LLM provider used that I not have configured for agent. | Friends of the Crustacean 🦞🤝 | Page 1

stiff yew Apr 29, 2026, 8:14 PM

#

I have a Primary LLM minimax m2.7 and fallback m2.5. for hearbeats I have same but other way around. But I see that I have 2 agents that have used Claude Pus provider. I do have this provider in my openclaw.json file but not configured for agent work. I cannot find why claude llm is used.

sinful mirageBOT Apr 29, 2026, 8:14 PM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

nocturne pulsarBOT Apr 29, 2026, 8:17 PM

#

First check: what does openclaw --version show?

Most likely places Claude/Anthropic can sneak in:

Session override — /model or Control UI model picker persists per session. Check the same chat/thread with:
```
/model status
/status
```
And globally:
```
openclaw sessions --all-agents --json
```
Sub-agents inherit caller model unless configured:

#

   agents: {
     defaults: {
       subagents: {
         model: "minimax/...",
       },
     },
   }

Heartbeat model settings do not control sub-agent model choice. Docs: https://docs.openclaw.ai/tools/subagents

Heartbeat has only a single model override:
```
agents: { defaults: { heartbeat: { model: "minimax/..." } } }
```
It is not a separate primary/fallback object. If fallback happens, check normal agents.defaults.model.fallbacks / per-agent fallbacks. Docs: https://docs.openclaw.ai/gateway/heartbeat

#

Claude in agents.defaults.models is only an allowlist/catalog, not active routing by itself. Active routing is agents.defaults.model.primary, fallbacks, per-agent agents.list[].model, session overrides, cron payloads, and sub-agent settings.
Agent runtime — check for:
```
agentRuntime: { id: "claude-cli" }
```
under agents.defaults or agents.list[].

Useful safe checks:

openclaw models status
openclaw config get agents.defaults.model
openclaw config get agents.defaults.heartbeat
openclaw config get agents.defaults.subagents
openclaw config get agents.list

#

openclaw logs --limit 1000 --plain | grep -Ei "anthropic|claude|fallback|modelOverride"

If you paste openclaw --version plus redacted output from those, I can help pinpoint exactly which layer selected Claude. Docs: https://docs.openclaw.ai/concepts/models / https://docs.openclaw.ai/concepts/model-failover

stiff yew Apr 29, 2026, 8:20 PM

#

"key": "agent:neon:telegram:direct:@neonnajef",
"updatedAt": 1777228378141,
"ageMs": 265543368,
"sessionId": "beebe403-f080-4815-9648-3fcb33c5aa08",
"totalTokens": null,
"totalTokensFresh": false,
"model": "gpt-5.4",
"contextTokens": 200000,
"agentId": "neon",
"kind": "direct"
},
{
"key": "agent:neon:telegram:group:@najef",
"updatedAt": 1777225480629,
"ageMs": 268440880,
"sessionId": "fa124050-395d-4ce9-8f6a-c6608c962968",
"totalTokens": null,
"totalTokensFresh": false,
"model": "gpt-5.4",
"contextTokens": 200000,
"agentId": "neon",
"kind": "group"
I do not even have configured OpenAI token or oath,

nocturne pulsarBOT Apr 29, 2026, 8:20 PM

#

That snippet is a session record, not proof that OpenAI was successfully called.

A few important bits:

"model": "gpt-5.4" can be a stored/stale session model value from when that session was created or selected.
It may also be only the model id, with provider omitted in that display. To know the actual provider, look for modelProvider, providerOverride, and modelOverride.
If you have no OpenAI auth configured, OpenClaw shouldn’t be able to complete an OpenAI call unless some compatible/provider alias is involved or it falls back elsewhere.

For these two sessions, try resetting the stale model state:

openclaw sessions --all-agents --json

Then in those Telegram chats/threads, send:

#

/reset
/model status

Also check the neon agent specifically:

openclaw config get agents.list
openclaw config get agents.defaults.model
openclaw config get agents.defaults.heartbeat
openclaw config get agents.defaults.subagents

The likely cause is one of:

agent:neon has its own agents.list[].model
the session has an old /model override

#

a subagent inherited/overrode model separately
heartbeat only set heartbeat.model, while normal runs still used another model/fallback

Please also paste openclaw --version — that matters a lot here. Docs: https://docs.openclaw.ai/concepts/models

stiff yew Apr 29, 2026, 8:27 PM

#

openclaw config get agents.list Only shows Minimax.

openclaw config get agents.defaults.model
"primary": "",
"fallbacks": [
"minimax/MiniMax-M2.5"

openclaw config get agents.defaults.heartbeat
"every": "25m",
"model": "minimax/MiniMax-M2.5",
"target": "none",
"lightContext": true

openclaw config get agents.defaults.subagents
"maxConcurrent": 8

OpenClaw 2026.4.22 (00bd2cf)

#

openclaw config get agents.defaults.subagents

najef@OpenClaw-Server:~$ openclaw --version
OpenClaw 2026.4.22 (00bd2cf)
najef@OpenClaw-Server:~$ ^C
najef@OpenClaw-Server:~$ openclaw logs --limit 1000 --plain | grep -Ei "anthropic|claude|fallback|modelOverride"
Gateway not reachable. Is it running and accessible?
Gateway target: ws://127.0.0.1:18789
Source: local loopback
Config: /home/najef/.openclaw/openclaw.json
Bind: loopback
Hint: run openclaw doctor.

nocturne pulsarBOT Apr 29, 2026, 8:28 PM

#

Aha, the smoking gun is:

"primary": ""

An empty primary is treated as unset, so OpenClaw falls back to its built-in default model — in your version that can show up as gpt-5.4. Your MiniMax entry is only a fallback, not the first model.

