docs(prd): add PRD for egress control plane

Out-of-band egress enforcement & cost-control plane: meter token usage
at the egress proxy, evaluate budgets with agent→bottle→parent→global
precedence, and force cutoff/freeze/kill without the agent in the loop.
Introduces a host-level SQLite ledger behind a thin repository API and a
host-only TUI dashboard. Closes the design discussion on #251.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NkwFXLFff9PYPy4wgVBJp9

.coveragerc

@@ -0,0 +1,18 @@
+[run]
+branch = True
+source = .
+
+[report]
+# Coverage policy: see docs/decisions/0004-coverage-policy.md.
+#
+# `omit` is reserved for genuinely interactive entry-point shells whose
+# bodies are `read_tty_line()` / curses prompt loops — there is no
+# behaviour to assert that a test wouldn't have to fake wholesale, so a
+# test here would inflate the number without buying confidence. This is
+# NOT a place to hide subprocess/backend orchestration: that code is
+# security-relevant and is measured via the integration suite instead
+# (run scripts/coverage.sh for the combined unit+integration number).
+omit =
+    bot_bottle/cli/tui.py
+    bot_bottle/cli/init.py
+    tests/*

.gitea/workflows/test.yml

@@ -39,8 +39,14 @@ jobs:
         with:
           python-version: "3.12"
 
+      - name: Install dev requirements
+        run: python3 -m pip install -r requirements-dev.txt
+
       - name: Run unit tests
-        run: python3 -m unittest discover -t . -s tests/unit -v
+        run: python3 -m coverage run -m unittest discover -t . -s tests/unit -v
+
+      - name: Report unit coverage
+        run: python3 -m coverage report -m
 
   integration:
     runs-on: ubuntu-latest
@@ -64,3 +70,32 @@ jobs:
 
       - name: Run integration tests
         run: python3 -m unittest discover -t . -s tests/integration -v
+
+  # Combined unit+integration coverage + the diff-coverage gate.
+  # See docs/decisions/0004-coverage-policy.md. The hard gate is diff
+  # coverage (new/changed lines >= 90%); the combined + critical reports
+  # are informational and degrade gracefully when the runner has no
+  # Docker (integration tests skip, those modules just read lower).
+  coverage:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install dev requirements
+        run: python3 -m pip install -r requirements-dev.txt
+
+      - name: Combined coverage (unit + integration)
+        run: PYTHON=python3 bash scripts/coverage.sh critical
+
+      - name: Diff-coverage gate (changed lines >= 90%)
+        run: |
+          git fetch --no-tags origin main:refs/remotes/origin/main
+          python3 scripts/diff_coverage.py --base origin/main --min 90

.gitea/workflows/update-badges.yml

@@ -8,6 +8,7 @@ on:
       - '**.py'
       - '.pylintrc'
       - 'pyrightconfig.json'
+      - '.coveragerc'
   workflow_dispatch:
 
 jobs:
@@ -45,10 +46,31 @@ jobs:
           echo "errors=$ERRORS" >> $GITHUB_OUTPUT
           echo "Pyright errors: $ERRORS"
 
+      - name: Run coverage and extract percentage
+        id: coverage
+        run: |
+          python -m coverage run -m unittest discover -t . -s tests/unit > /dev/null 2>&1 || true
+          PERCENT=$(python -m coverage report 2>/dev/null | grep '^TOTAL' | grep -oP '\d+(?=%)' | tail -1)
+          echo "percent=$PERCENT" >> $GITHUB_OUTPUT
+          echo "Coverage: $PERCENT%"
+
+      - name: Extract core (critical-module) coverage percentage
+        id: core_coverage
+        run: |
+          # Reuses the .coverage data from the previous step. The core list is
+          # the single source of truth in scripts/critical-modules.txt; every
+          # core module is unit-tested, so the unit-only run is accurate for it.
+          INCLUDE=$(grep -vE '^[[:space:]]*(#|$)' scripts/critical-modules.txt | paste -sd, -)
+          PERCENT=$(python -m coverage report --include="$INCLUDE" 2>/dev/null | grep '^TOTAL' | grep -oP '\d+(?=%)' | tail -1)
+          echo "percent=$PERCENT" >> $GITHUB_OUTPUT
+          echo "Core coverage: $PERCENT%"
+
       - name: Update badges in README
         run: |
           PYLINT_SCORE="${{ steps.pylint.outputs.score }}"
           PYRIGHT_ERRORS="${{ steps.pyright.outputs.errors }}"
+          COVERAGE_PERCENT="${{ steps.coverage.outputs.percent }}"
+          CORE_COVERAGE_PERCENT="${{ steps.core_coverage.outputs.percent }}"
 
           PYLINT_SCORE_ENCODED=$(echo "$PYLINT_SCORE" | sed 's|/|%2F|g')
 
@@ -58,9 +80,15 @@ jobs:
           if [ -n "$PYRIGHT_ERRORS" ]; then
             sed -i "s|/badge/pyright-[^)]*|/badge/pyright-${PYRIGHT_ERRORS}%20errors-brightgreen|" README.md
           fi
+          if [ -n "$COVERAGE_PERCENT" ]; then
+            sed -i "s|/badge/coverage-[^)]*|/badge/coverage-${COVERAGE_PERCENT}%25-brightgreen|" README.md
+          fi
+          if [ -n "$CORE_COVERAGE_PERCENT" ]; then
+            sed -i "s|/badge/core%20coverage-[^)]*|/badge/core%20coverage-${CORE_COVERAGE_PERCENT}%25-brightgreen|" README.md
+          fi
 
           echo "Updated badges:"
-          grep -E "pylint|pyright" README.md | head -2
+          grep -E "pylint|pyright|coverage" README.md | head -4
 
       - name: Commit and push badge updates
         run: |
@@ -73,7 +101,7 @@ jobs:
           else
             echo "Badge changes detected, committing..."
             git add README.md
-            MSG="chore: update quality badges"$'\n\n'"- Pylint: ${{ steps.pylint.outputs.score }}"$'\n'"- Pyright: ${{ steps.pyright.outputs.errors }} errors"$'\n\n'"[skip ci]"
+            MSG="chore: update quality badges"$'\n\n'"- Pylint: ${{ steps.pylint.outputs.score }}"$'\n'"- Pyright: ${{ steps.pyright.outputs.errors }} errors"$'\n'"- Coverage: ${{ steps.coverage.outputs.percent }}%"$'\n'"- Core coverage: ${{ steps.core_coverage.outputs.percent }}%"$'\n\n'"[skip ci]"
             git commit -m "$MSG"
             git push
           fi

.gitignore

@@ -22,3 +22,4 @@ venv/
 .pytest_cache/
 .mypy_cache/
 .ruff_cache/
+.coverage

Dockerfile.sidecars

@@ -62,6 +62,7 @@ COPY --from=gitleaks-src /usr/bin/gitleaks /usr/bin/gitleaks
 # top-level siblings (absolute imports), matching the prior
 # Dockerfile.egress / Dockerfile.supervise layout.
 COPY bot_bottle/egress_addon_core.py /app/egress_addon_core.py
+COPY bot_bottle/egress_dlp_config.py /app/egress_dlp_config.py
 COPY bot_bottle/egress_addon.py      /app/egress_addon.py
 COPY bot_bottle/dlp_detectors.py     /app/dlp_detectors.py
 COPY bot_bottle/yaml_subset.py       /app/yaml_subset.py

README.md

@@ -7,6 +7,8 @@
 [![test](https://gitea.dideric.is/didericis/bot-bottle/actions/workflows/test.yml/badge.svg?branch=main)](https://gitea.dideric.is/didericis/bot-bottle/actions?workflow=test.yml)
 [![pylint](https://img.shields.io/badge/pylint-9.93%2F10-brightgreen)](https://github.com/PyCQA/pylint)
 [![pyright](https://img.shields.io/badge/pyright-0%20errors-brightgreen)](https://github.com/microsoft/pyright)
+[![coverage](https://img.shields.io/badge/coverage-83%25-brightgreen)](https://coverage.readthedocs.io/)
+[![core coverage](https://img.shields.io/badge/core%20coverage-95%25-brightgreen)](https://gitea.dideric.is/didericis/bot-bottle/src/branch/main/docs/decisions/0004-coverage-policy.md)
 
 **Problem:** Developer wants to run a coding agent without supervision, but they don't want a prompt injected or misbehaving agent wrecking their environment or exfiltrating sensitive data.
 
@@ -14,7 +16,8 @@
 
 ## Features
 
-- **Per-bottle egress allowlist** — TLS-bumped HTTP/HTTPS chokepoint with a per-manifest host allowlist and request-body DLP scanner; DoH and arbitrary hosts blocked by default.
+- **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.
 - **Tokens the agent never sees** — host secrets live in a sidecar; the agent dials `http://sidecar: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.
@@ -106,8 +109,15 @@ egress:
   routes:
     - host: gitea.dideric.is
       auth:
-        scheme: token
+        scheme: token        # Bearer | token
         token_ref: BOT_BOTTLE_GITEA_TOKEN
+      matches:               # optional — restrict to specific paths/methods/headers
+        - paths:
+            - {type: prefix, value: /api/v1/}
+          methods: [GET, POST, PATCH, DELETE]
+      dlp:                   # optional — per-route detector overrides (default: all on)
+        outbound_detectors: [token_patterns, known_secrets]
+        inbound_detectors: false   # disable response scanning for this host
 ---
 
