•	OpenClaw version: 2026.3.24
•	OS: macOS
•	Hardware: Apple Silicon Mac Mini M4 Pro
•	Deployment mode: LaunchAgent-managed gateway, also reproduced by starting gateway manually in foreground
•	Slack used as primary communication channel
•	Multi-agent system with 37 agents

Current default config block
From ~/.openclaw/openclaw.json:

{
  "model": {
    "primary": "anthropic/claude-sonnet-4-6",
    "fallbacks": []
  },
  "models": {
    "anthropic/claude-sonnet-4-6": {
      "alias": "sonnet"
    },
    "vllm/mlx-community/Qwen3-14B-4bit": {
      "alias": "qwen-mlx"
    },
    "vllm/mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit": {
      "alias": "qwen-coder"
    }
  }
}
This appears clean and does not include any openrouter/... entries.

Critical reproduction test
We opened a brand new terminal tab and verified:
env | grep OPENROUTER
returned nothing.

Then we killed all gateway processes and started the gateway manually in the foreground with:
node entry.js gateway --port 18889

Even in that clean-shell foreground run, Jack still returned:
All models failed (4): openrouter/google/gemini-2.5-flash: HTTP 401 | openrouter/deepseek/deepseek-v3.2: HTTP 401 | openrouter/anthropic/claude-sonnet-4-6: HTTP 401 | openrouter/anthropic/claude-opus-4-6: HTTP 401

Why this seems unexpected
At this point, the OpenRouter model attempts do not appear to be coming from:
	•	current environment variables
	•	visible config files
	•	auth profiles
	•	cache
	•	known fallback chains

That suggests stale runtime/session state, model-selection persistence, or bundled provider catalog behavior that is still surfacing openrouter/... model refs after provider removal.

Additional observation
The installed binary contains hardcoded default model constants such as:
const DEFAULT_GEMINI_MODEL = "gemini-2.5-flash";

I do not know whether that is relevant, but I am including it because the regenerated state may be influenced by built-in provider catalogs/defaults.

Questions for maintainers
	1.	Is there a known bug where stale model/provider state survives complete provider removal?
	2.	Can bundled provider catalogs surface openrouter/... model refs even when the provider is no longer configured?
	3.	Is there a hidden session or runtime state location that should also be cleared?
	4.	Is there a supported way to fully disable or deregister a bundled provider plugin?
	5.	Should deleting per-agent models.json be avoided in this situation?

Reproduction steps
	1.	Start with OpenClaw using OpenRouter
	2.	Migrate to direct Anthropic API
	3.	Remove all visible OpenRouter config/env/auth references
	4.	Confirm defaults only allow Anthropic + local vLLM models
	5.	Start gateway from a clean shell
	6.	Message an affected agent/channel
	7.	Observe attempts to use openrouter/... models anyway