| name | watch-po |
| description | Event-driven detection of stalled or abnormal PO panes, reported to the main PO — approval-prompt stalls (POPROMPT), context threshold crossings (POCTX), forced model switches (POMODEL), pane disappearance (PODEAD) — via a persistent Monitor, optionally pushed to a notification topic. The standing sensor that fills the gaps between main-po-patrol rounds. Triggers - "arm the PO watch", "watch-po", "tell me when a PO stalls". |
watch-po — event watch for PO stalls and anomalies
English | 日本語
If main-po-patrol is the periodic round, this is the standing sensor
between rounds. Born from a real incident: two freshly rotated PO sessions
stalled on approval prompts during monitor re-arming, and the human owner
noticed before the main PO did — twice. The main PO should notice first.
Events (all deduplicated — emitted on transition only)
| Event | Meaning | Main PO's first move |
|---|
POPROMPT <label> <pane> | stalled waiting for approval/input | look at the pane → verify the command against the handover/canon, then approve or reject (no blind approvals) |
POCTX <label> <pane> <pct>% | context crossed the threshold (default 60%) | check for a mid-flight live gate → drive rotation (main-po-patrol §3) |
POMODEL <label> <pane> <from>-><to> | the statusline's model changed | measured platform behavior: models can be switched out silently. Owner decides continue vs. restore |
PODEAD <label> <pane> | pane disappeared | rebuild from the latest handover |
POBACK <label> <pane> | recovered from PROMPT/DEAD | record only |
WATCH ERROR <msg> | the watch itself is failing | check herdr |
Arming (from the main PO session)
- Read the PO panes from the roster (re-measure with
herdr pane list
right before arming — pane IDs drift; fix the roster first if it
disagrees).
- Start under a persistent Monitor:
bash <skill-dir>/scripts/watch-po.sh \
--pane <paneId>=<space>/po [--pane ...] \
--ctx-threshold 60 --interval 30 [--ntfy <topic>]
Labels must be <space>/po form — the human owner reads labels, not pane
IDs (pane IDs appear in parentheses as an aid). Give the herdr pane itself the
same label (herdr pane rename <pane> '<space>/po').
--ntfy <topic> pushes each event to ntfy.sh/<topic> as well (events are
deduplicated, so the volume is low).- After arming, add one ledger line — and put the watch on your own
handover's monitor re-arm list (monitors die with the session).
Discipline on receiving events
- POPROMPT first — a PO's clock is stopped. But approve only after
looking at the pane and matching the command against the handover/canon.
- POCTX is not "rotate immediately" — wait for the terminal if a live gate /
one-shot is mid-flight.
- POMODEL: never switch the model back on your own — the owner decides based
on load (a light-load PO may simply continue).
- Ledger one line per handled event. POBACK is record-only.
Constraints (know these)
- Detection greps the visible screen — a prompt that scrolled away in an
instant is missed (in practice approval prompts sit on screen until
answered, so they are not missed).
- The statusline must show the model name and a
[ctx:NN%] field — adapt the
regex in the script to your statusline format.
- POMODEL's baseline is the first observation after arming; a switch that
happened before arming is invisible — the patrol's model sweep covers
that.
- A rotation that replaces the pane (new-pane style) requires re-arming; the
same-pane
/clear style does not (pane ID unchanged).
Anti-goals
- Do not retire the patrol — the watch only sees "stopped". "Running but not
progressing" (declaration stalls, hollow output) is only caught by actually
reading panes on patrol.
- Do not add worker lanes as subjects — lanes belong to each PO's
herdr-event-watch.
