| name | analyze-ci-failures |
| metadata | {"version":"2026-08-14T00:00:00.000Z"} |
| description | Analyze collected vcpkg CI failure evidence, classify regressions and known failures, identify root causes, recommend actions, and generate a report by package and triplet.
USE FOR: explicit requests for CI regression triage, a failure inventory, baseline classification, action recommendations, or a formal CI failure report.
DO NOT USE FOR: fetching logs as the final output, targeted debugging of one failure, answering a one-off question about logs the user pasted or pointed at, general coding, creating ports, or modifying portfiles. When requested analysis lacks local evidence, run fetch-vcpkg-ci-logs as the acquisition stage first.
**ANALYSIS SKILL** USES: locally collected CI logs, vcpkg baselines, Azure DevOps metadata.
|
vcpkg CI Failures Analyzer
Overview
Consumes already-collected CI evidence, cross-references it with baselines, and produces a regression
report. If evidence has not yet been collected, invoke fetch-vcpkg-ci-logs first, then return to this
workflow. Do not duplicate its acquisition steps here.
MCP Tools
| Tool | Purpose |
|---|
github-mcp-server-get_file_contents | Read baseline files |
Prerequisites
- Local failed step logs and relevant extracted failure-log artifacts.
- Build metadata and full Azure build URL.
- For PR builds, PR metadata and full PR URL.
Evidence Acquisition
When the required evidence is not already local, delegate acquisition to a subagent using
fetch-vcpkg-ci-logs. Prefer a fast, low-cost model for this mechanical work. Ask
the subagent to return only the evidence inventory and local paths; keep regression analysis and report
generation in the main agent.
When the evidence is already available locally, proceed directly to analysis.
Workflow
OUTPUT RULE: This skill applies only when the user asked for triage or report output, so a report.md
is always required once it is selected. Write it as soon as step-log analysis completes, then enhance it
from artifacts. 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 analysis, read .\.github\skills\shared\azure-vcpkg-ci-notes.md.
Phase 1: Analyze failed step logs
-
Inventory evidence — Confirm the build metadata, failed step-log paths, artifact paths, and triplets
available from the acquisition step.
-
Scan step logs — Extract every REGRESSION: line and retain exact failure type keywords:
BUILD_FAILED, FILE_CONFLICTS, POST_BUILD_CHECKS_FAILED, and
CASCADED_DUE_TO_MISSING_DEPENDENCIES. Capture 2-3 surrounding lines for context.
-
PR feature-test logs — PR builds may not have REGRESSION: lines. Instead scan for FAIL:,
failed with, and feature-test error: lines. Capture verbatim:
- Compiler errors (missing headers, undefined symbols)
- Post-build check failures (file path issues, misplaced files)
- Version validation errors
- Platform-specific feature guard messages
Report each feature failure individually. Dependency ports that fail get their own entry.
-
Version validation — Inspect the "Validate version files" evidence for version database errors.
-
Write report immediately — Generate and save report.md using the step-log evidence.
Phase 2: Analyze artifacts and enhance
- Analyze artifacts — Read relevant
stdout-{triplet}.log tails and supporting logs. Classify per
references/vcpkg-failure-patterns.md.
- Baselines — Check both
ci.baseline.txt and ci.feature.baseline.txt.
- Enhance report — Add root causes, downstream impact, and actionable recommendations supported by
the evidence.
Report Requirements
Format per references/report-template.md:
- Full build URL:
[{buildNumber}](https://dev.azure.com/vcpkg/public/_build/results?buildId={buildId})
- For PRs:
[#{prNumber}](https://github.com/microsoft/vcpkg/pull/{prNumber})
- List every triplet by full name (e.g.,
x64-windows, arm64-linux) — never "N triplets". Only include triplets that actually had failures for this specific build.
- Use exact failure type keywords from logs:
BUILD_FAILED, POST_BUILD_CHECKS_FAILED, FILE_CONFLICTS, CASCADED_DUE_TO_MISSING_DEPENDENCIES — never paraphrase
- Include error messages verbatim — quote the exact text from logs for compiler errors, path issues, and validation failures
- Dependency ports' failures as separate entries
- Include baseline/known failures — report them with their failure types, but classify them separately from new regressions
Output Structure
Write reports under the session folder from the environment context, alongside the raw evidence, never
into the repository working tree:
{session-folder}/files/ci-failure-analysis/
├── ci-129315/ ← scheduled build
│ └── report.md
└── pr-51202/ ← PR build
└── report.md
Raw evidence stays where fetch-vcpkg-ci-logs placed it, in
{session-folder}/files/ci-failure-logs/; reference those paths from the report instead of copying them.
Critical Rules
- Do not fetch or download logs as part of this skill; use the acquisition skill first.
- Scan step logs first because
FILE_CONFLICTS may appear only there.
- Check both baseline files
- Never suggest
<= version constraints or VCPKG_BUILD_TYPE release
- Clearly distinguish proven root causes from hypotheses and requests for additional diagnostics.