Repository navigation
ci: each workflow once per PR commit, and no PR blind to a push path (S37) #875
Description
Part of #851 (W14, CI pool time).
What
Three ways a workflow's on: block either spends runner time twice or skips a PR. They are classified by speed doubles in the build-speed loop's tooling (pure text, 16 fixtures). The tool lives outside this repo.
| class | meaning | main 57de4a1 |
|---|---|---|
DOUBLE_ON_PR |
push with no branch filter plus pull_request: every commit to a PR branch starts the workflow twice |
15 |
PR_BLIND |
push on main only, and a push path missing from the pull_request paths: a PR that touches only that path runs nothing before merge |
3 |
DOUBLE_GLOB |
push on globbed branches (feat/*, fix/**, ...) plus pull_request: twice for PRs from matching branches only |
4 |
PUSH_ALL_ONLY |
push on every branch, no pull_request |
1 |
| OK | everything else | 99/122 |
Counted from the Actions API: over 24 h, 351 of 374 push runs on non-main branches came from DOUBLE_ON_PR workflows.
Fix, in two PRs
- ci: run the nine document gates on push to main only #867 restricts
pushto main for the nine unfiltered document gates. - The S37 PR (branch
ci/push-pr-once) handles the six path-filteredDOUBLE_ON_PRworkflows and the three that arePR_BLINDon main. Eachpull_requestpath list now holds every push path.
A push: branches: [main] change on its own would have added PR_BLIND workflows. Five of the six listed their own workflow file under push paths only, so the push run on the PR branch was the only run that tested an edit to the workflow. That is the hand-copied-list defect again: two path lists in one file, kept in step by hand.
| DOUBLE_ON_PR | PR_BLIND | OK | |
|---|---|---|---|
| main | 15 | 3 | 99/122 |
| S37 branch | 9 (#867's nine) | 0 | 108/122 |
| S37 + #867 (expected; not measured on a merged tree) | 0 | 0 | 117/122 |
Deferred
- A repo-level gate (
tools/check_*.pyfailing onDOUBLE_ON_PRorPR_BLIND) waits until ci: give the 102 push workflows a dispatch-safe concurrency group #870 merges. Both would editgates-can-fail.yml, and the checker should land on a tree where the count is already 0. DOUBLE_GLOB(brain-ci, codegen, fpga-formal, fpga-formal-comb): whether pushes tofeat/*andfix/**should run without a PR is the owner's call.
Not measured
- Runner minutes saved. The run count is the measured quantity; minutes per run vary by workflow.
- Whether any external tooling relies on a push run for a non-main branch (a status check required by branch protection on a non-main branch, for instance).
🤖 Generated with Claude Code
Activity
gHashTag commentedon Oct 5, 2026
A measurement for the
DOUBLE_GLOBdecision, brain-ci only. It is still the owner's call; this is only the cost.brain-ci's last 200 runs (2026-10-02 14:47Z to 2026-10-05 03:37Z) covered 146 commits. 45 of them ran twice, once as
pushand once aspull_request: 44 onfix/*branches, 1 onfeat/*. Events: 91 push, 95 pull_request, 13 workflow_dispatch, 1 schedule.What each extra run costs:
The push half of a double tests the branch head, and the pull_request half tests the merge with main. If brain-ci's
pushwere narrowed to[main, 'v*'], afix/*branch with no PR would get no brain-ci run until a PR is opened or someone dispatches it.workflow_dispatchis already on the workflow.🤖 Generated with Claude Code
gHashTag commentedon Oct 5, 2026
The other three
DOUBLE_GLOBworkflows, measured the same way as brain-ci above:gh run list --workflow <file> --limit 400, then the commits that ran on bothpushandpull_request.fpga-formal.ymlfix/, 2feat/)fpga-formal-comb.ymlcodegen.yml(ralph/dev-*)Wall clock includes queue time, so 235 min is an upper bound on runner time, not billed minutes.
So the glob doubles that cost anything are brain-ci's (45 of 146 commits, earlier comment).
fpga-formal.ymlis about 2.4 min a day over 97 days. The other two have never doubled in the window read. No change proposed for these three. The brain-ci decision is still the owner's.Not measured: whether the push runs on
fix/**gave anyone feedback before the PR was opened, which is the reason to keep them.🤖 Generated with Claude Code