docs(prd): add PRD 0049 — named/labelled agents

Draft PRD for prompting operators for a custom label and optional
ANSI color at agent launch time, storing both in metadata.json, and
surfacing the label (in color) in the dashboard's active-agents pane.

Closes #171

.codex/skills/quality-eval/SKILL.md

@@ -0,0 +1,76 @@
+---
+name: quality-eval
+description: Use when the user asks to objectively evaluate, score, rate, audit, or quality-gate code, codebases, files, pull requests, or snippets using a strict 5-dimension engineering rubric with scores and refactoring steps.
+metadata:
+  short-description: Score code quality with a strict rubric
+---
+
+# Quality Eval
+
+## Role
+
+Act as a Staff Software Engineer and automated quality gate. Evaluate code objectively against the rubric below, surface hidden anti-patterns, and provide a mathematical grade with atomic refactoring steps.
+
+## Evaluation Rules
+
+- Evaluate only against the five rubric dimensions.
+- Be candid. Do not inflate scores for politeness.
+- Avoid generic advice. Every recommendation must name a specific code location, behavior, or pattern and include a concrete improvement direction.
+- Inspect the code before scoring. For codebases, read enough representative files, tests, and architecture boundaries to justify the scope.
+- When exact line numbers are available, cite them.
+- Do not reveal private chain-of-thought. In the required `Chain of Thought Analysis` section, provide a concise, step-by-step audit rationale with observable findings and score justifications.
+
+## Rubric
+
+Score each dimension from 1 to 5 using these anchors:
+
+| Dimension | Score 1 (Fail) | Score 3 (Pass) | Score 5 (Exemplary) |
+| :--- | :--- | :--- | :--- |
+| **Architecture** | Spaghettified; tight coupling; violated separation of concerns. | Modular but relies on leaky abstractions or mixed domains. | Strict domain isolation; follows SOLID; clear dependency inversion. |
+| **Readability** | Cryptic naming; deep nesting (>3 levels); widespread DRY violations. | Idiomatic but features over-complex functions or sparse documentation. | Self-documenting; expressive naming; high cohesion; flat structure. |
+| **Resilience** | Swallows errors blindly; lacks contextual logging; fragile to bad input. | Basic try/catch blocks present but lacks granular, typed error handling. | Explicit error boundaries; contextual logging; structured failure modes. |
+| **Testability** | Hardcoded dependencies make mocking or isolated testing impossible. | Pure functions are testable, but side-effect heavy logic lacks test hooks. | Decoupled IO; deterministic execution; structured for unit and integration tests. |
+| **SecOps** | Hardcoded secrets; O(n^2) bottlenecks; zero input sanitization. | Safe from obvious flaws but lacks deep defensive optimization. | Validated inputs; optimized algorithmic complexity; zero security debt. |
+
+## Scoring Method
+
+1. Determine the evaluated scope and primary language.
+2. Identify concrete evidence for each dimension.
+3. Assign integer dimension scores from 1 to 5.
+4. Compute `composite_score` as the arithmetic mean of the five dimension scores, rounded to one decimal place.
+5. Include code snippets only when they make a refactoring step more actionable.
+
+## Required Output
+
+Structure every response into exactly these three Markdown sections:
+
+### 1. Chain of Thought Analysis
+
+Provide a concise step-by-step audit rationale. Name specific files, functions, patterns, anti-patterns, and rubric anchors. Keep it evidence-based and do not include hidden private reasoning.
+
+### 2. Normalized Score Report
+
+```json
+{
+  "evaluation_metadata": {
+    "target_scope": "string",
+    "primary_language": "string"
+  },
+  "metrics": {
+    "architecture_and_modularity": 0,
+    "readability_and_maintainability": 0,
+    "error_handling_and_resilience": 0,
+    "testability_and_mocking": 0,
+    "security_and_performance": 0
+  },
+  "composite_score": 0.0
+}
+```
+
+### 3. Atomic Refactoring Playbook
+
+* **High Priority (To lift Score 1/2 to 3):**
+  - [ ] Actionable, specific refactoring step with file/line/context reference.
+* **Medium Priority (To lift Score 3 to 4/5):**
+  - [ ] Optimization or architectural pattern implementation step.
+

.codex/skills/quality-eval/agents/openai.yaml

@@ -0,0 +1,3 @@
+display_name: Quality Eval
+short_description: Scores code quality with a strict five-dimension rubric and refactoring playbook.
+default_prompt: Evaluate this code objectively using the quality-eval rubric and return the three-section score report.

AGENTS.md

@@ -28,14 +28,28 @@ the container lifecycle and the copying of skills and env vars into it.
 - `bot-bottle.json` — legacy manifest of named agents (env / skills / prompt
   per agent), consumed by `cli.py`. See "Manifest" under
   "Intended design".
-- `docs/INDEX.md` — pointer to the research notes.
-- `docs/prds/` — product requirement docs.
-- `docs/research/` — research notes (empty for now, kept tracked via `.gitkeep`).
+- `docs/README.md` — docs overview; when to write which document.
+- `docs/prds/` — product requirement docs (see `docs/prds/README.md` for format).
+- `docs/research/` — research notes (see `docs/research/README.md`).
+- `docs/decisions/` — decision records (ADR-lite).
 
 ## Conventions
 
-- Product requirement docs live in `docs/prds/`.
-- Research notes live in `docs/research/`.
+- Three kinds of doc, each with its own conventions in-folder; see
+  `docs/README.md` for when to write which:
+  - **PRDs** (`docs/prds/`) — one feature per file, numbered
+    `NNNN-kebab.md`. A `Status:` line tracks lifecycle: Draft → Active
+    (shipped to `main`) → Superseded/Retargeted. Format in
+    `docs/prds/README.md`.
+  - **Research notes** (`docs/research/`) — opinionated investigations;
+    unnumbered kebab-case, freeform and verdict-first. See
+    `docs/research/README.md`.
+  - **Decision records** (`docs/decisions/`) — ADR-lite, numbered
+    `NNNN-kebab.md`, for policies and non-feature decisions. See
+    `docs/decisions/README.md`.
+- Keep decision rationale self-contained in the repo, not in Gitea
+  issue threads. Issues are an ephemeral inbox; the durable "why" lives
+  in a PRD, research note, or decision record.
 - Low dependencies by default. The project is Python, stdlib-first (no
   runtime pip dependencies in the package itself; the only language
   runtime is the Python 3.13 used by the CLI + sidecars). Ask before

Dockerfile.claude

@@ -23,7 +23,7 @@ FROM node:22-slim
 # tool (curl itself, plus anything that shells out to it) works
 # against pipelock's bumped TLS without the agent needing local DNS.
 RUN apt-get update \
-  && apt-get install -y --no-install-recommends git ca-certificates openssh-client socat curl dnsutils \
+  && apt-get install -y --no-install-recommends git ca-certificates openssh-client socat curl dnsutils python3 python3-pip python3-venv \
   && rm -rf /var/lib/apt/lists/*
 
 # Install claude-code globally. Pinned to the version verified in the v1

Dockerfile.codex

@@ -6,10 +6,10 @@
 FROM node:22-slim
 
 RUN apt-get update \
-  && apt-get install -y --no-install-recommends git ca-certificates openssh-client socat curl dnsutils \
+  && apt-get install -y --no-install-recommends git ca-certificates openssh-client socat curl dnsutils python3 python3-pip python3-venv \
   && rm -rf /var/lib/apt/lists/*
 
-RUN npm install -g --no-fund --no-audit @openai/codex@0.134.0 \
+RUN npm install -g --no-fund --no-audit @openai/codex@0.136.0 \
   && npm cache clean --force
 
 USER node

Dockerfile.sidecars

@@ -31,6 +31,7 @@
 #   9099  egress (mitmproxy, pipelock's upstream — not externally
 #                 addressed by the agent)
 #   9418  git-gate (git-daemon)
+#   9420  git-gate smart HTTP (smolmachines agent-facing transport)
 #   9100  supervise (MCP HTTP)
 
 # Stage 1: pipelock binary. The upstream pipelock image is a
@@ -81,6 +82,7 @@ COPY bot_bottle/yaml_subset.py       /app/yaml_subset.py
 COPY bot_bottle/supervise.py         /app/supervise.py
 COPY bot_bottle/supervise_server.py  /app/supervise_server.py
 COPY bot_bottle/sidecar_init.py      /app/sidecar_init.py
+COPY bot_bottle/git_http_backend.py  /app/git_http_backend.py
 COPY bot_bottle/egress_entrypoint.sh /app/egress-entrypoint.sh
 RUN chmod +x /app/egress-entrypoint.sh
 
@@ -97,7 +99,7 @@ RUN mkdir -p \
 
 # Documentation only — the compose renderer publishes whichever
 # subset the bottle uses.
-EXPOSE 8888 9099 9418 9100
+EXPOSE 8888 9099 9418 9420 9100
 
 # WORKDIR matches Dockerfile.supervise's prior layout so the
 # in-app same-dir import in supervise_server.py stays deterministic.

README.md

@@ -6,96 +6,26 @@
 
 [![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)
 
-Run multiple Claude Code agents on your own machine, each scoped to its own secrets, skills, and egress allowlist.
+**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.
 
-![pipelock and git-gate blocking exfil attempts against a live bottle](docs/demo.gif)
+**Solution:** Ephemeral, per agent "bottles" the agent cannot modify that scan all traffic for data exfiltration and limit capabilities and egress to only what the agent needs.
 
-Four prompts to the agent inside a real bottle:
-claude replies to `hello there` — proof api.anthropic.com routes
-through pipelock's bumped TLS end-to-end;
-asked to GET a non-allowlisted host, the agent's curl gets 403 back
-from pipelock;
-asked to POST a credential-shaped body to an allowlisted host, the
-same 403 — pipelock's DLP body scanner caught it;
-asked to commit and push an AKIA-shaped key, git-gate's gitleaks
-pre-receive hook rejects the ref.
-Run it yourself with `bash scripts/demo.sh`.
+## Features
 
-## Why "bot-bottle"?
-
-Each container is a bottle; Claude is the genie inside. The genie's
-powers are exactly what the manifest grants it — a specific set of
-skills, a specific set of secrets, and a specific set of hosts it can
-reach — nothing more. You uncork one bottle per agent
-(`./cli.py start <agent>`), many bottles run in parallel, and each is
-scoped to its task. When the session ends the bottle is destroyed and
-the genie does not persist.
-
-## Goals
-
-- Scope each agent to the minimum credentials and network egress its task actually needs
-- Run multiple agents in parallel, isolated from each other
-- Keep code, credentials, and agent activity on infrastructure I control — no third-party agent runtime
-
-## Project status
-
-bot-bottle is a self-hosted secure runtime for AI coding agents.
-Each agent runs in an isolated container or micro-VM-backed bottle with
-scoped secrets, allowlisted egress, TLS-aware proxying, DLP checks, and
-a git-gate that withholds upstream credentials and scans pushes before
-forwarding. The project includes a documented threat model, PRD-driven
-development history, Docker and smolmachines backends, dashboard and
-remediation flows, and unit/integration tests covering exfiltration and
-sandbox escape scenarios.
-
-## Security model
-
-Each agent runs in its own bottle: its own container, its own internal
-Docker network, and its own pipelock sidecar. Bottles don't share
-state, don't talk to each other, and only get the env vars, skills,
-SSH identities, and egress hosts the manifest grants them — nothing
-more. Any one agent only has the access it needs to do its job.
-
-The bottle limits both what an agent can see and where it can send
-it. Each bottle gets only the secrets and SSH identities the manifest
-grants it — a Gitea token but not a GitHub token, a deploy key but
-not a personal SSH key — so even a compromised or misbehaving agent
-only handles credentials it was already trusted with for its job.
-Egress flows through pipelock, which constrains where those
-credentials can travel: an agent with a Gitea token can reach
-`gitea.dideric.is`, not arbitrary attacker-controlled hosts. The same
-constraint blocks DNS-over-HTTPS as an exfil channel — a DoH resolver
-like `cloudflare-dns.com` would have to be on the allowlist for the
-agent to reach it at all. The container itself adds a layer between
-the agent and the host, but the v1 design leans more on secret
-minimization and egress allowlisting than on the container as a
-hardened boundary. On Linux hosts where [gVisor](https://gvisor.dev/)
-is registered with Docker, bot-bottle auto-detects it and launches
-every bottle under `runsc` for a userspace syscall barrier — no
-manifest configuration required. The broader v2 discussion lives in
-`docs/research/stronger-isolation-alternatives.md`.
-
-The egress proxy and OAuth-token handling below are the load-bearing
-pieces of v1.
+- **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.
+- **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.
+- **Trust boundary at `$HOME`** — bottles (credentials, egress, remotes) live only under `~/.bot-bottle/bottles/`. Repos may ship agents but not bottles, so a cloned repo can't redirect an env var to an attacker host.
+- **Composable bottles (`extends:`)** — keep provider/runtime policy in one base bottle (e.g. `claude.md`) and overlay task bottles on top.
+- **Parallel, isolated bottles** — each bottle is its own per-agent Docker `--internal` network; bottles don't share state or talk to each other.
+- **Provider templates (Claude, Codex)** — `Dockerfile.claude` / `Dockerfile.codex`, or a bottle-supplied Dockerfile. Claude auth via long-lived OAuth token; Codex via opt-in host device-auth forwarding.
+- **gVisor auto-detect** — on Linux hosts where `runsc` is registered with Docker, every bottle launches under it for a userspace syscall barrier; no manifest config required.
+- **Smolmachines backend (macOS)** — opt-in `BOT_BOTTLE_BACKEND=smolmachines` runs the agent in a libkrun micro-VM with the sidecar bundle still in Docker.
 
 ## Architecture
 
-A bottle is two containers per agent: an `agent` container, and a
-`sidecars` container that bundles pipelock + egress + git-gate +
-supervise behind a Python init supervisor (PRD 0024). They share a
-per-agent Docker `--internal` network; the agent has no default
-route off-box. All HTTP and HTTPS egress funnels through pipelock,
-where the egress allowlist, TLS interception, and request-body DLP
-scanner enforce the manifest before any byte leaves the host. The
-only egress that doesn't traverse pipelock is git-gate's SSH
-push/fetch to `bottle.git` upstreams — pipelock can't proxy SSH,
-so git-gate is its own L4-style egress path with gitleaks doing
-the pre-receive scan.
-
-The agent dials the bundle by the legacy short names (`pipelock`,
-`egress`, `git-gate`, `supervise`); the renderer registers those as
-docker-network aliases on the bundle so existing HTTPS_PROXY URLs
-and MCP endpoints resolve without an agent-side change.
+A bottle is two containers per agent: an `agent` container, and a `sidecars` container that bundles pipelock + cred-proxy + git-gate + supervise behind a Python init supervisor. They share a per-agent Docker `--internal` network; the agent has no default route off-box.
 
 ```
                             host  ( ./cli.py )
@@ -104,26 +34,21 @@ and MCP endpoints resolve without an agent-side change.
                                   ▼
    ┌─────────────────────────── bottle ──────────────────────────────────┐
    │                                                                     │
-   │   ┌──────────────────┐                                              │
-   │   │ agent image      │  HTTPS_PROXY                                 │
-   │   │ (claude-code,    │ ────────────────────────┐                    │
-   │   │  built locally)  │                         │                    │
-   │   │                  │   plain HTTP            │                    │
-   │   │ skills, env,     │  (token injection) ┌────▼─────────┐          │
-   │   │ ~/.gitconfig,    │ ──────────────────►│ cred-proxy   │          │
-   │   │ ~/.npmrc, tea    │                    │ (strips/inj  │          │
-   │   │                  │                    │  Authoriz.)  │          │
-   │   │ environ: URLs    │                    └─────┬────────┘          │
-   │   │ only, no real    │     HTTPS_PROXY          │                   │
-   │   │ tokens           │                          ▼                   │
-   │   │                  │                  ┌────────────────┐          │  HTTPS to
+   │   ┌──────────────────┐                   ┌──────────────┐           │
+   │   │ agent image      │   HTTP(S) proxy   │ cred-proxy   │           │
+   │   │ (claude-code,    │ ─────────────────►│ (strips/inj  │           │
+   │   │  codex, etc)     │                   │  Authoriz.)  │           │
+   │   │                  │                   └──────┬───────┘           │
+   │   │ environ: URLs    │                          │                   │
+   │   │ only, no real    │                          ▼                   │
+   │   │ tokens           │                  ┌────────────────┐          │  HTTPS to
    │   │                  │                  │ pipelock image │──────────┼──►  allowlisted
    │   │                  │                  │ (TLS bump, DLP │          │     hosts (incl.
    │   │                  │                  │  body scan,    │          │      cred-proxy
    │   │                  │                  │  allowlist)    │          │      upstreams)
    │   │                  │                  └────────────────┘          │
    │   │                  │                                              │
-   │   │                  │   git://         ┌────────────────┐          │  SSH push/fetch
+   │   │                  │    git proxy     ┌────────────────┐          │  SSH push/fetch
    │   │                  │ ────────────────►│ git-gate image │──────────┼──►  to bottle.git
    │   │                  │                  │ (gitleaks +    │          │      upstreams
    │   └──────────────────┘                  │  git daemon)   │          │     (direct — not
@@ -137,198 +62,25 @@ and MCP endpoints resolve without an agent-side change.
    └─────────────────────────────────────────────────────────────────────┘
 ```
 
-- **agent image** — built from the provider template Dockerfile
-  (`Dockerfile.claude` for Claude, `Dockerfile.codex` for Codex, or
-  `agent_provider.dockerfile`) on first run; runs the selected agent
-  CLI with the manifest-granted skills, env vars, and `~/.gitconfig`
-  (the latter for the git-gate's `insteadOf` rules when `bottle.git`
-  is set).
-- **pipelock image** — per-agent sidecar. Terminates the agent's
-  outbound HTTP/HTTPS, enforces the resolved allowlist, runs DLP
-  scanning. Design in `docs/prds/0001-per-agent-egress-proxy-via-pipelock.md`
-  and `docs/prds/0006-pipelock-tls-interception.md`.
-- **git-gate image** — per-agent sidecar built on `zricethezav/gitleaks`
-  (alpine + gitleaks + git-daemon + openssh-client). Runs
-  `git daemon` over `git://` as a bidirectional mirror of each
-  declared upstream. A pre-receive hook gitleaks-scans incoming
-  refs and forwards clean refs to the real upstream over SSH; an
-  access-hook runs `git fetch origin --prune` against the upstream
-  before every upload-pack so an agent fetch returns whatever the
-  upstream has *now* (fail-closed if unreachable). The agent's
-  `~/.gitconfig` rewrites the real URL to the gate via `insteadOf`,
-  so push, fetch, clone, and pull all route through. The agent
-  never sees the upstream credential. If the upstream's hostname
-  isn't resolvable from the gate container (e.g. a Tailscale-only
-  host whose public DNS points elsewhere), pin its IP via
-  `ExtraHosts: { "<hostname>": "<ip>" }` on the `bottle.git` entry —
-  the gate's `/etc/hosts` gets the override while the agent's
-  `insteadOf` rewrite still keys off the original hostname. Brought
-  up only when `bottle.git` has entries. Design in
-  `docs/prds/0008-git-gate.md`.
-- **cred-proxy image** — per-bottle sidecar (`python:3.13-alpine`
-  base, stdlib-only) that holds API tokens declared in
-  `bottle.cred_proxy.routes`. Each route names a `path`,
-  `upstream`, `auth_scheme`, and `token_ref` (host env var); the
-  agent dials `http://cred-proxy:9099<path>...` over plain HTTP
-  and the proxy strips any inbound `Authorization`, injects
-  `<auth_scheme> <token>` using the value held only in its own
-  container's environ, and forwards to the real upstream over
-  HTTPS. SSE responses stream back unbuffered. The cred-proxy's
-  outbound HTTPS routes through pipelock (it trusts pipelock's
-  per-bottle CA), so pipelock's egress allowlist + body scanner
-  apply to cred-proxy traffic the same way they apply to direct
-  agent traffic. Smart-HTTP push paths (`/git-receive-pack`,
-  `/info/refs?service=git-receive-pack`) are refused at the
-  proxy — push must go through `bottle.git` / git-gate where
-  gitleaks runs. Optional per-route `role` tags drive agent-side
-  rewrites: `anthropic-base-url`, `npm-registry`, `git-insteadof`,
-  `tea-login`. The agent's `printenv` shows only proxy URLs —
-  none of the real token values. Design in
-  `docs/prds/0010-cred-proxy.md`.
-
-When the agent exits, `cli.py` tears down every sidecar that was
-brought up and the two networks; nothing about a bottle persists
-between runs.
+When the agent exits, `cli.py` tears down every sidecar and both networks; nothing about a bottle persists between runs.
 
 ## Quickstart
 
-Requires Docker on the host and a long-lived Claude Code OAuth token in
-your shell env.
+Requires Docker on the host and a long-lived Claude Code OAuth token (`claude setup-token`) exported as `BOT_BOTTLE_CLAUDE_OAUTH_TOKEN`.
 
 ```sh
 ./cli.py start <agent>   # builds the image on first run, drops you into claude
 ```
 
-The container is removed automatically when the session ends. If the script
-is killed with SIGKILL the exit trap won't fire and the container may be
-left running; remove it with `docker rm -f <container-name>`.
-
-### Smolmachines backend (experimental, macOS-only)
-
-A second backend runs the agent in a smolvm micro-VM (libkrun) with the
-sidecar bundle still in Docker. Selected via
-`BOT_BOTTLE_BACKEND=smolmachines ./cli.py start <agent>`. Requires
-`smolvm` on PATH (`curl -sSL https://smolmachines.com/install.sh | sh`).
-
-The integration tests run against whichever backend the env var
-selects and skip cleanly when its prerequisites are missing.
-
-**One-time sudo on first launch (macOS):** smolmachines bottles
-each reserve a loopback alias from a pool (`127.0.0.16` ..
-`127.0.0.31`) and bind their bundle's port-forwards to it; the
-first `./cli.py start` after each reboot prompts for sudo to add
-missing aliases via `ifconfig lo0 alias`. Aliases persist until
-reboot; subsequent launches don't prompt. The agent's TSI
-allowlist is the alias's `/32`, so each bottle can only reach
-its own bundle's published ports — not other bottles' ports,
-not other host loopback services (postgres, dev servers, etc.).
-
-This enforcement requires a workaround for a smolvm 0.8.0 bug:
-the CLI's `--allow-cidr` flag is silently dropped when combined
-with `--from <smolmachine>`. The launcher patches smolvm's
-persistent state DB
-(`~/Library/Application Support/smolvm/server/smolvm.db`)
-directly between `machine create` and `machine start` to set
-the allowlist. The hack falls away automatically when smolvm
-honors the flag upstream — see the `loopback_alias` module's
-docstring for the investigation trail.
-
 ## Manifest
 
-Bottles and agents live as Markdown files with YAML frontmatter under
-`~/.bot-bottle/`. Each bottle is one file in `bottles/`, each agent
-is one file in `agents/`:
+Bottles and agents are Markdown files with YAML frontmatter under `~/.bot-bottle/`. The Markdown body is the system prompt. Bottles live in `~/.bot-bottle/bottles/`; agents may also be shipped by a repo at `<repo>/.bot-bottle/agents/<name>.md`.
 
-```
-~/.bot-bottle/
-├── bottles/
-│   ├── dev.md
-│   └── gitea-dev.md
-└── agents/
-    ├── implementer.md
-    └── researcher.md
-```
-
-The filename (without `.md`) is the entity's name. Filenames must
-match `[a-z][a-z0-9-]*`; files that don't are skipped with a warning.
-
-A repo can ship its own agent files alongside its code at
-`<repo>/.bot-bottle/agents/<name>.md`. Those agents reference
-bottles defined in `~/.bot-bottle/bottles/` (the only place
-bottles can come from); a `bottles/` subdir in a repo is ignored
-with a warning. **This is the trust boundary**: bottle infrastructure
-— credentials, egress allowlists, git remotes — comes from your home
-directory only. A cloned repo cannot redirect a host env var to an
-attacker-named upstream because it has no way to declare a bottle.
-
-### Bottle composition with `extends:`
-
-A bottle can inherit from another via `extends: <bottle-name>` so
-operators don't have to duplicate a whole bottle file to vary one
-field (PRD 0025). The parent's resolved config is the base; the
-child's declared fields overlay. Merge rules:
-
-- `env:` — dict merge, child wins on key collision.
-- `git.user:` — per-field overlay (child's non-empty `name` /
-  `email` wins; empty falls through to parent).
-- `git.remotes:` — dict merge by host, child wins on host collision.
-  An explicit `git.remotes: {}` clears the parent's remotes; omitting
-  `git.remotes` inherits the parent's remotes.
-- `agent_provider:`, `egress:`, `supervise:` — full replace when the
-  child declares the field.
-
-```yaml
----
-extends: dev          # inherit everything from bottles/dev.md
-egress:
-  routes:
-    - host: staging.example.com
-      auth:
-        scheme: Bearer
-        token_ref: STAGING_TOKEN
----
-```
-
-Cycles (`A extends B extends A`), self-references, and missing
-parents die at parse with a clear pointer. Bottles remain
-`$HOME`-only — `extends:` preserves the trust boundary above.
-
-### Provider base bottles
-
-Keep provider/runtime policy in one home-owned base bottle, then have
-task bottles extend it. That keeps provider egress/auth in one place
-without hiding security-relevant routes behind `agent_provider.template`.
-
-For example, `~/.bot-bottle/bottles/claude.md` can hold the Claude
-provider selection and Anthropic API egress:
+**Bottle** (`~/.bot-bottle/bottles/gitea-dev.md`):
 
 ````markdown
 ---
-agent_provider:
-  template: claude
-
-egress:
-  routes:
-    - host: api.anthropic.com
-      role: claude_code_oauth
-      auth:
-        scheme: Bearer
-        token_ref: BOT_BOTTLE_CLAUDE_OAUTH_TOKEN
-      pipelock:
-        tls_passthrough: true
----
-
-Common Claude provider boundary.
-````
-
-Task bottles can then inherit that provider boundary and add their own
-env/git configuration without repeating the Claude route.
-
-### Example bottle (`~/.bot-bottle/bottles/gitea-dev.md`)
-
-````markdown
----
-extends: claude
+extends: claude          # inherit the Claude provider boundary
 
 env:
   GIT_AUTHOR_NAME: didericis
@@ -343,19 +95,22 @@ git:
       Upstream: ssh://git@gitea.dideric.is:30009/didericis/bot-bottle.git
       IdentityFile: /Users/didericis/.ssh/id_ed25519_gitea
       KnownHostKey: ssh-ed25519 AAAA...
+
+egress:
+  routes:
+    - host: gitea.dideric.is
+      auth:
+        scheme: token
+        token_ref: BOT_BOTTLE_GITEA_TOKEN
+      pipelock:
+        ssrf_ip_allowlist: [100.78.141.42/32]
 ---
 
-The `gitea-dev` bottle. Backs my work on personal projects: provider
-auth through egress and gitea.dideric.is over SSH.
+The `gitea-dev` bottle. Provider auth via the inherited Claude route;
+gitea over SSH for push, token over HTTPS for the API.
 ````
 
-For a Codex-backed base bottle, set `agent_provider.template: codex`
-and use the `codex_auth` egress role for the OpenAI API route. The
-built-in Codex template uses `Dockerfile.codex`; set
-`agent_provider.dockerfile` to build the agent from a custom
-Dockerfile while keeping the bot-bottle sidecars in place.
-
-### Example agent (`~/.bot-bottle/agents/gitea-helper.md`)
+**Agent** (`~/.bot-bottle/agents/gitea-helper.md`):
 
 ````markdown
 ---
@@ -367,117 +122,12 @@ skills:
 You help maintain Gitea-hosted projects.
 ````
 
-The agent's Markdown body is its system prompt (whitespace
-stripped). The frontmatter declares the bottle to launch in and any
-skills to mount. You can also include Claude Code subagent fields
-(`name`, `description`, `model`, `color`, `memory`) in the
-frontmatter — bot-bottle ignores them at launch but doesn't
-reject them, so the same file can drop into `~/.claude/agents/` as a
-Claude Code subagent.
-
-Unknown top-level frontmatter keys die at load with a "did you mean"
-pointer; typos don't silently ghost into an empty config.
-
-The YAML subset the frontmatter accepts is bounded (flat keys,
-strings / ints / true-or-false bools / null / lists / one-level
-nested dicts). Anchors, multi-line block scalars, tags, and
-ambiguous bare strings (`yes` / `NO` / `2026-05-24` /
-`0x...`) all die with a clear pointer at the spec — quote your
-strings when in doubt. The full schema lives in
-`bot_bottle/yaml_subset.py` (~450 lines, stdlib-only, no PyYAML).
-
-Working examples live under `examples/`. Pipelock's design lives in
-`docs/prds/0001-per-agent-egress-proxy-via-pipelock.md` and the
-rationale in `docs/research/pipelock-assessment.md`. The trust
-boundary rationale lives in `docs/prds/0011-per-file-md-manifest.md`.
-
-## Auth: Claude OAuth token, not API key
-
-Bottles that use `agent_provider.template: claude` authenticate
-`claude` inside the container with the same Pro/Max subscription you
-already use on the host, via a long-lived OAuth token. No
-`ANTHROPIC_API_KEY` is needed.
-
-**Why a token instead of mounting `~/.claude.json`:** on macOS, Claude
-Code stores OAuth credentials in the encrypted Keychain, not in
-`~/.claude.json`. Mounting that file into a Linux container does not
-carry the credentials with it. Linux hosts keep credentials in
-`~/.claude/.credentials.json`, but to keep the launcher portable
-bot-bottle uses the env-var path on every host.
-
-**One-time setup on the host:**
-
-```sh
-claude setup-token   # browser login, prints a ~1-year OAuth token
-```
-
-Stash the token in your shell env (e.g. `~/.zshrc` or a secret manager)
-as `BOT_BOTTLE_CLAUDE_OAUTH_TOKEN`:
-
-```sh
-export BOT_BOTTLE_CLAUDE_OAUTH_TOKEN="<token>"
-```
-
-The Claude bottle reaches the Anthropic API only through the cred-proxy
-sidecar. To let `claude` authenticate, declare an egress route with
-`role: claude_code_oauth` and
-`token_ref: BOT_BOTTLE_CLAUDE_OAUTH_TOKEN`:
-
-```yaml
-egress:
-  routes:
-    - host: api.anthropic.com
-      role: claude_code_oauth
-      auth:
-        scheme: Bearer
-        token_ref: BOT_BOTTLE_CLAUDE_OAUTH_TOKEN
-      pipelock:
-        tls_passthrough: true
-```
-
-Routes that resolve to private or Tailscale addresses can opt into
-pipelock's SSRF destination allowlist explicitly:
-
-```yaml
-egress:
-  routes:
-    - host: gitea.dideric.is
-      auth:
-        scheme: token
-        token_ref: BOT_BOTTLE_GITEA_TOKEN
-      pipelock:
-        ssrf_ip_allowlist:
-          - 100.78.141.42/32
-```
-
-At launch, `cli.py` reads `BOT_BOTTLE_CLAUDE_OAUTH_TOKEN` from the host
-env and forwards it into the cred-proxy container's environ — never
-into the agent's. The agent receives `ANTHROPIC_BASE_URL` pointing at
-`http://cred-proxy:9099/anthropic` and a non-secret placeholder for
-`CLAUDE_CODE_OAUTH_TOKEN` (claude-code refuses to start without one;
-the proxy strips and replaces the header on every request). `printenv`
-inside the agent does not surface the real token, and the value is
-never written to disk or placed on argv on the host.
-
-A Claude bottle without a `claude_code_oauth` route has no path to the
-Anthropic API — there is no fallback that forwards the token directly
-to the agent. Caveats: the token is bound to your subscription tier
-(Pro/Max/Team/Enterprise), it does not work with `claude --bare`
-(which only reads `ANTHROPIC_API_KEY`), and if it leaks, regenerate
-via `claude setup-token` again. Reference:
-<https://code.claude.com/docs/en/authentication>.
+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 is an independent project and is not affiliated with,
-endorsed by, or sponsored by Anthropic, PBC. "Claude" and "Claude
-Code" are trademarks of Anthropic, PBC; the project name uses
-"claude" descriptively to indicate that the tool runs Claude Code
-inside a sandbox.
+bot-bottle is an independent project and is not affiliated with, endorsed by, or sponsored by Anthropic, PBC. "Claude" and "Claude Code" are trademarks of Anthropic, PBC; the project name uses "claude" descriptively to indicate that the tool runs Claude Code inside a sandbox.
 
 ## License
 
-Copyright 2026 Eric Bauerfeld
-
-Licensed under the Apache License, Version 2.0. See [LICENSE](LICENSE)
-for the full text.
+Copyright 2026 Eric Bauerfeld. Licensed under the Apache License, Version 2.0. See [LICENSE](LICENSE) for the full text.

bot-bottle.demo.json

@@ -4,14 +4,15 @@
       "env": {
         "FAKE_TOKEN": "ghp_aB3cD4eF5gH6iJ7kL8mN9oP0qR1sT2uV3wX4yZ"
       },
-      "git": [
-        {
-          "Name": "foo",
-          "Upstream": "ssh://git@upstream.invalid/path.git",
-          "IdentityFile": "~/.cache/bot-bottle-demo/fake-key",
-          "KnownHostKey": "ssh-ed25519 AAAAEXAMPLE"
+      "git-gate": {
+        "repos": {
+          "foo": {
+            "url": "ssh://git@upstream.invalid/path.git",
+            "identity": "~/.cache/bot-bottle-demo/fake-key",
+            "host_key": "ssh-ed25519 AAAAEXAMPLE"
+          }
         }
-      ]
+      }
     }
   },
 

bot_bottle/agent_provider.py

@@ -3,18 +3,42 @@
 The manifest owns the user-facing AgentProvider shape. This module is
 the launch-time table that turns a provider template into an executable
 command, default image, and prompt/auth behavior.
+
+Per PRD 0050 the per-provider implementations live under
+`bot_bottle/contrib/<template>/agent_provider.py`. This module exposes:
+
+  - `AgentProvider` (ABC) — the contract each plugin implements.
+  - `get_provider(template)` — lazy-imported registry; the analogue
+    of `bot_bottle/deploy_key_provisioner.get_provisioner`.
+  - `AgentProvisionPlan` (+ helper dataclasses) — declarative shape
+    each provider produces and the backends consume unchanged.
+  - `agent_provision_plan` / `runtime_for` — thin wrappers around the
+    registry kept so existing callers keep working without per-call
+    edits.
 """
 
 from __future__ import annotations
 
-from dataclasses import dataclass
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Literal
+from typing import TYPE_CHECKING, Literal
+
+from .egress import EgressRoute
+
+
+if TYPE_CHECKING:
+    from .backend import Bottle, BottlePlan
 
 
 PROVIDER_CLAUDE = "claude"
 PROVIDER_CODEX = "codex"
 PROVIDER_TEMPLATES = frozenset({PROVIDER_CLAUDE, PROVIDER_CODEX})
+
+# Hosts that egress injects the host ChatGPT bearer on when Codex
+# forward_host_credentials is enabled. Pipelock must pass these through
+# (no TLS MITM) or its header DLP blocks the injected JWT.
+CODEX_HOST_CREDENTIAL_HOSTS = ("api.openai.com", "chatgpt.com")
 PromptMode = Literal["append_file", "read_prompt_file"]
 
 
@@ -24,47 +48,176 @@ class AgentProviderRuntime:
     command: str
     image: str
     dockerfile: str
-    auth_role: str
-    placeholder_env: str
     prompt_mode: PromptMode
     bypass_args: tuple[str, ...]
     resume_args: tuple[str, ...]
     remote_control_args: tuple[str, ...]
 
 
-_REPO_ROOT = Path(__file__).resolve().parent.parent
+@dataclass(frozen=True)
+class AgentProvisionDir:
+    guest_path: str
+    mode: str = "700"
+    owner: str = "node:node"
 
 
-_RUNTIMES = {
-    PROVIDER_CLAUDE: AgentProviderRuntime(
-        template=PROVIDER_CLAUDE,
-        command="claude",
-        image="bot-bottle-claude:latest",
-        dockerfile=str(_REPO_ROOT / "Dockerfile.claude"),
-        auth_role="claude_code_oauth",
-        placeholder_env="CLAUDE_CODE_OAUTH_TOKEN",
-        prompt_mode="append_file",
-        bypass_args=("--dangerously-skip-permissions",),
-        resume_args=("--continue",),
-        remote_control_args=("--remote-control",),
-    ),
-    PROVIDER_CODEX: AgentProviderRuntime(
-        template=PROVIDER_CODEX,
-        command="codex",
-        image="bot-bottle-codex:latest",
-        dockerfile=str(_REPO_ROOT / "Dockerfile.codex"),
-        auth_role="codex_auth",
-        placeholder_env="OPENAI_API_KEY",
-        prompt_mode="read_prompt_file",
-        bypass_args=("--dangerously-bypass-approvals-and-sandbox",),
-        resume_args=("resume", "--last"),
-        remote_control_args=(),
-    ),
-}
+@dataclass(frozen=True)
+class AgentProvisionFile:
+    host_path: Path
+    guest_path: str
+    mode: str = "600"
+    owner: str = "node:node"
+
+
+@dataclass(frozen=True)
+class AgentProvisionCommand:
+    argv: tuple[str, ...]
+    error: str = ""
+
+
+@dataclass(frozen=True)
+class AgentProvisionPlan:
+    """Provider-owned guest setup.
+
+    Backends interpret this plan with their own copy/exec primitives.
+    Provider-specific content stays here so future provider plugins can
+    return the same shape without adding backend-plan fields.
+
+    `egress_routes` are provider-declared EgressRoutes that backends
+    pass to `Egress.prepare` and `PipelockProxy.prepare`. This keeps
+    provider logic out of the egress and pipelock modules — they merge
+    provider routes generically without knowing the provider type.
+
+    `hidden_env_names` is the set of env var names the provider injected
+    as non-secret placeholders. `print_util.visible_agent_env_names` uses
+    this to suppress them from the preflight summary so operators don't
+    mistake them for real credentials.
+    """
+
+    template: str
+    command: str
+    prompt_mode: PromptMode
+    image: str
+    dockerfile: str
+    guest_env: dict[str, str]
+    env_vars: dict[str, str] = field(default_factory=dict)
+    dirs: tuple[AgentProvisionDir, ...] = ()
+    files: tuple[AgentProvisionFile, ...] = ()
+    pre_copy: tuple[AgentProvisionCommand, ...] = ()
+    verify: tuple[AgentProvisionCommand, ...] = ()
+    egress_routes: tuple[EgressRoute, ...] = ()
+    hidden_env_names: frozenset[str] = field(default_factory=frozenset)
+    provisioned_env: dict[str, str] = field(default_factory=dict)
+
+
+class AgentProvider(ABC):
+    """Per-template plugin: produces the provision plan and applies
+    the provider-specific in-guest setup steps (skills, prompt, the
+    declarative `dirs`/`files`/`pre_copy`/`verify` apply loop, and
+    supervise MCP registration). Concrete subclasses live under
+    `bot_bottle/contrib/<template>/agent_provider.py`."""
+
+    @property
+    @abstractmethod
+    def runtime(self) -> AgentProviderRuntime:
+        """The static command / image / prompt-mode table for this
+        template."""
+
+    @abstractmethod
+    def provision_plan(
+        self,
+        *,
+        dockerfile: str,
+        state_dir: Path,
+        guest_home: str,
+        guest_env: dict[str, str] | None = None,
+        auth_token: str = "",
+        forward_host_credentials: bool = False,
+        host_env: dict[str, str] | None = None,
+        trusted_project_path: str = "",
+    ) -> AgentProvisionPlan:
+        """Build the declarative AgentProvisionPlan for one launch.
+        Backends call this during `prepare` and consume the result as
+        before."""
+
+    @abstractmethod
+    def provision_skills(self, plan: "BottlePlan", bottle: "Bottle") -> None:
+        """Copy each of the agent's named skills from the host into
+        the guest. No-op when the agent has no skills. The in-guest
+        layout is provider-specific (claude-code's
+        `~/.claude/skills/` today; future providers may differ)."""
+
+    @abstractmethod
+    def provision_prompt(self, plan: "BottlePlan", bottle: "Bottle") -> str | None:
+        """Copy the prompt file into the guest, fix ownership/mode,
+        and return the in-guest path iff the agent has a non-empty
+        prompt (drives the `--append-system-prompt-file` flag).
+
+        The file is copied either way so the path always exists."""
+
+    @abstractmethod
+    def provision(self, plan: "BottlePlan", bottle: "Bottle") -> None:
+        """Apply the provider's declarative
+        `dirs`/`pre_copy`/`files`/`verify` steps from
+        `plan.agent_provision`. Was called `provision_provider_auth`
+        on `BottleBackend` before PRD 0050."""
+
+    @abstractmethod
+    def provision_supervise_mcp(
+        self,
+        plan: "BottlePlan",
+        bottle: "Bottle",
+        supervise_url: str,
+    ) -> None:
+        """Register the per-bottle supervise sidecar as an MCP server
+        in the provider's in-guest config. Called by the backend after
+        the supervise sidecar is reachable. No-op when
+        `plan.supervise_plan is None`."""
+
+
+def get_provider(template: str) -> AgentProvider:
+    """Resolve a provider template name to its plugin instance.
+
+    Lazy-imports the contrib module so importing this module doesn't
+    pull provider-specific code paths in. Mirrors the contrib
+    convention PRD 0048 established for deploy key provisioners."""
+    if template == PROVIDER_CLAUDE:
+        from .contrib.claude.agent_provider import ClaudeAgentProvider
+        return ClaudeAgentProvider()
+    if template == PROVIDER_CODEX:
+        from .contrib.codex.agent_provider import CodexAgentProvider
+        return CodexAgentProvider()
+    raise ValueError(f"unknown agent provider template: {template!r}")
 
 
 def runtime_for(template: str) -> AgentProviderRuntime:
-    return _RUNTIMES[template]
+    return get_provider(template).runtime
+
+
+def agent_provision_plan(
+    *,
+    template: str,
+    dockerfile: str,
+    state_dir: Path,
+    guest_home: str,
+    guest_env: dict[str, str] | None = None,
+    auth_token: str = "",
+    forward_host_credentials: bool = False,
+    host_env: dict[str, str] | None = None,
+    trusted_project_path: str = "",
+) -> AgentProvisionPlan:
+    """Back-compat shim — `prepare` callers stay the same; the work
+    now lives on the provider plugin."""
+    return get_provider(template).provision_plan(
+        dockerfile=dockerfile,
+        state_dir=state_dir,
+        guest_home=guest_home,
+        guest_env=guest_env,
+        auth_token=auth_token,
+        forward_host_credentials=forward_host_credentials,
+        host_env=host_env,
+        trusted_project_path=trusted_project_path,
+    )
 
 
 def prompt_args(

bot_bottle/backend/__init__.py

@@ -32,15 +32,22 @@ manifest does not carry a backend field; the host picks.
 from __future__ import annotations
 
 import os
+import sys
 from abc import ABC, abstractmethod
 from contextlib import AbstractContextManager
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, Generic, Sequence, TypeVar
 
-from ..log import die
+from ..agent_provider import AgentProvisionPlan, get_provider
+from ..egress import EgressPlan
+from ..git_gate import GitGatePlan
+from ..log import die, info
 from ..manifest import GitEntry, Manifest
+from ..supervise import SupervisePlan
 from ..util import expand_tilde
+from ..workspace import WorkspacePlan
+from .print_util import print_multi, visible_agent_env_names
 from .util import host_skill_dir
 
 
@@ -65,15 +72,58 @@ class BottleSpec:
 @dataclass(frozen=True)
 class BottlePlan(ABC):
     """Base output of a backend's prepare step. Concrete subclasses
-    (e.g. DockerBottlePlan) add backend-specific resolved fields and
-    implement `print`."""
+    (e.g. DockerBottlePlan) add backend-specific resolved fields."""
 
     spec: BottleSpec
     stage_dir: Path
+    guest_home: str
+    git_gate_plan: GitGatePlan
+    egress_plan: EgressPlan
+    supervise_plan: SupervisePlan | None
+    agent_provision: AgentProvisionPlan
+    workspace_plan: WorkspacePlan
 
-    @abstractmethod
     def print(self, *, remote_control: bool) -> 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)
+
+        env_names = visible_agent_env_names(
+            sorted(
+                set(bottle.env.keys())
+                | set(self.agent_provision.guest_env.keys())
+            ),
+            hidden_env_names=self.agent_provision.hidden_env_names,
+        )
+
+        print(file=sys.stderr)
+        info(f"agent           : {spec.agent_name}")
+        info(f"provider        : {self.agent_provision.template}")
+        print_multi("env             ", env_names)
+        print_multi("skills          ", list(agent.skills))
+        info(f"bottle          : {agent.bottle}")
+
+        identity = manifest.git_identity_summary(spec.agent_name)
+        if identity:
+            info(f"  git identity  : {identity}")
+
+        git_lines = [
+            f"{u.name} → {u.upstream_host}:{u.upstream_port}"
+            for u in self.git_gate_plan.upstreams
+        ]
+        if git_lines:
+            print_multi("  git gate      ", git_lines)
+
+        if self.egress_plan.routes:
+            egress_lines = []
+            for r in self.egress_plan.routes:
+                auth = f" [auth:{r.auth_scheme}]" if r.auth_scheme else ""
+                egress_lines.append(f"{r.host}{auth}")
+            print_multi("  egress        ", egress_lines)
+        print(file=sys.stderr)
 
 
 @dataclass(frozen=True)
@@ -263,35 +313,44 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
     def launch(self, plan: PlanT) -> AbstractContextManager[Bottle]:
         """Build/run the bottle and yield a handle; tear down on exit."""
 