 The `gitea-dev` bottle. Provider auth via the inherited Claude route;
@@ -126,6 +136,26 @@ skills:
 You help maintain Gitea-hosted projects.
 ````
 
+**Egress route fields:**
+
+| Field | Required | Description |
+|---|---|---|
+| `host` | yes | Hostname to allowlist. One entry per host. |
+| `role` | no | Reserved for future use. The key is recognised but any value is currently rejected at load. Provider auth routes (e.g. Claude's `api.anthropic.com`) are injected automatically from `agent_provider.auth_token`, not via `role`. |
+| `auth.scheme` | when `auth` present | `Bearer` or `token`. Injected by the proxy; the agent never sees the value. |
+| `auth.token_ref` | when `auth` present | Env-var name holding the secret on the host. |
+| `matches` | no | Array of `{paths, methods, headers}` filters. A request must match at least one entry (if any are given) to be forwarded. |
+| `matches[].paths` | no | Array of `{type, value}`. `type` is `prefix` (default), `exact`, or `regex`. |
+| `matches[].methods` | no | Array of HTTP method strings, e.g. `[GET, POST]`. |
+| `matches[].headers` | no | Array of `{name, value, type}`. `type` is `exact` (default) or `regex`. |
+| `dlp` | no | Per-route DLP overrides. Omit to use defaults (all detectors on). |
+| `dlp.outbound_detectors` | no | `false` disables outbound scanning; list restricts to named detectors (`token_patterns`, `known_secrets`). |
+| `dlp.inbound_detectors` | no | `false` disables inbound scanning; list restricts to named detectors (`naive_injection_detection`). |
+| `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.
+
 More examples in `examples/`. Full design lives under `docs/prds/`; the trust-boundary rationale is in `docs/prds/0011-per-file-md-manifest.md`.
 
 ## Trademarks

bot_bottle/agent_provider.py

@@ -61,7 +61,6 @@ class AgentProviderRuntime:
     prompt_mode: PromptMode
     bypass_args: tuple[str, ...]
     resume_args: tuple[str, ...]
-    remote_control_args: tuple[str, ...]
 
 
 @dataclass(frozen=True)
@@ -240,7 +239,7 @@ class AgentProvider(ABC):
         BottleBackend.provision_workspace against the running bottle."""
         from .log import info
 
-        manifest_bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+        manifest_bottle = plan.manifest.bottle
         if manifest_bottle.git:
             from .git_gate import GIT_GATE_HOSTNAME, git_gate_render_gitconfig
             gate_host = getattr(plan, "git_gate_insteadof_host", GIT_GATE_HOSTNAME)
@@ -371,6 +370,15 @@ def build_agent_provision_plan(
     )
 
 
+def provider_startup_args(
+    provider_settings: dict[str, object] | None,
+) -> tuple[str, ...]:
+    raw = (provider_settings or {}).get("startup_args", ())
+    if not isinstance(raw, (list, tuple)):
+        return ()
+    return tuple(arg for arg in raw if isinstance(arg, str))
+
+
 def prompt_args(
     prompt_mode: PromptMode,
     prompt_path: str | None,
@@ -382,7 +390,7 @@ def prompt_args(
     if prompt_mode == "append_file":
         return ["--append-system-prompt-file", prompt_path]
     if prompt_mode == "read_prompt_file":
-        if argv and "resume" in argv:
+        if argv and ("resume" in argv or "remote-control" in argv):
             return []
         return [f"Read and follow the instructions in {prompt_path}."]
     if prompt_mode == "print_read_prompt_file":

bot_bottle/backend/__init__.py

@@ -45,7 +45,7 @@ from ..agent_provider import AgentProvisionPlan, get_provider, build_agent_provi
 from ..egress import EgressPlan
 from ..git_gate import GitGatePlan
 from ..log import die, info
-from ..manifest import ManifestGitEntry, Manifest
+from ..manifest import Manifest, ManifestIndex
 from ..supervise import SupervisePlan
 from ..util import expand_tilde
 from ..env import resolve_env, ResolvedEnv
@@ -61,7 +61,7 @@ class BottleSpec:
     Resolved values (image names, container name, scratch paths, runsc
     availability) live on the plan, not the spec."""
 
-    manifest: Manifest
+    manifest: ManifestIndex
     agent_name: str
     copy_cwd: bool
     user_cwd: str
@@ -72,6 +72,9 @@ class BottleSpec:
     identity: str = ""
     label: str = ""
     color: str = ""
+    # Ordered bottle names selected at launch (issue #269). When non-empty
+    # they are merged in order and replace the agent's `bottle:` field.
+    bottle_names: tuple[str, ...] = ()
 
 
 @dataclass(frozen=True)
@@ -80,6 +83,7 @@ class BottlePlan(ABC):
     (e.g. DockerBottlePlan) add backend-specific resolved fields."""
 
     spec: BottleSpec
+    manifest: Manifest
     stage_dir: Path
     git_gate_plan: GitGatePlan
 
@@ -108,13 +112,12 @@ class BottlePlan(ABC):
     def workspace_plan(self) -> WorkspacePlan:
         return workspace_plan(self.spec, guest_home=self.guest_home)
 
-    def print(self, *, remote_control: bool) -> None:
+    def print(self) -> None:
         """Render the y/N preflight summary to stderr."""
-        del remote_control
         spec = self.spec
-        manifest = spec.manifest
-        agent = manifest.agents[spec.agent_name]
-        bottle = manifest.bottle_for(spec.agent_name)
+        manifest = self.manifest
+        agent = manifest.agent
+        bottle = manifest.bottle
 
         env_names = visible_agent_env_names(
             sorted(
@@ -129,9 +132,13 @@ class BottlePlan(ABC):
         info(f"provider        : {self.agent_provision.template}")
         print_multi("env             ", env_names)
         print_multi("skills          ", list(agent.skills))
-        info(f"bottle          : {agent.bottle}")
+        effective_bottles = (
+            list(spec.bottle_names) if spec.bottle_names
+            else ([agent.bottle] if agent.bottle else [])
+        )
+        print_multi("bottle          ", effective_bottles)
 
-        identity = manifest.git_identity_summary(spec.agent_name)
+        identity = manifest.git_identity_summary()
         if identity:
             info(f"  git identity  : {identity}")
 
@@ -289,15 +296,14 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
             write_launch_metadata,
         )
 
-        self._validate(spec)
+        manifest = self._validate(spec)
 
         self._preflight()
 
-        manifest = spec.manifest
-        manifest_bottle = manifest.bottle_for(spec.agent_name)
+        manifest_bottle = manifest.bottle
         manifest_agent_provider = manifest_bottle.agent_provider
         agent_provider = get_provider(manifest_agent_provider.template)
-        resolved_env = resolve_env(manifest, spec.agent_name)
+        resolved_env = resolve_env(manifest)
         workspace = workspace_plan(spec, guest_home=agent_provider.guest_home)
 
         slug = mint_slug(spec)
@@ -313,7 +319,7 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
         else:
             agent_dockerfile_path = str(agent_provider.dockerfile)
 
-        agent_dir, prompt_file = prepare_agent_state_dir(slug, spec)
+        agent_dir, prompt_file = prepare_agent_state_dir(slug, manifest)
 
         agent_provision_plan = build_agent_provision_plan(
             template=manifest_agent_provider.template,
@@ -337,6 +343,7 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
 
         return self._resolve_plan(
             spec,
+            manifest=manifest,
             slug=slug,
             resolved_env=resolved_env,
             agent_provision_plan=agent_provision_plan,
@@ -355,18 +362,18 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
         """
         pass
 
-    def _validate(self, spec: BottleSpec) -> None:
-        """Cross-backend pre-launch checks. Confirms the agent exists,
-        the named skills are present on the host, and every git
-        IdentityFile resolves. Subclasses with additional preconditions
-        should override and call `super()._validate(spec)` first."""
-        manifest = spec.manifest
-        manifest.require_agent(spec.agent_name)
-        agent = manifest.agents[spec.agent_name]
-        bottle = manifest.bottle_for(spec.agent_name)
-        self._validate_skills(agent.skills)
-        self._validate_git_entries(bottle.git)
-        self._validate_agent_provider_dockerfile(spec)
+    def _validate(self, spec: BottleSpec) -> Manifest:
+        """Cross-backend pre-launch checks. Parses the selected agent and
+        its bottle (raising ManifestError on invalid content), confirms
+        skills are present on the host, and every git IdentityFile resolves.
+
+        Returns the loaded Manifest for the selected agent. Subclasses with
+        additional preconditions should override and call
+        `super()._validate(spec)` first."""
+        manifest = spec.manifest.load_for_agent(spec.agent_name, spec.bottle_names)
+        self._validate_skills(manifest.agent.skills)
+        self._validate_agent_provider_dockerfile(spec, manifest)
+        return manifest
 
     def _validate_skills(self, skills: Sequence[str]) -> None:
         """Each named skill must be a directory under the host's
@@ -380,18 +387,8 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
                     f"Create it under ~/.claude/skills/, then re-run."
                 )
 
-    def _validate_git_entries(self, entries: Sequence[ManifestGitEntry]) -> None:
-        """Each entry's IdentityFile must exist on the host (after
-        expanding leading ~) — the git-gate copies it in at start time
-        to authenticate the upstream push (PRD 0008). Shape is already
-        enforced by Manifest validation; this only checks presence."""
-        for entry in entries:
-            key = expand_tilde(entry.IdentityFile)
-            if not os.path.isfile(key):
-                die(f"git upstream key file not found for '{entry.Name}': {key}")
-
-    def _validate_agent_provider_dockerfile(self, spec: BottleSpec) -> None:
-        bottle = spec.manifest.bottle_for(spec.agent_name)
+    def _validate_agent_provider_dockerfile(self, spec: BottleSpec, manifest: Manifest) -> None:
+        bottle = manifest.bottle
         dockerfile = bottle.agent_provider.dockerfile
         if not dockerfile:
             return
@@ -399,15 +396,19 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
         if not path.is_absolute():
             path = Path(spec.user_cwd) / path
         if not path.is_file():
+            effective = (
+                ", ".join(spec.bottle_names) if spec.bottle_names else manifest.agent.bottle
+            )
             die(
                 f"agent_provider.dockerfile for bottle "
-                f"'{spec.manifest.agents[spec.agent_name].bottle}' not found: {path}"
+                f"'{effective}' not found: {path}"
             )
 
     @abstractmethod
     def _resolve_plan(self,
                       spec: BottleSpec,
                       *,
+                      manifest: Manifest,
                       slug: str,
                       resolved_env: ResolvedEnv,
                       agent_provision_plan: AgentProvisionPlan,
@@ -534,6 +535,11 @@ from .docker import DockerBottleBackend  # noqa: E402  # pylint: disable=wrong-i
 from .macos_container import MacosContainerBottleBackend  # noqa: E402  # pylint: disable=wrong-import-position
 from .smolmachines import SmolmachinesBottleBackend  # noqa: E402  # pylint: disable=wrong-import-position
 
+# Freezer is imported after the backend classes for the same reason:
+# Freezer.commit_slug constructs ActiveAgent, which must be fully
+# defined first.
+from .freeze import CommitCancelled, Freezer, get_freezer  # noqa: E402  # pylint: disable=wrong-import-position
+
 
 # The dict is heterogeneous: each value is a BottleBackend specialized
 # over its own plan type. Concrete plan types are erased here because
@@ -621,9 +627,12 @@ __all__ = [
     "BottleCleanupPlan",
     "BottlePlan",
     "BottleSpec",
+    "CommitCancelled",
     "ExecResult",
+    "Freezer",
     "enumerate_active_agents",
     "get_bottle_backend",
+    "get_freezer",
     "has_backend",
     "known_backend_names",
 ]

bot_bottle/backend/docker/backend.py

@@ -30,6 +30,7 @@ from ...egress import EgressPlan
 from ...env import ResolvedEnv
 from ...git_gate import GitGatePlan
 from ...supervise import SupervisePlan
+from ...manifest import Manifest
 from .. import ActiveAgent, BottleBackend, BottleSpec
 from . import cleanup as _cleanup
 from . import enumerate as _enumerate
@@ -63,6 +64,7 @@ class DockerBottleBackend(BottleBackend["DockerBottlePlan", "DockerBottleCleanup
         self,
         spec: BottleSpec,
         *,
+        manifest: Manifest,
         slug: str,
         resolved_env: ResolvedEnv,
         agent_provision_plan: AgentProvisionPlan,
@@ -73,6 +75,7 @@ class DockerBottleBackend(BottleBackend["DockerBottlePlan", "DockerBottleCleanup
     ) -> DockerBottlePlan:
         return _resolve_plan.resolve_plan(
             spec,
+            manifest=manifest,
             slug=slug,
             resolved_env=resolved_env,
             agent_provision_plan=agent_provision_plan,

bot_bottle/backend/docker/capability_apply.py

@@ -1,211 +0,0 @@
-"""capability_apply — host-side orchestrator for capability-block
-remediation (PRD 0016).
-
-On approval of a capability-block proposal, the dashboard calls
-apply_capability_change(slug, new_dockerfile) which:
-
-  1. Snapshots the agent's transcript dir to
-     ~/.bot-bottle/state/<slug>/transcript/ (best-effort).
-  2. Pushes the agent's working tree via `git push` (best-effort —
-     no upstream / no commits / no git repo all skip with a log).
-  3. Writes the new Dockerfile to
-     ~/.bot-bottle/state/<slug>/Dockerfile (PRD 0016 Phase 1
-     state). The next `cli.py start <agent>` picks it up.
-  4. Force-removes the agent container + all sidecars + the
-     per-bottle networks. Idempotent — missing resources are not
-     errors.
-
-Returns (before, after) Dockerfile contents so the dashboard can
-record / render the diff. (capability-block has no audit log per
-PRD 0013 — the per-bottle Dockerfile state is its own record.)
-
-This is "fire-and-forget" from the agent's perspective: by the time
-the dashboard writes the response file the supervise sidecar is
-gone, so the agent's tool call connection drops without ever
-receiving the response. The replacement agent (next manual
-`cli.py start`) sees the new Dockerfile and starts from there.
-v1 does not auto-relaunch — see PRD 0016's capability-block return
-semantics open question.
-"""
-
-from __future__ import annotations
-
-import shutil
-import subprocess
-
-from ...agent_provider import get_provider
-from ...log import info, warn
-from ...bottle_state import (
-    mark_preserved,
-    per_bottle_dockerfile,
-    transcript_snapshot_dir,
-    write_per_bottle_dockerfile,
-)
-from .sidecar_bundle import sidecar_bundle_container_name
-
-
-# Agent home inside the container (per the repo Dockerfile's
-# `USER node` + `WORKDIR /home/node`). Used to locate the transcript
-# dir + the workspace dir for git push.
-_AGENT_HOME_IN_CONTAINER = "/home/node"
-_AGENT_TRANSCRIPT_IN_CONTAINER = f"{_AGENT_HOME_IN_CONTAINER}/.claude"
-_AGENT_WORKSPACE_IN_CONTAINER = f"{_AGENT_HOME_IN_CONTAINER}/workspace"
-
-# Per-bottle resource name patterns (mirroring prepare.py).
-def _agent_container_name(slug: str) -> str:
-    return f"bot-bottle-{slug}"
-
-
-def _per_bottle_container_names(slug: str) -> list[str]:
-    """All container names that belong to this bottle. Missing
-    containers are silently skipped by the teardown helper, so it's
-    fine to include names that don't exist for a given bottle."""
-    return [
-        _agent_container_name(slug),
-        sidecar_bundle_container_name(slug),
-    ]
-
-
-def _per_bottle_network_names(slug: str) -> list[str]:
-    return [
-        f"bot-bottle-net-{slug}",
-        f"bot-bottle-egress-{slug}",
-    ]
-
-
-class CapabilityApplyError(RuntimeError):
-    """Raised when the apply fails in a way that should keep the
-    proposal pending (so the operator can retry). Best-effort
-    failures (transcript snapshot, git push) do not raise — they
-    just log and proceed."""
-
-
-# --- Public helpers --------------------------------------------------------
-
-
-def fetch_current_dockerfile(slug: str) -> str:
-    """Return the Dockerfile content the next `cli.py start <agent>`
-    would use for this bottle. If a per-bottle override exists, that
-    one; otherwise the repo's Dockerfile.
-
-    Used by the operator-edit verb to show the current source of
-    truth, and by apply_capability_change for the before-diff."""
-    override = per_bottle_dockerfile(slug)
-    if override is not None:
-        return override
-    repo_dockerfile = get_provider("claude").dockerfile
-    if repo_dockerfile.is_file():
-        return repo_dockerfile.read_text()
-    raise CapabilityApplyError(
-        f"no per-bottle Dockerfile for {slug} and no provider Dockerfile at "
-        f"{repo_dockerfile}"
-    )
-
-
-def apply_capability_change(slug: str, new_dockerfile: str) -> tuple[str, str]:
-    """End-to-end capability-block remediation. See module docstring
-    for the sequence. Returns (before, after) Dockerfile content."""
-    if not new_dockerfile.strip():
-        raise CapabilityApplyError("proposed Dockerfile is empty")
-    before = fetch_current_dockerfile(slug)
-
-    snapshot_transcript(slug)
-    _push_working_tree(slug)
-    write_per_bottle_dockerfile(slug, new_dockerfile)
-    # Set the preserve marker BEFORE teardown so cli.py's session-end
-    # cleanup sees it and keeps the state dir intact for the
-    # operator's `cli.py resume <identity>`. Without the marker the
-    # state dir would be deleted as part of normal session end.
-    mark_preserved(slug)
-    _teardown_bottle(slug)
-
-    return before, new_dockerfile
-
-
-# --- Internals -------------------------------------------------------------
-
-
-
-def snapshot_transcript(slug: str) -> None:
-    """`docker cp` /home/node/.claude out of the agent container into
-    ~/.bot-bottle/state/<slug>/transcript/. Best-effort: missing
-    container, missing dir, or cp error all log a warning and return.
-    The transcript is what `claude --resume` reads to pick up where
-    the agent left off.
-
-    Called from two places:
-      - capability-apply, before tearing the bottle down.
-      - cli.py's session-end path, before the launch context closes,
-        so a crash or normal exit also leaves a transcript on disk
-        (deleted along with the state dir on clean exit, kept on
-        crash or capability-block per the preserve marker)."""
-    container = _agent_container_name(slug)
-    dest = transcript_snapshot_dir(slug)
-    if dest.exists():
-        # Remove any prior snapshot so the new one is a clean copy.
-        shutil.rmtree(dest, ignore_errors=True)
-    dest.parent.mkdir(parents=True, exist_ok=True)
-    r = subprocess.run(
-        ["docker", "cp", f"{container}:{_AGENT_TRANSCRIPT_IN_CONTAINER}", str(dest)],
-        capture_output=True, text=True, check=False,
-    )
-    if r.returncode != 0:
-        warn(
-            f"transcript snapshot skipped "
-            f"({(r.stderr or '').strip() or 'no transcript dir in container?'})"
-        )
-        return
-    info(f"transcript snapshotted to {dest}")
-
-
-def _push_working_tree(slug: str) -> None:
-    """`docker exec <agent> git push` from /home/node/workspace.
-    Best-effort: not-a-git-repo, no upstream, nothing-to-push, no
-    network all log a warning and return. The replacement bottle
-    will pick up whatever's actually upstream."""
-    container = _agent_container_name(slug)
-    r = subprocess.run(
-        [
-            "docker", "exec", container, "sh", "-c",
-            f"cd {_AGENT_WORKSPACE_IN_CONTAINER} && "
-            f"git rev-parse --is-inside-work-tree >/dev/null 2>&1 && "
-            f"git push origin HEAD 2>&1 || true",
-        ],
-        capture_output=True, text=True, check=False,
-    )
-    if r.returncode != 0:
-        warn(
-            f"capability-apply: git push skipped "
-            f"({(r.stderr or '').strip() or 'docker exec failed'})"
-        )
-        return
-    output = (r.stdout or "").strip()
-    if output:
-        info(f"capability-apply: git push: {output}")
-    else:
-        info("capability-apply: git push ran (no output — likely not a git workspace)")
-
-
-def _teardown_bottle(slug: str) -> None:
-    """Force-remove all per-bottle docker resources. Idempotent —
-    `docker rm -f` / `docker network rm` silently ignore missing
-    names, so this can be called even mid-rebuild."""
-    info(f"capability-apply: tearing down bottle {slug}")
-    for name in _per_bottle_container_names(slug):
-        subprocess.run(
-            ["docker", "rm", "-f", name],
-            stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, check=False,
-        )
-    for net in _per_bottle_network_names(slug):
-        subprocess.run(
-            ["docker", "network", "rm", net],
-            stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, check=False,
-        )
-
-
-__all__ = [
-    "CapabilityApplyError",
-    "apply_capability_change",
-    "fetch_current_dockerfile",
-    "snapshot_transcript",
-]

bot_bottle/backend/docker/compose.py

@@ -28,11 +28,12 @@ from typing import Any
 from ...egress import (
     EGRESS_HOSTNAME,
     EGRESS_ROUTES_IN_CONTAINER,
+    egress_agent_env_entries,
+    egress_sidecar_env_entries,
 )
 from ...git_gate import GIT_GATE_HOSTNAME
 from ...log import die, warn
 from ...supervise import (
-    CURRENT_CONFIG_DIR_IN_AGENT,
     QUEUE_DIR_IN_CONTAINER,
     SUPERVISE_HOSTNAME,
     SUPERVISE_PORT,
@@ -134,9 +135,8 @@ def _sidecar_bundle_service(plan: DockerBottlePlan) -> dict[str, Any]:
     ep = plan.egress_plan
     volumes.append(_bind(ep.mitmproxy_ca_host_path, EGRESS_CA_IN_CONTAINER))
     if ep.routes:
-        volumes.append(_bind(ep.routes_path, EGRESS_ROUTES_IN_CONTAINER))
-        for token_env in sorted(ep.token_env_map.keys()):
-            env.append(token_env)
+        volumes.append(_bind(ep.routes_path.parent, str(Path(EGRESS_ROUTES_IN_CONTAINER).parent)))
+    env.extend(egress_sidecar_env_entries(ep))
 
     # --- git-gate -----------------------------------------------------
     gp = plan.git_gate_plan
@@ -220,6 +220,7 @@ def _agent_service(plan: DockerBottlePlan) -> dict[str, Any]:
     # never lands on argv or in the compose file.
     for name in sorted(plan.forwarded_env.keys()):
         env.append(name)
+    env.extend(egress_agent_env_entries(plan.egress_plan))
 
     service: dict[str, Any] = {
         "image": plan.image,
@@ -231,15 +232,6 @@ def _agent_service(plan: DockerBottlePlan) -> dict[str, Any]:
     if plan.use_runsc:
         service["runtime"] = "runsc"
 
-    volumes: list[dict[str, Any]] = []
-    if plan.supervise_plan is not None:
-        volumes.append(_bind(
-            plan.supervise_plan.current_config_dir,
-            CURRENT_CONFIG_DIR_IN_AGENT,
-        ))
-    if volumes:
-        service["volumes"] = volumes
-
     # The init supervisor inside the bundle owns intra-bundle
     # daemon ordering, so the agent only waits for the bundle
     # container itself.

bot_bottle/backend/docker/egress_apply.py

@@ -1,24 +1,21 @@
-"""Host-side helper for egress sidecar inspection (issue #198).
+"""Host-side helper for egress sidecar inspection and live updates.
 
-`_merge_single_route`, `add_route`, and `apply_routes_change` were
-removed when the egress-block MCP tool was dropped. The remaining
-helpers support runtime inspection and validation of the routes file
-without modifying it at runtime.
+The approve path uses this module to validate a proposed routes file,
+write it to the bottle's live egress state dir, and signal the sidecar
+bundle so the mitmproxy addon reloads it.
 """
 
 from __future__ import annotations
 
+import os
 import subprocess
 
 from ...egress import EGRESS_ROUTES_IN_CONTAINER
-from ...egress_addon_core import load_routes
+from ...log import warn
+from ..egress_apply import EgressApplicator, EgressApplyError
 from .sidecar_bundle import sidecar_bundle_container_name
 
 
-class EgressApplyError(RuntimeError):
-    pass
-
-
 def fetch_current_routes(slug: str) -> str:
     container = sidecar_bundle_container_name(slug)
     r = subprocess.run(
@@ -33,17 +30,31 @@ def fetch_current_routes(slug: str) -> str:
     return r.stdout
 
 
-def validate_routes_content(content: str) -> None:
-    try:
-        load_routes(content)
-    except ValueError as e:
-        raise EgressApplyError(
-            f"proposed routes.yaml is not valid: {e}"
-        ) from e
+class DockerEgressApplicator(EgressApplicator):
+    def _signal_bundle_reload(self, slug: str) -> None:
+        container = sidecar_bundle_container_name(slug)
+        result = subprocess.run(
+            ["docker", "kill", "--signal", "HUP", container],
+            capture_output=True, text=True, check=False, env=os.environ,
+        )
+        if result.returncode != 0:
+            last_error = (result.stderr or "").strip() or (result.stdout or "").strip()
+            warn(
+                f"egress: routes updated on disk for {slug}, but bundle reload failed: "
+                f"{last_error or 'docker kill failed'}"
+            )
+            raise EgressApplyError(
+                f"could not reload egress bundle {container}: "
+                f"{last_error or 'docker kill failed'}"
+            )
+
+
+applicator = DockerEgressApplicator()
 
 
 __all__ = [
+    "DockerEgressApplicator",
     "EgressApplyError",
+    "applicator",
     "fetch_current_routes",
-    "validate_routes_content",
 ]

bot_bottle/backend/docker/freezer.py

@@ -0,0 +1,23 @@
+"""DockerFreezer — snapshot a Docker bottle via `docker commit`."""
+
+from __future__ import annotations
+
+from .. import ActiveAgent
+from ..freeze import Freezer
+from .util import commit_container
+from ...log import info
+
+
+class DockerFreezer(Freezer):
+    """Freezes a Docker bottle by running `docker commit`."""
+
+    backend_name = "docker"
+
+    def _freeze(self, agent: ActiveAgent) -> str:
+        container = f"bot-bottle-{agent.slug}"
+        image_tag = f"bot-bottle-committed-{agent.slug}:latest"
+        commit_container(container, image_tag)
+        return image_tag
+
+    def _export_hint(self, slug: str, image_ref: str) -> None:
+        info(f"to export for migration:      docker save {image_ref} -o {slug}.tar")

bot_bottle/backend/docker/launch.py

@@ -47,6 +47,7 @@ from ...bottle_state import (
     bottle_state_dir,
     egress_state_dir,
     git_gate_state_dir,
+    read_committed_image,
 )
 from .compose import (
     bottle_plan_to_compose,
@@ -75,7 +76,7 @@ def launch(
     Teardown on exit."""
     stack = ExitStack()
 
-    _bottle_for_revoke = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+    _bottle_for_revoke = plan.manifest.bottle
     _git_gate_dir_for_revoke = git_gate_state_dir(plan.slug)
 
     def teardown() -> None:
@@ -91,12 +92,22 @@ def launch(
         )
 
     try:
-        # Step 1: agent image build. Sidecar images get built lazily by
-        # `docker compose up` via the renderer's `build:` directives.
-        docker_mod.build_image(
-            plan.image, _REPO_DIR,
-            dockerfile=plan.dockerfile_path,
-        )
+        # Step 1: agent image. Use a committed snapshot when one exists
+        # and is present in the local daemon; otherwise build from the
+        # Dockerfile. Sidecar images get built lazily by `docker compose
+        # up` via the renderer's `build:` directives.
+        committed = read_committed_image(plan.slug)
+        if committed and docker_mod.image_exists(committed):
+            info(f"using committed image {committed!r}")
+            plan = dataclasses.replace(
+                plan,
+                agent_provision=dataclasses.replace(plan.agent_provision, image=committed),
+            )
+        else:
+            docker_mod.build_image(
+                plan.image, _REPO_DIR,
+                dockerfile=plan.dockerfile_path,
+            )
 
         internal_network = network_mod.network_name_for_slug(plan.slug)
         egress_network = network_mod.network_egress_name_for_slug(plan.slug)
@@ -176,7 +187,7 @@ def launch(
             agent_command=plan.agent_command,
             agent_prompt_mode=plan.agent_prompt_mode,
             agent_provider_template=plan.agent_provider_template,
-            terminal_title=plan.spec.label or plan.spec.agent_name,
+            terminal_title=f"{plan.spec.label} ({plan.spec.agent_name})" if plan.spec.label else plan.spec.agent_name,
             terminal_color=plan.spec.color,
             agent_workdir=plan.workspace_plan.workdir,
         )

bot_bottle/backend/docker/resolve_plan.py

@@ -18,6 +18,7 @@ from .. import BottleSpec
 from ...env import ResolvedEnv
 from ...agent_provider import AgentProvisionPlan
 from ...egress import EgressPlan
+from ...manifest import Manifest
 from ...supervise import SupervisePlan
 from ...git_gate import GitGatePlan
 
@@ -31,6 +32,7 @@ def build_guest_env(resolved_env: ResolvedEnv) -> dict[str, str]:
 
 def resolve_plan(
     spec: BottleSpec,
+    manifest: Manifest,
     slug: str,
     resolved_env: ResolvedEnv,
     agent_provision_plan: AgentProvisionPlan,
@@ -48,6 +50,7 @@ def resolve_plan(
 
     return DockerBottlePlan(
         spec=spec,
+        manifest=manifest,
         stage_dir=stage_dir,
         slug=slug,
         forwarded_env=dict(resolved_env.forwarded),

bot_bottle/backend/docker/util.py

@@ -152,6 +152,21 @@ def build_image(ref: str, context: str, *, dockerfile: str = "") -> None:
 #         )
 
 
+def commit_container(container_name: str, image_tag: str) -> None:
+    """Run `docker commit <container_name> <image_tag>` to snapshot the
+    running container's filesystem state as a local Docker image."""
+    result = subprocess.run(
+        ["docker", "commit", container_name, image_tag],
+        capture_output=True, text=True, check=False,
+    )
+    if result.returncode != 0:
+        die(
+            f"docker commit {container_name!r} → {image_tag!r} failed: "
+            f"{(result.stderr or '').strip() or '<no stderr>'}"
+        )
+    info(f"committed {container_name!r} → {image_tag!r}")
+
+
 def image_id(ref: str) -> str:
     """Return the content-addressed image ID (e.g.
     `sha256:abcd...`) for `ref`. The smolmachines backend keys its

bot_bottle/backend/egress_apply.py

@@ -0,0 +1,54 @@
+"""Shared base class for host-side egress apply across backends.
+
+Each backend subclasses EgressApplicator and overrides _signal_bundle_reload
+with the backend-specific kill command.
+"""
+
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from pathlib import Path
+
+from ..bottle_state import egress_state_dir
+from ..egress import EGRESS_ROUTES_FILENAME
+from ..egress_addon_core import LOG_OFF, load_config
+
+
+class EgressApplyError(RuntimeError):
+    pass
+
+
+class EgressApplicator(ABC):
+    def apply_routes_change(self, slug: str, content: str) -> tuple[str, str]:
+        """Persist `content` to the live routes file and reload egress."""
+        self.validate_routes_content(content)
+        routes_path = self._routes_path(slug)
+        routes_path.parent.mkdir(parents=True, exist_ok=True)
+        before = routes_path.read_text(encoding="utf-8") if routes_path.exists() else ""
+        routes_path.write_text(content, encoding="utf-8")
+        routes_path.chmod(0o600)
+        self._signal_bundle_reload(slug)
+        return before, content
+
+    @staticmethod
+    def validate_routes_content(content: str) -> None:
+        try:
+            config = load_config(content)
+        except ValueError as e:
+            raise EgressApplyError(
+                f"proposed routes.yaml is not valid: {e}"
+            ) from e
+        if config.log != LOG_OFF:
+            raise EgressApplyError(
+                "proposed routes.yaml must not change egress logging"
+            )
+
+    @staticmethod
+    def _routes_path(slug: str) -> Path:
+        return egress_state_dir(slug) / EGRESS_ROUTES_FILENAME
+
+    @abstractmethod
+    def _signal_bundle_reload(self, slug: str) -> None: ...
+
+
+__all__ = ["EgressApplicator", "EgressApplyError"]

bot_bottle/backend/freeze.py

@@ -0,0 +1,100 @@
+"""Freezer — snapshot a running bottle to a resumable artifact.
+
+Follows the same pattern as BottleBackend: a shared base class with
+common post-freeze steps (write committed-image path, mark preserved,
+print resume hint) and backend-specific subclasses in their respective
+backend directories.
+
+Entry points:
+  Freezer.commit(agent)      — freeze by ActiveAgent
+  Freezer.commit_slug(slug)  — convenience wrapper for cmd_commit
+  get_freezer(backend_name)  — factory
+"""
+
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+
+from . import ActiveAgent
+from ..bottle_state import mark_preserved, write_committed_image
+from ..log import die, info
+
+
+class CommitCancelled(Exception):
+    """Raised by Freezer._freeze when the user declines a confirmation prompt."""
+
+
+class Freezer(ABC):
+    """Freezes a running bottle to a resumable artifact.
+
+    The base class owns the shared post-commit steps:
+      - write_committed_image  — records the artifact path in per-bottle state
+      - mark_preserved         — prevents teardown from removing the state dir
+      - resume hint            — printed to stderr after the snapshot
+
+    Subclasses implement _freeze with the backend-specific snapshot
+    operation and optionally override _export_hint for migration hints.
+    """
+
+    backend_name: str
+
+    def commit(self, agent: ActiveAgent) -> None:
+        """Freeze the bottle for `agent` to a resumable artifact.
+
+        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.
+
+        Raises CommitCancelled if the user declines an interactive
+        confirmation prompt (e.g. the macos-container stop prompt).
+        """
+        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}")
+        self._export_hint(agent.slug, image_ref)
+
+    @abstractmethod
+    def _freeze(self, agent: ActiveAgent) -> str:
+        """Backend-specific snapshot. Returns the image tag or artifact path
+        stored by write_committed_image. Raises CommitCancelled if the user
+        declines a stop-confirmation prompt."""
+
+    def _export_hint(self, slug: str, image_ref: str) -> None:
+        """Optionally print an export-for-migration hint after committing.
+        Overridden by backends that provide a meaningful export command."""
+
+    def commit_slug(self, slug: str) -> None:
+        """Convenience entry for cmd_commit when only a slug is available."""
+        from ..bottle_state import read_metadata
+        metadata = read_metadata(slug)
+        agent = ActiveAgent(
+            backend_name=self.backend_name,
+            slug=slug,
+            agent_name=metadata.agent_name if metadata else "",
+            started_at=metadata.started_at if metadata else "",
+            services=(),
+        )
+        self.commit(agent)
+
+
+def get_freezer(backend_name: str) -> Freezer:
+    """Return the Freezer for the named backend.
+
+    backend_name "" is treated as "docker" for backward compatibility
+    with state dirs written before the backend field was added."""
+    resolved = backend_name or "docker"
+    if resolved == "docker":
+        from .docker.freezer import DockerFreezer
+        return DockerFreezer()
+    if resolved == "macos-container":
+        from .macos_container.freezer import MacosContainerFreezer
+        return MacosContainerFreezer()
+    if resolved == "smolmachines":
+        from .smolmachines.freezer import SmolmachinesFreezer
+        return SmolmachinesFreezer()
+    die(
+        f"commit is only supported for docker, macos-container, and "
+        f"smolmachines; backend {backend_name!r} has no freezer"
+    )
+    raise AssertionError("unreachable")

bot_bottle/backend/macos_container/backend.py

@@ -11,6 +11,7 @@ from ...egress import EgressPlan
 from ...env import ResolvedEnv
 from ...git_gate import GitGatePlan
 from ...supervise import SupervisePlan
+from ...manifest import Manifest
 from .. import ActiveAgent, BottleBackend, BottleSpec
 from . import cleanup as _cleanup
 from . import enumerate as _enumerate
@@ -45,6 +46,7 @@ class MacosContainerBottleBackend(
         self,
         spec: BottleSpec,
         *,
+        manifest: Manifest,
         slug: str,
         resolved_env: ResolvedEnv,
         agent_provision_plan: AgentProvisionPlan,
@@ -55,6 +57,7 @@ class MacosContainerBottleBackend(
     ) -> MacosContainerBottlePlan:
         return _resolve_plan.resolve_plan(
             spec,
+            manifest=manifest,
             slug=slug,
             resolved_env=resolved_env,
             agent_provision_plan=agent_provision_plan,

bot_bottle/backend/macos_container/bottle.py

@@ -2,12 +2,41 @@
 
 from __future__ import annotations
 
+import os
 import subprocess
+import sys
 from typing import Callable, cast
 
 from ...agent_provider import PromptMode, prompt_args
 from .. import Bottle, ExecResult
 from ..terminal import exec_shell_script
+from . import pty_forward as _pty_forward
+
+
+_PTY_FORWARD_SCRIPT = _pty_forward.__file__
+_TERMINAL_ENV_NAMES = (
+    "TERM",
+    "COLORTERM",
+    "TERM_PROGRAM",
+    "TERM_PROGRAM_VERSION",
+    "KITTY_WINDOW_ID",
+    "KITTY_PID",
+    "WEZTERM_PANE",
+    "WEZTERM_UNIX_SOCKET",
+    "GHOSTTY_BIN_DIR",
+    "GHOSTTY_RESOURCES_DIR",
+    "ITERM_SESSION_ID",
+    "VTE_VERSION",
+    "KONSOLE_VERSION",
+    "ALACRITTY_WINDOW_ID",
+)
+
+
+def _terminal_env_names() -> tuple[str, ...]:
+    return tuple(
+        name for name in _TERMINAL_ENV_NAMES
+        if name == "TERM" or os.environ.get(name)
+    )
 
 
 class MacosContainerBottle(Bottle):
@@ -44,13 +73,24 @@ class MacosContainerBottle(Bottle):
                 argv=full_argv,
             )
         )
-        cmd = ["container", "exec"]
+        container_exec = ["container", "exec"]
         if tty:
-            cmd.extend(["--interactive", "--tty"])
+            container_exec.extend(["--interactive", "--tty"])
+            # Forward terminal capability hints so TUIs can enable modified-key
+            # protocols. Use bare env names: values stay in the child env, not
+            # on argv, and pty_forward supplies a TERM fallback when needed.
+            for name in _terminal_env_names():
+                container_exec.extend(["--env", name])
         if self.agent_workdir and self.agent_workdir != "/home/node":
-            cmd.extend(["--workdir", self.agent_workdir])
-        cmd.extend([self.name, self.agent_command, *full_argv])
-        return cmd
+            container_exec.extend(["--workdir", self.agent_workdir])
+        container_exec.extend([self.name, self.agent_command, *full_argv])
+        if tty:
+            # Wrap with the raw-mode forwarder: container exec does not put
+            # the host terminal into raw mode itself, so the line discipline
+            # buffers modifier-key sequences until CR.  The wrapper sets raw
+            # mode before exec and restores it on exit.
+            return [sys.executable, _PTY_FORWARD_SCRIPT, "--", *container_exec]
+        return container_exec
 
     def exec_agent(self, argv: list[str], *, tty: bool = True) -> int:
         agent_argv = self.agent_argv(argv, tty=tty)

bot_bottle/backend/macos_container/egress_apply.py

@@ -0,0 +1,39 @@
+"""Host-side egress apply for the macos-container backend.
+
+Uses `container kill --signal HUP` (Apple Container framework) instead
+of `docker kill` to signal the sidecar bundle.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+
+from ...log import warn
+from ..egress_apply import EgressApplicator, EgressApplyError
+from .launch import sidecar_container_name
+
+
+class MacOSContainerEgressApplicator(EgressApplicator):
+    def _signal_bundle_reload(self, slug: str) -> None:
+        container = sidecar_container_name(slug)
+        result = subprocess.run(
+            ["container", "kill", "--signal", "HUP", container],
+            capture_output=True, text=True, check=False, env=os.environ,
+        )
+        if result.returncode != 0:
+            last_error = (result.stderr or "").strip() or (result.stdout or "").strip()
+            warn(
+                f"egress: routes updated on disk for {slug}, but bundle reload failed: "
+                f"{last_error or 'container kill failed'}"
+            )
+            raise EgressApplyError(
+                f"could not reload egress bundle {container}: "
+                f"{last_error or 'container kill failed'}"
+            )
+
+
+applicator = MacOSContainerEgressApplicator()
+
+
+__all__ = ["MacOSContainerEgressApplicator", "EgressApplyError", "applicator"]

bot_bottle/backend/macos_container/freezer.py

@@ -0,0 +1,31 @@
+"""MacosContainerFreezer — snapshot a macOS container bottle.
+
+Apple Container removes containers when they stop, making stop-then-export
+impossible. Instead, commit_container execs into the running container and
+streams the root filesystem via tar. The bottle continues running after commit.
+"""
+
+from __future__ import annotations
+
+from .. import ActiveAgent
+from ..freeze import Freezer
+from .util import commit_container
+from ...log import info
+
+
+class MacosContainerFreezer(Freezer):
+    """Freezes a macOS-container bottle via exec-tar + image rebuild."""
+
+    backend_name = "macos-container"
+
+    def _freeze(self, agent: ActiveAgent) -> str:
+        container = f"bot-bottle-{agent.slug}"
+        image_tag = f"bot-bottle-committed-{agent.slug}:latest"
+        commit_container(container, image_tag)
+        return image_tag
+
+    def _export_hint(self, slug: str, image_ref: str) -> None:
+        info(
+            f"to export for migration:      "
+            f"container image save {image_ref} -o {slug}.tar"
+        )

bot_bottle/backend/macos_container/launch.py

@@ -12,14 +12,22 @@ from __future__ import annotations
 
 import dataclasses
 import os
-import shutil
 import subprocess
 from contextlib import ExitStack, contextmanager
 from pathlib import Path
 from typing import Callable, Generator
 
-from ...bottle_state import egress_state_dir, git_gate_state_dir
-from ...egress import EGRESS_ROUTES_IN_CONTAINER, egress_resolve_token_values
+from ...bottle_state import (
+    egress_state_dir,
+    git_gate_state_dir,
+    read_committed_image,
+)
+from ...egress import (
+    EGRESS_ROUTES_IN_CONTAINER,
+    egress_agent_env_entries,
+    egress_resolve_token_values,
+    egress_sidecar_env_entries,
+)
 from ...git_gate import revoke_git_gate_provisioned_keys
 from ...log import die, info, warn
 from ...supervise import QUEUE_DIR_IN_CONTAINER, SUPERVISE_PORT
@@ -68,7 +76,7 @@ def launch(
 ) -> Generator[MacosContainerBottle, None, None]:
     """Build, run, provision, and yield an Apple Container bottle."""
     stack = ExitStack()
-    bottle_for_revoke = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+    bottle_for_revoke = plan.manifest.bottle
     git_gate_dir_for_revoke = git_gate_state_dir(plan.slug)
 
     def teardown() -> None:
@@ -84,7 +92,7 @@ def launch(
 
     try:
         plan = _mint_certs(plan)
-        _build_images(plan)
+        plan = _build_images(plan)
 
         internal_network = internal_network_name(plan.slug)
         egress_network = egress_network_name(plan.slug)
@@ -112,7 +120,7 @@ def launch(
             agent_command=plan.agent_command,
             agent_prompt_mode=plan.agent_prompt_mode,
             agent_provider_template=plan.agent_provider_template,
-            terminal_title=plan.spec.label or plan.spec.agent_name,
+            terminal_title=f"{plan.spec.label} ({plan.spec.agent_name})" if plan.spec.label else plan.spec.agent_name,
             terminal_color=plan.spec.color,
             agent_workdir=plan.workspace_plan.workdir,
         )
@@ -135,17 +143,28 @@ def _mint_certs(plan: MacosContainerBottlePlan) -> MacosContainerBottlePlan:
     return dataclasses.replace(plan, egress_plan=egress_plan)
 
 
-def _build_images(plan: MacosContainerBottlePlan) -> None:
+def _build_images(plan: MacosContainerBottlePlan) -> MacosContainerBottlePlan:
     container_mod.build_image(
         SIDECAR_BUNDLE_IMAGE,
         _REPO_DIR,
         dockerfile=SIDECAR_BUNDLE_DOCKERFILE,
     )
+    committed = read_committed_image(plan.slug)
+    if committed and container_mod.image_exists(committed):
+        info(f"using committed image {committed!r}")
+        return dataclasses.replace(
+            plan,
+            agent_provision=dataclasses.replace(
+                plan.agent_provision,
+                image=committed,
+            ),
+        )
     container_mod.build_image(
         plan.image,
         _REPO_DIR,
         dockerfile=plan.dockerfile_path,
     )
+    return plan
 
 
 def _create_networks(
@@ -314,7 +333,6 @@ def _agent_run_argv(
         "container", "run",
         "--name", plan.container_name,
         "--detach",
-        "--rm",
         "--network", internal_network,
     ]
     for entry in _agent_env_entries(plan, sidecar_ip):
@@ -337,9 +355,7 @@ def _sidecar_daemons(plan: MacosContainerBottlePlan) -> tuple[str, ...]:
 
 
 def _sidecar_env_entries(plan: MacosContainerBottlePlan) -> tuple[str, ...]:
-    env: list[str] = []
-    if plan.egress_plan.routes:
-        env.extend(sorted(plan.egress_plan.token_env_map.keys()))
+    env: list[str] = list(egress_sidecar_env_entries(plan.egress_plan))
     if plan.git_gate_plan.upstreams:
         env.append(f"BOT_BOTTLE_GIT_GATE_READY_FILE={_GIT_GATE_READY_FILE}")
     if plan.supervise_plan is not None:
@@ -364,7 +380,7 @@ def _sidecar_mounts(
     ))
     if ep.routes:
         mounts.append((
-            str(_stage_routes_dir(plan)),
+            str(ep.routes_path.parent),
             str(Path(EGRESS_ROUTES_IN_CONTAINER).parent),
             True,
         ))
@@ -375,17 +391,6 @@ def _sidecar_mounts(
 
     return tuple(mounts)
 
-
-def _stage_routes_dir(plan: MacosContainerBottlePlan) -> Path:
-    routes_dir = plan.stage_dir / "macos-container-egress"
-    routes_dir.mkdir(parents=True, exist_ok=True)
-    shutil.copyfile(
-        plan.egress_plan.routes_path,
-        routes_dir / Path(EGRESS_ROUTES_IN_CONTAINER).name,
-    )
-    return routes_dir
-
-
 def _mount_spec(host_path: str, container_path: str, read_only: bool) -> str:
     spec = f"type=bind,source={host_path},target={container_path}"
     if read_only:
@@ -418,6 +423,7 @@ def _agent_env_entries(
         env.append(f"{name}={value}")
     for name in sorted(plan.forwarded_env.keys()):
         env.append(name)
+    env.extend(egress_agent_env_entries(plan.egress_plan))
     return tuple(env)
 
 

bot_bottle/backend/macos_container/pty_forward.py

@@ -0,0 +1,70 @@
+"""Host-side raw-mode wrapper for `container exec --interactive --tty`.
+
+Apple's `container exec --interactive --tty` does not set the host terminal to
+raw mode before starting its I/O relay.  Without raw mode the kernel line
+discipline buffers modifier-key escape sequences (e.g. Shift+Enter in
+modifyOtherKeys mode produces \\x1b[13;2~) until a carriage-return arrives, so
+they never reach Claude Code inside the container.
+
+This module sets the host terminal to raw mode, spawns the inner argv (the
+container exec command), and restores the original terminal attributes on
+exit.  When stdin is not a TTY (piped invocations, CI) it falls through to a
+bare subprocess.run so callers do not need to special-case non-interactive
+contexts.
+
+Usage (the `--` separator is the API contract — everything after it is the
+inner command):
+
+    python pty_forward.py -- container exec --interactive --tty <name> <cmd>
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+import termios
+import tty
+
+
+def _inner_env() -> dict[str, str]:
+    env = dict(os.environ)
+    env.setdefault("TERM", "xterm-256color")
+    return env
+
+
+def _run_inner(inner: list[str]) -> int:
+    return subprocess.run(inner, check=False, env=_inner_env()).returncode
+
+
+def main(argv: list[str]) -> int:
+    """Entry point.  ``argv`` shape: ``-- <inner-argv...>``."""
+    if len(argv) < 2 or argv[0] != "--":
+        sys.stderr.write(
+            "usage: python pty_forward.py -- <container-exec-argv...>\n"
+        )
+        return 2
+    inner = argv[1:]
+
+    try:
+        fd = sys.stdin.fileno()
+    except OSError:
+        return _run_inner(inner)
+
+    if not os.isatty(fd):
+        return _run_inner(inner)
+
+    try:
+        old = termios.tcgetattr(fd)
+    except termios.error:
+        return _run_inner(inner)
+
+    try:
+        tty.setraw(fd)
+        return _run_inner(inner)
+    finally:
+        termios.tcsetattr(fd, termios.TCSADRAIN, old)
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv[1:]))

bot_bottle/backend/macos_container/resolve_plan.py

@@ -9,6 +9,7 @@ from ...egress import EgressPlan
 from ...env import ResolvedEnv
 from ...git_gate import GitGatePlan
 from ...supervise import SupervisePlan
+from ...manifest import Manifest
 from .. import BottleSpec
 from . import util as container_mod
 from .bottle_plan import MacosContainerBottlePlan
@@ -24,6 +25,7 @@ def build_guest_env(resolved_env: ResolvedEnv) -> dict[str, str]:
 
 def resolve_plan(
     spec: BottleSpec,
+    manifest: Manifest,
     slug: str,
     resolved_env: ResolvedEnv,
     agent_provision_plan: AgentProvisionPlan,
@@ -34,6 +36,7 @@ def resolve_plan(
 ) -> MacosContainerBottlePlan:
     return MacosContainerBottlePlan(
         spec=spec,
+        manifest=manifest,
         stage_dir=stage_dir,
         slug=slug,
         forwarded_env=dict(resolved_env.forwarded),

bot_bottle/backend/macos_container/util.py

@@ -8,6 +8,7 @@ import ipaddress
 import platform
 import shutil
 import subprocess
+import tempfile
 import time
 from typing import Iterable
 
@@ -67,11 +68,63 @@ def build_image(ref: str, context: str, *, dockerfile: str = "") -> None:
     _ensure_builder_dns()
     args = [_CONTAINER, "build", "-t", ref, "--dns", dns_server()]
     if dockerfile:
+        # `container build` resolves -f relative to the current working
+        # directory, not the build context. Anchor a relative Dockerfile to
+        # the context so builds work from any cwd.
+        if not os.path.isabs(dockerfile):
+            dockerfile = os.path.join(context, dockerfile)
         args.extend(["-f", dockerfile])
     args.append(context)
     subprocess.run(args, check=True)
 
 
+def commit_container(container_name: str, image_tag: str) -> None:
+    """Snapshot a running Apple Container as a local image.
+
+    `container export` requires a stopped container, but Apple Container
+    removes containers when they stop, making stop-then-export impossible.
+    Instead, exec into the running container as root and stream the root
+    filesystem out via tar, then build a new image from that archive.
+    The bottle continues running after commit.
+    """
+    with tempfile.TemporaryDirectory(prefix="bot-bottle-container-commit.") as tmp:
+        rootfs_tar = os.path.join(tmp, "rootfs.tar")
+        dockerfile = os.path.join(tmp, "Dockerfile")
+        with open(rootfs_tar, "wb") as tar_out:
+            result = subprocess.run(
+                [
+                    _CONTAINER, "exec",
+                    "--user", "root",
+                    container_name,
+                    "tar", "--create",
+                    "--exclude=./proc",
+                    "--exclude=./sys",
+                    "--exclude=./dev",
+                    "--exclude=./run",
+                    "--file=-",
+                    "--directory=/",
+                    ".",
+                ],
+                stdout=tar_out,
+                stderr=subprocess.PIPE,
+                check=False,
+            )
+        if result.returncode != 0:
+            die(
+                f"container exec tar {container_name!r} failed: "
+                f"{(result.stderr or b'').decode().strip() or '<no stderr>'}"
+            )
+        with open(dockerfile, "w", encoding="utf-8") as f:
+            f.write(
+                "FROM scratch\n"
+                "ADD rootfs.tar /\n"
+                "USER node\n"
+                "WORKDIR /home/node\n"
+            )
+        build_image(image_tag, tmp, dockerfile=dockerfile)
+    info(f"committed {container_name!r} → {image_tag!r}")
+
+
 def _ensure_builder_dns() -> None:
     dns = dns_server()
     status = _builder_status()
@@ -218,6 +271,36 @@ def container_exists(name: str) -> bool:
     return name in {line.strip() for line in result.stdout.splitlines()}
 
 
+def container_is_running(name: str) -> bool:
+    """Return True if the named container is currently running.
+
+    `container list` without `--all` lists only running containers."""
+    result = subprocess.run(
+        [_CONTAINER, "list", "--quiet"],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        return False
+    return name in {line.strip() for line in result.stdout.splitlines()}
+
+
+def stop_container(name: str) -> None:
+    """Stop the named container without deleting it."""
+    result = subprocess.run(
+        [_CONTAINER, "stop", name],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        die(
+            f"container stop {name!r} failed: "
+            f"{(result.stderr or '').strip() or '<no stderr>'}"
+        )
+
+
 def force_remove_container(name: str) -> None:
     if container_exists(name):
         subprocess.run(

bot_bottle/backend/resolve_common.py

@@ -26,15 +26,25 @@ from ..bottle_state import (
 )
 from ..egress import Egress, EgressPlan
 from ..git_gate import GitGate, GitGatePlan
-from ..manifest import ManifestBottle
+from ..manifest import Manifest, ManifestBottle
 from ..supervise import Supervise, SupervisePlan
 from . import BottleSpec
 
 
 def mint_slug(spec: BottleSpec) -> str:
     """Return the bottle identity: the recorded identity for a resume,
-    or a freshly minted one for a new start."""
-    return spec.identity or bottle_identity(spec.agent_name)
+    or a freshly minted one for a new start.
+
+    When a label is provided it becomes the full slug (no random suffix),
+    so two launches with the same label collide by design.  When no label
+    is given the identity is minted with a random suffix to avoid
+    collisions between anonymous launches of the same agent."""
+    if spec.identity:
+        return spec.identity
+    if spec.label:
+        from .docker import util as docker_mod
+        return docker_mod.slugify(spec.label)
+    return bottle_identity(spec.agent_name)
 
 
 def write_launch_metadata(
@@ -53,14 +63,14 @@ def write_launch_metadata(
         backend=backend,
         label=spec.label,
         color=spec.color,
+        bottle_names=spec.bottle_names,
     ))
 
 
-def prepare_agent_state_dir(slug: str, spec: BottleSpec) -> tuple[Path, Path]:
+def prepare_agent_state_dir(slug: str, manifest: Manifest) -> tuple[Path, Path]:
     """Create the agent state subdir, write the prompt file.
     Returns (agent_dir, prompt_file)."""
-    manifest = spec.manifest
-    agent = manifest.agents[spec.agent_name]
+    agent = manifest.agent
     agent_dir = agent_state_dir(slug)
     agent_dir.mkdir(parents=True, exist_ok=True)
     prompt_file = agent_dir / "prompt.txt"

bot_bottle/backend/smolmachines/backend.py

@@ -18,6 +18,7 @@ from ...egress import EgressPlan
 from ...env import ResolvedEnv
 from ...git_gate import GitGatePlan
 from ...supervise import SupervisePlan
+from ...manifest import Manifest
 from .. import ActiveAgent, BottleBackend, BottleSpec
 from . import cleanup as _cleanup
 from . import enumerate as _enumerate
@@ -55,6 +56,7 @@ class SmolmachinesBottleBackend(
         self,
         spec: BottleSpec,
         *,
+        manifest: Manifest,
         slug: str,
         resolved_env: ResolvedEnv,
         agent_provision_plan: AgentProvisionPlan,
@@ -65,6 +67,7 @@ class SmolmachinesBottleBackend(
     ) -> SmolmachinesBottlePlan:
         return _resolve_plan.resolve_plan(
             spec,
+            manifest=manifest,
             slug=slug,
             resolved_env=resolved_env,
             agent_provision_plan=agent_provision_plan,

bot_bottle/backend/smolmachines/bottle.py

@@ -145,7 +145,12 @@ class SmolmachinesBottle(Bottle):
         script = exec_shell_script(agent_argv, self.terminal_title, self.terminal_color) if tty else None
         if script is None:
             return subprocess.run(agent_argv, check=False).returncode
-        return subprocess.run(["sh", "-lc", script], check=False).returncode
+        # Use sh -c (not -lc) so the script inherits PATH from the calling
+        # process. sh -l sources login-shell init files (e.g. /etc/profile)
+        # which may NOT include smolvm's location when it was installed via
+        # homebrew. The calling process (./cli.py) already has smolvm on PATH
+        # (provision steps succeed), so -c is sufficient.
+        return subprocess.run(["sh", "-c", script], check=False).returncode
 
     # smolvm/libkrun can SIGKILL an otherwise-normal exec during
     # early-VM provisioning. Retry once after a short settle so

bot_bottle/backend/smolmachines/egress_apply.py

@@ -0,0 +1,21 @@
+"""Egress apply for the smolmachines backend.
+
+The smolmachines sidecar bundle runs as a host-side Docker container,
+so egress signalling is identical to the docker backend.
+"""
+
+from __future__ import annotations
+
+from ..docker.egress_apply import (  # noqa: F401
+    DockerEgressApplicator,
+    EgressApplyError,
+    applicator,
+    fetch_current_routes,
+)
+
+__all__ = [
+    "DockerEgressApplicator",
+    "EgressApplyError",
+    "applicator",
+    "fetch_current_routes",
+]

bot_bottle/backend/smolmachines/freezer.py

@@ -0,0 +1,145 @@
