Spec Review Skill

Multi-perspective spec review using parallel subagent dispatch for comprehensive validation.

Reference: See reference/roles/ for reviewer personas, reference/report.md for YAML schemas, reference/playbook.md for edge case handling.

When to Use

• After spec-create to validate before implementation
• When spec feels incomplete or ambiguous
• Before task-dispatch for Initiatives (catches gate issues early)
• Standalone review of existing specs

Workflow

Step 1: Identify Spec

1. Parse spec name from argument (e.g., /spec.review auth-system)
2. If no argument: find most recent in ./specs/draft/
3. If no specs in draft: check ./specs/active/
4. Read spec documents: spec.md, context.md, tasks.yaml, validation.yaml

Step 2: Select Reviewers

Question 1: Select reviewers:

Header: Reviewers
Question: Which reviewers should analyze this spec?
multiSelect: true
Options:
- claude-opus: Claude Opus - native subagent, comprehensive, context-aware
- claude-sonnet: Claude Sonnet - faster native review
- openai-gpt5.2: OpenAI GPT-5.2 - base model
- openai-gpt5.2-codex: OpenAI GPT-5.2 Codex - code-specialized
- openai-gpt5.2-pro: OpenAI GPT-5.2 Pro - extended capabilities
- gemini-3-flash: Google Gemini 3 Flash - fast, efficient
- gemini-3-pro: Google Gemini 3 Pro - advanced reasoning

Default selection: claude-opus, openai-gpt5.2-pro, gemini-3-pro

Question 2: Select reasoning effort (if OpenCode reviewers selected):

Header: Reasoning
Question: What reasoning effort level for OpenCode reviewers?
multiSelect: false
Options:
- low: Quick responses, minimal deliberation
- medium: Balanced reasoning (Recommended)
- high: Deep analysis, thorough deliberation
- xhigh: Maximum reasoning (GPT-5.2 only)

Default: medium

Model mapping to commands:

• claude-opus → Task tool with model: "opus"
• claude-sonnet → Task tool with model: "sonnet"
• openai-gpt5.2 → opencode run --model "openai/gpt-5.2" --variant {reasoning}-medium
• openai-gpt5.2-codex → opencode run --model "openai/gpt-5.2-codex" --variant {reasoning}-medium
• openai-gpt5.2-pro → opencode run --model "openai/gpt-5.2" --variant {reasoning}-medium
• gemini-3-flash → opencode run --model "google/gemini-3-flash-preview" --variant {reasoning}-medium
• gemini-3-pro → opencode run --model "google/gemini-3-pro-preview" --variant {reasoning}-medium

Step 3: Dispatch Reviewers in Parallel

CRITICAL: Dispatch all selected reviewers in the same message for true parallelism.

Review Prompt Template:

You are reviewing a spec for completeness and feasibility.

## Spec Documents
[Include spec.md, context.md, tasks.yaml content]

## Review Focus
Evaluate against these gates:

1. **Completeness** - Are all requirements specified? Missing behaviors?
2. **Consistency** - Do documents contradict each other? Ambiguous terms?
3. **Feasibility** - Can tasks be implemented as described? Missing dependencies?
4. **Clarity** - Would a fresh developer understand what to build?

## Output Format
Return a YAML report:

```yaml
reviewer_report:
  reviewer: {REVIEWER_ID}
  gates:
    completeness:
      status: pass | fail
      issues: []
    consistency:
      status: pass | fail
      issues: []
    feasibility:
      status: pass | fail
      issues: []
    clarity:
      status: pass | fail
      issues: []
  issues:
    - severity: critical | high | medium
      gate: completeness
      area: ${TAXONOMY_AREA}
      description: "Clear description"
      suggestion: "How to fix"
  clarifying_questions:
    - area: ${TAXONOMY_AREA}
      question: "What needs clarification?"
  strengths:
    - "Positive observation"

**Dispatch by Type:**

**Claude reviewers (Task tool):**
```python
Task(
  subagent_type="general-purpose",
  model="opus",  # or "sonnet"
  prompt=review_prompt
)

OpenCode reviewers (Bash tool, background):

timeout 1200 opencode run --model "{MODEL_PATH}" --variant {reasoning}-medium "{review_prompt}"

Examples (with high reasoning):

