fix(cred_proxy): close git-push bypass + route through pipelock (PRD 0010)
test / unit (pull_request) Successful in 15s
test / integration (pull_request) Successful in 29s

Three coupled fixes that close a documented bypass of git-gate's
gitleaks pre-receive hook:

1. cred-proxy refuses git smart-HTTP push at runtime. Any path
   ending in /git-receive-pack or /info/refs?service=git-receive-pack
   returns 403 with a pointer at the bottle.git SSH path. Fetch
   (upload-pack) is still allowed — the bypass we're closing is
   push, where gitleaks is the load-bearing scanner. Hard guarantee.

2. The provisioner suppresses the cred-proxy `~/.gitconfig` insteadOf
   rewrite for any host already declared in bottle.git. git-gate is
   the canonical git path there; we don't write a competing rule
   that would let `git clone https://<host>/...` succeed in ways
   that confuse on push. Defense in depth — (1) is the hard guarantee.

3. cred-proxy routes its outbound HTTPS through pipelock. The
   sidecar's environ now sets HTTPS_PROXY=<pipelock-url>, and the
   image's entrypoint runs `update-ca-certificates` over the
   per-bottle pipelock CA (docker cp'd into
   /usr/local/share/ca-certificates/pipelock.crt before start) so
   the proxy's HTTPS client trusts pipelock's bumped certs.

   Consequence: pipelock's allowlist + body scanner now sit in the
   cred-proxy egress path the same way they sit in front of direct
   agent traffic. The cred-proxy upstream hosts (api.github.com,
   github.com, gitea hosts, registry.npmjs.org) come OFF
   pipelock's passthrough_domains. Only api.anthropic.com remains
   on passthrough (LLM body content legitimately trips DLP).

PRD 0010 updated to reflect all three. Tests adjusted: the
"cred-proxy hosts go on passthrough" assertion in
test_pipelock_allowlist flips to "they don't", a new
TestIsGitPushRequest exercises the smart-HTTP refusal predicate,
and the gitconfig renderer tests cover the per-host suppression
matrix.
This commit is contained in:
2026-05-13 21:09:33 -04:00
parent c8ab90d01d
commit 27b2d78b11
12 changed files with 329 additions and 63 deletions
@@ -35,8 +35,10 @@ def provision_cred_proxy(plan: DockerBottlePlan, target: str) -> None:
upstreams = plan.cred_proxy_plan.upstreams
if not upstreams:
return
bottle = plan.spec.manifest.bottle_for(plan.spec.agent_name)
git_gate_hosts = {g.UpstreamHost for g in bottle.git}
_provision_npmrc(plan, target, upstreams)
_provision_gitconfig(plan, target, upstreams)
_provision_gitconfig(plan, target, upstreams, git_gate_hosts)
_provision_tea_config(plan, target, upstreams)
@@ -82,29 +84,41 @@ def _provision_npmrc(
# --- git config -------------------------------------------------------------
def render_cred_proxy_gitconfig(upstreams: tuple[CredProxyUpstream, ...]) -> str:
def render_cred_proxy_gitconfig(
upstreams: tuple[CredProxyUpstream, ...],
git_gate_hosts: set[str] = frozenset(), # type: ignore[assignment]
) -> str:
"""Render the `~/.gitconfig` fragment for cred-proxy insteadOf
rewrites. Empty string when no github / gitea routes are declared.
github expands to two rewrites: https://github.com/... → /gh-git/...
(the git transport endpoint), and the agent's git client reaches
api.github.com over the same proxy via the /gh-api/ route, but
that's used by tools that call the GitHub API directly (gh, tea,
octokit) rather than `git` itself.
The rewrite is suppressed for any host that's also declared in
`bottle.git`. git-gate is the canonical git path on those hosts —
its pre-receive runs gitleaks before forwarding the push. A
cred-proxy https://<host>/ rewrite would route HTTPS git ops
around the gate. cred-proxy still refuses smart-HTTP push at
runtime (defense in depth), but suppressing the rewrite means
`git clone https://<host>/...` doesn't have a tempting shortcut
that just confuses on push.
Gitea entries get one rewrite per declared host, pointing at
/gitea/<host>/. The path component scopes the credential
so multiple gitea instances coexist on one proxy."""
github expands to one rewrite (https://github.com/... → /gh-git/...,
the git transport endpoint); /gh-api/ stays unmapped here because
tools call api.github.com directly rather than through git.
Gitea entries get one rewrite per declared host."""
rules: list[str] = []
for u in upstreams:
if u.kind == "github" and u.path == "/gh-git/":
if "github.com" in git_gate_hosts:
continue
rules.append(
f'[url "{cred_proxy_url()}/gh-git/"]\n'
f"\tinsteadOf = https://github.com/\n"
)
elif u.kind == "gitea":
# u.upstream is the configured gitea URL (e.g.
# https://gitea.dideric.is) and u.path is /gitea/<host>/.
# u.path is /gitea/<host>/; derive the host the same way
# the route table did so we match git_gate's UpstreamHost.
host = u.path[len("/gitea/"):].rstrip("/")
if host in git_gate_hosts:
continue
rules.append(
f'[url "{cred_proxy_url()}{u.path}"]\n'
f"\tinsteadOf = {u.upstream}/\n"
@@ -123,11 +137,13 @@ def _provision_gitconfig(
plan: DockerBottlePlan,
target: str,
upstreams: tuple[CredProxyUpstream, ...],
git_gate_hosts: set[str],
) -> None:
"""Append the cred-proxy insteadOf rules to ~/.gitconfig. Runs
after `provision_git`, so any git-gate rules already live in the
file; we append rather than overwrite."""
content = render_cred_proxy_gitconfig(upstreams)
file; we append rather than overwrite. Hosts already brokered by
git-gate are skipped — git-gate is the canonical git path there."""
content = render_cred_proxy_gitconfig(upstreams, git_gate_hosts)
if not content:
return
container_home = os.environ.get("CLAUDE_BOTTLE_CONTAINER_HOME", "/home/node")