openclaw - 💡(How to fix) Fix Bug: Session reset via Control UI injects blank user message, crashing MiniMax provider [1 comments, 2 participants]

Bug: Session reset via Control UI injects blank user message, crashing MiniMax provider

Summary

When a session is reset via sessions.patch from the Control UI, the new session is created with a blank user message (content: [{"type": "text", "text": ""}]). This causes MiniMax to reject the request with error 2013 ("invalid params, messages must not be empty"), triggering an unnecessary model fallback. OpenClaw's self-repair mechanism eventually cleans up the blank message, but only after the first turn has already failed and fallen back to a different model.

Steps to Reproduce

1. Open a per-peer session (e.g. agent:main:direct:bob) in the Control UI
2. Reset or patch the session (e.g. via model change or session reset in the UI)
3. The new session file is created with a blank user message before any real user input
4. If the model is MiniMax, it rejects the empty message array with HTTP 400

Expected Behaviour

A freshly reset session should not contain any user messages until the user actually sends one. The session file should start with only the session header, model_change, and thinking_level_change entries.

Actual Behaviour

The session file starts with a blank user message:

{
  "type": "message",
  "id": "2d8f982d",
  "timestamp": "2026-04-28T22:43:46.836Z",
  "message": {
    "role": "user",
    "content": [
      {
        "type": "text",
        "text": ""
      }
    ]
  }
}

This is immediately followed by MiniMax rejecting it:

embedded run agent end: isError=true model=MiniMax-M2.7 provider=minimax
error=LLM request rejected: invalid params, messages must not be empty (2013)

OpenClaw then:

1. Falls back to openai/gpt-5.4-mini (model fallback, reason=format)
2. Self-repairs the session file ("rewrote 1 assistant message(s), dropped 1 blank user message(s)")
3. A second repair pass drops another blank user message

Log Evidence

23:43:22 [ws] sessions.patch 105ms
23:43:47 [agent/embedded] embedded run agent end: isError=true model=MiniMax-M2.7 error=invalid params, messages must not be empty (2013)
23:43:48 [model-fallback/decision] decision=candidate_failed requested=minimax/MiniMax-M2.7 next=openai/gpt-5.4-mini
23:43:53 [agent/embedded] session file repaired: rewrote 1 assistant message(s), dropped 1 blank user message(s)
23:44:10 [model-fallback/decision] decision=candidate_succeeded candidate=openai/gpt-5.4-mini
23:44:17 [agent/embedded] session file repaired: dropped 1 blank user message(s)

Impact

• User-visible: Confusing experience — model fails, fallback kicks in on a different model, then appears to "recover". User sees an unexpected model responding.
• Self-repair masks the bug: The blank message is cleaned up automatically, so the session eventually works, but the root cause (blank message injection on reset) persists.
• Provider-specific: MiniMax rejects empty messages; other providers (OpenAI, Anthropic) may silently accept them, hiding the issue.

Environment

• OpenClaw: 2026.4.26 (be8c246)
• Node: v25.6.1
• OS: Ubuntu 24.04 (Linux 6.8.0-110-generic, x64)
• Primary model: minimax/MiniMax-M2.7 (via Ollama cloud proxy)
• Fallback model: openai/gpt-5.4-mini
• UI: Control UI (webchat)

Suggested Fix

Prevent sessions.patch (or the session reset codepath) from injecting a user message with empty content. Alternatively, filter out blank user messages before sending to the LLM provider, rather than relying on post-failure self-repair.