+"""SmolmachinesFreezer — snapshot a smolmachines bottle.
+
+`smolvm pack create --from-vm` requires the VM to be stopped, and smolvm
+removes VMs when stopped (same issue as Apple Container). Instead, exec
+into the running VM as root to write a gzip-compressed tar of the root
+filesystem to /var/tmp, then copy it to the host with `smolvm machine cp`,
+build a Docker image from the archive, convert it to a smolmachine artifact
+via the existing registry pipeline, and record the sidecar path. The VM
+stays running throughout."""
+
+from __future__ import annotations
+
+import tempfile
+from pathlib import Path
+
+from .. import ActiveAgent
+from ..freeze import Freezer
+from ..docker import util as docker_mod
+from .local_registry import crane_push_tarball, ephemeral_registry
+from .smolvm import machine_cp, machine_exec, pack_create
+from ...bottle_state import bottle_state_dir
+from ...log import die, info
+
+
+# Temp file written inside the VM during commit. Lives in /var/tmp
+# (on-disk, unlike tmpfs /tmp) to survive for machine_cp.
+_VM_COMMIT_TAR = "/var/tmp/.bot-bottle-commit.tar.gz"
+
+
+class SmolmachinesFreezer(Freezer):
+    """Freezes a smolmachines bottle via exec-tar + Docker image + smolmachine pack.
+
+    The VM is NOT stopped. We exec into the running VM to write a compressed
+    tar of the root filesystem to /var/tmp, copy it to the host with
+    machine_cp, build a Docker image (Docker's ADD decompresses .tar.gz
+    automatically), then run the same image→registry→pack_create pipeline
+    that _ensure_smolmachine uses for fresh builds."""
+
+    backend_name = "smolmachines"
+
+    def _freeze(self, agent: ActiveAgent) -> str:
+        machine = f"bot-bottle-{agent.slug}"
+        image_ref = f"bot-bottle-committed-{agent.slug}:latest"
+        output_dir = bottle_state_dir(agent.slug)
+        output_dir.mkdir(parents=True, exist_ok=True)
+        binary = output_dir / "committed-smolmachine"
+        sidecar = output_dir / "committed-smolmachine.smolmachine"
+        _snapshot_running_vm(machine, image_ref, binary)
+        return str(sidecar)
+
+    def _export_hint(self, slug: str, image_ref: str) -> None:
+        info(f"to export for migration:      cp {image_ref} {slug}.smolmachine")
+
+
+def _snapshot_running_vm(machine: str, image_ref: str, binary: Path) -> None:
+    """Exec-tar the running VM, build a Docker image, and pack to a smolmachine.
+
+    binary:  destination for the launcher (sibling .smolmachine is the artifact
+             that machine_create --from consumes, same convention as pack_create).
+    """
+    with tempfile.TemporaryDirectory(prefix="bot-bottle-vm-commit.") as tmp:
+        tmp_path = Path(tmp)
+        # Use .tar.gz — Docker ADD decompresses automatically and the
+        # compressed archive fits in the VM's /var/tmp more easily.
+        rootfs_tar_gz = tmp_path / "rootfs.tar.gz"
+        dockerfile = tmp_path / "Dockerfile"
+
+        _exec_tar_to_file(machine, rootfs_tar_gz)
+
+        dockerfile.write_text(
+            "FROM scratch\n"
+            "ADD rootfs.tar.gz /\n"
+            "USER node\n"
+            "WORKDIR /home/node\n"
+        )
+        docker_mod.build_image(image_ref, str(tmp_path), dockerfile=str(dockerfile))
+
+    image_tarball = binary.parent / "committed.image.tar"
+    docker_mod.save(image_ref, str(image_tarball))
+    try:
+        with ephemeral_registry() as handle:
+            digest = docker_mod.image_id(image_ref).split(":", 1)[-1][:16]
+            push_ref = f"{handle.push_endpoint}/bot-bottle-committed:{digest}"
+            pack_ref = f"{handle.pull_endpoint}/bot-bottle-committed:{digest}"
+            crane_push_tarball(handle, str(image_tarball), push_ref)
+            pack_create(pack_ref, binary)
+    finally:
+        image_tarball.unlink(missing_ok=True)
+
+
+def _exec_tar_to_file(machine: str, dest: Path) -> None:
+    """Snapshot the running VM's root filesystem to dest (.tar.gz).
+
+    Writes a gzip-compressed tar to _VM_COMMIT_TAR inside the VM via
+    machine_exec (same mechanism as provisioning), then copies it to the
+    host with machine_cp. This avoids binary-stdout piping through the
+    smolvm exec channel, which does not reliably handle large binary output.
+
+    A connectivity probe (machine_exec true) runs first so a concurrent-exec
+    limitation (smolvm may reject a second exec while -i -t is active) is
+    reported clearly rather than as a silent failure."""
+    # Connectivity probe — if smolvm rejects concurrent exec while an
+    # interactive session is running, fail clearly here.
+    probe = machine_exec(machine, ["true"])
+    if probe.returncode != 0:
+        die(
+            f"smolvm exec is not available for {machine!r} "
+            f"(exit {probe.returncode}: {probe.stderr.strip() or probe.stdout.strip() or '<no output>'}). "
+            f"If an interactive session is active, smolvm may not support concurrent exec."
+        )
+
+    # Create the compressed tar inside the VM.
+    # tar exits 1 when files change during archiving (normal for a live
+    # filesystem); only treat exit > 1 as fatal.
+    tar_result = machine_exec(
+        machine,
+        [
+            "tar", "--create", "--gzip",
+            "--exclude=./proc",
+            "--exclude=./sys",
+            "--exclude=./dev",
+            "--exclude=./run",
+            # /tmp and /var/tmp are ephemeral. Their stale contents
+            # (e.g. /tmp/claude-<uid>) have uid remapped by smolvm's
+            # pack process, causing Claude Code to refuse to use them
+            # on resume. Exclude both; _init_vm recreates them with
+            # mkdir -p + correct ownership on every boot.
+            "--exclude=./tmp",
+            "--exclude=./var/tmp",
+            f"--file={_VM_COMMIT_TAR}",
+            "--directory=/",
+            ".",
+        ],
+    )
+    if tar_result.returncode > 1:
+        die(
+            f"smolvm exec tar {machine!r} failed (exit {tar_result.returncode}): "
+            f"{tar_result.stderr.strip() or tar_result.stdout.strip() or '<no output>'}"
+        )
+
+    # Copy from VM to host, then clean up.
+    try:
+        machine_cp(f"{machine}:{_VM_COMMIT_TAR}", str(dest))
+    finally:
+        machine_exec(machine, ["rm", "-f", _VM_COMMIT_TAR])

bot_bottle/backend/smolmachines/launch.py

@@ -23,7 +23,9 @@ from typing import Callable, Generator
 
 from ...egress import (
     EGRESS_ROUTES_IN_CONTAINER,
+    egress_agent_env_entries,
     egress_resolve_token_values,
+    egress_sidecar_env_entries,
 )
 from ...supervise import QUEUE_DIR_IN_CONTAINER, SUPERVISE_PORT
 from ...util import expand_tilde
@@ -40,8 +42,12 @@ from ..docker.git_gate import (
     GIT_GATE_HOOK_IN_CONTAINER,
 )
 from ...git_gate import revoke_git_gate_provisioned_keys
-from ...log import warn
-from ...bottle_state import egress_state_dir, git_gate_state_dir
+from ...log import info, warn
+from ...bottle_state import (
+    egress_state_dir,
+    git_gate_state_dir,
+    read_committed_image,
+)
 from . import loopback_alias as _loopback
 from . import sidecar_bundle as _bundle
 from . import smolvm as _smolvm
@@ -85,14 +91,7 @@ def launch(
         plan = _start_bundle(plan, network, loopback_ip, stack)
         plan = _discover_urls(plan, loopback_ip)
 
-        # Build the agent image and pack it into a `.smolmachine`
-        # artifact (or hit the per-Dockerfile-digest cache). Runs
-        # here, not in prepare, so the docker-build output doesn't
-        # garble the dashboard's preflight modal.
-        agent_from_path = _ensure_smolmachine(
-            plan.agent_image,
-            dockerfile=plan.agent_dockerfile_path,
-        )
+        agent_from_path = _agent_from_path(plan)
 
         _launch_vm(plan, agent_from_path, loopback_ip, stack)
         _init_vm(plan)
@@ -104,7 +103,7 @@ def launch(
             agent_command=plan.agent_command,
             agent_prompt_mode=plan.agent_prompt_mode,
             agent_provider_template=plan.agent_provider_template,
-            terminal_title=plan.spec.label or plan.spec.agent_name,
+            terminal_title=f"{plan.spec.label} ({plan.spec.agent_name})" if plan.spec.label else plan.spec.agent_name,
             terminal_color=plan.spec.color,
             agent_workdir=plan.workspace_plan.workdir,
         )
@@ -130,7 +129,7 @@ def _teardown_smolmachines(
     except BaseException as exc:  # noqa: W0718 — teardown must not fail
         teardown_exc = exc
         warn(f"smolmachines teardown failed: {exc!r}")
-    bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+    bottle = plan.manifest.bottle
     revoke_git_gate_provisioned_keys(bottle, git_gate_state_dir(plan.slug))
     if teardown_exc is not None:
         raise teardown_exc
@@ -217,16 +216,23 @@ def _discover_urls(
         agent_supervise_url = f"http://{loopback_ip}:{supervise_host_port}/"
 
     existing_no_proxy = plan.guest_env.get("NO_PROXY", "localhost,127.0.0.1")
+    no_proxy = f"{existing_no_proxy},{loopback_ip}"
     guest_env = {
         **plan.guest_env,
         "HTTPS_PROXY": agent_proxy_url,
         "HTTP_PROXY":  agent_proxy_url,
-        "NO_PROXY":    f"{existing_no_proxy},{loopback_ip}",
+        "https_proxy": agent_proxy_url,
+        "http_proxy":  agent_proxy_url,
+        "NO_PROXY":    no_proxy,
+        "no_proxy":    no_proxy,
     }
     if agent_git_gate_host:
         guest_env["GIT_GATE_URL"] = f"http://{agent_git_gate_host}"
     if agent_supervise_url:
         guest_env["MCP_SUPERVISE_URL"] = agent_supervise_url
+    for entry in egress_agent_env_entries(plan.egress_plan):
+        name, value = entry.split("=", 1)
+        guest_env[name] = value
 
     return dataclasses.replace(
         plan,
@@ -275,10 +281,16 @@ def _init_vm(plan: SmolmachinesBottlePlan) -> None:
     All folded into one sh -c to avoid back-to-back exec calls
     immediately after machine_start (libkrun exec-channel race).
 
+    mkdir -p guards: when booting from a committed snapshot, /tmp and
+    /var/tmp are excluded from the archive (they're ephemeral and their
+    stale contents would have wrong uid after smolvm's uid remap). The
+    directories must be created before chown/chmod can set permissions.
+
     wait_exec_ready polls until the exec channel is ready for the
     subsequent provision calls, replacing the empirical sleep."""
     _smolvm.machine_exec(plan.machine_name, [
         "sh", "-c",
+        "mkdir -p /tmp /var/tmp && "
         "chown -R node:node /home/node && "
         "chown root:root /tmp /var/tmp && "
         "chmod 1777 /tmp /var/tmp",
@@ -308,12 +320,8 @@ def _bundle_launch_spec(
     ep = plan.egress_plan
     volumes.append((str(ep.mitmproxy_ca_host_path), EGRESS_CA_IN_CONTAINER, True))
     if ep.routes:
-        volumes.append((str(ep.routes_path), EGRESS_ROUTES_IN_CONTAINER, True))
-        # Bare-name entries for upstream-token slots. Their values
-        # come from the docker-run subprocess env (inherited from
-        # the operator's shell), never landing on argv.
-        for token_env in sorted(ep.token_env_map.keys()):
-            env.append(token_env)
+        volumes.append((str(ep.routes_path.parent), str(Path(EGRESS_ROUTES_IN_CONTAINER).parent), True))
+    env.extend(egress_sidecar_env_entries(ep))
 
     # --- git-gate ---------------------------------------------
     gp = plan.git_gate_plan
@@ -382,6 +390,30 @@ def _resolve_token_env(
     return egress_resolve_token_values(plan.egress_plan.token_env_map, effective_env)
 
 
+def _agent_from_path(plan: SmolmachinesBottlePlan) -> Path:
+    """Return the `.smolmachine` artifact used for `machine create --from`.
+
+    Prefer a committed VM artifact when one is recorded and still
+    present. If the file was removed, fall back to the normal image
+    build + pack cache path.
+    """
+    committed = read_committed_image(plan.slug)
+    if committed:
+        committed_path = Path(committed)
+        if committed_path.is_file():
+            info(f"using committed smolmachine {str(committed_path)!r}")
+            return committed_path
+
+    # Build the agent image and pack it into a `.smolmachine`
+    # artifact (or hit the per-Dockerfile-digest cache). Runs here,
+    # not in prepare, so the docker-build output doesn't garble the
+    # dashboard's preflight modal.
+    return _ensure_smolmachine(
+        plan.agent_image,
+        dockerfile=plan.agent_dockerfile_path,
+    )
+
+
 def _ensure_smolmachine(image_ref: str, *, dockerfile: str = "") -> Path:
     """Build the agent docker image and convert it into a
     `.smolmachine` artifact, caching the result under

bot_bottle/backend/smolmachines/resolve_plan.py

@@ -13,6 +13,7 @@ from __future__ import annotations
 from pathlib import Path
 
 from .. import BottleSpec
+from ...manifest import Manifest
 from ...env import ResolvedEnv
 from ...agent_provider import AgentProvisionPlan
 from ...egress import EgressPlan
@@ -46,6 +47,7 @@ def build_guest_env(resolved_env: ResolvedEnv) -> dict[str, str]:
 
 def resolve_plan(
     spec: BottleSpec,
+    manifest: Manifest,
     slug: str,
     resolved_env: ResolvedEnv,
     agent_provision_plan: AgentProvisionPlan,
@@ -67,6 +69,7 @@ def resolve_plan(
 
     return SmolmachinesBottlePlan(
         spec=spec,
+        manifest=manifest,
         stage_dir=stage_dir,
         slug=slug,
         bundle_subnet=subnet,

bot_bottle/backend/smolmachines/smolvm.py

@@ -25,6 +25,7 @@ smolvm binary."""
 
 from __future__ import annotations
 
+import json
 import shutil
 import subprocess
 import time
@@ -94,6 +95,16 @@ def pack_create(image: str, output: Path) -> None:
     _smolvm("pack", "create", "--image", image, "-o", str(output))
 
 
+def pack_create_from_vm(name: str, output: Path) -> None:
+    """`smolvm pack create --from-vm <name> -o <output>`.
+
+    Snapshots an existing persistent VM into a pack artifact. As
+    with `pack_create`, smolvm writes a launcher at `output` and the
+    bootable sidecar at `output.smolmachine`.
+    """
+    _smolvm("pack", "create", "--from-vm", name, "-o", str(output))
+
+
 # --- Machine lifecycle ---------------------------------------------------
 
 
@@ -143,6 +154,21 @@ def machine_create(
     _smolvm(*args)
 
 
+def machine_is_running(name: str) -> bool:
+    """Return True if the named VM is in the 'running' state."""
+    result = _smolvm("machine", "ls", "--json", check=False)
+    if result.returncode != 0:
+        return False
+    try:
+        machines = json.loads(result.stdout or "[]")
+    except ValueError:
+        return False
+    return any(
+        isinstance(m, dict) and m.get("name") == name and m.get("state") == "running"
+        for m in machines
+    )
+
+
 def machine_start(name: str) -> None:
     """`smolvm machine start --name NAME`."""
     _smolvm("machine", "start", "--name", name)

bot_bottle/backend/terminal.py

@@ -12,22 +12,11 @@ import shlex
 # uses true/24-bit colors for its own chrome, which would otherwise bypass
 # the palette entirely.
 _COLORS: dict[str, tuple[int, str, int, str, str]] = {
-    "black":          (0,  "#2d2d2d", 8,  "#5c5c5c", "#0a0a0a"),
-    "red":            (1,  "#c0392b", 9,  "#e74c3c", "#1a0707"),
-    "green":          (2,  "#27ae60", 10, "#2ecc71", "#071a09"),
-    "yellow":         (3,  "#d4ac0d", 11, "#f1c40f", "#1a1507"),
-    "blue":           (4,  "#2471a3", 12, "#3498db", "#07071a"),
-    "magenta":        (5,  "#7d3c98", 13, "#9b59b6", "#12071a"),
-    "cyan":           (6,  "#148f77", 14, "#1abc9c", "#071a1a"),
-    "white":          (7,  "#bdc3c7", 15, "#ecf0f1", "#111111"),
-    "bright-black":   (8,  "#5c5c5c", 0,  "#2d2d2d", "#111111"),
-    "bright-red":     (9,  "#e74c3c", 1,  "#c0392b", "#200808"),
-    "bright-green":   (10, "#2ecc71", 2,  "#27ae60", "#082008"),
-    "bright-yellow":  (11, "#f1c40f", 3,  "#d4ac0d", "#201808"),
-    "bright-blue":    (12, "#3498db", 4,  "#2471a3", "#080820"),
-    "bright-magenta": (13, "#9b59b6", 5,  "#7d3c98", "#160820"),
-    "bright-cyan":    (14, "#1abc9c", 6,  "#148f77", "#082020"),
-    "bright-white":   (15, "#ecf0f1", 7,  "#bdc3c7", "#151515"),
+    "red":     (9,  "#e74c3c", 1,  "#c0392b", "#200808"),
+    "green":   (10, "#2ecc71", 2,  "#27ae60", "#082008"),
+    "yellow":  (11, "#f1c40f", 3,  "#d4ac0d", "#201808"),
+    "blue":    (12, "#3498db", 4,  "#2471a3", "#080820"),
+    "magenta": (13, "#9b59b6", 5,  "#7d3c98", "#160820"),
 }
 
 # OSC 104 resets all indexed palette entries; OSC 111 resets default background.

bot_bottle/bottle_state.py

@@ -1,8 +1,7 @@
-"""Per-bottle persistent state (PRD 0016).
+"""Per-bottle persistent state.
 
-Holds the per-bottle Dockerfile override that capability-block
-remediation writes, the transcript snapshot the state-preservation
-helper saves before teardown, and the launch metadata that lets
+Holds optional per-bottle Dockerfile overrides, the transcript snapshot
+the state-preservation helper saves before teardown, and the launch metadata that lets
 `cli.py resume <identity>` reconstruct a bottle's spec. State
 lives at:
 
@@ -43,6 +42,7 @@ from . import supervise as _supervise
 # Directory layout: ~/.bot-bottle/state/<identity>/...
 _STATE_SUBDIR = "state"
 _PER_BOTTLE_DOCKERFILE_NAME = "Dockerfile"
+_COMMITTED_IMAGE_NAME = "committed-image"
 _TRANSCRIPT_SUBDIR = "transcript"
 # Per-sidecar scratch subdirs. PRD 0018 chunk 2: bind-mount sources
 # live here so chunk 3's `docker compose up` can find them at stable
@@ -60,7 +60,7 @@ _METADATA_NAME = "metadata.json"
 _LIVE_CONFIG_SUBDIR = "live-config"
 LIVE_CONFIG_ROUTES_NAME = "routes.yaml"
 LIVE_CONFIG_ALLOWLIST_NAME = "allowlist"
-# Empty marker file. capability_apply writes it before teardown so
+# Empty marker file. Session preservation writes it before teardown so
 # cli.py's session-end cleanup knows to preserve the state dir for
 # `cli.py resume <identity>`. Absent = clean up.
 _PRESERVE_MARKER = ".preserve"
@@ -111,6 +111,10 @@ class BottleMetadata:
     backend: str = ""
     label: str = ""
     color: str = ""
+    # Ordered bottle names selected at launch (issue #269). Empty tuple
+    # for state dirs written before this change; resume falls back to
+    # the agent's `bottle:` field in that case.
+    bottle_names: tuple[str, ...] = ()
 
 
 def metadata_path(identity: str) -> Path:
@@ -138,6 +142,10 @@ def read_metadata(identity: str) -> BottleMetadata | None:
     if not isinstance(raw, dict):
         return None
     raw_typed = cast(dict[str, object], raw)
+    raw_bottle_names = raw_typed.get("bottle_names", [])
+    bottle_names: tuple[str, ...] = ()
+    if isinstance(raw_bottle_names, list):
+        bottle_names = tuple(str(n) for n in raw_bottle_names if isinstance(n, str))
     return BottleMetadata(
         identity=str(raw_typed.get("identity", identity)),
         agent_name=str(raw_typed.get("agent_name", "")),
@@ -148,6 +156,7 @@ def read_metadata(identity: str) -> BottleMetadata | None:
         backend=str(raw_typed.get("backend", "")),
         label=str(raw_typed.get("label", "")),
         color=str(raw_typed.get("color", "")),
+        bottle_names=bottle_names,
     )
 
 
@@ -163,8 +172,7 @@ def per_bottle_dockerfile_path(identity: str) -> Path:
 
 def per_bottle_dockerfile(identity: str) -> str | None:
     """Return the per-bottle Dockerfile content if present, else
-    None. None means: use the repo's Dockerfile (the original
-    pre-capability-block behavior)."""
+    None. None means: use the provider or manifest Dockerfile."""
     p = per_bottle_dockerfile_path(identity)
     if p.is_file():
         return p.read_text()
@@ -179,6 +187,32 @@ def write_per_bottle_dockerfile(identity: str, content: str) -> Path:
     return p
 
 
+def committed_image_path(identity: str) -> Path:
+    return bottle_state_dir(identity) / _COMMITTED_IMAGE_NAME
+
+
+def write_committed_image(identity: str, image_tag: str) -> Path:
+    """Persist the committed image tag for `identity`. The next
+    `cli.py resume <identity>` will boot from this image instead of
+    rebuilding from the Dockerfile."""
+    path = committed_image_path(identity)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(image_tag.strip() + "\n")
+    path.chmod(0o644)
+    return path
+
+
+def read_committed_image(identity: str) -> str | None:
+    """Return the committed image tag for `identity`, or None if no
+    commit has been recorded. Used by the Docker launch step to skip
+    the Dockerfile build when a committed snapshot exists."""
+    path = committed_image_path(identity)
+    if not path.is_file():
+        return None
+    tag = path.read_text().strip()
+    return tag or None
+
+
 def per_bottle_image_tag(identity: str) -> str:
     """Image tag for a rebuilt bottle. Distinct from the base
     bot-bottle-claude:latest so per-bottle rebuilds don't collide in
@@ -222,9 +256,7 @@ def write_live_config(
 
 
 def transcript_snapshot_dir(identity: str) -> Path:
-    """Where capability_apply stashes the agent's transcript before
-    teardown, so the next `cli.py start <agent>` can offer to
-    resume from it."""
+    """Where agent session snapshots are kept for resume flows."""
     return bottle_state_dir(identity) / _TRANSCRIPT_SUBDIR
 
 
@@ -251,8 +283,7 @@ def git_gate_state_dir(identity: str) -> Path:
 
 
 def supervise_state_dir(identity: str) -> Path:
-    """State subdir for the supervise sidecar's current-config dir
-    (bind-mounted into the agent at /etc/bot-bottle/current-config).
+    """State subdir reserved for supervise sidecar bind-mount sources.
     The queue dir is intentionally NOT under here — it lives at
     ~/.bot-bottle/queue/<slug>/ alongside the audit logs, so it
     survives state-dir cleanup."""
@@ -274,9 +305,8 @@ def preserve_marker_path(identity: str) -> Path:
 
 def mark_preserved(identity: str) -> Path:
     """Mark this bottle's state for preservation across session
-    teardown. Written by capability_apply.apply_capability_change so
-    cli.py's session-end cleanup leaves the state dir intact for a
-    subsequent `cli.py resume`."""
+    teardown so cli.py's session-end cleanup leaves the state dir
+    intact for a subsequent `cli.py resume`."""
     path = preserve_marker_path(identity)
     path.parent.mkdir(parents=True, exist_ok=True)
     path.touch()
@@ -289,7 +319,7 @@ def is_preserved(identity: str) -> bool:
 
 def clear_preserve_marker(identity: str) -> None:
     """Idempotent removal. Called at fresh launch (start or resume)
-    so a marker left from a prior capability-block doesn't keep
+    so a marker left from a prior preserved session doesn't keep
     state alive past the next normal session-end."""
     try:
         preserve_marker_path(identity).unlink()
@@ -314,6 +344,7 @@ __all__ = [
     "bottle_state_dir",
     "cleanup_state",
     "clear_preserve_marker",
+    "committed_image_path",
     "egress_state_dir",
     "git_gate_state_dir",
     "is_preserved",
@@ -323,9 +354,11 @@ __all__ = [
     "per_bottle_dockerfile_path",
     "per_bottle_image_tag",
     "preserve_marker_path",
+    "read_committed_image",
     "read_metadata",
     "supervise_state_dir",
     "transcript_snapshot_dir",
+    "write_committed_image",
     "write_metadata",
     "write_per_bottle_dockerfile",
 ]

bot_bottle/cli/__init__.py

@@ -1,6 +1,6 @@
 """Main CLI dispatcher.
 
-Commands: cleanup, edit, info, init, list, resume, start, supervise
+Commands: cleanup, commit, edit, info, init, list, resume, start, supervise
 """
 
 from __future__ import annotations
@@ -12,6 +12,7 @@ from ..manifest import ManifestError
 from ._common import PROG
 from . import list as _list_mod
 from .cleanup import cmd_cleanup
+from .commit import cmd_commit
 from .edit import cmd_edit
 from .info import cmd_info
 from .init import cmd_init
@@ -23,6 +24,7 @@ cmd_list = _list_mod.cmd_list
 
 COMMANDS = {
     "cleanup": cmd_cleanup,
+    "commit": cmd_commit,
     "edit": cmd_edit,
     "info": cmd_info,
     "init": cmd_init,
@@ -37,6 +39,7 @@ def usage() -> None:
     sys.stderr.write(f"usage: {PROG} <command> [args...]\n\n")
     sys.stderr.write("Commands:\n")
     sys.stderr.write("  cleanup   stop and remove all active bot-bottle containers\n")
+    sys.stderr.write("  commit    snapshot a running bottle's container state to a Docker image\n")
     sys.stderr.write("  edit      open an agent in vim for editing\n")
     sys.stderr.write("  info      print env, skills, and prompt details for a named agent\n")
     sys.stderr.write("  init      interactively create a new agent and add it to bot-bottle.json\n")

bot_bottle/cli/cleanup.py

@@ -13,9 +13,8 @@ dirs are shared layout, so docker is the single owner of that
 bucket.
 
 State dirs with `.preserve` are intentionally never touched — they
-hold capability-block rebuilds or crash snapshots the operator may
-want to `resume`. Manual `rm -rf ~/.bot-bottle/state/<identity>`
-is the path for those.
+hold preserved sessions the operator may want to `resume`. Manual
+`rm -rf ~/.bot-bottle/state/<identity>` is the path for those.
 """
 
 from __future__ import annotations

bot_bottle/cli/commit.py

@@ -0,0 +1,53 @@
+"""commit: freeze a running bottle's state to a resumable artifact.
+
+Docker bottles are committed to a local Docker image. Macos-container
+bottles are exported and rebuilt as a local Apple Container image.
+Smolmachines bottles are packed from the running VM into a
+`.smolmachine` artifact. The resulting reference is stored in
+per-bottle state so the next `./cli.py resume <slug>` boots from the
+snapshot instead of rebuilding from the Dockerfile.
+"""
+
+from __future__ import annotations
+
+import argparse
+
+from ..backend import enumerate_active_agents
+from ..backend.freeze import CommitCancelled, get_freezer
+from ..bottle_state import read_metadata
+from ..log import die
+from ._common import PROG
+from . import tui
+
+
+def cmd_commit(argv: list[str]) -> int:
+    parser = argparse.ArgumentParser(prog=f"{PROG} commit", add_help=True)
+    parser.add_argument(
+        "slug",
+        nargs="?",
+        default=None,
+        help=(
+            "bottle slug from `cli.py list active` "
+            "(omit to pick interactively)"
+        ),
+    )
+    args = parser.parse_args(argv)
+
+    slug = args.slug
+    if slug is None:
+        active = enumerate_active_agents()
+        if not active:
+            die("no active bottles; start one with `./cli.py start`")
+        choices = [a.slug for a in active]
+        slug = tui.filter_select(choices, title="Select bottle to commit")
+        if slug is None:
+            return 0
+
+    metadata = read_metadata(slug)
+    backend = metadata.backend if metadata else ""
+
+    try:
+        get_freezer(backend).commit_slug(slug)
+    except CommitCancelled:
+        return 0
+    return 0

bot_bottle/cli/info.py

@@ -5,7 +5,7 @@ from __future__ import annotations
 import argparse
 
 from ..log import info
-from ..manifest import Manifest
+from ..manifest import ManifestIndex
 from ._common import PROG, USER_CWD
 
 
@@ -14,11 +14,12 @@ def cmd_info(argv: list[str]) -> int:
     parser.add_argument("name", help="agent name defined in bot-bottle.json")
     args = parser.parse_args(argv)
 
-    manifest = Manifest.resolve(USER_CWD)
-    manifest.require_agent(args.name)
+    names = ManifestIndex.resolve(USER_CWD)
+    names.require_agent(args.name)
+    manifest = names.load_for_agent(args.name)
 
-    agent = manifest.agents[args.name]
-    bottle = manifest.bottle_for(args.name)
+    agent = manifest.agent
+    bottle = manifest.bottle
     env_names = list(bottle.env.keys())
     prompt_first_line = agent.prompt.splitlines()[0] if agent.prompt else ""
 
@@ -31,7 +32,7 @@ def cmd_info(argv: list[str]) -> int:
         f"first line: {prompt_first_line or '(empty)'}"
     )
     info(f"bottle             : {agent.bottle}")
-    identity = manifest.git_identity_summary(args.name)
+    identity = manifest.git_identity_summary()
     if identity:
         info(f"  git identity  : {identity}")
     if bottle.git:

bot_bottle/cli/list.py

@@ -7,26 +7,15 @@ import os
 import sys
 
 from ..backend import enumerate_active_agents
-from ..manifest import Manifest
+from ..manifest import ManifestIndex
 from ._common import PROG, USER_CWD
 
 _ANSI_COLOR_CODES: dict[str, str] = {
-    "black": "\033[30m",
-    "red": "\033[31m",
-    "green": "\033[32m",
-    "yellow": "\033[33m",
-    "blue": "\033[34m",
-    "magenta": "\033[35m",
-    "cyan": "\033[36m",
-    "white": "\033[37m",
-    "bright-black": "\033[90m",
-    "bright-red": "\033[91m",
-    "bright-green": "\033[92m",
-    "bright-yellow": "\033[93m",
-    "bright-blue": "\033[94m",
-    "bright-magenta": "\033[95m",
-    "bright-cyan": "\033[96m",
-    "bright-white": "\033[97m",
+    "red": "\033[91m",
+    "green": "\033[92m",
+    "yellow": "\033[93m",
+    "blue": "\033[94m",
+    "magenta": "\033[95m",
 }
 _ANSI_RESET = "\033[0m"
 
@@ -51,8 +40,8 @@ def cmd_list(argv: list[str]) -> int:
     args = parser.parse_args(argv)
 
     if args.scope == "available":
-        manifest = Manifest.resolve(USER_CWD)
-        for name in manifest.agents.keys():
+        manifest = ManifestIndex.resolve(USER_CWD)
+        for name in manifest.all_agent_names:
             print(name)
         return 0
 
@@ -66,7 +55,7 @@ def cmd_list(argv: list[str]) -> int:
     # Tab-separated keeps the format stable for shell pipelines.
     for b in active:
         services = ",".join(b.services) if b.services else "-"
-        display_name = b.label if b.label else b.agent_name
+        display_name = f"{b.label} ({b.agent_name})" if b.label else b.agent_name
         colored_name = _ansi_label(display_name, b.color)
         print(f"{b.backend_name}\t{b.slug}\t{colored_name}\t{services}")
     return 0

bot_bottle/cli/resume.py

@@ -4,13 +4,12 @@ Reads ~/.bot-bottle/state/<identity>/metadata.json to recover the
 (agent_name, cwd, copy_cwd) the bottle was originally started with,
 then runs the same launch core as `start` — but pinned to the
 recorded identity so the new bottle picks up any per-bottle Dockerfile
-(from capability-block apply) and transcript snapshot under the same
-state dir.
+override and transcript snapshot under the same state dir.
 
-Use case: an agent calls capability-block, the dashboard approves
-and tears down the bottle, the operator runs
+Use case: an interrupted or preserved bottle needs to be relaunched;
+the operator runs
     ./cli.py resume <identity>
-to bring up the replacement with the new capabilities baked in.
+to bring up the replacement from the recorded state.
 """
 
 from __future__ import annotations
@@ -20,7 +19,7 @@ import argparse
 from ..backend import BottleSpec
 from ..bottle_state import read_metadata
 from ..log import die
-from ..manifest import Manifest
+from ..manifest import ManifestIndex
 from ._common import PROG, USER_CWD
 from .start import _launch_bottle
 
@@ -28,7 +27,6 @@ from .start import _launch_bottle
 def cmd_resume(argv: list[str]) -> int:
     parser = argparse.ArgumentParser(prog=f"{PROG} resume", add_help=True)
     parser.add_argument("--dry-run", action="store_true")
-    parser.add_argument("--remote-control", action="store_true")
     parser.add_argument(
         "identity",
         help="bottle identity from a prior `start` (see its session-end output)",
@@ -42,7 +40,7 @@ def cmd_resume(argv: list[str]) -> int:
             f"check ~/.bot-bottle/state/ or run `cli.py start` to create a new bottle"
         )
 
-    manifest = Manifest.resolve(USER_CWD)
+    manifest = ManifestIndex.resolve(USER_CWD)
     manifest.require_agent(metadata.agent_name)
 
     spec = BottleSpec(
@@ -51,11 +49,11 @@ def cmd_resume(argv: list[str]) -> int:
         copy_cwd=metadata.copy_cwd,
         user_cwd=metadata.cwd or USER_CWD,
         identity=metadata.identity,
+        bottle_names=tuple(metadata.bottle_names),
     )
     backend_name = metadata.backend or None
     return _launch_bottle(
         spec,
         dry_run=args.dry_run,
-        remote_control=args.remote_control,
         backend_name=backend_name,
     )

bot_bottle/cli/start.py

@@ -20,18 +20,19 @@ from ..agent_provider import runtime_for
 from ..backend import (
     Bottle,
     BottleSpec,
+    enumerate_active_agents,
     get_bottle_backend,
     known_backend_names,
 )
+from ..backend.docker import util as docker_mod
 from ..backend.docker.bottle_plan import DockerBottlePlan
 from ..bottle_state import (
     cleanup_state,
     is_preserved,
     mark_preserved,
 )
-# from ..backend.docker.capability_apply import snapshot_transcript
 from ..log import info
-from ..manifest import Manifest
+from ..manifest import Manifest, ManifestIndex
 from ._common import PROG, USER_CWD, read_tty_line
 from . import tui
 
@@ -40,7 +41,6 @@ def cmd_start(argv: list[str]) -> int:
     parser = argparse.ArgumentParser(prog=f"{PROG} start", add_help=True)
     parser.add_argument("--dry-run", action="store_true")
     parser.add_argument("--cwd", action="store_true", help="copy host cwd into the running bottle")
-    parser.add_argument("--remote-control", action="store_true")
     parser.add_argument(
         "--backend",
         choices=known_backend_names(),
@@ -60,12 +60,12 @@ def cmd_start(argv: list[str]) -> int:
 
     dry_run = args.dry_run or os.environ.get("BOT_BOTTLE_DRY_RUN") == "1"
 
-    manifest = Manifest.resolve(USER_CWD)
+    manifest = ManifestIndex.resolve(USER_CWD)
 
     agent_name: str | None = args.name
     if agent_name is None:
         agent_name = tui.filter_select(
-            sorted(manifest.agents.keys()),
+            manifest.all_agent_names,
             title="Select agent",
         )
         if agent_name is None:
@@ -73,7 +73,25 @@ def cmd_start(argv: list[str]) -> int:
 
     backend_name: str | None = args.backend
 
+    # Bottle multiselect: always show after agent selection so operators
+    # can compose bottles at launch time without editing agent manifests.
+    available_bottles = manifest.all_bottle_names
+    lineage_map = _bottle_lineage(manifest)
+    display_labels = [lineage_map.get(n, n) for n in available_bottles]
+    label_to_name = {lineage_map.get(n, n): n for n in available_bottles}
+    initial_bottle = _peek_agent_bottle(manifest, agent_name)
+    initial_labels = [lineage_map.get(initial_bottle, initial_bottle)] if initial_bottle else []
+    selected_labels = tui.filter_multiselect(
+        display_labels,
+        title="Select bottles",
+        initial=initial_labels,
+    )
+    if selected_labels is None:
+        return 0
+    bottle_names = tuple(label_to_name.get(lbl, lbl) for lbl in selected_labels)
+
     label, color = tui.name_color_modal(default_label=agent_name)
+    label, color = _resolve_unique_label(label, color)
 
     spec = BottleSpec(
         manifest=manifest,
@@ -82,11 +100,11 @@ def cmd_start(argv: list[str]) -> int:
         user_cwd=USER_CWD,
         label=label,
         color=color,
+        bottle_names=bottle_names,
     )
     return _launch_bottle(
         spec,
         dry_run=dry_run,
-        remote_control=args.remote_control,
         backend_name=backend_name,
     )
 
@@ -131,7 +149,7 @@ def prepare_with_preflight(
 
 
 def attach_agent(
-    bottle: Bottle, *, remote_control: bool = False, resume: bool = False,
+    bottle: Bottle, *, resume: bool = False,
     agent_provider_template: str = "claude",
     startup_args: tuple[str, ...] = (),
 ) -> int:
@@ -150,8 +168,6 @@ def attach_agent(
         "(Ctrl-D or 'exit' to leave; container will be removed)"
     )
     agent_args = list(runtime.bypass_args)
-    if remote_control:
-        agent_args.extend(runtime.remote_control_args)
     agent_args.extend(startup_args)
     if resume:
         agent_args.extend(runtime.resume_args)
@@ -191,6 +207,53 @@ def _identity_from_plan(plan: object) -> str:
     return getattr(plan, "slug", "")
 
 
+def _peek_agent_bottle(manifest: ManifestIndex, agent_name: str) -> str:
+    """Return the `bottle:` value from the named agent's frontmatter without
+    fully parsing the agent file, or "" when absent or unreadable.
+
+    Used to pre-populate the bottle multiselect with the agent's default
+    bottle so operators who haven't removed `bottle:` from their manifests
+    don't need to re-select it every time."""
+    if manifest.home_md is None:
+        # Eager mode (from_json_obj): agent is pre-parsed.
+        if agent_name in manifest.agents:
+            return manifest.agents[agent_name].bottle
+        return ""
+
+    from ..manifest_loader import scan_agent_names
+    from ..yaml_subset import YamlSubsetError, parse_frontmatter
+
+    home_agents = scan_agent_names(manifest.home_md / "agents")
+    cwd_agents: dict[str, Path] = {}
+    if manifest.cwd_md is not None:
+        cwd_agents = scan_agent_names(manifest.cwd_md / "agents")
+    merged = {**home_agents, **cwd_agents}
+    path = merged.get(agent_name)
+    if path is None:
+        return ""
+    try:
+        fm, _ = parse_frontmatter(path.read_text())
+        bottle = fm.get("bottle", "")
+        return str(bottle) if isinstance(bottle, str) else ""
+    except (OSError, YamlSubsetError):
+        return ""
+
+
+def _resolve_unique_label(label: str, color: str) -> tuple[str, str]:
+    """Re-prompt with a disclaimer until the label's slug is not already
+    in use among running bottles.  Passes through unchanged when no
+    collision is found on the first check."""
+    while True:
+        slug_candidate = docker_mod.slugify(label)
+        active_slugs = {a.slug for a in enumerate_active_agents()}
+        if slug_candidate not in active_slugs:
+            return label, color
+        label, color = tui.name_color_modal(
+            default_label=label,
+            disclaimer=f'"{label}" is already in use',
+        )
+
+
 def _text_prompt_yes() -> bool:
     """Default `prompt_yes` for CLI use: reads y/N from the
     controlling tty via stderr prompt + tty-line read."""
@@ -200,17 +263,118 @@ def _text_prompt_yes() -> bool:
     return reply in ("y", "Y", "yes", "YES")
 
 
-def _text_render_preflight(*, remote_control: bool):
+def _text_render_preflight():
     def _render(plan: DockerBottlePlan) -> None:
-        plan.print(remote_control=remote_control)
+        print(file=sys.stderr)
+        print(_manifest_to_yaml(plan.manifest), file=sys.stderr)
     return _render
 
 
+def _bottle_lineage(manifest: ManifestIndex) -> dict[str, str]:
+    """Return {bottle_name: lineage_label} for bottles that have an extends chain.
+
+    Bottles without a parent are omitted (the caller falls back to the bare name).
+    Labels show the chain root-first: e.g. 'dev -> bot-bottle-dev -> claude-dev'."""
+    if manifest.home_md is None:
+        return {}
+    bottles_dir = manifest.home_md / "bottles"
+    if not bottles_dir.is_dir():
+        return {}
+
+    from ..yaml_subset import YamlSubsetError, parse_frontmatter
+
+    extends_of: dict[str, str] = {}
+    for path in bottles_dir.glob("*.md"):
+        try:
+            fm, _ = parse_frontmatter(path.read_text())
+            parent = fm.get("extends", "")
+            if isinstance(parent, str) and parent:
+                extends_of[path.stem] = parent
+        except (OSError, YamlSubsetError):
+            pass
+
+    labels: dict[str, str] = {}
+    for name in extends_of:
+        chain = [name]
+        seen = {name}
+        cur = name
+        while cur in extends_of:
+            par = extends_of[cur]
+            if par in seen:
+                break
+            chain.append(par)
+            seen.add(par)
+            cur = par
+        labels[name] = " -> ".join(reversed(chain))
+
+    return labels
+
+
+def _manifest_to_yaml(manifest: Manifest) -> str:
+    """Serialize the resolved Manifest to a YAML string for preflight display."""
+    lines: list[str] = []
+
+    agent = manifest.agent
+    lines.append("agent:")
+    if agent.skills:
+        lines.append("  skills:")
+        for s in agent.skills:
+            lines.append(f"    - {s}")
+    if not agent.git_user.is_empty():
+        lines.append("  git-gate:")
+        lines.append("    user:")
+        if agent.git_user.name:
+            lines.append(f"      name: {agent.git_user.name}")
+        if agent.git_user.email:
+            lines.append(f"      email: {agent.git_user.email}")
+
+    bottle = manifest.bottle
+    lines.append("bottle:")
+
+    if bottle.agent_provider.template != "claude" or bottle.agent_provider.dockerfile:
+        lines.append("  agent_provider:")
+        lines.append(f"    template: {bottle.agent_provider.template}")
+        if bottle.agent_provider.dockerfile:
+            lines.append(f"    dockerfile: {bottle.agent_provider.dockerfile}")
+
+    if bottle.env:
+        lines.append("  env:")
+        for k, v in sorted(bottle.env.items()):
+            lines.append(f"    {k}: {v}")
+
+    has_git_gate = not bottle.git_user.is_empty() or bottle.git
+    if has_git_gate:
+        lines.append("  git-gate:")
+        if not bottle.git_user.is_empty():
+            lines.append("    user:")
+            if bottle.git_user.name:
+                lines.append(f"      name: {bottle.git_user.name}")
+            if bottle.git_user.email:
+                lines.append(f"      email: {bottle.git_user.email}")
+        if bottle.git:
+            lines.append("    repos:")
+            for entry in bottle.git:
+                lines.append(f"      {entry.Name}:")
+                lines.append(f"        url: {entry.Upstream}")
+
+    if bottle.egress.routes:
+        lines.append("  egress:")
+        lines.append("    routes:")
+        for r in bottle.egress.routes:
+            lines.append(f"      - host: {r.Host}")
+            if r.AuthScheme:
+                lines.append(f"        auth:")
+                lines.append(f"          scheme: {r.AuthScheme}")
+
+    lines.append(f"  supervise: {'true' if bottle.supervise else 'false'}")
+
+    return "\n".join(lines)
+
+
 def _launch_bottle(
     spec: BottleSpec,
     *,
     dry_run: bool,
-    remote_control: bool,
     backend_name: str | None = None,
 ) -> int:
     """Shared launch core for `start` and `resume`. Builds the plan,
@@ -222,7 +386,7 @@ def _launch_bottle(
         plan, identity = prepare_with_preflight(
             spec,
             stage_dir=stage_dir,
-            render_preflight=_text_render_preflight(remote_control=remote_control),
+            render_preflight=_text_render_preflight(),
             prompt_yes=_text_prompt_yes,
             dry_run=dry_run,
             backend_name=backend_name,
@@ -235,7 +399,6 @@ def _launch_bottle(
             agent_provider_template = getattr(plan, "agent_provider_template", "claude")
             exit_code = attach_agent(
                 bottle,
-                remote_control=remote_control,
                 agent_provider_template=agent_provider_template,
                 startup_args=plan.agent_provision.startup_args,
             )
@@ -245,12 +408,8 @@ def _launch_bottle(
             )
             # While the container is still alive: always snapshot the
             # transcript and — if the agent exited non-zero — mark
-            # the state for preservation. Capability-block already
-            # did both before triggering teardown from the dashboard;
-            # this picks up crashes / Ctrl-Cs / OOM kills the same
-            # way. snapshot_transcript is best-effort so the
-            # capability-block path's prior snapshot isn't clobbered
-            # when the container is already gone.
+            # the state for preservation. This picks up crashes /
+            # Ctrl-Cs / OOM kills before cleanup removes the state dir.
             if agent_provider_template == "claude":
                 capture_claude_session_state(identity, exit_code)
         return 0

bot_bottle/cli/supervise.py

@@ -2,8 +2,8 @@
 act on them (approve / modify / reject).
 
 Curses-based TUI; modify-then-approve shells out to $EDITOR. The
-approval handler wires to PRD 0016 (capability-block), which rebuilds
-the bottle Dockerfile. The egress-block tool was removed in issue #198.
+Egress proposals are queued for operator review as full routes.yaml
+updates.
 """
 
 from __future__ import annotations
@@ -20,17 +20,19 @@ from datetime import datetime, timezone
 from pathlib import Path
 
 from .. import supervise as _supervise
-# from ..bottle_state import read_metadata
-# from ..backend.docker.capability_apply import (
-#     CapabilityApplyError,
-#     apply_capability_change,
-# )
+from ..bottle_state import read_metadata
+from ..backend.docker.egress_apply import (
+    EgressApplyError,
+    applicator as _docker_applicator,
+)
+from ..backend.macos_container.egress_apply import (
+    applicator as _macos_applicator,
+)
+from ..backend.smolmachines.egress_apply import (
+    applicator as _smolmachines_applicator,
+)
 from ..log import Die, error, info
 
-
-class CapabilityApplyError(RuntimeError):
-    """Placeholder while capability_apply is disabled."""
-
 from ..supervise import (
     COMPONENT_FOR_TOOL,
     AuditEntry,
@@ -39,8 +41,10 @@ from ..supervise import (
     STATUS_APPROVED,
     STATUS_MODIFIED,
     STATUS_REJECTED,
-    TOOL_CAPABILITY_BLOCK,
-    archive_proposal,
+    TOOL_EGRESS_ALLOW,
+    TOOL_EGRESS_BLOCK,
+    TOOL_GITLEAKS_ALLOW,
+    TOOL_EGRESS_TOKEN_ALLOW,
     list_pending_proposals,
     render_diff,
     write_audit_entry,
@@ -51,6 +55,11 @@ from ._common import PROG
 
 _REFRESH_INTERVAL_MS = 1000
 
+# Proposal tools whose payload is a read-only report, not a file the operator
+# edits: modify is unavailable and approval requires a recorded reason for the
+# audit trail.
+_REPORT_ONLY_TOOLS: tuple[str, ...] = (TOOL_GITLEAKS_ALLOW, TOOL_EGRESS_TOKEN_ALLOW)
+
 
 @dataclass(frozen=True)
 class QueuedProposal:
@@ -63,7 +72,17 @@ class QueuedProposal:
 # Errors any remediation engine may raise. Caught by the TUI key
 # handlers and surfaced in the status line so a failed apply keeps
 # the proposal pending rather than crashing curses.
-ApplyError = (CapabilityApplyError,)
+ApplyError = (EgressApplyError,)
+
+
+def apply_routes_change(slug: str, content: str) -> tuple[str, str]:
+    meta = read_metadata(slug)
+    backend = meta.backend if meta is not None else ""
+    if backend == "macos-container":
+        return _macos_applicator.apply_routes_change(slug, content)
+    if backend == "smolmachines":
+        return _smolmachines_applicator.apply_routes_change(slug, content)
+    return _docker_applicator.apply_routes_change(slug, content)
 
 
 def discover_pending() -> list[QueuedProposal]:
@@ -113,8 +132,10 @@ def _detail_lines(
 
 
 def _suffix_for_tool(tool: str) -> str:
-    if tool == TOOL_CAPABILITY_BLOCK:
-        return ".dockerfile"
+    if tool in (TOOL_EGRESS_ALLOW, TOOL_EGRESS_BLOCK):
+        return ".yaml"
+    if tool in (TOOL_GITLEAKS_ALLOW, TOOL_EGRESS_TOKEN_ALLOW):
+        return ".txt"
     return ".txt"
 
 
@@ -129,19 +150,14 @@ def approve(
 ) -> None:
     """Apply the proposal, write the waiting response, and audit it."""
     status = STATUS_MODIFIED if final_file is not None else STATUS_APPROVED
+    file_to_apply = final_file if final_file is not None else qp.proposal.proposed_file
 
     diff_before, diff_after = "", ""
-    # if qp.proposal.tool == TOOL_CAPABILITY_BLOCK:
-    #     _meta = read_metadata(qp.proposal.bottle_slug)
-    #     if _meta is not None and not _meta.compose_project:
-    #         raise CapabilityApplyError(
-    #             "capability-block remediation is not supported for smolmachines "
-    #             "bottles. Reject this proposal or handle the capability change "
-    #             "manually, then restart the bottle."
-    #         )
-    #     diff_before, diff_after = apply_capability_change(
-    #         qp.proposal.bottle_slug, file_to_apply,
-    #     )
+    if qp.proposal.tool in (TOOL_EGRESS_ALLOW, TOOL_EGRESS_BLOCK):
+        diff_before, diff_after = apply_routes_change(
+            qp.proposal.bottle_slug,
+            file_to_apply,
+        )
 
     response = Response(
         proposal_id=qp.proposal.id,
@@ -154,9 +170,6 @@ def approve(
         qp, action=status, notes=notes,
         diff_before=diff_before, diff_after=diff_after,
     )
-    if qp.proposal.tool == TOOL_CAPABILITY_BLOCK:
-        archive_proposal(qp.queue_dir, qp.proposal.id)
-
 
 def reject(qp: QueuedProposal, *, reason: str) -> None:
     """Write a rejection response and an audit entry."""
@@ -170,6 +183,23 @@ def reject(qp: QueuedProposal, *, reason: str) -> None:
     _write_audit(qp, action=STATUS_REJECTED, notes=reason, diff_before="", diff_after="")
 
 
+def _approve_from_tui(
+    stdscr: "curses._CursesWindow",  # type: ignore
+    qp: QueuedProposal,
+    *,
+    final_file: str | None = None,
+    notes: str = "",
+) -> str:
+    """Approve from curses, prompting for any tool-specific audit note."""
+    if qp.proposal.tool in _REPORT_ONLY_TOOLS and final_file is None:
+        notes = _prompt(stdscr, "allow reason (false positive / legitimately needed): ")
+        if not notes:
+            return "approve aborted (empty reason)"
+    approve(qp, final_file=final_file, notes=notes)
+    verb = "modified+approved" if final_file is not None else "approved"
+    return _approval_status(qp, verb)
+
+
 def _write_audit(
     qp: QueuedProposal,
     *,
@@ -241,7 +271,10 @@ def cmd_supervise(argv: list[str]) -> int:
         return e.code if isinstance(e.code, int) else 1
     except Exception as e:  # noqa: W0718 — catch supervise crash for logging
         log_path = _write_crash_log(e)
-        error(f"supervise crashed: {type(e).__name__}: {e}")
+        error(
+            f"supervise crashed: {type(e).__name__}: {e}",
+            context={"error_type": type(e).__name__, "crash_log": str(log_path)},
+        )
         error(f"full traceback written to {log_path}")
         return 1
     return 0
@@ -286,7 +319,7 @@ def _list_once() -> int:
     return 0
 
 
-def _try_init_green() -> int:
+def _try_init_green() -> int:  # pragma: no cover
     """Initialise a green color pair and return its attr, or 0."""
     try:
         curses.start_color()
@@ -297,7 +330,7 @@ def _try_init_green() -> int:
         return 0
 
 
-def _main_loop(stdscr: "curses._CursesWindow") -> None:  # type: ignore
+def _main_loop(stdscr: "curses._CursesWindow") -> None:  # type: ignore  # pragma: no cover
     curses.curs_set(0)
     stdscr.timeout(_REFRESH_INTERVAL_MS)
     green_attr = _try_init_green()
@@ -353,18 +386,22 @@ def _main_loop(stdscr: "curses._CursesWindow") -> None:  # type: ignore
             _detail_view(stdscr, qp, green_attr=green_attr)
         elif key == ord("a"):
             try:
-                approve(qp)
-                status_line = _approval_status(qp, "approved")
+                status_line = _approve_from_tui(stdscr, qp)
             except ApplyError as e:
                 status_line = f"apply failed: {e}"
         elif key == ord("m"):
+            if qp.proposal.tool in _REPORT_ONLY_TOOLS:
+                status_line = f"modify unavailable for {qp.proposal.tool}"
+                continue
             edited = _modify(stdscr, qp)
             if edited is None:
                 status_line = "modify aborted (no change)"
             else:
                 try:
-                    approve(qp, final_file=edited, notes="operator modified before approving")
-                    status_line = _approval_status(qp, "modified+approved")
+                    status_line = _approve_from_tui(
+                        stdscr, qp, final_file=edited,
+                        notes="operator modified before approving",
+                    )
                 except ApplyError as e:
                     status_line = f"apply failed: {e}"
         elif key == ord("r"):
@@ -383,7 +420,7 @@ def _render(
     status_line: str,
     *,
     green_attr: int = 0,  # noqa: F841 — unused, but required by interface
-) -> None:
+) -> None:  # pragma: no cover
     stdscr.erase()
     h, w = stdscr.getmaxyx()
     header = f"bot-bottle supervise  ({len(pending)} pending)"
@@ -434,7 +471,7 @@ def _detail_view(
     qp: QueuedProposal,
     *,
     green_attr: int = 0,
-) -> None:
+) -> None:  # pragma: no cover
     """Render the full proposal. Scrollable. Press q to return."""
     lines = _detail_lines(qp, green_attr=green_attr)
     offset = 0
@@ -462,15 +499,20 @@ def _detail_view(
             offset = max(0, len(lines) - 1)
         elif key == ord("a"):
             try:
-                approve(qp)
+                _approve_from_tui(stdscr, qp)
             except ApplyError:
                 pass
             return
         elif key == ord("m"):
+            if qp.proposal.tool in _REPORT_ONLY_TOOLS:
+                return
             edited = _modify(stdscr, qp)
             if edited is not None:
                 try:
-                    approve(qp, final_file=edited, notes="operator modified before approving")
+                    _approve_from_tui(
+                        stdscr, qp, final_file=edited,
+                        notes="operator modified before approving",
+                    )
                 except ApplyError:
                     pass
             return
@@ -481,7 +523,7 @@ def _detail_view(
             return
 
 
-def _modify(stdscr: "curses._CursesWindow", qp: QueuedProposal) -> str | None:  # type: ignore
+def _modify(stdscr: "curses._CursesWindow", qp: QueuedProposal) -> str | None:  # type: ignore  # pragma: no cover
     """Suspend curses, open $EDITOR on the proposed file, return edited content."""
     suffix = _suffix_for_tool(qp.proposal.tool)
     curses.endwin()
@@ -492,7 +534,7 @@ def _modify(stdscr: "curses._CursesWindow", qp: QueuedProposal) -> str | None:
     return edited
 
 
-def _prompt(stdscr: "curses._CursesWindow", label: str) -> str:  # type: ignore
+def _prompt(stdscr: "curses._CursesWindow", label: str) -> str:  # type: ignore  # pragma: no cover
     """One-line input at the bottom of the screen."""
     curses.curs_set(1)
     h, _ = stdscr.getmaxyx()

bot_bottle/cli/tui.py

@@ -17,6 +17,43 @@ import sys
 from typing import Any, Optional
 
 
+def filter_multiselect(
+    items: list[str],
+    *,
+    title: str = "",
+    initial: Optional[list[str]] = None,
+    tty_path: str = "/dev/tty",
+) -> Optional[list[str]]:
+    """Render a multi-select picker over *items*.
+
+    Returns the ordered list of selected items, or ``None`` if the user
+    cancelled (Esc / ``q`` / Ctrl-C / Ctrl-D with no items).
+
+    Press Space to toggle the item under the cursor.
+    Press Enter to confirm the current selection.
+    Press Ctrl-D to confirm the current selection (returns even if empty).
+    Press Esc/q to cancel (returns None).
+
+    *initial* pre-populates the selection in insertion order. Items
+    added are appended; removed items leave the remaining order unchanged.
+    """
+    if not items:
+        return []
+
+    try:
+        tty_fd = open(tty_path, "r+b", buffering=0)
+    except OSError:
+        return None
+
+    try:
+        fd_dup = os.dup(tty_fd.fileno())
+        return _run_multiselect(
+            items, title=title, initial=list(initial or []), tty_fd=fd_dup
+        )
+    finally:
+        tty_fd.close()
+
+
 def filter_select(
     items: list[str],
     *,
@@ -221,25 +258,283 @@ def _addstr_safe(screen: Any, row: int, col: int, text: str, attr: int = curses.
         pass
 
 
+# ---------------------------------------------------------------------------
+# filter_multiselect internals
+# ---------------------------------------------------------------------------
+
+_KEY_SPACE = 32
+
+
+def _run_multiselect(
+    items: list[str], *, title: str, initial: list[str], tty_fd: int
+) -> Optional[list[str]]:
+    """Drive a curses multi-select session on *tty_fd*."""
+    os.environ.setdefault("TERM", "xterm-256color")
+
+    orig_stdin = sys.__stdin__
+    orig_stdout = sys.__stdout__
+
+    try:
+        import io
+        tty_text = io.TextIOWrapper(io.FileIO(tty_fd, mode='r+'), write_through=True)
+        sys.__stdin__ = tty_text   # type: ignore[assignment]
+        sys.__stdout__ = tty_text  # type: ignore[assignment]
+
+        screen = curses.initscr()
+        curses.noecho()
+        curses.cbreak()
+        screen.keypad(True)
+
+        try:
+            result = _multiselect_loop(screen, items, title=title, initial=initial)
+        finally:
+            screen.keypad(False)
+            curses.nocbreak()
+            curses.echo()
+            curses.endwin()
+    except Exception:  # noqa: W0718
+        return None
+    finally:
+        sys.__stdin__ = orig_stdin    # type: ignore[assignment]
+        sys.__stdout__ = orig_stdout  # type: ignore[assignment]
+
+    return result
+
+
+def _toggle_membership(items: list[str], item: str) -> None:
+    """Add `item` if absent, remove it if present (in place)."""
+    if item in items:
+        items.remove(item)
+    else:
+        items.append(item)
+
+
+def _handle_order_key(key: int, selected: list[str], order_cursor: int) -> int:
+    """Apply a keypress in 'order' focus: navigate, reorder, or remove the
+    item at `order_cursor`. Mutates `selected` in place and returns the new
+    order cursor."""
+    if key in (curses.KEY_UP, ord("k")):
+        if order_cursor > 0:
+            order_cursor -= 1
+    elif key in (curses.KEY_DOWN, ord("j")):
+        if order_cursor < len(selected) - 1:
+            order_cursor += 1
+    elif key == ord("K"):
+        # Move selected item up (earlier in order).
+        if order_cursor > 0:
+            i = order_cursor
+            selected[i - 1], selected[i] = selected[i], selected[i - 1]
+            order_cursor -= 1
+    elif key == ord("J"):
