Agent self audit

Bootstrap (pre-flight)

Before running any checks, execute the bootstrap sequence. Read ~/.claude/skills/self-audit/bootstrap.md for platform detection, command mapping, writability check, and first-run initialization. On T2 termination or T1 degradation, adjust audit mode accordingly before proceeding.

Self Audit v2

Mode Selection

0. Self-Audit Integrity. Run wc -l ~/.claude/skills/self-audit/SKILL.md. If >300 lines → WARNING: self-audit SKILL.md itself exceeds limit. Report as F-SLF-001.

When self-audit is invoked, determine mode by checking trigger words AND prior audit count:

Trigger Words	Prior Audits	Mode
quick audit / health ping	any	Quick
full audit / deep audit / weekly audit	any	Full
self-optimization / config check / token efficiency / skill management / optimize yourself / config review	<3 total prior audits	Quick
self-optimization / config check / token efficiency / skill management / optimize yourself / config review	>=3 total prior audits	Full
no explicit mode word	<5 prior audits	Quick
no explicit mode word	>=5 prior audits	Full

To determine prior audit count, read ~/.claude/memory/audit-log.md and count lines matching "| <date> |" format under the "## Pipeline Run Log" section.

Quick mode: items 1-6 + item 10-Quick (4 atomic env checks). Full mode: items 1-13 + trend analysis + predictive warnings.

Anti-Hallucination Gate

Read ~/.claude/skills/self-audit/anti-hallucination-gate.md before emitting ANY finding. All 4 gates (AH-1 through AH-4) must pass. Failure = suppress the finding.

Agent Roles

Boss (协调/决策/视觉审查) · Yushi (流程合规督查) · Gongjiang (代码/文档生成) Full roster & collaboration modes: ~/.claude/rules/team.md

Dynamic Paths

Resolve ~ at runtime. Core paths (full list in bootstrap.md Section 4):

• Main config: ~/.claude/CLAUDE.md, ~/.claude/settings.json
• Rules: ~/.claude/rules/coding.md, ~/.claude/rules/team.md, ~/.claude/rules/pitfalls-archive.md
• Memory: ~/.claude/memory/ (audit-log, environment, skill-usage-inventory, .archive/)
• Agents: ~/.claude/agents/yushi.md
• Skills: ~/.claude/skills/, ~/.agents/skills/

Overall Grade

Computed from worst finding across all items. ACCEPTED findings excluded.

Grade	Condition
A	0 Critical, 0 Warning. All items OK or N/A.
B	0 Critical, 1-2 Warnings. Minor issues, no systemic risk.
C	1 Critical OR 3+ Warnings.
D	2-3 Critical. Multiple serious issues.
F	4+ Critical, OR any item FAIL, OR F-SLF-001. Systemic failure.

Baseline Load

Before running checks, read audit-log entries. Branch by count:

• 0 entries: Bootstrap. Output 'No prior audits. Trends unavailable.' Set growth rates to 0.
• 1-2 entries: Partial trend. Compute from available entries. Flag limited sample.
• 3+ entries: Full trend. Compute (current - previous_avg) / previous_avg. Flag if any metric grew >10% per audit.

Extract: CLAUDE.md line count, total skill count, Critical/Warning/Suggestion counts from each entry.

Five-Type Taxonomy

Every finding MUST carry exactly one taxonomy tag AND one Finding ID. This enables structured trend analysis and reliable repetition detection across audits.

Tag	Meaning	Example
Correction	Defect that needs immediate fix	Broken symlink, missing file, API key plaintext
Repetition	Same Finding ID appeared in prior audit	CLAUDE.md growing same section as last audit
Role Redirect	Work done by wrong agent/role	Boss editing code directly, bypassing Gongjiang
Frustration Escalation	User had to ask >1 time for same thing	Same request repeated across messages
Workaround	Ad-hoc fix that should be structural	copy-paste validation instead of hook

