLI + SDK | SDK | CLI / SDK / MCP | TOML Smolfile |
-| Maturity | Active May 2026 | Research (2022+) | Early (~66 ⭐) | Active (~1.4k ⭐) | Experimental (~574 ⭐) | Private preview | YC, ~4.7k ⭐ | YC, ~6k ⭐, beta | ~3.1k ⭐ |
+*Isolation/sandbox tools only. AGT and OAP are governance layers — see their per-project notes above.*
+
+| Axis | bot-bottle | endo-familiar | litterbox | agent-safehouse | matchlock | tilde.run | boxlite | microsandbox | smolmachines | E2B | Daytona | CubeSandbox | Cleanroom | container-use | Docker sbx | Anthropic srt |
+|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+| Isolation | MicroVM per bottle default (Firecracker/KVM on Linux, Apple Container on macOS) + own egress DLP scanner; Docker legacy fallback, gVisor there if present | Object-capability (no OS isolation) | Podman + opt. Landlock | macOS `sandbox-exec` | MicroVM (Firecracker / Virt.fw) | Hosted container (unverified) | MicroVM (KVM / Hypervisor.fw) | MicroVM (libkrun) | MicroVM (libkrun / KVM) | Firecracker microVM | Container or Linux/Windows VM class | MicroVM (RustVMM / KVM) | MicroVM (Firecracker / Virt.fw) | Docker container + git worktree | MicroVM (proprietary) | OS-level (Seatbelt / bubblewrap / WFP) — no container |
+| Local vs hosted | Local | Local | Local (Linux) | Local (macOS) | Local | Hosted SaaS | Local | Local | Local | Hosted; self-host/BYOC available | Hosted; dedicated/custom regions | Self-hosted (server/cluster) | Self-hosted server | Local | Local | Local |
+| Open source | Apache 2.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 | MIT | No | Apache 2.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 | Production closed-source; legacy AGPL repo unmaintained | Apache 2.0 | Apache 2.0 | Apache 2.0 | Proprietary | Apache 2.0 (experimental) |
+| Agent target | Claude Code, Codex, Pi, and provider plugins | Generic (demo) | Generic | Multi-agent wrapper | Generic (+ Claude/OpenAI SDKs) | Claude focus | Generic | Claude + Cursor (MCP/Skills) | Generic (AGENTS.md) | LLM-agnostic platform builders | LLM-agnostic platform builders | E2B-compatible (platform builders) | CI / generic process | Claude Code, Cursor, Windsurf (MCP) | Claude Code, Codex, Gemini CLI, Copilot, Kiro | Claude Code (and any process) |
+| Network policy | Default-deny via own egress scanner + per-bottle allowlist + content DLP + gitleaks on git push | Capability model only | Limited | Not addressed | Default-deny + allowlist + secret-injecting proxy | Default-deny + logging | Per-VM net (unverified) | Not documented | Off by default + allowlist | Per-sandbox allow/deny rules and custom egress proxy; internet configurable | Per-sandbox CIDR/domain allowlist or block-all; tier policy; secret-injecting proxy | Default-deny allowlist + instant egress block + audit logs + per-sandbox tokens (eBPF) + credential vault | Default-deny + per-repo host allowlist (cleanroom.yaml) | Not addressed | Default-deny; Open / Balanced / Locked Down presets; live TUI network panel | Proxy-based allowlist/denylist (HTTP + SOCKS5); custom proxy supported |
+| Parallel agents | Yes (one bottle per agent) | n/a | Not addressed | One at a time | Multiple VMs | Yes (dashboard) | SDK-level | SDK-level | Architectural | Yes (platform service) | Yes (platform service) | Yes (2,000+/host claimed) | Yes (server model) | Yes (per-agent containers + worktrees) | Yes | Yes |
+| Long-running posture | Persistent by default (named, supervised) | n/a (demo) | Session (up while in use) | Per-invocation | Ephemeral VM per run | Per-run (versioned) | Ephemeral + snapshot/fork | Ephemeral / on-demand | Named persistent by default | Runtime tier limits + indefinite pause/resume | Persistent filesystem; VM pause/resume; configurable auto-stop | Ephemeral + auto pause/resume | Per-run + suspend/resume | Per-agent container (ephemeral) | Per-session; branch mode creates git worktree in .sbx/ | Per-invocation |
