CI was burning ~2,340 min/month — now fixed, here's the breakdown

[CalebisGross]

Summary

The ci/release-only change (#260, merged today) reduced CI from running on every PR and push to only running on release-please PRs and tag pushes. This issue documents the before/after numbers.

Before (March 1–20 actuals from Actions Usage Metrics)

Workflow	Runs	Linux min	Windows min	Total billed min
ci.yml	364	1,354	985	2,339
release-please.yml	83	83	—	83
release.yml	~19	31	3	51*
dependabot-updates	7	7	—	7
Total				~2,493

*Release also runs macOS builds (10x multiplier) — the 17 min macOS shown in the screenshot accounts for most of release.yml's cost.

Where 2,339 CI minutes came from

Trigger	Runs	Est. min	Notes
PR (pull_request)	213	~1,363	Every PR to main
Push to main	151	~966	Redundant — CI already passed on the PR
Total CI	364	~2,339

The push-to-main trigger alone was responsible for ~966 min — 39% of all CI — re-validating code that was already validated on the PR.

Each CI run spins up 7 jobs (test + build on both ubuntu and windows, plus lint and two gate jobs), billing ~6.4 min on average. Windows jobs are billed at 2x, which is why 985 billed minutes come from Windows despite each Windows job only running ~2 min.

Top branches by CI runs

Branch	Runs
main (push)	50
release-please--branches--main	18
All feature/fix branches combined	~296

After (projected with ci/release-only)

Workflow	Est. runs/month	Est. billed min
ci.yml (release-please PRs + tags only)	~24	~154
release-please.yml	~83	~83
release.yml	~19	~51
Total		~288

Estimated savings: ~2,200 min/month (88%)

What changed

PR #260 modified ci.yml to:

1. Remove push: branches: [main] trigger — no more redundant CI on merge
2. Add a should-run gate job that checks if the PR is from a release-please--* branch
3. All other jobs depend on should-run and skip if it outputs false
4. Keep push: tags: ["v*"] so CI runs as a final check alongside release builds
5. Keep workflow_dispatch as an escape hatch

Regular PRs now rely on local pre-commit hooks (go fmt, go vet, golangci-lint) for quality.

Remaining optimization opportunities

• Drop Windows from CI matrix: Windows jobs account for ~40% of CI minutes due to the 2x multiplier. The release workflow already builds and tests on Windows. Removing Windows from CI and relying on release builds would save another ~40% of remaining CI minutes (~60 min/month). Low risk for a pure-Go project with no Windows-specific code paths in CI.
• Release builds macOS 10x multiplier: Each macOS build minute costs 10 Linux minutes. The current macOS build is fast (~0.7 min raw = ~7 billed), but worth noting if release frequency increases.

Data source

• Screenshot: Actions Usage Metrics page, March 1–20, 2026
• Run counts: gh api repos/AppSprout-dev/mnemonic/actions/runs (deduplicated, 494 total runs)
• Job timings: gh run view on representative runs

[CalebisGross]

This was a post-mortem documenting the fix, not an open task. The CI change (#260) has been live since March 20 — CI only runs on release-please PRs and tag pushes now. Closing as resolved.