| name | execution-reality-audit |
| description | 用于某个实现 slice、wave、fix 或 family 执行后做 evidence-driven reality audit:先锚定 current truth,对照文档/issue/spec 声明与代码、测试、contract、probe、运行时现实,遇到不确定点优先补最小 proof-carrying test 或 probe;当前 scope 内可闭合的小缺口直接修,跨边界问题转 residual / handoff。用户说“执行后 review”、“对照文档和代码确认实现度”、“不确定就补测试”、“closeout 前做 reality audit”时触发。 |
| compatibility | Best in pi with read, edit, write, and bash available. When this skill is routed by extension autopilot for `review`, `autopilot_report` must also be available. |
| allowed-tools | read edit write bash |
| metadata | {"gene_schema":"pi-skill-gene/v1","gene_profile":{"summary":"Audit implementation claims against code, tests, contracts, and runtime reality.","signals":["review after execution","compare spec and code for implementation truth","add proof when uncertain","audit before closeout"],"strategy":["anchor current truth first","compare claims against code, tests, and runtime evidence","turn uncertainty into proof","repair local gaps or route residuals","keep review-vs-closeout ownership truthful","issue an honest evidence-backed verdict"],"avoid":["broad architecture wandering without a concrete claim","fake closure without evidence","blind cross-boundary patching","review before current truth is anchored","inventing a separate global closeout skill when the routed runtime uses a repo-local closeout prompt surface"],"validation":["findings are evidence-backed","new proof or reused proof is named","direct fixes and residuals are separated","final verdict and next handoff are explicit","routed autopilot sessions end with one compatible autopilot_report when that protocol is in scope"]}} |
Execution Reality Audit
在一个实现看起来“做完了”之后,用这个 skill 判断:claim 是否真的被 reality 支撑。
优先 evidence,次优先 prose。
Routed Autopilot Review Contract
Apply this section only when the repo is running under the extension-driven autopilot runtime.
Do not imply that every repo or every manual review turn has these runtime surfaces.
- deterministic routed ownership here is
review
execute routes back to execute-plan when more implementation is needed
master_plan, wave_plan, and replan route to plan-creator
closeout routes to the repo-local closeout prompt surface, not a separate global closeout skill
- skill-backed phases require
selectedTools that still include read and autopilot_report
- end the phase with exactly one
autopilot_report
- when the prompt provides an active slice, set
stepId to that slice ID
- do not ask whether to continue unless there is a real external blocker or approval boundary
- translate the review verdict into
autopilot_report.status: use completed when the wave passes review, continue for more execution in the same wave, needs_replan when the plan must change, done only when the overall objective is actually complete, and blocked / failed for unsafe stops
- if review accepts and the objective is not yet complete, hand off to
execute-plan or plan-creator; if the objective is complete, hand off to the repo-local closeout prompt surface
Trigger Signals
Use this skill when the user wants to:
- review an implementation slice, wave, fix, or family after execution
- compare docs / issue / spec claims against code and tests
- add the smallest proof when the implementation status is still uncertain
- run a reality audit before closeout or handoff
Core Strategy
- Anchor current truth first. Start from the active plan / status / workset, repo rules, contracts, or stronger control-plane truth before broad repo scanning.
- Define the review root. Identify the smallest real surface: current claim, governing docs, relevant code paths, tests, contracts, and any paired dependency or runtime signal.
- Compare claim vs reality. Classify findings as
confirmed, drift, or uncertain by comparing docs/specs with code, tests, contracts, probes, and runtime evidence.
- Turn uncertainty into proof. Prefer the smallest proof-carrying test or probe. Reuse existing gates and artifacts before creating new ones.
- Repair or route. Fix gaps that close within the current scope, owner boundary, and validation ladder. Route cross-boundary or unsafe gaps into residuals / handoff instead of patching blindly.
- Keep review vs closeout truthful. Accept or reject the reviewed claim here, but do not blur review into a fake closeout surface. If the next route is closeout, make that handoff explicit and repo-local.
- Issue an honest verdict. Close only when claims are supported by evidence. Separate what is confirmed, what was fixed now, and what must move to successor work.
AVOID
- AVOID using this skill before implementation work or before current truth is known.
- AVOID broad architecture roaming when there is no concrete execution claim.
- AVOID fake closure without evidence.
- AVOID creating tests just to look rigorous when existing proof surfaces already answer the question.
- AVOID blind repairs across owner or repo boundaries.
- AVOID routing closeout through a made-up global closeout skill when the runtime uses a repo-local closeout prompt surface.
Validation / Output Contract
Default output should include:
findings
evidence added
fixes landed
successor residuals
verdict (accept, accept_with_residuals, blocked-closeout, or successor-required)
- optional
next handoff (plan-creator, execute-plan, repo-local closeout prompt surface, or none)
When routed by extension autopilot, keep the same audit substance but express the final decision through exactly one autopilot_report, with:
summary for the short review verdict
evidence for confirmed proof
artifacts for touched files / tests / probes
risks for residuals and blockers
stepId aligned to the active slice when present
nextAction pointing at execute-plan, plan-creator, or the repo-local closeout prompt surface