npx skills add ...
npx skills add microsoft/vcpkg --skill analyze-ci-failures
npx skills add microsoft/vcpkg --skill analyze-ci-failures
Analyze vcpkg Azure DevOps CI failures. Downloads logs, identifies regression root causes, generates a report by package and triplet.
Fetches build metadata and failure logs via Azure DevOps REST API, cross-references with baselines, produces a regression report.
| Tool | Purpose |
|---|---|
github-mcp-server-get_file_contents | Read baseline files |
OUTPUT RULE: Write
report.mdas SOON as step-log analysis (Phase 1) completes. Do not wait for artifact downloads. You can update the file later. Your response MUST also contain the complete report content — not a summary.
URL RULE: Your response text (not just the report file) MUST include:
- For scheduled/manual builds: the full Azure DevOps URL
https://dev.azure.com/vcpkg/public/_build/results?buildId={buildId}- For PR builds: both the full PR URL
https://github.com/microsoft/vcpkg/pull/{prNumber}AND the Azure DevOps build URL
Before doing Azure-specific work, read .\.github\skills\shared\azure-vcpkg-ci-notes.md and follow its shared API and log-handling rules.
When you use the shared helper script, execute the .ps1 file directly by path. Do not open the script and inline its contents into a large PowerShell block.
Parse input — Extract buildId from Azure DevOps URL. For PR URLs:
See references/azure-devops-api.md. For PR inputs where you want the failed step logs immediately, you may instead use the shared helper script:
Fetch metadata — Build info, timeline, and artifacts list (parallelize these API calls).
Scan step logs — For each failed job, find "*** Test Modified Ports" task and fetch its log:
You may also use the shared helper script directly from a build id:
Captures all types: BUILD_FAILED, FILE_CONFLICTS, POST_BUILD_CHECKS_FAILED, CASCADED_DUE_TO_MISSING_DEPENDENCIES. Report every REGRESSION line using the EXACT failure type keyword from the log — never paraphrase. Also capture 2-3 lines around each error for root cause context. Extract the triplet name from each failed job (e.g., from job names like x64-windows Build or from log paths) — ensure ALL triplets with failures are listed in your report.
PR feature-test logs — PR builds may NOT have REGRESSION: lines. Instead scan for FAIL: or failed with lines showing per-feature failures. Capture and quote verbatim:
Report each feature failure individually. Dependency ports that fail get their own entry.
Version validation — Check "Validate version files" task. If failed, scan for version database errors and quote them verbatim. Fix: vcpkg x-add-version.
Write report immediately — Generate and save report.md using all step-log data. This ensures output exists even if later steps time out.
Invoke-WebRequest or curl shell commands for artifact ZIPs (not web_fetch). Only download "failure logs for {triplet}" — skip "file lists". Extract into ci-failure-analysis/{scope}/logs/{triplet}/. If download fails, still create the directory with a placeholder noting the URL.stdout-{triplet}.log last lines. Classify per references/vcpkg-failure-patterns.md. Update report with additional root cause detail.ci.baseline.txt and ci.feature.baseline.txt.Format per references/report-template.md:
[{buildNumber}](https://dev.azure.com/vcpkg/public/_build/results?buildId={buildId})[#{prNumber}](https://github.com/microsoft/vcpkg/pull/{prNumber})x64-windows, arm64-linux) — never "N triplets". Only include triplets that actually had failures for this specific build.BUILD_FAILED, POST_BUILD_CHECKS_FAILED, FILE_CONFLICTS, CASCADED_DUE_TO_MISSING_DEPENDENCIES — never paraphraseInvoke-WebRequest or curl shell commands for ZIPs — web_fetch can't download binariesPipelineArtifact — Container API won't workFILE_CONFLICTS only appear there<= version constraints or VCPKG_BUILD_TYPE releaselogs/{triplet}/ directory with a placeholder noting the download URL**`