菜单
Optional. Use when the user asks to review, compare, or interpret multiple local Go pprof bundles across optimization rounds. Classify bundles as cold, warm, or transition, extract exact deltas for target symbols, and produce a net-win or net-regression judgment with explicit confidence and caveats.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Use after neutral discovery and before Scope Ledger drafting when consequential design uncertainty, ownership placement, lifecycle/state responsibility, compatibility, migration, algorithm choice, operability, or reversibility could admit a materially different safe approach. Begin with a lightweight design-space probe when the lead currently sees one obvious design but an alternative may plausibly exist. Run the full comparison when the probe identifies two or more viable approaches or when choosing incorrectly would be materially consequential. Do not trigger when settled constraints clearly force one safe design.
Explain, trace, or deeply review existing code using inspected repository evidence without implementing changes. Use for code-understanding requests unless the user explicitly asks only for a concise, brief, quick, high-level, or initial overview. This skill explicitly owns thorough reviews, comprehensive explanations, audits, diagnosis, correctness analysis, architecture assessment, risk analysis, execution tracing, concurrency analysis, lifecycle analysis, and full subsystem understanding.
Use when actual scientific or model inputs, outputs, semantics, units, boundaries, classifications, calibration, uncertainty, tolerances, or claims change or require authoritative resolution. Independently derive and evaluate the scientifically defensible contract, compare competing interpretations, create independent golden vectors, and make evidence-qualified recommendations while leaving product policy, scope, architecture, and implementation authority with the user and lead. Do not trigger for mechanical work that cannot affect model behavior or claims.
Use before approval when testability or oracle quality could materially affect scope, behavior, architecture, observability, or validation planning, and after approval before implementation when the detailed planned evidence may not falsify a broken design. Apply to unclear parser, protocol, config, lifecycle, retained-state, performance, scientific/model, security, workflow-checker, or other material oracles. Do not trigger when accepted checks directly prove a localized mechanical change.
Use when reviewing or changing Codex workflow authority, approval, validation routing, review rules, templates, repo-managed skills, or enforcing scripts. Select conformance mode when checking implementation against accepted workflow policy. Select effectiveness mode when the user explicitly asks whether the workflow, its rules, or its architecture should be retained, changed, consolidated, replaced, or removed. Do not trigger for ordinary product documentation or runtime work whose workflow contract is unchanged.
Use after current-state discovery when material product, operator, compatibility, failure, classification, default, threshold, ordering, precedence, persistence, or test-oracle semantics still admit more than one plausible interpretation. Compare interpretations and make conditional recommendations, but unresolved material semantics remain user-owned and block concrete design selection and an unconditional Scope Ledger until explicitly resolved by the user, controlling authority, or evidence that eliminates every competing interpretation. Do not trigger merely because requirements are detailed or work is Non-trivial.
| name | pprof-impact-review |
| description | Optional. Use when the user asks to review, compare, or interpret multiple local Go pprof bundles across optimization rounds. Classify bundles as cold, warm, or transition, extract exact deltas for target symbols, and produce a net-win or net-regression judgment with explicit confidence and caveats. |
Use this skill for repo-local performance evidence review when multiple
profile bundles exist and the user wants to know whether a change helped,
regressed, or only partially succeeded. Fable-native counterpart to
codex-skills/pprof-impact-review/SKILL.md. Unlike the 7 skills required
by docs/fable-workflow.md's Dependency rigor and audit sections, this one
is optional on both sides of the contract — it is not in AGENTS.md's
required Skill Check trigger list either, so treat it as available
performance-evidence tooling, not a gating audit.
Build the bundle ledger first.
cold, warm, or transition.Inspect the standard profile set.
go tool pprof -top for:
inuse_space)alloc_space)logs/ over assumptions.Compare the target symbols explicitly.
alloc_space vs inuse_spaceProduce the judgment.
Typical commands:
go tool pprof -top logs/cpu-<stamp>.pprofgo tool pprof -top logs/mutex-<stamp>.pprofgo tool pprof -top logs/heap-<stamp>.pprofgo tool pprof -top -sample_index=alloc_space logs/allocs-<stamp>.pprofUse focused filtering when comparing specific symbols across bundles.