-    def provision(self, plan: PlanT, target: str) -> str | None:
+    def provision(self, plan: PlanT, bottle: "Bottle") -> str | None:
         """Copy host-side files (CA cert, prompt, skills, .git) into
         the running bottle. Called from `launch` after the container
-        / machine is up. `target` identifies the running instance in
-        backend-specific terms (Docker: resolved container name; fly:
-        machine id). Returns the in-container prompt path if a prompt
-        was provisioned, else None — the Bottle handle uses it to
-        decide whether to add provider-specific prompt args to the agent's
-        argv.
+        / machine is up. Returns the in-container prompt path if a
+        prompt was provisioned, else None — the Bottle handle uses it
+        to decide whether to add provider-specific prompt args to the
+        agent's argv.
 
-        Default orchestration: ca → prompt → skills → git →
-        supervise. CA install runs first so the agent's trust store
-        is rebuilt before anything inside the agent makes a TLS call.
-        Subclasses typically don't override this; they implement the
-        sub-methods below.
+        Default orchestration: ca → prompt → provider apply → skills
+        → workspace → git → supervise-mcp. CA install runs first so
+        the agent's trust store is rebuilt before anything inside the
+        agent makes a TLS call.
+
+        Per PRD 0050 the per-provider steps (prompt, skills,
+        declarative provision-plan apply, supervise MCP registration)
+        live on the `AgentProvider` plugin. The backend only owns the
+        steps that are about backend infrastructure (CA, workspace,
+        git) and surfaces the supervise sidecar URL its launch step
+        knows about via `supervise_mcp_url`.
 
         PRD 0017: cred-proxy's agent-side dotfile rewrites (~/.npmrc,
         ~/.gitconfig insteadOf, tea config) are gone. Egress-proxy is
         on the agent's HTTP_PROXY path so every tool that respects
         HTTPS_PROXY (claude-code, git over HTTPS, npm, curl) is
         intercepted without per-tool reconfiguration."""
-        self.provision_ca(plan, target)
-        prompt_path = self.provision_prompt(plan, target)
-        self.provision_skills(plan, target)
-        self.provision_git(plan, target)
-        self.provision_supervise(plan, target)
+        provider = get_provider(plan.agent_provision.template)
+        self.provision_ca(plan, bottle)
+        prompt_path = provider.provision_prompt(plan, bottle)
+        provider.provision(plan, bottle)
+        provider.provision_skills(plan, bottle)
+        self.provision_workspace(plan, bottle)
+        self.provision_git(plan, bottle)
+        provider.provision_supervise_mcp(
+            plan, bottle, self.supervise_mcp_url(plan),
+        )
         return prompt_path
 
-    def provision_ca(self, plan: PlanT, target: str) -> None:
+    def provision_ca(self, plan: PlanT, bottle: "Bottle") -> None:
         """Install the per-bottle CA into the agent's trust store so
         the agent trusts the bumped CONNECT cert egress (was
         pipelock, pre-PRD-0017) presents. Default impl is a no-op so
@@ -300,29 +359,26 @@ class BottleBackend(ABC, Generic[PlanT, CleanupT]):
         backend overrides to docker-cp the cert in and run
         `update-ca-certificates`."""
 
-    @abstractmethod
-    def provision_prompt(self, plan: PlanT, target: str) -> str | None:
-        """Copy the prompt file into the running bottle. Returns the
-        in-container path iff the agent has a non-empty prompt;
-        callers use the return value to decide whether to add
-        provider-specific prompt args to the agent's argv."""
+    def provision_workspace(self, plan: PlanT, bottle: "Bottle") -> None:
+        """Copy the operator workspace into the running bottle when
+        the backend cannot bake it into the agent image. Default is
+        no-op for backends like Docker that handle this before launch."""
 
     @abstractmethod
-    def provision_skills(self, plan: PlanT, target: str) -> None:
-        """Copy the agent's named skills from the host into the
-        running bottle. No-op when the agent has no skills."""
-
-    @abstractmethod
-    def provision_git(self, plan: PlanT, target: str) -> None:
+    def provision_git(self, plan: PlanT, bottle: "Bottle") -> None:
         """Copy the host's cwd `.git` directory into the running
         bottle if the user requested --cwd. No-op otherwise."""
 
-    def provision_supervise(self, plan: PlanT, target: str) -> None:
-        """Write the in-bottle Claude Code MCP config so the agent
-        discovers the per-bottle supervise sidecar (PRD 0013).
-        No-op when bottle.supervise is False or the backend doesn't
-        support the supervise sidecar yet. The Docker backend
-        overrides."""
+    def supervise_mcp_url(self, plan: PlanT) -> str:
+        """Return the agent-side URL of the per-bottle supervise
+        sidecar, or "" when this bottle has no sidecar. The provider
+        plugin's `provision_supervise_mcp` uses it to register the
+        MCP entry inside the guest.
+
+        Default returns "" so backends without supervise support
+        don't have to implement it. Docker and smolmachines override."""
+        del plan
+        return ""
 
     @abstractmethod
     def prepare_cleanup(self) -> CleanupT:
@@ -413,14 +469,20 @@ def enumerate_active_agents() -> list[ActiveAgent]:
     """All currently-running agents, across every available
     backend. Used by CLI `list active` and the dashboard's agents
     pane so neither has to know which backends exist. Skips
-    backends whose `is_available()` reports False. Ordered by
-    backend name, then by whatever each backend's
-    `enumerate_active` returns."""
+    backends whose `is_available()` reports False.
+
+    Sorted by `(started_at, slug)` so the list is stable across
+    dashboard refresh ticks — agents don't shift position while
+    the operator navigates with arrow keys. ISO 8601 timestamps
+    sort lexicographically in chronological order; `slug` is the
+    deterministic tiebreaker. Agents with missing metadata
+    (`started_at == ""`) sort first."""
     out: list[ActiveAgent] = []
     for name in known_backend_names():
         if not has_backend(name):
             continue
         out.extend(_BACKENDS[name].enumerate_active())
+    out.sort(key=lambda a: (a.started_at, a.slug))
     return out
 
 

bot_bottle/backend/docker/backend.py

@@ -9,6 +9,12 @@ This module is a thin façade. The real work lives in four siblings:
 
 The base class's `prepare` template runs cross-backend host-side
 validation before calling `_resolve_plan` here.
+
+Per PRD 0050 the per-provider provisioning steps (prompt, skills,
+the declarative provision-plan apply, supervise MCP registration)
+live on the `AgentProvider` plugin under `bot_bottle/contrib/`. The
+Docker backend only owns the steps that are about backend
+infrastructure: CA install and git copy-in.
 """
 
 from __future__ import annotations
@@ -18,7 +24,8 @@ from contextlib import contextmanager
 from pathlib import Path
 from typing import Generator, Sequence
 
-from .. import ActiveAgent, BottleBackend, BottleSpec
+from ...supervise import SUPERVISE_HOSTNAME, SUPERVISE_PORT
+from .. import ActiveAgent, Bottle, BottleBackend, BottleSpec
 from . import cleanup as _cleanup
 from . import enumerate as _enumerate
 from . import launch as _launch
@@ -28,9 +35,6 @@ from .bottle_cleanup_plan import DockerBottleCleanupPlan
 from .bottle_plan import DockerBottlePlan
 from .provision import ca as _ca
 from .provision import git as _git
-from .provision import prompt as _prompt
-from .provision import skills as _skills
-from .provision import supervise as _supervise_prov
 
 
 class DockerBottleBackend(BottleBackend["DockerBottlePlan", "DockerBottleCleanupPlan"]):
@@ -56,20 +60,19 @@ class DockerBottleBackend(BottleBackend["DockerBottlePlan", "DockerBottleCleanup
         with _launch.launch(plan, provision=self.provision) as bottle:
             yield bottle
 
-    def provision_ca(self, plan: DockerBottlePlan, target: str) -> None:
-        _ca.provision_ca(plan, target)
+    def provision_ca(self, plan: DockerBottlePlan, bottle: Bottle) -> None:
+        _ca.provision_ca(plan, bottle)
 
-    def provision_prompt(self, plan: DockerBottlePlan, target: str) -> str | None:
-        return _prompt.provision_prompt(plan, target)
+    def provision_git(self, plan: DockerBottlePlan, bottle: Bottle) -> None:
+        _git.provision_git(plan, bottle)
 
-    def provision_skills(self, plan: DockerBottlePlan, target: str) -> None:
-        _skills.provision_skills(plan, target)
-
-    def provision_git(self, plan: DockerBottlePlan, target: str) -> None:
-        _git.provision_git(plan, target)
-
-    def provision_supervise(self, plan: DockerBottlePlan, target: str) -> None:
-        _supervise_prov.provision_supervise(plan, target)
+    def supervise_mcp_url(self, plan: DockerBottlePlan) -> str:
+        """Docker bottles reach the supervise sidecar via the
+        compose-network alias `supervise:9100`. No per-bottle URL
+        plumbing needed; the alias resolves inside the bridge."""
+        if plan.supervise_plan is None:
+            return ""
+        return f"http://{SUPERVISE_HOSTNAME}:{SUPERVISE_PORT}/"
 
     def prepare_cleanup(self) -> DockerBottleCleanupPlan:
         return _cleanup.prepare_cleanup()

bot_bottle/backend/docker/bottle_plan.py

@@ -2,30 +2,25 @@
 
 Carries the Docker-specific resolved fields produced by
 DockerBottleBackend.prepare. The launch step consumes it without
-further resolution; show_plan-style rendering is the `print` method.
+further resolution; preflight rendering is inherited from BottlePlan.
 """
 
 from __future__ import annotations
 
-import sys
 from dataclasses import dataclass, field
 from pathlib import Path
 
 from ...agent_provider import PromptMode
-from ...egress import EgressPlan
-from ...git_gate import GitGatePlan
-from ...log import info
 from ...pipelock import PipelockProxyPlan
-from ...supervise import SupervisePlan
 from .. import BottlePlan
-from ..print_util import print_multi, visible_agent_env_names
 
 
 @dataclass(frozen=True)
 class DockerBottlePlan(BottlePlan):
     """Docker-specific resolved fields produced by
-    DockerBottleBackend.prepare. Inherits `spec` and `stage_dir` from
-    BottlePlan."""
+    DockerBottleBackend.prepare. Inherits `spec`, `stage_dir`,
+    `git_gate_plan`, `egress_plan`, `supervise_plan`, and
+    `agent_provision` from BottlePlan."""
 
     slug: str
     container_name: str
@@ -46,56 +41,16 @@ class DockerBottlePlan(BottlePlan):
     forwarded_env: dict[str, str] = field(repr=False)
     prompt_file: Path
     proxy_plan: PipelockProxyPlan
-    git_gate_plan: GitGatePlan
-    egress_plan: EgressPlan
-    # None when bottle.supervise is False. PRD 0013 supervise sidecar
-    # is opt-in via the manifest's bottle.supervise field.
-    supervise_plan: SupervisePlan | None
     use_runsc: bool
-    agent_command: str = "claude"
-    agent_prompt_mode: PromptMode = "append_file"
-    agent_provider_template: str = "claude"
 
-    def print(self, *, remote_control: bool) -> None:
-        """Render the y/N preflight summary to stderr — compact form
-        intended to fit on screen without scrolling. The full
-        structured shape (image, container, runtime, etc.) lives on
-        this dataclass for tooling that wants to introspect it."""
-        del remote_control  # not surfaced in the compact summary
-        spec = self.spec
-        manifest = spec.manifest
-        agent = manifest.agents[spec.agent_name]
-        bottle = manifest.bottle_for(spec.agent_name)
-        # The agent sees the union of literal env names (rendered into
-        # --env-file) and forwarded env names (`-e NAME` with the
-        # value arriving via subprocess env). The forwarded set holds
-        # the OAuth token (CLAUDE_CODE_OAUTH_TOKEN) and any host-env
-        # interpolations from the manifest; egress holds
-        # upstream tokens in its own environ, so no token forwarding
-        # from the agent to the proxy is needed.
-        env_names = visible_agent_env_names(
-            sorted(set(bottle.env.keys()) | set(self.forwarded_env.keys())),
-            agent_provider_template=self.agent_provider_template,
-        )
+    @property
+    def agent_command(self) -> str:
+        return self.agent_provision.command
 
-        print(file=sys.stderr)
-        info(f"agent           : {spec.agent_name}")
-        info(f"provider        : {self.agent_provider_template}")
-        print_multi("env             ", env_names)
-        print_multi("skills          ", list(agent.skills))
-        info(f"bottle          : {agent.bottle}")
+    @property
+    def agent_prompt_mode(self) -> PromptMode:
+        return self.agent_provision.prompt_mode
 
-        git_lines = [
-            f"{u.upstream_host}:{u.upstream_port}"
-            for u in self.git_gate_plan.upstreams
-        ]
-        if git_lines:
-            print_multi("  git gate      ", git_lines)
-
-        if self.egress_plan.routes:
-            egress_lines = []
-            for r in self.egress_plan.routes:
-                auth = f" [auth:{r.auth_scheme}]" if r.auth_scheme else ""
-                egress_lines.append(f"{r.host}{auth}")
-            print_multi("  egress        ", egress_lines)
-        print(file=sys.stderr)
+    @property
+    def agent_provider_template(self) -> str:
+        return self.agent_provision.template

bot_bottle/backend/docker/bottle_state.py

@@ -105,6 +105,10 @@ class BottleMetadata:
     # written before chunk 3 (resume / inspect should fall back to
     # deriving from identity in that case).
     compose_project: str = ""
+    # PRD 0040: backend name ("docker" or "smolmachines"). Empty string
+    # for state dirs written before PRD 0040; callers default to "docker"
+    # for backward compatibility.
+    backend: str = ""
 
 
 def metadata_path(identity: str) -> Path:
@@ -138,6 +142,7 @@ def read_metadata(identity: str) -> BottleMetadata | None:
         copy_cwd=bool(raw.get("copy_cwd", False)),
         started_at=str(raw.get("started_at", "")),
         compose_project=str(raw.get("compose_project", "")),
+        backend=str(raw.get("backend", "")),
     )
 
 

bot_bottle/backend/docker/compose.py

@@ -49,7 +49,7 @@ from ...egress import (
     EGRESS_HOSTNAME,
     EGRESS_ROUTES_IN_CONTAINER,
 )
-from ...git_gate import GIT_GATE_HOSTNAME, git_gate_aggregate_extra_hosts
+from ...git_gate import GIT_GATE_HOSTNAME
 from ...log import die, warn
 from ...pipelock import PIPELOCK_HOSTNAME
 from ...supervise import (
@@ -198,7 +198,6 @@ def _sidecar_bundle_service(plan: DockerBottlePlan) -> dict[str, Any]:
             env.append(token_env)
 
     # --- git-gate ----------------------------------------------------
-    extra_hosts: list[str] = []
     gp = plan.git_gate_plan
     if gp.upstreams:
         volumes += [
@@ -217,8 +216,6 @@ def _sidecar_bundle_service(plan: DockerBottlePlan) -> dict[str, Any]:
                     u.known_hosts_file,
                     f"{GIT_GATE_CREDS_DIR_IN_CONTAINER}/{u.name}-known_hosts",
                 ))
-        extra_map = git_gate_aggregate_extra_hosts(gp.upstreams)
-        extra_hosts = [f"{host}:{ip}" for host, ip in sorted(extra_map.items())]
 
     # --- supervise ---------------------------------------------------
     sp = plan.supervise_plan
@@ -261,8 +258,6 @@ def _sidecar_bundle_service(plan: DockerBottlePlan) -> dict[str, Any]:
         "environment": env,
         "volumes": volumes,
     }
-    if extra_hosts:
-        service["extra_hosts"] = extra_hosts
     return service
 
 
@@ -286,6 +281,8 @@ def _agent_service(plan: DockerBottlePlan) -> dict[str, Any]:
         f"SSL_CERT_FILE={AGENT_CA_BUNDLE}",
         f"REQUESTS_CA_BUNDLE={AGENT_CA_BUNDLE}",
     ]
+    for name, value in sorted(plan.agent_provision.guest_env.items()):
+        env.append(f"{name}={value}")
     # Forwarded vars (OAuth token, manifest host-interpolations):
     # bare name → inherits from compose-up process env, value
     # never lands on argv or in the compose file.

bot_bottle/backend/docker/launch.py

@@ -43,7 +43,8 @@ from pathlib import Path
 from typing import Callable, Generator
 
 from ...egress import egress_resolve_token_values
-from ...log import info
+from ...git_gate import revoke_git_gate_provisioned_keys
+from ...log import info, warn
 from . import network as network_mod
 from . import util as docker_mod
 from .bottle import DockerBottle
@@ -51,6 +52,7 @@ from .bottle_plan import DockerBottlePlan
 from .bottle_state import (
     bottle_state_dir,
     egress_state_dir,
+    git_gate_state_dir,
     pipelock_state_dir,
 )
 from .compose import (
@@ -84,13 +86,20 @@ def launch(
     Teardown on exit."""
     stack = ExitStack()
 
+    _bottle_for_revoke = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+    _git_gate_dir_for_revoke = git_gate_state_dir(plan.slug)
+
     def teardown() -> None:
         try:
             stack.close()
-        except BaseException:
-            # Teardown must not raise; swallow so the caller's
-            # __exit__ path can still propagate the original error.
-            pass
+        except BaseException as exc:
+            warn(
+                f"teardown failed for container {plan.container_name}"
+                f" (compose-down): {exc!r}"
+            )
+        revoke_git_gate_provisioned_keys(
+            _bottle_for_revoke, _git_gate_dir_for_revoke
+        )
 
     try:
         # Step 1: agent image build. Sidecar images get built lazily by
@@ -101,7 +110,7 @@ def launch(
         )
         if plan.derived_image:
             docker_mod.build_image_with_cwd(
-                plan.derived_image, plan.image, plan.spec.user_cwd
+                plan.derived_image, plan.image, plan.workspace_plan
             )
 
         # Networks: compose-managed. The names are derived
@@ -176,11 +185,10 @@ def launch(
         # Step 7: compose up. Token values + the OAuth placeholder
         # flow through subprocess env; the compose file holds only
         # bare names for the secret-carrying entries.
-        token_values: dict[str, str] = {}
-        if plan.egress_plan.routes:
-            token_values = egress_resolve_token_values(
-                plan.egress_plan.token_env_map, dict(os.environ),
-            )
+        effective_env = {**dict(os.environ), **plan.agent_provision.provisioned_env}
+        token_values = egress_resolve_token_values(
+            plan.egress_plan.token_env_map, effective_env,
+        )
         compose_env: dict[str, str] = {
             **os.environ,
             **plan.forwarded_env,
@@ -200,19 +208,21 @@ def launch(
             compose_dump_logs, project, compose_file, compose_log_path(state_dir),
         )
 
-        # Step 8: provision. Unchanged — uses `docker exec` against
-        # the agent container by its known name.
-        prompt_path = provision(plan, plan.container_name)
+        # Step 8: provision. Create the bottle first so provisioners
+        # can use bottle.exec / bottle.cp_in; set the prompt path
+        # returned by provision_prompt after the fact.
+        bottle = DockerBottle(
+            plan.container_name,
+            teardown,
+            None,
+            agent_command=plan.agent_command,
+            agent_prompt_mode=plan.agent_prompt_mode,
+        )
+        bottle._prompt_path = provision(plan, bottle)
 
         # Step 9: yield. exec_agent continues to use `docker exec -it`
         # — the agent runs `sleep infinity` per the renderer's
         # service spec.
-        yield DockerBottle(
-            plan.container_name,
-            teardown,
-            prompt_path,
-            agent_command=plan.agent_command,
-            agent_prompt_mode=plan.agent_prompt_mode,
-        )
+        yield bottle
     finally:
         teardown()

bot_bottle/backend/docker/prepare.py

@@ -12,15 +12,17 @@ from __future__ import annotations
 
 import os
 from datetime import datetime, timezone
+from dataclasses import replace
 from pathlib import Path
 
-from ...agent_provider import runtime_for
+from ...agent_provider import agent_provision_plan, runtime_for
 from ...egress import Egress
 from ...env import ResolvedEnv, resolve_env
 from ...git_gate import GitGate
 from ...log import die
 from ...pipelock import PipelockProxy
 from ...supervise import Supervise
+from ...workspace import workspace_plan as resolve_workspace_plan
 from .. import BottleSpec
 from . import util as docker_mod
 from .bottle_plan import DockerBottlePlan
@@ -61,6 +63,8 @@ def resolve_plan(
     bottle = manifest.bottle_for(spec.agent_name)
     provider = bottle.agent_provider
     provider_runtime = runtime_for(provider.template)
+    guest_home = "/home/node"
+    workspace_plan = resolve_workspace_plan(spec, guest_home=guest_home)
 
     # PRD 0016 follow-up: identity, not bare slug. A fresh `start`
     # mints a random-suffixed identity (so parallel runs of the same
@@ -78,6 +82,7 @@ def resolve_plan(
         copy_cwd=spec.copy_cwd,
         started_at=datetime.now(timezone.utc).isoformat(),
         compose_project=f"bot-bottle-{slug}",
+        backend="docker",
     ))
     # Clear any leftover preserve marker from a prior capability-block
     # so this fresh launch can be cleaned up at session-end unless
@@ -158,17 +163,45 @@ def resolve_plan(
     prompt_file.write_text("")
     prompt_file.chmod(0o600)
 
-    pipelock_dir = pipelock_state_dir(slug)
-    pipelock_dir.mkdir(parents=True, exist_ok=True)
-    proxy_plan = proxy.prepare(bottle, slug, pipelock_dir)
-
     git_gate_dir = git_gate_state_dir(slug)
     git_gate_dir.mkdir(parents=True, exist_ok=True)
     git_gate_plan = git_gate.prepare(bottle, slug, git_gate_dir)
 
+    resolved = resolve_env(manifest, spec.agent_name)
+    # Everything that should reach the bottle by-name (so its value
+    # never lands on argv or in env_file) goes into one dict. Nothing
+    # mutates the host os.environ.
+    forwarded_env: dict[str, str] = dict(resolved.forwarded)
+    _write_env_file(resolved, env_file)
+    prompt_file.write_text(agent.prompt)
+
+    use_runsc = docker_mod.runsc_available()
+    agent_provision = agent_provision_plan(
+        template=provider.template,
+        dockerfile=dockerfile_path,
+        state_dir=agent_dir,
+        guest_home=guest_home,
+        forward_host_credentials=provider.forward_host_credentials,
+        auth_token=provider.auth_token,
+        host_env=dict(os.environ),
+        trusted_project_path=workspace_plan.workdir,
+    )
+    guest_env = dict(agent_provision.guest_env)
+    for key, val in agent_provision.env_vars.items():
+        guest_env.setdefault(key, val)
+    agent_provision = replace(agent_provision, guest_env=guest_env)
+
+    pipelock_dir = pipelock_state_dir(slug)
+    pipelock_dir.mkdir(parents=True, exist_ok=True)
+    proxy_plan = proxy.prepare(
+        bottle, slug, pipelock_dir, agent_provision.egress_routes,
+    )
+
     egress_dir = egress_state_dir(slug)
     egress_dir.mkdir(parents=True, exist_ok=True)
-    egress_plan = egress.prepare(bottle, slug, egress_dir)
+    egress_plan = egress.prepare(
+        bottle, slug, egress_dir, agent_provision.egress_routes,
+    )
 
     supervise_plan = None
     if bottle.supervise:
@@ -196,37 +229,11 @@ def resolve_plan(
             slug, supervise_dir,
             dockerfile_content=dockerfile_content,
         )
-    resolved = resolve_env(manifest, spec.agent_name)
-    # Everything that should reach the bottle by-name (so its value
-    # never lands on argv or in env_file) goes into one dict. Nothing
-    # mutates the host os.environ.
-    forwarded_env: dict[str, str] = dict(resolved.forwarded)
-    # When the bottle declares an egress route with the
-    # `claude_code_oauth` role marker, claude-code's outbound
-    # Authorization gets stripped + re-injected by egress. The
-    # agent's environ still needs *something* claude-code recognises
-    # as a credential or it refuses to start; ship a non-secret
-    # placeholder. The placeholder isn't any real token value, so
-    # leaking it would tell an attacker only that egress is in
-    # front. Manifest validation enforces singleton on this role.
-    has_provider_auth = any(
-        provider_runtime.auth_role in r.roles for r in egress_plan.routes
-    )
-    if has_provider_auth:
-        forwarded_env[provider_runtime.placeholder_env] = "egress-placeholder"
-    if provider.template == "claude" and has_provider_auth:
-        # Belt-and-braces: turn off telemetry endpoints (statsig,
-        # error reporting) that egress can't gate by auth.
-        forwarded_env.setdefault("CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC", "1")
-        forwarded_env.setdefault("DISABLE_ERROR_REPORTING", "1")
-    _write_env_file(resolved, env_file)
-    prompt_file.write_text(agent.prompt)
-
-    use_runsc = docker_mod.runsc_available()
 
     return DockerBottlePlan(
         spec=spec,
         stage_dir=stage_dir,
+        guest_home=guest_home,
         slug=slug,
         container_name=container_name,
         container_name_pinned=container_name_pinned,
@@ -242,9 +249,8 @@ def resolve_plan(
         egress_plan=egress_plan,
         supervise_plan=supervise_plan,
         use_runsc=use_runsc,
-        agent_command=provider_runtime.command,
-        agent_prompt_mode=provider_runtime.prompt_mode,
-        agent_provider_template=provider.template,
+        agent_provision=agent_provision,
+        workspace_plan=workspace_plan,
     )
 
 

bot_bottle/backend/docker/provision/__init__.py

@@ -1,8 +1,11 @@
-"""Per-provisioner modules for the Docker backend.
+"""Backend-infrastructure provisioners for the Docker backend.
 
-Each module exports one top-level function:
-    provision_<thing>(plan: DockerBottlePlan, target: str) -> ...
+Per PRD 0050 the per-provider provisioning steps (prompt, skills,
+declarative provision-plan apply, supervise MCP registration) live on
+the `AgentProvider` plugin under `bot_bottle/contrib/`. The modules
+left in this subpackage handle only the steps that are
+backend-specific:
 
-`DockerBottleBackend.provision_*` methods delegate to these. The
-abstract `BottleBackend.provision_*` surface is unchanged; this
-subpackage exists only to keep `backend.py` from being a god-file."""
+  - ca.py   — install per-bottle CA bundle into the guest trust store
+  - git.py  — copy host cwd `.git` into the guest when --cwd is used
+"""

bot_bottle/backend/docker/provision/ca.py

@@ -31,33 +31,21 @@ stage dir; nothing in the agent ever sees it."""
 
 from __future__ import annotations
 
-import subprocess
-
+from ... import Bottle
 from ...util import AGENT_CA_PATH, log_ca_fingerprint, select_ca_cert
 from ..bottle_plan import DockerBottlePlan
 
 
-def provision_ca(plan: DockerBottlePlan, target: str) -> None:
+def provision_ca(plan: DockerBottlePlan, bottle: Bottle) -> None:
     """Copy the agent-facing CA cert into the agent, rebuild the
     trust bundle, emit a one-line fingerprint log. Called from
     `BottleBackend.provision` after the agent container is up."""
-    container = target
     cert_host_path, label = select_ca_cert(plan.egress_plan, plan.proxy_plan)
 
-    subprocess.run(
-        ["docker", "cp", str(cert_host_path), f"{container}:{AGENT_CA_PATH}"],
-        stdout=subprocess.DEVNULL,
-        check=True,
-    )
-    subprocess.run(
-        ["docker", "exec", "-u", "0", container, "chmod", "644", AGENT_CA_PATH],
-        stdout=subprocess.DEVNULL,
-        check=True,
-    )
-    subprocess.run(
-        ["docker", "exec", "-u", "0", container, "update-ca-certificates"],
-        stdout=subprocess.DEVNULL,
-        check=True,
+    bottle.cp_in(str(cert_host_path), AGENT_CA_PATH)
+    bottle.exec(
+        f"chmod 644 {AGENT_CA_PATH} && update-ca-certificates",
+        user="root",
     )
 
     log_ca_fingerprint(cert_host_path, label)

bot_bottle/backend/docker/provision/git.py

@@ -3,7 +3,7 @@
 Three concerns, all about git in the agent:
 
   1. If --cwd was passed AND the host cwd has a .git, copy that .git
-     into /home/node/workspace/.git so the agent operates on the
+     into the planned guest workspace so the agent operates on the
      user's repo.
   2. If the bottle declares `git` entries (PRD 0008), write a
      ~/.gitconfig with insteadOf rules so every git operation
@@ -18,73 +18,62 @@ Three concerns, all about git in the agent:
 
 from __future__ import annotations
 
-import os
-import subprocess
-from pathlib import Path
+import shlex
 
 from ....git_gate import GIT_GATE_HOSTNAME, git_gate_render_gitconfig
 from ....log import info
-from .. import util as docker_mod
+from ... import Bottle
 from ..bottle_plan import DockerBottlePlan
 
 
-def provision_git(plan: DockerBottlePlan, target: str) -> None:
+def provision_git(plan: DockerBottlePlan, bottle: Bottle) -> None:
     """Set up git inside the bottle. Runs all three subcases; each
     no-ops when its condition isn't met."""
-    _provision_cwd_git(plan, target)
-    _provision_git_gate_config(plan, target)
-    _provision_git_user(plan, target)
+    _provision_cwd_git(plan, bottle)
+    _provision_git_gate_config(plan, bottle)
+    _provision_git_user(plan, bottle)
 
 
-def _provision_cwd_git(plan: DockerBottlePlan, target: str) -> None:
+def _provision_cwd_git(plan: DockerBottlePlan, bottle: Bottle) -> None:
     """If --cwd was set and the host cwd has a .git directory, copy
     it into /home/node/workspace/.git and fix ownership. No-op
     otherwise."""
-    if not (plan.spec.copy_cwd and Path(plan.spec.user_cwd, ".git").is_dir()):
+    workspace = plan.workspace_plan
+    if not (workspace.enabled and workspace.copy_git and workspace.has_host_git_dir):
         return
-    container = target
-    info(f"copying {plan.spec.user_cwd}/.git -> {container}:/home/node/workspace/.git")
-    subprocess.run(
-        ["docker", "cp", f"{plan.spec.user_cwd}/.git", f"{container}:/home/node/workspace/.git"],
-        stdout=subprocess.DEVNULL,
-        check=True,
-    )
-    subprocess.run(
-        [
-            "docker", "exec", "-u", "0", container,
-            "chown", "-R", "node:node", "/home/node/workspace/.git",
-        ],
-        stdout=subprocess.DEVNULL,
-        check=True,
+    guest_workspace_git = f"{workspace.guest_path}/.git"
+    host_git = str(workspace.host_path / ".git")
+    info(f"copying {host_git} -> {bottle.name}:{guest_workspace_git}")
+    bottle.cp_in(host_git, guest_workspace_git)
+    bottle.exec(
+        f"chown -R {shlex.quote(workspace.owner)} {shlex.quote(guest_workspace_git)}",
+        user="root",
     )
 
 
-def _provision_git_gate_config(plan: DockerBottlePlan, target: str) -> None:
+def _provision_git_gate_config(plan: DockerBottlePlan, bottle: Bottle) -> None:
     """Write ~/.gitconfig in the bottle with the git-gate
     insteadOf rules. No-op when the bottle has no `git` entries."""
-    bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
-    if not bottle.git:
+    manifest_bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+    if not manifest_bottle.git:
         return
-    container = target
-    container_home = os.environ.get("BOT_BOTTLE_CONTAINER_HOME", "/home/node")
-    container_gitconfig = f"{container_home}/.gitconfig"
+    container_gitconfig = f"{plan.guest_home}/.gitconfig"
 
-    content = git_gate_render_gitconfig(bottle.git, GIT_GATE_HOSTNAME)
+    content = git_gate_render_gitconfig(manifest_bottle.git, GIT_GATE_HOSTNAME)
     config_file = plan.stage_dir / "agent_gitconfig"
     config_file.write_text(content)
     config_file.chmod(0o600)
 
-    info(f"writing {container_gitconfig} with {len(bottle.git)} insteadOf rule(s)")
-    subprocess.run(
-        ["docker", "cp", str(config_file), f"{container}:{container_gitconfig}"],
-        stdout=subprocess.DEVNULL,
-        check=True,
+    info(f"writing {container_gitconfig} with {len(manifest_bottle.git)} insteadOf rule(s)")
+    bottle.cp_in(str(config_file), container_gitconfig)
+    bottle.exec(
+        f"chown node:node {shlex.quote(container_gitconfig)} && "
+        f"chmod 644 {shlex.quote(container_gitconfig)}",
+        user="root",
     )
-    docker_mod.docker_exec_root(container, ["chown", "node:node", container_gitconfig])
-    docker_mod.docker_exec_root(container, ["chmod", "644", container_gitconfig])
 
 
-def _provision_git_user(plan: DockerBottlePlan, target: str) -> None:
+def _provision_git_user(plan: DockerBottlePlan, bottle: Bottle) -> None:
     """Apply `git config --global user.{name,email}` inside the
     bottle so the agent's commits are attributed to the operator-
     chosen identity instead of the agent image's default
@@ -99,23 +88,19 @@ def _provision_git_user(plan: DockerBottlePlan, target: str) -> None:
     Each field set independently — name-only or email-only
     configs only run the `git config` line for the field
     present."""
-    bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
-    gu = bottle.git_user
+    manifest_bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+    gu = manifest_bottle.git_user
     if gu.is_empty():
         return
     if gu.name:
         info(f"git config --global user.name = {gu.name!r}")
-        subprocess.run(
-            ["docker", "exec", "-u", "node", target,
-             "git", "config", "--global", "user.name", gu.name],
-            stdout=subprocess.DEVNULL,
-            check=True,
+        bottle.exec(
+            f"git config --global user.name {shlex.quote(gu.name)}",
+            user="node",
         )
     if gu.email:
         info(f"git config --global user.email = {gu.email!r}")
-        subprocess.run(
-            ["docker", "exec", "-u", "node", target,
-             "git", "config", "--global", "user.email", gu.email],
-            stdout=subprocess.DEVNULL,
-            check=True,
+        bottle.exec(
+            f"git config --global user.email {shlex.quote(gu.email)}",
+            user="node",
         )

bot_bottle/backend/docker/provision/prompt.py

@@ -1,43 +0,0 @@
-"""Copy the agent prompt into a running Docker bottle.
-
-The prompt file is always copied (so the in-container path always
-exists) but `--append-system-prompt-file` only fires when the agent
-actually has a prompt — the return value signals which case."""
-
-from __future__ import annotations
-
-import os
-import subprocess
-
-from ..bottle_plan import DockerBottlePlan
-
-
-def provision_prompt(plan: DockerBottlePlan, target: str) -> str | None:
-    """Copy the prompt file into the container, fix ownership/mode.
-    Returns the in-container path if the agent has a non-empty
-    prompt (drives --append-system-prompt-file), else None. The
-    file is copied either way so the path always exists."""
-    container = target
-    container_home = os.environ.get("BOT_BOTTLE_CONTAINER_HOME", "/home/node")
-    in_container_prompt_path = f"{container_home}/.bot-bottle-prompt.txt"
-
-    subprocess.run(
-        ["docker", "cp", str(plan.prompt_file), f"{container}:{in_container_prompt_path}"],
-        stdout=subprocess.DEVNULL,
-        check=True,
-    )
-    # `docker cp` preserves host UID; re-own/mode as root so node
-    # can read its own mode-600 prompt regardless of host UID.
-    subprocess.run(
-        ["docker", "exec", "-u", "0", container, "chown", "node:node", in_container_prompt_path],
-        stdout=subprocess.DEVNULL,
-        check=True,
-    )
-    subprocess.run(
-        ["docker", "exec", "-u", "0", container, "chmod", "600", in_container_prompt_path],
-        stdout=subprocess.DEVNULL,
-        check=True,
-    )
-
-    agent = plan.spec.manifest.agents[plan.spec.agent_name]
-    return in_container_prompt_path if agent.prompt else None

bot_bottle/backend/docker/provision/skills.py

@@ -1,62 +0,0 @@
-"""Copy host-side skill directories into a running Docker bottle.
-
-Skills are validated on the host before launch by the base class's
-`BottleBackend._validate_skills` (called from `prepare`); this module
-assumes that validation has already run. A skill disappearing between
-validation and copy still dies loudly rather than silently producing
-a partial container."""
-
-from __future__ import annotations
-
-import os
-import subprocess
-
-from ....log import die, info
-from ...util import host_skill_dir
-from ..bottle_plan import DockerBottlePlan
-
-
-def provision_skills(plan: DockerBottlePlan, target: str) -> None:
-    """Copy each of the agent's named skills from the host's
-    ~/.claude/skills/<name>/ into the container's equivalent path.
-    For each skill: ensure parent dir, wipe any prior copy, then
-    `docker cp <host>/. <container>:<dst>/` so the contents are
-    copied into a freshly-created destination dir. No-op when the
-    agent has no skills."""
-    agent = plan.spec.manifest.agents[plan.spec.agent_name]
-    if not agent.skills:
-        return
-
-    container = target
-    container_home = os.environ.get("BOT_BOTTLE_CONTAINER_HOME", "/home/node")
-    skills_dir = os.environ.get(
-        "BOT_BOTTLE_CONTAINER_SKILLS_DIR", f"{container_home}/.claude/skills"
-    )
-
-    subprocess.run(
-        ["docker", "exec", container, "mkdir", "-p", skills_dir],
-        stdout=subprocess.DEVNULL,
-        check=True,
-    )
-
-    for n in agent.skills:
-        src = host_skill_dir(n)
-        if not os.path.isdir(src):
-            die(f"skill '{n}' disappeared from host between validation and copy at {src}.")
-        dst = f"{skills_dir}/{n}"
-        info(f"copying skill {n} into {container}:{dst}")
-        subprocess.run(
-            ["docker", "exec", container, "rm", "-rf", dst],
-            stdout=subprocess.DEVNULL,
-            check=True,
-        )
-        subprocess.run(
-            ["docker", "exec", container, "mkdir", "-p", dst],
-            stdout=subprocess.DEVNULL,
-            check=True,
-        )
-        subprocess.run(
-            ["docker", "cp", f"{src}/.", f"{container}:{dst}/"],
-            stdout=subprocess.DEVNULL,
-            check=True,
-        )

bot_bottle/backend/docker/provision/supervise.py

@@ -1,65 +0,0 @@
-"""Supervise sidecar provisioning inside a running Docker bottle
-(PRD 0013).
-
-Registers the per-bottle supervise sidecar as an HTTP MCP server in
-the agent's claude-code config so the agent discovers the three
-stuck-recovery MCP tools (cred-proxy-block, pipelock-block,
-capability-block) at startup.
-
-Uses `claude mcp add` rather than writing JSON directly. claude-code
-owns the on-disk config format (`~/.claude.json` `mcpServers` shape,
-field names, scope semantics) and changes it between versions; the
-official command handles whatever the installed version expects.
-
-No-op when bottle.supervise is False — bottles that haven't opted
-into the supervise sidecar shouldn't get an MCP entry pointing at a
-sidecar that isn't running.
-"""
-
-from __future__ import annotations
-
-import subprocess
-
-from ....log import info, warn
-from ....supervise import SUPERVISE_HOSTNAME, SUPERVISE_PORT
-from ..bottle_plan import DockerBottlePlan
-
-
-_SUPERVISE_MCP_NAME = "supervise"
-
-
-def supervise_mcp_url() -> str:
-    return f"http://{SUPERVISE_HOSTNAME}:{SUPERVISE_PORT}/"
-
-
-def provision_supervise(plan: DockerBottlePlan, target: str) -> None:
-    """Run `claude mcp add` inside the agent container to register
-    the supervise sidecar in claude-code's user config. No-op when
-    bottle.supervise is False.
-
-    Failure is logged but not fatal: the bottle still works (you
-    just can't call supervise tools from the agent until the entry
-    is added manually). The operator sees the warning at launch."""
-    if plan.supervise_plan is None:
-        return
-    url = supervise_mcp_url()
-    argv = [
-        "docker", "exec", "-u", "node", target,
-        "claude", "mcp", "add",
-        "--scope", "user",
-        "--transport", "http",
-        _SUPERVISE_MCP_NAME,
-        url,
-    ]
-    info(f"registering supervise MCP server in agent claude config → {url}")
-    r = subprocess.run(argv, capture_output=True, text=True, check=False)
-    if r.returncode != 0:
-        warn(
-            f"`claude mcp add supervise` failed (exit {r.returncode}): "
-            f"{(r.stderr or r.stdout or '').strip()}. Inside the bottle, "
-            f"register manually with: "
-            f"claude mcp add --scope user --transport http supervise {url}"
-        )
-
-
-__all__ = ["provision_supervise", "supervise_mcp_url"]

bot_bottle/backend/docker/util.py

@@ -7,9 +7,11 @@ from __future__ import annotations
 import re
 import shutil
 import subprocess
+import tempfile
 from typing import Iterable, Iterator
 
 from ...log import die, info
+from ...workspace import WorkspacePlan
 
 
 # Cap on the suffix the container-name conflict logic will try before
@@ -116,35 +118,39 @@ def build_image(ref: str, context: str, *, dockerfile: str = "") -> None:
     subprocess.run(args, check=True)
 
 
-_TRUST_DIALOG_NODE_SCRIPT = (
-    'const fs=require("fs"),p=process.env.HOME+"/.claude.json",'
-    'c=JSON.parse(fs.readFileSync(p,"utf8"));'
-    'c.projects=c.projects||{};'
-    'c.projects[process.env.HOME+"/workspace"]={hasTrustDialogAccepted:true};'
-    'fs.writeFileSync(p,JSON.stringify(c,null,2));'
-)
-
-
-def build_image_with_cwd(derived: str, base: str, cwd: str) -> None:
-    """Build a thin derived image that copies <cwd> into
-    /home/node/workspace and adds a trust-dialog entry for it."""
+def build_image_with_cwd(
+    derived: str,
+    base: str,
+    workspace: WorkspacePlan,
+) -> None:
+    """Build a thin derived image that copies the workspace into
+    the plan's guest path and sets the plan's workdir."""
     import os
 
+    cwd = str(workspace.host_path)
     if not os.path.isdir(cwd):
         die(f"cwd not found at {cwd}")
-    info(f"building image {derived} from {base} with {cwd} -> /home/node/workspace")
-    dockerfile = (
-        f"FROM {base}\n"
-        f"COPY --chown=node:node . /home/node/workspace\n"
-        f"RUN node -e '{_TRUST_DIALOG_NODE_SCRIPT}'\n"
-        f"WORKDIR /home/node/workspace\n"
-    )
-    subprocess.run(
-        ["docker", "build", "-t", derived, "-f", "-", cwd],
-        input=dockerfile,
-        text=True,
-        check=True,
-    )
+    info(f"building image {derived} from {base} with {cwd} -> {workspace.guest_path}")
+    with tempfile.TemporaryDirectory(prefix="bot-bottle-cwd.") as tmp:
+        context_dir = os.path.join(tmp, "context")
+        staged_workspace = os.path.join(context_dir, "workspace")
+        shutil.copytree(
+            cwd,
+            staged_workspace,
+            symlinks=True,
+            ignore=shutil.ignore_patterns(".git"),
+        )
+        dockerfile = (
+            f"FROM {base}\n"
+            f"COPY --chown=node:node workspace/. {workspace.guest_path}\n"
+            f"WORKDIR {workspace.workdir}\n"
+        )
+        subprocess.run(
+            ["docker", "build", "-t", derived, "-f", "-", context_dir],
+            input=dockerfile,
+            text=True,
+            check=True,
+        )
 
 
 def image_id(ref: str) -> str:

bot_bottle/backend/print_util.py

@@ -9,7 +9,6 @@ from __future__ import annotations
 
 from typing import Sequence
 
-from ..agent_provider import runtime_for
 from ..log import info
 
 
@@ -30,16 +29,13 @@ def print_multi(label: str, values: Sequence[str]) -> None:
 
 
 def visible_agent_env_names(
-    env_names: Sequence[str], *, agent_provider_template: str,
+    env_names: Sequence[str], *, hidden_env_names: frozenset[str],
 ) -> list[str]:
     """Env names worth showing in launch summaries.
 
