Investigate CI Failure

Given a PR URL (e.g. https://github.com/openshift/lightspeed-service/pull/2825), diagnose why CI jobs failed.

Workflow

1. Extract PR info

Parse org, repo, and PR number from the URL. Fetch metadata with gh:

# PR metadata
gh api repos/{org}/{repo}/pulls/{pr} --jq '{title, state, user: .user.login, head_sha: .head.sha}'

# Changed files
gh api repos/{org}/{repo}/pulls/{pr}/files --jq '.[].filename'

2. Get check statuses

# All checks at a glance
gh pr checks {pr} --repo {org}/{repo}

# Detailed statuses with Prow URLs (use head SHA from step 1)
gh api repos/{org}/{repo}/statuses/{head_sha} \
  --jq '.[] | select(.state == "failure" or .state == "error") | {context, state, target_url}'

This gives you the list of failed jobs and their Prow dashboard URLs.

3. Construct GCS artifact URLs

From a Prow target_url like:

https://prow.ci.openshift.org/view/gs/test-platform-results/pr-logs/pull/{org}_{repo}/{pr}/{job_name}/{build_id}

Derive:

• Directory browser (for navigating artifact tree):

https://gcsweb-ci.apps.ci.l2s4.p1.openshiftapps.com/gcs/test-platform-results/pr-logs/pull/{org}{repo}/{pr}/{jobname}/{build_id}/

• Raw file content (for fetching logs and JSON):

https://storage.googleapis.com/test-platform-results/pr-logs/pull/{org}{repo}/{pr}/{jobname}/{build_id}/{path}

4. Triage the failure

For each failed job, fetch artifacts in this order:

4a. Quick status

GET storage.googleapis.com/.../finished.json

Check "passed": false and "result": "FAILURE".

4b. Build log (most useful)

GET storage.googleapis.com/.../build-log.txt

This is the main ci-operator build log. It can be large (200KB+). Search from the end for:

• failed / FAILED / error / ERROR
• step .* failed
• Python tracebacks (Traceback, AssertionError, FAILED tests/)
• Container crash indicators (CrashLoopBackOff, OOMKilled, Error from server)

4c. Artifact tree exploration

The build log alone often doesn't tell the full story. Browse the GCS artifact directory to find step-specific logs, cluster state, and pod logs:

GET gcsweb-ci.apps.ci.l2s4.p1.openshiftapps.com/gcs/.../artifacts/

Full artifact tree for an e2e job:

{build_id}/
├── build-log.txt                    ← main ci-operator log (start here)
├── finished.json                    ← pass/fail + metadata
├── artifacts/
│   ├── ci-operator.log              ← detailed ci-operator log
│   ├── junit_operator.xml           ← top-level JUnit results
│   ├── ci-operator-step-graph.json  ← step dependency graph
│   ├── ci-operator-metrics.json
│   ├── metadata.json
│   ├── build-logs/                  ← container image build logs
│   │   ├── lightspeed-service-api-amd64.log
│   │   ├── root-amd64.log
│   │   └── src-amd64.log
│   ├── build-resources/             ← CI namespace state
│   │   ├── pods.json                ← all pods in CI namespace
│   │   ├── events.json              ← k8s events (useful for crashes)
│   │   ├── builds.json
│   │   ├── imagestreams.json
│   │   └── clusterClaim.json
│   ├── release/                     ← cluster provisioning step
│   │   ├── build-log.txt
│   │   └── finished.json
│   └── e2e-ols-cluster/             ← test workflow steps
│       ├── ipi-install-rbac/        ← cluster RBAC setup
│       │   └── build-log.txt
│       ├── e2e/                     ← THE ACTUAL TEST STEP
│       │   ├── build-log.txt        ← test runner output (pytest)
│       │   ├── finished.json
│       │   └── artifacts/           ← per-provider test results
│       │       ├── junit_e2e_azure_openai.xml
│       │       ├── junit_e2e_openai.xml
│       │       ├── junit_e2e_watsonx.xml
│       │       ├── junit_e2e_rhelai_vllm.xml
│       │       ├── junit_e2e_rhoai_vllm.xml
│       │       ├── junit_e2e_*_tool_calling.xml
│       │       ├── junit_e2e_quota_limits.xml
│       │       └── {provider}/cluster/   ← cluster state per provider
│       │           ├── podlogs/
│       │           │   ├── lightspeed-app-server-*.log  ← OLS service logs
│       │           │   ├── lightspeed-postgres-server-*.log
│       │           │   └── lightspeed-console-plugin-*.log
│       │           ├── olsconfig.yaml    ← OLS config used
│       │           ├── pods.yaml
│       │           ├── deployments.yaml
│       │           ├── configmap.yaml
│       │           ├── services.yaml
│       │           └── routes.yaml
│       ├── gather-must-gather/      ← cluster diagnostics
│       │   └── artifacts/
│       │       ├── must-gather.tar  ← full must-gather (large, ~25MB)
│       │       ├── camgi.html       ← must-gather analysis report
│       │       └── event-filter.html
│       └── openshift-configure-cincinnati/

Where to look by failure type:

4d. Downloading artifacts locally

When you need to search across many files or the artifacts are too large for WebFetch, download them to a temp directory using gsutil or gcloud storage:

TMPDIR=$(mktemp -d)
# Download a specific subdirectory
gcloud storage cp -r \
  gs://test-platform-results/pr-logs/pull/{org}_{repo}/{pr}/{job_name}/{build_id}/artifacts/e2e-ols-cluster/e2e/artifacts/ \
  "$TMPDIR/"

The GCS bucket path mirrors the Prow URL: strip https://prow.ci.openshift.org/view/gs/ and prepend gs://.

When multiple jobs have failed, investigate each in a separate subagent (Task tool) to keep build-log context isolated and run fetches in parallel.

5. Cross-reference with PR changes

Compare the failure with the files changed in the PR. Common patterns:

Check if the same job fails on main branch (flaky test) by looking at job history:

https://prow.ci.openshift.org/job-history/gs/test-platform-results/pr-logs/directory/{job_name}

6. Report findings

Summarize:

1. Which jobs failed and which passed
2. Root cause for each failure (with relevant log excerpts)
3. Whether it's PR-related or infrastructure/flaky
4. Suggested fix if the failure is caused by the PR changes

Known CI jobs for this repo

Tool usage notes

• Use gh CLI for all GitHub API calls (PR metadata, statuses, checks, comments, files).
• Use WebFetch to browse GCS directories (gcsweb-ci.apps.ci.l2s4.p1.openshiftapps.com/gcs/...).
• Use WebFetch to fetch raw log/JSON content (storage.googleapis.com/test-platform-results/...).
• The Prow dashboard URL itself is JS-rendered and not useful via WebFetch — always use GCS URLs instead.
• Build logs can be very large. When fetched via WebFetch, they're saved to a temp file — read from the end to find failures quickly.