fix(release): publish only verified wheel

`container image tag` only accepts image-name[:tag] as its source and
rejects a bare 64-hex image ID ("cannot specify 64 byte hex string as
reference"), so pinning the agent image for the nested-containers layer
died before the build could start. Tag from the ref instead; the
existing post-tag inspection is what keeps the handoff fail-closed if
the ref moves between the two commands.

The unit test mocked subprocess and asserted the ID-as-source call
shape, so it never saw the CLI's rejection. It now pins the ref as the
source, and covers the bare-hex ID that `container image inspect`
actually reports plus the mid-flight tag move.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.gitea/workflows/package-release.yml

@@ -0,0 +1,68 @@
+# Package bot-bottle only after immutable OCI and Firecracker artifacts have
+# been published. The inputs are the release boundary: the resulting wheel
+# carries these exact identities and production startup never chooses another.
+
+name: package-release
+
+on:
+  workflow_dispatch:
+    inputs:
+      orchestrator_image:
+        description: Digest-pinned orchestrator OCI reference
+        required: true
+      gateway_image:
+        description: Digest-pinned gateway OCI reference
+        required: true
+      firecracker_orchestrator_version:
+        description: Published orchestrator rootfs package version
+        required: true
+      firecracker_orchestrator_sha256:
+        description: SHA-256 of orchestrator rootfs.ext4.gz
+        required: true
+      firecracker_gateway_version:
+        description: Published gateway rootfs package version
+        required: true
+      firecracker_gateway_sha256:
+        description: SHA-256 of gateway rootfs.ext4.gz
+        required: true
+
+jobs:
+  package:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout the release commit
+        uses: actions/checkout@v4
+
+      - name: Install package build tooling
+        run: python3 -m pip install --break-system-packages build
+
+      - name: Generate immutable release manifest
+        run: |
+          python3 scripts/generate_release_manifest.py \
+            --source-commit "$GITHUB_SHA" \
+            --orchestrator-image '${{ inputs.orchestrator_image }}' \
+            --gateway-image '${{ inputs.gateway_image }}' \
+            --firecracker-orchestrator-version '${{ inputs.firecracker_orchestrator_version }}' \
+            --firecracker-orchestrator-sha256 '${{ inputs.firecracker_orchestrator_sha256 }}' \
+            --firecracker-gateway-version '${{ inputs.firecracker_gateway_version }}' \
+            --firecracker-gateway-sha256 '${{ inputs.firecracker_gateway_sha256 }}' \
+            --output release-manifest.json
+
+      - name: Build package with assigned infrastructure pins
+        env:
+          BOT_BOTTLE_RELEASE_MANIFEST: ${{ github.workspace }}/release-manifest.json
+        run: python3 -m build --wheel
+
+      - name: Verify packaged manifest
+        run: |
+          test "$(find dist -maxdepth 1 -name '*.whl' -type f | wc -l)" -eq 1
+          python3 -m venv verify-venv
+          verify-venv/bin/pip install --no-deps dist/*.whl
+          verify-venv/bin/python -c \
+            'from bot_bottle.release_manifest import load_manifest; print(load_manifest())'
+
+      - name: Upload verified release wheel
+        uses: actions/upload-artifact@v3
+        with:
+          name: bot-bottle-release
+          path: dist/*.whl

Dockerfile.orchestrator

@@ -9,12 +9,9 @@
 #     Keeping the content in one place means future orchestrator deps (e.g.
 #     iroh) are added here once, not duplicated per backend.
 #
-# 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.orchestrator.fc). Keeping the
-# secret-dense control plane on a minimal dependency surface is the point
-# (PRD 0070's "secret concentration").
+# It stays deliberately lean: only the pinned FastAPI/Uvicorn control-plane
+# stack is installed here — none of the gateway's mitmproxy/git/gitleaks
+# (that's Dockerfile.gateway) and no buildah (that's firecracker-only).
 #
 # Shares an exact multi-architecture Python/trixie manifest with the gateway
 # image. The version-qualified tag keeps the human-readable upstream version;
@@ -25,6 +22,11 @@ FROM ${PYTHON_BASE_IMAGE}
 
 WORKDIR /app
 
+COPY requirements.orchestrator.lock /tmp/requirements.orchestrator.lock
+RUN pip install --no-cache-dir --require-hashes \
+        -r /tmp/requirements.orchestrator.lock \
+    && rm /tmp/requirements.orchestrator.lock
+
 # 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.

README.md

@@ -15,7 +15,7 @@
 ## Features
 
 - **Per-bottle egress allowlist** — TLS-bumped HTTP/HTTPS chokepoint with a per-manifest host allowlist; per-route path/method/header `matches` filtering; outbound DLP scanning for known tokens and secrets, inbound DLP scanning for prompt-injection attempts; DoH and arbitrary hosts blocked by default.
-- **Per-route token-match policy** — each egress route picks what happens when the outbound DLP catches a token via `dlp.outbound_on_match`: `supervise` (default) holds the request and surfaces it in `./cli.py supervise` for approval (an approved value is remembered for the life of the proxy); `redact` scrubs the value and forwards; `block` is a hard `403`. Cuts false-positive friction without weakening default-deny.
+- **Per-route token-match policy** — each egress route picks what happens when the outbound DLP catches a token via `dlp.outbound_on_match`: `supervise` (default) holds the request and surfaces it in `bot-bottle supervise` for approval (an approved value is remembered for the life of the proxy); `redact` scrubs the value and forwards; `block` is a hard `403`. Cuts false-positive friction without weakening default-deny.
 - **Tokens the agent never sees** — host secrets live in a gateway; the agent dials `http://gateway:9099/<path>` and the proxy strips inbound `Authorization` and injects the real token before forwarding. `printenv` in the agent shows proxy URLs only.
 - **Gitleaks-scanned push (git-gate)** — `bottle.git` remotes route through a per-bottle `git daemon` that gitleaks-scans incoming refs pre-receive and forwards clean refs upstream over SSH. The agent never holds the upstream credential.
 - **Manifest-scoped skills + secrets** — each bottle declares its skills, env, git identity, remotes, and egress routes; unknown keys die at load.
@@ -39,7 +39,7 @@ On the legacy Docker backend, the same logical bottle is two containers per agen
 The Docker topology looks like this:
 
 ```
-                            host  ( ./cli.py )
+                            host  ( bot-bottle )
                                   │
                           starts  │  stops
                                   ▼
@@ -71,9 +71,23 @@ When the agent exits, `cli.py` tears down every gateway and both networks; nothi
 
 ## Quickstart
 
-On compatible macOS hosts, the default backend requires Apple's `container` CLI and does not require Docker. The Firecracker backend (Linux) requires Docker on the host for the gateway plus the `firecracker` binary and KVM. The legacy Docker backend requires Docker. Claude bottles also need a long-lived Claude Code OAuth token (`claude setup-token`) exported as `BOT_BOTTLE_CLAUDE_OAUTH_TOKEN`.
+```sh
+curl -fsSL https://gitea.dideric.is/didericis/bot-bottle/raw/branch/main/install.sh | sh
+```
 
-Use `BOT_BOTTLE_BACKEND=docker ./cli.py start <agent>` on hosts where neither Apple Container nor KVM is available and Docker is the desired backend.
+The installer is a bootstrapper: it finds a suitable Python, installs bot-bottle with `pipx` (falling back to `pip --user`), creates `~/.bot-bottle`, and runs `bot-bottle doctor`. It is idempotent and never uses `sudo`. Python-native users can skip it entirely with `pipx install bot-bottle` or `uv tool install bot-bottle`.
+
+### Requirements
+
+**Python ≥ 3.11**, and this is the one that trips people up on macOS: the `python3` Apple ships at `/usr/bin/python3` is **3.9.6**, which is too old. Bare `python3` resolves to that stub far more often than people expect. `path_helper` builds a login shell's `PATH` from `/etc/paths` and then appends `/etc/paths.d/*`, and `/usr/bin` sits in the former — so even when `/opt/homebrew/bin` *is* on the `PATH` (via `/etc/paths.d/homebrew`), it comes after `/usr/bin` and loses. Prepending a newer Python is something your shell profile does, and a fresh account, a launchd job, or a CI runner has no such profile. So the installer looks past bare `python3` before giving up: it tries `python3`, then the versioned `python3.11`–`python3.14` names, then `/opt/homebrew/bin`, `/usr/local/bin`, `~/.local/bin`, and python.org framework builds — and tells you which one it picked when it isn't the obvious one. Point it somewhere specific with `BOT_BOTTLE_PYTHON=/path/to/python3`.
+
+**No `pipx` required.** If `pipx` is present the installer uses it and stays out of the way. If it isn't, bot-bottle installs into a private venv at `~/.bot-bottle/venv` (override with `BOT_BOTTLE_VENV`) and symlinks the entry point into `~/.local/bin`. There is deliberately no `pip install --user` path: Homebrew, python.org and Debian/Ubuntu interpreters are all externally managed (PEP 668), which blocks `--user` outright — so on a Mac it is never the fallback it appears to be. A venv is exempt from PEP 668, and `venv` is stdlib, so unlike `pipx` there is nothing to bootstrap first.
+
+**`git`**, because the default install spec is a `git+` URL. Set `BOT_BOTTLE_INSTALL_SPEC` to a wheel path or index name to avoid it.
+
+**A backend**, which the installer deliberately does *not* install for you — `doctor` reports what's missing afterwards. On compatible macOS hosts, the default backend requires Apple's `container` CLI and does not require Docker. The Firecracker backend (Linux) requires Docker on the host for the gateway plus the `firecracker` binary and KVM. The legacy Docker backend requires Docker. Claude bottles also need a long-lived Claude Code OAuth token (`claude setup-token`) exported as `BOT_BOTTLE_CLAUDE_OAUTH_TOKEN`.
+
+Use `BOT_BOTTLE_BACKEND=docker bot-bottle start <agent>` on hosts where neither Apple Container nor KVM is available and Docker is the desired backend.
 
 > **CI (macOS Apple Container):** the advisory `integration-macos` job in `.gitea/workflows/pre-release-test.yml` runs only on manual dispatch. It targets a self-hosted host-mode runner labelled `macos`; Apple Container cannot run inside the Linux pull-request runner. Provision an Apple Silicon host with the `container` CLI running and Python ≥ 3.11 plus `coverage` on the launchd service's explicit `PATH`. The infra container is a singleton (`bot-bottle-mac-infra`), so keep runner concurrency at 1. Its coverage is reported separately and never feeds the required pull-request gate.
 
@@ -166,10 +180,10 @@ On Linux, a KVM-capable host defaults to the Firecracker backend. It needs:
 - **`/dev/kvm`** present and accessible. Load `kvm-intel` or `kvm-amd` (and enable virtualization in BIOS/firmware). The invoking user must be in the `kvm` group: `sudo usermod -aG kvm "$USER"` then re-login. bot-bottle preflights this and reports exactly what's missing.
 - **`firecracker`** on `PATH`: grab a release from <https://github.com/firecracker-microvm/firecracker/releases>. Start flows print this pointer when the binary is missing.
 - **Docker** for the gateway and image build.
-- **A one-time privileged network setup** — the per-bottle TAP pool plus the fail-closed `nftables` isolation table. Run `./cli.py backend setup --backend=firecracker` for the host-appropriate config (a NixOS module, a `sudo` script elsewhere); `./cli.py backend status --backend=firecracker` reports what's present, including whether the pool range collides with an existing route. The pool defaults to `10.243.0.0/16` (an obscure RFC-1918 block that dodges docker/libvirt/LAN and, deliberately, Tailscale's `100.64.0.0/10` CGNAT range); override with `BOT_BOTTLE_FC_IP_BASE` if it clashes on your host.
+- **A one-time privileged network setup** — the per-bottle TAP pool plus the fail-closed `nftables` isolation table. Run `bot-bottle backend setup --backend=firecracker` for the host-appropriate config (a NixOS module, a `sudo` script elsewhere); `bot-bottle backend status --backend=firecracker` reports what's present, including whether the pool range collides with an existing route. The pool defaults to `10.243.0.0/16` (an obscure RFC-1918 block that dodges docker/libvirt/LAN and, deliberately, Tailscale's `100.64.0.0/10` CGNAT range); override with `BOT_BOTTLE_FC_IP_BASE` if it clashes on your host.
 
 ```sh
-BOT_BOTTLE_BACKEND=firecracker ./cli.py start <agent>
+BOT_BOTTLE_BACKEND=firecracker bot-bottle 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`.
@@ -177,9 +191,16 @@ BOT_BOTTLE_BACKEND=firecracker ./cli.py start <agent>
 > **CI:** Firecracker integration runs in the manually dispatched `.gitea/workflows/pre-release-test.yml` on a self-hosted runner labelled `kvm`; privileged KVM hosts never execute unreviewed PR code automatically. Provision it like a normal Firecracker host: `firecracker` on `PATH`, `/dev/kvm`, the cached guest kernel and static dropbear, and the persistent TAP/nft pool. The required pull-request workflow runs unit plus the complete Docker integration suite on `ubuntu-latest`; see `docs/ci.md`.
 
 ```sh
-./cli.py start <agent>   # builds the image on first run, drops you into claude
+bot-bottle start <agent>   # prepares the selected agent image, then attaches
 ```
 
+Packaged releases carry immutable orchestrator and gateway identities. Docker
+and Apple Container pull the package-selected OCI digests; Firecracker pulls
+the matching checksum-verified rootfs artifacts. A source checkout uses local
+infrastructure builds for development. Set `BOT_BOTTLE_INFRA_BUILD=local` to
+make that development override explicit when diagnosing a packaged release;
+production never falls back to a build after an artifact pull fails.
+
 ## Manifest
 
 Bottles and agents are Markdown files with YAML frontmatter under `~/.bot-bottle/`. The Markdown body is the system prompt. Bottles live in `~/.bot-bottle/bottles/`; agents may also be shipped by a repo at `<repo>/.bot-bottle/agents/<name>.md`.
@@ -253,7 +274,7 @@ You help maintain Gitea-hosted projects.
 | `dlp.outbound_on_match` | no | What to do when an outbound token is detected: `supervise` (default for manifest routes — hold for operator approval), `redact` (scrub the value and forward), or `block` (hard 403). Agent-provider routes (e.g. `api.anthropic.com`) default to `redact`. |
 | `git.fetch` | no | `true` permits smart HTTP clone/fetch (`git-upload-pack`) for this host. Push (`git-receive-pack`) remains blocked. |
 
-When an outbound DLP detector matches a token, the route's `dlp.outbound_on_match` policy decides what happens. Under the default `supervise`, the proxy queues an `egress-token-allow` proposal for the operator's `./cli.py supervise` TUI and holds the request open until it is answered (or `EGRESS_TOKEN_ALLOW_TIMEOUT_SECONDS`, default 300s, elapses — after which it fails closed). The operator never sees the raw token, only the host, method, path, and a redacted snippet; approving adds the value to an in-memory safelist for the life of the egress proxy. Under `redact`, the matched value is scrubbed from the body, headers, and path and the request is forwarded (failing closed if a match lands somewhere unredactable, like the hostname). Under `block` it stays a hard `403`. Structural blocks (CRLF injection) and not-in-allowlist host blocks are always hard `403`s regardless of policy.
+When an outbound DLP detector matches a token, the route's `dlp.outbound_on_match` policy decides what happens. Under the default `supervise`, the proxy queues an `egress-token-allow` proposal for the operator's `bot-bottle supervise` TUI and holds the request open until it is answered (or `EGRESS_TOKEN_ALLOW_TIMEOUT_SECONDS`, default 300s, elapses — after which it fails closed). The operator never sees the raw token, only the host, method, path, and a redacted snippet; approving adds the value to an in-memory safelist for the life of the egress proxy. Under `redact`, the matched value is scrubbed from the body, headers, and path and the request is forwarded (failing closed if a match lands somewhere unredactable, like the hostname). Under `block` it stays a hard `403`. Structural blocks (CRLF injection) and not-in-allowlist host blocks are always hard `403`s regardless of policy.
 
 More examples in `examples/`. Full design lives under `docs/prds/`; the trust-boundary rationale is in `docs/prds/0011-per-file-md-manifest.md`.
 

bot_bottle/agent_provider.py

@@ -226,7 +226,7 @@ class AgentProvider(ABC):
         initial task in a non-interactive (headless) session.
 
         Called only when ``--prompt`` is passed to
-        ``./cli.py start --headless``; the returned args are appended
+        ``bot-bottle start --headless``; the returned args are appended
         after the provider's ``bypass_args`` and ``startup_args``."""
 
     def provision_ca(self, bottle: "Bottle", plan: "BottlePlan") -> None:

bot_bottle/backend/__init__.py

@@ -30,6 +30,7 @@ if TYPE_CHECKING:
         BottleImages,
         BottlePlan,
         BottleSpec,
+        EnumerationError,
         ExecResult,
     )
     from .selection import (
@@ -59,6 +60,7 @@ _LAZY_MODULES: dict[str, str] = {
     "BottleImages": "base",
     "BottleBackend": "base",
     "BackendStatus": "base",
+    "EnumerationError": "base",
     "get_bottle_backend": "selection",
     "known_backend_names": "selection",
     "has_backend": "selection",
@@ -100,6 +102,7 @@ __all__ = [
     "BottlePlan",
     "BottleSpec",
     "ExecResult",
+    "EnumerationError",
     "CommitCancelled",
     "Freezer",
     "get_freezer",

bot_bottle/backend/base.py

@@ -42,6 +42,10 @@ class BackendStatus(enum.IntEnum):
     READY = 0
 
 
+class EnumerationError(RuntimeError):
+    """A backend could not produce an authoritative live-resource snapshot."""
+
+
 @dataclass(frozen=True)
 class BottleSpec:
     """CLI-supplied intent. Backend-agnostic — each backend's prepare
@@ -168,6 +172,10 @@ class BottleCleanupPlan(ABC):
         """True iff there is nothing to clean up; the CLI uses this to
         short-circuit before showing the y/N."""
 
+    @abstractmethod
+    def intersect(self, current: "BottleCleanupPlan") -> "BottleCleanupPlan":
+        """Resources both displayed to the operator and currently removable."""
+
 
 @dataclass(frozen=True)
 class ExecResult:
@@ -523,7 +531,7 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
         host-appropriate config or commands. Prints to stdout/stderr and
         returns a shell exit code (0 = nothing to report / success). A
         backend that needs no host setup prints a short note and returns
-        0. Invoked generically by `./cli.py backend setup [--backend=…]`
+        0. Invoked generically by `bot-bottle backend setup [--backend=…]`
         so operators can provision any backend without a
         backend-specific command. Classmethod (like `is_available`) —
         it's a host query, not per-bottle state."""
@@ -540,14 +548,14 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
         stderr. When quiet=True returns the status code silently —
         useful for cheap programmatic checks.
 
-        Invoked by `./cli.py backend status [--backend=…]` (quiet=False)
+        Invoked by `bot-bottle backend status [--backend=…]` (quiet=False)
         and by is_backend_ready() (caller-controlled)."""
 
     @classmethod
     @abstractmethod
     def teardown(cls) -> int:
         """Undo `setup()` — the inverse operation, surfaced as
-        `./cli.py backend teardown [--backend=…]` (uninstall). Symmetric
+        `bot-bottle backend teardown [--backend=…]` (uninstall). Symmetric
         with setup: where setup is advisory (prints the privileged
         commands / declarative config to apply), teardown prints the
         commands / config change to remove the host prerequisites. A

bot_bottle/backend/cleanup_control.py

@@ -0,0 +1,60 @@
+"""Shared destructive-cleanup execution and failure accounting."""
+
+from __future__ import annotations
+
+import os
+import shutil
+import subprocess
+from collections.abc import Sequence
+from pathlib import Path
+
+
+class CleanupError(RuntimeError):
+    """One or more approved cleanup mutations did not complete."""
+
+
+class CleanupFailures:
+    """Attempt every approved mutation, then fail with complete diagnostics."""
+
+    def __init__(self) -> None:
+        self._messages: list[str] = []
+
+    def run(self, argv: Sequence[str], description: str) -> None:
+        raw_timeout = os.environ.get(
+            "BOT_BOTTLE_CLEANUP_COMMAND_TIMEOUT_SECONDS", "120",
+        )
+        try:
+            timeout = float(raw_timeout)
+        except ValueError:
+            timeout = 120.0
+        try:
+            result = subprocess.run(
+                list(argv), capture_output=True, text=True, check=False,
+                timeout=max(timeout, 1.0),
+            )
+        except (OSError, subprocess.SubprocessError) as exc:
+            self._messages.append(f"{description}: {exc}")
+            return
+        if result.returncode != 0:
+            detail = (result.stderr or result.stdout).strip()
+            self._messages.append(
+                f"{description}: {detail or f'exit {result.returncode}'}"
+            )
+
+    def remove_tree(self, path: Path, description: str) -> None:
+        try:
+            shutil.rmtree(path)
+        except FileNotFoundError:
+            return
+        except OSError as exc:
+            self._messages.append(f"{description}: {exc}")
+
+    def record(self, message: str) -> None:
+        self._messages.append(message)
+
+    def raise_if_any(self) -> None:
+        if self._messages:
+            raise CleanupError("; ".join(self._messages))
+
+
+__all__ = ["CleanupError", "CleanupFailures"]

bot_bottle/backend/docker/bottle_cleanup_plan.py

@@ -46,6 +46,22 @@ class DockerBottleCleanupPlan(BottleCleanupPlan):
             and not self.orphan_state_dirs
         )
 
+    def intersect(self, current: BottleCleanupPlan) -> "DockerBottleCleanupPlan":
+        if not isinstance(current, DockerBottleCleanupPlan):
+            raise TypeError("cleanup plans must have the same backend type")
+        return DockerBottleCleanupPlan(
+            projects=tuple(x for x in self.projects if x in current.projects),
+            stray_containers=tuple(
+                x for x in self.stray_containers if x in current.stray_containers
+            ),
+            stray_networks=tuple(
+                x for x in self.stray_networks if x in current.stray_networks
+            ),
+            orphan_state_dirs=tuple(
+                x for x in self.orphan_state_dirs if x in current.orphan_state_dirs
+            ),
+        )
+
     def print(self) -> None:
         print(file=sys.stderr)
         for name in self.projects:

bot_bottle/backend/docker/cleanup.py

@@ -23,11 +23,12 @@ Active-agent enumeration lives in `backend/docker/enumerate.py`.
 
 from __future__ import annotations
 
-import shutil
 import subprocess
 
 from ...paths import bot_bottle_root
-from ...log import info, warn
+from ...log import info
+from .. import EnumerationError
+from ..cleanup_control import CleanupFailures
 from . import util as docker_mod
 from .bottle_cleanup_plan import DockerBottleCleanupPlan
 from ...bottle_state import bottle_state_dir, is_preserved
@@ -36,15 +37,17 @@ from .compose import COMPOSE_PROJECT_PREFIX, list_compose_projects
 
 def _list_prefixed_containers() -> list[str]:
     """All bot-bottle-prefixed containers, running or stopped."""
-    result = subprocess.run(
-        ["docker", "ps", "-a",
-         "--filter", f"name=^{COMPOSE_PROJECT_PREFIX}",
-         "--format", "{{.Names}}\t{{.Label \"com.docker.compose.project\"}}"],
-        capture_output=True, text=True, check=False,
-    )
+    try:
+        result = subprocess.run(
+            ["docker", "ps", "-a",
+             "--filter", f"name=^{COMPOSE_PROJECT_PREFIX}",
+             "--format", "{{.Names}}\t{{.Label \"com.docker.compose.project\"}}"],
+            capture_output=True, text=True, check=False,
+        )
+    except OSError as exc:
+        raise EnumerationError(f"docker ps failed: {exc}") from exc
     if result.returncode != 0:
-        warn(f"docker ps failed: {result.stderr.strip()}")
-        return []
+        raise EnumerationError(f"docker ps failed: {result.stderr.strip()}")
     out: list[str] = []
     for line in (result.stdout or "").splitlines():
         if not line:
@@ -63,15 +66,19 @@ def _list_prefixed_networks() -> list[str]:
     to a compose project. Compose-managed networks have a
     `com.docker.compose.project` label; bare ones (from pre-compose
     code paths) don't."""
-    result = subprocess.run(
-        ["docker", "network", "ls",
-         "--filter", f"name={COMPOSE_PROJECT_PREFIX}",
-         "--format", "{{.Name}}\t{{.Label \"com.docker.compose.project\"}}"],
-        capture_output=True, text=True, check=False,
-    )
+    try:
+        result = subprocess.run(
+            ["docker", "network", "ls",
+             "--filter", f"name={COMPOSE_PROJECT_PREFIX}",
+             "--format", "{{.Name}}\t{{.Label \"com.docker.compose.project\"}}"],
+            capture_output=True, text=True, check=False,
+        )
+    except OSError as exc:
+        raise EnumerationError(f"docker network ls failed: {exc}") from exc
     if result.returncode != 0:
-        warn(f"docker network ls failed: {result.stderr.strip()}")
-        return []
+        raise EnumerationError(
+            f"docker network ls failed: {result.stderr.strip()}"
+        )
     out: list[str] = []
     for line in (result.stdout or "").splitlines():
         if not line:
@@ -120,7 +127,10 @@ def prepare_cleanup() -> DockerBottleCleanupPlan:
     `enumerate_active_agents()` so the orphan-state-dir bucket
     doesn't include slugs whose non-docker bottle is still up."""
     docker_mod.require_docker()
-    projects = list_compose_projects()
+    projects = list_compose_projects(
+        warn_on_error=False,
+        raise_on_error=True,
+    )
     project_set = set(projects)
     # Late import to avoid a circular at module-load time —
     # the backend package's __init__ imports this module.
@@ -140,40 +150,30 @@ def cleanup(plan: DockerBottleCleanupPlan) -> None:
     """Remove everything in the plan. Projects first (whose `compose
     down` reaps their containers + networks atomically), then stray
     legacy resources, then orphan state dirs."""
+    failures = CleanupFailures()
     for project in plan.projects:
         info(f"docker compose down  ({project})")
-        result = subprocess.run(
+        failures.run(
             ["docker", "compose", "-p", project, "down", "--volumes"],
-            capture_output=True, text=True, check=False,
+            f"docker compose down failed for {project}",
         )
-        if result.returncode != 0:
-            warn(
-                f"compose down failed for {project}: "
-                f"{result.stderr.strip()}"
-            )
 
     for name in plan.stray_containers:
         info(f"removing stray container {name}")
-        subprocess.run(
+        failures.run(
             ["docker", "rm", "-f", name],
-            stdout=subprocess.DEVNULL,
-            stderr=subprocess.DEVNULL,
-            check=False,
+            f"removing stray container {name}",
         )
 
     for name in plan.stray_networks:
         info(f"removing stray network {name}")
-        subprocess.run(
+        failures.run(
             ["docker", "network", "rm", name],
-            stdout=subprocess.DEVNULL,
-            stderr=subprocess.DEVNULL,
-            check=False,
+            f"removing stray network {name}",
         )
 
     for identity in plan.orphan_state_dirs:
         path = bottle_state_dir(identity)
         info(f"removing orphan state dir {path}")
-        try:
-            shutil.rmtree(path, ignore_errors=True)
-        except OSError as e:
-            warn(f"failed to remove {path}: {e}")
+        failures.remove_tree(path, f"removing orphan state dir {path}")
+    failures.raise_if_any()

bot_bottle/backend/docker/compose.py

@@ -16,6 +16,7 @@ from pathlib import Path
 from typing import Any
 
 from ...log import die, warn
+from ..base import EnumerationError
 
 
 # --- Lifecycle helpers (PRD 0018 chunk 3) ----------------------------------
@@ -52,19 +53,20 @@ def slug_from_compose_project(project: str) -> str:
 
 
 def list_compose_projects(
-    *, include_stopped: bool = True, warn_on_error: bool = True,
+    *,
+    include_stopped: bool = True,
+    warn_on_error: bool = True,
+    raise_on_error: bool = False,
 ) -> list[str]:
     """All compose project names starting with `bot-bottle-`.
     `include_stopped=True` (default) runs `docker compose ls --all`
     so exited projects appear too; pass False to get only projects
     with at least one running container.
 
-    Returns [] on docker daemon errors or malformed output rather
-    than raising — callers should treat the empty list as "no
-    projects discoverable", not "no projects exist". `warn_on_error`
-    stays true for explicit operator commands like cleanup, but active
-    discovery paths set it false so dashboard refreshes don't spam
-    stderr while Docker Desktop is stopped."""
+    Best-effort callers get ``[]`` on Docker errors or malformed output.
+    Enumeration callers pass ``raise_on_error=True`` so a failed query is not
+    reported as an authoritative empty result.
+    """
     argv = ["docker", "compose", "ls", "--format", "json"]
     if include_stopped:
         argv.insert(3, "--all")
@@ -72,19 +74,30 @@ def list_compose_projects(
         result = subprocess.run(
             argv, capture_output=True, text=True, check=False,
         )
-    except FileNotFoundError:
-        # docker binary not on PATH — same shape as a daemon-down
-        # error from the caller's POV: no projects discoverable.
+    except OSError as exc:
+        # Not only "not found": docker on PATH but not executable by this
+        # user raises PermissionError. Either way the query never ran, so an
+        # enumeration caller must not read the empty result as authoritative.
+        if raise_on_error:
+            raise EnumerationError(
+                f"docker compose ls failed: docker unavailable ({exc})"
+            ) from exc
         return []
     if result.returncode != 0:
+        message = f"docker compose ls failed: {result.stderr.strip()}"
+        if raise_on_error:
+            raise EnumerationError(message)
         if warn_on_error:
-            warn(f"docker compose ls failed: {result.stderr.strip()}")
+            warn(message)
         return []
     try:
         projects = json.loads(result.stdout or "[]")
     except json.JSONDecodeError as e:
+        message = f"docker compose ls returned malformed JSON: {e}"
+        if raise_on_error:
+            raise EnumerationError(message) from e
         if warn_on_error:
-            warn(f"docker compose ls returned malformed JSON: {e}")
+            warn(message)
         return []
     names: list[str] = []
     for p in projects:
@@ -97,7 +110,10 @@ def list_compose_projects(
 
 
 def list_active_slugs(
-    *, include_stopped: bool = False, warn_on_error: bool = True,
+    *,
+    include_stopped: bool = False,
+    warn_on_error: bool = True,
+    raise_on_error: bool = False,
 ) -> list[str]:
     """Slugs (project name minus prefix) of currently-running
     bottles. Used by the dashboard's operator-edit verbs to choose
@@ -108,6 +124,7 @@ def list_active_slugs(
             for p in list_compose_projects(
                 include_stopped=include_stopped,
                 warn_on_error=warn_on_error,
+                raise_on_error=raise_on_error,
             )
         ) if slug
     )

bot_bottle/backend/docker/consolidated_launch.py

@@ -68,6 +68,11 @@ def _network_container_ips(network: str) -> list[str]:
         "docker", "network", "inspect", "--format",
         "{{range .Containers}}{{.IPv4Address}} {{end}}", network,
     ])
+    if proc.returncode != 0:
+        detail = proc.stderr.strip() or f"exit {proc.returncode}"
+        raise ConsolidatedLaunchError(
+            f"could not inspect addresses on gateway network {network}: {detail}"
+        )
     ips: list[str] = []
     for entry in proc.stdout.split():
         ips.append(entry.split("/", 1)[0])

bot_bottle/backend/docker/enumerate.py

@@ -1,9 +1,8 @@
 """Active-agent enumeration for the docker backend.
 
 Returns `ActiveAgent` records the CLI `active` command and the
-dashboard agents pane consume. Empty when docker isn't reachable
-— gated by `has_backend('docker')` at the cross-backend caller
-so this module trusts that docker is available when called.
+dashboard agents pane consume. Docker query failures raise rather
+than masquerading as an authoritative empty result.
 
 The parser (`_parse_services_by_project`) is exposed for direct
 unit testing; the docker `docker ps` invocation is in
@@ -13,17 +12,18 @@ from __future__ import annotations
 
 import subprocess
 
-from .. import ActiveAgent
+from .. import ActiveAgent, EnumerationError
 from ...bottle_state import read_metadata
 from .compose import compose_project_name, list_active_slugs
 
 
 def enumerate_active() -> list[ActiveAgent]:
-    """All currently-running docker-backed agents. Caller is
-    responsible for gating on `has_backend('docker')` if it
-    matters; if docker is missing the `docker ps` call below
-    returns an empty list silently."""
-    slugs = list_active_slugs(include_stopped=False, warn_on_error=False)
+    """All currently-running docker-backed agents."""
+    slugs = list_active_slugs(
+        include_stopped=False,
+        warn_on_error=False,
+        raise_on_error=True,
+    )
     if not slugs:
         return []
     services_by_project = _query_services_by_project()
@@ -74,8 +74,9 @@ def _query_services_by_project() -> dict[str, set[str]]:
             ],
             capture_output=True, text=True, check=False,
         )
-    except FileNotFoundError:
-        return {}
+    except OSError as exc:
+        # Missing, or on PATH but not executable by this user (PermissionError).
+        raise EnumerationError(f"docker ps failed: docker unavailable ({exc})") from exc
     if r.returncode != 0:
-        return {}
+        raise EnumerationError(f"docker ps failed: {r.stderr.strip()}")
     return _parse_services_by_project(r.stdout or "")

bot_bottle/backend/docker/gateway.py

@@ -89,6 +89,10 @@ class DockerGateway(Gateway):
         when no dockerfile is configured (a pre-pulled image). BOT_BOTTLE_NO_CACHE
         forces a full rebuild (parity with `start --no-cache`)."""
         if self._dockerfile is None:
+            proc = run_docker(["docker", "pull", self.image_ref])
+            if proc.returncode != 0:
+                raise GatewayError(
+                    f"gateway image pull failed: {proc.stderr.strip()}")
             return
         context = self._build_context or resources.build_root()
         argv = ["docker", "build", "-t", self.image_ref,

bot_bottle/backend/docker/infra.py

@@ -34,6 +34,7 @@ from .orchestrator import (
     ORCHESTRATOR_NETWORK,
 )
 from ... import resources
+from ... import release_manifest
 from ...gateway import (
     GATEWAY_IMAGE,
     GATEWAY_NAME,
@@ -90,6 +91,14 @@ class DockerInfraService(InfraService):
         self._orchestrator_name = orchestrator_name
         self._orchestrator_label = orchestrator_label
         self._gateway_name = gateway_name
+        resolved_orchestrator, orchestrator_local = release_manifest.oci_image(
+            "orchestrator", orchestrator_image)
+        resolved_gateway, gateway_local = release_manifest.oci_image(
+            "gateway", gateway_image)
+        self.orchestrator_image = resolved_orchestrator
+        self.gateway_image = resolved_gateway
+        self._orchestrator_local = orchestrator_local
+        self._gateway_local = gateway_local
 
     def orchestrator(self) -> DockerOrchestrator:
         """The control-plane service. Cheap to reconstruct — `ensure_built`
@@ -104,6 +113,8 @@ class DockerInfraService(InfraService):
             repo_root=self._repo_root,
             host_root=self._host_root,
             root_mount_source=self._root_mount_source,
+            dockerfile=(
+                "Dockerfile.orchestrator" if self._orchestrator_local else None),
         )
 
     def gateway(self) -> DockerGateway:
@@ -119,6 +130,7 @@ class DockerInfraService(InfraService):
             control_network=self.control_network,
             build_context=self._repo_root,
             ca_mount_source=self._gateway_ca_mount_source,
+            dockerfile="Dockerfile.gateway" if self._gateway_local else None,
         )
 
     def ensure_running(
@@ -132,8 +144,8 @@ class DockerInfraService(InfraService):
         gateway = self.gateway()
         # Build both images (cache-aware; a no-op when nothing changed) before
         # bringing either plane up.
-        orchestrator.ensure_built()
-        gateway.ensure_built()
+        orchestrator.ensure_available()
+        gateway.ensure_available()
 
         orchestrator.ensure_running(startup_timeout=startup_timeout)
 

bot_bottle/backend/docker/orchestrator.py

@@ -125,6 +125,10 @@ class DockerOrchestrator(Orchestrator):
         no-op when nothing changed). No-op when no dockerfile is configured (a
         pre-pulled image). BOT_BOTTLE_NO_CACHE forces a full rebuild."""
         if self._dockerfile is None:
+            proc = run_docker(["docker", "pull", self.image_ref])
+            if proc.returncode != 0:
+                raise GatewayError(
+                    f"orchestrator image pull failed: {proc.stderr.strip()}")
             return
         argv = ["docker", "build", "-t", self.image_ref,
                 "-f", str(self._repo_root / self._dockerfile),

bot_bottle/backend/docker/setup.py

@@ -8,7 +8,7 @@ pointer, and `status()` reports whether docker is usable.
 This is intentionally minimal; a richer version (daemon config checks,
 gVisor/runsc install guidance, rootless-docker hints) is tracked
 separately. Reached via `DockerBottleBackend.setup` / `.status`, which
-the generic `./cli.py backend {setup,status}` dispatches to.
+the generic `bot-bottle backend {setup,status}` dispatches to.
 """
 
 from __future__ import annotations
@@ -69,7 +69,7 @@ def teardown() -> int:
     sys.stderr.write(
         "Docker backend: nothing to undo — it provisions no privileged host "
         "state (networks and the gateway are per-launch and are "
-        "removed by `./cli.py cleanup`). Docker itself is left installed.\n"
+        "removed by `bot-bottle cleanup`). Docker itself is left installed.\n"
     )
     return 0
 
@@ -89,5 +89,5 @@ def status() -> int:
     runsc = _docker_on_path() and _util.runsc_available()
     sys.stderr.write(f"gVisor runsc runtime: {'registered' if runsc else 'not registered (optional)'}\n")
     if not ok:
-        sys.stderr.write("\nRun: ./cli.py backend setup --backend=docker\n")
+        sys.stderr.write("\nRun: bot-bottle backend setup --backend=docker\n")
     return 0 if ok else 1

bot_bottle/backend/firecracker/bottle_cleanup_plan.py

@@ -27,3 +27,11 @@ class FirecrackerBottleCleanupPlan(BottleCleanupPlan):
     @property
     def empty(self) -> bool:
         return not (self.vm_pids or self.run_dirs)
+
+    def intersect(self, current: BottleCleanupPlan) -> "FirecrackerBottleCleanupPlan":
+        if not isinstance(current, FirecrackerBottleCleanupPlan):
+            raise TypeError("cleanup plans must have the same backend type")
+        return FirecrackerBottleCleanupPlan(
+            vm_pids=tuple(x for x in self.vm_pids if x in current.vm_pids),
+            run_dirs=tuple(x for x in self.run_dirs if x in current.run_dirs),
+        )

bot_bottle/backend/firecracker/cleanup.py

@@ -11,10 +11,9 @@ Reaps *orphans* only — resources with no live VM behind them:
     — a VMM left lingering after its dir was removed.
 