Finding ID System

Check Item	Prefix	IDs
1. CLAUDE.md	F-CMD-	001: exceeding threshold, 002: duplicate content (Jaccard >0.45), 003: compression candidate (prose >0.50), 004: fragment section <15 lines
2. Skill Health	F-SKL-	001: duplicate across sources, 002: broken symlink, 003: >25 total, 004: stale+better alt exists (Item 13), 005: active+better alt exists (Item 13), 006: skill category gap (Item 13)
3. Security	F-SEC-	001: plaintext API key, 002: skipDangerousPrompt, 003: unverified plugin
4. Memory	F-MEM-	001: MEMORY.md >200 lines, 002: missing frontmatter, 003: stale files
5. Model Routing	F-MDL-	001: identical tiers, 002: outdated model names
6. Updates	F-UPD-	001: >3 outdated skills, 002: security update, 003: breaking change
7. Skill Utilization	F-USE-	001: zero-usage >30d, 002: similar skills >2, 003: total >25
8. Archive	F-ARC-	001: restore candidate, 002: stale archive
9. Yushi Audit	F-YUS-	001: spot-check fail, 002: oversight quality decline, 003: missed SKILL VIOLATION
10. Environment	F-ENV-	001: LibreOffice missing (Quick), 002: encoding conflict (Quick), 003: disk <20GB (Quick), 004: network unreachable (Full), 005: critical pkg missing (Full), 006: RAM <8GB (Quick), 007: stale baseline >90d (Full)
11. Solo Ratio	F-SOL-	001: ratio >20%, 002: rising trend, 003: >50% critical
12. Skill Violation	F-SKP-	001: same skill skipped 2+ times, 002: Protocol Gate text missing
--. Bootstrap	B-	001: yushi agent not found (suggestion only, not a finding)
Rule: Every finding MUST include its Finding ID. For repetition, match by ID (e.g., F-MDL-001), not by prose similarity.

Accepted Exceptions

Some findings are known and deliberately accepted by the user. These should NOT trigger repetition counting or promotion.

Before counting repetitions, read ~/.claude/memory/audit-log.md and check the "## Accepted Exceptions" table. Skip any finding whose ID appears there.

Users add exceptions by editing audit-log.md directly:

| F-SEC-001 | API key plaintext | Accepted 2026-05-16 | User confirmed: not changing |

If a finding is in Accepted Exceptions, report it as "ACCEPTED (not counted)" with severity reduced to Suggestion.

Checks

1. CLAUDE.md Size & Trend

Quick: wc -l ~/.claude/CLAUDE.md — line count only. Do NOT read the file. (12 tokens) Full: Read companion ~/.claude/skills/self-audit/compress-config.md. Parse CLAUDE.md sections by ## headers. Compute 5-dim health score: D1 Section Balance(25%), D2 Structure Granularity(20%), D3 Prose Density(25%), D4 Reference Ratio(15%), D5 Line Count Modifier(15%). A-F grade per config thresholds. Jaccard dedup check (>=0.45, excl list). Min-section gate (<15 lines → F-CMD-004). 3-audit trend, +30d predict. Emit F-CMD-001004 per trigger conditions.

Symptom	Cause	Verify	Fix
Growth +10-20/wk	Yushi depositing pitfalls	Grep "Pitfall" in CLAUDE.md	caveman-compress; cap at 20
Spike +30+	New feature/rules added	Grep new "## " headers	Split to domain file under rules/
Jaccard >0.45 (check excl list in config)	Duplicate content	Diff sections	Merge; keep specific version
Fragment section <15 lines	Over-granular ## split	Check section count vs total lines	Merge with adjacent same-domain section

2. Skill Health

Quick: ls ~/.claude/skills/ | wc -l + ls ~/.agents/skills/ | wc -l. Count only. Do NOT read lockfile. Do NOT check duplicates. (~15 tokens) Full: Read lockfile. Count enabled/disabled. Stale-skill detection (>60d unused). Overlap analysis between user-level and plugin-provided skills. Taxonomy tag, Finding ID.

3. Security Issues

Quick: grep -c "sk-" ~/.claude/settings.json (plaintext key check) + grep "skipDangerousModePermissionPrompt" ~/.claude/settings.json (dangerous mode check). Do NOT read full file. Two-line output. (~30 tokens) Full: Read full settings.json. Plugin source verification. Permission audit. Taxonomy tag, Finding ID.

4. Memory System

Quick: ls ~/.claude/memory/ | wc -l (file count). wc -l ~/.claude/memory/MEMORY.md (MEMORY.md size). Do NOT read file contents. (~20 tokens) Full: Read all memory files. Per-file staleness report (>60d). Frontmatter completeness check. Memory-to-pitfall cross-reference. Taxonomy tag, Finding ID.

5. Model Routing

Quick: grep "ANTHROPIC_DEFAULT_" ~/.claude/settings.json — check if 3 tiers are distinct. Do NOT read full file. (~15 tokens) Full: Read full settings. Cost-efficiency analysis. Model allocation optimization suggestions. Taxonomy tag, Finding ID.

6. Plugin & Skill Updates

Quick: npx skills check 2>&1 || echo N/A — count only. Do NOT run update. (~20 tokens) Full: Run full check + update. Per-plugin changelog. Breaking-change warnings. Update priority order. Taxonomy tag, Finding ID.

7. Skill Utilization

Full only (items 7-11 are Full mode only). Read ~/.claude/memory/skill-usage-inventory.md. Cross-reference installed skills vs invocation count. Flag skills with 0 usage after 30d, flag >2 similar skills, flag total >25. Add invocation trend (3-audit), per-skill ROI estimate, consolidation recommendations. Taxonomy tag, Finding ID.

