docs(prds): expand PRD 0008 to bidirectional mirror scope
test / unit (pull_request) Successful in 11s
test / integration (pull_request) Successful in 18s

The gate now fronts every git operation, not just push. Fetch
(clone, pull, ls-remote) is mirrored via git daemon's
--access-hook running 'git fetch origin --prune' against the
real upstream before each upload-pack; fail-closed if upstream
is unreachable so the agent never serves stale data.

Push path is unchanged in concept (gitleaks gate → forward) but
the hook now pushes to 'origin' rather than 'upstream', matching
the remote name the entrypoint configures.
This commit is contained in:
2026-05-12 21:26:19 -04:00
parent bea433015f
commit ae7e22065f
+69 -39
View File
@@ -6,11 +6,18 @@
## Summary
Per-bottle sidecar that fronts the agent's git remotes, runs
gitleaks against incoming refs via a `pre-receive` hook, and only
forwards to the real upstream on a clean scan. Upstream push
credentials live in the gate, not the agent — so a misbehaving
agent cannot push a secret-bearing commit past it.
Per-bottle sidecar that fronts the agent's git remotes as a
transparent mirror. Push is gated: gitleaks scans incoming refs
via a `pre-receive` hook, and only clean refs get forwarded to
the real upstream. Fetch is mirrored: every `upload-pack` first
runs `git fetch origin --prune` against the upstream via the
daemon's `--access-hook`, so an agent fetch returns whatever the
upstream has *now* (fail-closed if the upstream is unreachable).
Upstream credentials live in the gate, not the agent — so a
misbehaving agent cannot push a secret-bearing commit past it
and cannot acquire push access by inspecting the agent's own
filesystem.
## Problem
@@ -30,12 +37,19 @@ the agent can disable is not a gate.
## Goals / Success Criteria
Integration test: spin up a bottle whose only push path for a
declared upstream is the gate. Drop a synthetic high-entropy
secret into a commit, run `git push` from inside the agent,
observe a non-zero exit and a gitleaks finding in the gate's
stderr. Repeat with a clean commit, observe exit 0 and the commit
landing on the real upstream.
Two integration tests, both with the gate as the only git path
for a declared upstream:
1. **Push:** drop a synthetic high-entropy secret into a commit,
run `git push` from inside the agent, observe a non-zero exit
and a gitleaks finding in the response. Repeat with a clean
commit and observe exit 0 + the commit landing on the real
upstream.
2. **Fetch:** clone the upstream through the gate (`git clone`
against the gate URL), observe the upstream's content. Push
a new commit to the upstream out-of-band, refetch through the
gate, observe the new commit. The gate must never serve stale
data — every fetch refreshes from upstream first.
## Non-goals
@@ -71,12 +85,19 @@ landing on the real upstream.
uses to push upstream. The agent gets no parallel `bottle.ssh`
entry for those upstreams.
- **Agent-side URL rewrite.** Provisioner emits `~/.gitconfig`
with `[url "<gate-url>"] insteadOf = <real-url>` so `git push
origin` from inside the agent transparently hits the gate.
with `[url "<gate-url>"] insteadOf = <real-url>` so every git
operation against the declared upstream (push, fetch, clone,
pull, ls-remote) transparently hits the gate.
- **Pre-receive gitleaks hook.** Baked into the gate image. On a
hit the hook exits non-zero and the push fails; on clean it
shells out `git push <upstream> <ref>:<ref>` using the
gate-resident credential.
shells out `git push origin <ref>:<ref>` using the gate-resident
credential.
- **Access-hook upstream refresh.** `git daemon --access-hook` runs
`git fetch origin --prune` against the upstream before every
`upload-pack` request, so a fetch through the gate is observably
equivalent to a fetch against the real upstream. Failure to reach
the upstream is fail-closed: the access hook exits non-zero and
the agent's fetch fails.