+        # Move selected item down (later in order).
+        if order_cursor < len(selected) - 1:
+            i = order_cursor
+            selected[i], selected[i + 1] = selected[i + 1], selected[i]
+            order_cursor += 1
+    elif key in (curses.KEY_ENTER, _KEY_ENTER_ALT, ord("\r"), _KEY_SPACE):
+        # Remove item from selection while in order mode.
+        del selected[order_cursor]
+        if order_cursor >= len(selected) and order_cursor > 0:
+            order_cursor -= 1
+    return order_cursor
+
+
+def _multiselect_loop(
+    screen: Any, items: list[str], *, title: str, initial: list[str]
+) -> Optional[list[str]]:
+    query = ""
+    cursor = 0
+    selected: list[str] = [s for s in initial if s in items]
+    # focus = "filter": navigate + toggle items in the filterable list
+    # focus = "order":  navigate + reorder items in the selected list
+    focus = "filter"
+    order_cursor = 0
+
+    while True:
+        filtered = _filter_items(items, query)
+
+        if not filtered:
+            cursor = 0
+        elif cursor >= len(filtered):
+            cursor = len(filtered) - 1
+
+        if not selected:
+            order_cursor = 0
+            if focus == "order":
+                focus = "filter"
+        elif order_cursor >= len(selected):
+            order_cursor = len(selected) - 1
+
+        try:
+            _render_multiselect(
+                screen, filtered, cursor,
+                query=query, title=title, selected=selected,
+                focus=focus, order_cursor=order_cursor,
+            )
+        except curses.error:
+            return None
+
+        try:
+            key = screen.getch()
+        except KeyboardInterrupt:
+            return None
+
+        if key in (_KEY_ESC, _KEY_CTRL_C, ord("q")):
+            return None
+
+        if key == _KEY_CTRL_D:
+            return list(selected)
+
+        # Tab toggles between filter and order focus.
+        if key == ord("\t"):
+            if focus == "filter" and selected:
+                focus = "order"
+                order_cursor = 0
+            else:
+                focus = "filter"
+            continue
+
+        if focus == "filter":
+            if key in (curses.KEY_ENTER, _KEY_ENTER_ALT, ord("\r")):
+                return list(selected)
+
+            elif key == _KEY_SPACE:
+                if filtered:
+                    _toggle_membership(selected, filtered[cursor])
+
+            elif key in (curses.KEY_UP, ord("k")):
+                if cursor > 0:
+                    cursor -= 1
+
+            elif key in (curses.KEY_DOWN, ord("j")):
+                if cursor < len(filtered) - 1:
+                    cursor += 1
+
+            elif key in (curses.KEY_BACKSPACE, _KEY_BACKSPACE_WIN, 127):
+                query = query[:-1]
+                new_filtered = _filter_items(items, query)
+                if cursor >= len(new_filtered):
+                    cursor = max(0, len(new_filtered) - 1)
+
+            elif 32 <= key <= 126 and key != _KEY_SPACE:
+                query += chr(key)
+                cursor = 0
+
+        else:  # focus == "order"
+            order_cursor = _handle_order_key(key, selected, order_cursor)
+
+
+def _render_multiselect(
+    screen: Any,
+    filtered: list[str],
+    cursor: int,
+    *,
+    query: str,
+    title: str,
+    selected: list[str],
+    focus: str = "filter",
+    order_cursor: int = 0,
+) -> None:
+    screen.erase()
+    rows, cols = screen.getmaxyx()
+    min_rows = 7
+
+    if rows < min_rows:
+        raise curses.error("terminal too small")
+
+    sep = "─" * min(cols - 1, 40)
+    row = 0
+
+    if title and row < rows - 1:
+        _addstr_safe(screen, row, 0, title[:cols - 1], curses.A_BOLD)
+        row += 1
+
+    # Filter line — dim when focus is on the order panel.
+    filter_label = f"Filter: {query}"
+    filter_hint = "  [Tab: reorder]" if focus == "filter" and selected else ""
+    filter_attr = curses.A_DIM if focus == "order" else curses.A_NORMAL
+    if row < rows - 1:
+        _addstr_safe(screen, row, 0, (filter_label + filter_hint)[:cols - 1], filter_attr)
+        row += 1
+
+    if row < rows - 1:
+        _addstr_safe(screen, row, 0, sep)
+        row += 1
+
+    # Compute how many rows the bottom order panel needs.
+    # Cap the visible selected list to keep the filter list legible.
+    order_rows = min(len(selected), max(1, (rows - row) // 3)) if selected else 0
+    # Bottom reserved: sep + order_rows + sep + help = order_rows + 3
+    bottom_reserved = order_rows + 3
+
+    list_start = row
+    list_rows = rows - list_start - bottom_reserved
+    if list_rows < 1:
+        list_rows = 1
+
+    selected_set = set(selected)
+    filter_dim = focus == "order"
+    scroll = max(0, cursor - list_rows + 1)
+    visible = filtered[scroll: scroll + list_rows]
+
+    for idx, item in enumerate(visible):
+        abs_idx = scroll + idx
+        mark = "[*]" if item in selected_set else "[ ]"
+        prefix = "> " if (abs_idx == cursor and focus == "filter") else "  "
+        line = (prefix + mark + " " + item)[:cols - 1]
+        item_attr = curses.A_DIM if filter_dim else (
+            curses.A_REVERSE if abs_idx == cursor else curses.A_NORMAL
+        )
+        if row < rows - bottom_reserved:
+            _addstr_safe(screen, row, 0, line, item_attr)
+        row += 1
+
+    # Separator before the order panel.
+    if row < rows - (order_rows + 2):
+        _addstr_safe(screen, row, 0, sep)
+        row += 1
+
+    # Order panel.
+    order_scroll = max(0, order_cursor - order_rows + 1)
+    order_visible = selected[order_scroll: order_scroll + order_rows]
+    for idx, item in enumerate(order_visible):
+        abs_idx = order_scroll + idx
+        is_active = focus == "order" and abs_idx == order_cursor
+        prefix = "> " if is_active else "  "
+        line = (prefix + item)[:cols - 1]
+        attr = curses.A_REVERSE if is_active else curses.A_NORMAL
+        if row < rows - 2:
+            _addstr_safe(screen, row, 0, line, attr)
+        row += 1
+
+    if row < rows - 1:
+        _addstr_safe(screen, row, 0, sep)
+        row += 1
+
+    if focus == "filter":
+        help_line = "[↑↓/jk] move  [Space] toggle  [Enter] confirm  [Tab] reorder  [Esc/q] cancel"
+    else:
+        help_line = "[↑↓/jk] cursor  [K/J] reorder  [Space/Enter] remove  [Tab] back  [Ctrl-D] done"
+    if row < rows:
+        _addstr_safe(screen, min(rows - 1, row), 0, help_line[:cols - 1])
+
+    screen.refresh()
+
+
 # ---------------------------------------------------------------------------
 # name_color_modal — two-step label + color picker
 # ---------------------------------------------------------------------------
 
 _ANSI_COLORS = [
-    "red", "green", "blue", "yellow", "magenta", "cyan", "white", "black",
-    "bright-red", "bright-green", "bright-blue", "bright-yellow",
-    "bright-magenta", "bright-cyan", "bright-white", "bright-black",
+    "red", "green", "yellow", "blue", "magenta",
 ]
 
 _CURSES_COLOR_MAP: dict[str, int] = {
-    "black": curses.COLOR_BLACK,
     "red": curses.COLOR_RED,
     "green": curses.COLOR_GREEN,
     "yellow": curses.COLOR_YELLOW,
     "blue": curses.COLOR_BLUE,
     "magenta": curses.COLOR_MAGENTA,
-    "cyan": curses.COLOR_CYAN,
-    "white": curses.COLOR_WHITE,
 }
 
 _COLOR_NONE = "(none)"
@@ -248,11 +543,15 @@ _COLOR_NONE = "(none)"
 def name_color_modal(
     default_label: str,
     *,
+    disclaimer: str = "",
     tty_path: str = "/dev/tty",
 ) -> tuple[str, str]:
     """Present a two-step curses modal: first edit the agent label,
     then optionally pick a color.
 
+    ``disclaimer`` is shown below the input field — use it to surface
+    an error from a previous attempt (e.g. name already in use).
+
     Returns ``(label, color)`` where ``color`` is one of the 16 ANSI
     color name strings or ``""`` for no color. Falls back to
     ``(default_label, "")`` on any error (terminal too small, not a tty).
@@ -264,14 +563,14 @@ def name_color_modal(
 
     try:
         fd_dup = os.dup(tty_fd.fileno())
-        return _run_name_color(default_label, tty_fd=fd_dup)
+        return _run_name_color(default_label, tty_fd=fd_dup, disclaimer=disclaimer)
     except Exception:  # noqa: BLE001  # pylint: disable=broad-exception-caught
         return default_label, ""
     finally:
         tty_fd.close()
 
 
-def _run_name_color(default_label: str, *, tty_fd: int) -> tuple[str, str]:
+def _run_name_color(default_label: str, *, tty_fd: int, disclaimer: str = "") -> tuple[str, str]:
     import io
     orig_stdin = sys.__stdin__
     orig_stdout = sys.__stdout__
@@ -286,7 +585,7 @@ def _run_name_color(default_label: str, *, tty_fd: int) -> tuple[str, str]:
         curses.cbreak()
         screen.keypad(True)
         try:
-            label = _label_step(screen, default_label)
+            label = _label_step(screen, default_label, disclaimer=disclaimer)
             color = _color_step(screen, label)
         finally:
             screen.keypad(False)
@@ -299,14 +598,14 @@ def _run_name_color(default_label: str, *, tty_fd: int) -> tuple[str, str]:
     return label, color
 
 
-def _label_step(screen: Any, default_label: str) -> str:
+def _label_step(screen: Any, default_label: str, *, disclaimer: str = "") -> str:
     """Step 1: edit the label. First printable key replaces the
     pre-fill; subsequent keys append. Enter confirms."""
     text = default_label
     replaced = False  # True once the user has typed their first char
 
     while True:
-        _render_label(screen, text)
+        _render_label(screen, text, disclaimer=disclaimer)
         try:
             key = screen.getch()
         except KeyboardInterrupt:
@@ -330,7 +629,7 @@ def _label_step(screen: Any, default_label: str) -> str:
                 text += chr(key)
 
 
-def _render_label(screen: Any, text: str) -> None:
+def _render_label(screen: Any, text: str, *, disclaimer: str = "") -> None:
     screen.erase()
     rows, cols = screen.getmaxyx()
     sep = "─" * min(cols - 1, 40)
@@ -338,8 +637,12 @@ def _render_label(screen: Any, text: str) -> None:
     _addstr_safe(screen, 1, 0, sep)
     _addstr_safe(screen, 2, 0, text[:cols - 1], curses.A_REVERSE)
     _addstr_safe(screen, 3, 0, sep)
-    if rows > 5:
-        _addstr_safe(screen, 5, 0, "[any key] edit  [Enter] confirm", curses.A_DIM)
+    row = 4
+    if disclaimer and rows > row + 1:
+        _addstr_safe(screen, row, 0, disclaimer[:cols - 1], curses.A_BOLD)
+        row += 1
+    if rows > row + 1:
+        _addstr_safe(screen, row, 0, "[any key] edit  [Enter] confirm", curses.A_DIM)
     screen.refresh()
 
 
@@ -379,13 +682,10 @@ def _init_color_pairs() -> dict[str, int]:
         curses.use_default_colors()
         pair_idx = 2  # pair 1 reserved for other uses
         for name in _ANSI_COLORS:
-            base = name.replace("bright-", "")
-            fg = _CURSES_COLOR_MAP.get(base, curses.COLOR_WHITE)
+            fg = _CURSES_COLOR_MAP.get(name, curses.COLOR_WHITE)
             try:
                 curses.init_pair(pair_idx, fg, -1)
-                attr = curses.color_pair(pair_idx)
-                if name.startswith("bright-"):
-                    attr |= curses.A_BOLD
+                attr = curses.color_pair(pair_idx) | curses.A_BOLD
                 attrs[name] = attr
                 pair_idx += 1
             except curses.error:

bot_bottle/contrib/claude/Dockerfile

@@ -21,7 +21,7 @@ FROM node:22-slim
 # to it) works against egress's bumped TLS without the agent needing
 # local DNS.
 RUN apt-get update \
-  && apt-get install -y --no-install-recommends git ca-certificates curl \
+  && apt-get install -y --no-install-recommends git ca-certificates curl ripgrep \
   && rm -rf /var/lib/apt/lists/*
 
 # App-specific deps. Python isn't required by claude-code itself

bot_bottle/contrib/claude/agent_provider.py

@@ -20,6 +20,7 @@ from ...agent_provider import (
     AgentProvisionDir,
     AgentProvisionFile,
     AgentProvisionPlan,
+    provider_startup_args,
 )
 from ...backend.docker import util as docker_mod
 from ...egress import EgressRoute
@@ -42,41 +43,19 @@ def _prompt_path(guest_home: str) -> str:
 
 
 _STATUS_LINE_COLORS = {
-    "black": "\033[30m",
-    "red": "\033[31m",
-    "green": "\033[32m",
-    "yellow": "\033[33m",
-    "blue": "\033[34m",
-    "magenta": "\033[35m",
-    "cyan": "\033[36m",
-    "white": "\033[37m",
-    "bright-black": "\033[90m",
-    "bright-red": "\033[91m",
-    "bright-green": "\033[92m",
-    "bright-yellow": "\033[93m",
-    "bright-blue": "\033[94m",
-    "bright-magenta": "\033[95m",
-    "bright-cyan": "\033[96m",
-    "bright-white": "\033[97m",
+    "red": "\033[91m",
+    "green": "\033[92m",
+    "yellow": "\033[93m",
+    "blue": "\033[94m",
+    "magenta": "\033[95m",
 }
 
 _CLAUDE_THEME_COLORS = {
-    "black": "black",
-    "red": "red",
-    "green": "green",
-    "yellow": "yellow",
-    "blue": "blue",
-    "magenta": "magenta",
-    "cyan": "cyan",
-    "white": "white",
-    "bright-black": "blackBright",
-    "bright-red": "redBright",
-    "bright-green": "greenBright",
-    "bright-yellow": "yellowBright",
-    "bright-blue": "blueBright",
-    "bright-magenta": "magentaBright",
-    "bright-cyan": "cyanBright",
-    "bright-white": "whiteBright",
+    "red": "redBright",
+    "green": "greenBright",
+    "yellow": "yellowBright",
+    "blue": "blueBright",
+    "magenta": "magentaBright",
 }
 
 
@@ -112,7 +91,6 @@ _RUNTIME = AgentProviderRuntime(
     prompt_mode="append_file",
     bypass_args=("--dangerously-skip-permissions",),
     resume_args=("--continue",),
-    remote_control_args=("--remote-control",),
 )
 
 
@@ -137,8 +115,9 @@ class ClaudeAgentProvider(AgentProvider):
         color: str = "",
         provider_settings: dict[str, object] | None = None,
     ) -> AgentProvisionPlan:
-        del forward_host_credentials, host_env, provider_settings
+        del forward_host_credentials, host_env
         resolved_guest_env = dict(guest_env or {})
+        startup_args = provider_startup_args(provider_settings)
         guest_home = self.guest_home
         trusted_path = trusted_project_path or guest_home
 
@@ -221,6 +200,7 @@ class ClaudeAgentProvider(AgentProvider):
             env_vars=env_vars,
             guest_env=resolved_guest_env,
             has_prompt=has_prompt,
+            startup_args=startup_args,
             dirs=dirs,
             files=tuple(files),
             egress_routes=egress_routes,
@@ -233,7 +213,7 @@ class ClaudeAgentProvider(AgentProvider):
         when the agent has no skills."""
         from ...backend.util import host_skill_dir
 
-        agent = plan.spec.manifest.agents[plan.spec.agent_name]
+        agent = plan.manifest.agent
         if not agent.skills:
             return
         skills_dir = _skills_dir(plan.guest_home)
@@ -262,7 +242,7 @@ class ClaudeAgentProvider(AgentProvider):
             f"chown node:node {prompt_path} && chmod 600 {prompt_path}",
             user="root",
         )
-        agent = plan.spec.manifest.agents[plan.spec.agent_name]
+        agent = plan.manifest.agent
         return prompt_path if plan.agent_provision.has_prompt or agent.prompt else None
 
     def provision(self, plan: "BottlePlan", bottle: "Bottle") -> None:

bot_bottle/contrib/codex/Dockerfile

@@ -1,12 +1,12 @@
 # bot-bottle Codex provider image.
 #
 # Mirrors the default Claude image shape: Node LTS, git/network tooling,
-# non-root node user, and the provider CLI installed globally.
+# non-root node user, and the provider CLI installed for that user.
 
 FROM node:22-slim
 
 RUN apt-get update \
-  && apt-get install -y --no-install-recommends git ca-certificates curl \
+  && apt-get install -y --no-install-recommends git ca-certificates curl procps ripgrep \
   && rm -rf /var/lib/apt/lists/*
 
 # App-specific deps. Python isn't required by codex itself
@@ -17,12 +17,15 @@ RUN apt-get update \
   && apt-get install -y --no-install-recommends python3 python3-pip python3-venv \
   && rm -rf /var/lib/apt/lists/*
 
-RUN npm install -g --no-fund --no-audit @openai/codex@0.136.0 \
-  && npm cache clean --force
-
 USER node
 WORKDIR /home/node
 
-RUN mkdir -p /home/node/.codex
+ENV PATH="/home/node/.local/bin:${PATH}"
+
+# Remote-control support requires the standalone Codex install layout
+# under ~/.codex/packages/standalone/current. The npm package can run
+# the TUI, but remote-control commands expect this installer-owned path.
+RUN mkdir -p /home/node/.codex \
+  && curl -fsSL https://chatgpt.com/codex/install.sh | sh
 
 CMD ["codex"]

bot_bottle/contrib/codex/agent_provider.py

@@ -22,6 +22,7 @@ from ...agent_provider import (
     AgentProvisionCommand,
     AgentProvisionFile,
     AgentProvisionPlan,
+    provider_startup_args,
 )
 from .codex_auth import codex_host_access_token, write_codex_dummy_auth_file
 from ...egress import CODEX_HOST_CREDENTIAL_TOKEN_REF, EgressRoute
@@ -54,7 +55,6 @@ _RUNTIME = AgentProviderRuntime(
     prompt_mode="read_prompt_file",
     bypass_args=("--dangerously-bypass-approvals-and-sandbox",),
     resume_args=("resume", "--last"),
-    remote_control_args=(),
 )
 
 
@@ -79,8 +79,9 @@ class CodexAgentProvider(AgentProvider):
         color: str = "",
         provider_settings: dict[str, object] | None = None,
     ) -> AgentProvisionPlan:
-        del auth_token, label, color, provider_settings
+        del auth_token, label, color
         resolved_guest_env = dict(guest_env or {})
+        startup_args = provider_startup_args(provider_settings)
         guest_home = self.guest_home
         trusted_path = trusted_project_path or guest_home
 
@@ -163,6 +164,7 @@ class CodexAgentProvider(AgentProvider):
             env_vars=env_vars,
             guest_env=resolved_guest_env,
             has_prompt=has_prompt,
+            startup_args=startup_args,
             dirs=tuple(dirs),
             files=tuple(files),
             pre_copy=tuple(pre_copy),
@@ -177,7 +179,7 @@ class CodexAgentProvider(AgentProvider):
         skills."""
         from ...backend.util import host_skill_dir
 
-        agent = plan.spec.manifest.agents[plan.spec.agent_name]
+        agent = plan.manifest.agent
         if not agent.skills:
             return
         skills_dir = _skills_dir(plan.guest_home)
@@ -206,7 +208,7 @@ class CodexAgentProvider(AgentProvider):
             f"chown node:node {prompt_path} && chmod 600 {prompt_path}",
             user="root",
         )
-        agent = plan.spec.manifest.agents[plan.spec.agent_name]
+        agent = plan.manifest.agent
         return prompt_path if plan.agent_provision.has_prompt or agent.prompt else None
 
     def provision(self, plan: "BottlePlan", bottle: "Bottle") -> None:
@@ -261,8 +263,8 @@ class CodexAgentProvider(AgentProvider):
             return
         info(f"registering supervise MCP server in agent codex config → {supervise_url}")
         r = bottle.exec(
-            f"codex mcp add --transport http "
-            f"{_SUPERVISE_MCP_NAME} {supervise_url}",
+            f"codex mcp add {_SUPERVISE_MCP_NAME} --url "
+            f"{shlex.quote(supervise_url)}",
             user="node",
         )
         if r.returncode != 0:
@@ -270,7 +272,7 @@ class CodexAgentProvider(AgentProvider):
                 f"`codex mcp add supervise` failed (exit {r.returncode}): "
                 f"{(r.stderr or r.stdout or '').strip()}. Inside the bottle, "
                 f"register manually with: "
-                f"codex mcp add --transport http supervise {supervise_url}"
+                f"codex mcp add supervise --url {shlex.quote(supervise_url)}"
             )
 
 

bot_bottle/contrib/gitea/deploy_key_provisioner.py

@@ -19,7 +19,12 @@ import urllib.error
 import urllib.request
 from pathlib import Path
 
-from ...deploy_key_provisioner import DeployKeyProvisioner
+from ...deploy_key_provisioner import DeployKeyCollisionError, DeployKeyProvisioner
+
+# Timeout for ssh-keygen and Gitea API HTTP calls. A hung Gitea instance at
+# prepare time would stall bottle launch indefinitely without this bound.
+_API_TIMEOUT_SECS = 30
+_KEYGEN_TIMEOUT_SECS = 10
 
 
 class GiteaDeployKeyProvisioner(DeployKeyProvisioner):
@@ -46,6 +51,7 @@ class GiteaDeployKeyProvisioner(DeployKeyProvisioner):
                 check=True,
                 stdout=subprocess.DEVNULL,
                 stderr=subprocess.DEVNULL,
+                timeout=_KEYGEN_TIMEOUT_SECS,
             )
             private_key = key_path.read_bytes()
             public_key = key_path.with_suffix(".pub").read_text().strip()
@@ -67,10 +73,15 @@ class GiteaDeployKeyProvisioner(DeployKeyProvisioner):
             method="POST",
         )
         try:
-            with urllib.request.urlopen(req) as resp:
+            with urllib.request.urlopen(req, timeout=_API_TIMEOUT_SECS) as resp:
                 body = json.loads(resp.read())
         except urllib.error.HTTPError as exc:
             _body = _read_error_body(exc)
+            if exc.code == 422:
+                raise DeployKeyCollisionError(
+                    f"deploy key collision for {owner_repo!r} "
+                    f"(title={title!r}): key title or content already registered — {_body}"
+                ) from exc
             raise RuntimeError(
                 f"failed to create deploy key for {owner_repo}: "
                 f"HTTP {exc.code} — {_body}"
@@ -93,7 +104,7 @@ class GiteaDeployKeyProvisioner(DeployKeyProvisioner):
             method="DELETE",
         )
         try:
-            with urllib.request.urlopen(req):
+            with urllib.request.urlopen(req, timeout=_API_TIMEOUT_SECS):
                 pass
         except urllib.error.HTTPError as exc:
             if exc.code == 404:

bot_bottle/contrib/pi/agent_provider.py

@@ -21,6 +21,7 @@ from ...agent_provider import (
     AgentProvisionDir,
     AgentProvisionFile,
     AgentProvisionPlan,
+    provider_startup_args,
 )
 from ...egress import EgressRoute
 from ...log import die, info
@@ -165,7 +166,6 @@ _RUNTIME = AgentProviderRuntime(
     prompt_mode="append_system_prompt",
     bypass_args=(),
     resume_args=(),
-    remote_control_args=(),
 )
 
 
@@ -199,6 +199,7 @@ class PiAgentProvider(AgentProvider):
         models_payload, base_url, api_key_env, models, provider_name = (
             _pi_models_json(settings)
         )
+        extra_startup_args = provider_startup_args(provider_settings)
         models_file = state_dir / "pi-models.json"
         models_file.write_text(json.dumps(models_payload, indent=2) + "\n")
         models_file.chmod(0o600)
@@ -219,6 +220,7 @@ class PiAgentProvider(AgentProvider):
             startup_args=(
                 "--models",
                 ",".join(f"{provider_name}/{model}" for model in models),
+                *extra_startup_args,
             ),
             dirs=(AgentProvisionDir(f"{guest_home}/.pi/agent"),),
             files=(AgentProvisionFile(models_file, _models_path(guest_home)),),
@@ -232,7 +234,7 @@ class PiAgentProvider(AgentProvider):
     def provision_skills(self, plan: "BottlePlan", bottle: "Bottle") -> None:
         from ...backend.util import host_skill_dir
 
-        agent = plan.spec.manifest.agents[plan.spec.agent_name]
+        agent = plan.manifest.agent
         if not agent.skills:
             return
         skills_dir = _skills_dir(plan.guest_home)

bot_bottle/deploy_key_provisioner.py

@@ -11,6 +11,10 @@ from __future__ import annotations
 from abc import ABC, abstractmethod
 
 
+class DeployKeyCollisionError(RuntimeError):
+    """Raised when a deploy key title or public key already exists on the repo."""
+
+
 class DeployKeyProvisioner(ABC):
     """Manages a single deploy-key lifecycle on a remote forge."""
 

bot_bottle/dlp_detectors.py

@@ -15,6 +15,8 @@ import gzip
 import re
 import typing
 import unicodedata
+from math import log2
+from collections import Counter
 from urllib.parse import quote as url_quote
 
 try:
@@ -78,16 +80,27 @@ TOKEN_PATTERNS: tuple[tuple[str, re.Pattern[str]], ...] = (
 )
 
 
-def scan_token_patterns(text: str, *, location: str = "body") -> ScanResult | None:
+def scan_token_patterns(
+    text: str,
+    *,
+    location: str = "body",
+    safe_tokens: typing.AbstractSet[str] | None = None,
+) -> ScanResult | None:
     normalized = _normalize_text(text)
     for name, pattern in TOKEN_PATTERNS:
-        m = pattern.search(normalized)
-        if m is not None:
+        for m in pattern.finditer(normalized):
+            value = m.group(0)
+            # A value the supervisor has approved (PRD 0062) is no longer a
+            # block — keep scanning so a second, un-approved token in the
+            # same request is still caught.
+            if safe_tokens is not None and value in safe_tokens:
+                continue
             return ScanResult(
                 severity="block",
                 reason=f"{name} found in {location}",
                 location=location,
-                context=_snippet(text, m.start(), m.end()),
+                context=_snippet(normalized, m.start(), m.end()),
+                matched=value,
             )
     return None
 
@@ -96,20 +109,21 @@ def redact_tokens(
     text: str,
     *,
     env: typing.Mapping[str, str] | None = None,
+    sensitive_prefixes: tuple[str, ...] = ("EGRESS_TOKEN_",),
 ) -> str:
     """Replace token pattern matches and (if env given) provisioned secrets with REDACT."""
     for _, pattern in TOKEN_PATTERNS:
         text = pattern.sub(REDACT, text)
     if env is not None:
         for key, value in env.items():
-            if key.startswith("EGRESS_TOKEN_") and value:
+            if any(key.startswith(p) for p in sensitive_prefixes) and value:
                 for variant in _encoded_variants(value):
                     text = text.replace(variant, REDACT)
     return text
 
 
 # ---------------------------------------------------------------------------
-# Known secrets detector (Phase 1b)
+# Known secrets detector
 # ---------------------------------------------------------------------------
 
 def _encoded_variants(secret: str) -> list[str]:
@@ -150,26 +164,179 @@ def _encoded_variants(secret: str) -> list[str]:
     return variants
 
 
+# ---------------------------------------------------------------------------
+# Fragmentation-resistant helpers
+# ---------------------------------------------------------------------------
+
+# Minimum length of alnum projection for projection-based checks to run.
+# Short secrets produce too many false positives in projection space.
+_ALNUM_MIN_LEN = 8
+
+# Minimum window length for the partial-substring sliding scan.
+PARTIAL_MATCH_MIN_LEN = 12
+
+
+def _alnum_projection(text: str) -> str:
+    """Return text with every non-alphanumeric character stripped.
+
+    Used for fragmentation-resistant matching: separator-injected secrets
+    (spaces, hyphens, dots inserted between characters) are identical to
+    their originals in alnum projection space.
+    """
+    return "".join(c for c in text if c.isalnum())
+
+
+def _find_partial_window(secret_alnum: str, text_alnum: str, min_len: int) -> int | None:
+    """Return the position in text_alnum where any min_len-char window of
+    secret_alnum first appears, or None.
+
+    Slides a window of width min_len across secret_alnum and searches for
+    each window in text_alnum.  The first hit position is returned.
+    """
+    if len(secret_alnum) < min_len or len(text_alnum) < min_len:
+        return None
+    for i in range(len(secret_alnum) - min_len + 1):
+        window = secret_alnum[i:i + min_len]
+        pos = text_alnum.find(window)
+        if pos >= 0:
+            return pos
+    return None
+
+
 def scan_known_secrets(
     text: str,
     *,
     location: str = "body",
     env: typing.Mapping[str, str] | None = None,
+    sensitive_prefixes: tuple[str, ...] = ("EGRESS_TOKEN_",),
+    safe_tokens: typing.AbstractSet[str] | None = None,
 ) -> ScanResult | None:
     if env is None:
         return None
+
+    # Pre-compute alnum projection of the scan text once; reused per secret.
+    text_alnum: str | None = None
+
     for key, value in env.items():
-        if not key.startswith("EGRESS_TOKEN_") or not value:
+        if not any(key.startswith(p) for p in sensitive_prefixes) or not value:
             continue
+
+        # Pass 1: exact match across encoded variants (original behaviour).
+        approved_exact = False
         for variant in _encoded_variants(value):
             pos = text.find(variant)
             if pos >= 0:
+                # The supervisor approves the exact encoded variant found
+                # (PRD 0062); a different encoding of the same secret is a
+                # fresh block.
+                if safe_tokens is not None and variant in safe_tokens:
+                    approved_exact = True
+                    continue
                 return ScanResult(
                     severity="block",
                     reason=f"provisioned secret from {key} found in {location}",
                     location=location,
                     context=_snippet(text, pos, pos + len(variant)),
+                    matched=variant,
                 )
+        if approved_exact:
+            # Exact match was found and approved; projection passes would
+            # fire on the same value, so skip them for this secret.
+            continue
+
+        # Pass 2 & 3: fragmentation-resistant projection checks.
+        secret_alnum = _alnum_projection(value)
+        if len(secret_alnum) < _ALNUM_MIN_LEN:
+            continue
+
+        if text_alnum is None:
+            text_alnum = _alnum_projection(text)
+
+        # Pass 2: full alnum-projection exact match (catches separator injection).
+        pos2 = text_alnum.find(secret_alnum)
+        if pos2 >= 0:
+            return ScanResult(
+                severity="block",
+                reason=(
+                    f"provisioned secret from {key} found in {location} "
+                    f"(fragmented match — separator injection)"
+                ),
+                location=location,
+                context=_snippet(text_alnum, pos2, pos2 + len(secret_alnum)),
+            )
+
+        # Pass 3: sliding-window partial match (catches chunked-substring leaks).
+        pos3 = _find_partial_window(secret_alnum, text_alnum, PARTIAL_MATCH_MIN_LEN)
+        if pos3 is not None:
+            return ScanResult(
+                severity="block",
+                reason=(
+                    f"provisioned secret from {key} found in {location} "
+                    f"(partial match — at least {PARTIAL_MATCH_MIN_LEN} consecutive "
+                    f"alphanumeric chars)"
+                ),
+                location=location,
+                context=_snippet(text_alnum, pos3, pos3 + PARTIAL_MATCH_MIN_LEN),
+            )
+
+    return None
+
+
+# ---------------------------------------------------------------------------
+# Entropy detector (warn-only)
+# ---------------------------------------------------------------------------
+
+# Sliding window size and step for the entropy scan.
+ENTROPY_WINDOW = 64
+ENTROPY_STEP = 32
+
+# Bits-per-character threshold.  Random ASCII printable ≈ 6.6 bits; random
+# lowercase hex ≈ 4 bits; random base64url ≈ 6 bits.  5.5 sits above
+# typical structured data (JSON, URLs) while staying below truly random
+# content.
+ENTROPY_BLOCK_THRESHOLD = 5.5
+
+
+def _shannon_entropy(text: str) -> float:
+    if not text:
+        return 0.0
+    counts = Counter(text)
+    n = len(text)
+    return -sum((c / n) * log2(c / n) for c in counts.values())
+
+
+def scan_entropy(
+    text: str,
+    *,
+    location: str = "body",
+    window: int = ENTROPY_WINDOW,
+    threshold: float = ENTROPY_BLOCK_THRESHOLD,
+) -> ScanResult | None:
+    """Warn-only detector: flag windows of `window` chars with Shannon entropy
+    above `threshold` bits per character.
+
+    Never blocks; always returns severity='warn'.  Disabled by default —
+    routes must opt in via dlp.outbound_detectors=['entropy'].
+    """
+    if not text:
+        return None
+    step = max(1, window // 2)
+    end = len(text)
+    # Scan overlapping windows; also check the final tail if shorter than window.
+    positions = list(range(0, end - window + 1, step))
+    if end < window:
+        positions = [0]
+    elif (end - window) % step != 0:
+        positions.append(end - window)
+    for i in positions:
+        chunk = text[i:i + window]
+        if _shannon_entropy(chunk) >= threshold:
+            return ScanResult(
+                severity="warn",
+                reason=f"high-entropy content in {location} (possible encrypted exfil)",
+                location=location,
+                context=_snippet(text, i, i + len(chunk)),
+            )
     return None
 
 
@@ -265,6 +432,14 @@ _CRLF_ENCODED_RE = re.compile(r"%0[dD]%0[aA]", re.ASCII)
 _CRLF_HEADER_INJECT_RE = re.compile(r"\r\n[A-Za-z][A-Za-z0-9\-]+\s*:", re.ASCII)
 
 
+def strip_crlf(text: str) -> str:
+    """Remove URL-encoded and literal CRLF injection sequences from a request
+    surface (PRD 0062 redact policy). Used to scrub the request line / headers
+    so the request can be forwarded instead of hard-blocked."""
+    text = _CRLF_ENCODED_RE.sub("", text)
+    return _CRLF_HEADER_INJECT_RE.sub(lambda m: m.group(0)[2:], text)
+
+
 def scan_crlf_injection(text: str) -> ScanResult | None:
     if _CRLF_ENCODED_RE.search(text):
         return ScanResult(
@@ -280,12 +455,20 @@ def scan_crlf_injection(text: str) -> ScanResult | None:
 
 
 __all__ = [
+    "ENTROPY_BLOCK_THRESHOLD",
+    "ENTROPY_WINDOW",
+    "ENTROPY_STEP",
+    "PARTIAL_MATCH_MIN_LEN",
     "REDACT",
     "SNIPPET_CONTEXT",
     "TOKEN_PATTERNS",
+    "_alnum_projection",
+    "_shannon_entropy",
     "redact_tokens",
     "scan_crlf_injection",
+    "scan_entropy",
     "scan_known_secrets",
     "scan_naive_injection",
     "scan_token_patterns",
+    "strip_crlf",
 ]

bot_bottle/egress.py

@@ -10,12 +10,14 @@ specific and lives on concrete subclasses (see
 from __future__ import annotations
 
 import dataclasses
+import secrets
 from abc import ABC
 from dataclasses import dataclass
 from pathlib import Path
 from typing import TYPE_CHECKING
 
 from .egress_addon_core import (
+    ON_MATCH_REDACT,
     HeaderMatch as CoreHeaderMatch,
     MatchEntry as CoreMatchEntry,
     PathMatch as CorePathMatch,
@@ -31,6 +33,51 @@ CODEX_HOST_CREDENTIAL_TOKEN_REF = "BOT_BOTTLE_CODEX_HOST_ACCESS_TOKEN"
 EGRESS_HOSTNAME = "egress"
 
 EGRESS_ROUTES_IN_CONTAINER = "/etc/egress/routes.yaml"
+EGRESS_ROUTES_FILENAME = Path(EGRESS_ROUTES_IN_CONTAINER).name
+
+_CANARY_ENV_WORDS = (
+    "ACCORD",
+    "ANCHOR",
+    "ATLAS",
+    "CANON",
+    "CIPHER",
+    "EMBER",
+    "FALCON",
+    "HARBOR",
+    "LANTERN",
+    "MARBLE",
+    "NOVA",
+    "ORBIT",
+    "PIVOT",
+    "RADIUS",
+    "SUMMIT",
+    "VECTOR",
+)
+
+
+def _random_canary_env() -> str:
+    first = secrets.choice(_CANARY_ENV_WORDS)
+    remaining = tuple(word for word in _CANARY_ENV_WORDS if word != first)
+    second = secrets.choice(remaining)
+    return f"{first}_{second}_SECRET"
+
+
+def egress_sidecar_env_entries(plan: "EgressPlan") -> tuple[str, ...]:
+    """Return sidecar env entries needed by egress across all backends."""
+    env: list[str] = []
+    if plan.routes:
+        env.extend(sorted(plan.token_env_map.keys()))
+    if plan.canary and plan.canary_env:
+        env.append(f"{plan.canary_env}={plan.canary}")
+        env.append(f"BOT_BOTTLE_SENSITIVE_PREFIXES={plan.canary_env}")
+    return tuple(env)
+
+
+def egress_agent_env_entries(plan: "EgressPlan") -> tuple[str, ...]:
+    """Return agent-visible egress env entries shared by all backends."""
+    if plan.canary and plan.canary_env:
+        return (f"{plan.canary_env}={plan.canary}",)
+    return ()
 
 
 @dataclass(frozen=True)
@@ -63,6 +110,8 @@ class EgressPlan:
     mitmproxy_ca_host_path: Path = Path()
     mitmproxy_ca_cert_only_host_path: Path = Path()
     log: int = 0
+    canary: str = ""
+    canary_env: str = ""
 
 
 def egress_manifest_routes(
@@ -94,6 +143,7 @@ def egress_manifest_routes(
             git_fetch=r.GitFetch,
             outbound_detectors=r.OutboundDetectors,
             inbound_detectors=r.InboundDetectors,
+            outbound_on_match=r.OutboundOnMatch,
         ))
     return tuple(out)
 
@@ -104,12 +154,27 @@ def egress_routes_for_bottle(
 ) -> tuple[EgressRoute, ...]:
     manifest = egress_manifest_routes(bottle)
     provisioned_hosts = {pr.host.lower() for pr in provider_routes}
-    merged = list(provider_routes) + [
+    merged = list(_default_provider_on_match(provider_routes)) + [
         r for r in manifest if r.host.lower() not in provisioned_hosts
     ]
     return _assign_token_slots(merged)
 
 
+def _default_provider_on_match(
+    provider_routes: tuple[EgressRoute, ...],
+) -> tuple[EgressRoute, ...]:
+    """Provider routes (the agent talking to its own LLM API) default to the
+    `redact` on-match policy (PRD 0062): high-volume conversation payloads are
+    the worst source of token-shaped false positives, so a match is scrubbed
+    and forwarded rather than hard-blocked or queued for the operator. A
+    provider that sets `outbound_on_match` explicitly keeps its choice."""
+    return tuple(
+        r if r.outbound_on_match
+        else dataclasses.replace(r, outbound_on_match=ON_MATCH_REDACT)
+        for r in provider_routes
+    )
+
+
 def _assign_token_slots(
     routes: list[EgressRoute],
 ) -> tuple[EgressRoute, ...]:
@@ -145,6 +210,17 @@ def egress_token_env_map(
     return out
 
 
+def _yaml_str_escape(s: str) -> str:
+    """Escape a string for use inside a YAML double-quoted scalar."""
+    return (
+        s.replace("\\", "\\\\")
+         .replace('"', '\\"')
+         .replace("\n", "\\n")
+         .replace("\r", "\\r")
+         .replace("\t", "\\t")
+    )
+
+
 def _route_to_yaml_fields(r: Route) -> dict[str, object]:
     fields: dict[str, object] = {"host": r.host}
     if r.auth_scheme and r.token_env:
@@ -176,7 +252,11 @@ def _route_to_yaml_fields(r: Route) -> dict[str, object]:
         fields["matches"] = matches_data
     if r.git_fetch:
         fields["git"] = {"fetch": True}
-    if r.outbound_detectors is not None or r.inbound_detectors is not None:
+    if (
+        r.outbound_detectors is not None
+        or r.inbound_detectors is not None
+        or r.outbound_on_match
+    ):
         dlp: dict[str, object] = {}
         if r.outbound_detectors is not None:
             dlp["outbound_detectors"] = (
@@ -188,6 +268,8 @@ def _route_to_yaml_fields(r: Route) -> dict[str, object]:
                 False if not r.inbound_detectors
                 else list(r.inbound_detectors)
             )
+        if r.outbound_on_match:
+            dlp["outbound_on_match"] = r.outbound_on_match
         fields["dlp"] = dlp
     return fields
 
@@ -201,12 +283,12 @@ def _render_match_entry(entry: dict[str, object]) -> list[str]:
         for pd in entry["paths"]:  # type: ignore[union-attr]
             pd_dict: dict[str, str] = pd  # type: ignore[assignment]
             if "type" in pd_dict:
-                lines.append(f'          - type: "{pd_dict["type"]}"')
-                lines.append(f'            value: "{pd_dict["value"]}"')
+                lines.append(f'          - type: "{_yaml_str_escape(pd_dict["type"])}"')
+                lines.append(f'            value: "{_yaml_str_escape(pd_dict["value"])}"')
             else:
-                lines.append(f'          - value: "{pd_dict["value"]}"')
+                lines.append(f'          - value: "{_yaml_str_escape(pd_dict["value"])}"')
     if "methods" in entry:
-        methods_str = ", ".join(f'"{m}"' for m in entry["methods"])  # type: ignore[union-attr]
+        methods_str = ", ".join(f'"{_yaml_str_escape(m)}"' for m in entry["methods"])  # type: ignore[union-attr]
         prefix = "      - " if first_key else "        "
         lines.append(f'{prefix}methods: [{methods_str}]')
         first_key = False
@@ -216,8 +298,8 @@ def _render_match_entry(entry: dict[str, object]) -> list[str]:
         first_key = False
         for hd in entry["headers"]:  # type: ignore[union-attr]
             hd_dict: dict[str, str] = hd  # type: ignore[assignment]
-            lines.append(f'          - name: "{hd_dict["name"]}"')
-            lines.append(f'            value: "{hd_dict["value"]}"')
+            lines.append(f'          - name: "{_yaml_str_escape(hd_dict["name"])}"')
+            lines.append(f'            value: "{_yaml_str_escape(hd_dict["value"])}"')
     if first_key:
         lines.append("      - {}")
     return lines
@@ -237,10 +319,10 @@ def egress_render_routes(
         return "\n".join(lines) + "\n"
     for r in routes:
         f = _route_to_yaml_fields(r)
-        lines.append(f'  - host: "{f["host"]}"')
+        lines.append(f'  - host: "{_yaml_str_escape(str(f["host"]))}"')
         if "auth_scheme" in f:
-            lines.append(f'    auth_scheme: "{f["auth_scheme"]}"')
-            lines.append(f'    token_env: "{f["token_env"]}"')
+            lines.append(f'    auth_scheme: "{_yaml_str_escape(str(f["auth_scheme"]))}"')
+            lines.append(f'    token_env: "{_yaml_str_escape(str(f["token_env"]))}"')
         if "matches" in f:
             lines.append("    matches:")
             for entry in f["matches"]:  # type: ignore[union-attr]
@@ -259,6 +341,8 @@ def egress_render_routes(
                 elif isinstance(dv, list):
                     items_str = ", ".join(f'"{x}"' for x in dv)
                     lines.append(f"      {dk}: [{items_str}]")
+                elif isinstance(dv, str):
+                    lines.append(f'      {dk}: "{_yaml_str_escape(dv)}"')
     return "\n".join(lines) + "\n"
 
 
@@ -295,20 +379,27 @@ class Egress(ABC):
     ) -> EgressPlan:
         routes = egress_routes_for_bottle(bottle, provider_routes)
         log = bottle.egress.Log
-        routes_path = stage_dir / "egress_routes.yaml"
+        routes_path = stage_dir / EGRESS_ROUTES_FILENAME
         routes_path.write_text(egress_render_routes(routes, log=log))
         routes_path.chmod(0o600)
+        # Generate a per-session fake secret under a plausible random env name.
+        # The sidecar marks that exact env name as sensitive for known-secret
+        # scanning; the agent receives the same name/value as exfil bait.
+        canary = secrets.token_urlsafe(32)
         return EgressPlan(
             slug=slug,
             routes_path=routes_path,
             routes=routes,
             token_env_map=egress_token_env_map(routes),
             log=log,
+            canary=canary,
+            canary_env=_random_canary_env(),
         )
 
 __all__ = [
     "CODEX_HOST_CREDENTIAL_TOKEN_REF",
     "EGRESS_HOSTNAME",
+    "EGRESS_ROUTES_FILENAME",
     "EGRESS_ROUTES_IN_CONTAINER",
     "Egress",
     "EgressPlan",
@@ -317,5 +408,7 @@ __all__ = [
     "egress_render_routes",
     "egress_resolve_token_values",
     "egress_routes_for_bottle",
+    "egress_agent_env_entries",
+    "egress_sidecar_env_entries",
     "egress_token_env_map",
 ]

bot_bottle/egress_addon.py

@@ -5,7 +5,7 @@ egress container."""
 
 from __future__ import annotations
 
-import dataclasses
+import asyncio
 import json
 import os
 import signal
@@ -17,9 +17,15 @@ from mitmproxy import http  # type: ignore[import-not-found]  # pylint: disable=
 from egress_addon_core import (  # type: ignore[import-not-found]  # pylint: disable=import-error
     LOG_BLOCKS,
     LOG_FULL,
+    DEFAULT_OUTBOUND_ON_MATCH,
+    ON_MATCH_BLOCK,
+    ON_MATCH_REDACT,
     Config,
+    Route,
+    ScanResult,
     build_inbound_scan_text,
     build_outbound_scan_text,
+    build_token_allow_payload,
     decide,
     decide_git_fetch,
     is_git_fetch_request,
@@ -27,28 +33,61 @@ from egress_addon_core import (  # type: ignore[import-not-found]  # pylint: dis
     load_config,
     match_route,
     outbound_scan_headers,
+    route_to_yaml_dict,
     scan_inbound,
     scan_outbound,
 )
 
 try:
-    from dlp_detectors import redact_tokens  # type: ignore[import-not-found]
+    from dlp_detectors import redact_tokens, strip_crlf  # type: ignore[import-not-found]
 except ImportError:  # pragma: no cover - host-side path
-    from bot_bottle.dlp_detectors import redact_tokens  # type: ignore[import-not-found]
+    from bot_bottle.dlp_detectors import (  # type: ignore[import-not-found]
+        redact_tokens,
+        strip_crlf,
+    )
+
+try:
+    import supervise as _sv  # type: ignore[import-not-found]
+except ImportError:  # pragma: no cover - host-side path
+    from bot_bottle import supervise as _sv  # type: ignore[import-not-found]
 
 
 DEFAULT_ROUTES_PATH = "/etc/egress/routes.yaml"
 
 INTROSPECT_HOST = "_egress.local"
 
+# Seconds the egress proxy holds a token-blocked request open waiting for the
+# operator's supervisor decision (PRD 0062), overridable via env.
+DEFAULT_TOKEN_ALLOW_TIMEOUT_SECONDS = 300.0
+# Filesystem poll cadence while awaiting the operator's response.
+TOKEN_ALLOW_POLL_INTERVAL_SECONDS = 0.5
+
+# Fixed operator guidance attached to every token-allow proposal.
+_TOKEN_ALLOW_JUSTIFICATION = (
+    "egress DLP blocked an outbound request carrying a detected token. "
+    "Approve only if this value is a false positive or a credential this "
+    "request legitimately needs; the value is then allowed for the life of "
+    "this bottle's egress proxy."
+)
+
 
 class EgressAddon:
     def __init__(self) -> None:
         self.routes_path = os.environ.get("EGRESS_ROUTES", DEFAULT_ROUTES_PATH)
         self.config: Config = Config(routes=())
+        # Tokens the operator has approved this session (PRD 0062). In-memory
+        # only — a restart re-prompts. Mutated only from the asyncio loop that
+        # runs the addon hooks, so no lock is needed.
+        self.safe_tokens: set[str] = set()
+        self._supervise_queue_dir = os.environ.get("SUPERVISE_QUEUE_DIR", "").strip()
+        self._supervise_slug = os.environ.get("SUPERVISE_BOTTLE_SLUG", "").strip()
+        self._token_allow_timeout = _token_allow_timeout_from_env(os.environ)
         self._reload(initial=True)
         self._install_sighup()
 
+    def _supervise_available(self) -> bool:
+        return bool(self._supervise_queue_dir and self._supervise_slug)
+
     def _reload(self, *, initial: bool = False) -> None:
         try:
             text = Path(self.routes_path).read_text(encoding="utf-8")
@@ -82,7 +121,7 @@ class EgressAddon:
     def _serve_introspection(self, flow: http.HTTPFlow, path: str) -> None:
         if path == "/allowlist":
             payload = json.dumps(
-                {"routes": [dataclasses.asdict(r) for r in self.config.routes]},
+                {"routes": [route_to_yaml_dict(r) for r in self.config.routes]},
                 indent=2,
             ).encode("utf-8")
             flow.response = http.Response.make(
@@ -121,31 +160,42 @@ class EgressAddon:
         )
 
     def _log_request(self, flow: http.HTTPFlow) -> None:
+        headers = {
+            k: redact_tokens(v, env=os.environ)
+            for k, v in flow.request.headers.items()
+            if k.lower() != "authorization"
+        }
+        body = redact_tokens(flow.request.get_text(strict=False) or "", env=os.environ)
         sys.stderr.write(
             json.dumps({
                 "event": "egress_request",
                 "host": redact_tokens(flow.request.pretty_host, env=os.environ),
                 "method": flow.request.method,
                 "path": redact_tokens(flow.request.path, env=os.environ),
-                "headers": dict(flow.request.headers),
-                "body": flow.request.get_text(strict=False) or "",
+                "headers": headers,
+                "body": body,
             })
             + "\n"
         )
 
     def _log_response(self, flow: http.HTTPFlow) -> None:
+        headers = {
+            k: redact_tokens(v, env=os.environ)
+            for k, v in flow.response.headers.items()
+        }
+        body = redact_tokens(flow.response.get_text(strict=False) or "", env=os.environ)
         sys.stderr.write(
             json.dumps({
                 "event": "egress_response",
                 "host": flow.request.pretty_host,
                 "status": flow.response.status_code,
-                "headers": dict(flow.response.headers),
-                "body": flow.response.get_text(strict=False) or "",
+                "headers": headers,
+                "body": body,
             })
             + "\n"
         )
 
-    def request(self, flow: http.HTTPFlow) -> None:
+    async def request(self, flow: http.HTTPFlow) -> None:
         request_path, _, query = flow.request.path.partition("?")
 
         if flow.request.pretty_host == INTROSPECT_HOST:
@@ -157,21 +207,11 @@ class EgressAddon:
         # Hostname is included to catch DNS-tunnelling exfiltration attempts.
         route = match_route(self.config.routes, flow.request.pretty_host)
         if route is not None:
-            body = flow.request.get_text(strict=False) or ""
-            scan_text = build_outbound_scan_text(
-                flow.request.pretty_host,
-                request_path,
-                query,
-                outbound_scan_headers(route, dict(flow.request.headers)),
-                body,
-            )
-            dlp_result = scan_outbound(route, scan_text, os.environ)
-            if dlp_result is not None and dlp_result.severity == "block":
-                ctx = self._req_ctx(flow)
-                if dlp_result.context:
-                    ctx = {**ctx, "context": dlp_result.context}
-                self._block(flow, f"egress DLP: {dlp_result.reason}", ctx=ctx)
+            if not await self._handle_outbound_dlp(flow, route):
                 return
+            # The redact policy may have rewritten the request line; recompute
+            # 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(
@@ -221,6 +261,202 @@ class EgressAddon:
         if self.config.log >= LOG_FULL:
             self._log_request(flow)
 
+    def _block_dlp(self, flow: http.HTTPFlow, result: ScanResult) -> None:
+        ctx = self._req_ctx(flow)
+        if result.context:
+            ctx = {**ctx, "context": result.context}
+        self._block(flow, f"egress DLP: {result.reason}", ctx=ctx)
+
+    async def _handle_outbound_dlp(
+        self,
+        flow: http.HTTPFlow,
+        route: Route,
+    ) -> bool:
+        """Scan the outbound request and apply the route's on-match policy
+        (PRD 0062). Returns True if the request may be forwarded, False if a
+        403 response has been written to `flow`.
+
+        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, os.environ,
+                safe_tokens=self.safe_tokens, crlf_text=crlf_text,
+            )
+            if result is None or result.severity != "block":
+                return True
+
+            policy = route.outbound_on_match or DEFAULT_OUTBOUND_ON_MATCH
+
+            # 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):
+                    if self.config.log >= LOG_BLOCKS:
+                        sys.stderr.write(json.dumps({
+                            "event": "egress_redacted",
+                            "reason": f"egress DLP: {result.reason}",
+                            **self._req_ctx(flow),
+                        }) + "\n")
+                    return True
+                self._block(
+                    flow,
+                    f"egress DLP: {result.reason}; redaction could not remove "
+                    "all matches (e.g. a match in the hostname)",
+                    ctx=self._req_ctx(flow),
+                )
+                return False
+
+            # Structural blocks (CRLF, no safelist-able value) cannot be
+            # supervised — there is nothing to approve and remember — so under
+            # block/supervise they are a hard 403.
+            if policy == ON_MATCH_BLOCK or not result.matched:
+                self._block_dlp(flow, result)
+                return False
+
+            # supervise (default): hold the request for operator approval.
+            # Fall back to a hard 403 when supervise isn't wired for the bottle.
+            if not self._supervise_available():
+                self._block_dlp(flow, result)
+                return False
+            approved = await self._supervise_token_block(flow, request_path, result)
+            if not approved:
+                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) -> bool:
+        """Scrub detected tokens (and CRLF injection sequences) from the mutable
+        request surfaces (body, headers, path/query) and re-scan. 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=os.environ)
+            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=os.environ))
+            if redacted != value:
+                flow.request.headers[name] = redacted
+        redacted_path = strip_crlf(redact_tokens(flow.request.path, env=os.environ))
+        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, os.environ, crlf_text=crlf_text)
+        return result is None or result.severity != "block"
+
+    async def _supervise_token_block(
+        self,
+        flow: http.HTTPFlow,
+        request_path: str,
+        result: ScanResult,
+    ) -> bool:
+        """Route a token DLP block to the operator's supervisor queue and wait.
+
+        Returns True if the operator approved (the matched value is added to
+        `self.safe_tokens` and the caller re-scans); False if the request must
+        be blocked (a 403 response has been written to `flow`)."""
+        host = flow.request.pretty_host
+        payload = build_token_allow_payload(
+            redact_tokens(host, env=os.environ),
+            flow.request.method,
+            redact_tokens(request_path, env=os.environ),
+            result,
+        )
+        proposal = _sv.Proposal.new(
+            bottle_slug=self._supervise_slug,
+            tool=_sv.TOOL_EGRESS_TOKEN_ALLOW,
+            proposed_file=payload,
+            justification=_TOKEN_ALLOW_JUSTIFICATION,
+            current_file_hash=_sv.sha256_hex(payload),
+        )
+        queue_dir = Path(self._supervise_queue_dir)
+        try:
+            _sv.write_proposal(queue_dir, proposal)
+        except OSError as e:
+            sys.stderr.write(
+                f"egress: could not queue token-allow proposal: {e}; "
+                "blocking request\n"
+            )
+            self._block(flow, f"egress DLP: {result.reason}", ctx=self._req_ctx(flow))
+            return False
+
+        sys.stderr.write(json.dumps({
+            "event": "egress_token_supervise",
+            "reason": f"egress DLP: {result.reason}",
+            "proposal": proposal.id,
+            **self._req_ctx(flow),
+        }) + "\n")
+
+        response = await self._await_token_response(queue_dir, proposal.id)
+        _sv.archive_proposal(queue_dir, proposal.id)
+
+        if response is not None and response.status in (
+            _sv.STATUS_APPROVED, _sv.STATUS_MODIFIED,
+        ):
+            self.safe_tokens.add(result.matched)
+            if self.config.log >= LOG_BLOCKS:
+                sys.stderr.write(json.dumps({
+                    "event": "egress_token_allowed",
+                    "reason": f"egress DLP: {result.reason}",
+                    "proposal": proposal.id,
+                    **self._req_ctx(flow),
+                }) + "\n")
+            return True
+
+        if response is None:
+            reason = (
+                f"egress DLP: {result.reason}; supervisor approval timed out "
+                f"after {self._token_allow_timeout:g}s"
+            )
+        else:
+            reason = f"egress DLP: {result.reason}; supervisor rejected the request"
+        self._block(flow, reason, ctx=self._req_ctx(flow))
+        return False
+
+    async def _await_token_response(
+        self,
+        queue_dir: Path,
+        proposal_id: str,
+    ) -> "_sv.Response | None":
+        """Poll the queue dir for the operator's response without blocking the
+        proxy event loop. Returns the Response, or None on timeout."""
+        loop = asyncio.get_running_loop()
+        deadline = loop.time() + self._token_allow_timeout
+        while True:
+            try:
+                return _sv.read_response(queue_dir, proposal_id)
+            except (OSError, ValueError, KeyError):
+                # Not written yet, or a partial/malformed write — retry until
+                # the deadline, then fail closed.
+                pass
+            if loop.time() >= deadline:
+                return None
+            await asyncio.sleep(TOKEN_ALLOW_POLL_INTERVAL_SECONDS)
+
     def response(self, flow: http.HTTPFlow) -> None:
         """DLP inbound scan on response headers and body."""
         route = match_route(self.config.routes, flow.request.pretty_host)
@@ -272,7 +508,12 @@ class EgressAddon:
         message = flow.websocket.messages[-1]  # type: ignore[union-attr]
         content = message.content.decode("utf-8", errors="replace")
         if message.from_client:
-            result = scan_outbound(route, content, os.environ)
+            # A WebSocket data frame is not an HTTP request line, so CRLF is
+            # not an injection vector here — scan only for credential leakage.
+            result = scan_outbound(
+                route, content, os.environ,
+                safe_tokens=self.safe_tokens, crlf_text="",
+            )
             if result is not None and result.severity == "block":
                 sys.stderr.write(f"egress DLP: {result.reason}\n")
                 flow.kill()  # type: ignore[union-attr]
@@ -286,4 +527,23 @@ class EgressAddon:
