[BUG] Claude Code on the web: autofix sessions stopped receiving bot PR review comments (Copilot/Codex)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

On Claude Code on the web (managed cloud sessions), with a PR subscribed via subscribe_pr_activity and the "Autofix pull requests" setting ON, the session no longer receives <github-webhook-activity> events for bot-authored PR reviews/comments — e.g. GitHub Copilot (copilot-pull-request-reviewer[bot]) and Codex (chatgpt-codex-connector[bot]). The reviews post to GitHub normally, but never reach the session, so autofix never acts on them.

The drop is selective, which is the key signal:

• ✅ Human-authored comments are delivered.
• ✅ CI / check-status events are delivered.
• ❌ Bot-authored reviews and review comments are silently dropped (Copilot, Codex; also vercel[bot] / supabase[bot] comments).

This looks like an anti-feedback-loop sender-type filter (meant to stop a session reacting to its own bot comments) that is now scoped too broadly and also excludes third-party review bots. These events used to be delivered and autofix acted on them until ~mid-May 2026.

What Should Happen?

A subscribed / autofix-enabled web session should receive bot-authored PR review comments from third-party reviewers (GitHub Copilot, Codex, etc.) as <github-webhook-activity> events — as it did until ~mid-May 2026 — so it can address them automatically. Any loop-prevention filter should exclude only the session's own bot identity, not all bot/App-authored review and comment events.

Error Messages/Logs

No error is surfaced. The failure is silent — dropped events produce no message, no notification, and nothing in the session transcript.

Steps to Reproduce

1. In Claude Code on the web, have a session subscribed to a PR (autofix setting ON, or subscribe_pr_activity).
2. On that PR, trigger a GitHub Copilot review and a Codex review so each posts review comments.
3. Separately, post a comment as a human user, and let CI run.
4. Observe which events reach the session.

Observed on a draft PR (times UTC, 2026-05-27):

| Time | Event on GitHub | Author | Reached session? |
|---|---|---|---|
| 04:29:52 | preview-deploy comment | vercel[bot] | ❌ |
| 04:29:52 | branch-ignored notice | supabase[bot] | ❌ |
| 04:34:28 | review + 3 inline comments | copilot-pull-request-reviewer[bot] | ❌ |
| 04:40:24 | review + 2 inline comments | chatgpt-codex-connector[bot] | ❌ |
| 04:41:36 | top-level comment | human (repo owner) | ✅ |

Only the human comment produced a <github-webhook-activity> event in the session. All four bot events were absent from the session despite existing on the PR (confirmed by reading the PR via the GitHub API).

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Claude Code on the web (cloud sessions), 2026-05-27

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Non-interactive/CI environment

Additional Information

• Surface: Claude Code on the web (managed remote execution / cloud sessions with GitHub integration). The Platform/OS/Terminal dropdowns don't map cleanly to the web product; closest values selected.
• Related: #52474 looks like the same auto-fix event-delivery subsystem on the desktop app, but there all events are dropped (including human reviews and CI). This report is web/cloud and the drop is selective (only bot-authored reviews/comments go missing while human comments + CI are delivered). Possibly a shared root cause — cross-linking for triage.
• Same-subsystem feature request: #50555 (PR-activity subscription permissions).
• Suggested direction: scope the loop-prevention filter to the session's own bot identity, or allowlist known review apps (copilot-pull-request-reviewer[bot], chatgpt-codex-connector[bot]); and add a visible "event received / filtered" indicator for diagnosability (echoes #52474's request).
• Ruled out: GitHub's 2026-05-19 Copilot-review change is not the cause — the Copilot/Codex reviews post to GitHub normally; the gap is delivery into the session.