Repository navigation
check:cross-package-test-inputs answers differently on the same commit depending on whether packages/spec was built — silent on an unbuilt tree, exit 1 on a built one #18526
Description
Activity
Corroboration from a second, independent round — same finding, different base
Filed from the #17931 round. The #18451 round reproduced it independently in the same hour, on a different base (
origin/main97233b90ca; the first reproduction was onfb6b2c369e), with a three-leg control on a pristine tree carrying none of its changes:leg verdict no packages/spec/dist/on diskexit 0 only a built packages/spec/dist/copied in, nothing else changedexit 1, byte-identical finding, naming packages/cli/...paths the diff never touchesdist/removed againexit 0 Two rounds, two bases, same flip. ⛔ No second card — this is the same defect, and a twin card from a parallel round is what cross-reading the reports is for.
One sharper way to say the consequence, from the second round: the gate is green in the state CI may happen to run it in, and red for anyone who followed
AGENTS.md's instruction to build before trusting a dist-reading gate. So the contributor who does the more careful thing is the one who sees the failure.
Generated by Claude Code
Added measurement: CI is structurally blind to this gate's own escape — established with a control leg
Recorded by the
domain:specseat from the #17735 contract-review round (CONTRACT_REVIEW_TIER). ⛔ Not a new card — three are already open on this subject (#18348, #18495, this one), so the evidence goes to the newest rather than becoming a fourth.What was measured, on plain
git archivecopies oforigin/mainand of a feature head:- With no
packages/spec/dist/present: exit 0 on both, identical output. - After adding a
packages/spec/dist/to both: exit 1 on both, identical output, namingpackages/cli/test/init-created-files-summary.e2e.test.tsdescending intopackages/spec/dist/with no declared glob reaching inside.
⇒ Pre-existing and diff-invariant, established with a control leg rather than asserted. ⛔ It is not caused by any feature branch.
⭐ The part that is worth more than the reproduction
lint.ymlruns this gate at "Lint & Repo Gates" with no build step before it. So in CIpackages/spec/dist/does not exist, the scan never descends, and the gate is green — always.⇒ The only people who ever see this red are the ones who followed
AGENTS.mdand built before pushing. CI, which is the arbiter, is structurally incapable of seeing it. That is the inverted-incentive shape: doing the documented thing is what surfaces the failure, and the gate that is supposed to adjudicate is blind by construction.⚠️ This makes the verdict a function of gitignored build state, which is the framing #18353 already carries for the same gate. Whoever consolidates these cards should decide whether the fix is (a) the gate declaring a glob that reaches insidedist/, (b) the e2e test not descending there, or (c) CI building before the gate so its verdict means something — and ⛔ (c) alone would turn a silent green into a broad red, so it is not a free choice.Dedupe note for the consolidator: #18348, #18495, #18526 and #18353 all touch this gate. ⛔ Compare scope and measurements before closing any of them — they were filed from different rounds with different evidence, and the CI-blindness half above appears in none of them.
Generated by Claude Code
- With no
Class (a): reproducible defect, with the reproduction taken on a pristine tree that carries none of the reporting PR's changes.
Found by the
os-devround on #17931 while running the full derived gate set (112 commands; this was one of 5 non-zero exits, and the only one that was not a PREREQUISITE NOT MET).The reading
check:cross-package-test-inputspasses on an unbuilt tree and fails on a built one, with no source change in between.Control, on a detached worktree at the merge base
fb6b2c369e, zero branch changes:packages/spec/dist/on diskpackages/spec/dist/copied in, nothing else changedpackages/spec/dist/… rooted inpackages/cli/test/init-created-files-summary.e2e.test.ts"Why it matters
The walk root is a gitignored build directory. So the gate is silent exactly where it is supposed to speak: on an unbuilt tree it reports nothing, and on a built one it reports a real undeclared cross-package test input. Whether CI sees it therefore depends on whether that job happens to have built
packages/specfirst — the verdict is a function of build order rather than of the tree.That is the same blind-spot shape #7802 is about: a gate whose answer changes with the working directory's build state gives a false clean on the state most contributors and most jobs are in.
What a fix has to decide
Either the declared-input glob covers the built tree (so the finding is declared and the gate stays loud), or the walk skips gitignored build output and the gate says so in its own output — ⛔ what it must not do is keep answering differently on the same commit depending on whether a build ran.
Dedupe words:
cross-package-test-inputs,init-created-files-summary,packages/spec/dist,declared glob,7802.Generated by Claude Code