Classify files in a large getsentry/sentry Scraps migration, confirm the proposed noise set with the user, mark the approved files as viewed, and report the substantive files left to review.
This skill uses Claude-specific invocation metadata and $ARGUMENTS; invoke it manually on other Agent Skills hosts.
Requires: authenticated gh and uv CLIs.
1. Resolve the PR
Accept a getsentry/sentry PR number or full URL from $ARGUMENTS. With no argument, the script resolves the pull request for the current branch. Ask for a PR only if that lookup fails.
Do not run this workflow against another repository or a PR that is not a large Scraps design-system migration.
2. Classify without mutation
Run from this skill directory so bundled paths remain skill-root-relative:
uv run scripts/classify_pr_files.py [<pr>]
The non-interactive script emits JSON to stdout and exits nonzero on fatal or partial mutation failure. It does not modify GitHub state unless --mark-viewed is explicitly supplied.
Successful and partial results use this shape:
{
"status": "success",
"repository": "getsentry/sentry",
"pr": 12345,
"head_sha": "abc123...",
"approval_token": "def456...",
"summary": {
"total": 139,
"noise": 128,
"substantive": 11,
"marked_viewed": 0,
"failed_to_mark": 0
},
"substantive": [{"path": "...", "classification": "substantive", "reason": "..."}],
"noise": [{"path": "...", "classification": "noise", "reason": "..."}],
"failed_to_mark": []
}
Fatal failures emit {"status": "error", "error": "..."} and exit nonzero. The script does not emit progress output or expose gh stderr.
Classification reasons:
| Reason | Classification | Meaning |
|---|---|---|
import-path-only |
noise | import or re-export declarations differ only by module path |
known-noise-file |
noise | generated codeowners baseline or snapshot mock |
pure-rename |
noise | GitHub reports a rename with zero changed lines |
destination-dir |
substantive | file is in a component directory being moved into components/core/ |
patch-unavailable |
substantive | GitHub omitted the patch, so the script fails closed |
has-substantive-changes |
substantive | changes are not a recognized mechanical migration |
3. Confirm the mutation
Show the user the noise files and substantive files with their reasons and the count summary. Retain the approval_token, then ask whether to mark that exact proposed noise set as viewed.
Do not run the mutation command until the user explicitly approves the displayed set. Approval applies only to that classification result.
4. Mark the approved files
After approval, rerun the same PR with:
uv run scripts/classify_pr_files.py [<pr>] --mark-viewed --approval-token <token>
The token binds approval to the PR head SHA and exact noise-path set. If the PR or classification changed, stop and show the new classification for fresh approval. The mutation uses GraphQL variables for PR-controlled paths and is idempotent: marking an already viewed file leaves it viewed.
If status is partial, report every path in failed_to_mark; do not claim the batch succeeded. Rerun only after resolving the reported GitHub CLI or permission failure.
5. Report
Show the substantive files as a table with the reason column. End with the total, noise, substantive, marked, and failed counts.
Do not perform the substantive code review unless the user also asks for it.
Fallback
If classification fails, use gh pr diff <pr> --repo getsentry/sentry to classify files manually with the table above. Treat ambiguous files as substantive.
If marking fails, leave the files unmodified and report the failure. Do not construct an inline GraphQL mutation or interpolate a filename into a shell command.