Bug Report: Model Configuration Inconsistency

Summary

Models can be added to openclaw.json config but cannot be switched to via session_status tool due to undocumented allowlist restriction.

Environment

• OpenClaw Version: 2026.2.3-1
• OS: Linux 6.8.0-94-generic (x64)
• Node: v22.22.0

Steps to Reproduce

1. Add a model to config using gateway config.patch:

{
  "agents": {
    "defaults": {
      "models": {
        "venice/venice-uncensored": {}
      }
    }
  }
}

2. Verify model appears in config:

grep -A 3 "venice-uncensored" ~/.openclaw/openclaw.json

Result: Model is present in config ✅

3. Attempt to switch to the model using session_status:

session_status(model="venice/venice-uncensored")

Expected Behavior

Option A (Permissive):

• If model is in config, session_status should accept it
• Model switch succeeds

Option B (Restrictive):

• If model is not on allowlist, config.patch should reject it with clear error
• User knows immediately which models are supported

Actual Behavior

• Config accepts the model without error ❌
• Gateway restarts successfully ❌
• session_status rejects with: Model "venice/venice-uncensored" is not allowed. ❌
• No documentation about which models are approved ❌
• No way to discover the allowlist ❌

Approved Models (discovered through testing)

Only these two models work:

1. anthropic/claude-sonnet-4-5
2. openrouter/google/gemini-2.5-flash

All other attempts (Venice, Llama, other OpenRouter models) fail with "not allowed" error.

Config State vs Runtime State

Config says: Model is configured and available

"models": {
  "anthropic/claude-sonnet-4-5": { "alias": "sonnet" },
  "openrouter/google/gemini-2.5-flash": { "alias": "flash" },
  "venice/venice-uncensored": {}
}

Runtime says: Model is not allowed

Error: Model "venice/venice-uncensored" is not allowed.

Impact

User Experience:

• Confusing: "I added it to config, why doesn't it work?"
• Time-wasting: Multiple restart cycles trying to diagnose
• False promise: OpenClaw appears model-agnostic but isn't in practice

Functional:

• Cannot access other models even with valid API keys
• Limits legitimate use cases (testing, research, different capabilities)
• Undermines "multi-model" value proposition

Suggested Fixes

Short-term:

1. Document the allowlist clearly in config schema
2. Add validation to config.patch that rejects unsupported models with helpful error
3. Provide command to list approved models: openclaw models list --allowed

Long-term:

1. Remove allowlist restriction (trust user's config choices)
2. Or make allowlist configurable via security policy setting
3. Or provide clear pathway to add models to allowlist

Additional Context

Tested models that failed:

• meta-llama/llama-3.1-405b-instruct
• meta-llama/llama-3.3-70b-instruct
• venice/venice-uncensored
• venice/hermes-3-llama-3.1-405b
• venice/llama-3.3-70b

All were added to config successfully but rejected at runtime.

Reproducibility

Consistent across multiple attempts, multiple models, multiple config patches.

Related Files:

• Config: ~/.openclaw/openclaw.json
• Models definition: ~/.openclaw/agents/main/agent/models.json
• Tool: session_status (built-in)

Expected vs Actual Matrix:

Environment Details:

🦞 OpenClaw 2026.2.3-1 (d84eb46)
Runtime: agent=main | host=openclaw-01 | os=Linux 6.8.0-94-generic (x64) | node=v22.22.0
Default model: anthropic/claude-sonnet-4-5
Configured providers: anthropic (token), openrouter (token), venice (token)