 A run dir with a *live* firecracker process is a running bottle and is
-left strictly alone: it is neither killed nor removed. (The backend's
-`enumerate_active` registry is still a stub — #354 — so a live process
-is the only reliable "this bottle is in use" signal we have. Once the
-registry lands, registry-orphaned-but-running VMs can be reaped too.)
+left strictly alone: it is neither killed nor removed. Active-agent
+enumeration uses this same process snapshot, so cleanup and generic
+backend consumers agree about which bottles are running.
 
 TAP slots free themselves (the flock drops when the launcher exits), so
 there is nothing to reclaim there.
@@ -22,14 +21,16 @@ there is nothing to reclaim there.
 
 from __future__ import annotations
 
+from collections.abc import Sequence
 import os
-import shutil
 import signal
 import subprocess
 from pathlib import Path
 
 from ...log import info
-from . import util
+from .. import EnumerationError
+from ..cleanup_control import CleanupError, CleanupFailures
+from . import lifecycle_lock, util
 from .bottle_cleanup_plan import FirecrackerBottleCleanupPlan
 
 
@@ -37,7 +38,7 @@ def _run_root() -> Path:
     return util.cache_dir() / "run"
 
 
-def _run_dir_of(cmd: str, run_root: Path) -> Path | None:
+def _run_dir_of(args: Sequence[str], run_root: Path) -> Path | None:
     """The bottle run dir a firecracker cmdline belongs to, or None.
 
     A bottle VM is launched with `--config-file <run_root>/<slug>/config.json`,
@@ -45,15 +46,35 @@ def _run_dir_of(cmd: str, run_root: Path) -> Path | None:
     the run root. Anything else (a builder VM, the infra VM elsewhere) is
     not ours to reap here.
     """
-    toks = cmd.split()
-    for i, tok in enumerate(toks):
-        if tok == "--config-file" and i + 1 < len(toks):
-            parent = Path(toks[i + 1]).parent
+    for i, arg in enumerate(args):
+        if arg == "--config-file" and i + 1 < len(args):
+            parent = Path(args[i + 1]).parent
             if parent.parent == run_root:
                 return parent
     return None
 
 
+def _decode_cmdline(raw: bytes) -> tuple[str, ...]:
+    """Decode Linux's NUL-delimited argv without losing embedded spaces."""
+    return tuple(
+        value.decode(errors="surrogateescape")
+        for value in raw.split(b"\0") if value
+    )
+
+
+def _process_args(pid: int) -> tuple[str, ...] | None:
+    """Read one process's lossless argv, or None when it exited meanwhile."""
+    try:
+        raw = Path(f"/proc/{pid}/cmdline").read_bytes()
+    except FileNotFoundError:
+        return None
+    except OSError as exc:
+        raise EnumerationError(
+            f"could not inspect Firecracker pid {pid}: {exc}"
+        ) from exc
+    return _decode_cmdline(raw)
+
+
 def _scan_processes(run_root: Path) -> tuple[set[str], list[int]]:
     """Inspect running firecracker VMs under ``run_root``.
 
@@ -62,23 +83,34 @@ def _scan_processes(run_root: Path) -> tuple[set[str], list[int]]:
       * ``orphan_pids`` — firecracker pids whose run dir no longer exists
         (a lingering VMM to kill).
     """
-    result = subprocess.run(
-        ["pgrep", "-a", "firecracker"],
-        capture_output=True, text=True, check=False,
-    )
-    if result.returncode != 0:
+    try:
+        result = subprocess.run(
+            ["pgrep", "firecracker"],
+            capture_output=True, text=True, check=False,
+        )
+    except OSError as exc:
+        raise EnumerationError(
+            f"could not enumerate Firecracker processes: {exc}"
+        ) from exc
+    if result.returncode == 1:
+        # pgrep's documented "no processes matched" result.
         return set(), []
+    if result.returncode != 0:
+        detail = (result.stderr or "").strip() or f"exit {result.returncode}"
+        raise EnumerationError(
+            f"could not enumerate Firecracker processes: {detail}"
+        )
     live: set[str] = set()
     orphan_pids: list[int] = []
     for line in result.stdout.splitlines():
-        parts = line.split(None, 1)
-        if len(parts) != 2:
-            continue
         try:
-            pid = int(parts[0])
+            pid = int(line.strip())
         except ValueError:
             continue
-        run_dir = _run_dir_of(parts[1], run_root)
+        args = _process_args(pid)
+        if args is None:
+            continue
+        run_dir = _run_dir_of(args, run_root)
         if run_dir is None:
             continue
         if run_dir.is_dir():
@@ -114,12 +146,54 @@ def prepare_cleanup() -> FirecrackerBottleCleanupPlan:
 
 
 def cleanup(plan: FirecrackerBottleCleanupPlan) -> None:
-    for pid in plan.vm_pids:
+    """Revalidate the preview under the launch lock, then remove its survivors."""
+    with lifecycle_lock.hold():
+        fresh = prepare_cleanup()
+        approved_pids = set(plan.vm_pids).intersection(fresh.vm_pids)
+        approved_dirs = set(plan.run_dirs).intersection(fresh.run_dirs)
+        failures = CleanupFailures()
+        for pid in sorted(approved_pids):
+            try:
+                _terminate_orphan(pid, _run_root())
+            except CleanupError as exc:
+                failures.record(str(exc))
+        for path in sorted(approved_dirs):
+            info(f"rm -rf {path}")
+            failures.remove_tree(Path(path), f"removing Firecracker run dir {path}")
+        failures.raise_if_any()
+
+
+def _terminate_orphan(pid: int, run_root: Path) -> None:
+    """Signal exactly the process identity that still owns an orphan config."""
+    try:
+        pidfd = os.pidfd_open(pid)
+    except ProcessLookupError:
+        return
+    except OSError as exc:
+        raise EnumerationError(
+            f"could not pin Firecracker pid {pid} for cleanup: {exc}"
+        ) from exc
+    try:
+        try:
+            raw = Path(f"/proc/{pid}/cmdline").read_bytes()
+        except FileNotFoundError:
+            return
+        except OSError as exc:
+            raise EnumerationError(
+                f"could not revalidate Firecracker pid {pid}: {exc}"
+            ) from exc
+        args = _decode_cmdline(raw)
+        run_dir = _run_dir_of(args, run_root)
+        if run_dir is None or run_dir.is_dir():
+            return
         info(f"kill firecracker VM pid {pid}")
         try:
-            os.kill(pid, signal.SIGTERM)
+            signal.pidfd_send_signal(pidfd, signal.SIGTERM)
         except ProcessLookupError:
-            pass
-    for path in plan.run_dirs:
-        info(f"rm -rf {path}")
-        shutil.rmtree(path, ignore_errors=True)
+            return
+        except OSError as exc:
+            raise CleanupError(
+                f"could not signal Firecracker pid {pid}: {exc}"
+            ) from exc
+    finally:
+        os.close(pidfd)

bot_bottle/backend/firecracker/enumerate.py

@@ -1,14 +1,32 @@
 """Active-agent enumeration for the Firecracker backend.
 
-The backend is disabled during the companion-container removal (#385) — it can't
-launch bottles, so there are none to enumerate. Real enumeration returns
-with the backend's consolidated relaunch (#354).
+Running bottles are the Firecracker processes whose ``--config-file`` points
+at an existing per-bottle run directory. The same authoritative process scan
+protects cleanup from deleting live VMs; operational scan failures propagate
+as ``EnumerationError`` instead of masquerading as an empty host.
 """
 
 from __future__ import annotations
 
+from ...bottle_state import read_metadata
 from .. import ActiveAgent
+from .cleanup import live_run_dirs
 
 
 def enumerate_active() -> list[ActiveAgent]:
-    return []
+    out: list[ActiveAgent] = []
+    for run_dir in live_run_dirs():
+        slug = run_dir.name
+        metadata = read_metadata(slug)
+        out.append(ActiveAgent(
+            backend_name="firecracker",
+            slug=slug,
+            agent_name=metadata.agent_name if metadata else "?",
+            started_at=metadata.started_at if metadata else "",
+            # Firecracker uses the shared gateway, so there are no
+            # per-bottle gateway service containers to report.
+            services=(),
+            label=metadata.label if metadata else "",
+            color=metadata.color if metadata else "",
+        ))
+    return out

bot_bottle/backend/firecracker/infra.py

@@ -61,7 +61,7 @@ class FirecrackerInfraService(InfraService):
                 return url
             # Clear stale/hung/OUTDATED VMs holding either link before booting fresh.
             infra_vm.stop()
-            infra_vm.ensure_built()
+            infra_vm.ensure_available()
             # Orchestrator first — the gateway daemons reach the control plane at
             # startup. Each service owns its own boot; the orchestrator (which
             # holds the signing key) mints the role-scoped `gateway` JWT for the

bot_bottle/backend/firecracker/infra_artifact.py

@@ -41,6 +41,8 @@ from ... import resources
 from ...log import die, info
 from . import util
 
+ARTIFACT_HTTP_TIMEOUT_SECONDS = 30.0
+
 # 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"
@@ -54,6 +56,7 @@ _BUILD_INPUTS = {
         "image-build-args.json",
         "Dockerfile.orchestrator",
         "Dockerfile.orchestrator.fc",
+        "requirements.orchestrator.lock",
     ),
     "gateway": (
         "image-build-args.json",
@@ -163,7 +166,9 @@ 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:
+        with urllib.request.urlopen(
+            _open(url), timeout=ARTIFACT_HTTP_TIMEOUT_SECONDS,
+        ) as resp, open(tmp, "wb") as out:
             shutil.copyfileobj(resp, out, _CHUNK)
     except urllib.error.HTTPError as e:
         tmp.unlink(missing_ok=True)
@@ -190,7 +195,9 @@ def _sha256_file(path: Path) -> str:
     return h.hexdigest()
 
 
-def ensure_artifact_gz(version: str, *, role: str) -> Path:
+def ensure_artifact_gz(
+    version: str, *, role: str, expected_sha256: str | None = None,
+) -> Path:
     """The verified, cached `rootfs.ext4.gz` for `role` at `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
@@ -217,6 +224,11 @@ def ensure_artifact_gz(version: str, *, role: str) -> Path:
         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()
+        if expected_sha256 is not None and expected != expected_sha256:
+            die(
+                f"infra candidate packaged checksum mismatch ({role}) for {version}:\n"
+                f"    packaged {expected_sha256}\n    candidate {expected}"
+            )
         actual = _sha256_file(gz)
         if actual != expected:
             die(
@@ -230,7 +242,13 @@ def ensure_artifact_gz(version: str, *, role: str) -> Path:
     gz = root / _GZ_NAME
     ok = root / ".verified"
     if gz.is_file() and ok.is_file():
-        return gz
+        actual = _sha256_file(gz)
+        recorded = ok.read_text(encoding="utf-8").strip()
+        wanted = expected_sha256 or recorded
+        if actual == recorded == wanted:
+            return gz
+        gz.unlink(missing_ok=True)
+        ok.unlink(missing_ok=True)
 
     info(f"pulling infra rootfs artifact {_package(role)}/{version}")
     _download(artifact_url(version, _GZ_NAME, role=role), gz)
@@ -238,6 +256,14 @@ def ensure_artifact_gz(version: str, *, role: str) -> Path:
     _download(artifact_url(version, _SHA_NAME, role=role), sha)
 
     expected = sha.read_text().split()[0].strip().lower()
+    if expected_sha256 is not None and expected != expected_sha256:
+        gz.unlink(missing_ok=True)
+        sha.unlink(missing_ok=True)
+        die(
+            f"infra artifact published checksum mismatch ({role}) for {version}:\n"
+            f"    packaged  {expected_sha256}\n"
+            f"    published {expected}"
+        )
     actual = _sha256_file(gz)
     if actual != expected:
         gz.unlink(missing_ok=True)
@@ -248,15 +274,22 @@ def ensure_artifact_gz(version: str, *, role: str) -> Path:
             f"    actual   {actual}\n"
             f"    refusing to boot an unverified rootfs."
         )
-    ok.write_text("ok\n")
+    ok.write_text(actual + "\n")
     return gz
 
 
-def materialize_ext4(version: str, dest: Path, *, role: str) -> None:
+def materialize_ext4(
+    version: str,
+    dest: Path,
+    *,
+    role: str,
+    expected_sha256: str | None = None,
+) -> None:
     """Ensure the verified `role` 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, role=role)
+    gz = ensure_artifact_gz(
+        version, role=role, expected_sha256=expected_sha256)
     tmp = dest.with_suffix(dest.suffix + ".part")
     info(f"expanding {role} infra rootfs -> {dest}")
     with gzip.open(gz, "rb") as src, open(tmp, "wb") as out:

bot_bottle/backend/firecracker/infra_vm.py

@@ -43,6 +43,7 @@ from pathlib import Path
 from typing import Generator
 
 from ... import resources
+from ... import release_manifest
 from ...log import die, info
 from ..docker import util as docker_mod
 from . import firecracker_vm, infra_artifact, netpool, util
@@ -108,6 +109,18 @@ def _role_version(role: str) -> str:
     return infra_artifact.infra_artifact_version(role_init(role), role)
 
 
+def _role_release(role: str) -> tuple[str, str | None]:
+    manifest = release_manifest.packaged_manifest()
+    if manifest is None or release_manifest.local_build_requested():
+        return _role_version(role), None
+    artifact = (
+        manifest.firecracker_orchestrator
+        if role == "orchestrator"
+        else manifest.firecracker_gateway
+    )
+    return artifact.version, artifact.sha256
+
+
 def ensure_built() -> None:
     """Ensure both infra rootfs artifacts are available before boot.
 
@@ -117,11 +130,16 @@ def ensure_built() -> None:
     `BOT_BOTTLE_INFRA_BUILD=local` instead builds the images from source with
     host Docker (the orchestrator-fc image is `FROM` the orchestrator image, so
     it must exist first) — for iterating on the Dockerfiles."""
-    if infra_artifact.local_build_requested():
+    if release_manifest.local_build_requested():
         build_infra_images_with_docker()
         return
     for role in infra_artifact.ROLES:
-        infra_artifact.ensure_artifact_gz(_role_version(role), role=role)
+        version, expected = _role_release(role)
+        infra_artifact.ensure_artifact_gz(
+            version, role=role, expected_sha256=expected)
+
+
+ensure_available = ensure_built
 
 
 def build_infra_images_with_docker() -> None:
@@ -204,7 +222,7 @@ def boot_vm(
     Records the PID."""
     if not netpool.tap_present(slot.iface):
         die(f"infra link {slot.iface} not present.\n"
-            f"    ./cli.py backend setup --backend=firecracker")
+            f"    bot-bottle backend setup --backend=firecracker")
 
     run_dir.mkdir(parents=True, exist_ok=True)
     rootfs = run_dir / "rootfs.ext4"
@@ -214,7 +232,9 @@ def boot_vm(
     else:
         # Prebuilt artifact already carries the role's build slack; expand it to
         # a fresh writable rootfs for this boot.
-        infra_artifact.materialize_ext4(_role_version(role), rootfs, role=role)
+        version, expected = _role_release(role)
+        infra_artifact.materialize_ext4(
+            version, rootfs, role=role, expected_sha256=expected)
     private_key, pubkey = _stable_keypair()
 
     info(f"booting {role} VM on {slot.iface} (guest {slot.guest_ip})")
@@ -264,7 +284,8 @@ def _version_file() -> Path:
 def expected_version() -> str:
     """The combined marker for the running pair: both per-plane artifact
     versions, so a change to either rootfs dislodges the adopted pair."""
-    return " ".join(f"{role}={_role_version(role)}" for role in infra_artifact.ROLES)
+    return " ".join(
+        f"{role}={_role_release(role)[0]}" for role in infra_artifact.ROLES)
 
 
 def adoptable(key: Path, url: str, want: str) -> bool:

bot_bottle/backend/firecracker/isolation_probe.py

@@ -108,7 +108,7 @@ def verify_isolation(private_key: Path, guest_ip: str) -> None:
         die(f"ISOLATION FAILURE: the VM reached the host canary "
             f"{canary_ip}:{canary_port}. The egress boundary is not in "
             f"force — refusing to run the agent (fail-closed). Verify the "
-            f"nft table with: ./cli.py backend setup --backend=firecracker")
+            f"nft table with: bot-bottle backend setup --backend=firecracker")
     if result.returncode == 2:
         die("isolation probe inconclusive: the guest has no python3/bash/nc "
             "to run the connectivity test. Refusing to continue "

bot_bottle/backend/firecracker/launch.py

@@ -46,7 +46,7 @@ from ...log import die, info, warn
 from ...supervisor.types 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 . import firecracker_vm, image_builder, isolation_probe, lifecycle_lock, netpool, util
 from .bottle import FirecrackerBottle
 from .bottle_plan import FirecrackerBottlePlan
 from ...orchestrator.store.config_store import resolve_teardown_timeout
@@ -164,25 +164,29 @@ def launch(
         )
 
         # 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)
-        # Remove the run dir on teardown so the per-bottle rootfs.ext4 (~1G)
-        # doesn't leak. Registered before vm.terminate below so it runs *after*
-        # it (ExitStack is LIFO): the VM is gone before we rm its rootfs.
-        stack.callback(lambda: shutil.rmtree(run_dir, ignore_errors=True))
-        rootfs = run_dir / "rootfs.ext4"
-        util.build_rootfs_ext4(agent_base, rootfs)
-        private_key, pubkey = util.generate_keypair(run_dir)
+        # Cleanup takes the same lock while refreshing its process snapshot.
+        # Hold it until the VMM exists so a newly-created run dir can never be
+        # mistaken for an orphan in the build-before-boot window.
+        with lifecycle_lock.hold():
+            run_dir = util.cache_dir() / "run" / plan.slug
+            run_dir.mkdir(parents=True, exist_ok=True)
+            # Remove the run dir on teardown so the per-bottle rootfs.ext4 (~1G)
+            # doesn't leak. Registered before vm.terminate below so it runs
+            # *after* it (ExitStack is LIFO).
+            stack.callback(lambda: shutil.rmtree(run_dir, ignore_errors=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,
-        )
+            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)
         persist_env_var_secret(private_key, slot.guest_ip, ctx.env_var_secret)

bot_bottle/backend/firecracker/lifecycle_lock.py

@@ -0,0 +1,30 @@
+"""Serialize Firecracker run-directory creation with orphan cleanup."""
+
+from __future__ import annotations
+
+import fcntl
+from contextlib import contextmanager
+from pathlib import Path
+from typing import Generator
+
+from . import util
+
+
+def _lock_path() -> Path:
+    return util.cache_dir() / "run.lifecycle.lock"
+
+
+@contextmanager
+def hold() -> Generator[None]:
+    """Exclude cleanup while a launch directory lacks a visible VMM."""
+    path = _lock_path()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with path.open("a", encoding="utf-8") as handle:
+        fcntl.flock(handle, fcntl.LOCK_EX)
+        try:
+            yield
+        finally:
+            fcntl.flock(handle, fcntl.LOCK_UN)
+
+
+__all__ = ["hold"]

bot_bottle/backend/firecracker/netpool.py

@@ -3,7 +3,7 @@ config renderers (shell command + NixOS module) shown to operators.
 
 The Firecracker backend needs a privileged one-time network setup:
 a pool of point-to-point TAP devices (owned by the invoking user, so
-`./cli.py start` never needs root) and a dedicated nftables table that
+`bot-bottle start` never needs root) and a dedicated nftables table that
 isolates every VM. The pool parameters live in exactly one place —
 `netpool.defaults.env`, a plain KEY=VALUE file next to this module —
 and every consumer reads *that*: this module (below), the shell script
@@ -177,14 +177,20 @@ def gw_slot() -> Slot:
 # --- fail-closed verification ---------------------------------------
 
 def _run_ok(argv: list[str]) -> bool:
-    """Run a probe command, treating a missing binary as failure
+    """Run a probe command, treating an unavailable binary as failure
     (rather than crashing) so callers can stay fail-closed."""
     try:
         return subprocess.run(
             argv, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
             check=False,
         ).returncode == 0
-    except FileNotFoundError:
+    except OSError:
+        # Not only "missing". A name on PATH that isn't executable by this
+        # user raises PermissionError, and CPython reports that EACCES in
+        # preference to the ENOENT from the other PATH entries — which is
+        # how `doctor` came to die with a traceback on a fresh macOS
+        # account. Any OSError means the probe couldn't run, which for a
+        # fail-closed check is indistinguishable from "not present".
         return False
 
 
@@ -244,7 +250,9 @@ def overlapping_routes() -> list[RouteConflict]:
             ["ip", "-json", "route", "show", "table", "all"],
             capture_output=True, text=True, check=False,
         )
-    except FileNotFoundError:
+    except OSError:
+        # Missing, or present-but-not-executable for this user; either way
+        # there are no routes we can enumerate. See _run_ok.
         return []
     if proc.returncode != 0 or not proc.stdout.strip():
         return []
@@ -307,11 +315,11 @@ def allocate(slug: str) -> tuple[Slot, IO[str]]:
         return s, handle
     die(f"Firecracker TAP pool exhausted ({pool_size()} slots, all in "
         f"use). Stop a running bottle or raise BOT_BOTTLE_FC_POOL_SIZE "
-        f"and re-run `./cli.py backend setup --backend=firecracker`.")
+        f"and re-run `bot-bottle backend setup --backend=firecracker`.")
     raise AssertionError("unreachable")
 
 
-# --- config renderers (shown by `./cli.py backend setup`) -----------
+# --- config renderers (shown by `bot-bottle backend setup`) -----------
 
 # The persistent unit is the portable install: the same systemd oneshot
 # on every systemd distro (Debian/Ubuntu/Fedora/RHEL/Arch/…).

bot_bottle/backend/firecracker/publish_infra.py

@@ -34,6 +34,7 @@ from pathlib import Path
 from . import infra_artifact, infra_vm, util
 
 _CHUNK = 1 << 20
+_REGISTRY_HTTP_TIMEOUT_SECONDS = 30.0
 
 _GZ_NAME = "rootfs.ext4.gz"
 _SHA_NAME = "rootfs.ext4.gz.sha256"
@@ -91,7 +92,9 @@ def _put(url: str, body: "bytes | Path", token: str) -> None:
         req.add_header("Authorization", f"token {token}")
     req.add_header("Content-Type", "application/octet-stream")
     try:
-        with urllib.request.urlopen(req) as resp:
+        with urllib.request.urlopen(
+            req, timeout=_REGISTRY_HTTP_TIMEOUT_SECONDS,
+        ) as resp:
             print(f"  uploaded {url} (HTTP {resp.status})")
     except urllib.error.HTTPError as e:
         if e.code == 409:
@@ -112,7 +115,9 @@ def _delete(url: str, token: str) -> None:
     if token:
         req.add_header("Authorization", f"token {token}")
     try:
-        with urllib.request.urlopen(req):
+        with urllib.request.urlopen(
+            req, timeout=_REGISTRY_HTTP_TIMEOUT_SECONDS,
+        ):
             pass
     except urllib.error.HTTPError as e:
         if e.code != 404:
@@ -151,7 +156,10 @@ def _try_download_published(role: str, role_dir: Path) -> str | None:
     version = _role_version(role)
     sha_url = infra_artifact.artifact_url(version, _SHA_NAME, role=role)
     try:
-        with urllib.request.urlopen(infra_artifact._open(sha_url)):
+        with urllib.request.urlopen(
+            infra_artifact._open(sha_url),
+            timeout=_REGISTRY_HTTP_TIMEOUT_SECONDS,
+        ):
             pass
     except urllib.error.HTTPError as e:
         if e.code == 404:
@@ -195,7 +203,10 @@ def _publish_bundle(role: str, role_dir: Path, token: str) -> str:
     # present, a re-publish 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:
+        with urllib.request.urlopen(
+            infra_artifact._open(sha_url),
+            timeout=_REGISTRY_HTTP_TIMEOUT_SECONDS,
+        ) as resp:
             remote_sha = resp.read().decode("utf-8").split()[0].strip().lower()
     except urllib.error.HTTPError as e:
         if e.code != 404:

bot_bottle/backend/firecracker/setup.py

@@ -8,7 +8,7 @@ bundled setup script. `status()` reports what's present, including
 whether the pool range collides with an existing route.
 
 Called through `FirecrackerBottleBackend.setup` / `.status`, which the
-generic `./cli.py backend {setup,status}` command dispatches to.
+generic `bot-bottle backend {setup,status}` command dispatches to.
 """
 
 from __future__ import annotations
@@ -34,7 +34,7 @@ _UNIT_PATH = Path("/etc/systemd/system") / netpool.SYSTEMD_UNIT
 
 def _owner() -> str:
     # Under `sudo`, USER is root but SUDO_USER is the real invoker — the
-    # TAPs must be owned by them so `./cli.py start` stays rootless.
+    # TAPs must be owned by them so `bot-bottle start` stays rootless.
     return os.environ.get("SUDO_USER") or os.environ.get("USER") or "youruser"
 
 
@@ -161,7 +161,7 @@ def _setup_systemd() -> None:
         if rc == 0:
             sys.stderr.write(
                 f"Installed and started {netpool.SYSTEMD_UNIT}. Verify with "
-                f"`./cli.py backend status --backend=firecracker`.\n"
+                f"`bot-bottle backend status --backend=firecracker`.\n"
             )
         else:
             sys.stderr.write(
@@ -181,7 +181,7 @@ def _setup_systemd() -> None:
     )
     sys.stderr.write(
         f"\n(Or re-run this as root to install it directly: "
-        f"sudo ./cli.py backend setup --backend=firecracker)\n"
+        f"sudo bot-bottle backend setup --backend=firecracker)\n"
     )
 
 
@@ -334,7 +334,7 @@ def status() -> int:
         sys.stderr.write(f"range overlap: none (base {netpool.ip_base()})\n")
     _report_persistence()
     if not ok:
-        sys.stderr.write("\nRun: ./cli.py backend setup --backend=firecracker\n")
+        sys.stderr.write("\nRun: bot-bottle backend setup --backend=firecracker\n")
     return 0 if ok else 1
 
 

bot_bottle/backend/firecracker/util.py

@@ -9,7 +9,7 @@ generation.
 
 The privileged network setup (TAP pool + nft table) is a one-time
 operator step — see `netpool.py`, `scripts/firecracker-netpool.sh`,
-and `./cli.py backend setup --backend=firecracker`.
+and `bot-bottle backend setup --backend=firecracker`.
 """
 
 from __future__ import annotations
@@ -132,18 +132,18 @@ def _require_network_pool() -> None:
              f"{netpool.pool_size()} slots) overlaps existing routes: "
              f"{detail}. This can shadow or be shadowed by that route; "
              f"set BOT_BOTTLE_FC_IP_BASE to a free range and re-run "
-             f"./cli.py backend setup --backend=firecracker.")
+             f"bot-bottle backend setup --backend=firecracker.")
     missing = netpool.missing_taps()
     if missing:
         die(f"network pool incomplete — missing TAP devices: "
-            f"{', '.join(missing)}.\n    ./cli.py backend setup --backend=firecracker")
+            f"{', '.join(missing)}.\n    bot-bottle backend setup --backend=firecracker")
     if shutil.which("nft") is not None and not netpool.nft_table_present():
         # nft is queryable and says the table is absent — that's a
         # definite, catchable misconfiguration; fail early.
         warn(f"isolation table `inet {netpool.NFT_TABLE}` not found via nft. "
              "If this is a permissions issue it will be re-checked "
              "empirically after boot; otherwise run: "
-             "./cli.py backend setup --backend=firecracker")
+             "bot-bottle backend setup --backend=firecracker")
 
 
 # --- rootfs pipeline (rootless) -------------------------------------

bot_bottle/backend/freeze.py

@@ -43,7 +43,7 @@ class Freezer(ABC):
 
         Calls _freeze for the backend-specific snapshot, then writes the
         committed image reference to per-bottle state and marks the bottle
-        preserved so the next `./cli.py resume` boots from the snapshot.
+        preserved so the next `bot-bottle resume` boots from the snapshot.
 
         Raises CommitCancelled if the user declines an interactive
         confirmation prompt (e.g. the macos-container stop prompt).
@@ -51,7 +51,7 @@ class Freezer(ABC):
         image_ref = self._freeze(agent)
         write_committed_image(agent.slug, image_ref)
         mark_preserved(agent.slug)
-        info(f"to resume from this snapshot: ./cli.py resume {agent.slug}")
+        info(f"to resume from this snapshot: bot-bottle resume {agent.slug}")
         self._export_hint(agent.slug, image_ref)
 
     @abstractmethod

bot_bottle/backend/macos_container/bottle_cleanup_plan.py

@@ -25,3 +25,11 @@ class MacosContainerBottleCleanupPlan(BottleCleanupPlan):
     @property
     def empty(self) -> bool:
         return not self.containers and not self.networks
+
+    def intersect(self, current: BottleCleanupPlan) -> "MacosContainerBottleCleanupPlan":
+        if not isinstance(current, MacosContainerBottleCleanupPlan):
+            raise TypeError("cleanup plans must have the same backend type")
+        return MacosContainerBottleCleanupPlan(
+            containers=tuple(x for x in self.containers if x in current.containers),
+            networks=tuple(x for x in self.networks if x in current.networks),
+        )

bot_bottle/backend/macos_container/cleanup.py

@@ -4,7 +4,9 @@ from __future__ import annotations
 
 import subprocess
 
-from ...log import info, warn
+from .. import EnumerationError
+from ..cleanup_control import CleanupFailures
+from ...log import info
 from . import util as container_mod
 from .bottle_cleanup_plan import MacosContainerBottleCleanupPlan
 
@@ -19,8 +21,8 @@ def _list_prefixed_containers() -> list[str]:
         check=False,
     )
     if result.returncode != 0:
-        warn(f"container list failed: {result.stderr.strip()}")
-        return []
+        detail = result.stderr.strip() or f"exit {result.returncode}"
+        raise EnumerationError(f"container list failed: {detail}")
     return sorted(
         name for name in (line.strip() for line in result.stdout.splitlines())
         if name.startswith(_PREFIX)
@@ -35,7 +37,8 @@ def _list_prefixed_networks() -> list[str]:
         check=False,
     )
     if result.returncode != 0:
-        return []
+        detail = result.stderr.strip() or f"exit {result.returncode}"
+        raise EnumerationError(f"container network list failed: {detail}")
     return sorted(
         name for name in (line.strip() for line in result.stdout.splitlines())
         if name.startswith(_PREFIX)
@@ -51,19 +54,17 @@ def prepare_cleanup() -> MacosContainerBottleCleanupPlan:
 
 
 def cleanup(plan: MacosContainerBottleCleanupPlan) -> None:
+    failures = CleanupFailures()
     for name in plan.containers:
         info(f"container delete --force {name}")
-        subprocess.run(
+        failures.run(
             ["container", "delete", "--force", name],
-            stdout=subprocess.DEVNULL,
-            stderr=subprocess.DEVNULL,
-            check=False,
+            f"deleting container {name}",
         )
     for name in plan.networks:
         info(f"container network delete {name}")
-        subprocess.run(
+        failures.run(
             ["container", "network", "delete", name],
-            stdout=subprocess.DEVNULL,
-            stderr=subprocess.DEVNULL,
-            check=False,
+            f"deleting network {name}",
         )
+    failures.raise_if_any()

bot_bottle/backend/macos_container/enumerate.py

@@ -5,7 +5,7 @@ from __future__ import annotations
 import subprocess
 
 from ...bottle_state import read_metadata
-from .. import ActiveAgent
+from .. import ActiveAgent, EnumerationError
 from .infra import INFRA_NAME, ORCHESTRATOR_NAME
 
 # The name every agent container carries: `bot-bottle-<slug>`. Exported
@@ -20,17 +20,18 @@ CONTAINER_NAME_PREFIX = "bot-bottle-"
 _INFRA_NAMES = frozenset({INFRA_NAME, ORCHESTRATOR_NAME})
 
