workflow: a late spawn failure leaks the Bots it already made #2785

Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment on this issue.

Written-by: FlyBot Lead
Claimed-by: FlyBot Lead
Claim-Id: c7b0191a904142ccac1611e93a33ff4f
Branch: dev/5-late-spawn-release
Base: 3db12f0
At: 2026-11-10T06:20:44Z

Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment.

Written-by: FlyBot Lead
Bot: FlyBot Product

What happens

boundedSpawn (plugins/workflow/src/service.js:423-443, added for #2748) bounds the wait for a host spawn. When the host answers after the deadline with success, it releases the Bots that answer created:

Promise.resolve(pending)
  .then(ids => releaseAll(Object.values(ids ?? {}), release, warn))
  .catch(error => warn(`flybot-workflow: a late start could not release its Bots: ${error?.message ?? error}`))

A late answer that fails does not release anything: the .then handler above only runs on success, and the .catch merely warns. spawnAll (plugins/workflow/src/ephemeral.js:46-62) already put the ids it had made onto that error (:61 error.spawned = made), so the rejection carries exactly the Bots that now leak:

} catch (error) {
  error.spawned = made
  throw error
}

The early-answer path does not leak: startLive's catch (service.js:1299) calls releaseAll(error.spawned ?? [], …) on the same error. Same error object, released when it arrives before the deadline, dropped when it arrives after it.

The leak is permanent, not deferred:

• the start is refused before beginRun(), so no run ever holds those ids in its ephemeral map;
• reap() only walks lane.active / lane.last (service.js:1403-1470), and orphanIds() only reads state.orphans (:2421); this path writes neither.

So Bots created at hidden: true, ephemeral: true live until the process exits — while changelog.d/2748-workflow-stuck-run.fixed.md and docs/subsystems/workflow.md both state that a late spawn's Bots are released.

Reproduction

At b89ea5ecef80a996bc912f4e90c538d0f0ff2cbb (base here), with a spawned host whose create never answers inside SPAWN_DEADLINE_MS and then rejects with error.spawned set:

• late failure branch: released: [] plus flybot-workflow: a late start could not release its Bots: … — fails;
• the same partial failure answered before the deadline: released: ["eph-plan"] — passes.

# tests 2 / # pass 1 / # fail 1, failing line not ok 1 - a late spawn failure still releases the Bots it made.

Expected

A start refused at the deadline releases every Bot the pending spawn had already created, whether the host's late answer is a success or a failure. The refusal itself (error.code === 'spawn-timeout') does not change, and the existing warn stays so a failed release is still visible.

Fix shape (verified, 3 lines)

In the deadline branch, release late.spawned when the late answer rejects, then warn:

.catch(late => releaseAll(Array.isArray(late?.spawned) ? late.spawned : [], release, warn)
  .then(() => warn(`flybot-workflow: a late start was refused: ${late?.message ?? late}`)))

The sentence changes from could not release its Bots to was refused, because after this fix the release has already happened when that line prints; releaseAll keeps warning once per Bot it really could not remove, so a failed release stays visible.

Verified at the same base: 2/2 green, plugins/workflow full suite 515/515.

Scope

• plugins/workflow/src/service.js only (the deadline branch of boundedSpawn), plus a regression test.
• No core/, no client, no patches/ds-bot mirror ⇒ no generate, so this does not take the single generate lane.
• Risk tier T2: plugins/** is a T2 row of .github/agent/risk-rules.json, and this change adds no diffPatterns hit (no DSH_HOME, no child_process, no fenced-prose line), so nothing bumps it.

Dedup

Read live main at the base above, no-cache, and searched issues and PRs for spawn-timeout, SPAWN_DEADLINE, spawned, 2748: #2735 is the "refused start puts the card back in the queue" half, #2752 (open) is "a failed run does not drain", #2753/#2750 are Stop release retry limits. No open issue or PR covers a late failure leaking the Bots it made.