feat: prep reviewer repo at init and load repo skills (#1480)

* feat: prep reviewer repo at init and load repo skills

Clone + checkout the PR head during reviewer agent init so SkillsMiddleware
can discover the repo's .agents/skills and .claude/skills from disk at its
one-shot scan, and so the LLM no longer narrates the clone mid-run.

Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>

* fix: load reviewer skills from trusted base sha and fetch PR pull ref

Address review findings: skills are now extracted from the PR base sha via
git archive into a dir outside the checkout (prevents PR-authored SKILL.md
prompt injection), and repo prep fetches refs/pull/<n>/head with a strict
checkout so fork PRs fail loudly instead of silently reviewing the default
branch.

* fix: drop ref from skill-extraction log to satisfy CodeQL

---------

Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
This commit is contained in:
Johannes du Plessis 2026-06-10 09:54:32 -07:00 • committed by GitHub
parent 16eb15bab3
commit 8f596a421d
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 331 additions and 1 deletions

View file

@ -69,6 +69,7 @@ from .utils.api_standards_skill import fetch_api_standards_skill
from .utils.github_app import get_github_app_installation_token_with_expiry
from .utils.github_token import cache_github_token_for_thread
from .utils.model import DEFAULT_LLM_REASONING, make_model, provider_model_kwargs
from .utils.repo_prep import materialize_trusted_skills, prepare_review_repo
from .utils.sandbox_paths import aresolve_sandbox_work_dir
REVIEWER_PROMPT_TEMPLATE = """You are a specialized code reviewer agent. Your job is to review one GitHub PR and publish a single review.
@ -87,12 +88,17 @@ Re-review (user message says "A new commit has been pushed"):
GH_TOKEN=dummy gh api repos/{repo_owner}/{repo_name}/compare/<last_reviewed_sha>...<head_sha> -H "Accept: application/vnd.github.v3.diff"
```
Clone the repo so you can grep for full file context:
The repo is normally already cloned and checked out at the PR head in
`{working_dir}` — `cd` there and grep for full file context. If that directory
is missing (prep can occasionally fail), clone it yourself:
```
GH_TOKEN=dummy gh repo clone {repo_owner}/{repo_name} && cd {repo_name} && git checkout <head_sha>
```
If a skills section appears below, the repo ships reviewer-relevant skills. Read
the `SKILL.md` that matches the area you're reviewing and apply it.
Tools: `add_finding`, `update_finding`, `list_findings`, `publish_review`,
`resolve_finding_thread`, `reply_to_finding_thread`.
Call `publish_review` once at the end.
@ -697,6 +703,26 @@ async def get_reviewer_agent(config: RunnableConfig) -> Pregel:
base_sha = str(config["configurable"].get("base_sha", "") or "")
head_sha = str(config["configurable"].get("head_sha", "") or "")
pr_number = config["configurable"].get("pr_number")
# Prep the repo on the sandbox before the first model call so the LLM does
# not narrate `gh repo clone`, and so SkillsMiddleware can discover the
# repo's skills at its one-shot scan. Skills are materialized from the PR
# base sha (trusted), never the PR head (author-controlled).
repo_ready = await prepare_review_repo(
sandbox_backend,
work_dir=work_dir,
repo_owner=repo_owner,
repo_name=repo_name,
head_sha=head_sha,
pr_number=pr_number if isinstance(pr_number, int) else None,
base_sha=base_sha,
)
skill_sources: list[str] = []
if repo_ready and repo_name:
skill_sources = await materialize_trusted_skills(
sandbox_backend, repo_dir=f"{work_dir}/{repo_name}", trusted_ref=base_sha
)
pr_url = str(config["configurable"].get("pr_url", "") or "")
last_reviewed_sha = str(config["configurable"].get("last_reviewed_sha", "") or "")
is_re_review = bool(config["configurable"].get("re_review"))
@ -945,6 +971,7 @@ async def get_reviewer_agent(config: RunnableConfig) -> Pregel:
],
subagents=[_general_purpose_subagent(reviewer_subagent_model)],
backend=sandbox_backend,
skills=skill_sources or None,
middleware=[
SanitizeToolInputsMiddleware(),
ModelCallLimitMiddleware(run_limit=MODEL_CALL_RECURSION_LIMIT, exit_behavior="end"),

160
agent/utils/repo_prep.py Normal file
View file

@ -0,0 +1,160 @@
"""Deterministic repo prep for the reviewer sandbox.
The reviewer reviews a single PR, so we clone its repo and check out the PR
head during agent init -- before the first model call -- instead of asking the
