lessan/jobs

Was macht jobs?

Use runjob to hand bounded work to an installed headless model CLI and to keep a machine-wide execution ledger. Confirm that runjob is on PATH; if it is not, point the user to the repository's installation steps instead of improvising a machine-wide install.

Launching a job can consume quota or incur provider charges. Treat an explicit request to dispatch or hand off work as authorization for that launch. A request to explain, inspect, draft, or plan is not launch authorization.

Choose the execution shape

• Use runjob run for one bounded job. It runs in the selected working directory and does not create a branch or worktree. For work that may edit or commit, use a disposable Git worktree or the fleet mode.
• Use runjob fleet run for several jobs, dependency ordering, automatic isolated worktrees, resumable sessions, or per-engine concurrency caps. Read references/fleet.md before creating fleet files.
• Use runjob log, runjob report, and runjob dash to inspect work already in flight. These are read-only; runjob repair is an explicit ledger mutation.

Do not substitute this skill for an in-session subagent when the user asked for delegation within the current conversation.

One-shot jobs

Use a prompt file for substantial instructions so the dispatch is reviewable:

runjob run --engine codex --effort high \
  --cwd /absolute/path/to/worktree --file /absolute/path/to/prompt.md

Omit --model or --effort when the user has not selected them and the engine has a sensible CLI default. Do not invent current model names. Read references/engines.md when engine-specific flags, authentication, or sandbox behavior matters.

runjob run returns after launch unless --wait is supplied. Keep the printed uid. Inspect it with:

runjob log --uid UID
runjob log --active
runjob report UID --json
runjob dash

Logs and prompts may contain source code or sensitive instructions. They live under ${DEV_JOBS_HOME:-~/.config/dev-jobs}; do not publish or attach them without a separate review.

Completion and handoff

Report the uid, engine/model actually requested, working directory, and whether the job is merely launched or has completed. Do not describe a process launch as a successful result. When waiting, verify the terminal ledger state and summarize the job's actual output, including failures or residual work.

For machine-readable job endings, use the exact protocol in references/report-trailer.md. Treat its fields as the job's claims; independently verify commits, tests, and modified paths before merging or publishing anything.