harness-analyzer

name

harness-analyzer

description

Analyze all 14 SEA harness zone families deterministically and produce a HarnessDiagnosis JSON artifact. Use when you want to understand the current structural health of the agent harness before deciding whether to mutate it. Trigger words: "audit harness for evolution", "diagnose harness", "analyze harness", "harness health check", "what zones need mutation", "harness score". Do NOT trigger for general harness questions — use for evolution-cycle analysis only.

harness-analyzer Skill

Purpose

Produces a HarnessDiagnosis JSON by running deterministic structural checks across all 14 SEA harness zone families. This is the first step in the slow-loop evolution cycle.

Role in the two-loop model:

fast loop  →  task execution (unchanged)
slow loop  →  [harness-analyzer] → harness-transformer → harness-candidate-pr-loop

Trigger

Use this skill when the user says:

• "audit harness for evolution"
• "diagnose harness"
• "analyze harness health"
• "what zones need mutation"
• "harness readiness score"

Execution Steps

1. Resolve workspace root using $CLAUDE_PROJECT_DIR (never use parents[N] path hacking):

   import os
   harness_dir = os.environ['CLAUDE_PROJECT_DIR']
   

2. Run the analyzer:

   python ${CLAUDE_PLUGIN_ROOT}/scripts/harness_core/analyzer.py \
     --harness-dir $CLAUDE_PROJECT_DIR \
     --json-output ${CLAUDE_PLUGIN_ROOT}/state/harness-report.json
   

3. Parse the output and produce a HarnessDiagnosis struct (see schemas/diagnosis.json).

4. Emit HARNESS_EVOLUTION_STATE.md using the state template if the diagnosis reveals mutation_pressure >= 0.3 or any critical/high severity findings with mutation_candidate: true.

5. Report a concise summary:

   • Overall score
   • Top 3 findings by severity
   • Whether mutation is recommended (pressure threshold)
   • Zone scores for zones with score < 0.8

Output Format

Harness Diagnosis Summary
─────────────────────────
Overall score:     0.742  (threshold for mutation: < 0.85)
Mutation pressure: 0.41   (threshold: >= 0.30 → recommend mutation)

Top findings:
  🔴 [governed-speed] gs-001: services/policy-gateway/ is missing  [CANDIDATE]
  🟠 [invariants] inv-003: 2 invariant file(s) not in REF-012 §10  [CANDIDATE]
  🟡 [thin-contract] thin-002: AGENTS.md is 267 lines (threshold: 250)  [CANDIDATE]

Zone scores (< 0.80):
  governance:           0.65
  surface_coverage:     0.75

Recommendation: mutation recommended → route to harness-transformer

Constraints

• Read-only: This skill must not modify any harness files.
• No LLM mutation: The analyzer is deterministic Python — it does not use LLM judgment to score zones in V1.
• Workspace root via env: Always use $CLAUDE_PROJECT_DIR, never compute relative paths.
• Max retries: 1: If the analyzer script fails, retry once with --quiet flag. If it fails again, report the error and stop.

Error Handling

If analyzer.py exits non-zero:

1. Report the stderr output to the user.
2. Do NOT produce a partial diagnosis.
3. Do NOT route to harness-transformer — a failed diagnosis is not a valid trigger for mutation.

Next Skill

If mutation_pressure >= 0.30, route to harness-transformer skill. Otherwise, report "harness is healthy, no mutation needed" and stop.