[ccch1mneyyy]

Why

voyager serve ran one job at a time, so several agents asking ChatGPT Pro through one browser queued behind each other's hour-long replies. With the companion Voyager change, the person can turn on Run jobs in parallel in the popup's Local bridge card; serve then runs that many jobs at once (up to 10, or fewer with "maxParallelJobs" in ~/.voyager/config.json). Switch off, or an older extension: nothing changes, one job at a time and busy (exit 5) for the next.

Companion PR (extension + @voyager-crew/bridge-protocol 0.5.0): voyager-crew/voyager#1067

What changes

submit ── admission.ts ──▶ busy?  capacity = min(extension maxTasks, maxParallelJobs)
                                   conflict key: same conversation / any project change
                                   voyager call runs only alone (it borrows a free tab)
job ── OpsClient.call(op)
        ├─ page op (open, select, attach, send, approve, project.*, close)
        │     └─ PageTurns.take(job) … slot.run … slot.idle(job) → give back
        └─ waitReply / read / download → slot.run directly (overlaps other jobs)
      ask: ops.exclusive(model.select → composer.attach → composer.send)
TaskSlot: one entry per job; a poll settles every entry, then sends one message
          (an abort first, then the oldest task due)

• JobRunner keeps a map of active jobs; voyager status lists activeJobs and maxJobs ("parallel jobs: up to 10 at once" / "off").
• PluginState.update() serializes read-modify-write per plugin, so project-upload fingerprints of jobs side by side no longer overwrite each other.
• contrib/web-pro-kit: drive_request.py and drive_mirror.py push take a lock on this computer ($VOYAGER_HOME/locks/, never in Drive) for what they publish; two agents could otherwise lose a request or leave a mirror whose MIRROR.json names the wrong archives. drive_wait.py already watches one task folder each; the runbook now says one task ID and one waiter per job, and one mirror per commit.
• README, README_ZH, the agent skill and CONTRIBUTING describe the switch, the cap and the conflict rule.

Before merge

This depends on @voyager-crew/bridge-protocol 0.5.0 (maxTasks, activeJobs, maxJobs), which is staged by the Voyager PR. package.json already asks for ^0.5.0, and bun.lock is left at 0.4.0 until 0.5.0 is on npm, so CI's bun install --frozen-lockfile fails until then: after publishing, run bun install and commit the lockfile. Kept as a draft for that reason.

Testing

• bun run verify (format, lint, typecheck, 333 tests) and bun run build, against the unreleased protocol package via a local file: link.
• New tests: test/jobs/parallel.test.ts (capacity from maxTasks, maxParallelJobs, conflict keys, voyager call alone, page ops never overlap across jobs while replies do; mutation-checked by disabling the page turns), test/jobs/pageTurns.test.ts, slot tests for several jobs, serialized plugin-state updates. The fake extension now runs several tasks at once and records a timeline.
• Kit: every script's --selftest passes; drive_request.py's new two-writer check fails without the lock (Version: 2) and passes with it.
• Live, Chrome 155 with the Voyager branch build and this serve: three voyager ask --model extra-high at once all succeeded, awaited together, each reply carrying only its own token; a second ask into the same conversation was busy with error.details.conflict; with the switch off the second of two asks was busy (maxJobs: 1). Not tried live: GPT-6 Pro itself (quota) and more than three jobs at once.

[Nagi-ovo]

Thanks so much!