ci-diagnostics

name	ci-diagnostics
description	Diagnose Proton CI failures and performance comparison results from GitHub checks and uploaded reports. Make sure to use this skill whenever CI checks are mentioned, a PR has red or failing checks, the user pastes a CI URL, or asks about test failures in the pipeline, even if they just ask 'why is CI failing'.

CI Diagnostics

Inputs

• PR number or URL
• optional check name substring
• optional direct report URL

First step: collect check status

For a PR:

gh pr view "$PR" --json title,body,url
gh pr checks "$PR"
REPO=$(gh repo view --json nameWithOwner --jq .nameWithOwner)
SHA=$(gh pr view "$PR" --json commits --jq '.commits[-1].oid')
gh api "repos/$REPO/commits/$SHA/status"

Use the commit status payload to find:

• failing or pending contexts
• target_url links for uploaded HTML reports, raw logs, or performance artifacts

Proton report layout

CI uploads reports under:

<pr-number>/<commit-sha>/<normalized-check-name>...

The normalization logic is defined in:

• tests/ci/upload_result_helper.py
• tests/ci/performance_comparison_check.py

Normalize a check name with lowercase and replacements for spaces, (, ), and ,.

Failure triage workflow

1. List failing contexts from gh pr checks or commit statuses.
2. Open each target_url report first.
3. If the report is sparse, inspect the linked raw log.
4. For test reports, summarize:
   • failing test names
   • first common error signature
   • whether the failure looks deterministic, flaky, infra, or environment-specific
5. Map failures back to touched areas in the diff.

Performance comparison workflow

Performance comparison artifacts upload:

• report.html
• all-queries.html
• all-query-metrics.tsv
• queries.rep
• images/flamegraphs

If you have a report.html URL, inspect sibling artifacts by replacing the filename in the same prefix.

When reviewing perf results:

• start from the summary in report.html
• inspect all-query-metrics.tsv for the biggest client_time regressions
• distinguish broad regressions from a few outlier queries
• correlate with touched execution paths, joins, windows, aggregations, or storage reads

Output expectations

Always report:

1. failing checks
2. best report/log URL for each failing check
3. likely failure class: code bug, flaky test, infra, dependency, or timeout/resource limit
4. smallest next debugging action

For performance changes also report:

1. whether the regression is broad or narrow
2. the most affected workload family
3. whether more local benchmarking is needed before code changes