| name | autonomous-development |
| description | This skill should be used when the user asks about "dev-auto", "autonomous development", "orchestrator", "worktree lifecycle", "structured output", or when running dev-auto-* commands or any autonomous loop that relies on orchestrator-owned worktrees, explicit inputs, and structured JSON outputs. Auto-safe development contract for orchestrated dev/review/fix workflows.
|
For humans: The safety contract for autonomous execution — what operations are allowed, who owns the worktree lifecycle, how results are reported. The guardrails that prevent destructive autonomous behavior. Short file, worth reading in full.
Autonomous Development Contract (Auto-Safe)
This skill defines the shared safety and output contract for autonomous dev/review/fix commands invoked by an orchestrator.
<core_invariants>
- Orchestrator owns worktree lifecycle. Commands must not create or remove worktrees.
- Inputs are explicit. Do not infer branch/worktree from Linear or other sources.
- Emit structured JSON only as the final assistant message (no prose after).
- Avoid global Graphite operations; use only allowlisted commands.
</core_invariants>
Auto-safe guardrails
- Worktree: assume
cwd is already the correct issue worktree; verify before edits.
- Branch: ensure the current branch matches the orchestrator-provided
branchName; if not, emit a structured failure result.
- Iteration: do not run sub-issue loops; the orchestrator controls issue and review/fix iteration.
Graphite allowlist (v0)
- Allowed:
gt ls, gt create, gt track, gt modify, gt sync --no-restack, gt restack --upstack, gt ss --draft, gt ss -u.
- Not allowed:
gt sync without --no-restack, global restacks, reparenting unrelated stacks, publishing PRs unless explicitly authorized.
Output contract (v0)
- Final assistant message must be JSON only matching the phase schema.
- Prefer emitting a structured
status: "failed" or status: "deferred" over crashing.