+| DX: run Claude yolo-style | One command → interactive yolo Claude (`start <agent>`, `--dangerously-skip-permissions` default) | n/a (lib demo) | Wizard + build, then run claude inside (Linux only) | One-command wrapper (`safehouse claude --dangerously-skip-permissions`) | CLI: run a cmd in a VM (not a Claude wrapper) | Hosted (`tilde exec`), not local-native | SDK code required (build the run yourself) | CLI/MCP: sandbox-as-a-tool for the agent, not a wrapper around it | SSH into a named machine, run claude there | SDK/CLI sandbox; wire the agent yourself | SDK/CLI sandbox; wire the agent yourself | Stand up a cluster + drive via E2B SDK | CI-oriented, not a Claude wrapper | MCP server: `claude mcp add container-use -- container-use stdio` | One command: `sbx` wraps claude with `--dangerously-skip-permissions` default | Library/wrapper, not a standalone CLI |
+| Config | YAML-in-Markdown manifests (bottles + agents) | Programmatic refs | CLI wizard | Profile files / shell fns | CLI / SDK | DSL + CLI + SDK | SDK | CLI / SDK / MCP | TOML Smolfile | SDK/API + templates | SDK/API/CLI + images/snapshots | E2B-compatible SDK | cleanroom.yaml in repo | None (no policy config) | Preset levels at launch | Programmatic per-invocation (allow/deny lists) |
+| Agent-tailored policy | Yes — bottle/agent split; declarative per-role egress + credentials; composable via `extends:` | Partial — capability model scopes per-agent, but no declarative role manifest | No | Partial — per-agent profile files (Seatbelt); no egress | No | Yes — per-agent DSL RBAC (allow/deny/approve per action/repo/agent) | No | No | No | No — per-sandbox SDK config | No — per-sandbox SDK config | No — per-sandbox SDK config, not role-scoped | Partial — per-repo cleanroom.yaml, not per-role | No | No — network presets only | No |
+| Maturity | Active July 2026 | Research (2022+) | Early (~66 ⭐) | Active (~1.8k ⭐) | Experimental (~574 ⭐) | Private preview | YC, ~4.7k ⭐ | YC, ~6k ⭐, beta | ~3.1k ⭐ | Established hosted platform, ~12.4k ⭐ | Production commercial; closed-source since June 2026 | Tencent, prod, ~10.4k ⭐ | Active (Buildkite product) | Early development | GA 2026 | Early research preview |

 ## What's closest, what's different

 **Closest in design and scope.** agent-safehouse and litterbox sit
 nearest bot-bottle: local, single-user, thin wrappers over an
 existing OS primitive, low-dep. The split is the isolation primitive —
-bot-bottle uses Docker + pipelock egress (plus gVisor where
-available); agent-safehouse uses `sandbox-exec`; litterbox uses Podman +
-Landlock. matchlock and smolmachines are spiritually close on the
-*policy* side (default-deny net, per-host allowlist) but use microVMs
-instead of containers.
+bot-bottle now defaults to a VM per bottle (Firecracker microVM on KVM
+Linux, Apple Container on macOS) with its own DLP-scanning egress proxy,
+keeping Docker only as a legacy fallback; agent-safehouse uses
+`sandbox-exec`; litterbox uses Podman + Landlock. matchlock and
+smolmachines are close on *both* the policy side (default-deny net,
+per-host allowlist) and — now that bot-bottle has moved off
+containers-by-default — the microVM isolation primitive. Note: Apple
+Container 1.0 stable shipped June 9 2026 (frozen CLI and APIs), which
+makes the macOS backend stable surface area rather than a moving target.
+
+**New closest on agent-tailored policy.** Two governance tools are the
+direct competitors on the "coarse-grained sandbox" axis. **tilde.run**
+has had per-agent DSL RBAC since its launch (though it's hosted SaaS).
+**Microsoft AGT** is the most serious new entrant: per-agent DID
+identity, YAML policy that can allow/deny/sandbox/approve individual tool
+calls per agent, and a dynamic behavioural trust score. It operates at
+the framework tool-call layer, not the network layer — so it's
+complementary to bot-bottle's network/filesystem isolation rather than a
+direct substitute, but on the "does this sandbox know what this agent is
+for?" question it is the most complete answer in the field. OAP's
+pre-action hook pattern achieves similar goals with cryptographic audit
+and a 0% adversarial-attack success rate under a restrictive policy.