Set the primary explicitly:

openclaw config set agents.defaults.model.primary "minimax/MiniMax-M2.7"
openclaw config set agents.defaults.model.fallbacks '["minimax/MiniMax-M2.5"]' --strict-json
openclaw gateway restart

Then reset the stale sessions/chats:

#

/reset
/model status

Also, since subagents has no model configured, subagents inherit whatever the caller/session resolved to. If you want them forced to MiniMax too:

openclaw config set agents.defaults.subagents.model "minimax/MiniMax-M2.7"

For those exact sessions, check the provider too:

openclaw sessions --all-agents --json

#

Look for modelProvider, providerOverride, modelOverride near those agent:neon... entries. The old gpt-5.4 is probably stale session state from before the MiniMax config.

#

Yep — now the immediate blocker is simpler: your CLI can’t reach the Gateway, so openclaw logs can’t inspect the live run history.

Run these next:

openclaw status --all
openclaw gateway status
openclaw doctor

If the gateway is just stopped:

openclaw gateway start

or:

#

openclaw gateway restart

Then re-run:

openclaw logs --limit 1000 --plain | grep -Ei "anthropic|claude|fallback|modelOverride|gpt-5.4|minimax"

But still fix this config first:

"primary": ""

That means “no primary configured”, so OpenClaw can resolve to the built-in default/stale session model. Set it explicitly:

#

openclaw config set agents.defaults.model.primary "minimax/MiniMax-M2.7"
openclaw config set agents.defaults.model.fallbacks '["minimax/MiniMax-M2.5"]' --strict-json
openclaw config set agents.defaults.subagents.model "minimax/MiniMax-M2.7"

Then restart gateway and /reset the affected chats.

stiff yew Apr 29, 2026, 8:30 PM

#

Where does it get gpt-5.4? I do not have configured this model.. I cannot believe it just picks models that are not even configured

nocturne pulsarBOT Apr 29, 2026, 8:31 PM

#

I get why that feels scary, but it’s not inventing a random model from your config.

gpt-5.4 is coming from OpenClaw’s built-in default model fallback path because this is effectively unset:

agents.defaults.model.primary: ""

So the chain becomes:

primary is empty → “no primary configured”
OpenClaw uses its internal default primary, in your version showing as gpt-5.4
then your configured fallbacks are considered after that

Your config currently says: “no primary, fallback to MiniMax M2.5” — not “primary MiniMax M2.7”.

Set the primary explicitly:

#

openclaw config set agents.defaults.model.primary "minimax/MiniMax-M2.7"
openclaw config set agents.defaults.model.fallbacks '["minimax/MiniMax-M2.5"]' --strict-json
openclaw gateway restart

Then reset affected sessions:

/reset

If you want no subagent inheritance surprises:

openclaw config set agents.defaults.subagents.model "minimax/MiniMax-M2.7"

#

And yes: without OpenAI auth it should not be able to successfully call OpenAI. The session list can still display the resolved default model. The important fix is: don’t leave primary empty. Docs: https://docs.openclaw.ai/concepts/models

stiff yew Apr 29, 2026, 8:32 PM

#

why is the fallback command line different that the primary one?

nocturne pulsarBOT Apr 29, 2026, 8:32 PM

#

An error occurred while processing your request. You can retry your request, or contact us through our help center at help.openai.com if the error persists. Please include the request ID req_e59c6ae9645d45aa95e2064c92ea850a in your message.

stiff yew Apr 29, 2026, 8:32 PM

#

why is the fallback command line different that the primary one?

nocturne pulsarBOT Apr 29, 2026, 8:32 PM

#

Because they’re different data types:

primary: "minimax/MiniMax-M2.7"        // one string
fallbacks: ["minimax/MiniMax-M2.5"]    // array/list of strings

So the CLI has to write them differently:

openclaw config set agents.defaults.model.primary "minimax/MiniMax-M2.7"

…but fallbacks needs JSON array syntax:

#

openclaw config set agents.defaults.model.fallbacks '["minimax/MiniMax-M2.5"]' --strict-json

You can also set the whole model block in one go if that feels cleaner:

openclaw config set agents.defaults.model '{"primary":"minimax/MiniMax-M2.7","fallbacks":["minimax/MiniMax-M2.5"]}' --strict-json

Then restart/reset sessions. Docs: https://docs.openclaw.ai/gateway/configuration

#LLM provider used that I not have configured for agent.