[Bug]: Context handoff after an interrupt leads to Opus 5.5 safeguard block (reasoning_extraction)#23881

Before submitting

• I searched existing issues and did not find a duplicate. (Closest is #18601, which is a different trigger and is closed.)
• I included enough detail to reproduce or investigate the problem.

Area

Not sure

Steps to reproduce

1. Use the Claude provider with claude-opus-5-5.
2. Run a long multi-step task that produces many tool calls and reasoning items.
3. Interrupt the run, then send a follow-up message.
4. T3 Code performs a provider context handoff. The new session calls t3_thread_read on its own thread to recover context, and reads back the earlier activity, including reasoning items.

Expected behavior

The thread picks up the follow-up message and continues normally.

Actual behavior

Right after the self-read, the response is blocked with API Error: Opus 5.5's safeguards flagged this message ... Details: [reasoning_extraction]. After that, every later message in the thread gets the same block, even unrelated ones, so the thread can't be used.

It looks like reading the earlier session's reasoning items back into the new context is what trips the classifier. Leaving reasoning items out of the handoff and self-read context might avoid it.

Impact

Blocks work completely

Version or commit

Unknown

Environment

macOS, T3 Code with the Claude provider, model claude-opus-5-5

Logs or stack traces

Workaround

Start a new thread and restate the task directly, without having the agent read the old thread.

🤖 Generated with Claude Code