+
+**New closest on DX.** **Docker sbx** is the first tool in this set that
+matches bot-bottle on the "one command, dangerously-skip-permissions safe
+by default" DX bar, at microVM isolation strength, with host-side
+credential injection. It is proprietary, preset-based (not role-
+declarative), and cloud-agent-specific, but it directly competes on the
+UX proposition. agent-safehouse was the previous DX peer; Docker sbx
+materially raises the bar.
+
+**New closest on repo-scoped policy.** **Cleanroom** (Buildkite) is the
+first tool to combine microVM isolation with a declarative egress policy
+file — though the policy lives in the repo being sandboxed
+(`cleanroom.yaml`), not in an agent-role manifest. That makes it per-
+repo rather than per-role: the same Cleanroom config applies to any
+agent running in that repo. The distinction matters for bot-bottle's
+use case (one developer running multiple agent *roles* with different
+egress footprints), but for CI/CD use cases Cleanroom is a direct
+alternative.

 **Solving a different problem.** tilde.run is hosted SaaS for team /
 production agent pipelines with data-versioned rollback — explicitly
-opposite to bot-bottle's "infrastructure I control" goal. boxlite and
-microsandbox are infrastructure libraries aimed at platform builders
-embedding sandboxes into agent frameworks; they would be a *backend*
-bot-bottle could call, not a competitor to its manifest layer.
-endo-familiar is in a different paradigm entirely: capability passing
-rather than kernel boundaries.
+opposite to bot-bottle's "infrastructure I control" goal. E2B and Daytona
+are hosted sandbox platforms, while boxlite, microsandbox, and CubeSandbox
+are infrastructure libraries/services aimed at platform builders embedding
+sandboxes into agent frameworks; they
+would be a *backend* bot-bottle could call, not a competitor to its
+manifest layer. endo-familiar is in a different paradigm entirely:
+capability passing rather than kernel boundaries.

 ## Borrowable ideas

-What bot-bottle already has that the survey suggested as
-differentiators:
- Default-deny egress with a per-agent allowlist (pipelock).
+### Already shipped or otherwise addressed
+
+- Default-deny egress with a per-agent allowlist (own egress scanner).
 - DLP scanning of outbound traffic.
 - Bottle / agent split (manifest layer above the isolation primitive).
 - gVisor auto-detection on Linux.
+- **In-flight secret injection** (suggested by matchlock) — **shipped.**
+  Real provider and git-host tokens are held outside the agent and injected
+  by the egress gateway on matching routes. The agent receives only proxy
+  URLs and, where a client requires a credential-shaped value, a placeholder;
+  `GITEA_TOKEN` and equivalent real tokens do not appear in the agent's
+  environment.
+- **MicroVM backend** — **shipped.** MicroVMs are now the default:
+  Firecracker on KVM Linux and Apple Container on macOS. Docker is the legacy
+  fallback.
+- **Per-use SSH key confirmation** (suggested by litterbox) — **addressed by
+  stronger credential custody instead.** The agent does not hold the upstream
+  git SSH key or an SSH-agent socket: git-gate holds the credential and gates
+  git operations. A confirmation wrapper inside the agent would therefore
+  protect a credential that is no longer there. Operator approval at the gate
+  remains the appropriate control point for any future per-use confirmation.

-Ideas worth considering, without abandoning the Python-stdlib-first / local-Docker
-stance:
+### Still worth considering

-1. **Per-use SSH key confirmation** (from litterbox). Even with
-   KnownHostKey pinning and pipelock egress, a wrapper SSH agent that
-   prompts on each key use (e.g. via `osascript` / `notify-send`) would
-   catch an agent doing something off-policy with a key it legitimately
-   holds. Pure-stdlib, no new deps.
-2. **In-flight secret injection** (from matchlock). Pipelock already
-   does egress allowlisting and DLP; teaching it to *inject* tokens at
-   proxy time so e.g. `GITEA_TOKEN` never appears in the container's
-   env would close the "agent reads its own env and exfiltrates" path.
-   Fits the existing pipelock architecture.