 
-class EnumerationError(RuntimeError):
-    """container list failed; the resulting live set is not authoritative."""
-
-
 def enumerate_active() -> list[ActiveAgent]:
-    result = subprocess.run(
-        ["container", "list", "--quiet"],
-        capture_output=True,
-        text=True,
-        check=False,
-    )
+    try:
+        result = subprocess.run(
+            ["container", "list", "--quiet"],
+            capture_output=True,
+            text=True,
+            check=False,
+        )
+    except FileNotFoundError as exc:
+        raise EnumerationError(
+            "container list failed: container CLI not found"
+        ) from exc
     if result.returncode != 0:
         raise EnumerationError(
             f"container list failed: "

bot_bottle/backend/macos_container/gateway.py

@@ -84,6 +84,7 @@ class MacosGateway(Gateway):
         egress_network: str = GATEWAY_EGRESS_NETWORK,
         control_network: str = CONTROL_NETWORK,
         repo_root: Path | None = None,
+        local_build: bool = True,
     ) -> None:
         self.image_ref = image_ref
         self.name = name
@@ -93,6 +94,7 @@ class MacosGateway(Gateway):
         # Build context: the repo root in a checkout, a staged copy from the
         # installed wheel otherwise (bot_bottle.resources).
         self._repo_root = repo_root if repo_root is not None else resources.build_root()
+        self._local_build = local_build
         # Set by `connect_to_orchestrator`: the URL the daemons resolve policy
         # against + the pre-minted `gateway` token they present. The gateway
         # never mints, so it never holds the signing key (#469).
@@ -101,8 +103,11 @@ class MacosGateway(Gateway):
 
     def ensure_built(self) -> None:
         """Build the data-plane image from `Dockerfile.gateway`."""
-        container_mod.build_image(
-            self.image_ref, str(self._repo_root), dockerfile="Dockerfile.gateway")
+        if self._local_build:
+            container_mod.build_image(
+                self.image_ref, str(self._repo_root), dockerfile="Dockerfile.gateway")
+        else:
+            container_mod.pull_image(self.image_ref)
 
     def connect_to_orchestrator(self, orchestrator_url: str, gateway_token: str) -> None:
         """Bind the gateway to this orchestrator and (re)start it, dual-homed on

bot_bottle/backend/macos_container/infra.py

@@ -25,6 +25,7 @@ from __future__ import annotations
 from pathlib import Path
 
 from ... import resources
+from ... import release_manifest
 from ...orchestrator.lifecycle import (
     DEFAULT_PORT,
     DEFAULT_STARTUP_TIMEOUT_SECONDS,
@@ -86,6 +87,14 @@ class MacosInfraService(InfraService):
         self._orchestrator_name = orchestrator_name
         self._gateway_name = gateway_name
         self._db_volume = db_volume
+        resolved_orchestrator, orchestrator_local = release_manifest.oci_image(
+            "orchestrator", orchestrator_image)
+        resolved_gateway, gateway_local = release_manifest.oci_image(
+            "gateway", gateway_image)
+        self.orchestrator_image = resolved_orchestrator
+        self.gateway_image = resolved_gateway
+        self._orchestrator_local = orchestrator_local
+        self._gateway_local = gateway_local
 
     def orchestrator(self) -> MacosOrchestrator:
         """The control-plane service on the host-only control network. Cheap to
@@ -98,6 +107,7 @@ class MacosInfraService(InfraService):
             control_network=self.control_network,
             repo_root=self._repo_root,
             db_volume=self._db_volume,
+            local_build=self._orchestrator_local,
         )
 
     def gateway(self) -> MacosGateway:
@@ -112,6 +122,7 @@ class MacosInfraService(InfraService):
             egress_network=self.egress_network,
             control_network=self.control_network,
             repo_root=self._repo_root,
+            local_build=self._gateway_local,
         )
 
     def ensure_running(
@@ -127,8 +138,8 @@ class MacosInfraService(InfraService):
 
         orchestrator = self.orchestrator()
         gateway = self.gateway()
-        orchestrator.ensure_built()
-        gateway.ensure_built()
+        orchestrator.ensure_available()
+        gateway.ensure_available()
 
         orchestrator.ensure_running(startup_timeout=startup_timeout)
 

bot_bottle/backend/macos_container/orchestrator.py

@@ -63,6 +63,7 @@ class MacosOrchestrator(Orchestrator):
         control_network: str = CONTROL_NETWORK,
         repo_root: Path | None = None,
         db_volume: str = ORCHESTRATOR_DB_VOLUME,
+        local_build: bool = True,
     ) -> None:
         self.image_ref = image_ref
         self.name = name
@@ -73,6 +74,7 @@ class MacosOrchestrator(Orchestrator):
         # staged copy from the installed wheel otherwise (bot_bottle.resources).
         self._repo_root = repo_root if repo_root is not None else resources.build_root()
         self._db_volume = db_volume
+        self._local_build = local_build
 
     def url(self) -> str:
         """The orchestrator's control-network address (host CLI + registration),
@@ -114,8 +116,12 @@ class MacosOrchestrator(Orchestrator):
         """Build the control-plane image. The source is bind-mounted so a code
         change takes effect without a rebuild; the image still carries the
         package for its entrypoint."""
-        container_mod.build_image(
-            self.image_ref, str(self._repo_root), dockerfile="Dockerfile.orchestrator")
+        if self._local_build:
+            container_mod.build_image(
+                self.image_ref, str(self._repo_root),
+                dockerfile="Dockerfile.orchestrator")
+        else:
+            container_mod.pull_image(self.image_ref)
 
     def ensure_running(
         self, *, startup_timeout: float = DEFAULT_STARTUP_TIMEOUT_SECONDS,
@@ -146,11 +152,6 @@ class MacosOrchestrator(Orchestrator):
             "--dns", container_mod.dns_server(),
             # Container-only DB volume: exactly one kernel writes bot-bottle.db.
             "--volume", f"{self._db_volume}:{_DB_ROOT_IN_CONTAINER}",
-            # Live control-plane source (a code change takes effect on relaunch).
-            "--mount",
-            container_mod.bind_mount_spec(
-                str(self._repo_root), _SRC_IN_CONTAINER, readonly=True),
-            "--env", f"PYTHONPATH={_SRC_IN_CONTAINER}",
             "--env", f"BOT_BOTTLE_ROOT={_DB_ROOT_IN_CONTAINER}",
             # Detect a real control-plane code change and recreate.
             "--env", f"BOT_BOTTLE_SOURCE_HASH={current_hash}",
@@ -161,6 +162,14 @@ class MacosOrchestrator(Orchestrator):
             # Dockerfile.orchestrator ENTRYPOINT is `-m bot_bottle.orchestrator`.
             "--host", "0.0.0.0", "--port", str(self.port), "--broker", "stub",
         ]
+        if self._local_build:
+            image_index = argv.index(self.image_ref)
+            argv[image_index:image_index] = [
+                "--mount",
+                container_mod.bind_mount_spec(
+                    str(self._repo_root), _SRC_IN_CONTAINER, readonly=True),
+                "--env", f"PYTHONPATH={_SRC_IN_CONTAINER}",
+            ]
         result = container_mod.run_container_argv(
             argv, env={**os.environ, ORCHESTRATOR_TOKEN_ENV: _signing_key})
         if result.returncode != 0:

bot_bottle/backend/macos_container/setup.py

@@ -5,7 +5,7 @@ Like Docker, this backend needs no privileged network-pool provisioning
 running. `setup()` points at the install/`container system start` steps;
 `status()` reports readiness. Reached via
 `MacosContainerBottleBackend.setup` / `.status`, dispatched from the
-generic `./cli.py backend {setup,status}`.
+generic `bot-bottle backend {setup,status}`.
 """
 
 from __future__ import annotations
@@ -75,5 +75,5 @@ def status() -> int:
         if not _service_running():
             ok = False
     if not ok:
-        sys.stderr.write("\nRun: ./cli.py backend setup --backend=macos-container\n")
+        sys.stderr.write("\nRun: bot-bottle backend setup --backend=macos-container\n")
     return 0 if ok else 1

bot_bottle/backend/macos_container/util.py

@@ -101,6 +101,19 @@ def build_image(
     subprocess.run(args, check=True)
 
 
+def pull_image(ref: str) -> None:
+    """Acquire an immutable OCI reference through Apple Container."""
+    result = subprocess.run(
+        [_CONTAINER, "image", "pull", ref],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        detail = (result.stderr or result.stdout or "").strip()
+        raise RuntimeError(f"container image pull failed for {ref}: {detail}")
+
+
 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
@@ -689,6 +702,13 @@ def pinned_local_image_ref(ref: str) -> str:
     nested-containers layer. A content-derived tag prevents another concurrent
     build from moving the provider's ordinary ``:latest`` tag between those
     two builds.
+
+    Unlike Docker, `container image tag` only accepts ``image-name[:tag]`` as
+    its source and rejects a bare image ID ("cannot specify 64 byte hex string
+    as reference"), so the mutable ``ref`` is what gets tagged here. The
+    post-tag inspection below is what keeps that fail-closed: if ``ref`` moved
+    between the two commands, the new tag will not resolve to the ID this call
+    derived its name from.
     """
     image = image_id(ref)
     digest = image.removeprefix("sha256:")
@@ -701,14 +721,14 @@ def pinned_local_image_ref(ref: str) -> str:
         repository = repository[:last_colon]
     pinned_ref = f"{repository}:sha256-{digest}"
     result = subprocess.run(
-        [_CONTAINER, "image", "tag", image, pinned_ref],
+        [_CONTAINER, "image", "tag", ref, pinned_ref],
         capture_output=True,
         text=True,
         check=False,
     )
     if result.returncode != 0:
         die(
-            f"could not tag exact local image {image}: "
+            f"could not tag exact local image {image} via {ref!r}: "
             f"{(result.stderr or result.stdout or '').strip() or '<no detail>'}"
         )
     if image_id(pinned_ref) != image:

bot_bottle/backend/provision_gateway.py

@@ -63,12 +63,23 @@ def provision_git_gate(
     transport.exec(["chmod", "+x", "/etc/git-gate/access-hook"])
     creds = _creds_dir(bottle_id)
     transport.exec(["mkdir", "-p", creds])
+    transport.exec(["chmod", "700", creds])
+    credential_paths: list[str] = []
     for u in plan.upstreams:
         if u.identity_file:
-            transport.cp_into(u.identity_file, f"{creds}/{u.name}-key")
+            key_path = f"{creds}/{u.name}-key"
+            transport.cp_into(u.identity_file, key_path)
+            credential_paths.append(key_path)
         known_hosts = str(u.known_hosts_file)
         if known_hosts and known_hosts != ".":
-            transport.cp_into(known_hosts, f"{creds}/{u.name}-known_hosts")
+            known_hosts_path = f"{creds}/{u.name}-known_hosts"
+            transport.cp_into(known_hosts, known_hosts_path)
+            credential_paths.append(known_hosts_path)
+    # Copy-mode behavior differs across Docker, Apple Container, and SSH.
+    # Apply the security contract inside the gateway so every backend produces
+    # the same private credential namespace.
+    if credential_paths:
+        transport.exec(["chmod", "600", *credential_paths])
     # Init the bare repos + per-repo credential config for this namespace.
     script = git_gate_render_provision(bottle_id, plan.upstreams)
     transport.exec(["sh", "-c", script])

bot_bottle/backend/selection.py

@@ -86,7 +86,7 @@ def _print_vm_install_instructions() -> None:
         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")
+        info("Configure the host:  bot-bottle backend setup")
 
 
 def _auto_select_backend(prompt: bool = True) -> str:

bot_bottle/cli/commands/backend.py

@@ -1,9 +1,9 @@
 """`backend` CLI command — generic host setup/status across backends.
 
-`./cli.py backend setup    [--backend=NAME]` provisions (or points at how
+`bot-bottle backend setup    [--backend=NAME]` provisions (or points at how
 to provision) the chosen backend's one-time host prerequisites.
-`./cli.py backend status   [--backend=NAME]` reports readiness.
-`./cli.py backend teardown [--backend=NAME]` undoes setup (uninstall).
+`bot-bottle backend status   [--backend=NAME]` reports readiness.
+`bot-bottle backend teardown [--backend=NAME]` undoes setup (uninstall).
 
 All dispatch to the backend's `setup()` / `status()` / `teardown()`
 classmethods, so there are no backend-specific commands — swapping

bot_bottle/cli/commands/cleanup.py

@@ -1,7 +1,7 @@
 """cleanup: stop and remove all orphaned bot-bottle resources.
 
 Walks every registered backend (docker, firecracker, macos-container)
-so a single `./cli.py cleanup` reaps every backend's leftovers — a
+so a single `bot-bottle cleanup` reaps every backend's leftovers — a
 firecracker bottle's VM processes and run dirs won't survive a
 docker-only cleanup pass (issue addressed alongside #77).
 
@@ -22,6 +22,7 @@ from __future__ import annotations
 import sys
 
 from ...backend import get_bottle_backend, has_backend, known_backend_names
+from ...backend.cleanup_control import CleanupError
 from ...log import info
 from ...util import read_tty_line
 
@@ -52,10 +53,20 @@ def cmd_cleanup(_argv: list[str]) -> int:
         info("cleanup: skipped")
         return 0
 
-    for name, backend, plan in prepared:
-        if plan.empty:
+    # Confirmation authorizes a fresh authoritative snapshot, not blind use of
+    # identities that may have changed while the operator reviewed the preview.
+    failures: list[str] = []
+    for name, backend, displayed in prepared:
+        current = backend.prepare_cleanup()
+        approved = displayed.intersect(current)
+        if approved.empty:
             continue
-        backend.cleanup(plan)
+        try:
+            backend.cleanup(approved)
+        except CleanupError as exc:
+            failures.append(f"{name}: {exc}")
+    if failures:
+        raise CleanupError("cleanup incomplete: " + "; ".join(failures))
     info("cleanup: done")
     return 0
 

bot_bottle/cli/commands/commit.py

@@ -4,7 +4,7 @@ Docker bottles are committed to a local Docker image. Macos-container
 bottles are exported and rebuilt as a local Apple Container image.
 Firecracker bottles stream the guest rootfs out over SSH and rebuild a
 local Docker image. The resulting reference is stored in per-bottle
-state so the next `./cli.py resume <slug>` boots from the snapshot
+state so the next `bot-bottle resume <slug>` boots from the snapshot
 instead of rebuilding from the Dockerfile.
 """
 
@@ -37,7 +37,7 @@ def cmd_commit(argv: list[str]) -> int:
     if slug is None:
         active = enumerate_active_agents()
         if not active:
-            die("no active bottles; start one with `./cli.py start`")
+            die("no active bottles; start one with `bot-bottle start`")
         choices = [a.slug for a in active]
         slug = tui.filter_select(choices, title="Select bottle to commit")
         if slug is None:

bot_bottle/cli/commands/resume.py

@@ -8,7 +8,7 @@ override and transcript snapshot under the same state dir.
 
 Use case: an interrupted or preserved bottle needs to be relaunched;
 the operator runs
-    ./cli.py resume <identity>
+    bot-bottle resume <identity>
 to bring up the replacement from the recorded state.
 """
 

bot_bottle/cli/commands/start.py

@@ -203,19 +203,19 @@ def _start_headless(
     if not os.isatty(stdin_fd):
         die(
             "--headless requires a PTY on stdin; run via:\n"
-            "  script -q /dev/null ./cli.py start ..."
+            "  script -q /dev/null bot-bottle start ..."
         )
 
     agent_name = args.name
     if not agent_name:
-        die("--headless requires an agent name: ./cli.py start <agent> --headless")
+        die("--headless requires an agent name: bot-bottle start <agent> --headless")
     manifest.require_agent(agent_name)  # raises ManifestError if unknown
 
     prompt = args.prompt
     if not prompt:
         die(
             "--headless requires --prompt: "
-            "./cli.py start <agent> --headless --prompt 'Do the thing'"
+            "bot-bottle start <agent> --headless --prompt 'Do the thing'"
         )
 
     if args.bottle:
@@ -319,9 +319,9 @@ def attach_agent(
 
     `resume=True` adds `--continue` so claude picks up its most
     recent session non-interactively (no session-picker prompt).
-    First-attach paths (`./cli.py start`) leave it False.
+    First-attach paths (`bot-bottle start`) leave it False.
 
-    Used as the inner step of `./cli.py start`."""
+    Used as the inner step of `bot-bottle start`."""
     runtime = runtime_for(agent_provider_template)
     info(
         f"attaching interactive {agent_provider_template} session "
@@ -354,7 +354,7 @@ def settle_state(identity: str) -> None:
     if not identity:
         return
     if is_preserved(identity):
-        info(f"to resume this bottle: ./cli.py resume {identity}")
+        info(f"to resume this bottle: bot-bottle resume {identity}")
         return
     cleanup_state(identity)
 

bot_bottle/cli/commands/supervise.py

@@ -110,7 +110,7 @@ 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.label}]"
-    return f"{base}; resume: ./cli.py resume {qp.label}"
+    return f"{base}; resume: bot-bottle resume {qp.label}"
 
 
 def _detail_lines(

bot_bottle/gateway/__init__.py

@@ -115,10 +115,12 @@ class Gateway(abc.ABC):
 
     name: str
 
+    def ensure_available(self) -> None:
+        """Acquire the exact image/rootfs selected for this application."""
+        self.ensure_built()
+
     def ensure_built(self) -> None:
-        """Ensure the gateway's image / rootfs exists, building it if needed.
-        Default: nothing to build (e.g. a stub or a pre-pulled image). Call
-        before `connect_to_orchestrator`."""
+        """Local-build implementation hook retained for backend compatibility."""
         return
 
     @abc.abstractmethod

bot_bottle/gateway/bootstrap.py

@@ -136,10 +136,18 @@ def _pump(name: str, stream: IO[bytes]) -> None:
     """Read lines from `stream`, prefix with `[name]`, write to
     stdout. Runs in its own thread per child; daemon=True so a
     blocked read doesn't keep the process alive after main exits."""
-    for raw in iter(stream.readline, b""):
-        line = raw.decode("utf-8", errors="replace").rstrip("\n")
-        sys.stdout.write(f"[{name}] {line}\n")
-        sys.stdout.flush()
+    try:
+        for raw in iter(stream.readline, b""):
+            line = raw.decode("utf-8", errors="replace").rstrip("\n")
+            sys.stdout.write(f"[{name}] {line}\n")
+            sys.stdout.flush()
+    except (OSError, ValueError) as exc:
+        # The manager closes a dead child's pipe after wait() and before a
+        # restart. A pump can be between readline calls at that exact moment;
+        # closed-stream errors are normal completion, not uncaught thread
+        # failures. Preserve genuinely unexpected I/O diagnostics.
+        if not stream.closed:
+            _log(f"{name} output pump stopped: {type(exc).__name__}: {exc}")
 
 
 def _spawn(spec: _DaemonSpec) -> subprocess.Popen[bytes]:

bot_bottle/gateway/bounded_http.py

@@ -0,0 +1,137 @@
+"""Shared resource boundaries for gateway stdlib HTTP services."""
+
+from __future__ import annotations
+
+import http.server
+import io
+import socket
+import threading
+import time
+from dataclasses import dataclass
+from typing import Any, Protocol
+
+
+class Readable(Protocol):
+    def read(self, size: int = -1, /) -> bytes: ...
+
+
+class Writable(Protocol):
+    def write(self, data: bytes, /) -> object: ...
+
+
+@dataclass(frozen=True)
+class BodyReadError(Exception):
+    status: int
+    message: str
+
+
+def read_declared_body(
+    stream: Readable,
+    connection: socket.socket,
+    raw_length: str | None,
+    *,
+    maximum: int,
+    timeout_seconds: float,
+    require_length: bool,
+) -> bytes:
+    """Validate and read exactly one declared body under a read deadline."""
+    output = io.BytesIO()
+    copy_declared_body(
+        stream, output, connection, raw_length, maximum=maximum,
+        timeout_seconds=timeout_seconds, require_length=require_length,
+    )
+    return output.getvalue()
+
+
+def copy_declared_body(
+    stream: Readable,
+    output: Writable,
+    connection: socket.socket,
+    raw_length: str | None,
+    *,
+    maximum: int,
+    timeout_seconds: float,
+    require_length: bool,
+) -> int:
+    """Copy one declared body to a sink without retaining it in memory."""
+    if raw_length is None:
+        if require_length:
+            raise BodyReadError(411, "Content-Length required")
+        raw_length = "0"
+    try:
+        length = int(raw_length)
+    except ValueError as exc:
+        raise BodyReadError(400, "invalid Content-Length") from exc
+    if length < 0:
+        raise BodyReadError(400, "invalid Content-Length")
+    if length > maximum:
+        raise BodyReadError(413, "request body too large")
+    previous_timeout = connection.gettimeout()
+    deadline = time.monotonic() + timeout_seconds
+    remaining = length
+    try:
+        while remaining:
+            timeout = deadline - time.monotonic()
+            if timeout <= 0:
+                raise BodyReadError(408, "request body read timed out")
+            connection.settimeout(timeout)
+            chunk = stream.read(min(remaining, 64 * 1024))
+            if not chunk:
+                raise BodyReadError(400, "incomplete request body")
+            output.write(chunk)
+            remaining -= len(chunk)
+    except TimeoutError as exc:
+        raise BodyReadError(408, "request body read timed out") from exc
+    finally:
+        connection.settimeout(previous_timeout)
+    return length
+
+
+class BoundedThreadingHTTPServer(http.server.ThreadingHTTPServer):
+    """ThreadingHTTPServer with a hard cap on in-flight request threads."""
+
+    daemon_threads = True
+
+    def __init__(  # pylint: disable=consider-using-with
+        self, *args, max_workers: int = 32, **kwargs,  # type: ignore[no-untyped-def]
+    ):
+        if max_workers < 1:
+            raise ValueError("max_workers must be positive")
+        self._request_slots = threading.BoundedSemaphore(max_workers)
+        super().__init__(*args, **kwargs)
+
+    def process_request(
+        self, request: Any, client_address: Any,
+    ) -> None:
+        if not self._request_slots.acquire(  # pylint: disable=consider-using-with
+            blocking=False,
+        ):
+            try:
+                request.sendall(
+                    b"HTTP/1.1 503 Service Unavailable\r\n"
+                    b"Content-Length: 0\r\nConnection: close\r\n\r\n"
+                )
+            finally:
+                self.shutdown_request(request)
+            return
+        try:
+            super().process_request(request, client_address)
+        except BaseException:
+            self._request_slots.release()
+            raise
+
+    def process_request_thread(
+        self, request: Any, client_address: Any,
+    ) -> None:
+        try:
+            super().process_request_thread(request, client_address)
+        finally:
+            self._request_slots.release()
+
+
+__all__ = [
+    "BodyReadError",
+    "BoundedThreadingHTTPServer",
+    "copy_declared_body",
+    "read_declared_body",
+]

bot_bottle/gateway/egress/addon.py

@@ -16,7 +16,7 @@ import typing
 from mitmproxy import http  # type: ignore[import-not-found]  # pylint: disable=import-error
 
 from bot_bottle.constants import IDENTITY_HEADER
-from bot_bottle.gateway.egress.dlp_detectors import redact_tokens, strip_crlf
+from bot_bottle.gateway.egress.dlp_detectors import redact_tokens
 from bot_bottle.gateway.egress.dlp_config import (
     DEFAULT_OUTBOUND_ON_MATCH,
     ON_MATCH_BLOCK,
@@ -25,19 +25,19 @@ from bot_bottle.gateway.egress.dlp_config import (
 from bot_bottle.gateway.egress.context import resolve_client_context
 from bot_bottle.gateway.egress.dlp import (
     build_inbound_scan_text,
-    build_outbound_scan_text,
     build_token_allow_payload,
-    outbound_scan_headers,
     scan_inbound,
     scan_outbound,
 )
+from bot_bottle.gateway.egress.outbound_pipeline import redact_request, scan_request
 from bot_bottle.gateway.egress.matching import (
     decide,
-    decide_git_fetch,
-    is_git_fetch_request,
-    is_git_push_request,
     match_route,
 )
+from bot_bottle.gateway.egress.request_pipeline import (
+    evaluate_route_policy,
+    git_block_reason,
+)
 from bot_bottle.gateway.egress.schema import route_to_yaml_dict
 from bot_bottle.gateway.egress.types import (
     LOG_BLOCKS,
@@ -389,19 +389,9 @@ class EgressAddon:
         self._passthrough_conns.discard(conn_id)
 
     async def request(self, flow: http.HTTPFlow) -> None:
+        config, slug, env = self._request_context(flow)
         request_path, _, query = flow.request.path.partition("?")
 
-        # Reuse the context stashed by http_connect for HTTPS flows (one
-        # orchestrator round-trip per connection). Plain-HTTP flows have no
-        # prior CONNECT stash, so resolve now and stash for response/websocket.
-        meta = getattr(flow, "metadata", None)
-        if isinstance(meta, dict) and _FLOW_CTX_KEY in meta:
-            config, slug, env = meta[_FLOW_CTX_KEY]
-            self._request_token(flow)  # strip identity headers; token already resolved
-        else:
-            config, slug, env = self._resolve_flow(flow)
-            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.
@@ -422,56 +412,66 @@ class EgressAddon:
             # the path/query the git checks below rely on.
             request_path, _, query = flow.request.path.partition("?")
 
-        if is_git_push_request(request_path, query):
-            self._block(
-                flow,
-                "egress: git push over HTTPS is not supported; "
-                "use the bottle.git SSH path (gitleaks-scanned by "
-                "git-gate's pre-receive hook).",
-                ctx=self._req_ctx(flow),
-            )
+        if not self._allow_git_request(flow, config, request_path, query):
             return
 
-        if is_git_fetch_request(request_path, query):
-            git_decision = decide_git_fetch(
-                config.routes, flow.request.pretty_host,
-            )
-            if git_decision.action == "block":
-                self._block(
-                    flow,
-                    git_decision.reason,
-                    ctx=self._req_ctx(flow),
-                )
-                return
+        self._apply_route_policy(flow, config, route, request_path, env)
 
+    def _request_context(
+        self, flow: http.HTTPFlow,
+    ) -> tuple[Config, str, "typing.Mapping[str, str]"]:
+        """Resolve one bottle context, reusing the HTTPS CONNECT snapshot."""
+        meta = getattr(flow, "metadata", None)
+        if isinstance(meta, dict) and _FLOW_CTX_KEY in meta:
+            config, slug, env = meta[_FLOW_CTX_KEY]
+            self._request_token(flow)
+            return config, slug, env
+        config, slug, env = self._resolve_flow(flow)
+        self._stash_flow_ctx(flow, config, slug, env)
+        return config, slug, env
+
+    def _allow_git_request(
+        self, flow: http.HTTPFlow, config: Config,
+        request_path: str, query: str,
+    ) -> bool:
+        """Apply the HTTPS Git push/fetch boundary before general routing."""
+        reason = git_block_reason(
+            config.routes, flow.request.pretty_host, request_path, query,
+        )
+        if not reason:
+            return True
+        self._block(flow, reason, ctx=self._req_ctx(flow))
+        return False
+
+    def _apply_route_policy(
+        self, flow: http.HTTPFlow, config: Config, route: Route | None,
+        request_path: str, env: "typing.Mapping[str, str]",
+    ) -> None:
+        """Strip agent auth, evaluate the route, then inject gateway auth."""
         # Strip agent-set Authorization after DLP scan so smuggled tokens
         # are caught above; the route may inject gateway-owned auth below.
         # Routes with preserve_auth=True pass the header through as-is so the
         # agent's own credentials (e.g. registry bearer tokens) reach the upstream.
-        if route is None or not route.preserve_auth:
+        result = evaluate_route_policy(
+            config,
+            route,
+            host=flow.request.pretty_host,
+            request_path=request_path,
+            method=flow.request.method,
+            headers=dict(flow.request.headers),
+            env=env,
+        )
+        if result.strip_authorization:
             flow.request.headers.pop("authorization", None)
 
-        # Build headers mapping for match evaluation
-        req_headers = {k.lower(): v for k, v in flow.request.headers.items()}
-
-        decision = decide(
-            config.routes,
-            flow.request.pretty_host,
-            request_path,
-            env,
-            request_method=flow.request.method,
-            request_headers=req_headers,
-            deny_reason=config.deny_reason,
-        )
-
-        if decision.action == "block":
-            self._block(flow, decision.reason, ctx=self._req_ctx(flow))
+        if result.block_reason:
+            self._block(flow, result.block_reason, ctx=self._req_ctx(flow))
             return
 
-        if decision.inject_authorization is not None:
-            flow.request.headers["authorization"] = decision.inject_authorization
+        if result.inject_authorization is not None:
+            flow.request.headers["authorization"] = result.inject_authorization
 
-        if config.log >= LOG_FULL:
+        if result.log_request:
             self._log_request(flow, env)
 
     def _block_dlp(self, flow: http.HTTPFlow, result: ScanResult) -> None:
@@ -495,20 +495,12 @@ class EgressAddon:
         Loops so the supervise policy can re-scan after each approval — a
         second, un-approved token in the same request is still caught."""
         while True:
-            request_path, _, query = flow.request.path.partition("?")
-            body = flow.request.get_text(strict=False) or ""
-            headers = outbound_scan_headers(route, dict(flow.request.headers))
-            scan_text = build_outbound_scan_text(
-                flow.request.pretty_host, request_path, query, headers, body,
-            )
-            # CRLF is scanned only over the request line + headers, never the
-            # body (see scan_outbound) — a body is not an injection vector.
-            crlf_text = build_outbound_scan_text(
-                flow.request.pretty_host, request_path, query, headers, "",
-            )
-            result = scan_outbound(
-                route, scan_text, env,
-                safe_tokens=self._safe_tokens_for(slug), crlf_text=crlf_text,
+            request_path, _, _ = flow.request.path.partition("?")
+            result = scan_request(
+                flow.request,
+                route,
+                env,
+                safe_tokens=self._safe_tokens_for(slug),
             )
             if result is None or result.severity != "block":
                 return True
@@ -518,7 +510,7 @@ class EgressAddon:
             # redact scrubs every detection (tokens and structural CRLF) and
             # forwards; it fails closed only if a match survives the scrub.
             if policy == ON_MATCH_REDACT:
-                if self._redact_outbound(flow, route, env):
+                if redact_request(flow.request, route, env):
                     if self._flow_log(flow) >= LOG_BLOCKS:
                         sys.stderr.write(json.dumps({
                             "event": "egress_redacted",
@@ -551,41 +543,6 @@ class EgressAddon:
                 return False  # _supervise_token_block wrote the 403 response
             # loop: the approved value is now in safe_tokens; re-scan.
 
-    def _redact_outbound(
-        self, flow: http.HTTPFlow, route: Route, env: "typing.Mapping[str, str]",
-    ) -> bool:
-        """Scrub detected tokens (and CRLF injection sequences) from the mutable
-        request surfaces (body, headers, path/query) and re-scan. `env` is the
-        per-bottle env overlay. Returns True if the request is now clean; False
-        if a block-severity match remains on a surface redaction cannot rewrite
-        (the hostname) so the caller fails closed."""
-        body = flow.request.get_text(strict=False)
-        if body:
-            redacted_body = redact_tokens(body, env=env)
-            if redacted_body != body:
-                flow.request.text = redacted_body
-        for name, value in list(flow.request.headers.items()):
-            if name.lower() == "host":
-                continue  # routing-critical; never a legitimate token
-            redacted = strip_crlf(redact_tokens(value, env=env))
-            if redacted != value:
-                flow.request.headers[name] = redacted
-        redacted_path = strip_crlf(redact_tokens(flow.request.path, env=env))
-        if redacted_path != flow.request.path:
-            flow.request.path = redacted_path
-
-        request_path, _, query = flow.request.path.partition("?")
-        new_body = flow.request.get_text(strict=False) or ""
-        headers = outbound_scan_headers(route, dict(flow.request.headers))
-        scan_text = build_outbound_scan_text(
-            flow.request.pretty_host, request_path, query, headers, new_body,
-        )
-        crlf_text = build_outbound_scan_text(
-            flow.request.pretty_host, request_path, query, headers, "",
-        )
-        result = scan_outbound(route, scan_text, env, crlf_text=crlf_text)
-        return result is None or result.severity != "block"
-
     async def _supervise_token_block(
         self,
         flow: http.HTTPFlow,

bot_bottle/gateway/egress/outbound_pipeline.py

@@ -0,0 +1,83 @@
+"""Outbound DLP request scanning and redaction for the egress pipeline."""
+
+from __future__ import annotations
+
+from typing import ItemsView, Mapping, Protocol
+
+from .dlp import (
+    build_outbound_scan_text,
+    outbound_scan_headers,
+    scan_outbound,
+)
+from .dlp_detectors import redact_tokens, strip_crlf
+from .types import Route, ScanResult
+
+
+class MutableHeaders(Protocol):
+    def items(self) -> ItemsView[str, str]: ...
+    def __getitem__(self, name: str, /) -> str: ...
+    def __setitem__(self, name: str, value: str, /) -> None: ...
+
+
+class MutableRequest(Protocol):
+    pretty_host: str
+    path: str
+    headers: MutableHeaders
+    text: str
+
+    def get_text(self, strict: bool = False) -> str | None: ...
+
+
+def scan_request(
+    request: MutableRequest,
+    route: Route,
+    env: Mapping[str, str],
+    *,
+    safe_tokens: set[str] | None = None,
+) -> ScanResult | None:
+    """Scan all mutable outbound request surfaces in their canonical order."""
+    request_path, _, query = request.path.partition("?")
+    headers = outbound_scan_headers(route, dict(request.headers.items()))
+    body = request.get_text(strict=False) or ""
+    scan_text = build_outbound_scan_text(
+        request.pretty_host, request_path, query, headers, body,
+    )
+    # Bodies cannot alter HTTP framing, so CRLF detection is deliberately
+    # restricted to the request line and headers.
+    crlf_text = build_outbound_scan_text(
+        request.pretty_host, request_path, query, headers, "",
+    )
+    return scan_outbound(
+        route,
+        scan_text,
+        env,
+        safe_tokens=safe_tokens,
+        crlf_text=crlf_text,
+    )
+
+
+def redact_request(
+    request: MutableRequest,
+    route: Route,
+    env: Mapping[str, str],
+) -> bool:
+    """Redact mutable request surfaces and return whether the result is clean."""
+    body = request.get_text(strict=False)
+    if body:
+        redacted_body = redact_tokens(body, env=env)
+        if redacted_body != body:
+            request.text = redacted_body
+    for name, value in list(request.headers.items()):
+        if name.lower() == "host":
+            continue
+        redacted = strip_crlf(redact_tokens(value, env=env))
+        if redacted != value:
+            request.headers[name] = redacted
+    redacted_path = strip_crlf(redact_tokens(request.path, env=env))
+    if redacted_path != request.path:
+        request.path = redacted_path
+    result = scan_request(request, route, env)
+    return result is None or result.severity != "block"
+
+
+__all__ = ["MutableRequest", "redact_request", "scan_request"]

bot_bottle/gateway/egress/request_pipeline.py

@@ -0,0 +1,92 @@
+"""Framework-neutral request policy stages for the egress adapter.
+
+The mitmproxy addon owns flow mutation and response construction. This module
+owns the ordered Git and route-policy decisions so those rules remain directly
+testable without a live proxy flow.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Mapping, Sequence
+
+from .matching import (
+    decide,
+    decide_git_fetch,
+    is_git_fetch_request,
+    is_git_push_request,
+)
+from .types import LOG_FULL, Config, Route
+
+GIT_PUSH_BLOCK_REASON = (
+    "egress: git push over HTTPS is not supported; "
+    "use the bottle.git SSH path (gitleaks-scanned by "
+    "git-gate's pre-receive hook)."
+)
+
+
+@dataclass(frozen=True)
+class RoutePolicyResult:
+    """The flow mutations and outcome produced by general route policy."""
+
+    block_reason: str = ""
+    strip_authorization: bool = False
+    inject_authorization: str | None = None
+    log_request: bool = False
+
+
+def git_block_reason(
+    routes: Sequence[Route],
+    host: str,
+    request_path: str,
+    query: str,
+) -> str:
+    """Return the HTTPS Git policy denial, or ``""`` when allowed."""
+    if is_git_push_request(request_path, query):
+        return GIT_PUSH_BLOCK_REASON
+    if not is_git_fetch_request(request_path, query):
+        return ""
+    decision = decide_git_fetch(routes, host)
+    return decision.reason if decision.action == "block" else ""
+
+
+def evaluate_route_policy(
+    config: Config,
+    route: Route | None,
+    *,
+    host: str,
+    request_path: str,
+    method: str,
+    headers: Mapping[str, str],
+    env: Mapping[str, str],
+) -> RoutePolicyResult:
+    """Evaluate authorization stripping, matching, injection, and logging."""
+    strip_authorization = route is None or not route.preserve_auth
+    effective_headers = {
+        name.lower(): value
+        for name, value in headers.items()
+        if not (strip_authorization and name.lower() == "authorization")
+    }
+    decision = decide(
+        config.routes,
+        host,
+        request_path,
+        env,
+        request_method=method,
+        request_headers=effective_headers,
+        deny_reason=config.deny_reason,
+    )
+    return RoutePolicyResult(
+        block_reason=decision.reason if decision.action == "block" else "",
+        strip_authorization=strip_authorization,
+        inject_authorization=decision.inject_authorization,
+        log_request=config.log >= LOG_FULL,
+    )
+
+
+__all__ = [
+    "GIT_PUSH_BLOCK_REASON",
+    "RoutePolicyResult",
+    "evaluate_route_policy",
+    "git_block_reason",
+]

bot_bottle/gateway/git_gate/http_backend.py

@@ -21,12 +21,19 @@ from __future__ import annotations
 import os
 import subprocess
 import sys
+import tempfile
+import threading
 import typing
-from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from http.server import BaseHTTPRequestHandler
 from pathlib import Path
 from urllib.parse import urlsplit
 
 from bot_bottle.constants import GIT_GATE_TIMEOUT_SECS, IDENTITY_HEADER
+from bot_bottle.gateway.bounded_http import (
+    BodyReadError,
+    BoundedThreadingHTTPServer,
+    copy_declared_body,
+)
 from bot_bottle.gateway.policy_resolver import PolicyResolveError, PolicyResolver
 