Symptom	Cause	Verify	Fix
Zero usage after 30d	Wrong domain, never triggered	Check invocation log	Disable or force first use
Similar skills >2	Install without comparison	Compare SKILL.md files	Consolidate

8. Archive Recovery

Full only. List ~/.claude/memory/.archive/ (archived memory count). Read ~/.claude/rules/pitfalls-archive.md (archived pitfalls). Flag archived items recurring in recent conversations. Add restoration candidate report, archive age distribution, cross-reference with active pitfalls. Taxonomy tag, Finding ID.

9. Yushi Audit

Full only. Pre-check: test -f ~/.claude/agents/yushi.md. If missing → SKIP, queue Bootstrap Suggestion B-001.

Read latest Yushi report from audit-log or conversation history. Randomly pick 3 items Yushi claimed to check. Personally re-verify each item: Did Yushi actually read the required files? Are conclusions consistent with data? Report PASS/FAIL per item.

Full only. Spot-check Yushi work quality. If >1 FAIL -> flag as Yushi oversight quality decline. Taxonomy tag.

10. Environment Snapshot

Quick (4 atomic checks, ~65 tokens): Run inline checks for (a) disk space <20GB → F-ENV-003, (b) RAM <8GB → F-ENV-006, (c) LibreOffice missing → F-ENV-001, (d) PYTHONIOENCODING unset/GBK → F-ENV-002. Use python one-liners for disk/RAM, test -f for LibreOffice path, echo $PYTHONIOENCODING for encoding. Output: 4-line status. See ~/.claude/skills/self-audit/environment-checks.md for exact commands.

Full: Read ~/.claude/skills/self-audit/environment-checks.md. Execute: version parsing (8 tools), pip package check (7 packages), network connectivity (read ANTHROPIC_BASE_URL from settings.json → TCP connect), baseline comparison (load environment-baseline.md, diff against current). Flag: F-ENV-001~007 per trigger conditions. Write new snapshot to ~/.claude/memory/environment.md if tool versions changed or disk/RAM ±20%.

11. Solo vs Pipeline Ratio

Scan current session transcript. Count Boss direct Write/Edit calls vs Gongjiang invocations. Compute solo ratio = Boss edits / (Boss edits + Gongjiang invocations).

Full only. Report ratio with verdict: PASS (<=20%), WARNING (20-50%), FAIL (>50%). Cross-reference against team.md Gate Check rules.

Symptom	Cause	Verify	Fix
Solo ratio >20%	Small tasks degraded to solo	Check transcript for Edit/Write	Enforce gate check
Solo ratio rising trend	Pipeline avoidance habit	Compare 3-audit trend	Add PreToolUse hook reminder

12. Yushi F-SKP Cross-Validation (F-YUS-)

Full only. Pre-check: test -f ~/.claude/agents/yushi.md. If missing → SKIP (B-001 already queued by Item 9).

Full only. Cross-validates Yushi detects SKILL VIOLATION correctly. Does NOT execute promotions.

Procedure:

1. Read audit-log.md, scan last 5 Pipeline Run Log entries
2. For each entry with "SKILL VIOLATION" or "skill" in Key Finding → check if Yushi produced F-SKP entry in same run
3. If SKILL VIOLATION present but no Yushi F-SKP entry → F-YUS-003(WARN): "Yushi missed SKILL VIOLATION in run <timestamp>"
4. Do NOT classify violations or execute promotions. This item only validates Yushi detection coverage.

Distinct from Check 9: Check 9 random-spot-checks Yushi's general oversight quality; this check specifically validates SKILL VIOLATION detection coverage.

Boundary: F1 gate (Step 3 + Yushi Promotion Review) handles promotion execution. This item only checks coverage.

13. Skill Market Intelligence

Full only. 24h TTL-gated market scan. Read ~/.claude/skills/self-audit/market-intelligence.md. Scan 3 skill marketplace repos via GitHub API (gh api), score repos (stars 35% + recency 30% + activity 20% + maintenance 15%), cross-reference installed skills vs market. Emit F-SKL-004 (stale+better alt), F-SKL-005 (active+better alt), F-SKL-006 (category gap). Cache results in ~/.claude/memory/skill-market-intelligence.json. Skip if gh auth status fails (no GitHub auth) or network unavailable without cache. Recommendations only — never auto-install.

Repetition Detection & Structural Promotion

Load Accepted Exceptions

Read audit-log "## Accepted Exceptions" → build skip-list.

Match Findings by ID

Match each finding by Finding ID against prior 3 audits. ID in skip-list → ACCEPTED (Suggestion).

