harness-audit

name	harness-audit
description	Check system setup health across 7 categories. Use for /harness-audit, "verify my setup", "check harness". NOT for project quality (/audit) or eval scoring (/eval-score).
model	sonnet
effort	low
allowed-tools	Read, Bash, Glob, Grep

Harness Audit: Control Plane Health Score

EXECUTE this skill now. Follow the workflow steps below using the provided $ARGUMENTS. Do NOT describe, summarize, or explain this skill — run it.

Constants

• CLAUDE_DIR: C:/Users/gurusharan.gupta/.claude
• REGISTRY_DIR: C:/Users/gurusharan.gupta/Agents/Claude Code

---

Category 1: Tool Coverage (4 checks)

1. Skills installed — ls ~/.claude/skills/ has ≥ 8 skill directories
2. Core skills present — scan, audit, research, execplan, init-project all exist in ~/.claude/skills/
3. Bin scripts valid — bin/scan.sh and bin/install.sh exist and are executable
4. Manifest exists — manifest.json present and valid JSON (check with python -m json.tool manifest.json)

Category 2: Context Efficiency (4 checks)

5. Global CLAUDE.md size — word count ≤ 500 words (token efficient)
6. Project CLAUDE.md size — word count ≤ 300 words if present
7. Skills have context:fork — count skills with context: fork in frontmatter; PASS if ≥ 70% of exploration/registry skills have it (scan, audit, dashboard, register, context-budget, verify, save-session = 7; PASS if ≥ 5 have it)
8. Skill size limit — no individual SKILL.md exceeds 200 lines

Category 3: Quality Gates (4 checks)

9. Model frontmatter — all SKILL.md files have a model: field in frontmatter
10. Effort frontmatter — all SKILL.md files have an effort: field
11. Eval criteria exist — eval/criteria/ has ≥ 1 non-template JSON file
12. Principles registered — principles/_index.json exists and has ≥ 3 entries

Category 4: Session Persistence (4 checks)

13. save-session skill — ~/.claude/skills/save-session/SKILL.md exists
14. resume-session skill — ~/.claude/skills/resume-session/SKILL.md exists
15. session-data dir — ~/.claude/session-data/ directory exists (create if missing)
16. Context budget skill — ~/.claude/skills/context-budget/SKILL.md exists

Category 5: Eval Coverage (4 checks)

17. Eval scaffolds — eval/scaffolds/ has ≥ 1 directory
18. Eval history — eval/history/ has ≥ 1 score entry
19. Criteria coverage — each workflow in workflows/ has a matching criteria file in eval/criteria/
20. Score recency — most recent score in eval/history/ is within 30 days (check file mtime)

Category 6: Security Guardrails (4 checks)

21. Hooks managed-only — ~/.claude/settings.json does not contain any hooks key (managed-only env)
22. Bypass disabled — settings.json does not have bypassPermissionsMode: true
23. MCP whitelist — if .mcp.json exists, all servers are in the known whitelist (context7, playwright, or other explicitly approved)
24. No secrets in skills — no SKILL.md file contains patterns matching (password|secret|api_key|token)\s*=\s*['\"][^'\"]{8,}

Category 7: Cost Efficiency (4 checks)

25. Mechanical skills use Haiku — scan, register, dashboard, context-budget, save-session, verify all have model: haiku
26. Model routing documented — ~/.claude/CLAUDE.md contains either a model routing table OR a reference to rules/performance.md (check for "rules/" or "model routing" or "Model Routing" in the file)
27. Subagent model set — settings.json has CLAUDE_CODE_SUBAGENT_MODEL=haiku in env block
28. Token cap set — settings.json has MAX_THINKING_TOKENS in env block

---

Scoring Workflow

For each check:

• Use Read/Glob/Bash/Grep to evaluate the condition
• Record PASS or FAIL with a one-line reason

After all checks, compute:

• Per-category score: (passed / 4) × 10 → 0–10
• Overall score: average of 7 category scores → 0–10

Output Format

## Harness Audit Report — <date>

| # | Category | Score | Checks |
|---|----------|-------|--------|
| 1 | Tool Coverage | X/10 | 4 PASS / 0 FAIL |
| 2 | Context Efficiency | X/10 | 3 PASS / 1 FAIL |
| 3 | Quality Gates | X/10 | ... |
| 4 | Session Persistence | X/10 | ... |
| 5 | Eval Coverage | X/10 | ... |
| 6 | Security Guardrails | X/10 | ... |
| 7 | Cost Efficiency | X/10 | ... |

**Overall Score: X.X / 10**

### Failures
<list each failed check with: check name, what was found, how to fix>

### Grade
- 9.0–10: Excellent — control plane is production-ready
- 7.0–8.9: Good — minor gaps, address before scaling
- 5.0–6.9: Fair — notable gaps affecting efficiency
- < 5.0: Poor — critical issues require immediate attention

Stop and report even if some checks error (e.g., file missing). A missing file is a FAIL, not an error.