-    Provider auth placeholders (`OPENAI_API_KEY`,
-    `CLAUDE_CODE_OAUTH_TOKEN`) are implementation details: they are
-    non-secret dummy values that satisfy the provider CLI while egress
-    injects the real upstream Authorization header. Showing them in
-    preflight makes the operator think a real key is entering the
-    agent, so hide only that provider-owned placeholder.
+    Provider-injected placeholder env vars are implementation details:
+    they are non-secret dummy values that satisfy provider CLIs while
+    egress injects the real Authorization header. The plan's
+    `hidden_env_names` carries exactly which names to suppress.
     """
-    hidden = {runtime_for(agent_provider_template).placeholder_env}
-    return sorted({name for name in env_names if name not in hidden})
+    return sorted({name for name in env_names if name and name not in hidden_env_names})

bot_bottle/backend/smolmachines/backend.py

@@ -1,5 +1,11 @@
 """SmolmachinesBottleBackend — the smolmachines implementation of
-BottleBackend (PRD 0023)."""
+BottleBackend (PRD 0023).
+
+Per PRD 0050 the per-provider provisioning steps (prompt, skills,
+the declarative provision-plan apply, supervise MCP registration)
+live on the `AgentProvider` plugin under `bot_bottle/contrib/`. The
+smolmachines backend only owns the steps that are about backend
+infrastructure: CA install (no-op for now), workspace, git copy-in."""
 
 from __future__ import annotations
 
@@ -7,7 +13,7 @@ from contextlib import contextmanager
 from pathlib import Path
 from typing import Generator, Sequence
 
-from .. import ActiveAgent, BottleBackend, BottleSpec
+from .. import ActiveAgent, Bottle, BottleBackend, BottleSpec
 from . import cleanup as _cleanup
 from . import enumerate as _enumerate
 from . import launch as _launch
@@ -18,9 +24,7 @@ from .bottle_cleanup_plan import SmolmachinesBottleCleanupPlan
 from .bottle_plan import SmolmachinesBottlePlan
 from .provision import ca as _ca
 from .provision import git as _git
-from .provision import prompt as _prompt
-from .provision import skills as _skills
-from .provision import supervise as _supervise
+from .provision import workspace as _workspace
 
 
 class SmolmachinesBottleBackend(
@@ -52,29 +56,26 @@ class SmolmachinesBottleBackend(
             yield bottle
 
     def provision_ca(
-        self, plan: SmolmachinesBottlePlan, target: str
+        self, plan: SmolmachinesBottlePlan, bottle: Bottle
     ) -> None:
-        _ca.provision_ca(plan, target)
+        _ca.provision_ca(plan, bottle)
 
-    def provision_prompt(
-        self, plan: SmolmachinesBottlePlan, target: str
-    ) -> str | None:
-        return _prompt.provision_prompt(plan, target)
-
-    def provision_skills(
-        self, plan: SmolmachinesBottlePlan, target: str
+    def provision_workspace(
+        self, plan: SmolmachinesBottlePlan, bottle: Bottle
     ) -> None:
-        _skills.provision_skills(plan, target)
+        _workspace.provision_workspace(plan, bottle)
 
     def provision_git(
-        self, plan: SmolmachinesBottlePlan, target: str
+        self, plan: SmolmachinesBottlePlan, bottle: Bottle
     ) -> None:
-        _git.provision_git(plan, target)
+        _git.provision_git(plan, bottle)
 
-    def provision_supervise(
-        self, plan: SmolmachinesBottlePlan, target: str
-    ) -> None:
-        _supervise.provision_supervise(plan, target)
+    def supervise_mcp_url(self, plan: SmolmachinesBottlePlan) -> str:
+        """The smolmachines guest reaches the supervise sidecar via a
+        host-published random port the launch step pinned earlier
+        (`http://<loopback_ip>:<random_port>/`). `agent_supervise_url`
+        on the plan is "" when the bottle has no sidecar."""
+        return plan.agent_supervise_url
 
     def prepare_cleanup(self) -> SmolmachinesBottleCleanupPlan:
         return _cleanup.prepare_cleanup()

bot_bottle/backend/smolmachines/bottle.py

@@ -45,19 +45,11 @@ _HOME_FOR = {
 }
 
 
-def _env_flags_for(user: str) -> list[str]:
+def _env_assignments_for(user: str, env: Mapping[str, str]) -> list[str]:
     home = _HOME_FOR.get(user, f"/home/{user}")
-    return ["-e", f"HOME={home}", "-e", f"USER={user}"]
-
-
-def _guest_env_flags(env: Mapping[str, str]) -> list[str]:
-    """Render `{K: V}` into a flat `-e K=V` argv slice for
-    `smolvm machine exec`. `smolvm machine create -e` set env
-    on PID 1 but it doesn't propagate to fresh exec process
-    trees, so we have to re-pass them every call."""
-    out: list[str] = []
+    out = [f"HOME={home}", f"USER={user}"]
     for k, v in env.items():
-        out += ["-e", f"{k}={v}"]
+        out.append(f"{k}={v}")
     return out
 
 
@@ -98,9 +90,8 @@ class SmolmachinesBottle(Bottle):
         flags = ["smolvm", "machine", "exec", "--name", self.name]
         if tty:
             flags += ["-i", "-t"]
-        flags += _env_flags_for("node")
-        flags += _guest_env_flags(self._guest_env)
-        agent_tail = [self.agent_command]
+        agent_tail = ["env", *_env_assignments_for("node", self._guest_env),
+                      self.agent_command]
         provider_prompt_args = prompt_args(
             self._agent_prompt_mode, self._prompt_path, argv=argv,
         )
@@ -148,16 +139,16 @@ class SmolmachinesBottle(Bottle):
         on both backends. Pass `user="root"` for tests that need
         root.
 
-        `runuser -u <user> -- /bin/sh -c <script>` switches UID
-        without invoking a login shell; HOME / USER are set via
-        `smolvm -e` (see `_env_flags_for`)."""
-        argv = (
-            _env_flags_for(user)
-            + _guest_env_flags(self._guest_env)
-            + ["--", "runuser", "-u", user, "--", "/bin/sh", "-c", script]
-        )
-        # _smolvm.machine_exec expects argv (the bit after `--`);
-        # the -e flags go before, so call smolvm directly.
+        `runuser -u <user> -- env ... /bin/sh -c <script>` switches UID
+        without invoking a login shell, then sets HOME / USER and the
+        bottle env in the child process."""
+        argv = [
+            "--", "runuser", "-u", user, "--",
+            "env", *_env_assignments_for(user, self._guest_env),
+            "/bin/sh", "-c", script,
+        ]
+        # Call smolvm directly because this path needs the host-side
+        # subprocess capture shape used by the Docker backend.
         r = subprocess.run(
             ["smolvm", "machine", "exec", "--name", self.name] + argv,
             capture_output=True, text=True, check=False,

bot_bottle/backend/smolmachines/bottle_plan.py

@@ -8,25 +8,20 @@ in chunk 4."""
 
 from __future__ import annotations
 
-import sys
 from dataclasses import dataclass
 from pathlib import Path
 
 from ...agent_provider import PromptMode
-from ...egress import EgressPlan
-from ...git_gate import GitGatePlan
-from ...log import info
 from ...pipelock import PipelockProxyPlan
-from ...supervise import SupervisePlan
 from .. import BottlePlan
-from ..print_util import print_multi, visible_agent_env_names
 
 
 @dataclass(frozen=True)
 class SmolmachinesBottlePlan(BottlePlan):
     """Resolved fields the launch step needs to bring up the bottle.
 
-    Inherits `spec` and `stage_dir` from BottlePlan."""
+    Inherits `spec`, `stage_dir`, `git_gate_plan`, `egress_plan`,
+    `supervise_plan`, and `agent_provision` from BottlePlan."""
 
     slug: str
     # Per-bottle docker subnet for the sidecar bundle container.
@@ -68,7 +63,7 @@ class SmolmachinesBottlePlan(BottlePlan):
     # empty when the agent has no prompt — claude-code reads it
     # via --append-system-prompt-file only when non-empty.
     prompt_file: Path
-    # Inner Plans for the four bundle daemons. The same shape the
+    # Inner Plans for the sidecar bundle daemons. The same shape the
     # docker backend uses — same `.prepare()` calls produced
     # them — but our launch step doesn't populate the
     # docker-specific network fields (internal_network,
@@ -77,11 +72,6 @@ class SmolmachinesBottlePlan(BottlePlan):
     # per-bottle bridge with a pinned IP. The unused fields stay
     # at their dataclass defaults.
     proxy_plan: PipelockProxyPlan
-    git_gate_plan: GitGatePlan
-    egress_plan: EgressPlan
-    # None when bottle.supervise is False, matching the docker
-    # backend's convention.
-    supervise_plan: SupervisePlan | None
     # Agent-side endpoints. On Docker Desktop the docker bridge
     # IPs aren't reachable from the smolvm guest (TSI uses macOS
     # networking; docker container IPs live in the daemon's VM),
@@ -93,40 +83,19 @@ class SmolmachinesBottlePlan(BottlePlan):
     agent_proxy_url: str = ""
     agent_git_gate_host: str = ""
     agent_supervise_url: str = ""
-    agent_command: str = "claude"
-    agent_prompt_mode: PromptMode = "append_file"
-    agent_provider_template: str = "claude"
-    agent_dockerfile_path: str = ""
 
-    def print(self, *, remote_control: bool) -> None:
-        """Compact y/N preflight. Same shape as the Docker
-        backend's so operators see one format across backends."""
-        del remote_control  # not surfaced in the compact summary
-        spec = self.spec
-        manifest = spec.manifest
-        agent = manifest.agents[spec.agent_name]
-        bottle = manifest.bottle_for(spec.agent_name)
+    @property
+    def agent_command(self) -> str:
+        return self.agent_provision.command
 
-        env_names = visible_agent_env_names(
-            sorted(bottle.env.keys()),
-            agent_provider_template=self.agent_provider_template,
-        )
-        upstreams = [
-            f"{g.Name} → {g.Upstream}" for g in bottle.git
-        ]
-        # Use the resolved egress_plan (lowercase `host` on the
-        # plan-level EgressRoute) rather than `bottle.egress.routes`,
-        # which is the manifest's capitalized-attr form.
-        routes = [r.host for r in self.egress_plan.routes]
+    @property
+    def agent_prompt_mode(self) -> PromptMode:
+        return self.agent_provision.prompt_mode
 
-        print(file=sys.stderr)
-        info(f"agent           : {spec.agent_name}")
-        info(f"provider        : {self.agent_provider_template}")
-        print_multi("env             ", env_names)
-        print_multi("skills          ", list(agent.skills))
-        info(f"bottle          : {agent.bottle}")
-        if upstreams:
-            print_multi("  git gate      ", upstreams)
-        if routes:
-            print_multi("  egress        ", routes)
-        print(file=sys.stderr)
+    @property
+    def agent_provider_template(self) -> str:
+        return self.agent_provision.template
+
+    @property
+    def agent_dockerfile_path(self) -> str:
+        return self.agent_provision.dockerfile

bot_bottle/backend/smolmachines/launch.py

@@ -21,12 +21,14 @@ from __future__ import annotations
 
 import dataclasses
 import os
-import time
 from contextlib import ExitStack, contextmanager
 from pathlib import Path
 from typing import Callable, Generator
 
-from ...egress import EGRESS_ROUTES_IN_CONTAINER, egress_resolve_token_values
+from ...egress import (
+    EGRESS_ROUTES_IN_CONTAINER,
+    egress_resolve_token_values,
+)
 from ...pipelock import (
     PIPELOCK_CA_CERT_IN_CONTAINER,
     PIPELOCK_CA_KEY_IN_CONTAINER,
@@ -45,13 +47,15 @@ from ..docker.git_gate import (
     GIT_GATE_CREDS_DIR_IN_CONTAINER,
     GIT_GATE_ENTRYPOINT_IN_CONTAINER,
     GIT_GATE_HOOK_IN_CONTAINER,
-    GIT_GATE_PORT as _GIT_GATE_PORT,
 )
 from ..docker.pipelock import (
     BUNDLE_LOCAL_PIPELOCK_URL,
     PIPELOCK_PORT as _PIPELOCK_PORT_STR,
     pipelock_tls_init,
 )
+from ...git_gate import revoke_git_gate_provisioned_keys
+from ...log import warn
+from ..docker.bottle_state import git_gate_state_dir
 from . import loopback_alias as _loopback
 from . import sidecar_bundle as _bundle
 from . import smolvm as _smolvm
@@ -77,6 +81,7 @@ _SMOLMACHINE_CACHE_DIR = Path.home() / ".cache" / "bot-bottle" / "smolmachines"
 # them up post-start. Pipelock's port is an env-overridable string
 # in docker.pipelock; coerce to int here.
 _PIPELOCK_PORT = int(_PIPELOCK_PORT_STR)
+_GIT_HTTP_PORT = 9420
 _SUPERVISE_PORT = SUPERVISE_PORT
 
 
@@ -91,210 +96,230 @@ def launch(
     via the ExitStack."""
     stack = ExitStack()
     try:
-        # 1. Reserve a loopback alias for this bottle. macOS only
-        # routes 127.0.0.1 by default; the per-bottle alias is
-        # what bundles the docker port-publishes and TSI allowlist
-        # against, so this bottle can't reach other bottles' (or
-        # other host services') ports on the loopback. Lazy
-        # sudo-driven on first use per boot. No-op on Linux.
-        _loopback.ensure_pool()
-        loopback_ip = _loopback.allocate(plan.slug)
+        loopback_ip, network = _allocate_resources(plan, stack)
+        plan = _mint_certs(plan)
+        plan = _start_bundle(plan, network, loopback_ip, stack)
+        plan = _discover_urls(plan, loopback_ip)
 
-        # 2. Per-bottle docker bridge.
-        network = _bundle.bundle_network_name(plan.slug)
-        _bundle.create_bundle_network(network, plan.bundle_subnet, plan.bundle_gateway)
-        stack.callback(_bundle.remove_bundle_network, network)
-
-        # 2. Mint per-bottle CAs and update the inner Plans with
-        # their launch-time paths. pipelock always runs in the
-        # bundle; egress's CA is only minted when the bottle
-        # declares routes (otherwise egress runs idle without
-        # MITM and the CA files would be unused).
-        ca_cert_host, ca_key_host = pipelock_tls_init(plan.proxy_plan.yaml_path.parent)
-        proxy_plan = dataclasses.replace(
-            plan.proxy_plan,
-            ca_cert_host_path=ca_cert_host,
-            ca_key_host_path=ca_key_host,
-        )
-        egress_plan = plan.egress_plan
-        if egress_plan.routes:
-            egress_ca_host, egress_ca_cert_only = egress_tls_init(
-                plan.egress_plan.routes_path.parent,
-            )
-            egress_plan = dataclasses.replace(
-                egress_plan,
-                mitmproxy_ca_host_path=egress_ca_host,
-                mitmproxy_ca_cert_only_host_path=egress_ca_cert_only,
-                pipelock_ca_host_path=ca_cert_host,
-                # On smolmachines, egress's upstream is pipelock
-                # on the bundle's localhost — they're in the same
-                # container's network namespace.
-                pipelock_proxy_url=BUNDLE_LOCAL_PIPELOCK_URL,
-            )
-        plan = dataclasses.replace(
-            plan, proxy_plan=proxy_plan, egress_plan=egress_plan,
-        )
-
-        # 3. Build the BundleLaunchSpec from the (now-resolved)
-        # inner Plans: daemon subset, env, bind-mounts, and the
-        # loopback alias to bind published ports against. The
-        # spec's ports_to_publish list expands depending on which
-        # daemons the agent needs to reach from the smolvm guest.
-        bundle_spec = _bundle_launch_spec(plan, network, loopback_ip)
-        token_env = _resolve_token_env(plan, os.environ)
-        _bundle.ensure_bundle_image(bundle_spec.image)
-        _bundle.start_bundle(bundle_spec, env={**os.environ, **token_env})
-        stack.callback(_bundle.stop_bundle, plan.slug)
-
-        # 4. Discover the host-side ports docker assigned for the
-        # bundle's published container ports, and bind the
-        # agent's URLs to `<loopback_ip>:<host port>`. Docker
-        # container IPs (192.168.x.x in the daemon's bridge)
-        # aren't reachable from the smolvm guest on macOS — TSI
-        # uses macOS networking, and macOS sees the daemon's
-        # bridge via the published-port loopback forward only.
-        #
-        # Proxy hop order matches the docker backend: when the
-        # bottle declares egress routes, the agent's first hop is
-        # egress (for token injection), then pipelock. Without
-        # routes, the agent dials pipelock directly. Whichever
-        # one is "agent-facing" is the daemon whose port we
-        # publish on host loopback; the other stays bundle-
-        # internal as the upstream proxy.
-        if plan.egress_plan.routes:
-            agent_facing_port = _EGRESS_PORT
-        else:
-            agent_facing_port = _PIPELOCK_PORT
-        agent_facing_host_port = _bundle.bundle_host_port(
-            plan.slug, agent_facing_port, host_ip=loopback_ip,
-        )
-        agent_proxy_url = f"http://{loopback_ip}:{agent_facing_host_port}"
-        agent_git_gate_host = ""
-        if plan.git_gate_plan.upstreams:
-            git_gate_host_port = _bundle.bundle_host_port(
-                plan.slug, _GIT_GATE_PORT, host_ip=loopback_ip,
-            )
-            agent_git_gate_host = f"{loopback_ip}:{git_gate_host_port}"
-        agent_supervise_url = ""
-        if plan.supervise_plan is not None:
-            supervise_host_port = _bundle.bundle_host_port(
-                plan.slug, _SUPERVISE_PORT, host_ip=loopback_ip,
-            )
-            agent_supervise_url = f"http://{loopback_ip}:{supervise_host_port}/"
-
-        # Stamp the URLs onto the plan + guest_env. provision_git
-        # and provision_supervise read the plan fields; the agent
-        # reads guest_env on every exec_agent.
-        #
-        # NO_PROXY has to include the per-bottle loopback alias —
-        # otherwise claude's HTTPS_PROXY catches direct calls to
-        # the supervise URL (`http://<alias>:<port>/`) and proxies
-        # them through egress, which has no route for the alias
-        # and rejects with "Failed to connect". The git-gate URL
-        # uses git://, not affected by HTTP_PROXY, so the alias
-        # only has to be in NO_PROXY for the MCP / supervise
-        # path. Append rather than overwrite so prepare.py's
-        # `localhost,127.0.0.1` baseline stays in place.
-        existing_no_proxy = plan.guest_env.get("NO_PROXY", "localhost,127.0.0.1")
-        guest_env = {
-            **plan.guest_env,
-            "HTTPS_PROXY": agent_proxy_url,
-            "HTTP_PROXY":  agent_proxy_url,
-            "NO_PROXY":    f"{existing_no_proxy},{loopback_ip}",
-        }
-        if agent_git_gate_host:
-            guest_env["GIT_GATE_URL"] = f"git://{agent_git_gate_host}"
-        if agent_supervise_url:
-            guest_env["MCP_SUPERVISE_URL"] = agent_supervise_url
-        plan = dataclasses.replace(
-            plan,
-            guest_env=guest_env,
-            agent_proxy_url=agent_proxy_url,
-            agent_git_gate_host=agent_git_gate_host,
-            agent_supervise_url=agent_supervise_url,
-        )
-
-        # 5. 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:
-        # both the curses-endwin path and the tmux pane-routing
-        # path redirect stderr around `launch` already.
+        # 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_ref,
             dockerfile=plan.agent_dockerfile_path,
         )
 
-        # smolvm VM. --from carries the pre-packed .smolmachine
-        # artifact; --allow-cidr + -e carry the per-bottle TSI
-        # allowlist + env. The allowlist is the per-bottle
-        # loopback alias — narrowing it to one /32 keeps the
-        # agent from reaching other host loopback services or
-        # other bottles' published ports. Smolfile isn't usable
-        # here — smolvm 0.8.0 makes `--from` and `--smolfile`
-        # mutually exclusive.
-        _smolvm.machine_create(
+        _launch_vm(plan, agent_from_path, loopback_ip, stack)
+        _init_vm(plan)
+
+        bottle = SmolmachinesBottle(
             plan.machine_name,
-            from_path=agent_from_path,
-            allow_cidrs=[f"{loopback_ip}/32"],
-            env=plan.guest_env,
-        )
-        stack.callback(_smolvm.machine_delete, plan.machine_name)
-        # Workaround smolvm 0.8.0: `--allow-cidr` is silently
-        # dropped when combined with `--from`. Patch the persisted
-        # state DB to set the allowlist before start so the booted
-        # VM's TSI actually enforces. See loopback_alias's module
-        # docstring for the investigation that led here.
-        _loopback.force_allowlist(plan.machine_name, [f"{loopback_ip}/32"])
-        _smolvm.machine_start(plan.machine_name)
-        stack.callback(_smolvm.machine_stop, plan.machine_name)
-
-        # 6. Repair filesystem ownership + perms that smolvm's
-        # pack process remapped to the host invoker's uid (501
-        # on macOS) rather than preserving the image's expected
-        # ownership.
-        #
-        #  - /home/node → node:node so the node user can write
-        #    its own dotfiles (claude appendFileSync on
-        #    ~/.claude.json otherwise bails with ENOENT/EPERM
-        #    and the TUI hangs without surfacing the error).
-        #  - /tmp + /var/tmp → root:root mode 1777 so non-root
-        #    processes can create their per-uid scratch dirs
-        #    (claude-code creates /tmp/claude-<uid>/ as soon as
-        #    it spawns a Bash tool call).
-        #
-        # All folded into one sh -c so we only pay one
-        # machine_exec round trip — back-to-back exec calls
-        # right after machine_start hit a SIGKILL race in
-        # libkrun's exec channel (see provision_ca for the
-        # other half of this same workaround).
-        _smolvm.machine_exec(plan.machine_name, [
-            "sh", "-c",
-            "chown -R node:node /home/node && "
-            "chown root:root /tmp /var/tmp && "
-            "chmod 1777 /tmp /var/tmp",
-        ])
-
-        # Wait briefly for the VM to settle. Back-to-back smolvm
-        # machine_exec calls immediately after machine_start
-        # occasionally SIGKILL the in-VM child at ~100ms (looks
-        # like a VM warm-up race in libkrun's exec channel).
-        # 1.5s is empirically enough to dodge it; provisioning
-        # already takes seconds so the wait is amortized.
-        time.sleep(1.5)
-
-        # 7. Provision (CA / prompt / skills / git / supervise).
-        prompt_path = provision(plan, plan.machine_name)
-
-        yield SmolmachinesBottle(
-            plan.machine_name,
-            prompt_path=prompt_path,
+            prompt_path=None,
             guest_env=plan.guest_env,
             agent_command=plan.agent_command,
             agent_prompt_mode=plan.agent_prompt_mode,
         )
+        bottle._prompt_path = provision(plan, bottle)
+
+        yield bottle
     finally:
+        _teardown_smolmachines(stack, plan)
+
+
+def _teardown_smolmachines(
+    stack: ExitStack,
+    plan: SmolmachinesBottlePlan,
+) -> None:
+    """Unwind the ExitStack, then revoke any provisioned deploy keys.
+
+    ExitStack errors are caught and logged (non-fatal) so that key
+    revocation always runs. Revocation errors propagate — a stranded
+    deploy key is a security concern the operator must address."""
+    teardown_exc: BaseException | None = None
+    try:
         stack.close()
+    except BaseException as exc:
+        teardown_exc = exc
+        warn(f"smolmachines teardown failed: {exc!r}")
+    bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+    revoke_git_gate_provisioned_keys(bottle, git_gate_state_dir(plan.slug))
+    if teardown_exc is not None:
+        raise teardown_exc
+
+
+def _allocate_resources(
+    plan: SmolmachinesBottlePlan,
+    stack: ExitStack,
+) -> tuple[str, str]:
+    """Reserve a loopback alias and create the per-bottle docker bridge.
+
+    macOS only routes 127.0.0.1 by default; the per-bottle alias
+    scopes TSI's allowlist to this bottle's published ports so the
+    agent can't reach other bottles' or host services' ports on
+    loopback. No-op on Linux."""
+    _loopback.ensure_pool()
+    loopback_ip = _loopback.allocate(plan.slug)
+    network = _bundle.bundle_network_name(plan.slug)
+    _bundle.create_bundle_network(network, plan.bundle_subnet, plan.bundle_gateway)
+    stack.callback(_bundle.remove_bundle_network, network)
+    return loopback_ip, network
+
+
+def _mint_certs(plan: SmolmachinesBottlePlan) -> SmolmachinesBottlePlan:
+    """Mint per-bottle CAs and return the plan with CA paths filled.
+
+    Pipelock always runs in the bundle. Egress's CA is only minted
+    when the bottle declares routes — otherwise egress runs idle
+    without MITM and the CA files would be unused."""
+    ca_cert_host, ca_key_host = pipelock_tls_init(plan.proxy_plan.yaml_path.parent)
+    proxy_plan = dataclasses.replace(
+        plan.proxy_plan,
+        ca_cert_host_path=ca_cert_host,
+        ca_key_host_path=ca_key_host,
+    )
+    egress_plan = plan.egress_plan
+    if egress_plan.routes:
+        egress_ca_host, egress_ca_cert_only = egress_tls_init(
+            plan.egress_plan.routes_path.parent,
+        )
+        egress_plan = dataclasses.replace(
+            egress_plan,
+            mitmproxy_ca_host_path=egress_ca_host,
+            mitmproxy_ca_cert_only_host_path=egress_ca_cert_only,
+            pipelock_ca_host_path=ca_cert_host,
+            # On smolmachines, egress's upstream is pipelock on the
+            # bundle's localhost — they're in the same container's
+            # network namespace.
+            pipelock_proxy_url=BUNDLE_LOCAL_PIPELOCK_URL,
+        )
+    return dataclasses.replace(plan, proxy_plan=proxy_plan, egress_plan=egress_plan)
+
+
+def _start_bundle(
+    plan: SmolmachinesBottlePlan,
+    network: str,
+    loopback_ip: str,
+    stack: ExitStack,
+) -> SmolmachinesBottlePlan:
+    """Build the BundleLaunchSpec, resolve token env, start the
+    sidecar bundle container, and register teardown."""
+    bundle_spec = _bundle_launch_spec(plan, network, loopback_ip)
+    token_env = _resolve_token_env(plan, dict(os.environ))
+    _bundle.ensure_bundle_image(bundle_spec.image)
+    _bundle.start_bundle(bundle_spec, env={**os.environ, **token_env})
+    stack.callback(_bundle.stop_bundle, plan.slug)
+    return plan
+
+
+def _discover_urls(
+    plan: SmolmachinesBottlePlan,
+    loopback_ip: str,
+) -> SmolmachinesBottlePlan:
+    """Discover host-side ports for published container ports and
+    return the plan with URLs + guest_env stamped in.
+
+    Docker container IPs (192.168.x.x in the daemon's bridge)
+    aren't reachable from the smolvm guest on macOS — TSI uses
+    macOS networking, and macOS sees the daemon's bridge via the
+    published-port loopback forward only.
+
+    Proxy hop order: when the bottle declares egress routes, the
+    agent's first hop is egress (for token injection), then
+    pipelock. Without routes, the agent dials pipelock directly.
+    NO_PROXY includes the per-bottle loopback alias so the
+    supervise + git-gate URLs bypass HTTPS_PROXY."""
+    if plan.egress_plan.routes:
+        agent_facing_port = _EGRESS_PORT
+    else:
+        agent_facing_port = _PIPELOCK_PORT
+    agent_facing_host_port = _bundle.bundle_host_port(
+        plan.slug, agent_facing_port, host_ip=loopback_ip,
+    )
+    agent_proxy_url = f"http://{loopback_ip}:{agent_facing_host_port}"
+
+    agent_git_gate_host = ""
+    if plan.git_gate_plan.upstreams:
+        git_gate_host_port = _bundle.bundle_host_port(
+            plan.slug, _GIT_HTTP_PORT, host_ip=loopback_ip,
+        )
+        agent_git_gate_host = f"{loopback_ip}:{git_gate_host_port}"
+
+    agent_supervise_url = ""
+    if plan.supervise_plan is not None:
+        supervise_host_port = _bundle.bundle_host_port(
+            plan.slug, _SUPERVISE_PORT, host_ip=loopback_ip,
+        )
+        agent_supervise_url = f"http://{loopback_ip}:{supervise_host_port}/"
+
+    existing_no_proxy = plan.guest_env.get("NO_PROXY", "localhost,127.0.0.1")
+    guest_env = {
+        **plan.guest_env,
+        "HTTPS_PROXY": agent_proxy_url,
+        "HTTP_PROXY":  agent_proxy_url,
+        "NO_PROXY":    f"{existing_no_proxy},{loopback_ip}",
+    }
+    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
+
+    return dataclasses.replace(
+        plan,
+        guest_env=guest_env,
+        agent_proxy_url=agent_proxy_url,
+        agent_git_gate_host=agent_git_gate_host,
+        agent_supervise_url=agent_supervise_url,
+    )
+
+
+def _launch_vm(
+    plan: SmolmachinesBottlePlan,
+    agent_from_path: Path,
+    loopback_ip: str,
+    stack: ExitStack,
+) -> None:
+    """Create, patch, and start the smolvm VM; register teardown.
+
+    --allow-cidr is the per-bottle loopback alias so the guest can
+    only reach this bottle's bundle ports. force_allowlist patches
+    smolvm 0.8.0's silent-drop of --allow-cidr when combined with
+    --from. Smolfile isn't usable here — smolvm 0.8.0 makes --from
+    and --smolfile mutually exclusive."""
+    _smolvm.machine_create(
+        plan.machine_name,
+        from_path=agent_from_path,
+        allow_cidrs=[f"{loopback_ip}/32"],
+        env=plan.guest_env,
+    )
+    stack.callback(_smolvm.machine_delete, plan.machine_name)
+    # Workaround smolvm 0.8.0: `--allow-cidr` is silently dropped
+    # when combined with `--from`. Patch the persisted state DB
+    # before start so the booted VM's TSI actually enforces.
+    _loopback.force_allowlist(plan.machine_name, [f"{loopback_ip}/32"])
+    _smolvm.machine_start(plan.machine_name)
+    stack.callback(_smolvm.machine_stop, plan.machine_name)
+
+
+def _init_vm(plan: SmolmachinesBottlePlan) -> None:
+    """Repair filesystem ownership and wait for exec channel readiness.
+
+    Ownership repair: smolvm's pack process remaps files to the host
+    invoker's uid (501 on macOS). /home/node must be node:node so
+    Claude Code can write ~/.claude.json; /tmp + /var/tmp need root
+    mode 1777 so non-root processes can create per-uid scratch dirs.
+    All folded into one sh -c to avoid back-to-back exec calls
+    immediately after machine_start (libkrun exec-channel race).
+
+    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",
+        "chown -R node:node /home/node && "
+        "chown root:root /tmp /var/tmp && "
+        "chmod 1777 /tmp /var/tmp",
+    ])
+    _smolvm.wait_exec_ready(plan.machine_name)
 
 
 def _bundle_launch_spec(
@@ -305,10 +330,10 @@ def _bundle_launch_spec(
     Daemons in the CSV:
       - egress + pipelock are always present (pipelock is the
         agent's first hop; egress is its upstream).
-      - git-gate is conditional on plan.git_gate_plan.upstreams.
+      - git-gate + git-http are conditional on plan.git_gate_plan.upstreams.
       - supervise is conditional on plan.supervise_plan.
 
-    Env + volumes are the union of the four daemons' needs, with
+    Env + volumes are the union of the sidecar daemons' needs, with
     daemon-private values only (HTTPS_PROXY is scoped to the
     egress process by egress_entrypoint.sh — see PRD 0024's bundle
     bind-address PR)."""
@@ -320,10 +345,9 @@ def _bundle_launch_spec(
     # is "agent-facing" gets its port published on the host
     # loopback (see `_ensure_smolmachine`'s discovery loop) and the
     # other stays bundle-internal. The bundle is NOT reachable by
-    # bridge IP from the smolvm guest, so the
-    # PRD-0023-chunk-3 EGRESS_LISTEN_HOST=127.0.0.1 mitigation
-    # isn't needed: the agent can only dial whatever daemon's
-    # host port we publish, period.
+    # bridge IP from the smolvm guest on macOS — TSI uses macOS
+    # networking, and macOS sees the daemon's bridge via the
+    # published-port loopback forward only.
 
     # --- pipelock ---------------------------------------------
     pp = plan.proxy_plan
@@ -350,10 +374,9 @@ def _bundle_launch_spec(
             env.append(token_env)
 
     # --- git-gate ---------------------------------------------
-    extra_hosts: list[str] = []
     gp = plan.git_gate_plan
     if gp.upstreams:
-        daemons.append("git-gate")
+        daemons += ["git-gate", "git-http"]
         volumes += [
             (str(gp.entrypoint_script), GIT_GATE_ENTRYPOINT_IN_CONTAINER, True),
             (str(gp.hook_script), GIT_GATE_HOOK_IN_CONTAINER, True),
@@ -395,7 +418,7 @@ def _bundle_launch_spec(
     else:
         ports_to_publish = [_PIPELOCK_PORT]
     if gp.upstreams:
-        ports_to_publish.append(_GIT_GATE_PORT)
+        ports_to_publish.append(_GIT_HTTP_PORT)
     if sp is not None:
         ports_to_publish.append(_SUPERVISE_PORT)
 
@@ -414,15 +437,13 @@ def _bundle_launch_spec(
 
 
 def _resolve_token_env(
-    plan: SmolmachinesBottlePlan, host_env: object
+    plan: SmolmachinesBottlePlan, host_env: dict[str, str],
 ) -> dict[str, str]:
     """Resolve the egress token env-var values from the host's
     environ so they reach the bundle's process env via docker's
     `-e NAME` inheritance. Empty when no routes declare auth."""
-    ep = plan.egress_plan
-    if not ep.routes:
-        return {}
-    return egress_resolve_token_values(ep.token_env_map, dict(host_env))
+    effective_env = {**host_env, **plan.agent_provision.provisioned_env}
+    return egress_resolve_token_values(plan.egress_plan.token_env_map, effective_env)
 
 
 def _ensure_smolmachine(image_ref: str, *, dockerfile: str = "") -> Path:

bot_bottle/backend/smolmachines/loopback_alias.py

@@ -45,6 +45,7 @@ alias gets handed to a new bottle."""
 
 from __future__ import annotations
 
+import fcntl
 import json
 import os
 import platform
@@ -83,6 +84,14 @@ _POOL_START = 16
 _POOL_END = 31  # inclusive
 
 
+# File lock that serialises concurrent allocate() calls so two
+# simultaneous launches can't read the same docker state and claim
+# the same alias. Narrowed to the allocate() call itself; docker run
+# runs after the lock is released. Once the container is running it
+# appears in docker state and future allocate() calls will see it.
+_ALLOC_LOCK_PATH = Path.home() / ".cache" / "bot-bottle" / "smolmachines.lock"
+
+
 # Loopback aliases pool: 127.0.0.<start>..127.0.0.<end>.
 def _pool_addresses() -> list[str]:
     return [f"127.0.0.{i}" for i in range(_POOL_START, _POOL_END + 1)]
@@ -179,9 +188,20 @@ def allocate(slug: str) -> str:
     On non-macOS the whole `127.0.0.0/8` is loopback by default;
     `127.0.0.1` is fine to share and we skip the alias dance.
     This still returns a deterministic address so launch.py's
-    callers don't have to branch on platform."""
+    callers don't have to branch on platform.
+
+    An exclusive file lock serialises concurrent calls so two
+    simultaneous launches don't read the same docker state and
+    claim the same alias."""
     if not _is_macos():
         return "127.0.0.1"
+    _ALLOC_LOCK_PATH.parent.mkdir(parents=True, exist_ok=True)
+    with open(_ALLOC_LOCK_PATH, "w") as lf:
+        fcntl.flock(lf, fcntl.LOCK_EX)
+        return _allocate_locked()
+
+
+def _allocate_locked() -> str:
     in_use = _aliases_in_use()
     for ip in _pool_addresses():
         if ip not in in_use:

bot_bottle/backend/smolmachines/prepare.py

@@ -12,9 +12,10 @@ from __future__ import annotations
 
 import os
 from datetime import datetime, timezone
+from dataclasses import replace
 from pathlib import Path
 
-from ...agent_provider import runtime_for
+from ...agent_provider import agent_provision_plan, runtime_for
 from ...backend import BottleSpec
 from ...backend.docker.bottle_state import (
     BottleMetadata,
@@ -27,9 +28,11 @@ from ...backend.docker.bottle_state import (
     write_metadata,
 )
 from ...egress import Egress
+from ...env import resolve_env
 from ...git_gate import GitGate
 from ...pipelock import PipelockProxy
 from ...supervise import Supervise
+from ...workspace import workspace_plan as resolve_workspace_plan
 from .bottle_plan import SmolmachinesBottlePlan
 from .util import smolmachines_bundle_subnet, smolmachines_preflight
 
@@ -58,6 +61,8 @@ def resolve_plan(
     bottle = manifest.bottle_for(spec.agent_name)
     provider = bottle.agent_provider
     provider_runtime = runtime_for(provider.template)
+    guest_home = "/home/node"
+    workspace_plan = resolve_workspace_plan(spec, guest_home=guest_home)
 
     slug = spec.identity or bottle_identity(spec.agent_name)
 
@@ -69,72 +74,34 @@ def resolve_plan(
         cwd=spec.user_cwd if spec.copy_cwd else "",
         copy_cwd=spec.copy_cwd,
         started_at=datetime.now(timezone.utc).isoformat(),
-        # No compose project for smolmachines bottles; chunk 4
-        # will give dashboard discovery a backend-specific path.
         compose_project="",
+        backend="smolmachines",
     ))
 
     subnet, gateway, bundle_ip = smolmachines_bundle_subnet(slug)
 
-    # Agent's env: the prepare-time view doesn't yet know the
-    # host loopback ports the bundle's daemons get published on
-    # (those come from docker AFTER `docker run` returns), so
-    # HTTPS_PROXY / GIT_GATE_URL / MCP_SUPERVISE_URL are
-    # populated in launch.py and stamped onto guest_env there.
-    # What we set here is the part that doesn't depend on
-    # bundle bringup — bottle.env literals, the empty-NO_PROXY
-    # safe default, and the TLS trust env trio
-    # (NODE_EXTRA_CA_CERTS / SSL_CERT_FILE / REQUESTS_CA_BUNDLE)
-    # pointing at Debian's update-ca-certificates output bundle.
+    # Agent's env: resolve through resolve_env() so ?prompt entries
+    # are prompted and ${HOST_VAR} entries are interpolated — matching
+    # the Docker backend's contract. Forwarded (secret/interpolated)
+    # values still reach the guest as -e K=V smolvm flags because
+    # smolvm 0.8.0 has no env-file or stdin injection path; this is
+    # the known argv-exposure gap documented in PRD 0038.
+    # HTTPS_PROXY / GIT_GATE_URL / MCP_SUPERVISE_URL are populated
+    # in launch.py after bundle bringup.
+    resolved = resolve_env(manifest, spec.agent_name)
     guest_env: dict[str, str] = {
-        **bottle.env,
+        **resolved.literals,
+        **resolved.forwarded,
         "NO_PROXY":    "localhost,127.0.0.1",
         "NODE_EXTRA_CA_CERTS": "/etc/ssl/certs/ca-certificates.crt",
         "SSL_CERT_FILE":       "/etc/ssl/certs/ca-certificates.crt",
         "REQUESTS_CA_BUNDLE":  "/etc/ssl/certs/ca-certificates.crt",
     }
 
-    # Inner Plans for the four bundle daemons. The ABCs are
-    # platform-neutral — `.prepare()` writes config files + returns
-    # a Plan dataclass with no backend-specific assumptions. State
-    # dirs are still keyed by slug under the docker backend's
-    # bottle_state layout (shared on-host convention; not a docker
-    # dependency).
-    pipelock_dir = pipelock_state_dir(slug)
-    pipelock_dir.mkdir(parents=True, exist_ok=True)
-    proxy_plan = PipelockProxy().prepare(bottle, slug, pipelock_dir)
-
     git_gate_dir = git_gate_state_dir(slug)
     git_gate_dir.mkdir(parents=True, exist_ok=True)
     git_gate_plan = GitGate().prepare(bottle, slug, git_gate_dir)
 
-    egress_dir = egress_state_dir(slug)
-    egress_dir.mkdir(parents=True, exist_ok=True)
-    egress_plan = Egress().prepare(bottle, slug, egress_dir)
-
-    # Claude-code refuses to start without *something* it
-    # recognises as a credential. When the bottle has an egress
-    # route carrying the `claude_code_oauth` role marker, egress
-    # strips + re-injects the real Authorization header on the
-    # outbound leg using a token held in egress's own environ — so
-    # the agent gets a non-secret placeholder here (matches the
-    # docker backend's forwarded_env logic in
-    # bot_bottle/backend/docker/prepare.py).
-    has_provider_auth = any(
-        provider_runtime.auth_role in r.roles for r in egress_plan.routes
-    )
-    if has_provider_auth:
-        guest_env[provider_runtime.placeholder_env] = "egress-placeholder"
-    if provider.template == "claude" and has_provider_auth:
-        guest_env.setdefault("CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC", "1")
-        guest_env.setdefault("DISABLE_ERROR_REPORTING", "1")
-
-    supervise_plan = None
-    if bottle.supervise:
-        supervise_dir = supervise_state_dir(slug)
-        supervise_dir.mkdir(parents=True, exist_ok=True)
-        supervise_plan = Supervise().prepare(slug, supervise_dir)
-
     # Prompt file is always written (mode 0o600) so the in-VM
     # path always exists. Content is the agent's `prompt`
     # field (markdown body) — empty for agents with no prompt.
@@ -162,26 +129,64 @@ def resolve_plan(
     else:
         image_default = provider_runtime.image
     agent_image_ref = os.environ.get("BOT_BOTTLE_IMAGE", image_default)
+    agent_provision = agent_provision_plan(
+        template=provider.template,
+        dockerfile=agent_dockerfile_path,
+        state_dir=agent_dir,
+        guest_home=guest_home,
+        guest_env=guest_env,
+        forward_host_credentials=provider.forward_host_credentials,
+        auth_token=provider.auth_token,
+        host_env=dict(os.environ),
+        trusted_project_path=workspace_plan.workdir,
+    )
+    merged_guest_env = dict(agent_provision.guest_env)
+    for key, val in agent_provision.env_vars.items():
+        merged_guest_env.setdefault(key, val)
+    agent_provision = replace(agent_provision, guest_env=merged_guest_env)
+
+    # Inner Plans for the four bundle daemons. The ABCs are
+    # platform-neutral — `.prepare()` writes config files + returns
+    # a Plan dataclass with no backend-specific assumptions. State
+    # dirs are still keyed by slug under the docker backend's
+    # bottle_state layout (shared on-host convention; not a docker
+    # dependency).
+    pipelock_dir = pipelock_state_dir(slug)
+    pipelock_dir.mkdir(parents=True, exist_ok=True)
+    proxy_plan = PipelockProxy().prepare(
+        bottle, slug, pipelock_dir, agent_provision.egress_routes,
+    )
+
+    egress_dir = egress_state_dir(slug)
+    egress_dir.mkdir(parents=True, exist_ok=True)
+    egress_plan = Egress().prepare(
+        bottle, slug, egress_dir, agent_provision.egress_routes,
+    )
+
+    supervise_plan = None
+    if bottle.supervise:
+        supervise_dir = supervise_state_dir(slug)
+        supervise_dir.mkdir(parents=True, exist_ok=True)
+        supervise_plan = Supervise().prepare(slug, supervise_dir)
 
     return SmolmachinesBottlePlan(
         spec=spec,
         stage_dir=stage_dir,
+        guest_home=guest_home,
         slug=slug,
         bundle_subnet=subnet,
         bundle_gateway=gateway,
         bundle_ip=bundle_ip,
         machine_name=machine_name,
         agent_image_ref=agent_image_ref,
-        guest_env=guest_env,
+        guest_env=agent_provision.guest_env,
         prompt_file=prompt_file,
         proxy_plan=proxy_plan,
         git_gate_plan=git_gate_plan,
         egress_plan=egress_plan,
         supervise_plan=supervise_plan,
-        agent_command=provider_runtime.command,
-        agent_prompt_mode=provider_runtime.prompt_mode,
-        agent_provider_template=provider.template,
-        agent_dockerfile_path=agent_dockerfile_path,
+        agent_provision=agent_provision,
+        workspace_plan=workspace_plan,
     )
 
 

bot_bottle/backend/smolmachines/provision/__init__.py

@@ -1,14 +1,12 @@
-"""Provisioning helpers for the smolmachines backend (PRD 0023
-chunk 4).
+"""Backend-infrastructure provisioners for the smolmachines backend.
 