 
@@ -77,6 +84,10 @@ def resolve_sandbox_root(
 
 # Bound memory use while still allowing ordinary git push packfiles.
 MAX_BODY_BYTES = 100 * 1024 * 1024
+REQUEST_BODY_TIMEOUT_SECONDS = 30.0
+MAX_REQUEST_WORKERS = 16
+MAX_BODY_WORKERS = 2
+_BODY_WORK_SLOTS = threading.BoundedSemaphore(MAX_BODY_WORKERS)
 
 
 class GitHttpHandler(BaseHTTPRequestHandler):
@@ -184,27 +195,40 @@ class GitHttpHandler(BaseHTTPRequestHandler):
             value = self.headers.get(header)
             if value:
                 env[variable] = value
-        raw_length = self.headers.get("content-length", "0") or "0"
+        if not _BODY_WORK_SLOTS.acquire(blocking=False):
+            self.send_error(503, "git request capacity exhausted")
+            return
         try:
-            length = int(raw_length)
-        except ValueError:
-            self.send_error(400, "Bad Content-Length")
-            return
-        if length < 0:
-            self.send_error(400, "Negative Content-Length")
-            return
-        if length > MAX_BODY_BYTES:
-            self.send_error(413, "Request body too large")
-            return
-        body = self.rfile.read(length) if length else b""
-        proc = subprocess.run(
-            ["git", "http-backend"],
-            input=body,
-            env=env,
-            capture_output=True,
-            check=False,
-            timeout=GIT_GATE_TIMEOUT_SECS,
-        )
+            with tempfile.TemporaryFile() as body:
+                try:
+                    copy_declared_body(
+                        self.rfile,
+                        body,
+                        self.connection,
+                        self.headers.get("content-length"),
+                        maximum=MAX_BODY_BYTES,
+                        timeout_seconds=REQUEST_BODY_TIMEOUT_SECONDS,
+                        require_length=False,
+                    )
+                except BodyReadError as exc:
+                    self.send_error(exc.status, exc.message)
+                    return
+                body.seek(0)
+                try:
+                    proc = subprocess.run(
+                        ["git", "http-backend"],
+                        stdin=body,
+                        env=env,
+                        capture_output=True,
+                        check=False,
+                        timeout=GIT_GATE_TIMEOUT_SECS,
+                    )
+                except (OSError, subprocess.SubprocessError) as exc:
+                    self.log_message("git http-backend unavailable: %s", exc)
+                    self.send_error(503, "git backend unavailable")
+                    return
+        finally:
+            _BODY_WORK_SLOTS.release()
         self._write_cgi_response(proc.stdout)
 
     def _repo_dir(self, sandbox_root: Path, path: str) -> Path | None:
@@ -273,7 +297,9 @@ def main() -> int:
             "(no single-tenant flat-root fallback)\n"
         )
         return 1
-    server = ThreadingHTTPServer(("0.0.0.0", port), GitHttpHandler)
+    server = BoundedThreadingHTTPServer(
+        ("0.0.0.0", port), GitHttpHandler, max_workers=MAX_REQUEST_WORKERS,
+    )
     # Resolve each request's sandbox namespace by source IP against the
     # orchestrator control plane.
     server.policy_resolver = PolicyResolver(orch_url)  # type: ignore[attr-defined]

bot_bottle/gateway/git_gate/render.py

@@ -385,7 +385,7 @@ PY
       ;;
   esac
   echo "git-gate: queued # gitleaks:allow supervisor approval $proposal_id" >&2
-  echo "git-gate: approve with './cli.py supervise' to continue this push" >&2
+  echo "git-gate: approve with 'bot-bottle supervise' to continue this push" >&2
   waited=0
   while [ "$waited" -lt "$timeout" ]; do
     status=$(PYTHONPATH="/app${PYTHONPATH:+:$PYTHONPATH}" python3 - "$proposal_id" <<'PY'

bot_bottle/gateway/supervisor/mcp_dispatch.py

@@ -0,0 +1,92 @@
+"""Framework-neutral MCP method and tool dispatch."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from typing import Callable, Protocol
+
+from bot_bottle.gateway.egress.schema import load_config, route_to_yaml_dict
+from bot_bottle.gateway.policy_resolver import PolicyResolveError, PolicyResolver
+from bot_bottle.supervisor import types as _sv
+
+
+class Request(Protocol):
+    @property
+    def method(self) -> str: ...
+
+    @property
+    def params(self) -> dict[str, object]: ...
+
+
+class MethodNotFoundError(Exception):
+    """Raised when a JSON-RPC method has no MCP handler."""
+
+
+class RouteResolutionError(Exception):
+    """The caller's live route table could not be resolved authoritatively."""
+
+
+Handler = Callable[[dict[str, object]], object]
+
+
+@dataclass(frozen=True)
+class Handlers:
+    initialize: Handler
+    tools_list: Handler
+    list_routes: Handler
+    check_proposal: Handler
+    propose: Handler
+
+
+def dispatch(request: Request, handlers: Handlers) -> object:
+    """Route one parsed request without depending on the HTTP server."""
+    if request.method == "initialize":
+        return handlers.initialize(request.params)
+    if request.method == "notifications/initialized":
+        return None
+    if request.method == "tools/list":
+        return handlers.tools_list(request.params)
+    if request.method != "tools/call":
+        raise MethodNotFoundError(request.method)
+
+    tool = request.params.get("name")
+    if tool == _sv.TOOL_LIST_EGRESS_ROUTES:
+        return handlers.list_routes(request.params)
+    if tool == _sv.TOOL_CHECK_PROPOSAL:
+        return handlers.check_proposal(request.params)
+    return handlers.propose(request.params)
+
+
+def resolved_routes_payload(
+    resolver: PolicyResolver,
+    source_ip: str,
+    identity_token: str,
+) -> dict[str, object]:
+    """Render an authoritatively resolved route table for the calling bottle."""
+    try:
+        policy, bottle_id, _tokens = resolver.resolve_policy_and_bottle_id(
+            source_ip, identity_token,
+        )
+    except PolicyResolveError as exc:
+        raise RouteResolutionError("orchestrator unavailable") from exc
+    if not bottle_id:
+        raise RouteResolutionError("request source is not attributed to a bottle")
+    try:
+        config = load_config(policy or "")
+    except ValueError as exc:
+        raise RouteResolutionError("resolved policy is invalid") from exc
+    body = json.dumps(
+        {"routes": [route_to_yaml_dict(route) for route in config.routes]},
+        indent=2,
+    )
+    return {"content": [{"type": "text", "text": body}], "isError": False}
+
+
+__all__ = [
+    "Handlers",
+    "MethodNotFoundError",
+    "RouteResolutionError",
+    "dispatch",
+    "resolved_routes_payload",
+]

bot_bottle/gateway/supervisor/server.py

@@ -51,17 +51,27 @@ from __future__ import annotations
 import http.server
 import json
 import os
-import socketserver
 import sys
 import time
 import typing
 from dataclasses import dataclass
 
 from bot_bottle.constants import IDENTITY_HEADER
-from bot_bottle.gateway.egress.context import resolve_client_context
-from bot_bottle.gateway.egress.schema import load_config, route_to_yaml_dict
+from bot_bottle.gateway.bounded_http import (
+    BodyReadError,
+    BoundedThreadingHTTPServer,
+    read_declared_body,
+)
+from bot_bottle.gateway.egress.schema import load_config
 from bot_bottle.gateway.egress.types import LOG_OFF
 from bot_bottle.gateway.policy_resolver import PolicyResolveError, PolicyResolver
+from bot_bottle.gateway.supervisor.mcp_dispatch import (
+    Handlers as DispatchHandlers,
+    MethodNotFoundError,
+    RouteResolutionError,
+    dispatch,
+    resolved_routes_payload,
+)
 from bot_bottle.supervisor import types as _sv
 
 
@@ -565,6 +575,8 @@ def format_unknown_proposal_text(proposal_id: str) -> str:
 # Max request body the server accepts. 1 MB is well above any realistic
 # routes.yaml proposal.
 MAX_BODY_BYTES = 1 * 1024 * 1024
+REQUEST_BODY_TIMEOUT_SECONDS = 10.0
+MAX_REQUEST_WORKERS = 32
 
 
 class MCPHandler(http.server.BaseHTTPRequestHandler):
@@ -587,19 +599,18 @@ class MCPHandler(http.server.BaseHTTPRequestHandler):
         self._write_text(405, "use POST for MCP requests\n")
 
     def do_POST(self) -> None:
-        length_header = self.headers.get("Content-Length")
-        if length_header is None:
-            self._write_text(411, "Content-Length required\n")
-            return
         try:
-            length = int(length_header)
-        except ValueError:
-            self._write_text(400, "invalid Content-Length\n")
+            body = read_declared_body(
+                self.rfile,
+                self.connection,
+                self.headers.get("Content-Length"),
+                maximum=MAX_BODY_BYTES,
+                timeout_seconds=REQUEST_BODY_TIMEOUT_SECONDS,
+                require_length=True,
+            )
+        except BodyReadError as exc:
+            self._write_text(exc.status, exc.message + "\n")
             return
-        if length < 0 or length > MAX_BODY_BYTES:
-            self._write_text(413, "request body too large\n")
-            return
-        body = self.rfile.read(length)
 
         try:
             req = parse_jsonrpc(body)
@@ -611,6 +622,11 @@ class MCPHandler(http.server.BaseHTTPRequestHandler):
 
         try:
             result = self._dispatch(req, config)
+        except MethodNotFoundError as e:
+            self._write_jsonrpc(
+                jsonrpc_error(req.id, ERR_METHOD_NOT_FOUND, f"method not found: {e}"),
+            )
+            return
         except _RpcClientError as e:
             self._write_jsonrpc(jsonrpc_error(req.id, e.code, e.message))
             return
@@ -633,41 +649,42 @@ class MCPHandler(http.server.BaseHTTPRequestHandler):
         self._write_jsonrpc(jsonrpc_result(req.id, result))
 
     def _dispatch(self, req: JsonRpcRequest, config: ServerConfig) -> object:
-        method = req.method
-        if method == "initialize":
-            return handle_initialize(req.params)
-        if method == "notifications/initialized":
-            return None  # ack-only
-        if method == "tools/list":
-            return handle_tools_list(req.params)
-        if method == "tools/call":
-            # `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()
-            resolver = self._resolver_or_fail()
-            source_ip = self.client_address[0]
-            token = self._identity_token()
-            # `check-proposal` is a non-blocking read of the calling bottle's
-            # own queue — attributed by (source_ip, identity_token) like a
-            # proposal, but it never queues or blocks.
-            if req.params.get("name") == _sv.TOOL_CHECK_PROPOSAL:
-                return handle_check_proposal(
-                    req.params, resolver=resolver,
-                    source_ip=source_ip, identity_token=token,
-                )
-            # The control plane attributes the proposal to the source-IP + token
-            # resolved bottle, so the one shared queue holds each bottle's
-            # proposal under its own id — no slug is asserted by this daemon.
-            return handle_tools_call(
-                req.params, config, resolver=resolver,
-                source_ip=source_ip, identity_token=token,
+        def check(params: dict[str, object]) -> object:
+            return handle_check_proposal(
+                params,
+                resolver=self._resolver_or_fail(),
+                source_ip=self.client_address[0],
+                identity_token=self._identity_token(),
             )
-        raise _RpcClientError(ERR_METHOD_NOT_FOUND, f"method not found: {method}")
+
+        def propose(params: dict[str, object]) -> object:
+            return handle_tools_call(
+                params,
+                config,
+                resolver=self._resolver_or_fail(),
+                source_ip=self.client_address[0],
+                identity_token=self._identity_token(),
+            )
+
+        def list_routes(_params: dict[str, object]) -> object:
+            try:
+                return resolved_routes_payload(
+                    self._resolver_or_fail(),
+                    self.client_address[0],
+                    self._identity_token(),
+                )
+            except RouteResolutionError as exc:
+                raise _RpcInternalError(
+                    f"could not resolve live egress routes: {exc}"
+                ) from exc
+
+        return dispatch(req, DispatchHandlers(
+            initialize=handle_initialize,
+            tools_list=handle_tools_list,
+            list_routes=list_routes,
+            check_proposal=check,
+            propose=propose,
+        ))
 
     def _identity_token(self) -> str:
         """The agent's per-bottle identity token from the request header (the
@@ -686,20 +703,6 @@ class MCPHandler(http.server.BaseHTTPRequestHandler):
             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: 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()
-        conf, _slug, _tokens = resolve_client_context(
-            resolver, self.client_address[0], self._identity_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 _write_jsonrpc(self, body: bytes) -> None:
         self.send_response(200)
         self.send_header("Content-Type", "application/json")
@@ -719,7 +722,7 @@ class MCPHandler(http.server.BaseHTTPRequestHandler):
             self.wfile.write(encoded)
 
 
-class MCPServer(socketserver.ThreadingMixIn, http.server.HTTPServer):
+class MCPServer(BoundedThreadingHTTPServer):
     allow_reuse_address = True
     daemon_threads = True
     config: ServerConfig = ServerConfig()
@@ -728,6 +731,9 @@ class MCPServer(socketserver.ThreadingMixIn, http.server.HTTPServer):
     # closed per request (see `_resolver_or_fail`).
     policy_resolver: "PolicyResolver | None" = None
 
+    def __init__(self, *args, **kwargs):  # type: ignore[no-untyped-def]
+        super().__init__(*args, max_workers=MAX_REQUEST_WORKERS, **kwargs)
+
 
 # --- Entry point -----------------------------------------------------------
 

bot_bottle/orchestrator/__init__.py

@@ -44,7 +44,7 @@ if TYPE_CHECKING:
     from ..gateway import Gateway, GatewayError
     from .lifecycle import Orchestrator
     from .service import OrchestratorCore
-    from .server import OrchestratorServer, dispatch, make_server
+    from .server import OrchestratorServer, create_app, make_server
 
 
 # Facade name -> submodule that defines it. Lazy so importing a leaf (or the
@@ -67,8 +67,8 @@ _LAZY: dict[str, str] = {
     "GatewayError": "..gateway",
     "Orchestrator": ".lifecycle",
     "OrchestratorCore": ".service",
+    "create_app": ".server",
     "OrchestratorServer": ".server",
-    "dispatch": ".server",
     "make_server": ".server",
 }
 
@@ -100,7 +100,7 @@ __all__ = [
     "GatewayError",
     "Orchestrator",
     "OrchestratorCore",
+    "create_app",
     "OrchestratorServer",
-    "dispatch",
     "make_server",
 ]

bot_bottle/orchestrator/__main__.py

@@ -1,6 +1,7 @@
 """Run the orchestrator control plane as a plain process (PRD 0070 dev-harness).
 
-    python -m bot_bottle.orchestrator [--host H] [--port P] [--db PATH]
+    BOT_BOTTLE_ORCHESTRATOR_TOKEN=<signing-key> \
+        python -m bot_bottle.orchestrator [--host H] [--port P] [--db PATH]
 
 The PRD sequences the orchestrator as a plain-process dev-harness first, so
 the consolidation core (registry + attribution + HTTP control plane + live
@@ -16,12 +17,13 @@ import secrets
 from pathlib import Path
 
 from .. import log
-from .store.store_manager import StoreManager
+from ..trust_domain import CONTROL_PLANE
 from .broker import LaunchBroker, StubBroker
-from .server import make_server
 from .docker_broker import DockerBroker
-from .store.registry_store import RegistryStore, default_db_path
+from .server import make_server
 from .service import OrchestratorCore
+from .store.store_manager import StoreManager
+from .store.registry_store import RegistryStore, default_db_path
 
 
 def main(argv: list[str] | None = None) -> int:
@@ -38,6 +40,11 @@ def main(argv: list[str] | None = None) -> int:
         help="launch broker: 'stub' records requests; 'docker' runs containers",
     )
     args = parser.parse_args(argv)
+    if not CONTROL_PLANE.key_from_env():
+        log.die(
+            f"{CONTROL_PLANE.key_env} is required; refusing to start the "
+            "orchestrator without caller authentication"
+        )
 
     registry = RegistryStore(args.db)
     registry.migrate()
@@ -55,17 +62,14 @@ def main(argv: list[str] | None = None) -> int:
     orchestrator = OrchestratorCore(registry, broker, secret)
 
     server = make_server(orchestrator, host=args.host, port=args.port)
-    bound_host, bound_port = server.server_address[0], server.server_address[1]
     log.info(
         "orchestrator control plane listening",
-        context={"host": bound_host, "port": bound_port, "db": str(registry.db_path)},
+        context={"host": args.host, "port": args.port, "db": str(registry.db_path)},
     )
     try:
-        server.serve_forever()
+        server.run()
     except KeyboardInterrupt:
         log.info("orchestrator shutting down")
-    finally:
-        server.server_close()
     return 0
 
 

bot_bottle/orchestrator/api.py

@@ -0,0 +1,353 @@
+"""FastAPI control-plane routes for the orchestrator."""
+# pyright: reportUnusedFunction=false
+
+from __future__ import annotations
+
+import asyncio
+import math
+import sys
+from fastapi import FastAPI, HTTPException
+from fastapi.exceptions import RequestValidationError
+from fastapi.responses import JSONResponse
+from pydantic import BaseModel, ConfigDict, StrictStr
+from starlette.types import ASGIApp, Message, Receive, Scope, Send
+
+from ..orchestrator_auth import ROLE_CLI, ROLES
+from ..supervisor.types import TOOLS
+from ..trust_domain import CONTROL_PLANE
+from .http_contract import (
+    MAX_BODY_BYTES,
+    ORCHESTRATOR_AUTH_HEADER,
+    REQUEST_BODY_TIMEOUT_SECONDS,
+)
+from .service import OrchestratorCore
+
+_GATEWAY_ROUTES = frozenset({
+    ("POST", "/resolve"),
+    ("POST", "/supervise/propose"),
+    ("POST", "/supervise/poll"),
+})
+
+
+class _StrictModel(BaseModel):
+    model_config = ConfigDict(extra="ignore", strict=True)
+
+
+class LaunchBody(_StrictModel):
+    source_ip: StrictStr
+    image_ref: StrictStr = ""
+    metadata: StrictStr = ""
+    policy: StrictStr = ""
+    tokens: dict[StrictStr, StrictStr] = {}
+    env_var_secret: StrictStr = ""
+
+
+class PolicyBody(_StrictModel):
+    policy: StrictStr
+
+
+class ReprovisionBody(_StrictModel):
+    env_var_secret: StrictStr
+
+
+class ReconcileBody(_StrictModel):
+    live_source_ips: list[StrictStr]
+    grace_seconds: float | None = None
+
+
+class IdentityBody(_StrictModel):
+    source_ip: StrictStr
+    identity_token: StrictStr = ""
+
+
+class AttributeBody(_StrictModel):
+    source_ip: StrictStr
+    identity_token: StrictStr
+
+
+class RespondBody(_StrictModel):
+    proposal_id: StrictStr
+    bottle_slug: StrictStr
+    decision: StrictStr
+    notes: StrictStr = ""
+    final_file: StrictStr | None = None
+
+
+class ProposeBody(IdentityBody):
+    tool: StrictStr
+    proposed_file: StrictStr
+    justification: StrictStr
+
+
+class PollBody(IdentityBody):
+    proposal_id: StrictStr
+
+
+class ControlPlaneBoundary:
+    """Reject unauthenticated and oversized requests before reading a body."""
+
+    def __init__(self, app: ASGIApp, signing_key: str) -> None:
+        self.app = app
+        self.signing_key = signing_key
+
+    async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
+        if scope["type"] != "http":
+            await self.app(scope, receive, send)
+            return
+        method = scope["method"]
+        route = scope["path"].rstrip("/") or "/"
+        if not (method == "GET" and route == "/health"):
+            headers = dict(scope["headers"])
+            presented = headers.get(
+                ORCHESTRATOR_AUTH_HEADER.encode(), b"",
+            ).decode(errors="ignore")
+            role = CONTROL_PLANE.verify(presented, self.signing_key)
+            if role is None:
+                await self._reject(
+                    scope, send, 401, "control-plane authentication required",
+                )
+                return
+            allowed = ROLES if (method, route) in _GATEWAY_ROUTES else {ROLE_CLI}
+            if role not in allowed:
+                await self._reject(scope, send, 403, "insufficient role for this route")
+                return
+            scope.setdefault("state", {})["role"] = role
+        raw_length = dict(scope["headers"]).get(b"content-length")
+        if raw_length is not None:
+            try:
+                length = int(raw_length)
+            except ValueError:
+                await self._reject(scope, send, 400, "invalid Content-Length")
+                return
+            if length < 0:
+                await self._reject(scope, send, 400, "invalid Content-Length")
+                return
+            if length > MAX_BODY_BYTES:
+                await self._reject(scope, send, 413, "request body too large")
+                return
+        try:
+            body = await self._read_body(receive)
+        except _BodyTooLarge:
+            await self._reject(scope, send, 413, "request body too large")
+            return
+        except TimeoutError:
+            await self._reject(scope, send, 408, "request body read timed out")
+            return
+        try:
+            await self.app(scope, self._replay_body(body), send)
+        except Exception as exc:  # noqa: BLE001 - redact control-plane failures
+            sys.stderr.write(
+                f"orchestrator: {method} {route} failed "
+                f"[error_type={type(exc).__name__}]\n"
+            )
+            sys.stderr.flush()
+            await self._reject(scope, send, 500, "internal error")
+
+    @staticmethod
+    async def _reject(
+        scope: Scope, send: Send, status: int, error: str,
+    ) -> None:
+        response = JSONResponse({"error": error}, status_code=status)
+        await response(scope, ControlPlaneBoundary._empty_receive, send)
+
+    @staticmethod
+    async def _empty_receive() -> Message:
+        return {"type": "http.disconnect"}
+
+    @staticmethod
+    async def _read_body(receive: Receive) -> bytes:
+        body = bytearray()
+        async with asyncio.timeout(REQUEST_BODY_TIMEOUT_SECONDS):
+            while True:
+                message = await receive()
+                if message["type"] != "http.request":
+                    break
+                body.extend(message.get("body", b""))
+                if len(body) > MAX_BODY_BYTES:
+                    raise _BodyTooLarge
+                if not message.get("more_body", False):
+                    break
+        return bytes(body)
+
+    @staticmethod
+    def _replay_body(body: bytes) -> Receive:
+        sent = False
+
+        async def replay() -> Message:
+            nonlocal sent
+            if sent:
+                return {"type": "http.disconnect"}
+            sent = True
+            return {"type": "http.request", "body": body, "more_body": False}
+
+        return replay
+
+
+class _BodyTooLarge(Exception):
+    """The streamed request exceeded the control-plane body limit."""
+
+
+def _required(value: str, name: str) -> str:
+    if not value:
+        raise HTTPException(400, f"{name} (string) is required")
+    return value
+
+
+def create_app(orch: OrchestratorCore, *, signing_key: str) -> FastAPI:
+    """Build the authenticated orchestrator ASGI application."""
+    key = signing_key.strip()
+    if not key:
+        raise ValueError(
+            "orchestrator control-plane signing key is required; "
+            "refusing to start without caller authentication"
+        )
+    app = FastAPI(
+        title="bot-bottle orchestrator",
+        docs_url=None,
+        redoc_url=None,
+        openapi_url=None,
+    )
+    app.add_middleware(ControlPlaneBoundary, signing_key=key)
+
+    @app.exception_handler(RequestValidationError)
+    async def invalid_request(
+        _request: object, exc: RequestValidationError,
+    ) -> JSONResponse:
+        errors = exc.errors()
+        location = errors[0].get("loc", ()) if errors else ()
+        field = str(location[1]) if len(location) > 1 else ""
+        suffix = f": {field}" if field else ""
+        return JSONResponse(
+            {"error": f"invalid request body{suffix}"},
+            status_code=400,
+        )
+
+    @app.get("/health")
+    def health() -> dict[str, str]:
+        return {"status": "ok"}
+
+    @app.get("/gateway")
+    def gateway() -> dict[str, object]:
+        return orch.gateway_status()
+
+    @app.get("/bottles")
+    def bottles() -> dict[str, object]:
+        return {"bottles": [record.redacted() for record in orch.registry.all()]}
+
+    @app.post("/bottles", status_code=201)
+    def launch(body: LaunchBody) -> dict[str, str]:
+        rec = orch.launch_bottle(
+            _required(body.source_ip, "source_ip"),
+            image_ref=body.image_ref,
+            metadata=body.metadata,
+            policy=body.policy,
+            tokens=dict(body.tokens),
+            env_var_secret=body.env_var_secret,
+        )
+        return {"bottle_id": rec.bottle_id, "identity_token": rec.identity_token}
+
+    @app.put("/bottles/{bottle_id}/policy")
+    def set_policy(bottle_id: str, body: PolicyBody) -> dict[str, object]:
+        if orch.set_policy(bottle_id, body.policy):
+            return {"updated": True}
+        raise HTTPException(404, "no such bottle")
+
+    @app.post("/bottles/{bottle_id}/reprovision_gateway")
+    def reprovision(bottle_id: str, body: ReprovisionBody) -> dict[str, object]:
+        secret = _required(body.env_var_secret, "env_var_secret")
+        if orch.reprovision_from_secret(bottle_id, secret):
+            return {"reprovisioned": True}
+        raise HTTPException(404, "no stored secrets for this bottle")
+
+    @app.delete("/bottles/{bottle_id}")
+    def teardown(bottle_id: str) -> dict[str, object]:
+        if orch.teardown_bottle(bottle_id):
+            return {"torn_down": True}
+        raise HTTPException(404, "no such bottle")
+
+    @app.post("/reconcile")
+    def reconcile(body: ReconcileBody) -> dict[str, object]:
+        if any(not ip for ip in body.live_source_ips):
+            raise HTTPException(400, "live_source_ips must contain non-empty strings")
+        kwargs: dict[str, float] = {}
+        if body.grace_seconds is not None:
+            if not math.isfinite(body.grace_seconds) or body.grace_seconds < 0:
+                raise HTTPException(
+                    400, "grace_seconds must be a non-negative finite number",
+                )
+            kwargs["grace_seconds"] = body.grace_seconds
+        return {"reaped": orch.reconcile(body.live_source_ips, **kwargs)}
+
+    @app.post("/attribute")
+    def attribute(body: AttributeBody) -> dict[str, str]:
+        rec = orch.attribute(body.source_ip, body.identity_token)
+        if rec is None:
+            raise HTTPException(403, "unattributed")
+        return {"bottle_id": rec.bottle_id}
+
+    @app.get("/supervise/proposals")
+    def proposals() -> dict[str, object]:
+        return {"proposals": orch.supervise_pending()}
+
+    @app.post("/supervise/respond")
+    def respond(body: RespondBody) -> dict[str, object]:
+        ok, error = orch.supervise_respond(
+            _required(body.proposal_id, "proposal_id"),
+            bottle_slug=_required(body.bottle_slug, "bottle_slug"),
+            decision=_required(body.decision, "decision"),
+            notes=body.notes,
+            final_file=body.final_file,
+        )
+        if not ok:
+            raise HTTPException(409, error)
+        return {"responded": True}
+
+    @app.post("/supervise/propose", status_code=201)
+    def propose(body: ProposeBody) -> dict[str, str]:
+        source_ip = _required(body.source_ip, "source_ip")
+        if body.tool not in TOOLS:
+            raise HTTPException(400, f"tool (string) must be one of {TOOLS}")
+        rec = orch.resolve(source_ip, body.identity_token)
+        if rec is None:
+            raise HTTPException(403, "unattributed")
+        proposal_id = orch.supervise_queue_proposal(
+            rec.bottle_id,
+            tool=body.tool,
+            proposed_file=_required(body.proposed_file, "proposed_file"),
+            justification=_required(body.justification, "justification"),
+        )
+        return {"proposal_id": proposal_id}
+
+    @app.post("/supervise/poll")
+    def poll(body: PollBody) -> dict[str, object]:
+        rec = orch.resolve(
+            _required(body.source_ip, "source_ip"), body.identity_token,
+        )
+        if rec is None:
+            raise HTTPException(403, "unattributed")
+        return orch.supervise_poll_response(
+            rec.bottle_id, _required(body.proposal_id, "proposal_id"),
+        )
+
+    @app.post("/resolve")
+    def resolve(body: IdentityBody) -> dict[str, object]:
+        rec = orch.resolve(
+            _required(body.source_ip, "source_ip"), body.identity_token,
+        )
+        if rec is None:
+            raise HTTPException(403, "unattributed")
+        return {
+            "bottle_id": rec.bottle_id,
+            "policy": rec.policy,
+            "tokens": orch.tokens_for(rec.bottle_id),
+        }
+
+    return app
+
+
+__all__ = [
+    "ControlPlaneBoundary",
+    "MAX_BODY_BYTES",
+    "ORCHESTRATOR_AUTH_HEADER",
+    "create_app",
+]

bot_bottle/orchestrator/client.py

@@ -21,7 +21,7 @@ from dataclasses import dataclass
 from ..log import debug
 from ..orchestrator_auth import ROLE_CLI
 from ..trust_domain import CONTROL_PLANE
-from .server import ORCHESTRATOR_AUTH_HEADER
+from .http_contract import ORCHESTRATOR_AUTH_HEADER
 
 DEFAULT_TIMEOUT_SECONDS = 5.0
 

bot_bottle/orchestrator/http_contract.py

@@ -0,0 +1,11 @@
+"""Dependency-free constants shared by orchestrator HTTP clients and server."""
+
+ORCHESTRATOR_AUTH_HEADER = "x-bot-bottle-orchestrator-auth"
+MAX_BODY_BYTES = 1 * 1024 * 1024
+REQUEST_BODY_TIMEOUT_SECONDS = 10.0
+
+__all__ = [
+    "MAX_BODY_BYTES",
+    "ORCHESTRATOR_AUTH_HEADER",
+    "REQUEST_BODY_TIMEOUT_SECONDS",
+]

bot_bottle/orchestrator/lifecycle.py

@@ -63,10 +63,12 @@ class Orchestrator(abc.ABC):
     # orchestrator never starts without its signing key.
     provisioning: ControlPlaneProvisioning = ControlPlaneProvisioning()
 
+    def ensure_available(self) -> None:
+        """Acquire the exact image/rootfs selected for this application."""
+        self.ensure_built()
+
     def ensure_built(self) -> None:
-        """Ensure the orchestrator's image / rootfs exists, building it if
-        needed. Default: nothing to build (e.g. a pre-pulled image). Call before
-        `ensure_running`."""
+        """Local-build implementation hook retained for backend compatibility."""
         return
 
     @abc.abstractmethod

bot_bottle/orchestrator/server.py

@@ -1,464 +1,80 @@
-"""Orchestrator HTTP control plane (PRD 0070).
-
-The backend-agnostic control-plane RPC (CLI / console -> orchestrator) over
-**HTTP** — the universal transport chosen in 0070 (works on every host; no
-vsock / unix-socket portability caveats):
-
-    GET    /health                    -> 200 {"status": "ok"}
-    GET    /gateway                   -> 200 {"configured", ["name","running"]}
-    GET    /bottles                   -> 200 {"bottles": [ <redacted record>, ...]}
-    POST   /bottles                   -> 201 {"bottle_id","identity_token"}  (launch)
-                                         body: {"source_ip", ["image_ref"],
-                                                ["metadata"], ["policy"],
-                                                ["tokens"], ["env_var_secret"]}
-    PUT    /bottles/<bottle_id>/policy -> 200 {"updated": true} | 404  (live reload)
-                                         body: {"policy"}
-    POST   /bottles/<bottle_id>/reprovision_gateway
-                                      -> 200 {"reprovisioned": true} | 404
-                                         body: {"env_var_secret"}
-    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   /supervise/propose         -> 201 {"proposal_id"} | 403  (agent)
-                                         body: {"source_ip","identity_token",
-                                                "tool","proposed_file","justification"}
-    POST   /supervise/poll            -> 200 {"status", ["notes"],["final_file"]} | 403
-                                         body: {"source_ip","identity_token",
-                                                "proposal_id"}
-
-The `/supervise/propose` + `/supervise/poll` pair is the **agent** half of the
-supervise flow: the data plane (supervise / egress / git-gate) queues a proposal
-and polls for its response over RPC instead of opening `bot-bottle.db` directly.
-`poll` is idempotent — it never archives, so a dropped connection can't lose an
-operator decision (the row is reaped when the bottle is torn down / reconciled).
-Both attribute the caller by `(source_ip, identity_token)` exactly like
-`/resolve`, so a bottle can only ever queue or read its own proposals.
-
-`POST /bottles` / `DELETE` drive the full launch lifecycle: they mint (or
-tear down) the bottle in the registry AND broker the backend-native launch
-via the orchestrator. Register/deregister without a launch are internal to
-`OrchestratorCore`, not exposed here.
-
-Routing/handling is the pure function `dispatch()` so it is unit-testable
-without a socket; `Handler` / `OrchestratorServer` / `make_server` are a
-thin stdlib adapter around it. Listing redacts identity tokens — they are
-returned only once, to the caller that launches the bottle.
-"""
+"""Uvicorn transport for the FastAPI orchestrator control plane."""
 
 from __future__ import annotations
 
-import http.server
-import json
-import math
 import os
-import socketserver
-import sys
-import typing
-from urllib.parse import urlsplit
+import socket
+import threading
+
+import uvicorn
 
-from ..orchestrator_auth import ROLE_CLI, ROLES
 from ..trust_domain import CONTROL_PLANE
-from ..supervisor.types import TOOLS
+from .api import create_app
+from .http_contract import MAX_BODY_BYTES, ORCHESTRATOR_AUTH_HEADER
 from .service import OrchestratorCore
 
-# JSON body payload type (parsed request / rendered response).
-Json = dict[str, object]
-
-# The request header carrying the caller's role-scoped control-plane token (a
-# signed JWT naming the caller's role — see orchestrator_auth). The role gates which
-# routes the caller may reach: the data plane holds a `gateway` token good only
-# for the agent-facing lookups; the host CLI holds a `cli` token for the
-# operator/mutating routes. An agent that can merely *reach* the port holds no
-# token at all, and a compromised gateway holds only `gateway` — neither can
-# drive the operator routes (approve proposals, rewrite policy, read tokens).
-ORCHESTRATOR_AUTH_HEADER = "x-bot-bottle-orchestrator-auth"
-
-# The routes the data plane (role `gateway`) is allowed to reach — exactly the
-# per-request lookups PolicyResolver makes. Every other authenticated route is
-# operator-only. `cli` is a superset role: it may reach any route.
-_GATEWAY_ROUTES: frozenset[tuple[str, str]] = frozenset({
-    ("POST", "/resolve"),
-    ("POST", "/supervise/propose"),
-    ("POST", "/supervise/poll"),
-})
+MAX_REQUESTS = 32
+KEEP_ALIVE_TIMEOUT_SECONDS = 10
 
 
-def _allowed_roles(method: str, route: str) -> frozenset[str]:
-    """The roles permitted on `(method, route)`: `gateway` or `cli` on the
-    data-plane routes, `cli`-only everywhere else."""
-    if (method, route) in _GATEWAY_ROUTES:
-        return ROLES
-    return frozenset({ROLE_CLI})
+class OrchestratorServer:
+    """Small lifecycle wrapper around Uvicorn with an eagerly bound socket."""
 