                     sys.stderr.write(f"egress DLP warn: {result.reason}\n")
 
 
+def _token_allow_timeout_from_env(env: "os._Environ[str]") -> float:
+    """Read EGRESS_TOKEN_ALLOW_TIMEOUT_SECONDS; fall back to the default on an
+    unset or invalid value (a bad value should not wedge egress at boot)."""
+    raw = env.get("EGRESS_TOKEN_ALLOW_TIMEOUT_SECONDS", "").strip()
+    if not raw:
+        return DEFAULT_TOKEN_ALLOW_TIMEOUT_SECONDS
+    try:
+        value = float(raw)
+    except ValueError:
+        value = 0.0
+    if value <= 0:
+        sys.stderr.write(
+            "egress: invalid EGRESS_TOKEN_ALLOW_TIMEOUT_SECONDS="
+            f"{raw!r}; using default {DEFAULT_TOKEN_ALLOW_TIMEOUT_SECONDS:g}s\n"
+        )
+        return DEFAULT_TOKEN_ALLOW_TIMEOUT_SECONDS
+    return value
+
+
 addons = [EgressAddon()]

bot_bottle/egress_addon_core.py

@@ -21,6 +21,32 @@ try:
 except ImportError:  # pragma: no cover - host-side path
     from .yaml_subset import YamlSubsetError, parse_yaml_subset
 
+# DLP detector-config parsing lives in a sibling module (also flat-bundled
+# into the sidecar — see Dockerfile.sidecars). Re-exported below so existing
+# `from egress_addon_core import ON_MATCH_*` callers keep working.
+try:
+    from egress_dlp_config import (  # type: ignore[import-not-found]
+        DEFAULT_OUTBOUND_ON_MATCH,
+        INBOUND_DETECTOR_NAMES,
+        ON_MATCH_BLOCK,
+        ON_MATCH_REDACT,
+        ON_MATCH_SUPERVISE,
+        OUTBOUND_DETECTOR_NAMES,
+        OUTBOUND_ON_MATCH_VALUES,
+        parse_dlp_block,
+    )
+except ImportError:  # pragma: no cover - host-side path
+    from .egress_dlp_config import (
+        DEFAULT_OUTBOUND_ON_MATCH,
+        INBOUND_DETECTOR_NAMES,
+        ON_MATCH_BLOCK,
+        ON_MATCH_REDACT,
+        ON_MATCH_SUPERVISE,
+        OUTBOUND_DETECTOR_NAMES,
+        OUTBOUND_ON_MATCH_VALUES,
+        parse_dlp_block,
+    )
+
 
 # ---------------------------------------------------------------------------
 # Match types (Gateway API HTTPRoute vocabulary, PRD 0053)
@@ -34,9 +60,6 @@ VALID_METHODS = frozenset({
     "CONNECT",
 })
 
-OUTBOUND_DETECTOR_NAMES = frozenset({"token_patterns", "known_secrets"})
-INBOUND_DETECTOR_NAMES = frozenset({"naive_injection_detection"})
-
 
 @dataclass(frozen=True)
 class PathMatch:
@@ -69,6 +92,8 @@ class Route:
     git_fetch: bool = False
     outbound_detectors: tuple[str, ...] | None = None
     inbound_detectors: tuple[str, ...] | None = None
+    # "" means unset → DEFAULT_OUTBOUND_ON_MATCH. See OUTBOUND_ON_MATCH_VALUES.
+    outbound_on_match: str = ""
 
 
 LOG_OFF = 0    # no logging
@@ -95,6 +120,11 @@ class ScanResult:
     reason: str
     location: str = ""  # where the match was found, e.g. "body", "authorization header"
     context: str = ""   # surrounding text with the match replaced by REDACT
+    # Raw substring the detector matched. Used inside the sidecar to key the
+    # supervisor-approved "safe tokens" set (PRD 0062); never logged or written
+    # to a proposal file. Empty for structural detectors (CRLF) that carry no
+    # safelist-able value.
+    matched: str = ""
 
 
 # ---------------------------------------------------------------------------
@@ -214,61 +244,6 @@ def _parse_match_entry(idx: int, k: int, raw: object) -> MatchEntry:
     return MatchEntry(paths=paths, methods=methods, headers=headers)
 
 
-def _parse_detectors(
-    idx: int,
-    host: str,
-    raw_dict: dict[str, object],
-) -> tuple[tuple[str, ...] | None, tuple[str, ...] | None]:
-    """Parse the optional `dlp` block on a route, returning
-    (outbound_detectors, inbound_detectors)."""
-    dlp_raw = raw_dict.get("dlp")
-    if dlp_raw is None:
-        return None, None
-    label = f"route[{idx}] ({host})"
-    if not isinstance(dlp_raw, dict):
-        raise ValueError(f"{label}: 'dlp' must be an object")
-    dlp = typing.cast(dict[str, object], dlp_raw)
-
-    def _parse_detector_field(
-        field: str,
-        valid_names: frozenset[str],
-    ) -> tuple[str, ...] | None:
-        val = dlp.get(field)
-        if val is None:
-            return None
-        if val is False:
-            return ()
-        if not isinstance(val, list):
-            raise ValueError(
-                f"{label}: dlp.{field} must be false, a list, or omitted"
-            )
-        items = typing.cast(list[object], val)
-        names: list[str] = []
-        for j, item in enumerate(items):
-            if not isinstance(item, str):
-                raise ValueError(
-                    f"{label}: dlp.{field}[{j}] must be a string"
-                )
-            if item not in valid_names:
-                raise ValueError(
-                    f"{label}: dlp.{field}[{j}] {item!r} is not a valid "
-                    f"detector name; valid names: {', '.join(sorted(valid_names))}"
-                )
-            names.append(item)
-        return tuple(names)
-
-    outbound = _parse_detector_field("outbound_detectors", OUTBOUND_DETECTOR_NAMES)
-    inbound = _parse_detector_field("inbound_detectors", INBOUND_DETECTOR_NAMES)
-
-    for k in dlp:
-        if k not in ("outbound_detectors", "inbound_detectors"):
-            raise ValueError(
-                f"{label}: dlp has unknown key {k!r}; accepted keys "
-                f"are 'outbound_detectors', 'inbound_detectors'"
-            )
-    return outbound, inbound
-
-
 def parse_routes(payload: object) -> tuple[Route, ...]:
     if not isinstance(payload, dict):
         raise ValueError("routes payload: top-level must be an object")
@@ -337,7 +312,7 @@ def _parse_one(idx: int, raw: object) -> Route:
                 )
 
     # dlp detectors
-    outbound_detectors, inbound_detectors = _parse_detectors(
+    outbound_detectors, inbound_detectors, outbound_on_match = parse_dlp_block(
         idx, host, raw_dict,
     )
 
@@ -356,16 +331,60 @@ def _parse_one(idx: int, raw: object) -> Route:
         git_fetch=git_fetch,
         outbound_detectors=outbound_detectors,
         inbound_detectors=inbound_detectors,
+        outbound_on_match=outbound_on_match,
     )
 
 
-def load_routes(text: str) -> tuple[Route, ...]:
-    """Parse YAML text → routes."""
-    try:
-        payload = parse_yaml_subset(text)
-    except YamlSubsetError as e:
-        raise ValueError(f"routes payload: invalid YAML: {e}") from e
-    return parse_routes(payload)
+def _path_match_to_dict(pm: PathMatch) -> dict[str, object]:
+    d: dict[str, object] = {"value": pm.value}
+    if pm.type != "prefix":
+        d["type"] = pm.type
+    return d
+
+
+def _header_match_to_dict(hm: HeaderMatch) -> dict[str, object]:
+    d: dict[str, object] = {"name": hm.name, "value": hm.value}
+    if hm.type != "exact":
+        d["type"] = hm.type
+    return d
+
+
+def _match_entry_to_dict(me: MatchEntry) -> dict[str, object]:
+    d: dict[str, object] = {}
+    if me.paths:
+        d["paths"] = [_path_match_to_dict(p) for p in me.paths]
+    if me.methods:
+        d["methods"] = list(me.methods)
+    if me.headers:
+        d["headers"] = [_header_match_to_dict(h) for h in me.headers]
+    return d
+
+
+def route_to_yaml_dict(r: Route) -> dict[str, object]:
+    """Serialize a Route to YAML-schema-compatible dict.
+
+    Uses the same field names the YAML parser accepts, so the output
+    can be round-tripped directly into an `allow` or `egress-block`
+    proposal without translation. Fields that are empty/default are
+    omitted so the agent doesn't copy irrelevant keys."""
+    d: dict[str, object] = {"host": r.host}
+    if r.auth_scheme:
+        d["auth_scheme"] = r.auth_scheme
+        d["token_env"] = r.token_env
+    if r.matches:
+        d["matches"] = [_match_entry_to_dict(m) for m in r.matches]
+    if r.git_fetch:
+        d["git"] = {"fetch": True}
+    dlp: dict[str, object] = {}
+    if r.outbound_detectors is not None:
+        dlp["outbound_detectors"] = list(r.outbound_detectors)
+    if r.inbound_detectors is not None:
+        dlp["inbound_detectors"] = list(r.inbound_detectors)
+    if r.outbound_on_match:
+        dlp["outbound_on_match"] = r.outbound_on_match
+    if dlp:
+        d["dlp"] = dlp
+    return d
 
 
 def parse_config(payload: object) -> "Config":
@@ -640,43 +659,103 @@ def scan_outbound(
     route: Route,
     body: str | bytes,
     environ: typing.Mapping[str, str],
+    *,
+    safe_tokens: typing.AbstractSet[str] | None = None,
+    crlf_text: str | None = None,
 ) -> ScanResult | None:
     # Lazy import to avoid circular deps and keep dlp_detectors optional
     # at import time (the sidecar copies it flat alongside this file).
     try:
         from dlp_detectors import (  # type: ignore[import-not-found]
             scan_crlf_injection,
+            scan_entropy,
             scan_known_secrets,
             scan_token_patterns,
         )
     except ImportError:  # pragma: no cover - host-side path
         from .dlp_detectors import (  # type: ignore[import-not-found]
             scan_crlf_injection,
+            scan_entropy,
             scan_known_secrets,
             scan_token_patterns,
         )
 
-    text = body if isinstance(body, str) else body.decode("utf-8", errors="replace")
+    # Binary bodies: latin-1 is a bijective byte↔codepoint mapping that
+    # preserves every byte value, so ASCII-range secret strings remain
+    # findable by str.find / regex.  Prefer strict UTF-8 for valid text bodies.
+    if isinstance(body, bytes):
+        try:
+            text = body.decode("utf-8")
+        except UnicodeDecodeError:
+            text = body.decode("latin-1")
+    else:
+        text = body
 
-    # CRLF injection is never legitimate — runs unconditionally, not gated
-    # by outbound_detectors config.
-    result = scan_crlf_injection(text)
+    # CRLF injection is only an attack in the request line + headers, never the
+    # body: an HTTP body is delimited by Content-Length, so CRLF bytes there
+    # cannot split the request. Scanning the body produces false positives on
+    # legitimate form-encoded / multi-line content. Callers pass the
+    # body-excluded surfaces as `crlf_text`; `None` falls back to the full text
+    # for backward-compatible callers (host-side tests, websocket frames).
+    crlf_target = text if crlf_text is None else crlf_text
+    result = scan_crlf_injection(crlf_target)
     if result is not None:
         return result
 
     if _detector_enabled(route.outbound_detectors, "token_patterns"):
-        result = scan_token_patterns(text, location="body")
+        result = scan_token_patterns(text, location="body", safe_tokens=safe_tokens)
         if result is not None:
             return result
 
     if _detector_enabled(route.outbound_detectors, "known_secrets"):
-        result = scan_known_secrets(text, location="body", env=environ)
+        # BOT_BOTTLE_SENSITIVE_PREFIXES lets operators add extra env prefixes
+        # beyond EGRESS_TOKEN_* without changing the manifest schema.
+        extra_raw = environ.get("BOT_BOTTLE_SENSITIVE_PREFIXES", "")
+        extra = tuple(p for p in extra_raw.split(",") if p)
+        sensitive_prefixes = ("EGRESS_TOKEN_",) + extra
+        result = scan_known_secrets(
+            text, location="body", env=environ,
+            sensitive_prefixes=sensitive_prefixes, safe_tokens=safe_tokens,
+        )
+        if result is not None:
+            return result
+
+    # Entropy scanning requires explicit opt-in: it is NOT part of the
+    # default "all detectors" set because it produces false positives on
+    # legitimate base64 / binary payloads.  Routes must list "entropy" in
+    # dlp.outbound_detectors to enable it.
+    if (
+        route.outbound_detectors is not None
+        and "entropy" in route.outbound_detectors
+    ):
+        result = scan_entropy(text, location="body")
         if result is not None:
             return result
 
     return None
 
 
+def build_token_allow_payload(
+    host: str,
+    method: str,
+    path: str,
+    result: ScanResult,
+) -> str:
+    """Render the human-readable supervisor proposal body for an outbound
+    token block (PRD 0062). Carries the host/method/path, the detector
+    reason, and the redacted context snippet — never the raw token value."""
+    lines = [
+        "egress blocked an outbound request carrying a detected token",
+        f"host: {host}",
+        f"method: {method}",
+        f"path: {path}",
+        f"detector: {result.reason}",
+    ]
+    if result.context:
+        lines.append(f"context: {result.context}")
+    return "\n".join(lines) + "\n"
+
+
 def scan_inbound(
     route: Route,
     body: str | bytes,
@@ -698,8 +777,17 @@ def scan_inbound(
 
 __all__ = [
     "LOG_BLOCKS",
+    "route_to_yaml_dict",
     "LOG_FULL",
     "LOG_OFF",
+    "ON_MATCH_BLOCK",
+    "ON_MATCH_REDACT",
+    "ON_MATCH_SUPERVISE",
+    "OUTBOUND_ON_MATCH_VALUES",
+    "DEFAULT_OUTBOUND_ON_MATCH",
+    "OUTBOUND_DETECTOR_NAMES",
+    "INBOUND_DETECTOR_NAMES",
+    "parse_dlp_block",
     "Config",
     "Decision",
     "HeaderMatch",
@@ -709,13 +797,13 @@ __all__ = [
     "ScanResult",
     "build_inbound_scan_text",
     "build_outbound_scan_text",
+    "build_token_allow_payload",
     "decide",
     "decide_git_fetch",
     "evaluate_matches",
     "is_git_push_request",
     "is_git_fetch_request",
     "load_config",
-    "load_routes",
     "match_route",
     "outbound_scan_headers",
     "parse_config",

bot_bottle/egress_dlp_config.py

@@ -0,0 +1,92 @@
+"""DLP detector-config parsing for egress routes (PRD 0053, PRD 0062).
+
+A route's optional `dlp:` block names which outbound/inbound detectors run
+and what the proxy does when an outbound detector matches a token
+(`outbound_on_match`). This module owns parsing and validating that block,
+kept apart from the request-time scan/decision flow in `egress_addon_core`
+so each half reads top-to-bottom without scrolling past the other.
+
+Stdlib-only; ships flat into the sidecar bundle image alongside
+`egress_addon_core.py` — see `Dockerfile.sidecars`."""
+
+from __future__ import annotations
+
+import typing
+
+OUTBOUND_DETECTOR_NAMES = frozenset({"token_patterns", "known_secrets", "entropy"})
+INBOUND_DETECTOR_NAMES = frozenset({"naive_injection_detection"})
+
+# Per-route policy for what the proxy does when an outbound DLP detector
+# matches a token (PRD 0062).
+ON_MATCH_BLOCK = "block"          # hard 403, never overridable
+ON_MATCH_REDACT = "redact"        # scrub the matched value, forward the request
+ON_MATCH_SUPERVISE = "supervise"  # queue for operator approval, hold the request
+OUTBOUND_ON_MATCH_VALUES = (ON_MATCH_BLOCK, ON_MATCH_REDACT, ON_MATCH_SUPERVISE)
+# Unset resolves to supervise (fall back to block when supervise is not wired).
+DEFAULT_OUTBOUND_ON_MATCH = ON_MATCH_SUPERVISE
+
+
+def parse_dlp_block(
+    idx: int,
+    host: str,
+    raw_dict: dict[str, object],
+) -> tuple[tuple[str, ...] | None, tuple[str, ...] | None, str]:
+    """Parse the optional `dlp` block on a route, returning
+    (outbound_detectors, inbound_detectors, outbound_on_match)."""
+    dlp_raw = raw_dict.get("dlp")
+    if dlp_raw is None:
+        return None, None, ""
+    label = f"route[{idx}] ({host})"
+    if not isinstance(dlp_raw, dict):
+        raise ValueError(f"{label}: 'dlp' must be an object")
+    dlp = typing.cast(dict[str, object], dlp_raw)
+
+    def _parse_detector_field(
+        field: str,
+        valid_names: frozenset[str],
+    ) -> tuple[str, ...] | None:
+        val = dlp.get(field)
+        if val is None:
+            return None
+        if val is False:
+            return ()
+        if not isinstance(val, list):
+            raise ValueError(
+                f"{label}: dlp.{field} must be false, a list, or omitted"
+            )
+        items = typing.cast(list[object], val)
+        names: list[str] = []
+        for j, item in enumerate(items):
+            if not isinstance(item, str):
+                raise ValueError(
+                    f"{label}: dlp.{field}[{j}] must be a string"
+                )
+            if item not in valid_names:
+                raise ValueError(
+                    f"{label}: dlp.{field}[{j}] {item!r} is not a valid "
+                    f"detector name; valid names: {', '.join(sorted(valid_names))}"
+                )
+            names.append(item)
+        return tuple(names)
+
+    outbound = _parse_detector_field("outbound_detectors", OUTBOUND_DETECTOR_NAMES)
+    inbound = _parse_detector_field("inbound_detectors", INBOUND_DETECTOR_NAMES)
+
+    on_match = ""
+    on_match_raw = dlp.get("outbound_on_match")
+    if on_match_raw is not None:
+        if not isinstance(on_match_raw, str) or on_match_raw not in OUTBOUND_ON_MATCH_VALUES:
+            raise ValueError(
+                f"{label}: dlp.outbound_on_match must be one of "
+                f"{', '.join(OUTBOUND_ON_MATCH_VALUES)} (got {on_match_raw!r})"
+            )
+        on_match = on_match_raw
+
+    for k in dlp:
+        if k not in ("outbound_detectors", "inbound_detectors", "outbound_on_match"):
+            raise ValueError(
+                f"{label}: dlp has unknown key {k!r}; accepted keys "
+                f"are 'outbound_detectors', 'inbound_detectors', "
+                f"'outbound_on_match'"
+            )
+    return outbound, inbound, on_match

bot_bottle/env.py

@@ -114,7 +114,7 @@ def _read_secret_silent(name: str, prompt_body: str) -> str:
     return value
 
 
-def resolve_env(manifest: Manifest, agent: str) -> ResolvedEnv:
+def resolve_env(manifest: Manifest) -> ResolvedEnv:
     """Iterate the agent's env entries:
       - secret:       prompt at runtime; carry value in forwarded
       - interpolated: read $HOST_VAR from os.environ; carry value in forwarded
@@ -124,7 +124,7 @@ def resolve_env(manifest: Manifest, agent: str) -> ResolvedEnv:
     backend injects forwarded values via its launcher's env parameter."""
     forwarded: dict[str, str] = {}
     literals: dict[str, str] = {}
-    bottle = manifest.bottle_for(agent)
+    bottle = manifest.bottle
     for name, raw in bottle.env.items():
         if not name:
             continue

bot_bottle/git_gate.py

@@ -43,10 +43,10 @@ from .manifest import ManifestBottle, ManifestGitEntry
 # Short network alias for git-gate inside the sidecar bundle. The
 # agent's `.gitconfig` insteadOf rewrites resolve through this name.
 GIT_GATE_HOSTNAME = "git-gate"
-# Bound half-open git client sessions. If an agent/tool runner is
-# interrupted during push, git daemon should reap the receive-pack
-# child instead of keeping the gate wedged indefinitely.
-GIT_GATE_DAEMON_TIMEOUT_SECS = 15
+# Shared timeout (seconds) for all git-gate subprocess and CGI calls:
+# git daemon (--timeout/--init-timeout), the access-hook subprocess in
+# git_http_backend, and the git http-backend CGI subprocess.
+GIT_GATE_TIMEOUT_SECS = 15
 
 
 @dataclass(frozen=True)
@@ -112,6 +112,15 @@ def git_gate_upstreams_for_bottle(bottle: ManifestBottle) -> tuple[GitGateUpstre
     )
 
 
+def _gitconfig_validate_value(field: str, value: str) -> None:
+    """Raise ValueError if value contains characters that break gitconfig line syntax."""
+    if "\n" in value or "\r" in value:
+        raise ValueError(
+            f"git-gate: {field} contains a newline, which would inject "
+            f"arbitrary gitconfig keys; rejecting manifest entry"
+        )
+
+
 def git_gate_render_gitconfig(
     entries: tuple[ManifestGitEntry, ...], gate_host: str, *, scheme: str = "git",
 ) -> str:
@@ -136,6 +145,7 @@ def git_gate_render_gitconfig(
         "# fetch-from-upstream-before-every-upload-pack via access-hook).\n",
     ]
     for entry in entries:
+        _gitconfig_validate_value(f"repos[{entry.Name!r}].url", entry.Upstream)
         out.append(f'[url "{scheme}://{gate_host}/{entry.Name}.git"]\n')
         out.append(f"\tinsteadOf = {entry.Upstream}\n")
         if entry.RemoteKey and entry.RemoteKey != entry.UpstreamHost:
@@ -148,6 +158,7 @@ def git_gate_render_gitconfig(
                 f"ssh://{entry.UpstreamUser}@{entry.RemoteKey}{port}/"
                 f"{entry.UpstreamPath}"
             )
+            _gitconfig_validate_value(f"repos[{entry.Name!r}].url (resolved alias)", alias)
             out.append(f"\tinsteadOf = {alias}\n")
     return "".join(out)
 
@@ -217,8 +228,8 @@ def git_gate_render_entrypoint(upstreams: tuple[GitGateUpstream, ...]) -> str:
         "",
         "exec git daemon \\",
         "  --reuseaddr \\",
-        f"  --timeout={GIT_GATE_DAEMON_TIMEOUT_SECS} \\",
-        f"  --init-timeout={GIT_GATE_DAEMON_TIMEOUT_SECS} \\",
+        f"  --timeout={GIT_GATE_TIMEOUT_SECS} \\",
+        f"  --init-timeout={GIT_GATE_TIMEOUT_SECS} \\",
         "  --base-path=/git \\",
         "  --export-all \\",
         "  --enable=receive-pack \\",
@@ -247,6 +258,164 @@ cat > "$refs_file"
 
 zero=0000000000000000000000000000000000000000
 
+supervise_gitleaks_allow() {
+  log_opts=$1
+  ref=$2
+  report_file=$(mktemp)
+  if ! gitleaks git \
+      --log-opts="$log_opts" \
+      --no-banner \
+      --redact \
+      --ignore-gitleaks-allow \
+      --report-format=json \
+      --report-path="$report_file" \
+      --exit-code 0 \
+      1>&2; then
+    rm -f "$report_file"
+    echo "git-gate: gitleaks inline-suppression scan failed for $ref" >&2
+    return 1
+  fi
+
+  proposal_id=$(
+    GITLEAKS_ALLOW_REF="$ref" python3 - "$report_file" <<'PY'
+import datetime
+import hashlib
+import json
+import os
+import sys
+import uuid
+from pathlib import Path
+
+report_path = Path(sys.argv[1])
+queue_dir = os.environ.get("SUPERVISE_QUEUE_DIR", "")
+slug = os.environ.get("SUPERVISE_BOTTLE_SLUG", "")
+if not queue_dir or not slug:
+    sys.exit(2)
+
+try:
+    raw = json.loads(report_path.read_text() or "[]")
+except json.JSONDecodeError:
+    sys.exit(3)
+if not isinstance(raw, list):
+    sys.exit(3)
+if not raw:
+    sys.exit(0)
+
+ref = os.environ.get("GITLEAKS_ALLOW_REF", "")
+lines = [
+    "gitleaks inline suppression requires supervisor approval",
+    f"ref: {ref}",
+    "",
+]
+for i, finding in enumerate(raw, 1):
+    if not isinstance(finding, dict):
+        continue
+    file_path = finding.get("File", "")
+    line_no = finding.get("StartLine", finding.get("Line", ""))
+    rule_id = finding.get("RuleID", "")
+    commit = finding.get("Commit", "")
+    line = finding.get("Line", "")
+    lines.extend([
+        f"finding {i}:",
+        f"  file: {file_path}",
+        f"  line: {line_no}",
+        f"  rule: {rule_id}",
+        f"  commit: {commit}",
+        f"  code: {line}",
+        "",
+    ])
+
+payload = "\n".join(lines).rstrip() + "\n"
+proposal_id = str(uuid.uuid4())
+proposal = {
+    "id": proposal_id,
+    "bottle_slug": slug,
+    "tool": "gitleaks-allow",
+    "proposed_file": payload,
+    "justification": (
+        "git-gate found gitleaks findings hidden by # gitleaks:allow; "
+        "approve only for dummy test fixtures or confirmed false positives"
+    ),
+    "arrival_timestamp": datetime.datetime.now(
+        datetime.timezone.utc
+    ).isoformat(),
+    "current_file_hash": hashlib.sha256(payload.encode("utf-8")).hexdigest(),
+}
+queue = Path(queue_dir)
+queue.mkdir(parents=True, exist_ok=True)
+path = queue / f"{proposal_id}.proposal.json"
+tmp = path.with_suffix(path.suffix + ".tmp")
+with tmp.open("w", encoding="utf-8") as f:
+    json.dump(proposal, f, indent=2)
+    f.write("\n")
+os.chmod(tmp, 0o600)
+os.replace(tmp, path)
+print(proposal_id)
+PY
+  )
+  rc=$?
+  rm -f "$report_file"
+  if [ "$rc" -eq 0 ] && [ -z "$proposal_id" ]; then
+    return 0
+  fi
+  if [ "$rc" -ne 0 ]; then
+    echo "git-gate: cannot route # gitleaks:allow finding to supervisor; refusing push" >&2
+    return 1
+  fi
+
+  queue_dir=${SUPERVISE_QUEUE_DIR:-}
+  response_file="$queue_dir/${proposal_id}.response.json"
+  timeout=${SUPERVISE_GITLEAKS_ALLOW_TIMEOUT_SECONDS:-300}
+  case "$timeout" in
+    ''|*[!0-9]*)
+      echo "git-gate: invalid SUPERVISE_GITLEAKS_ALLOW_TIMEOUT_SECONDS=$timeout" >&2
+      return 1
+      ;;
+  esac
+  echo "git-gate: queued # gitleaks:allow supervisor approval $proposal_id" >&2
+  echo "git-gate: approve with './cli.py supervise' to continue this push" >&2
+  waited=0
+  while [ "$waited" -lt "$timeout" ]; do
+    if [ -f "$response_file" ]; then
+      status=$(python3 - "$response_file" <<'PY'
+import json
+import sys
+try:
+    with open(sys.argv[1], encoding="utf-8") as f:
+        raw = json.load(f)
+except (OSError, json.JSONDecodeError):
+    sys.exit(1)
+status = raw.get("status")
+if not isinstance(status, str):
+    sys.exit(1)
+print(status)
+PY
+      ) || status=""
+      case "$status" in
+        approved|modified)
+          mkdir -p "$queue_dir/processed"
+          mv -f "$queue_dir/${proposal_id}.proposal.json" "$queue_dir/processed/" 2>/dev/null || true
+          mv -f "$queue_dir/${proposal_id}.response.json" "$queue_dir/processed/" 2>/dev/null || true
+          echo "git-gate: supervisor approved # gitleaks:allow for $ref" >&2
+          return 0
+          ;;
+        rejected)
+          echo "git-gate: supervisor rejected # gitleaks:allow for $ref" >&2
+          return 1
+          ;;
+        *)
+          echo "git-gate: invalid supervisor response for # gitleaks:allow" >&2
+          return 1
+          ;;
+      esac
+    fi
+    sleep 1
+    waited=$((waited + 1))
+  done
+  echo "git-gate: supervisor approval timed out for # gitleaks:allow; refusing push" >&2
+  return 1
+}
+
 # Phase 1: gitleaks scan each ref's incoming commits.
 while IFS=' ' read -r old new ref; do
   [ -z "$ref" ] && continue
@@ -268,6 +437,9 @@ while IFS=' ' read -r old new ref; do
     echo "git-gate: gitleaks rejected push to $ref" >&2
     exit 1
   fi
+  if ! supervise_gitleaks_allow "$log_opts" "$ref"; then
+    exit 1
+  fi
 done < "$refs_file"
 
 # Phase 2: forward each ref to the upstream (`origin`, configured

bot_bottle/git_http_backend.py

@@ -16,6 +16,8 @@ from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 from pathlib import Path
 from urllib.parse import urlsplit
 
+from .git_gate import GIT_GATE_TIMEOUT_SECS
+
 
 DEFAULT_PORT = 9420
 
@@ -47,6 +49,7 @@ class GitHttpHandler(BaseHTTPRequestHandler):
                 [hook_path, "upload-pack", str(repo_dir), peer, peer],
                 capture_output=True,
                 check=False,
+                timeout=GIT_GATE_TIMEOUT_SECS,
             )
             if hook.returncode != 0:
                 detail = (hook.stderr or hook.stdout).decode(
@@ -110,6 +113,7 @@ class GitHttpHandler(BaseHTTPRequestHandler):
             env=env,
             capture_output=True,
             check=False,
+            timeout=GIT_GATE_TIMEOUT_SECS,
         )
         self._write_cgi_response(proc.stdout)
 
@@ -148,7 +152,13 @@ class GitHttpHandler(BaseHTTPRequestHandler):
             key, _, value = line.decode("latin1").partition(":")
             value = value.strip()
             if key.lower() == "status":
-                status = int(value.split()[0])
+                try:
+                    status = int(value.split()[0])
+                except (ValueError, IndexError):
+                    self.log_message(
+                        "malformed CGI Status header %r; using 500", value,
+                    )
+                    status = 500
             else:
                 headers.append((key, value))
         self.send_response(status)

bot_bottle/log.py

@@ -1,21 +1,107 @@
-"""Tiny logging wrappers. All output goes to stderr."""
+"""Tiny logging wrappers. All output goes to stderr.
+
+Two capabilities layer onto the bare wrappers (issue #252):
+
+  - **Levels.** `debug` / `info` / `warn` / `error` carry an ordered
+    severity. Output is gated by `BOT_BOTTLE_LOG_LEVEL` (debug | info |
+    warn | error; default `info`). A message emits when its severity is
+    at or above the threshold, so `debug` is silent by default and
+    `error` always surfaces (nothing sits above it) — which keeps the
+    fatal `die` path visible regardless of the configured level.
+
+  - **Context.** Every wrapper takes an optional `context` mapping that
+    renders as a parseable ` [k=v ...]` suffix (keys sorted; values with
+    whitespace/quotes are quoted), so failures can be filtered and
+    correlated instead of being flat strings.
+
+With no `context` and the default level, output is byte-identical to the
+original `bot-bottle: <msg>` / `bot-bottle: warning: <msg>` /
+`bot-bottle: error: <msg>` lines — the 100+ existing call sites are
+unaffected.
+"""
 
 from __future__ import annotations
 
+import os
 import sys
-from typing import NoReturn
+from typing import Mapping, NoReturn
+
+# Ordered severities. Gaps left between values so intermediate levels
+# can be added later without renumbering.
+DEBUG = 10
+INFO = 20
+WARN = 30
+ERROR = 40
+
+_LEVEL_NAMES: dict[str, int] = {
+    "debug": DEBUG,
+    "info": INFO,
+    "warn": WARN,
+    "warning": WARN,
+    "error": ERROR,
+}
+
+# Default threshold when BOT_BOTTLE_LOG_LEVEL is unset or unrecognised.
+_DEFAULT_THRESHOLD = INFO
+
+_LOG_LEVEL_ENV = "BOT_BOTTLE_LOG_LEVEL"
 
 
-def info(msg: str) -> None:
-    print(f"bot-bottle: {msg}", file=sys.stderr)
+def _threshold() -> int:
+    """Resolve the active level threshold from the environment.
+
+    Read per-call (not cached) so the level can be changed at runtime
+    and so tests can patch `os.environ` without a reload. Unknown values
+    fall back to the default rather than raising — logging must never be
+    the thing that crashes the process."""
+    raw = os.environ.get(_LOG_LEVEL_ENV, "")
+    return _LEVEL_NAMES.get(raw.strip().lower(), _DEFAULT_THRESHOLD)
 
 
-def warn(msg: str) -> None:
-    print(f"bot-bottle: warning: {msg}", file=sys.stderr)
+def _format_context(context: Mapping[str, object] | None) -> str:
+    """Render a context mapping as a ` [k=v k2=v2]` suffix.
+
+    Keys are sorted for stable, diffable output. Values that are empty or
+    contain whitespace or a quote are wrapped in double quotes (with inner
+    quotes escaped) so each `k=v` pair stays parseable. Empty/None context
+    renders as the empty string."""
+    if not context:
+        return ""
+    parts: list[str] = []
+    for key in sorted(context):
+        value = str(context[key])
+        if value == "" or any(ch.isspace() for ch in value) or '"' in value:
+            value = '"' + value.replace('"', '\\"') + '"'
+        parts.append(f"{key}={value}")
+    return " [" + " ".join(parts) + "]"
 
 
-def error(msg: str) -> None:
-    print(f"bot-bottle: error: {msg}", file=sys.stderr)
+def _emit(
+    level: int,
+    label: str,
+    msg: str,
+    context: Mapping[str, object] | None,
+) -> None:
+    if level < _threshold():
+        return
+    prefix = f"{label}: " if label else ""
+    sys.stderr.write(f"bot-bottle: {prefix}{msg}{_format_context(context)}\n")
+
+
+def debug(msg: str, *, context: Mapping[str, object] | None = None) -> None:
+    _emit(DEBUG, "debug", msg, context)
+
+
+def info(msg: str, *, context: Mapping[str, object] | None = None) -> None:
+    _emit(INFO, "", msg, context)
+
+
+def warn(msg: str, *, context: Mapping[str, object] | None = None) -> None:
+    _emit(WARN, "warning", msg, context)
+
+
+def error(msg: str, *, context: Mapping[str, object] | None = None) -> None:
+    _emit(ERROR, "error", msg, context)
 
 
 class Die(SystemExit):
@@ -31,6 +117,6 @@ class Die(SystemExit):
         self.message = message
 
 
-def die(msg: str) -> NoReturn:
-    error(msg)
+def die(msg: str, *, context: Mapping[str, object] | None = None) -> NoReturn:
+    error(msg, context=context)
     raise Die(1, msg)

bot_bottle/manifest.py

@@ -19,7 +19,7 @@ Bottle schema (frontmatter):
     repos:      { <name>: <git-gate-entry>, ... }  # optional
   egress: { routes: [ <egress-route>, ... ] }
     # route keys: host, matches, auth, role, dlp
-  supervise:    <bool>                          # optional
+  supervise:    <bool>                          # optional (default true)
 
 Agent schema (frontmatter):
   bottle:        <bottle-name>          # required
@@ -36,10 +36,23 @@ Bottles can ONLY live under $HOME. A bottles/ dir under $CWD is a
 warn at load time and contributes nothing. The trust boundary is
 expressed as filesystem layout rather than resolver logic.
 
-Validation runs once at load. Manifest.from_json_obj is preserved
-as a programmatic entry point (used by tests) that takes a dict
-with the same field names — useful for building manifests without
-on-disk files.
+Two types are exported:
+
+  ManifestIndex  — the multi-agent/bottle collection returned by
+                   resolve() and from_json_obj(). Used for agent
+                   selection (all_agent_names), validation
+                   (require_agent), and lazy loading (load_for_agent).
+                   This is the pre-preflight form.
+
+  Manifest       — a single-agent/bottle value type holding exactly
+                   one agent: ManifestAgent and one bottle:
+                   ManifestBottle (with the agent's git-gate.user
+                   already overlaid). Returned by load_for_agent().
+                   This is the post-preflight form passed to backends.
+
+ManifestIndex.from_json_obj is preserved as a programmatic entry
+point (used by tests) that takes a dict with the same field names —
+useful for building manifests without on-disk files.
 """
 
 from __future__ import annotations
@@ -71,6 +84,7 @@ __all__ = [
     "ManifestEgressConfig",
     "ManifestAgent",
     "ManifestBottle",
+    "ManifestIndex",
     "Manifest",
 ]
 
@@ -97,13 +111,11 @@ class ManifestBottle:
     # identity without any git-gate.repos upstreams, and vice versa.
     git_user: ManifestGitUser = field(default_factory=ManifestGitUser)
     egress: ManifestEgressConfig = field(default_factory=ManifestEgressConfig)
-    # Opt-in per-bottle stuck-recovery sidecar (PRD 0013). When true,
-    # the launch step brings up a supervise sidecar that exposes MCP
-    # tools to the agent (egress-block, capability-block) plus mounts
-    # the current-config dir read-only into the agent at
-    # /etc/bot-bottle/current-config. False (the default) skips the
-    # sidecar and mount.
-    supervise: bool = False
+    # Per-bottle stuck-recovery sidecar (PRD 0013). When true (the
+    # default, issue #249), the launch step brings up a supervise
+    # sidecar that exposes egress MCP tools to the agent. Set
+    # `supervise: false` to skip the sidecar.
+    supervise: bool = True
 
     @classmethod
     def from_dict(cls, name: str, raw: object) -> "ManifestBottle":
@@ -176,7 +188,7 @@ class ManifestBottle:
             else ManifestEgressConfig()
         )
 
-        supervise_raw = d.get("supervise", False)
+        supervise_raw = d.get("supervise", True)
         if not isinstance(supervise_raw, bool):
             raise ManifestError(
                 f"bottle '{name}' supervise must be a boolean "
@@ -189,14 +201,123 @@ class ManifestBottle:
         )
 
 
+def _merge_git_user(
+    agent_user: ManifestGitUser, base_user: ManifestGitUser
+) -> ManifestGitUser:
+    """Merge the agent's git.user over the bottle's, agent-wins-on-non-empty."""
+    if agent_user.is_empty():
+        return base_user
+    return ManifestGitUser(
+        name=agent_user.name or base_user.name,
+        email=agent_user.email or base_user.email,
+    )
+
+
+def _resolve_effective_bottle_eager(
+    agent_name: str,
+    agent: "ManifestAgent",
+    bottle_names: "tuple[str, ...]",
+    bottles: "Mapping[str, ManifestBottle]",
+) -> "ManifestBottle":
+    """Return the effective ManifestBottle for the eager (from_json_obj) path.
+
+    When bottle_names is non-empty they are merged in order. When empty, falls
+    back to agent.bottle. Raises ManifestError when neither is set."""
+    from .manifest_extends import merge_bottles_runtime
+
+    if bottle_names:
+        resolved: list[ManifestBottle] = []
+        for bn in bottle_names:
+            if bn not in bottles:
+                available = ", ".join(sorted(bottles.keys())) or "(none)"
+                raise ManifestError(
+                    f"bottle '{bn}' not defined. Available: {available}"
+                )
+            resolved.append(bottles[bn])
+        return merge_bottles_runtime(resolved)
+
+    if not agent.bottle:
+        raise ManifestError(
+            f"agent '{agent_name}' has no 'bottle' field and no bottles were "
+            f"selected at launch. Select at least one bottle or add "
+            f"'bottle: <name>' to the agent manifest."
+        )
+    return bottles[agent.bottle]
+
+
+def _resolve_effective_bottle_lazy(
+    agent_name: str,
+    agent_bottle: str,
+    bottle_names: "tuple[str, ...]",
+    bottles_dir: "Path",
+) -> "ManifestBottle":
+    """Return the effective ManifestBottle for the lazy (from_md_dirs) path.
+
+    When bottle_names is non-empty they are resolved from disk and merged in
+    order. When empty, falls back to agent_bottle. Raises ManifestError when
+    neither is set."""
+    from .manifest_extends import merge_bottles_runtime
+    from .manifest_loader import load_bottle_chain_from_dir
+
+    if bottle_names:
+        resolved = [load_bottle_chain_from_dir(bn, bottles_dir) for bn in bottle_names]
+        return merge_bottles_runtime(resolved)
+
+    if not agent_bottle:
+        raise ManifestError(
+            f"agent '{agent_name}' has no 'bottle' field and no bottles were "
+            f"selected at launch. Select at least one bottle or add "
+            f"'bottle: <name>' to the agent manifest."
+        )
+    return load_bottle_chain_from_dir(agent_bottle, bottles_dir)
+
+
 @dataclass(frozen=True)
 class Manifest:
+    """Single-agent/bottle value type. Returned by ManifestIndex.load_for_agent().
+
+    `bottle` is the effective bottle with the agent's git-gate.user already
+    overlaid per-field (agent wins on non-empty). Backends and provisioners
+    use this directly — no agent_name lookup needed."""
+
+    agent: ManifestAgent
+    bottle: ManifestBottle
+
+    def git_identity_summary(self) -> str | None:
+        """One-line effective git identity with per-field provenance, e.g.
+        `name=claude (agent), email=eric@dideric.is (bottle)`.
+        Returns None when neither agent nor bottle sets an identity."""
+        over = self.agent.git_user        # agent's declared git_user (pre-merge)
+        merged = self.bottle.git_user     # effective git_user (post-merge)
+        if merged.is_empty():
+            return None
+        parts: list[str] = []
+        if merged.name:
+            parts.append(f"name={merged.name} ({'agent' if over.name else 'bottle'})")
+        if merged.email:
+            parts.append(f"email={merged.email} ({'agent' if over.email else 'bottle'})")
+        return ", ".join(parts)
+
+
+@dataclass(frozen=True)
+class ManifestIndex:
+    """Multi-agent/bottle collection. The pre-preflight form.
+
+    In lazy mode (from resolve()/from_md_dirs()) only filenames are scanned;
+    no file content is read. In eager mode (from from_json_obj()) all agents
+    and bottles are pre-parsed. Call load_for_agent() to get a single-value
+    Manifest ready for backend use."""
+
     bottles: Mapping[str, ManifestBottle]
     agents: Mapping[str, ManifestAgent]
+    # Set by from_md_dirs; None in from_json_obj (test/programmatic) mode.
+    # Stores the manifest root dirs so load_for_agent can locate files later.
+    home_md: Path | None = field(default=None)
+    cwd_md: Path | None = field(default=None)
 
     @classmethod
-    def resolve(cls, cwd: str, *, missing_ok: bool = False) -> "Manifest":
-        """Walk the per-file manifest tree and build a Manifest.
+    def resolve(cls, cwd: str, *, missing_ok: bool = False) -> "ManifestIndex":
+        """Walk the per-file manifest tree and build a ManifestIndex.
 
         Layout (PRD 0011):
           $HOME/.bot-bottle/bottles/<name>.md  — bottles (home-only)
@@ -209,7 +330,7 @@ class Manifest:
         boundary.
 
         If `missing_ok` is true, a missing `$HOME/.bot-bottle/`
-        returns an empty manifest instead of dying. This is for
+        returns an empty index instead of dying. This is for
         passive UI surfaces like the dashboard, which can still
         monitor already-running agents without launch config.
 
@@ -248,25 +369,16 @@ class Manifest:
         cls,
         home_dir: Path,
         cwd_dir: Path | None,
-    ) -> "Manifest":
-        """Programmatic entry point. Loads bottles from
-        `<home_dir>/bottles/`, home agents from `<home_dir>/agents/`,
-        and (if `cwd_dir` is passed) cwd agents from
-        `<cwd_dir>/agents/`. Cwd agents override home agents on
-        name collision. A `bottles/` subdir under `cwd_dir` is
-        logged as a warning and ignored.
+    ) -> "ManifestIndex":
+        """Return a names-only ManifestIndex. No file content is read; only
+        filenames are scanned for the agent selector.  Full parsing happens
+        later, per-agent, via `load_for_agent`.
 
-        Used by tests to build a Manifest from fixture directories
+        A `bottles/` subdir under `cwd_dir` is logged as a warning and
+        ignored — the filesystem layout IS the trust boundary.
+
+        Used by tests to build a ManifestIndex from fixture directories
         without touching `os.environ`."""
-        bottles_dir = home_dir / "bottles"
-        from .manifest_loader import load_agents_from_dir, load_bottles_from_dir
-
-        bottles = load_bottles_from_dir(bottles_dir)
-
-        bottle_names = set(bottles.keys())
-        agents_dir = home_dir / "agents"
-        agents = load_agents_from_dir(agents_dir, bottle_names, source="$HOME")
-
         if cwd_dir is not None:
             stale_bottles = cwd_dir / "bottles"
             if stale_bottles.is_dir():
@@ -280,17 +392,11 @@ class Manifest:
                         f"live under $HOME/.bot-bottle/bottles/ "
                         f"(PRD 0011). Move them or delete."
                     )
-            cwd_agents_dir = cwd_dir / "agents"
-            cwd_agents = load_agents_from_dir(
-                cwd_agents_dir, bottle_names, source="$CWD"
-            )
-            agents = {**agents, **cwd_agents}
-
-        return cls(bottles=bottles, agents=agents)
+        return cls(bottles={}, agents={}, home_md=home_dir, cwd_md=cwd_dir)
 
     @classmethod
-    def from_json_obj(cls, obj: object) -> "Manifest":
-        """Validate and build a Manifest from a raw JSON-like dict."""
+    def from_json_obj(cls, obj: object) -> "ManifestIndex":
+        """Validate and build a ManifestIndex from a raw JSON-like dict."""
         d = as_json_object(obj, "manifest")
         raw_bottles_obj = _section_dict(d.get("bottles"), "manifest 'bottles'")
         raw_agents = _section_dict(d.get("agents"), "manifest 'agents'")
@@ -311,75 +417,151 @@ class Manifest:
         }
         return cls(bottles=bottles, agents=agents)
 
+    @property
+    def all_bottle_names(self) -> list[str]:
+        """Sorted list of all discoverable bottle names.
+
+        In names-only mode (from resolve/from_md_dirs) this scans bottle
+        filenames without reading their content. In eager mode (from
+        from_json_obj) it returns the pre-parsed bottles' names."""
+        if self.home_md is not None:
+            from .manifest_loader import scan_bottle_names
+            return scan_bottle_names(self.home_md / "bottles")
+        return sorted(self.bottles.keys())
+
+    @property
+    def all_agent_names(self) -> list[str]:
+        """Sorted list of all discoverable agent names.
+
+        In names-only mode (from resolve/from_md_dirs) this scans agent
+        filenames without reading their content. In eager mode (from
+        from_json_obj) it returns the pre-parsed agents' names."""
+        if self.home_md is not None:
+            from .manifest_loader import scan_agent_names
+            home_names = set(scan_agent_names(self.home_md / "agents").keys())
+            cwd_names: set[str] = set()
+            if self.cwd_md is not None:
+                cwd_names = set(scan_agent_names(self.cwd_md / "agents").keys())
+            return sorted(home_names | cwd_names)
+        return sorted(self.agents.keys())
+
+    def load_for_agent(
+        self,
+        agent_name: str,
+        bottle_names: "tuple[str, ...] | None" = None,
+    ) -> "Manifest":
+        """Parse the named agent and its bottle; return a single-value Manifest.
+
+        `bottle_names` is an ordered list of bottles selected at launch time.
+        When non-empty they are resolved and merged in order (index 0 = base;
+        later entries override). When empty or None, falls back to the agent's
+        own `bottle:` field. Raises ManifestError when neither is set.
+
+        In lazy mode (from resolve/from_md_dirs) the agent file and its
+        bottle chain are read from disk for the first time here.  In eager
+        mode (from_json_obj) the data is already parsed; this just filters
+        down to the requested agent and its bottle.
+
+        The returned Manifest.bottle has the agent's git-gate.user already
+        overlaid (agent wins on non-empty, per-field).
+
+        Always raises ManifestError if the agent is unknown or invalid.
+        Backends call this at preflight inside _validate."""
+        effective_bottle_names: tuple[str, ...] = bottle_names or ()
+
+        if self.home_md is None:
+            # Eager manifest (from_json_obj): data already parsed; filter to
+            # the one requested agent and its bottle so the returned Manifest
+            # always holds exactly one agent and one bottle regardless of path.
+            if agent_name not in self.agents:
+                available = ", ".join(sorted(self.agents.keys())) or "(none)"
+                raise ManifestError(
+                    f"agent '{agent_name}' not defined. Available: {available}"
+                )
+            agent = self.agents[agent_name]
+            raw_bottle = _resolve_effective_bottle_eager(
+                agent_name, agent, effective_bottle_names, self.bottles
+            )
+            merged = _merge_git_user(agent.git_user, raw_bottle.git_user)
+            bottle = raw_bottle if merged == raw_bottle.git_user else replace(raw_bottle, git_user=merged)
+            return Manifest(agent=agent, bottle=bottle)
+
+        from .manifest_loader import scan_agent_names
+        from .manifest_schema import validate_agent_frontmatter_keys
+        from .yaml_subset import YamlSubsetError, parse_frontmatter
+
+        # Locate the agent file; cwd wins over home on name collision.
+        home_agents = scan_agent_names(self.home_md / "agents")
+        cwd_agents: dict[str, Path] = {}
+        if self.cwd_md is not None:
+            cwd_agents = scan_agent_names(self.cwd_md / "agents")
+        merged_agents = {**home_agents, **cwd_agents}
+
+        if agent_name not in merged_agents:
+            available = ", ".join(sorted(merged_agents.keys())) or "(none)"
+            raise ManifestError(
+                f"agent '{agent_name}' not defined. Available: {available}"
+            )
+
+        agent_path = merged_agents[agent_name]
+        try:
+            fm, body = parse_frontmatter(agent_path.read_text())
+        except OSError as e:
+            raise ManifestError(f"could not read {agent_path}: {e}") from e
+        except YamlSubsetError as e:
+            raise ManifestError(f"{agent_path}: {e}") from e
+
+        validate_agent_frontmatter_keys(agent_path, fm.keys())
+
+        # Determine the effective bottle name(s).
+        agent_bottle = fm.get("bottle") or ""
+        bottles_dir = self.home_md / "bottles"
+        raw_bottle = _resolve_effective_bottle_lazy(
+            agent_name, str(agent_bottle), effective_bottle_names, bottles_dir
+        )
+        effective_bottle_name = (
+            effective_bottle_names[-1] if effective_bottle_names
+            else str(agent_bottle)
+        )
+
+        # Build and validate the full ManifestAgent.
+        agent_dict: dict[str, object] = {
+            "skills": fm.get("skills", []),
+            "prompt": body.strip(),
+        }
+        if agent_bottle:
+            agent_dict["bottle"] = agent_bottle
+        if "git-gate" in fm:
+            agent_dict["git-gate"] = fm["git-gate"]
+        # Pass the effective bottle name as the known-bottles set so agents
+        # that have bottle: set are validated; agents without bottle: pass {}
+        # since bottle_names were already resolved above.
+        known = {effective_bottle_name} if effective_bottle_name else set()
+        agent = ManifestAgent.from_dict(agent_name, agent_dict, known)
+
+        merged_user = _merge_git_user(agent.git_user, raw_bottle.git_user)
+        bottle = raw_bottle if merged_user == raw_bottle.git_user else replace(raw_bottle, git_user=merged_user)
+        return Manifest(agent=agent, bottle=bottle)
+
     def has_agent(self, name: str) -> bool:
         return name in self.agents
 
     def require_agent(self, name: str) -> None:
+        """Check that `name` is a discoverable agent. In names-only mode
+        this checks whether the .md file exists; in eager mode it checks
+        the pre-parsed agents dict. Does NOT parse file content."""
         if self.has_agent(name):
             return
-        available = ", ".join(self.agents.keys())
-        if available:
-            msg = f"agent '{name}' not defined in bot-bottle.json. Available: {available}"
-            raise ManifestError(msg)
-        raise ManifestError(
-            f"agent '{name}' not defined in bot-bottle.json (manifest is empty)."
-        )
-
-    def has_bottle(self, name: str) -> bool:
-        return name in self.bottles
-
-    def require_bottle(self, name: str) -> None:
-        if self.has_bottle(name):
-            return
-        available = ", ".join(self.bottles.keys())
-        if available:
-            raise ManifestError(
-                f"bottle '{name}' not defined in bot-bottle.json. "
-                f"Available bottles: {available}"
+        if self.home_md is not None:
+            # Names-only mode: check file existence without parsing.
+            home_path = self.home_md / "agents" / f"{name}.md"
+            cwd_path = (
+                self.cwd_md / "agents" / f"{name}.md"
+                if self.cwd_md else None
             )
-        raise ManifestError(f"bottle '{name}' not defined in bot-bottle.json (no bottles defined).")
-
-    def _effective_git_user(self, agent_name: str) -> ManifestGitUser:
-        """Merge the agent's git.user over the referenced bottle's,
-        per-field, agent-wins-on-non-empty (issue #94). Same overlay
-        the `extends:` resolver applies between bottles
-        (`_merge_bottles`)."""
-        agent = self.agents[agent_name]
-        base = self.bottles[agent.bottle].git_user
-        over = agent.git_user
-        if over.is_empty():
-            return base
-        return ManifestGitUser(
-            name=over.name or base.name,
-            email=over.email or base.email,
+            if home_path.is_file() or (cwd_path and cwd_path.is_file()):
+                return
+        available = ", ".join(self.all_agent_names) or "(none)"
+        raise ManifestError(
+            f"agent '{name}' not defined. Available: {available}"
         )
-
-    def bottle_for(self, agent_name: str) -> ManifestBottle:
-        """Resolve the Bottle the named agent references, with the
-        agent's git.user overlaid on top. The validator guarantees both
-        lookups succeed for a manifest built via from_json_obj.
-
-        The overlay lives here, the single point both backends call to
-        resolve an agent's bottle, so the docker / smolmachines git
-        provisioners pick up the merged identity unchanged."""
-        bottle = self.bottles[self.agents[agent_name].bottle]
-        merged = self._effective_git_user(agent_name)
-        if merged == bottle.git_user:
-            return bottle
-        return replace(bottle, git_user=merged)
-
-    def git_identity_summary(self, agent_name: str) -> str | None:
-        """One-line effective git identity with per-field provenance
-        for launch summaries, e.g.
-        `name=claude (agent), email=eric@dideric.is (bottle)`.
-        Returns None when neither agent nor bottle sets an identity."""
-        over = self.agents[agent_name].git_user
-        merged = self._effective_git_user(agent_name)
-        if merged.is_empty():
-            return None
-        parts: list[str] = []
-        if merged.name:
-            parts.append(f"name={merged.name} ({'agent' if over.name else 'bottle'})")
-        if merged.email:
-            parts.append(f"email={merged.email} ({'agent' if over.email else 'bottle'})")
-        return ", ".join(parts)

bot_bottle/manifest_agent.py

@@ -109,7 +109,8 @@ class ManifestAgentProvider:
 
 @dataclass(frozen=True)
 class ManifestAgent:
-    bottle: str
+    # Optional: when empty the operator selects bottles at launch time.
+    bottle: str = ""
     skills: tuple[str, ...] = ()
     prompt: str = ""
     # Per-agent git identity (issue #94). Overlays the referenced
@@ -129,18 +130,20 @@ class ManifestAgent:
                 f"allowed keys are {allowed}."
             )
 
-        bottle = d.get("bottle")
-        if not isinstance(bottle, str) or not bottle:
-            raise ManifestError(
-                f"agent '{name}' must declare a 'bottle' field naming a "
-                f"defined bottle"
-            )
-        if bottle not in bottle_names:
-            available = ", ".join(sorted(bottle_names)) or "(none defined)"
-            raise ManifestError(
-                f"agent '{name}' references bottle '{bottle}', which is not defined. "
-                f"Available: {available}"
-            )
+        bottle_raw = d.get("bottle")
+        bottle = ""
+        if bottle_raw is not None:
+            if not isinstance(bottle_raw, str) or not bottle_raw:
+                raise ManifestError(
+                    f"agent '{name}' bottle must be a non-empty string when declared"
+                )
+            if bottle_raw not in bottle_names:
+                available = ", ".join(sorted(bottle_names)) or "(none defined)"
+                raise ManifestError(
+                    f"agent '{name}' references bottle '{bottle_raw}', which is not defined. "
+                    f"Available: {available}"
+                )
+            bottle = bottle_raw
 
         skills: tuple[str, ...] = ()
         skills_raw = d.get("skills")
@@ -199,13 +202,10 @@ def _parse_provider_settings(
 ) -> dict[str, object]:
     if raw is None:
         return {}
-    if template != "pi":
-        raise ManifestError(
-            f"bottle '{bottle_name}' agent_provider.settings is only "
-            "supported for template 'pi'"
-        )
     settings = as_json_object(raw, f"bottle '{bottle_name}' agent_provider.settings")
-    allowed = {
+
+    common_allowed = {"startup_args"}
+    pi_allowed = {
         "provider",
         "base_url",
         "api",
@@ -218,12 +218,37 @@ def _parse_provider_settings(
         "supports_developer_role",
         "supports_reasoning_effort",
     }
+    if template == "pi":
+        allowed = common_allowed | pi_allowed
+    elif template in ("claude", "codex"):
+        allowed = common_allowed
+    elif template not in PROVIDER_TEMPLATES:
+        return dict(settings)
+    else:
+        allowed = common_allowed
+
     for key in settings:
         if key not in allowed:
             raise ManifestError(
                 f"bottle '{bottle_name}' agent_provider.settings has unknown "
                 f"key {key!r}; allowed: {', '.join(sorted(allowed))}"
             )
+    startup_args = settings.get("startup_args")
+    if startup_args is not None:
+        if not isinstance(startup_args, list):
+            raise ManifestError(
+                f"bottle '{bottle_name}' agent_provider.settings.startup_args "
+                f"must be an array of strings"
+            )
+        for i, arg in enumerate(startup_args):
+            if not isinstance(arg, str) or not arg:
+                raise ManifestError(
+                    f"bottle '{bottle_name}' agent_provider.settings."
+                    f"startup_args[{i}] must be a non-empty string"
+                )
+    if template != "pi":
+        return dict(settings)
+
     for key in ("provider", "base_url", "api", "api_key", "api_key_env"):
         value = settings.get(key)
         if value is not None and (not isinstance(value, str) or not value):

bot_bottle/manifest_egress.py

@@ -21,6 +21,9 @@ VALID_METHODS = frozenset({
 OUTBOUND_DETECTOR_NAMES = frozenset({"token_patterns", "known_secrets"})
 INBOUND_DETECTOR_NAMES = frozenset({"naive_injection_detection"})
 
+# What the proxy does on an outbound token match (PRD 0062).
+OUTBOUND_ON_MATCH_VALUES = ("block", "redact", "supervise")
+
 
 def validate_egress_routes(
     bottle_name: str,
@@ -67,6 +70,7 @@ class ManifestEgressRoute:
     GitFetch: bool = False
     OutboundDetectors: tuple[str, ...] | None = None
     InboundDetectors: tuple[str, ...] | None = None
+    OutboundOnMatch: str = ""
 
     @classmethod
     def from_dict(cls, bottle_name: str, idx: int, raw: object) -> "ManifestEgressRoute":
@@ -161,8 +165,9 @@ class ManifestEgressRoute:
         # --- dlp ---
         outbound_detectors: tuple[str, ...] | None = None
         inbound_detectors: tuple[str, ...] | None = None
+        outbound_on_match = ""
         if "dlp" in d:
-            outbound_detectors, inbound_detectors = _parse_dlp_block(
+            outbound_detectors, inbound_detectors, outbound_on_match = _parse_dlp_block(
