Skip to content

Add hosted PR performance calibration workflow - #2087

Merged
justinjoy merged 1 commit into
mainfrom
codex/2085-hosted-perf-evidence
Oct 5, 2026
Merged

justinjoy merged 1 commit into
mainfrom
codex/2085-hosted-perf-evidence

Conversation

@justinjoy

Copy link
Copy Markdown
Collaborator

Summary

Add a manually dispatched, hosted Ubuntu calibration campaign for an open PR. It freezes the PR base, head, and merge SHAs, then records six sequential runner observations. Each observation builds and measures two immutable revisions, with balanced base/merge order and same-revision controls.

The collector preserves raw CRDT and CSPA samples, correctness, CoV, target verdicts, build and host metadata, logs, and partial artifacts. It classifies target misses separately from skips and failures; incomplete or ineligible campaigns exit unsuccessfully.

This supplies the evidence needed for #2085 and the failing Perf Suite on #2077. It does not change the required PR performance check or its targets. The workflow uses GitHub-hosted ubuntu-latest, not the dedicated perf runner.

Validation

  • 17 focused calibration tests passed in a uv virtual environment.
  • actionlint .github/workflows/perf-hosted-calibration.yml passed.
  • Python compilation and git diff --check passed.
  • Independent Reviewer, Architect, and Critic approved commit 89422921457b820ad9244c44ef06389fd3b0f9b6.

Next step

After this workflow is available on main, dispatch it with PR number 2077. Review eligible paired evidence before changing the required performance gate or #2077 source. A skipped or incomplete campaign does not count as a timing result.

Refs #2085; diagnostic for #2077.

@justinjoy
justinjoy merged commit 84487d5 into main Oct 5, 2026
11 of 17 checks passed
@justinjoy
justinjoy deleted the codex/2085-hosted-perf-evidence branch October 5, 2026 09:12
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