+    def __init__(self, config: uvicorn.Config) -> None:
+        self._server = uvicorn.Server(config)
+        self._stopped = threading.Event()
+        self._socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+        self._socket.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+        self._socket.bind((config.host, config.port))
+        self._socket.listen(config.backlog)
+        self.server_address = self._socket.getsockname()
 
-def _parse_json_object(body: bytes) -> Json:
-    """Parse a JSON object body. Raises ValueError for non-objects / bad JSON."""
-    if not body:
-        return {}
-    obj = json.loads(body)  # raises json.JSONDecodeError (a ValueError)
-    if not isinstance(obj, dict):
-        raise ValueError("request body must be a JSON object")
-    return obj
-
-
-def dispatch(  # pylint: disable=too-many-return-statements,too-many-branches
-    orch: OrchestratorCore, method: str, path: str, body: bytes, *, role: str | None = ROLE_CLI,
-) -> 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.
-
-    `role` is the caller's verified control-plane role (`gateway` or `cli`), or
-    None for an unauthenticated request; an open-mode server (no signing key
-    configured — see `OrchestratorServer`) passes `cli`. Every route except
-    `GET /health` requires a role: a missing role is 401, and a role that
-    doesn't cover the route is 403 — so a `gateway` data-plane token can reach
-    `/resolve` + `/supervise/{propose,poll}` but not the operator routes
-    (rewrite policy, read injected tokens, approve its own supervise proposals).
-    The source-IP + identity-token checks inside `/resolve` and `/attribute`
-    authenticate the *bottle* a request is about, not the *caller*, so this role
-    gate is what protects the caller-privileged routes. Defaults `cli` 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"}
-
-    # Role gate — every route below is a trusted-caller operation. Deny before
-    # touching the registry / broker / supervise store.
-    if role is None:
-        return 401, {"error": "control-plane authentication required"}
-    if role not in _allowed_roles(method, route):
-        return 403, {"error": "insufficient role for this route"}
-
-    if method == "GET" and route == "/gateway":
-        return 200, orch.gateway_status()
-
-    if method == "GET" and route == "/bottles":
-        return 200, {"bottles": [r.redacted() for r in orch.registry.all()]}
-
-    if method == "POST" and route == "/bottles":
+    def run(self) -> None:
         try:
-            data = _parse_json_object(body)
-        except ValueError as e:
-            return 400, {"error": f"invalid JSON: {e}"}
-        source_ip = data.get("source_ip")
-        if not isinstance(source_ip, str) or not source_ip:
-            return 400, {"error": "source_ip (string) is required"}
-        image_ref = data.get("image_ref")
-        metadata = data.get("metadata")
-        policy = data.get("policy")
-        raw_tokens = data.get("tokens")
-        tokens = {
-            k: v for k, v in raw_tokens.items() if isinstance(k, str) and isinstance(v, str)
-        } if isinstance(raw_tokens, dict) else {}
-        env_var_secret = data.get("env_var_secret", "")
-        rec = orch.launch_bottle(
-            source_ip,
-            image_ref=image_ref if isinstance(image_ref, str) else "",
-            metadata=metadata if isinstance(metadata, str) else "",
-            policy=policy if isinstance(policy, str) else "",
-            tokens=tokens,
-            env_var_secret=env_var_secret if isinstance(env_var_secret, str) else "",
-        )
-        return 201, {"bottle_id": rec.bottle_id, "identity_token": rec.identity_token}
+            self._server.run(sockets=[self._socket])
+        finally:
+            self._stopped.set()
 
-    if method == "PUT" and route.startswith("/bottles/") and route.endswith("/policy"):
-        bottle_id = route[len("/bottles/"):-len("/policy")]
-        try:
-            data = _parse_json_object(body)
-        except ValueError as e:
-            return 400, {"error": f"invalid JSON: {e}"}
-        policy = data.get("policy")
-        if not isinstance(policy, str):
-            return 400, {"error": "policy (string) is required"}
-        if orch.set_policy(bottle_id, policy):
-            return 200, {"updated": True}
-        return 404, {"error": "no such bottle"}
+    def serve_forever(self) -> None:
+        self.run()
 
-    if (
-        method == "POST"
-        and route.startswith("/bottles/")
-        and route.endswith("/reprovision_gateway")
-    ):
-        bottle_id = route[len("/bottles/") : -len("/reprovision_gateway")]
-        try:
-            data = _parse_json_object(body)
-        except ValueError as e:
-            return 400, {"error": f"invalid JSON: {e}"}
-        env_var_secret = data.get("env_var_secret")
-        if not isinstance(env_var_secret, str) or not env_var_secret:
-            return 400, {"error": "env_var_secret (string) is required"}
-        if orch.reprovision_from_secret(bottle_id, env_var_secret):
-            return 200, {"reprovisioned": True}
-        return 404, {"error": "no stored secrets for this bottle"}
+    def shutdown(self) -> None:
+        self._server.should_exit = True
+        self._stopped.wait(timeout=5)
 
-    if method == "DELETE" and route.startswith("/bottles/"):
-        bottle_id = route[len("/bottles/"):]
-        if orch.teardown_bottle(bottle_id):
-            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"}
-        if any(not isinstance(ip, str) or not ip for ip in raw_ips):
-            return 400, {"error": "live_source_ips must contain non-empty strings"}
-        live = raw_ips
-        grace = data.get("grace_seconds")
-        kwargs: dict[str, float] = {}
-        if grace is not None:
-            if isinstance(grace, bool) or not isinstance(grace, (int, float)):
-                return 400, {"error": "grace_seconds must be a non-negative finite number"}
-            parsed_grace = float(grace)
-            if not math.isfinite(parsed_grace) or parsed_grace < 0:
-                return 400, {"error": "grace_seconds must be a non-negative finite number"}
-            kwargs["grace_seconds"] = parsed_grace
-        return 200, {"reaped": orch.reconcile(live, **kwargs)}
-
-    if method == "POST" and route == "/attribute":
-        try:
-            data = _parse_json_object(body)
-        except ValueError as e:
-            return 400, {"error": f"invalid JSON: {e}"}
-        source_ip = data.get("source_ip")
-        token = data.get("identity_token")
-        if not isinstance(source_ip, str) or not isinstance(token, str):
-            return 400, {"error": "source_ip and identity_token (strings) required"}
-        rec = orch.attribute(source_ip, token)
-        if rec is None:
-            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 == "/supervise/propose":
-        # Agent half: queue a proposal, attributed to the caller resolved from
-        # (source_ip, identity_token) — never a caller-supplied slug — so the
-        # data plane can't forge attribution. Fail-closed 403 when unattributed.
-        try:
-            data = _parse_json_object(body)
-        except ValueError as e:
-            return 400, {"error": f"invalid JSON: {e}"}
-        source_ip = data.get("source_ip")
-        token = data.get("identity_token")
-        tool = data.get("tool")
-        proposed_file = data.get("proposed_file")
-        justification = data.get("justification")
-        if not isinstance(source_ip, str) or not source_ip:
-            return 400, {"error": "source_ip (string) is required"}
-        if not isinstance(tool, str) or tool not in TOOLS:
-            return 400, {"error": f"tool (string) must be one of {TOOLS}"}
-        if not isinstance(proposed_file, str) or not proposed_file:
-            return 400, {"error": "proposed_file (string) is required"}
-        if not isinstance(justification, str) or not justification:
-            return 400, {"error": "justification (string) is required"}
-        rec = orch.resolve(source_ip, token if isinstance(token, str) else "")
-        if rec is None:
-            return 403, {"error": "unattributed"}
-        proposal_id = orch.supervise_queue_proposal(
-            rec.bottle_id, tool=tool, proposed_file=proposed_file,
-            justification=justification,
-        )
-        return 201, {"proposal_id": proposal_id}
-
-    if method == "POST" and route == "/supervise/poll":
-        # Agent half: non-blocking read of the caller's own proposal decision.
-        # Attributed like /propose, and scoped to the resolved bottle id, so a
-        # guessed proposal_id can never read another bottle's response.
-        try:
-            data = _parse_json_object(body)
-        except ValueError as e:
-            return 400, {"error": f"invalid JSON: {e}"}
-        source_ip = data.get("source_ip")
-        token = data.get("identity_token")
-        proposal_id = data.get("proposal_id")
-        if not isinstance(source_ip, str) or not source_ip:
-            return 400, {"error": "source_ip (string) is required"}
-        if not isinstance(proposal_id, str) or not proposal_id:
-            return 400, {"error": "proposal_id (string) is required"}
-        rec = orch.resolve(source_ip, token if isinstance(token, str) else "")
-        if rec is None:
-            return 403, {"error": "unattributed"}
-        return 200, orch.supervise_poll_response(rec.bottle_id, proposal_id)
-
-    if method == "POST" and route == "/resolve":
-        # The per-request lookup the multi-tenant gateway makes: returns the
-        # 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:
-            return 400, {"error": f"invalid JSON: {e}"}
-        source_ip = data.get("source_ip")
-        token = data.get("identity_token")
-        if not isinstance(source_ip, str) or not source_ip:
-            return 400, {"error": "source_ip (string) is required"}
-        rec = orch.resolve(source_ip, token if isinstance(token, str) else "")
-        if rec is None:
-            return 403, {"error": "unattributed"}
-        # tokens are the in-memory per-bottle egress auth values the gateway
-        # injects; served here, never persisted.
-        return 200, {
-            "bottle_id": rec.bottle_id,
-            "policy": rec.policy,
-            "tokens": orch.tokens_for(rec.bottle_id),
-        }
-
-    return 404, {"error": "not found"}
-
-
-class Handler(http.server.BaseHTTPRequestHandler):
-    """Thin stdlib adapter: read the body, call `dispatch`, write JSON."""
-
-    # Quiet by default (the orchestrator has its own logging); opt back into
-    # stdlib access logging with BOT_BOTTLE_ORCHESTRATOR_DEBUG.
-    def log_message(self, format: str, *args: typing.Any) -> None:  # noqa: A002
-        if os.environ.get("BOT_BOTTLE_ORCHESTRATOR_DEBUG"):
-            super().log_message(format, *args)
-
-    def _serve(self, method: str) -> None:
-        """Read the request body, dispatch it, and write the JSON reply. A
-        dispatch failure (e.g. a broker error) returns a 500 rather than
-        crashing the connection, so one bad request can't take the control
-        plane down for the caller."""
-        server = self.server
-        assert isinstance(server, OrchestratorServer)
-        length = int(self.headers.get("Content-Length") or 0)
-        body = self.rfile.read(length) if length > 0 else b""
-        role = server.role_for(self.headers.get(ORCHESTRATOR_AUTH_HEADER, ""))
-        try:
-            status, payload = dispatch(
-                server.orchestrator, method, self.path, body, role=role)
-        except Exception as e:  # noqa: BLE001 — the control plane must stay up
-            # Do not echo exception messages to the caller or logs: broker and
-            # persistence exceptions can contain request data. The operation,
-            # route, and exception type are enough to correlate a traceback.
-            sys.stderr.write(
-                f"orchestrator: {method} {self.path} failed "
-                f"[error_type={type(e).__name__}]\n"
-            )
-            sys.stderr.flush()
-            status, payload = 500, {"error": "internal error"}
-        data = json.dumps(payload).encode()
-        self.send_response(status)
-        self.send_header("Content-Type", "application/json")
-        self.send_header("Content-Length", str(len(data)))
-        self.end_headers()
-        self.wfile.write(data)
-
-    def do_GET(self) -> None:
-        self._serve("GET")
-
-    def do_POST(self) -> None:
-        self._serve("POST")
-
-    def do_PUT(self) -> None:
-        self._serve("PUT")
-
-    def do_DELETE(self) -> None:
-        self._serve("DELETE")
-
-
-class OrchestratorServer(socketserver.ThreadingMixIn, http.server.HTTPServer):
-    """Threading HTTP server that carries the orchestrator for its handlers.
-
-    Holds the per-host control-plane *signing key* (from
-    `$BOT_BOTTLE_ORCHESTRATOR_TOKEN`, injected by the launcher into the
-    orchestrator process only) and verifies each request's role-scoped token
-    against it. When a key is set, every route but `/health` requires a valid
-    token whose role covers the route; when it is unset the server runs **open**
-    (full `cli` access) and says so loudly at startup — a fail-visible fallback
-    for tests and any backend that hasn't wired the key yet (e.g. Firecracker,
-    whose nft boundary already blocks agents from the control-plane port)."""
-
-    daemon_threads = True
-    allow_reuse_address = True
-
-    def __init__(self, address: tuple[str, int], orchestrator: OrchestratorCore) -> None:
-        self.orchestrator = orchestrator
-        # The control-plane trust domain's signing key, as injected into THIS
-        # (the owning) process by the launcher (#476). Unset → open mode below.
-        self._signing_key = CONTROL_PLANE.key_from_env()
-        if not self._signing_key:
-            sys.stderr.write(
-                "orchestrator: WARNING — no control-plane signing key "
-                f"(${CONTROL_PLANE.key_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 role_for(self, presented: str) -> str | None:
-        """The role the request is authorized as, or None if unauthenticated.
-        Open mode (no signing key) grants full `cli` access — the fail-visible
-        fallback. Otherwise verify the presented signed token; a missing/invalid
-        token yields None (→ 401), a valid one yields its `gateway`/`cli`
-        role (→ per-route 401/403 in `dispatch`)."""
-        if not self._signing_key:
-            return ROLE_CLI
-        return CONTROL_PLANE.verify(presented, self._signing_key)
+    def server_close(self) -> None:
+        self._socket.close()
 
 
 def make_server(
-    orchestrator: OrchestratorCore, host: str = "127.0.0.1", port: int = 0
+    orchestrator: OrchestratorCore,
+    host: str = "127.0.0.1",
+    port: int = 0,
+    *,
+    signing_key: str | None = None,
 ) -> OrchestratorServer:
-    """Build (but do not start) a control-plane server. `port=0` binds an
-    ephemeral port — read `server.server_address` for the actual one."""
-    return OrchestratorServer((host, port), orchestrator)
+    """Build a bounded Uvicorn server around the orchestrator application."""
+    key = CONTROL_PLANE.key_from_env() if signing_key is None else signing_key
+    app = create_app(orchestrator, signing_key=key)
+    config = uvicorn.Config(
+        app,
+        host=host,
+        port=port,
+        access_log=bool(os.environ.get("BOT_BOTTLE_ORCHESTRATOR_DEBUG")),
+        log_level="info",
+        limit_concurrency=MAX_REQUESTS,
+        timeout_keep_alive=KEEP_ALIVE_TIMEOUT_SECONDS,
+        server_header=False,
+    )
+    return OrchestratorServer(config)
 
 
 __all__ = [
-    "dispatch", "Handler", "OrchestratorServer", "make_server", "Json",
+    "KEEP_ALIVE_TIMEOUT_SECONDS",
+    "MAX_BODY_BYTES",
+    "MAX_REQUESTS",
     "ORCHESTRATOR_AUTH_HEADER",
+    "OrchestratorServer",
+    "create_app",
+    "make_server",
 ]

bot_bottle/orchestrator/service.py

@@ -371,10 +371,12 @@ class OrchestratorCore:
         if not encrypted:
             return False
         try:
-            self._tokens[bottle_id] = {k: decrypt_value(env_var_secret, v)
-                                       for k, v in encrypted.items()}
+            decrypted = {
+                k: decrypt_value(env_var_secret, v) for k, v in encrypted.items()
+            }
         except ValueError:
             return False
+        self._tokens[bottle_id] = decrypted
         return True
 
     # --- consolidated gateway ----------------------------------------------

bot_bottle/orchestrator/store/registry_store.py

@@ -129,6 +129,10 @@ _MIGRATIONS = TableMigrations(
         # v5 — index for fast per-bottle lookups and bulk DELETE on teardown.
         "CREATE INDEX IF NOT EXISTS idx_bottled_agent_secrets_id "
         "ON bottled_agent_secrets (bottled_agent_id, type)",
+        # v6 — unauthenticated legacy ciphertext must never be selected by
+        # attacker-controlled blob contents. Existing local agents are
+        # intentionally reprovisioned instead of retaining downgrade support.
+        "DELETE FROM bottled_agent_secrets",
     ],
 )
 

bot_bottle/orchestrator/store/secret_store.py

@@ -12,9 +12,15 @@ reattachment path reads ENV_VAR_SECRET from the running agent container via
 ``POST /bottles/<id>/reprovision_gateway``; the orchestrator decrypts the
 stored rows and re-populates ``_tokens``.
 
-Encryption scheme: HMAC-SHA256 used as a PRF in CTR mode (stdlib-only,
-no external deps). Each value is encrypted independently. The output blob is
-``nonce (16 bytes) || ciphertext`` encoded as URL-safe base64 (no padding).
+Encryption scheme: encrypt-then-MAC using independent HMAC-SHA256-derived
+encryption and authentication subkeys (stdlib-only, no external deps). Each
+value is encrypted independently. New output blobs are:
+
+``version || nonce (16 bytes) || ciphertext || tag (32 bytes)``
+
+encoded as URL-safe base64 (no padding). Unversioned legacy ciphertext is
+rejected; the registry migration clears those rows rather than allowing blob
+contents to select an unauthenticated decoder.
 
     keystream_block_i = HMAC-SHA256(key, nonce || i.to_bytes(4, "big"))
     ciphertext_i      = plaintext_i XOR keystream_block_i[:len(plaintext_i)]
@@ -30,6 +36,8 @@ import secrets
 _KEY_BYTES = 32    # 256-bit key from ENV_VAR_SECRET
 _NONCE_BYTES = 16  # 128-bit random nonce per encrypt call
 _BLOCK = 32        # HMAC-SHA256 output width == one keystream block
+_TAG_BYTES = 32
+_VERSION = b"BBSE1"
 
 # Env-var name the agent container receives at startup.
 ENV_VAR_SECRET_NAME = "ENV_VAR_SECRET"
@@ -41,7 +49,13 @@ def new_env_var_secret() -> str:
 
 
 def _b64dec(s: str) -> bytes:
-    return base64.urlsafe_b64decode(s + "=" * (-len(s) % 4))
+    return base64.b64decode(
+        s + "=" * (-len(s) % 4), altchars=b"-_", validate=True,
+    )
+
+
+def _subkey(key: bytes, purpose: bytes) -> bytes:
+    return hmac.new(key, b"bot-bottle-secret-store:" + purpose, hashlib.sha256).digest()
 
 
 def _keystream(key: bytes, nonce: bytes, block_index: int) -> bytes:
@@ -53,37 +67,53 @@ def _keystream(key: bytes, nonce: bytes, block_index: int) -> bytes:
 def encrypt_value(secret_b64: str, plaintext: str) -> str:
     """Encrypt a single string value with *secret_b64* (the ENV_VAR_SECRET).
 
-    Returns a URL-safe base64 blob ``nonce || ciphertext`` suitable for
+    Returns a URL-safe base64 authenticated blob suitable for
     the ``bottled_agent_secrets.value`` column."""
     key = _b64dec(secret_b64)
+    encryption_key = _subkey(key, b"encryption")
+    authentication_key = _subkey(key, b"authentication")
     pt = plaintext.encode()
     nonce = secrets.token_bytes(_NONCE_BYTES)
     ct = bytearray()
     for i in range(0, len(pt), _BLOCK):
         chunk = pt[i : i + _BLOCK]
-        ks = _keystream(key, nonce, i)[: len(chunk)]
+        ks = _keystream(encryption_key, nonce, i // _BLOCK)[: len(chunk)]
         ct.extend(p ^ k for p, k in zip(chunk, ks))
-    return base64.urlsafe_b64encode(nonce + bytes(ct)).rstrip(b"=").decode()
+    authenticated = _VERSION + nonce + bytes(ct)
+    tag = hmac.new(authentication_key, authenticated, hashlib.sha256).digest()
+    return base64.urlsafe_b64encode(authenticated + tag).rstrip(b"=").decode()
 
 
 def decrypt_value(secret_b64: str, blob_b64: str) -> str:
     """Decrypt a blob produced by :func:`encrypt_value`.
 
     Returns the original plaintext string. Raises ``ValueError`` for malformed
-    input or a key mismatch (wrong key produces garbage, not an error, unless
-    the plaintext is non-UTF-8 — treat all such failures as wrong key)."""
+    input, authentication failure, or a key mismatch."""
     key = _b64dec(secret_b64)
     try:
         blob = _b64dec(blob_b64)
-    except Exception as exc:
+    except (ValueError, TypeError) as exc:
         raise ValueError(f"invalid ciphertext blob: {exc}") from exc
-    if len(blob) < _NONCE_BYTES:
+    if not blob.startswith(_VERSION):
+        raise ValueError("unsupported ciphertext format")
+    minimum = len(_VERSION) + _NONCE_BYTES + _TAG_BYTES
+    if len(blob) < minimum:
         raise ValueError("ciphertext blob too short")
-    nonce, ciphertext = blob[:_NONCE_BYTES], blob[_NONCE_BYTES:]
+    authenticated, supplied_tag = blob[:-_TAG_BYTES], blob[-_TAG_BYTES:]
+    authentication_key = _subkey(key, b"authentication")
+    expected_tag = hmac.new(
+        authentication_key, authenticated, hashlib.sha256,
+    ).digest()
+    if not hmac.compare_digest(supplied_tag, expected_tag):
+        raise ValueError("ciphertext authentication failed")
+    nonce_start = len(_VERSION)
+    nonce = blob[nonce_start : nonce_start + _NONCE_BYTES]
+    ciphertext = blob[nonce_start + _NONCE_BYTES : -_TAG_BYTES]
+    encryption_key = _subkey(key, b"encryption")
     pt = bytearray()
     for i in range(0, len(ciphertext), _BLOCK):
         chunk = ciphertext[i : i + _BLOCK]
-        ks = _keystream(key, nonce, i)[: len(chunk)]
+        ks = _keystream(encryption_key, nonce, i // _BLOCK)[: len(chunk)]
         pt.extend(c ^ k for c, k in zip(chunk, ks))
     try:
         return bytes(pt).decode()
@@ -91,4 +121,9 @@ def decrypt_value(secret_b64: str, blob_b64: str) -> str:
         raise ValueError(f"decryption produced non-UTF-8 output (wrong key?): {exc}") from exc
 
 
-__all__ = ["ENV_VAR_SECRET_NAME", "new_env_var_secret", "encrypt_value", "decrypt_value"]
+__all__ = [
+    "ENV_VAR_SECRET_NAME",
+    "new_env_var_secret",
+    "encrypt_value",
+    "decrypt_value",
+]

bot_bottle/release-manifest.json

@@ -0,0 +1 @@
+{"schema": 1, "development": true}

bot_bottle/release_manifest.py

@@ -0,0 +1,149 @@
+"""Immutable infrastructure identities assigned when bot-bottle is packaged."""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+from . import resources
+
+_MANIFEST_NAME = "release-manifest.json"
+_OCI_RE = re.compile(r"^[^@\s]+@sha256:([0-9a-f]{64})$")
+_SHA_RE = re.compile(r"^[0-9a-f]{64}$")
+_COMMIT_RE = re.compile(r"^[0-9a-f]{40}$")
+
+
+class ReleaseManifestError(RuntimeError):
+    """The installed package has no usable infrastructure release identity."""
+
+
+@dataclass(frozen=True)
+class FirecrackerArtifact:
+    version: str
+    sha256: str
+
+
+@dataclass(frozen=True)
+class ReleaseManifest:
+    source_commit: str
+    orchestrator_image: str
+    gateway_image: str
+    firecracker_orchestrator: FirecrackerArtifact
+    firecracker_gateway: FirecrackerArtifact
+
+
+def manifest_path() -> Path:
+    override = os.environ.get("BOT_BOTTLE_RELEASE_MANIFEST", "").strip()
+    return Path(override) if override else Path(__file__).resolve().parent / _MANIFEST_NAME
+
+
+def _artifact(value: Any, role: str) -> FirecrackerArtifact:
+    if not isinstance(value, dict):
+        raise ReleaseManifestError(f"release manifest firecracker.{role} must be an object")
+    version = value.get("version")
+    sha256 = value.get("sha256")
+    if not isinstance(version, str) or not version.strip():
+        raise ReleaseManifestError(
+            f"release manifest firecracker.{role}.version must be non-empty")
+    if not isinstance(sha256, str) or _SHA_RE.fullmatch(sha256) is None:
+        raise ReleaseManifestError(
+            f"release manifest firecracker.{role}.sha256 must be 64 lowercase hex")
+    return FirecrackerArtifact(version.strip(), sha256)
+
+
+def parse_manifest(data: Any) -> ReleaseManifest:
+    if not isinstance(data, dict) or data.get("schema") != 1:
+        raise ReleaseManifestError("release manifest must use schema 1")
+    commit = data.get("source_commit")
+    if not isinstance(commit, str) or _COMMIT_RE.fullmatch(commit) is None:
+        raise ReleaseManifestError(
+            "release manifest source_commit must be 40 lowercase hex")
+    oci = data.get("oci")
+    if not isinstance(oci, dict):
+        raise ReleaseManifestError("release manifest oci must be an object")
+    images: dict[str, str] = {}
+    for role in ("orchestrator", "gateway"):
+        ref = oci.get(role)
+        if not isinstance(ref, str) or _OCI_RE.fullmatch(ref) is None:
+            raise ReleaseManifestError(
+                f"release manifest oci.{role} must be digest-pinned")
+        images[role] = ref
+    firecracker = data.get("firecracker")
+    if not isinstance(firecracker, dict):
+        raise ReleaseManifestError("release manifest firecracker must be an object")
+    return ReleaseManifest(
+        source_commit=commit,
+        orchestrator_image=images["orchestrator"],
+        gateway_image=images["gateway"],
+        firecracker_orchestrator=_artifact(
+            firecracker.get("orchestrator"), "orchestrator"),
+        firecracker_gateway=_artifact(firecracker.get("gateway"), "gateway"),
+    )
+
+
+def load_manifest() -> ReleaseManifest:
+    path = manifest_path()
+    try:
+        data = json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError) as exc:
+        raise ReleaseManifestError(
+            f"bot-bottle was packaged without valid infrastructure pins "
+            f"({path}): {exc}") from exc
+    return parse_manifest(data)
+
+
+def packaged_manifest() -> ReleaseManifest | None:
+    """Release identity, or ``None`` for checkouts/ad-hoc development wheels."""
+    if resources.is_source_checkout() and not os.environ.get(
+        "BOT_BOTTLE_RELEASE_MANIFEST"
+    ):
+        return None
+    path = manifest_path()
+    try:
+        raw = json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return load_manifest()  # raise the detailed, fail-closed error
+    if isinstance(raw, dict) and raw.get("development") is True:
+        return None
+    return load_manifest()
+
+
+def local_build_requested() -> bool:
+    return os.environ.get("BOT_BOTTLE_INFRA_BUILD", "").strip().lower() == "local"
+
+
+def oci_image(role: str, local_default: str) -> tuple[str, bool]:
+    """Return ``(reference, should_build_locally)`` for a fixed OCI service."""
+    override_name = f"BOT_BOTTLE_{role.upper()}_IMAGE"
+    override = os.environ.get(override_name, "").strip()
+    manifest = packaged_manifest()
+    local = local_build_requested() or manifest is None
+    if override:
+        if not local and _OCI_RE.fullmatch(override) is None:
+            raise ReleaseManifestError(
+                f"{override_name} must be digest-pinned outside local build mode")
+        return override, local
+    if local:
+        return local_default, True
+    assert manifest is not None
+    return (
+        manifest.orchestrator_image if role == "orchestrator"
+        else manifest.gateway_image,
+        False,
+    )
+
+
+__all__ = [
+    "FirecrackerArtifact",
+    "ReleaseManifest",
+    "ReleaseManifestError",
+    "load_manifest",
+    "local_build_requested",
+    "oci_image",
+    "packaged_manifest",
+    "parse_manifest",
+]

bot_bottle/resources.py

@@ -44,6 +44,7 @@ BUNDLED_RESOURCES: tuple[str, ...] = (
     "Dockerfile.orchestrator",
     "Dockerfile.orchestrator.fc",
     "requirements.gateway.lock",
+    "requirements.orchestrator.lock",
     "nix/firecracker-netpool.nix",
     "scripts/firecracker-netpool.sh",
 )

bot_bottle/store/db_store.py

@@ -2,7 +2,9 @@
 
 from __future__ import annotations
 
+import os
 import sqlite3
+import stat
 from contextlib import contextmanager
 from pathlib import Path
 
@@ -19,9 +21,62 @@ class DbStore:
     def __init__(self, db_path: Path, migrations: TableMigrations) -> None:
         self.db_path = db_path
         self._migrations = migrations
-        self.db_path.parent.mkdir(parents=True, exist_ok=True)
+        self._secure_parent()
+        if self.db_path.exists():
+            self._chmod()
+
+    def _secure_parent(self) -> None:
+        """Create and verify the private parent directory."""
+        parent = self.db_path.parent
+        parent.mkdir(mode=0o700, parents=True, exist_ok=True)
+        if parent.is_symlink():
+            raise PermissionError(f"database directory must not be a symlink: {parent}")
+        parent.chmod(0o700)
+        parent_stat = parent.lstat()
+        if not stat.S_ISDIR(parent_stat.st_mode):
+            raise PermissionError(f"database parent is not a directory: {parent}")
+        if stat.S_IMODE(parent_stat.st_mode) != 0o700:
+            raise PermissionError(f"database directory is not mode 0700: {parent}")
+
+    def _secure_db_file(self) -> None:
+        """Create the database without a permissive filesystem window.
+
+        SQLite otherwise creates a missing database using the process umask.
+        This store contains control-plane identity tokens, so both creation and
+        repair are fail-closed rather than best-effort.
+        """
+        flags = os.O_RDWR | os.O_CREAT | os.O_NOFOLLOW
+        fd = os.open(
+            self.db_path,
+            flags,
+            stat.S_IRUSR | stat.S_IWUSR,
+        )
+        try:
+            self._secure_open_file(fd)
+        finally:
+            os.close(fd)
+
+    def _chmod(self) -> None:
+        """Enforce and verify the private database mode after every write."""
+        fd = os.open(self.db_path, os.O_RDWR | os.O_NOFOLLOW)
+        try:
+            self._secure_open_file(fd)
+        finally:
+            os.close(fd)
+
+    def _secure_open_file(self, fd: int) -> None:
+        """Pin, validate, and secure an opened database filesystem object."""
+        file_stat = os.fstat(fd)
+        if not stat.S_ISREG(file_stat.st_mode):
+            raise PermissionError(
+                f"database must be a regular file: {self.db_path}"
+            )
+        os.fchmod(fd, stat.S_IRUSR | stat.S_IWUSR)
+        if stat.S_IMODE(os.fstat(fd).st_mode) != 0o600:
+            raise PermissionError(f"database is not mode 0600: {self.db_path}")
 
     def _connect(self) -> sqlite3.Connection:
+        self._secure_db_file()
         conn = sqlite3.connect(self.db_path)
         conn.row_factory = sqlite3.Row
         return conn
@@ -51,16 +106,9 @@ class DbStore:
         return version == len(self._migrations.migrations)
 
     def migrate(self) -> None:
-        """Apply any pending migrations and set permissions on the DB file."""
+        """Apply any pending migrations to the already-secured DB file."""
         with self._connection() as conn:
             self._migrations.apply(conn)
-        self._chmod()
-
-    def _chmod(self) -> None:
-        try:
-            self.db_path.chmod(0o600)
-        except OSError:
-            pass
 
 
 __all__ = ["DbStore", "DbVersionError"]

bot_bottle/trust_domain.py

@@ -40,7 +40,7 @@ from .paths import (
 
 class ProvisioningError(RuntimeError):
     """A control-plane auth invariant would be violated (e.g. starting the
-    orchestrator without its signing key — which would run OPEN)."""
+    orchestrator without its signing key)."""
 
 
 @dataclass(frozen=True)
@@ -67,9 +67,8 @@ class TrustDomain:
 
     def key_from_env(self, environ: Mapping[str, str] | None = None) -> str:
         """The signing key as the owning process sees it — read from `key_env`
-        (default `os.environ`). "" when unset; the caller decides whether that is
-        fatal (`ControlPlaneProvisioning`) or the open-mode fallback
-        (`OrchestratorServer`)."""
+        (default `os.environ`). ``""`` when unset; owning services reject that
+        value rather than start without authentication."""
         env = os.environ if environ is None else environ
         return env.get(self.key_env, "").strip()
 

docs/demo.tape

@@ -1,4 +1,4 @@
-# VHS tape — drives `./cli.py start demo` interactively and asks
+# VHS tape — drives `bot-bottle start demo` interactively and asks
 # claude (the AI) to run four probes via natural-language prompts.
 # Setup (manifest + dummy SSH key + image pre-warm) and teardown
 # happen outside the tape; record via `bash scripts/demo-record.sh`,
@@ -29,7 +29,7 @@ Show
 # defaults), one git upstream (unreachable on purpose so gitleaks runs
 # before the gate would forward), and a FAKE_TOKEN env var shaped like
 # a GitHub PAT.
-Type "./cli.py start demo"
+Type "bot-bottle start demo"
 Enter
 Sleep 8s
 

docs/prds/0011-per-file-md-manifest.md

@@ -132,7 +132,7 @@ Each test runs against a temporary `$HOME` and a temporary `$CWD`:
   can revisit, but the v1 of this PRD is one file = one bottle.
 
 - **Hot-reload.** Changes to manifest files take effect at next
-  `./cli.py start`; we do not watch the directory.
+  `bot-bottle start`; we do not watch the directory.
 
 ## Scope
 

docs/prds/0018-compose-per-instance.md

@@ -290,7 +290,7 @@ After this PRD:
 
 ### Cleanup CLI
 
-`./cli.py cleanup` switches from "list every container with prefix
+`bot-bottle cleanup` switches from "list every container with prefix
 `bot-bottle-` and every network with prefix `bot-bottle-net-`
 or `bot-bottle-egress-`" to:
 
@@ -369,7 +369,7 @@ Sized for one PR each, in order.
    `docker compose up -d` + attach + teardown. Per-sidecar `start()`/
    `stop()` lifecycle methods deleted in the same chunk. Compose-
    log dump on teardown added.
-4. **Cleanup CLI on compose.** Switch `./cli.py cleanup` to
+4. **Cleanup CLI on compose.** Switch `bot-bottle cleanup` to
    `docker compose ls`-based discovery; keep prefix-scan as
    fallback for one release.
 5. **Dashboard.** Decide on the discovery question (open question

docs/prds/0019-active-agents-in-dashboard.md

@@ -37,7 +37,7 @@ Two rough edges in the current dashboard:
    shows only pending proposals. If no agent has called a tool,
    the screen reads "no pending proposals" — even when five
    bottles are quietly working. The operator has to `docker
-   compose ls` (or `./cli.py cleanup -n` to see the y/N preview)
+   compose ls` (or `bot-bottle cleanup -n` to see the y/N preview)
    to find out what's actually live.
 
 2. **`e` / `p` re-discover-and-disambiguate every invocation.**
@@ -82,12 +82,12 @@ the "operator wants to make an unprompted change" case.
   global across bottles. Filtering ("show me only this agent's
   proposals") might be a follow-up but isn't this PRD.
 - **Agent lifecycle from the dashboard.** Starting / stopping
-  agents stays in `./cli.py start` / `./cli.py cleanup`. The
+  agents stays in `bot-bottle start` / `bot-bottle cleanup`. The
   dashboard reads state; it doesn't change it.
 - **Preserved-but-not-running bottles.** The active-agents list
   is strictly "what's running now" (cross-referenced from
   `docker compose ls`). Preserved state dirs without a live
-  project don't appear — `./cli.py resume <identity>` is the
+  project don't appear — `bot-bottle resume <identity>` is the
   path for those.
 - **A separate per-agent detail view.** The agent rows are
   one-line summaries. Pressing Enter on a proposal still drops
@@ -125,7 +125,7 @@ the "operator wants to make an unprompted change" case.
 - Changes to proposal handling (`a` / `m` / `r` / Enter all
   unchanged).
 - Changes to the queue-dir / supervise sidecar protocol.
-- New CLI surface beyond what's in `./cli.py dashboard`.
+- New CLI surface beyond what's in `bot-bottle dashboard`.
 - Touching the manifest, compose renderer, launch lifecycle.
 
 ## Proposed design

docs/prds/0020-start-and-attach-from-dashboard.md

@@ -14,8 +14,8 @@
 Today the dashboard is read-only: it surfaces pending proposals
 and active agents (PRD 0019) but can't *start* an agent or
 *re-enter* one. The operator's path is split — they launch
-agents from one terminal (`./cli.py start <name>`), and watch
-them from another (`./cli.py dashboard`).
+agents from one terminal (`bot-bottle start <name>`), and watch
+them from another (`bot-bottle dashboard`).
 
 This PRD collapses that split. The dashboard becomes the
 operator's single surface: pressing a key opens an agent picker,
@@ -31,7 +31,7 @@ claude session AND the dashboard process. Exit claude → back to
 dashboard, bottle still running. Start another agent → two
 bottles up at once. Quit the dashboard → bottles continue
 running. Teardown is **always explicit**: the operator presses
-`x` on an agent, or runs `./cli.py cleanup` later.
+`x` on an agent, or runs `bot-bottle cleanup` later.
 
 ## Problem
 
@@ -45,7 +45,7 @@ Two real frictions today:
    open and the dashboard's "active agents" pane is hopelessly
    behind reality because they just spawned three in a row.
 
-2. **`./cli.py start` ties the bottle to a single claude
+2. **`bot-bottle start` ties the bottle to a single claude
    session.** The start command's `ExitStack` brings the bottle
    up, runs claude, and tears down on Ctrl-D — fine for a one-
    shot session, wrong for "let me bounce in and out of this
@@ -60,7 +60,7 @@ captures full-merged logs per bottle (PRD 0018). It already
 
 ## Goals / Success Criteria
 
-1. From inside `./cli.py dashboard`, pressing `n` (new) opens
+1. From inside `bot-bottle dashboard`, pressing `n` (new) opens
    an agent picker listing every agent defined in the manifest.
    Selecting one runs `prepare → preflight → launch`.
 2. The preflight Y/N summary renders cleanly — either as a
@@ -83,7 +83,7 @@ captures full-merged logs per bottle (PRD 0018). It already
    state cleanup) without quitting the dashboard.
 7. Quitting the dashboard (`q`) leaves every running bottle
    running. Bottle teardown is always explicit (per-bottle `x`
-   or `./cli.py cleanup`). The next `./cli.py dashboard`
+   or `bot-bottle cleanup`). The next `bot-bottle dashboard`
    invocation re-discovers them via `list_active_slugs()` and
    surfaces re-attach for any it can reconstruct context for
    (see "Cross-dashboard re-attach" below).
@@ -94,13 +94,13 @@ captures full-merged logs per bottle (PRD 0018). It already
   embedded-emulator option from the research doc is out of
   scope. The handoff (option 1) is the v1; option 2 is a
   separate PRD if and when handoff is observably insufficient.
-- **Adopting bottles started by an out-of-dashboard `./cli.py
+- **Adopting bottles started by an out-of-dashboard `bot-bottle
   start` invocation.** Those have their own ExitStack-owner and
   the dashboard treats them as read-only-watch (already does
   today). Re-attach only applies to bottles the *current
   dashboard process* started.
 - **Resurrecting an out-of-process bottle into a new dashboard
-  with full re-attach.** A bottle started by `./cli.py start`
+  with full re-attach.** A bottle started by `bot-bottle start`
   in another terminal — or by a previous dashboard run, now
   exited — appears in the agents pane (already does, PRD 0019)
   and can be re-attached via `docker exec -it claude` because
@@ -109,12 +109,12 @@ captures full-merged logs per bottle (PRD 0018). It already
   context object to drive teardown — e.g., the
   ExitStack-tracked CA + state cleanup `_settle_state` performs
   today. Cross-dashboard re-attach uses the existing
-  `./cli.py cleanup` for teardown, not an `x` keypress (see
+  `bot-bottle cleanup` for teardown, not an `x` keypress (see
   open questions).
 - **Multi-window UI.** Single curses window, two existing
   panes (proposals + agents); the agent picker is a modal, not
   a third pane.
-- **Removing `./cli.py start`.** Stays as the script-friendly /
+- **Removing `bot-bottle start`.** Stays as the script-friendly /
   legacy entry point. The dashboard is the new default.
 
 ## Scope
@@ -140,7 +140,7 @@ captures full-merged logs per bottle (PRD 0018). It already
 
 ### Out of scope
 
-- Changes to `./cli.py start` itself. It keeps its current
+- Changes to `bot-bottle start` itself. It keeps its current
   shape; the dashboard reuses its internal pieces (backend.
   prepare / backend.launch) without reaching through the CLI
   layer.
@@ -157,7 +157,7 @@ captures full-merged logs per bottle (PRD 0018). It already
 Today's flow:
 
 ```
-./cli.py start agent
+bot-bottle start agent
   └─ with backend.launch(plan) as bottle:        ← bottle alive while inside `with`
        bottle.exec_agent([...], tty=True)       ← blocks until claude exits
      # context exits → compose down → state cleanup
@@ -166,7 +166,7 @@ Today's flow:
 The proposed dashboard-driven flow:
 
 ```
-./cli.py dashboard
+bot-bottle dashboard
   └─ bottles: dict[str, tuple[ContextManager, DockerBottle]] = {}
 
      # operator presses `n`, picks agent
@@ -205,7 +205,7 @@ Two shifts:
    evaluation, state-dir reap) doesn't fire on a quit-while-
    running bottle. It DOES fire when the operator explicitly
    stops via `x`, because that calls `cm.__exit__`. For
-   bottles a previous dashboard quit on, `./cli.py cleanup`
+   bottles a previous dashboard quit on, `bot-bottle cleanup`
    is the path — its compose-down + state-reap logic
    already covers the case.
 
@@ -213,7 +213,7 @@ Two shifts:
 
 When the dashboard discovers a bottle in `discover_active_agents`
 that it didn't itself start (a previous-dashboard or external
-`./cli.py start` bottle), Enter still attaches via `docker exec
+`bot-bottle start` bottle), Enter still attaches via `docker exec
 -it … claude` — the agent container is running `sleep infinity`
 exactly the same way regardless of who started it. The only
 thing the current dashboard lacks for those bottles is the
@@ -221,8 +221,8 @@ launch-context object needed to drive a clean teardown via
 `x`.
 
