fix(relay): count in-flight queue jobs in the running-box gate

[madarco]

Problem

agentbox <agent> -i "<prompt>" --max-running N (the running-box concurrency gate) could start more jobs than the cap. Observed live on Vercel: two jobs submitted back to back with --max-running 1 both started immediately, creating two boxes.

Root cause in packages/relay/src/queue.ts:

• The running-box gate counts boxes via defaultCountRunningBoxes, which reads ~/.agentbox/state.json. A just-started job's box doesn't appear there until the worker finishes provisioning (~25s cloud, image pull docker). During that window the gate saw 0 running, so the scheduler's per-tick "start as many as fit" loop re-selected the same free slot for the next job.
• The 3s result cache compounded it by returning the stale pre-start count within a single tick.
• The working-agent gate (--max-working) was already correct — it counts in-flight running jobs. Only the running-box gate lacked that term.

Provider-agnostic (docker affected identically; just easier to hit on cloud where provisioning is slower).

Fix

• Extract countInFlightCreateJobs(jobs, accountedBoxIds) — the in-flight term the working-agent gate already used — and share it across both gates.
• defaultCountRunningBoxes now adds that term, computed fresh each call (cheap loadQueue read), while only the expensive box-state portion (state.json + docker inspect) stays cached for 3s. Dedup by state.json box ids; dead-pid workers skipped.
• defaultCountWorkingBoxes is behavior-preserving (now calls the shared helper).

Verification

• Unit (pnpm --filter @agentbox/relay test): +5 countInFlightCreateJobs tests (in-flight counted, deduped once box accounted, non-running ignored, dead-pid skipped, multi-sum). Existing working-gate + selector tests stay green. 125 relay tests pass. lint + full-workspace typecheck clean.
• Docker e2e: two -i jobs with --max-running 1 → submit-time count shows 1/1, job A runs (one container), job B stays queued; destroying A's box freed the slot and B started within a tick. No over-start, no leftover containers.

Note

The original cloud repro (--provider vercel) depends on #34 (cloud -i support). This fix is independent and lands on main; the docker path exhibits and is fixed for the identical bug.

---

Note

Medium Risk
Changes core queue scheduling and concurrency accounting; behavior is well-tested but affects when jobs start across providers.

Overview
Fixes --max-running over-starting when multiple queued jobs are picked in one scheduler tick: the running-box gate only looked at state.json, so a job just flipped to running still looked like zero occupancy until provisioning finished.

defaultCountRunningBoxes now adds the same in-flight term the working-agent gate already used, via a shared countInFlightCreateJobs. Only the expensive box-state read stays on the 3s cache; queue manifests are re-read each call so a freshly started job counts immediately. defaultCountWorkingBoxes is refactored to call the helper (behavior unchanged). Five unit tests cover dedup, dead PIDs, and status filtering.

^{Reviewed by Cursor Bugbot for commit 7c2e033. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agentbox-web	Ready	Preview, Comment	May 31, 2026 1:02pm