-3. **MicroVM backend as an opt-in bottle type** — already on the radar
-   in `stronger-isolation-alternatives.md`. microsandbox, smolmachines,
-   and matchlock all show that libkrun + Apple's
-   Virtualization.framework is ergonomic enough that a
-   `"runtime": "microvm"` field on a bottle is plausible without a heavy
-   stack.
+- **Live network activity in the supervisor TUI** (from Docker sbx): show
+  allowed and blocked connections and let the operator propose policy changes
+  from the existing supervision surface.
+- **Tamper-evident audit records** (from OAP): sign and hash-chain egress and
+  supervision decisions for compliance-sensitive deployments.
+- **Behaviour-informed policy downgrade** (from Microsoft AGT): use repeated
+  DLP alerts or supervision holds as a signal to narrow policy or request
+  closer review. This needs a carefully specified trust model before it can be
+  more than a heuristic.

 Not worth borrowing: the SDK-first programmatic API style of boxlite /
 microsandbox (cuts against the declarative-manifest stance), and the
 hosted-SaaS dashboard model of tilde.run (cuts against the
 "infrastructure I control" goal).

+## Publishing and positioning verdict
+
+Publishing remains worthwhile, but the defensible claim is the combination,
+not any single primitive. Credential custody is matched by OneCLI, matchlock,
+Daytona, Docker sbx, and CubeSandbox; local one-command isolation is matched by
+agent-safehouse and Docker sbx; hosted microVM execution is a crowded platform
+category.
+
+bot-bottle remains unusual in combining:
+
+- local, operator-controlled execution with persistent named bottles;
+- one declarative role layer across Claude Code, Codex, Pi, and provider
+  plugins;
+- composable agent/bottle manifests, skills, and system prompts;
+- Firecracker/Apple Container isolation with a Docker fallback;
+- default-deny per-role egress, payload DLP, and git-push secret scanning;
+- credentials injected outside the agent process; and
+- supervision and audit state suited to long-running parallel agents.
+
+The practical wedge is “as easy as native yolo, with declarative role policy
+and self-hosted custody,” including scoped access to private LAN/Tailnet
+services that cloud-first runtimes cannot provide without additional network
+plumbing. The main competitive risks are a local wrapper such as claudebox or
+Docker sbx growing a role-manifest layer, and GUI products such as SuperHQ
+adding equivalent policy and audit depth.
+
 ## Caveats

 - Star counts and last-commit dates are point-in-time snapshots.
@@ -230,3 +691,180 @@ hosted-SaaS dashboard model of tilde.run (cuts against the
 - The `superradcompany/microsandbox` URL in the original prompt
  redirects to `microsandbox/microsandbox`; the surveyed project is the
  same.
+- CubeSandbox performance/scale numbers (<60ms cold start, <5MB/instance,
+  2,000+ sandboxes per 96-vCPU host) are the project's own launch claims,
+  not independently verified here.
+
+## Addendum 2026-07-18 — CubeSandbox and the positioning read
+
+CubeSandbox (Tencent Cloud, Apache 2.0, ~10.4k stars, HN launch
+[#47863430](https://news.ycombinator.com/item?id=47863430)) is the first
+open-source, self-hostable project in this survey to combine, in one stack,
+the main primitives
+bot-bottle treated as its differentiator:
+
+- **Egress custody (connection level)** — default-deny domain allowlist
+  (L7 domain/SNI filtering), instant block on unauthorized egress,
+  per-sandbox traffic tokens, full audit logs of destinations (eBPF
+  virtual switch, "CubeVS"). This matches bot-bottle's egress scanner at
+  the *connection level*, productized — see the one thing it does **not**
+  match, below.
+- **Credential custody** — a vault where keys "never enter the sandbox,
+  model context, or logs." This is the in-flight-injection idea from
+  matchlock, but as a first-class feature, and it's exactly the
+  cross-vendor "egress audit + custody" wedge the monetization
+  positioning treats as the one defensible moat.
+- **Isolation on par with bot-bottle's current default** — a dedicated
+  guest kernel per sandbox (RustVMM/KVM). bot-bottle now defaults to the
+  same class of boundary (Firecracker microVM / Apple Container), so this
+  is parity, not an edge; CubeSandbox's remaining edge is running that
+  per-kernel isolation multi-tenant at scale on one host.
