PYR0D-353: Count only live workers against the away spend cap, and show headroom at entry #41

Merged

john merged 9 commits from fm/fm-away-spend-cap-live into main 2026-10-02 06:39:05 +00:00

Conversation 0 Commits 9 Files changed 10 +820 -40

john commented 2026-10-01 16:29:24 +00:00

Owner

Copy link

Story: PYR0D-353 - Away mode: count only live workers against the away spend cap, and show headroom at entry

What was wrong

On 2026-10-01 away mode did no work for 4h21m. The away spend cap in bin/fm-spawn.sh counted every task record that wasn't a secondmate, including workers that had already exited and were only parked on a PR waiting for a merge. At entry, 7 records counted against the default cap of 4 while one worker was running. The captain's words held merges, so the parked records could never drain, and every spawn was refused all night. This is upstream bug kunchenguid/firstmate#6139, carried here until a sync brings upstream's fix. The diagnosis is in firstmate's scout report fm-afk-idle-diagnosis.

The return brief made it worse: it reported "the away session took no action" and "0 outcome(s)". On the daemon path nothing writes the outcome store it reads, so it could not know.

What changed

• New bin/fm-away-spend-lib.sh holds one shared count of live spenders: records whose worker agent is still running. Liveness is read first: an exited worker parked on a PR, a stopped worker and a record whose endpoint is gone do not count; live workers count, including ones paused on their own job or waiting at a gate. The spawn cap, the return brief's "tasks live" count and the entry read-back all use it. If the count cannot be read, the cap refuses rather than letting work through.
• The away entry read-back gains a headroom line: the cap and how many workers count against it now.
• On the daemon path the return brief says the path records no outcomes, instead of claiming no action was taken.
• The afk skill relays the headroom line at entry.

Tests

Colocated tests cover each change and were proven red against the old code:

• Cap 4 with seven records (six parked, one live) allows a spawn.
• Four live workers still refuse.
• A relaunch stays exempt.
• The read-back shows headroom.
• The return brief wording on the daemon path.

The no-mistakes pipeline passed review, test, document and lint at daa7a831.

What stays open

This fixes the first of the two causes. The second is that nothing gives firstmate a turn when the fleet goes idle, and the captain's away words are never re-read on the Claude daemon path. That belongs to the supervision-host adoption (D148 A, fm-adopt-away-host), starting with the Node module-type fix.

Decisions made in review

• Live agents waiting at a gate or on the captain still count. They hold a session and resume without a spawn, so they never pass the cap check again. The first adversarial review (F2/F3) asked for this. Only exited and PR-parked workers, the ones that stalled the 10/01 night, are freed.
• A relaunch is exempt only while its own record counts. Restarting a dead worker takes a slot like any spawn (first review F4). Relaunches are firstmate-driven and serial, so the concurrent-relaunch race is noted, not fixed.
• Launch grace window. A record written within the grace window counts even if its endpoint reads dead, so a just-spawned worker whose agent has not started cannot let a second spawn past the cap.
• Unverified backends (Zellij, Orca, cmux). A capture failure there is labelled "backend target gone" but proves no death, so it counts unless an authoritative check proves the endpoint gone (second review B1).

Review

Two adversarial reviews by other models (Claude Opus at daa7a831, Codex gpt-6-luna at bf7e1ecd), with every blocking item fixed. A third Codex check of the B1 rework at 878752c8 approved. CI is green at 878752c8.

