AgentSkillsCN

Review Memory Coherence

审查记忆一致性

SKILL.md
--- frontmatter
id: review-memory-coherence
summary: Validates history entry coherence (type correctness, topic unity, linking).

ENVIRONMENT: Use only cat/grep/echo. DO NOT run npm/node commands.

TASK: Check history entry coherence.

FILES:

  • DIFF.txt: Code changes being committed
  • ENTRIES.txt: History entries (marked as [TYPE])

RULES:

  1. ENTRY TYPE CORRECTNESS:

    • "fix"/"incident" entries are for BUGS/FIXES (something broke, we fixed it)
    • "decision"/"feature"/"refactor"/"investigation"/"note"/"meta" are for FEATURES/CHANGES
    • If a fix/incident entry describes a NEW FEATURE → flag as "wrong_entry_type"
    • If a non-fix entry describes a BUG FIX → flag as "wrong_entry_type"
  2. TOPIC COHERENCE:

    • Each entry should cover ONE logical change
    • If entry mixes multiple UNRELATED changes → flag as "multiple_topics"
    • Exception: Related changes (e.g., fix + test for that fix) are OK together
    • If multiple topics are properly linked via "## Related" section → OK
  3. Check each entry and report issues.

MANDATORY: Produce COHERENCE.json as a JSON object (no extra text): { "entry_count": 5, "issues": [ { "file": "path/to/entry.md", "issue_type": "wrong_entry_type | multiple_topics | missing_links", "description": "brief description", "suggestion": "brief fix" } ], "all_coherent": true }

  • entry_count: integer count of entries checked (DO NOT list file paths)
  • If no issues found, set all_coherent=true and issues=[]
  • Keep descriptions and suggestions BRIEF (under 50 chars each)
  • Be pragmatic: minor bundling of closely-related fixes is fine