| name | pr-failure-scan |
| description | Analyze a dotnet/runtime PR's CI failures, skip failures already known to Build Analysis, find matching Known Build Errors, and create or draft new KBEs for the remaining failures. Supports dry-run output to local markdown files instead of creating GitHub issues.
|
PR Failure Scan Skill
Use this skill for requests like:
/pr-failure-scan <PR URL>
/pr-failure-scan <PR URL> dryrun
create KBEs for PR <number>
This skill is PR-targeted and local-account-aware. It complements the
scheduled CI failure scan workflow but does not try to reproduce the workflow's
safe-output or follow-up PR behavior.
Shared instructions
Before triaging any failure candidate, read and follow:
.github/workflows/shared/create-kbe.instructions.md
That file owns the shared KBE lookup flow, KBE body template, signature
specificity guidance, verification checklist, and the rules for identifying
"existing KBE that Build Analysis likely missed".
Step 0: Parse input
Accept one of:
- a PR number,
- a PR URL,
- text containing a PR reference,
- an optional trailing
dryrun token.
Interpret dryrun case-insensitively. The source repository is always
dotnet/runtime.
Examples:
/pr-failure-scan 123456
/pr-failure-scan https://github.com/dotnet/runtime/pull/123456
/pr-failure-scan https://github.com/dotnet/runtime/pull/123456 dryrun
Step 1: Resolve identities and require explicit permission for non-owned PRs
- Resolve the target PR metadata: title, author login, URL, head SHA, base
branch, state, draft status.
- Resolve the currently authenticated GitHub user (
gh api user is fine).
- Compare the PR author's login with the authenticated user's login.
If they differ, you must ask for explicit confirmation before doing the rest
of the run. Running the command itself is not enough permission.
The confirmation must warn the user that this skill can create GitHub issues
under their account, and that analyzing an unknown PR carries risk because the
resulting issue content and decisions will be attributed to them.
If the user declines, stop immediately.
Step 2: Gather PR CI context
Use the latest completed check runs for the PR head SHA.
Collect all of:
- The
Build Analysis check run payload from the GitHub REST API.
- The non-success CI check runs for the PR head SHA.
- The AzDO build URLs linked from those check runs.
The local skill has a different candidate-source model than the scheduled
workflow:
- The workflow starts from a fixed main-branch pipeline list.
- This skill starts from one PR and must use the PR's Build Analysis results to
decide what is already known, what is still unknown, and which failed
pipelines were excluded from Build Analysis.
Step 3: Decide which failures are in scope
Build the candidate list from two sources:
- Build Analysis unknowns on analyzed pipelines.
- Parse the
Build Analysis check text for Create issue in this repo links
and the surrounding failure description.
- Each such entry is an in-scope candidate unless you later find an existing
matching KBE.
- Failed pipelines explicitly excluded from Build Analysis.
- Parse the Build Analysis warning section listing pipelines excluded from
analysis.
- For each excluded pipeline that also failed on this PR, inspect its AzDO
build/timeline/logs and derive concrete failure candidates yourself.
Skip any failure that Build Analysis already treated as known. The skill should
not re-triage failures already recognized by Build Analysis.
If Build Analysis and the raw PR checks disagree, prefer the raw CI evidence and
report the disagreement in the final output.
Step 4: Analyze each candidate failure
For each candidate failure, do the following:
- Gather the most concrete available evidence:
- the Build Analysis excerpt or create-issue link context,
- the AzDO failed timeline record,
- the relevant build or task log,
- for Helix-submitted legs, the Helix console log or failure details when
accessible.
- Extract the narrowest signature that fits the shared rules.
- Run the shared KBE lookup flow from
.github/workflows/shared/create-kbe.instructions.md.
- Classify the candidate into exactly one of these buckets:
- new KBE needed
- existing KBE matched but Build Analysis likely missed it
- unhandled
PR-specific guidance for excluded or partially-analyzed pipelines
For PR-targeted runs, it is acceptable to prepare a draft-only KBE candidate
for a Build Analysis unknown or excluded failure even when the evidence is not
yet strong enough for fully automated filing, as long as you clearly mark why
the case still needs human review.
Use this only for the local skill. Do not feed this relaxed draft behavior
back into the scheduled workflow rules.
