PR-activity webhook events render as if the user sent them, and non-actionable bot comment-edits are relayed with full raw payloads

Summary

When a Claude Code Remote session is subscribed to PR activity (via the PR-watching / "subscribe to PR activity" feature), incoming GitHub webhook events are injected into the conversation as <wake reason="external-event"><event source="github" ...> blocks. Two problems with this:

1. Misattribution in the UI. These injected blocks are not visually distinguished from a message the user actually typed. A large block of raw JSON dropped in by a webhook event reads, in the transcript, as if the user pasted it themselves. This is confusing and clutters the conversation — users have reported it looking like "a bug" because it appears to be their own message.

1. Non-actionable bot noise is relayed in full, repeatedly. Status-comment bots (e.g. a CI/deploy-preview bot) edit the same PR comment multiple times as a build progresses (queued → building → ready/skipped). Each edit fires a distinct issue_comment.edited webhook, and each one is relayed as a full wake event with the entire comment body inlined — including large opaque content (e.g. base64-encoded internal config blobs) that has zero informational value to the agent or the user. Over the life of a PR this can inject many kilobytes of pure noise into the context window across several edits, none of which ever requires a response.

Example (sanitized)

<wake reason="external-event" current-time="2026-08-14T05:25:22Z">
  <event source="github" kind="issue_comment.edited" from="system" trust="relay" untrusted-keys="author,file,comment,review_comments">
    <!-- An existing GitHub comment on the PR was updated in place. Review the new content and determine what action (if any) is needed to address it. -->
    {"author":"<deploy-preview-bot>","comment":"[vc]: #<long base64 config blob>\nThe latest updates on your projects...\n\n| Project | Deployment | Actions | Updated (UTC) |\n...","pr":"<org>/<repo>#<number>"}
  </event>
</wake>

Suggested fixes

• Visually distinguish (in the UI/rendering layer) turns that originate from an injected external event vs. an actual user-typed message, so they can never be confused with something the user sent.
• Recognize known noisy bot authors (deploy-preview/status bots that repeatedly edit the same status comment) and either suppress repeat edits of the same comment, strip/truncate clearly non-actionable binary/base64 content before relaying, or collapse them into a lighter-weight notification instead of the full raw payload.

Impact

Fills the context window with non-actionable content and makes it hard for users to follow their own conversation, since injected events are indistinguishable from their own messages.

<wake reason="external-event" current-time="2026-08-14T11:29:36Z">
  <event source="github" kind="pull_request_review.submitted" from="system" trust="relay" untrusted-keys="author,file,comment,review_comments">
    <!-- A GitHub review was submitted on the PR with no summary text. Its inline comments have already been delivered as separate events. -->
    {"author":"toumix","author_association":"OWNER","author_type":"User","notes":["1 review comment(s) already delivered as separate events"],"pr":"toumix/desire#68","review_id":4936712226,"review_state":"commented","state":"commented"}
  </event>
</wake>