Bug Report: Context Serialization Fails When Switching from Reasoning to Non-Reasoning Models

Environment

• OpenClaw Version: 2026.5.5 (b1abf9d)
• OS: macOS 26.4.1 (arm64)
• Node: 22.22.2

Summary

When switching from a reasoning-capable model (e.g., venice/claude-sonnet-4-6) to a non-reasoning model (e.g., openai/gpt-5.5) mid-session, OpenClaw fails to sanitize reasoning items from the context payload before sending to OpenAI. OpenAI rejects the request with a 400 validation error.

Error Message

400 Item 'msg_01df1e7d03fcdcc20069fb58de4f908194b630002ad65511ec' of type 'message' was provided without its required 'reasoning' item: 'rs_01df1e7d03fcdcc20069fb58d8f8e88194a0e1d118390470f'.

Reproduction Steps

1. Start a session on a reasoning-capable model (e.g., venice/claude-sonnet-4-6)
2. Have a few conversation turns (reasoning items are created in session context)
3. Switch to a non-reasoning model via dropdown or session_status (e.g., openai/gpt-5.5)
4. Send a new user message
5. Result: Assistant turn fails with [assistant turn failed before producing content] and the above 400 error

Expected Behavior

OpenClaw should detect when switching to a non-reasoning model and:

• Strip/omit reasoning items from the context payload
• OR convert reasoning items to text-only format
• OR trigger a sanitizing compaction that removes incompatible items

Actual Behavior

OpenClaw sends the full context including reasoning item references to OpenAI, which doesn't support the reasoning format and rejects the request.

Session Evidence

{
  "type": "message",
  "id": "c314cd87",
  "timestamp": "2026-05-06T16:14:39.281Z",
  "message": {
    "role": "assistant",
    "content": [{"type": "text", "text": "[assistant turn failed before producing content]"}],
    "api": "openai-responses",
    "provider": "openai",
    "model": "gpt-5.5",
    "stopReason": "error",
    "errorMessage": "400 Item 'msg_01df1e7d03fcdcc20069fb58de4f908194b630002ad65511ec' of type 'message' was provided without its required 'reasoning' item: 'rs_01df1e7d03fcdcc20069fb58d8f8e88194a0e1d118390470f'."
  }
}

Impact

• Model switching mid-session is broken when reasoning capability differs
• Users cannot switch from Venice (reasoning) → OpenAI (non-reasoning) without restarting session
• Conversations lose continuity

Suggested Fix

Add context sanitization logic in the model request builder:

1. Check if target model supports reasoning (via model capabilities)
2. If not, strip reasoning items from message payload
3. Preserve text content but omit reasoning references

Workaround

Restart the session (lose history) when switching between reasoning and non-reasoning models.

Related Models

• Reasoning-capable: venice/claude-sonnet-4-6, venice/claude-opus-4-*
• Non-reasoning: openai/gpt-5.5, openai/gpt-5-mini, openai/gpt-4o