                 label, d.get("dlp"),
             )
 
@@ -201,6 +206,7 @@ class ManifestEgressRoute:
             GitFetch=git_fetch,
             OutboundDetectors=outbound_detectors,
             InboundDetectors=inbound_detectors,
+            OutboundOnMatch=outbound_on_match,
         )
 
 
@@ -323,7 +329,7 @@ def _parse_header_match(
 def _parse_dlp_block(
     route_label: str,
     raw: object,
-) -> tuple[tuple[str, ...] | None, tuple[str, ...] | None]:
+) -> tuple[tuple[str, ...] | None, tuple[str, ...] | None, str]:
     label = f"{route_label} dlp"
     d = as_json_object(raw, label)
 
@@ -358,13 +364,24 @@ def _parse_dlp_block(
     outbound = _parse_field("outbound_detectors", OUTBOUND_DETECTOR_NAMES)
     inbound = _parse_field("inbound_detectors", INBOUND_DETECTOR_NAMES)
 
+    on_match = ""
+    on_match_raw = d.get("outbound_on_match")
+    if on_match_raw is not None:
+        if not isinstance(on_match_raw, str) or on_match_raw not in OUTBOUND_ON_MATCH_VALUES:
+            raise ManifestError(
+                f"{label} outbound_on_match must be one of "
+                f"{', '.join(OUTBOUND_ON_MATCH_VALUES)} (got {on_match_raw!r})"
+            )
+        on_match = on_match_raw
+
     for k in d:
-        if k not in ("outbound_detectors", "inbound_detectors"):
+        if k not in ("outbound_detectors", "inbound_detectors", "outbound_on_match"):
             raise ManifestError(
                 f"{label} has unknown key {k!r}; accepted keys are "
-                f"'outbound_detectors', 'inbound_detectors'"
+                f"'outbound_detectors', 'inbound_detectors', "
+                f"'outbound_on_match'"
             )
-    return outbound, inbound
+    return outbound, inbound, on_match
 
 
 LOG_LEVELS = frozenset({0, 1, 2})

bot_bottle/manifest_extends.py

@@ -5,16 +5,72 @@ from __future__ import annotations
 from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
-    from .manifest import ManifestBottle, ManifestGitEntry
+    from .manifest import ManifestBottle
     from .manifest_egress import ManifestEgressConfig
 
 
+def merge_bottles_runtime(bottles: "list[ManifestBottle]") -> "ManifestBottle":
+    """Merge an ordered list of pre-resolved ManifestBottle objects.
+
+    Index 0 is the base; each subsequent entry is applied on top using
+    the same field-merge rules as the file-based extends machinery:
+      env: dict merge, later wins; git_user: per-field overlay, later
+      wins on non-empty; git (repos): union by name, later wins; egress
+      routes: concatenate; agent_provider, supervise: later replaces.
+    """
+    if not bottles:
+        raise ValueError("merge_bottles_runtime requires at least one bottle")
+    result = bottles[0]
+    for override in bottles[1:]:
+        result = _merge_two_bottles_runtime(result, override)
+    return result
+
+
+def _merge_two_bottles_runtime(base: "ManifestBottle", override: "ManifestBottle") -> "ManifestBottle":
+    from .manifest import ManifestBottle, ManifestGitUser
+    from .manifest_egress import ManifestEgressConfig
+
+    merged_env = {**base.env, **override.env}
+
+    merged_git_user = ManifestGitUser(
+        name=override.git_user.name or base.git_user.name,
+        email=override.git_user.email or base.git_user.email,
+    )
+
+    # git repos: union keyed by Name, override wins per-name.
+    base_repos_by_name = {entry.Name: entry for entry in base.git}
+    override_repos_by_name = {entry.Name: entry for entry in override.git}
+    merged_repos_names = list(base_repos_by_name) + [
+        n for n in override_repos_by_name if n not in base_repos_by_name
+    ]
+    merged_git = tuple(
+        override_repos_by_name.get(n, base_repos_by_name[n])
+        for n in merged_repos_names
+    )
+
+    merged_routes = base.egress.routes + override.egress.routes
+    merged_egress = ManifestEgressConfig(routes=merged_routes, Log=override.egress.Log)
+
+    return ManifestBottle(
+        env=merged_env,
+        agent_provider=override.agent_provider,
+        git=merged_git,
+        git_user=merged_git_user,
+        egress=merged_egress,
+        supervise=override.supervise,
+    )
+
+
 def resolve_bottles(raws: dict[str, dict[str, object]]) -> dict[str, ManifestBottle]:
     """Apply `extends:` chains and return resolved ManifestBottle objects."""
     cache: dict[str, ManifestBottle] = {}
+    # Per-bottle effective git-gate.repos, as raw dicts keyed by repo name.
+    # Threaded alongside `cache` so a child can field-merge against its
+    # parent's repos without reconstructing them from parsed entries.
+    repos_cache: dict[str, dict[str, object]] = {}
     for name in raws:
         if name not in cache:
-            _resolve_one_bottle(name, raws, cache, ())
+            _resolve_one_bottle(name, raws, cache, repos_cache, ())
     return cache
 
 
@@ -22,6 +78,7 @@ def _resolve_one_bottle(
     name: str,
     raws: dict[str, dict[str, object]],
     cache: dict[str, ManifestBottle],
+    repos_cache: dict[str, dict[str, object]],
     seen: tuple[str, ...],
 ) -> ManifestBottle:
     from .manifest import ManifestBottle, ManifestError
@@ -41,39 +98,145 @@ def _resolve_one_bottle(
     if parent_name_raw is None:
         bottle = ManifestBottle.from_dict(name, child_raw)
         cache[name] = bottle
+        repos_cache[name] = _resolve_repos_raw({}, child_raw)
         return bottle
 
-    if not isinstance(parent_name_raw, str):
+    # Normalize to list, accepting both str and list[str].
+    raw_list: list[object]
+    if isinstance(parent_name_raw, str):
+        raw_list = [parent_name_raw]
+    elif isinstance(parent_name_raw, list):
+        raw_list = parent_name_raw
+    else:
         raise ManifestError(
-            f"bottle '{name}' extends must be a string "
+            f"bottle '{name}' extends must be a string or list of strings "
             f"(was {type(parent_name_raw).__name__})"
         )
-    parent_name: str = parent_name_raw
-    if parent_name == name:
-        raise ManifestError(
-            f"bottle '{name}' extends itself; remove the "
-            f"self-reference"
-        )
-    if parent_name not in raws:
-        avail = ", ".join(sorted(raws.keys())) or "(none)"
-        raise ManifestError(
-            f"bottle '{name}' extends '{parent_name}' which is not "
-            f"defined. Available bottles: {avail}"
-        )
-    parent = _resolve_one_bottle(parent_name, raws, cache, seen + (name,))
-    bottle = _merge_bottles(parent, child_raw, name)
+
+    # Validate each entry before resolving any of them.
+    parent_names: list[str] = []
+    for i, pname in enumerate(raw_list):
+        if not isinstance(pname, str):
+            raise ManifestError(
+                f"bottle '{name}' extends[{i}] must be a string "
+                f"(was {type(pname).__name__})"
+            )
+        parent_names.append(pname)
+        if pname == name:
+            raise ManifestError(
+                f"bottle '{name}' extends itself; remove the self-reference"
+            )
+        if pname not in raws:
+            avail = ", ".join(sorted(raws.keys())) or "(none)"
+            raise ManifestError(
+                f"bottle '{name}' extends '{pname}' which is not "
+                f"defined. Available bottles: {avail}"
+            )
+
+    combined_parent, combined_repos_raw = _fold_parents(
+        parent_names, raws, cache, repos_cache, seen + (name,)
+    )
+    merged_repos_raw = _resolve_repos_raw(combined_repos_raw, child_raw)
+    bottle = _merge_bottles(combined_parent, child_raw, merged_repos_raw, name)
     cache[name] = bottle
+    repos_cache[name] = merged_repos_raw
     return bottle
 
 
+def _fold_parents(
+    parent_names: list[str],
+    raws: dict[str, dict[str, object]],
+    cache: dict[str, ManifestBottle],
+    repos_cache: dict[str, dict[str, object]],
+    seen: tuple[str, ...],
+) -> tuple[ManifestBottle, dict[str, object]]:
+    """Resolve each parent and fold them left-to-right.
+
+    Later parents win over earlier ones on conflict.  The `seen` tuple
+    carries the current bottle's name so cycle detection works across
+    every parent edge in the multi-parent graph."""
+    first = parent_names[0]
+    effective = _resolve_one_bottle(first, raws, cache, repos_cache, seen)
+    effective_repos_raw = repos_cache[first]
+    for pname in parent_names[1:]:
+        later = _resolve_one_bottle(pname, raws, cache, repos_cache, seen)
+        later_repos_raw = repos_cache[pname]
+        effective, effective_repos_raw = _fold_two_bottles(
+            effective, effective_repos_raw, later, later_repos_raw
+        )
+    return effective, effective_repos_raw
+
+
+def _fold_two_bottles(
+    earlier: ManifestBottle,
+    earlier_repos_raw: dict[str, object],
+    later: ManifestBottle,
+    later_repos_raw: dict[str, object],
+) -> tuple[ManifestBottle, dict[str, object]]:
+    """Combine two resolved parent bottles; later wins over earlier."""
+    from .manifest import ManifestBottle, ManifestGitUser
+    from .manifest_egress import ManifestEgressConfig
+    from .manifest_git import parse_git_gate_config
+    from .manifest_util import as_json_object
+
+    merged_env = {**earlier.env, **later.env}
+
+    merged_git_user = ManifestGitUser(
+        name=later.git_user.name or earlier.git_user.name,
+        email=later.git_user.email or earlier.git_user.email,
+    )
+
+    # Repos: union by name; for same-name entries, later wins per-field.
+    # Unlike _resolve_repos_raw, an empty later_repos_raw means "no repos
+    # declared" — it does NOT clear the earlier parent's repos.
+    names = list(earlier_repos_raw) + [
+        n for n in later_repos_raw if n not in earlier_repos_raw
+    ]
+    merged_repos_raw: dict[str, object] = {
+        n: {
+            **as_json_object(earlier_repos_raw.get(n, {}), "earlier parent repo"),
+            **as_json_object(later_repos_raw.get(n, {}), "later parent repo"),
+        }
+        for n in names
+    }
+    if merged_repos_raw:
+        merged_git, _ = parse_git_gate_config("_fold", {"repos": merged_repos_raw})
+    else:
+        merged_git = ()
+
+    # Egress: routes concatenate; scalar fields use last-wins.
+    merged_egress = ManifestEgressConfig(
+        routes=earlier.egress.routes + later.egress.routes,
+        Log=later.egress.Log,
+    )
+
+    return ManifestBottle(
+        env=merged_env,
+        agent_provider=later.agent_provider,
+        git=merged_git,
+        git_user=merged_git_user,
+        egress=merged_egress,
+        supervise=later.supervise,
+    ), merged_repos_raw
+
+
 def _merge_bottles(
     parent: ManifestBottle,
     child_raw: dict[str, object],
+    merged_repos_raw: dict[str, object],
     name: str,
 ) -> ManifestBottle:
     """Apply PRD 0025 merge rules."""
     from .manifest import ManifestBottle, ManifestGitUser
     from .manifest_egress import validate_egress_routes
+    from .manifest_util import as_json_object
+
+    # git-gate.repos: when the child declares repos, inject the already
+    # name-merged repo set (computed by _resolve_repos_raw) so the child
+    # parses with the full inherited+overridden list (issue #237).
+    if _child_declares_git_gate_repos(child_raw):
+        git_raw = as_json_object(child_raw.get("git-gate", {}), "child git-gate")
+        child_raw = {**child_raw, "git-gate": {**git_raw, "repos": merged_repos_raw}}
 
     # Parse the child's declared fields into a ManifestBottle (with the
     # usual defaults for anything missing). Validation runs the same
@@ -92,11 +255,11 @@ def _merge_bottles(
         email=child.git_user.email or parent.git_user.email,
     )
 
-    # git-gate.repos: missing means inherit; an explicit empty object
-    # clears; otherwise parent and child merge by UpstreamHost with
-    # child entries replacing duplicate hosts.
+    # git-gate.repos: when declared, child.git already holds the merged
+    # set (an explicit empty dict clears parent, leaving child.git empty).
+    # When omitted, the parent's entries are inherited verbatim.
     if _child_declares_git_gate_repos(child_raw):
-        merged_git = _merge_git_remotes(parent.git, child.git) if child.git else ()
+        merged_git = child.git
     else:
         merged_git = parent.git
 
@@ -130,6 +293,45 @@ def _merge_bottles(
     )
 
 
+def _resolve_repos_raw(
+    parent_repos: dict[str, object],
+    child_raw: dict[str, object],
+) -> dict[str, object]:
+    """Compute a bottle's effective git-gate.repos as raw dicts.
+
+    Repos are keyed by name. When the child omits git-gate.repos it
+    inherits the parent's set verbatim; an explicit empty dict clears it.
+    Otherwise parent and child unite by name, with same-name entries
+    field-merged (parent fields are defaults, child fields win)."""
+    from .manifest_util import as_json_object
+
+    if not _child_declares_git_gate_repos(child_raw):
+        return parent_repos
+    child_repos = _declared_repos_raw(child_raw)
+    if not child_repos:
+        return {}
+    # Parent entries keep their order; child-only names are appended.
+    names = list(parent_repos) + [n for n in child_repos if n not in parent_repos]
+    return {
+        name: {
+            **as_json_object(parent_repos.get(name, {}), "parent git-gate repo"),
+            **as_json_object(child_repos.get(name, {}), "child git-gate repo"),
+        }
+        for name in names
+    }
+
+
+def _declared_repos_raw(child_raw: dict[str, object]) -> dict[str, object]:
+    """Return the child's explicitly declared git-gate.repos as raw dicts,
+    or an empty dict when none are declared."""
+    from .manifest_util import as_json_object
+
+    if not _child_declares_git_gate_repos(child_raw):
+        return {}
+    git_raw = as_json_object(child_raw.get("git-gate", {}), "child git-gate")
+    return as_json_object(git_raw.get("repos", {}), "child git-gate.repos")
+
+
 def _child_declares_git_gate_repos(child_raw: dict[str, object]) -> bool:
     from .manifest_util import as_json_object
 
@@ -140,16 +342,6 @@ def _child_declares_git_gate_repos(child_raw: dict[str, object]) -> bool:
     return "repos" in git_obj
 
 
-def _merge_git_remotes(
-    parent: tuple[ManifestGitEntry, ...],
-    child: tuple[ManifestGitEntry, ...],
-) -> tuple[ManifestGitEntry, ...]:
-    by_host = {entry.UpstreamHost: entry for entry in parent}
-    for entry in child:
-        by_host[entry.UpstreamHost] = entry
-    return tuple(by_host.values())
-
-
 def _merge_egress(
     parent: ManifestEgressConfig,
     child: ManifestEgressConfig,

bot_bottle/manifest_loader.py

@@ -8,21 +8,19 @@ from typing import TYPE_CHECKING
 from .log import warn
 from .manifest_schema import (
     entity_name_from_path,
-    validate_agent_frontmatter_keys,
     validate_bottle_frontmatter_keys,
 )
+from .manifest_util import ManifestError
 from .yaml_subset import YamlSubsetError, parse_frontmatter
 
 if TYPE_CHECKING:
-    from .manifest import ManifestAgent, ManifestBottle
+    from .manifest import ManifestBottle
 
 
 def check_stale_json(dir_path: Path, md_dir: Path, label: str) -> None:
     """Die if `<dir_path>/bot-bottle.json` exists but `md_dir` does
     not. The manifest format changed in PRD 0011 and we do not want
     to silently leave the JSON content unused."""
-    from .manifest import ManifestError
-
     legacy = dir_path / "bot-bottle.json"
     if legacy.is_file() and not md_dir.exists():
         raise ManifestError(
@@ -34,15 +32,13 @@ def check_stale_json(dir_path: Path, md_dir: Path, label: str) -> None:
         )
 
 
-def load_bottles_from_dir(bottles_dir: Path) -> dict[str, ManifestBottle]:
-    """Walk `<bottles_dir>/*.md`, parse each as a bottle, and return
-    `{name: Bottle}`. Missing dir returns an empty dict."""
-    from .manifest import ManifestError
-    from .manifest_extends import resolve_bottles
+def scan_bottle_names(bottles_dir: Path) -> list[str]:
+    """Scan `<bottles_dir>/*.md` for valid filenames and return sorted bottle names.
 
-    raws: dict[str, dict[str, object]] = {}
+    No file content is read. Invalid filenames are skipped with a warning."""
+    result: list[str] = []
     if not bottles_dir.is_dir():
-        return {}
+        return result
     for path in sorted(bottles_dir.glob("*.md")):
         name = entity_name_from_path(path)
         if name is None:
@@ -51,31 +47,17 @@ def load_bottles_from_dir(bottles_dir: Path) -> dict[str, ManifestBottle]:
                 f"[a-z][a-z0-9-]*.md (got {path.name!r})"
             )
             continue
-        try:
-            fm, _body = parse_frontmatter(path.read_text())
-        except OSError as e:
-            raise ManifestError(f"could not read {path}: {e}") from e
-        except YamlSubsetError as e:
-            raise ManifestError(f"{path}: {e}") from e
-        validate_bottle_frontmatter_keys(path, fm.keys())
-        raws[name] = fm
-    return resolve_bottles(raws)
+        result.append(name)
+    return result
 
 
-def load_agents_from_dir(
-    agents_dir: Path,
-    bottle_names: set[str],
-    *,
-    source: str,  # noqa: F841 — unused, but required by interface
-) -> dict[str, ManifestAgent]:
-    """Walk `<agents_dir>/*.md`, parse each as an agent, and return
-    `{name: Agent}`. The Markdown body becomes the agent's prompt.
-    Missing dir returns an empty dict."""
-    from .manifest import ManifestAgent, ManifestError
+def scan_agent_names(agents_dir: Path) -> dict[str, Path]:
+    """Scan `<agents_dir>/*.md` for valid filenames and return `{name: path}`.
 
-    out: dict[str, ManifestAgent] = {}
+    No file content is read. Invalid filenames are skipped with a warning."""
+    result: dict[str, Path] = {}
     if not agents_dir.is_dir():
-        return out
+        return result
     for path in sorted(agents_dir.glob("*.md")):
         name = entity_name_from_path(path)
         if name is None:
@@ -84,22 +66,47 @@ def load_agents_from_dir(
                 f"[a-z][a-z0-9-]*.md (got {path.name!r})"
             )
             continue
+        result[name] = path
+    return result
+
+
+def load_bottle_chain_from_dir(
+    bottle_name: str, bottles_dir: Path
+) -> ManifestBottle:
+    """Load `bottle_name` and its full `extends:` chain from `bottles_dir`,
+    returning the resolved ManifestBottle.
+
+    Only the files in the extends chain are read — unrelated bottle files
+    are never touched. Raises ManifestError on parse or validation failure."""
+    from .manifest_extends import resolve_bottles
+
+    raws: dict[str, dict[str, object]] = {}
+    to_load = [bottle_name]
+    while to_load:
+        name = to_load.pop()
+        if name in raws:
+            continue
+        path = bottles_dir / f"{name}.md"
+        if not path.is_file():
+            avail = ", ".join(
+                p.stem for p in sorted(bottles_dir.glob("*.md")) if p.is_file()
+            ) or "(none)"
+            raise ManifestError(
+                f"bottle '{name}' not found at {path}. "
+                f"Available: {avail}"
+            )
         try:
-            fm, body = parse_frontmatter(path.read_text())
+            fm, _body = parse_frontmatter(path.read_text())
         except OSError as e:
             raise ManifestError(f"could not read {path}: {e}") from e
         except YamlSubsetError as e:
             raise ManifestError(f"{path}: {e}") from e
-        validate_agent_frontmatter_keys(path, fm.keys())
-        # Build the dict Agent.from_dict expects. The body becomes
-        # prompt; Claude Code passthrough fields stay in fm and get
-        # ignored by Agent.from_dict (reads bottle/skills/git-gate/prompt).
-        agent_dict: dict[str, object] = {
-            "bottle": fm.get("bottle"),
-            "skills": fm.get("skills", []),
-            "prompt": body.strip(),
-        }
-        if "git-gate" in fm:
-            agent_dict["git-gate"] = fm["git-gate"]
-        out[name] = ManifestAgent.from_dict(name, agent_dict, bottle_names)
-    return out
+        validate_bottle_frontmatter_keys(path, fm.keys())
+        raws[name] = dict(fm)
+        parent = fm.get("extends")
+        if isinstance(parent, str):
+            to_load.append(parent)
+        elif isinstance(parent, list):
+            to_load.extend(p for p in parent if isinstance(p, str))
+
+    return resolve_bottles(raws)[bottle_name]

bot_bottle/manifest_schema.py

@@ -18,8 +18,8 @@ _FILENAME_RX = re.compile(r"^[a-z][a-z0-9-]*$")
 BOTTLE_KEYS = frozenset(
     {"env", "extends", "agent_provider", "git-gate", "egress", "supervise"}
 )
-AGENT_KEYS_REQUIRED = frozenset({"bottle"})
-AGENT_KEYS_OPTIONAL = frozenset({"skills", "git-gate"})
+AGENT_KEYS_REQUIRED: frozenset[str] = frozenset()
+AGENT_KEYS_OPTIONAL = frozenset({"bottle", "skills", "git-gate"})
 
 # Claude Code subagent fields bot-bottle ignores at launch but does
 # not reject. This lets the same file double as

bot_bottle/supervise.py

@@ -2,11 +2,10 @@
 
 The supervise plane is the per-bottle MCP sidecar plus its host-side
 queue/audit support. The sidecar (bot_bottle.supervise_server)
-sits on the bottle's internal network and exposes three MCP tools the
-agent calls when it hits a stuck-recovery category:
+sits on the bottle's internal network and exposes MCP tools the agent
+calls when it needs an operator-reviewed egress change:
 
-  * egress-block — agent proposes a new routes.yaml
-  * capability-block — agent proposes a new agent Dockerfile
+  * egress-block / allow — agent proposes a new routes.yaml
 
 Each tool call: the agent passes the full proposed file plus a
 justification text. The sidecar validates the proposal syntactically,
@@ -48,28 +47,35 @@ from pathlib import Path
 SUPERVISE_HOSTNAME = "supervise"
 SUPERVISE_PORT = 9100
 
-TOOL_CAPABILITY_BLOCK = "capability-block"
+TOOL_EGRESS_BLOCK = "egress-block"
+TOOL_EGRESS_ALLOW = "egress-allow"
+TOOL_GITLEAKS_ALLOW = "gitleaks-allow"
+# Written directly by the egress addon (not an agent-facing MCP tool) when an
+# outbound DLP token block is routed to the operator for override (PRD 0062).
+TOOL_EGRESS_TOKEN_ALLOW = "egress-token-allow"
 TOOL_LIST_EGRESS_ROUTES = "list-egress-routes"
 TOOLS: tuple[str, ...] = (
-    TOOL_CAPABILITY_BLOCK,
+    TOOL_EGRESS_ALLOW,
+    TOOL_EGRESS_BLOCK,
+    TOOL_GITLEAKS_ALLOW,
+    TOOL_EGRESS_TOKEN_ALLOW,
     TOOL_LIST_EGRESS_ROUTES,
 )
 
 # The supervise sidecar uses these to query egress's
 # introspection endpoint for the `list-egress-routes` MCP
 # tool. The hostname + port match egress's docker network
-# alias + listen port (see bot_bottle.egress.EGRESS_HOSTNAME
-# and backend.docker.egress.EGRESS_PORT — the values
-# are inlined here so the in-container supervise_server doesn't
-# need to import the egress package).
-EGRESS_FORWARD_PROXY = "http://egress:9099"
+# listen port (see backend.docker.egress.EGRESS_PORT). The supervise
+# daemon runs inside the sidecar bundle alongside egress, so loopback
+# is the stable address across docker, smolmachines, and Apple
+# Container backends.
+EGRESS_FORWARD_PROXY = "http://127.0.0.1:9099"
 EGRESS_INTROSPECT_URL = "http://_egress.local/allowlist"
 
-# capability-block has no on-disk config the operator edits in place
-# (the Dockerfile is rebuilt, not patched), so it has no audit log
-# here — those changes are captured by git history + the rebuild
-# record laid down in PRD 0016. egress-block was removed in issue #198.
-COMPONENT_FOR_TOOL: dict[str, str] = {}
+COMPONENT_FOR_TOOL: dict[str, str] = {
+    TOOL_EGRESS_ALLOW: "egress",
+    TOOL_EGRESS_BLOCK: "egress",
+}
 
 STATUS_APPROVED = "approved"
 STATUS_MODIFIED = "modified"
@@ -81,8 +87,6 @@ STATUSES: tuple[str, ...] = (STATUS_APPROVED, STATUS_MODIFIED, STATUS_REJECTED)
 ACTION_OPERATOR_EDIT = "operator-edit"
 
 QUEUE_DIR_IN_CONTAINER = "/run/supervise/queue"
-CURRENT_CONFIG_DIR_IN_AGENT = "/etc/bot-bottle/current-config"
-
 DEFAULT_POLL_INTERVAL_SEC = 0.5
 
 
@@ -425,59 +429,39 @@ def sha256_hex(content: str) -> str:
 # --- Sidecar plan + abstract lifecycle -------------------------------------
 
 
-# Filename of the staged Dockerfile inside the agent's read-only
-# current-config mount. The capability-block tool's description
-# points the agent at this exact path so it can read the current
-# Dockerfile and propose modifications.
-#
-# routes.yaml + allowlist used to live here too; PRD 0017 chunk 3
-# moved them behind the `list-egress-routes` MCP tool (live
-# state from egress's introspection endpoint) so the agent
-# always sees current data rather than a launch-time snapshot.
-CURRENT_CONFIG_DOCKERFILE = "Dockerfile"
-
-
 @dataclass(frozen=True)
 class SupervisePlan:
     """Output of Supervise.prepare; consumed by .start.
 
     `queue_dir` is the host directory bind-mounted into the sidecar
-    at /run/supervise/queue. `current_config_dir` is the host
-    directory bind-mounted (read-only) into the *agent* container
-    at /etc/bot-bottle/current-config — currently holds only the
-    Dockerfile snapshot (routes.yaml + allowlist moved to the
-    `list-egress-routes` MCP tool). `internal_network` is
-    empty at prepare time; the backend's launch step fills it via
-    dataclasses.replace before calling .start."""
+    at /run/supervise/queue. `internal_network` is empty at prepare
+    time; the backend's launch step fills it via dataclasses.replace
+    before calling .start."""
 
     slug: str
     queue_dir: Path
-    current_config_dir: Path
     internal_network: str = ""
 
 
 class Supervise(ABC):
     """Per-bottle supervise sidecar. Encapsulates the host-side
-    prepare (queue dir + current-config staging); the sidecar's
-    start/stop lifecycle is backend-specific."""
+    prepare (queue dir staging); the sidecar's start/stop lifecycle
+    is backend-specific."""
 
     def prepare(
         self,
         slug: str,
         stage_dir: Path,
     ) -> SupervisePlan:
-        """Stage the per-bottle queue dir on the host and the
-        current-config dir under `stage_dir`. Returns the plan;
-        `internal_network` must be set by the launch step before
+        """Stage the per-bottle queue dir on the host. Returns the
+        plan; `internal_network` must be set by the launch step before
         .start runs."""
+        del stage_dir
         queue_dir = queue_dir_for_slug(slug)
         queue_dir.mkdir(parents=True, exist_ok=True)
-        current_config_dir = stage_dir / "current-config"
-        current_config_dir.mkdir(parents=True, exist_ok=True)
         return SupervisePlan(
             slug=slug,
             queue_dir=queue_dir,
-            current_config_dir=current_config_dir,
         )
 
 # --- Helpers ---------------------------------------------------------------
@@ -528,8 +512,6 @@ __all__ = [
     "ACTION_OPERATOR_EDIT",
     "AuditEntry",
     "COMPONENT_FOR_TOOL",
-    "CURRENT_CONFIG_DIR_IN_AGENT",
-    "CURRENT_CONFIG_DOCKERFILE",
     "DEFAULT_POLL_INTERVAL_SEC",
     "Proposal",
     "QUEUE_DIR_IN_CONTAINER",
@@ -545,7 +527,10 @@ __all__ = [
     "TOOLS",
     "EGRESS_FORWARD_PROXY",
     "EGRESS_INTROSPECT_URL",
-    "TOOL_CAPABILITY_BLOCK",
+    "TOOL_EGRESS_ALLOW",
+    "TOOL_EGRESS_BLOCK",
+    "TOOL_GITLEAKS_ALLOW",
+    "TOOL_EGRESS_TOKEN_ALLOW",
     "TOOL_LIST_EGRESS_ROUTES",
     "archive_proposal",
     "audit_dir",

bot_bottle/supervise_server.py

@@ -1,8 +1,8 @@
 """Supervise sidecar HTTP server (PRD 0013).
 
-Per-bottle MCP server exposing tools the agent calls to propose config
-changes when stuck. The egress-block tool was removed in issue #198;
-the remaining tools are `capability-block` and `list-egress-routes`.
+Per-bottle MCP server exposing tools the agent calls to propose egress
+config changes when stuck. The tools are `egress-allow`,
+`egress-block`, and `list-egress-routes`.
 
 Each queued tool call:
 
@@ -44,9 +44,15 @@ import urllib.request
 from dataclasses import dataclass
 from pathlib import Path
 
-# Same-directory import inside the bundle container; `supervise.py`
-# is COPYed alongside this file by Dockerfile.sidecars.
-import supervise as _sv
+try:
+    # Same-directory imports inside the bundle container; these files are
+    # COPYed flat under /app by Dockerfile.sidecars.
+    from egress_addon_core import LOG_OFF, load_config
+    import supervise as _sv
+except ModuleNotFoundError:
+    # Package imports for host-side tests and tooling.
+    from .egress_addon_core import LOG_OFF, load_config
+    from . import supervise as _sv
 
 
 # --- JSON-RPC / MCP plumbing ----------------------------------------------
@@ -84,19 +90,19 @@ def parse_jsonrpc(body: bytes) -> JsonRpcRequest:
     try:
         raw = json.loads(body)
     except json.JSONDecodeError as e:
-        raise _RpcError(ERR_PARSE, f"parse error: {e}") from e
+        raise _RpcClientError(ERR_PARSE, f"parse error: {e}") from e
     if not isinstance(raw, dict):
-        raise _RpcError(ERR_INVALID_REQUEST, "request must be a JSON object")
+        raise _RpcClientError(ERR_INVALID_REQUEST, "request must be a JSON object")
     if raw.get("jsonrpc") != JSONRPC_VERSION:
-        raise _RpcError(ERR_INVALID_REQUEST, "jsonrpc field must be '2.0'")
+        raise _RpcClientError(ERR_INVALID_REQUEST, "jsonrpc field must be '2.0'")
     method = raw.get("method")
     if not isinstance(method, str):
-        raise _RpcError(ERR_INVALID_REQUEST, "method must be a string")
+        raise _RpcClientError(ERR_INVALID_REQUEST, "method must be a string")
     params = raw.get("params", {})
     if params is None:
         params = {}
     if not isinstance(params, dict):
-        raise _RpcError(ERR_INVALID_PARAMS, "params must be an object")
+        raise _RpcClientError(ERR_INVALID_PARAMS, "params must be an object")
     rpc_id = raw.get("id", _NO_ID)
     is_notification = rpc_id is _NO_ID
     return JsonRpcRequest(
@@ -111,12 +117,23 @@ _NO_ID = object()
 
 
 class _RpcError(Exception):
+    """Base class for all typed RPC errors that surface as JSON-RPC error responses."""
     def __init__(self, code: int, message: str):
         super().__init__(message)
         self.code = code
         self.message = message
 
 
+class _RpcClientError(_RpcError):
+    """Caller sent a bad request; returned verbatim, no server-side logging."""
+
+
+class _RpcInternalError(_RpcError):
+    """Server-side fault; logged at ERROR with cause, always returns ERR_INTERNAL."""
+    def __init__(self, message: str) -> None:
+        super().__init__(ERR_INTERNAL, message)
+
+
 def jsonrpc_result(request_id: object, result: object) -> bytes:
     payload = {"jsonrpc": JSONRPC_VERSION, "id": request_id, "result": result}
     return (json.dumps(payload) + "\n").encode("utf-8")
@@ -142,8 +159,9 @@ TOOL_DEFINITIONS: list[dict[str, object]] = [
             "allowlist. Returns JSON with one entry per allowed host, "
             "each carrying its matches rules (if any) and whether "
             "the proxy injects Authorization for the route. Use this "
-            "before composing an `egress-block` proposal so the new "
-            "routes file extends the live one rather than replacing it."
+            "before composing an `egress-allow` or `egress-block` proposal so "
+            "the new routes file extends the live one rather than "
+            "replacing it."
         ),
         "inputSchema": {
             "type": "object",
@@ -152,41 +170,97 @@ TOOL_DEFINITIONS: list[dict[str, object]] = [
         },
     },
     {
-        "name": _sv.TOOL_CAPABILITY_BLOCK,
+        "name": _sv.TOOL_EGRESS_ALLOW,
         "description": (
-            "Call when the bottle is missing a tool, skill, permission, "
-            "or env var you need — something that lives in the agent "
-            "Dockerfile rather than in the egress routes. "
-            "Read the current Dockerfile from "
-            "/etc/bot-bottle/current-config/Dockerfile, compose a "
-            "modified version, and pass the full new file plus a "
-            "justification. On approval the supervisor rebuilds the "
-            "bottle from the new Dockerfile and starts a replacement on "
-            "the same branch (wired in PRD 0016; v1 acknowledges only)."
+            "Request operator approval to change the bottle's egress "
+            "allowlist. Pass the full proposed routes.yaml content, not "
+            "just the new host, plus a justification. Use "
+            "`list-egress-routes` first so the proposal preserves existing "
+            "routes."
         ),
         "inputSchema": {
             "type": "object",
             "properties": {
-                "dockerfile": {
+                "routes_yaml": {
                     "type": "string",
-                    "description": "Full proposed Dockerfile content.",
+                    "description": (
+                        "Full proposed /etc/egress/routes.yaml content. "
+                        "Each route entry accepts these keys:\n"
+                        "  host: <hostname>  (required)\n"
+                        "  auth_scheme: Bearer|token  (must pair with token_env)\n"
+                        "  token_env: <ENV_VAR_NAME>  (must pair with auth_scheme)\n"
+                        "  matches:  (optional list of match entries)\n"
+                        "    - paths: [{type: prefix|exact|regex, value: /...}]\n"
+                        "      methods: [GET, POST, ...]\n"
+                        "      headers: [{name: X-Hdr, value: val, type: exact|regex}]\n"
+                        "  git:  (optional; omit to block git clone/fetch)\n"
+                        "    fetch: true\n"
+                        "  dlp:  (optional DLP scanner overrides)\n"
+                        "    outbound_detectors: [token_patterns, known_secrets]\n"
+                        "    inbound_detectors: [naive_injection_detection]\n"
+                        "    outbound_on_match: block|redact|supervise  (default supervise)\n"
+                        "Omit any key that should use its default. "
+                        "`list-egress-routes` returns routes in this same format."
+                    ),
                 },
                 "justification": {
                     "type": "string",
-                    "description": "Why this capability is needed.",
+                    "description": "Why this egress route is needed.",
                 },
             },
-            "required": ["dockerfile", "justification"],
+            "required": ["routes_yaml", "justification"],
+        },
+    },
+    {
+        "name": _sv.TOOL_EGRESS_BLOCK,
+        "description": (
+            "Request operator approval to change the bottle's egress "
+            "allowlist after a blocked outbound request. Pass the full "
+            "proposed routes.yaml content plus a justification. Use "
+            "`list-egress-routes` first so the proposal preserves existing "
+            "routes."
+        ),
+        "inputSchema": {
+            "type": "object",
+            "properties": {
+                "routes_yaml": {
+                    "type": "string",
+                    "description": (
+                        "Full proposed /etc/egress/routes.yaml content. "
+                        "Each route entry accepts these keys:\n"
+                        "  host: <hostname>  (required)\n"
+                        "  auth_scheme: Bearer|token  (must pair with token_env)\n"
+                        "  token_env: <ENV_VAR_NAME>  (must pair with auth_scheme)\n"
+                        "  matches:  (optional list of match entries)\n"
+                        "    - paths: [{type: prefix|exact|regex, value: /...}]\n"
+                        "      methods: [GET, POST, ...]\n"
+                        "      headers: [{name: X-Hdr, value: val, type: exact|regex}]\n"
+                        "  git:  (optional; omit to block git clone/fetch)\n"
+                        "    fetch: true\n"
+                        "  dlp:  (optional DLP scanner overrides)\n"
+                        "    outbound_detectors: [token_patterns, known_secrets]\n"
+                        "    inbound_detectors: [naive_injection_detection]\n"
+                        "    outbound_on_match: block|redact|supervise  (default supervise)\n"
+                        "Omit any key that should use its default. "
+                        "`list-egress-routes` returns routes in this same format."
+                    ),
+                },
+                "justification": {
+                    "type": "string",
+                    "description": "Why this egress route is needed.",
+                },
+            },
+            "required": ["routes_yaml", "justification"],
         },
     },
 ]
 
 
-# Map each non-egress tool to the input field that carries the agent's
-# payload (stored in Proposal.proposed_file). egress-block builds its
-# payload from structured input fields in `handle_egress_block`.
+# Map each proposal tool to the input field that carries the agent's
+# payload (stored in Proposal.proposed_file).
 PROPOSED_FILE_FIELD: dict[str, str] = {
-    _sv.TOOL_CAPABILITY_BLOCK: "dockerfile",
+    _sv.TOOL_EGRESS_ALLOW: "routes_yaml",
+    _sv.TOOL_EGRESS_BLOCK: "routes_yaml",
 }
 
 
@@ -198,13 +272,22 @@ def validate_proposed_file(tool: str, content: str) -> None:
     catches obvious paste-errors / wrong-tool selections before they
     enter the queue."""
     if not content.strip():
-        raise _RpcError(ERR_INVALID_PARAMS, f"{tool}: proposed file is empty")
-    if tool == _sv.TOOL_CAPABILITY_BLOCK:
-        # Dockerfiles are too varied to validate syntactically beyond
-        # non-empty. The operator reads the diff in the TUI.
-        pass
+        raise _RpcClientError(ERR_INVALID_PARAMS, f"{tool}: proposed file is empty")
+    if tool in (_sv.TOOL_EGRESS_ALLOW, _sv.TOOL_EGRESS_BLOCK):
+        try:
+            config = load_config(content)
+        except ValueError as e:
+            raise _RpcClientError(
+                ERR_INVALID_PARAMS,
+                f"{tool}: proposed routes.yaml is not valid: {e}",
+            ) from e
+        if config.log != LOG_OFF:
+            raise _RpcClientError(
+                ERR_INVALID_PARAMS,
+                f"{tool}: proposed routes.yaml must not change egress logging",
+            )
     else:
-        raise _RpcError(ERR_INVALID_PARAMS, f"unknown tool {tool!r}")
+        raise _RpcClientError(ERR_INVALID_PARAMS, f"unknown tool {tool!r}")
 
 
 # --- MCP handlers ----------------------------------------------------------
@@ -277,17 +360,17 @@ def handle_tools_call(
     doesn't need operator approval."""
     name = params.get("name")
     if not isinstance(name, str):
-        raise _RpcError(ERR_INVALID_PARAMS, "tools/call missing 'name'")
+        raise _RpcClientError(ERR_INVALID_PARAMS, "tools/call missing 'name'")
     if name == _sv.TOOL_LIST_EGRESS_ROUTES:
         return handle_list_egress_routes(typing.cast(dict[str, object], params.get("arguments", {})), config)
 
     args_raw = params.get("arguments", {})
     if not isinstance(args_raw, dict):
-        raise _RpcError(ERR_INVALID_PARAMS, "tools/call 'arguments' must be an object")
+        raise _RpcClientError(ERR_INVALID_PARAMS, "tools/call 'arguments' must be an object")
 
     justification = args_raw.get("justification")
     if not isinstance(justification, str) or not justification.strip():
-        raise _RpcError(
+        raise _RpcClientError(
             ERR_INVALID_PARAMS,
             f"{name}: 'justification' is required and must be a non-empty string",
         )
@@ -296,13 +379,13 @@ def handle_tools_call(
         file_field = PROPOSED_FILE_FIELD[name]
         proposed_file = args_raw.get(file_field)
         if not isinstance(proposed_file, str):
-            raise _RpcError(
+            raise _RpcClientError(
                 ERR_INVALID_PARAMS,
                 f"{name}: '{file_field}' is required and must be a string",
             )
         validate_proposed_file(name, proposed_file)
     else:
-        raise _RpcError(ERR_INVALID_PARAMS, f"unknown tool {name!r}")
+        raise _RpcClientError(ERR_INVALID_PARAMS, f"unknown tool {name!r}")
 
     proposal = _sv.Proposal.new(
         bottle_slug=config.bottle_slug,
@@ -311,7 +394,10 @@ def handle_tools_call(
         justification=justification,
         current_file_hash=_sv.sha256_hex(proposed_file),
     )
-    _sv.write_proposal(config.queue_dir, proposal)
+    try:
+        _sv.write_proposal(config.queue_dir, proposal)
+    except OSError as e:
+        raise _RpcInternalError(f"failed to write proposal to queue: {e}") from e
     sys.stderr.write(
         f"supervise: queued proposal {proposal.id} ({name}) "
         f"for bottle {config.bottle_slug}; waiting for operator...\n"
@@ -331,7 +417,10 @@ def handle_tools_call(
             "content": [{"type": "text", "text": text}],
             "isError": False,
         }
-    _sv.archive_proposal(config.queue_dir, proposal.id)
+    try:
+        _sv.archive_proposal(config.queue_dir, proposal.id)
+    except OSError as e:
+        raise _RpcInternalError(f"failed to archive proposal: {e}") from e
 
     text = format_response_text(response)
     return {
@@ -365,9 +454,8 @@ def format_pending_response_text(timeout_seconds: float) -> str:
 # --- HTTP transport --------------------------------------------------------
 
 
-# Max request body the server accepts. Generous because Dockerfile
-# proposals can be a few KB; routes.json is small. 1 MB is well above
-# any realistic config file.
+# Max request body the server accepts. 1 MB is well above any realistic
+# routes.yaml proposal.
 MAX_BODY_BYTES = 1 * 1024 * 1024
 
 
@@ -407,7 +495,7 @@ class MCPHandler(http.server.BaseHTTPRequestHandler):
 
         try:
             req = parse_jsonrpc(body)
-        except _RpcError as e:
+        except _RpcClientError as e:
             self._write_jsonrpc(jsonrpc_error(None, e.code, e.message))
             return
 
@@ -415,11 +503,19 @@ class MCPHandler(http.server.BaseHTTPRequestHandler):
 
         try:
             result = self._dispatch(req, config)
-        except _RpcError as e:
+        except _RpcClientError as e:
             self._write_jsonrpc(jsonrpc_error(req.id, e.code, e.message))
             return
-        except Exception as e:  # noqa: W0718 — catch-all for RPC dispatch errors
-            sys.stderr.write(f"supervise: internal error: {e}\n")
+        except _RpcInternalError as e:
+            cause = e.__cause__
+            detail = f": {cause}" if cause else ""
+            sys.stderr.write(f"supervise: internal error: {e.message}{detail}\n")
+            sys.stderr.flush()
+            self._write_jsonrpc(jsonrpc_error(req.id, ERR_INTERNAL, "internal error"))
+            return
+        except Exception as e:  # noqa: W0718 — unexpected errors
+            sys.stderr.write(f"supervise: unexpected error: {type(e).__name__}: {e}\n")
+            sys.stderr.flush()
             self._write_jsonrpc(jsonrpc_error(req.id, ERR_INTERNAL, "internal error"))
             return
 
@@ -438,7 +534,7 @@ class MCPHandler(http.server.BaseHTTPRequestHandler):
             return handle_tools_list(req.params)
         if method == "tools/call":
             return handle_tools_call(req.params, config)
-        raise _RpcError(ERR_METHOD_NOT_FOUND, f"method not found: {method}")
+        raise _RpcClientError(ERR_METHOD_NOT_FOUND, f"method not found: {method}")
 
     def _write_jsonrpc(self, body: bytes) -> None:
         self.send_response(200)

docs/decisions/0004-coverage-policy.md

@@ -0,0 +1,96 @@
+# ADR 0004: Risk-weighted coverage, not a single global target
+
+- **Status:** Accepted
+- **Date:** 2026-06-25
+- **Deciders:** didericis
+
+## Context
+
+bot-bottle is a security tool: it sandboxes agents, scans egress for
+secret exfiltration, strips credentials, and gates git pushes. A latent
+bug in that logic is expensive, so test coverage there genuinely
+matters. But the repo also contains code where coverage is a poor
+signal:
+
+- **Interactive entry-point shells** — `cli/init.py` (a `read_tty_line()`
+  prompt loop) and `cli/tui.py` (a curses picker). Their bodies are I/O;
+  a unit test has to fake the entire terminal conversation, so it
+  inflates the number without asserting behaviour that would otherwise
+  go unchecked.
+- **Subprocess / backend orchestration** — the docker / smolmachines /
+  macos-container backends shell out to `docker`, `container`, `smolvm`.
+  Mock-heavy unit tests here mostly re-assert the argv you already
+  wrote (the test passes whether or not the real teardown works), while
+  many of the missed *branches* are failure paths you cannot provoke
+  against a real daemon on cue.
+
+Chasing a single global percentage (e.g. 90%) pushes the most test
+effort onto the least safety-relevant code — exactly backwards — and
+invites performative tests written to colour a line rather than to catch
+a regression (Goodhart's law).
+
+## Decision
+
+Coverage is **risk-weighted**, measured over the **combined unit +
+integration** suites, with three rules:
+
+1. **Critical modules target ≥ 90%.** The security/logic core —
+   `egress_addon{,_core}.py`, `dlp_detectors.py`, `egress.py`,
+   `manifest*.py`, `git_gate.py`, `git_http_backend.py`, `supervise.py`,
+   `yaml_subset.py`, `bottle_state.py` — is Docker-independent and
+   unit-testable, so it carries the high bar. We ratchet toward 90% as
+   these modules are touched; new gaps in them are not acceptable.
+
+2. **Subprocess/backend orchestration is covered by the integration
+   suite, not omitted.** `scripts/coverage.sh` runs unit + integration
+   under one coverage measurement so these modules are scored where they
+   are actually exercised. They stay *visible* — hiding the code that
+   tears down sandboxes and wires networks is the one place we will not
+   omit.
+
+3. **Interactive entry-point shells are omitted** (`.coveragerc`), with a
+   rationale comment. This is the only sanctioned use of `omit` besides
+   `tests/*`.
+
+The forward-looking guard is a **diff-coverage gate**
+(`scripts/diff_coverage.py`): new/changed executable lines on a branch
+must be ≥ 90% covered. This catches regressions where they are
+introduced without forcing a back-fill crusade through legacy glue. The
+gate skips lines in omitted files (there is no coverage data for them),
+so the omit list cannot launder *new* logic into the dark: anything that
+needs real testing must live outside the interactive shells to be
+scored at all.
+
+The **global percentage is informational**, not a CI gate — it would
+otherwise be hostage to the CI runner's Docker availability and to the
+omit list.
+
+## Consequences
+
+- The number we report (`scripts/coverage.sh`) means "coverage of the
+  code we consider testable, across both suites" — a dip is a real
+  regression in code we control, not noise from added CLI glue.
+- No incentive to write mock-the-mock tests for orchestration to defend
+  a global figure.
+- The omit list needs governance: an entry must be a genuinely
+  interactive shell, justified in the `.coveragerc` comment and here.
+  `cli/init.py` and `cli/tui.py` qualify; backend orchestration does
+  not.
+- CI must run the integration suite under coverage to score the
+  orchestration modules; where the runner lacks Docker those tests skip
+  and their modules read low — accepted, because the *enforced* gates
+  (critical-module standard + diff coverage) are Docker-independent.
+- "We're at N%" is now a curated figure; outsiders should read the
+  policy, not just the badge.
+
+## Links
+
+- PRs #290 (cover the egress adapter), and the coverage-policy PR that
+  introduces this record.
+- `.coveragerc`, `scripts/coverage.sh`, `scripts/diff_coverage.py`.
+- `scripts/critical-modules.txt` — the single source of truth for the
+  core-module list; read by both `scripts/coverage.sh` and the
+  `update-badges.yml` "core coverage" badge so they cannot drift.
+- The README carries a `core coverage` badge (auto-updated from that
+  list) — the headline number, distinct from the informational global
+  `coverage` badge.

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

@@ -0,0 +1,159 @@
+# PRD 0060: Commit bottle state to an image
+
+- **Status:** Active
+- **Author:** Claude
+- **Created:** 2026-06-20
+- **Issue:** #194
+
+## Summary
+
+Add a `commit` CLI command that freezes a running bottle's state to a
+resumable local artifact. Docker bottles are stored as Docker images;
+smolmachines bottles are stored as `.smolmachine` artifacts. Operators
+can then resume the bottle from that exact filesystem snapshot, or
+export the artifact to migrate work to a different host.
+
+## Problem
+
+When a long-running agent session is interrupted — by a host reboot, a
+network failure, or a planned infrastructure migration — the in-progress
+container state is lost. `cli.py resume` rebuilds the agent image from
+the Dockerfile and reprovi-sions the bottle, but that returns the guest
+to its initial state, not to wherever the agent was mid-task.
+
+There is no mechanism today to capture "what's installed / configured
+inside the running container right now" and make it reproducible. The
+`capability-block` flow writes a new Dockerfile and marks the bottle for
+resume, but that only applies when the agent itself has requested a
+capability change; it doesn't help the operator who wants to take a
+snapshot before a planned host reboot or hardware migration.
+
+## Goals / Success Criteria
+
+- `./cli.py 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
+  snapshot instead of rebuilding from the Dockerfile.
+- `mark_preserved` is called so the state dir survives the normal
+  session-end cleanup.
+- A backend-specific export hint is printed so operators know how to
+  migrate the snapshot.
+- The command errors clearly on unsupported backends.
+
+## Non-goals
+
+- macOS-container backend support.
+- Automatic commit on agent exit.
+- Image push to a remote registry.
+- Storing the image tag in the manifest or sharing it between operators.
+
+## Design
+
+### Docker image tag
+
+`bot-bottle-committed-<slug>:latest` — namespaced under `bot-bottle-`
+to match existing image naming conventions; `committed` distinguishes it
+from the build-time image (`bot-bottle-claude:latest`) and the
+capability-block rebuild image (`bot-bottle-rebuilt-<identity>:latest`).
+
+### State storage
+
+A new plain-text file `committed-image` is added to the per-bottle state
+directory:
+
+```
+~/.bot-bottle/state/<identity>/
+    metadata.json
+    Dockerfile            (capability-block override; optional)
+    committed-image       (committed artifact reference; optional)
+    transcript/
+```
+
+`bottle_state.committed_image_path(identity)` returns the path.
+`write_committed_image` / `read_committed_image` are the read/write
+helpers, matching the existing `per_bottle_dockerfile` pattern. Docker
+stores a Docker tag in this file; smolmachines stores the absolute path
+to the committed `.smolmachine` artifact.
+
+### `commit` command
+
+```
+./cli.py commit [<slug>]
+```
+
+1. Resolve slug (arg or interactive picker from `enumerate_active_agents`).
+2. Check metadata and branch by backend.
+3. For Docker, derive container name `bot-bottle-<slug>` and run
+   `docker commit <container> bot-bottle-committed-<slug>:latest`.
+4. For smolmachines, derive machine name `bot-bottle-<slug>` and run
+   `smolvm pack create --from-vm <machine> -o ~/.bot-bottle/state/<slug>/committed-smolmachine`.
+5. Write the Docker image tag or smolmachine artifact path to
+   `~/.bot-bottle/state/<slug>/committed-image`.
+6. Call `mark_preserved(<slug>)` so the state dir survives session-end.
+7. Print the resume hint and a backend-specific export example.
+
+### Resume from committed image
+
+`bot_bottle/backend/docker/launch.py` already rebuilds the agent image
+at the top of the `launch` context manager. The change is a check
+immediately before that step:
+
+```python
+committed = read_committed_image(plan.slug)
+if committed and docker_mod.image_exists(committed):
+    info(f"using committed image {committed!r}")
+    plan = dataclasses.replace(
+        plan,
+        agent_provision=dataclasses.replace(plan.agent_provision, image=committed),
+    )
+else:
+    docker_mod.build_image(plan.image, _REPO_DIR, dockerfile=plan.dockerfile_path)
+```
+
+Replacing `agent_provision.image` propagates to `plan.image` (a
+property) and from there to the Compose spec renderer's `_agent_service`
+→ `image:` field, so the container boots from the committed snapshot.
+The build step is skipped entirely when a committed image is found and
+exists locally.
+
+If the committed image has been deleted from the local daemon (e.g.
+after `docker rmi` or a `docker system prune`), the launch falls back
+to a normal Dockerfile build, matching the pre-commit behavior.
+
+### Resume from committed smolmachine
+
+`bot_bottle/backend/smolmachines/launch.py` checks the committed
+reference before the normal Docker build -> pack cache path:
+
+```python
+committed = read_committed_image(plan.slug)
+if committed and Path(committed).is_file():
+    return Path(committed)
+return _ensure_smolmachine(plan.agent_image, dockerfile=plan.agent_dockerfile_path)
+```
+
+The returned path is passed to `smolvm machine create --from`, so the
+resumed VM boots from the committed snapshot. If the artifact has been
+deleted, launch falls back to the normal build and pack flow.
+
+## Testing strategy
+
+- Unit tests for `write_committed_image` / `read_committed_image` in
+  `tests/unit/test_bottle_state.py`, using the existing `_FakeHomeMixin`
+  pattern.
+- Unit tests for `commit_container` in `tests/unit/test_docker_util_image.py`,
+  mocking `subprocess.run` and asserting on the `docker commit` argv.
+- Unit tests for `cmd_commit` argument parsing, Docker commit,
+  smolmachines pack, and the unsupported backend error path, mocking
+  `enumerate_active_agents`, `commit_container`, and
+  `pack_create_from_vm`.
+- Unit tests for the launch-step committed-image branch: patch
+  `read_committed_image` to return a tag, patch `image_exists` to return
+  True, and assert that `build_image` is not called and `plan.image` is
+  overridden.
+- Unit tests for the smolmachines launch-step committed-artifact branch:
+  patch `read_committed_image` to return an existing path and assert the
+  normal `_ensure_smolmachine` path is skipped.

docs/prds/0061-gitleaks-inline-suppression-supervision.md

@@ -0,0 +1,101 @@
+# PRD 0061: Gitleaks inline suppression supervision
+
+- **Status:** Active
+- **Author:** codex
+- **Created:** 2026-06-10
+- **Issue:** #208
+
+## Summary
+
+Route every `# gitleaks:allow` inline suppression through the existing
+supervisor approval queue before git-gate forwards a push upstream. The
+annotation remains available for test fixtures and confirmed false positives,
+but it no longer lets an agent silently bypass the gate's secret scan.
+
+## Problem
+
+Gitleaks supports inline suppression comments. With the default behaviour,
+a pushed line that contains a secret-like value and `# gitleaks:allow` is
+ignored by gitleaks and accepted by git-gate. That is convenient for dummy
+credentials in tests, but it also gives an agent a one-line bypass for a real
+secret. Once the commit lands, the diff alone cannot prove whether the
+suppression was legitimate.
+
+## Goals / Success Criteria
+
+1. Git-gate continues to run the normal gitleaks scan for every incoming ref.
+2. After the normal scan passes, git-gate runs a second scan with
+   `--ignore-gitleaks-allow` and a JSON report so suppressed findings become
+   visible.
+3. If that second scan reports no suppressed findings, the push proceeds
+   unchanged.
+4. If it reports suppressed findings, git-gate creates a `gitleaks-allow`
+   supervisor proposal containing the ref, file path, line number, rule,
+   commit, and flagged line for each finding.
+5. The push proceeds only when the supervisor explicitly approves the
+   proposal; rejection, malformed responses, missing supervisor configuration,
+   and timeout all refuse the push.
+6. The supervisor TUI requires a reason when approving a `gitleaks-allow`
+   proposal, so the audit trail records whether the approval was for a test
+   fixture or a false positive.
+
+## Non-goals
+
+- Replacing gitleaks or changing the main secret-detection rule set.
+- Removing support for `# gitleaks:allow`.
+- Automatically classifying fixture files or false positives.
+- Adding new supervisor transport or authentication mechanisms.
+
+## Design
+
+### Git-gate flow
+
+`git_gate_render_hook()` emits a `supervise_gitleaks_allow` shell helper.
+For each incoming ref, git-gate first runs the existing gitleaks command. If
+that scan passes, it runs:
+
+```sh
+gitleaks git \
+  --log-opts="$log_opts" \
+  --no-banner \
+  --redact \
+  --ignore-gitleaks-allow \
+  --report-format=json \
+  --report-path="$report_file" \
+  --exit-code 0
+```
+
+The second pass keeps the push path non-interactive while producing a report
+of findings that would otherwise have been hidden by inline suppression.
+
+### Supervisor proposal
+
+When the JSON report contains findings, an embedded Python helper writes a
+proposal into `SUPERVISE_QUEUE_DIR` using the existing proposal schema. The
+proposal uses:
+
+- `tool: "gitleaks-allow"`
+- a text payload with the ref and each finding's file, line, rule, commit,
+  and redacted code line
+- a justification that tells the operator to approve only dummy test fixtures
+  or confirmed false positives
+
+Git-gate then waits for `<proposal-id>.response.json` for
+`SUPERVISE_GITLEAKS_ALLOW_TIMEOUT_SECONDS`, defaulting to 300 seconds.
+`approved` and `modified` responses allow the push; `rejected`, invalid
