review-pr
Review an AgentDescent PR (or local change) at a tip SHA and write a Verifier report with Tip SHA / Ran / Not run / Verdict PASS|FAIL|INC that the review gate can read.
ThinkFlowLab · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add ThinkFlowLab/agentdescent --skill review-pr技能说明
name: review-pr description: Review an AgentDescent PR (or local change) at a tip SHA and write a Verifier report with Tip SHA / Ran / Not run / Verdict PASS|FAIL|INC that the review gate can read.
AgentDescent review-pr
Produces a Verifier report for ThinkFlowLab/agentdescent. The report is evidence,
not merge authority: what may happen after a PASS is AGENTS.md › Merge policy.
Establish the tip
- Work in a fresh clone (
git clone https://github.com/ThinkFlowLab/agentdescent), check out the PR head (gh pr checkout <n>), recordgit rev-parse HEAD. If the head moves during the review, start over. - Read the rules from
origin/main, not a stale copy: AGENTS.md, CODE_STANDARDS.md, LESSONS.md, and this file (git show origin/main:.claude/skills/review-pr/SKILL.md). - Inventory the diff against the merge base (
git diff origin/main...HEAD --stat) and get the tier:python3 .github/scripts/risk_gate.py --pr <n>. The PR description is narrative, not evidence.
Re-run the gates yourself
pip install -e ".[dev,mcp]"
pytest -q # the whole suite, not only the new tests
python -m tools.gen_api_docs --check # when agentdescent/__init__.py or a public signature changed
python -m tools.changelog_fragments check
mkdocs build --strict # when docs/ or mkdocs.yml changed (needs .[docs])
- Fails on
main: copy only the new/changed test files onto the merge base and run them; they must fail there and pass at the tip. - Ark real-model checks only when the change touches a model path and
ARK_API_KEYis already in the environment; otherwise list them underNot run:.
Instant FAIL
- A key, token or credential anywhere in the diff.
- A network or model call reachable from
pytest -q. docs/api.mdedited by hand or out of sync;CHANGELOG.mdedited in a PR.- A claimed fix with no test that fails on
main. - An unconditional
@pytest.mark.skip/xfailhiding a failure. - A port that departs from its source's released code without saying so.
Verdicts
| Verdict | Meaning |
|---|---|
| PASS | No actionable findings; every gate you list under Ran: was run by you at this tip and is green |
| FAIL | An actionable defect (correctness, a hard rule, a missing fail-on-main test, a dishonest claim) |
| INC | Incomplete: a material check not run, the tip moved, scope unclear. INC is not PASS |
Report shape (the review gate parses this)
Post it as one PR comment from the Birfy login (self-review counts). Keep the
block intact; the last Tip SHA: block in the comment is the one read, the
newest report at a tip wins, and the tip may be written above the verdict or
inline in it (Verdict: PASS (tip b0a30224)). A newer comment or review at that
same tip with no Verdict: line (a withdrawal, an added requirement, an open
question) also supersedes an earlier PASS there, so the tip reads unreviewed
until a new PASS is posted. A comment or review body carries no thread id, so a
verdict-less report attaches to the tip its author most recently named - if that
is the tip being judged it supersedes, even when the prose is about an earlier
tip. Anything written at that tip after a PASS therefore voids it: a fresh PASS
must be the last report written there. A PASS without a Ran: entry or without
a Not run: line still counts, but the gate marks it weak: no Ran:/Not run:
so it never passes unseen.
Self-review (Birfy)
Tip SHA: <full 40-hex head>
Ran:
- pytest -q → pass (<N> passed, <M> skipped)
- python -m tools.gen_api_docs --check → pass
- CI at this tip → green (build, test 3.9-3.13)
Not run:
- Ark real-model run — no model path changed
Findings:
- none
Verdict: PASS
Then check it counts: python3 .github/scripts/review_gate.py --pr <n>.
Merge lane by tier
The tier is the diff's (risk_gate.py), never the PR text. Once the PASS counts,
run this from the PR clone and do exactly what it prints:
python3 .github/scripts/merge_lane.py --pr <n> --repo ThinkFlowLab/agentdescent \
--tip <the PASS's Tip SHA>
The lane reads its two trust inputs — .github/agent/review-policy.json and
.github/agent/risk-rules.json — from the PR's base commit
(git show <base>:<path>), never from the clone it runs in, so a branch that
edits either file cannot lend itself a PASS or a lower tier. If the clone cannot
serve a file from base the lane prints no lane and exits 1 without a command:
git fetch origin and run it again; never substitute the branch's copy.
Independently of the rules file, a diff touching .github/** or
.claude/skills/** is T3.
auto-merge (T0/T1 with a compliant PASS) is the pinned
gh pr merge … --auto --squash --match-head-commit <tip> — the only sanctioned
auto-merge. awaiting-approval (T2/T3, including a T3 that only the size cap
produced) adds the label, leaves auto-merge unset and stops for Danyang's click.
wait means no compliant PASS names the tip: post one or re-review, never merge.