Claude Code

Claude Code Skill

This skill runs coding tasks as persistent, detached jobs using the @anthropic-ai/claude-agent-sdk. Unlike raw exec claude calls, jobs survive agent turn timeouts, gateway restarts, and API rate limits.

Starting a coding task

Use exec to start a detached job. It returns immediately with a jobId.

node {{skill_dir}}/scripts/run.mjs start \
  --prompt "Read TASK.md and implement. Run tests." \
  --cwd /path/to/project \
  --job-id task-137

Arguments:

• --prompt (required): The task description for Claude.
• --cwd (required): The working directory for the coding task.
• --job-id (required): A unique identifier for this job.
• --model (optional): Override the model (defaults to SDK default).

Returns JSON:

{ "jobId": "task-137", "pid": 12345, "status": "running" }

Checking status

Poll job status to see if it's still running, completed, or failed.

node {{skill_dir}}/scripts/run.mjs status --job-id task-137

Returns JSON:

{
  "jobId": "task-137",
  "pid": 12345,
  "status": "running",
  "startedAt": "2025-01-01T00:00:00.000Z",
  "endedAt": null,
  "error": null,
  "rateLimited": false
}

Status values: running, completed, failed, killed, not_found

If rateLimited is true, the job hit API rate limits — this is a temporary condition, not a hang.

Getting logs

Read the last N lines of streaming output from the job.

node {{skill_dir}}/scripts/run.mjs logs --job-id task-137 --tail 50

Getting the result

Once status is completed, retrieve the final result.

node {{skill_dir}}/scripts/run.mjs result --job-id task-137

Returns JSON:

{
  "jobId": "task-137",
  "status": "completed",
  "result": "...",
  "cost_usd": 0.05,
  "duration_ms": 120000,
  "num_turns": 8
}

Listing all jobs

node {{skill_dir}}/scripts/run.mjs list

Killing a job

node {{skill_dir}}/scripts/run.mjs kill --job-id task-137

Best Practices

• Unique job IDs: Use descriptive IDs like issue-42 or feat-auth-v2 to track jobs.
• Poll status: After starting a job, check status periodically. Jobs can take 5-30 minutes.
• Check logs on failure: If status is failed, read logs to understand what went wrong.
• Rate limits are normal: If rateLimited is true, the SDK is handling retries automatically.
• Verification (Mandatory): Always include "Run npm run lint and npm test to verify" in your prompt. The repo enforces CI checks; unverified code will fail to commit.
• Conventional Commits: Ensure the agent knows to use feat:, fix:, chore: prefixes, or the commit hook will reject it.

Environment

• ANTHROPIC_API_KEY — required (the skill checks for this and fails fast with a clear error if missing).
• CLAUDE_SKILL_JOBS_DIR — optional, defaults to <skill_dir>/jobs.
• CLAUDE_SKILL_MODEL — optional, overrides the default model.