Bug: auto-mode safety review request drops reasoning_content, DeepSeek rejects it and the action is held

Issue details

With a custom openai-chat-completions connection to DeepSeek (deepseek-v4-flash, thinking mode) and the default auto permission mode, shell commands are sometimes held with:

{"error":{"type":"tool_review_held","tool_name":"shell","message":"Safety reviewer unavailable; action held","reason":"review_unavailable","review_cause":"transport_permanent",...}}

The review request itself fails with HTTP 400 from DeepSeek:

The `reasoning_content` in the thinking mode must be passed back to the API.

The review request is built from the pending assistant tool call (build_review in src/gateway/chat_completions.zig). In the captured requests that assistant message never has reasoning_content, while the main-loop requests in the same session do replay it. DeepSeek doesn't reject every such request, but when it does, the command the user asked for doesn't run.

Environment: fx v0.0.14 (build_review unchanged on current main), Linux, fx acp, custom connection with "protocol": "openai-chat-completions", base_url DeepSeek, model deepseek-v4-flash.

Steps to reproduce

1. Configure the connection above and leave permission_mode at the default.
2. Ask something that makes the model think before running a shell command, e.g. give it a setup command to run and then ask it to run the tests.
3. Some runs: the review call returns 400 and the shell call fails with review_unavailable.

Seen in 4 of 18 runs across a few prompts. All 38 review requests we captured were missing reasoning_content, whether they got 400 or 200.

Trace

Not attached (captured from network logs, not /trace).

Sensitive information

• I reviewed the information above and removed secrets and other sensitive data.