How the analysis works
Every calculation, threshold and limit. Results are per workload, from the timings you paste, never a universal promise.
What it reads
Two formats, detected from the first character. Everything runs in your browser; nothing is sent anywhere.
gh api repos/OWNER/REPO/actions/runs/RUN_ID/jobs > run.json queue = started_at − created_atstep = step.completed_at − step.started_atjob test (ubuntu-latest) # header: "job …", "## …" or "name:"Queued 41snpm test 7m 46s # 7m 46s · 466s · 7:46 · 0:07:46Phases
Each step goes to the first phase whose keywords match its name, in this order.
| Phase | Matched by |
|---|---|
| Queue | “Queued” / “Queue” lines; in API JSON, started_at − created_at |
| Cache restore | cache, restore, actions/cache, rust-cache, sccache, ccache |
| Install | npm ci, yarn install, pnpm install, pip install, poetry install, uv sync, bundle install, go mod download, cargo fetch, apt-get install |
| Setup | Set up job, checkout, setup-node/python/go/java, toolchain, docker login, initialize containers |
| Test | test, spec, jest, vitest, pytest, rspec, playwright, cypress, e2e, nextest, miri |
| Build | build, compile, tsc, webpack, vite, make, docker build, gradle, mvn, bazel |
| Other | Anything else: lint, upload, deploy, post steps |
Rules and thresholds
Rules run per job. T is the job's total: the sum of its steps, queue included.
| Fix | Fires when |
|---|---|
| Undersized runner | build + test ≥ 5 min and ≥ 60% of the job |
| Serial tests | test ≥ 3 min, ≥ 40% of the job, one step ≥ 80% of test time |
| Cold caches | cache steps ≥ 30s or ≥ 10% of the job |
| Slow installs | install ≥ 60s or ≥ 15% of the job |
| Queue wait | queued ≥ 60s or ≥ 20% of the job |
| Slow setup | setup ≥ 60s or ≥ 15% of the job |
Undersized runner, cold caches and serial tests are runner-bound cases: the analyzer lists runner options beside them, judged on the criteria on the runners page.
Ranking
time_at_stake(fix) = Σ over flagged jobs of time in the fix's phasesorder = time_at_stake, largest firstTime at stake is time already spent, not a saving. Undersized runner counts build and test, so it can overlap serial tests; both stay listed.
Baseline comparison
match jobs by name (case-insensitive, spaces collapsed)difference = candidate_total − baseline_totalpercent = difference ÷ baseline_totalJobs on one side only are listed and left out of totals. The baseline name you type is printed above every number.
Assumptions and limits
- Results describe the run you pasted. Another commit, cache state or time of day can differ.
- Step names decide phases. A step named “build-and-test” counts as test; rename it or split it for a sharper read.
- Totals add job times. Jobs that run in parallel overlap in wall-clock time, so the total is compute time, not wait time.
- Text pasted from the job page has no queue time unless you add a “Queued” line.
- Thresholds are rules of thumb chosen to flag large shares, not industry standards.
- No fix here removes or skips tests. Each fix page lists how to verify the verdict stays the same.