feat: adversarial spec review loop + skill chaining (v0.9.1.0)

[garrytan]

Summary

• Adversarial spec review loop: /office-hours and /plan-ceo-review now dispatch an independent reviewer subagent that checks documents on 5 dimensions (completeness, consistency, clarity, scope, feasibility) with a convergence guard, quality score (1-10), and JSONL metrics. Implemented as a {{SPEC_REVIEW_LOOP}} resolver in gen-skill-docs.ts — any skill can use it.
• Visual sketch phase: For UI ideas, /office-hours generates a rough HTML wireframe using design principles from {{DESIGN_METHODOLOGY}} and DESIGN.md, renders via $B, and presents a screenshot for iteration before writing the design doc.
• Skill chaining: /plan-ceo-review and /plan-eng-review now detect when no design doc exists and offer to run /office-hours first. Implemented via benefits-from: frontmatter + {{BENEFITS_FROM}} resolver. One-hop-max, never blocks, max one offer per session.
• Spec review metrics: Every review logs iterations, issues found/fixed, remaining, and quality score to ~/.gstack/analytics/spec-review.jsonl.

Test Coverage

Tests: 370 → 439 (+69 new)

CODE PATH COVERAGE
===========================
[+] scripts/gen-skill-docs.ts
    ├── generateSpecReviewLoop()  [★★★ TESTED] 7 assertions — dimensions, Agent, iterations, quality, metrics, convergence, graceful failure
    ├── generateDesignSketch()    [★★★ TESTED] 6 assertions — DESIGN.md, wireframe, $B goto, screenshot, rough aesthetic, skip conditions
    ├── generateBenefitsFrom()    [★★★ TESTED] 4 assertions — CEO offer, eng offer, graceful decline, negative (qa has no offer)
    └── benefitsFrom parsing     [★★★ TESTED] via generated output validation

[+] office-hours/SKILL.md (generated)
    ├── spec review section      [★★★ TESTED] structure + content validation
    └── visual sketch section    [★★★ TESTED] structure + content validation

[+] E2E tests
    ├── office-hours-spec-review [★★★ TESTED] Agent tool called, spec review mentioned
    └── plan-ceo-review-benefits [★★★ TESTED] skill chaining offer understood

Pre-Landing Review

No issues found — template-only changes with no attack surface expansion.

Eval Results

• Routing E2E: 11/11 pass ($1.65)
• E2E: 36/37 pass — 1 pre-existing failure in /design-consultation preview (unrelated)
• Our 2 new E2E tests: both pass

TODOS

No TODO items completed in this PR.

Test plan

• All skill validation tests pass (439 tests, 0 failures)
• gen-skill-docs freshness check passes (--dry-run)
• E2E routing tests pass (11/11)
• New E2E tests pass (spec-review + benefits-from)

🤖 Generated with Claude Code