Builder Loop

Deliver a non-trivial change as a few coherent, validated checkpoints followed by one integrated agent-review gate. The goal: catch design flaws before coding and implementation flaws with focused proof, without repeatedly reviewing the same change.

Use only when the user explicitly requests checkpointed planning/building or named review gates. Do not infer it from generic requests such as "do this end-to-end" or from a change merely spanning layers. Skip for trivial edits and prefer one to three checkpoints for most work.

The review engine is the repo's agent-review skill. Read .agents/skills/agent-review/SKILL.md for its contract and commands. This skill orchestrates when to invoke it; that skill defines how.

Pipeline

Keep this checklist in the available task tracker:

• CP0 — Plan & validate locally (no production code yet)
• CP1..N — Implement each checkpoint (focused project gate)
• CP-final — Full validation + bounded full-change agent-review + handoff

Create a branch before any code (never work on the default branch).

CP0 — Plan & validate before coding

1. Decompose the task into the fewest coherent checkpoints that compile and are

independently verifiable, each with its own gate (typecheck / lint / focused tests / build). Prefer one to three checkpoints and order dependencies clearly.

1. Explore the real code first (read the actual files / call sites), then record the

plan in the task tracker. Use a scratch doc (e.g. PLAN-<slug>.md) only when the plan needs durable detail beyond the tracker. Cover verified current state with file:line refs, design decisions, per-checkpoint scope + gate, and risks.

1. Validate the plan against reality:

- When the user or active agent instructions explicitly request parallel review, use independent subagents to pressure-test separate concerns. Otherwise perform the same checks locally against the real code. See references/playbook.md for the review rubric. - Resolve concrete gaps against the real code. Do not run agent-review on the plan doc: the runner sees changed files, not the unchanged implementation the plan describes.

1. If you created a scratch plan doc, delete it after validation and before CP1. It

is a working artifact, not a review or delivery artifact.

Do not start CP1 while a concrete design decision remains unresolved. Ask the user when resolving it would expand or materially change the requested task.

CP1..N — Implement each checkpoint

Group checkpoints into coherent compilable units (e.g. all backend, then all iOS) — reviewing a half-written layer with no consumer produces noise. For each unit:

1. Implement the checkpoint. Keep edits at the right ownership boundary; update docs

in the same change (per AGENTS.md).

1. Run the checkpoint gate: the project's typecheck / lint / focused tests, or the

build for client code. Fix until green. See references/playbook.md for the gate matrix.

1. Do not run agent-review automatically at each checkpoint. Focused proof is the

checkpoint gate; the integrated change is reviewed at CP-final. - Run a checkpoint review only when the user explicitly requested per-checkpoint review for that checkpoint. - When requested, use the same bounded process as CP-final below; never restart a complete multi-round sequence after every fix.

1. Commit the checkpoint (only when the user authorizes commits). Move to the next.

A finding outside touched files is in scope only when the requested behavior would otherwise be incorrect (e.g. new copy that leaves another affected surface stale). Classify other adjacent findings as follow-up work.

CP-final — Full validation, full-branch review, handoff

1. Update colocated docs and the session changelog (.agents/sessions/...).
2. Run the full gate: complete typecheck / lint / test suite and a clean build.
3. Run one bounded agent-review landing review over the complete change: use branch

mode for a committed branch or local mode when work remains uncommitted. - Start with round 1 using the combined correctness, security, structural- simplicity, modularity, and reuse contract. - Verify findings against the real code and fix only in-scope blockers. - After an accepted fix changes the source, rerun relevant proof and advance to the next convergence round. Allow round 3 only when round 2 produces another accepted blocker whose fix changes the source. - Stop after round 3 and report unresolved findings; do not loop until clean.

1. If the user requested publishing, use the available GitHub workflow to open or

finish the PR and confirm CI is green and the PR is mergeable.

1. Stop at ready. Do not merge or deploy — report merge/ship-readiness and let

the owner decide (per AGENTS.md).

Reference