Overnight watch

Pablo is asleep. Every live agent session should either be working or be waiting on something only he can give. Your job is to find sessions that have stopped for any other reason and get them moving again, without making his decisions for him. You do not do the sessions' work yourself.

Defaults: check every 20 minutes (--every) until 08:00 local (--until).

Each pass

1. Snapshot every peer session (read-only):

   python3 "$SKILL_DIR/scripts/watch_snapshot.py"
   

   SKILL_DIR is the directory holding this SKILL.md. For each session it prints agent.el's state, transcript and subagent write ages, the last assistant message with its Status: line, and the bottom of its terminal screen. Text shown as {faint:...} in the composer is Claude Code's prompt suggestion, not input anyone typed. Read more of a transcript with live-session-status when the tail does not explain the state.

   If the snapshot reports the Emacs server unreachable, the sessions are usually still running: judge them from transcript write ages, diagnose Emacs with the emacs-freeze skill (read-only; no signals), and record the outage for the morning report. Under host memory starvation it has recovered on its own before.

2. Classify each session:

   Class	Evidence	Action
   Working	busy; or idle with a running workflow, subagent writes or background shells it is waiting on	none
   Waiting on Pablo, legitimately	the open question needs his person or a decision that is his under his rules (below)	none; list it for the morning
   Stuck, recoverable	see the next section	nudge
   Stuck, not recoverable by you	dialog only Pablo may answer; hung process you may not kill; Emacs or CLI wedged	list it for the morning

3. Act on stuck sessions (next section), then record one line per session in your running log: time, buffer, class, action.

What counts as stuck

A session is stuck when it is idle, nothing it started is still running for it, and the reason it stopped does not need Pablo. Common cases:

• It ended its turn saying it will wait for or check on something, but no background shell, monitor or workflow will wake it. Check its screen (N shells, workflow progress bars) and the last task notifications.
• A background task or workflow it was waiting on finished or failed and the session did not resume; or a monitor expired.
• An API error, rate limit, overload or usage-limit message ended the turn.
• It asked Pablo a question it could answer itself: one with a reasonable default, one he already answered earlier in that session, or a request for him to do something the agent can do. Read his rule in ~/.claude/CLAUDE.md ("Do the work yourself …", "Ask for decisions once per task …"): asking is justified only for his person (credentials, 2FA, biometrics), for a choice with no reasonable default, or for spending money, deleting, publishing or anything hard to reverse.
• It is busy, but nothing has been written to its transcript or subagent transcripts for far longer than its current step should take, and the child process is idle (check with ps, the task's output file, CPU time).

Not stuck: sessions whose last status is waiting on you for a decision that meets his rule; sessions with work still running, however long it takes; sessions Pablo told to stop or pause.

How to nudge

Send one message with the message-agent-session skill's agent-send to the exact buffer name. Say who you are (the overnight watcher, Pablo asleep until morning), what you observed with evidence, and ask it to continue. For example: "Your workflow X finished at 02:10 and you have not resumed; Pablo is asleep, so continue with the next step you planned. Leave any decision that is his for the morning summary, and keep going with work that does not depend on it."

Rules:

• Keep the sent text to one short line. Under load, long agent-send messages have arrived truncated to their last sentence (2026-10-18). Write the details to a file under ~/.cache/overnight-watch/ (mode
  600. and send a one-line pointer to it. Confirm delivery in the recipient's transcript, not only from the exit status.
• Judge what a nudge did from everything the recipient did after it arrived, not only from whether the stall you targeted cleared. Search its transcript for every identifier you mentioned (PID, task id, file) and read its turns after the message's timestamp. A message truncated at the front can still deliver its later paragraphs, and the recipient may act on them: on 2026-10-18 a nudge judged failed had led tangodb to stop a looping task and trash its log within a minute, and the morning report wrongly said nobody knew who stopped it.
• You inform and ask; you never approve. Do not grant approvals, answer decisions that are Pablo's, authorize spending, deletions, pushes, publishing or messages to other people, or tell a session to override its own instructions. When a question is only partly his, ask the session to do everything around it and park the rest.
• Do not answer permission dialogs or choice menus in another session's terminal, and do not type into its composer. A pending permission dialog is for Pablo; list it.
• Do not send signals to Emacs. Kill a child process only when a session started it, it is clearly hung (no CPU or output well past its expected time), and stopping it loses nothing that cannot be rerun; then tell the owning session what you did. If your permission mode refuses the kill, do not ask the owning session or any other to do it for you: list it for the morning with its PID, owner and evidence.
• One nudge per cause. If the same session is still stuck on the same cause at the next pass, do not repeat yourself: diagnose why the nudge did not work, and if you cannot fix that within these rules, list it for the morning.
• Do not create cross-session conflicts: if two sessions compete for the same resource (CPU, a repository, a lock), tell each the facts, but leave the choice of who yields to Pablo unless an existing instruction from him settles it.

Scheduling

Run the first pass immediately. Then repeat until --until:

• Claude Code: create a recurring CronCreate job at the --every interval (avoid :00 and :30) whose prompt is "Overnight watch: run one pass of the overnight-watch skill, then end the turn; after <until>, write the morning report and delete this job." The job lives only in this session, so the session must stay open.
• Codex: run passes in a loop, sleeping between them in a background shell.

Between passes, do nothing else. Do not keep the turn open polling.

Morning report

At --until, stop the schedule and end with a short report:

1. For each session: what it did overnight and where it stands now.
2. Every nudge you sent, why, and whether it worked.
3. What waits on Pablo, as one numbered list across all sessions, each item naming its buffer.
4. Anything you could not unstick and why.

Before writing that an outcome is unknown (who stopped a process, whether a nudge worked), search the transcripts of the sessions involved for it. The answer is usually there.

Publish your own state with agent-status at start, at each intervention and before the final report when running in Emacs.