openclaw - 💡(How to fix) Fix Bug: New session fails with anthropic-messages API models due to empty messages array [1 comments, 2 participants]

Description

When a new session starts with a model using the anthropic-messages API format (e.g., minimax-M2.7 via minimax-portal), the LLM request fails with "messages must not be empty". The fallback model (using openai-completions format) then succeeds.

Root Cause

During new session initialization, OpenClaw constructs the LLM request with:

• messages: [] (empty array)
• prompt: "" (empty string)
• systemPrompt: populated (52KB in our case)

The Anthropic Messages API requires at least one message in the messages array, causing minimax to reject the request with error code 2013.

Models using openai-completions API format (e.g., xiaomi/mimo-v2.5) do not have this requirement and succeed with empty messages.

Evidence

1. Trajectory Data (prompt.submitted)

{
  "messages": [],
  "prompt": "",
  "systemPrompt length": 52110
}

2. Gateway Logs

11:39:48 [agent/embedded] embedded run agent end: runId=a3db82ec 
  model=MiniMax-M2.7 provider=minimax-portal 
  error=LLM request rejected: invalid params, messages must not be empty (2013)

11:39:49 [model-fallback/decision] model fallback decision: 
  decision=candidate_failed requested=minimax-portal/MiniMax-M2.7 
  reason=format next=xiaomi/mimo-v2.5

11:39:54 [agent/embedded] session file repaired: 
  rewrote 1 assistant message(s), dropped 1 blank user message(s)

3. Session File Corruption

Session files contain blank user messages created during initialization:

{"type": "message", "message": {"role": "user", "content": [{"type": "text", "text": ""}]}}

The repair mechanism (repairUserEntryWithBlankTextContent) detects and removes these, but only AFTER the LLM call has already failed.

Reproduction

1. Configure a model with "api": "anthropic-messages" as the primary model
2. Start a new session (e.g., via /new command or first message in a group chat)
3. Observe the "messages must not be empty" error
4. Session falls back to the first openai-completions model in fallbacks list

Why Manual /model Switch Works

When a user manually sends /model minimax-portal/MiniMax-M2.7:

1. The session already exists with conversation history
2. The /model command itself becomes a user message in the messages array
3. messages is not empty → minimax succeeds

Impact

• Users cannot use minimax-M2.7 (or any anthropic-messages model) as the default primary model
• The model only works after manual /model switch in an existing session
• This affects any provider using the anthropic-messages API format

Suggested Fix

Ensure that when constructing the initial LLM request for a new session, the user message (the first message that triggered the session) is included in the messages array before making the API call.

The issue appears to be in the session initialization flow where context.compiled produces empty messages and prompt fields.

Environment

• OpenClaw version: 2026.4.26 (be8c246)
• Primary model: minimax-portal/MiniMax-M2.7 (api: "anthropic-messages")
• Fallback: xiaomi/mimo-v2.5 (api: "openai-completions")
• Channel: Feishu group chat