LLM to narrate ``gh repo clone`` mid-run. Pre-cloning also lets ``SkillsMiddleware``
discover the repo's ``.agents/skills`` / ``.claude/skills`` at its one-shot
``before_agent`` scan.
Best-effort: any failure leaves the sandbox usable (the review still works off
the fetched diff) and returns ``False`` so callers can skip skill wiring.
"""
from __future__ import annotations
import asyncio
import logging
import posixpath
import shlex
from collections.abc import Sequence
from deepagents.backends.protocol import SandboxBackendProtocol
logger = logging.getLogger(__name__)
CLONE_TIMEOUT_SECONDS = 240
DEFAULT_SKILL_DIRS = (".agents/skills", ".claude/skills")
TRUSTED_SKILLS_DIRNAME = ".review-skills"
def _prep_command(
work_dir: str,
repo_owner: str,
repo_name: str,
head_sha: str,
pr_number: int | None,
base_sha: str,
) -> str:
repo_dir = posixpath.join(work_dir, repo_name)
q_work_dir = shlex.quote(work_dir)
q_repo_dir = shlex.quote(repo_dir)
q_full_name = shlex.quote(f"{repo_owner}/{repo_name}")
q_repo_name = shlex.quote(repo_name)
q_head = shlex.quote(head_sha) if head_sha else ""
lines = [
"set -e",
f"if [ -d {q_repo_dir}/.git ]; then",
f" cd {q_repo_dir} && GH_TOKEN=dummy git fetch --all --quiet",
"else",
f" cd {q_work_dir} && GH_TOKEN=dummy gh repo clone {q_full_name} && cd {q_repo_name}",
"fi",
]
if base_sha:
q_base = shlex.quote(base_sha)
lines.append(f"GH_TOKEN=dummy git fetch origin {q_base} --quiet 2>/dev/null || true")
if q_head:
# Direct sha fetch covers same-repo PRs; the pull ref covers fork PRs
# whose head commit is not reachable from origin's branches.
lines.append(f"GH_TOKEN=dummy git fetch origin {q_head} --quiet 2>/dev/null || true")
if pr_number is not None:
pull_ref = shlex.quote(f"refs/pull/{pr_number}/head")
lines.append(f"GH_TOKEN=dummy git fetch origin {pull_ref} --quiet 2>/dev/null || true")
# Strict on purpose: a failed checkout must fail the prep so callers
# know the tree is NOT at the PR head.
lines.append(f"git checkout {q_head} --quiet")
return "\n".join(lines)
async def prepare_review_repo(
sandbox_backend: SandboxBackendProtocol,
*,
work_dir: str,
repo_owner: str,
repo_name: str,
head_sha: str,
pr_number: int | None = None,
base_sha: str = "",
) -> bool:
"""Clone-or-fetch the repo and check out ``head_sha`` in the sandbox.
Returns ``True`` only when the repo is prepped at ``work_dir/repo_name``
and (when ``head_sha`` is given) actually checked out at the PR head.
"""
if not repo_owner or not repo_name:
return False
command = _prep_command(work_dir, repo_owner, repo_name, head_sha, pr_number, base_sha)
try:
result = await asyncio.to_thread(
sandbox_backend.execute, command, timeout=CLONE_TIMEOUT_SECONDS
)
except Exception: # noqa: BLE001
logger.warning("Failed to prep review repo %s/%s", repo_owner, repo_name, exc_info=True)
return False
exit_code = getattr(result, "exit_code", None)
if exit_code not in (0, None):
logger.warning(
"Review repo prep for %s/%s exited %s: %s",
repo_owner,
repo_name,
exit_code,
getattr(result, "output", ""),
)
return False
logger.info(
"Prepped review repo %s/%s at %s (head=%s)",
repo_owner,
repo_name,
posixpath.join(work_dir, repo_name),
head_sha or "<none>",
)
return True
async def materialize_trusted_skills(
sandbox_backend: SandboxBackendProtocol,
*,
repo_dir: str,
trusted_ref: str,
skill_dirs: Sequence[str] = DEFAULT_SKILL_DIRS,
) -> list[str]:
"""Extract skill dirs from ``trusted_ref`` into a path outside the checkout.
Skills are sourced from the PR's base sha -- never the PR head, which the
PR author controls -- so a PR cannot inject instructions into the reviewer
prompt by adding or editing a ``SKILL.md``. Returns the extracted source
dirs (with a trailing slash, as SkillsMiddleware expects).
"""
if not trusted_ref:
return []
dest_root = posixpath.join(posixpath.dirname(repo_dir), TRUSTED_SKILLS_DIRNAME)
q_repo_dir = shlex.quote(repo_dir)
q_ref = shlex.quote(trusted_ref)
sources: list[str] = []
for skill_dir in skill_dirs:
dest = posixpath.join(dest_root, skill_dir)
q_dest = shlex.quote(dest)
q_dir = shlex.quote(skill_dir)
depth = len(skill_dir.split("/"))
command = (
f"cd {q_repo_dir} && "
f"git cat-file -e {q_ref}:{q_dir} 2>/dev/null && "
f"rm -rf {q_dest} && mkdir -p {q_dest} && "
f"git archive {q_ref} {q_dir} | tar -x --strip-components={depth} -C {q_dest} && "
f"echo {q_dest}"
)
try:
