ql-deep-review — whole-feature post-implementation review

Purpose

quantum-loop's built-in two-stage review gate (ql-review: spec-compliance → code-quality) operates on ONE story at a time inside ONE worktree. It does not detect:

1. Cross-story divergence (e.g., story A uses 'google' as a secret key while story B uses 'google-api-key' for the same constant).
2. Post-merge regressions (test that was green in isolation breaks after integration).
3. Drift from original user intent (paraphrase chain from intent → design → PRD → plan → code).
4. Low-signal comments that look like findings but lack evidence (CRA actionability is 0.9-19.2% per Chowdhury 2604.03196; human baseline is ~60%).

ql-deep-review closes these gaps with a whole-feature review that runs AFTER all stories in a wave / feature pass the per-story gate.

When to use

• After ql-execute emits COMPLETE for a wave and before merging the feature branch to master.
• After cherry-picking or merging a foreign branch whose conflict-resolution changed semantics.
• Manually, when suspicion of cross-story drift is high (e.g., follow-on work after a long autonomous run).

What it does NOT do

• Does not replace the per-story two-stage gate. Run ql-review per story, then this.
• Does not auto-fix findings. Produces a structured report; user or orchestrator drives action.
• Does not block merge autonomously. Emits a verdict + confidence; user decides.

Risk scoring (0-100)

Risk factors and weights (inspired by soliton's risk-adaptive dispatch):

Score → dispatch tier:

• 0-30 LOW: 2 reviewers (code-reviewer, synthesizer). Target turnaround 2-3 min.
• 31-60 MEDIUM: 4 reviewers (+security-reviewer, test-engineer). 4-6 min.
• 61-80 HIGH: 6 reviewers (+critic, architect). 6-10 min.
• 81-100 CRITICAL: 7 reviewers (+cross-provider critic using codex or gemini via /ask). 10-15 min plus manual inspection.

Reviewer agents (dispatched per tier)

All agents are invoked via the Agent tool in parallel. Each receives:

• BASE_SHA..HEAD_SHA — whole-feature diff scope
• PRD_PATH — path to the feature PRD
• STORY_LIST — JSON list of stories executed with their IDs and status
• INTENT_SNAPSHOT — verbatim user intent (from quantum.json.userIntent if present)
• CHANGED_FILES — file-list manifest

Tier-core reviewers (always dispatched)

• oh-my-claudecode:code-reviewer — severity-rated findings (CRITICAL / HIGH / MEDIUM / LOW) with line-level evidence.
• soliton:synthesizer — risk-adaptive PR-style review; contributes a reviewer-side risk score and a READY_TO_MERGE / NEEDS_REWORK / BLOCKED verdict.

Tier-MEDIUM additions

• oh-my-claudecode:security-reviewer — OWASP Top 10 + secret exposure + input validation; hard-dispatched when sensitive-paths factor > 0.
• oh-my-claudecode:test-engineer — test-quality audit: AC-to-test mapping, over-mock detection (Hora & Robbes 2026), missing edge cases.

Tier-HIGH additions

• oh-my-claudecode:critic — multi-perspective adversarial critique; self-audit + Realist Check.
• oh-my-claudecode:architect — architectural review: SOLID, layering, cross-cutting concerns.

Tier-CRITICAL additions

• Cross-provider critic — via omc ask codex --agent-prompt critic (Codex reviews Claude's output) OR omc ask gemini. Different failure modes → higher catch rate.

Actionability filter (the Chowdhury 2604.03196 fix)

Every finding returned by a reviewer MUST include:

1. file (string, path)
2. line or line_start + line_end (integer)
3. evidence_type: one of code-reference / command-output / spec-citation / test-failure / diff-hunk
4. severity: critical / high / medium / low / info
5. confidence: 0-100

Findings missing any required field are moved to a suppressed[] array with reason "no actionable evidence." Surface count to the user; do not silently drop.

Synthesis

Dedup

Group findings by (file, line_start, severity); merge identical claims from different reviewers by concatenating agents array. Increases confidence when multiple reviewers agree (per MARS 2509.20502).

Conflict detection

Two findings on the same (file, line) with opposed verdicts (e.g., one says "introduce abstraction", another says "remove abstraction") are flagged in a conflicts[] block for user arbitration.

Hallucination check

For every finding that cites a file / symbol / API:

• Verify the file exists: [ -f "$file" ].
• Verify the symbol is reachable (grep for declaration).
• Verify commands in suggested_fix actually match project toolchain. Findings that fail this check move to suppressed[] with reason "reviewer hallucinated target."

Meta-review

The orchestrator (or this skill's own synthesis step) produces:

• Overall verdict: APPROVE / APPROVE_WITH_COMMENTS / REQUEST_CHANGES / BLOCKS_MERGE.
• Critical blockers (severity=critical with confidence ≥80).
• High-priority issues (severity=high with confidence ≥70).
• Kudos (explicitly captured positive signals — what was done well).
• Suppressed findings count (transparency about what was dropped).

Output format

Emits a single JSON artifact at quantum.reviews[<feature-id>].deepReview:

{
  "feature_id": "<prd-id or feature-slug>",
  "base_sha": "<before-first-story-commit>",
  "head_sha": "<after-last-story-commit>",
  "files_changed": 12,
  "stories_included": ["US-001", "US-002", ...],
  "timestamp": "<ISO 8601>",
  "risk_score": 47,
  "tier": "MEDIUM",
  "reviewers_dispatched": ["code-reviewer", "synthesizer", "security-reviewer", "test-engineer"],
  "findings": [
    {
      "id": "F-001",
      "agents": ["code-reviewer", "synthesizer"],
      "severity": "high",
      "confidence": 88,
      "category": "correctness",
      "file": "src/auth/session.ts",
      "line_start": 42, "line_end": 48,
      "evidence_type": "code-reference",
      "description": "<what>",
      "suggested_fix": "<how>",
      "cites": ["PRD AC-3", "tests/auth.test.ts:100"]
    }
  ],
  "conflicts": [],
  "suppressed": [{"agent": "architect", "reason": "no line citation", "count": 2}],
  "kudos": ["Clean separation of concerns in the new token-refresh flow"],
  "verdict": "APPROVE_WITH_COMMENTS",
  "blockers": [],
  "high_priority": ["F-001", "F-003"]
}

Also emits a human-readable markdown summary to docs/reviews/<feature-id>-deep-review.md.

Anti-rationalization guards