Flag for Promotion (3rd repetition)

1st → Suggestion. 2nd → Warning. 3rd → CRITICAL (PROMOTION PENDING):

PROMOTION PENDING: <Finding ID> (<description>) — 3rd repetition on <dates>.
Gate: Yushi → Boss confirm → write to coding.md.

Append to audit-log "## Promotion Pending" table. Yushi handles execution; do NOT write directly to coding.md.

Cross-Item Conflict Detection

1. CLAUDE.md bloat vs memory starve: MEM<5 & CLAUDE>150 → content aggregating in CLAUDE.
2. Skill inflation: Skills +>3/30d & 0 utilization → install without use. Cap/disable.
3. Solo + no safety prompt: Solo >30% & skipDangerous → dangerous autonomy. Require gate-check prompt.

Output Format

All findings emitted in ONE batched report. No one-at-a-time output.

# Self Audit Report --- YYYY-MM-DD HH:MM
**Mode**: [Quick | Full]
**Overall**: [A/B/C/D/F]

## Findings
| # | Item | Rating | Taxonomy | Detail |
|---|------|--------|----------|--------|
| 1 | CLAUDE.md Size | OK/WARN/CRIT | <tag> | <line count> lines, <summary> |
| 2 | Skill Health | ... | ... | ... |
...through item N...
## Trend Summary
- CLAUDE.md: <count>-><count>-><count> (3 audits), direction [growing/stable/shrinking]
- Skills: <count>-><count>-><count>, direction [...]
- Solo ratio: <pct>-><pct>-><pct>, direction [...]
- Critical findings: <count>-><count>-><count>
## Predictive Warnings
[Full mode only] Based on trends, predict issues in next 30d:
- <item>: expected to reach <threshold> by <date> --- pre-emptive fix: <action>
## Structural Promotions
[Show what was promoted, to what rule, written where]
- Before: <finding> appeared in audit N-2, N-1, N as <severity>
- After: promoted to hard pitfall #X in rules/coding.md
- Documented: audit-log.md "## Structural Promotions" table
## Bootstrap Suggestions
[Conditional: appears when yushi.md missing. Provides setup guidance + minimal template path + dismissal instructions. See bootstrap.md Section 6.]
## Managed Items (no longer checked)
[List findings previously promoted. These are now hard rules, not audit items.]
- pitfall #N: <issue> — promoted on <date>
## Cross-Item Conflicts
[If conflicts detected]
- <conflict type>: <detail> -> <resolution>
## Actions Taken
- [List fixes actually executed this run]

Recording

1. Append ONE line to ~/.claude/memory/audit-log.md under "## Pipeline Run Log":

   | YYYY-MM-DD HH:MM | self-audit (<mode>) | self-audit | <PASS/FAIL> | <N> findings | <key finding IDs> |
   

   Include Finding IDs in the key finding column: e.g., "F-CMD-001(CRIT), F-SEC-001(ACCEPTED)"

2. If environment changed, update ~/.claude/memory/environment.md with new snapshot.

3. If any 3rd-repetition promotion was FLAGGED (PROMOTION PENDING): a. Append to "## Promotion Pending" in audit-log.md (already done in Step 3) b. Yushi handles execution (see yushi.md "Promotion Gate") c. After Yushi executes, VERIFY pitfall was written to coding.md with grep

4. If any finding was marked ACCEPTED, ensure its Finding ID is in "## Accepted Exceptions" in audit-log.md.

5. If the "## Accepted Exceptions" section does not exist in audit-log.md, create it with header:

   ## Accepted Exceptions
   | Finding ID | Description | Accepted Date | Reason |
   |------------|-------------|---------------|--------|
   

SkillOpt-Sleep Integration

When SkillOpt-Sleep is installed, self-audit serves as the health scoring bridge for its validation gate.

How it works

1. SkillOpt-Sleep proposes an edit to CLAUDE.md or a SKILL.md
2. Before gating, it calls python ~/.claude/skills/self-audit/health_score.py
3. Self-audit returns a JSON health score: {"score": 0.75, "grade": "B", ...}
4. The gate combines: combined = 0.6 x task_replay_score + 0.4 x health_score
5. Edit is accepted only if combined score > baseline

Scoring table

Self-Audit Grade	Numeric Score	Meaning
A	1.0	All items OK
B	0.8	Minor warnings
C	0.6	1 critical or 3+ warnings
D	0.4	2-3 criticals
F	0.0	4+ criticals or systemic failure

Each Critical finding: -0.1 (max -0.3). Each Warning: -0.05 (max -0.15).

Manual invocation

python ~/.claude/skills/self-audit/health_score.py --verbose

Handoff

希小台 Handoff

• From: self-audit
• To: Boss
• Key Output: Self-audit report with <N> findings
• File: tap-comms/handoff/selfaudit-to-boss.md