- **Plan rendering / dry-run.** `bottle_plan.py` and the y/N
preflight surface the gate sidecar (name, listed upstreams,
which credential it holds per upstream).
@@ -86,10 +107,10 @@ landing on the real upstream.
- Push policy beyond gitleaks. No commit-author allowlist, no
branch-name policy, no signed-commit enforcement. gitleaks is
the single rule for v1.
- Fetch routing. Fetch can continue going through ssh-gate as
today, with the agent holding a read-scoped deploy key. Routing
fetch through the git-gate is a follow-up; this PRD is
push-side only. (Open question: revisit.)
- Fetch caching / stale-while-revalidate. Every `upload-pack`
refresh is a synchronous round-trip to the upstream; there is
no TTL cache, no background refresh. If the upstream is slow,
the agent's fetch is slow.
- Quarantine / replay. A rejected push is discarded; we do not
stash it for the user to inspect.
- Non-Docker backends. Implementation lands for Docker only; the
@@ -116,20 +137,34 @@ Mirror the existing sidecar layout:
`stop` is idempotent `docker rm -f`. Container name:
`claude-bottle-git-gate-<slug>`.
Gate image: a minimal `git` + `gitleaks` + `openssh-server`
image, pinned by digest (declared next to `PIPELOCK_IMAGE` and
the socat image constant). For each declared upstream the gate
hosts a bare repo at a stable local path (`/git/<name>.git`)
with `hooks/pre-receive` wired to gitleaks. On a clean scan the
hook (or a `post-receive` companion) does `git push <upstream>
<ref>:<ref>` using the credential the gate holds for that
upstream.
Gate image: `git-daemon` + `openssh-client` over a
`zricethezav/gitleaks` base (alpine + gitleaks), pinned by digest.
For each declared upstream the gate hosts a bare repo at
`/git/<name>.git` with `remote.origin.url` set to the real
upstream (via `git remote add --mirror=fetch`), `hooks/pre-receive`
wired to gitleaks-then-`git push origin`, and the bare repo's
config carrying per-upstream credential paths.
Inside the bottle, the agent's `.gitconfig` rewrites the real
upstream URL to the gate's local URL via `insteadOf`. A `git
push origin main` therefore pushes to the gate; the gate scans;
on success the gate pushes to the real upstream. The agent never
sees the upstream push credential.
upstream URL to the gate's `git://` URL via `insteadOf`. Every
git operation against the declared upstream therefore hits the
gate.
For pushes, the pre-receive hook gitleaks-scans the incoming
refs and, on clean, pushes each accepted ref to the real
upstream using the credential the gate holds.
For fetches (clone, pull, fetch, ls-remote), `git daemon`'s
`--access-hook=<path>` runs `git fetch origin --prune` against
the real upstream before the upload-pack service serves the
client. The bare repo therefore reflects the upstream's current
state at the moment the agent's fetch begins; if the upstream
is unreachable, the access hook exits non-zero and the agent's
fetch fails — same observable behavior as if the agent were
talking to the upstream directly.
The agent never sees the upstream credential under either
operation.
### Existing code touched
@@ -167,17 +202,12 @@ exposes it as, and the credential the gate uses to push upstream
### External dependencies
- A minimal `git` + `gitleaks` + `openssh-server` image, pinned
by digest.
- `gitleaks` binary, version pinned in the image build.
- `zricethezav/gitleaks` base image, pinned by digest. The base
ships gitleaks + git; the gate Dockerfile adds `git-daemon` and
`openssh-client` on top.
- No new Python packages.
## Future work
- **Fetch through the gate.** A v2 could route fetch through the
gate too, so the agent holds no upstream credentials at all.
Today fetch falls back to ssh-gate; pushing through git-gate
alone is the v1 win.
- **Smolmachines colocation.** The eventual smolmachines backend
may pack pipelock + ssh-gate + git-gate into a single microVM,
or split git-gate off because it holds push creds and the