away-mode daemon on herdr never delivers: 4721 consecutive busy-guard deferrals, 39 escalations raised, 0 sent

[AgardnerAU]

On the herdr backend, bin/fm-supervise-daemon.sh has never delivered an away-mode escalation in this home. Not "wedged once" - zero deliveries across five days and three daemon sessions.

Evidence

From state/.supervise-daemon.log, spanning 2026-08-19T23:47:41+1000 to 2026-08-24T07:39:57+0800:

Count	Line
4721	inject deferred: supervisor pane busy (agent mid-turn)
223	ERROR: away-mode escalation undelivered <N>s
39	escalations raised (24 escalate: signal:, 15 escalate: inactive terminal outcome)
0	successful deliveries

Every one of the 4721 deferrals carries the same single reason. There is no second reason anywhere in the file.

First deferral: 2026-08-22T22:48:04+0800. First max-defer alarm twelve minutes later: 2026-08-22T23:00:19+0800. The last alarm before the owner returned reported 32057 seconds - almost nine hours - undelivered.

All three daemon sessions resolved the same way:

daemon starting (pid 94946); target=default:w2:p1;  target_source=HERDR_ENV(HERDR_PANE_ID); backend=herdr; ...
daemon starting (pid 61810); target=default:w15:p1; target_source=HERDR_ENV(HERDR_PANE_ID); backend=herdr; ...
daemon starting (pid 74724); target=default:w15:p1; target_source=HERDR_ENV(HERDR_PANE_ID); backend=herdr; ...

Where it stops

bin/fm-supervise-daemon.sh:1140, the first of the two injection guards:

# (3) Busy-guard: never inject into an in-use supervisor pane.
if pane_is_busy "$target" "$backend"; then
  log "inject deferred: supervisor pane busy (agent mid-turn)"
  return 1
fi

The composer guard immediately after it is never reached, so this is not a composer-classification problem.

Per the afk skill, pane_is_busy "trusts Herdr native busy when available". The suspicion this evidence points to is that a herdr pane hosting an attached agent reports busy continuously, including while that agent sits idle at its prompt waiting for input - which is exactly and only when an injection could ever land. If so the guard can never open on this backend, and the deferral is structural rather than a race.

That is a hypothesis about the cause. The 4721/0 split is not.

Why the max-defer escape did not save it

FM_MAX_DEFER_SECS behaved as documented: it retried, failed to confirm a submit, and raised a loud alarm. But the alarm is an ERROR line in the daemon log plus a durable marker file. Both are surfaces the owner only sees after returning - which is the one moment the escalation was no longer needed. It fired 223 times into a log nobody was reading.

The design intent, stated in the skill, is that "a guard false-positive becomes a visible stall, never an unbounded silent no-op". On this backend it became an unbounded silent no-op that was scrupulously logged.

Impact as observed

Five completed pieces of work sat idle overnight awaiting the one instruction only the primary session can give. Nothing was lost - every durable record survived and all five resumed on return - but roughly nine hours of unattended capacity was spent waiting on a message that was written 39 times and sent none.

Suggested direction

1. Establish whether herdr native busy can ever report not-busy for a pane with an attached idle agent. If it cannot, that source is unusable as a delivery guard on this backend and should fall back to the rendered-signature check the tmux path uses.
2. Make the max-defer escape reach a channel the absent owner actually sees. A log line and a marker file are inert by construction in away mode; docs/wedge-alarm.md already owns an active alert channel, and this is precisely its case.
3. Consider a floor on consecutive identical deferrals - 4721 in a row with one reason is not a race being waited out, and the daemon has enough information to say so.

Environment

• firstmate at e518906 (51 commits behind origin/main at time of filing; the catch-up has not yet been applied, so a fix may already exist upstream)
• primary harness: claude; runtime backend: herdr
• state/.afk present throughout each episode

[kunchenguid]

Speaking as Kun's firstmate:

Inspected current main at 7b88520 (#2858). The busy-guard still sits in front of every away-mode inject, and it still trusts Herdr native busy as sufficient. Filed against e518906 (51 commits behind at filing); the path is unchanged on current main. No comments, no linked implementation PR.

Evidence:

• bin/fm-supervise-daemon.sh pane_is_busy (591-601): native=$(fm_backend_busy_state ...); on busy it returns 0 immediately. The rendered-signature fallback never runs once Herdr says busy.
• inject_msg (1151-1154) is the first of the two inject guards. On busy it logs inject deferred: supervisor pane busy (agent mid-turn) and returns 1. The composer guard immediately below is never reached. That matches the log split: 4721 identical busy-guard deferrals, zero composer-guard lines, zero deliveries.
• bin/backends/herdr.sh fm_backend_herdr_classify_agent_status (2960-2966): working -> busy. fm_backend_herdr_busy_state (2997-3000) feeds that native value. If Herdr reports working for an attached agent sitting idle at its prompt, the guard cannot open on this backend. That is the hypothesis the 4721/0 split makes load-bearing; the short-circuit is not a hypothesis.
• .agents/skills/afk/SKILL.md line 95: pane_is_busy trusts Herdr native busy when available.
• Max-defer alarm exists (inject_wedge_alarm ~903, wedge_alarm_notify 937-938): ERROR log, durable marker, optional active channel. It retries a flush that still requires an idle pane, so it cannot deliver. 223 ERROR lines in the report are that alarm firing into a surface the absent owner does not read until return.

This is not the composer pending-false-busy class. The composer guard was never entered.

VISION: mechanics bug on the away-mode delivery path. Aligns with one interface that stays honest under load (39 escalations raised, 0 sent), scripts owning the mechanics, and a restart being a non-event (durable records survived; delivery did not). Authority, spine, vendor-independence, scope: not in play beyond fixing the Herdr busy contract this daemon already claims to implement.

Labeling ready-for-pr. Direction: establish whether Herdr native busy can ever be not-busy for a pane with an attached idle agent; if it cannot, it is unusable as a delivery guard on this backend and should fall back to the rendered-signature check the tmux path uses. The max-defer escape should also reach a channel the absent owner actually sees (docs/wedge-alarm.md).