fix(IW-517): the GitLab CI gate reads every pipeline at a commit #559

mph · 2026-09-14T12:45:41Z

mph commented

2026-09-14 12:45:41 +00:00

Summary

The GitLab CI gate judged a commit on one pipeline and could report Green
while another pipeline on the same commit was still running. iw phase-pr,
iw phase-merge and iw ci-gate all share that gate, so on GitLab the
"CI is green, safe to merge" contract could be false.

GitLabClient.fetchCiSnapshot built its CiSnapshot from the merge request's
head_pipeline, which names at most one pipeline. GitLab routinely starts
several pipelines for one push (a push pipeline and a merge_request_event
pipeline). Every pipeline but one was invisible to the gate, whatever its state.

GitHub and Forgejo enumerate runs and were never affected.

Defects fixed

Phase 1 — the GitLab snapshot holds at most one pipeline (PR #554, merged)

GitLabClient now reads GET projects/:id/merge_requests/:iid/pipelines?per_page=100
and maps every entry to a CiRun. head_pipeline is no longer read.
A parse failure on any entry returns Left rather than defaulting a field.
A full 100-entry page sets incomplete, which the model already refuses to judge.
A GitLab CiRun carries its pipeline id as workflowLabel, so two pipelines at
one commit survive dedupe as two separate answers.
core/model/CiGate.scala is unchanged by design — the model was already correct;
it could not judge a pipeline it was never handed.
docs/plans/2026-08-11-ci-green-gate.md §G2 records the decision this reverses.

Behaviour changes a GitLab user meets

A stale canceled sibling pipeline at the same sha now reads as Red and stops
the phase. This is the safe direction.
A merge request with exactly 100 pipelines is refused rather than paginated.

Testing

5 regression tests added in core/test/GitLabCiTest.scala; the 4 pre-existing
fetchCiSnapshot tests were repaired for the new endpoint, not weakened. A
reviewer confirmed this by mutation.
./mill core.test.testOnly 'iw.tests.GitLabCiTest' — exit 0, 187 tests.
./iw ./test unit — exit 0. ./iw ./test lint — exit 0.
./iw ./test — exit 1 on test/start.bats "the tmux session does not carry
iw's internal plumbing". That failure reproduces at the phase baseline in a
clean worktree, asserts on tmux pane environment, and has no causal path to
this change. Filed as #556.

Follow-ups filed

#556 — test/start.bats tmux-plumbing test fails at baseline on this machine
#557 — GitLabClient.parseMrPipeline has no production caller; removal needs approval
#558 — GitLabClient.encodedRepo under-escapes repository slugs

Release notes

project-management/issues/IW-517/release-notes.md (Czech) describes the change
for users of phase-pr, phase-merge and ci-gate.

🤖 Generated with Claude Code

## Summary The GitLab CI gate judged a commit on one pipeline and could report `Green` while another pipeline on the same commit was still running. `iw phase-pr`, `iw phase-merge` and `iw ci-gate` all share that gate, so on GitLab the "CI is green, safe to merge" contract could be false. `GitLabClient.fetchCiSnapshot` built its `CiSnapshot` from the merge request's `head_pipeline`, which names at most one pipeline. GitLab routinely starts several pipelines for one push (a `push` pipeline and a `merge_request_event` pipeline). Every pipeline but one was invisible to the gate, whatever its state. GitHub and Forgejo enumerate runs and were never affected. ## Defects fixed **Phase 1 — the GitLab snapshot holds at most one pipeline** (PR #554, merged) - `GitLabClient` now reads `GET projects/:id/merge_requests/:iid/pipelines?per_page=100` and maps every entry to a `CiRun`. `head_pipeline` is no longer read. - A parse failure on any entry returns `Left` rather than defaulting a field. - A full 100-entry page sets `incomplete`, which the model already refuses to judge. - A GitLab `CiRun` carries its pipeline id as `workflowLabel`, so two pipelines at one commit survive `dedupe` as two separate answers. - `core/model/CiGate.scala` is unchanged by design — the model was already correct; it could not judge a pipeline it was never handed. - `docs/plans/2026-08-11-ci-green-gate.md` §G2 records the decision this reverses. ## Behaviour changes a GitLab user meets - A stale `canceled` sibling pipeline at the same sha now reads as Red and stops the phase. This is the safe direction. - A merge request with exactly 100 pipelines is refused rather than paginated. ## Testing - 5 regression tests added in `core/test/GitLabCiTest.scala`; the 4 pre-existing `fetchCiSnapshot` tests were repaired for the new endpoint, not weakened. A reviewer confirmed this by mutation. - `./mill core.test.testOnly 'iw.tests.GitLabCiTest'` — exit 0, 187 tests. - `./iw ./test unit` — exit 0. `./iw ./test lint` — exit 0. - `./iw ./test` — exit 1 on `test/start.bats` "the tmux session does not carry iw's internal plumbing". That failure reproduces at the phase baseline in a clean worktree, asserts on tmux pane environment, and has no causal path to this change. Filed as #556. ## Follow-ups filed - #556 — `test/start.bats` tmux-plumbing test fails at baseline on this machine - #557 — `GitLabClient.parseMrPipeline` has no production caller; removal needs approval - #558 — `GitLabClient.encodedRepo` under-escapes repository slugs ## Release notes `project-management/issues/IW-517/release-notes.md` (Czech) describes the change for users of `phase-pr`, `phase-merge` and `ci-gate`. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

mph added 15 commits

2026-09-14 12:45:41 +00:00

workflow(dx): Create diagnostic analysis for IW-517 6c10e40f70

phase-pr reports CI green on GitLab because it queries pipelines by branch ref, never seeing merge_request_event pipelines

Artifacts:
- analysis.md

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0179JhvXg4eopsQNmDLUSuim

chore(IW-517): update review-state to analysis_ready 7d9e608398

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0179JhvXg4eopsQNmDLUSuim

workflow(dx): Create task index for IW-517 5e46e7f58a

phase-pr reports CI green on GitLab because it queries pipelines by branch ref, never seeing merge_request_event pipelines

Artifacts:
- tasks.md
- review-state.json (tasks_ready)

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0179JhvXg4eopsQNmDLUSuim

workflow(dx): Replace the IW-517 analysis with the reproduced one 42baee0593

The delegated analysis pass returned after the first draft was committed. It
reproduces the defect against the real GitLabClient.fetchCiSnapshot and the
real CiGate.verdict, and it measures a second cause the draft only guessed at:
dedupe groups GitLab runs by (workflowLabel, event), and with workflowLabel a
constant and event the pipeline source, two same-source pipelines at one commit
collapse to the newest. Reading the full pipeline list is therefore necessary
but not sufficient.

Estimate narrows to 4-7h; tasks.md follows.

Artifacts:
- analysis.md
- tasks.md

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0179JhvXg4eopsQNmDLUSuim

docs(IW-517): intake verdict — agent-ready a364621c20

Independent review re-ran the pasted evidence and returned sound with no
blocking finding. Seven notes recorded, two marked: the Testing Strategy
understates the work D7 creates, and a stale canceled sibling pipeline
becomes Red for GitLab users.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0179JhvXg4eopsQNmDLUSuim

Phase 1: context ready 88ec448ca5

Issue: IW-517
Phase: 1 - The GitLab snapshot holds at most one pipeline, so a second pipeline on the same commit is invisible

This commit marks the checkpoint for phase 1 context generation.
Use dx-fix IW-517 to return to this point.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

chore(IW-517): update review-state to context_ready 6fe981f48d

workflow(dx): Generate Phase 1 tasks for IW-517 cb473bbe66

The GitLab snapshot holds at most one pipeline, so a second pipeline on the same commit is invisible

Artifacts:
- phase-01-tasks.md

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

chore(IW-517): update review-state to tasks_ready 902e61a28e

Phase 1: The GitLab CI gate reads every pipeline at a commit (#554 ) b08b4d4114

chore(IW-517): update review-state for phase 01 891aa8dda0

chore(IW-517): mark phase 01 complete in tasks.md 28303fca8e

docs(IW-517): add release notes e1e5db34d3

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

chore(IW-517): update review-state to phase_merged d8589d2db1

workflow: Triage lessons for IW-517

CI / changes (pull_request) Successful in 2s

Details

CI / contract (pull_request) Has been skipped

Details

CI / format (pull_request) Successful in 8s

Details

CI / lint (pull_request) Successful in 8s

Details

CI / test (pull_request) Has been cancelled

Details

CI / dashboard-build (pull_request) Has been cancelled

Details

CI / compile (pull_request) Has been cancelled

Details

5cfd45fed0

Follow-ups filed: 3 (#556, #557, #558), commented: 0, struck: 0
Lessons: 0 approved, 0 rejected, 0 still pending (no LESSONS.md captured)

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011EmLSy5eWo2imTeFHvySMc

mph added 2 commits

2026-09-14 12:46:59 +00:00

chore(IW-517): update review-state to all_complete e9b69aee31

chore(IW-517): update review-state

CI / changes (pull_request) Successful in 2s

Details

CI / contract (pull_request) Has been skipped

Details

CI / format (pull_request) Successful in 7s

Details

CI / lint (pull_request) Successful in 8s

Details