-Each method maps onto one of `BottleBackend`'s `provision_*`
-overrides. They run after the VM is up + the bundle is reachable
-and copy host-side state (prompt, skills, .git, CA cert,
-supervise MCP config) into the guest via `smolvm machine cp` /
-`smolvm machine exec`.
+Per PRD 0050 the per-provider provisioning steps (prompt, skills,
+declarative provision-plan apply, supervise MCP registration) live on
+the `AgentProvider` plugin under `bot_bottle/contrib/`. The modules
+left in this subpackage handle only the steps that are
+backend-specific:
 
-Chunk 4a ships `provision_prompt` and `provision_skills` — the
-two that don't depend on agent-image tooling (claude-code,
-update-ca-certificates) beyond `cp` and `mkdir`. provision_ca /
-provision_git / provision_supervise land once the agent-image
-gap is solved."""
+  - ca.py        — install per-bottle CA bundle into the guest trust store
+  - git.py       — copy host cwd `.git` into the guest when --cwd is used
+  - workspace.py — copy the operator workspace into the guest
+"""

bot_bottle/backend/smolmachines/provision/ca.py

@@ -2,8 +2,8 @@
 trust store (PRD 0023 chunk 4d).
 
 Mirrors `backend.docker.provision.ca`: select the right CA (egress
-when the bottle has routes, else pipelock), `smolvm machine cp` it
-to Debian's `/usr/local/share/ca-certificates/` path,
+when the bottle has routes, else pipelock), copy it to Debian's
+`/usr/local/share/ca-certificates/` path,
 `update-ca-certificates` to rebuild the trust bundle, and log the
 fingerprint once. The selected cert depends on the agent's
 HTTP_PROXY target — same logic as the docker backend, since the
@@ -15,6 +15,8 @@ flag exists; the VM init is root), so we don't need the explicit
 
 from __future__ import annotations
 
+import time
+
 from ....log import die
 from ...util import (
     AGENT_CA_BUNDLE,
@@ -22,17 +24,20 @@ from ...util import (
     log_ca_fingerprint,
     select_ca_cert,
 )
-from .. import smolvm as _smolvm
+from ... import Bottle, ExecResult
 from ..bottle_plan import SmolmachinesBottlePlan
 
 
-def provision_ca(plan: SmolmachinesBottlePlan, target: str) -> None:
+_SIGKILL_EXIT = 128 + 9
+
+
+def provision_ca(plan: SmolmachinesBottlePlan, bottle: Bottle) -> None:
     """Copy the agent-facing CA cert into the guest, rebuild the
     trust bundle, emit a one-line fingerprint log. Called from
     `BottleBackend.provision` after the smolvm guest is up."""
     cert_host_path, label = select_ca_cert(plan.egress_plan, plan.proxy_plan)
 
-    _smolvm.machine_cp(str(cert_host_path), f"{target}:{AGENT_CA_PATH}")
+    bottle.cp_in(str(cert_host_path), AGENT_CA_PATH)
     # Mode 0644 — readable to non-root tools in the guest.
     # update-ca-certificates rebuilds the bundle at AGENT_CA_BUNDLE,
     # which is what curl / Python ssl / OpenSSL-based tools read by
@@ -40,22 +45,21 @@ def provision_ca(plan: SmolmachinesBottlePlan, target: str) -> None:
     # REQUESTS_CA_BUNDLE) on the guest_env covers Node + Python
     # `requests` / libraries that don't load the system bundle.
     #
-    # chown + chmod + update-ca-certificates run in one
-    # `sh -c` so we only pay one machine_exec round trip; the
-    # `&&` chaining surfaces the first failure as the return
-    # code.
-    r = _smolvm.machine_exec(target, [
-        "sh", "-c",
-        f"chown root:root {AGENT_CA_PATH} && "
-        f"chmod 644 {AGENT_CA_PATH} && "
-        f"update-ca-certificates",
-    ])
-    if r.returncode != 0 or "1 added" not in (r.stdout or ""):
+    r = _install_ca(bottle)
+    if r.returncode == _SIGKILL_EXIT:
+        # smolvm/libkrun can SIGKILL an otherwise-normal exec
+        # during early-VM provisioning. `update-ca-certificates`
+        # is idempotent, so retry the same install once after a
+        # short settle delay before treating it as fatal.
+        time.sleep(1.0)
+        r = _install_ca(bottle)
+
+    if r.returncode != 0:
         # update-ca-certificates not adding our cert is fatal —
         # claude-code's TLS handshake against the egress-MITM'd
         # api.anthropic.com would fail downstream. Bail early
-        # with what we can see (output is captured by smolvm so
-        # we can surface it).
+        # with what we can see (output is captured so we can
+        # surface it).
         die(
             f"update-ca-certificates didn't add the agent CA "
             f"(exit {r.returncode}): "
@@ -66,6 +70,23 @@ def provision_ca(plan: SmolmachinesBottlePlan, target: str) -> None:
     log_ca_fingerprint(cert_host_path, label)
 
 
+def _install_ca(bottle: Bottle) -> ExecResult:
+    # chown + chmod + update-ca-certificates + bundle
+    # verification run in one exec so we only pay one
+    # round trip; the `&&` chaining surfaces the first failure
+    # as the return code. The verify check is more stable than
+    # requiring "1 added" in stdout: a retry after a
+    # partially-completed first run may legitimately report "0
+    # added" while the cert is already installed.
+    return bottle.exec(
+        f"chown root:root {AGENT_CA_PATH} && "
+        f"chmod 644 {AGENT_CA_PATH} && "
+        f"update-ca-certificates && "
+        f"openssl verify -CAfile {AGENT_CA_BUNDLE} {AGENT_CA_PATH}",
+        user="root",
+    )
+
+
 # Re-exported for the launch/provision_ca caller + tests. The path
 # constants live in the shared `backend.util` (Debian's
 # `update-ca-certificates` layout is the same in both backends).

bot_bottle/backend/smolmachines/provision/git.py

@@ -4,7 +4,7 @@
 Three concerns, all about git in the agent:
 
   1. If --cwd was passed AND the host cwd has a .git, copy that
-     .git into /home/node/workspace/.git so the agent operates on
+     .git into the planned guest workspace so the agent operates on
      the user's repo.
   2. If the bottle declares `git` entries (PRD 0008), write a
      ~/.gitconfig with insteadOf rules so every git operation
@@ -18,7 +18,7 @@ Three concerns, all about git in the agent:
 Differs from `backend.docker.provision.git` in one address detail:
 the TSI-allowlisted guest can only reach the bundle's pinned IP
 (no DNS resolver in the /32 allowlist), so the insteadOf URLs
-are `git://<bundle_ip>:<port>/<name>.git` rather than the
+are `http://<bundle_ip>:<port>/<name>.git` rather than the
 docker backend's `git://git-gate/<name>.git`. The render itself
 is the shared `git_gate_render_gitconfig` on the platform-neutral
 git_gate module."""
@@ -26,71 +26,66 @@ git_gate module."""
 from __future__ import annotations
 
 import os
+import shlex
 import tempfile
 from pathlib import Path
 
 from ....git_gate import git_gate_render_gitconfig
 from ....log import info
-from .. import smolvm as _smolvm
+from ... import Bottle
 from ..bottle_plan import SmolmachinesBottlePlan
 
 
-# `node` is the agent user from the repo Dockerfile. Override via
-# BOT_BOTTLE_GUEST_HOME mirrors the docker backend's
-# BOT_BOTTLE_CONTAINER_HOME knob — same purpose, different
-# transport.
-_DEFAULT_GUEST_HOME = "/home/node"
-
-
-def _guest_home() -> str:
-    return os.environ.get("BOT_BOTTLE_GUEST_HOME", _DEFAULT_GUEST_HOME)
-
-
-def provision_git(plan: SmolmachinesBottlePlan, target: str) -> None:
+def provision_git(plan: SmolmachinesBottlePlan, bottle: Bottle) -> None:
     """Set up git inside the guest. Runs all three subcases; each
     no-ops when its condition isn't met."""
-    _provision_cwd_git(plan, target)
-    _provision_git_gate_config(plan, target)
-    _provision_git_user(plan, target)
+    _provision_cwd_git(plan, bottle)
+    _provision_git_gate_config(plan, bottle)
+    _provision_git_user(plan, bottle)
 
 
-def _provision_cwd_git(plan: SmolmachinesBottlePlan, target: str) -> None:
+def _provision_cwd_git(plan: SmolmachinesBottlePlan, bottle: Bottle) -> None:
     """If --cwd was set and the host cwd has a .git directory, copy
     it into <guest_home>/workspace/.git and fix ownership. No-op
     otherwise."""
-    if not (plan.spec.copy_cwd and Path(plan.spec.user_cwd, ".git").is_dir()):
+    workspace = plan.workspace_plan
+    if not (workspace.enabled and workspace.copy_git and workspace.has_host_git_dir):
         return
-    guest_workspace_git = f"{_guest_home()}/workspace/.git"
-    info(f"copying {plan.spec.user_cwd}/.git -> {target}:{guest_workspace_git}")
-    # mkdir -p the workspace dir so `machine cp` lands the .git
+    guest_workspace_git = f"{workspace.guest_path}/.git"
+    host_git = str(workspace.host_path / ".git")
+    info(f"copying {host_git} -> {bottle.name}:{guest_workspace_git}")
+    # mkdir -p the workspace dir so cp_in lands the .git
     # directly there even on first-time bottles.
-    _smolvm.machine_exec(target, ["mkdir", "-p", f"{_guest_home()}/workspace"])
-    _smolvm.machine_cp(
-        f"{plan.spec.user_cwd}/.git", f"{target}:{guest_workspace_git}",
-    )
-    # `machine cp` lands files as root; the agent runs as node so
+    bottle.exec(f"mkdir -p {shlex.quote(workspace.guest_path)}", user="root")
+    bottle.cp_in(host_git, guest_workspace_git)
+    # cp_in lands files as root; the agent runs as node so
     # the workspace tree must be chowned over.
-    _smolvm.machine_exec(
-        target, ["chown", "-R", "node:node", guest_workspace_git],
+    bottle.exec(
+        f"chown -R {shlex.quote(workspace.owner)} {shlex.quote(guest_workspace_git)}",
+        user="root",
     )
 
 
-def _provision_git_gate_config(plan: SmolmachinesBottlePlan, target: str) -> None:
+def _provision_git_gate_config(
+    plan: SmolmachinesBottlePlan, bottle: Bottle
+) -> None:
     """Write ~/.gitconfig in the guest with the git-gate insteadOf
     rules. No-op when the bottle has no `git` entries."""
-    bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
-    if not bottle.git:
+    manifest_bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+    if not manifest_bottle.git:
         return
 
-    # `127.0.0.1:<host port>` form: the bundle's git-gate port
-    # is published on host loopback at launch time so the
-    # smolvm guest (which can only reach macOS networking via
+    # `<loopback alias>:<host port>` form: the bundle's git-gate
+    # HTTP port is published on host loopback at launch time so
+    # the smolvm guest (which can only reach macOS networking via
     # TSI, not the docker bridge IP) can dial it. launch.py
     # populates `plan.agent_git_gate_host` after bundle bringup.
-    content = git_gate_render_gitconfig(bottle.git, plan.agent_git_gate_host)
+    content = git_gate_render_gitconfig(
+        manifest_bottle.git, plan.agent_git_gate_host, scheme="http",
+    )
 
-    guest_gitconfig = f"{_guest_home()}/.gitconfig"
-    # Stage the file under the plan's stage_dir so `machine cp`
+    guest_gitconfig = f"{plan.guest_home}/.gitconfig"
+    # Stage the file under the plan's stage_dir so cp_in
     # has a stable host path. The plan's stage_dir is cleaned up
     # by start.py's session-end teardown.
     with tempfile.NamedTemporaryFile(
@@ -101,41 +96,38 @@ def _provision_git_gate_config(plan: SmolmachinesBottlePlan, target: str) -> Non
         config_file = Path(f.name)
     os.chmod(config_file, 0o600)
 
-    info(f"writing {guest_gitconfig} with {len(bottle.git)} insteadOf rule(s)")
-    _smolvm.machine_cp(str(config_file), f"{target}:{guest_gitconfig}")
-    _smolvm.machine_exec(target, ["chown", "node:node", guest_gitconfig])
-    _smolvm.machine_exec(target, ["chmod", "644", guest_gitconfig])
+    info(f"writing {guest_gitconfig} with {len(manifest_bottle.git)} insteadOf rule(s)")
+    bottle.cp_in(str(config_file), guest_gitconfig)
+    bottle.exec(
+        f"chown node:node {shlex.quote(guest_gitconfig)} && "
+        f"chmod 644 {shlex.quote(guest_gitconfig)}",
+        user="root",
+    )
 
 
 def _provision_git_user(
-    plan: SmolmachinesBottlePlan, target: str,
+    plan: SmolmachinesBottlePlan, bottle: Bottle,
 ) -> None:
     """Apply `git config --global user.{name,email}` inside the
     guest as the node user so --global lands in the same
     `/home/node/.gitconfig` that `_provision_git_gate_config`
     writes to. No-op when the bottle didn't declare `git.user`.
 
-    Runs via `runuser -u node --`; HOME is forced via smolvm's
-    `-e` flag because runuser (without -l) inherits root's
-    HOME=/root, which would put --global in the wrong file."""
-    bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
-    gu = bottle.git_user
+    SmolmachinesBottle.exec(user="node") automatically sets
+    HOME=/home/node so --global writes to /home/node/.gitconfig."""
+    manifest_bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
+    gu = manifest_bottle.git_user
     if gu.is_empty():
         return
-    env = {"HOME": _guest_home(), "USER": "node"}
     if gu.name:
         info(f"git config --global user.name = {gu.name!r}")
-        _smolvm.machine_exec(
-            target,
-            ["runuser", "-u", "node", "--",
-             "git", "config", "--global", "user.name", gu.name],
-            env=env,
+        bottle.exec(
+            f"git config --global user.name {shlex.quote(gu.name)}",
+            user="node",
         )
     if gu.email:
         info(f"git config --global user.email = {gu.email!r}")
-        _smolvm.machine_exec(
-            target,
-            ["runuser", "-u", "node", "--",
-             "git", "config", "--global", "user.email", gu.email],
-            env=env,
+        bottle.exec(
+            f"git config --global user.email {shlex.quote(gu.email)}",
+            user="node",
         )

bot_bottle/backend/smolmachines/provision/prompt.py

@@ -1,42 +0,0 @@
-"""Copy the agent prompt into a running smolmachines bottle.
-
-The prompt file is always copied (so the in-guest path always
-exists) but `--append-system-prompt-file` only fires when the
-agent actually has a prompt — the return value signals which
-case, mirroring the docker backend's contract.
-
-`smolvm machine cp` lands files as root inside the VM; the claude
-process runs as `node`, so we chown + chmod the prompt after the
-copy. Same flow as the docker backend's provision_prompt."""
-
-from __future__ import annotations
-
-import os
-
-from .. import smolvm as _smolvm
-from ..bottle_plan import SmolmachinesBottlePlan
-
-
-# `node` is the agent user from the repo Dockerfile.
-# BOT_BOTTLE_GUEST_HOME mirrors the docker backend's
-# BOT_BOTTLE_CONTAINER_HOME knob.
-_DEFAULT_GUEST_HOME = "/home/node"
-
-
-def provision_prompt(plan: SmolmachinesBottlePlan, target: str) -> str | None:
-    """Copy the prompt file into the running smolvm guest, fix
-    ownership/mode. Returns the in-guest path if the agent has a
-    non-empty prompt (drives --append-system-prompt-file), else
-    None. The file is copied either way so the path always
-    exists — mirrors the docker backend's behavior."""
-    guest_home = os.environ.get("BOT_BOTTLE_GUEST_HOME", _DEFAULT_GUEST_HOME)
-    in_guest_prompt_path = f"{guest_home}/.bot-bottle-prompt.txt"
-
-    _smolvm.machine_cp(str(plan.prompt_file), f"{target}:{in_guest_prompt_path}")
-    # machine cp lands as root, source's 0o600 mode is preserved —
-    # node can't read its own prompt without these two.
-    _smolvm.machine_exec(target, ["chown", "node:node", in_guest_prompt_path])
-    _smolvm.machine_exec(target, ["chmod", "600", in_guest_prompt_path])
-
-    agent = plan.spec.manifest.agents[plan.spec.agent_name]
-    return in_guest_prompt_path if agent.prompt else None

bot_bottle/backend/smolmachines/provision/skills.py

@@ -1,63 +0,0 @@
-"""Copy host-side skill directories into a running smolmachines
-bottle.
-
-Skills are validated on the host before launch by
-`BottleBackend._validate_skills`; this module assumes that
-validation has already run. A skill that disappears between
-validation and copy still dies loudly rather than silently
-producing a partial guest."""
-
-from __future__ import annotations
-
-import os
-
-from ....log import die, info
-from ...util import host_skill_dir
-from .. import smolvm as _smolvm
-from ..bottle_plan import SmolmachinesBottlePlan
-
-
-# In-guest path mirrors the docker backend's claude-skills
-# convention (~/.claude/skills/<name>/) under the node user's
-# home — same path as the real bot-bottle image's
-# /home/node/.claude/skills (pre-created in the Dockerfile).
-_DEFAULT_SKILLS_DIR = "/home/node/.claude/skills"
-
-
-def provision_skills(plan: SmolmachinesBottlePlan, target: str) -> None:
-    """Copy each of the agent's named skills from the host's
-    ~/.claude/skills/<name>/ into the guest's equivalent path.
-    For each skill: `mkdir -p` the destination, `smolvm machine cp`
-    the host source dir over, then chown the result to node:node so
-    the agent can read it. No-op when the agent has no skills.
-
-    smolvm machine cp on a directory copies recursively (same
-    semantics as `cp -r`); unlike docker cp's trailing-slash
-    convention, smolvm doesn't need the `/.` suffix dance.
-
-    machine cp lands files as root inside the VM, so we chown each
-    skill tree over to node:node after the copy — same pattern as
-    the docker backend's provision_prompt."""
-    agent = plan.spec.manifest.agents[plan.spec.agent_name]
-    if not agent.skills:
-        return
-
-    skills_dir = os.environ.get(
-        "BOT_BOTTLE_GUEST_SKILLS_DIR", _DEFAULT_SKILLS_DIR,
-    )
-
-    _smolvm.machine_exec(target, ["mkdir", "-p", skills_dir])
-
-    for name in agent.skills:
-        src = host_skill_dir(name)
-        if not os.path.isdir(src):
-            die(
-                f"skill {name!r} disappeared from host between "
-                f"validation and copy at {src}."
-            )
-        dst = f"{skills_dir}/{name}"
-        info(f"copying skill {name} into {target}:{dst}")
-        # Wipe any prior copy so re-runs don't accumulate.
-        _smolvm.machine_exec(target, ["rm", "-rf", dst])
-        _smolvm.machine_cp(src, f"{target}:{dst}")
-        _smolvm.machine_exec(target, ["chown", "-R", "node:node", dst])

bot_bottle/backend/smolmachines/provision/supervise.py

@@ -1,67 +0,0 @@
-"""Supervise sidecar provisioning inside a running smolmachines
-bottle (PRD 0023 chunk 4d; PRD 0013 supervise plane).
-
-Registers the per-bottle supervise sidecar as an HTTP MCP server
-in the agent's claude-code config so the agent discovers the
-stuck-recovery MCP tools (pipelock-block, capability-block) at
-startup.
-
-Mirrors `backend.docker.provision.supervise` — same `claude mcp
-add` call, just dispatched via `smolvm machine exec` instead of
-`docker exec`, and against `<bundle_ip>:<port>` instead of the
-short `supervise` alias (no DNS in the TSI-allowlisted guest)."""
-
-from __future__ import annotations
-
-from ....log import info, warn
-from .. import smolvm as _smolvm
-from ..bottle_plan import SmolmachinesBottlePlan
-
-
-_SUPERVISE_MCP_NAME = "supervise"
-
-
-def provision_supervise(plan: SmolmachinesBottlePlan, target: str) -> None:
-    """Run `claude mcp add` inside the guest to register the
-    supervise sidecar in claude-code's user config. No-op when
-    bottle.supervise is False.
-
-    The URL is the agent-side endpoint launch.py populated after
-    bundle bringup — `http://127.0.0.1:<host port>/` rather than
-    the bundle's docker bridge IP, because that bridge isn't
-    reachable from the smolvm guest on macOS.
-
-    Failure is logged but not fatal: the bottle still works (you
-    just can't call supervise tools from the agent until the entry
-    is added manually). The operator sees the warning at launch."""
-    if plan.supervise_plan is None:
-        return
-    url = plan.agent_supervise_url
-    info(f"registering supervise MCP server in agent claude config → {url}")
-    # `claude mcp add --scope user` writes to ~/.claude.json. The
-    # agent is the `node` user; smolvm machine_exec runs as root
-    # by default, so we have to switch user explicitly and set
-    # HOME so the config lands in /home/node/.claude.json (where
-    # the agent's claude actually reads it from).
-    r = _smolvm.machine_exec(
-        target,
-        [
-            "runuser", "-u", "node", "--",
-            "env", "HOME=/home/node",
-            "claude", "mcp", "add",
-            "--scope", "user",
-            "--transport", "http",
-            _SUPERVISE_MCP_NAME,
-            url,
-        ],
-    )
-    if r.returncode != 0:
-        warn(
-            f"`claude mcp add supervise` failed (exit {r.returncode}): "
-            f"{(r.stderr or r.stdout or '').strip()}. Inside the bottle, "
-            f"register manually with: "
-            f"claude mcp add --scope user --transport http supervise {url}"
-        )
-
-
-__all__ = ["provision_supervise"]

bot_bottle/backend/smolmachines/provision/workspace.py

@@ -0,0 +1,32 @@
+"""Copy the operator workspace into a smolmachines guest."""
+
+from __future__ import annotations
+
+import shlex
+
+from ....log import info
+from ... import Bottle
+from ..bottle_plan import SmolmachinesBottlePlan
+
+
+def provision_workspace(plan: SmolmachinesBottlePlan, bottle: Bottle) -> None:
+    """Copy host cwd contents to the planned guest workspace."""
+    workspace = plan.workspace_plan
+    if not (workspace.enabled and workspace.copy_contents):
+        return
+
+    guest_parent = workspace.guest_path.rsplit("/", 1)[0] or "/"
+    guest_path_q = shlex.quote(workspace.guest_path)
+    guest_parent_q = shlex.quote(guest_parent)
+    owner_q = shlex.quote(workspace.owner)
+    mode_q = shlex.quote(workspace.mode)
+    info(f"copying {workspace.host_path} -> {bottle.name}:{workspace.guest_path}")
+    bottle.exec(
+        f"rm -rf {guest_path_q} && mkdir -p {guest_parent_q}",
+        user="root",
+    )
+    bottle.cp_in(str(workspace.host_path), workspace.guest_path)
+    bottle.exec(
+        f"chown -R {owner_q} {guest_path_q} && chmod {mode_q} {guest_path_q}",
+        user="root",
+    )

bot_bottle/backend/smolmachines/smolvm.py

@@ -27,11 +27,13 @@ from __future__ import annotations
 
 import shutil
 import subprocess
+import time
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Mapping, Sequence
 
 
+
 _SMOLVM = "smolvm"
 
 
@@ -197,6 +199,34 @@ def machine_exec(
     )
 
 
+def wait_exec_ready(name: str, *, timeout: float = 5.0) -> None:
+    """Poll `machine exec true` until exit 0 or `timeout` elapses.
+
+    Replaces `time.sleep(1.5)` after `machine_start`: libkrun's exec
+    channel needs a brief warm-up before back-to-back exec calls are
+    safe. Polling exits as soon as the channel is ready and fails
+    loudly if the VM never responds."""
+    deadline = time.monotonic() + timeout
+    delay = 0.1
+    while time.monotonic() < deadline:
+        r = machine_exec(name, ["true"])
+        if r.returncode == 0:
+            return
+        remaining = deadline - time.monotonic()
+        if remaining <= 0:
+            break
+        time.sleep(min(delay, remaining))
+        delay = min(delay * 2, 0.5)
+    argv = ["smolvm", "machine", "exec", "--name", name, "--", "true"]
+    raise SmolvmError(
+        argv,
+        subprocess.CompletedProcess(
+            args=argv, returncode=-1, stdout="",
+            stderr=f"exec channel not ready after {timeout:.0f}s — VM may have failed to boot.",
+        ),
+    )
+
+
 def machine_cp(src: str, dst: str) -> None:
     """`smolvm machine cp SRC DST`. Path syntax: `machine:path` to
     reference a path inside the VM, bare path for the host. Both

bot_bottle/cli/__init__.py

@@ -1,34 +1,35 @@
 """Main CLI dispatcher.
 
-Commands: cleanup, dashboard, edit, info, init, list, resume, start
+Commands: cleanup, edit, info, init, list, resume, start, supervise
 """
 
 from __future__ import annotations
 
 import sys
 
-from ..log import Die, die
+from ..log import Die, die, error
+from ..manifest import ManifestError
 from ._common import PROG
 from . import list as _list_mod
 from .cleanup import cmd_cleanup
-from .dashboard import cmd_dashboard
 from .edit import cmd_edit
 from .info import cmd_info
 from .init import cmd_init
 from .resume import cmd_resume
 from .start import cmd_start
+from .supervise import cmd_supervise
 
 cmd_list = _list_mod.cmd_list
 
 COMMANDS = {
     "cleanup": cmd_cleanup,
-    "dashboard": cmd_dashboard,
     "edit": cmd_edit,
     "info": cmd_info,
     "init": cmd_init,
     "list": cmd_list,
     "resume": cmd_resume,
     "start": cmd_start,
+    "supervise": cmd_supervise,
 }
 
 
@@ -36,13 +37,13 @@ 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("  dashboard view + approve/modify/reject pending supervise proposals (PRD 0013)\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")
     sys.stderr.write("  list      list available agents or active containers\n")
     sys.stderr.write("  resume    re-launch a bottle by its identity (continues state from PRD 0016)\n")
-    sys.stderr.write("  start     boot a container for a named agent and attach an interactive session\n\n")
+    sys.stderr.write("  start     boot a container for a named agent and attach an interactive session\n")
+    sys.stderr.write("  supervise view + approve/modify/reject pending supervise proposals (PRD 0013)\n\n")
     sys.stderr.write(f"Run '{PROG} <command> --help' for command-specific usage.\n")
 
 
@@ -63,6 +64,11 @@ def main(argv: list[str] | None = None) -> int:
         die(f"unknown command: {command}")
     try:
         return handler(rest) or 0
+    except ManifestError as e:
+        # Manifest/config problems surface as a catchable exception;
+        # print the reason and exit non-zero (same UX die() used to give).
+        error(str(e))
+        return 1
     except Die as e:
         return e.code if isinstance(e.code, int) else 1
     except KeyboardInterrupt:

bot_bottle/cli/info.py

@@ -31,6 +31,9 @@ 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)
+    if identity:
+        info(f"  git identity  : {identity}")
     if bottle.git:
         for e in bottle.git:
             info(

bot_bottle/cli/resume.py

@@ -52,8 +52,10 @@ def cmd_resume(argv: list[str]) -> int:
         user_cwd=metadata.cwd or USER_CWD,
         identity=metadata.identity,
     )
+    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

@@ -2,10 +2,8 @@
 interactive claude-code session. The container is torn down when the
 session ends.
 
-The launch core is shared with `cli.py resume <identity>` and (PRD
-0020 chunk 1+) the dashboard's in-process start flow: see the
-public helpers `prepare_with_preflight`, `attach_agent`, and the
-private orchestrator `_launch_bottle`.
+The launch core is shared with `cli.py resume <identity>` through
+the private orchestrator `_launch_bottle`.
 """
 
 from __future__ import annotations
@@ -71,7 +69,7 @@ def cmd_start(argv: list[str]) -> int:
     )
 
 
-# --- Public helpers shared with the dashboard (PRD 0020) -----------------
+# --- Launch helpers ------------------------------------------------------
 
 
 def prepare_with_preflight(
@@ -84,14 +82,11 @@ def prepare_with_preflight(
     backend_name: str | None = None,
 ) -> tuple[DockerBottlePlan | None, str]:
     """Run `backend.prepare`, render the preflight summary via the
-    injected callable, prompt y/N via the injected callable. The CLI
-    binds these to stderr/stdin; the dashboard binds them to a
-    curses modal.
+    injected callable, prompt y/N via the injected callable.
 
     `backend_name` selects which backend prepares the plan
-    (`None` → `$BOT_BOTTLE_BACKEND` → `docker`). Dashboard
-    passes the value from its new-agent backend-picker modal; the
-    CLI passes whatever `--backend` resolved to.
+    (`None` → `$BOT_BOTTLE_BACKEND` → `docker`). The CLI passes
+    whatever `--backend` resolved to.
 
     Returns `(plan, identity)`. `plan` is None on dry-run or
     operator-N, but `identity` is set as soon as `backend.prepare`
@@ -122,16 +117,10 @@ def attach_agent(
     agent process's exit code.
 
     `resume=True` adds `--continue` so claude picks up its most
-    recent session non-interactively (no session-picker prompt) —
-    the right shape for the dashboard's Enter re-attach (PRD 0020
-    chunk 3), where a bottle typically has exactly one session.
-    First-attach paths (`./cli.py start`, the dashboard's new-agent
-    flow) leave it False.
+    recent session non-interactively (no session-picker prompt).
+    First-attach paths (`./cli.py start`) leave it False.
 
-    Used as the inner step of `./cli.py start` (one-shot) and by the
-    dashboard, which calls it from inside a `curses.endwin → … →
-    stdscr.refresh()` handoff so the curses surface gets out of the
-    terminal's way while the agent has it."""
+    Used as the inner step of `./cli.py start`."""
     runtime = runtime_for(agent_provider_template)
     info(
         f"attaching interactive {agent_provider_template} session "
@@ -148,8 +137,7 @@ def attach_agent(
 def capture_claude_session_state(identity: str, exit_code: int) -> None:
     """Inside the launch context, while the container is still
     alive: snapshot the transcript and mark for preservation if
-    claude crashed. Public for the dashboard's death-handling path
-    (PRD 0020 open question 3)."""
+    claude crashed."""
     # FIXME: this captures Claude-specific session state. A follow-up
     # spike should explore freezing provider-neutral container state
     # instead of relying on each agent's transcript layout.
@@ -162,9 +150,7 @@ def capture_claude_session_state(identity: str, exit_code: int) -> None:
 
 def settle_state(identity: str) -> None:
     """Post-teardown housekeeping: print the resume hint if the
-    state was preserved, otherwise reap the per-bottle state dir.
-    Public so the dashboard's explicit-stop path calls the same
-    settlement the CLI uses on context exit."""
+    state was preserved, otherwise reap the per-bottle state dir."""
     if not identity:
         return
     if is_preserved(identity):

