
11 results

Implementation agent that executes one milestone from the active run's milestones/ directory at a time, following the spec strictly and stopping at checkpoints. Use when break-n-bake workflow is in bake phase. Never runs more than one milestone per invocation without an explicit user OK.

Heavy-cognitive-load agent that transforms a raw user prompt plus scout report into a layered spec directory under the active run (`.bnb/<slug>/`). Use when break-n-bake workflow enters break phase. Produces spec/, milestones/, quality/, validation/ (seed), README, questions-before-start.

Corrective agent that fixes blocker-severity failures identified by Validator. Runs at end-of-run fix pass or when bake is halted mid-milestone by blockers. Cannot modify test files, lint/type configs, or any content under the active run's quality/ directory — enforced by hooks and snapshot verification.

Read-only validator that runs stack-appropriate test/lint/typecheck commands after a bake milestone, executes the per-run validation layer, and produces a structured severity-classified report. Invoked by orchestrator immediately after Baker finishes. Has no Write or Edit tools — physically cannot modify code.

Routes oversized prompts and large refactors into a structured spec-first workflow instead of letting Claude start coding blindly. Activate for long prompts (≥800 words or ≥5 numbered requirements or ≥3 distinct concerns), refactor/migration keywords, or short prompts that touch cross-cutting concerns (auth, routing, state, schema, build). Suggest `/break-n-bake:break` and wait for yes/no before proceeding.