build

# Build Execute an implementation plan with builder agents and automated review chains. **Announce:** "Using build to execute this plan." ## The Iron Law ``` NO TASK MARKED COMPLETE WITHOUT BUILDER SELF-REVIEW AND PASSING TESTS ``` ## Step 1: Load and Review Plan 1. Read the plan file at $ARGUMENTS 2. Review critically — identify concerns or questions 3. If concerns: raise them before starting 4. Count tasks, identify dependencies, determine execution mode 5. Check if `--reviewed` is in $ARGUMENTS — if so, enable review chain for all tasks 6. Parse `--checkpoint=N` from $ARGUMENTS (default `N=3`). `--checkpoint=0` or `--no-checkpoint` disables the human-wait between batches (use only for small/low-risk plans). ## Step 2: Choose Execution Mode **Standard** (default): Builder implements each task with self-review. No separate reviewer agents. - Spawn builder (Sonnet) per task - Builder self-reviews using its built-in checklist - Batches of N tasks (default 3, set via `--checkpoint=N`) with a human checkpoint; `--checkpoint=0` runs straight through with no wait **Reviewed** (`--reviewed` flag, or user says "with reviews"): For high-risk, security-sensitive, or unfamiliar codebases. - Same as Standard, plus: one task-reviewer (Sonnet) after each task — reads the diff once and emits two verdicts (Spec + Quality) - Leaner dispatch: pass plan file path + task line range + git diff only — no narrative summaries - Max 2 review cycles per task — escalate to user if unresolved ...