Summary

OpenClaw gateway on macOS/launchd still fails to register the Ollama provider at runtime even after:

• removing models.providers.ollama from ~/.openclaw/openclaw.json
• setting OLLAMA_API_KEY=ollama-local
• restarting the gateway cleanly

Startup logs show the session/agent model as ollama/qwen3.5:4b, but live requests still fail with Unknown model: ollama/qwen3.5:4b and silently fall back to openai-codex/gpt-5.4.

Environment

• OpenClaw: 2026.3.31
• OS: macOS (LaunchAgent / launchd)
• Gateway label: ai.openclaw.gateway
• Ollama local API working directly outside gateway
• Local test model: ollama/qwen3.5:4b

Expected behavior

If:

1. models.providers.ollama is absent (implicit behavior)
2. LaunchAgent env contains OLLAMA_API_KEY=ollama-local
3. gateway is restarted

then ollama/qwen3.5:4b should register and serve requests without fallback.

Actual behavior

Gateway startup shows Ollama as the configured model:

[lcm] Compaction summarization model: ollama/qwen3.5:4b (default)
agent model: ollama/qwen3.5:4b

But live requests fail immediately:

FailoverError: Unknown model: ollama/qwen3.5:4b. Ollama requires authentication to be registered as a provider. Set OLLAMA_API_KEY="ollama-local" (any value works) or run "openclaw configure".

and then:

Model "ollama/qwen3.5:4b" not found. Fell back to "openai-codex/gpt-5.4".

Repro steps

1. Ensure local Ollama is running and the model is available directly.
2. Remove models.providers.ollama from ~/.openclaw/openclaw.json.
3. Set LaunchAgent env:
   • OLLAMA_API_KEY=ollama-local
4. Restart gateway.
5. Request ollama/qwen3.5:4b in a live session.
6. Inspect gateway log.

Evidence

Config state

• models.providers no longer contains ollama
• active provider keys in config: anthropic, google, minimax-portal

LaunchAgent env

Persisted plist env now contains:

"OLLAMA_API_KEY" => "ollama-local"

Fresh startup log

2026-04-05T01:51:47 [lcm] Compaction summarization model: ollama/qwen3.5:4b (default)
2026-04-05T01:51:48 agent model: ollama/qwen3.5:4b

Fresh failure log after restart

2026-04-05T01:52:27 lane task error: lane=main durationMs=858 error="FailoverError: Unknown model: ollama/qwen3.5:4b. Ollama requires authentication to be registered as a provider. Set OLLAMA_API_KEY=\"ollama-local\" (any value works) or run \"openclaw configure\". See: https://docs.openclaw.ai/providers/ollama"

2026-04-05T01:52:27 model_fallback_decision: candidate_failed requestedProvider=ollama requestedModel=qwen3.5:4b reason=model_not_found nextCandidateProvider=openai-codex nextCandidateModel=gpt-5.4

2026-04-05T01:52:31 Model "ollama/qwen3.5:4b" not found. Fell back to "openai-codex/gpt-5.4".

Notes

This may be one of:

• LaunchAgent env not being used by Ollama provider registration path
• wrong env key being checked by provider bootstrap
• implicit Ollama provider registration not occurring in gateway runtime on macOS launchd
• session/agent startup path believing the model exists before provider registration actually succeeds

Impact

This is confusing because session_status / startup logs can imply Ollama is active, while actual serving silently falls back to a cloud model.