Feature Pipeline

Each feature generates a full document chain in .claude/pipeline/.

Document Chain

.claude/pipeline/{feature}/
├── 01-plan.md                  Planner: requirements + numbered ACs + 4-lens scores
├── 01.5-plan-critique.md       Plan Challenger: 6-vector attack, APPROVED/REVISE/REJECT
├── 02-design.md                Designer: design decisions + specs
├── 02.5-spec-critique.md       Spec Challenger: 8-vector attack, APPROVED/REVISE/REJECT
├── 03-dev-notes.md             Developer: implementation + 6-question analysis + self-review
├── 04-qa-report.md             QA Tester: test map + acceptance criteria verification
├── 05-browser-qa.md            Browser QA: health score (0-100) + screenshots
├── 03.5-spec-verification.md   Spec Verifier: every AC vs code, PASS/FAIL gate before review
├── 06-review.md                Reviewer: 4-specialist findings + auto-fixes
├── 07-ship.md                  Shipper: PR URL + release notes
└── coherence-report.md         Coherence Auditor: Coordination Score 0-100%

Standalone Mode Outputs

.claude/pipeline/project-audit/     00-backlog.md + iterations/
.claude/pipeline/browser-qa/        browser-qa-report.md
.claude/pipeline/security-audit/    security-audit.md
.claude/pipeline/debug-{bug}/       investigation.md
.claude/pipeline/health/            health-report.md
.claude/pipeline/canary/            canary-report.md
.claude/pipeline/review/            review-report.md
.claude/pipeline/qa-audit/          qa-report.md
.claude/pipeline/think/             design-doc.md
.claude/pipeline/arch-review/       architecture-review.md
.claude/pipeline/design-review/     design-review.md

Handoff Protocol

Each agent:

Reads the previous agent's output file
Does its work
Writes its own output file
The next agent reads that file

Every decision is documented. Every handoff is traceable.

Parallel Verification Fleet

Since v1.15, the three verification stages (04-qa-report.md, 05-browser-qa.md, 03.5-spec-verification.md) are produced concurrently — the orchestrator dispatches qa-tester, browser-qa, and spec-verifier in a single message because all three are read-only and independent. Failures merge into one fix list for developer; only failed members re-run. Write-capable agents never run in parallel.

Second Opinion

After any mode completes, the orchestrator offers an independent review:

Codex CLI available: different AI model challenges the work
No Codex: fresh Claude subagent with no session memory

The user decides what to act on.

Templates Architecture