Agent mode discards session summary: cleanup phase gates updateCommentLink on commentId, agent mode hard-codes commentId: undefined

[Kevin-Whelan-Stagwell]

Summary

When the action runs in agent mode (prompt: set, no @claude trigger), the session summary the SDK emits never reaches GitHub. The Claude session runs to completion, generates a review, and the output is discarded at action cleanup.

Verified on v1 (action built from v1.0 tag per action.yml).

Observed behavior

Workflow config (our .github/workflows/claude-code-review.yml — automated PR-review workflow):

on:
  pull_request:
    types: [opened, synchronize]
jobs:
  claude-review:
    # ...
    steps:
      - uses: actions/checkout@v4
      - uses: anthropics/claude-code-action@v1
        with:
          claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
          classify_inline_comments: "false"
          prompt: |
            Review this PR. Report blockers, should-fixes, and nits.
            If clean, say "no blockers found" plainly.

Across 7 PRs on our repo over two days (N=7):

• 5 clean PRs (docs-only, small code, tests, etc.): silent. Zero bot output on any of pulls/<n>/reviews, issues/<n>/comments, pulls/<n>/comments. Session succeeded (subtype: "success", is_error: false), turn count 15–22, cost $0.23–$0.49 per run. No error anywhere.
• 1 PR with a hard-blocker pattern (deliberate throw): posted 3 inline findings as a review. Summary text still not present.
• 1 PR with show_full_output: "true" enabled: revealed the SDK's final assistant text event contained a complete 1098-character clean-PR review ending "No blockers found. This is a clean docs-only change." — emitted 117ms before "type": "result". Nothing reached GitHub.

Root cause (source reading at ref=v1)

src/entrypoints/run.ts cleanup phase (~line 244 in current v1):

if (commentId && context && isEntityContext(context) && githubToken && octokit) {
  try {
    await updateCommentLink({
      commentId, ..., outputFile: executionFile, ...
    });
  } catch (error) {
    console.error("Error updating comment with job link:", error);
  }
}

updateCommentLink is the function that reads executionFile (where the SDK wrote the summary), formats it, and updates a GitHub comment with it. It only runs if commentId is truthy.

src/modes/agent/index.ts (end of prepareAgentMode):

return {
  commentId: undefined, // No tracking comment in agent mode
  branchInfo: { ... },
  mcpConfig: ourMcpConfig,
  claudeArgs,
};

Agent mode explicitly returns commentId: undefined. Cleanup sees falsy → skips updateCommentLink → summary never published.

src/modes/tag/index.ts:prepareTagMode doesn't have this issue — it calls createInitialComment(octokit.rest, context) up front, captures commentData.id, and the tracking comment becomes the summary destination.

Inline comments are not affected

src/entrypoints/post-buffered-inline-comments.ts reads /tmp/inline-comments-buffer.jsonl and posts independently of commentId. When the session uses the inline-comment MCP tool (hard-blocker patterns trigger this), those findings do post. The summary does not.

Current workaround

track_progress: "true" forces tag mode even when prompt: is set (per src/modes/detector.ts:18–29), which flows commentId correctly. Verified working on our repo — tracking comment appears at session start and gets rewritten with the 1413-character final review 71 seconds later.

Side effect: a transient "Claude is working on this…" placeholder comment appears before the rewrite. That's fine for short sessions but isn't ideal for the automation-review use case where users just want the final output.

Feature request

A documented option for "post the final summary to the PR even in agent mode, without the tracking-comment dance." Something like:

post_summary: "always"   # or post_summary_on_clean: "true"

Or equivalently: a config that makes agent mode create a lightweight summary-only comment at cleanup (not a tracking comment at start) if one doesn't already exist.

Either way, the underlying fix is narrow: in run.ts cleanup, when in agent mode with an executionFile that has extractable summary content, post it to issues/<pr>/comments as a new comment rather than silently discarding it.

Why this matters for review automation

PR-review workflows are a natural fit for agent mode (one prompt, run on every push, no @claude trigger ceremony). The silent-on-clean default surprises maintainers who expected "clean PR → short 'no blockers' summary" and got nothing. It also breaks cross-check patterns where the reviewing human depends on being able to observe bot findings alongside their own.

Happy to open a PR if the approach makes sense to you — the fix is small and localized to run.ts and agent/index.ts. Let me know.