verifier-pass2 - SKILL.md Agent Skill

name: verifier-pass2 description: | Two-pass subagent validation. Pass-1 (Code Review agent) flags findings; Pass-2 (this skill) validates each finding adversarially with high confidence before action. Two-pass pattern modeled after Anthropic's official code-review plugin. Triggered by: pass-2 verify, validate review, second pass, adversarial verify, claim check license: MIT metadata: version: "1.0.0" author: mykcs category: meta-review created: 2026-06-02 source: rich-audit Layer 3 evolution (Context7) pattern_origin: anthropics/claude-code official code-review plugin tags: - review - verification - meta-cognition - two-pass - adversarial user-invocable: true

verifier-pass2 Skill

Purpose

Given a list of findings (from any reviewer agent or audit tool), adversarially verify each one with high confidence. Pass-1 surfaces, Pass-2 validates.

When to Use

IF	THEN
Code reviewer returned > 3 findings	Run pass-2 on each before fixing
Audit tool flagged MED/HIGH issues	Run pass-2 to confirm before acting
User asks "is this finding real?"	Single-finding mode
Reviewer said "this is wrong" but evidence is weak	Adversarial re-check

When NOT to Use

Single trivial typo (overhead > benefit)
Findings the user already explicitly accepted
L1 PHANTOM / L2 MISSING (these are mechanical, not interpretive)

Algorithm

for finding in findings:
    # 1. Re-read the cited file:line with fresh eyes
    # 2. Reproduce the failure mode if possible (run, query, etc.)
    # 3. Check for: false positive, regression risk, alternative explanation
    # 4. Output verdict: { real: bool, confidence: high|medium|low, evidence: [...] }

Output Format

Per-finding JSON:

{
  "id": "F001",
  "original_finding": "...",
  "verdict": {
    "real": true,
    "confidence": "high",
    "evidence": [
      "file:line — quoted text",
      "reproduction: ran `cmd`, observed X"
    ],
    "alternative_explanations_considered": [
      "Could be X — but Y rules it out"
    ]
  },
  "actionable": true,
  "suggested_action": "..."
}

Adversarial Bias

Default to real: false if uncertain. Most audit findings are noise.

Common false positive patterns to watch for:

Path mentions in comments (the audit tool flagging git add -A text in a comment)
Glob patterns in documentation (~/.claude/rules/behavioral-*.md is a pattern, not a missing file)
Stale "as of YYYY-MM" references in skill metadata
Frontmatter examples that don't reflect current usage

Example: 2026-06-02 rich-audit pass-2

#	Finding (pass-1)	Verdict (pass-2)	Why
F001	`smart-push.sh` uses `git add -A`	FALSE POSITIVE (initially) → TRUE (after external fix)	Code at L432 was `git add -- $ALL_FILES`; L435 was the `git add -A` fallback that was later changed to `git add -u` by an external linter
F002	`CASE-HEALER-...` references non-existent `CASE-SKILL-DIR-UNIFICATION-20260527`	TRUE	File genuinely missing; replaced with `...-CONSOLIDATED`
F003	`behavioral-workflow.md` references `behavioral-*.md` glob	FALSE POSITIVE	Glob pattern is intentional documentation convention
F004	L2 memory missing = 84	FALSE POSITIVE	BSD sed `\1` bug; Python check returns 0

Integration

Pair with code-review agent: code-review finds → verifier-pass2 validates → only validated findings get fixed.

Limits

This skill does not fix anything. It only validates.
For complex reproductions (e.g. "does this CV cache bust actually work?"), invoke verify skill instead.
Do not chain more than 2 passes (pass-3 = diminishing returns).