**Story:** PYR0D-353 - Away mode: count only live workers against the away spend cap, and show headroom at entry ## What was wrong On 2026-10-01 away mode did no work for 4h21m. The away spend cap in `bin/fm-spawn.sh` counted every task record that wasn't a secondmate, including workers that had already exited and were only parked on a PR waiting for a merge. At entry, 7 records counted against the default cap of 4 while one worker was running. The captain's words held merges, so the parked records could never drain, and every spawn was refused all night. This is upstream bug kunchenguid/firstmate#6139, carried here until a sync brings upstream's fix. The diagnosis is in firstmate's scout report `fm-afk-idle-diagnosis`. The return brief made it worse: it reported "the away session took no action" and "0 outcome(s)". On the daemon path nothing writes the outcome store it reads, so it could not know. ## What changed - New `bin/fm-away-spend-lib.sh` holds one shared count of live spenders: records whose worker agent is still running. Liveness is read first: an exited worker parked on a PR, a stopped worker and a record whose endpoint is gone do not count; live workers count, including ones paused on their own job or waiting at a gate. The spawn cap, the return brief's "tasks live" count and the entry read-back all use it. If the count cannot be read, the cap refuses rather than letting work through. - The away entry read-back gains a headroom line: the cap and how many workers count against it now. - On the daemon path the return brief says the path records no outcomes, instead of claiming no action was taken. - The afk skill relays the headroom line at entry. ## Tests Colocated tests cover each change and were proven red against the old code: - Cap 4 with seven records (six parked, one live) allows a spawn. - Four live workers still refuse. - A relaunch stays exempt. - The read-back shows headroom. - The return brief wording on the daemon path. The no-mistakes pipeline passed review, test, document and lint at daa7a831. ## What stays open This fixes the first of the two causes. The second is that nothing gives firstmate a turn when the fleet goes idle, and the captain's away words are never re-read on the Claude daemon path. That belongs to the supervision-host adoption (D148 A, `fm-adopt-away-host`), starting with the Node module-type fix. ## Decisions made in review - **Live agents waiting at a gate or on the captain still count.** They hold a session and resume without a spawn, so they never pass the cap check again. The first adversarial review (F2/F3) asked for this. Only exited and PR-parked workers, the ones that stalled the 10/01 night, are freed. - **A relaunch is exempt only while its own record counts.** Restarting a dead worker takes a slot like any spawn (first review F4). Relaunches are firstmate-driven and serial, so the concurrent-relaunch race is noted, not fixed. - **Launch grace window.** A record written within the grace window counts even if its endpoint reads dead, so a just-spawned worker whose agent has not started cannot let a second spawn past the cap. - **Unverified backends (Zellij, Orca, cmux).** A capture failure there is labelled "backend target gone" but proves no death, so it counts unless an authoritative check proves the endpoint gone (second review B1). ## Review Two adversarial reviews by other models (Claude Opus at daa7a831, Codex gpt-6-luna at bf7e1ecd), with every blocking item fixed. A third Codex check of the B1 rework at 878752c8 approved. CI is green at 878752c8.

john added 3 commits 2026-10-01 16:29:24 +00:00

Count only live workers against the away spend cap, and say so honestly in the entry read-back and return brief ... f30b4f7a5a

The away spend cap counted every task record, so exited workers parked on a PR
waiting for a merge held every slot and no spawn could run all night
(kunchenguid/firstmate#6139). The cap, the return brief's tasks-live count, and
a new headroom line in the entry read-back now share one count of records whose
worker endpoint is alive and not parked, held, or done. On the away daemon path,
which writes no outcome store, the return brief says it records no outcomes
instead of claiming the away session took no action.

no-mistakes(review): Fail spend count closed and align brief path detection 99dfca1923

no-mistakes(document): Relay away entry spend headroom in afk skill daa7a8314f

john added 1 commit 2026-10-01 17:30:04 +00:00

Rework the away spend count after review: count paused, gated and blocked workers, cap a dead worker's relaunch, count a read with no result, and decide the daemon path from its flag ... 422459768f

A worker that declared a wait on its own job, a pipeline gate or blocker
firstmate answers while away, and an unreadable endpoint all resume without a
spawn, so they count. Only done and failed records and records whose endpoint is
gone leave the cap. A relaunch is exempt only while its record counts. A record
whose state read left no result counts as spend. The return brief reads whether
the away daemon ran from its mode flag instead of re-deriving it from the
harness and config, which drops the engine-lib helper that collided with
upstream's function of the same name.

john added 3 commits 2026-10-01 19:08:53 +00:00

no-mistakes(review): Count away spend by agent liveness before run verdict f71f9e40e4

no-mistakes(review): Count just-spawned dead endpoints within a launch grace window aa4f78f030

no-mistakes(test): Fix afk test fixtures for liveness-first spend count bf7e1ecdbf

john added 2 commits 2026-10-01 22:56:56 +00:00

Keep an unverified-backend worker in the away spend count when its capture failed ... ff1bc5eca8

fm-crew-state.sh reports "backend target gone" for any capture failure on a backend with no liveness classifier (zellij, orca, cmux). That proves no death, so the spend count treats it as spending there; a verified dead endpoint, a torn-down worktree, and the other authoritative gone reasons still free the record.

no-mistakes(document): Say stopped workers do not count against away spend cap 878752c823

john merged commit 6b0fa30426 into main 2026-10-02 06:39:05 +00:00

john referenced this pull request from a commit 2026-10-02 06:39:08 +00:00

PYR0D-353: Count only live workers against the away spend cap, and show headroom at entry (#41)

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

john/firstmate!41

No description provided.