Summary

As of the 2026.4.x release, OpenClaw refuses to use the ollama API type for any provider that isn't the built-in Ollama URL. Custom provider entries in openclaw.json with "api": "ollama" now fail immediately with:

No API provider registered for api: ollama

This worked in prior releases (confirmed working as of 2026.3.24). We run a local proxy (toolbridge) that implements the native Ollama /api/chat wire format — api: "ollama" was the correct and natural configuration for it. That is now broken.

What api: "ollama" gave us

• think: false injected at the native /api/chat level — the only place Ollama honours it for Qwen3.x models. The OpenAI-compat endpoint silently ignores it, causing multi-minute thinking loops.
• options.num_ctx set automatically from contextWindow — no duplication in the Modelfile required.
• Proper Ollama streaming format, tool passing, etc.

The fallback paths are all broken too

Switching to openai-completions doesn't help. The injectNumCtxForOpenAICompat flag — the only other mechanism to get options.num_ctx injected — is gated behind isOllamaCompatProvider(), which hardcodes port 11434 or the literal provider id "ollama":

function isOllamaCompatProvider(model) {
    if (providerId === "ollama") return true;
    if (hostname === "localhost" && parsed.port === "11434") return true;
    return providerHintsOllama && isOllamaPort && isOllamaCompatPath;
}

Any proxy on any other port, with any other name, is excluded. So:

• api: "ollama" → rejected outright
• api: "openai-completions" + injectNumCtxForOpenAICompat: true → silently does nothing
• api: "openai-responses" → same; no num_ctx, no think suppression

The only remaining option is to bake PARAMETER num_ctx into the Modelfile — which defeats the point of having runtime-configurable context windows.

Requested fixes

1. Restore api: "ollama" support for arbitrary baseUrl providers. A proxy that implements /api/chat is Ollama-compatible regardless of what port it runs on. The operator has already opted in by setting api: "ollama".

2. If keeping the restriction, make injectNumCtxForOpenAICompat: true act as an explicit override for isOllamaCompatProvider(). If the operator has set that flag, they know what they're doing.

3. Document the change. This appears to be a silent regression in the 2026.4.x window with no mention in the changelog.

Environment

• lastTouchedVersion: 2026.4.5
• lastRunVersion (wizard): 2026.3.24
• Platform: Linux/WSL2
• Local proxy: toolbridge (Python aiohttp, implements /api/chat and /v1/chat/completions)