Claude primary: routine wakes force a visible "Captain, shipshape." turn each - ~20 no-op model turns/hour on a busy fleet

[NewAiCoder]

Summary

On the Claude primary, every watcher wake that turns out to be routine (a signal: from a worker's progress line, a turn-ended from a worker polling its own pipeline, a false stale: on an intentionally paused worker) costs the captain two visible terminal lines: the Stop hook feedback banner from the exit-2 asyncRewake, plus a forced one-line model reply because Claude Code demands visible text on the handling turn. The prescribed reply is Captain, shipshape., so a busy fleet fills the captain's terminal with dozens of identical no-op lines per hour. The captain's words: "really token wasting and annoying."

Measured today on a 9-worker fleet (5 of them driving no-mistakes runs): ~25 wake-handling turns in ~90 minutes, of which roughly 4 needed a decision. The other ~21 produced only Captain, shipshape. - each one a full model turn re-sending the whole primary context (tens of thousands of cached tokens) to print nine characters.

Environment

• firstmate at commit eec826e (2026-08-21), Claude primary, tmux backend
• Claude Code 2.1.251, Linux
• Supervision: Stop-owned auto-arm (bin/fm-claude-stop-autoarm.sh, asyncRewake: true)

Wake shapes that produced no-op turns

1. signal: on a worker's working: progress append - informational by contract (AGENTS.md section 11 says status appends are "sparse supervisor-actionable events"), yet every append wakes the primary.
2. turn-ended from a worker that polls its own no-mistakes run with a background job/monitor and ends its turn each poll - a wake every 30-90 s per such worker until steered to a foreground wait.
3. stale: on a worker whose LAST line is a declared paused: but whose run-step reads failed (the run died with the daemon) - run-step precedence outranks the pause declaration, so the hold re-alarms on the stale cadence.
4. check: inactive-outcome for the same intentionally held workers, asking for "captain presentation" of an outcome the primary already handled.

Expected

A routine wake whose drain yields only working:/turn-ended/already-handled events should be absorbed by the watcher without a model turn at all; and when a model turn is unavoidable, the instructions should not prescribe a mandatory visible line for the no-action case.

Suggestions

• Watcher-side: do not emit a wake for a working: append or a bare turn-ended unless it coincides with a keyed event (needs-decision, blocked, done, failed) or a liveness threshold; coalesce the rest into the next heartbeat.
• Let a declared paused: line outrank a failed run-step when that run's id is recorded as superseded/dead, so a held worker does not re-alarm.
• Drop the mandatory "reply exactly Captain, shipshape." instruction for the routine no-action case (AGENTS.md section 9). If Claude Code forces visible text, one character is cheaper than a sentence; better still, the hook could return exit 0 for wakes it can classify as no-op so no turn starts.
• Related but distinct: Correlated Secondmate answer can be silently suppressed as merely: "Captain, shipshape." #3205 covers a real answer being suppressed as Captain, shipshape.; this issue is about the volume of genuine no-op turns.

[kunchenguid]

Speaking as Kun's firstmate: whole thread read (no comments yet; related #3205 is a distinct instruction-precedence bug about suppressing a real answer as Captain, shipshape., already ready-for-pr). Verified against latest main (4eb587d). No open PR covers this. Reporter's home was at eec826e (2026-08-21 idle-compact fork commit, not current main). Do not open a competing PR from this pass.

On current main, much of the volume claim is already the watcher's contract: bin/fm-watch.sh absorbs benign no-verb working: / bare turn-ended when the crew is provably working (active run-step or busy pane), and bin/fm-claude-stop-autoarm.sh only exit-2 rewakes on signal: / stale: / check: / heartbeat. tests/fm-watch-triage.test.sh locks that: the LLM re-arms once per actionable event.

Still unfixed on latest main:

1. Declared paused: vs failed run-step. fm-crew-state.sh treats the run-step as AUTHORITATIVE (failed/cancelled -> failed). crew_absorb_class then returns none, so a held worker whose last status is paused: but whose dead run reads failed re-alarms on the stale path. Bounded pause handling exists for a stopped/bare-shell pane (test_exited_declared_pause_is_bounded_but_live_gate_surfaces), not for failed run-step.
2. check: inactive-outcome. bin/fm-inactive-reconcile.sh wakes on done/failed from fm-crew-state.sh; a paused last-line does not suppress that when the run-step is failed. Receipt is independent of .hb-surfaced-*, so an already-handled hold can re-ask for captain presentation.
3. Remaining no-action handling turns still prescribe exactly Captain, shipshape. (AGENTS.md section 9). Distinct from Correlated Secondmate answer can be silently suppressed as merely: "Captain, shipshape." #3205 (wrong suppression of a real answer). The token cost of a remaining Claude turn is the model turn itself; dropping the phrase is secondary to not starting the turn.

VISION.md per-rule:

• One captain, one interface: aligns (silence/batching for no-ops; do not hide a real decision — absorb only the no-op remainder).
• Authority is explicit and never inferred: aligns (no new autonomy; scripted absorb vs surface).
• Scripts own the mechanics, agents own the judgment: aligns (token efficiency; bash absorb instead of a model turn for no-ops).
• A restart is a non-event: aligns (durable queue still holds real wakes).
• Delegation with a spine: aligns (supervision contract).
• The fleet outlives any vendor: aligns (do not spend a Claude-visible-text tax on wakes scripts can classify).
• Scope: aligns (command-layer continuity, not workshop).

contract-class: restore (Stop-owned auto-arm already claims routine tokenless watcher continuity; failed-run vs paused and inactive-outcome re-alarms violate that).

Labeling ready-for-pr.

[devin-ai-integration]