stark-build

Invoke the skill with the path to a spec that contains an accepted-base commit header, for example with optional flags like --dry-run or --agent claude. The runner parses the task list, dispatches one fresh writer per task, and commits each task only after its done-when check passes.

Completion is defined as a runnable check executed by the runner, never a model verdict. Checks and existing tests live in a protected-path list enforced by hooks and post-dispatch diff comparison, so a writer that edits its own gate aborts the run.

The dispatch fallback supports the Claude, Codex, and Gemini headless CLIs, selected with --agent auto|claude|codex|gemini. Claude receives hook-based path protection, while other writers rely on workspace isolation plus the runner's before-and-after state manifest comparison.

A crashed task is re-dispatched exactly once, resuming from progress and commits rather than zero. If it still fails, the runner logs a blocked deviation and moves on, because measured data shows repeated attempts increase cheating rather than fix rates.

No. The skill runs exactly one cross-vendor advisory review and at most one fix round covering critical, high, and medium findings. Anything still open after that round stays on the draft PR for the human to decide.

If all tasks are green but the spec's held-out verification command is red, the runner logs the divergence, keeps the PR as a draft, and reports it plainly. There is no automatic fix loop for end-to-end failures.