openclaw - 💡(How to fix) Fix [Bug]: embedded agent (memory flush) fails when session history contains DeepSeek thinking-mode messages [1 comments, 2 participants]

Bug: embedded agent (memory flush) fails when session history contains DeepSeek thinking-mode messages

Bug type

Regression / Integration bug

Summary

When a session contains messages from DeepSeek in thinking mode (which produce reasoning_content fields), any subsequent embedded agent run (e.g. memory flush before compaction) that inherits the session history will fail with a 400 error from DeepSeek:

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

The fallback model (claude-sonnet) then also fails with:

400 消息内容不能为空

This causes memory flush to silently fail — important context from the session is never written to memory files.

Steps to reproduce

1. Start a session with DeepSeek thinking mode enabled (e.g. thinkingDefault: low with a DeepSeek model)
2. Have a multi-turn conversation that produces reasoning_content in the assistant messages
3. Wait for or trigger a memory flush (compaction or session end)
4. Observe embedded agent failure in logs

Expected behavior

The embedded agent should either:

• Strip reasoning_content from inherited session history before sending to DeepSeek (since it's a background task, not a continuation of the thinking chain), OR
• Use a non-thinking model for embedded agent runs regardless of the parent session's model

Actual behavior

The embedded agent inherits the full session history including reasoning_content fields, then calls DeepSeek without passing those fields back, causing a 400 rejection. The fallback to sonnet also fails because the message body is left empty after stripping the reasoning content.

Log evidence

Embedded agent failed before reply: All models failed (2): 
  deepseek/deepseek-v4-flash: 400 The `reasoning_content` in the thinking mode must be passed back to the API. (format) 
  openai/claude-sonnet-4-6: 400 消息内容不能为空 (format) 
  LLM request failed: provider rejected the request schema or tool payload.

OpenClaw version

2026.4.25 (aa36ee6)

Operating system

macOS (arm64)

Install method

npm global

Model

deepseek/deepseek-v4-flash (thinking mode)

Impact and severity

High — memory flush silently fails every time a DeepSeek thinking-mode session triggers compaction. Long-running sessions lose all memory consolidation. The failure is silent from the user's perspective (no error shown in chat).

Workaround

Use a non-thinking model as the default (e.g. sonnet). New sessions without DeepSeek thinking history are not affected.

Additional information

DeepSeek's API requires that reasoning_content from a previous assistant turn be passed back in the next request when using thinking mode. This is a DeepSeek API constraint, not a bug in DeepSeek. The fix should be in OpenClaw's embedded agent message history builder: either strip reasoning_content fields when constructing the history for background/embedded runs, or ensure they are properly forwarded.