Description

Independent agents (created via openclaw agents add) receive empty responses from OpenAI-compatible proxy providers (e.g., yunwu-openai), while the same provider/model works correctly via sessions_spawn from the main agent.

Environment

• OpenClaw version: 2026.2.25 (4b5d4a4)
• OS: macOS (Darwin 25.2.0 arm64)
• Node: v24.13.0
• Provider: yunwu-openai (OpenAI-compatible proxy for Gemini models)
• Models tested: gemini-3.1-pro-preview, gemini-3-flash-preview-nothinking

Steps to Reproduce

1. Create an independent agent:

   openclaw agents add thoth --model "yunwu-openai/gemini-3.1-pro-preview" \
     --workspace ~/.openclaw/agents/thoth/workspace \
     --agent-dir ~/.openclaw/agents/thoth/agent --non-interactive

2. Configure Feishu account for the agent, add binding with match.accountId
3. Copy models.json from main agent to thoth agent dir
4. Send a message to the agent via Feishu

Expected Behavior

Agent responds with model-generated content.

Actual Behavior

Agent runs successfully (isError=false) but produces replies=0. Session JSONL shows:

{
  "role": "assistant",
  "content": [],
  "provider": "yunwu-openai",
  "model": "gemini-3.1-pro-preview",
  "usage": {
    "input": 0,
    "output": 0
  },
  "stopReason": "stop"
}

Root Cause Analysis

Direct curl test to the same API returns valid content:

{
  "choices": [{"message": {"role": "assistant", "content": "Hello! How can I help you today?"}}],
  "usage": {
    "prompt_tokens": 4,
    "completion_tokens": 222,
    "total_tokens": 226,
    "input_tokens": 0,
    "output_tokens": 0
  }
}

The provider returns two sets of token counts:

• prompt_tokens / completion_tokens: correct values (4 / 222)
• input_tokens / output_tokens: always 0

OpenClaw appears to read input_tokens / output_tokens (both 0) and may interpret this as an empty/invalid response, discarding the actual content.

Key Observations

1. sessions_spawn with the same model works — the main agent can successfully call yunwu-openai/gemini-3.1-pro-preview via subagent
2. deepseek/deepseek-chat works for independent agents — switching the model to DeepSeek produces normal responses
3. All yunwu-openai models fail for independent agents (tested gemini-3.1-pro, gemini-3-flash)
4. The API itself is functional (confirmed via curl)

Workaround

Use non-OpenAI-compatible providers (e.g., deepseek/deepseek-chat, zhihui/claude-opus-4-6) for independent agents.

Related Issues

• GLM-5 model switching fails for main agent (works for other agents) #15716 (GLM-5 model switching fails — "Only 5 tokens sent, suggests malformed/empty request")
• Agents silently mutate openclaw.json config (model/auth settings changed without user consent) #24237 (Agents silently mutate openclaw.json config)