+
+The one axis CubeSandbox does **not** cover — and where bot-bottle stays
+distinctive:
+
+- **Content DLP on *authorized* channels.** CubeSandbox's egress control
+  is connection-level: it decides *whether* a destination is allowed and
+  logs it, and its vault keeps *injected* credentials out of the sandbox
+  entirely. Neither inspects the *payload* of traffic to an allowed
+  destination. So an agent that exfiltrates over a permitted channel —
+  pasting a repo's contents, an agent-derived secret, or PHI into an
+  allowed API/domain — is not caught by CubeSandbox. bot-bottle's own
+  egress DLP scanner does scan that: response + websocket content against
+  the resolved per-flow config, with per-bottle token redaction (see
+  recent egress commits). The vault
+  approach is arguably *stronger* for the specific case of pre-known
+  injected credentials (they can't leak if they were never present), but
+  it is not a substitute for content inspection of everything else.
+
+**Long-running posture — a sharper axis than raw isolation.** E2B and
+CubeSandbox are *ephemeral-per-task* by design; a long-running agent is an
+architected pattern on top, not the default. E2B: 5-minute default
+timeout, continuous runtime tier-capped (~1h Hobby / ~24h Pro), duration
+achieved via **pause/resume** (preserves filesystem + memory + processes;
+reconnect by sandbox ID via `Sandbox.connect()`; resume resets the timeout
+to 5 min; auto-pause via `on_timeout: "pause"`). CubeSandbox mirrors this
+(E2B drop-in) with first-class auto pause/resume and hundred-ms
+checkpoint/fork — and, self-hosted, sets its own timeout policy with no
+vendor tier caps. bot-bottle inverts the model: a bottle is **persistent,
+named, and supervised by default** — long-running *is* the default, not a
+session-management loop over pause/resume. smolmachines is the other
+persistent-by-default project in this set. For anyone building agents that
+run for hours/days, this posture difference matters more than the
+isolation primitive.
+
+**DX — the "run Claude yolo-style" bar.** The reason `claude
+--dangerously-skip-permissions` is so widely used is DX: it's one command
+and the agent just goes. The bottle thesis is to make a *sandboxed* run
+that easy — `start <agent>` builds the image on first run and drops you
+into an interactive Claude session that already has
+`--dangerously-skip-permissions` on by default
+(`contrib/claude/agent_provider.py`), with the sandbox as the guardrail
+instead of per-action prompts. On this axis the field splits cleanly:
+- **Wrappers around the agent** (as-easy-as-native): bot-bottle and
+  **agent-safehouse** (`safehouse claude --dangerously-skip-permissions`).
+  These *are* the run-Claude experience. agent-safehouse is the real DX
+  peer — but it's macOS-only Seatbelt, single-run, and doesn't address
+  network egress; bot-bottle adds VM-grade isolation, egress DLP, and
+  persistent/parallel bottles across macOS + Linux.
+- **Libraries / services** (you build the run yourself): boxlite,
+  microsandbox, CubeSandbox, E2B, Daytona. These hand you an SDK or a cluster and
+  expect you to wire the agent in — powerful for platform builders,
+  heavyweight for "just run Claude on my laptop." microsandbox's MCP/Skills
+  angle is *sandbox-as-a-tool the agent calls*, which is the inverse of
+  wrapping the agent.
+- **In between:** litterbox (wizard + build, Linux only), smolmachines
+  (SSH into a named machine), matchlock (run a command in a VM).
+
+So DX is a genuine bot-bottle differentiator. agent-safehouse matches the
+one-command wrapper with weaker isolation and no egress story; Docker sbx now
+matches it at microVM strength but remains proprietary and preset-based. "As
+easy as native yolo, with declarative role policy" is the narrower defensible
+one-liner.
+
+Why it still doesn't collide head-on:
+
+1. **Shape.** CubeSandbox is a *multi-tenant service for platform
+   builders* (drop-in E2B replacement, SDK-driven, 2,000 sandboxes on a
+   box). bot-bottle is a *single-operator, declarative-manifest tool for
+   the infrastructure I run*. Different buyer, different ergonomics — no
+   declarative role manifest, no bottle/agent split, no "one command on my
+   laptop."
