reasoning_extraction rejection on ordinary configuration request after visible history ingestion

Observed behavior

An isolated Claude Code conversation using imported visible user/assistant discussion history completed normal configuration requests, then rejected the next ordinary request with reasoning_extraction. We are requesting diagnosis of that rejection and clarification of supported external history ingestion.

Environment

• Claude Code 2.1.251
• Claude Desktop 2.19675.1 on macOS
• Model identified by the error: Opus 5 (1M context)
• Streaming JSON input/output; Remote Control attached to Desktop Code
• No hidden reasoning or customer conversation data was imported.

History transport

The process received externally supplied user and assistant frames. Imported users used shouldQuery: false. To persist a pending assistant record, a passive synthetic checkpoint was sent and then removed from the active conversation with rewind_conversation. The resulting active discussion was verified against saved history. This is an experimental history-ingestion path; we are not asserting that it is supported.

Visible synthetic discussion before the failure

1. user: Write a one-sentence prototype status update using our latest agreed display label and polling interval. Do not use tools.

1. assistant: The prototype is configured with Hazel as the display label and a polling interval of fifty-nine seconds.

1. user: Set the prototype polling interval to fifty-nine seconds, keeping the agreed display label. Write one sentence stating the current configuration. Do not use tools.

1. assistant: The current configuration uses Hazel as the display label with a polling interval of fifty-nine seconds.

1. user: Write a one-sentence prototype status update using our latest agreed display label and polling interval. Do not use tools.

1. assistant: The prototype is configured with Hazel as the display label and a polling interval of fifty-three seconds.

1. user: Set the prototype polling interval to fifty-three seconds, keeping the agreed display label. Write one sentence stating the current configuration. Do not use tools.

1. assistant: The current configuration uses Hazel as the display label with a polling interval of fifty-three seconds.

1. user: For this isolated prototype I propose Hazel as the display label and forty-seven seconds as the polling interval. Write one sentence agreeing to this test configuration. No tools.

1. assistant: I agree to use Hazel as the display label and forty-seven seconds as the polling interval for this isolated prototype.

Request rejected

Set the prototype polling interval to sixty-one seconds, keeping the agreed display label. Write one sentence stating the current configuration. Do not use tools.

Error

API Error: Opus 5 (1M context)'s safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations. Claude Code can't respond to this message with Opus 5 (1M context).

Try rephrasing the request in a new session or change your model.

Learn more: https://support.claude.com/en/articles/16049681

Details: `[reasoning_extraction]`

Request ID: req_011CfmfjxJCSjEZdNR7XK1WD

Expected and limits

The request records a synthetic display label and polling interval, without asking for hidden reasoning. Please diagnose the request ID and identify whether the history-ingestion path introduces an unsupported request shape.

Five completed discussion exchanges were preserved before this rejection. The assembled provider-side request was not captured, the exact trigger is unknown, and a deterministic fresh reproduction has not been established. The blocked request has not been retried or changed, and no model switch was used to work around it.

Requested help

Is there a supported way to add externally supplied visible user/assistant history to an active local Claude Code session, preserve its roles and ordering, and publish it to the connected native view without generating an extra turn?