+responses, invalid timeout configuration, or timeout refuse it.
+
+### Supervisor UI
+
+`TOOL_GITLEAKS_ALLOW` is added to the supervisor tool registry. The curses
+supervisor renders the proposal as text and allows approval or rejection.
+Modification is unavailable for this proposal type because there is no file
+patch to apply. Approval from the TUI prompts for a non-empty reason and
+writes that reason to the response/audit path.
+
+### Tests
+
+Unit tests assert that the rendered git-gate hook includes the second gitleaks
+pass, supervisor queue fields, and fail-closed messages. Supervisor tests cover
+the new tool constant, proposal archiving, and the required TUI approval
+reason.

docs/prds/0062-egress-supervisor-token-override.md

@@ -0,0 +1,210 @@
+# PRD 0062: Supervisor override for egress token blocks
+
+- **Status:** Active
+- **Author:** claude
+- **Created:** 2026-06-24
+- **Issue:** #261
+
+## Summary
+
+Give each egress route a policy for what happens when an outbound DLP detector
+matches a token, via `dlp.outbound_on_match: block | redact | supervise`
+(default `supervise`):
+
+- **`supervise`** (default) — route the block through the existing supervisor
+  approval queue instead of returning `403` immediately. The proxy holds the
+  request open until the operator approves or rejects it. On approval the
+  matched token is added to an in-memory "safe tokens" set so the request — and
+  any later request carrying the same token — flows through without
+  re-prompting.
+- **`redact`** — scrub the matched value(s) from the request and forward it,
+  no operator in the loop. For routes where a token-shaped value is noise the
+  upstream doesn't need (telemetry/log sinks). Fails closed if a match lands on
+  a surface redaction can't rewrite (the hostname).
+- **`block`** — the original hard `403`; never overridable. For routes where a
+  detected token must always stop.
+
+The motivating goal is reducing friction from false positives without weakening
+the default-deny posture: supervise keeps a human in the loop, redact is an
+explicit per-route opt-in, and block stays available for sensitive routes.
+
+## Problem
+
+The outbound DLP detectors (`token_patterns`, `known_secrets`) are
+deliberately aggressive: any string that looks like a credential is blocked
+before it leaves the bottle. That is the right default, but it produces false
+positives — a token-shaped value that is not actually a secret, or a credential
+the agent legitimately needs to send to a declared host. Today the only
+recovery is for the operator to notice the `egress DLP` 403 in the logs and
+hand-edit the route's `dlp.outbound_detectors`, which disables the detector for
+the whole route rather than allowing the one value.
+
+The operator has no in-the-loop signal that a token block happened and no
+fine-grained way to say "this specific value is fine."
+
+## Goals / Success Criteria
+
+1. An outbound DLP **token** block (a `ScanResult` carrying a matched secret
+   value) creates a supervisor proposal instead of an immediate `403`.
+2. The egress proxy holds the blocked request open, polling for the operator's
+   response up to a bounded timeout.
+3. The proposal shows the operator the host, method, path, the detector reason,
+   and a **redacted** context snippet — never the raw token value.
+4. On `approved`/`modified`, the matched token value is added to an in-memory
+   safe-tokens set and the request proceeds normally; later requests carrying
+   the same value skip the block.
+5. On `rejected`, timeout, malformed response, or missing supervisor wiring,
+   the request fails closed with the same `403` as today.
+6. Structural blocks that carry no token value (CRLF injection) and the
+   route-not-allowlisted / git blocks are unchanged — they stay hard `403`s and
+   keep their existing agent-driven `allow` / `egress-block` MCP path.
+7. The proxy event loop is not stalled while waiting: the wait is asynchronous,
+   so other flows keep being served.
+
+## Non-goals
+
+- Persisting the safe-tokens set across egress restarts. It lives in process
+  memory only; a restart re-prompts. (The issue explicitly defers persistence.)
+- Supervising inbound (prompt-injection) blocks or WebSocket frame blocks.
+  WebSocket frames still honour the safe-tokens set for already-approved values
+  but cannot wait for approval (there is no response surface after upgrade).
+- Generalising an approved secret across encodings. The safe-tokens set matches
+  the exact value the detector found.
+- Replacing the per-route `dlp.outbound_detectors` override. That remains the
+  way to turn a detector off wholesale.
+- Making `redact` the default. Silent redaction of a true false positive
+  corrupts legitimate data, so it is opt-in per route; `supervise` (human in
+  the loop) stays the default.
+
+## Scope
+
+### In scope
+
+The minimum cut that ships, in build order:
+
+1. **Core** — `ScanResult.matched`; thread `safe_tokens` through
+   `scan_outbound` / the token detectors; `build_token_allow_payload`.
+2. **Supervise + TUI** — `TOOL_EGRESS_TOKEN_ALLOW`; TUI suffix, modify guard,
+   required approval reason.
+3. **Addon glue** — async `request`, safe-tokens set, proposal write + async
+   poll, allow/block decision; pass `safe_tokens` into the WebSocket path.
+4. **On-match policy** — `dlp.outbound_on_match` through manifest → render →
+   addon; `redact` surface scrub with fail-closed re-scan; policy dispatch in
+   the addon's outbound handler.
+5. **Tests + docs** — core/supervise/TUI/manifest/render unit tests; README
+   egress + supervisor notes.
+
+### Out of scope
+
+The deferrals enumerated under **Non-goals** — restart persistence, inbound /
+WebSocket-frame supervision, cross-encoding generalisation, replacing
+`dlp.outbound_detectors`, and making `redact` the default.
+
+## Proposed Design
+
+### New services / components
+
+A new proposal tool constant `egress-token-allow` (`TOOL_EGRESS_TOKEN_ALLOW`)
+is added to `supervise.TOOLS`, and the egress addon gains an in-memory
+safe-tokens set plus the policy-dispatch path that drives it.
+
+On an outbound block the addon dispatches on the resolved policy:
+
+- **Structural blocks always 403.** A `ScanResult` with no `matched` value
+  (CRLF injection) is a hard `403` regardless of policy — there is nothing to
+  redact or safelist.
+- **`redact`** runs `redact_tokens` over the body, non-`host` header values,
+  and path/query, then re-scans. If the re-scan is clean the (rewritten)
+  request is forwarded; if a block-severity match remains (e.g. in the
+  hostname, or a unicode-evasion token redaction can't reach) it fails closed
+  with a `403`.
+- **`block`** writes the `403` immediately.
+- **`supervise`** runs the queue-and-wait loop, falling back to `block` when
+  supervise isn't wired for the bottle.
+
+For `supervise`, the addon writes the proposal directly to
+`SUPERVISE_QUEUE_DIR` (the queue is bind-mounted into the sidecar bundle and
+shared by every daemon, exactly as git-gate's `gitleaks-allow` proposal in PRD
+0061 does). The proposal's `proposed_file` is a human-readable text payload
+built by `build_token_allow_payload`:
+
+```
+egress blocked an outbound request carrying a detected token
+host: api.example.com
+method: POST
+path: /v1/ingest
+detector: OpenAI API key found in body
+context: ...before ******** after...
+```
+
+The justification tells the operator to approve only if the value is a false
+positive or a credential the request legitimately needs. The addon then polls
+`<proposal-id>.response.json` for `EGRESS_TOKEN_ALLOW_TIMEOUT_SECONDS` (default
+300). `approved`/`modified` allow the request and add the value to the
+safe-tokens set; `rejected`, malformed responses, and timeout fail the request
+closed. The proposal + response are archived to `processed/` after a decision.
+Because the wait happens inside mitmproxy's asyncio loop, the addon's `request`
+hook is async and polls with `asyncio.sleep`, so concurrent flows are
+unaffected.
+
+### Existing code touched
+
+- **Policy threading.** `dlp.outbound_on_match` is a per-route enum threaded
+  from the bottle manifest (`manifest_egress`) through the resolved route
+  (`egress.EgressRoute`), the rendered `routes.yaml` (`egress_render_routes`),
+  and the addon's `Route` (`egress_addon_core`). Unset renders nothing and
+  resolves to `supervise` at request time. The `list-egress-routes`
+  introspection endpoint round-trips it so the agent's proposals preserve it.
+- **Provider-route default.** Agent-provider routes (the agent talking to its
+  own LLM API — `api.anthropic.com`, the Codex backend, etc.) are the worst
+  source of token-shaped false positives because the whole conversation payload
+  flows through them. `egress_routes_for_bottle` fills `outbound_on_match=redact`
+  on any provider route that doesn't set it explicitly; a provider that sets the
+  policy keeps its choice, and manifest routes are unaffected (they default to
+  `supervise`).
+- **Scanners.** `scan_outbound` (and the token detectors `scan_token_patterns`
+  / `scan_known_secrets` it calls) accept a `safe_tokens` set. A match whose
+  value is in `safe_tokens` is skipped, so an approved token no longer blocks;
+  the scanners keep searching past a safelisted match so a second, un-approved
+  secret in the same request is still caught. The WebSocket path is passed the
+  same `safe_tokens` set.
+- **Supervisor UI.** `cli/supervise.py` renders `egress-token-allow` like
+  `gitleaks-allow`: the text payload is shown, modify is unavailable (there is
+  no file patch to edit), and approval prompts for a non-empty reason recorded
+  in the response notes. There is no on-disk config diff, so — like
+  `gitleaks-allow` and `capability-block` — it writes no egress audit-log entry.
+- **Failure handling.** If `SUPERVISE_QUEUE_DIR` / `SUPERVISE_BOTTLE_SLUG` are
+  unset (supervise disabled for the bottle), the addon skips the queue and
+  returns the existing `403`. Any error writing the proposal or reading the
+  response also fails closed.
+
+### Data model changes
+
+- New per-route manifest field `dlp.outbound_on_match: block | redact |
+  supervise`, rendered into `routes.yaml` (omitted when unset).
+- `ScanResult` gains a `matched: str = ""` field carrying the raw substring the
+  detector matched. The token detectors populate it; the structural CRLF
+  detector leaves it empty. The value stays inside the egress sidecar process —
+  never written to a log line (logs use the redacted `context`) nor to the
+  proposal file.
+- Proposal text payload (above) plus `<proposal-id>.response.json` in
+  `SUPERVISE_QUEUE_DIR`, archived to `processed/` after a decision.
+- New env var `EGRESS_TOKEN_ALLOW_TIMEOUT_SECONDS` (default 300).
+
+### External dependencies
+
+None. Reuses the existing supervisor queue (`SUPERVISE_QUEUE_DIR`) and the
+mitmproxy addon framework already in the egress sidecar.
+
+## Open questions
+
+- Should `known_secrets` (provisioned `EGRESS_TOKEN_*` exfiltration) be
+  override-able at all, or only `token_patterns`? This PRD allows both —
+  approval is an explicit operator decision and the safe-tokens set matches the
+  exact found value — but a future revision could restrict `known_secrets` to
+  reject-only.
+
+## References
+
+- Issue #261
+- PRD 0061 — `gitleaks-allow` supervisor proposal pattern this reuses.

docs/prds/0063-strengthen-outbound-exfil-detection.md

@@ -0,0 +1,189 @@
+# PRD 0063: Strengthen outbound exfiltration detection
+
+- **Status:** Active
+- **Author:** claude
+- **Created:** 2026-06-25
+- **Issue:** #259
+
+## Summary
+
+The current DLP system has two detection tiers: weak (vendor token regexes)
+and strong (exact matching of provisioned `EGRESS_TOKEN_*` values in multiple
+encodings). This PRD adds four complementary hardening layers in ranked order
+of impact:
+
+1. **Canary tokens** — inject a per-session fake secret into the agent's
+   environment; any outbound appearance is a zero-false-positive exfil signal.
+2. **Broadened known-value scanning** — extend the strong tier beyond
+   `EGRESS_TOKEN_*` via a `BOT_BOTTLE_SENSITIVE_PREFIXES` sidecar env var.
+3. **Fragmentation-resistant matching** — catch separator injection and
+   long-substring leaks via alnum projection and sliding-window substring
+   search.
+4. **Entropy scoring** — warn-only flag on high-entropy blobs as a
+   complementary signal for encrypted exfil.
+
+## Problem
+
+After PRD 0056 the outbound scan covers all HTTP request surfaces (host, path,
+query, headers, body). But the content-based strong tier only matches
+`EGRESS_TOKEN_*` values and their encodings. Three gaps remain:
+
+| Gap | How an agent exploits it |
+|-----|--------------------------|
+| No canary | There is no "can't be false-positive" signal |
+| Narrow prefix filter | Sensitive values provisioned under other env prefixes (MCP keys, API keys injected via the agent provider) are invisible to `scan_known_secrets` |
+| Fragmentation blindness | Inserting `-`, space, or `\n` between secret characters turns an exact-match into a miss |
+
+## Goals / Success Criteria
+
+1. Each launched bottle has a unique canary token in the agent's environment
+   under a randomized `WORD_WORD_SECRET` env var name. The egress sidecar gets
+   the same env var and registers that exact name through
+   `BOT_BOTTLE_SENSITIVE_PREFIXES`. Any outbound appearance of the canary
+   blocks the request as a known-secret match.
+2. `scan_known_secrets` accepts a `sensitive_prefixes` parameter (default:
+   `("EGRESS_TOKEN_",)`). `scan_outbound` reads
+   `BOT_BOTTLE_SENSITIVE_PREFIXES` from `environ` and merges those prefixes
+   in, so operators can mark additional env vars as scanned values without
+   changing the manifest schema.
+3. For every secret that passes exact-match, a secondary alnum-projection pass
+   checks for the secret with all non-alphanumeric characters stripped. This
+   catches separator-injection evasion (`MY-SECRET` → body contains
+   `MY SECRET`).
+4. A sliding-window partial-match pass checks for long-enough contiguous
+   substrings of the secret's alnum projection in the text's alnum projection.
+   Any match ≥ `PARTIAL_MATCH_MIN_LEN` (12 chars) blocks with reason
+   `"partial match"`.
+5. A new `scan_entropy` detector flags outbound text windows with Shannon
+   entropy ≥ `ENTROPY_BLOCK_THRESHOLD` (5.5 bits/char) at **warn** severity
+   only. It is registered under the new detector name `"entropy"` in
+   `OUTBOUND_DETECTOR_NAMES` and disabled by default (routes must opt in).
+6. Binary request bodies are decoded via `latin-1` instead of
+   `utf-8 errors="replace"`, preserving every byte value and allowing
+   ASCII-range secrets to be found within binary payloads.
+7. All new behaviour is unit-tested; existing tests pass unchanged.
+
+## Non-goals
+
+- Rolling per-host buffer for split-across-requests detection (state in the
+  stateless addon is complex; deferred).
+- Additional vendor regexes.
+- ML / embedding-based detection.
+- Entropy-based hard blocks (warn only per the issue).
+
+## Design
+
+### Canary token flow
+
+```
+Egress.prepare()
+  canary = secrets.token_urlsafe(32)
+  canary_env = <random WORD_WORD_SECRET>
+  EgressPlan(canary=canary, canary_env=canary_env, ...)
+
+Docker compose render:
+  sidecar env: <canary_env>=<canary>
+  sidecar env: BOT_BOTTLE_SENSITIVE_PREFIXES=<canary_env>
+  agent env:   <canary_env>=<canary>      ← visible to agent as a "secret"
+
+macos-container launch: same literals added to sidecar + agent env entries
+```
+
+The sidecar uses `BOT_BOTTLE_SENSITIVE_PREFIXES` to make the random canary env
+name part of the existing `scan_known_secrets` detector without adding a
+manifest schema field.
+
+### Broadened known-value scanning
+
+`scan_known_secrets` gains a `sensitive_prefixes` parameter:
+
+```python
+def scan_known_secrets(
+    text: str,
+    *,
+    location: str = "body",
+    env: Mapping[str, str] | None = None,
+    sensitive_prefixes: tuple[str, ...] = ("EGRESS_TOKEN_",),
+) -> ScanResult | None:
+```
+
+`scan_outbound` reads `BOT_BOTTLE_SENSITIVE_PREFIXES` (comma-separated list
+of additional prefixes) from `environ` and appends them:
+
+```python
+extra = tuple(
+    p for p in environ.get("BOT_BOTTLE_SENSITIVE_PREFIXES", "").split(",") if p
+)
+sensitive_prefixes = ("EGRESS_TOKEN_",) + extra
+```
+
+`redact_tokens` receives the same treatment for consistent redaction.
+
+### Fragmentation-resistant matching
+
+A new helper `_alnum_projection(text)` strips all non-alphanumeric characters.
+`scan_known_secrets` runs two passes per secret:
+
+1. **Exact pass** — existing encoded-variant loop (unchanged).
+2. **Alnum-projection pass** — if the secret's alnum projection has ≥ 8 chars,
+   check if it appears in the text's alnum projection. Match → block with
+   `"fragmented match (separator injection)"` reason.
+3. **Partial-substring pass** — if the secret's alnum projection has ≥
+   `PARTIAL_MATCH_MIN_LEN` chars (12), slide a window of that length across the
+   secret's projection and look for each window in the text's alnum projection.
+   First match → block with `"partial match"` reason.
+
+All three passes run only for the `"known_secrets"` detector; the token-pattern
+and entropy detectors are unchanged.
+
+### Entropy scoring
+
+New public function:
+
+```python
+def scan_entropy(
+    text: str,
+    *,
+    location: str = "body",
+    window: int = ENTROPY_WINDOW,           # 64
+    threshold: float = ENTROPY_BLOCK_THRESHOLD,  # 5.5
+) -> ScanResult | None:
+```
+
+Slides a window of `window` characters across `text` in steps of `window // 2`.
+If any window's Shannon entropy exceeds `threshold`, returns a **warn**-severity
+`ScanResult`. Never blocks.
+
+`OUTBOUND_DETECTOR_NAMES` gains `"entropy"`. Routes opt in via their `dlp`
+block; entropy scanning is **off by default** to avoid false-positive noise on
+legitimate binary payloads.
+
+### Binary body handling
+
+In `scan_outbound`, the bytes → str decoding changes from:
+
+```python
+body.decode("utf-8", errors="replace")
+```
+
+to:
+
+```python
+body.decode("utf-8") if body is str else body.decode("latin-1")
+```
+
+`latin-1` is a bijective byte↔codepoint mapping; every byte value is preserved
+as its corresponding Latin-1 code point, so ASCII-range secret strings remain
+intact and `str.find` / regex still locate them correctly. The fallback from
+strict UTF-8 is tried first so valid UTF-8 bodies are decoded faithfully.
+
+## Implementation
+
+Delivered in three commits on the same branch:
+
+1. **DLP detector changes** — `_alnum_projection`, fragmentation passes,
+   `scan_entropy`, broadened `scan_known_secrets`, updated `scan_outbound` and
+   `redact_tokens`; all accompanying unit tests.
+2. **Canary injection** — `EgressPlan.canary`, `Egress.prepare()`,
+   Docker compose + macos-container backend injection.
+3. **PRD flip** — `Status: Draft → Active`.

docs/prds/0064-log-full-credential-redaction.md

@@ -0,0 +1,85 @@
+# PRD 0064: LOG_FULL egress logging credential redaction
+
+- **Status:** Active
+- **Author:** claude
+- **Created:** 2026-06-25
+- **Issue:** #257
+
+## Summary
+
+The `LOG_FULL` egress logging path (`_log_request` and `_log_response` in `egress_addon.py`) writes request/response headers and bodies to stderr without redaction and includes the sidecar-injected upstream `Authorization` header verbatim. This PR applies `redact_tokens` to header values and bodies in both log functions and strips the injected `Authorization` header from request logs entirely.
+
+## Problem
+
+`LOG_FULL` (log level 2) is intended for debugging egress traffic. When active it calls `_log_request` and `_log_response`. Both functions have two related bugs:
+
+1. **Injected `Authorization` header exposure.** `_log_request` is called *after* the sidecar injects upstream credentials (`flow.request.headers["authorization"] = decision.inject_authorization`). The full header dict — including the live credential — is serialized to stderr. Any log collector that ingests the egress container's stderr will receive the upstream bearer token in plaintext.
+
+2. **Unredacted bodies and header values.** Neither `_log_request` nor `_log_response` passes body or header values through `redact_tokens`. By contrast, `_req_ctx` (used for block/warn events) already calls `redact_tokens` on path and host. Any provisioned secret or recognized token pattern that appears in a request body, response body, or non-Authorization header value will be logged verbatim under `LOG_FULL`.
+
+These two bugs compose: an agent that enables `LOG_FULL` and simultaneously triggers a request that carries a known token gains a write path from credentials → egress logs.
+
+## Goals / Success Criteria
+
+- `_log_request` never logs the `authorization` header in any form.
+- `_log_request` applies `redact_tokens(value, env=os.environ)` to every other header value before serializing.
+- `_log_request` applies `redact_tokens(body, env=os.environ)` to the request body before logging.
+- `_log_response` applies `redact_tokens(value, env=os.environ)` to every response header value before logging.
+- `_log_response` applies `redact_tokens(body, env=os.environ)` to the response body before logging.
+- Unit tests cover each of the five cases above.
+
+## Non-goals
+
+- Redacting host or path in the full-log path (already covered by `_req_ctx` for block/warn events; `_log_request` already calls `redact_tokens` on host and path).
+- Suppressing `LOG_FULL` or adding a new log level.
+- Changing the outbound DLP scan logic.
+
+## Design
+
+### `_log_request`
+
+```python
+def _log_request(self, flow: http.HTTPFlow) -> None:
+    headers = {
+        k: redact_tokens(v, env=os.environ)
+        for k, v in flow.request.headers.items()
+        if k.lower() != "authorization"
+    }
+    body = redact_tokens(flow.request.get_text(strict=False) or "", env=os.environ)
+    sys.stderr.write(
+        json.dumps({
+            "event": "egress_request",
+            "host": redact_tokens(flow.request.pretty_host, env=os.environ),
+            "method": flow.request.method,
+            "path": redact_tokens(flow.request.path, env=os.environ),
+            "headers": headers,
+            "body": body,
+        })
+        + "\n"
+    )
+```
+
+The `authorization` key is excluded because by the time `_log_request` is called the sidecar has already injected the upstream credential (`decision.inject_authorization`). Logging it would write a live bearer token to stderr on every allowed request. There is no safe subset to log — the value is always a live credential or empty.
+
+### `_log_response`
+
+```python
+def _log_response(self, flow: http.HTTPFlow) -> None:
+    headers = {
+        k: redact_tokens(v, env=os.environ)
+        for k, v in flow.response.headers.items()
+    }
+    body = redact_tokens(flow.response.get_text(strict=False) or "", env=os.environ)
+    sys.stderr.write(
+        json.dumps({
+            "event": "egress_response",
+            "host": flow.request.pretty_host,
+            "status": flow.response.status_code,
+            "headers": headers,
+            "body": body,
+        })
+        + "\n"
+    )
+```
+
+Response headers don't carry injected credentials, so no header name is suppressed — only the values are scrubbed by `redact_tokens`.

docs/prds/0065-multi-parent-extends.md

@@ -0,0 +1,166 @@
+# PRD 0065: Multi-parent `extends:` for bottles
+
+- **Status:** Active
+- **Author:** didericis
+- **Created:** 2026-06-25
+- **Issue:** #268
+- **Extends:** PRD 0025 (`0025-bottle-extends.md`)
+
+## Summary
+
+Allow a bottle's `extends:` field to accept either a single bottle name (existing
+behavior) or a list of bottle names (new). Multiple parents are resolved
+independently and folded left-to-right into a single effective parent before the
+child is merged on top. This lets orthogonal concerns (base env, networking/egress,
+agent provider) live in separate bottles and be composed without forcing them into a
+linear chain.
+
+## Problem
+
+PRD 0025 shipped single-parent `extends:` and listed "No multi-parent inheritance"
+as a non-goal. In practice, users want to compose multiple orthogonal bottles — a
+base environment, a networking profile, and an agent-provider override — without
+creating a three-level linear chain that couples unrelated parents to each other.
+The linear chain workaround has two problems:
+
+1. **Ordering constraint.** `networking extends base` works, but then
+   `agent extends networking` can't also pick up `base` without going through
+   `networking`, coupling two unrelated concerns.
+
+2. **Quadratic duplication.** N orthogonal bottles require O(N²) chain variants
+   (one chain per permutation of applied concerns).
+
+Multi-parent `extends:` removes both constraints: each orthogonal concern stays in
+its own bottle, and the child bottle is the only place that names the combination.
+
+## Goals / Success Criteria
+
+- `extends:` accepts a list of strings in addition to a plain string.
+- Backward compat: existing single-string `extends:` is unchanged.
+- Parents are resolved left-to-right; later entries win on conflict.
+- Child wins over all parents (unchanged from PRD 0025).
+- Cycle detection covers multi-parent graphs, not just linear chains.
+- Diamond inheritance: a shared ancestor is resolved once (via the existing cache).
+- Invalid list entries (non-string, undefined bottle, self-reference) die at parse
+  with clear messages.
+- `manifest_loader.py`'s `load_bottle_chain_from_dir` enqueues all parents from a
+  list `extends:` so the resolver sees every bottle in the graph.
+
+## Non-goals
+
+- No change to the agent-vs-bottle trust boundary (PRD 0025 "Alternatives
+  considered" option 2 stays rejected).
+- No MRO / C3 linearization. Left-to-right fold is sufficient for the expected use
+  cases.
+- No preflight display of per-field provenance across multiple parents (same open
+  question as PRD 0025; remains a follow-up).
+
+## Design
+
+### Schema
+
+`extends:` now accepts either form:
+
+```yaml
+# single parent (unchanged)
+extends: base
+
+# multiple parents (new)
+extends: [base, networking]
+```
+
+Both forms are normalized to a list internally. A list with one element behaves
+identically to the string form.
+
+### Merge rules for multi-parent fold
+
+Parents are folded pairwise left-to-right before the child merge. For each step in
+the fold, the "earlier" bottle is the running accumulator and the "later" bottle is
+the next parent. Rules per field:
+
+| Field              | Fold rule                                                    |
+|--------------------|--------------------------------------------------------------|
+| `env`              | dict merge; later wins on key collision                      |
+| `git-gate.user`    | per-field overlay; later's non-empty fields win              |
+| `git-gate.repos`   | union by name; for same-name entries, later wins per-field   |
+| `egress.routes`    | concatenate (earlier first, later appended)                  |
+| `egress.log`       | later wins (last-wins)                                       |
+| `agent_provider`   | later wins (last-wins)                                       |
+| `supervise`        | later wins (last-wins)                                       |
+
+After the fold, the combined parent is merged against the child using the existing
+PRD 0025 rules (child always wins). The child's `egress.routes` appends to the
+combined parent's concatenated routes; `validate_egress_routes` runs once on the
+final merged set and catches duplicate hosts.
+
+### Algorithm
+
+```
+extends: [p1, p2, p3]
+
+fold:
+  combined = resolve(p1)
+  combined = fold_two(combined, resolve(p2))
+  combined = fold_two(combined, resolve(p3))
+
+merge:
+  result = _merge_bottles(combined, child_raw, name)
+```
+
+`fold_two(earlier, later)` applies the rules in the table above. Cycle detection
+(the `seen` tuple) is passed to each parent resolution call unchanged — if any
+parent's chain circles back to the current bottle, it is caught. The `cache` dict
+ensures a shared ancestor is only resolved once across all parents.
+
+### Error cases
+
+| Condition                              | Error message shape                                              |
+|----------------------------------------|------------------------------------------------------------------|
+| `extends` is not a string or list      | `extends must be a string or list of strings (was <type>)`       |
+| A list entry is not a string           | `extends[<i>] must be a string (was <type>)`                     |
+| A list entry names an undefined bottle | `extends '<name>' which is not defined. Available bottles: ...`  |
+| A list entry is the bottle itself      | `extends itself; remove the self-reference`                      |
+| Cycle through any parent edge          | `is in an extends cycle: <chain>`                                |
+
+## Implementation
+
+### `bot_bottle/manifest_extends.py`
+
+- `_resolve_one_bottle`: accept `str | list[str]` for `extends`; normalize to list;
+  validate each entry; for a single-entry list fall through to the existing
+  single-parent path; for multiple entries call `_fold_parents` then
+  `_merge_bottles`.
+- `_fold_parents(parent_names, raws, cache, repos_cache, seen)`: resolve each
+  parent and fold pairwise left-to-right; return `(effective_bottle,
+  effective_repos_raw)`.
+- `_fold_two_bottles(earlier, earlier_repos_raw, later, later_repos_raw)`: apply
+  the fold rules above; return `(folded_bottle, folded_repos_raw)`.
+
+### `bot_bottle/manifest_loader.py`
+
+- `load_bottle_chain_from_dir`: when `extends` is a list, enqueue all parent names
+  for loading (previously only `isinstance(parent, str)` was handled).
+
+### `tests/unit/test_manifest_extends.py`
+
+- `TestExtendsErrors.test_non_string_extends_dies`: update to use an integer
+  `extends` value (a list is now valid).
+- New class `TestExtendsMultiParent` covering all cases listed in the issue.
+
+## Testing strategy
+
+Unit tests via `ManifestIndex.from_json_obj` (same resolver surface used by all
+paths). No integration test changes needed — downstream code consumes the already-
+merged bottle and is unchanged.
+
+Test cases:
+- Two-parent list: env union, egress routes concat, git repos union
+- Last-parent-wins on scalar (supervise, agent_provider)
+- Child wins over all parents on conflict
+- Diamond: two parents share an ancestor; ancestor resolved once
+- Single-element list: identical to string form
+- Non-string extends value → ManifestError
+- Non-string list entry → ManifestError
+- Undefined bottle in list → ManifestError
+- Self-reference in list → ManifestError
+- Cycle through multi-parent edge → ManifestError

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

@@ -0,0 +1,216 @@
+# PRD 0066: Separate agent and bottle selection
+
+- **Status:** Active
+- **Author:** claude
+- **Created:** 2026-06-25
+- **Issue:** #269
+
+## Summary
+
+Agents and bottles are two separate concerns: agents carry a system prompt and
+skills; bottles carry infrastructure configuration (egress, git-gate, env,
+agent provider). Today an agent's manifest file hard-codes a single `bottle:`
+reference, which prevents the same agent prompt from being reused across
+projects that need different bottle configurations. This PRD decouples them: at
+launch time, after choosing the agent, the operator picks an ordered list of
+bottles via a multi-select picker. The selected bottles are merged in order
+(later entries override earlier ones) to produce the effective bottle for the
+session.
+
+## Problem
+
+The current `bottle: <name>` field on an agent manifest file binds the agent
+permanently to one bottle. To use the same system prompt with a different bottle
+(e.g. `claude-implementer` at home vs. at a client site that needs a different
+egress policy), the operator must duplicate the agent file and change the
+`bottle:` field. Duplicate agent files drift out of sync.
+
+## Goals / Success Criteria
+
+1. `bottle:` in an agent's frontmatter becomes optional. Existing manifests with
+   `bottle:` continue to work unchanged (backward compat).
+2. After selecting an agent (via the existing single-select picker), a new
+   multi-select bottle picker appears showing all available bottles.
+3. The multi-select picker pre-populates with the agent's `bottle:` value when
+   present.
+4. Confirming with one or more bottles selected uses those bottles, merged in
+   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`
+   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
+   selection, an ordered "Selected: ..." summary line, Ctrl-D to confirm, and
+   Esc/q to cancel the whole start operation.
+9. Unit tests cover: multi-select widget (filter, toggle, confirm, cancel),
+   the `cmd_start` bottle-picker step, and the manifest `load_for_agent`
+   runtime-bottle-merge path.
+
+## Non-goals
+
+- 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`.
+- Removing the `bottle:` key from the agent schema (it stays, now optional).
+
+## Design
+
+### `bot_bottle/cli/tui.py` — `filter_multiselect`
+
+```python
+def filter_multiselect(
+    items: list[str],
+    *,
+    title: str = "",
+    initial: list[str] | None = None,
+    tty_path: str = "/dev/tty",
+) -> list[str] | None:
+    """Multi-select variant of filter_select.
+
+    Returns the ordered list of selected items, or None on cancel.
+    Press Space/Enter to toggle the item under the cursor.
+    Press Ctrl-D to confirm. Press Esc/q to cancel.
+    """
+```
+
+Layout:
+
+```
+Select bottles
+Filter: _
+─────────────────────────────────────────
+> [*] claude
+  [ ] dev
+  [ ] codex
+─────────────────────────────────────────
+Selected (in order): claude
+─────────────────────────────────────────
+[↑↓/jk] move  [Space] toggle  [Ctrl-D] done  [Esc] cancel
+```
+
+`initial` pre-populates the ordered selection. `None` means no pre-selection.
+Items added are appended in insertion order; items removed leave the remaining
+order unchanged.
+
+### `bot_bottle/manifest_schema.py` — optional `bottle:`
+
+`bottle` moves from `AGENT_KEYS_REQUIRED` to `AGENT_KEYS_OPTIONAL`.
+
+### `bot_bottle/manifest_agent.py` — optional `bottle:`
+
+`ManifestAgent.bottle` changes from `str` (required) to `str = ""`.
+`from_dict` no longer requires the key to be present; the bottle-exists
+validation is skipped when the key is absent.
+
+### `bot_bottle/manifest_loader.py` — `scan_bottle_names`
+
+```python
+def scan_bottle_names(bottles_dir: Path) -> list[str]:
+    """Scan <bottles_dir>/*.md and return sorted bottle names."""
+```
+
+### `bot_bottle/manifest.py` — `ManifestIndex` changes
+
+**`all_bottle_names` property** — analogous to `all_agent_names`; scans
+`home_md / "bottles"` in lazy mode, returns `sorted(self.bottles.keys())` in
+eager mode.
+
+**`load_for_agent(agent_name, bottle_names: tuple[str, ...] = ())`** — new
+`bottle_names` parameter. When non-empty, the listed bottles are resolved and
+merged in order (index 0 is the base; each subsequent bottle is applied on top
+using the same field-merge rules as `extends:`). The result replaces the bottle
+that `agent.bottle` would have provided. When empty, falls back to `agent.bottle`.
+Raises ManifestError if neither `bottle_names` nor `agent.bottle` is set.
+
+### `bot_bottle/manifest_extends.py` — `merge_bottles_runtime`
+
+```python
+def merge_bottles_runtime(bottles: list[ManifestBottle]) -> ManifestBottle:
+    """Merge an ordered list of pre-resolved ManifestBottle objects.
+
+    Index 0 is the base; each subsequent entry overrides the previous using
+    the same rules as the file-based extends machinery:
+      - env: dict merge, later wins
+      - git_user: per-field overlay, later wins on non-empty
+      - git (repos): union by name, later wins per-name
+      - egress.routes: concatenate
+      - agent_provider, supervise: later bottle's value replaces earlier
+    """
+```
+
+This function operates on already-parsed `ManifestBottle` objects, so it does
+not need to touch the raw-dict path.
+
+### `bot_bottle/backend/__init__.py` — `BottleSpec` + `_validate`
+
+`BottleSpec` gains `bottle_names: tuple[str, ...] = ()`.
+
+`BottleBackend._validate` passes `spec.bottle_names` to `load_for_agent`:
+
+```python
+manifest = spec.manifest.load_for_agent(spec.agent_name, spec.bottle_names)
+```
+
+The preflight print updates `info(f"bottle: {agent.bottle}")` to display the
+effective bottle name(s). When `spec.bottle_names` is non-empty those are
+shown; when empty and `agent.bottle` is set, the agent's `bottle:` is shown.
+
+### `bot_bottle/bottle_state.py` — persist bottle names
+
+`BottleMetadata` gains `bottle_names: tuple[str, ...] = ()`. `read_metadata`
+reads this from JSON (default `()`). `write_launch_metadata` passes
+`spec.bottle_names` through.
+
+### `bot_bottle/cli/start.py` — bottle multiselect step
+
+After agent selection, before the name/color modal:
+
+```python
+available_bottle_names = manifest.all_bottle_names
+# Peek at agent's bottle default for pre-population
+initial_bottle = _peek_agent_bottle(manifest, agent_name)
+initial = [initial_bottle] if initial_bottle else []
+
+bottle_names_list = tui.filter_multiselect(
+    available_bottle_names,
+    title="Select bottles",
+    initial=initial,
+)
+if bottle_names_list is None:
+    return 0  # user cancelled
+bottle_names = tuple(bottle_names_list)
+```
+
+`_peek_agent_bottle` reads the agent file's frontmatter without full parsing,
+returning the `bottle:` value or `""` when absent.
+
+`BottleSpec` is built with `bottle_names=bottle_names`.
+
+### `bot_bottle/cli/resume.py` — bottle names from metadata
+
+```python
+spec = BottleSpec(
+    ...
+    bottle_names=tuple(metadata.bottle_names),
+)
+```
+
+## Implementation chunks
+
+1. **Schema + model** — `manifest_schema.py`, `manifest_agent.py` (optional
+   `bottle:`), `manifest_loader.py` (`scan_bottle_names`), `manifest.py`
+   (`all_bottle_names`, `load_for_agent` signature), `manifest_extends.py`
+   (`merge_bottles_runtime`), `bottle_state.py` (`bottle_names` field),
+   `resolve_common.py` (thread through).
+2. **Backend** — `BottleSpec.bottle_names`, `_validate`, preflight print.
+3. **TUI** — `filter_multiselect` in `tui.py` + unit tests.
+4. **CLI wiring** — `start.py` bottle picker step, `resume.py` metadata load.
+5. **Tests** — `test_cli_start_selector.py` bottle-picker cases,
+   `test_manifest_agent.py` optional-bottle cases, new
+   `test_manifest_bottle_merge.py` for `merge_bottles_runtime`.
+
+## Open questions
+
+None.

docs/prds/prd-new-egress-control-plane.md

@@ -0,0 +1,247 @@
+# PRD prd-new: Egress control plane — metering, budgets, and forced cutoff
+
+- **Status:** Draft
+- **Author:** didericis
+- **Created:** 2026-06-25
+- **Issue:** #251
+
+## Summary
+
+Add an **out-of-band egress enforcement & observability plane**: meter every
+agent's token usage at the egress proxy, decrement budgets without the agent's
+cooperation, and forcibly cut a bottle's egress when a budget is exhausted —
+either automatically or on command from a host-level dashboard. The trigger
+(usage threshold) and the action (route-drop / freeze / kill) both live in the
+egress plane and run with no agent in the loop. This is distinct from the
+supervise sidecar (PRD 0013), which is agent-initiated and therefore cannot
+enforce a cost cutoff on a runaway agent. State (usage ledger, budgets, audit)
+moves into a host-level SQLite database behind a thin repository API, the first
+SQL store in an otherwise flat-file repo.
+
+## Problem
+
+bot-bottle can't currently do two things the cost-overrun case demands:
+
+1. **Forced egress shutdown on limit.** When an agent crosses a token
+   threshold, kill its egress automatically — no human in the loop.
+2. **Remote (host-level) management.** Drive agents from a single surface:
+   see usage, cut egress, stop bottles, to prevent cost overruns.
+
+The existing supervise sidecar (PRD 0013) is **entirely agent-initiated**: every
+action begins with the agent voluntarily calling an MCP tool and an operator
+approving it. A runaway or expensive agent — exactly the cost-overrun case —
+will never call `egress-block` on itself. Supervision is therefore a
+**collaborative recovery** mechanism, not an **enforcement** mechanism; making
+it mandatory (#249) would not deliver forced cost-cutoff.
+
+The requirement forces a distinction the current design blurs:
+
+- **Plane A — enforcement / observability (this PRD).** System → infrastructure.
+  Meter usage, cut egress on threshold or command, account for cost.
+  Out-of-band; independent of the agent. **Unconditional** — an enforcement
+  plane you can opt out of isn't enforcement.
+- **Plane B — agent-facing recovery (the existing supervise sidecar).**
+  Agent → operator, approval-gated. Useful interactively; meaningless for a
+  headless agent with no operator watching its queue. Remains optional.
+
+This PRD builds Plane A. It reframes the "always-on control" invariant of #249
+as "the egress control plane is always present" — a more defensible property
+than "every agent runs the agent-facing supervisor." Unsupervised
+(headless/CI/ephemeral) agents stay first-class: still subject to the mandatory
+meter + kill switch, they simply lack the agent-facing proposal tools they
+couldn't use anyway.
+
+## Goals / Success Criteria
+
+- The egress proxy meters every request to a metered API host (e.g.
+  `api.anthropic.com`) and records authoritative token usage per bottle and per
+  agent provider, with no agent cooperation.
+- A budget can be set at four scopes with deterministic precedence
+  (**agent → bottle → parent bottle → global host budget**); the
+  most-specific applicable budget governs.
+- When usage crosses a budget, the bottle's configured **cutoff policy**
+  (`cutoff` | `freeze` | `kill`) fires automatically, executed host-side on the
+  egress plane — never via the supervise queue.
+- An operator can, from a single **host-level TUI dashboard**, see live per-bottle
+  usage against budget and command a cutoff/stop on demand.
+- Host budgets, default cutoff policy, and per-provider limits are declared in a
+  new host-level `~/.bot-bottle/settings.yml`, parseable by `yaml_subset.py`.
+- All usage, budget state, and enforcement actions persist in a host-level
+  SQLite DB behind a thin repository API, so the store can later be swapped for
+  a cross-host cloud service.
+
+## Non-goals
+
+- **Remote control / cross-host control plane.** Web + mobile remote control,
+  cross-host budgets, and the authn/transport they require are explicitly
+  deferred. v1 is a **host-only TUI** with no remote surface.
+- **Dollar-denominated budgets.** Budgets are token counts keyed by agent
+  provider, not currency. Price tables are out of scope.
+- **Migrating existing flat-file state into SQLite.** Resume `metadata.json`,
+  transcripts, Dockerfile overrides, the supervise queue, and audit logs stay on
+  the filesystem. Only the *new* metering/budget/enforcement ledger is SQL.
+- **Making the supervise sidecar (Plane B) mandatory.** Out of scope here; this
+  PRD is the answer to "what should be unconditional" (Plane A), leaving #249's
+  Plane-B question open.
+- **Per-request hard pre-send blocking as the primary mechanism.** The gate is
+  budget-crossing detected at/after metering; a pre-flight estimator (below) is a
+  refinement, not the core enforcement path.
+
+## Design
+
+### Two measurements: gate vs. account
+
+There are two distinct needs, and they want different signals:
+
+- **Account (authoritative).** Decrement the real budget from the API
+  **response**, which already carries authoritative usage (Anthropic
+  `input_tokens` / `output_tokens`, OpenAI `usage`). The egress addon already
+  has a `response(flow)` hook (`bot_bottle/egress_addon.py:460`), so the real
+  number is available with no extra network call. **Caveat:** agent traffic is
+  mostly streaming SSE, so the response path must tail the stream for the final
+  usage event rather than parse a single JSON body — scoped explicitly as work.
+- **Gate (estimate).** To block *before* sending, only the request is available,
+  so an estimator / provider `count_tokens` endpoint is the only option.
+
+Calling `count_tokens` for accounting would be both less accurate *and* an extra
+metered egress call per request, so accounting uses response `usage` and the
+estimator is reserved for the optional pre-flight gate.
+
+### `count_tokens` on agent providers
+
+Add an abstract `count_tokens(request) -> int` to the `AgentProvider`
+abstraction (`bot_bottle/agent_provider.py`):
+
+- **Default** is a good-enough stdlib estimator. Prefer stdlib only; a small
+  pip dependency *for the sidecar* is acceptable for the fallback if stdlib
+  proves too inaccurate (this does not relax the package's stdlib-first stance —
+  it would be a sidecar-only dep, like the bundle already carries).
+- **Built-in `claude`** uses Anthropic's token-counting endpoint;
+  **built-in `codex`** uses OpenAI's. These are exact for the gate but cost a
+  metered call, so they are gate-only; accounting still comes from the response.
+
+### Budgets and precedence
+
+Budgets are token counts keyed by **agent provider name** (the same names
+bottles already use). Four scopes, most-specific wins:
+
+```
+agent  →  bottle  →  parent bottle  →  global (host)
+```
+
+The global host budget is the highest-priority feature to ship (the cross-host
+control plane will eventually consume it); per-agent and per-bottle budgets
+override it for finer control. A budget can also be supplied **at bottle
+launch** (`--budget` or equivalent), overriding the settings.yml defaults for
+that run. Enforcement evaluates the effective budget as the
+nearest-defined scope at decrement time.
+
+### `~/.bot-bottle/settings.yml`
+
+New **host-level** settings file (the `~/.bot-bottle/` root, *not* the per-repo
+`.bot-bottle/` — host budgets must not be committed per-repo). Parsed by
+`yaml_subset.py`, so it must stay within that bounded subset (flat mappings,
+scalars; no anchors, no multi-line block scalars). Shape:
+
+```yaml
+budget:
+  claude: 5000000      # token budget keyed by agent provider
+  codex: 2000000
+shutdown: cutoff       # default cutoff policy: cutoff | freeze | kill
+```
+
+### Forced cutoff and cutoff policy
+
+On budget exhaustion (or an operator command), the configured per-bottle cutoff
+policy fires. The three policies map onto primitives that already exist:
+
+- **`cutoff`** (default) — drop the bottle's `routes.yaml` to empty and reload
+  (or isolate the bottle from the egress network); the agent/bottle keeps
+  running but can no longer reach metered hosts. This is the route-drop already
+  available on the egress plane (`bot_bottle/backend/egress_apply.py`).
+- **`freeze`** — commit/snapshot state, then kill the agent/bottle; resumable
+  later via `bot_bottle/backend/freeze.py`.
+- **`kill`** — tear the bottle down without saving state (backend teardown).
+
+The trigger lives in the metering path and the action in the egress/backend
+plane; **neither touches the supervise proposal queue** (design constraint from
+#251).
+
+### Host-level SQLite store
+
+**Decision: introduce SQLite now, narrowly.**
+
+- **The dependency objection doesn't apply.** `sqlite3` is in the Python stdlib,
+  so it does not break the AGENTS.md stdlib-first / no-runtime-pip stance — same
+  category as the hand-rolled `yaml_subset.py`, except the stdlib already ships
+  the whole engine.
+- **It fits the problem.** A *global* token budget decremented concurrently by N
+  egress sidecars (today `~/.bot-bottle/` already has `state/`, `audit/`,
+  `queue/` written by parallel bottles) is a read-modify-write race. Over JSON
+  that means hand-rolled file locking; SQLite gives atomic transactions + WAL for
+  free. The per-agent/per-bottle precedence rollup plus "sum across all bottles"
+  is a `GROUP BY`, not an N-directory rescan.
+- **It rehearses the cloud swap.** "Wrap operations in an API so we can swap to a
+  cloud service" maps directly onto a thin repository/DAO over SQLite → Postgres
+  later. A JSON-file store is a worse rehearsal than SQL.
+
+**Costs (real but bounded):** a new paradigm in a flat-file repo needs a
+`schema_version` table + idempotent startup migrations; SQLite serializes
+writers, so WAL mode + `busy_timeout` are required (a non-issue at a handful of
+bottles); test fixtures need temp DBs.
+
+**Scope of the store:** one DB at `~/.bot-bottle/bot-bottle.db` behind a thin
+repository API. Only the **new** metering/budget/enforcement-audit ledger lives
+there. Existing per-bottle blobs (resume `metadata.json`, transcripts,
+Dockerfile overrides, supervise queue) stay on the filesystem — migrating them
+now is churn for no benefit and they lack the concurrency/aggregation problem.
+
+### Host-level controller + dashboard
+
+A single **host-level controller** owns the meter, budget evaluation, and the
+cutoff actions across all bottles (cf. `bot_bottle/cli/supervise.py`'s
+cross-bottle view), rather than a per-bottle daemon. v1 ships one host-level
+**TUI dashboard** that reads live usage-vs-budget from the SQLite store and
+offers on-demand cutoff/stop. The existing supervisor UI should eventually fold
+into this same dashboard; this PRD lays the host-level surface it will move to.
+
+## Implementation chunks
+
+Ordered, individually mergeable:
+
+1. **SQLite repository foundation.** `~/.bot-bottle/bot-bottle.db`, schema +
+   `schema_version` migrations, WAL + `busy_timeout`, thin repository API,
+   temp-DB test fixtures. No behavior wired yet.
+2. **Metering at the egress proxy.** Parse authoritative response `usage`
+   (including SSE final-usage tailing) in the egress addon `response` hook;
+   write per-bottle / per-provider usage rows to the ledger.
+3. **`settings.yml` + budget model.** Host-level `~/.bot-bottle/settings.yml`
+   parsed by `yaml_subset.py`; budget precedence (agent → bottle → parent →
+   global) and the `--budget` launch flag.
+4. **Forced cutoff + cutoff policy.** Wire the threshold trigger to the
+   `cutoff` / `freeze` / `kill` primitives on the egress/backend plane; record
+   enforcement actions to the audit ledger.
+5. **Host-level TUI dashboard.** Live usage-vs-budget view + on-demand
+   cutoff/stop, reading the store.
+6. **`count_tokens` pre-flight gate (optional refinement).** Abstract method +
+   stdlib estimator default; Anthropic/OpenAI endpoints for built-in
+   claude/codex; optional pre-send block.
+
+## Open questions
+
+- **SSE usage tailing robustness.** Buffering streamed responses to extract the
+  final usage event without breaking the agent's own stream consumption — how
+  much of the body must the addon hold, and what's the failure mode if the
+  stream is interrupted mid-flight?
+- **Crossing mid-request.** A single response can push usage past budget only
+  *after* it's already been delivered. Is post-hoc cutoff (next request blocked)
+  sufficient, or is a pre-flight estimator gate (chunk 6) required for v1?
+- **Provider name ↔ metered host mapping.** How does the proxy attribute a
+  flow to an agent-provider budget key — by destination host, by bottle
+  identity, or both?
+- **Parent-bottle budget semantics.** For `bottle extends` (PRD 0025 / 0065)
+  chains, does "parent bottle" mean the manifest parent, the launching bottle,
+  or the full ancestry summed?
+- **Dashboard ↔ controller transport (even host-only).** In-process, a local
+  socket, or polling the SQLite store directly? Picks the seam the future remote
+  control plane will extend.

docs/research/local-vs-remote-agent-execution.md

@@ -22,7 +22,7 @@ escapes**, and **whether credentials are short-lived and scoped**.
 - Outbound: Docker containers have full internet access by default; no egress monitoring on most home networks
 - Lateral movement: compromised container can reach the LAN — NAS, other machines, internal services
 - Notable: CVE-2025-59536 (CVSS 8.7, Feb 2026) — a poisoned `.claude/settings.json` in a repo gives RCE when Claude Code opens it. `--dangerously-skip-permissions` removes the last gate.
-- Supply chain: MCP servers, skills, and npm packages pulled during agent execution. ~20% of ClawHub skills were found malicious in early 2026.
+- Supply chain: MCP servers, skills, and npm packages pulled during agent execution. A Jan 2026 large-scale empirical study of a 98,380-skill snapshot confirmed 157 malicious skills, ~71% of them credential harvesters. Exfiltration was overwhelmingly naive — plaintext HTTP to hardcoded endpoints; under 10% used any code obfuscation, and concealment was mostly at the documentation level, not the code level. ([Malicious Agent Skills in the Wild](https://arxiv.org/html/2602.06547v1), arXiv:2602.06547)
 
 **What local topology protects:**
 - No inbound attack surface — nothing listening on a public port

examples/bottles/claude.md

@@ -1,14 +1,14 @@
 ---
 agent_provider:
   template: claude
-
-egress:
-  routes:
-    - host: api.anthropic.com
-      role: claude_code_oauth
-      auth:
-        scheme: Bearer
-        token_ref: BOT_BOTTLE_CLAUDE_OAUTH_TOKEN
+  # auth_token names the host env var holding the Claude OAuth token. The
+  # provider injects a provider-owned api.anthropic.com egress route that
+  # re-injects this token as the Bearer header; the agent only ever sees a
+  # placeholder CLAUDE_CODE_OAUTH_TOKEN. DLP defaults (token_patterns,
+  # known_secrets outbound; naive_injection_detection inbound) apply to
+  # that route. To scan additional hosts, declare them under egress.routes
+  # with per-route matches/dlp (see README "Egress route fields").
+  auth_token: BOT_BOTTLE_CLAUDE_OAUTH_TOKEN
 ---
 
 Common Claude provider boundary. Drop this file into

requirements-dev.txt

@@ -4,3 +4,4 @@
 
 pylint>=3.0.0
 pyright>=1.1.300
+coverage>=7.0.0

scripts/coverage.sh

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+# Combined unit + integration coverage (see docs/decisions/0004-coverage-policy.md).
+#
+# Runs the unit suite, then appends the integration suite (which skips
+# cleanly when Docker / the backend CLIs are unavailable), and prints one
+# combined report. The integration suite is what scores the subprocess /
+# backend orchestration modules, so the number here is the policy's
+# yardstick — not the unit-only badge.
+#
+# Usage:
+#   scripts/coverage.sh            # combined report
+#   scripts/coverage.sh critical   # also report just the critical modules
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+PY="${PYTHON:-python3}"
+
+# Critical security/logic core held to the high bar by ADR 0004. The list
+# lives in one place (scripts/critical-modules.txt) so this report and the
+# README "core coverage" badge can't drift; comma-join it for --include.
+CRITICAL=$(grep -vE '^[[:space:]]*(#|$)' scripts/critical-modules.txt | paste -sd, -)
+
+rm -f .coverage
+
+echo "== unit ==" >&2
+"$PY" -m coverage run -m unittest discover -t . -s tests/unit
+
+echo "== integration (skips without Docker) ==" >&2
+"$PY" -m coverage run --append -m unittest discover -t . -s tests/integration
+
+echo "== combined report ==" >&2
+"$PY" -m coverage report -m
+
+if [ "${1:-}" = "critical" ]; then
+    echo "== critical modules (ADR 0004 target: 90%) ==" >&2
+    "$PY" -m coverage report --include="$CRITICAL"
+fi

scripts/critical-modules.txt

@@ -0,0 +1,23 @@
+# Critical security/logic core held to the >=90% coverage bar by
+# docs/decisions/0004-coverage-policy.md.
+#
+# SINGLE SOURCE OF TRUTH: scripts/coverage.sh (the `critical` report) and
+# .gitea/workflows/update-badges.yml (the "core coverage" badge) both read
+# this file. Add a module here when it becomes part of the core; a coverage
+# number that silently stops measuring a module is worse than no badge.
+#
+# One module path per line, relative to the repo root. Blank lines and
+# `#` comments are ignored.
+bot_bottle/egress_addon.py
+bot_bottle/egress_addon_core.py
+bot_bottle/dlp_detectors.py
+bot_bottle/egress.py
+bot_bottle/manifest.py
+bot_bottle/manifest_egress.py
+bot_bottle/manifest_agent.py
+bot_bottle/manifest_schema.py
+bot_bottle/git_gate.py
+bot_bottle/git_http_backend.py
+bot_bottle/supervise.py
+bot_bottle/yaml_subset.py
+bot_bottle/bottle_state.py

scripts/diff_coverage.py

@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+"""Diff-coverage gate (see docs/decisions/0004-coverage-policy.md).
+
+Fails if too few of the *added/changed* executable lines on this branch
+are covered. Stdlib-only by design — the project carries no runtime deps
+and we are not adding `diff-cover` to satisfy a check.
+
+Reads coverage data already produced by a `coverage run` (e.g. via
+`scripts/coverage.sh`): it shells out to `coverage json` for per-line
+data and to `git diff` for the changed lines. Lines in omitted files
+(the interactive shells) have no coverage data and are skipped, by
+policy.
+
+Usage:
+    scripts/coverage.sh                 # produce .coverage first
+    python3 scripts/diff_coverage.py    # gate against origin/main, min 90%
+    python3 scripts/diff_coverage.py --base main --min 85
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+_HUNK_RE = re.compile(r"^@@ -\d+(?:,\d+)? \+(\d+)(?:,(\d+))? @@")
+
+
+def _run(cmd: list[str]) -> str:
+    return subprocess.run(
+        cmd, check=True, capture_output=True, text=True,
+    ).stdout
+
+
+def added_lines_by_file(base: str) -> dict[str, set[int]]:
+    """Map each changed .py file to the set of line numbers added/changed
+    relative to `base`, parsed from a zero-context unified diff."""
+    diff = _run(["git", "diff", "--unified=0", f"{base}...HEAD", "--", "*.py"])
