build-steplisted
Install: claude install-skill aberson/skill-mesh
# Build Step
> **Judging doctrine:** the reviewer-gate invariants here (producer never grades itself, weak/local models never gate, a deterministic Claude gate consolidates) live in [`_shared/judge-core.md`](../_shared/judge-core.md) §5–§6 — this skill instantiates them for the dev→review→merge loop.
Execute a single build step end-to-end: developer writes code, reviewer(s) gate it,
approved changes merge back to the project.
Three independent knobs control how it runs:
| Knob | Flag | Options | Default |
|---|---|---|---|
| Isolation | `--isolation` | `worktree`, `docker` | `worktree` |
| Review style | `--reviewers` | `auto`, `code`, `deep`, `runtime`, `full` | `auto` |
| UI evidence | `--ui` | flag (present = on) | off |
Any combination is valid. Examples:
```bash
/build-step --problem "Add user search endpoint" --issue 12
/build-step --problem "Fix login form" --isolation docker --reviewers code
/build-step --problem "Fix broken dashboard" --reviewers runtime --ui --start-cmd "npm run dev" --url http://localhost:3000
/build-step --problem "Refactor auth middleware" --reviewers full --ui --start-cmd "uv run python -m myapp" --url http://localhost:8000
```
---
## Arguments
| Arg | Required | Default | Description |
|---|---|---|---|
| `--problem` | yes | -- | What to build or fix |
| `--issue` | no | -- | GitHub issue number |
| `--acceptance` | no | -- | Optional acceptance target (the step's `Done when:`, forwarded by `/build-phase`'s Step 0 extract + Step 2 disp