result = await asyncio.to_thread(sandbox_backend.execute, command)
except Exception: # noqa: BLE001
logger.warning("Failed to extract trusted skills %s", skill_dir, exc_info=True)
continue
output = getattr(result, "output", "") or ""
if dest in output.splitlines():
sources.append(f"{dest}/")
return sources

143
tests/test_repo_prep.py Normal file
View file

@ -0,0 +1,143 @@
from __future__ import annotations
from deepagents.backends.protocol import ExecuteResponse
from agent.utils.repo_prep import materialize_trusted_skills, prepare_review_repo
class _FakeSandboxBackend:
def __init__(
self,
*,
exit_code: int = 0,
raise_exc: bool = False,
output: str = "",
outputs: list[str] | None = None,
) -> None:
self._exit_code = exit_code
self._raise = raise_exc
self._output = output
self._outputs = outputs
self.commands: list[str] = []
@property
def id(self) -> str:
return "fake-sandbox"
def execute(self, command: str, *, timeout: int | None = None) -> ExecuteResponse:
del timeout
if self._raise:
raise RuntimeError("sandbox unreachable")
self.commands.append(command)
output = self._output
if self._outputs is not None:
output = self._outputs[len(self.commands) - 1]
return ExecuteResponse(output=output, exit_code=self._exit_code, truncated=False)
async def test_prepare_review_repo_clones_and_checks_out_head() -> None:
backend = _FakeSandboxBackend()
ok = await prepare_review_repo(
backend,
work_dir="/work",
repo_owner="acme",
repo_name="widget",
head_sha="abc123",
pr_number=42,
base_sha="def456",
)
assert ok is True
assert len(backend.commands) == 1
cmd = backend.commands[0]
assert "gh repo clone acme/widget" in cmd
assert "/work/widget/.git" in cmd
assert "git fetch origin def456" in cmd
assert "git fetch origin refs/pull/42/head" in cmd
assert "git checkout abc123 --quiet" in cmd
assert "git checkout abc123 --quiet 2>/dev/null || true" not in cmd
async def test_prepare_review_repo_skips_pull_ref_without_pr_number() -> None:
backend = _FakeSandboxBackend()
ok = await prepare_review_repo(
backend,
work_dir="/work",
repo_owner="acme",
repo_name="widget",
head_sha="abc123",
)
assert ok is True
assert "refs/pull" not in backend.commands[0]
async def test_prepare_review_repo_skips_checkout_without_head() -> None:
backend = _FakeSandboxBackend()
ok = await prepare_review_repo(
backend,
work_dir="/work",
repo_owner="acme",
repo_name="widget",
head_sha="",
)
assert ok is True
assert "git checkout" not in backend.commands[0]
async def test_prepare_review_repo_requires_owner_and_name() -> None:
backend = _FakeSandboxBackend()
ok = await prepare_review_repo(
backend, work_dir="/work", repo_owner="", repo_name="widget", head_sha="abc"
)
assert ok is False
assert backend.commands == []
async def test_prepare_review_repo_returns_false_on_nonzero_exit() -> None:
backend = _FakeSandboxBackend(exit_code=1)
ok = await prepare_review_repo(
backend, work_dir="/work", repo_owner="acme", repo_name="widget", head_sha="abc"
)
assert ok is False
async def test_prepare_review_repo_returns_false_on_exception() -> None:
backend = _FakeSandboxBackend(raise_exc=True)
ok = await prepare_review_repo(
backend, work_dir="/work", repo_owner="acme", repo_name="widget", head_sha="abc"
)
assert ok is False
async def test_materialize_trusted_skills_extracts_from_trusted_ref() -> None:
backend = _FakeSandboxBackend(outputs=["/work/.review-skills/.agents/skills\n", ""])
sources = await materialize_trusted_skills(
backend, repo_dir="/work/widget", trusted_ref="def456"
)
assert sources == ["/work/.review-skills/.agents/skills/"]
assert len(backend.commands) == 2
for cmd in backend.commands:
assert "git cat-file -e def456:" in cmd
assert "git archive def456" in cmd
async def test_materialize_trusted_skills_empty_without_ref() -> None:
backend = _FakeSandboxBackend()
sources = await materialize_trusted_skills(backend, repo_dir="/work/widget", trusted_ref="")
assert sources == []
assert backend.commands == []
async def test_materialize_trusted_skills_empty_when_none_exist() -> None:
backend = _FakeSandboxBackend(output="")
sources = await materialize_trusted_skills(
backend, repo_dir="/work/widget", trusted_ref="def456"
)
assert sources == []
async def test_materialize_trusted_skills_handles_exception() -> None:
backend = _FakeSandboxBackend(raise_exc=True)
sources = await materialize_trusted_skills(
backend, repo_dir="/work/widget", trusted_ref="def456"
)
assert sources == []