[aviadr1]

Hi guys, I'm Aviad, maintainer of CI-esta and pytest-selective packages, and CTO of tapin.io
these packages are useful for projects where CI time is annoying. this is an outreach offer that will decrease CI times.
these tools are used by us everyday saving thousands of dollars each month in CI bills, so this isnt slop.
I appreciate humanize is a small library that isnt plagued much by CI times, but would be honored if you accept this PR.

if it helps, I have a fork of this repo where the PR is merged: https://github.com/gigaverse-app/humanize-pytest-selective/

Best,
Aviad

TLDR

Humanize runs its full compatibility matrix for documentation and tooling-only changes. This PR skips unaffected complete jobs with pytest-selective and adds CI-esta reports so the decisions and runner usage are visible. The existing 27-job Python/OS matrix, tox commands, coverage and Test successful check are preserved.

I benchmarked previous Humanize PRs before proposing this. Matched replays of a docs dependency update and a mypy update used 94.1% and 93.1% fewer core-workflow runner-seconds. A separate historical model estimates 26.7% lower usage across 19 PRs with five successful workflows, including planning and report overhead. Measured replays and estimates are separated below.

Existing action SHAs are unchanged. The new CI-esta and pytest-selective actions are pinned to the exact revisions used in the measured runs. No application runtime dependency is added.

Changes proposed in this pull request:

• Select complete test, benchmark, package and routing jobs from their declared inputs. Keep lint and docs checks.
• Retain full manual and weekly checkpoints, plus a test:full label override.
• Reject failed planning, invalid decisions, failed jobs and unexpected skips at Test successful.
• Add a read-only CI-esta observer with summaries and downloadable artifacts. It runs independently after Test, does not execute source PR code, and posts no comments.
• Run benchmark tests in trial forks without uploading to the upstream CodSpeed account; preserve the upstream CodSpeed command and action SHA.

BEFORE
PR -> full Python/OS matrix + benchmark + package + lint + docs

AFTER
PR -> input plan -> affected complete jobs -> Test successful
                -> unaffected jobs skipped explicitly
                -> lint + docs retained
completed Test -> independent CI-esta report + artifacts

Measurements on real Humanize PRs

Each pair ran on an identical replay revision, with unchanged historical application code and the PR's original action SHAs. All five core workflows passed: Test, Benchmark, Release, Lint and Docs. All accepted runs are attempt 1. Lower runner usage is better.

Historical PR	Full core runner-seconds	Selected core runner-seconds	Reduction	Jobs full → selected	Separate report seconds full / selected
#494: docs dependencies	1613	95	94.1%	36 → 7	15 / 10
#495: mypy update	1633	113	93.1%	36 → 8	11 / 11
#501: Turkish translation	1670	1585	5.1%	36 → 35	13 / 12

Exact replay revisions and elapsed times

PR	Same replay head for both arms	Test creation to successful gate, full → selected
#494	f94abc51d2c79afa864414a3d62359ef8f8d97b2	304s → 19s
#495	e09a48c359d68ed4f8f2189b105c41176454ca6e	293s → 39s
#501	b9064f43c9b978202ad7285d7b1bc35ef3ea168e	485s → 394s

Historical workload estimate

The sample contains 20 recent merged PRs from 16 September through 2 November 2026, selected by merge time from the 100 most recently updated closed PRs. Seven would skip the runtime matrix. Every sampled source change retains it.

Cohort	Original observed runner-seconds	Estimated candidate including reports	Estimated reduction
19 PRs with five successful workflows	31,603	23,165.5	26.7%
All 20, including #435's failed benchmark	33,130	23,254	29.8%

Per-PR historical data and source runs

PR	Retained lanes	Original seconds	Estimated candidate seconds	Original Test run
#501	tests, benchmark, package	1892	1927.0	41497253715, attempt 2
#499	tests, benchmark, package, routing	1517	1566.5	41284940161, attempt 1
#494	none	1590	110.5	41257820554, attempt 2
#495	routing	1572	116.0	41257830275, attempt 2
#483	benchmark, routing	1620	143.0	40197931159, attempt 1
#482	tests, benchmark, routing	1607	1645.0	40197000419, attempt 1
#435	none	1527	88.5	39084699946, attempt 2
#410	package	1585	113.0	35782518080, attempt 2
#465	tests, benchmark, package	1784	1819.0	39643285767, attempt 2
#461	tests, benchmark, package	1774	1809.0	39417822505, attempt 2
#464	tests, benchmark, package	2084	2119.0	39471647729, attempt 2
#471	tests, benchmark, package	1891	1926.0	39771719558, attempt 2
#444	none	1673	121.5	37392712979, attempt 1
#440	tests, benchmark, package, routing	1578	1627.5	37146282621, attempt 1
#422	none	1517	86.5	36814887274, attempt 1
#421	tests, routing	1504	1481.0	36814877716, attempt 1
#295	tests, benchmark, package	1552	1587.0	28497194164, attempt 2
#346	tests, benchmark, package	1589	1624.0	28808853031, attempt 1
#402	tests, benchmark, package	1656	1691.0	35126593363, attempt 1
#381	tests, benchmark, package	1618	1653.0	35121695374, attempt 2

Checks and maintainer controls

Scenario	Before	After
Unreferenced documentation	Full runtime matrix	Runtime/benchmark/package/routing work skipped; docs and lint retained
Source, translations, tests or execution inputs	Full affected checks	Complete checks inside every affected job
Missing history or uncertain plan	Full suite	Full suite fallback; a failed planner also fails the required gate
Explicit full checkpoint	Manual full suite	Manual + weekly full suite; test:full applies on the next push/reopen
Required status	Test successful	Same name, validates explicit boolean decisions and prerequisite outcomes
Existing action refs	SHA pins	Identical SHA pins
Publishing	Upstream-only publishing	Same conditions; fork benchmarks execute without upstream upload

About the tooling

Disclosure: this contribution comes from the Gigaverse team, which maintains CI-esta and pytest-selective, and was prepared with Codex. CI-esta is open source under Apache-2.0 (license).

The underlying tooling has been used in Gigaverse's production CI for about five months across a Django monorepo and a Python microservices repository. It was extracted into the public packages more recently. The measured production history and approach are documented in Faster, cheaper CI: from the trenches.

The new refs are fixed to CI-esta 1.14.0 (b694847) and pytest-selective 1.4.0 (286d0ae), the revisions resolved by the benchmark runs. The routing-test dependency is also pinned to pytest-selective 1.4.0.