skill-evolution-advisor

name	skill-evolution-advisor
description	经验 + skill 演化记忆：capture lightweight learnings, redacted diagnostic summaries, optional private local-detailed records, confirmed habits/gotchas, and confirmation-gated retrospective improvement prompts.

Skill Evolution Advisor

Preference contract: explicitly neutral with an empty eligible catalog. Governance and diagnostic privacy/detail constraints remain deterministic owner rules rather than soft research preferences.

Use this skill when a real workflow exposes friction in the research system, when the user corrects the agent, or when lightweight memory should capture a confirmed habit, recurring issue, or skill defect.

Workflow

1. For a user preference, capture only an explicit correction or repeated same-shape edit. Store one short normalized preference, the short verbatim observation, and its exact target skill and operation; leave it pending.
2. Accumulate no more than two pending preference observations at task close and ask naturally whether to remember them. Confirmation or dismissal must use the one-time public kb review snapshot; direct learning review/promotion is retired. source=user never counts as confirmation.
3. Do not recall or broadcast the full memory at session start. A preference affects work only while its human-signed, current-message confirmation receipt and runtime binding remain current; each routed consumer receives only its exact task-eligible selected subset. Recall recurring issues only for an explicit memory/diagnostic task.
4. Keep skill-defect entries record-only; do not auto-edit skills or roadmap files from them.
5. For deeper retrospectives, record which skills were used, separate observed friction from inferred causes, and generate an AI-ready patch prompt.
6. When the user explicitly asks to remember a failure, record one local redacted diagnostic issue even if automatic diagnostics are off.
7. For automatic failures, read the effective workspace/per-skill mode and detail policy first. errors-only is deterministic capture only; developer may request a short Agent retrospective only within the configured task budget. redacted is the default; local-detailed must be explicit and never means raw.
8. Keep issues.yaml as the redacted index. Private detail is bounded, digest-bound, owner-only, and excluded from export, public output, versioning, sync, install, and update. Do not expose its path, digest, frames, or content to the user.
9. When private protocol returns run_diagnostic_retrospective, validate that it carries only one safe issue_id and expected_detail_digest. Read the current detail through the owner, distinguish observation from hypothesis, write only the closed four-field analysis JSON under kb/.runtime/ as a current-user 0600 regular file, and apply through the owner with the same ID/digest. On digest drift, discard the stale analysis and prepare again. Never label a hypothesis confirmed.
10. Never copy raw stdout/stderr, traceback text, arguments, secrets, environment values, absolute paths, user/source/evidence text, or arbitrary excerpts into either diagnostic artifact. The only allowed user-message excerpt anywhere in this skill is the bounded verbatim preference observation. A defect remains record-only until a developer separately changes code.

Natural-language interaction

Ordinary users do not need a new pseudo CLI. Interpret requests such as “记下刚才的问题”“列出待复审的 skill 问题”“把这条问题标为已解决”“生成本地脱敏预览”“对待分析的详细诊断做复盘” and use the owner operations privately. Export preview requires explicit authorization in the current user message and remains summary-only. Summarize outcomes in natural language; do not reveal owner commands, flags, issue/detail paths, fingerprints, digests, or raw context.

Automatic runtime capture is best-effort and must never replace the original business exit status. If diagnostic recording itself fails, preserve the original operation result and let recovery restore the diagnostic file.

Commands

Lightweight learnings:

python3 .agents/skills/skill-evolution-advisor/scripts/learnings.py log --category recurring-issue --text "..." --source agent
python3 .agents/skills/skill-evolution-advisor/scripts/learnings.py log --category user-preference --text "..." --observation "<short verbatim correction>" --source user --skill report-author --operation weekly
python3 .agents/skills/skill-evolution-advisor/scripts/learnings.py recall --kind all --limit 5
python3 .agents/skills/skill-evolution-advisor/scripts/learnings.py recall --kind defects

The compatibility review and promote subcommands reject user-preference entries without writing. Use the unified kb review snapshot for both confirmation and dismissal.

Deep retrospective:

python3 .agents/skills/skill-evolution-advisor/scripts/create_retrospective.py --slug routing-gap --skill unit-analyst --target-skill research-orchestrator --task-summary "..." --observed-issue "..." --suggestion "..." --stdout-prompt --root kb/memory/skill-evolution

默认落盘路径为 kb/memory/skill-evolution/retrospectives/<timestamp>-<slug>.md；--stdout-prompt 会在写入后同时打印可交给后续 agent 的改进 prompt。

Local diagnostics (owner-only; never paste these commands into user-visible output):

${RESEARCH_PYTHON:-python3} .agents/skills/skill-evolution-advisor/scripts/diagnostics.py policy --skill unit-analyst
${RESEARCH_PYTHON:-python3} .agents/skills/skill-evolution-advisor/scripts/diagnostics.py record --category skill-defect --severity medium --skill unit-analyst --summary "short safe summary" --source user
${RESEARCH_PYTHON:-python3} .agents/skills/skill-evolution-advisor/scripts/diagnostics.py list --status pending
${RESEARCH_PYTHON:-python3} .agents/skills/skill-evolution-advisor/scripts/diagnostics.py detail --id diag-...
${RESEARCH_PYTHON:-python3} .agents/skills/skill-evolution-advisor/scripts/diagnostics.py apply-retrospective --id diag-... --expected-detail-digest <sha256> --analysis-file diagnostic-analysis.json
${RESEARCH_PYTHON:-python3} .agents/skills/skill-evolution-advisor/scripts/diagnostics.py review --id diag-... --status resolved
${RESEARCH_PYTHON:-python3} .agents/skills/skill-evolution-advisor/scripts/diagnostics.py export-preview --authorized

detail and apply-retrospective are private Agent-owner steps, never user-facing commands. The analysis file must be a bounded current-user 0600 regular JSON file below kb/.runtime/ with exactly explanation, reproduction, optimization_candidates, and next_validation; it must not contain raw child output, user/source/evidence text, secrets, environment values, or absolute paths.

启动澄清（Agent 用）

• 记录类型：习惯、坑还是 skill 缺陷？默认按内容分类，skill 缺陷仅记录。
• 只记录还是同时生成复盘 prompt？默认只记录为 pending 待复审。