Job Babysitter Skill for Claude

Job Babysitter

Purpose

Stop manually polling long-running background jobs. Instead of dozens-to-hundreds of

ls -lh / ps checks while guessing at completion, start one background watcher that

detects the terminal state via plateau heuristics, then routes a verdict — done,

needs-attention, or blocked — with the exact next command.

A night-shift nurse for background jobs: it checks vitals on a schedule and escalates

only when something is actually wrong.

When to use

Use when a job will run long enough that babysitting it by hand wastes attention:

• Media encodes / transcodes (ffmpeg, video-transcribe, audio extraction)
• Embedding or vector-DB builds (qmd embed, index builds)
• Batch agent / LLM pipelines run in the background
• Browser / scrape daemons (real-browser, agent-browser) prone to hanging

Do NOT use for jobs that finish in seconds, or where a single Bash call already

returns the result.

Core principle: stay thin, lean on the harness

This skill orchestrates Claude Code's own primitives — do not reimplement them:

• Start the watcher with run_in_background: true. When it exits, the harness

re-invokes the agent automatically — no manual polling loop needed.

• The watcher (scripts/watch_job.py) owns the deterministic part: poll with backoff,

detect plateau, distinguish done from stuck, emit a verdict JSON.

• The skill's value is the per-job-type heuristics, the safe-recovery playbook,

and notification routing — all in references/playbook.md.

Workflow

1. Identify the job's signals

Determine what can be watched, in order of reliability:

• PID — the process ID (most reliable completion signal). Get it from the job's

launch, pgrep, or ps.

• Output file — a file that grows as the job progresses (e.g. ffmpeg target).
• Log file — a log that gets appended (e.g. an embed progress log).

Read references/playbook.md § "Completion heuristics by job type" to pick flags for

the specific job type (ffmpeg, embed, batch, browser).

2. Launch the watcher in the background

Run with run_in_background: true. Always pass --pid when known; add file/log

signals as corroboration. Write the verdict to a known path.

scripts/watch_job.py \
  --label "lab05 stream encode" \
  --pid <PID> \
  --output-file /path/to/output.mp4 \
  --plateau-bytes 65536 --plateau-polls 5 --stuck-after 120 \
  --max-wait 7200 \
  --verdict-out /tmp/job-babysitter-<label>.json

The watcher prints a one-line JSON heartbeat per poll (tail it for live progress) and

writes the final verdict JSON to --verdict-out on exit.

Tuning lives in the playbook; sensible defaults: --interval 10 (backs off to 60),

--plateau-polls 4, --stuck-after 300, --max-wait 7200.

3. On watcher exit, read the verdict and route it

The harness re-invokes the agent when the background watcher finishes. Read the

verdict JSON. It has status ∈ {done, needs-attention, blocked}, a reason,

suggested_next, elapsed time, and final size.

• done → verify the output is real (see the job-type "Done check" in the playbook,

e.g. ffprobe for media, count match for embeds), then proceed with the original task.

• needs-attention → the job plateaued while still alive (possibly wedged). Follow

the recovery playbook: diagnose read-only FIRST. **Never kill or run destructive

recovery (pkill, WAL checkpoint, VACUUM) without asking the user.**

• blocked → the watcher gave up after --max-wait. Report honestly: "gave up

waiting" ≠ "failed". Offer to re-check or extend the ceiling.

4. Notify per the chosen channel

Default to in-session resume. If the user picked a channel (Telegram, voice/TTS,

desktop notification), route per references/playbook.md § "Notification routing".

Always include the status emoji, label, elapsed time, and the exact next command.

Guardrails (non-negotiable)

• Never act on a single slow poll. "Stuck" requires plateau AND elapsed past

--stuck-after — the watcher already enforces this before returning needs-attention.

• Ask before any destructive recovery — pkill, kill, WAL checkpoint, VACUUM,

daemon restart. Diagnose read-only first.

• Report honestly. Distinguish done from "gave up waiting" from "wedged". Never

imply a success the watcher did not observe.

• Poll with backoff, not tight loops — the watcher handles this; never wrap it in

a manual fast-polling loop.

Resources

• scripts/watch_job.py — background watcher: plateau detection, stuck-vs-done logic,

verdict JSON. Stdlib only, Python 3.11+.

• references/playbook.md — per-job-type completion heuristics, the safe-recovery

table, and notification routing. Load when picking watcher flags or handling a

needs-attention/blocked verdict.