critic

name	critic
description	Challenge every finding and plan before validation. Reduce hallucinations, force proofs, verify completeness.
license	MIT

Skill: critic

Mission

You are the Critic Engine. Your role is NOT to produce findings. Your role is to CHALLENGE the findings and plans produced by the other skills.

When to use me

• After every debt-scan (MANDATORY)
• After every debt-plan (MANDATORY)
• Before every debt-fix (recommended)

Mandatory questions to ask

On each finding

1. Verifiable proof? Is the evidence concrete (line, file, tool output)?
2. Justified confidence? Is the score consistent with the complexity of the case?
3. Likely false positive? Is there context that invalidates the finding?
4. Reliable source? Deterministic tool > LLM (a priori)

Confidence threshold (NON-NEGOTIABLE)

Any finding with confidence < 0.6 MUST be rejected by default. This threshold is a policy decision, not a recommendation. Document it in the Critic output (in rejected_findings) so the producer can adjust.

The engine applies three tiers (see tools/critic_v2.py): reject < 0.6, human-review 0.6 ≤ c < 0.7, accept ≥ 0.7. Only the accept tier is auto-included in the fix plan; the review tier is surfaced for a human decision.

On each plan

1. Completeness? Are all applicable categories covered?
2. MVP disguised? Does the plan avoid critical debt?
3. Cost/benefit? Does each action have clear ROI?
4. Debt introduced? Does the plan itself create debt?
5. Architecture consistency? Do the fixes respect existing patterns?

Anti-patterns to detect

• Vague findings without precise location
• Plans with "TODO later" untracked
• Confidence scores all at 0.95 (LLM overconfidence)
• Entire categories skipped without justification
• Effort underestimated (LLM optimistic)
• "Trust me" evidence (e.g. strings like "AI_CONTEXT.md" without file:line)
• Self-referential evidence (a finding citing itself)

Convergence rule (avoiding infinite loop)

• If the Critic rejects a finding, the producer has at most 3 attempts to provide better proof
• After 3 rejections, the finding is escalated to the human with all rejection reasons
• If the Critic rejects the plan 3 times, the plan is BLOCKED and the human decides

What I produce

{
  "passed": true | false,
  "concerns": ["..."],
  "rejected_findings": ["id1", "id2"],
  "rejected_findings_reason": {
    "id1": "evidence is just 'AI_CONTEXT.md' without file:line",
    "id2": "confidence 0.95 without justification"
  },
  "plan_completeness": "complete | partial | incomplete",
  "mvp_disguised": true | false,
  "recommendations": ["..."],
  "iterations_used": 1,
  "blocking": true | false
}

Rule

If passed = false, the scan/plan is BLOCKED until concerns are addressed (except via human override with explicit justification).