feat(ci): scitex-dev ci why — the ecosystem CI-failure-reading primitive (joins ci runner) - #345
feat(ci): scitex-dev ci why — the ecosystem CI-failure-reading primitive (joins ci runner)#345ywatanabe1989 wants to merge 2 commits into
Conversation
…re-reading primitive
Reading CI status is one word (`failure`); reading why has been tens of
thousands of lines. Add the reusable SSOT primitive that inverts that
price — it fetches a failing run's log once and distils it to a few
hundred bytes (failing test ids / assertion lines / a setup `##[error]`)
— plus the thin CLI verb any consumer's own verb can mirror.
Primitive scitex_dev.ci.why (importable by any project):
explain_ci_run(target, *, run_gh=None, repo=None) -> list[RunFailures]
+ parser helpers clean_log_line / split_log_by_job / parse_job_context
/ parse_failed_log (four-tier: pytest summary -> FAILURES E-lines ->
##[error] annotations -> tail fallback)
+ the injectable run_gh gh-seam and CIWhyError (UNKNOWN never reads as
green). Everything but the gh-seam is pure/string-based.
CLI: scitex-dev ecosystem ci why <target> (a new ci subgroup under
ecosystem) resolves a PR# / run id / branch / nothing to the failing
run(s); --json and -R/--repo; tri-state exit (0 green / 1 red / 2
could-not-read).
Lifted generically from sac's shipped _ci_why.py (no sac specifics).
Tests carry no mocks: the parser runs off real-shaped GitHub-Actions log
strings, the resolver through an injected run_gh callable, and the CLI
end-to-end against a real fake gh on PATH.
…ci runner)
Per operator ruling the CLI verb lives on the EXISTING top-level `ci`
group (which already holds `ci runner`), not under `ecosystem`. scitex-dev
drops the redundant middle `dev` because "dev" is already its package
name — consumers use `<cli> dev ci why`, but scitex-dev itself is just
`scitex-dev ci why`.
The library primitive scitex_dev.ci.why is UNCHANGED. Only the verb
wiring moved:
- new scitex_dev/ci/_why_cli.py::register_ci_why_command attaches `why`
to the ci group in _root.py (which now captures the group returned by
register_ci_runner_commands, then registers `why` beside `runner`);
- the `ecosystem ci` subgroup + its Discovery category entry are
reverted;
- the ci group docstring lists the new `why` verb.
Tests updated: invoke `ci why` (not `ecosystem ci why`) and assert
`ci runner` + `ci why` coexist. Everything else (four-tier parser, run_gh
seam, CIWhyError, --json, tri-state exit, no-mocks) unchanged.
ywatanabe1989
changed the title
|
Reviewed — sound, merge once develop is green. (GitHub won't let me formally approve: the whole fleet pushes under one account, so it reads as my own PR. Recording the review here instead.) WHAT I VERIFIED rather than took from the description:
Those two are the difference between a CI-reading tool and a CI-reassurance tool, and this gets them right. WHY IT MATTERS MORE THAN WHEN IT WAS WRITTEN: today I merged a PR that broke develop, then merged four more on top of the red for an hour without noticing — because I had convinced myself the runners were dead and stopped reading the gate at all. This primitive is the direct countermeasure: it makes the reason as cheap as the status. It also implicitly proved itself during that incident, since the only reason I could bisect the break was that the log named the failing assertion. Blocked only on develop going green, which is my own breakage (#355 §1/§4 violations, #359 orphaned tests, #354 an ungenerated import gate). Three fix PRs in flight; this one is unrelated and not a cause. Sorry it sat open — that's on me, not on the work, which was ready and had already been reworked twice to my naming rulings. |
|
Closing under the 3-day PR rule (operator: 「PRの賞味期限は３日にしましょう」「３日より古いものは強制close」「腐ってる、新しく作った方が早い」). Opened 2026-07-16, now 3 days old and DIRTY — develop has moved substantially underneath it (audit-violation fix, versioning fastpath, PS-142 truncation fix, hosted-runner migration all landed since). NOT ABANDONED — the intent is recorded and will be rebuilt fresh off current develop: Rebuilding costs less than resolving a 3-day-old conflict set, and the result is based on code that actually exists now. That is the whole point of the rule. |
1 participant
What
The ecosystem CI-failure-reading primitive + its thin CLI verb
scitex-dev ci why.Reading CI status is one word (
failure); reading why has been tens of thousands of lines, so a bounded-context agent gets steered to the cheap word and the word replaces the reason. This inverts the price: fetch a failing run's--log-failedonce, distil it to a few hundred bytes.The primitive (SSOT) —
scitex_dev.ci.whyImportable by ANY project (sac, the umbrella, …), which supply only a thin verb over it:
explain_ci_run(target, *, run_gh=None, repo=None) -> list[RunFailures]— a PR can front >1 failing run, hence a list.explain_run(run_id, ...) -> RunFailures,resolve_run_ids(target, ...) -> list[str]clean_log_line,split_log_by_job,parse_job_context,parse_failed_log— four tiers: pytestshort test summaryFAILED ids ->FAILURESblockElines ->##[error]annotations -> tail fallback.run_ghgh-seam (the only network touch) +CIWhyError(UNKNOWN is raised loudly, never swallowed into a reassuring "no failures").The CLI verb
scitex-dev ci why <target>— thewhyverb attached to the existing top-levelcigroup, alongsideci runner(both coexist; runner is untouched).--json,-R/--repo. Tri-state exit: 0 green / 1 red (reason printed) / 2 could-not-read.Watch it fail (real live runs)
scitex-dev run
29312551895(workflowtests) — pytest tier, all 3 matrix legs:sac run
29446283736(workflowquality, a setup failure with no pytest) —##[error]tier:Both extract the reason, not the whole log.
Tests — no mocks
##[error]setup-failure log).explain_*through an injectedrun_ghcallable (offline).ghon PATH (a real subprocess; PATH restored, no monkeypatch — PA-306). A coexistence test assertsci runnerandci whyboth resolve.Full slice green (156 passed — the new primitive/CLI tests plus the pre-existing
ci/runnersuite and the ecosystem registry/categories regression).Notes
_ci_why.py— no sac specifics..github/workflowsreusable migration), fix(ci-template): isolate per-job git global config to stop ~/.gitconfig lock collisions #314 (ci_template templates) — zero file overlap; this only READS logs.audit-cliflags only PRE-EXISTING free-form-help commands (ci runner,linter,ecosystem dashboard); the newci whyis spec-built and passes.