Imported from missingbulb/ClaudiniteCanary (.claudinite/shared/packs/basics/skills/ci-performance-evaluation/SKILL.md). Install upstream with npx skills add missingbulb/ClaudiniteCanary --skill ci-performance-evaluation. Copyright stays with the author.

Evaluating CI performance

The order below is the method, and the order is the point: each step tells you whether the next one is worth doing. Most of what makes a suite slow is invisible from the top and obvious two steps down, and most of what looks slow at the top is not where the time is.

1. Take the step breakdown before forming any theory

A job's own step timings are recorded — never guess from the total. List the run's jobs and read each step's started_at/completed_at; the difference per step is the breakdown. Do this for several recent runs, not one, so you know the spread before you call anything a regression.

State the split before going further ("checkout 4s, setup 5s, tests 48s, sweep 2s"). Only the dominant step is worth profiling; time spent on a 2s step is time wasted whatever you find.

When setup and install dominate rather than the tests, stop here: that is a caching finding, not a profiling one. Give the setup action its cache: input where a lockfile exists, warm the cache from a run that may write the default-branch scope, and read the scoping rules in git-github-advanced before calling a cache that never hits a key problem. None of steps 2–8 can recover time spent before the tests start. (1)

2. Reproduce the dominant step locally — and check the local run is comparable

Profile locally, but establish the gap first: a local run and a runner differ in core count, in warm caches, and — the one that has actually bitten here — in git commit signing. A session with signing configured pays a round trip per fixture commit that a runner never pays, so a local profile can be dominated by a cost CI does not have. Compare a local total against the CI step before trusting local numbers, and when they disagree, find out why before optimizing.

3. Find the long pole: time each test file serially

Time each file on its own, sum the results, and sort. That gives two numbers that matter: the serial total (the work) and the slowest single file (the floor no amount of parallelism gets under). Compare the serial total against the wall clock to see how much parallelism you are already getting.

4. Read the distribution of per-test durations, not just the top of the list

Get every test's duration and bucket them. The shape tells you which kind of problem you have:

• A few tests far above the rest — specific slow cases. Read them; a test that waits seconds is usually waiting on something (see step 5).
• A broad band of similar durations — a fixed per-test cost, almost always in a shared fixture helper, and by far the more valuable find because it multiplies by every test in the band. Measure the helper directly rather than any one test.

5. Hunt the out-of-process waits

Time a suite spends waiting is not time it spends working, and it is the cheapest to remove. Look for: subprocess spawns per test; sleeps and retry backoffs with hardcoded durations; calls to a signing service, a network, a real git remote; anything with a timeout.

A test that waits is often reporting a bug, not just being slow. A test that sets a 0.3s timeout and takes 10s is telling you the timeout does not work — the mechanism under test is failing to kill what it claims to kill, and the 10s is the real lifetime leaking through. Chase that reading before optimizing the test away: the same discipline found a code_work_timeout that had never killed anything. Make a retry backoff injectable so tests scale the real timings down rather than skip the path.

6. Check the parallelism headroom before adding parallelism

Compare wall clock against user+sys CPU time. Well under the core count means the suite is waiting, so more workers help and oversubscribing beyond core count is right. At or near the core count it is CPU-bound, and more workers buy nothing — including intra-file concurrency on the longest file, which is the tempting move precisely when it is useless.

7. A/B every change against the whole suite, and revert what does not show

A change that halves one file routinely shows zero at suite level, because the cores were already busy with other files. The file-level number is not the result; the suite-level number is. Run both arms at least twice — a single pair inside normal variance is not a measurement. Complexity that buys nothing measurable does not ship, and saying so is a finding, not a failure.

8. Verify in real CI, and report the two numbers separately

Local and CI improvements are different claims. Land the change, watch a real run, and give both — "local 106s → 71s, CI 53s → 48s" — rather than letting the flattering one stand for both. When they differ, explain the mechanism; a gap you cannot explain means the profile was measuring something CI does not do.

Reporting

Say where the time goes as a breakdown that adds up, name what you changed and what it bought, and name what you deliberately left alone and why. A finding that an area is already fast is a result worth stating — it stops the next person re-profiling it.