review-session
Review an agent session for evidence-backed friction and the smallest measurable remedies.
- License 许可
- MIT
- 最近更新
- 2026-11-12
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- review-session
- license
- MIT
- metadata
- { "summary": "Turns observed session failures into bounded remedies without inventing causes or applying unrequested policy changes.", "category": "Improvement" }
- description
- Review an agent session for evidence-backed friction and the smallest measurable remedies.
- disable-model-invocation
- true
02 资源文件 13
- 参考文档agents/openai.yaml43 B
- 参考文档evals/budget.json873 B
- 其他evals/cases-initial.jsonl16.6 KB
- 其他evals/cases.jsonl20.5 KB
- 参考文档evals/fixtures/README.md212 B
- 参考文档evals/matrix.json1.2 KB
- 参考文档evals/results/README.md1.4 KB
- 参考文档evals/rubric.md3.2 KB
- 参考文档evals/run.md4.0 KB
- 参考文档evals/score-input.schema.json635 B
- 脚本evals/score.mjs180 B
- 参考文档evals/suite-manifest-i1.json11.6 KB
- 参考文档evals/suite-manifest-i2.json15.3 KB
03 SKILL.md 正文
Review Session
Trace each recommendation to an observed event. Apply remedies only when the user requests application.
Scope
Use when explicitly invoked to review a session's effectiveness.
Do not use for:
- Diagnosing a live defect: use
diagnose; reproduce and discriminate causes if unavailable. - Authoring a skill: tell the user to run
create-skill. If it is not installed, confirm the skill's job and compare it with a cheaper mechanism. - Editing persistent project instructions: tell the user to run
write-agent-instructions. If it is not installed, verify the instruction defect and test a minimal change. - General opinions about agent quality without session evidence: request the missing trace or limit the review to available facts.
Workflow
1. Reconstruct the useful sequence
Review the session the user names. With none named, review the current conversation and say that turns removed by compaction are not covered. Read its tool results, artifacts, and the relevant repository state. Identify the intended outcome and what actually happened. Separate observations from causal hypotheses. Keep gaps, unlogged time, and inaccessible evidence explicit. Quote only the minimum context needed to identify a failure. Redact secrets and unrelated private data.
Done when every reported friction point has an evidence pointer and a consequence.
2. Find the owning mechanism
Inspect existing checks, scripts, instruction layers, and local conventions. Distinguish an absent mechanism from one that failed or was never run. Use a deterministic check for a mechanical defect. Prefer wiring or repairing an existing check over adding prose. Use a bounded instruction change only for demonstrated instruction-related behavior. Leave unresolved product choices with their owner.
Done when each proposed remedy addresses the observed failure in the appropriate layer.
3. Test the proposed effect
Replay a safe existing check against the original failing input when possible. Verify check descriptions against runtime output rather than source spelling. Show how the smallest remedy would change that result. Define a downstream task for behavioral remedies. Report an unavailable replay as a limit rather than an accomplished prevention. Compare expected consequence and effort. Rank at most three remedies and identify the first useful one.
Done when recommendations have a checkable effect and no unsupported universal guarantee.
4. Save the review
Follow the requested location or use docs/session-reviews/<topic>.md.
Include the observed issue, evidence, remedy, effect check, and unresolved limit.
Keep review-only work to the report.
Apply an explicitly requested remedy within its bounded scope.
Verify any applied change with its owning workflow.
Done when the review is inspectable and application status is unambiguous.
Worked example
Issue:
retrysreached main; the formatter passed it. Evidence: the CI log runs only the formatter;check-config.mjsexists and is not in the CI command. Remedy: addcheck-config.mjsto the CI command. Effect check: replayed the checker on the misspelled file; it exits non-zero. Limit: the delay before the fix is unlogged; no cause is claimed.
Verification
Check recommendations against actual events and existing mechanisms. Check that the report contains no invented causal story or unrequested policy edit. Recommend the smallest remedy whose effect the evidence can support.
04 同仓库更多 Skills
write-agent-instructions
Make the smallest evidence-backed change to persistent project instructions and test downstream behavior.
verify-work
Use when a finished, runnable artifact (a script, config, parser, prompt, or CLI, not a diff under review) must survive hostile input before shipping: the user asks to "verify this script", "find the loopholes", "test this properly against bad input", or asks "is this ready to ship" about one. Also when a shipped artifact failed on an input nobody tried, or when verification so far is a happy-path smoke test.
understand-codebase
Use when the user asks to "understand this codebase", "how does this feature work in this repo?", "trace this feature through the code", or needs onboarding, an architecture explanation, or answers about unfamiliar code. Also use for read-only code-path questions before a change.
tdd
Use when writing or changing production behavior: implementing a feature, fixing a bug whose cause is already known, or altering stored, transformed, migrated, serialized, or externally transmitted data. Also use when a user says "use TDD", "test first", or "red-green-refactor", and when code is being written before a failing test. For a failure whose cause is unknown, diagnose first.
simplify-code
Use when the user asks to "simplify this code", "remove the fluff", or "clean up this over-engineered module", or wants working code that is bloated, duplicated, over-abstracted, or dependency-heavy reduced without changing behavior.
review-code
Use when a user asks to "review this branch", "review this PR", "review my changes", or "review since main". Also use when asked whether a branch, PR, or diff is ready to merge or ship, and for work-in-progress diffs, staged changes, suspected regressions in a patch, or checking implementation against a specification.