bot_bottle/cli/supervise.py

@@ -0,0 +1,577 @@
+"""supervise: list pending supervise proposals across all bottles and
+act on them (approve / modify / reject).
+
+Curses-based TUI; modify-then-approve shells out to $EDITOR. The
+approval handlers wire to the per-tool remediation engines:
+PRD 0014 (egress, retargeted from cred-proxy in PRD 0017
+chunk 3) writes routes.yaml + SIGHUPs egress; PRD 0015
+(pipelock) writes the allowlist + restarts pipelock; PRD 0016
+(capability) rebuilds the bottle Dockerfile.
+"""
+
+from __future__ import annotations
+
+import argparse
+import curses
+import os
+import subprocess
+import sys
+import tempfile
+import traceback
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+
+from .. import supervise as _supervise
+from ..backend.docker.bottle_state import read_metadata
+from ..backend.docker.capability_apply import (
+    CapabilityApplyError,
+    apply_capability_change,
+)
+from ..backend.docker.egress_apply import EgressApplyError, add_route
+from ..backend.docker.pipelock_apply import (
+    PipelockApplyError,
+    apply_allowlist_change,
+    fetch_current_allowlist,
+    parse_allowlist_content,
+    render_allowlist_content,
+)
+from ..log import Die, error, info
+from ..supervise import (
+    COMPONENT_FOR_TOOL,
+    AuditEntry,
+    Proposal,
+    Response,
+    STATUS_APPROVED,
+    STATUS_MODIFIED,
+    STATUS_REJECTED,
+    TOOL_CAPABILITY_BLOCK,
+    TOOL_EGRESS_BLOCK,
+    TOOL_PIPELOCK_BLOCK,
+    archive_proposal,
+    list_pending_proposals,
+    render_diff,
+    write_audit_entry,
+    write_response,
+)
+from ._common import PROG
+
+
+_REFRESH_INTERVAL_MS = 1000
+
+
+@dataclass(frozen=True)
+class QueuedProposal:
+    """A pending proposal plus the queue dir it was found in."""
+
+    proposal: Proposal
+    queue_dir: Path
+
+
+# Errors any remediation engine may raise. Caught by the TUI key
+# handlers and surfaced in the status line so a failed apply keeps
+# the proposal pending rather than crashing curses.
+ApplyError = (EgressApplyError, PipelockApplyError, CapabilityApplyError)
+
+
+def discover_pending() -> list[QueuedProposal]:
+    """Walk ~/.bot-bottle/queue/* and collect pending proposals."""
+    queue_root = _supervise.bot_bottle_root() / "queue"
+    if not queue_root.is_dir():
+        return []
+    out: list[QueuedProposal] = []
+    for slug_dir in sorted(queue_root.iterdir()):
+        if not slug_dir.is_dir():
+            continue
+        for proposal in list_pending_proposals(slug_dir):
+            out.append(QueuedProposal(proposal=proposal, queue_dir=slug_dir))
+    out.sort(key=lambda q: q.proposal.arrival_timestamp)
+    return out
+
+
+def _approval_status(qp: QueuedProposal, verb: str) -> str:
+    """Status-line text after a successful approval."""
+    base = f"{verb} {qp.proposal.tool} for [{qp.proposal.bottle_slug}]"
+    if qp.proposal.tool == TOOL_CAPABILITY_BLOCK:
+        return f"{base}; resume: ./cli.py resume {qp.proposal.bottle_slug}"
+    return base
+
+
+def _detail_lines(
+    qp: QueuedProposal,
+    *,
+    green_attr: int = 0,
+) -> list[tuple[str, int]]:
+    """Return the detail-view body as (text, curses-attr) tuples."""
+    p = qp.proposal
+    out: list[tuple[str, int]] = [
+        (f"bottle: {p.bottle_slug}", 0),
+        (f"tool: {p.tool}", 0),
+        (f"id: {p.id}", 0),
+        (f"arrived: {p.arrival_timestamp}", 0),
+        (f"queue: {qp.queue_dir}", 0),
+        ("", 0),
+        ("justification:", 0),
+    ]
+    out.extend(("  " + line, 0) for line in p.justification.splitlines() or [""])
+    out.extend([
+        ("", 0),
+        (_proposed_payload_label(p.tool) + ":", 0),
+    ])
+    out.extend((line, 0) for line in p.proposed_file.splitlines() or [""])
+    if p.tool == TOOL_PIPELOCK_BLOCK:
+        host = _failed_url_host(p.proposed_file)
+        if host:
+            out.append(("", 0))
+            out.append((host, green_attr))
+    return out
+
+
+def _failed_url_host(url: str) -> str:
+    """Best-effort hostname extraction from a pipelock-block proposal."""
+    import urllib.parse
+
+    try:
+        return urllib.parse.urlsplit(url.strip()).hostname or ""
+    except ValueError:
+        return ""
+
+
+def _proposed_payload_label(tool: str) -> str:
+    if tool == TOOL_PIPELOCK_BLOCK:
+        return "failed URL"
+    return "proposed file"
+
+
+def _suffix_for_tool(tool: str) -> str:
+    if tool == TOOL_CAPABILITY_BLOCK:
+        return ".dockerfile"
+    return ".txt"
+
+
+# --- Operator actions ------------------------------------------------------
+
+
+def approve(
+    qp: QueuedProposal,
+    *,
+    notes: str = "",
+    final_file: str | None = None,
+) -> None:
+    """Apply the proposal, write the waiting response, and audit it."""
+    status = STATUS_MODIFIED if final_file is not None else STATUS_APPROVED
+    file_to_apply = final_file if final_file is not None else qp.proposal.proposed_file
+
+    diff_before, diff_after = "", ""
+    if qp.proposal.tool == TOOL_EGRESS_BLOCK:
+        diff_before, diff_after = add_route(
+            qp.proposal.bottle_slug, file_to_apply,
+        )
+    elif qp.proposal.tool == TOOL_PIPELOCK_BLOCK:
+        diff_before, diff_after = _apply_pipelock_url(
+            qp.proposal.bottle_slug, file_to_apply,
+        )
+    elif 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,
+        )
+
+    response = Response(
+        proposal_id=qp.proposal.id,
+        status=status,
+        notes=notes,
+        final_file=final_file,
+    )
+    write_response(qp.queue_dir, response)
+    _write_audit(
+        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."""
+    response = Response(
+        proposal_id=qp.proposal.id,
+        status=STATUS_REJECTED,
+        notes=reason,
+        final_file=None,
+    )
+    write_response(qp.queue_dir, response)
+    _write_audit(qp, action=STATUS_REJECTED, notes=reason, diff_before="", diff_after="")
+
+
+def _apply_pipelock_url(slug: str, failed_url: str) -> tuple[str, str]:
+    """Merge a pipelock-block failed URL's host into the allowlist."""
+    import urllib.parse
+
+    parsed = urllib.parse.urlsplit(failed_url.strip())
+    host = parsed.hostname or ""
+    if not host:
+        raise PipelockApplyError(
+            f"proposed failed_url has no extractable host: {failed_url!r}"
+        )
+    current = fetch_current_allowlist(slug)
+    hosts = parse_allowlist_content(current)
+    if host not in hosts:
+        hosts.append(host)
+    return apply_allowlist_change(slug, render_allowlist_content(hosts))
+
+
+def _write_audit(
+    qp: QueuedProposal,
+    *,
+    action: str,
+    notes: str,
+    diff_before: str,
+    diff_after: str,
+) -> None:
+    """Audit log for egress / pipelock tools."""
+    component = COMPONENT_FOR_TOOL.get(qp.proposal.tool)
+    if component is None:
+        return
+    write_audit_entry(AuditEntry(
+        timestamp=datetime.now(timezone.utc).isoformat(),
+        bottle_slug=qp.proposal.bottle_slug,
+        component=component,
+        operator_action=action,
+        operator_notes=notes,
+        justification=qp.proposal.justification,
+        diff=render_diff(diff_before, diff_after, label=component),
+    ))
+
+
+# --- $EDITOR integration --------------------------------------------------
+
+
+def edit_in_editor(content: str, *, suffix: str = ".tmp") -> str | None:
+    """Open `content` in $EDITOR and return edited content, if changed."""
+    editor = os.environ.get("EDITOR", "vim")
+    with tempfile.NamedTemporaryFile(
+        mode="w", suffix=suffix, delete=False, prefix="supervise-modify.",
+    ) as f:
+        f.write(content)
+        path = f.name
+    try:
+        subprocess.run([editor, path], check=False)
+        with open(path) as f:
+            edited = f.read()
+        return edited if edited != content else None
+    finally:
+        try:
+            os.unlink(path)
+        except OSError:
+            pass
+
+
+# --- TUI -------------------------------------------------------------------
+
+
+def cmd_supervise(argv: list[str]) -> int:
+    parser = argparse.ArgumentParser(prog=f"{PROG} supervise", add_help=True)
+    parser.add_argument(
+        "--once", action="store_true",
+        help="list pending proposals once and exit (no TUI)",
+    )
+    args = parser.parse_args(argv)
+
+    if args.once:
+        return _list_once()
+    try:
+        curses.wrapper(_main_loop)
+    except KeyboardInterrupt:
+        return 130
+    except Die as e:
+        if e.message:
+            error(e.message)
+        else:
+            error("supervise exited on a fatal error (no detail captured).")
+        return e.code if isinstance(e.code, int) else 1
+    except Exception as e:
+        log_path = _write_crash_log(e)
+        error(f"supervise crashed: {type(e).__name__}: {e}")
+        error(f"full traceback written to {log_path}")
+        return 1
+    return 0
+
+
+def _write_crash_log(exc: BaseException) -> Path:
+    """Persist `exc`'s traceback to a stable file under ~/.bot-bottle/."""
+    stamp = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
+    body = "".join(
+        traceback.format_exception(type(exc), exc, exc.__traceback__)
+    )
+    entry = f"=== supervise crash {stamp} ===\n{body}\n"
+    try:
+        log_dir = _supervise.bot_bottle_root() / "logs"
+        log_dir.mkdir(parents=True, exist_ok=True)
+        path = log_dir / "supervise-crash.log"
+        with path.open("a", encoding="utf-8") as fh:
+            fh.write(entry)
+        return path
+    except OSError:
+        fd, tmp = tempfile.mkstemp(
+            prefix="bot-bottle-supervise-crash-", suffix=".log",
+        )
+        with os.fdopen(fd, "w", encoding="utf-8") as fh:
+            fh.write(entry)
+        return Path(tmp)
+
+
+def _list_once() -> int:
+    pending = discover_pending()
+    if not pending:
+        info("no pending proposals")
+        return 0
+    for qp in pending:
+        sys.stdout.write(
+            f"{qp.proposal.arrival_timestamp}  "
+            f"[{qp.proposal.bottle_slug}]  "
+            f"{qp.proposal.tool}  "
+            f"{qp.proposal.id}\n"
+        )
+        sys.stdout.write(f"    {qp.proposal.justification}\n")
+    return 0
+
+
+def _try_init_green() -> int:
+    """Initialise a green color pair and return its attr, or 0."""
+    try:
+        curses.start_color()
+        curses.use_default_colors()
+        curses.init_pair(1, curses.COLOR_GREEN, -1)
+        return curses.color_pair(1)
+    except curses.error:
+        return 0
+
+
+def _main_loop(stdscr: "curses._CursesWindow") -> None:
+    curses.curs_set(0)
+    stdscr.timeout(_REFRESH_INTERVAL_MS)
+    green_attr = _try_init_green()
+    selected = 0
+    status_line = ""
+    seen_ids: set[str] = set()
+
+    while True:
+        pending = discover_pending()
+        if selected >= len(pending):
+            selected = max(0, len(pending) - 1)
+
+        live_ids = {qp.proposal.id for qp in pending}
+        newly_arrived = live_ids - seen_ids
+        if seen_ids and newly_arrived:
+            try:
+                curses.beep()
+            except curses.error:
+                pass
+            for i, qp in enumerate(pending):
+                if qp.proposal.id in newly_arrived:
+                    selected = i
+                    break
+        seen_ids = live_ids
+
+        _render(
+            stdscr, pending, selected, status_line,
+            green_attr=green_attr,
+        )
+
+        try:
+            key = stdscr.getch()
+        except KeyboardInterrupt:
+            return
+
+        if key == -1:
+            continue
+
+        status_line = ""
+
+        if key in (ord("q"), 27):
+            return
+
+        if not pending:
+            continue
+        qp = pending[selected]
+
+        if key in (curses.KEY_DOWN, ord("j")):
+            selected = min(selected + 1, len(pending) - 1)
+        elif key in (curses.KEY_UP, ord("k")):
+            selected = max(selected - 1, 0)
+        elif key in (curses.KEY_ENTER, 10, 13):
+            _detail_view(stdscr, qp, green_attr=green_attr)
+        elif key == ord("a"):
+            try:
+                approve(qp)
+                status_line = _approval_status(qp, "approved")
+            except ApplyError as e:
+                status_line = f"apply failed: {e}"
+        elif key == ord("m"):
+            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")
+                except ApplyError as e:
+                    status_line = f"apply failed: {e}"
+        elif key == ord("r"):
+            reason = _prompt(stdscr, "reject reason: ")
+            if reason:
+                reject(qp, reason=reason)
+                status_line = f"rejected {qp.proposal.tool} for [{qp.proposal.bottle_slug}]"
+            else:
+                status_line = "reject aborted (empty reason)"
+
+
+def _render(
+    stdscr: "curses._CursesWindow",
+    pending: list[QueuedProposal],
+    selected: int,
+    status_line: str,
+    *,
+    green_attr: int = 0,
+) -> None:
+    stdscr.erase()
+    h, w = stdscr.getmaxyx()
+    header = f"bot-bottle supervise  ({len(pending)} pending)"
+    stdscr.addnstr(0, 0, header, w - 1, curses.A_BOLD)
+    stdscr.hline(1, 0, curses.ACS_HLINE, w)
+
+    row = 2
+    if not pending:
+        stdscr.addnstr(
+            row, 2,
+            "no pending proposals; agents will queue here when they call a "
+            "supervise tool",
+            w - 4,
+        )
+    else:
+        for i, qp in enumerate(pending):
+            if row >= h - 3:
+                break
+            p = qp.proposal
+            ts_short = (
+                p.arrival_timestamp.split("T", 1)[1][:8]
+                if "T" in p.arrival_timestamp else p.arrival_timestamp
+            )
+            cursor = "> " if i == selected else "  "
+            line = (
+                f"{cursor}{ts_short}  "
+                f"[{p.bottle_slug}]  {p.tool:<18}  {p.id[:8]}  "
+                f"{_proposed_payload_label(p.tool)}"
+            )
+            attr = curses.A_REVERSE if i == selected else curses.A_NORMAL
+            stdscr.addnstr(row, 0, line, w - 1, attr)
+            row += 1
+            if row >= h - 3:
+                break
+            if p.justification:
+                stdscr.addnstr(row, 4, p.justification[: max(0, w - 5)], w - 5)
+                row += 1
+
+    footer = "[j/k] move  [Enter] view  [a] approve  [m] modify  [r] reject  [q] quit"
+    stdscr.hline(h - 2, 0, curses.ACS_HLINE, w)
+    stdscr.addnstr(h - 1, 0, footer, w - 1, curses.A_DIM)
+    if status_line:
+        stdscr.addnstr(h - 3, 0, status_line, w - 1, curses.A_BOLD)
+    stdscr.refresh()
+
+
+def _detail_view(
+    stdscr: "curses._CursesWindow",
+    qp: QueuedProposal,
+    *,
+    green_attr: int = 0,
+) -> None:
+    """Render the full proposal. Scrollable. Press q to return."""
+    lines = _detail_lines(qp, green_attr=green_attr)
+    offset = 0
+    while True:
+        stdscr.erase()
+        h, w = stdscr.getmaxyx()
+        for i, (text, attr) in enumerate(lines[offset:offset + h - 1]):
+            stdscr.addnstr(i, 0, text, w - 1, attr)
+        stdscr.addnstr(
+            h - 1, 0,
+            "[j/k] scroll  [g/G] top/bottom  [a] approve  [m] modify  [r] reject  [q] back",
+            w - 1, curses.A_DIM,
+        )
+        stdscr.refresh()
+        key = stdscr.getch()
+        if key in (ord("q"), 27):
+            return
+        if key in (curses.KEY_DOWN, ord("j")):
+            offset = min(offset + 1, max(0, len(lines) - 1))
+        elif key in (curses.KEY_UP, ord("k")):
+            offset = max(offset - 1, 0)
+        elif key == ord("g"):
+            offset = 0
+        elif key == ord("G"):
+            offset = max(0, len(lines) - 1)
+        elif key == ord("a"):
+            try:
+                approve(qp)
+            except ApplyError:
+                pass
+            return
+        elif key == ord("m"):
+            edited = _modify(stdscr, qp)
+            if edited is not None:
+                try:
+                    approve(qp, final_file=edited, notes="operator modified before approving")
+                except ApplyError:
+                    pass
+            return
+        elif key == ord("r"):
+            reason = _prompt(stdscr, "reject reason: ")
+            if reason:
+                reject(qp, reason=reason)
+            return
+
+
+def _modify(stdscr: "curses._CursesWindow", qp: QueuedProposal) -> str | None:
+    """Suspend curses, open $EDITOR on the proposed file, return edited content."""
+    suffix = _suffix_for_tool(qp.proposal.tool)
+    curses.endwin()
+    try:
+        edited = edit_in_editor(qp.proposal.proposed_file, suffix=suffix)
+    finally:
+        stdscr.refresh()
+    return edited
+
+
+def _prompt(stdscr: "curses._CursesWindow", label: str) -> str:
+    """One-line input at the bottom of the screen."""
+    curses.curs_set(1)
+    h, _ = stdscr.getmaxyx()
+    stdscr.move(h - 2, 0)
+    stdscr.clrtoeol()
+    stdscr.addstr(h - 2, 0, label)
+    stdscr.refresh()
+    curses.echo()
+    try:
+        raw = stdscr.getstr(h - 2, len(label), 200)
+    finally:
+        curses.noecho()
+        curses.curs_set(0)
+    return raw.decode("utf-8", errors="replace").strip()
+
+
+__all__ = [
+    "QueuedProposal",
+    "approve",
+    "cmd_supervise",
+    "discover_pending",
+    "edit_in_editor",
+    "reject",
+]

bot_bottle/codex_auth.py

@@ -0,0 +1,325 @@
+"""Host Codex auth helpers.
+
+Reads the host's Codex ChatGPT/device-login auth state and returns only
+the short-lived access token needed by egress. This module deliberately
+does not expose refresh tokens or raw auth payloads.
+"""
+
+from __future__ import annotations
+
+import base64
+import json
+import os
+from copy import deepcopy
+from datetime import datetime, timezone
+from pathlib import Path
+
+from .log import die
+from .util import expand_tilde
+
+
+def codex_auth_path(host_env: dict[str, str] | None = None) -> Path:
+    env = os.environ if host_env is None else host_env
+    home = env.get("CODEX_HOME")
+    if home:
+        return Path(expand_tilde(home)) / "auth.json"
+    return Path.home() / ".codex" / "auth.json"
+
+
+def codex_host_access_token(
+    host_env: dict[str, str] | None = None,
+    *,
+    now: datetime | None = None,
+) -> str:
+    path = codex_auth_path(host_env)
+    if not path.is_file():
+        die(
+            f"codex host credentials: auth file missing at {path}. "
+            "Run `codex login --device-auth` on the host or disable "
+            "agent_provider.forward_host_credentials."
+        )
+    raw = _read_auth_object(path)
+
+    auth_mode = raw.get("auth_mode")
+    if not isinstance(auth_mode, str) or auth_mode == "api_key":
+        die(
+            "codex host credentials: host Codex auth is not user/device "
+            "auth. Run `codex login --device-auth` on the host."
+        )
+
+    tokens = raw.get("tokens")
+    if not isinstance(tokens, dict):
+        die(f"codex host credentials: {path} is missing tokens")
+    access = tokens.get("access_token")
+    if not isinstance(access, str) or not access:
+        die(
+            f"codex host credentials: {path} is missing tokens.access_token. "
+            "Run `codex login --device-auth` on the host."
+        )
+
+    exp = _jwt_exp(access)
+    if exp is None:
+        die("codex host credentials: tokens.access_token is not a JWT with exp")
+    check_now = now or datetime.now(timezone.utc)
+    if exp <= check_now:
+        die(
+            "codex host credentials: host Codex access token is expired. "
+            "Run `codex login --device-auth` on the host and restart the bottle."
+        )
+    return access
+
+
+def codex_dummy_auth_json(
+    host_env: dict[str, str] | None = None,
+    *,
+    now: datetime | None = None,
+) -> str:
+    """Return a non-secret `auth.json` that keeps Codex in the host's
+    auth branch while egress owns the real bearer token.
+
+    The dummy access/id tokens carry the *host* token's real `exp` so
+    Codex's proactive refresh lifecycle (it refreshes when its local
+    access token is at/past expiry) tracks the real token instead of
+    firing after an artificial TTL. Codex cannot refresh inside the
+    bottle — the refresh token is a placeholder and the OpenAI token
+    endpoint is off-route — so a shorter dummy exp would drop Codex to
+    the sign-in screen the moment it lapsed, even while egress still
+    holds a valid bearer."""
+    path = codex_auth_path(host_env)
+    access = codex_host_access_token(host_env, now=now)
+    raw = _read_auth_object(path)
+    host_exp = _jwt_exp(access)
+    exp_ts = int(host_exp.timestamp()) if host_exp is not None else None
+    dummy = _redact_codex_auth(deepcopy(raw), now=now, exp_ts=exp_ts)
+    return json.dumps(dummy, indent=2, sort_keys=True) + "\n"
+
+
+def write_codex_dummy_auth_file(
+    path: Path,
+    host_env: dict[str, str] | None = None,
+    *,
+    now: datetime | None = None,
+) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(codex_dummy_auth_json(host_env, now=now))
+    path.chmod(0o600)
+
+
+def _read_auth_object(path: Path) -> dict:
+    try:
+        raw = json.loads(path.read_text())
+    except (OSError, json.JSONDecodeError) as e:
+        die(f"codex host credentials: could not read valid JSON at {path}: {e}")
+    if not isinstance(raw, dict):
+        die(f"codex host credentials: {path} must contain a JSON object")
+    return raw
+
+
+def _dummy_exp(now: datetime | None, exp_ts: int | None) -> int:
+    if exp_ts is not None:
+        return exp_ts
+    check_now = now or datetime.now(timezone.utc)
+    return int(check_now.timestamp()) + 3600
+
+
+def _dummy_timestamp(now: datetime | None = None) -> str:
+    check_now = now or datetime.now(timezone.utc)
+    if check_now.tzinfo is None:
+        check_now = check_now.replace(tzinfo=timezone.utc)
+    check_now = check_now.astimezone(timezone.utc)
+    return check_now.isoformat(timespec="milliseconds").replace("+00:00", "Z")
+
+
+def _dummy_jwt(now: datetime | None = None, *, exp_ts: int | None = None) -> str:
+    return _encode_dummy_jwt({
+        "exp": _dummy_exp(now, exp_ts),
+        "sub": "bot-bottle-placeholder",
+    })
+
+
+def _dummy_jwt_from_host(
+    value: object, *, now: datetime | None = None, exp_ts: int | None = None,
+) -> str:
+    if not isinstance(value, str):
+        return _dummy_jwt(now, exp_ts=exp_ts)
+    parts = value.split(".")
+    if len(parts) < 2:
+        return _dummy_jwt(now, exp_ts=exp_ts)
+    try:
+        payload = json.loads(_b64url_decode(parts[1]))
+    except (ValueError, json.JSONDecodeError):
+        return _dummy_jwt(now, exp_ts=exp_ts)
+    if not isinstance(payload, dict):
+        return _dummy_jwt(now, exp_ts=exp_ts)
+    return _encode_dummy_jwt(_redact_jwt_payload(payload, now=now, exp_ts=exp_ts))
+
+
+def _encode_dummy_jwt(payload: dict) -> str:
+    def enc(obj: dict) -> str:
+        raw = json.dumps(obj, separators=(",", ":")).encode()
+        return base64.urlsafe_b64encode(raw).decode().rstrip("=")
+
+    return f"{enc({'alg': 'none', 'typ': 'JWT'})}.{enc(payload)}.placeholder"
+
+
+def _redact_jwt_payload(
+    payload: dict,
+    *,
+    now: datetime | None = None,
+    exp_ts: int | None = None,
+) -> dict:
+    out = _redact_claims(payload)
+    if not isinstance(out, dict):
+        out = {}
+    out["exp"] = _dummy_exp(now, exp_ts)
+    out.setdefault("sub", "bot-bottle-placeholder")
+    return out
+
+
+def _redact_claims(value: object) -> object:
+    if isinstance(value, dict):
+        out: dict[str, object] = {}
+        for key, inner in value.items():
+            lower = key.lower()
+            if key == "https://api.openai.com/profile":
+                out[key] = _redact_profile_claim(inner)
+            elif key == "https://api.openai.com/auth":
+                out[key] = _redact_auth_claim(inner)
+            elif lower == "email":
+                out[key] = "bot-bottle@example.invalid"
+            elif lower == "email_verified":
+                out[key] = True
+            elif lower in {"exp", "iat", "nbf", "auth_time", "pwd_auth_time"}:
+                out[key] = inner if isinstance(inner, (int, float)) else 0
+            elif lower in {"aud", "scp", "amr"}:
+                out[key] = inner if isinstance(inner, list) else []
+            elif isinstance(inner, bool):
+                out[key] = inner
+            elif isinstance(inner, dict):
+                out[key] = {}
+            elif isinstance(inner, list):
+                out[key] = []
+            else:
+                out[key] = "bot-bottle-placeholder"
+        return out
+    if isinstance(value, list):
+        return []
+    return "bot-bottle-placeholder"
+
+
+def _redact_profile_claim(value: object) -> dict:
+    profile = value if isinstance(value, dict) else {}
+    return {
+        "email": "bot-bottle@example.invalid",
+        "email_verified": bool(profile.get("email_verified", True)),
+    }
+
+
+def _redact_auth_claim(value: object) -> dict:
+    auth = value if isinstance(value, dict) else {}
+    out: dict[str, object] = {}
+    for key, inner in auth.items():
+        lower = key.lower()
+        if lower == "chatgpt_plan_type" and isinstance(inner, str) and inner:
+            out[key] = inner
+        elif lower == "chatgpt_account_id" and isinstance(inner, str) and inner:
+            # Current Codex uses the selected account id when building
+            # ChatGPT requests. Keep that non-secret identifier aligned
+            # with the host while egress owns the real bearer token.
+            out[key] = inner
+        elif lower == "localhost" and isinstance(inner, bool):
+            out[key] = inner
+        elif isinstance(inner, bool):
+            out[key] = inner
+        elif isinstance(inner, list):
+            out[key] = []
+        elif isinstance(inner, dict):
+            out[key] = {}
+        else:
+            out[key] = "bot-bottle-placeholder"
+    out.setdefault("chatgpt_plan_type", "unknown")
+    out.setdefault("user_id", "bot-bottle-placeholder")
+    out.setdefault("chatgpt_user_id", "bot-bottle-placeholder")
+    out.setdefault("chatgpt_account_id", "bot-bottle-placeholder")
+    return out
+
+
+def _redact_codex_auth(
+    value: object, *, now: datetime | None = None, exp_ts: int | None = None,
+) -> object:
+    auth = value if isinstance(value, dict) else {}
+    out: dict[str, object] = {}
+    for key, inner in auth.items():
+        lower = key.lower()
+        if lower == "auth_mode" and isinstance(inner, str) and inner:
+            out[key] = inner
+        elif lower == "openai_api_key":
+            out[key] = None
+        elif lower == "last_refresh":
+            # Codex parses this as a timestamp on startup. Keep the
+            # schema valid without copying host-side session metadata.
+            out[key] = _dummy_timestamp(now)
+        elif lower == "tokens":
+            out[key] = _redact_token_block(inner, now=now, exp_ts=exp_ts)
+        else:
+            out[key] = _redact_unknown_auth_value(inner)
+    return out
+
+
+def _redact_token_block(
+    value: object, *, now: datetime | None = None, exp_ts: int | None = None,
+) -> dict[str, object]:
+    tokens = value if isinstance(value, dict) else {}
+    out: dict[str, object] = {}
+    for key, inner in tokens.items():
+        lower = key.lower()
+        if lower in {"access_token", "id_token"}:
+            out[key] = _dummy_jwt_from_host(inner, now=now, exp_ts=exp_ts)
+        elif lower == "account_id" and isinstance(inner, str) and inner:
+            # Current Codex uses this non-secret selected account id
+            # while egress owns the real bearer token.
+            out[key] = inner
+        else:
+            out[key] = _redact_unknown_auth_value(inner)
+    return out
+
+
+def _redact_unknown_auth_value(value: object) -> object:
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, dict):
+        return {}
+    if isinstance(value, list):
+        return []
+    if value is None:
+        return None
+    return "bot-bottle-placeholder"
+
+
+def _jwt_exp(token: str) -> datetime | None:
+    parts = token.split(".")
+    if len(parts) < 2:
+        return None
+    try:
+        payload = json.loads(_b64url_decode(parts[1]))
+    except (ValueError, json.JSONDecodeError):
+        return None
+    if not isinstance(payload, dict):
+        return None
+    exp = payload.get("exp")
+    if not isinstance(exp, (int, float)):
+        return None
+    return datetime.fromtimestamp(exp, timezone.utc)
+
+
+def _b64url_decode(value: str) -> str:
+    padded = value + ("=" * (-len(value) % 4))
+    return base64.urlsafe_b64decode(padded.encode("ascii")).decode("utf-8")
+
+
+__all__ = [
+    "codex_auth_path",
+    "codex_dummy_auth_json",
+    "codex_host_access_token",
+    "write_codex_dummy_auth_file",
+]

bot_bottle/contrib/claude/agent_provider.py

@@ -0,0 +1,226 @@
+"""Claude agent provider plugin (PRD 0050, contrib).
+
+The Claude-specific behavior previously inlined under
+`agent_provider.agent_provision_plan` (claude.json trust marker,
+api.anthropic.com egress route, OAuth-token placeholder), plus
+the `claude mcp add` invocation that registers the supervise
+sidecar in claude-code's user config (PRD 0013)."""
+
+from __future__ import annotations
+
+import json
+import os
+import shlex
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+from ...agent_provider import (
+    AgentProvider,
+    AgentProviderRuntime,
+    AgentProvisionFile,
+    AgentProvisionPlan,
+)
+from ...egress import EgressRoute
+from ...log import die, info, warn
+
+
+if TYPE_CHECKING:
+    from ...backend import Bottle, BottlePlan
+
+
+_REPO_ROOT = Path(__file__).resolve().parents[3]
+
+_SUPERVISE_MCP_NAME = "supervise"
+
+
+def _skills_dir(guest_home: str) -> str:
+    return f"{guest_home}/.claude/skills"
+
+
+def _prompt_path(guest_home: str) -> str:
+    return f"{guest_home}/.bot-bottle-prompt.txt"
+
+_RUNTIME = AgentProviderRuntime(
+    template="claude",
+    command="claude",
+    image="bot-bottle-claude:latest",
+    dockerfile=str(_REPO_ROOT / "Dockerfile.claude"),
+    prompt_mode="append_file",
+    bypass_args=("--dangerously-skip-permissions",),
+    resume_args=("--continue",),
+    remote_control_args=("--remote-control",),
+)
+
+
+class ClaudeAgentProvider(AgentProvider):
+    @property
+    def runtime(self) -> AgentProviderRuntime:
+        return _RUNTIME
+
+    def provision_plan(
+        self,
+        *,
+        dockerfile: str,
+        state_dir: Path,
+        guest_home: str,
+        guest_env: dict[str, str] | None = None,
+        auth_token: str = "",
+        forward_host_credentials: bool = False,
+        host_env: dict[str, str] | None = None,
+        trusted_project_path: str = "",
+    ) -> AgentProvisionPlan:
+        del forward_host_credentials, host_env  # Codex-only knobs
+        resolved_guest_env = dict(guest_env or {})
+        trusted_path = trusted_project_path or guest_home
+
+        env_vars: dict[str, str] = {
+            "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
+            "DISABLE_ERROR_REPORTING": "1",
+        }
+        claude_config = state_dir / "claude.json"
+        claude_projects = {guest_home: {"hasTrustDialogAccepted": True}}
+        claude_projects[trusted_path] = {"hasTrustDialogAccepted": True}
+        claude_config.write_text(json.dumps({
+            "hasCompletedOnboarding": True,
+            "theme": "dark",
+            "bypassPermissionsModeAccepted": True,
+            "projects": claude_projects,
+        }, indent=2) + "\n")
+        claude_config.chmod(0o600)
+        files = (
+            AgentProvisionFile(claude_config, f"{guest_home}/.claude.json"),
+        )
+        egress_routes = (EgressRoute(
+            host="api.anthropic.com",
+            auth_scheme="Bearer" if auth_token else "",
+            token_ref=auth_token,
+            tls_passthrough=True,
+        ),)
+        hidden_env_names: frozenset[str] = frozenset()
+        if auth_token:
+            env_vars["CLAUDE_CODE_OAUTH_TOKEN"] = "egress-placeholder"
+            hidden_env_names = frozenset({"CLAUDE_CODE_OAUTH_TOKEN"})
+
+        return AgentProvisionPlan(
+            template=_RUNTIME.template,
+            command=_RUNTIME.command,
+            prompt_mode=_RUNTIME.prompt_mode,
+            image=_RUNTIME.image,
+            dockerfile=dockerfile,
+            env_vars=env_vars,
+            guest_env=resolved_guest_env,
+            files=files,
+            egress_routes=egress_routes,
+            hidden_env_names=hidden_env_names,
+        )
+
+    def provision_skills(self, plan: "BottlePlan", bottle: "Bottle") -> None:
+        """Copy each named skill tree from `~/.claude/skills/<name>/`
+        on the host into the guest's claude-code skills dir. No-op
+        when the agent has no skills."""
+        from ...backend.util import host_skill_dir
+
+        agent = plan.spec.manifest.agents[plan.spec.agent_name]
+        if not agent.skills:
+            return
+        skills_dir = _skills_dir(plan.guest_home)
+        bottle.exec(f"mkdir -p {skills_dir}", user="root")
+        for name in agent.skills:
+            src = host_skill_dir(name)
+            if not os.path.isdir(src):
+                die(
+                    f"skill {name!r} disappeared from host between "
+                    f"validation and copy at {src}."
+                )
+            dst = f"{skills_dir}/{name}"
+            info(f"copying skill {name} into {bottle.name}:{dst}")
+            bottle.exec(f"rm -rf {dst} && mkdir -p {dst}", user="root")
+            bottle.cp_in(f"{src}/.", f"{dst}/")
+            bottle.exec(f"chown -R node:node {dst}", user="root")
+
+    def provision_prompt(self, plan: "BottlePlan", bottle: "Bottle") -> str | None:
+        """Copy the prompt file into the guest, fix ownership/mode.
+        Returns the in-guest path iff the agent has a non-empty
+        prompt (drives `--append-system-prompt-file`); the file is
+        copied either way so the path always exists."""
+        prompt_path = _prompt_path(plan.guest_home)
+        bottle.cp_in(str(plan.prompt_file), prompt_path)
+        bottle.exec(
+            f"chown node:node {prompt_path} && chmod 600 {prompt_path}",
+            user="root",
+        )
+        agent = plan.spec.manifest.agents[plan.spec.agent_name]
+        return prompt_path if agent.prompt else None
+
+    def provision(self, plan: "BottlePlan", bottle: "Bottle") -> None:
+        """Apply the claude-side declarative provision steps from
+        `plan.agent_provision` — today that's the `claude.json`
+        trust-marker file. Hot-replace this with a richer flow as
+        claude-code's harness shape evolves."""
+        provision = plan.agent_provision
+        for d in provision.dirs:
+            path = shlex.quote(d.guest_path)
+            _exec(bottle, f"mkdir -p {path}", f"could not create {d.guest_path}")
+            _exec(
+                bottle,
+                f"chown {shlex.quote(d.owner)} {path}",
+                f"could not chown {d.guest_path}",
+            )
+            _exec(
+                bottle,
+                f"chmod {shlex.quote(d.mode)} {path}",
+                f"could not chmod {d.guest_path}",
+            )
+        for command in provision.pre_copy:
+            _exec(bottle, shlex.join(command.argv), command.error)
+        for f in provision.files:
+            bottle.cp_in(str(f.host_path), f.guest_path)
+            path = shlex.quote(f.guest_path)
+            _exec(
+                bottle,
+                f"chown {shlex.quote(f.owner)} {path}",
+                f"could not chown {f.guest_path}",
+            )
+            _exec(
+                bottle,
+                f"chmod {shlex.quote(f.mode)} {path}",
+                f"could not chmod {f.guest_path}",
+            )
+        for command in provision.verify:
+            _exec(bottle, shlex.join(command.argv), command.error)
+
+    def provision_supervise_mcp(
+        self,
+        plan: "BottlePlan",
+        bottle: "Bottle",
+        supervise_url: str,
+    ) -> None:
+        """Run `claude mcp add` inside the agent guest to register the
+        supervise sidecar in claude-code's user config (~/.claude.json).
+
+        Failure is logged but not fatal — the bottle still works without
+        the entry; the operator can register it manually."""
+        if plan.supervise_plan is None:
+            return
+        info(f"registering supervise MCP server in agent claude config → {supervise_url}")
+        r = bottle.exec(
+            f"claude mcp add --scope user --transport http "
+            f"{_SUPERVISE_MCP_NAME} {supervise_url}",
+            user="node",
+        )
+        if r.returncode != 0:
+            warn(
+                f"`claude mcp add supervise` failed (exit {r.returncode}): "
+                f"{(r.stderr or r.stdout or '').strip()}. Inside the bottle, "
+                f"register manually with: "
+                f"claude mcp add --scope user --transport http supervise {supervise_url}"
+            )
+
+
+def _exec(bottle: "Bottle", script: str, error: str) -> None:
+    result = bottle.exec(script, user="root")
+    if result.returncode != 0:
+        detail = (result.stderr or result.stdout).strip()
+        if detail:
+            detail = f": {detail}"
+        die(f"agent provider provisioning: {error}{detail}")

bot_bottle/contrib/codex/agent_provider.py

@@ -0,0 +1,271 @@
+"""Codex agent provider plugin (PRD 0050, contrib).
+
+The Codex-specific behavior previously inlined under
+`agent_provider.agent_provision_plan` (config.toml trust marker,
+chatgpt.com / api.openai.com egress routes, optional host-credential
+forwarding with dummy-auth.json + verify), plus the `codex mcp add`
+invocation that registers the supervise sidecar in Codex's
+~/.codex/config.toml (PRD 0050)."""
+
+from __future__ import annotations
+
+import os
+import shlex
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+from ...agent_provider import (
+    CODEX_HOST_CREDENTIAL_HOSTS,
+    AgentProvider,
+    AgentProviderRuntime,
+    AgentProvisionCommand,
+    AgentProvisionDir,
+    AgentProvisionFile,
+    AgentProvisionPlan,
+)
+from ...codex_auth import codex_host_access_token, write_codex_dummy_auth_file
+from ...egress import CODEX_HOST_CREDENTIAL_TOKEN_REF, EgressRoute
+from ...log import die, info, warn
+
+
+if TYPE_CHECKING:
+    from ...backend import Bottle, BottlePlan
+
+
+_REPO_ROOT = Path(__file__).resolve().parents[3]
+
+_SUPERVISE_MCP_NAME = "supervise"
+
+
+def _skills_dir(guest_home: str) -> str:
+    # Codex agents still read skills from the claude-code convention
+    # (~/.claude/skills/) — the bot-bottle-codex image follows the
+    # same layout. If Codex grows native skill discovery later,
+    # change here.
+    return f"{guest_home}/.claude/skills"
+
+
+def _prompt_path(guest_home: str) -> str:
+    return f"{guest_home}/.bot-bottle-prompt.txt"
+
+_RUNTIME = AgentProviderRuntime(
+    template="codex",
+    command="codex",
+    image="bot-bottle-codex:latest",
+    dockerfile=str(_REPO_ROOT / "Dockerfile.codex"),
+    prompt_mode="read_prompt_file",
+    bypass_args=("--dangerously-bypass-approvals-and-sandbox",),
+    resume_args=("resume", "--last"),
+    remote_control_args=(),
+)
+
+
+class CodexAgentProvider(AgentProvider):
+    @property
+    def runtime(self) -> AgentProviderRuntime:
+        return _RUNTIME
+
+    def provision_plan(
+        self,
+        *,
+        dockerfile: str,
+        state_dir: Path,
+        guest_home: str,
+        guest_env: dict[str, str] | None = None,
+        auth_token: str = "",
+        forward_host_credentials: bool = False,
+        host_env: dict[str, str] | None = None,
+        trusted_project_path: str = "",
+    ) -> AgentProvisionPlan:
+        del auth_token  # Claude-only knob
+        resolved_guest_env = dict(guest_env or {})
+        trusted_path = trusted_project_path or guest_home
+
+        env_vars: dict[str, str] = {
+            "CODEX_CA_CERTIFICATE": "/etc/ssl/certs/ca-certificates.crt",
+        }
+        auth_dir = resolved_guest_env.get("CODEX_HOME", f"{guest_home}/.codex")
+        if forward_host_credentials:
+            env_vars["CODEX_HOME"] = auth_dir
+
+        dirs = [AgentProvisionDir(auth_dir)]
+        files: list[AgentProvisionFile] = []
+        pre_copy: list[AgentProvisionCommand] = []
+        verify: list[AgentProvisionCommand] = []
+        provisioned_env: dict[str, str] = {}
+
+        config_path = f"{auth_dir}/config.toml"
+        config_file = state_dir / "codex-config.toml"
+        toml_path = trusted_path.replace("\\", "\\\\").replace('"', '\\"')
+        config_file.write_text(
+            f'[projects."{toml_path}"]\n'
+            'trust_level = "trusted"\n'
+        )
+        config_file.chmod(0o600)
+        files.append(AgentProvisionFile(config_file, config_path))
+
+        egress_routes: list[EgressRoute] = []
+        for host in CODEX_HOST_CREDENTIAL_HOSTS:
+            egress_routes.append(EgressRoute(
+                host=host,
+                auth_scheme="Bearer" if forward_host_credentials else "",
+                token_ref=CODEX_HOST_CREDENTIAL_TOKEN_REF if forward_host_credentials else "",
+                tls_passthrough=True,
+            ))
+
+        if forward_host_credentials:
+            _host_env = host_env or dict(os.environ)
+            provisioned_env[CODEX_HOST_CREDENTIAL_TOKEN_REF] = (
+                codex_host_access_token(_host_env)
+            )
+            auth_file = state_dir / "codex-auth.json"
+            write_codex_dummy_auth_file(auth_file, _host_env)
+            files.append(AgentProvisionFile(auth_file, f"{auth_dir}/auth.json"))
+            pre_copy.append(AgentProvisionCommand((
+                "find", auth_dir,
+                "-maxdepth", "1",
+                "-type", "f",
+                "(",
+                "-name", "*.sqlite",
+                "-o", "-name", "*.sqlite-*",
+                "-o", "-name", "*.codex-repair-*.bak",
+                ")",
+                "-delete",
+            ), "codex host credentials: could not reset runtime db files"))
+            verify.append(AgentProvisionCommand((
+                "runuser", "-u", "node", "--",
+                "env",
+                f"HOME={guest_home}",
+                f"CODEX_HOME={auth_dir}",
+                "codex", "login", "status",
+            ), (
+                "codex host credentials: dummy auth was copied into the "
+                "guest, but Codex did not accept it"
+            )))
+
+        return AgentProvisionPlan(
+            template=_RUNTIME.template,
+            command=_RUNTIME.command,
+            prompt_mode=_RUNTIME.prompt_mode,
+            image=_RUNTIME.image,
+            dockerfile=dockerfile,
+            env_vars=env_vars,
+            guest_env=resolved_guest_env,
+            dirs=tuple(dirs),
+            files=tuple(files),
+            pre_copy=tuple(pre_copy),
+            verify=tuple(verify),
+            egress_routes=tuple(egress_routes),
+            provisioned_env=provisioned_env,
+        )
+
+    def provision_skills(self, plan: "BottlePlan", bottle: "Bottle") -> None:
+        """Copy each named skill tree from `~/.claude/skills/<name>/`
+        on the host into the guest. No-op when the agent has no
+        skills."""
+        from ...backend.util import host_skill_dir
+
+        agent = plan.spec.manifest.agents[plan.spec.agent_name]
+        if not agent.skills:
+            return
+        skills_dir = _skills_dir(plan.guest_home)
+        bottle.exec(f"mkdir -p {skills_dir}", user="root")
+        for name in agent.skills:
+            src = host_skill_dir(name)
+            if not os.path.isdir(src):
+                die(
+                    f"skill {name!r} disappeared from host between "
+                    f"validation and copy at {src}."
+                )
+            dst = f"{skills_dir}/{name}"
+            info(f"copying skill {name} into {bottle.name}:{dst}")
+            bottle.exec(f"rm -rf {dst} && mkdir -p {dst}", user="root")
+            bottle.cp_in(f"{src}/.", f"{dst}/")
+            bottle.exec(f"chown -R node:node {dst}", user="root")
+
+    def provision_prompt(self, plan: "BottlePlan", bottle: "Bottle") -> str | None:
+        """Copy the prompt file into the guest, fix ownership/mode.
+        Codex reads it via the agent's `Read and follow the
+        instructions in <path>.` bootstrap (see `prompt_args`); the
+        file is copied either way so the path always exists."""
+        prompt_path = _prompt_path(plan.guest_home)
+        bottle.cp_in(str(plan.prompt_file), prompt_path)
+        bottle.exec(
+            f"chown node:node {prompt_path} && chmod 600 {prompt_path}",
+            user="root",
+        )
+        agent = plan.spec.manifest.agents[plan.spec.agent_name]
+        return prompt_path if agent.prompt else None
+
+    def provision(self, plan: "BottlePlan", bottle: "Bottle") -> None:
+        """Apply the codex-side declarative provision steps from
+        `plan.agent_provision`: the `~/.codex/` dir + config.toml
+        trust marker, plus the dummy-auth.json drop + `codex login
+        status` verify when host-credential forwarding is on."""
+        provision = plan.agent_provision
+        for d in provision.dirs:
+            path = shlex.quote(d.guest_path)
+            _exec(bottle, f"mkdir -p {path}", f"could not create {d.guest_path}")
+            _exec(
+                bottle,
+                f"chown {shlex.quote(d.owner)} {path}",
+                f"could not chown {d.guest_path}",
+            )
+            _exec(
+                bottle,
+                f"chmod {shlex.quote(d.mode)} {path}",
+                f"could not chmod {d.guest_path}",
+            )
+        for command in provision.pre_copy:
+            _exec(bottle, shlex.join(command.argv), command.error)
+        for f in provision.files:
+            bottle.cp_in(str(f.host_path), f.guest_path)
+            path = shlex.quote(f.guest_path)
+            _exec(
+                bottle,
+                f"chown {shlex.quote(f.owner)} {path}",
+                f"could not chown {f.guest_path}",
+            )
+            _exec(
+                bottle,
+                f"chmod {shlex.quote(f.mode)} {path}",
+                f"could not chmod {f.guest_path}",
+            )
+        for command in provision.verify:
+            _exec(bottle, shlex.join(command.argv), command.error)
+
+    def provision_supervise_mcp(
+        self,
+        plan: "BottlePlan",
+        bottle: "Bottle",
+        supervise_url: str,
+    ) -> None:
+        """Run `codex mcp add` inside the agent guest to register the
+        supervise sidecar in Codex's user config (~/.codex/config.toml).
+
+        Mirrors the Claude provider's `claude mcp add` flow — failure
+        is logged but not fatal."""
+        if plan.supervise_plan is None:
+            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}",
+            user="node",
+        )
+        if r.returncode != 0:
+            warn(
+                f"`codex mcp add supervise` failed (exit {r.returncode}): "
+                f"{(r.stderr or r.stdout or '').strip()}. Inside the bottle, "
+                f"register manually with: "
+                f"codex mcp add --transport http supervise {supervise_url}"
+            )
+
+
+def _exec(bottle: "Bottle", script: str, error: str) -> None:
+    result = bottle.exec(script, user="root")
+    if result.returncode != 0:
+        detail = (result.stderr or result.stdout).strip()
+        if detail:
+            detail = f": {detail}"
+        die(f"agent provider provisioning: {error}{detail}")

bot_bottle/contrib/gitea/deploy_key_provisioner.py

@@ -0,0 +1,121 @@
+"""Gitea deploy-key provisioner (PRD 0048, contrib).
+
+Generates ed25519 keypairs via `ssh-keygen` and registers / deletes
+them using the Gitea deploy-key HTTP API. No new Python dependencies —
+only stdlib `urllib.request` and `subprocess`."""
+
+from __future__ import annotations
+
+import json
+import subprocess
+import tempfile
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+from ...deploy_key_provisioner import DeployKeyProvisioner
+
+
+class GiteaDeployKeyProvisioner(DeployKeyProvisioner):
+    """Manages deploy keys on a Gitea instance."""
+
+    def __init__(self, *, token: str, api_url: str) -> None:
+        self._token = token
+        self._api_url = api_url.rstrip("/")
+
+    def create(self, owner_repo: str, title: str) -> tuple[str, bytes]:
+        """Generate an ed25519 keypair, register the public half as a
+        repo deploy key, and return `(key_id, private_key_bytes)`.
+
+        The key is registered with `read_only=False` because git-gate
+        needs push access to forward gitleaks-scanned refs upstream."""
+        with tempfile.TemporaryDirectory() as tmpdir:
+            key_path = Path(tmpdir) / "key"
+            subprocess.run(
+                [
+                    "ssh-keygen", "-t", "ed25519",
+                    "-f", str(key_path),
+                    "-N", "",
+                ],
+                check=True,
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.DEVNULL,
+            )
+            private_key = key_path.read_bytes()
+            public_key = key_path.with_suffix(".pub").read_text().strip()
+
+        owner, repo = _split_owner_repo(owner_repo)
+        url = f"{self._api_url}/api/v1/repos/{owner}/{repo}/keys"
+        payload = json.dumps({
+            "key": public_key,
+            "read_only": False,
+            "title": title,
+        }).encode()
+        req = urllib.request.Request(
+            url,
+            data=payload,
+            headers={
+                "Authorization": f"token {self._token}",
+                "Content-Type": "application/json",
+            },
+            method="POST",
+        )
+        try:
+            with urllib.request.urlopen(req) as resp:
+                body = json.loads(resp.read())
+        except urllib.error.HTTPError as exc:
+            _body = _read_error_body(exc)
+            raise RuntimeError(
+                f"failed to create deploy key for {owner_repo}: "
+                f"HTTP {exc.code} — {_body}"
+            ) from exc
+        except urllib.error.URLError as exc:
+            raise RuntimeError(
+                f"failed to create deploy key for {owner_repo}: {exc.reason}"
+            ) from exc
+
+        return str(body["id"]), private_key
+
+    def delete(self, owner_repo: str, key_id: str) -> None:
+        """Delete the deploy key. HTTP 404 (already gone) is success.
+        All other errors raise RuntimeError so teardown halts."""
+        owner, repo = _split_owner_repo(owner_repo)
+        url = f"{self._api_url}/api/v1/repos/{owner}/{repo}/keys/{key_id}"
+        req = urllib.request.Request(
+            url,
+            headers={"Authorization": f"token {self._token}"},
+            method="DELETE",
+        )
+        try:
+            with urllib.request.urlopen(req):
+                pass
+        except urllib.error.HTTPError as exc:
+            if exc.code == 404:
+                return
+            _body = _read_error_body(exc)
+            raise RuntimeError(
+                f"failed to delete deploy key {key_id} for {owner_repo}: "
+                f"HTTP {exc.code} — {_body}"
+            ) from exc
+        except urllib.error.URLError as exc:
+            raise RuntimeError(
+                f"failed to delete deploy key {key_id} for {owner_repo}: "
+                f"{exc.reason}"
+            ) from exc
+
+
+def _split_owner_repo(owner_repo: str) -> tuple[str, str]:
+    """Split `'owner/repo'` into `('owner', 'repo')`."""
+    parts = owner_repo.split("/", 1)
+    if len(parts) != 2 or not all(parts):
+        raise ValueError(
+            f"expected 'owner/repo' format, got {owner_repo!r}"
+        )
+    return parts[0], parts[1]
+
+
+def _read_error_body(exc: urllib.error.HTTPError) -> str:
+    try:
+        return exc.read().decode("utf-8", errors="replace")
+    except Exception:
+        return ""

bot_bottle/deploy_key_provisioner.py

@@ -0,0 +1,52 @@
+"""Deploy-key provisioner interface and factory (PRD 0048).
+
+The core defines the abstract contract; concrete implementations live
+in `bot_bottle/contrib/<provider>/deploy_key_provisioner.py`. The
+factory `get_provisioner` imports contrib modules lazily so that a
+missing optional dependency in one provider doesn't break unrelated
+features."""
+
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+
+
+class DeployKeyProvisioner(ABC):
+    """Manages a single deploy-key lifecycle on a remote forge."""
+
+    @abstractmethod
+    def create(self, owner_repo: str, title: str) -> tuple[str, bytes]:
+        """Generate a keypair and register the public half as a
+        deploy key on the forge.
+
+        `owner_repo` is the `<owner>/<repo>` path (no `.git` suffix).
+        `title` is the human-readable label shown in the forge UI.
+
+        Returns `(key_id, private_key_bytes)` where `key_id` is opaque
+        to the caller and is only ever passed back to `delete`."""
+
+    @abstractmethod
+    def delete(self, owner_repo: str, key_id: str) -> None:
+        """Delete the registered deploy key.
+
+        Must not raise if the key is already absent (HTTP 404 is
+        success). Must raise for all other failures so teardown halts."""
+
+
+def get_provisioner(
+    provider: str, token: str, api_url: str
+) -> DeployKeyProvisioner:
+    """Instantiate the contrib provisioner for `provider`.
+
+    Raises `ManifestError` for unknown providers so the error surfaces
+    at parse time rather than at runtime."""
+    if provider == "gitea":
+        from bot_bottle.contrib.gitea.deploy_key_provisioner import (
+            GiteaDeployKeyProvisioner,
+        )
+        return GiteaDeployKeyProvisioner(token=token, api_url=api_url)
+    from .manifest_util import ManifestError
+    raise ManifestError(
+        f"unknown provisioned_key provider: {provider!r}; "
+        f"available: gitea"
+    )

bot_bottle/egress.py

@@ -24,12 +24,19 @@ flow (PRD 0014) at egress and renames the MCP tool.
 
 from __future__ import annotations
 
+import dataclasses
 from abc import ABC, abstractmethod
 from dataclasses import dataclass
 from pathlib import Path
+from typing import TYPE_CHECKING
 
+from .egress_addon_core import Route
 from .log import die
-from .manifest import Bottle
+
+if TYPE_CHECKING:
+    from .manifest import Bottle
+
+CODEX_HOST_CREDENTIAL_TOKEN_REF = "BOT_BOTTLE_CODEX_HOST_ACCESS_TOKEN"
 
 
 # DNS name agents will dial for the per-bottle egress sidecar.
@@ -48,32 +55,30 @@ EGRESS_ROUTES_IN_CONTAINER = "/etc/egress/routes.yaml"
 
 
 @dataclass(frozen=True)
-class EgressRoute:
-    """One resolved route on the egress sidecar.
+class EgressRoute(Route):
+    """Host-side extension of the addon's `Route`.
 
