ca-full-cycle

CA Full-Cycle — Research · Plan · Implement · Review

One human gate before code, one human loop after it. Everything between runs alone.

┌──────────┐   ┌────────┐   ┌───────────┐   ┌──────────┐   ┌────────┐
│ RESEARCH │ → │  PLAN  │ → │ IMPLEMENT │ → │  REVIEW  │ → │   QA   │
└──────────┘   └────────┘   └───────────┘   └──────────┘   └────────┘
 human + AI       AI            AI              AI           human
 (the gate)    top tier     worker waves    independent     fix loop

The human's job is to make the problem understood (Research) and to judge the result (QA). The pipeline's job is everything else — and it never stops in the middle to ask whether it may continue.

Critical Rules (read before acting)

Loading this skill's files. References live under references/ in this skill's own directory. Resolve them relative to the skill directory — never the workspace root. A card or a section you open is read to its end. Sub-agents open the long references only by section, never whole; the orchestrator reads its current phase reference whole.

1. One gate before code. The brief confirmation closes Research — the single pre-execution approval. From there to the end of Review the pipeline is autonomous: never ask permission to plan, to dispatch a wave, to run the Reviewer, or to fix a gap the machinery caught. The human returns at QA. The only mid-pipeline escalations are the bounded-failure valves (3 strikes), a worker's spec-ambiguity stop, and a discovery that invalidates the brief — those go to the user; nothing else does.
2. Tests derive from the brief's ACs and assert brief-defined outcomes — they never mirror the implementation. The gate (the test runner) decides a task is done, not self-assessment. Never weaken, skip, or delete a test to make it pass. A red that turns green on re-run with no code change is a flaky finding to report — never a pass to absorb.
3. One atomic, pathspec-limited commit per task. Never batch tasks into one commit.
4. The orchestrator never implements a clustered plan. From 4 tasks up, Implement is delegated to workers — one per cluster, all clusters of a wave in parallel; pulling a cluster into the planning window is a violation, not a shortcut. A plan of ≤3 tasks runs inline in the planning window (light path) — and rules 5–6 still hold.
5. Verification is structural, never prompted. A fresh wave verifier closes every wave (gate + delivery check + one discrimination mutant); a fresh, independent Reviewer closes the feature (author ≠ reviewer, evidence-or-zero). Neither is optional, neither is offered as a question. Declared limit: the wave verifier anchors on the plan, the Reviewer on the brief — a wrong plan is caught by the brief-anchored close; a wrong brief is caught only by human QA. That is what QA is for.
6. Artifacts in English, chat in the user's language. Everything under .ca-plans/ plus slugs, payloads, summaries and verdicts is English — agents are the only readers, and English re-reads are cheaper. The QA script is chat, so it follows the user; product text (UI labels, error messages) is quoted as-is.