+    out: dict[str, set[int]] = {}
+    current: str | None = None
+    new_line = 0
+    for line in diff.splitlines():
+        if line.startswith("+++ b/"):
+            current = line[6:]
+            out.setdefault(current, set())
+            continue
+        hunk = _HUNK_RE.match(line)
+        if hunk:
+            new_line = int(hunk.group(1))
+            continue
+        if current is None:
+            continue
+        if line.startswith("+") and not line.startswith("+++"):
+            out[current].add(new_line)
+            new_line += 1
+        elif line.startswith("-") and not line.startswith("---"):
+            # Deletion: does not advance the new-file cursor.
+            continue
+    return out
+
+
+def coverage_json() -> dict[str, object]:
+    """Render the existing .coverage data to JSON and load it."""
+    with tempfile.NamedTemporaryFile("r", suffix=".json", delete=True) as fh:
+        _run([sys.executable, "-m", "coverage", "json", "-o", fh.name])
+        return json.load(open(fh.name, encoding="utf-8"))
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--base", default="origin/main",
+                    help="git ref to diff against (default: origin/main)")
+    ap.add_argument("--min", type=float, default=90.0,
+                    help="minimum %% of changed executable lines covered")
+    args = ap.parse_args()
+
+    if not Path(".coverage").exists():
+        print("diff-coverage: no .coverage data; run scripts/coverage.sh first",
+              file=sys.stderr)
+        return 2
+
+    added = added_lines_by_file(args.base)
+    files = coverage_json().get("files", {})
+    if not isinstance(files, dict):
+        files = {}
+
+    total = 0
+    covered = 0
+    misses: list[str] = []
+    for path, lines in sorted(added.items()):
+        info = files.get(path)
+        if not isinstance(info, dict):
+            # Omitted file or not measured (e.g. a test file) — skip by policy.
+            continue
+        executed = set(info.get("executed_lines", []))
+        missing = set(info.get("missing_lines", []))
+        executable = lines & (executed | missing)
+        for ln in sorted(executable):
+            total += 1
+            if ln in executed:
+                covered += 1
+            else:
+                misses.append(f"{path}:{ln}")
+
+    if total == 0:
+        print("diff-coverage: no measured changed lines to check — pass")
+        return 0
+
+    pct = 100.0 * covered / total
+    print(f"diff-coverage: {covered}/{total} changed lines covered ({pct:.1f}%)")
+    if misses:
+        print("uncovered changed lines:", file=sys.stderr)
+        for m in misses:
+            print(f"  {m}", file=sys.stderr)
+    if pct + 1e-9 < args.min:
+        print(f"diff-coverage: below {args.min:.0f}% threshold", file=sys.stderr)
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tests/fixtures.py

@@ -10,7 +10,7 @@ import tempfile
 from pathlib import Path
 from typing import Any, Callable
 
-from bot_bottle.manifest import Manifest
+from bot_bottle.manifest import ManifestIndex
 
 
 def fixture_minimal_dict() -> dict[str, Any]:
@@ -62,16 +62,16 @@ def fixture_with_git_dict() -> dict[str, Any]:
     }
 
 
-def fixture_minimal() -> Manifest:
-    return Manifest.from_json_obj(fixture_minimal_dict())
+def fixture_minimal() -> ManifestIndex:
+    return ManifestIndex.from_json_obj(fixture_minimal_dict())
 
 
-def fixture_with_egress() -> Manifest:
-    return Manifest.from_json_obj(fixture_with_egress_dict())
+def fixture_with_egress() -> ManifestIndex:
+    return ManifestIndex.from_json_obj(fixture_with_egress_dict())
 
 
-def fixture_with_git() -> Manifest:
-    return Manifest.from_json_obj(fixture_with_git_dict())
+def fixture_with_git() -> ManifestIndex:
+    return ManifestIndex.from_json_obj(fixture_with_git_dict())
 
 
 def write_fixture(fn: Callable[[], dict[str, Any]]) -> Path:

tests/integration/test_macos_container_launch.py

@@ -29,7 +29,7 @@ from bot_bottle.backend.macos_container.util import (
     dns_server as _container_dns_server,
     is_available as _container_available,
 )
-from bot_bottle.manifest import Manifest
+from bot_bottle.manifest import ManifestIndex
 
 
 _AGENT_PROMPT = "You are a launch smoke-test agent. Be brief."
@@ -52,8 +52,8 @@ def _minimal_agent_dockerfile(path: Path) -> None:
     )
 
 
-def _minimal_manifest(dockerfile: Path) -> Manifest:
-    return Manifest.from_json_obj({
+def _minimal_manifest(dockerfile: Path) -> ManifestIndex:
+    return ManifestIndex.from_json_obj({
         "bottles": {
             "dev": {
                 "agent_provider": {

tests/integration/test_sandbox_escape.py

@@ -31,7 +31,7 @@ from pathlib import Path
 
 from bot_bottle.backend import BottleSpec, get_bottle_backend
 from bot_bottle.bottle_state import cleanup_state
-from bot_bottle.manifest import Manifest
+from bot_bottle.manifest import ManifestIndex
 from tests._docker import skip_unless_docker
 
 
@@ -92,17 +92,16 @@ class TestSandboxEscape(unittest.TestCase):
                     "on PATH: curl -sSL https://smolmachines.com/install.sh | sh"
                 )
 
-        # Throwaway "identity file" so the manifest's _validate_git_entries
-        # passes (it only checks `os.path.isfile`, not that the content is
-        # a real SSH key). Test 5 reaches gitleaks before any SSH attempt
-        # anyway.
+        # Throwaway static key for the git-gate fixture. It need not
+        # be a real SSH key: test 5 reaches gitleaks before any SSH
+        # attempt anyway.
         fd, kp = tempfile.mkstemp(prefix="sandbox-test-key.")
         os.close(fd)
         cls._key_path = Path(kp)
         cls._key_path.write_text("placeholder\n")
         cls._key_path.chmod(0o600)
 
-        manifest = Manifest.from_json_obj({
+        manifest = ManifestIndex.from_json_obj({
             "bottles": {
                 "dev": {
                     # Three fake secrets — different shapes — land
@@ -124,7 +123,10 @@ class TestSandboxEscape(unittest.TestCase):
                     "git-gate": {"repos": {
                         "throwaway": {
                             "url": "ssh://git@unreachable.invalid:22/throwaway.git",
-                            "identity": str(cls._key_path),
+                            "key": {
+                                "provider": "static",
+                                "path": str(cls._key_path),
+                            },
                         },
                     }},
                 },

tests/integration/test_sidecar_bundle_compose.py

@@ -22,15 +22,15 @@ from pathlib import Path
 from unittest.mock import patch
 
 from bot_bottle.backend import BottleSpec, get_bottle_backend
-from bot_bottle.manifest import Manifest
+from bot_bottle.manifest import ManifestIndex
 from tests._docker import skip_unless_docker
 
 
-def _manifest() -> Manifest:
+def _manifest() -> ManifestIndex:
     """Bottle with supervise on so the bundle exercises egress +
     supervise. Git is off because a meaningful git-gate test needs
     a real upstream and SSH keys — out of scope for a bundle smoke."""
-    return Manifest.from_json_obj({
+    return ManifestIndex.from_json_obj({
         "bottles": {
             "dev": {
                 "supervise": True,

tests/integration/test_smolmachines_launch.py

@@ -35,15 +35,15 @@ from pathlib import Path
 
 from bot_bottle.backend import BottleSpec, get_bottle_backend
 from bot_bottle.backend.smolmachines.smolvm import is_available as _smolvm_available
-from bot_bottle.manifest import Manifest
+from bot_bottle.manifest import ManifestIndex
 from tests._docker import skip_unless_docker
 
 
 _AGENT_PROMPT = "You are demo. Be brief."
 
 
-def _minimal_manifest() -> Manifest:
-    return Manifest.from_json_obj({
+def _minimal_manifest() -> ManifestIndex:
+    return ManifestIndex.from_json_obj({
         "bottles": {
             "dev": {
                 "egress": {
@@ -198,6 +198,7 @@ class TestSmolmachinesLaunch(unittest.TestCase):
         # connect fails, which is the property chunk 3 will
         # preserve once egress is actually running.
         r = self.bottle.exec(
+            "env -u HTTPS_PROXY -u HTTP_PROXY -u https_proxy -u http_proxy "
             f"curl -s --show-error --max-time 3 http://{self.plan.bundle_ip}:9099 "
             "2>&1 || true"
         )

tests/unit/test_agent_provider.py

@@ -74,7 +74,7 @@ class TestAgentProviderRuntime(unittest.TestCase):
                 instance_name="bot-bottle-test",
                 prompt_file=prompt_file,
                 label="review-api",
-                color="bright-cyan",
+                color="cyan",
             )
             prompt = prompt_file.read_text()
             config = Path(tmp, "codex-config.toml").read_text()
@@ -168,6 +168,34 @@ class TestAgentProviderRuntime(unittest.TestCase):
         self.assertEqual("~/.claude/statusline.sh", settings["statusLine"]["command"])
         self.assertEqual("custom:bot-bottle-research-ui", settings["theme"])
 
+    def test_claude_plan_uses_startup_args_from_provider_settings(self):
+        with tempfile.TemporaryDirectory(prefix="bb-provider.") as tmp:
+            plan = build_agent_provision_plan(
+                template="claude",
+                dockerfile="",
+                state_dir=Path(tmp),
+                instance_name="bot-bottle-test",
+                prompt_file=Path(tmp) / "prompt.txt",
+                provider_settings={
+                    "startup_args": ["--model", "opus"],
+                },
+            )
+        self.assertEqual(("--model", "opus"), plan.startup_args)
+
+    def test_codex_plan_uses_startup_args_from_provider_settings(self):
+        with tempfile.TemporaryDirectory(prefix="bb-provider.") as tmp:
+            plan = build_agent_provision_plan(
+                template="codex",
+                dockerfile="",
+                state_dir=Path(tmp),
+                instance_name="bot-bottle-test",
+                prompt_file=Path(tmp) / "prompt.txt",
+                provider_settings={
+                    "startup_args": ["--model", "gpt-5-codex"],
+                },
+            )
+        self.assertEqual(("--model", "gpt-5-codex"), plan.startup_args)
+
     def test_codex_forward_host_credentials_populates_egress_routes(self):
         with tempfile.TemporaryDirectory(prefix="bb-provider.") as tmp:
             home = Path(tmp) / "host-codex"
@@ -394,6 +422,24 @@ class TestAgentProviderRuntime(unittest.TestCase):
         self.assertNotIn("OPENROUTER_API_KEY", plan.guest_env)
         self.assertTrue(provider["compat"]["supportsReasoningEffort"])
 
+    def test_pi_plan_appends_startup_args_from_provider_settings(self):
+        with tempfile.TemporaryDirectory(prefix="bb-provider.") as tmp:
+            plan = build_agent_provision_plan(
+                template="pi",
+                dockerfile="",
+                state_dir=Path(tmp),
+                instance_name="bot-bottle-test",
+                prompt_file=Path(tmp) / "prompt.txt",
+                provider_settings={
+                    "models": ["qwen3:14b"],
+                    "startup_args": ["--no-stream"],
+                },
+            )
+        self.assertEqual(
+            ("--models", "ollama/qwen3:14b", "--no-stream"),
+            plan.startup_args,
+        )
+
     def test_pi_prompt_mode_appends_system_prompt_interactively(self):
         self.assertEqual(
             ["--append-system-prompt", "/home/node/.bot-bottle-prompt.txt"],

tests/unit/test_backend_freezer.py

@@ -0,0 +1,216 @@
+"""Unit: Freezer class hierarchy."""
+
+from __future__ import annotations
+
+import tempfile
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+from bot_bottle import supervise, bottle_state
+from bot_bottle.backend import ActiveAgent
+from bot_bottle.backend.freeze import get_freezer
+from bot_bottle.backend.docker.freezer import DockerFreezer
+from bot_bottle.backend.macos_container.freezer import MacosContainerFreezer
+from bot_bottle.backend.smolmachines.freezer import SmolmachinesFreezer
+
+
+class _FakeHomeMixin:
+    def _setup_fake_home(self):
+        self._tmp = tempfile.TemporaryDirectory(prefix="freezer-test.")
+        original = supervise.bot_bottle_root
+
+        def fake_root() -> Path:
+            return Path(self._tmp.name) / ".bot-bottle"
+
+        supervise.bot_bottle_root = fake_root  # type: ignore[assignment]
+        self._restore = lambda: setattr(supervise, "bot_bottle_root", original)
+
+    def _teardown_fake_home(self):
+        self._restore()
+        self._tmp.cleanup()
+
+
+def _make_agent(slug: str, backend: str = "docker") -> ActiveAgent:
+    return ActiveAgent(
+        backend_name=backend,
+        slug=slug,
+        agent_name="dev",
+        started_at="t",
+        services=(),
+    )
+
+
+class TestGetFreezer(unittest.TestCase):
+    def test_docker(self):
+        self.assertIsInstance(get_freezer("docker"), DockerFreezer)
+
+    def test_empty_backend_gives_docker(self):
+        self.assertIsInstance(get_freezer(""), DockerFreezer)
+
+    def test_macos_container(self):
+        self.assertIsInstance(get_freezer("macos-container"), MacosContainerFreezer)
+
+    def test_smolmachines(self):
+        self.assertIsInstance(get_freezer("smolmachines"), SmolmachinesFreezer)
+
+    def test_unknown_backend_dies(self):
+        with patch("bot_bottle.backend.freeze.die", side_effect=SystemExit("die")):
+            with self.assertRaises(SystemExit):
+                get_freezer("unknown-backend")
+
+
+class TestFreezerBaseCommit(_FakeHomeMixin, unittest.TestCase):
+    """The base Freezer.commit() owns the shared post-freeze steps."""
+
+    def setUp(self):
+        self._setup_fake_home()
+
+    def tearDown(self):
+        self._teardown_fake_home()
+
+    def test_writes_committed_image_and_marks_preserved(self):
+        slug = "dev-abc12"
+        bottle_state.write_metadata(bottle_state.BottleMetadata(
+            identity=slug, agent_name="dev", cwd="", copy_cwd=False,
+            started_at="t", backend="docker",
+        ))
+        freezer = get_freezer("docker")
+        agent = _make_agent(slug)
+
+        with patch.object(freezer, "_freeze", return_value="bot-bottle-committed-dev-abc12:latest"), \
+             patch("bot_bottle.backend.freeze.info"):
+            freezer.commit(agent)
+
+        self.assertEqual(
+            "bot-bottle-committed-dev-abc12:latest",
+            bottle_state.read_committed_image(slug),
+        )
+        self.assertTrue(bottle_state.is_preserved(slug))
+
+    def test_commit_slug_passes_correct_slug_to_freeze(self):
+        slug = "dev-abc12"
+        bottle_state.write_metadata(bottle_state.BottleMetadata(
+            identity=slug, agent_name="dev", cwd="", copy_cwd=False,
+            started_at="t", backend="docker",
+        ))
+        freezer = get_freezer("docker")
+        captured = {}
+
+        def capture_freeze(agent: ActiveAgent) -> str:
+            captured["slug"] = agent.slug
+            return "some-ref"
+
+        with patch.object(freezer, "_freeze", side_effect=capture_freeze), \
+             patch("bot_bottle.backend.freeze.info"):
+            freezer.commit_slug(slug)
+
+        self.assertEqual(slug, captured["slug"])
+
+
+class TestDockerFreezer(_FakeHomeMixin, unittest.TestCase):
+    def setUp(self):
+        self._setup_fake_home()
+
+    def tearDown(self):
+        self._teardown_fake_home()
+
+    def test_commits_container_and_records_image(self):
+        slug = "dev-abc12"
+        bottle_state.write_metadata(bottle_state.BottleMetadata(
+            identity=slug, agent_name="dev", cwd="", copy_cwd=False,
+            started_at="t", backend="docker",
+        ))
+        freezer = DockerFreezer()
+        agent = _make_agent(slug)
+
+        with patch("bot_bottle.backend.docker.freezer.commit_container") as mock_commit, \
+             patch("bot_bottle.backend.freeze.info"), \
+             patch("bot_bottle.backend.docker.freezer.info"):
+            freezer.commit(agent)
+
+        mock_commit.assert_called_once_with(
+            f"bot-bottle-{slug}",
+            f"bot-bottle-committed-{slug}:latest",
+        )
+        self.assertEqual(
+            f"bot-bottle-committed-{slug}:latest",
+            bottle_state.read_committed_image(slug),
+        )
+        self.assertTrue(bottle_state.is_preserved(slug))
+
+
+class TestMacosContainerFreezer(_FakeHomeMixin, unittest.TestCase):
+    def setUp(self):
+        self._setup_fake_home()
+
+    def tearDown(self):
+        self._teardown_fake_home()
+
+    def _write_meta(self, slug: str) -> None:
+        bottle_state.write_metadata(bottle_state.BottleMetadata(
+            identity=slug, agent_name="dev", cwd="", copy_cwd=False,
+            started_at="t", backend="macos-container",
+        ))
+
+    def test_commits_running_container_without_stopping(self):
+        """Commit should exec-tar the running container, not stop it."""
+        slug = "dev-abc12"
+        self._write_meta(slug)
+        freezer = MacosContainerFreezer()
+        agent = _make_agent(slug, "macos-container")
+
+        with patch("bot_bottle.backend.macos_container.freezer.commit_container") as mock_commit, \
+             patch("bot_bottle.backend.freeze.info"), \
+             patch("bot_bottle.backend.macos_container.freezer.info"):
+            freezer.commit(agent)
+
+        mock_commit.assert_called_once_with(
+            f"bot-bottle-{slug}",
+            f"bot-bottle-committed-{slug}:latest",
+        )
+        self.assertEqual(
+            f"bot-bottle-committed-{slug}:latest",
+            bottle_state.read_committed_image(slug),
+        )
+        self.assertTrue(bottle_state.is_preserved(slug))
+
+
+class TestSmolmachinesFreezer(_FakeHomeMixin, unittest.TestCase):
+    def setUp(self):
+        self._setup_fake_home()
+
+    def tearDown(self):
+        self._teardown_fake_home()
+
+    def _write_meta(self, slug: str) -> None:
+        bottle_state.write_metadata(bottle_state.BottleMetadata(
+            identity=slug, agent_name="dev", cwd="", copy_cwd=False,
+            started_at="t", backend="smolmachines",
+        ))
+
+    def test_snapshots_running_vm_without_stopping(self):
+        """Commit should exec-tar the running VM, not stop it."""
+        slug = "dev-abc12"
+        self._write_meta(slug)
+        freezer = SmolmachinesFreezer()
+        agent = _make_agent(slug, "smolmachines")
+
+        with patch("bot_bottle.backend.smolmachines.freezer._snapshot_running_vm") as mock_snap, \
+             patch("bot_bottle.backend.freeze.info"), \
+             patch("bot_bottle.backend.smolmachines.freezer.info"):
+            freezer.commit(agent)
+
+        expected_binary = bottle_state.bottle_state_dir(slug) / "committed-smolmachine"
+        mock_snap.assert_called_once_with(
+            f"bot-bottle-{slug}",
+            f"bot-bottle-committed-{slug}:latest",
+            expected_binary,
+        )
+        expected_sidecar = str(expected_binary.with_suffix(".smolmachine"))
+        self.assertEqual(expected_sidecar, bottle_state.read_committed_image(slug))
+        self.assertTrue(bottle_state.is_preserved(slug))
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/unit/test_backend_prepare.py

@@ -16,12 +16,13 @@ from bot_bottle import bottle_state
 from bot_bottle import supervise
 from bot_bottle.backend import BottleSpec
 from bot_bottle.backend.docker import DockerBottleBackend
+from bot_bottle.backend.resolve_common import mint_slug
 from bot_bottle.backend.smolmachines import SmolmachinesBottleBackend
-from bot_bottle.manifest import Manifest
+from bot_bottle.manifest import ManifestIndex
 
 
-def _manifest() -> Manifest:
-    return Manifest.from_json_obj({
+def _manifest() -> ManifestIndex:
+    return ManifestIndex.from_json_obj({
         "bottles": {
             "dev": {
                 "env": {
@@ -115,5 +116,36 @@ class TestSmolmachinesPrepare(_FakeStateMixin, unittest.TestCase):
         )
 
 
+class TestMintSlug(unittest.TestCase):
+    def _spec(self, *, label: str = "", identity: str = "") -> BottleSpec:
+        manifest = _manifest()
+        return BottleSpec(
+            manifest=manifest,
+            agent_name="demo",
+            copy_cwd=False,
+            user_cwd="/tmp",
+            label=label,
+            identity=identity,
+        )
+
+    def test_no_label_uses_agent_name_with_random_suffix(self) -> None:
+        slug = mint_slug(self._spec(label=""))
+        self.assertTrue(slug.startswith("demo-"), slug)
+        # random suffix present — slug is longer than just "demo"
+        self.assertGreater(len(slug), len("demo-"))
+
+    def test_label_becomes_exact_slug(self) -> None:
+        slug = mint_slug(self._spec(label="my-run"))
+        self.assertEqual("my-run", slug)
+
+    def test_label_with_spaces_slugified_no_suffix(self) -> None:
+        slug = mint_slug(self._spec(label="My Feature Run"))
+        self.assertEqual("my-feature-run", slug)
+
+    def test_identity_takes_precedence_over_label(self) -> None:
+        slug = mint_slug(self._spec(label="my-run", identity="fixed-id"))
+        self.assertEqual("fixed-id", slug)
+
+
 if __name__ == "__main__":
     unittest.main()

tests/unit/test_backend_terminal.py

@@ -11,14 +11,14 @@ class TestPalettePrintf(unittest.TestCase):
     def test_known_color_returns_printf(self):
         cmd = palette_printf("red")
         self.assertTrue(cmd.startswith("printf '"))
-        self.assertIn("\\033]4;1;", cmd)   # normal red
-        self.assertIn("\\033]4;9;", cmd)   # bright red
+        self.assertIn("\\033]4;9;", cmd)   # bright-red slot
+        self.assertIn("\\033]4;1;", cmd)   # normal-red slot
         self.assertIn("\\033]11;", cmd)    # default background tint
 
-    def test_bright_variant_sets_both_slots(self):
-        cmd = palette_printf("bright-blue")
-        self.assertIn("\\033]4;12;", cmd)  # bright-blue
-        self.assertIn("\\033]4;4;", cmd)   # blue
+    def test_color_sets_both_palette_slots(self):
+        cmd = palette_printf("blue")
+        self.assertIn("\\033]4;12;", cmd)  # bright-blue slot
+        self.assertIn("\\033]4;4;", cmd)   # normal-blue slot
 
     def test_unknown_color_returns_empty(self):
         self.assertEqual("", palette_printf(""))
@@ -26,10 +26,7 @@ class TestPalettePrintf(unittest.TestCase):
 
     def test_all_named_colors_produce_output(self):
         colors = [
-            "black", "red", "green", "yellow",
-            "blue", "magenta", "cyan", "white",
-            "bright-black", "bright-red", "bright-green", "bright-yellow",
-            "bright-blue", "bright-magenta", "bright-cyan", "bright-white",
+            "red", "green", "yellow", "blue", "magenta",
         ]
         for color in colors:
             with self.subTest(color=color):
@@ -65,7 +62,7 @@ class TestExecShellScript(unittest.TestCase):
         self.assertFalse(agent_part.startswith("exec "))
 
     def test_title_and_color_both_appear(self):
-        script = exec_shell_script(self._ARGV, terminal_title="bot", terminal_color="cyan")
+        script = exec_shell_script(self._ARGV, terminal_title="bot", terminal_color="magenta")
         assert script is not None
         self.assertIn("bot", script)
         self.assertIn("\\033]4;", script)

tests/unit/test_backend_workspace.py

@@ -17,11 +17,11 @@ from bot_bottle import supervise
 from bot_bottle.backend import Bottle, BottleSpec, ExecResult
 from bot_bottle.backend.docker import DockerBottleBackend
 from bot_bottle.backend.smolmachines import SmolmachinesBottleBackend
-from bot_bottle.manifest import Manifest
+from bot_bottle.manifest import ManifestIndex
 
 
-def _manifest() -> Manifest:
-    return Manifest.from_json_obj({
+def _manifest() -> ManifestIndex:
+    return ManifestIndex.from_json_obj({
         "bottles": {"dev": {}},
         "agents": {
             "demo": {

tests/unit/test_bottle_state.py

@@ -115,8 +115,8 @@ class TestBottleIdentity(unittest.TestCase):
 
 
 class TestPreserveMarker(_FakeHomeMixin, unittest.TestCase):
-    """The .preserve marker is how capability_apply tells cli.py's
-    session-end cleanup to keep the state dir instead of removing it."""
+    """The .preserve marker tells cli.py's session-end cleanup to keep
+    the state dir instead of removing it."""
 
     def setUp(self):
         self._setup_fake_home()
@@ -277,5 +277,56 @@ class TestBottleMetadataBackend(_FakeHomeMixin, unittest.TestCase):
         self.assertEqual("", loaded.backend)
 
 
+class TestCommittedImage(_FakeHomeMixin, unittest.TestCase):
+    """write_committed_image / read_committed_image round-trip."""
+
+    def setUp(self):
+        self._setup_fake_home()
+
+    def tearDown(self):
+        self._teardown_fake_home()
+
+    def test_returns_none_when_absent(self):
+        self.assertIsNone(bottle_state.read_committed_image("dev"))
+
+    def test_write_then_read_roundtrip(self):
+        bottle_state.write_committed_image("dev", "bot-bottle-committed-dev:latest")
+        self.assertEqual(
+            "bot-bottle-committed-dev:latest",
+            bottle_state.read_committed_image("dev"),
+        )
+
+    def test_strips_trailing_newline_on_read(self):
+        path = bottle_state.committed_image_path("dev")
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_text("bot-bottle-committed-dev:latest\n\n")
+        self.assertEqual(
+            "bot-bottle-committed-dev:latest",
+            bottle_state.read_committed_image("dev"),
+        )
+
+    def test_isolated_per_slug(self):
+        bottle_state.write_committed_image("dev", "bot-bottle-committed-dev:latest")
+        bottle_state.write_committed_image("api", "bot-bottle-committed-api:latest")
+        self.assertEqual(
+            "bot-bottle-committed-dev:latest",
+            bottle_state.read_committed_image("dev"),
+        )
+        self.assertEqual(
+            "bot-bottle-committed-api:latest",
+            bottle_state.read_committed_image("api"),
+        )
+
+    def test_path_under_state_dir(self):
+        path = bottle_state.committed_image_path("dev")
+        self.assertTrue(str(path).endswith("/.bot-bottle/state/dev/committed-image"))
+
+    def test_empty_content_returns_none(self):
+        path = bottle_state.committed_image_path("dev")
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_text("   \n")
+        self.assertIsNone(bottle_state.read_committed_image("dev"))
+
+
 if __name__ == "__main__":
     unittest.main()

tests/unit/test_cli_commit.py

@@ -0,0 +1,143 @@
+"""Unit: cli.py commit command."""
+
+from __future__ import annotations
+
+import tempfile
+import unittest
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+from bot_bottle.cli.commit import cmd_commit
+from bot_bottle import supervise
+from bot_bottle import bottle_state
+from bot_bottle.backend.freeze import CommitCancelled
+
+
+class _FakeHomeMixin:
+    def _setup_fake_home(self):
+        self._tmp = tempfile.TemporaryDirectory(prefix="cli-commit-test.")
+        original = supervise.bot_bottle_root
+
+        def fake_root() -> Path:
+            return Path(self._tmp.name) / ".bot-bottle"
+
+        supervise.bot_bottle_root = fake_root  # type: ignore[assignment]
+        self._restore = lambda: setattr(supervise, "bot_bottle_root", original)
+
+    def _teardown_fake_home(self):
+        self._restore()
+        self._tmp.cleanup()
+
+
+class TestCmdCommitSlugArg(_FakeHomeMixin, unittest.TestCase):
+    """cmd_commit with an explicit slug delegates to get_freezer."""
+
+    def setUp(self):
+        self._setup_fake_home()
+
+    def tearDown(self):
+        self._teardown_fake_home()
+
+    def _write_meta(self, slug: str, backend: str) -> None:
+        bottle_state.write_metadata(bottle_state.BottleMetadata(
+            identity=slug, agent_name="dev", cwd="", copy_cwd=False,
+            started_at="t", backend=backend,
+        ))
+
+    def test_commits_docker_bottle(self):
+        slug = "dev-abc12"
+        self._write_meta(slug, "docker")
+
+        with patch("bot_bottle.cli.commit.get_freezer") as mock_gf:
+            mock_freezer = MagicMock()
+            mock_gf.return_value = mock_freezer
+            rc = cmd_commit([slug])
+
+        self.assertEqual(0, rc)
+        mock_gf.assert_called_once_with("docker")
+        mock_freezer.commit_slug.assert_called_once_with(slug)
+
+    def test_empty_backend_passed_to_get_freezer(self):
+        """Old state dirs without a backend field pass '' to get_freezer."""
+        slug = "dev-abc12"
+        self._write_meta(slug, "")
+
+        with patch("bot_bottle.cli.commit.get_freezer") as mock_gf:
+            mock_freezer = MagicMock()
+            mock_gf.return_value = mock_freezer
+            rc = cmd_commit([slug])
+
+        self.assertEqual(0, rc)
+        mock_gf.assert_called_once_with("")
+
+    def test_commits_macos_container_bottle(self):
+        slug = "dev-abc12"
+        self._write_meta(slug, "macos-container")
+
+        with patch("bot_bottle.cli.commit.get_freezer") as mock_gf:
+            mock_freezer = MagicMock()
+            mock_gf.return_value = mock_freezer
+            rc = cmd_commit([slug])
+
+        self.assertEqual(0, rc)
+        mock_gf.assert_called_once_with("macos-container")
+        mock_freezer.commit_slug.assert_called_once_with(slug)
+
+    def test_commits_smolmachines_bottle(self):
+        slug = "dev-abc12"
+        self._write_meta(slug, "smolmachines")
+
+        with patch("bot_bottle.cli.commit.get_freezer") as mock_gf:
+            mock_freezer = MagicMock()
+            mock_gf.return_value = mock_freezer
+            rc = cmd_commit([slug])
+
+        self.assertEqual(0, rc)
+        mock_gf.assert_called_once_with("smolmachines")
+
+    def test_returns_zero_on_commit_cancelled(self):
+        slug = "dev-abc12"
+        self._write_meta(slug, "macos-container")
+
+        with patch("bot_bottle.cli.commit.get_freezer") as mock_gf:
+            mock_freezer = MagicMock()
+            mock_freezer.commit_slug.side_effect = CommitCancelled
+            mock_gf.return_value = mock_freezer
+            rc = cmd_commit([slug])
+
+        self.assertEqual(0, rc)
+
+
+class TestCmdCommitNoActiveBottles(_FakeHomeMixin, unittest.TestCase):
+    def setUp(self):
+        self._setup_fake_home()
+
+    def tearDown(self):
+        self._teardown_fake_home()
+
+    def test_dies_when_no_active_bottles_and_no_slug(self):
+        with patch(
+            "bot_bottle.cli.commit.enumerate_active_agents", return_value=[],
+        ), patch(
+            "bot_bottle.cli.commit.die", side_effect=SystemExit("die"),
+        ) as mock_die:
+            with self.assertRaises(SystemExit):
+                cmd_commit([])
+
+        mock_die.assert_called_once()
+
+    def test_returns_zero_when_picker_cancelled(self):
+        active = MagicMock()
+        active.slug = "dev-abc12"
+        with patch(
+            "bot_bottle.cli.commit.enumerate_active_agents", return_value=[active],
+        ), patch(
+            "bot_bottle.cli.commit.tui.filter_select", return_value=None,
+        ):
+            rc = cmd_commit([])
+
+        self.assertEqual(0, rc)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/unit/test_cli_dispatch.py

@@ -0,0 +1,82 @@
+"""Unit: top-level CLI dispatch in bot_bottle.cli.main (ADR 0004).
+
+`cli/__init__.py` is dispatch + exit-code mapping, not interactive I/O,
+so it carries real unit tests rather than being omitted like the
+`cli/init` / `cli/tui` shells."""
+
+from __future__ import annotations
+
+import io
+import unittest
+from unittest.mock import patch
+
+import bot_bottle.cli as climod
+from bot_bottle.cli import main
+from bot_bottle.log import Die
+from bot_bottle.manifest import ManifestError
+
+
+class TestMainDispatch(unittest.TestCase):
+    def test_no_args_prints_usage_returns_2(self) -> None:
+        with patch("sys.stderr", io.StringIO()):
+            self.assertEqual(2, main([]))
+
+    def test_help_flags_return_0(self) -> None:
+        with patch("sys.stderr", io.StringIO()):
+            self.assertEqual(0, main(["-h"]))
+            self.assertEqual(0, main(["--help"]))
+
+    def test_unknown_command_dies(self) -> None:
+        with patch("sys.stderr", io.StringIO()):
+            with self.assertRaises(Die):
+                main(["definitely-not-a-command"])
+
+    def test_handler_return_code_passthrough(self) -> None:
+        def handler(_rest: list[str]) -> int:
+            return 7
+
+        with patch.dict(climod.COMMANDS, {"x": handler}):
+            self.assertEqual(7, main(["x"]))
+
+    def test_handler_none_return_becomes_0(self) -> None:
+        def handler(_rest: list[str]) -> int | None:
+            return None
+
+        with patch.dict(climod.COMMANDS, {"x": handler}):
+            self.assertEqual(0, main(["x"]))
+
+    def test_args_forwarded_to_handler(self) -> None:
+        seen: list[list[str]] = []
+
+        def handler(rest: list[str]) -> int:
+            seen.append(rest)
+            return 0
+
+        with patch.dict(climod.COMMANDS, {"x": handler}):
+            main(["x", "a", "b"])
+        self.assertEqual([["a", "b"]], seen)
+
+    def test_manifest_error_maps_to_1(self) -> None:
+        def boom(_rest: list[str]) -> int:
+            raise ManifestError("bad manifest")
+
+        with patch.dict(climod.COMMANDS, {"x": boom}), patch("sys.stderr", io.StringIO()):
+            self.assertEqual(1, main(["x"]))
+
+    def test_die_maps_to_its_code(self) -> None:
+        def boom(_rest: list[str]) -> int:
+            raise Die(3)
+
+        with patch.dict(climod.COMMANDS, {"x": boom}):
+            self.assertEqual(3, main(["x"]))
+
+    def test_keyboard_interrupt_maps_to_130(self) -> None:
+        def boom(_rest: list[str]) -> int:
+            raise KeyboardInterrupt()
+
+        with patch.dict(climod.COMMANDS, {"x": boom}):
+            self.assertEqual(130, main(["x"]))
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/unit/test_cli_start_selector.py

@@ -1,7 +1,8 @@
-"""Unit: cmd_start selector dispatch (PRD 0051).
+"""Unit: cmd_start selector dispatch (PRD 0051, issue #269).
 
 Tests that cmd_start calls filter_select only when the agent name is
-absent, skips it when the agent is explicit, and returns 0 on cancel.
+absent, shows the bottle multiselect after agent selection, and skips
+pickers when both are explicitly set.
 
 All actual launch work is stubbed so no container is created.
 """
@@ -10,15 +11,24 @@ from __future__ import annotations
 
 import os
 import unittest
+from collections.abc import Mapping, Sequence
 from unittest.mock import MagicMock, patch
 
 import bot_bottle.cli.start as start_mod
 import bot_bottle.cli.tui as tui_mod
+from bot_bottle.backend import ActiveAgent
 
 
-def _make_manifest(agent_names: list[str]):
+def _make_manifest(
+    agent_names: list[str],
+    bottle_names: list[str] | None = None,
+    agent_bottle: str = "",
+):
     manifest = MagicMock()
-    manifest.agents = {name: MagicMock() for name in agent_names}
+    manifest.agents = {name: MagicMock(bottle=agent_bottle) for name in agent_names}
+    manifest.all_agent_names = sorted(agent_names)
+    manifest.all_bottle_names = sorted(bottle_names or [])
+    manifest.home_md = None  # eager mode so _peek_agent_bottle uses agents dict
     return manifest
 
 
@@ -26,27 +36,27 @@ class TestCmdStartSelector(unittest.TestCase):
     """Drive cmd_start with a minimal set of stubs."""
 
     def setUp(self):
-        # Stub Manifest.resolve so no on-disk manifest is needed.
-        self._manifest = _make_manifest(["researcher", "implementer"])
+        self._manifest = _make_manifest(["researcher", "implementer"], ["claude", "dev"])
         self._resolve_patch = patch(
-            "bot_bottle.cli.start.Manifest.resolve",
+            "bot_bottle.cli.start.ManifestIndex.resolve",
             return_value=self._manifest,
         )
         self._resolve_patch.start()
 
-        # Stub _launch_bottle so no real container work happens.
         self._launch_patch = patch(
             "bot_bottle.cli.start._launch_bottle",
             return_value=0,
         )
         self._launch_mock = self._launch_patch.start()
 
-        # Stub filter_select to avoid opening /dev/tty.
-        self._tui_patch = patch.object(tui_mod, "filter_select")
-        self._tui_mock = self._tui_patch.start()
+        # Stub filter_select (agent picker) and filter_multiselect (bottle picker).
+        self._agent_picker_patch = patch.object(tui_mod, "filter_select")
+        self._agent_picker_mock = self._agent_picker_patch.start()
+
+        self._bottle_picker_patch = patch.object(tui_mod, "filter_multiselect")
+        self._bottle_picker_mock = self._bottle_picker_patch.start()
+        self._bottle_picker_mock.return_value = ["claude"]  # default: one bottle selected
 
-        # Ensure BOT_BOTTLE_BACKEND is absent so omitted --backend
-        # flows through to the resolver default.
         self._env_patch = patch.dict(os.environ, {}, clear=False)
         self._env_patch.start()
         os.environ.pop("BOT_BOTTLE_BACKEND", None)
@@ -54,50 +64,108 @@ class TestCmdStartSelector(unittest.TestCase):
     def tearDown(self):
         self._resolve_patch.stop()
         self._launch_patch.stop()
-        self._tui_patch.stop()
+        self._agent_picker_patch.stop()
+        self._bottle_picker_patch.stop()
         self._env_patch.stop()
 
     # ------------------------------------------------------------------
-    # Both explicit — no picker shown
+    # Agent explicit — agent picker skipped; bottle picker always shown
     # ------------------------------------------------------------------
 
-    def test_both_explicit_skips_picker(self):
-        self._tui_mock.return_value = "researcher"
+    def test_explicit_agent_skips_agent_picker(self):
         rc = start_mod.cmd_start(["--backend=docker", "researcher"])
         self.assertEqual(0, rc)
-        self._tui_mock.assert_not_called()
+        self._agent_picker_mock.assert_not_called()
+        self._bottle_picker_mock.assert_called_once()
         self._launch_mock.assert_called_once()
-        _, kwargs = self._launch_mock.call_args
-        self.assertEqual("docker", kwargs["backend_name"])
+
+    def test_explicit_agent_bottle_picker_shows_available_bottles(self):
+        start_mod.cmd_start(["researcher"])
+        call_kwargs = self._bottle_picker_mock.call_args
+        self.assertEqual(["claude", "dev"], call_kwargs[0][0])
+        self.assertIn("bottle", call_kwargs[1]["title"].lower())
 
     # ------------------------------------------------------------------
-    # Agent absent → agent picker fires; backend explicit
+    # Agent absent → agent picker fires; bottle picker always follows
     # ------------------------------------------------------------------
 
     def test_agent_absent_shows_agent_picker(self):
-        self._tui_mock.return_value = "researcher"
+        self._agent_picker_mock.return_value = "researcher"
         rc = start_mod.cmd_start(["--backend=docker"])
         self.assertEqual(0, rc)
-        self._tui_mock.assert_called_once()
-        call_kwargs = self._tui_mock.call_args
+        self._agent_picker_mock.assert_called_once()
+        call_kwargs = self._agent_picker_mock.call_args
         self.assertEqual(["implementer", "researcher"], call_kwargs[0][0])
         self.assertIn("agent", call_kwargs[1]["title"].lower())
+        # Bottle picker must also fire after agent selection.
+        self._bottle_picker_mock.assert_called_once()
 
-    def test_agent_picker_cancel_returns_0(self):
-        self._tui_mock.return_value = None
+    def test_agent_picker_cancel_skips_bottle_picker(self):
+        self._agent_picker_mock.return_value = None
         rc = start_mod.cmd_start(["--backend=docker"])
         self.assertEqual(0, rc)
+        self._bottle_picker_mock.assert_not_called()
+        self._launch_mock.assert_not_called()
+
+    def test_bottle_picker_cancel_returns_0(self):
+        self._bottle_picker_mock.return_value = None
+        rc = start_mod.cmd_start(["researcher"])
+        self.assertEqual(0, rc)
         self._launch_mock.assert_not_called()
 
     # ------------------------------------------------------------------
-    # Agent explicit, backend absent → no picker
+    # Bottle selection is forwarded to BottleSpec
     # ------------------------------------------------------------------
 
-    def test_backend_absent_uses_default_without_picker(self):
-        rc = start_mod.cmd_start(["researcher"])
-        self.assertEqual(0, rc)
-        self._tui_mock.assert_not_called()
+    def test_selected_bottles_forwarded_to_spec(self):
+        self._bottle_picker_mock.return_value = ["claude", "dev"]
+        start_mod.cmd_start(["researcher"])
         self._launch_mock.assert_called_once()
+        spec = self._launch_mock.call_args[0][0]
+        self.assertEqual(("claude", "dev"), spec.bottle_names)
+
+    def test_empty_bottle_selection_forwarded(self):
+        self._bottle_picker_mock.return_value = []
+        start_mod.cmd_start(["researcher"])
+        self._launch_mock.assert_called_once()
+        spec = self._launch_mock.call_args[0][0]
+        self.assertEqual((), spec.bottle_names)
+
+    # ------------------------------------------------------------------
+    # Agent default bottle pre-populates the picker
+    # ------------------------------------------------------------------
+
+    def test_agent_bottle_prepopulates_bottle_picker(self):
+        manifest = _make_manifest(
+            ["implementer"], ["claude", "dev"], agent_bottle="claude"
+        )
+        with patch(
+            "bot_bottle.cli.start.ManifestIndex.resolve", return_value=manifest
+        ):
+            start_mod.cmd_start(["implementer"])
+        call_kwargs = self._bottle_picker_mock.call_args
+        self.assertEqual(["claude"], call_kwargs[1]["initial"])
+
+    def test_no_agent_bottle_empty_initial(self):
+        manifest = _make_manifest(["researcher"], ["claude", "dev"], agent_bottle="")
+        with patch(
+            "bot_bottle.cli.start.ManifestIndex.resolve", return_value=manifest
+        ):
+            start_mod.cmd_start(["researcher"])
+        call_kwargs = self._bottle_picker_mock.call_args
+        self.assertEqual([], call_kwargs[1]["initial"])
+
+    # ------------------------------------------------------------------
+    # Backend wiring
+    # ------------------------------------------------------------------
+
+    def test_explicit_backend_forwarded(self):
+        start_mod.cmd_start(["--backend=docker", "researcher"])
+        _, kwargs = self._launch_mock.call_args
+        self.assertEqual("docker", kwargs["backend_name"])
+
+    def test_absent_backend_uses_default(self):
+        start_mod.cmd_start(["researcher"])
         _, kwargs = self._launch_mock.call_args
         self.assertIsNone(kwargs["backend_name"])
 
@@ -108,30 +176,185 @@ class TestCmdStartSelector(unittest.TestCase):
         finally:
             os.environ.pop("BOT_BOTTLE_BACKEND", None)
         self.assertEqual(0, rc)
-        self._tui_mock.assert_not_called()
 
-    # ------------------------------------------------------------------
-    # Both absent → only agent picker
-    # ------------------------------------------------------------------
-
-    def test_both_absent_shows_only_agent_picker(self):
-        self._tui_mock.return_value = "researcher"
+    def test_both_absent_shows_agent_picker_then_bottle_picker(self):
+        self._agent_picker_mock.return_value = "researcher"
         rc = start_mod.cmd_start([])
         self.assertEqual(0, rc)
-        self._tui_mock.assert_called_once()
-        title = self._tui_mock.call_args[1]["title"].lower()
-        self.assertIn("agent", title)
+        self._agent_picker_mock.assert_called_once()
+        self._bottle_picker_mock.assert_called_once()
         self._launch_mock.assert_called_once()
-        _, kwargs = self._launch_mock.call_args
-        self.assertIsNone(kwargs["backend_name"])
 
-    def test_both_absent_agent_cancel_skips_backend_picker(self):
-        self._tui_mock.side_effect = [None]
+    def test_both_absent_agent_cancel_skips_bottle_and_launch(self):
+        self._agent_picker_mock.return_value = None
         rc = start_mod.cmd_start([])
         self.assertEqual(0, rc)
-        self.assertEqual(1, self._tui_mock.call_count)
+        self._agent_picker_mock.assert_called_once()
+        self._bottle_picker_mock.assert_not_called()
         self._launch_mock.assert_not_called()
 
 
+def _active_agent(slug: str) -> ActiveAgent:
+    return ActiveAgent(
+        backend_name="docker",
+        slug=slug,
+        agent_name="demo",
+        started_at="2026-01-01T00:00:00+00:00",
+        services=(),
+    )
+
+
+class TestCmdStartLabelCollision(unittest.TestCase):
+    """cmd_start re-prompts when the label's slug is already running."""
+
+    def setUp(self):
+        self._manifest = _make_manifest(["researcher"], ["claude"])
+        patch("bot_bottle.cli.start.ManifestIndex.resolve", return_value=self._manifest).start()
+        self._launch_mock = patch(
+            "bot_bottle.cli.start._launch_bottle", return_value=0,
+        ).start()
+        # Stub the bottle picker to always return a selection.
+        patch.object(tui_mod, "filter_multiselect", return_value=["claude"]).start()
+        self.addCleanup(patch.stopall)
+
+    def test_no_collision_proceeds_without_reprompt(self):
+        with (
+            patch.object(tui_mod, "name_color_modal", return_value=("researcher", "")) as modal,
+            patch("bot_bottle.cli.start.enumerate_active_agents", return_value=[]),
+        ):
+            rc = start_mod.cmd_start(["researcher"])
+        self.assertEqual(0, rc)
+        modal.assert_called_once()
+        self._launch_mock.assert_called_once()
+
+    def test_collision_reprompts_with_disclaimer(self):
+        collision_agent = _active_agent("researcher")
+        call_count = 0
+
+        def _modal(default_label: str, *, disclaimer: str = "", **_kw: object) -> tuple[str, str]:
+            nonlocal call_count
+            call_count += 1
+            if call_count == 1:
+                return "researcher", ""
+            return "researcher-2", ""
+
+        with (
+            patch.object(tui_mod, "name_color_modal", side_effect=_modal) as modal,
+            patch(
+                "bot_bottle.cli.start.enumerate_active_agents",
+                side_effect=[[collision_agent], []],
+            ),
+        ):
+            rc = start_mod.cmd_start(["researcher"])
+
+        self.assertEqual(0, rc)
+        self.assertEqual(2, modal.call_count)
+        second_call_kwargs = modal.call_args_list[1][1]
+        self.assertIn("researcher", second_call_kwargs.get("disclaimer", ""))
+        self.assertIn("already in use", second_call_kwargs.get("disclaimer", ""))
+
+
+class TestBottleLineage(unittest.TestCase):
+    """Unit tests for _bottle_lineage."""
+
+    def test_returns_empty_in_eager_mode(self):
+        manifest = _make_manifest(["agent"], ["base", "dev"])
+        # home_md is None in eager mode → no file reads, returns {}
+        result = start_mod._bottle_lineage(manifest)
+        self.assertEqual({}, result)
+
+    def test_reads_extends_chain_from_files(self):
+        import tempfile
+        from pathlib import Path
+
+        with tempfile.TemporaryDirectory() as tmp:
+            bottles_dir = Path(tmp) / "bottles"
+            bottles_dir.mkdir()
+            (bottles_dir / "base.md").write_text("---\n{}\n---\n")
+            (bottles_dir / "mid.md").write_text("---\nextends: base\n---\n")
+            (bottles_dir / "leaf.md").write_text("---\nextends: mid\n---\n")
+
+            manifest = MagicMock()
+            manifest.home_md = Path(tmp)
+
+            result = start_mod._bottle_lineage(manifest)
+
+        self.assertNotIn("base", result)          # no parent → not in map
+        self.assertEqual("base -> mid", result["mid"])
+        self.assertEqual("base -> mid -> leaf", result["leaf"])
+
+    def test_cycle_protection(self):
+        import tempfile
+        from pathlib import Path
+
+        with tempfile.TemporaryDirectory() as tmp:
+            bottles_dir = Path(tmp) / "bottles"
+            bottles_dir.mkdir()
+            (bottles_dir / "a.md").write_text("---\nextends: b\n---\n")
+            (bottles_dir / "b.md").write_text("---\nextends: a\n---\n")
+
+            manifest = MagicMock()
+            manifest.home_md = Path(tmp)
+
+            result = start_mod._bottle_lineage(manifest)
+
+        # Cycle must not hang; each should get a two-element chain.
+        for name in ("a", "b"):
+            self.assertIn(name, result)
+            self.assertIn("->", result[name])
+
+
+class TestManifestToYaml(unittest.TestCase):
+    """Unit tests for _manifest_to_yaml."""
+
+    def _make_manifest_obj(
+        self,
+        *,
+        skills: Sequence[str] = (),
+        env: Mapping[str, str] | None = None,
+        supervise: bool = True,
+        agent_provider_template: str = "claude",
+    ):
+        from bot_bottle.manifest import Manifest, ManifestBottle
+        from bot_bottle.manifest_agent import ManifestAgent, ManifestAgentProvider
+
+        agent = ManifestAgent(skills=tuple(skills))
+        bottle = ManifestBottle(
+            env=env or {},
+            supervise=supervise,
+            agent_provider=ManifestAgentProvider(template=agent_provider_template),
+        )
+        return Manifest(agent=agent, bottle=bottle)
+
+    def test_includes_agent_section(self):
+        m = self._make_manifest_obj(skills=["researcher"])
+        yaml = start_mod._manifest_to_yaml(m)
+        self.assertIn("agent:", yaml)
+        self.assertIn("- researcher", yaml)
+
+    def test_includes_bottle_section(self):
+        m = self._make_manifest_obj(env={"FOO": "bar"})
+        yaml = start_mod._manifest_to_yaml(m)
+        self.assertIn("bottle:", yaml)
+        self.assertIn("FOO: bar", yaml)
+
+    def test_supervise_rendered(self):
+        m_true = self._make_manifest_obj(supervise=True)
+        m_false = self._make_manifest_obj(supervise=False)
+        self.assertIn("supervise: true", start_mod._manifest_to_yaml(m_true))
+        self.assertIn("supervise: false", start_mod._manifest_to_yaml(m_false))
+
+    def test_non_claude_provider_shown(self):
+        m = self._make_manifest_obj(agent_provider_template="codex")
+        yaml = start_mod._manifest_to_yaml(m)
+        self.assertIn("agent_provider:", yaml)
+        self.assertIn("template: codex", yaml)
+
+    def test_default_claude_provider_omitted(self):
+        m = self._make_manifest_obj(agent_provider_template="claude")
+        yaml = start_mod._manifest_to_yaml(m)
+        self.assertNotIn("agent_provider:", yaml)
+
+
 if __name__ == "__main__":
     unittest.main()

tests/unit/test_cli_start_settle.py

@@ -29,8 +29,8 @@ class _FakeHomeMixin:
 
 
 class TestCaptureSessionState(_FakeHomeMixin, unittest.TestCase):
-    # snapshot_transcript is commented out (capability_apply is disabled);
-    # capture_claude_session_state now only handles the preserve marker.
+    # capture_claude_session_state handles the preserve marker for
+    # non-zero agent exits.
     def setUp(self):
         self._setup_fake_home()
 
@@ -102,6 +102,27 @@ class TestAttachAgent(unittest.TestCase):
             bottle.argv,
         )
 
+    def test_remote_control_is_provider_startup_arg(self):
+        class Bottle:
+            argv: list[str] = []
+
+            def exec_agent(self, argv: list[str], *, tty: bool = True) -> int:
+                self.argv = list(argv)
+                return 0
+
+        bottle = Bottle()
+        exit_code = start_mod.attach_agent(
+            bottle,  # type: ignore[arg-type]
+            agent_provider_template="codex",
+            startup_args=("remote-control",),
+        )
+
+        self.assertEqual(0, exit_code)
+        self.assertEqual(
+            ["--dangerously-bypass-approvals-and-sandbox", "remote-control"],
+            bottle.argv,
+        )
+
 
 if __name__ == "__main__":
     unittest.main()

tests/unit/test_cli_tui.py

@@ -1,4 +1,4 @@
-"""Unit tests for bot_bottle.cli.tui — filter_select internals.
+"""Unit tests for bot_bottle.cli.tui — filter_select and filter_multiselect.
 
 We test the pure-Python logic (_filter_items, cursor movement, confirm,
 cancel) by exercising the internal helpers directly, without spinning up
@@ -8,8 +8,15 @@ a real curses session (which requires a TTY).
 from __future__ import annotations
 
 import unittest
+from typing import Any, Optional
 
-from bot_bottle.cli.tui import _filter_items, filter_select
+from bot_bottle.cli.tui import _filter_items, _multiselect_loop, filter_multiselect, filter_select
+
+_KEY_SPACE = 32
+_KEY_ENTER = 10
+
+_KEY_ESC = 27
+_KEY_CTRL_D = 4
 
 
 class TestFilterItems(unittest.TestCase):
@@ -46,5 +53,124 @@ class TestFilterSelectEmptyItems(unittest.TestCase):
         self.assertIsNone(result)
 
 
+class TestFilterMultiselectEmptyItems(unittest.TestCase):
+    def test_returns_empty_list_for_empty_items(self):
+        # No TTY needed — short-circuits before opening tty.
+        result = filter_multiselect([], title="Select", tty_path="/dev/null")
+        self.assertEqual([], result)
+
+    def test_returns_none_when_tty_unavailable(self):
+        result = filter_multiselect(["a", "b"], tty_path="/nonexistent/tty")
+        self.assertIsNone(result)
+
+
+class TestMultiselectLoopReordering(unittest.TestCase):
+    """Exercise _multiselect_loop key handling without a real curses terminal.
+
+    We drive the loop via a fake screen that feeds a pre-recorded key sequence
+    and records what was drawn — we only need the return value, so the fake
+    screen's getch() raises StopIteration after the key list is exhausted, and
+    the loop is expected to return before that via Ctrl-D.
+    """
+
+    def _run(self, keys: list[int], items: list[str], initial: list[str]) -> Optional[list[str]]:
+        """Run _multiselect_loop with a synthetic screen feeding `keys`."""
+        key_iter = iter(keys)
+
+        class FakeScreen:
+            def erase(self) -> None: pass
+            def getmaxyx(self) -> tuple[int, int]: return (40, 80)
+            def refresh(self) -> None: pass
+            def getch(self) -> int: return next(key_iter)
+            def addstr(self, *a: Any) -> None: pass
+            def keypad(self, *a: Any) -> None: pass
+
+        return _multiselect_loop(FakeScreen(), items, title="", initial=initial)  # type: ignore[arg-type]
+
+    def test_ctrl_d_confirms_initial_selection(self):
+        result = self._run([_KEY_CTRL_D], ["a", "b", "c"], ["a", "b"])
+        self.assertEqual(["a", "b"], result)
+
+    def test_esc_cancels(self):
+        result = self._run([_KEY_ESC], ["a", "b"], ["a"])
+        self.assertIsNone(result)
+
+    def test_tab_then_K_moves_item_up(self):
+        # Start: selected = ["a", "b", "c"]
+        # Tab → order mode (order_cursor=0 on "a")
+        # ↓ → order_cursor=1 (on "b")
+        # K → swap b and a → ["b", "a", "c"], order_cursor=0
+        # Ctrl-D → confirm
+        DOWN = ord("j")
+        result = self._run(
+            [ord("\t"), DOWN, ord("K"), _KEY_CTRL_D],
+            ["a", "b", "c"],
+            ["a", "b", "c"],
+        )
+        self.assertEqual(["b", "a", "c"], result)
+
+    def test_tab_then_J_moves_item_down(self):
+        # selected = ["a", "b", "c"], focus order, cursor=0
+        # J → swap a and b → ["b", "a", "c"], cursor=1
+        # Ctrl-D → confirm
+        result = self._run(
+            [ord("\t"), ord("J"), _KEY_CTRL_D],
+            ["a", "b", "c"],
+            ["a", "b", "c"],
+        )
+        self.assertEqual(["b", "a", "c"], result)
+
+    def test_K_at_top_is_no_op(self):
+        # cursor already at 0, K should not change order
+        result = self._run(
+            [ord("\t"), ord("K"), _KEY_CTRL_D],
+            ["a", "b"],
+            ["a", "b"],
+        )
+        self.assertEqual(["a", "b"], result)
+
+    def test_J_at_bottom_is_no_op(self):
+        DOWN = ord("j")
+        result = self._run(
+            [ord("\t"), DOWN, ord("J"), _KEY_CTRL_D],
+            ["a", "b"],
+            ["a", "b"],
+        )
+        self.assertEqual(["a", "b"], result)
+
+    def test_tab_back_to_filter_then_confirm(self):
+        # Tab → order, Tab → filter, Ctrl-D confirms unchanged
+        result = self._run(
+            [ord("\t"), ord("\t"), _KEY_CTRL_D],
+            ["a", "b"],
+            ["a", "b"],
+        )
+        self.assertEqual(["a", "b"], result)
+
+    def test_space_toggles_item_on(self):
+        # Space on an unselected item selects it; Ctrl-D confirms.
+        result = self._run([_KEY_SPACE, _KEY_CTRL_D], ["a", "b"], [])
+        self.assertEqual(["a"], result)
+
+    def test_space_toggles_item_off(self):
+        # Space on a selected item deselects it; Ctrl-D confirms empty.
+        result = self._run([_KEY_SPACE, _KEY_CTRL_D], ["a", "b"], ["a"])
+        self.assertEqual([], result)
+
+    def test_enter_confirms_without_toggle(self):
+        # Enter immediately confirms the current selection without toggling.
+        result = self._run([_KEY_ENTER], ["a", "b"], ["a"])
+        self.assertEqual(["a"], result)
+
+    def test_enter_confirms_empty_selection(self):
+        result = self._run([_KEY_ENTER], ["a", "b"], [])
+        self.assertEqual([], result)
+
+    def test_space_then_enter_confirms(self):
+        # Space selects "a", Enter confirms.
+        result = self._run([_KEY_SPACE, _KEY_ENTER], ["a", "b"], [])
+        self.assertEqual(["a"], result)
+
+
 if __name__ == "__main__":
     unittest.main()