-    `host` matches the request's hostname (case-insensitive). The
-    optional `path_allowlist` constrains the URL path; empty tuple
-    means no path-level filtering. The `auth_scheme` / `token_env` /
-    `token_ref` triple is the credential-injection config; empty
-    strings mean "no auth injection" (the manifest's nested `auth`
-    block was omitted).
+    Inherits `host`, `path_allowlist`, `auth_scheme`, and `token_env`
+    from `egress_addon_core.Route` — those are the fields that cross the
+    YAML wire into the sidecar. The three fields below are host-only and
+    are never serialised to the addon.
 
-    `token_env` is the env-var slot inside the egress container
-    (e.g. `EGRESS_TOKEN_0`); `token_ref` is the host env var
-    the CLI reads at launch and forwards into the container's environ
-    under `token_env`. Routes that share a `token_ref` coalesce to
-    one `token_env` slot.
+    `token_ref` is the host env var the CLI reads at launch and forwards
+    into the container's environ under `token_env`. Routes that share a
+    `token_ref` coalesce to one `token_env` slot.
 
-    `roles` carries the manifest route's optional role markers (see
-    `manifest.EGRESS_ROLES`). The launch step reads these for
-    side effects like the claude-code OAuth placeholder env."""
+    `roles` carries the manifest route's role tuple (reserved for
+    future use; always empty today).
+
+    `tls_passthrough` signals that pipelock must not TLS-MITM this
+    host — either because the manifest declared `pipelock.tls_passthrough:
+    true` (lifted in `egress_manifest_routes`) or because a provider
+    route set it (e.g. egress injects its own Bearer on that host
+    after the agent boundary and pipelock's header DLP would block it)."""
 
-    host: str
-    path_allowlist: tuple[str, ...] = ()
-    auth_scheme: str = ""
-    token_env: str = ""
     token_ref: str = ""
     roles: tuple[str, ...] = ()
+    tls_passthrough: bool = False
 
 
 @dataclass(frozen=True)
@@ -130,55 +135,60 @@ class EgressPlan:
 def egress_manifest_routes(
     bottle: Bottle,
 ) -> tuple[EgressRoute, ...]:
-    """Lift each `bottle.egress.routes[]` manifest entry into a
-    resolved EgressRoute. Order is preserved so route lookup at
-    the proxy is stable.
-
-    Token-env slots are assigned per distinct `token_ref`: the first
-    authenticated route with `token_ref` "GH_PAT" gets
-    `EGRESS_TOKEN_0`; a second route with the same `token_ref`
-    shares slot 0. Unauthenticated routes (`auth` omitted) contribute
-    no slot.
-
-    This is the effective set the addon enforces. Provider runtime
-    routes are intentionally not injected implicitly; every allowed
-    host must come from the home-owned bottle manifest."""
+    """Lift each `bottle.egress.routes[]` manifest entry into an EgressRoute.
+    Order is preserved. Token slots are not assigned here — slot assignment
+    is a final step in `egress_routes_for_bottle` after provider and manifest
+    routes are merged."""
     out: list[EgressRoute] = []
-    slot_for_token: dict[str, str] = {}
     for r in bottle.egress.routes:
-        if r.AuthScheme and r.TokenRef:
-            token_env = slot_for_token.get(r.TokenRef)
-            if token_env is None:
-                token_env = f"EGRESS_TOKEN_{len(slot_for_token)}"
-                slot_for_token[r.TokenRef] = token_env
-            out.append(EgressRoute(
-                host=r.Host,
-                path_allowlist=r.PathAllowlist,
-                auth_scheme=r.AuthScheme,
-                token_env=token_env,
-                token_ref=r.TokenRef,
-                roles=r.Role,
-            ))
-        else:
-            out.append(EgressRoute(
-                host=r.Host,
-                path_allowlist=r.PathAllowlist,
-                roles=r.Role,
-            ))
+        out.append(EgressRoute(
+            host=r.Host,
+            path_allowlist=r.PathAllowlist,
+            auth_scheme=r.AuthScheme,
+            token_ref=r.TokenRef,
+            roles=r.Role,
+            tls_passthrough=r.Pipelock.TlsPassthrough,
+        ))
     return tuple(out)
 
 
 def egress_routes_for_bottle(
     bottle: Bottle,
+    provider_routes: tuple[EgressRoute, ...] = (),
 ) -> tuple[EgressRoute, ...]:
-    """Effective egress routes. This is what gets rendered into
-    routes.yaml + what the addon enforces.
+    """Effective egress routes for the agent.
 
-    Operators that want to allow a host declare it directly in
-    `bottle.egress.routes` as an authenticated route or bare-pass entry
-    (`- host: <name>`). The legacy `bottle.egress.allowlist`
-    folding is gone — egress is the single allowlist surface."""
-    return egress_manifest_routes(bottle)
+    Provider routes own their hosts outright; manifest routes for hosts
+    not claimed by any provider are appended. Token slots are assigned
+    in a final pass over the merged list in order, so provisioned routes
+    get the lower slot numbers."""
+    manifest = egress_manifest_routes(bottle)
+    provisioned_hosts = {pr.host.lower() for pr in provider_routes}
+    merged = list(provider_routes) + [
+        r for r in manifest if r.host.lower() not in provisioned_hosts
+    ]
+    return _assign_token_slots(merged)
+
+
+def _assign_token_slots(
+    routes: list[EgressRoute],
+) -> tuple[EgressRoute, ...]:
+    """Assign EGRESS_TOKEN_N slots to authenticated routes in order.
+
+    Routes sharing a token_ref share a slot. Unauthenticated routes
+    (no auth_scheme / token_ref) keep token_env empty."""
+    slot_for_ref: dict[str, str] = {}
+    out: list[EgressRoute] = []
+    for r in routes:
+        if r.auth_scheme and r.token_ref:
+            slot = slot_for_ref.get(r.token_ref)
+            if slot is None:
+                slot = f"EGRESS_TOKEN_{len(slot_for_ref)}"
+                slot_for_ref[r.token_ref] = slot
+            out.append(dataclasses.replace(r, token_env=slot))
+        else:
+            out.append(r)
+    return tuple(out)
 
 
 def egress_token_env_map(
@@ -193,7 +203,7 @@ def egress_token_env_map(
     silently picking one."""
     out: dict[str, str] = {}
     for r in routes:
-        if not r.token_env:
+        if not (r.auth_scheme and r.token_ref and r.token_env):
             continue
         existing = out.get(r.token_env)
         if existing is not None and existing != r.token_ref:
@@ -206,35 +216,43 @@ def egress_token_env_map(
     return out
 
 
+def _route_to_yaml_fields(r: Route) -> dict:
+    """Return the addon-visible fields for one route.
+
+    Single authoritative mapping between EgressRoute (host-side) and
+    egress_addon_core.Route (sidecar-side). When a field is added to
+    the addon's Route that must appear in the YAML, add it here and
+    in egress_addon_core._parse_one together."""
+    fields: dict = {"host": r.host}
+    if r.auth_scheme and r.token_env:
+        fields["auth_scheme"] = r.auth_scheme
+        fields["token_env"] = r.token_env
+    if r.path_allowlist:
+        fields["path_allowlist"] = list(r.path_allowlist)
+    return fields
+
+
 def egress_render_routes(
     routes: tuple[EgressRoute, ...],
 ) -> str:
     """Serialize the route table for the addon to read.
 
-    YAML content — no token values, no host env-var names. The only
-    thing the addon needs at runtime is the host → path_allowlist
-    + auth_scheme + in-container env-var mapping. The actual token
-    values arrive via the container's environ.
-
-    Authenticated routes carry `auth_scheme` + `token_env`;
-    unauthenticated routes omit both keys (the addon's parser
-    enforces both-or-neither). Hand-rolled YAML in the style of
-    `pipelock_render_yaml` so the addon's parser
-    (`yaml_subset.parse_yaml_subset`) round-trips it cleanly."""
+    YAML content — no token values, no host env-var names. Fields are
+    determined by `_route_to_yaml_fields`, which is the single point of
+    truth for the EgressRoute → egress_addon_core.Route mapping."""
     lines: list[str] = ["routes:"]
     if not routes:
-        # `routes:` with an empty list on the same line — the parser
-        # needs SOMETHING here. Empty inline list is the cleanest.
         lines[0] = "routes: []"
         return "\n".join(lines) + "\n"
     for r in routes:
-        lines.append(f'  - host: "{r.host}"')
-        if r.auth_scheme and r.token_env:
-            lines.append(f'    auth_scheme: "{r.auth_scheme}"')
-            lines.append(f'    token_env: "{r.token_env}"')
-        if r.path_allowlist:
+        f = _route_to_yaml_fields(r)
+        lines.append(f'  - host: "{f["host"]}"')
+        if "auth_scheme" in f:
+            lines.append(f'    auth_scheme: "{f["auth_scheme"]}"')
+            lines.append(f'    token_env: "{f["token_env"]}"')
+        if "path_allowlist" in f:
             lines.append("    path_allowlist:")
-            for p in r.path_allowlist:
+            for p in f["path_allowlist"]:
                 lines.append(f'      - "{p}"')
     return "\n".join(lines) + "\n"
 
@@ -274,18 +292,23 @@ class Egress(ABC):
     sidecar's start/stop lifecycle is backend-specific and lives on
     concrete subclasses."""
 
-    def prepare(self, bottle: Bottle, slug: str, stage_dir: Path) -> EgressPlan:
-        """Lift `bottle.egress.routes` into resolved routes,
-        render the routes file (mode 600) under `stage_dir`, and
+    def prepare(
+        self,
+        bottle: Bottle,
+        slug: str,
+        stage_dir: Path,
+        provider_routes: tuple[EgressRoute, ...] = (),
+    ) -> EgressPlan:
+        """Lift `bottle.egress.routes` + `provider_routes` into resolved
+        routes, render the routes file (mode 600) under `stage_dir`, and
         return the plan. Pure host-side, no docker subprocess. The
         token-env map records the mapping the launch step uses to
-        forward values from the host's environ into the sidecar's
-        environ.
+        forward values from the host's environ into the sidecar's environ.
 
         Returned plan is incomplete: the launch step must fill
         `internal_network` / `egress_network` / `pipelock_proxy_url`
         via `dataclasses.replace` before passing it to `.start`."""
-        routes = egress_routes_for_bottle(bottle)
+        routes = egress_routes_for_bottle(bottle, provider_routes)
         routes_path = stage_dir / "egress_routes.yaml"
         routes_path.write_text(egress_render_routes(routes))
         routes_path.chmod(0o600)
@@ -297,6 +320,7 @@ class Egress(ABC):
         )
 
 __all__ = [
+    "CODEX_HOST_CREDENTIAL_TOKEN_REF",
     "EGRESS_HOSTNAME",
     "EGRESS_ROUTES_IN_CONTAINER",
     "Egress",

bot_bottle/git_gate.py

@@ -29,22 +29,24 @@ backend-specific and lives on concrete subclasses (see
 
 from __future__ import annotations
 
+import dataclasses
+import os
+import shlex
 from abc import ABC, abstractmethod
-from dataclasses import dataclass, field
+from dataclasses import dataclass
 from pathlib import Path
-from typing import Mapping
 
-from .log import die
+from .log import info
 from .manifest import Bottle, GitEntry
 
 
 # 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"
-
-
-def _empty_str_map() -> dict[str, str]:
-    return {}
+# 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
 
 
 @dataclass(frozen=True)
@@ -60,10 +62,7 @@ class GitGateUpstream:
     KnownHostKey string from the manifest; the gate's start step
     materialises it into a known_hosts file if non-empty.
 
-    `extra_hosts` is a `{hostname: ip}` map the backend injects into
-    the gate container's `/etc/hosts` via `--add-host` so the gate
-    can resolve upstream hostnames that aren't reachable via the
-    container's default DNS (e.g. Tailscale-only hosts)."""
+    the gate credential paths inside the running sidecar."""
 
     name: str
     upstream_url: str
@@ -72,7 +71,6 @@ class GitGateUpstream:
     identity_file: str
     known_host_key: str
     known_hosts_file: Path = Path()
-    extra_hosts: Mapping[str, str] = field(default_factory=_empty_str_map)
 
 
 @dataclass(frozen=True)
@@ -109,46 +107,19 @@ def git_gate_upstreams_for_bottle(bottle: Bottle) -> tuple[GitGateUpstream, ...]
             upstream_port=e.UpstreamPort,
             identity_file=e.IdentityFile,
             known_host_key=e.KnownHostKey,
-            extra_hosts=dict(e.ExtraHosts),
         )
         for e in bottle.git
     )
 
 
-def git_gate_aggregate_extra_hosts(
-    upstreams: tuple[GitGateUpstream, ...],
-) -> dict[str, str]:
-    """Merge every upstream's `extra_hosts` into a single
-    `{hostname: ip}` map for `--add-host` on the gate container. Two
-    entries naming the same hostname with different IPs is a manifest
-    bug — the gate has one /etc/hosts — so die loudly with the
-    conflicting names rather than silently picking one."""
-    merged: dict[str, str] = {}
-    source: dict[str, str] = {}
-    for u in upstreams:
-        for host, ip in u.extra_hosts.items():
-            existing = merged.get(host)
-            if existing is None:
-                merged[host] = ip
-                source[host] = u.name
-            elif existing != ip:
-                die(
-                    f"git-gate ExtraHosts conflict: '{host}' maps to "
-                    f"'{existing}' in upstream '{source[host]}' and to "
-                    f"'{ip}' in upstream '{u.name}'. The gate has one "
-                    f"/etc/hosts; pick one IP."
-                )
-    return merged
-
-
 def git_gate_render_gitconfig(
-    entries: tuple[GitEntry, ...], gate_host: str
+    entries: tuple[GitEntry, ...], gate_host: str, *, scheme: str = "git",
 ) -> str:
     """Render the agent's ~/.gitconfig content for git-gate
     `insteadOf` rewrites. Pure host-side, no docker / smolvm;
     exposed for tests + reuse across backends.
 
-    `gate_host` is the part of the URL between `git://` and the
+    `gate_host` is the part of the URL between `<scheme>://` and the
     repo path — backends differ here:
       - docker:        `git-gate` (the short network alias)
       - smolmachines:  `<bundle_ip>:<port>` (no DNS in the
@@ -165,7 +136,7 @@ def git_gate_render_gitconfig(
         "# fetch-from-upstream-before-every-upload-pack via access-hook).\n",
     ]
     for entry in entries:
-        out.append(f'[url "git://{gate_host}/{entry.Name}.git"]\n')
+        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:
             port = (
@@ -233,20 +204,20 @@ def git_gate_render_entrypoint(upstreams: tuple[GitGateUpstream, ...]) -> str:
         "  git -C \"$repo\" config git-gate.identityFile \"$keyfile\"",
         "  git -C \"$repo\" config git-gate.knownHosts \"$hostsfile\"",
         "  git -C \"$repo\" config receive.denyCurrentBranch ignore",
+        "  git -C \"$repo\" config http.receivepack true",
         "  install -m 755 /etc/git-gate/pre-receive \"$repo/hooks/pre-receive\"",
         "}",
         "",
         "mkdir -p /git",
     ]
     for u in upstreams:
-        # Single-quote args so URL/path content (containing : and /)
-        # passes through ash unmangled. Names came through the manifest
-        # validator so they don't contain a single quote.
-        lines.append(f"init_repo '{u.name}' '{u.upstream_url}'")
+        lines.append(f"init_repo {shlex.quote(u.name)} {shlex.quote(u.upstream_url)}")
     lines.extend([
         "",
         "exec git daemon \\",
         "  --reuseaddr \\",
+        f"  --timeout={GIT_GATE_DAEMON_TIMEOUT_SECS} \\",
+        f"  --init-timeout={GIT_GATE_DAEMON_TIMEOUT_SECS} \\",
         "  --base-path=/git \\",
         "  --export-all \\",
         "  --enable=receive-pack \\",
@@ -280,7 +251,14 @@ while IFS=' ' read -r old new ref; do
   [ -z "$ref" ] && continue
   [ "$new" = "$zero" ] && continue
   if [ "$old" = "$zero" ]; then
-    log_opts="$new"
+    # New ref: scan only the commits this push introduces — those
+    # reachable from $new but not from any ref the gate already has.
+    # Everything already on the gate arrived via upstream mirror-fetch
+    # or a previously gitleaks-scanned push, so it's already-upstream
+    # or already-scanned; re-scanning it (the old `$new` full-ancestry
+    # range) only resurfaces historical findings and blocks every new
+    # branch. See PRD 0028 / issue #106.
+    log_opts="$new --not --all"
   else
     log_opts="$old..$new"
   fi
@@ -300,7 +278,7 @@ if [ ! -f "$hostsfile" ]; then
   echo "git-gate: add KnownHostKey to the bottle.git entry and restart the bottle" >&2
   exit 1
 fi
-ssh_cmd="ssh -i $keyfile -o UserKnownHostsFile=$hostsfile -o StrictHostKeyChecking=yes -o IdentitiesOnly=yes"
+ssh_cmd="ssh -i $keyfile -o UserKnownHostsFile=$hostsfile -o StrictHostKeyChecking=yes -o IdentitiesOnly=yes -o BatchMode=yes -o ConnectTimeout=10"
 
 while IFS=' ' read -r old new ref; do
   [ -z "$ref" ] && continue
@@ -355,7 +333,7 @@ if [ -z "$keyfile" ] || [ ! -f "$hostsfile" ]; then
   echo "git-gate: missing credentials for $repo_dir; refusing fetch" >&2
   exit 1
 fi
-ssh_cmd="ssh -i $keyfile -o UserKnownHostsFile=$hostsfile -o StrictHostKeyChecking=yes -o IdentitiesOnly=yes"
+ssh_cmd="ssh -i $keyfile -o UserKnownHostsFile=$hostsfile -o StrictHostKeyChecking=yes -o IdentitiesOnly=yes -o BatchMode=yes -o ConnectTimeout=10"
 
 echo "git-gate: refreshing $repo_dir from upstream" >&2
 if ! GIT_SSH_COMMAND="$ssh_cmd" git -C "$repo_dir" fetch origin --prune >&2; then
@@ -382,6 +360,80 @@ exit 0
 """
 
 
+def _provision_dynamic_key(
+    entry: GitEntry,
+    slug: str,
+    stage_dir: Path,
+) -> str:
+    """Generate a fresh ed25519 keypair, register the public half with
+    the forge, and persist the private key + key ID under `stage_dir`.
+
+    Returns the host-side path to the private key file so the caller
+    can inject it into the GitGateUpstream as `identity_file`."""
+    from .deploy_key_provisioner import get_provisioner
+    pk = entry.ProvisionedKey
+    assert pk is not None
+    token = os.environ.get(pk.token_env)
+    if token is None:
+        raise RuntimeError(
+            f"git-gate.repos[{entry.Name!r}] provisioned_key.token_env"
+            f" = {pk.token_env!r}: env var is not set"
+        )
+    api_url = pk.api_url or f"https://{entry.UpstreamHost}"
+    provisioner = get_provisioner(pk.provider, token, api_url)
+
+    owner_repo = entry.UpstreamPath
+    if owner_repo.endswith(".git"):
+        owner_repo = owner_repo[:-4]
+    title = f"bot-bottle:{slug}:{entry.Name}"
+
+    info(f"provisioning deploy key for git-gate.repos[{entry.Name!r}]")
+    key_id, private_key_bytes = provisioner.create(owner_repo, title)
+
+    key_file = stage_dir / f"{entry.Name}-key"
+    key_file.write_bytes(private_key_bytes)
+    key_file.chmod(0o600)
+
+    id_file = stage_dir / f"{entry.Name}-deploy-key-id"
+    id_file.write_text(key_id)
+    id_file.chmod(0o600)
+
+    info(f"provisioned deploy key {key_id} for git-gate.repos[{entry.Name!r}]")
+    return str(key_file)
+
+
+def revoke_git_gate_provisioned_keys(bottle: Bottle, stage_dir: Path) -> None:
+    """Revoke all deploy keys provisioned for `bottle` during prepare.
+
+    Called at teardown after containers stop. Raises if any revocation
+    fails — a stranded key is a security concern that the operator must
+    address manually."""
+    from .deploy_key_provisioner import get_provisioner
+    for entry in bottle.git:
+        if entry.ProvisionedKey is None:
+            continue
+        pk = entry.ProvisionedKey
+        id_file = stage_dir / f"{entry.Name}-deploy-key-id"
+        if not id_file.exists():
+            continue
+        key_id = id_file.read_text().strip()
+        token = os.environ.get(pk.token_env)
+        if token is None:
+            raise RuntimeError(
+                f"git-gate.repos[{entry.Name!r}] provisioned_key.token_env"
+                f" = {pk.token_env!r}: env var is not set;"
+                f" cannot revoke deploy key {key_id}"
+            )
+        api_url = pk.api_url or f"https://{entry.UpstreamHost}"
+        provisioner = get_provisioner(pk.provider, token, api_url)
+        owner_repo = entry.UpstreamPath
+        if owner_repo.endswith(".git"):
+            owner_repo = owner_repo[:-4]
+        info(f"revoking deploy key {key_id} for git-gate.repos[{entry.Name!r}]")
+        provisioner.delete(owner_repo, key_id)
+        info(f"revoked deploy key {key_id} for git-gate.repos[{entry.Name!r}]")
+
+
 class GitGate(ABC):
     """The per-agent git-gate. Encapsulates the host-side prepare
     (upstream lift + entrypoint/hook render); the sidecar's
@@ -393,10 +445,21 @@ class GitGate(ABC):
         entrypoint, pre-receive hook, and access-hook scripts (mode
         600) under `stage_dir`. Pure host-side, no docker subprocess.
 
+        For `provisioned_key` entries, also generates and registers
+        a fresh deploy key via the forge API and writes the private key
+        + key ID to `stage_dir`.
+
         Returned plan is incomplete: the launch step must fill
         `internal_network` / `egress_network` via `dataclasses.replace`
         before passing the plan to `.start`."""
-        upstreams = git_gate_upstreams_for_bottle(bottle)
+        upstreams_list = list(git_gate_upstreams_for_bottle(bottle))
+        for i, entry in enumerate(bottle.git):
+            if entry.ProvisionedKey is not None:
+                key_file = _provision_dynamic_key(entry, slug, stage_dir)
+                upstreams_list[i] = dataclasses.replace(
+                    upstreams_list[i], identity_file=key_file
+                )
+        upstreams = tuple(upstreams_list)
         entrypoint = stage_dir / "git_gate_entrypoint.sh"
         entrypoint.write_text(git_gate_render_entrypoint(upstreams))
         entrypoint.chmod(0o600)
@@ -429,7 +492,6 @@ class GitGate(ABC):
                     identity_file=u.identity_file,
                     known_host_key=u.known_host_key,
                     known_hosts_file=known_hosts_file,
-                    extra_hosts=dict(u.extra_hosts),
                 )
             )
         return GitGatePlan(

bot_bottle/git_http_backend.py

@@ -0,0 +1,175 @@
+"""Tiny smart-HTTP wrapper for git-gate repos.
+
+Used by the smolmachines backend where `git://` push traffic over the
+host-published Docker port can hang before receive-pack reaches hooks.
+The wrapper serves the same `/git/*.git` bare repos through
+`git http-backend`, so pre-receive and upstream forwarding remain the
+git-gate enforcement point.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+from urllib.parse import urlsplit
+
+
+DEFAULT_PORT = 9420
+
+# Body-size cap matching supervise_server.py's 1 MiB limit.
+MAX_BODY_BYTES = 1 * 1024 * 1024
+
+
+class GitHttpHandler(BaseHTTPRequestHandler):
+    server_version = "bot-bottle-git-http/1"
+
+    def do_GET(self) -> None:
+        self._run_backend()
+
+    def do_POST(self) -> None:
+        self._run_backend()
+
+    def _run_backend(self) -> None:
+        parsed = urlsplit(self.path)
+        if self._is_upload_pack(parsed.path, parsed.query):
+            repo_dir = self._repo_dir(parsed.path)
+            if repo_dir is None:
+                self.send_error(404)
+                return
+            hook_path = os.environ.get(
+                "GIT_GATE_ACCESS_HOOK", "/etc/git-gate/access-hook",
+            )
+            peer = self.client_address[0]
+            hook = subprocess.run(
+                [hook_path, "upload-pack", str(repo_dir), peer, peer],
+                capture_output=True,
+                check=False,
+            )
+            if hook.returncode != 0:
+                detail = (hook.stderr or hook.stdout).decode(
+                    "utf-8", errors="replace",
+                ).rstrip()
+                if detail:
+                    for line in detail.splitlines():
+                        self.log_message("access-hook denied %s: %s",
+                                         parsed.path, line)
+                else:
+                    self.log_message(
+                        "access-hook denied %s: exit=%d (no output)",
+                        parsed.path, hook.returncode,
+                    )
+                self.send_response(403)
+                self.send_header("Content-Type", "text/plain; charset=utf-8")
+                self.end_headers()
+                self.wfile.write(hook.stderr or hook.stdout)
+                return
+        env = os.environ.copy()
+        env.update({
+            "GIT_PROJECT_ROOT": os.environ.get("GIT_PROJECT_ROOT", "/git"),
+            "GIT_HTTP_EXPORT_ALL": "1",
+            "REQUEST_METHOD": self.command,
+            "PATH_INFO": parsed.path,
+            "QUERY_STRING": parsed.query,
+            "CONTENT_TYPE": self.headers.get("content-type", ""),
+            "CONTENT_LENGTH": self.headers.get("content-length", "0"),
+            "REMOTE_ADDR": self.client_address[0],
+            "REMOTE_PORT": str(self.client_address[1]),
+            "REMOTE_USER": "",
+            "SERVER_NAME": self.server.server_name,
+            "SERVER_PORT": str(self.server.server_port),
+            "SERVER_PROTOCOL": self.request_version,
+        })
+        for header, variable in (
+            ("accept", "HTTP_ACCEPT"),
+            ("content-encoding", "HTTP_CONTENT_ENCODING"),
+            ("git-protocol", "HTTP_GIT_PROTOCOL"),
+            ("user-agent", "HTTP_USER_AGENT"),
+        ):
+            value = self.headers.get(header)
+            if value:
+                env[variable] = value
+        raw_length = self.headers.get("content-length", "0") or "0"
+        try:
+            length = int(raw_length)
+        except ValueError:
+            self.send_error(400, "Bad Content-Length")
+            return
+        if length < 0:
+            self.send_error(400, "Negative Content-Length")
+            return
+        if length > MAX_BODY_BYTES:
+            self.send_error(413, "Request body too large")
+            return
+        body = self.rfile.read(length) if length else b""
+        proc = subprocess.run(
+            ["git", "http-backend"],
+            input=body,
+            env=env,
+            capture_output=True,
+            check=False,
+        )
+        self._write_cgi_response(proc.stdout)
+
+    def _repo_dir(self, path: str) -> Path | None:
+        root = Path(os.environ.get("GIT_PROJECT_ROOT", "/git")).resolve()
+        relative = path.lstrip("/").split(".git", 1)[0] + ".git"
+        candidate = (root / relative).resolve()
+        if root not in (candidate, *candidate.parents):
+            return None
+        if not candidate.is_dir():
+            return None
+        return candidate
+
+    @staticmethod
+    def _is_upload_pack(path: str, query: str) -> bool:
+        if path.endswith("/git-upload-pack"):
+            return True
+        if path.endswith("/info/refs"):
+            return any(
+                pair == "service=git-upload-pack"
+                for pair in query.split("&")
+            )
+        return False
+
+    def _write_cgi_response(self, raw: bytes) -> None:
+        head, sep, body = raw.partition(b"\r\n\r\n")
+        line_sep = b"\r\n"
+        if not sep:
+            head, sep, body = raw.partition(b"\n\n")
+            line_sep = b"\n"
+        status = 200
+        headers: list[tuple[str, str]] = []
+        for line in head.split(line_sep):
+            if not line:
+                continue
+            key, _, value = line.decode("latin1").partition(":")
+            value = value.strip()
+            if key.lower() == "status":
+                status = int(value.split()[0])
+            else:
+                headers.append((key, value))
+        self.send_response(status)
+        for key, value in headers:
+            self.send_header(key, value)
+        self.end_headers()
+        self.wfile.write(body)
+
+    def log_message(self, fmt: str, *args: object) -> None:
+        sys.stdout.write(fmt % args + "\n")
+        sys.stdout.flush()
+
+
+def main() -> int:
+    port = int(os.environ.get("GIT_HTTP_PORT", str(DEFAULT_PORT)))
+    server = ThreadingHTTPServer(("0.0.0.0", port), GitHttpHandler)
+    sys.stdout.write(f"git-http listening on 0.0.0.0:{port}\n")
+    sys.stdout.flush()
+    server.serve_forever()
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

bot_bottle/log.py

@@ -14,11 +14,23 @@ def warn(msg: str) -> None:
     print(f"bot-bottle: warning: {msg}", file=sys.stderr)
 
 
+def error(msg: str) -> None:
+    print(f"bot-bottle: error: {msg}", file=sys.stderr)
+
+
 class Die(SystemExit):
     """Raised by die() so callers (and tests) can distinguish a deliberate
-    fatal exit from an unrelated SystemExit."""
+    fatal exit from an unrelated SystemExit.
+
+    Carries the human-facing message so a caller that suppressed stderr
+    — e.g. the curses dashboard, whose alternate screen is wiped when the
+    terminal is restored — can re-surface the reason after the fact."""
+
+    def __init__(self, code: int = 1, message: str = "") -> None:
+        super().__init__(code)
+        self.message = message
 
 
 def die(msg: str) -> NoReturn:
-    print(f"bot-bottle: error: {msg}", file=sys.stderr)
-    raise Die(1)
+    error(msg)
+    raise Die(1, msg)

bot_bottle/manifest_agent.py

@@ -0,0 +1,166 @@
+"""Agent configuration manifest dataclasses."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import cast
+
+from .agent_provider import PROVIDER_TEMPLATES
+from .manifest_util import ManifestError, as_json_object
+from .manifest_git import GitUser
+from .manifest_schema import AGENT_MODEL_KEYS
+
+
+@dataclass(frozen=True)
+class AgentProvider:
+    """Provider/template for the agent process inside a bottle.
+
+    `template` selects a built-in launch/runtime contract. `dockerfile`
+    optionally points at a custom agent-image Dockerfile while leaving
+    bot-bottle's sidecar infrastructure intact.
+
+    `auth_token` names the host env var that holds the provider's OAuth
+    token (Claude only). The provisioner injects a provider-owned egress
+    route for api.anthropic.com that re-injects this token as the Bearer
+    header, and sets a placeholder CLAUDE_CODE_OAUTH_TOKEN in the agent
+    so the Claude Code CLI starts.
+
+    `forward_host_credentials` forwards the host Codex auth token into
+    the egress sidecar (Codex only).
+    """
+
+    template: str = "claude"
+    dockerfile: str = ""
+    auth_token: str = ""
+    forward_host_credentials: bool = False
+
+    @classmethod
+    def from_dict(cls, bottle_name: str, raw: object) -> "AgentProvider":
+        d = as_json_object(raw, f"bottle '{bottle_name}' agent_provider")
+        for k in d:
+            if k not in {"template", "dockerfile", "auth_token", "forward_host_credentials"}:
+                raise ManifestError(
+                    f"bottle '{bottle_name}' agent_provider has unknown key {k!r}; "
+                    f"allowed: template, dockerfile, auth_token, forward_host_credentials"
+                )
+        template = d.get("template", "claude")
+        if not isinstance(template, str) or not template:
+            raise ManifestError(
+                f"bottle '{bottle_name}' agent_provider.template must be a "
+                f"non-empty string"
+            )
+        if template not in PROVIDER_TEMPLATES:
+            raise ManifestError(
+                f"bottle '{bottle_name}' agent_provider.template {template!r} "
+                f"is not one of {', '.join(sorted(PROVIDER_TEMPLATES))}"
+            )
+        dockerfile = d.get("dockerfile", "")
+        if not isinstance(dockerfile, str):
+            raise ManifestError(
+                f"bottle '{bottle_name}' agent_provider.dockerfile must be a "
+                f"string (was {type(dockerfile).__name__})"
+            )
+        auth_token = d.get("auth_token", "")
+        if not isinstance(auth_token, str):
+            raise ManifestError(
+                f"bottle '{bottle_name}' agent_provider.auth_token must be a "
+                f"string (was {type(auth_token).__name__})"
+            )
+        if auth_token and template != "claude":
+            raise ManifestError(
+                f"bottle '{bottle_name}' agent_provider.auth_token is only "
+                f"supported for template 'claude'"
+            )
+        forward_host_credentials = d.get("forward_host_credentials", False)
+        if not isinstance(forward_host_credentials, bool):
+            raise ManifestError(
+                f"bottle '{bottle_name}' agent_provider.forward_host_credentials "
+                f"must be a boolean (was {type(forward_host_credentials).__name__})"
+            )
+        if forward_host_credentials and template != "codex":
+            raise ManifestError(
+                f"bottle '{bottle_name}' agent_provider.forward_host_credentials "
+                "is currently only supported for template 'codex'"
+            )
+        return cls(
+            template=template,
+            dockerfile=dockerfile,
+            auth_token=auth_token,
+            forward_host_credentials=forward_host_credentials,
+        )
+
+
+@dataclass(frozen=True)
+class Agent:
+    bottle: str
+    skills: tuple[str, ...] = ()
+    prompt: str = ""
+    # Per-agent git identity (issue #94). Overlays the referenced
+    # bottle's git-gate.user per-field at `Manifest.bottle_for`. Only
+    # `user` is allowed at the agent level; `repos` stays bottle-only
+    # because it carries credentials and host trust.
+    git_user: GitUser = GitUser()
+
+    @classmethod
+    def from_dict(cls, name: str, raw: object, bottle_names: set[str]) -> "Agent":
+        d = as_json_object(raw, f"agent '{name}'")
+        unknown = set(d.keys()) - AGENT_MODEL_KEYS
+        if unknown:
+            allowed = ", ".join(sorted(AGENT_MODEL_KEYS))
+            raise ManifestError(
+                f"agent '{name}' has unknown key(s) {sorted(unknown)}; "
+                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 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}"
+            )
+
+        skills: tuple[str, ...] = ()
+        skills_raw = d.get("skills")
+        if skills_raw is not None:
+            if not isinstance(skills_raw, list):
+                raise ManifestError(f"agent '{name}' skills must be an array (was {type(skills_raw).__name__})")
+            collected: list[str] = []
+            skills_list = cast(list[object], skills_raw)
+            for i, skill in enumerate(skills_list):
+                if not isinstance(skill, str):
+                    raise ManifestError(
+                        f"agent '{name}' skills[{i}] must be a string "
+                        f"(was {type(skill).__name__})"
+                    )
+                collected.append(skill)
+            skills = tuple(collected)
+
+        prompt_raw = d.get("prompt")
+        if prompt_raw is None:
+            prompt = ""
+        elif isinstance(prompt_raw, str):
+            prompt = prompt_raw
+        else:
+            raise ManifestError(f"agent '{name}' prompt must be a string (was {type(prompt_raw).__name__})")
+
+        # git-gate: agents may declare only `git-gate.user` (name/email).
+        # `git-gate.repos` is bottle-only — it carries credentials and host trust.
+        git_user = GitUser()
+        git_raw = d.get("git-gate")
+        if git_raw is not None:
+            gd = as_json_object(git_raw, f"agent '{name}' git-gate")
+            for k in gd.keys():
+                if k != "user":
+                    raise ManifestError(
+                        f"agent '{name}' git-gate.{k} is not allowed at the "
+                        f"agent level; only git-gate.user (name/email) may be "
+                        f"set on an agent. git-gate.repos is bottle-only "
+                        f"(it carries credentials and host trust)."
+                    )
+            if "user" in gd:
+                git_user = GitUser.from_dict(name, gd["user"])
+
+        return cls(bottle=bottle, skills=skills, prompt=prompt, git_user=git_user)

