[Bug]: PR watch wakes the agent for CodeRabbit's "Review triggered" acknowledgement, which has nothing to act on #32275

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

Real-world:

1. On a repository with CodeRabbit, have an agent call watch_pull_request on a PR and end its turn.
2. Post @coderabbitai review on the PR (the agent often does this itself after a push).
3. CodeRabbit replies with a new comment that starts "Actions performed / Review triggered." and later edits that same comment to "Review finished." or "Action not completed / Review rate limited."

Deterministic: call evaluatePullRequestWatch (apps/server/src/orchestration-v2/pullRequestWatch.ts) with a watch whose watermark precedes a new coderabbitai[bot] comment whose body starts with CodeRabbit's reply marker and "Actions performed ... Review triggered." It returns a remarks change.

Expected behavior

The acknowledgement does not wake the agent: it only says a review was queued. The watermark moves past it, and the agent is woken once when the comment is edited with an outcome (finished, or not completed and why), or when the review itself arrives.

Actual behavior

The acknowledgement wakes the agent as "1 new comment". The agent reads it, finds nothing to do, and ends the turn. On one watched merge request this note was re-sent in 8 wakes over about 6 hours across repeated review requests.

Impact

Minor bug or occasional failure

Version or commit

0.1.8-nightly.20261106.4262 (c55db3e); code read at main 6577c6b, unchanged at 240446e.

Environment

macOS and Linux, T3 server with Claude and Codex providers; GitHub and GitLab PRs reviewed by CodeRabbit.

Logs or stack traces

No raw wake text kept; the deterministic case above fails on main.

Workaround

None in T3. Telling the agent not to post review commands avoids it but loses re-reviews.

Proposed fix (for the PR, once triaged)

Treat a bot comment whose first 1,000 characters carry CodeRabbit's reply marker followed by "Action(s) performed ... Review triggered." or "Full review triggered." as progress: advance the watermark, do not wake, and let its later edit with the outcome wake once. The bound keeps a long body from stalling the evaluator. A bot reply that answers a question, or one created already as "not completed" or "rate limited", still wakes. Tests: "CodeRabbit's review triggered reply is not news" and "the acknowledgement's outcome edit wakes once" (both fail on main).