 For v1 we surface this honestly: pressing `x` on a non-owned
-agent shows a status hint pointing at `./cli.py cleanup` (or
-`./cli.py cleanup` targeted at the slug if we add that flag
+agent shows a status hint pointing at `bot-bottle cleanup` (or
+`bot-bottle cleanup` targeted at the slug if we add that flag
 later). The agent stays alive; the operator handles teardown
 out-of-band. Enter (re-attach) works for both owned and
 non-owned bottles.
@@ -288,7 +288,7 @@ agents pane.
 
 `x` on a non-owned agent (discovered via `list_active_slugs`
 but not in `bottles` dict): no-op with status hint pointing
-at `./cli.py cleanup` (the existing path that tears down
+at `bot-bottle cleanup` (the existing path that tears down
 ANY bot-bottle compose project plus reaps state dirs).
 
 ### Dashboard quit
@@ -300,7 +300,7 @@ the `docker compose` project keeps running. The next dashboard
 invocation discovers the bottles via `list_active_slugs` and
 surfaces re-attach.
 
-This is a real departure from today's `./cli.py start`
+This is a real departure from today's `bot-bottle start`
 semantics (which couples bottle lifetime to the process via
 ExitStack). It's intentional: the dashboard is a watching +
 acting surface, not a lifetime owner.
@@ -322,7 +322,7 @@ Sized for one PR each.
    dashboard's ExitStack; handoff invokes `attach_agent`.
 3. **Re-attach via Enter on owned agents-pane row.** Looks up
    the slug in the dashboard's `bottles` map; if present →
-   handoff; else → status-line hint pointing at `./cli.py
+   handoff; else → status-line hint pointing at `bot-bottle
    resume`.
 4. **Explicit per-bottle stop (`x` keybinding).** Pop the
    bottle's `close` callback off the stack, call it, refresh.
@@ -369,7 +369,7 @@ Sized for one PR each.
    bottles dict goes out of scope without invoking `__exit__`,
    so the `docker compose` projects keep running. Bottle
    teardown is always explicit: per-bottle `x` (for
-   dashboard-owned), or `./cli.py cleanup` (for everything).
+   dashboard-owned), or `bot-bottle cleanup` (for everything).
 
 ## Open questions
 

docs/prds/0021-dashboard-tmux-split-pane.md

@@ -46,7 +46,7 @@ window, two panes, no terminal handoff.
 
 ## Goals / Success Criteria
 
-1. When the operator runs `./cli.py dashboard` from inside a
+1. When the operator runs `bot-bottle dashboard` from inside a
    tmux session (`$TMUX` set), the dashboard establishes a
    two-pane layout: dashboard in the left pane, an initially-
    empty right pane reserved for claude sessions.
@@ -313,7 +313,7 @@ Sized small.
 3. **Dashboard launched OUTSIDE tmux but tmux is installed.**
    Should the dashboard auto-exec itself inside a fresh tmux
    session to get the split-pane experience? Convenient but
-   surprising (`./cli.py dashboard` shouldn't silently
+   surprising (`bot-bottle dashboard` shouldn't silently
    change what session you're in). v1 leaves this off —
    operators who want split-pane mode start tmux themselves
    and then run the dashboard.
@@ -332,7 +332,7 @@ Sized small.
    PRD-0019 focus indicator?
 
 6. **Concurrent dashboards in different tmux windows.**
-   Multiple `./cli.py dashboard` invocations in different
+   Multiple `bot-bottle dashboard` invocations in different
    tmux windows would each create their own right pane —
    probably fine, each has its own state, but worth
    verifying that `tmux list-panes` is scoped to the right

docs/prds/0026-agent-provider-templates.md

@@ -14,7 +14,7 @@ Today bot-bottle is hard-wired around Claude Code assumptions. When Claude runs
 
 ## Goals / Success Criteria
 
-- A Codex agent can be started from the dashboard and via `./cli.py start` alongside a Claude agent.
+- A Codex agent can be started from the dashboard and via `bot-bottle start` alongside a Claude agent.
 - The manifest can express the agent provider/template and, where needed, a custom agent Dockerfile.
 - Claude-specific default egress/auth behavior is no longer implicit; provider-specific auth is expressed through explicit bottle egress routes and roles.
 - The launcher preserves required infrastructure behavior for sidecars, egress, pipelock, supervisor MCP, CA handling, git, and shell basics.

docs/prds/0048-ssh-deploy-key-provisioning.md

@@ -30,7 +30,7 @@ across every bottle spin-up. This has several consequences:
   to grant that access.
 - **Manual rotation burden.** Operators must manage key files on disk, keeping
   them secure, rotating them on a schedule, and distributing them across hosts
-  that run `./cli.py start`.
+  that run `bot-bottle start`.
 
 ## Goals / Success Criteria
 

docs/prds/0049-strip-dashboard-to-supervisor-tui.md

@@ -6,7 +6,7 @@
 
 ## Summary
 
-The `./cli.py dashboard` command has grown from its PRD 0013 roots
+The `bot-bottle dashboard` command has grown from its PRD 0013 roots
 (triage supervise proposals) into a parallel-agent control surface
 (PRDs 0019/0020/0021): an active-agents pane, agent picker + start,
 re-attach, per-bottle stop, tmux split-pane handoff, operator-
@@ -21,7 +21,7 @@ proposals, approve / modify / reject each one, write audit entries,
 deliver the response that unblocks the agent's tool call. Everything
 that's about *starting / re-entering / stopping* bottles, or about
 *operator-initiated* config edits, comes out. The command is renamed
-`./cli.py supervise` so the name matches what it does after the cut.
+`bot-bottle supervise` so the name matches what it does after the cut.
 
 Future agent-management UX is explicitly punted: if and when a
 control surface for parallel agents resurfaces, the working
@@ -41,9 +41,9 @@ Three concrete pains, all downstream of the dashboard's growth:
   ExitStack-free bottle ownership are intricate enough that
   shipping the next polish increment costs more than it returns.
 2. **No clear ownership of "starts and stops bottles".** Today
-  that responsibility is split: `./cli.py start` owns one-shot
+  that responsibility is split: `bot-bottle start` owns one-shot
   sessions; the dashboard owns multi-session bottles it started
-  itself; `./cli.py cleanup` owns everything else. The dashboard
+  itself; `bot-bottle cleanup` owns everything else. The dashboard
   tracking its own `bottles: dict[str, (cm, bottle, identity)]`
   that doesn't survive a quit is a confusing third lane.
 3. **Wrong target shape for a "manage many agents" UI.** The
@@ -97,12 +97,12 @@ problem is everything that got bolted onto that core after.
   dashboard. After this PRD they don't exist anywhere — operators
   who need ad-hoc edits use the same path the agents do (call the
   supervise tool from inside the bottle) or hand-edit the host-
-  side files and restart the sidecar. Adding a `./cli.py routes
+  side files and restart the sidecar. Adding a `bot-bottle routes
   edit <slug>` verb is a follow-up if the loss bites.
-- **Removing `./cli.py start` or changing its semantics.** Start
+- **Removing `bot-bottle start` or changing its semantics.** Start
   remains the one-shot launch path. PRD 0020's bottle-outlives-
   process model is removed; the only path to a long-running
-  bottle is `./cli.py start` (foreground) plus `cli.py cleanup`
+  bottle is `bot-bottle start` (foreground) plus `cli.py cleanup`
   for teardown.
 - **Removing the supervise-sidecar protocol or any of the three
   block-remediation engines.** PRDs 0013–0016 stay Active. The
@@ -122,8 +122,8 @@ problem is everything that got bolted onto that core after.
 
 ### In scope
 
-- **Rename the subcommand.** `./cli.py dashboard` becomes
-  `./cli.py supervise`. The module moves from `bot_bottle/cli/
+- **Rename the subcommand.** `bot-bottle dashboard` becomes
+  `bot-bottle supervise`. The module moves from `bot_bottle/cli/
   dashboard.py` to `bot_bottle/cli/supervise.py`. The dispatcher
   in `bot_bottle/cli/__init__.py` and the help text both update.
 - **Strip the curses loop to proposal-only.** The remaining
@@ -167,7 +167,7 @@ problem is everything that got bolted onto that core after.
 
 - Any new feature in the supervise TUI. The cut is purely
   subtractive (except for the rename).
-- Behavior changes in `./cli.py start`, `cli.py cleanup`,
+- Behavior changes in `bot-bottle start`, `cli.py cleanup`,
   `cli.py resume`, `cli.py list`, `cli.py info`, `cli.py edit`,
   `cli.py init` — unchanged.
 - Changes to the supervise sidecar (`supervise_server.py`,
@@ -181,7 +181,7 @@ problem is everything that got bolted onto that core after.
 
 ### Final shape of the TUI
 
-After this PRD the `./cli.py supervise` curses surface is:
+After this PRD the `bot-bottle supervise` curses surface is:
 
 ```
 bot-bottle supervise  (3 pending)
@@ -307,8 +307,8 @@ The PR closes issue #174.
 
 1. **`e` / `p` operator-initiated edits — gone for good or
   moved to a separate CLI verb?** The PRD removes them with no
-  replacement. The simplest replacement is `./cli.py routes
-  edit <slug>` and `./cli.py pipelock edit <slug>`, sharing
+  replacement. The simplest replacement is `bot-bottle routes
+  edit <slug>` and `bot-bottle pipelock edit <slug>`, sharing
   the existing `apply_routes_change` / `apply_allowlist_change`
   engines. If the loss is felt within the first parallel
   run after this lands, that follow-up is a small PR. Leaving

docs/prds/0051-launch-selector.md

@@ -7,12 +7,12 @@
 
 ## Summary
 
-When `./cli.py start` is run without an agent name, or without a backend
+When `bot-bottle start` is run without an agent name, or without a backend
 explicitly specified, the user currently gets an argparse error (missing
 positional) or falls through to the `docker` default silently. This PRD
 adds a terminal UI that appears in those gaps: a filter-select screen
 built with `curses` that lets the operator pick the agent and/or backend
-interactively rather than memorising names or consulting `./cli.py list`.
+interactively rather than memorising names or consulting `bot-bottle list`.
 
 ## Problem
 
@@ -29,15 +29,15 @@ visible.
 
 ## Goals / Success Criteria
 
-1. `./cli.py start` (no arguments) shows an interactive agent selector;
+1. `bot-bottle start` (no arguments) shows an interactive agent selector;
    the selected name is used exactly as if it had been passed on the
    command line.
-2. `./cli.py start <name>` (no `--backend`, no `BOT_BOTTLE_BACKEND`)
+2. `bot-bottle start <name>` (no `--backend`, no `BOT_BOTTLE_BACKEND`)
    shows an interactive backend selector; the selected backend is used
    exactly as if `--backend=<selected>` had been passed.
-3. `./cli.py start <name> --backend=<b>` (both explicit) shows neither
+3. `bot-bottle start <name> --backend=<b>` (both explicit) shows neither
    screen — no behavioural change from today.
-4. `./cli.py start` (no arguments, no env backend) shows the agent
+4. `bot-bottle start` (no arguments, no env backend) shows the agent
    selector first, then the backend selector.
 5. The filter-select widget is a standalone utility
    (`bot_bottle/cli/tui.py`) shared by both selectors.
@@ -57,7 +57,7 @@ visible.
 - No pagination beyond what fits in the terminal window (scroll via
   cursor movement is sufficient for typical agent counts).
 - No multi-select; exactly one item is chosen per invocation.
-- No changes to `./cli.py resume`, `./cli.py list`, or any other
+- No changes to `bot-bottle resume`, `bot-bottle list`, or any other
   subcommand.
 
 ## Design
@@ -83,7 +83,7 @@ def filter_select(
 
 The widget renders to the tty file descriptor opened via `curses.initscr`
 (or `curses.newterm` on the tty fd so stdout remains clean for callers
-that pipe `./cli.py`).
+that pipe `bot-bottle`).
 
 Layout (full-width, minimal):
 
@@ -140,7 +140,7 @@ agent picker can populate itself from the real manifest. The same
 `filter_select` opens `/dev/tty` and feeds it as the input file to
 `curses.wrapper`-equivalent code (using `curses.newterm` to avoid
 clobbering the caller's stdout/stderr). This keeps the picker
-composable — callers can pipe `./cli.py` output without the curses
+composable — callers can pipe `bot-bottle` output without the curses
 draw sequences contaminating the pipe.
 
 ## Implementation chunks

docs/prds/0060-commit-bottle-state.md

@@ -30,12 +30,12 @@ snapshot before a planned host reboot or hardware migration.
 
 ## Goals / Success Criteria
 
-- `./cli.py commit [<slug>]` takes a snapshot of the running agent and
+- `bot-bottle commit [<slug>]` takes a snapshot of the running agent and
   stores it as a local artifact.
 - Without a slug argument the command shows the same interactive picker
   as `start` (the list of active slugs).
 - The committed artifact reference is stored in per-bottle state so
-  that the next `./cli.py resume <slug>` automatically uses the
+  that the next `bot-bottle resume <slug>` automatically uses the
   snapshot instead of rebuilding from the Dockerfile.
 - `mark_preserved` is called so the state dir survives the normal
   session-end cleanup.
@@ -81,7 +81,7 @@ to the committed `.smolmachine` artifact.
 ### `commit` command
 
 ```
-./cli.py commit [<slug>]
+bot-bottle commit [<slug>]
 ```
 
 1. Resolve slug (arg or interactive picker from `enumerate_active_agents`).

docs/prds/0066-separate-agent-bottle-selection.md

@@ -37,7 +37,7 @@ egress policy), the operator must duplicate the agent file and change the
    selection order, as the effective bottle for the session.
 5. Confirming with an empty selection falls back to the agent's `bottle:` field.
    If neither is set, a ManifestError is raised pointing the operator at the fix.
-6. The ordered bottle list is stored in launch metadata so `./cli.py resume`
+6. The ordered bottle list is stored in launch metadata so `bot-bottle resume`
    uses the same bottles.
 7. The preflight summary (`y/N` screen) shows the effective bottle name(s).
 8. The multi-select picker supports incremental filtering, Space/Enter to toggle
@@ -52,7 +52,7 @@ egress policy), the operator must duplicate the agent file and change the
 - Reordering the selection list from within the picker (order = insertion order;
   drag-and-drop is out of scope).
 - Storing bottle selection history / MRU.
-- Changes to `./cli.py edit`, `./cli.py list`, or `./cli.py info`.
+- Changes to `bot-bottle edit`, `bot-bottle list`, or `bot-bottle info`.
 - Removing the `bottle:` key from the agent schema (it stays, now optional).
 
 ## Design

docs/prds/0068-smolmachines-linux.md

@@ -74,7 +74,7 @@ macOS-only for v1. Three concrete blockers:
 
 ## Goals / Success Criteria
 
-- `BOT_BOTTLE_BACKEND=smolmachines ./cli.py start <agent>` launches,
+- `BOT_BOTTLE_BACKEND=smolmachines bot-bottle start <agent>` launches,
   runs, and tears down a bottle on a Linux host with `/dev/kvm`.
 - The TSI allowlist is enforced on Linux: PRD 0022's
   `tests/integration/test_sandbox_escape.py` passes against

docs/prds/0077-macos-container-ci-runner.md

@@ -39,7 +39,7 @@ end-to-end runner that would catch the *next* macOS-only launch regression.)
 PR #470 (#414) already made the integration suite backend-agnostic:
 `skip_unless_selected_backend_available()` gates on the *selected* backend's
 own `is_backend_ready()` rather than `docker_available()`, and each
-integration job runs `./cli.py backend status --backend=<name>` as a preflight
+integration job runs `bot-bottle backend status --backend=<name>` as a preflight
 that fails loudly when the backend is missing. That is the machinery this job
 plugs into; this PRD supplies the runner and the job.
 
@@ -98,7 +98,7 @@ Modeled on `integration-firecracker`:
 - `concurrency: { group: integration-macos-infra, cancel-in-progress: false }`
   to serialize runs against the singleton.
 - **Preflight** — `command -v container`, `container system status`, then
-  `./cli.py backend status --backend=macos-container`; any failure exits
+  `bot-bottle backend status --backend=macos-container`; any failure exits
   non-zero so a misprovisioned runner fails loudly instead of silently
   skipping.
 - Run the integration suite under coverage with

docs/prds/0078-install-script.md

@@ -16,7 +16,7 @@ verifies host prerequisites after install.
 ## Problem
 
 There is currently no install path for new users. The only way to run
-bot-bottle is to clone the repo and invoke `./cli.py`. This blocks any
+bot-bottle is to clone the repo and invoke `bot-bottle`. This blocks any
 public demo: readers want `curl | sh` or `pipx install`, not a manual
 clone-and-configure flow. There is also no single command that tells a
 user whether their host is actually ready to run a bottle.

docs/prds/0079-control-plane-auth-provisioning.md

@@ -56,8 +56,7 @@ key.
 - Rewriting the HMAC primitive: `orchestrator_auth.mint/verify` gain an optional
   `roles=` arg (default unchanged) so a key can carry a different role set;
   nothing else changes.
-- Network topology, the plane split (#469), or the server's open-mode fallback
-  for tests.
+- Network topology or the plane split (#469).
 
 ## Design
 

docs/prds/0082-authoritative-failure-boundaries.md

@@ -0,0 +1,197 @@
+# PRD 0082: Authoritative failure boundaries
+
+- **Status:** Draft
+- **Author:** codex
+- **Created:** 2026-07-27
+- **Issue:** #444
+
+## Summary
+
+Make every security- or lifecycle-sensitive snapshot distinguish authoritative
+empty state from unavailable state, and make every destructive or
+resource-consuming boundary revalidate the assumptions it acts on. This
+finishes the focused quality work begun under #444 without broad rewrites:
+cleanup cannot act on stale identities, policy introspection cannot publish a
+fabricated empty policy, gateway servers bound untrusted work, and daemon
+shutdown does not emit uncaught background-thread failures. Shared
+control-plane storage and gateway credential provisioning also enforce their
+filesystem security contract before sensitive data is written.
+
+## Problem
+
+Several paths are individually fail-closed but compose into unsafe or
+misleading behavior:
+
+1. `cleanup` prepares a plan, waits indefinitely for operator confirmation,
+   then kills stored PIDs and removes stored paths without checking that those
+   identities still describe the same orphan. A PID may be reused or a run
+   directory may become active during the prompt.
+2. The supervisor reuses egress's deny-all fallback for
+   `list-egress-routes`. Deny-all is correct for enforcement, but presenting it
+   as a successful empty route table can cause a later replace-all proposal to
+   discard live routes.
+3. The supervisor and Git HTTP services accept bounded declared body sizes but
+   use blocking reads and unbounded request threads. An untrusted bottle can
+   exhaust the shared gateway with slow or parallel requests.
+4. macOS cleanup enumerates containers and networks independently and treats a
+   failed query as an empty class, so a partial snapshot can still become a
+   destructive plan.
+5. Gateway log-pump threads race stream closure during shutdown and emit
+   uncaught exceptions even when shutdown otherwise succeeds.
+6. Firecracker discovers VMs through whitespace-split `pgrep -a` output.
+   A configured cache path containing spaces can hide a live VM from the
+   snapshot and make its run directory appear orphaned.
+7. Docker cleanup asks compose for its project snapshot in best-effort mode.
+   A transient query failure can therefore become an empty stopped-project
+   set and authorize deletion of associated state directories.
+8. Firecracker artifact downloads and registry publication have no network
+   deadline, so an unresponsive registry can hold setup or release work
+   indefinitely.
+9. Authenticated secret blobs select the unauthenticated legacy decoder when
+   their in-band version prefix is changed, allowing storage tampering to
+   bypass tag verification.
+10. Cleanup executes the entire post-confirmation snapshot rather than the
+    intersection with what the operator saw, and mutation failures are not
+    reflected in the command result.
+11. Git smart-HTTP can retain sixteen 100 MiB request bodies concurrently,
+    cleanup mutations have no subprocess deadline, and Firecracker signalling
+    failures bypass shared mutation accounting.
+12. SQLite creates the shared control-plane database before its mode is
+    restricted, then suppresses permission-repair failures. Gateway transports
+    also differ in whether copied deploy-key modes are preserved.
+
+These are one design problem: state used to authorize deletion, replacement,
+or resource allocation must be authoritative at the point of use.
+
+## Goals / Success Criteria
+
+- Cleanup never signals a PID or recursively deletes a path solely because it
+  appeared in a pre-confirmation snapshot.
+- Firecracker cleanup proves immediately before action that a PID is still the
+  same Firecracker process and that a run directory is still orphaned.
+- Firecracker process discovery reads NUL-delimited argv from `/proc`; paths
+  are never reconstructed from whitespace-delimited process listings.
+- All backend cleanup discovery primitives raise a typed enumeration error on
+  operational failure. No backend may independently continue from a partial
+  snapshot.
+- Shared cleanup control flow lives in the backend layer; concrete backends
+  override resource-specific discovery and validation primitives rather than
+  each implementing a bespoke failure policy.
+- `list-egress-routes` returns an MCP error when attribution or policy
+  resolution is unavailable. A genuine, authoritatively resolved empty policy
+  remains a successful empty list.
+- Supervisor and Git HTTP request bodies have total read deadlines, and each
+  service bounds concurrent request work. Limits apply to authenticated
+  callers because bottles themselves are untrusted.
+- Gateway child-output pumping treats expected stream closure during shutdown
+  as completion while preserving diagnostics for unexpected failures.
+- Artifact pull, existence-check, and publication requests use explicit
+  network deadlines.
+- Persisted secrets accept only the authenticated format. The schema migration
+  intentionally clears legacy rows; local agents are reprovisioned rather
+  than retaining a ciphertext-controlled downgrade path.
+- Cleanup executes only resources present in both the displayed and current
+  authoritative plans, attempts every approved mutation, and returns failure
+  when any mutation does not complete.
+- Git request bodies spool to disk behind a separate heavy-work semaphore;
+  cleanup commands have configurable deadlines; Firecracker signalling
+  failures aggregate while identity-verification uncertainty still aborts.
+- The shared database directory and file are private before SQLite writes any
+  control-plane state; an inability to enforce those modes aborts startup.
+- Gateway credential directories and files receive explicit private modes
+  inside the gateway, independent of Docker, Apple Container, or SSH copy
+  semantics.
+- Unit tests cover PID/path reuse, partial backend enumeration, transient
+  policy resolution failure, slow bodies, concurrency saturation, and stream
+  closure races.
+
+## Non-goals
+
+- Further decomposition solely to reduce module line counts.
+- Replacing gateway stdlib HTTP services with a web framework.
+- Changing egress matching, DLP decisions, proposal semantics, or backend
+  launch behavior beyond the synchronization required for safe cleanup.
+- Making cleanup silently skip uncertain resources. Uncertainty is an
+  operator-visible failure.
+
+## Design
+
+### Shared backend control flow
+
+Follow the backend architecture rule used by gateway attachment: shared
+behavior lives above concrete backends; subclasses provide primitives, not
+control flow.
+
+Cleanup remains previewable, but confirmation authorizes a *new authoritative
+evaluation*, not blind execution of the displayed object. The shared flow:
+
+1. asks each available backend for a preview;
+2. displays the union and asks for confirmation;
+3. refreshes each non-empty backend plan;
+4. validates destructive identities immediately before action;
+5. aborts loudly if the refreshed plan or any identity cannot be proven safe.
+
+Backend-specific primitives define how to identify a resource. Firecracker
+uses process start identity plus canonical config/run paths; container
+backends use authoritative CLI queries and stable resource names/labels.
+
+Container engines expose destructive name-based commands without a portable
+compare-and-delete operation. Cleanup therefore refreshes after confirmation
+and requires every discovery query to succeed, minimizing but not claiming to
+eliminate the final name-reuse race. A future engine-specific stable-ID
+primitive may close that residual window without moving control flow back
+into each backend.
+
+### Enforcement state versus introspection state
+
+Egress enforcement retains its deny-all fallback because uncertainty must not
+grant network access. Supervisor introspection uses a strict resolver path:
+unattributed callers and resolver failures become typed MCP errors, while a
+successfully resolved policy containing zero routes returns `routes: []`.
+
+### Gateway resource boundaries
+
+Both stdlib servers set a per-connection body deadline before reading and use a
+bounded request executor or semaphore. Saturated capacity fails quickly with a
+service-unavailable response. Existing size caps remain independent:
+supervisor proposals retain the 1 MiB cap and Git pack requests retain their
+larger protocol-appropriate cap.
+
+### Shutdown diagnostics
+
+The gateway output pump catches only stream-closure exceptions expected after
+the supervisor closes child pipes. Other I/O failures remain visible and are
+reported through the supervisor's normal diagnostic channel.
+
+### Shared filesystem security
+
+The common SQLite store owns database creation for every backend. It creates
+the parent directory and an empty database with private modes before opening
+SQLite, repairs existing modes, verifies the resulting state, and propagates
+every enforcement failure. Backend launchers do not duplicate this policy.
+
+The backend-neutral gateway provisioner likewise applies directory and file
+modes after transport copies complete. This avoids relying on copy behavior
+that differs among Docker, Apple Container, and Firecracker's SSH transport.
+
+## Implementation chunks
+
+1. Existing fail-closed security and backend enumeration fixes.
+2. FastAPI orchestrator transport and bounded control-plane bodies.
+3. Egress request-policy and outbound-DLP pipeline extraction.
+4. Supervisor MCP dispatch extraction.
+5. Shared cleanup refresh/revalidation plus authoritative macOS discovery.
+6. Strict supervisor introspection and bounded supervisor/Git HTTP work.
+7. Gateway shutdown log-pump closure handling.
+8. Lossless Firecracker process identities, authoritative Docker cleanup
+   queries, and bounded Firecracker artifact transfers.
+9. Mandatory authenticated secret storage, shared cleanup-plan intersection
+   and mutation accounting, and contained Git backend process failures.
+10. Disk-spooled and separately bounded Git bodies, cleanup command deadlines,
+    and classified Firecracker signalling failures.
+11. Fail-closed shared database creation and backend-neutral gateway credential
+    permissions.
+
+## Open questions
+
+None.

docs/prds/0083-packaged-infra-artifacts.md

@@ -0,0 +1,225 @@
+# PRD 0083: Packaged infrastructure artifacts
+
+- **Status:** Active
+- **Author:** Codex
+- **Created:** 2026-07-27
+- **Issue:** #536
+
+## Summary
+
+Publish the orchestrator and gateway infrastructure before packaging
+bot-bottle, record their immutable identities in the application package, and
+make every production backend acquire those exact artifacts instead of
+building infrastructure during bottle startup. Docker and Apple Container pull
+digest-pinned OCI images; Firecracker pulls checksum-verified rootfs artifacts
+selected by the same packaged release manifest.
+
+## Problem
+
+Starting a bottle currently invokes an infrastructure build on Docker and
+Apple Container even when a healthy orchestrator is already running. Both
+infra composers call `ensure_built()` before their health/currentness checks,
+and both implementations unconditionally invoke their OCI builder. Layer
+caching can make this less expensive, but startup still depends on a working
+builder, build context, base-image registry, and package installation network.
+
+The Apple Container orchestrator also bind-mounts the installed source over the
+package baked into its image. A pinned image alone would therefore not pin the
+code that actually runs.
+
+Firecracker already downloads fixed rootfs artifacts, but it derives their
+versions from the package contents at launch. There is no package-level record
+that says which already-published infrastructure release belongs to an
+installed bot-bottle release.
+
+The result is three different delivery contracts:
+
+- Docker and Apple Container build mutable `:latest` images on the launch host.
+- Firecracker computes an artifact version locally and downloads it.
+- An installed wheel contains Dockerfiles specifically so production startup
+  can rebuild its own infrastructure.
+
+That makes startup slower and less reliable, prevents a packaged application
+from identifying the infrastructure it was tested with, and weakens the
+supply-chain boundary between release production and runtime.
+
+## Goals / Success criteria
+
+- A packaged bot-bottle release contains one validated release manifest with
+  immutable orchestrator and gateway identities.
+- Docker and Apple Container acquire and run the manifest's digest-pinned OCI
+  images without invoking `docker build` or `container build`.
+- Firecracker acquires the manifest's versioned, checksum-verified
+  orchestrator and gateway rootfs artifacts rather than deriving versions at
+  launch.
+- All three backends expose the same backend-neutral `ensure_available`
+  lifecycle contract.
+- A healthy running orchestrator is retained only when its recorded release
+  identity matches the installed package.
+- Apple Container runs the package baked into the orchestrator image; no host
+  source overlays production code.
+- Missing, mutable, malformed, or internally inconsistent release metadata
+  fails closed before infrastructure starts.
+- Source-checkout development retains an explicit local-build mode.
+- Publishing produces the infrastructure first and the installable Python
+  package last, so the package cannot name artifacts that were not published.
+
+## Non-goals
+
+- Changing agent/provider image builds. User Dockerfiles continue to build
+  through the backend-specific agent-image paths.
+- Installing Docker, Apple Container, Firecracker, or other host prerequisites.
+- Replacing Gitea's OCI or generic-package registries.
+- Adding image signing or a transparency log. Immutable digests and existing
+  Firecracker SHA-256 verification are the integrity boundary for this version.
+- Forcing a network download when an immutable, verified artifact is already
+  cached locally. "Pull" means acquire the packaged identity, with safe cache
+  reuse.
+- Supporting live source overlays in packaged production installs.
+
+## Design
+
+### Release manifest
+
+`bot_bottle/release-manifest.json` is generated during package production and
+shipped as package data:
+
+```json
+{
+  "schema": 1,
+  "source_commit": "<40 lowercase hex>",
+  "oci": {
+    "orchestrator": "gitea.dideric.is/didericis/bot-bottle-orchestrator@sha256:<64 hex>",
+    "gateway": "gitea.dideric.is/didericis/bot-bottle-gateway@sha256:<64 hex>"
+  },
+  "firecracker": {
+    "orchestrator": {
+      "version": "<artifact version>",
+      "sha256": "<rootfs.ext4.gz sha256>"
+    },
+    "gateway": {
+      "version": "<artifact version>",
+      "sha256": "<rootfs.ext4.gz sha256>"
+    }
+  }
+}
+```
+
+The runtime loader accepts only schema 1, digest-qualified OCI references,
+non-empty Firecracker versions, 64-character lowercase SHA-256 values, and a
+full source commit. It returns immutable value objects rather than passing raw
+dictionaries through backend code.
+
+The repository carries no pretend production pins. A source checkout or ad-hoc
+Git wheel carries a `development: true` marker and selects local infrastructure
+builds, preserving the contributor and current `install.sh` paths. The release
+packaging workflow must replace that marker with the validated schema above;
+it refuses mutable or incomplete release inputs.
+
+### Packaging order
+
+The release job operates on one tested commit:
+
+1. Build the multi-architecture orchestrator and gateway OCI images.
+2. Push them and resolve their registry manifest digests.
+3. Build Firecracker's orchestrator and gateway rootfs artifacts from the same
+   source and pinned OCI bases.
+4. Publish both generic-package artifacts and retain their versions and
+   compressed-file checksums.
+5. Generate `release-manifest.json` from those published identities.
+6. Build the wheel with that manifest. Do not produce a source distribution:
+   rebuilding an sdist outside this release boundary could replace the pins
+   with the development manifest.
+7. Install the wheel in a clean environment and assert that every manifest
+   identity is valid and retrievable.
+8. Publish only that verified wheel.
+
+The build hook receives the manifest as a file input. It must not query a
+registry or choose versions itself: package builds stay deterministic and
+network-independent once their release inputs exist.
+
+### Backend-neutral lifecycle
+
+Rename the host-side control-plane lifecycle operation from `ensure_built()` to
+`ensure_available()`. The production meaning is "make the package-selected
+artifact locally available and verified." Infrastructure composers call it
+before starting either plane.
+
+The gateway receives the same contract so startup does not continue to rebuild
+the adjacent half of the fixed infrastructure pair.
+
+A source checkout is itself a development boundary and selects the existing
+builders. `BOT_BOTTLE_INFRA_BUILD=local` provides the same explicit override
+for release debugging. Both paths deliberately bypass release metadata.
+Packaged production startup never silently falls back from a failed pull to a
+local build.
+
+### Docker
+
+The Docker backend runs `docker pull` with each digest-qualified manifest
+reference. Docker may reuse its content-addressed local store. Containers are
+created from the immutable reference, and the running container's image ID is
+compared with the locally resolved ID for currentness.
+
+The orchestrator source-hash label is replaced by a release-identity label.
+The label is diagnostic; the image ID comparison remains authoritative.
+
+### Apple Container
+
+The Apple Container backend pulls the same multi-architecture OCI references.
+Because Apple Container's accepted syntax for pull, tag, inspect, and run is
+not identical to Docker's, one utility owns normalization from a digest
+reference to the local runnable name. Integration coverage on the macOS runner
+guards that adapter.
+
+Production orchestrator launch removes the build-root bind mount, `PYTHONPATH`,
+and `BOT_BOTTLE_SOURCE_HASH`. It runs only the code baked into the pulled
+image. Currentness is based on the packaged release identity/image ID.
+
+### Firecracker
+
+The publishing tool may continue computing content-derived artifact versions;
+that is a producer concern. Runtime launch instead reads the version and
+expected compressed checksum from the release manifest.
+
+The artifact cache validates the packaged expected checksum even when a
+previous `.verified` marker exists. The marker records the checksum it
+validated, rather than an unqualified `ok`, so changing package metadata cannot
+adopt an artifact verified against a different expectation.
+
+The existing candidate-directory and explicit local-build paths remain for CI
+and development. Candidate metadata must match the packaged identity unless
+the caller selected local development mode.
+
+### Overrides and failure behavior
+
+`BOT_BOTTLE_ORCHESTRATOR_IMAGE` and `BOT_BOTTLE_GATEWAY_IMAGE` remain useful
+for testing but production accepts them only when they are digest-qualified.
+Mutable tag overrides require explicit local-build mode.
+
+Artifact acquisition errors identify the packaged reference and backend. No
+backend catches an acquisition failure and rebuilds from local source.
+
+## Testing strategy
+
+- Unit-test manifest parsing, immutability validation, missing-manifest errors,
+  package-data inclusion, and deterministic build-hook injection.
+- Assert Docker and Apple production preparation pulls and never builds.
+- Assert explicit local mode retains existing build behavior.
+- Assert the Apple orchestrator command has no source bind mount or
+  `PYTHONPATH`.
+- Assert running-image mismatches recreate infrastructure while matching,
+  healthy instances are retained.
+- Assert Firecracker uses packaged versions/checksums and rejects mismatches in
+  downloaded, cached, and candidate artifacts.
+- Extend wheel-install coverage to inspect the installed manifest.
+- Run the full unit and Docker integration gates; run advisory Firecracker and
+  macOS integration before publishing a release.
+
+## Rollout
+
+Land runtime and publishing support together, but do not publish an installable
+release until its infrastructure artifacts have been published and the
+generated manifest has passed clean-install validation. Existing source
+checkouts use explicit local mode. Existing mutable local images are ignored by
+production acquisition and may be pruned independently.

docs/research/agent-sandbox-landscape.md

@@ -32,9 +32,17 @@ 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.
 
+Updated 2026-07-27 after a scan of recent Show HN launches: **Black LLAB,
+Eve, CloudRouter, Nucleus, yolo-cage, and Sandbox Agent SDK** added as a
+dated entrant cohort. They sharpen the comparison on three axes the original
+table underweighted: the browser/preview loop, parallel-agent operator UX, and
+a provider-neutral automation/session API.
+
 ## Summary
 
-The main table compares bot-bottle against fifteen isolation/sandbox tools.
+The main table compares bot-bottle against fifteen canonical
+isolation/sandbox tools; a later section evaluates six recent HN entrants
+without widening an already unwieldy table.
 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
@@ -542,6 +550,199 @@ them.
   framework runtime is not compromised.
 - **Maturity**: Specification + reference implementation, 2026.
 
+## Recent HN entrants (added 2026-07-27)
+
+These are grouped by launch date rather than promoted into the main table.
+Several are young or sparsely documented, and putting them beside mature
+runtime platforms with false precision would obscure the useful comparison.
+The HN launch posts are the evidence snapshot; feature claims should be
+rechecked against their repositories before relying on them for a security
+decision.
+
+### Black LLAB
+
+- **Source**: https://github.com/isaacdear/black-llab ;
+  HN launch https://news.ycombinator.com/item?id=47402394
+- **Isolation/locality**: Local Docker environment, with an isolated container
+  created for each agent task. Shared host kernel; no stronger boundary is
+  claimed.
+- **Agent integration**: General local/cloud model workspace. Its headline is
+  dynamic routing of simple prompts to local models and complex prompts to
+  hosted models, with code execution and web scraping inside the task
+  container.
+- **Network/credentials**: No default-deny egress, payload inspection, or
+  host-side credential injection documented in the launch.
+- **Competitive read**: Superficial overlap ("a container per agent task"),
+  but not a direct security-policy competitor. Its useful challenge is the
+  integrated model-selection UX, which bot-bottle intentionally leaves to the
+  selected agent provider.
+- **Maturity**: Early solo project; HN launch received 1 point.
+
+### Eve
+
+- **Source**: https://eve.new/ ;
+  HN launch https://news.ycombinator.com/item?id=47721255
+- **Isolation/locality**: Managed, hosted Linux sandbox per user/session
+  (claimed 2 vCPU, 4 GB RAM, 10 GB disk), with filesystem, code execution,
+  headless Chromium, and service connectors.
+- **Agent integration**: End-user OpenClaw-style agent product. An orchestrator
+  routes subtasks to specialist models and can run parallel subagents that
+  coordinate through a shared filesystem. Web UI and iMessage are primary
+  interaction surfaces.
+- **Network/credentials**: Broad connectors are a product feature; the launch
+  does not document bot-bottle-style default-deny route policy, content DLP,
+  or credentials held outside the sandbox.
+- **Competitive read**: Adjacent, not direct. Eve sells a managed colleague;
+  bot-bottle lets an operator run existing coding-agent CLIs under local
+  containment. Eve nevertheless demonstrates the appeal of background work,
+  live progress, browser capability, and mobile notification.
+- **Maturity**: Commercial hosted product; HN launch received 71 points and
+  39 comments.
+
+### CloudRouter
+
+- **Source**: https://github.com/manaflow-ai/manaflow/tree/main/packages/cloudrouter ;
+  HN launch https://news.ycombinator.com/item?id=47006393
+- **Isolation/locality**: Claude Code or Codex runs locally and provisions
+  remote cloud VMs/GPUs for execution. Project files are uploaded to the VM;
+  each machine exposes auth-protected VNC, VS Code, and Jupyter surfaces.
+- **Agent integration**: A skill plus CLI lets the coding agent itself start,
+  command, inspect, and tear down machines. Browser automation is integrated,
+  including snapshots and screenshots. Parallel disposable compute is the
+  central workflow.
+- **Network/credentials**: The launch emphasizes remote resource isolation and
+  authenticated UI endpoints, not default-deny guest egress, payload DLP, or
+  proxy-held application credentials.
+- **Competitive read**: The closest recent workflow competitor. It directly
+  addresses parallel coding agents, environmental conflict, and closing the
+  browser/test loop, but trades local custody for elastic cloud compute.
+  Cloud VMs and GPUs could be a future bot-bottle backend; they do not replace
+  its manifest/policy layer.
+- **Maturity**: Active open-source monorepo project; HN launch received
+  138 points and 36 comments.
+
+### Nucleus
+
+- **Source**: https://github.com/coproduct-opensource/nucleus ;
+  HN launch https://news.ycombinator.com/item?id=46855770
+- **Isolation/locality**: Firecracker microVM with an enforcing MCP tool proxy.
+- **Agent integration/config**: Compositional permission envelope for
+  read/write/run actions. The envelope is non-escalating and can tighten or
+  terminate, with scoped approval tokens for gated operations.
+- **Network/credentials**: Default-deny egress, DNS allowlist, iptables drift
+  detection, time/budget caps, and hash-chained audit logging are claimed.
+  Remote append-only audit storage and attestation were roadmap items at
+  launch.
+- **Competitive read**: Direct on security architecture, especially
+  non-escalating policy and tamper-evident audit. It is an early execution/tool
+  proxy rather than a provider-neutral, one-command coding-agent product. Its
+  tool-level action envelope is semantically finer than bot-bottle's network
+  boundary; bot-bottle is stronger on turnkey agent/provider integration,
+  credential custody, Git mediation, and long-running operator workflow.
+- **Maturity**: Early OSS experiment; HN launch received 3 points.
+
+### yolo-cage
+
+- **Source**: https://github.com/borenstein/yolo-cage ;
+  HN launch https://news.ycombinator.com/item?id=46706796
+- **Isolation/locality**: Local sandbox for running multiple coding agents in
+  YOLO mode. The launch discussion describes a VM boundary.
+- **Agent integration**: Built around the native Claude Code experience and
+  motivated by running many agents in parallel without permission-prompt
+  fatigue.
+- **Network/Git/credentials**: Strict egress filtering, configurable HTTP
+  middleware, and mediated `git`/`gh` dispatch are the main value. The launch
+  discussion explicitly identifies provider credential handling as unfinished
+  and difficult because Claude state spans multiple host paths.
+- **Competitive read**: The closest new threat-model competitor. It shares
+  bot-bottle's premise that filesystem isolation alone is insufficient and
+  that Git plus authorized HTTP channels need mediation. bot-bottle currently
+  leads on cross-provider support, proxy-held Claude/Codex/forge credentials,
+  typed per-role manifests, content DLP, and supervision. yolo-cage's simpler
+  pitch and narrower Claude-first setup may be easier to explain.
+- **Maturity**: Early local tool; HN launch received 60 points and 76 comments.
+
+### Sandbox Agent SDK
+
+- **Source**: https://github.com/rivet-dev/sandbox-agent ;
+  HN launch https://news.ycombinator.com/item?id=46795584
+- **Isolation/locality**: Does not provide the isolation primitive. It runs
+  inside E2B, Daytona, Modal, Cloudflare Containers, Agent Computer, BoxLite,
+  Docker, or another sandbox provider. Embedded mode can also run locally
+  without a sandbox.
+- **Agent integration**: Provider-neutral Rust server/SDK exposing a common
+  HTTP/SSE/OpenAPI interface across Claude Code, Codex, OpenCode, Cursor, Amp,
+  and Pi, plus a universal event/session schema for external storage and
+  replay. It also exposes filesystem, managed-process, terminal, MCP, skills,
+  custom-tool, and computer-use APIs. TypeScript is the primary SDK surface.
+- **Network/credentials**: Delegated to the chosen sandbox provider.
+- **Credential posture**: Its documented convenience command extracts real
+  OpenAI/Anthropic credentials from local agent configuration and passes them
+  as environment variables into the sandbox. That is materially weaker than
+  bot-bottle's host-side credential custody, but it is an integration choice,
+  not a structural limitation: a sandbox provider could put a credential
+  proxy underneath the same SDK.
+- **Competitive read**: A serious architectural threat despite not supplying
+  isolation. Sandbox Agent is trying to standardize the boundary *above* the
+  sandbox: one client protocol, session model, and UI/control surface across
+  every coding agent and runtime. If that boundary becomes the ecosystem
+  standard, users and application builders may choose a sandbox provider plus
+  Sandbox Agent rather than a vertically integrated launcher. bot-bottle's
+  manifests would then be valuable chiefly as a local policy/backend
+  implementation unless they expose an equally usable control contract.
+- **Maturity**: Apache 2.0, ~1.5k stars and 426 commits at the 2026-07-27
+  check; HN launch received 41 points.
+
+#### Why the Sandbox Agent architecture is strategically different
+
+The manifest and the universal control protocol solve different layers:
+
+- A bot-bottle manifest is a **trusted launch-time policy composition**. It
+  selects the agent role, isolation backend, image, skills, egress routes,
+  credentials, Git mediation, and supervision policy. Crucially, identity and
+  secret references live on the host side of the trust boundary.
+- Sandbox Agent is a **runtime control and observation protocol**. A remote
+  client creates sessions, sends messages, handles permissions, configures
+  skills/MCP, manipulates files/processes/desktops, and streams normalized
+  events. It deliberately delegates sandbox lifecycle, Git management,
+  storage, network policy, and credential security to other products.
+
+That makes it complementary in a component diagram but competitive in product
+architecture. The layer that becomes the stable integration point tends to own
+the ecosystem. Three plausible threat paths matter:
+
+1. **Standard control plane, interchangeable runtimes.** Applications integrate
+   once with Sandbox Agent and treat E2B, Daytona, BoxLite, Docker, or a future
+   local microVM as replaceable compute. A provider that bundles adequate
+   egress and credential custody makes bot-bottle's end-to-end launcher less
+   necessary.
+2. **Policy grows upward.** Sandbox Agent already configures permissions,
+   skills, MCP, custom tools, filesystem/process access, and computer use. If
+   it adds a declarative, host-verifiable policy document, the overlap with
+   agent/bottle manifests becomes substantial even if enforcement remains
+   delegated.
+3. **UI and session ownership.** Its universal transcript schema, Inspector,
+   React components, event replay, and remote terminal/computer APIs can become
+   the natural basis for desktop, web, and mobile agent managers. bot-bottle's
+   security layer could remain stronger while losing the operator surface and
+   distribution channel.
+
+The counter-position is not to claim that manifests and an API are mutually
+exclusive. The defensible split is:
+
+- bot-bottle owns the trusted policy and enforcement plane outside the agent;
+- a provider-neutral protocol owns agent process control and normalized
+  events; and
+- the operator UI consumes both.
+
+This suggests an explicit compatibility decision rather than parallel,
+accidental protocol design: evaluate running Sandbox Agent inside a bottle and
+exposing it only through the authenticated bot-bottle control plane. If its
+schema is suitable, adopting it could turn a threat into an integration while
+keeping manifests as the higher-trust policy source. If it is unsuitable,
+bot-bottle should still publish a stable provider-neutral session/event API so
+frontends do not depend on Claude/Codex/Pi-specific process behavior.
+
 ## Comparison table
 
 *Isolation/sandbox tools only. AGT and OAP are governance layers — see their per-project notes above.*
@@ -616,6 +817,70 @@ 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.
 
+**Recent entrants change two parts of this read.** yolo-cage is closer to the
+actual threat model than agent-safehouse or litterbox: it combines a VM-style
+boundary with mediated Git and filtered HTTP specifically for parallel coding
+agents. Sandbox Agent SDK is the more important strategic entrant even though
+it supplies no isolation. It can become the standard agent-control layer above
+all of these runtimes, including a future bot-bottle backend. CloudRouter is
+the clearest workflow challenge because its browser/desktop/GPU loop makes
+parallel agents visibly more capable, not merely safer.
+
+## Gap evaluation after the 2026-07-27 entrant scan
+
+### Material gaps
+
+1. **A stable provider-neutral control and event protocol.** This is the
+   largest newly visible gap. bot-bottle normalizes launch/provisioning across
+   providers, but an external UI or orchestrator still lacks one documented
+   contract for creating a Claude/Codex/Pi session, sending input, handling
+   permission/supervision events, streaming normalized output, reconnecting,
+   and replaying history. Sandbox Agent SDK addresses exactly this layer and
+   is already portable across many sandbox providers.
+2. **Browser/preview closure.** CloudRouter and Eve make a browser or desktop
+   part of the standard agent environment and expose screenshots/live viewing
+   to the operator. bot-bottle can run dev servers and supports nested
+   containers, but it does not present a first-class browser/computer-use
+   primitive or an auth-protected preview surface. For coding agents expected
+   to verify UI work, this is a real product gap.
+3. **Unified parallel-session operator UX.** Named persistent bottles and
+   supervision provide the substrate, but the recent products make task
+   switching, live progress, notifications, terminal attach, diffs, and
+   session history the product. Security depth will not compensate for a
+   visibly rougher daily loop.
+4. **Normalized transcript persistence and replay.** bot-bottle preserves
+   provider-specific state for resume; it does not expose a provider-neutral
+   event record suitable for audit, replay, analytics, or a web/mobile client.
+   This is both a UX gap and an audit gap.
+
+### Important, but not necessarily bot-bottle features
+
+- **Cloud VM/GPU provisioning.** Valuable for elastic workloads and could be a
+  backend, but it conflicts with the local-custody default and should not
+  displace core policy work.
+- **Automatic model routing.** Black LLAB and Eve sell task-to-model routing.
+  bot-bottle's provider-template boundary can host that choice without making
+  it part of the trusted sandbox policy.
+- **A thousand SaaS connectors.** This broadens capability and blast radius.
+  The bot-bottle-native answer should remain explicit, scoped forge/egress
+  associations rather than connector count as a goal.
+- **SDK-driven sandbox lifecycle as the primary configuration model.** Useful
+  for platform builders, but not a replacement for reviewable, host-owned
+  manifests. A control API and a declarative policy source are compatible;
+  neither should silently become the other.
+
+### Areas where bot-bottle remains ahead
+
+- real provider and forge credentials remain outside the agent process rather
+  than being extracted into its environment;
+- authorized HTTP payloads are scanned, not merely destination-filtered;
+- Git writes traverse a distinct gate with secret scanning and host-held
+  upstream credentials;
+- role policy is host-owned, composable, and separate from untrusted repo
+  content; and
+- local Firecracker/Apple Container execution preserves operator custody
+  without requiring a hosted sandbox platform.
+
 ## Borrowable ideas
 
 ### Already shipped or otherwise addressed
@@ -642,6 +907,19 @@ capability passing rather than kernel boundaries.
 
 ### Still worth considering
 
+- **Sandbox Agent compatibility or an equivalent stable protocol (highest
+  priority):** spike running its server inside a bottle behind bot-bottle's
+  authenticated control plane. Compare its session/event schema, permission
+  model, restore semantics, and provider coverage with current provider
+  adapters. Adopt compatibility if it preserves the host-owned trust boundary;
+  otherwise specify bot-bottle's own stable API before building another UI.
+- **First-class browser/preview loop** (from CloudRouter and Eve): give a
+  bottle an optional browser/computer-use capability plus an operator-visible,
+  authenticated preview/screenshot surface. Treat its network access as part
+  of the bottle policy, not an implicit bypass.
+- **Provider-neutral transcript/event persistence** (from Sandbox Agent SDK):
+  retain enough normalized structure for replay and audit while preserving the
+  provider-native state needed for exact resume.
 - **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.
@@ -652,10 +930,11 @@ capability passing rather than kernel boundaries.
   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).
+Not worth borrowing: SDK-first *policy configuration* as used by boxlite /
+microsandbox (cuts against the reviewable declarative-manifest stance), and
+the hosted-SaaS custody model of tilde.run (cuts against the "infrastructure I
+control" goal). A provider-neutral runtime-control API is a separate concern
+and is worth borrowing.
 
 ## Publishing and positioning verdict
 
@@ -679,9 +958,15 @@ bot-bottle remains unusual in combining:
 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.
+plumbing. The main competitive risks are now:
+
+- a local wrapper such as yolo-cage, claudebox, or Docker sbx growing a
+  role-manifest and credential-custody layer;
+- Sandbox Agent SDK becoming the standard control/session boundary and making
+  the runtime beneath it interchangeable; and
+- GUI products such as SuperHQ or CloudRouter adding equivalent policy and
+  audit depth before bot-bottle closes the browser/preview and
+  parallel-session UX gaps.
 
 ## Caveats
 

docs/research/built-in-supervisor-design.md

@@ -6,7 +6,7 @@ Can bot-bottle grow a built-in supervisor — TUI inventory plus PR-feedback rou
 
 ## Context
 
-bot-bottle today is a fleet *executor*: `./cli.py start <agent>` brings up one bottle (agent container + pipelock + optional git-gate + optional cred-proxy on a per-bottle internal network), and `cli.py` tears it down when the session ends. There is no inventory view, no idle-detection, no automated reaction to PR or CI events. In parallel use, a human is the supervisor — opening one terminal per bottle, switching between them, and watching upstream PR state by hand.
+bot-bottle today is a fleet *executor*: `bot-bottle start <agent>` brings up one bottle (agent container + pipelock + optional git-gate + optional cred-proxy on a per-bottle internal network), and `cli.py` tears it down when the session ends. There is no inventory view, no idle-detection, no automated reaction to PR or CI events. In parallel use, a human is the supervisor — opening one terminal per bottle, switching between them, and watching upstream PR state by hand.
 
 A separate survey of the broader ecosystem ([agent control dashboards research, mid-2026](https://gitea.dideric.is/didericis/consilium-research/src/branch/main/developer-workflow/agent-control-dashboards-2026-05-24.md)) sorts dashboards into five tiers (session managers, parallel runners, Kanban boards, mission-control SPAs, observability backends). The earlier first-pass conclusion was that a full SPA tier conflicts with bot-bottle's isolation model. This doc reconsiders the smaller question: a TUI supervisor in the existing Python CLI.
 
@@ -25,13 +25,13 @@ A supervisor doesn't have to be heavy. A TUI built into the existing Python CLI,
 
 Three layers, each independently useful, in order of ambition:
 
-### 1. `./cli.py status` — read-only inventory
+### 1. `bot-bottle status` — read-only inventory
 
 Reads `docker ps` filtered by a bottle label and tails each bottle's session log. Reports per bottle: name, agent, uptime, last-activity timestamp, token spend if available, associated PR/branch if recorded.
 
 No new daemons. No new ports. No new credentials. ~100 lines.
 
-### 2. `./cli.py watch` — TUI over the same data
+### 2. `bot-bottle watch` — TUI over the same data
 
 Same data as `status`, rendered with auto-refresh and keyboard shortcuts that shell out to the existing `cli.py attach / stop / start` commands.
 
@@ -39,7 +39,7 @@ Library choice: prefer the stdlib `curses` module to stay stdlib-first; fall bac
 
 This is the Claude Squad / tmux-agent-status pattern, applied to bottles instead of tmux sessions. The whole category exists *because* a TUI is the lightweight shape that doesn't require what the SPA tier requires.
 
-### 3. `./cli.py supervise` — PR feedback router
+### 3. `bot-bottle supervise` — PR feedback router
 
 The optional, more ambitious layer. The bottle manifest gains an optional field:
 
@@ -49,7 +49,7 @@ pr_watch:
   branch: agent/task-42
 ```
 
