review-thread
Audit a coding thread, run, or PR against its claims and evidence; return findings and a continuation prompt.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- review-thread
- description
- Audit a coding thread, run, or PR against its claims and evidence; return findings and a continuation prompt.
02 资源文件 1
- 参考文档references/local-history.md1.6 KB
03 SKILL.md 正文
Review a thread
Review completed or ongoing work from outside its implementation context. The selected reviewer can be any available model with the required capabilities; this method has no dependency on a named model or provider. Optimize useful correction and completion time, not the volume of findings.
Locate the candidate
Use the supplied thread ID and the host's native history/export or local read-only history store. Recover the user's outcome, accepted decisions, latest corrections and claimed result. Prefer a bounded snapshot with pointers over loading every earlier conversation or handing its full history to a worker. On this workstation, consult local-history.md only when the supplied ID cannot be opened directly in the current host.
Verify the actual root, branch, commit/diff and PR. Sidebar links and model selection labels can be stale. Associate each finding with the inspected revision. If work changes during the review, record that and inspect the relevant delta before claiming the newest candidate is faulty or accepted. An unsuccessful limited search means “not located”, not “does not exist”.
Keep product files, remote records and active sessions unchanged during audit. Write private review artifacts when useful. A request to review does not itself authorize implementation, external messages, merge or production changes.
Test the claims
Compare the requested behavior to the diff and the evidence already available. Prioritize failures that prevent completion: omitted scope, wrong data/state, authorization, broken interactions, visual direction, and unsupported delivery claims. Inspect the exact provider/runtime contract when correctness hinges on it. Tests that mock away a service restriction cannot settle that question.
Separate confirmed defects, corrected findings and unverified claims. Confirm check exit codes and what each check covers. Reuse evidence while its revision, environment and inputs still apply; rerun only checks needed to resolve an actual uncertainty. Do not demand another full audit by default.
For visual claims, open approved references and the actual rendered evidence with a vision-capable reviewer. Check route, role, loaded data, viewport and candidate revision before judging fidelity. Desktop footage is not mobile evidence. If browser access or vision is unavailable, report that limit and continue the review supported by available evidence.
If runtime verification needs data changes, use an authorized isolated fixture and recoverable cleanup. Do not mutate existing user data merely to obtain a screenshot. Keep credentials and auth state out of reports and uploads.
Inspect orchestration only where it explains the result
Use observed task/session records to determine whether workers actually ran, which models were used, whether independent work overlapped, and where retries or repeated investigation occurred. Distinguish a bad parent instruction, worker error, environment failure and missing verification. Attribute causes only when the trace supports them; one failed run is not a universal model ranking. A plan to delegate is not a completed delegation.
Measure time to the requested usable result. Separate interruptions and provider waits when timestamps support it. Report per-request usage without double counting cumulative snapshots. Distinguish cached tokens, catalog/API cost, subscription quota and cash paid; mark unavailable measurements as unknown.
Return a correction the implementing thread can use
Lead with whether the inspected candidate meets the request. For each material finding, give the trigger, observed consequence, exact source/revision or evidence, and the smallest repair acceptance criterion. Explicitly retain resolved findings rather than assigning them again. Label missing proof as a verification gap rather than inventing a defect.
Provide a short continuation prompt containing the goal, current root/revision, open findings, useful evidence paths, focused checks and authorized endpoint. Preserve working code and settled decisions. Do not broaden it into a redesign of the development system, a new benchmark or a mandatory provider change. The verdict and continuation packet complete a review request. Ask for input only when a missing fact prevents a material conclusion; do not end by asking whether to begin an unrequested implementation stage.
04 同仓库更多 Skills
working-backwards
Define a software feature through PRFAQ and staged documents when the user requests working backwards before coding.
flow-implement
Implement a requested software change through verification, corrections, and its authorized delivery endpoint.
development-steward
Run the scheduled, read-only weekly maintenance review for the six explicitly opted-in repositories and prepare one concise private report for Check-in.
grill-with-docs
Run one recoverable interview for functional, visual, or mixed software grills; reuse equivalent answers and offer the shared HTML questionnaire.
global-agent-guardrails
Install, audit, test, or update the Development System's global destructive-command guard for Codex, Claude Code and T3 Code. Use when configuring agent safety hooks or investigating whether dangerous shell operations are blocked.
impeccable
Design or refine frontend UI using project context and a focused design playbook; also handles visual critique and polish.