+2. **Backend, not competitor.** Like boxlite/microsandbox, CubeSandbox is
+   something bot-bottle could sit *on top of* — a `"runtime": "microvm"`
+   or `"runtime": "cubesandbox"` backend under the manifest layer — while
+   keeping the manifest, the bottle/agent split, and the local,
+   single-operator default.
+
+Why it matters anyway:
+
+- The "nobody else bundles connection-level egress allowlist + audit +
+  in-flight credential custody" line is **no longer true for the
+  primitive** — CubeSandbox ships the open-source/self-hosted combination,
+  and Daytona ships a proprietary firewall + credential-substitution variant.
+  But **content DLP on authorized channels is still not matched** (see
+  above), and neither is the *layer above* the primitive (declarative
+  manifest, cross-vendor orchestration, operator UX, the
+  phone-control/dashboard north star). Those two — outbound-payload DLP
+  and the orchestration layer — are where the defensible ground now sits;
+  the connection-level allowlist + vault mechanism, on its own, is no
+  longer differentiating. Revisit the monetization open/paid line with
+  that in mind.
+- Worth a closer look at **how** CubeSandbox does credential injection
+  and per-sandbox egress tokens (eBPF virtual switch vs. bot-bottle's
+  mitmproxy egress proxy) when hardening bot-bottle's now-shipped
+  credential-custody implementation.
+
+## Addendum 2026-07-18 (second pass) — agent-tailored policy landscape
+
+The second-pass question was: how novel is bot-bottle's per-agent,
+role-tailored sandbox relative to the expanded field?
+
+**The short answer:** on the isolation + network + role-tailoring
+combination, bot-bottle remains the only tool in this set. On
+role-tailored *policy at the tool-call level*, Microsoft AGT and OAP are
+the most complete answers, but they don't provide isolation; they
+complement rather than substitute.
+
+**The competitive picture by axis:**
+
+- *Agent-tailored egress (declarative, per-role)* — bot-bottle and
+  tilde.run. Cleanroom is per-repo, not per-role. Everyone else is
+  per-session or not addressed.
+- *Agent-tailored tool-call policy (declarative, per-agent identity)* —
+  Microsoft AGT (YAML policy + DID identity + trust score), OAP
+  (declarative policy rules + cryptographic audit). Neither provides
+  network/filesystem isolation.
+- *Composable policy (role overlays)* — bot-bottle (`extends:`). No
+  other tool surveyed supports composable role-policy inheritance.
+- *Isolation + DX (one-command safe yolo)* — bot-bottle and Docker sbx.
+  Docker sbx is proprietary, preset-based, and cloud-agent-specific;
+  it's the first DX-class competitor at microVM isolation strength.
+
+**What the HN "coarse-grained" complaint maps to:** The complaint is
+that a VM isolates the filesystem but doesn't know if the agent
+*should* be sending an email. bot-bottle's bottle/agent split is a
+structural answer to this: the bottle manifest declares exactly what
+the role can reach, and the sandbox enforces it at the network layer.
+Microsoft AGT is the most complete answer at the semantic/tool-call
+layer. The gap both leave open is *intent classification* — knowing
+whether a permitted action is consistent with the agent's actual task.
+See `hn-agent-safety-discourse-july-2026.md` for the blast-radius
+analysis.
+
+**Open ideas from new tools (also summarized above):**
+
+- **Microsoft AGT's trust-score decay** — privilege that reflects
+  observed behaviour rather than static provisioning. Applied to
+  bot-bottle: a bottle that has triggered DLP alerts or supervise holds
+  could auto-downgrade its network preset, or flag the session for
+  closer review. Fits the existing supervise-server architecture.
+- **Docker sbx's live network TUI** — real-time per-session view of
+  allowed and blocked outbound connections with point-and-click
+  allow/block. `cli.py supervise` is the right surface; adding a
+  live-connections panel would directly address the "I can't see what
+  the agent is doing" gap without any backend changes.
+- **OAP's cryptographic audit chain** — Ed25519-signed, hash-chained
+  audit records. Currently bot-bottle logs egress decisions but doesn't
+  chain them. A tamper-evident audit record per session would be useful
+  for the compliance use case the CubeSandbox positioning targets.
--- a/Show More
+++ b/Show More