-`./cli.py supervise` polls the named upstream for new review comments and CI failures on `branch`. When one fires, it surfaces as a desktop notification or a flash in the TUI. The human decides what to do with the feedback — there is no autonomous loop that feeds the comment back into a bottle's next prompt (see "Where to be conservative" for why).
+`bot-bottle supervise` polls the named upstream for new review comments and CI failures on `branch`. When one fires, it surfaces as a desktop notification or a flash in the TUI. The human decides what to do with the feedback — there is no autonomous loop that feeds the comment back into a bottle's next prompt (see "Where to be conservative" for why).
 
 The polling token is a **host** token (the same `GH_PAT` / Gitea token the host already keeps in shell env), not a bottle credential. The supervisor never holds bottle secrets.
 
@@ -62,7 +62,7 @@ The load-bearing question is whether the supervisor introduces the privileged-ch
 | Reaching into running bottles | Supervisor reads `docker ps` and host-side log files. The host already sees both — Docker is the trust boundary, the supervisor is on the host side of it. |
 | Holding bottle credentials | The polling token is a host token. The supervisor never receives `bottle.cred_proxy.routes` entries; it has no path to them. |
 | Bridging between bottles | The supervisor does not relay state from bottle A to bottle B. It relays *upstream PR state* to a bottle's next prompt — and only if the manifest opts in. |
-| New attack surface | All "control" actions go through `./cli.py start <agent>`, which already enforces the manifest. The supervisor is an automated caller of the existing CLI, not a parallel control plane. |
+| New attack surface | All "control" actions go through `bot-bottle start <agent>`, which already enforces the manifest. The supervisor is an automated caller of the existing CLI, not a parallel control plane. |
 
 The boundary stays at the bottle wall. The supervisor looks outward at git/PR state and downward at Docker; it does not look *inward* through pipelock.
 

docs/research/claude-code-pane-in-dashboard.md

@@ -13,11 +13,11 @@ What's the cheapest path to that, and where does it bottom out?
 
 ## What "interact" means
 
-Today the flow is bimodal. `./cli.py start <agent>` brings the
+Today the flow is bimodal. `bot-bottle start <agent>` brings the
 bottle up and immediately drops you into an interactive
 `docker exec -it bot-bottle-<slug> claude ...` — claude-code
 owns the whole terminal until you Ctrl-D out, at which point the
-bottle tears down. The dashboard (`./cli.py dashboard`) is a
+bottle tears down. The dashboard (`bot-bottle dashboard`) is a
 *separate* invocation that watches across bottles but never
 exposes the claude TUI itself.
 
@@ -107,7 +107,7 @@ What's not good:
 
 This is the v1 the project's existing code-shape strongly
 prefers. It clears the bar of "let me talk to claude-code
-without quitting `./cli.py dashboard`."
+without quitting `bot-bottle dashboard`."
 
 ## Option 2: Embedded emulator
 

docs/research/malicious-commit-scanning.md

@@ -84,7 +84,7 @@ dangerous changes before they left*.
    (`supervise_gitleaks_allow`, [`git_gate_render.py`](../../bot_bottle/git_gate_render.py)).
    Extend the same flow to **high-risk file classes**: any commit touching
    CI/build/deploy scripts, auth/crypto code, egress config, or
-   adding/changing dependencies → route to `./cli.py supervise`. This is
+   adding/changing dependencies → route to `bot-bottle supervise`. This is
    attribution/policy, not detection, and it's the strongest thing here —
    a human on exactly the temporal-escape surfaces.
 4. **LLM semantic diff-review — the behavioral backstop.** The only
@@ -146,7 +146,7 @@ preserves the bottom-up distribution funnel.
   **governed code-egress review**, not "we resell inference" (the
   monetization notes explicitly warn against reselling compute).
 - **The web-console supervise/review flow — the strongest anchor.** Turn
-  the CLI `./cli.py supervise` approval into a real review surface:
+  the CLI `bot-bottle supervise` approval into a real review surface:
   rendered diff + finding context, approve/reject, **who-approved audit
   trail, RBAC on approvers, mobile/phone-control** (ties to the
   dashboard/vault north star). This is "central enforcement +

docs/research/manifest-format-and-grouping.md

@@ -100,7 +100,7 @@ resolver globs each directory.
   becomes file ops (mkdir, mv, rm) instead of editing one file.
   Power users prefer that; new users may not.
 - Discovery requires `ls`, not "grep one file." Tooling helps
-  (e.g. `./cli.py list`) but the manifest is no longer a single
+  (e.g. `bot-bottle list`) but the manifest is no longer a single
   artifact to email or ship.
 - Atomicity: swapping a bottle name across agents touches
   multiple files. Git handles this fine; a one-shot text editor
@@ -364,7 +364,7 @@ wins; none of the body-prose or dependency story.
   warns / ignores / breaks. If it warns, we'd want a different
   field name (e.g. `bot-bottle-bottle`) or a namespaced block.
 - **Migration story.** Is the project willing to ship a one-shot
-  `./cli.py migrate-manifest` command that does the JSON → MD
+  `bot-bottle migrate-manifest` command that does the JSON → MD
   conversion? Or do users just rewrite by hand from the new docs?
 - **Bottle file body content.** If most bottle .md files have an
   empty body, is the MD-with-frontmatter format still warranted?

docs/research/polish-priorities.md

@@ -34,7 +34,7 @@ on top of working onboarding.
 
 A first-time user today goes through five steps: install Docker,
 install `uv`, set `BOT_BOTTLE_CLAUDE_OAUTH_TOKEN`, write
-`bot-bottle.json`, run `./cli.py start`. One of those is
+`bot-bottle.json`, run `bot-bottle start`. One of those is
 "author a JSON manifest." Polished tools in this category let
 users skip that step on day one. The fix is an `init` subcommand
 that drops a working `bot-bottle.json` with a default `coder`

docs/research/remote-docker-vm-isolation.md

@@ -131,7 +131,7 @@ The minimum-viable workflow, no bot-bottle code changes:
 3. SSH in.
 4. `git clone` bot-bottle on the VM, drop a manifest in place,
    inject `BOT_BOTTLE_CLAUDE_OAUTH_TOKEN` via the provider's secrets path.
-5. `./cli.py start <agent>` — the existing launcher handles the rest.
+5. `bot-bottle start <agent>` — the existing launcher handles the rest.
 6. On exit: destroy the VM. No host artifacts persist.
 
 For the "VPN pivot" failure mode, see

docs/research/sandbox-agent-sdk-architecture.md

@@ -0,0 +1,536 @@
+# Sandbox Agent SDK and bot-bottle: protocol versus product
+
+This note asks whether [Sandbox Agent SDK](https://github.com/rivet-dev/sandbox-agent)
+and bot-bottle compete for the same architectural layer, whether bot-bottle
+can productize the turnkey ecosystem/DX layer above it, and how far the
+Docker/OCI analogy actually holds.
+
+Research conducted 2026-07-27. Sandbox Agent SDK was at the `0.4.x` line,
+Apache 2.0, and documented support for Claude Code, Codex, OpenCode, Cursor,
+Amp, and Pi at the time of review.
+
+## Summary
+
+**The projects are complementary at the component boundary and competitive at
+the product boundary.** Sandbox Agent SDK normalizes how software controls a
+coding-agent process inside an arbitrary sandbox. bot-bottle decides what
+sandbox to create, what trusted role and policy it receives, how credentials
+and Git access cross the boundary, how traffic is constrained, and how an
+operator launches and supervises the result.
+
+The Docker analogy is useful with one correction:
+
+- Sandbox Agent SDK is not equivalent to Linux container APIs or OCI itself.
+  It is closer to a **containerd shim plus a portable exec/session API for
+  coding agents**. It adapts incompatible agent processes to one HTTP/SSE
+  contract.
+- A future independent agent-session specification would be the closer OCI
+  analogue.
+- bot-bottle can credibly occupy the **Docker Engine / Compose / Desktop**
+  layer: packaging, policy composition, lifecycle, networking, credentials,
+  storage, operator UX, and a one-command experience above interchangeable
+  agent adapters and isolation runtimes.
+
+That is a viable position, but “turnkey wrapper” undersells it. A thin wrapper
+is replaceable. The valuable product is a **turnkey, policy-first coding-agent
+runtime** whose manifest compiles trusted operator intent into multiple
+enforcement planes. Sandbox Agent SDK may be one internal process-control
+component of that product.
+
+The recommended direction is:
+
+1. Keep the bot-bottle manifest as the host-owned source of trusted policy.
+2. Spike Sandbox Agent SDK as the in-bottle provider/session adapter.
+3. Expose a stable, provider-neutral bot-bottle control API, compatible with
+   Sandbox Agent where practical.
+4. Keep security decisions and authoritative audit outside the sandbox.
+5. Build the ecosystem around policy packs, agent images, skills, backends,
+   operator UI, and trusted integrations—not around a proprietary transcript
+   protocol.
+
+## What each project is today
+
+### Sandbox Agent SDK
+
+Sandbox Agent is a Rust server that runs alongside the coding agent. A client
+connects over HTTP, streams events over SSE, and uses one API across agent
+implementations. Its documented surface includes:
+
+- creating and restoring agent sessions;
+- sending messages and streaming normalized events;
+- handling permissions;
+- configuring MCP servers, skills, and custom tools;
+- filesystem and managed-process APIs;
+- interactive terminal access;
+- computer-use/desktop operations;
+- a universal session/transcript schema;
+- an Inspector UI, React components, CLI, TypeScript SDK, and OpenAPI spec.
+
+It can run in embedded mode or inside E2B, Daytona, Modal, Cloudflare
+Containers, Agent Computer, BoxLite, Docker, and other environments. It
+explicitly leaves these concerns to the caller or sandbox provider:
+
+- sandbox creation and lifecycle;
+- Git repository management;
+- durable session storage;
+- network policy;
+- isolation strength; and
+- secure credential delivery.
+
+Its documented credential convenience path extracts real provider credentials
+from local agent configuration and passes them into the sandbox environment.
+That is convenient but is not an acceptable security boundary for bot-bottle.
+
+Sources:
+
+- [Sandbox Agent repository and architecture](https://github.com/rivet-dev/sandbox-agent)
+- [Sandbox Agent documentation](https://sandboxagent.dev/docs)
+- [HTTP API](https://sandboxagent.dev/docs/api-reference)
+- [Universal session/transcript schema](https://sandboxagent.dev/docs/session-transcript-schema)
+
+### bot-bottle
+
+bot-bottle is a host-side launch, policy, and enforcement system for existing
+coding-agent CLIs. Its current architecture includes:
+
+- agent and bottle manifests with composition via `extends:`;
+- a host-only trust boundary for roles, identity, and secret references;
+- provider templates and plugins for Claude Code, Codex, Pi, and custom
+  providers;
+- Firecracker on KVM Linux and Apple Container on macOS, with Docker fallback;
+- image construction and provider-specific provisioning;
+- default-deny inspected egress with path/method/header policy;
+- payload DLP on authorized channels;
+- real credentials held outside the agent and injected by the gateway;
+- Git mediation, upstream credential custody, and gitleaks scanning;
+- a per-host authenticated orchestrator and shared gateway;
+- named bottle lifecycle, resume, supervision, and audit state; and
+- a CLI/TUI intended to make full-permission agents operationally tolerable.
+
+The provider layer currently normalizes launch-time concerns—command, image,
+prompt delivery, files, skills, environment, verification, and provider-owned
+egress routes. It does **not** yet expose a stable provider-neutral runtime
+contract for sessions, messages, transcripts, terminals, or normalized events.
+That is the gap Sandbox Agent directly illuminates.
+
+Sources in this repository:
+
+- [`README.md`](../../README.md)
+- [`0070-per-host-orchestrator.md`](../prds/0070-per-host-orchestrator.md)
+- [`0026-agent-provider-templates.md`](../prds/0026-agent-provider-templates.md)
+- [`0053-user-provider-plugins.md`](../prds/0053-user-provider-plugins.md)
+- [`agent_provider.py`](../../bot_bottle/agent_provider.py)
+
+## The layer model
+
+The cleanest architecture has four layers:
+
+| Layer | Responsibility | Likely owner |
+|---|---|---|
+| Operator product | Install, select a role, launch, observe, intervene, resume, review changes | bot-bottle |
+| Trusted policy and lifecycle | Compose manifest, choose backend/image, hold credentials, enforce egress/Git, persist authoritative audit | bot-bottle |
+| Agent control protocol | Start provider process, create session, send input, stream normalized events, terminal/computer operations | Sandbox Agent or a compatible protocol |
+| Isolation primitive | VM/container/process boundary, filesystem, CPU/memory, networking substrate | Firecracker, Apple Container, Docker, E2B, Daytona, BoxLite, etc. |
+
+The important boundary is between trusted policy/lifecycle and agent control.
+The agent-control daemon runs in the environment being treated as untrusted.
+It can report what the agent says happened, but it cannot authoritatively prove
+that policy was enforced. Egress decisions, credential custody, Git scanning,
+bottle identity, and security audit must remain outside it.
+
+### Proposed composition
+
+```text
+operator UI / CLI / API
+          |
+          v
+bot-bottle orchestrator  (trusted)
+  - resolves manifest
+  - owns bottle identity and lifecycle
+  - stores authoritative audit
+  - authenticates clients
+          |
+          +--------------------------+
+          |                          |
+          v                          v
+isolation backend              shared gateway  (trusted)
+Firecracker / Apple / Docker   - egress policy + DLP
+          |                    - credential injection
+          |                    - Git mediation
+          v
+bottle / guest  (untrusted)
+  - Sandbox Agent server
+  - Claude Code / Codex / Pi subprocess
+  - workspace, skills, MCP configuration
+```
+
+The bot-bottle manifest would compile into both sides:
+
+- **outside the bottle:** backend, network, egress, credentials, Git,
+  supervision, identity, and authoritative lifecycle;
+- **inside the bottle:** selected provider, prompt, skills, MCP configuration,
+  startup arguments, and non-secret session metadata.
+
+Sandbox Agent should never receive real secrets merely because its API offers
+a credential extraction helper. Provider and forge requests should continue
+to use bot-bottle's placeholder/proxy pattern.
+
+## How accurate is the Docker/OCI analogy?
+
+### The useful part
+
+The container ecosystem separates low-level execution from a product that
+ordinary developers operate. OCI defines interoperable image, runtime, and
+distribution specifications. Docker Engine adds a daemon, API, CLI, object
+model, images, networks, volumes, and lifecycle; Docker Desktop and related
+products add installation, updates, UI, integrations, policy, and team
+workflows.
+
+The same separation can exist for coding agents:
+
+| Container ecosystem | Agent-sandbox ecosystem |
+|---|---|
+| OCI/runtime contract | A future open agent session/event contract |
+| `runc` / runtime adapter | Claude/Codex/Pi adapter |
+| containerd shim and task/exec API | Sandbox Agent server and HTTP/SSE session API |
+| containerd / CRI-style lifecycle | Sandbox-provider lifecycle APIs |
+| Docker Engine / Compose | bot-bottle orchestrator + manifests + backends + gateway |
+| Docker Desktop / Hub ecosystem | bot-bottle desktop/mobile UX, policy packs, agent images, skills, trusted integrations |
+
+Sandbox Agent makes coding-agent processes portable in roughly the way a shim
+makes runtimes consumable through a common lifecycle interface. bot-bottle can
+make the entire safe-agent system usable without asking the operator to
+assemble that plumbing.
+
+Official container references:
+
+- [Open Container Initiative](https://opencontainers.org/)
+- [OCI Runtime Specification](https://github.com/opencontainers/runtime-spec)
+- [Docker Engine architecture](https://docs.docker.com/engine/)
+- [Docker alternative runtimes and containerd shims](https://docs.docker.com/engine/daemon/alternative-runtimes/)
+
+### Where the analogy breaks
+
+1. **Sandbox Agent is an implementation, not an independent standard.**
+   Its OpenAPI document is public, but the project currently owns the server,
+   adapters, schema, and evolution. OCI is an independently governed set of
+   specifications with multiple implementations.
+2. **It sits above, not below, the isolation boundary.** Linux namespaces,
+   cgroups, VMs, and OCI runtimes create the boundary. Sandbox Agent controls a
+   process after some other system has created that boundary.
+3. **It reaches into product territory.** Inspector, React components,
+   computer-use APIs, skills/MCP configuration, transcripts, and restoration
+   are not merely low-level primitives. Sandbox Agent can continue growing
+   upward into the same UI and orchestration space bot-bottle might occupy.
+4. **Coding agents are semantically uneven.** Normalizing a container
+   lifecycle is easier than claiming full behavioral parity across Claude
+   Code, Codex, Cursor, Amp, OpenCode, and Pi. A universal schema can become a
+   lowest common denominator or accumulate provider-specific escape hatches.
+5. **The security contract is not standardized.** An agent-session API says
+   little about whether credentials are visible, egress is controlled, Git is
+   mediated, or audit is trustworthy. Those are core bot-bottle concerns.
+
+The positioning should therefore say “Docker-like product layer above an open
+agent-control protocol,” not “Sandbox Agent is OCI” or “bot-bottle implements
+OCI for agents.”
+
+## Can bot-bottle be the turnkey product layer?
+
+Yes, if it owns substantially more than launch syntax.
+
+The turnkey promise is:
+
+> Choose a trusted role, point it at a project, and run any supported coding
+> agent with full permissions. bot-bottle builds the environment, isolates it,
+> supplies only the capabilities it needs, keeps credentials outside, mediates
+> external writes, and gives the operator one place to watch and intervene.
+
+That product has several defensible jobs:
+
+### 1. Packaging and reproducibility
+
+- provider and toolchain images;
+- pinned, verified build inputs;
+- skills and MCP configuration;
+- role/bottle composition;
+- cached startup and portable environment definitions; and
+- compatibility testing across agents and backends.
+
+### 2. Trusted policy compilation
+
+The manifest is valuable because one reviewable document compiles into:
+
+- an isolation plan;
+- gateway routes and DLP policy;
+- credential slots;
+- Git-gate repositories and identities;
+- provider configuration;
+- supervision behavior; and
+- operator-facing preflight.
+
+Sandbox Agent's runtime configuration does not replace this. The policy must
+be resolved before an untrusted guest or agent-control daemon exists.
+
+### 3. Security enforcement
+
+- dedicated-kernel isolation where available;
+- no direct guest route to the internet;
+- credentials injected outside the agent;
+- content inspection on allowed destinations;
+- Git secrets scanning and upstream-key custody;
+- fail-closed policy resolution; and
+- authoritative host-side audit.
+
+This is the strongest current differentiation from a generic
+“Sandbox Agent + Docker/E2B” assembly.
+
+### 4. Lifecycle and operations
+
+- install and host preflight;
+- image build/update;
+- start, stop, resume, cleanup, and migration;
+- concurrent named agents;
+- state recovery after crashes;
+- live supervision and policy remediation; and
+- backend selection without changing the role definition.
+
+### 5. Ecosystem and DX
+
+A product layer can support:
+
+- curated provider images;
+- signed policy/bottle packs;
+- reusable role templates;
+- skills and MCP bundles;
+- backend plugins;
+- an authenticated desktop/web/mobile operator client;
+- browser/preview integration;
+- normalized transcripts and change review; and
+- team policy distribution and compliance exports.
+
+The analogy to Docker is strongest here: users adopt the coherent workflow and
+ecosystem, not because the low-level process API is proprietary.
+
+## Business and product positioning
+
+“Turnkey wrapper” is understandable internally but weak externally. It implies
+that the hard work lives underneath and that another wrapper can replace it.
+Prefer one of:
+
+- **The policy-first runtime for coding agents**
+- **Run any coding agent with full permissions, without giving it your host or
+  credentials**
+- **A turnkey local control plane for isolated coding agents**
+- **Docker-like packaging and operations for coding agents, with the security
+  boundary outside the agent**
+
+The open/product split could resemble the container ecosystem:
+
+### Open foundation
+
+- manifest schema and composition;
+- local CLI and core orchestrator;
+- provider adapters;
+- Firecracker/Apple Container/Docker backends;
+- gateway policy format and enforcement;
+- Sandbox Agent compatibility;
+- local audit and supervision; and
+- conformance tests for providers/backends/policy.
+
+### Productizable ecosystem/DX
+
+- polished desktop and mobile clients;
+- fleet/remote-host management;
+- signed and curated role/image/policy registry;
+- team policy distribution and administrative controls;
+- durable searchable transcripts and audit exports;
+- SSO, RBAC, retention, and tamper-evident audit;
+- managed update/compatibility channels;
+- remote browser/preview relay;
+- enterprise support; and
+- optional managed build/cache infrastructure.
+
+OCI itself is not the thing Docker sells. Interoperability expands the market;
+the product captures value through reliable packaging, workflow, distribution,
+management, and trust. bot-bottle should follow that logic rather than trying
+to make its session protocol the moat.
+
+## Strategic threat from Sandbox Agent
+
+Sandbox Agent is a real threat for three reasons:
+
+1. **It can become the integration default.** A frontend or agent platform can
+   integrate one API and choose among many agents and sandbox vendors.
+2. **It can own session data and UI.** The universal event schema, Inspector,
+   React components, restoration, terminal, and computer-use APIs give it a
+   natural path toward the operator surface.
+3. **Sandbox providers can move upward.** If E2B, Daytona, BoxLite, or another
+   runtime combines Sandbox Agent with adequate network policy and credential
+   custody, it can offer much of the turnkey stack.
+
+The threat is not that its manifest syntax is better. It currently has no
+equivalent trusted policy composition. The threat is that **the ecosystem may
+standardize around its API before bot-bottle has a stable external control
+surface**. In that world bot-bottle is evaluated as one sandbox provider,
+while the SDK and its consumers own the user relationship.
+
+## Why bot-bottle can still win its layer
+
+Sandbox Agent's scope exclusions align with bot-bottle's deepest work:
+
+- it does not choose or operate the sandbox provider;
+- it does not mediate Git;
+- it does not own network policy;
+- it does not securely deliver credentials;
+- it does not durably store sessions; and
+- it cannot make guest-generated telemetry authoritative.
+
+Those are not incidental features. Together they define the trusted system
+around an untrusted coding agent. bot-bottle also has a narrower and coherent
+initial customer: a developer or small operator who wants existing agent CLIs
+to run locally with broad permissions and bounded consequences.
+
+The durable advantage is therefore:
+
+> Sandbox Agent makes agents controllable. bot-bottle makes them safe and
+> operable.
+
+That sentence remains true only if bot-bottle closes its operator-DX gaps.
+Security without a browser/preview loop, stable API, normalized session view,
+and good parallel-task UX risks becoming an invisible backend feature.
+
+## Integration options
+
+### Option A — Embed Sandbox Agent inside each bottle
+
+bot-bottle launches Sandbox Agent as the provider process supervisor and
+connects it to the host orchestrator through a bottle-scoped authenticated
+channel.
+
+**Benefits**
+
+- immediate provider-neutral session API;
+- more supported agents;
+- normalized streaming and transcripts;
+- terminal, filesystem, process, and computer-use primitives;
+- Inspector/React ecosystem; and
+- less provider-specific reverse engineering in bot-bottle.
+
+**Risks**
+
+- `0.x` API/schema churn;
+- extra binary and release-supply-chain dependency;
+- lowest-common-denominator normalization;
+- conflict with provider-native resume state;
+- an in-guest daemon is attacker-controlled after guest compromise;
+- duplicate orchestration responsibilities; and
+- upstream can move into policy/lifecycle and compete more directly.
+
+**Security rule**
+
+Treat every event and state claim from Sandbox Agent as untrusted telemetry.
+Never delegate egress authorization, credential release, bottle identity,
+authoritative audit, or Git policy to it.
+
+### Option B — Implement a Sandbox Agent-compatible endpoint
+
+bot-bottle maps the external protocol onto its existing provider adapters and
+process model without running the upstream server.
+
+**Benefits**
+
+- ecosystem compatibility with tighter component control;
+- no in-guest daemon dependency; and
+- room to preserve bot-bottle-native lifecycle semantics.
+
+**Risks**
+
+- large and continuing compatibility burden;
+- “full feature coverage” is expensive across all providers;
+- accidental protocol fork; and
+- effort diverted from policy and UX differentiation.
+
+### Option C — Define an independent bot-bottle session API
+
+Build only the control surface bot-bottle needs.
+
+**Benefits**
+
+- clean fit with the trust model and persistent named bottles;
+- no upstream dependency; and
+- deliberate support for supervision and security events.
+
+**Risks**
+
+- recreates a fast-growing open-source project;
+- no existing client ecosystem;
+- slower browser/desktop/mobile work; and
+- increases the chance that Sandbox Agent becomes the de facto standard first.
+
+### Recommendation
+
+Start with **Option A as a bounded compatibility spike**, not a product
+commitment. Do not begin with a clean-room competing protocol.
+
+The spike should answer:
+
+1. Can Claude Code, Codex, and Pi retain exact native resume behavior?
+2. Can Sandbox Agent run without receiving real provider credentials?
+3. Can its server be reached through a bottle-scoped authenticated channel
+   without exposing the orchestrator or broadening guest egress?
+4. Which permission events overlap or conflict with bot-bottle supervision?
+5. Can normalized events be stored while clearly separating untrusted
+   transcript telemetry from authoritative gateway/Git audit?
+6. Can manifest skills, MCP servers, prompt, and startup arguments compile
+   deterministically into its configuration?
+7. Does its versioning policy permit a compatibility contract bot-bottle can
+   support?
+8. What image-size, startup-time, and update burden does the binary add?
+
+If the answers are favorable, adopt it behind a bot-bottle-owned interface and
+pin/test the supported version. If not, implement the smallest compatible
+subset needed by external clients before inventing a wholly separate API.
+
+## Product roadmap implications
+
+The competitor scan and this architecture comparison reorder the likely work:
+
+1. **Provider-neutral control/session compatibility spike**
+2. **Stable authenticated external bot-bottle API**
+3. **Normalized transcript/event persistence**
+4. **Parallel-session operator UI**
+5. **Browser/preview/computer-use capability**
+6. **Policy/image/skill distribution and signing**
+7. **Remote host/fleet management**
+
+This does not mean pausing security work. It means exposing the shipped
+security work through a product surface that can compete with the SDK-plus-
+sandbox ecosystem.
+
+## Decision
+
+Treat Sandbox Agent SDK as a potentially standard **agent process-control
+layer**, not as a sandbox replacement and not as a minor complementary
+library. Position bot-bottle one layer above it:
+
+- manifests express trusted role and environment policy;
+- bot-bottle compiles and enforces that policy across host, gateway, Git, and
+  isolation backends;
+- Sandbox Agent or a compatible protocol controls the selected agent process;
+  and
+- bot-bottle owns the turnkey operator experience.
+
+The Docker analogy is strategically sound when stated as:
+
+> Sandbox Agent can be the portable task/exec protocol; bot-bottle can be the
+> opinionated engine, Compose-like policy layer, and Desktop-like operator
+> product.
+
+It is not sound when stated as:
+
+> Sandbox Agent is OCI and bot-bottle is Docker.
+
+There is no independent OCI-equivalent agent specification yet, and Sandbox
+Agent already reaches into UI/session territory. Compatibility should be
+pursued quickly, while the trusted manifest/enforcement plane and operator
+experience remain the parts bot-bottle deliberately owns.

