work

Work

Purpose

Use this skill when Enoch should manage persistent work instead of treating every request as a single foreground chat turn.

The work skill covers three core execution modes:

• queue: /task FIFO background jobs;
• backlog: /backlog deferred idle-time work with priority;
• cron: /cron recurring scheduled jobs;

It also covers:

• single-message task status updates through the configured chat provider;
• automatic skill-only learning artifacts after successful work.

Boundary

Worker state is local runtime state under .enoch/:

• .enoch/task_queue.json
• .enoch/backlog.json
• .enoch/cron.json
• .enoch/artifacts/learning/

Do not treat every successful job as inheritable. Enoch records an inheritable learning artifact only when completed work changes a skill package under src/<agent>/skills/<skill-name>/.

Operation

When work is queued:

1. Preserve the request and any conversation context snapshot.
2. Keep task execution non-blocking for the active chat conversation.
3. Update one chat status message with queued, running, paused, completed, failed, elapsed time, latest update, and review URLs.
4. Run queued work through the same authorized repository workflow as foreground /do work.
5. Promote backlog items only when the task queue is idle.
6. Run cron checks in an independent scheduler rather than after chat polling. Claim due occurrences atomically and enqueue them at the front of the task queue. Keep at most one pending, running, or paused task per schedule. Interval targets are fixed-rate: after downtime, coalesce missed targets into one immediate task and advance to the first anchored target still in the future.
7. When agent runtime authentication, quota, or rate limits are unavailable, move the active task to paused, stop the worker before it consumes later tasks, and warn the human. /task resume <id|all> moves selected paused tasks back to the front with the same ids and context after access is available again.
8. Keep the agent-instance branch as the resident control worktree. Give every code task its own linked worktree and branch from the latest available authoritative revision supplied by the VCS provider, and run the agent runtime, tests, commits, publication, and review handoff there. Keep .enoch queue, memory, and event state in the resident worktree. Remove successful task worktrees after handoff; preserve failed or paused worktrees for inspection and recovery.
9. /task retry <id> retries only a failed task by creating a new task with a new id and parent_task_id; never rewrite the original failure. Preserve the request, context, source, provenance, and any recoverable task workspace/revision. Before new execution, reconcile recorded review identities with the configured review provider; reuse a validated open or landed review instead of duplicating work. If a retry fails, retry that latest failed task so the causal chain remains linear.
10. Give each running task a worker lease. Recovery must not requeue a task while its owner process is alive, and only the lease owner may publish a terminal task transition or final status message.
11. Classify failures before deciding whether to retry. Automatically retry only explicit transient failures such as network interruption, rate limiting, or temporary upstream unavailability, with bounded backoff and at most three attempts. Treat dirty worktrees, validation failures, task timeouts, permission or configuration errors, and unknown failures as non-retryable. Record attempt, failure code, failure class, and retry disposition in task events. Keep /task retry <id> as the explicit human override.
12. /task resume <id|all> resumes only paused tasks without changing their ids or causal history.