[Cowork web] Safeguards repeatedly halt routine tool calls mid-reply in one session (upload, issue filing, session rename)
Status Open
Maintainer reply None cached
Activity 0 comments · opened Oct 6, 2026
Concerns Cowork on claude.ai (cloud session, web client), not the CLI. Model: Opus 5.5.
What happened
In one Cowork session on 6 Oct 2026, the in-stream safety check stopped the assistant's reply five times, each time on an ordinary action:
- Starting a browser session to upload the user's own video to their own YouTube channel.
- Calling a connector to open a GitHub issue reporting (1).
- Navigating a browser session to claude.ai to rename the session itself.
- The same rename, retried later.
- A plain "Try again" turn, which returned
API Error: Opus 5.5's safeguards flagged this message, details[reasoning_extraction], request[id removed].
Each stop withholds the rest of the reply and cancels unfinished tool calls, with no reason given. Once it starts in a session it keeps firing on unrelated, benign steps, so the session cannot finish or close its work. A fresh session doing the same rename succeeded at once.
Expected
- Routine actions on the user's own accounts are not halted.
- When a stop happens, the user is told what category triggered it, and a single false positive does not poison the rest of the session.
Steps
Not reproducible on demand. The session id can be provided privately on request.