Skip to content

Pin each Fox CI job to one 8-CPU block, up to 4 at once - #41

Merged
quolu merged 4 commits into
mainfrom
dorothy/fox-pool-max1-20261114
Nov 14, 2026
Merged

quolu merged 4 commits into
mainfrom
dorothy/fox-pool-max1-20261114

Conversation

@quolu

@quolu quolu commented Nov 14, 2026 •

Copy link
Copy Markdown
Collaborator

Shiratsuru assignment (chart bot-38094207 / ci-fox-concurrency f2), changed by the Owner's ruling 2026-11-14: instead of max 1 concurrent job on Fox, each job gets one 8-CPU block of Fox's 32 logical CPUs, at most 4 jobs at once.

  • Diagram first: docs/feature-map/runner-capacity.md "Fox per-job CPU budget": rule, reason (CPU oversubscription 2026-11-14, Elegg's evidence: 3-4 full suites at once, each sized to 32 CPUs, FACTORY_CI_JOBS=32, Aiterm Windows-only timing failures), mechanism, probe, rollback. Hook row, contract row and Why paragraph updated.
  • FACTORY_RUNNERS windows-native maxRunning: 4, cpusPerJob: 8; factory-current-state.md regenerated.
  • bin/windows-factory-runner-job-started.ps1: Runner.Worker affinity = the 8-CPU block overlapping the fewest other Runner.Workers (lowest index on ties), chosen under the named mutex dotagents-factory-runner-cpuset; BelowNormal kept; any error still only warns (exit 0).
  • Probe on Fox (Node v24.22.1): a process started under an 8-CPU affinity reports availableParallelism() 8 (os.cpus().length 32), so FACTORY_CI_JOBS and node --test follow the block.
  • Tests: hook default equals contract cpusPerJob; child sees 8; a second job avoids a block already in use (Windows); scaler/current-docs/profile unchanged and passing.

Apply: restage on Fox and run windows-factory-runner-instances.ps1 -Mode Apply (hook + scaler config) while no Runner.Worker exists (Owner approval K-PX9P2R). Rollback: previous hook (git 9d04fb6) and config (no maxRunning), task restart while idle.

🤖 Generated with Claude Code

Shiratsuru decision 2026-11-14 from Elegg's evidence: 3-4 full Windows
suites ran at once on Fox, each sizing itself to all 32 CPUs, and
Aiterm hit Windows-only timing failures. Same mechanism as the Mac cap
(#38): windows-native maxRunning 1; the feature map gets the
"Fox concurrency cap" rule, reason and reopen condition.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@quolu
quolu marked this pull request as draft November 14, 2026 05:55
Owner ruling 2026-11-14 replaces the max-1 cap: the promise is parallel
CI, and one job using all 32 CPUs is the irregularity. The job-started
hook now sets the Runner.Worker affinity (inherited by steps) to the
8-CPU block overlapping the fewest other Runner.Workers, under a named
mutex; the scaler caps Fox at 4 concurrent jobs (32 / 8). Probe on Fox:
Node under an 8-CPU affinity reports availableParallelism 8, so
FACTORY_CI_JOBS and node --test follow the block.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@quolu quolu changed the title Cap the Fox factory runner pool at 1 concurrent listener Pin each Fox CI job to one 8-CPU block, up to 4 at once Nov 14, 2026
@quolu
quolu marked this pull request as ready for review November 14, 2026 05:57
quolu and others added 2 commits November 14, 2026 06:02
The generator now renders each seat's cpusPerJob from the deployment contract
instead of the stale "no CPU assignment" line; other seats stay unassigned.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…election test

join('; ') put a ';' between try {} and finally {}, which PowerShell rejects
as a parse error, so the test failed on Fox before reaching the hook.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@quolu
quolu merged commit e647e76 into main Nov 14, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant