Investigate a failing CI check¶
The healthy parent passes the regression test. The one-line broken demo commit fails in a dedicated workflow run. See evidence provenance for excerpt boundaries. The installed, healthy ribbit investigates the broken source; it does not execute or patch that source.
Install ribbit and configure a stronger object-capable profile using model setup. npm installation supplies the executable, not this repository's example files. Obtain the saved assets from the demo branch and record the checkout SHA for reproducibility. The broken branch is only an evidence source; all ribbit commands below resolve the installed executable on PATH:
git clone --branch demo/failing-ci-projection https://github.com/funsaized/ribbit.git ribbit-evidence
cd ribbit-evidence
git checkout fa32d61a23849b9419c2e29292dc1ad72ac4ea10
git rev-parse HEAD
From that checkout run:
mkdir -p ci-investigation/commands
cp -R fixtures/tutorials/failing-ci/. ci-investigation/
cp examples/commands/diagnose-ci.yaml ci-investigation/commands/
cp examples/flows/diagnose-ci.yaml ci-investigation/diagnose-ci.yaml
cd ci-investigation
ribbit read ci.log changes.diff src/builtins/exact.ts tests/release/projection-baseline.test.fixture --output records > evidence.records
Expected output: evidence.records starts with the v1 record header and contains four records with source paths. Check the exit status before accepting the file; a valid prefix can precede failure. Inspect the captured ci.log for (fail) projection keeps annotations after classify and the missing annotations path; compare the diff's yield { id: r.id, value } with the parent's yield { ...r, value }. The source excerpt labels original line 113.
The command definition refers to diagnose.schema.json, already copied into the working directory. Interpret the saved evidence first:
Required properties: the JSON has failureSignature, observations, hypotheses, sources, likelyChange, and nextChecks; cited paths and lines must exist in the saved evidence. This is a schema/format check, not a guarantee that the diagnosis is correct. Model prose varies; no captured live model answer is supplied. Examine the proposed change location yourself. The fixture's known failure is annotation/source loss on projection, not an arbitrary nonzero exit.
Once that pipeline makes sense, run the equivalent named flow from the same directory:
ribbit flow plan diagnose-ci.yaml
ribbit flow run diagnose-ci.yaml --output json > flow-diagnosis.json
Expected output: the plan shows one exact read and one model diagnose-ci stage on stronger. Required properties: successful exit, same six JSON fields, source-backed references. The flow shares a request budget; separate shell processes do not. stdout is data, stderr is diagnostics. Keep evidence.records, the captured diff/log, and the output for review. To reproduce the failure in the deliberately broken checkout, use its development test command bun test tests/release/projection-baseline.test.ts. All investigation commands use the installed healthy ribbit from PATH. The .fixture suffix only prevents the offline evidence copy of the test from running in routine tests; no patch is applied automatically.