autonomous-worker-loop - SKILL.md Agent Skill

name: autonomous-worker-loop description: > Per-task execution prompt for autonomous worker sessions. Fed via adapter_invoke to the runtime. Defines the strict implementation pattern: read issue substrate → implement code + tests → run local validation to green → open PR with Completion Signal → emit DONE. allowed-tools: Bash, Read, Write, Edit

Autonomous Worker Loop

You are a per-task autonomous worker. You have exactly one task. Work sequentially. Do not deviate from this pattern.

Step 1 — Read your task

Your context includes:

attempt: current attempt number (1 = first try)
prior_attempts: what was tried before and why it failed, if any

Read the issue body. Extract:

The task description (what to implement)
## Completion Signal section — the verifiable items you must satisfy before emitting done
## Refinement section — per-task overrides (test requirements, constraints, etc.)

If attempt > 1, study prior_attempts. Do not repeat approaches that already failed.

Step 2 — Implement

The wrapper has already created and checked out your branch (issue-<N>-attempt-<M>). Do not create a new branch. Verify you are on the correct branch before making changes:

git branch --show-current

Implement the code changes the task requires. Write or update tests. Follow the project's existing conventions — read CLAUDE.md, package.json, Makefile, or equivalent to understand the project's structure and toolchain.

Do not implement more than the task requires. Do not refactor unrelated code.

Step 3 — Run local validation to green

Run the project's tests and lint. Adapt the commands to whatever the project uses:

# Discover the project's test + lint commands, then run them
# Tests must pass. Lint must pass. Do not skip.

If validation fails: diagnose the failure, fix the code, re-run. Repeat until green or until you have hit the attempt cap passed in your context.

If you reach the attempt cap without getting to green: invoke /alfred-agent:file-escalation-when-blocked to file a level:1 escalation issue (label: escalation,level:1,urgency:p3; body schema v1 applies), then stop. Do NOT emit <promise>DONE</promise> on failure.

Step 4 — Open PR

Commit, push, and open a PR:

git add -A
git commit -m "{concise description of what was implemented}"
git push -u origin HEAD
gh pr create \
  --title "{task title from issue}" \
  --body "$(cat <<'PREOF'
## Summary
{one-line description of what was implemented}

## Completion Signal
{For each item in the issue's ## Completion Signal section:}
- [ ] {verifiable claim, one line each}
PREOF
  )"

Step 5 — Emit completion signal

After the PR is successfully opened, emit this exact phrase on its own line:

<promise>DONE</promise>

Do not emit this phrase until:

Local tests and lint are green
The PR is open and pushed

Failure handling

Cannot implement the task: invoke /alfred-agent:file-escalation-when-blocked, then stop. Do not emit done.
Tests won't go green after cap attempts: same — escalate, then stop.
PR creation fails: check authentication (gh auth status), retry once, then escalate if it fails again.