docs/research/testing-clean-install-on-macos.md

@@ -0,0 +1,305 @@
+# Testing a clean bot-bottle install on macOS
+
+How do you exercise `install.sh` (and, ideally, a first `bot-bottle start`)
+the way a brand-new user would — on a pristine macOS environment you can
+throw away afterward — *without* permanently polluting your daily-driver
+Mac? The user's framing: is there a VM or boundary that avoids creating a
+separate account, or is spinning up and tearing down a throwaway macOS
+user on the CLI easy enough to just do that?
+
+## Summary
+
+There is no lightweight, in-place macOS sandbox that hands you a clean home
+directory and wipeable system state without *either* a VM or a separate
+user account. `sandbox-exec` (Seatbelt) is deprecated and confines a
+process, not an environment; App Sandbox is for shipping apps, not for
+provisioning a fresh dev host. So the real choice is exactly the two the
+user named: **a disposable macOS VM** or **a throwaway user account** —
+and which one is right turns on a detail specific to *this* project.
+
+bot-bottle's default macOS backend is Apple's `container`, which runs each
+container in its own lightweight VM via `Virtualization.framework`
+([`README.md:27`](../README.md), [`apple-container-backend.md`](apple-container-backend.md)).
+That means a full end-to-end test — install *and* `bot-bottle start` —
+needs virtualization to work wherever bot-bottle runs. Inside a macOS guest
+VM that requires **nested virtualization, which Apple gates to M3 or newer
+chips on macOS 15+**. On M1/M2 you cannot run the Apple Container backend
+(or Docker Desktop, same reason) inside a macOS VM at all.
+
+The recommendation splits on what you're testing and what silicon you have:
+
+- **Install-script correctness only** (does `curl | sh` → pipx → config dir
+  → `doctor`'s Python/config checks pass?): a **disposable Tart VM** is the
+  cleanest boundary and works on any Apple Silicon Mac. `doctor` will report
+  the backend as not-ready inside the VM on M1/M2, which is fine — you're
+  testing the installer, not the runtime.
+- **Full runtime** (actually launch a bottle) on **M3/M4**: a **disposable
+  Tart VM from a golden base image, cloned per run** is the gold standard —
+  a genuine kernel/state boundary that wipes to nothing.
+- **Full runtime** on **M1/M2**, or when you'd rather not fight nested virt:
+  a **throwaway admin user via `sysadminctl`** is the pragmatic pick. It
+  tests the real backend because the backend runs on the host hypervisor —
+  but it is a *hygiene* boundary, not a security one, and it does **not**
+  clean the system-level footprint (see below).
+
+Prefer the VM. Reach for the throwaway user only when nested virt is off the
+table and you accept an imperfect wipe.
+
+## Why "a boundary without a separate user" doesn't really exist on macOS
+
+macOS has no namespace/overlay story like Linux `unshare` + tmpfs. The
+options that sound like in-place sandboxes don't fit:
+
+| Mechanism | Why it doesn't give you a clean, wipeable env |
+|---|---|
+| `sandbox-exec` / Seatbelt | Officially deprecated; confines *one process's* syscalls against a profile. It cannot present a fresh `$HOME` or a pristine `/usr/local`, and it won't let the Apple Container system service work. |
+| App Sandbox | Entitlement-based confinement for signed `.app` bundles, not a provisioning tool for a CLI dev environment. |
+| A second `$HOME` via `HOME=/tmp/foo` | Redirects only what honors `$HOME`. `install.sh` mostly does (it writes `~/.bot-bottle` and pipx/pip `--user` paths), but the Apple `container` install lands in `/usr/local` + a **system service**, and Homebrew lands in `/opt/homebrew` — all outside any `$HOME` you set. You'd get a false sense of "clean." |
+| APFS snapshot rollback (`tmutil localsnapshot`) | You can't roll the live boot volume back to a local snapshot without booting to Recovery; it's not a per-run userspace undo. |
+
+So the honest answer to "is there some boundary that avoids a separate
+user?": yes — a **VM** — and it's the *stronger* boundary anyway. The only
+lighter-weight option is the separate user, with the caveats below.
+
+## What a clean install actually touches (the footprint that decides "wipeable")
+
+Grounding the teardown story in what `install.sh` and the backend create:
+
+| Artifact | Location | In `$HOME`? | Survives user deletion? |
+|---|---|---|---|
+| Config / state / db | `~/.bot-bottle/{agents,bottles,contrib,state,db}` ([`install.sh:80-83`](../install.sh), [`bot_bottle/paths.py:59`](../bot_bottle/paths.py)) | ✅ | ❌ removed with home |
+| pipx venv + shim | `~/.local/pipx/venvs/bot-bottle`, shim in `~/.local/bin` ([`install.sh:87-89`](../install.sh)) | ✅ | ❌ removed with home |
+| private venv fallback (no pipx) | `~/.bot-bottle/venv` + symlink in `~/.local/bin` ([`install.sh`](../install.sh)) | ✅ | ❌ removed with home |
+| PATH / token exports | shell profile (`~/.zprofile`, etc.); `BOT_BOTTLE_CLAUDE_OAUTH_TOKEN` ([`README.md:74`](../README.md)) | ✅ | ❌ removed with home |
+| **Apple `container` install** | `/usr/local/...` + notarized `.pkg` receipts | ❌ | ✅ **stays** |
+| Apple `container` **service state** | per-user: `~/Library/Application Support/com.apple.container/` (`container system start`) | ✅ | ❌ removed with home |
+| **Homebrew** (if used for `container`/python) | `/opt/homebrew` | ❌ | ✅ **stays** |
+| Rosetta 2 (needed for image builds) | system | ❌ | ✅ **stays** |
+
+The bold rows are the crux: **deleting the throwaway user does not uninstall
+the Apple Container runtime, Homebrew, or Rosetta.** A VM, by contrast, wipes
+100% of the above by definition — that's its entire advantage for this task.
+
+The service row is the exception, and a live harness run corrected it: the
+Apple `container` service is **per-user**, not a host-wide launchd service.
+`container system status` reports an `appRoot` under
+`~/Library/Application Support/com.apple.container/`, and a freshly created
+account sees `container system service: NOT running` even while the creating
+admin's is running. That cuts both ways — the state is genuinely removed with
+the home, so the reset is *more* complete than this table first claimed, but
+it also means **no brand-new user can run a bottle until they run
+`container system start` once**. `doctor` correctly fails until they do, which
+is why `test` judges backend readiness separately from install correctness.
+
+## Option A — Disposable Tart VM (recommended)
+
+[Tart](https://tart.run) is a CLI-first macOS/Linux VM manager built on
+`Virtualization.framework`, purpose-built for exactly this "does it work on
+a clean macOS, without my settings/permissions/data" workflow. Keep one
+pristine *golden* image, clone a throwaway per run, delete it after.
+
+```sh
+brew install cirruslabs/cli/tart
+
+# One-time: build a golden base (either a prebuilt image or a vanilla IPSW).
+tart clone ghcr.io/cirruslabs/macos-tahoe-base:latest golden   # ~25 GB pull
+# — or a truly vanilla install you click through once —
+# tart create golden --from-ipsw latest --disk-size 60
+
+# Per test run: clone → boot → test → destroy.
+tart clone golden test-run
+tart run test-run &
+ssh admin@"$(tart ip test-run)"
+#   inside the guest:
+#   curl -fsSL https://gitea.dideric.is/didericis/bot-bottle/raw/branch/main/install.sh | sh
+#   bot-bottle doctor
+tart stop test-run
+tart delete test-run          # back to pristine; golden is untouched
+```
+
+Cloning is cheap (sparse files), so the golden image is your reset button —
+every `tart clone` is a fresh macOS. This is the closest thing to a Linux
+`docker run --rm` for a whole Mac.
+
+**The nested-virt caveat (read before relying on it for runtime tests).**
+The Apple Container backend inside the guest needs
+`Virtualization.framework` to work *inside* the VM. Apple enables nested
+virtualization only on **M3 or newer**, on **macOS 15 (Sequoia) or later**;
+M2 and earlier are excluded by Apple, confirmed by Apple DTS. Consequences:
+
+- **M3/M4 host:** full runtime works in the guest. `bot-bottle doctor`
+  reports the backend ready and `start` can launch a bottle. Gold standard.
+- **M1/M2 host:** the guest can install bot-bottle and pass the Python /
+  config-dir checks, but `doctor`'s backend check will fail and you cannot
+  launch a bottle in the VM. Still perfectly good for testing *the
+  installer*; not for the runtime.
+- **M4-specific:** a known bug blocks pre-Ventura guests on M4; use a
+  current macOS guest (which you want anyway, since Apple `container`
+  targets macOS 26 Tahoe).
+
+UTM is the GUI equivalent on the same framework (and was first to expose
+nested virt) if you'd rather click; Tart wins for a scriptable
+spin-up/tear-down loop.
+
+## Option B — Throwaway user via `sysadminctl` (pragmatic fallback)
+
+Creating and deleting a user from the CLI is genuinely a two-liner, and it
+tests the **real** backend on any Apple Silicon Mac because the backend runs
+on the host hypervisor — no nested virt needed.
+
+```sh
+# Create a self-contained admin user (admin needed for the container service).
+sudo sysadminctl -addUser bbtest -fullName "bot-bottle test" \
+    -password 'throwaway' -admin
+
+# Log into that account (fast-user-switch or the login window), then run the
+# installer as bbtest exactly as a new user would. When done:
+
+sudo sysadminctl -deleteUser bbtest -secure   # -secure erases the home dir
+```
+
+Honest accounting of what this does and doesn't buy you:
+
+- **Boundary strength:** it's a *hygiene / fresh-`$HOME`* boundary, **not a
+  security boundary.** Same kernel, same admin group; an admin test user can
+  touch system state. If the point is "clean environment," fine. If the point
+  is "contain something untrusted," this is the wrong tool — use a VM.
+- **Wipe completeness:** `-secure` erases the home dir (so `~/.bot-bottle`,
+  the pipx venv, and profile exports go away), but as the footprint table
+  shows, the **Apple Container runtime, its launchd system service,
+  Homebrew, and Rosetta persist.** For a truly repeatable "did a *system with
+  nothing installed* work?" test, that residue defeats the purpose — the
+  second run isn't clean.
+- **Operational gotchas:** don't pass real passwords on the command line (they
+  land in `ps` and history — this is a throwaway credential, so it's
+  tolerable here). Deletion must run as root from a normally-booted, admin-
+  logged-in session; the Terminal needs **Full Disk Access** or you'll hit
+  error `-14120` and a half-deleted account. Prefer letting the system place
+  the home dir (don't pass `-home`), or deletion can orphan it.
+
+Use this when you're on M1/M2, you specifically want to exercise the live
+backend, and you can tolerate the system-level runtime staying installed
+between runs (or you uninstall Apple `container` / brew by hand to reset).
+
+## Honorable mentions
+
+- **External bootable macOS volume.** A fresh macOS on an external SSD (or a
+  separate APFS volume) is bare-metal disposable: no nested-virt limit, real
+  backend works, and you `diskutil` the volume away to reset. Cost is reboot
+  friction per run — good for an occasional thorough pass, poor for a tight
+  loop.
+- **Rented / cloud Mac.** AWS EC2 Mac (dedicated Mac minis), Scaleway Apple
+  silicon, or MacStadium give a genuinely throwaway host you release when
+  done. Overkill for local iteration, but this is essentially what the
+  project's own advisory `integration-macos` CI job needs — a self-hosted
+  Apple Silicon runner with the `container` CLI, Python ≥ 3.11, and coverage
+  on the launchd service's PATH ([`README.md:78`](../README.md)). If you end
+  up standing up a cloud Mac for install testing, it doubles as that runner.
+
+## Recommendation
+
+Default to a **disposable Tart VM** — it's the only option that wipes the
+*entire* footprint (including the Apple Container system service that a user
+deletion leaves behind), it's a real boundary, and the spin-up/tear-down
+loop is a two-command `tart clone` / `tart delete`. Confirm your chip first:
+on **M3/M4** it tests install *and* runtime end-to-end; on **M1/M2** it still
+cleanly tests `install.sh` + `doctor`'s Python/config path, and you fall back
+to a **throwaway `sysadminctl` admin user** for live-backend testing —
+accepting that it's a hygiene boundary and that you'll manually uninstall the
+Apple Container runtime / Homebrew between runs to get back to truly clean.
+
+There is no third, lighter-weight "in-place boundary without a user" that
+actually delivers a clean, wipeable macOS — the VM *is* that answer, and it's
+the better one.
+
+## Harness
+
+The throwaway-user loop is scripted in
+[`scripts/macos-install-test.sh`](../../scripts/macos-install-test.sh):
+`up` creates the account, `run` pipes *this checkout's* `install.sh` into it
+headlessly (so a PR is verifiable before it lands) and lets the installer run
+`doctor`, `down` deletes the account and its home (the full reset), and
+`deep-reset` additionally uninstalls the host `container` runtime. It leans on
+the footprint analysis above — the reset is just user deletion because
+everything `install.sh` writes is user-home-local.
+
+There are two one-shot cycles, because "does the installer work" and "can a new
+user actually run a bottle" are different questions:
+
+```sh
+sudo ./scripts/macos-install-test.sh test        # up → run → status → down
+sudo ./scripts/macos-install-test.sh test-ready  # ... + prereqs before status
+```
+
+`test` models a macOS system **without** the prerequisites set up for this user
+— which is the default state of every new account, since the `container`
+service is per-user. It asserts the install is sound and *reports* backend
+readiness without failing on it, because install.sh provides no backend and so
+cannot regress one.
+
+`test-ready` models a system **with** them, then demands doctor go fully green,
+backend included. Both variants pass as of this writing.
+
+### The per-user prerequisite is two steps, not one
+
+Running it revealed that "set up the backend for this account" is more than
+starting a service:
+
+1. **`container system start`** — the service is per-user. The run confirms it
+   directly: the throwaway account's `appRoot` is
+   `/Users/bbtest/Library/Application Support/com.apple.container/`, its own,
+   and starting it left the admin's service untouched.
+2. **A guest kernel**, which also lives in that per-user app root. A fresh
+   account has none, so `container system start` prompts to download one —
+   and *only* prompts, since the flags default to asking. Headless callers
+   must pass `--enable-kernel-install` or the command dies on
+   `failed to read user input`.
+
+Neither step is done by `bot-bottle backend setup --backend=macos-container`,
+which only checks and then tells you to run `container system start` yourself.
+So a new account's real path to a working backend is:
+`container system start --enable-kernel-install`.
+
+The harness must also enter the user's launchd domain via
+`launchctl asuser <uid>` to do any of this. `container system start` registers
+`com.apple.container.apiserver` as a per-user launchd agent and talks to it
+over XPC; from plain `sudo -u` the caller is still in root's bootstrap
+namespace, the lookup crosses domains, and the apiserver answers
+`invalidState: "unauthorized request"` even though the agent started fine.
+
+It refuses to start against an existing account (a reused home is not a clean
+install), and it tears the account down from an `EXIT`/`INT` trap armed the
+moment the account exists, so a failed or Ctrl-C'd run still leaves the machine
+clean. Its verdict is deliberately stricter than the installer's own: note that
+`install.sh` exits **0** when it finishes but `doctor` reports unmet
+prerequisites, so "the installer succeeded" is not the assertion — `test` fails
+if the install fails, if `bot-bottle` never reached the new user's `PATH`, or if
+`doctor` is unhappy. `BB_TEST_KEEP=1` skips the teardown to poke at a failure.
+
+### What a fresh account actually inherits
+
+Expect the first honest run on a developer Mac to fail at the *Python* gate,
+and expect that to be correct. A new account's `PATH` is just `/etc/paths`
+(`/usr/local/bin:/System/Cryptexes/App/usr/bin:/usr/bin:/bin:/usr/sbin:/sbin`),
+which notably does **not** include `/opt/homebrew/bin`. Homebrew's `shellenv`
+line lives in the *installing* user's `~/.zprofile` and is not inherited, so a
+throwaway user resolves `python3` to `/usr/bin/python3` — the Command Line
+Tools stub, still **3.9.6** on macOS 26 — and `install.sh` correctly dies on its
+`3.11+` requirement. Your own shell resolving `python3` to a 3.14 Homebrew
+build says nothing about what a new user sees; that gap is exactly what this
+harness exists to expose.
+
+## Sources
+
+- [Apple Containers on macOS: technical comparison with Docker — The New Stack](https://thenewstack.io/apple-containers-on-macos-a-technical-comparison-with-docker/)
+- [How to Set Up Apple Containerization on macOS 26 — Stéphane Paquet](https://spaquet.medium.com/how-to-set-up-apple-containerization-on-macos-26-f870cc8c26cd)
+- [Install Apple Container CLI (macOS 15/26) — 4sysops](https://4sysops.com/archives/install-apple-container-cli-running-containers-natively-on-macos-15-sequoia-and-macos-26-tahoe/)
+- [Nested virtualization on Apple Silicon (M3+, macOS 15) — UTM issue #6700](https://github.com/utmapp/UTM/issues/6700)
+- [macOS 15 Sequoia nested virtualization for M3+ — Parallels Forums](https://forum.parallels.com/threads/macos-15-sequoia-nested-virtualization-for-m3-macs.364397/)
+- [M2 nested virtualization restriction (Apple DTS) — Apple Developer Forums](https://developer.apple.com/forums/thread/756723)
+- [M4 can't virtualize older macOS — Yahoo/Tech](https://tech.yahoo.com/computing/articles/m4-mac-computers-cant-virtualize-175122301.html)
+- [Tart — macOS/Linux VMs on Apple Silicon (Cirrus Labs)](https://tart.run/quick-start/)
+- [Tart GitHub](https://github.com/cirruslabs/tart)
+- [macOS VMs in a single command — frr.dev](https://www.frr.dev/posts/tart-macos-vms-from-terminal/)
+- [sysadminctl reference — SS64](https://ss64.com/mac/sysadminctl.html)
+- [User management from the macOS command line — macnotes](https://macnotes.wordpress.com/2019/03/28/user-management-create-remove-change-password-secure-token-from-macos-command-line/)

install.sh

@@ -8,14 +8,20 @@
 #   pipx install bot-bottle        # from a checkout or a published index
 #   uv tool install bot-bottle
 #
-# This script is a thin bootstrapper: it checks prerequisites, installs the
-# package with pipx (falling back to pip --user), creates the config dir, and
-# runs `bot-bottle doctor`. It is idempotent (safe to re-run) and never uses
-# sudo. It does NOT install Docker or a VM backend for you — `doctor` reports
-# what's missing after install.
+# This script is a thin bootstrapper: it finds a Python 3.11+ interpreter,
+# installs the package with pipx (falling back to a private venv), creates the
+# config dir, and runs `bot-bottle doctor`. It is idempotent (safe to re-run)
+# and never uses sudo. It does NOT install Docker or a VM backend for you —
+# `doctor` reports what's missing after install.
+#
+# Env:
+#   BOT_BOTTLE_PYTHON        interpreter to install with (skips the search)
+#   BOT_BOTTLE_INSTALL_SPEC  pip/git spec to install instead of the default
+#   BOT_BOTTLE_VENV          where the non-pipx install lives (~/.bot-bottle/venv)
 set -eu
 
 PACKAGE_SPEC="${BOT_BOTTLE_INSTALL_SPEC:-git+https://gitea.dideric.is/didericis/bot-bottle.git}"
+VENV="${BOT_BOTTLE_VENV:-${HOME}/.bot-bottle/venv}"
 MIN_PYTHON_MAJOR=3
 MIN_PYTHON_MINOR=11
 
@@ -28,17 +34,97 @@ die() {
     exit 1
 }
 
-# --- prerequisites -----------------------------------------------------------
+# --- prerequisites: find an interpreter new enough ----------------------------
 
-command -v python3 >/dev/null 2>&1 \
-    || die "python3 ${MIN_PYTHON_MAJOR}.${MIN_PYTHON_MINOR}+ is required but was not found"
-
-python3 - "$MIN_PYTHON_MAJOR" "$MIN_PYTHON_MINOR" <<'PY' || die "python3 ${MIN_PYTHON_MAJOR}.${MIN_PYTHON_MINOR} or newer is required"
+# Is $1 an interpreter that exists and meets the floor?
+python_ok() {
+    [ -n "${1:-}" ] || return 1
+    command -v "$1" >/dev/null 2>&1 || return 1
+    "$1" - "$MIN_PYTHON_MAJOR" "$MIN_PYTHON_MINOR" <<'PY' >/dev/null 2>&1
 import sys
 
 want = (int(sys.argv[1]), int(sys.argv[2]))
 raise SystemExit(0 if sys.version_info[:2] >= want else 1)
 PY
+}
+
+python_version() {
+    "$1" -c 'import sys; print("%d.%d.%d" % sys.version_info[:3])' 2>/dev/null
+}
+
+# `python3` on PATH is often NOT the newest interpreter installed, and on macOS
+# it is usually the oldest: a fresh login shell's PATH is just /etc/paths, so
+# python3 resolves to the Command Line Tools stub (3.9.x) while the usable
+# 3.11+ build sits in /opt/homebrew/bin or a python.org framework directory,
+# reachable only via a line in the *installing* user's shell profile. A new
+# account inherits none of that. Look past PATH before giving up, so the common
+# case installs instead of dead-ending on a version error.
+find_python() {
+    for candidate in \
+        "${BOT_BOTTLE_PYTHON:-}" \
+        python3 \
+        python3.14 python3.13 python3.12 python3.11 \
+        /opt/homebrew/bin/python3 \
+        /usr/local/bin/python3 \
+        "${HOME}/.local/bin/python3" \
+        /Library/Frameworks/Python.framework/Versions/*/bin/python3
+    do
+        # An unmatched glob arrives here literally; python_ok rejects it.
+        if python_ok "$candidate"; then
+            command -v "$candidate"
+            return 0
+        fi
+    done
+    return 1
+}
+
+# An explicit choice that doesn't work is an error, not a reason to quietly
+# search elsewhere and install somewhere the caller didn't ask for.
+if [ -n "${BOT_BOTTLE_PYTHON:-}" ] && ! python_ok "${BOT_BOTTLE_PYTHON}"; then
+    if command -v "${BOT_BOTTLE_PYTHON}" >/dev/null 2>&1; then
+        die "BOT_BOTTLE_PYTHON=${BOT_BOTTLE_PYTHON} is $(python_version "${BOT_BOTTLE_PYTHON}"), "\
+"below the ${MIN_PYTHON_MAJOR}.${MIN_PYTHON_MINOR} floor. Unset it to search for a newer one."
+    fi
+    die "BOT_BOTTLE_PYTHON=${BOT_BOTTLE_PYTHON} is not an executable interpreter."
+fi
+
+PYTHON="$(find_python || true)"
+
+if [ -z "${PYTHON}" ]; then
+    path_python="$(command -v python3 2>/dev/null || true)"
+    if [ -n "${path_python}" ]; then
+        found="the python3 on your PATH is ${path_python} ($(python_version "${path_python}")), which is too old"
+    else
+        found="no python3 was found on your PATH"
+    fi
+    case "$(uname -s)" in
+        Darwin) fix="  brew install python@3.12
+  # or install from https://www.python.org/downloads/macos/
+  # macOS itself ships only /usr/bin/python3, which is too old" ;;
+        *) fix="  sudo apt install python3.12    # Debian/Ubuntu
+  sudo dnf install python3.12    # Fedora/RHEL" ;;
+    esac
+    die "bot-bottle needs python3 ${MIN_PYTHON_MAJOR}.${MIN_PYTHON_MINOR} or newer, and none was found.
+  ${found}.
+  Also checked: python3.11-3.14, /opt/homebrew/bin, /usr/local/bin,
+  ~/.local/bin, and python.org framework builds.
+
+Install a newer Python, then re-run this installer:
+${fix}
+
+Already have one somewhere? Point at it directly:
+  BOT_BOTTLE_PYTHON=/path/to/python3 sh install.sh"
+fi
+
+# Be explicit when the interpreter isn't the obvious one, so nobody is left
+# wondering which Python their install ended up on.
+path_python="$(command -v python3 2>/dev/null || true)"
+if [ "${PYTHON}" != "${path_python}" ]; then
+    say "using ${PYTHON} ($(python_version "${PYTHON}"))"
+    if [ -n "${path_python}" ]; then
+        say "note: 'python3' on your PATH is ${path_python} ($(python_version "${path_python}")), which is below the ${MIN_PYTHON_MAJOR}.${MIN_PYTHON_MINOR} floor"
+    fi
+fi
 
 # Installing a `git+` spec (the default) shells out to git under the hood,
 # whether via pipx or pip. Fail early with a clear message rather than deep
@@ -51,30 +137,6 @@ case "${PACKAGE_SPEC}" in
         ;;
 esac
 
-# The pip fallback needs a usable pip. Externally-managed interpreters
-# (PEP 668, common on Debian/Ubuntu/Homebrew) reject `pip install --user`;
-# pipx sidesteps that, so recommend it when pip can't be used.
-if ! command -v pipx >/dev/null 2>&1; then
-    python3 -m pip --version >/dev/null 2>&1 || die \
-        "neither pipx nor a usable 'python3 -m pip' was found. Install pipx "\
-"(recommended): 'python3 -m pip install --user pipx' or your OS package manager."
-    if python3 - <<'PY'
-import os
-import sys
-import sysconfig
-
-# PEP 668: an EXTERNALLY-MANAGED marker in the stdlib dir means pip refuses
-# to install into this interpreter without --break-system-packages.
-marker = os.path.join(sysconfig.get_path("stdlib"), "EXTERNALLY-MANAGED")
-raise SystemExit(0 if os.path.exists(marker) else 1)
-PY
-    then
-        die "this Python is externally managed (PEP 668), so 'pip install --user' is "\
-"blocked. Install pipx and re-run: 'python3 -m pip install --user --break-system-packages pipx', "\
-"then 'pipx ensurepath'."
-    fi
-fi
-
 # --- config directories ------------------------------------------------------
 
 mkdir -p \
@@ -84,32 +146,50 @@ mkdir -p \
 
 # --- install -----------------------------------------------------------------
 
+BIN_DIR="${HOME}/.local/bin"
+
 if command -v pipx >/dev/null 2>&1; then
-    say "installing with pipx"
-    pipx install --force "${PACKAGE_SPEC}"
+    # --python pins the venv to the interpreter we vetted. Without it pipx uses
+    # whichever Python it was itself installed with, which is not necessarily
+    # the one that passed the version check above.
+    say "installing with pipx (python: ${PYTHON})"
+    pipx install --python "${PYTHON}" --force "${PACKAGE_SPEC}"
+    # Ask pipx where it puts entry points rather than assuming ~/.local/bin.
+    pipx_bin="$(pipx environment --value PIPX_BIN_DIR 2>/dev/null || true)"
+    [ -n "${pipx_bin}" ] && BIN_DIR="${pipx_bin}"
 else
-    say "pipx not found; installing with 'python3 -m pip install --user'"
-    python3 -m pip install --user --upgrade "${PACKAGE_SPEC}"
+    # No `pip install --user` fallback: PEP 668 makes it unusable on nearly
+    # every interpreter a Mac offers (Homebrew and python.org are both
+    # externally managed), and on Debian/Ubuntu too. A private venv sidesteps
+    # that entirely — PEP 668 does not apply inside a venv — and `venv` is
+    # stdlib, so unlike pipx there is nothing to bootstrap first.
+    say "pipx not found; installing into a managed venv at ${VENV}"
+    "${PYTHON}" -m venv --clear "${VENV}" || die \
+        "could not create a virtualenv at ${VENV} using ${PYTHON}. On Debian/Ubuntu "\
+"the venv module ships separately: 'sudo apt install python3-venv'."
+    "${VENV}/bin/python" -m pip install --upgrade "${PACKAGE_SPEC}"
+    # Expose the entry point outside the venv, the way pipx would.
+    mkdir -p "${BIN_DIR}"
+    ln -sf "${VENV}/bin/bot-bottle" "${BIN_DIR}/bot-bottle"
 fi
 
 # --- locate the entry point --------------------------------------------------
 
-# The pip --user scripts directory is platform-specific: ~/.local/bin on Linux,
-# but ~/Library/Python/<X.Y>/bin on a python.org macOS interpreter. Ask the
-# interpreter for its own user-scheme scripts dir instead of hardcoding.
-USER_SCRIPTS="$(python3 - <<'PY'
-import sysconfig
-print(sysconfig.get_path("scripts", sysconfig.get_preferred_scheme("user")))
-PY
-)"
-
 if command -v bot-bottle >/dev/null 2>&1; then
     BOT_BOTTLE_BIN="bot-bottle"
-elif [ -n "${USER_SCRIPTS}" ] && [ -x "${USER_SCRIPTS}/bot-bottle" ]; then
-    BOT_BOTTLE_BIN="${USER_SCRIPTS}/bot-bottle"
-    say "note: add ${USER_SCRIPTS} to your PATH to run 'bot-bottle' directly"
+elif [ -x "${BIN_DIR}/bot-bottle" ]; then
+    BOT_BOTTLE_BIN="${BIN_DIR}/bot-bottle"
+    # Name the file the user's own login shell actually reads. ~/.profile is
+    # the safe default for non-zsh: bash falls back to it, and suggesting
+    # ~/.bash_profile could shadow an existing ~/.profile.
+    case "${SHELL:-}" in
+        */zsh) profile="~/.zprofile" ;;
+        *) profile="~/.profile" ;;
+    esac
+    say "note: add ${BIN_DIR} to your PATH to run 'bot-bottle' directly:"
+    say "    echo 'export PATH=\"${BIN_DIR}:\$PATH\"' >> ${profile}"
 else
-    die "bot-bottle was installed but is not on PATH; add ${USER_SCRIPTS:-your user scripts dir} to PATH and re-run"
+    die "bot-bottle was installed but no entry point turned up in ${BIN_DIR}"
 fi
 
 # --- verify ------------------------------------------------------------------

pyproject.toml

@@ -36,6 +36,7 @@ include = ["bot_bottle*"]
 # files shipped under bot_bottle/; test_pyproject.py asserts they exist.
 [tool.setuptools.package-data]
 bot_bottle = [
+    "release-manifest.json",
     "gateway/egress/entrypoint.sh",
     "contrib/claude/Dockerfile",
     "contrib/claude/package.json",