• opencode run --model "openai/gpt-5.2" --variant high-medium "{prompt}"
• opencode run --model "google/gemini-3-pro-preview" --variant high-medium "{prompt}"
• opencode run --model "openai/gpt-5.2-codex" --variant high-medium "{prompt}"

Step 4: Synthesize Reviews

After all reviewers complete:

1. Parse reports - Extract YAML from all outputs
2. Merge issues:
   • Deduplicate by description similarity
   • Combine issues flagged by multiple reviewers (higher confidence)
   • Note which reviewer(s) found each issue
3. Aggregate gates:
   • Gate fails if ANY reviewer fails it
   • Record which reviewer(s) failed each gate
4. Prioritize questions:
   • Group by taxonomy area
   • Rank: Scope > Behavior > Data Model > Constraints > Edge Cases > Integration > Terminology

Step 5: Present Review

Gate Summary Table:

| Gate         | Status | Claude | GPT-5.2 Pro | Gemini-3 Pro |
|--------------|--------|--------|-------------|--------------|
| Completeness | FAIL   | fail   | pass        | fail         |
| Consistency  | PASS   | pass   | pass        | pass         |
| Feasibility  | FAIL   | fail   | fail        | pass         |
| Clarity      | PASS   | pass   | pass        | pass         |

Issues by Severity:

## Critical (must fix before implementation)
- [C1] Missing error handling for auth timeout (Completeness)
  Found by: claude-opus, opencode-gemini3-pro
  Suggestion: Add error case to spec.md#edge-cases

## High (should fix)
- [H1] Task T003 depends on undefined API contract (Feasibility)
  Found by: claude-opus, opencode-gpt5.2-pro
  Suggestion: Define API in context.md or defer task

## Medium (consider)
- [M1] Term "session" used inconsistently (Consistency)
  Found by: opencode-gpt5.2-pro
  Suggestion: Add to terminology section

Step 6: Clarifying Questions

Use AskUserQuestion with questions grouped by taxonomy area:

Header: Scope
Question: The spec mentions "user roles" but doesn't define them. What roles exist?
multiSelect: false
Options:
- Admin/User: Two-tier system
- Role-based: Granular permissions
- Defer: Address later

Record answers for validation.yaml update.

Step 7: Update Validation

Add clarification session to validation.yaml:

clarification_sessions:
  - id: S00${N}
    timestamp: ${ISO_TIMESTAMP}
    source: spec-review
    reviewers: [claude-opus, opencode-gpt5.2]
    questions:
      - id: Q001
        question: "${QUESTION}"
        answer: "${ANSWER}"
        area: ${TAXONOMY_AREA}
        doc_updates:
          - file: spec.md
            section: ${SECTION}
            action: modified

Update markers section:

• Close resolved markers (status: resolved)
• Add new markers for deferred questions (status: open)

Step 8: Recommend Action

All gates pass:

Review complete. All gates passed.

Recommendation: Ready for /spec.promote or /implement

Issues found:

Review complete. 2 gates failed.

Recommendation:
1. Address critical/high issues
2. Re-run /spec.review

Gates

Edge Cases

OpenCode timeout (> 20 minutes):

• Continue with completed reviews
• Note in output: "[Reviewer] timed out, partial results"
• Still usable but recommend re-running

One reviewer fails:

• Parse what you can
• Report partial results with clear indication
• "Claude review: complete, GPT-5.2 Pro: failed to parse, Gemini 3 Pro: complete"

No reviewers selected:

• Default to claude-opus only
• Warn: "Consider adding external reviewers for diverse perspectives"

Spec not found:

• List available specs in ./specs/draft/ and ./specs/active/
• Ask user to specify

OpenCode command syntax:

• GPT-5.2 Pro: opencode run --model "openai/gpt-5.2" --variant {reasoning}-medium {query}
• Gemini 3 Pro: opencode run --model "google/gemini-3-pro-preview" --variant {reasoning}-medium {query}

Integration

Command: /spec.review [spec-name]

Related skills:

• spec-create - Creates specs to review
• clarify - Resolves markers found during review
• task-dispatch - Next step after review passes

Reference

• reference/roles/claude-reviewer.md - Claude reviewer persona
• reference/roles/opencode-reviewer.md - OpenCode reviewer persona
• reference/report.md - YAML report schemas
• reference/playbook.md - Edge case handling