mirror of
https://github.com/stablyai/orca.git
synced 2026-09-28 16:02:45 +00:00
Jinwoo-H
0d5ab3f50e
pr.yml now calls the reusable e2e.yml twice in one workflow run: the advisory `e2e` job and the blocking `orchestration_e2e` job. Both start e2e.yml's `build` job, and both would upload `e2e-build-out`. upload-artifact v4+ fails outright when two jobs in one run upload the same name, so a PR touching orchestration source AND any other routed source would kill whichever build finished second -- taking the new required check red for a reason that has nothing to do with the PR. `playwright-traces-changed` had the same collision on failure in both lanes. Add an optional `artifact_suffix` workflow_call input, default '', and append it to every artifact name e2e.yml uploads or downloads -- not just the two that can collide today, because a bare name left behind would fail only when a spec fails in both lanes, hiding the signal exactly when it matters. The advisory caller passes nothing; schedule and dispatch runs pass no inputs at all, so both keep the names they have always used. `orchestration_e2e` passes `-orchestration`. release-e2e-dispatch-contract's "hands the built relay artifact to every E2E run command" pinned the literal `e2e-build-out` on the upload and both downloads. It now compares the downloads against the upload's own name, which is what the test was really guarding: upload and download must agree. A mutant that renames one download still fails it. Verified: actionlint resolves the caller/callee input contract (a deliberately typo'd input name is reported), and both new assertions were mutation-checked.