Codex Desktop automation + DeepSeek Responses: intermittent 400 'reasoning_text must be passed back' on tool-call continuations

Description

Codex Desktop automations (heartbeat/cron) with a DeepSeek Responses provider intermittently fail mid-turn with:

The `reasoning_text` in the thinking mode must be passed back to the API.
{"type": "invalid_request_error"}

Failure happens during tool-call continuations inside an automation run (e.g. an automation that executes a script then reports back). The same conversation works most of the time; the 400 appears intermittently on a later continuation request. Interactive turns in the same session succeed — only automation turns hit it.

Environment

• macOS 27.0, Apple Silicon
• Codex Desktop 26.825.51511, bundled app-server 0.151.0-alpha.7.2
• model_provider = "deepseek", base_url = "https://api.deepseek.com", wire_api = "responses", model = "deepseek-v4-flash"

Evidence

Captured at the proxy (request body before forwarding). Reasoning items sent by Codex look like:

{
  "type": "reasoning",
  "content": [{"type": "text", "text": "..."}],
  "encrypted_content": "...(38 bytes)",
  "summary": []
}

• summary is empty; the plaintext thinking lives in content[0].text.
• encrypted_content is present but tiny (38 bytes, not the real content).
• On the failing continuation, a reasoning item arrives without usable plaintext (content empty/missing), and DeepSeek's stateless Responses endpoint rejects with reasoning_text must be passed back.

DeepSeek's thinking-mode contract requires the prior reasoning to be replayed in plaintext; it does not accept OpenAI-style encrypted reasoning blobs.

Expected behavior

• Codex should replay the prior reasoning in a form DeepSeek accepts (plaintext in every reasoning item it sends), including on automation tool-call continuations.
• Related: #24500 covers the same DeepSeek contract on the chat-completions path (reasoning_content); this is the Responses/app-server variant (reasoning_text).

Workaround verified locally

A localhost proxy that scans request input for reasoning items lacking plaintext content[0].text and injects a placeholder text part ("(thinking unavailable)") makes the automation run complete without this 400. This confirms the failure is a missing-plaintext-reasoning serialization issue in Codex, not the scheduler or provider auth.

Start by tracing the Responses/app-server serialization path for reasoning items during automation tool-call continuations, using the provided DeepSeek configuration and automation reproduction. Inspect the outgoing request input for missing plaintext reasoning content; done means every continuation sends DeepSeek-acceptable plaintext reasoning and the intermittent 400 no longer occurs.