GitHub Issue Draft: GLM-5 Model Switching Fails for Main Agent

Summary

GLM-5 (zai/glm-5) works perfectly when configured as an agent's model in agents.list, but switching the main agent to GLM-5 fails via both /model command and config.patch.

Environment

• OpenClaw version: 2026.2.9 (33c75cb)
• Platform: macOS (Darwin 25.3.0 arm64)
• Provider: ZAI (api.z.ai)
• Model: GLM-5

What Works ✅

Stilgar agent (configured in agents.list with zai/glm-5):

• Responds normally
• Shows correct model in session history: "api": "openai-completions", "provider": "zai", "model": "glm-5"
• Handles 35K+ token contexts without issue

Direct API test (curl to ZAI API):

curl -X POST https://api.z.ai/api/coding/paas/v4/chat/completions \
  -H "Authorization: Bearer $ZAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5",
    "messages": [{"role": "user", "content": "Hello"}],
    "stream": true
  }'

✅ Works perfectly, returns reasoning_content in stream

What Fails ❌

Method 1: /model Command

User: /model glm
System: Model switched to glm (zai/glm-5).
User: hello
[No response]

Logs:

2026-02-13T18:55:09.350Z No reply from agent.

Session status after switch:

🧠 Model: zai/glm-5
🧮 Tokens: 5 in / 412 out

Only 5 tokens sent — suggests malformed/empty request.

Method 2: config.patch Agent Model

Changed agents.list main agent model from anthropic/claude-opus-4-6 to zai/glm-5:

{
  "agents": {
    "list": [
      {"id":"main","model":"zai/glm-5"},
      {"id":"leto","workspace":"...","model":"openrouter/minimax/minimax-m2.5"},
      {"id":"stilgar","workspace":"...","model":"zai/glm-5"}
    ]
  }
}

Result: Config saves, gateway restarts (SIGUSR1), but session_status still shows:

🧠 Model: anthropic/claude-sonnet-4-5

Even after full openclaw gateway stop && openclaw gateway start, session stays on Sonnet.

Configuration

ZAI Provider Config

{
  "models": {
    "providers": {
      "zai": {
        "baseUrl": "https://api.z.ai/api/coding/paas/v4",
        "apiKey": "...",
        "api": "openai-completions",
        "models": [
          {
            "id": "glm-5",
            "name": "GLM 5",
            "reasoning": true,
            "input": ["text"],
            "contextWindow": 128000,
            "maxTokens": 16384
          }
        ]
      }
    }
  }
}

Model Alias

{
  "agents": {
    "defaults": {
      "models": {
        "zai/glm-5": {
          "alias": "glm"
        }
      }
    }
  }
}

Hypothesis

1. /model command accepts the switch but fails to initialize the ZAI provider correctly for the main session
2. config.patch updates the config file but SIGUSR1 hot reload doesn't pick up agent model changes
3. Something about the main session preserves its original model binding across restarts

Tested Parameters

Verified ZAI API accepts all these parameters (none cause 400):

• ✅ tools
• ✅ stream_options
• ✅ reasoning_effort
• ✅ prediction
• ✅ Anthropic cache control markers

Expected Behavior

Switching the main agent to zai/glm-5 via either /model glm or config.patch should work the same way it does for Stilgar's agent config.

Workaround

None found. Only way to use GLM-5 is to create a separate agent with the model pre-configured in agents.list.

Additional Context

• Stilgar's agent has been running on zai/glm-5 successfully for days
• Same ZAI API key, same provider config
• Main agent defaults to openrouter/auto in agents.defaults.model.primary — might be fallback related?