gstack 1.91.24.0 keeps cancelled PR eval work and caps stuck CI runs

1 week ago

What changed in gstack 1.91.24.0: a cancelled PR eval run now hands its finished results to the next push, PR evals run only the paid cases a change can affect, stuck paid shards end at a per-runner ceiling, and a weekly report names flaky and slow free tests. Skill behavior is unchanged; /cso contained runtime and scanner runs are not available yet. https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#191240---2026-10-05 https://github.com/garrytan/gstack/commits/main

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Cancelled runs keep their work
  2. 0:15Narrower PR selection
  3. 0:29Per-runner ceilings
  4. 0:39Weekly test health
  5. 0:50Three CI fixes
  6. 1:02/cso not available yet
Show transcript

Cancelled runs keep their work

gstack logo

gstack 1.91.24.0 keeps a cancelled PR eval run's finished work

1.91.22.0
0 results

handed to the next push

1.91.24.0
Every finished shard

failures included

ReuseSelectionCeilingsHealthFixes/cso
garrytan/gstack · CHANGELOG.md

In gstack 1.91.24.0, a cancelled pull request eval run hands its finished passes and failures to the next push. So a re-push re-runs only what changed. Skills behave exactly as before.

Narrower PR selection

PR evals run only the paid cases a change can affect

Mapped to their cases

Free tests, free fixtures, free-only workflows, and files traced through imports and build outputs

Still the full gate

package.json, the CI image, setup, the eval policy

ReuseSelectionCeilingsHealthFixes/cso
garrytan/gstack · CHANGELOG.md

Pull request evals now run only the paid gate cases a change can affect, traced through imports, bin names and build outputs. Edits to shared inputs, such as package.json or the CI image, still run the full gate.

Per-runner ceilings

A stuck paid shard ends as TIMEOUT instead of holding a runner for four hours

254 34–82

minutes, gate census job ceiling per slice

ReuseSelectionCeilingsHealthFixes/cso
garrytan/gstack · CHANGELOG.md

Every paid lane now has a ceiling sized to its runner. A stuck shard ends as an honest timeout, or as not run, instead of holding a runner for four hours.

Weekly test health

A weekly report names the free tests that are getting flaky or slow

bun run test:healthMondays, on main
flakyfails in over 5% of main runs
ReuseSelectionCeilingsHealthFixes/cso
garrytan/gstack · CHANGELOG.md

A new weekly test health report runs every Monday. It lists free tests that turn flaky, failing in more than five percent of main runs, and tests that are getting slow.

Three CI fixes

The weekly census survives branch dispatches, and Windows tests finish sooner

Weekly censusA branch dispatch can no longer cancel main's census
Windows free lane8.1 min on one runner, now about 4.6 min on six
zsh portability testsAbout 42 skipped zsh arms now all run in CI
ReuseSelectionCeilingsHealthFixes/cso
garrytan/gstack · CHANGELOG.md

A branch dispatch can no longer cancel main's weekly census. Windows free tests finish in under five minutes instead of eight. And every zsh portability test now runs in CI.

/cso not available yet

/cso contained runtime and scanner runs are not available yet

CHANGELOG.mdFull 1.91.24.0 notes, with the retired scripts and their replacements
ReuseSelectionCeilingsHealthFixes/cso
garrytan/gstack · CHANGELOG.md

The /cso docs now say contained runtime and scanner runs are not available yet. Zero runtime profiles are qualified, and the scanner catalog is empty. The full notes for this release are in the changelog.