tools: claude: update workflow with systemd changes

[pzmarzly]

systemd project recently adopted this workflow, improving it in the process.
New workflow is more secure, and has more stable output format.

Thanks @daandemeyer!

I kept some of the differences:

• we want to re-use /review-pr skill as much as possible
• we don't mind burning tokens on 3rd party PRs automatically
• we're using openrouter backend

Demo:

• new PR Test systemd's fork pzmarzly/demo--claude-bot-reviews#12 (comment)
• existing PR Cross-fork test pzmarzly/demo--claude-bot-reviews#8 (comment)
• in systemd repo Introduce several more helper functions to read sysfs attributes and use them where applicable systemd/systemd#41065 (comment)

[github-actions]

Claude: suggestion: listComments uses per_page: 100 without pagination. On a long-lived PR with 100+ top-level comments, the tracking comment could be missed, causing a duplicate. Consider using github.paginate():

Suggested change

	per_page: 100,
	const issueComments = await github.paginate(
	github.rest.issues.listComments,
	{ owner, repo, issue_number: prNumber, per_page: 100 }
	);

Then use issueComments directly (it returns a flat array, no .data wrapper).

[github-actions]

Claude: suggestion: When JSON.parse fails, the error is logged but execution continues silently — the tracking comment ends up saying "no issues found" which hides the real problem. Consider using core.warning() so the workflow run status reflects the parse failure:

Suggested change

	console.log(`Failed to parse structured output: ${e.message}`);
	} catch (e) {
	core.warning(`Failed to parse structured output: ${e.message}`);
	}

[github-actions]

Claude: suggestion: The schema requires line as a non-nullable integer. Issues that aren't tied to a specific line (commit message problems, general documentation observations, CLAUDE.md improvement suggestions) can't be represented — Claude will either fabricate a line number or drop the finding.

Consider making line optional by removing it from the required array and adding a guard in the post job to skip createReviewComment when line is absent (those findings would still appear in the tracking comment summary).

[github-actions]

Claude: nit: When a re-run is triggered while a previous "re-reviewing" banner is already present, the new update prepends another > **Claude is re-reviewing…** on top of the previous one. Multiple rapid synchronize events (e.g. force-push + immediate amend) will accumulate layers of nested blockquote headers.

Consider stripping any existing re-reviewing prefix from prevBody before wrapping again, or replacing the body with just the status line + the last actual review content.

[pzmarzly]

It should eventually fix itself though, once workflow succeeds

[github-actions]

Claude: nit: "${{ needs.review.result }}" is injected directly into the JavaScript source via GitHub Actions expression expansion. While needs.*.result is a GitHub-controlled enum and safe here, the rest of the file consistently uses environment variables for expression values (e.g. process.env.PR_NUMBER, process.env.HEAD_SHA). For consistency:

env:
    REVIEW_RESULT: ${{ needs.review.result }}

if (process.env.REVIEW_RESULT !== "success") {

[github-actions]

Claude: review of facebook/bpfilter #472 (8ac27f6)

Suggestions

• Marker injection fails for single-line summaries — .github/workflows/claude-pr-review.yaml:388 — If Claude's summary has no newline, replace(/\n/, ...) never matches and the <!-- claude-pr-review --> marker is never inserted, causing duplicate tracking comments on re-runs
• Setup comment search not paginated — .github/workflows/claude-pr-review.yaml:46 — listComments uses per_page: 100 without pagination; PRs with 100+ comments will miss the tracking comment
• JSON parse failure silently swallowed — .github/workflows/claude-pr-review.yaml:345 — When JSON.parse fails, the error is logged but the workflow reports "no issues found" instead of signaling a warning
• Schema requires line as non-nullable integer — .github/workflows/claude-pr-review.yaml:120 — Issues not tied to a specific line (commit messages, CLAUDE.md suggestions) can't be represented; consider making line optional

Nits

• Re-review banner nests on repeated triggers — .github/workflows/claude-pr-review.yaml:54-59 — Multiple rapid synchronize events accumulate nested blockquote "re-reviewing" headers
• Expression interpolation in JS source — .github/workflows/claude-pr-review.yaml:319 — ${{ needs.review.result }} injected directly into JS; rest of file uses process.env.* pattern for consistency

[pzmarzly]

@daandemeyer do you mind taking a look at the issues that the bot found?

[qdeslandes]

Reviewing AI instructions really is not fun. I had a quick look and it LGTM, letting it run on PRs will be the proper way to actually check if it does what we expect it to do. I'll let @daandemeyer have a look at your comment if he wants to.