bot_bottle/manifest_egress.py

@@ -0,0 +1,286 @@
+"""Egress routing manifest dataclasses and helpers."""
+
+from __future__ import annotations
+
+import ipaddress
+from dataclasses import dataclass, field
+from typing import cast
+
+from .manifest_util import ManifestError, as_json_object
+
+
+# Auth schemes for the egress route's optional `auth` block.
+# Same values cred-proxy accepts today; `token` sidesteps the Gitea
+# token-not-Bearer quirk (go-gitea/gitea#16734).
+EGRESS_AUTH_SCHEMES = ("Bearer", "token")
+
+
+def validate_egress_routes(
+    bottle_name: str,
+    routes: tuple[EgressRoute, ...],
+) -> None:
+    """Cross-validation for `bottle.egress.routes`: hosts must be unique.
+
+    The proxy matches by exact-host (v1); duplicate hosts leave the
+    route choice ambiguous so we reject them up front.
+
+    No cross-validation against `bottle.git-gate.repos` is performed.
+    git-gate (SSH push/fetch) and egress (HTTPS) broker different
+    protocols; declaring both for the same host is a legitimate dev
+    setup."""
+    seen_hosts: dict[str, None] = {}
+    for r in routes:
+        key = r.Host.lower()
+        if key in seen_hosts:
+            raise ManifestError(
+                f"bottle '{bottle_name}' egress.routes has duplicate host "
+                f"{r.Host!r}; each host must be unique on the proxy."
+            )
+        seen_hosts[key] = None
+
+
+@dataclass(frozen=True)
+class PipelockRoutePolicy:
+    """Per-route pipelock policy overrides.
+
+    `TlsPassthrough` adds the route host to pipelock's
+    `tls_interception.passthrough_domains`, so pipelock still enforces
+    the hostname allowlist but does not MITM/decrypt request bodies or
+    headers for that host.
+
+    `SsrfIpAllowlist` adds explicit IPs/CIDRs to pipelock's SSRF
+    allowlist for private/internal destinations behind this route.
+    """
+
+    TlsPassthrough: bool = False
+    SsrfIpAllowlist: tuple[str, ...] = ()
+
+    @classmethod
+    def from_dict(
+        cls, bottle_name: str, idx: int, raw: object,
+    ) -> "PipelockRoutePolicy":
+        label = f"bottle '{bottle_name}' egress.routes[{idx}] pipelock"
+        d = as_json_object(raw, label)
+        for k in d:
+            if k not in ("tls_passthrough", "ssrf_ip_allowlist"):
+                raise ManifestError(
+                    f"{label} has unknown key {k!r}; "
+                    f"only 'tls_passthrough' and 'ssrf_ip_allowlist' "
+                    f"are accepted"
+                )
+        tls_passthrough_raw = d.get("tls_passthrough", False)
+        if not isinstance(tls_passthrough_raw, bool):
+            raise ManifestError(
+                f"{label}.tls_passthrough must be a boolean "
+                f"(was {type(tls_passthrough_raw).__name__})"
+            )
+        ssrf_raw = d.get("ssrf_ip_allowlist", [])
+        if not isinstance(ssrf_raw, list):
+            raise ManifestError(
+                f"{label}.ssrf_ip_allowlist must be an array "
+                f"(was {type(ssrf_raw).__name__})"
+            )
+        ssrf_ip_allowlist: list[str] = []
+        for j, item in enumerate(ssrf_raw):
+            if not isinstance(item, str) or not item:
+                raise ManifestError(
+                    f"{label}.ssrf_ip_allowlist[{j}] must be a non-empty "
+                    f"string (was {type(item).__name__})"
+                )
+            try:
+                ipaddress.ip_network(item, strict=False)
+            except ValueError as e:
+                raise ManifestError(
+                    f"{label}.ssrf_ip_allowlist[{j}] must be an IP address "
+                    f"or CIDR (was {item!r}): {e}"
+                )
+            ssrf_ip_allowlist.append(item)
+        return cls(
+            TlsPassthrough=tls_passthrough_raw,
+            SsrfIpAllowlist=tuple(ssrf_ip_allowlist),
+        )
+
+
+@dataclass(frozen=True)
+class EgressRoute:
+    """One route on the per-bottle egress sidecar (PRD 0017).
+
+    `Host` matches the request's hostname (case-insensitive). The
+    optional `PathAllowlist` constrains the URL path to a set of
+    prefixes; empty tuple means no path-level filtering. The optional
+    `AuthScheme` / `TokenRef` pair drives credential injection:
+    when set, the proxy strips any inbound Authorization and injects
+    `<AuthScheme> <value-of-host-env-named-by-TokenRef>`. When the
+    manifest's `auth` block is omitted both fields are empty strings —
+    no Authorization is written, no token forwarded.
+
+    `Role` is reserved for future use; all role strings are currently
+    rejected by the validator.
+
+    Validation rules (enforced in `from_dict`):
+      - `host` required, non-empty.
+      - `path_allowlist` optional, list of absolute path prefixes.
+      - `auth` optional. If present, MUST carry both `scheme` and
+        `token_ref` as non-empty strings; an empty `auth: {}` is an
+        error rather than a synonym for "no auth" (omit `auth` for
+        that case).
+      - `role` optional, reserved — any non-empty value is rejected.
+    """
+
+    Host: str
+    PathAllowlist: tuple[str, ...] = ()
+    AuthScheme: str = ""
+    TokenRef: str = ""
+    Role: tuple[str, ...] = ()
+    Pipelock: PipelockRoutePolicy = field(default_factory=PipelockRoutePolicy)
+
+    @classmethod
+    def from_dict(cls, bottle_name: str, idx: int, raw: object) -> "EgressRoute":
+        label = f"bottle '{bottle_name}' egress.routes[{idx}]"
+        d = as_json_object(raw, label)
+        host = d.get("host")
+        if not isinstance(host, str) or not host:
+            raise ManifestError(f"{label} missing required string field 'host'")
+
+        path_allow_raw = d.get("path_allowlist")
+        prefixes: tuple[str, ...] = ()
+        if path_allow_raw is not None:
+            if not isinstance(path_allow_raw, list):
+                raise ManifestError(
+                    f"{label} path_allowlist must be an array "
+                    f"(was {type(path_allow_raw).__name__})"
+                )
+            path_list = cast(list[object], path_allow_raw)
+            collected: list[str] = []
+            for j, p in enumerate(path_list):
+                if not isinstance(p, str):
+                    raise ManifestError(
+                        f"{label} path_allowlist[{j}] must be a string "
+                        f"(was {type(p).__name__})"
+                    )
+                if not p.startswith("/"):
+                    raise ManifestError(
+                        f"{label} path_allowlist[{j}] {p!r} must be an "
+                        f"absolute path prefix starting with '/'"
+                    )
+                collected.append(p)
+            prefixes = tuple(collected)
+
+        auth_scheme = ""
+        token_ref = ""
+        if "auth" in d:
+            auth_raw = d.get("auth")
+            auth_d = as_json_object(auth_raw, f"{label} auth")
+            if not auth_d:
+                raise ManifestError(
+                    f"{label} auth is empty ({{}}); omit the 'auth' key "
+                    f"entirely if this route is unauthenticated. Otherwise "
+                    f"both 'scheme' and 'token_ref' are required."
+                )
+            auth_scheme_raw = auth_d.get("scheme")
+            if not isinstance(auth_scheme_raw, str) or not auth_scheme_raw:
+                raise ManifestError(
+                    f"{label} auth.scheme is required when 'auth' is set "
+                    f"(non-empty string)"
+                )
+            if auth_scheme_raw not in EGRESS_AUTH_SCHEMES:
+                raise ManifestError(
+                    f"{label} auth.scheme {auth_scheme_raw!r} is not one of "
+                    f"{', '.join(EGRESS_AUTH_SCHEMES)}"
+                )
+            token_ref_raw = auth_d.get("token_ref")
+            if not isinstance(token_ref_raw, str) or not token_ref_raw:
+                raise ManifestError(
+                    f"{label} auth.token_ref is required when 'auth' is set "
+                    f"(name of the host env var holding the token value)"
+                )
+            for k in auth_d:
+                if k not in ("scheme", "token_ref"):
+                    raise ManifestError(
+                        f"{label} auth has unknown key {k!r}; "
+                        f"only 'scheme' and 'token_ref' are accepted"
+                    )
+            auth_scheme = auth_scheme_raw
+            token_ref = token_ref_raw
+
+        role_raw = d.get("role")
+        roles: tuple[str, ...] = ()
+        if role_raw is None:
+            roles = ()
+        elif isinstance(role_raw, str):
+            roles = (role_raw,)
+        elif isinstance(role_raw, list):
+            role_list = cast(list[object], role_raw)
+            collected_roles: list[str] = []
+            for r in role_list:
+                if not isinstance(r, str):
+                    raise ManifestError(f"{label} role items must be strings (got {type(r).__name__})")
+                collected_roles.append(r)
+            roles = tuple(collected_roles)
+        else:
+            raise ManifestError(
+                f"{label} role must be a string or a list of strings "
+                f"(was {type(role_raw).__name__})"
+            )
+        if roles:
+            raise ManifestError(
+                f"{label} role {roles[0]!r} is not accepted; "
+                f"the 'role' field is reserved for future use"
+            )
+
+        pipelock = (
+            PipelockRoutePolicy.from_dict(bottle_name, idx, d["pipelock"])
+            if "pipelock" in d
+            else PipelockRoutePolicy()
+        )
+
+        for k in d:
+            if k not in ("host", "path_allowlist", "auth", "role", "pipelock"):
+                raise ManifestError(
+                    f"{label} has unknown key {k!r}; accepted keys are "
+                    f"'host', 'path_allowlist', 'auth', 'role', 'pipelock'"
+                )
+
+        return cls(
+            Host=host,
+            PathAllowlist=prefixes,
+            AuthScheme=auth_scheme,
+            TokenRef=token_ref,
+            Role=roles,
+            Pipelock=pipelock,
+        )
+
+
+@dataclass(frozen=True)
+class EgressConfig:
+    """Per-bottle egress configuration. Today this is just the
+    route table; the nesting under `egress:` leaves room for
+    per-bottle proxy settings (port override, log level, etc.) in
+    follow-ups."""
+
+    routes: tuple[EgressRoute, ...] = ()
+
+    @classmethod
+    def from_dict(cls, bottle_name: str, raw: object) -> "EgressConfig":
+        d = as_json_object(raw, f"bottle '{bottle_name}' egress")
+        routes_raw = d.get("routes")
+        routes: tuple[EgressRoute, ...] = ()
+        if routes_raw is not None:
+            if not isinstance(routes_raw, list):
+                raise ManifestError(
+                    f"bottle '{bottle_name}' egress.routes must be an array "
+                    f"(was {type(routes_raw).__name__})"
+                )
+            routes_list = cast(list[object], routes_raw)
+            routes = tuple(
+                EgressRoute.from_dict(bottle_name, i, entry)
+                for i, entry in enumerate(routes_list)
+            )
+            validate_egress_routes(bottle_name, routes)
+        for k in d:
+            if k != "routes":
+                raise ManifestError(
+                    f"bottle '{bottle_name}' egress has unknown key {k!r}; "
+                    f"only 'routes' is accepted"
+                )
+        return cls(routes=routes)

bot_bottle/manifest_extends.py

@@ -0,0 +1,142 @@
+"""Internal bottle `extends:` resolution for manifests."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from .manifest import Bottle, GitEntry
+
+
+def resolve_bottles(raws: dict[str, dict[str, object]]) -> dict[str, Bottle]:
+    """Apply `extends:` chains and return resolved Bottle objects."""
+    cache: dict[str, Bottle] = {}
+    for name in raws:
+        if name not in cache:
+            _resolve_one_bottle(name, raws, cache, ())
+    return cache
+
+
+def _resolve_one_bottle(
+    name: str,
+    raws: dict[str, dict[str, object]],
+    cache: dict[str, Bottle],
+    seen: tuple[str, ...],
+) -> Bottle:
+    from .manifest import Bottle, ManifestError
+
+    if name in cache:
+        return cache[name]
+    if name in seen:
+        chain = " -> ".join(seen + (name,))
+        raise ManifestError(f"bottle '{name}' is in an extends cycle: {chain}")
+    raw = raws[name]
+    parent_name_raw = raw.get("extends")
+    # Strip `extends:` before passing to Bottle.from_dict so it
+    # is not accidentally treated as a real Bottle field by future
+    # schema additions. It is only meaningful here.
+    child_raw = {k: v for k, v in raw.items() if k != "extends"}
+
+    if parent_name_raw is None:
+        bottle = Bottle.from_dict(name, child_raw)
+        cache[name] = bottle
+        return bottle
+
+    if not isinstance(parent_name_raw, str):
+        raise ManifestError(
+            f"bottle '{name}' extends must be a string "
+            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)
+    cache[name] = bottle
+    return bottle
+
+
+def _merge_bottles(
+    parent: Bottle,
+    child_raw: dict[str, object],
+    name: str,
+) -> Bottle:
+    """Apply PRD 0025 merge rules."""
+    from .manifest import Bottle, GitUser
+    from .manifest_egress import validate_egress_routes
+
+    # Parse the child's declared fields into a Bottle (with the
+    # usual defaults for anything missing). Validation runs the same
+    # way it would for a leaf bottle: typos / wrong types die here.
+    child = Bottle.from_dict(name, child_raw)
+
+    # env: dict merge, child wins on collision.
+    merged_env = {**parent.env, **child.env}
+
+    # git-gate.user: per-field overlay. Each non-empty field on child
+    # wins; empties fall through to parent. The default GitUser()
+    # is two empty strings, so a child that omits git-gate.user
+    # inherits the parent's user verbatim.
+    merged_git_user = GitUser(
+        name=child.git_user.name or parent.git_user.name,
+        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.
+    if _child_declares_git_gate_repos(child_raw):
+        merged_git = _merge_git_remotes(parent.git, child.git) if child.git else ()
+    else:
+        merged_git = parent.git
+
+    # Presence-driven full-replace for the remaining list-valued +
+    # scalar fields.
+    merged_egress = child.egress if "egress" in child_raw else parent.egress
+    merged_agent_provider = (
+        child.agent_provider
+        if "agent_provider" in child_raw
+        else parent.agent_provider
+    )
+    merged_supervise = (
+        child.supervise if "supervise" in child_raw else parent.supervise
+    )
+    validate_egress_routes(name, merged_egress.routes)
+
+    return Bottle(
+        env=merged_env,
+        agent_provider=merged_agent_provider,
+        git=merged_git,
+        git_user=merged_git_user,
+        egress=merged_egress,
+        supervise=merged_supervise,
+    )
+
+
+def _child_declares_git_gate_repos(child_raw: dict[str, object]) -> bool:
+    from .manifest_util import as_json_object
+
+    git_raw = child_raw.get("git-gate")
+    if git_raw is None:
+        return False
+    git_obj = as_json_object(git_raw, "child git-gate")
+    return "repos" in git_obj
+
+
+def _merge_git_remotes(
+    parent: tuple[GitEntry, ...],
+    child: tuple[GitEntry, ...],
+) -> tuple[GitEntry, ...]:
+    by_host = {entry.UpstreamHost: entry for entry in parent}
+    for entry in child:
+        by_host[entry.UpstreamHost] = entry
+    return tuple(by_host.values())

bot_bottle/manifest_git.py

@@ -0,0 +1,301 @@
+"""Git-related manifest dataclasses and helpers."""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from typing import Optional
+
+from .manifest_util import ManifestError, as_json_object
+
+# Shell-safe characters for git-gate repo names. Names are embedded in
+# the generated entrypoint shell script (shlex.quote is the primary
+# defence; this regex is belt-and-suspenders and documents intent).
+_GIT_NAME_RE = re.compile(r"^[A-Za-z0-9._-]+$")
+
+
+def _opt_str(value: object, label: str) -> str:
+    if value is None:
+        return ""
+    if not isinstance(value, str):
+        raise ManifestError(f"{label} must be a string (was {type(value).__name__})")
+    return value
+
+
+def parse_git_upstream(url: str, label: str) -> tuple[str, str, str, str]:
+    """Parse `ssh://user@host[:port]/path` into (user, host, port, path).
+    Dies if `url` doesn't match the ssh:// shape v1 supports. Default
+    port is 22 (matches OpenSSH)."""
+    if not url.startswith("ssh://"):
+        raise ManifestError(f"{label} must be an ssh:// URL (was {url!r})")
+    rest = url[len("ssh://"):]
+    if "@" not in rest:
+        raise ManifestError(f"{label} must include a user (e.g. ssh://git@host/path.git); was {url!r}")
+    user, _, hostpart = rest.partition("@")
+    if not user:
+        raise ManifestError(f"{label} user is empty in {url!r}")
+    if "/" not in hostpart:
+        raise ManifestError(f"{label} must include a path (e.g. ssh://git@host/path.git); was {url!r}")
+    hostport, _, path = hostpart.partition("/")
+    if not path:
+        raise ManifestError(f"{label} path is empty in {url!r}")
+    if ":" in hostport:
+        host, _, port = hostport.partition(":")
+        if not port.isdigit():
+            raise ManifestError(f"{label} port must be numeric in {url!r}")
+    else:
+        host = hostport
+        port = "22"
+    if not host:
+        raise ManifestError(f"{label} host is empty in {url!r}")
+    return (user, host, port, path)
+
+
+def validate_unique_git_names(bottle_name: str, git: tuple[GitEntry, ...]) -> None:
+    seen: dict[str, None] = {}
+    for g in git:
+        if g.Name in seen:
+            raise ManifestError(
+                f"bottle '{bottle_name}' git-gate.repos has duplicate name '{g.Name}'; "
+                f"each entry maps to a distinct bare repo on the gate."
+            )
+        seen[g.Name] = None
+
+
+@dataclass(frozen=True)
+class ProvisionedKeyConfig:
+    """Configuration for automatic deploy-key lifecycle management
+    (PRD 0048). Used when a git-gate.repos entry opts out of a
+    static identity file and instead wants a fresh SSH keypair
+    generated at spin-up and revoked at teardown.
+
+    `provider` names the contrib sub-package to load (e.g. `gitea`).
+    `token_env` is the name of a host-side env var carrying the API
+    token; the value is read at provision time, never stored on the
+    plan. `api_url` is the forge's HTTP API root; if empty, it is
+    derived from the upstream URL's host at provision time."""
+
+    provider: str
+    token_env: str
+    api_url: str = ""
+
+
+@dataclass(frozen=True)
+class GitEntry:
+    """One upstream the per-agent git-gate (PRD 0008) is allowed to
+    talk to. `Upstream` is the real remote URL the agent would push to
+    if there were no gate; the gate hosts a bare repo at /git/<Name>.git
+    and `IdentityFile` is the SSH key the gate uses to push that repo
+    upstream after gitleaks passes. The agent itself never holds the
+    upstream credential.
+
+    The Upstream URL is parsed once at construction and the pieces are
+    stashed in the `Upstream*` fields so the git-gate render step
+    doesn't have to re-parse.
+
+    Manifest source: `git-gate.repos.<Name>` (PRD 0047/0048). Exactly
+    one of `identity` (static key path) or `provisioned_key` (automatic
+    lifecycle) must be present. The internal field names are stable."""
+
+    Name: str
+    Upstream: str
+    IdentityFile: str = ""
+    KnownHostKey: str = ""
+    ProvisionedKey: Optional[ProvisionedKeyConfig] = None
+    RemoteKey: str = ""
+    UpstreamUser: str = ""
+    UpstreamHost: str = ""
+    UpstreamPort: str = ""
+    UpstreamPath: str = ""
+
+    @classmethod
+    def from_repos_entry(
+        cls, bottle_name: str, repo_name: str, raw: object
+    ) -> "GitEntry":
+        """Parse one entry from `git-gate.repos.<repo_name>`.
+
+        YAML keys: `url` (required), exactly one of `identity` or
+        `provisioned_key` (required), `host_key` (optional).
+        The repo_name becomes `Name`."""
+        if not repo_name:
+            raise ManifestError(
+                f"bottle '{bottle_name}' git-gate.repos has an empty key"
+            )
+        if not _GIT_NAME_RE.match(repo_name):
+            raise ManifestError(
+                f"bottle '{bottle_name}' git-gate.repos name {repo_name!r} is invalid; "
+                f"allowed characters: A-Z a-z 0-9 . _ -"
+            )
+        label = f"git-gate.repos[{repo_name!r}]"
+        d = as_json_object(raw, f"bottle '{bottle_name}' {label}")
+        for k in d:
+            if k not in {"url", "identity", "provisioned_key", "host_key"}:
+                raise ManifestError(
+                    f"bottle '{bottle_name}' {label} has unknown key {k!r}; "
+                    f"allowed: url, identity, provisioned_key, host_key"
+                )
+        upstream = d.get("url")
+        if not isinstance(upstream, str) or not upstream:
+            raise ManifestError(
+                f"bottle '{bottle_name}' {label} missing required string field 'url'"
+            )
+
+        has_identity = "identity" in d
+        has_provisioned = "provisioned_key" in d
+        if has_identity and has_provisioned:
+            raise ManifestError(
+                f"bottle '{bottle_name}' {label} must set exactly one of "
+                f"'identity' or 'provisioned_key'; got both."
+            )
+        if not has_identity and not has_provisioned:
+            raise ManifestError(
+                f"bottle '{bottle_name}' {label} must set exactly one of "
+                f"'identity' or 'provisioned_key'; got neither."
+            )
+
+        ident = ""
+        provisioned_key: Optional[ProvisionedKeyConfig] = None
+        if has_identity:
+            raw_ident = d.get("identity")
+            if not isinstance(raw_ident, str) or not raw_ident:
+                raise ManifestError(
+                    f"bottle '{bottle_name}' {label} 'identity' must be a non-empty string"
+                )
+            ident = raw_ident
+        else:
+            provisioned_key = _parse_provisioned_key_config(
+                bottle_name, label, d["provisioned_key"]
+            )
+
+        khk = _opt_str(
+            d.get("host_key"),
+            f"bottle '{bottle_name}' {label} host_key",
+        )
+        user, host, port, path = parse_git_upstream(
+            upstream, f"bottle '{bottle_name}' {label} url"
+        )
+        return cls(
+            Name=repo_name,
+            Upstream=upstream,
+            IdentityFile=ident,
+            KnownHostKey=khk,
+            ProvisionedKey=provisioned_key,
+            RemoteKey=host,
+            UpstreamUser=user,
+            UpstreamHost=host,
+            UpstreamPort=port,
+            UpstreamPath=path,
+        )
+
+
+def _parse_provisioned_key_config(
+    bottle_name: str, label: str, raw: object
+) -> ProvisionedKeyConfig:
+    d = as_json_object(raw, f"bottle '{bottle_name}' {label}.provisioned_key")
+    for k in d:
+        if k not in {"provider", "token_env", "api_url"}:
+            raise ManifestError(
+                f"bottle '{bottle_name}' {label}.provisioned_key has unknown key {k!r}; "
+                f"allowed: provider, token_env, api_url"
+            )
+    provider = d.get("provider")
+    if not isinstance(provider, str) or not provider:
+        raise ManifestError(
+            f"bottle '{bottle_name}' {label}.provisioned_key missing required "
+            f"string field 'provider'"
+        )
+    token_env = d.get("token_env")
+    if not isinstance(token_env, str) or not token_env:
+        raise ManifestError(
+            f"bottle '{bottle_name}' {label}.provisioned_key missing required "
+            f"string field 'token_env'"
+        )
+    api_url_raw = d.get("api_url", "")
+    if not isinstance(api_url_raw, str):
+        raise ManifestError(
+            f"bottle '{bottle_name}' {label}.provisioned_key 'api_url' must be a string"
+        )
+    return ProvisionedKeyConfig(
+        provider=provider,
+        token_env=token_env,
+        api_url=api_url_raw,
+    )
+
+
+@dataclass(frozen=True)
+class GitUser:
+    """Per-bottle `git config --global user.name` / `user.email`
+    pair (issue #86). The agent's commits inside the bottle are
+    attributed to this identity rather than the agent image's
+    image-baked default (no user, or whatever the image dropped
+    in). Either or both fields can be set independently.
+
+    `from_dict` is forgiving on shape (a single missing field is
+    fine — we just skip that config line at provisioning) but
+    strict on types (string-or-die)."""
+
+    name: str = ""
+    email: str = ""
+
+    @classmethod
+    def from_dict(cls, bottle_name: str, raw: object) -> "GitUser":
+        d = as_json_object(raw, f"bottle '{bottle_name}' git-gate.user")
+        for k in d.keys():
+            if k not in {"name", "email"}:
+                raise ManifestError(
+                    f"bottle '{bottle_name}' git-gate.user has unknown key {k!r}; "
+                    f"allowed: name, email"
+                )
+        name = d.get("name", "")
+        email = d.get("email", "")
+        if not isinstance(name, str):
+            raise ManifestError(
+                f"bottle '{bottle_name}' git-gate.user.name must be a string "
+                f"(was {type(name).__name__})"
+            )
+        if not isinstance(email, str):
+            raise ManifestError(
+                f"bottle '{bottle_name}' git-gate.user.email must be a string "
+                f"(was {type(email).__name__})"
+            )
+        if not name and not email:
+            raise ManifestError(
+                f"bottle '{bottle_name}' git-gate.user is set but neither "
+                f"name nor email is non-empty; remove the block or "
+                f"fill at least one field."
+            )
+        return cls(name=name, email=email)
+
+    def is_empty(self) -> bool:
+        return not self.name and not self.email
+
+
+def parse_git_gate_config(
+    bottle_name: str,
+    raw: object,
+) -> tuple[tuple[GitEntry, ...], GitUser]:
+    d = as_json_object(raw, f"bottle '{bottle_name}' git-gate")
+    for k in d.keys():
+        if k not in {"user", "repos"}:
+            raise ManifestError(
+                f"bottle '{bottle_name}' git-gate has unknown key {k!r}; "
+                f"allowed: user, repos"
+            )
+
+    git_user = (
+        GitUser.from_dict(bottle_name, d["user"])
+        if "user" in d
+        else GitUser()
+    )
+
+    git: tuple[GitEntry, ...] = ()
+    repos_raw = d.get("repos")
+    if repos_raw is not None:
+        repos = as_json_object(repos_raw, f"bottle '{bottle_name}' git-gate.repos")
+        git = tuple(
+            GitEntry.from_repos_entry(bottle_name, name, entry)
+            for name, entry in repos.items()
+        )
+        validate_unique_git_names(bottle_name, git)
+
+    return git, git_user

bot_bottle/manifest_loader.py

@@ -0,0 +1,105 @@
+"""Internal per-file Markdown manifest loader."""
+
+from __future__ import annotations
+
+from pathlib import Path
+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 .yaml_subset import YamlSubsetError, parse_frontmatter
+
+if TYPE_CHECKING:
+    from .manifest import Agent, Bottle
+
+
+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(
+            f"found {legacy} but {md_dir} does not exist. The manifest "
+            f"format changed in PRD 0011 — rewrite the JSON content "
+            f"as per-file Markdown under {md_dir}/bottles/ and "
+            f"{md_dir}/agents/. See README.md for the schema. "
+            f"({label})"
+        )
+
+
+def load_bottles_from_dir(bottles_dir: Path) -> dict[str, Bottle]:
+    """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
+
+    raws: dict[str, dict[str, object]] = {}
+    if not bottles_dir.is_dir():
+        return {}
+    for path in sorted(bottles_dir.glob("*.md")):
+        name = entity_name_from_path(path)
+        if name is None:
+            warn(
+                f"skipping {path}: filename must match "
+                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}")
+        except YamlSubsetError as e:
+            raise ManifestError(f"{path}: {e}")
+        validate_bottle_frontmatter_keys(path, fm.keys())
+        raws[name] = fm
+    return resolve_bottles(raws)
+
+
+def load_agents_from_dir(
+    agents_dir: Path,
+    bottle_names: set[str],
+    *,
+    source: str,
+) -> dict[str, Agent]:
+    """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 Agent, ManifestError
+
+    out: dict[str, Agent] = {}
+    if not agents_dir.is_dir():
+        return out
+    for path in sorted(agents_dir.glob("*.md")):
+        name = entity_name_from_path(path)
+        if name is None:
+            warn(
+                f"skipping {path}: filename must match "
+                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}")
+        except YamlSubsetError as e:
+            raise ManifestError(f"{path}: {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] = Agent.from_dict(name, agent_dict, bottle_names)
+    return out

bot_bottle/manifest_schema.py

@@ -0,0 +1,70 @@
+"""Internal manifest schema policy helpers."""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+# Filename-as-key uses kebab-case ASCII. The first character is a
+# letter so we don't conflict with hidden files / Markdown special
+# names (`.md`, `_template.md`, etc.). Filenames that fail this
+# pattern are skipped with a warning rather than crashing the load.
+_FILENAME_RX = re.compile(r"^[a-z][a-z0-9-]*$")
+
+
+# Frontmatter keys we accept on each entity. Anything not in these
+# sets dies with a "did you mean" pointer: typos should not silently
+# ghost into an empty config.
+BOTTLE_KEYS = frozenset(
+    {"env", "extends", "agent_provider", "git-gate", "egress", "supervise"}
+)
+AGENT_KEYS_REQUIRED = frozenset({"bottle"})
+AGENT_KEYS_OPTIONAL = frozenset({"skills", "git-gate"})
+
+# Claude Code subagent fields bot-bottle ignores at launch but does
+# not reject. This lets the same file double as
+# `~/.claude/agents/*.md` without modification.
+CLAUDE_CODE_AGENT_PASSTHROUGH_KEYS = frozenset({
+    "name", "description", "model", "color", "memory",
+})
+AGENT_KEYS = (
+    AGENT_KEYS_REQUIRED | AGENT_KEYS_OPTIONAL | CLAUDE_CODE_AGENT_PASSTHROUGH_KEYS
+)
+AGENT_MODEL_KEYS = AGENT_KEYS | frozenset({"prompt"})
+
+
+def entity_name_from_path(path: Path) -> str | None:
+    """Return the entity name implied by the filename, or None if the
+    filename does not fit the [a-z][a-z0-9-]* convention."""
+    if path.suffix != ".md":
+        return None
+    stem = path.stem
+    if not _FILENAME_RX.match(stem):
+        return None
+    return stem
+
+
+def validate_bottle_frontmatter_keys(path: Path, keys: object) -> None:
+    _validate_frontmatter_keys("bottle", path, keys, BOTTLE_KEYS)
+
+
+def validate_agent_frontmatter_keys(path: Path, keys: object) -> None:
+    _validate_frontmatter_keys("agent", path, keys, AGENT_KEYS)
+
+
+def _validate_frontmatter_keys(
+    kind: str,
+    path: Path,
+    keys: object,
+    allowed_keys: frozenset[str],
+) -> None:
+    from .manifest_util import ManifestError
+
+    key_set = set(keys)
+    unknown = key_set - allowed_keys
+    if unknown:
+        allowed = ", ".join(sorted(allowed_keys))
+        raise ManifestError(
+            f"{kind} file {path}: unknown frontmatter key(s) "
+            f"{sorted(unknown)}; allowed keys are {allowed}."
+        )

bot_bottle/manifest_util.py

@@ -0,0 +1,24 @@
+"""Shared manifest primitives used by all manifest sub-modules."""
+
+from __future__ import annotations
+
+from typing import cast
+
+
+class ManifestError(Exception):
+    """A manifest file (or the manifest tree) is invalid."""
+
+
+def as_json_object(value: object, label: str) -> dict[str, object]:
+    """Assert that `value` is a JSON object (str-keyed dict) and return
+    a view typed as `dict[str, object]` so downstream `.get(...)` calls
+    have a typed surface."""
+    if not isinstance(value, dict):
+        raise ManifestError(f"{label} must be a JSON object (was {type(value).__name__})")
+    items = cast(dict[object, object], value)
+    out: dict[str, object] = {}
+    for k, v in items.items():
+        if not isinstance(k, str):
+            raise ManifestError(f"{label} keys must be strings (found {type(k).__name__})")
+        out[k] = v
+    return out

bot_bottle/pipelock.py

@@ -19,9 +19,8 @@ from __future__ import annotations
 
 from dataclasses import dataclass
 from pathlib import Path
-from typing import cast
 
-from .egress import EGRESS_HOSTNAME, egress_routes_for_bottle
+from .egress import EGRESS_HOSTNAME, EgressRoute, egress_routes_for_bottle
 from .supervise import SUPERVISE_HOSTNAME
 from .manifest import Bottle
 
@@ -50,14 +49,17 @@ PIPELOCK_HOSTNAME = "pipelock"
 # --- Allowlist resolution --------------------------------------------------
 
 
-def pipelock_effective_allowlist(bottle: Bottle) -> list[str]:
+def pipelock_effective_allowlist(
+    bottle: Bottle,
+    provider_routes: tuple[EgressRoute, ...] = (),
+) -> list[str]:
     """Hostnames pipelock allows. Sorted for stability.
 
-    Always mirrors `egress_routes_for_bottle(bottle)` — egress is the
-    single allowlist surface, and pipelock's allowlist is the downstream
-    copy for defense-in-depth + DLP body scanning. For bottles without
-    any `egress.routes[]` declared, this is empty except for supervise
-    sidecar traffic when `supervise: true`.
+    Always mirrors `egress_routes_for_bottle(bottle, provider_routes)` —
+    egress is the single allowlist surface, and pipelock's allowlist is
+    the downstream copy for defense-in-depth + DLP body scanning. For
+    bottles without any `egress.routes[]` declared, this is empty except
+    for supervise sidecar traffic when `supervise: true`.
 
     The supervise sidecar's hostname is auto-added when supervise
     is enabled (sibling-sidecar traffic that flows through pipelock
@@ -65,7 +67,7 @@ def pipelock_effective_allowlist(bottle: Bottle) -> list[str]:
     `bottle.git` do NOT contribute here — git traffic flows
     through git-gate (PRD 0008), not pipelock."""
     seen: dict[str, None] = {}
-    for r in egress_routes_for_bottle(bottle):
+    for r in egress_routes_for_bottle(bottle, provider_routes):
         if r.host:
             seen.setdefault(r.host, None)
     if bottle.supervise:
@@ -98,19 +100,23 @@ def pipelock_seed_phrase_detection_enabled(bottle: Bottle) -> bool:
     return False
 
 
-def pipelock_effective_tls_passthrough(bottle: Bottle) -> list[str]:
+def pipelock_effective_tls_passthrough(
+    bottle: Bottle,
+    provider_routes: tuple[EgressRoute, ...] = (),
+) -> list[str]:
     """Hostnames pipelock should pass through (no TLS MITM).
 
-    A route opts in with `pipelock.tls_passthrough: true`. This is
-    useful for provider API routes where egress injects the
-    Authorization header after the agent boundary; pipelock still
-    enforces the host allowlist but does not decrypt and scan that
-    provider request.
+    A manifest route opts in with `pipelock.tls_passthrough: true`
+    (lifted into `EgressRoute.tls_passthrough` in `egress_manifest_routes`).
+    Provider routes that set `tls_passthrough=True` (e.g. Codex credential
+    routes where egress injects the host bearer after the agent boundary)
+    are also included. Both arrive via `egress_routes_for_bottle` — no
+    provider-specific branching needed here.
     """
     seen: dict[str, None] = {host: None for host in DEFAULT_TLS_PASSTHROUGH}
-    for route in bottle.egress.routes:
-        if route.Pipelock.TlsPassthrough:
-            seen.setdefault(route.Host, None)
+    for route in egress_routes_for_bottle(bottle, provider_routes):
+        if route.tls_passthrough:
+            seen.setdefault(route.host, None)
     return sorted(seen.keys())
 
 
@@ -142,6 +148,7 @@ def pipelock_build_config(
     ca_cert_path: str = "",
     ca_key_path: str = "",
     ssrf_ip_allowlist: tuple[str, ...] = (),
+    provider_routes: tuple[EgressRoute, ...] = (),
 ) -> dict[str, object]:
     """Build the structured pipelock config dict the sidecar will load.
 
@@ -171,7 +178,7 @@ def pipelock_build_config(
         "version": 1,
         "mode": "strict",
         "enforce": True,
-        "api_allowlist": pipelock_effective_allowlist(bottle),
+        "api_allowlist": pipelock_effective_allowlist(bottle, provider_routes),
         "forward_proxy": {"enabled": True},
     }
     if not pipelock_seed_phrase_detection_enabled(bottle):
@@ -205,7 +212,7 @@ def pipelock_build_config(
             "enabled": True,
             "ca_cert": ca_cert_path,
             "ca_key": ca_key_path,
-            "passthrough_domains": pipelock_effective_tls_passthrough(bottle),
+            "passthrough_domains": pipelock_effective_tls_passthrough(bottle, provider_routes),
         }
     effective_ssrf_ip_allowlist = pipelock_effective_ssrf_ip_allowlist(
         bottle, ssrf_ip_allowlist,
@@ -215,6 +222,180 @@ def pipelock_build_config(
     return cfg
 
 
+_PIPELOCK_TOP_LEVEL_KEYS = {
+    "version",
+    "mode",
+    "enforce",
+    "api_allowlist",
+    "seed_phrase_detection",
+    "forward_proxy",
+    "dlp",
+    "request_body_scanning",
+    "tls_interception",
+    "ssrf",
+}
+
+
+def _pipelock_render_error(section: str, key: str, expected: str) -> ValueError:
+    return ValueError(
+        f"pipelock_render_yaml: {section}.{key} must be {expected}"
+    )
+
+
+def _reject_unknown_keys(
+    section: str,
+    obj: dict[str, object],
+    allowed: set[str],
+) -> None:
+    for key in sorted(set(obj) - allowed):
+        raise ValueError(f"pipelock_render_yaml: {section}.{key} is unsupported")
+
+
+def _required_dict(
+    obj: dict[str, object],
+    section: str,
+    key: str,
+) -> dict[str, object]:
+    value = obj.get(key)
+    if not isinstance(value, dict):
+        raise _pipelock_render_error(section, key, "a mapping")
+    return value
+
+
+def _required_bool(obj: dict[str, object], section: str, key: str) -> bool:
+    value = obj.get(key)
+    if not isinstance(value, bool):
+        raise _pipelock_render_error(section, key, "a boolean")
+    return value
+
+
+def _required_int(obj: dict[str, object], section: str, key: str) -> int:
+    value = obj.get(key)
+    if isinstance(value, bool) or not isinstance(value, int):
+        raise _pipelock_render_error(section, key, "an integer")
+    return value
+
+
+def _required_str(obj: dict[str, object], section: str, key: str) -> str:
+    value = obj.get(key)
+    if not isinstance(value, str):
+        raise _pipelock_render_error(section, key, "a string")
+    return value
+
+
+def _required_str_list(
+    obj: dict[str, object],
+    section: str,
+    key: str,
+) -> list[str]:
+    value = obj.get(key)
+    if not isinstance(value, list) or not all(isinstance(v, str) for v in value):
+        raise _pipelock_render_error(section, key, "a list of strings")
+    return value
+
+
+def _optional_str_list(
+    obj: dict[str, object],
+    section: str,
+    key: str,
+) -> list[str]:
+    if key not in obj:
+        return []
+    return _required_str_list(obj, section, key)
+
+
+def _optional_bool(
+    obj: dict[str, object],
+    section: str,
+    key: str,
+) -> bool | None:
+    if key not in obj:
+        return None
+    return _required_bool(obj, section, key)
+
+
+def _optional_str(
+    obj: dict[str, object],
+    section: str,
+    key: str,
+) -> str | None:
+    if key not in obj:
+        return None
+    return _required_str(obj, section, key)
+
+
+def _validate_pipelock_render_config(cfg: dict[str, object]) -> dict[str, object]:
+    _reject_unknown_keys("config", cfg, _PIPELOCK_TOP_LEVEL_KEYS)
+    normalized: dict[str, object] = {
+        "version": _required_int(cfg, "config", "version"),
+        "mode": _required_str(cfg, "config", "mode"),
+        "enforce": _required_bool(cfg, "config", "enforce"),
+        "api_allowlist": _required_str_list(cfg, "config", "api_allowlist"),
+    }
+
+    if "seed_phrase_detection" in cfg:
+        spd = _required_dict(cfg, "config", "seed_phrase_detection")
+        _reject_unknown_keys("seed_phrase_detection", spd, {"enabled"})
+        normalized["seed_phrase_detection"] = {
+            "enabled": _required_bool(spd, "seed_phrase_detection", "enabled"),
+        }
+
+    fp = _required_dict(cfg, "config", "forward_proxy")
+    _reject_unknown_keys("forward_proxy", fp, {"enabled"})
+    normalized["forward_proxy"] = {
+        "enabled": _required_bool(fp, "forward_proxy", "enabled"),
+    }
+
+    dlp = _required_dict(cfg, "config", "dlp")
+    _reject_unknown_keys("dlp", dlp, {"include_defaults", "scan_env"})
+    normalized["dlp"] = {
+        "include_defaults": _required_bool(dlp, "dlp", "include_defaults"),
+        "scan_env": _required_bool(dlp, "dlp", "scan_env"),
+    }
+
+    rbs = _required_dict(cfg, "config", "request_body_scanning")
+    _reject_unknown_keys(
+        "request_body_scanning",
+        rbs,
+        {"action", "scan_headers", "header_mode"},
+    )
+    normalized_rbs: dict[str, object] = {
+        "action": _required_str(rbs, "request_body_scanning", "action"),
+    }
+    scan_headers = _optional_bool(rbs, "request_body_scanning", "scan_headers")
+    if scan_headers is not None:
+        normalized_rbs["scan_headers"] = scan_headers
+    header_mode = _optional_str(rbs, "request_body_scanning", "header_mode")
+    if header_mode is not None:
+        normalized_rbs["header_mode"] = header_mode
+    normalized["request_body_scanning"] = normalized_rbs
+
+    if "tls_interception" in cfg:
+        tls = _required_dict(cfg, "config", "tls_interception")
+        _reject_unknown_keys(
+            "tls_interception",
+            tls,
+            {"enabled", "ca_cert", "ca_key", "passthrough_domains"},
+        )
+        normalized["tls_interception"] = {
+            "enabled": _required_bool(tls, "tls_interception", "enabled"),
+            "ca_cert": _required_str(tls, "tls_interception", "ca_cert"),
+            "ca_key": _required_str(tls, "tls_interception", "ca_key"),
+            "passthrough_domains": _optional_str_list(
+                tls, "tls_interception", "passthrough_domains",
+            ),
+        }
+
+    if "ssrf" in cfg:
+        ssrf = _required_dict(cfg, "config", "ssrf")
+        _reject_unknown_keys("ssrf", ssrf, {"ip_allowlist"})
+        normalized["ssrf"] = {
+            "ip_allowlist": _required_str_list(ssrf, "ssrf", "ip_allowlist"),
+        }
+
+    return normalized
+
+
 def pipelock_render_yaml(cfg: dict[str, object]) -> str:
     """Render a pipelock config dict (as produced by
     `pipelock_build_config`) as YAML. Hand-rolled so we don't take a
@@ -222,31 +403,38 @@ def pipelock_render_yaml(cfg: dict[str, object]) -> str:
     def _bool(b: object) -> str:
         return "true" if b else "false"
 
+    cfg = _validate_pipelock_render_config(cfg)
     lines: list[str] = []
     lines.append(f"version: {cfg['version']}")
     lines.append(f"mode: {cfg['mode']}")
     lines.append(f"enforce: {_bool(cfg['enforce'])}")
     lines.append("")
     lines.append("api_allowlist:")
-    for h in cast(list[str], cfg["api_allowlist"]):
+    api_allowlist = cfg["api_allowlist"]
+    assert isinstance(api_allowlist, list)
+    for h in api_allowlist:
         lines.append(f'  - "{h}"')
     lines.append("")
     if "seed_phrase_detection" in cfg:
         lines.append("seed_phrase_detection:")
-        spd = cast(dict[str, object], cfg["seed_phrase_detection"])
+        spd = cfg["seed_phrase_detection"]
+        assert isinstance(spd, dict)
         lines.append(f"  enabled: {_bool(spd['enabled'])}")
         lines.append("")
     lines.append("forward_proxy:")
-    fp = cast(dict[str, object], cfg["forward_proxy"])
+    fp = cfg["forward_proxy"]
+    assert isinstance(fp, dict)
     lines.append(f"  enabled: {_bool(fp['enabled'])}")
     lines.append("")
     lines.append("dlp:")
-    dlp = cast(dict[str, object], cfg["dlp"])
+    dlp = cfg["dlp"]
+    assert isinstance(dlp, dict)
     lines.append(f"  include_defaults: {_bool(dlp['include_defaults'])}")
     lines.append(f"  scan_env: {_bool(dlp['scan_env'])}")
     lines.append("")
     lines.append("request_body_scanning:")
-    rbs = cast(dict[str, object], cfg["request_body_scanning"])
+    rbs = cfg["request_body_scanning"]
+    assert isinstance(rbs, dict)
     lines.append(f'  action: "{rbs["action"]}"')
     if "scan_headers" in rbs:
         lines.append(f"  scan_headers: {_bool(rbs['scan_headers'])}")
@@ -255,11 +443,13 @@ def pipelock_render_yaml(cfg: dict[str, object]) -> str:
     if "tls_interception" in cfg:
         lines.append("")
         lines.append("tls_interception:")
-        tls = cast(dict[str, object], cfg["tls_interception"])
+        tls = cfg["tls_interception"]
+        assert isinstance(tls, dict)
         lines.append(f"  enabled: {_bool(tls['enabled'])}")
         lines.append(f'  ca_cert: "{tls["ca_cert"]}"')
         lines.append(f'  ca_key: "{tls["ca_key"]}"')
-        passthrough = cast(list[str], tls.get("passthrough_domains", []))
+        passthrough = tls["passthrough_domains"]
+        assert isinstance(passthrough, list)
         if passthrough:
             lines.append("  passthrough_domains:")
             for d in passthrough:
@@ -267,9 +457,12 @@ def pipelock_render_yaml(cfg: dict[str, object]) -> str:
     if "ssrf" in cfg:
         lines.append("")
         lines.append("ssrf:")
-        ssrf = cast(dict[str, object], cfg["ssrf"])
+        ssrf = cfg["ssrf"]
+        assert isinstance(ssrf, dict)
         lines.append("  ip_allowlist:")
-        for ip in cast(list[str], ssrf["ip_allowlist"]):
+        ip_allowlist = ssrf["ip_allowlist"]
+        assert isinstance(ip_allowlist, list)
+        for ip in ip_allowlist:
             lines.append(f'    - "{ip}"')
     return "\n".join(lines) + "\n"
 
@@ -319,7 +512,11 @@ class PipelockProxy:
     (`PIPELOCK_CA_CERT_IN_CONTAINER` / `PIPELOCK_CA_KEY_IN_CONTAINER`)."""
 
     def prepare(
-        self, bottle: Bottle, slug: str, stage_dir: Path
+        self,
+        bottle: Bottle,
+        slug: str,
+        stage_dir: Path,
+        provider_routes: tuple[EgressRoute, ...] = (),
     ) -> PipelockProxyPlan:
         """Write the pipelock yaml config (mode 600) under `stage_dir`
         and return the plan for launch. Pure host-side, no docker
@@ -342,6 +539,7 @@ class PipelockProxy:
             bottle,
             ca_cert_path=PIPELOCK_CA_CERT_IN_CONTAINER,
             ca_key_path=PIPELOCK_CA_KEY_IN_CONTAINER,
+            provider_routes=provider_routes,
         )
         yaml_path.write_text(pipelock_render_yaml(cfg))
         yaml_path.chmod(0o600)

bot_bottle/sidecar_init.py

@@ -20,7 +20,7 @@ sick daemon."
 
