Validate PR Description
Check a PR title and body against the three things this repository's gates read: the conventional-commit title, the issue-linking keyword, and the template sections the spec-coverage job parses. Catch them before submission, where the fix costs one edit instead of one CI round.
Migrated from .claude/commands/validate-pr-description.md under ADR-064, which
makes skills the single user-invocable surface. user-invocable: true is what
fires it as /validate-pr-description in both Claude Code and Copilot CLI.
Triggers
validate my PR description, check my PR title, is this PR body compliant,
validate PR metadata
Inputs
| Input | Source | Required |
|---|---|---|
| PR title | $0, or stated in conversation |
Yes |
| PR body | $1 as a path to a file, or pasted text |
Yes |
| Target repository | current checkout | Yes |
Process
Phase 1: Resolve the inputs
Read the title and body from the arguments when given. When they are not, ask for both rather than inferring them from the branch or the last commit: a PR body is usually longer than any commit message, and validating the wrong text reports a pass that does not describe the PR.
When the body arrives as a path, read the file. When it arrives as pasted text, work from the text directly.
Phase 2: Run the three checks
| # | Check | Rule | Failure meaning |
|---|---|---|---|
| 1 | Conventional commit title | <type>(<scope>)?: <description>, type one of feat, fix, docs, style, refactor, perf, test, chore, ci, build, revert |
Release tooling cannot classify the change |
| 2 | Issue-linking keyword | body holds at least one of Closes #N, Fixes #N, Resolves #N, or a past-tense variant |
The issue stays open after merge, and the PR fails the repository's issue-linkage rule |
| 3 | Template compliance | Summary non-empty, Changes has at least one item, Type of Change has at least one marked checkbox | The spec-coverage job reports FAIL on an unpopulated section |
Prefer the repository's own validator when it ships one, because it is the same code the gate runs. Resolve the script root through a plugin-root variable so the invocation works in a plugin install and not only in the upstream checkout:
SCRIPTS_DIR="${COPILOT_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-.claude}}/skills/github/scripts/pr"
uv run python "$SCRIPTS_DIR/validate_pr_description.py" \
--title "[title]" \
--body-file "[path-to-body.md]"
When that script is absent, validate manually against the table above. Say which path you took: a manual pass and a validator pass are not the same evidence.
Phase 2b: Check file mentions against the diff
When the body is a file and the checkout ships scripts/validation/pr_description.py,
run its local mode. It runs the same file-mention check the required
Validate PR check runs, so a pass here rules out a File mentioned but not in diff failure. It does not cover the dash or closing-keyword checks. Local mode
reads committed HEAD, so commit first:
uv run --frozen python -I scripts/validation/pr_description.py --body-file "[path-to-body.md]" --base origin/main
Exit 1 is a CRITICAL finding, usually File mentioned but not in diff. Report the
path and the fix: remove it, move it out of the change heading (## Changes,
## What changed, and the like), or wrap it in a code fence. With no change
heading, a bold or bulleted path counts anywhere, so remove or fence it. Do not
suggest the description-validation-bypass label; a human maintainer decides
that. Exit 2 means the check could not run: an unreadable or non-UTF-8 body
file, an unknown or invalid base ref, or git unavailable. For a missing
origin/main, pass --base with the ref the PR targets.
When the script is absent, or the body is pasted text with no file, say this
check was not run.
Phase 3: Report
Report per check, not as a single verdict. For each failure give the offending text and the corrected form, so the author edits rather than re-derives.
A closing keyword is a claim that the PR closes the issue. Verify it against the
diff before recommending one. When the diff does not deliver what the issue asks,
recommend Refs #N and say why, rather than upgrading a reference the merge will
act on.
Verification
- Title checked against the conventional-commit pattern, and the type named
- Body searched for every accepted issue-linking keyword, not just
Closes - All three template sections inspected, with the failing one named
- Validator path stated: the repository script ran, or the check was manual
- File mentions checked against the diff with
pr_description.py --body-file, or the check stated as not run - Every failure reported with its offending text and a corrected form
- Any recommended closing keyword checked against the diff first
Anti-Patterns
| Avoid | Why | Instead |
|---|---|---|
| Inferring the title from the branch name or last commit | Validates text that is not the PR's, and reports a pass the PR does not have | Ask for the title and body |
| Reporting one aggregate verdict | The author cannot tell which of the three checks failed | Report per check |
Recommending Closes #N because the PR mentions the issue |
A closing keyword acts on merge; an unsupported one closes live work | Verify against the diff, downgrade to Refs #N with a reason |
| Rewriting the body wholesale | The author loses their own wording to a style preference | Name the failing section and the minimal edit |
| Calling a manual pass a validator pass | Two different levels of evidence read as one | State which ran |
Extension Points
- New title types. The type list mirrors the repository's commit convention. When that convention gains a type, update the Phase 2 table and the validator together, not one of them.
- New template sections. The three checked sections are the ones the spec-coverage job parses. A new parsed section is a new row, not new prose.
- Other forges. The keyword list is GitHub and GitLab behavior. On Azure DevOps and Bitbucket the same keywords are traceability markers with no auto-close, so a missing keyword is a weaker finding there.