Typical cases where a draft-only candidate may still be useful:
- Build Analysis already surfaced a
Create issue in this repo link for the
failure.
- The pipeline is excluded from Build Analysis, but the PR clearly contains a
repeated failure family worth turning into a reviewable draft.
- The failure has a recognizable issue shape, but the exact log line still
needs to be copied from a deeper log.
For such draft-only cases:
- keep the issue in the new KBE needed bucket only for this local skill,
- make the uncertainty explicit in the draft body,
- explain the missing evidence in the final report.
If a candidate is too weak even for a useful draft, place it in unhandled
with a concrete reason.
Dry-run drafting guidance
When producing a dry-run draft, optimize for a human-reviewable issue proposal
that is close to what a developer would want to file from the PR, even if the
draft still needs final cleanup before live filing.
In particular:
- Prefer a descriptive draft title that names the concrete failure shape over a
generic
Hang: / Test failure: title when that better communicates the
issue for a human reviewer.
- If the failure happens during xUnit discovery, prefer the discovery-time
signature over a later native-crash bucket when both are visible and the
discovery signature is the more stable reusable match.
- For infra-shaped submission failures, prefer a multi-token array signature
over one long literal line when the issue is really identified by a small set
of stable tokens.
Step 5: Write draft files and optionally create live issues
Draft files for all new KBEs
For every item in the new KBE needed bucket, always write a markdown draft
file before deciding whether to create a GitHub issue. This applies to both
dryrun and live mode.
- Write one markdown file per
new KBE needed item.
- Prefer the session artifact directory if it is available. Otherwise, use a
temporary directory outside the repo. Do not write draft files into the repo
checkout.
- Report the exact file path for every draft.
Each draft file should contain:
- proposed title,
- proposed labels,
- a human-review-oriented draft body,
- any explicit note when the draft still needs more evidence before live filing.
These dry-run files are review artifacts, not necessarily byte-for-byte live
issue bodies. Keep them easy to compare and edit on disk.
Prefer this layout for dry-run files:
# Draft KBE issue
- Proposed title: `...`
- Proposed labels: `Known Build Error`, `blocking-clean-ci`
## Draft issue body
## Build Information
Build: ...
Build error leg or test failing: ...
## Error Message
```json
...
```
**Failure details:**
```text
...
```
**Affected legs:**
- ...
**Console Log:**
- ...
**First build in window:**
- ...
**Recommended action:**
...
For dry-run artifacts, prefer the layout above over adding extra sections like
Pull request: or ## Error Details unless they are truly necessary for
understanding the draft.
When linking to a draft file in a user-facing message or confirmation question,
include both:
- a clickable local-file link when the UI supports one, and
- the plain absolute path.
Dry-run mode
In dryrun mode:
- Do not create any GitHub issues.
- Stop after writing the draft files and producing the final report.
Live mode
In live mode:
- First prepare the complete batch of proposed issues.
- Write all draft files as described above.
- Process the proposed issues one by one. For each KBE:
- Ask the user whether to create this specific issue.
- The question must include:
- the proposed issue title,
- a short description of what the KBE is about,
- the list of failures it covers, or a shortened summary if the list is
long,
- a clickable draft-file link when possible plus the plain absolute path,
- a clear statement that the GitHub issue will be created on behalf of the
currently authenticated user.
- Only create that issue if the user explicitly confirms.
- If the user declines, skip that issue and continue to the next proposed
KBE.
Use the shared KBE template guidance for the body.
Every live-created issue body must include a visible AI disclosure note because
the issue is being filed under a developer account. A concise note is enough,
for example:
[!NOTE]
This issue draft was prepared with GitHub Copilot assistance and reviewed by
the submitting developer.
Step 6: Final output format
The final response must contain all four sections below.
1. New issues created or draft files
List either:
- the created issue URLs, or
- the dry-run draft file paths.
2. New issues and the failures they cover
For every newly created or newly drafted KBE, provide:
- title,
- URL or file path,
- short description of what the issue is about,
- the specific failure or failure family it covers.
3. Existing KBE matches that Build Analysis likely missed
For every such failure, provide:
- failing leg / build context,
- matched KBE number and URL,
- why it appears to match,
- why Build Analysis likely missed it.
4. Failures not handled
For every unhandled failure, provide:
- failing leg / build context,
- why it was not handled,
- what evidence or follow-up would be needed to handle it.