 Daemon subset is env-driven. The compose renderer narrows it via
 `BOT_BOTTLE_SIDECAR_DAEMONS=egress,pipelock` for bottles that
-don't use git-gate or supervise. Default: all four.
+don't use git-gate or supervise. Default: all daemons.
 
 Stdlib-only by design — adding supervisord/s6/runit for four
 daemons is heavier than this script.
@@ -98,6 +98,7 @@ _DAEMONS: tuple[_DaemonSpec, ...] = (
          "--listen", "0.0.0.0:8888"),
     ),
     _DaemonSpec("git-gate", ("/bin/sh", "/git-gate-entrypoint.sh")),
+    _DaemonSpec("git-http", ("python3", "/app/git_http_backend.py")),
     _DaemonSpec("supervise", ("python3", "/app/supervise_server.py")),
 )
 
@@ -162,6 +163,10 @@ class _Supervisor:
         # Names of children that have been logged as having exited
         # so we only log each death once across watch-loop ticks.
         self._logged_dead: set[str] = set()
+        # Signal handlers add daemon names here and return quickly.
+        # The main watch loop drains the set, so repeated restart
+        # requests for one daemon coalesce into one restart.
+        self._restart_requested: set[str] = set()
 
     def start_all(self) -> None:
         for spec in self.specs:
@@ -172,6 +177,7 @@ class _Supervisor:
         if self.shutdown_at is not None:
             return
         self.shutdown_at = time.monotonic()
+        self._restart_requested.clear()
         _log(f"shutting down ({reason}); forwarding SIGTERM")
         for _, p in self.procs:
             if p.poll() is None:
@@ -180,6 +186,24 @@ class _Supervisor:
                 except ProcessLookupError:
                     pass
 
+    def request_restart(self, daemon_name: str) -> bool:
+        """Queue a daemon restart for the main loop to process.
+
+        Signal handlers use this non-blocking path instead of doing
+        subprocess lifecycle work directly. Requests coalesce by
+        daemon name: one pending restart is enough to make the daemon
+        reread the latest config from disk.
+
+        Returns True iff a daemon by that name is known to the
+        supervisor and shutdown has not started."""
+        if self.shutdown_at is not None:
+            _log(f"restart {daemon_name} skipped; supervisor is shutting down")
+            return False
+        if not any(spec.name == daemon_name for spec, _ in self.procs):
+            return False
+        self._restart_requested.add(daemon_name)
+        return True
+
     def tick(self) -> bool:
         """One iteration of the watch loop. Returns True when every
         child has exited and the supervisor can return.
@@ -187,6 +211,8 @@ class _Supervisor:
         A child dying unexpectedly is logged but does NOT initiate
         shutdown — see the module docstring's failure-policy
         section. Shutdown is signal-driven only."""
+        self._drain_restart_requests()
+
         for spec, p in self.procs:
             rc = p.poll()
             if rc is None or spec.name in self._logged_dead:
@@ -219,14 +245,37 @@ class _Supervisor:
                             except ProcessLookupError:
                                 pass
 
-        return all(p.poll() is not None for _, p in self.procs)
+        done = all(p.poll() is not None for _, p in self.procs)
+        if done:
+            for _, p in self.procs:
+                if p.stdout is not None:
+                    p.stdout.close()
+        return done
 
     def exit_code(self) -> int:
-        """Worst child returncode wins. On graceful shutdown every
-        child is signal-killed (negative returncode) and max()
-        returns 0; if some child crashed nonzero before the signal
-        the operator gets that code on container exit."""
-        return max((p.returncode for _, p in self.procs), default=0)
+        """Positive child failures win; otherwise report success.
+
+        Python represents signal-terminated children as negative
+        return codes. A signal-only graceful shutdown should not leak
+        that platform-specific detail into the container exit status,
+        but a positive crash before shutdown should remain visible."""
+        positives = [
+            p.returncode for _, p in self.procs
+            if p.returncode is not None and p.returncode > 0
+        ]
+        return max(positives, default=0)
+
+    def _drain_restart_requests(self) -> None:
+        if self.shutdown_at is not None:
+            self._restart_requested.clear()
+            return
+        requested = tuple(sorted(self._restart_requested))
+        self._restart_requested.clear()
+        for daemon_name in requested:
+            if self.shutdown_at is not None:
+                self._restart_requested.clear()
+                return
+            self.restart_daemon(daemon_name)
 
     def forward_signal(self, sig: int, daemon_name: str) -> bool:
         """Forward a signal to one named child. Used by the SIGHUP
@@ -291,6 +340,8 @@ class _Supervisor:
                     except ProcessLookupError:
                         pass
                     p.wait()
+            if p.stdout is not None:
+                p.stdout.close()
             self._logged_dead.discard(daemon_name)
             new_proc = _spawn(spec)
             self.procs[idx] = (spec, new_proc)
@@ -322,7 +373,7 @@ def main(argv: Sequence[str] | None = None) -> int:
     # supervisor restarts the pipelock daemon in place (other
     # daemons keep running — specifically supervise, whose MCP
     # socket would drop on a whole-container `docker restart`).
-    signal.signal(signal.SIGUSR1, lambda *_: sup.restart_daemon("pipelock"))
+    signal.signal(signal.SIGUSR1, lambda *_: sup.request_restart("pipelock"))
 
     while not sup.tick():
         time.sleep(_POLL_INTERVAL)

bot_bottle/supervise.py

@@ -12,8 +12,8 @@ agent calls when it hits a stuck-recovery category:
 Each tool call: the agent passes the full proposed file plus a
 justification text. The sidecar validates the proposal syntactically,
 writes it to the host's per-bottle queue dir, and holds the tool-call
-connection open. The operator's TUI dashboard
-(bot_bottle.cli.dashboard) sees the proposal, accepts
+connection open. The operator's supervise TUI
+(bot_bottle.cli.supervise) sees the proposal, accepts
 approve / modify / reject, and writes a response file alongside the
 proposal. The sidecar sees the response and returns `{status, notes}`
 to the agent.

bot_bottle/supervise_server.py

@@ -35,6 +35,7 @@ import json
 import os
 import socketserver
 import sys
+import time
 import typing
 import urllib.error
 import urllib.parse
@@ -63,6 +64,10 @@ ERR_METHOD_NOT_FOUND = -32601
 ERR_INVALID_PARAMS = -32602
 ERR_INTERNAL = -32603
 
+DEFAULT_RESPONSE_TIMEOUT_SECONDS = 30.0
+MIN_RESPONSE_POLL_INTERVAL_SECONDS = 0.05
+EGRESS_LIST_TIMEOUT_SECONDS = 5.0
+
 
 @dataclass(frozen=True)
 class JsonRpcRequest:
@@ -412,6 +417,7 @@ def _validate_and_bundle_egress_route(
 class ServerConfig:
     bottle_slug: str
     queue_dir: Path
+    response_timeout_seconds: float = DEFAULT_RESPONSE_TIMEOUT_SECONDS
 
 
 def handle_initialize(_params: dict[str, object]) -> dict[str, object]:
@@ -442,7 +448,7 @@ def handle_list_egress_routes(
     })
     opener = urllib.request.build_opener(proxy_handler)
     try:
-        with opener.open(_sv.EGRESS_INTROSPECT_URL, timeout=5) as resp:
+        with opener.open(_sv.EGRESS_INTROSPECT_URL, timeout=EGRESS_LIST_TIMEOUT_SECONDS) as resp:
             body = resp.read().decode("utf-8")
     except (urllib.error.URLError, OSError) as e:
         return {
@@ -520,7 +526,20 @@ def handle_tools_call(
         f"for bottle {config.bottle_slug}; waiting for operator...\n"
     )
     sys.stderr.flush()
-    response = _sv.wait_for_response(config.queue_dir, proposal.id)
+    deadline = time.monotonic() + config.response_timeout_seconds
+    try:
+        response = _sv.wait_for_response(
+            config.queue_dir,
+            proposal.id,
+            poll_interval=MIN_RESPONSE_POLL_INTERVAL_SECONDS,
+            deadline=deadline,
+        )
+    except TimeoutError:
+        text = format_pending_response_text(config.response_timeout_seconds)
+        return {
+            "content": [{"type": "text", "text": text}],
+            "isError": False,
+        }
     _sv.archive_proposal(config.queue_dir, proposal.id)
 
     text = format_response_text(response)
@@ -542,6 +561,16 @@ def format_response_text(response: "_sv.Response") -> str:
     return "\n".join(lines)
 
 
+def format_pending_response_text(timeout_seconds: float) -> str:
+    return "\n".join([
+        "status: pending",
+        (
+            "notes: operator response timed out after "
+            f"{timeout_seconds:g}s; proposal remains queued"
+        ),
+    ])
+
+
 # --- HTTP transport --------------------------------------------------------
 
 
@@ -654,10 +683,15 @@ def serve(
     queue_dir: Path,
     port: int = _sv.SUPERVISE_PORT,
     bind: str = "0.0.0.0",
+    response_timeout_seconds: float = DEFAULT_RESPONSE_TIMEOUT_SECONDS,
 ) -> typing.NoReturn:
     queue_dir.mkdir(parents=True, exist_ok=True)
     server = MCPServer((bind, port), MCPHandler)
-    server.config = ServerConfig(bottle_slug=bottle_slug, queue_dir=queue_dir)
+    server.config = ServerConfig(
+        bottle_slug=bottle_slug,
+        queue_dir=queue_dir,
+        response_timeout_seconds=response_timeout_seconds,
+    )
     sys.stderr.write(
         f"supervise listening on {bind}:{port}; "
         f"slug={bottle_slug!r}; queue={queue_dir}; "
@@ -682,9 +716,37 @@ def main(argv: list[str]) -> int:
     queue_dir = Path(os.environ.get("SUPERVISE_QUEUE_DIR", _sv.QUEUE_DIR_IN_CONTAINER))
     port = int(os.environ.get("SUPERVISE_PORT", str(_sv.SUPERVISE_PORT)))
     bind = os.environ.get("SUPERVISE_BIND", "0.0.0.0")
-    serve(bottle_slug=bottle_slug, queue_dir=queue_dir, port=port, bind=bind)
+    try:
+        response_timeout_seconds = _response_timeout_from_env(os.environ)
+    except ValueError as e:
+        sys.stderr.write(f"supervise: {e}\n")
+        return 2
+    serve(
+        bottle_slug=bottle_slug,
+        queue_dir=queue_dir,
+        port=port,
+        bind=bind,
+        response_timeout_seconds=response_timeout_seconds,
+    )
     return 0  # serve() does not return
 
 
+def _response_timeout_from_env(env: typing.Mapping[str, str]) -> float:
+    raw = env.get("SUPERVISE_RESPONSE_TIMEOUT_SECONDS", "").strip()
+    if not raw:
+        return DEFAULT_RESPONSE_TIMEOUT_SECONDS
+    try:
+        value = float(raw)
+    except ValueError as e:
+        raise ValueError(
+            "SUPERVISE_RESPONSE_TIMEOUT_SECONDS must be a positive number"
+        ) from e
+    if value <= 0:
+        raise ValueError(
+            "SUPERVISE_RESPONSE_TIMEOUT_SECONDS must be a positive number"
+        )
+    return value
+
+
 if __name__ == "__main__":
     raise SystemExit(main(sys.argv))

bot_bottle/util.py

@@ -5,9 +5,18 @@ level deeper, under their backend package."""
 
 from __future__ import annotations
 
+import ipaddress
 import os
 
 
+def is_ip_literal(value: str) -> bool:
+    try:
+        ipaddress.ip_address(value)
+    except ValueError:
+        return False
+    return True
+
+
 def expand_tilde(path: str) -> str:
     """Expand a leading '~' to $HOME. Leaves paths without a leading
     tilde unchanged. Falls back to the empty string if $HOME is unset

bot_bottle/workspace.py

@@ -0,0 +1,52 @@
+"""Backend-neutral plan for porting the operator workspace."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Protocol
+
+
+WORKSPACE_DIRNAME = "workspace"
+DEFAULT_WORKSPACE_OWNER = "node:node"
+DEFAULT_WORKSPACE_MODE = "755"
+
+
+class WorkspaceSpec(Protocol):
+    copy_cwd: bool
+    user_cwd: str
+
+
+@dataclass(frozen=True)
+class WorkspacePlan:
+    """Resolved workspace contract shared by all bottle backends."""
+
+    enabled: bool
+    host_path: Path
+    guest_home: str
+    guest_path: str
+    workdir: str
+    owner: str = DEFAULT_WORKSPACE_OWNER
+    mode: str = DEFAULT_WORKSPACE_MODE
+    copy_contents: bool = True
+    copy_git: bool = True
+    has_host_git_dir: bool = False
+
+
+def workspace_plan(spec: WorkspaceSpec, *, guest_home: str) -> WorkspacePlan:
+    """Resolve the in-bottle workspace path from CLI intent."""
+    host_path = Path(spec.user_cwd).expanduser()
+    if spec.copy_cwd:
+        guest_path = f"{guest_home.rstrip('/')}/{WORKSPACE_DIRNAME}"
+        workdir = guest_path
+    else:
+        guest_path = guest_home
+        workdir = guest_home
+    return WorkspacePlan(
+        enabled=spec.copy_cwd,
+        host_path=host_path,
+        guest_home=guest_home,
+        guest_path=guest_path,
+        workdir=workdir,
+        has_host_git_dir=(host_path / ".git").is_dir(),
+    )

docs/INDEX.md

@@ -1 +0,0 @@
-Research notes live in `research/`. Product requirement docs live in `prds/`.

docs/README.md

@@ -0,0 +1,18 @@
+# Docs
+
+How this project records what it builds and why — and a guide to
+picking the right document for what you're capturing.
+
+## When to write which document
+
+| Artifact | For |
+|---|---|
+| **PRD** (`docs/prds/`) | A feature: what to build, scope, success criteria. |
+| **Research note** (`docs/research/`) | A landscape/tradeoff investigation. |
+| **Decision record** (`docs/decisions/`) | A decision that isn't itself a feature — a policy, a convention, a "we will / won't do this," or a load-bearing choice made inside a larger PRD that deserves to be discoverable on its own. |
+
+A decision that's fully specified by a PRD doesn't need duplicating in
+a decision record. Write one when the *decision* would otherwise be
+buried in prose, lost in an issue thread, or have no in-repo home at
+all (small requests that don't merit a PRD; non-feature choices like
+merge strategy or a trust posture).

docs/decisions/0001-merge-prs-with-rebase.md

@@ -0,0 +1,47 @@
+# ADR 0001: Merge PRs with rebase, not merge commits
+
+- **Status:** Accepted
+- **Date:** 2026-05-28
+- **Deciders:** didericis
+
+## Context
+
+PRs need a merge strategy. Gitea offers merge-commit, squash, rebase,
+and rebase-merge. The project uses [Conventional
+Commits](https://www.conventionalcommits.org/) enforced by a
+`commit-msg` hook, and PRDs typically land as a multi-commit PR where
+each commit is meaningful on its own (e.g. PR #95: a `docs(prd)` commit,
+a `feat(manifest)` implementation commit, and a `docs(manifest)`
+commit). The history should stay readable and the individual
+conventional commits should survive onto `main`.
+
+## Decision
+
+Merge PRs with **rebase** (Gitea's `rebase` style; `Do: "rebase"` via
+the API). The branch's commits are replayed onto `main` with no merge
+commit, producing a linear history that preserves each commit verbatim.
+
+## Consequences
+
+- **Linear history**, no merge bubbles; `git log --oneline` reads as a
+  straight sequence of conventional commits.
+- **Each commit is preserved** (unlike squash, which would collapse the
+  PRD/impl/docs commits into one and lose the staged structure).
+- **Commit SHAs are rewritten at merge.** The replayed commits on `main`
+  get new SHAs, and the source branch is deleted, so a link to a file
+  by *branch name* (`/src/branch/<feature>/…`) dies at merge. This is
+  why links to not-yet-merged files are pinned to a **commit SHA**
+  (`/src/commit/<sha>/…`), which stays reachable via the retained
+  `refs/pull/<n>/head` ref. See
+  `docs/research/issue-tracking-vs-in-repo-decision-history.md`.
+- **Trade-off accepted:** without a merge commit, the "these commits
+  landed together as PR #N" grouping is not recorded in git itself — it
+  lives in forge state (the PR). That is a mild concession against the
+  keep-history-in-the-repo posture; the conventional-commit scopes and
+  PRD references in the messages keep changes traceable without it.
+
+## Links
+
+- `docs/research/issue-tracking-vs-in-repo-decision-history.md` — the
+  commit-pinning consequence above.
+- Observed practice: PRs #92, #93 merged with rebase; #95 to follow.

docs/decisions/0002-agent-identity-claimed-not-vouched.md

@@ -0,0 +1,48 @@
+# ADR 0002: Agent-set git identity is claimed, not vouched
+
+- **Status:** Accepted
+- **Date:** 2026-05-28
+- **Deciders:** didericis
+
+## Context
+
+PRD 0027 lifts `git.user` (name/email) to the agent layer, so an agent
+file may declare its own commit identity. Agent files can live in
+`$CWD/.bot-bottle/agents/` — i.e. they can be supplied by a cloned,
+less-trusted repository. That raises the question of whether a
+repo-supplied agent setting its own git identity is a security concern,
+and whether agent identity should be gated differently for `$CWD`
+agents than for `$HOME` agents.
+
+This record exists because the decision is a **trust posture** worth
+finding on its own, separate from the feature PRD that introduced it.
+The full analysis lives in PRD 0027; the decision is summarized here.
+
+## Decision
+
+Allow agents to set `git.user`, and treat an agent-declared identity as
+**claimed, not vouched**. No `$CWD`-vs-`$HOME` gating on the identity
+field. `git.remotes` stays bottle-only (home-only).
+
+## Consequences
+
+- A cloned repo's agent file can present any commit author name/email,
+  including one that reads like a real person's. This is accepted: git
+  authorship is **not a credential** (push auth is the bottle's remote
+  key/token), is **already forgeable** from inside the bottle at runtime
+  (`git config user.email …`), and was never a trust anchor.
+- If attribution integrity ever matters, the answer is commit
+  **signing** (SSH/GPG), not the author field — so this decision closes
+  no door that was open.
+- `git.remotes` is deliberately *not* lifted to the agent layer: it
+  carries credentials and host trust (IdentityFile, KnownHostKey) and
+  remains a bottle-only, home-only concern.
+- Revisit if a future change ever makes commit identity load-bearing
+  (e.g. enforced signing keyed on author), at which point gating
+  `$CWD`-supplied identities would matter.
+
+## Links
+
+- PRD 0027 (`docs/prds/0027-agent-git-user-identity.md`) — full trust
+  analysis and schema.
+- Issue #94, PR #95 — the feature this decision was made for.

docs/decisions/0003-system-prompts-stay-user-directed.md

@@ -0,0 +1,43 @@
+# ADR 0003: Keep agent system prompts user-directed, not auto-generated from config
+
+- **Status:** Accepted
+- **Date:** 2026-05-29
+- **Deciders:** didericis
+
+## Context
+
+A bottle already declares exactly what an agent can reach: egress routes
+(allowlisted hosts + auth) and git config (remotes + identity). We
+considered deriving an agent's system prompt — or a section of it —
+automatically from those configs, so an agent would be told up front
+what it has access to (e.g. "you can reach `gitea.dideric.is` over the
+git remote and its API"). The question surfaced while hand-writing that
+exact line into the `claude-implementer` prompt.
+
+## Decision
+
+System prompts stay **user-directed** — authored by the operator. We do
+not auto-generate prompt content from a bottle's egress / git config.
+
+## Consequences
+
+- The operator controls what the agent is *told* about its environment,
+  independently of what the bottle *grants*. Sometimes we may want to
+  withhold that information from the agent directly — keep the prompt
+  silent about an allowlisted host even though egress permits it.
+- The agent can still infer its access on its own (attempt a request,
+  read its env, `git remote -v`, the gitconfig), so auto-injection is a
+  convenience, not a capability the agent depends on.
+- Cost accepted: operators must restate access in the prompt when they
+  want the agent to know it (as we did for the Gitea instance), and the
+  prompt can drift from the config. That decoupling of "what the bottle
+  grants" from "what the agent is told" is the point.
+- Revisit if keeping prompts in sync with configs becomes a real pain.
+  An *opt-in* helper that emits a capability summary the operator
+  chooses to include would honor this decision; silent auto-injection
+  would not.
+
+## Links
+
+- ADR 0002 (`0002-agent-identity-claimed-not-vouched.md`) — related
+  agent-trust posture (what the agent is granted vs. what it can claim).

docs/decisions/README.md

@@ -0,0 +1,42 @@
+# Decision records
+
+Short, durable records of decisions — one file per decision. This is a
+lightweight [Architecture Decision Record](https://adr.github.io/)
+practice: capture *what was decided and why* in a versioned file so the
+reasoning lives in the clone, not in a Gitea issue thread or a chat log
+that disappears when the host does.
+
+See `docs/research/issue-tracking-vs-in-repo-decision-history.md` for
+the rationale behind keeping decision history in-repo, and
+[`docs/README.md`](../README.md) for when to write a decision record
+vs. a PRD or research note.
+
+## Format
+
+One Markdown file per decision, numbered sequentially and zero-padded
+(`0001-…`, `0002-…`), matching the PRD numbering style. Keep it short —
+the discipline is writing it down, not the ceremony.
+
+```markdown
+# ADR 0000: <short imperative title>
+
+- **Status:** Proposed | Accepted | Superseded by ADR NNNN
+- **Date:** YYYY-MM-DD
+- **Deciders:** <who>
+
+## Context
+What forced the decision; the constraints in play.
+
+## Decision
+What we decided, stated plainly.
+
+## Consequences
+What follows — the good, and the costs/trade-offs accepted.
+
+## Links
+PRDs, research notes, issues/PRs. Gitea links are convenience
+pointers; the reasoning above must stand without them.
+```
+
+The records are the index: `ls docs/decisions/` or skim the titles.
+No hand-maintained list to keep in sync.

docs/prds/0001-per-agent-egress-proxy-via-pipelock.md

@@ -1,6 +1,6 @@
 # PRD 0001: Per-agent egress proxy via pipelock
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-08
 

docs/prds/0002-test-pipeline-on-gitea-actions.md

@@ -1,6 +1,6 @@
 # PRD 0002: Test pipeline on Gitea Actions
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-08
 

docs/prds/0003-bottle-backend-abstraction.md

@@ -1,6 +1,6 @@
 # PRD 0003: Bottle Backend abstraction
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-10
 

docs/prds/0004-split-out-provisioners.md

@@ -1,6 +1,6 @@
 # PRD 0004: Split out provisioners
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-11
 

docs/prds/0006-pipelock-tls-interception.md

@@ -1,6 +1,6 @@
 # PRD 0006: pipelock native TLS interception
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-12
 

docs/prds/0008-git-gate.md

@@ -1,6 +1,6 @@
 # PRD 0008: Git gate
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-12
 
@@ -83,12 +83,7 @@ for a declared upstream:
 - **Manifest field.** `bottle.git` — a list of git remotes the
   bottle is allowed to talk to, each with the credential the gate
   uses to push upstream. The agent gets no parallel `bottle.ssh`
-  entry for those upstreams. Each entry may also carry an
-  `ExtraHosts: { hostname: ip }` map, surfaced to the gate as
-  `--add-host` so the gate can resolve upstreams whose public DNS
-  doesn't point at the reachable IP (e.g. Tailscale-only hosts).
-  The agent-side `insteadOf` rewrite keys off the original hostname,
-  so the manifest's `Upstream` URL stays human-readable.
+  entry for those upstreams.
 - **Agent-side URL rewrite.** Provisioner emits `~/.gitconfig`
   with `[url "<gate-url>"] insteadOf = <real-url>` so every git
   operation against the declared upstream (push, fetch, clone,

docs/prds/0009-remove-ssh-gate.md

@@ -1,6 +1,6 @@
 # PRD 0009: Remove ssh-gate and bottle.ssh
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-13
 
@@ -88,8 +88,7 @@ the unused path.
 - **Pipelock interaction.** Drop the SSH-derived branch from
   pipelock's `ssrf.ip_allowlist` build. With no `bottle.ssh`
   there is no per-upstream IP carve-out to render; git-gate
-  has its own egress network and pulls in upstream resolution
-  via `ExtraHosts` plus DNS.
+  has its own egress network.
 - **Tests.** Delete the ssh-gate unit + integration suites,
   the ssh fixtures in `tests/fixtures.py`, and the
   shadow-route assertions in `test_manifest_git.py`. Adjust

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

@@ -1,6 +1,6 @@
 # PRD 0011: Per-file Markdown manifest
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-24
 
@@ -274,8 +274,6 @@ git:
       Name: bot-bottle
       Upstream: ssh://git@gitea.dideric.is:30009/didericis/bot-bottle.git
       IdentityFile: ~/.ssh/gitea-delos-2.pem
-      ExtraHosts:
-        gitea.dideric.is: 100.78.141.42
       KnownHostKey: ssh-rsa AAAAB3...
 egress:
   allowlist:

docs/prds/0012-stuck-agent-recovery-flow.md

@@ -1,6 +1,6 @@
 # PRD 0012: Stuck-agent recovery flow
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-24
 
@@ -22,7 +22,7 @@ A real stuck agent recovers end-to-end in each of the three categories: a **cred
 
 - Live attach or in-place mutation of running containers. The whole design exists to avoid this.
 - Agent-to-agent communication. Re-stated from the project's existing non-goals; the recovery flow is human→agent only.
-- Auditing or forensic replay of agent runs. Git/forge history is the audit log; this PRD does not add a separate run log.
+- Auditing or forensic replay of agent runs. Git/Gitea history is the audit log; this PRD does not add a separate run log.
 - Reducing time-to-unstuck below some target. Faster than kill-and-restart is implicit, but no specific SLO is in scope.
 
 ## Stuck categories

docs/prds/0013-supervise-plane-foundation.md

@@ -1,6 +1,6 @@
 # PRD 0013: Supervise plane foundation
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-25
 - **Parent:** PRD 0012

docs/prds/0015-pipelock-block-remediation.md

@@ -1,6 +1,6 @@
 # PRD 0015: pipelock block remediation
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-25
 - **Parent:** PRD 0012

docs/prds/0016-capability-block-remediation.md

@@ -1,6 +1,6 @@
 # PRD 0016: capability block remediation
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-25
 - **Parent:** PRD 0012

docs/prds/0017-egress-proxy-via-mitmproxy.md

@@ -1,6 +1,6 @@
 # PRD 0017: Egress-proxy — universal MITM with path filtering + auth injection
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-25
 - **Supersedes:** the cred-proxy sidecar (PRD 0010) — hard cutover.

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

@@ -1,6 +1,6 @@
 # PRD 0018: One Compose project per bottle instance
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-25
 

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

@@ -1,6 +1,6 @@
 # PRD 0019: Active agents in the dashboard, agent-scoped edit verbs
 
-- **Status:** Draft
+- **Status:** Superseded by [PRD 0049](0049-strip-dashboard-to-supervisor-tui.md)
 - **Author:** didericis
 - **Created:** 2026-05-26
 

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

@@ -1,6 +1,6 @@
 # PRD 0020: Start and attach to agents from inside the dashboard
 
-- **Status:** Draft
+- **Status:** Superseded by [PRD 0049](0049-strip-dashboard-to-supervisor-tui.md)
 - **Author:** didericis
 - **Created:** 2026-05-26
 

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

@@ -1,6 +1,6 @@
 # PRD 0021: Dashboard as left tmux pane, selected agent as right pane
 
-- **Status:** Draft
+- **Status:** Superseded by [PRD 0049](0049-strip-dashboard-to-supervisor-tui.md)
 - **Author:** didericis
 - **Created:** 2026-05-26
 

docs/prds/0022-sandbox-escape-integration-test.md

@@ -1,6 +1,6 @@
 # PRD 0022: End-to-end sandbox-escape integration test
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-26
 

docs/prds/0023-smolmachines-backend.md

@@ -1,6 +1,6 @@
 # PRD 0023: smolmachines bottle backend
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-26
 

docs/prds/0024-consolidate-sidecar-bundle.md

@@ -1,6 +1,6 @@
 # PRD 0024: Consolidate per-bottle sidecars into a single bundle
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-26
 

docs/prds/0025-bottle-extends.md

@@ -1,6 +1,6 @@
 # PRD 0025: Bottle composition via `extends:`
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** didericis
 - **Created:** 2026-05-27
 - **Issue:** #88
@@ -39,6 +39,41 @@ trust boundary*: only `$HOME` bottles can declare it, only `$HOME`
 bottles can be its target. Cloned repos still cannot author
 bottle-equivalent config.
 
+## Alternatives considered
+
+The question raised in issue #88 was *where composition should live*.
+Three points in that design space, recorded here so the decision
+stands on its own without the issue thread:
+
+1. **Duplicate bottles (status quo).** Copy `dev.md` to `staging.md`
+   and edit. Zero new mechanism, but every shared field drifts: a
+   route added to `dev` is silently missing from `staging`. This is
+   the pain that prompted #88.
+
+2. **Agent-side `bottle_config:` override (the original #88
+   proposal).** Let an agent file carry an inline block that merges
+   over its referenced bottle. Ergonomically attractive — one file,
+   no second bottle — but it **breaks the trust boundary**: agent
+   files can come from `$CWD/.bot-bottle/agents/` in a cloned repo, so
+   a clone could redeclare egress routes, env mappings, and git
+   remotes — i.e. grant itself bottle-equivalent authority over
+   credentials and network egress. The home-only-bottle invariant
+   exists precisely to stop this.
+
+3. **Bottle-side `extends:` (chosen).** Move composition to the
+   bottle layer, where it inherits the home-only property for free:
+   only `$HOME` bottles can declare `extends:`, and only `$HOME`
+   bottles can be its target. Identical duplication relief to option
+   2, none of its trust erosion. The cost is that an override requires
+   a (home-owned) child bottle rather than an inline agent block —
+   which is the *point*: the override authority stays in `$HOME`.
+
+`extends:` wins because it solves the duplication pain entirely on the
+trusted side of the agent-vs-bottle boundary. (PRD 0027 later lifts a
+deliberately narrow, non-credential field — `git.user` — to the agent
+layer, on the separate reasoning that commit identity is not a
+capability; egress, credentials, and remotes stay bottle-only.)
+
 ## Goals / Success Criteria
 
 - Add `extends: <bottle-name>` to the bottle frontmatter schema.
@@ -58,9 +93,9 @@ bottle-equivalent config.
 
 ## Non-goals
 
-- **No agent-side `bottle_config:`.** That's the design issue #88
-  considered and weighed against; this PRD is the alternative
-  picked in the issue's design discussion. Don't reintroduce it.
+- **No agent-side `bottle_config:`.** Option 2 under "Alternatives
+  considered" — weighed and rejected on trust grounds. Don't
+  reintroduce it.
 - **No additive list merges** (e.g., `routes: append` keyword).
   The `extends:` design uses full-replace for list-valued fields
   (see "Merge rules"); if a use case shows up that genuinely
@@ -126,8 +161,7 @@ expectation. (Same model as shell `export` precedence.)
 `git.remotes` is also keyed, so it follows dict-style inheritance:
 children can override one host without restating every remote. The
 remote entry is replaced as a whole on host collision because
-`Upstream`, `IdentityFile`, `KnownHostKey`, and `ExtraHosts` are
-tightly coupled.
+`Upstream`, `IdentityFile`, and `KnownHostKey` are tightly coupled.
 
 The `git.user` dataclass-overlay (each non-empty field wins
 individually) is so a parent can declare `git.user.name` and a
@@ -167,7 +201,7 @@ Bottles continue to be loaded from `$HOME/.bot-bottle/bottles/`
 only (`Manifest.from_md_dirs` is unchanged). The `extends:` field
 references another file in that same directory. No cwd-readable
 file gains the ability to declare or modify bottle config — the
-attack surface from issue #88's comment thread stays closed.
+attack surface from option 2 ("Alternatives considered") stays closed.
 
 If a future change ever introduces cwd-loaded bottles, the
 `extends:` resolver should be gated to forbid a `$CWD` bottle

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

@@ -1,6 +1,6 @@
 # PRD 0026: Agent Provider Templates
 
-- **Status:** Draft
+- **Status:** Active
 - **Author:** codex
 - **Created:** 2026-05-28
 
@@ -86,7 +86,10 @@ agent_provider:
 
 ## Open questions
 
-- The initial Codex auth role is `codex_auth`; it provides a non-secret `OPENAI_API_KEY` placeholder to the agent while egress holds the real token.
+- `codex_auth` is retained as a placeholder marker for follow-up Codex
+  credential-injection work. The Codex template should not inject an
+  `OPENAI_API_KEY` placeholder; Codex bottles use device/ChatGPT login
+  state instead.
 - Existing state-folder transcript capture is Claude-specific and should remain gated to Claude until the follow-up state/transcript refactor.
 
 ## References

docs/prds/0027-agent-git-user-identity.md

@@ -0,0 +1,226 @@
+# PRD 0027: Agent-level git user identity
+
+- **Status:** Active
+- **Author:** didericis
+- **Created:** 2026-05-28
+- **Issue:** #94
+
+## Summary
+
+Let an **agent** file declare `git.user` (name / email). At launch
+the agent's `git.user` overlays the referenced bottle's `git.user`
+per-field (agent wins on non-empty fields), mirroring the
+`extends:` overlay from PRD 0025. `git.remotes` stays bottle-only.
+
+Solves the "I need a whole separate bottle just to change the commit
+name" coupling: today commit attribution — a purpose/presentation
+concern — can only be varied by authoring a new security boundary
+(a bottle). After this change, two agents (e.g. `implementer` and
+`reviewer`) can share one bottle and still commit under distinct
+identities.
+
+## Problem
+
+`git.user.name` / `git.user.email` is a bottle-only field. Agent
+frontmatter is validated against a strict allowlist (`_AGENT_KEYS`
+in `manifest.py`); a `git:` block in an agent file dies at parse
+with "unknown frontmatter key(s)". So the only way to give an agent
+a distinct commit identity is to give it a distinct *bottle*.
+
+That couples identity to the trust boundary. Bottles are home-only
+and define security (egress, credentials, provider). Forcing a new
+bottle per commit-name means either (a) duplicating a bottle just to
+flip one string, or (b) leaning on PRD 0025 `extends:` to make a
+near-clone whose only delta is `git.user`. Both are heavier than the
+concern deserves.
+
+## Why this is safe (trust analysis)
+
+Allowing agent files — which can live in `$CWD/.bot-bottle/agents/`
+and therefore be supplied by a cloned repo — to set `git.user` does
+**not** weaken the security model, because git author identity is
+not a credential or a capability:
+
+- **Push auth is separate.** Authentication to a remote is the
+  bottle's `git.remotes` IdentityFile / token. `user.name` /
+  `user.email` grant zero access.
+- **It's already forgeable.** The agent inside the bottle can run
+  `git config user.email <anything>` or `git commit --author=...`
+  at runtime regardless of the manifest. The manifest field is only
+  a *default*; allowing agents to set it adds no capability that
+  isn't already reachable one layer down.
+- **Authorship was never a trust anchor.** If attribution integrity
+  matters, that is a commit-*signing* concern (SSH/GPG), which a
+  name/email field cannot provide either way.
+
+The one residual is cosmetic impersonation (a cloned repo's agent
+file could set an identity that reads like a real person's). Commits
+still push under the bottle's credentials, and the author string was
+never vouched-for, so this is presentation, not escalation. We
+document that an agent identity is *claimed, not vouched*.
+
+`git.remotes` is explicitly **not** lifted to the agent layer — that
+block carries credentials and host trust (IdentityFile, KnownHostKey)
+and stays a bottle-only, home-only concern.
+
+## Goals / Success Criteria
+
+- Add an optional `git.user` block to the agent frontmatter schema
+  (name and/or email), reusing the existing `GitUser` validator.
+- An agent's declared `git.user` overlays the referenced bottle's
+  `git.user` per-field; each non-empty agent field wins, empties
+  fall through to the bottle. Identical overlay semantics to the
+  bottle `extends:` merge (PRD 0025).
+- `git.remotes` in an agent file is rejected at parse with a clear
+  "bottle-only" pointer.
+- The overlay is applied at `Manifest.bottle_for()` — the single
+  point both backends call to resolve an agent's bottle — so the
+  docker and smolmachines provisioners need no changes.
+- Existing agent files continue to parse identically — `git.user`
+  is opt-in.
+- **Identity provenance is surfaced.** The y/N preflight (both
+  backends) and `cli.py info <agent>` print the effective git
+  identity with a per-field `(agent)` / `(bottle)` annotation so the
+  operator can see which level each field came from.
+- **The example collapses to demonstrate the pattern.** The bundled
+  example replaces the identity-only-bottle shape with one shared
+  bottle + per-agent `git.user`, showing the intended usage.
+
+## Non-goals
+
+- **No agent-level `git.remotes`.** Credentials and host trust stay
+  bottle-only and home-only.
+- **No other agent-level bottle fields.** This PRD lifts `git.user`
+  and nothing else; agents do not gain `egress`, `env`,
+  `agent_provider`, etc. (that is the issue #88 design PRD 0025
+  deliberately rejected).
+- **No commit signing.** Attribution integrity via SSH/GPG is a
+  separate concern, out of scope here.
+- **No CWD-vs-HOME identity gating.** Because the field is
+  non-enforcing (see trust analysis), CWD and HOME agents are
+  treated the same. If a future change makes identity load-bearing,
+  revisit.
+
+## Design
+
+### Schema
+
+A new optional `git:` block on agent files, carrying only `user`:
+
+```yaml
+---
+bottle: claude-dev
+git:
+  user:
+    name: claude-implementer
+    email: eric+claude-implementer@dideric.is
+---
+
+You are a feature-implementation agent ...
+```
+
+- `git.user` accepts `name` and/or `email` (string-or-die; at least
+  one non-empty), validated by the **existing** `GitUser.from_dict`.
+- `git.remotes` (or any `git` key other than `user`) in an agent
+  file dies at parse: "git.remotes is bottle-only".
+- `git` is added to `_AGENT_KEYS`; the agent md-loader threads the
+  raw `git` block into the dict `Agent.from_dict` consumes.
+
+### Merge rule
+
+When `Manifest.bottle_for(agent)` resolves the agent's bottle, it
+returns the bottle with `git_user` overlaid by the agent's
+`git_user`:
+
+| Field             | Merge                                          |
+|-------------------|------------------------------------------------|
+| `git_user.name`   | agent's if non-empty, else bottle's            |
+| `git_user.email`  | agent's if non-empty, else bottle's            |
+
+This is the same per-field overlay `_merge_bottles` already applies
+for `extends:` (`manifest.py:1212`). Empty string = "not set", the
+same predicate the provisioner's `is_empty()` uses. All other bottle
+fields are returned unchanged.
+
+### Where the overlay lives
+
+```
+def bottle_for(self, agent_name) -> Bottle:
+    agent = self.agents[agent_name]
+    bottle = self.bottles[agent.bottle]
+    if agent.git_user.is_empty():
+        return bottle
+    merged = GitUser(
+        name=agent.git_user.name or bottle.git_user.name,
+        email=agent.git_user.email or bottle.git_user.email,
+    )
+    return dataclasses.replace(bottle, git_user=merged)
+```
+
+Both `provision/git.py` paths (docker + smolmachines) already call
+`bottle_for(agent_name)` and read `bottle.git_user`, so they pick up
+the merged identity with no edits. The dashboard / `info` / preflight
+surfaces that print bottle config go through the same resolution.
+
+### Provenance display
+
+A sibling `Manifest.git_identity_summary(agent_name) -> str | None`
+returns the effective identity annotated per field, e.g.:
+
+```
+identity        : name=claude-implementer (agent), email=eric+claude-implementer@dideric.is (bottle)
+```
+
+`(agent)` when the agent's `git.user` supplied that field, `(bottle)`
+when it fell through to the referenced bottle. Returns `None` when no
+identity is set at either level (callers omit the line). Both
+`bottle_plan.py` preflights and `cli/info.py` print it; `info` today
+prints `git remotes` but not the identity, so this adds the line.
+
+### `Agent` dataclass
+
+`Agent` gains `git_user: GitUser = GitUser()`. `Agent.from_dict`
+parses the optional `git.user` and rejects non-`user` git keys.
+
+## Implementation chunks
+
+1. **PRD (this commit).** Sets the design.
+2. **Schema + overlay + tests.**
+   - Add `"git"` to `_AGENT_KEYS`.
+   - Add `git_user: GitUser = GitUser()` to `Agent`; parse it in
+     `Agent.from_dict`, reusing `GitUser.from_dict`; reject any
+     `git` key other than `user` with a bottle-only message.
+   - Thread `fm.get("git")` through `_load_agents_from_dir`'s
+     `agent_dict`.
+   - Apply the per-field overlay in `Manifest.bottle_for()`.
+   - Add `Manifest.git_identity_summary()` and print it in both
+     `bottle_plan.py` preflights and `cli/info.py`.
+   - Unit tests: agent name+email overlay; agent name-only (email
+     falls through to bottle); agent email-only; agent identity with
+     a bottle that declares none; agent `git.remotes` dies; agent
+     with no `git` block unchanged; bottle-only behavior preserved;
+     provenance summary returns the right `(agent)`/`(bottle)` tags
+     and `None` when unset.
+3. **Docs + example.** README manifest section: note `git.user` is
+   allowed on agents and overlays the bottle; `git.remotes` stays
+   bottle-only. Collapse the example — `examples/agents/implementer.md`
+   carries its own `git.user` against the shared `dev` bottle,
+   demonstrating per-agent identity without an identity-only bottle.
+
+## Testing strategy
+
+- **Unit (must):** the overlay matrix above, the parse-time reject
+  for agent `git.remotes`, and confirmation that `bottle_for` leaves
+  every non-`git_user` bottle field untouched.
+- **No integration changes needed:** provisioners consume the
+  already-merged `Bottle` via `bottle_for`; existing docker /
+  smolmachines git-provisioning tests cover the consumption path.
+
+## Open questions
+
+- **Should a `git.user`-declaring agent be loadable as a Claude Code
+  subagent file too?** The CC-passthrough keys (`name`, `model`, …)
+  let one file double as `~/.claude/agents/*.md`; `git` is not a CC
+  key, so a file carrying it won't round-trip as a CC subagent. Not a
+  blocker (bot-bottle ignores unknown CC keys; CC ignores `git`), but
+  worth noting. Out of scope.