Missing OpenRouter models in Local Mode

[JoseArroyave]

Most OpenRouter models missing in Local Mode

Description

When using OpenRouter in Local Mode, many models retrieved from the OpenRouter API do not appear in the system.

This happens due to two related issues in the pricing sync and caching logic:

1. OpenRouter models are not stored with provider = OpenRouter.
2. The pricing cache assumes model_name is unique, causing collisions when multiple providers share the same model.

Together, these issues prevent many OpenRouter models from appearing in the available models list.

---

Problem 1: Incorrect provider assignment

The pricing sync retrieves models from the OpenRouter API. Example model ID:

openai/gpt-4.1

The provider is derived from the vendor prefix:

canonical: gpt-4.1
provider: OpenAI

The database entry becomes:

model_name | provider
gpt-4.1   | OpenAI

However, users access this model via OpenRouter, not directly through the vendor.

When models are filtered by provider:

activeProviders = ['ollama', 'openrouter']

The check:

activeProviders.has(m.provider.toLowerCase())

evaluates to:

openrouter.has(openai) → false

So the model is excluded even though it is available through OpenRouter.

---

Problem 2: Cache key collision

To properly support OpenRouter, the same model may exist for multiple providers:

model_name | provider
gpt-4.1   | OpenAI
gpt-4.1   | OpenRouter

However, pricingCache.reload() uses only model_name as the cache key:

this.cache.set(row.model_name, row);

When duplicates exist:

1. gpt-4.1 | OpenAI is cached
2. gpt-4.1 | OpenRouter is cached
3. The second entry overwrites the first

Which entry survives depends on query order.

---

Impact

• Many OpenRouter models are missing
• available-models returns incomplete results
• Provider filtering behaves inconsistently
• Local Mode users cannot access most OpenRouter models

---

Affected components

• syncPricing
• syncAllModels
• pricingCache.reload
• available-models endpoint
• provider filtering logic

[SebConejo]

@guillaumegay13 can you work on this ?

[guillaumegay13]

@JoseArroyave I'm a bit confused here. It's not expected to have all the same models than in OpenRouter right now.

Which model are you trying to add/get?
Are models missing within a specific provider (OpenAI) ?
Did you try to add it as custom provider ?

Models pricing are currently aligned on OpenRouter pricing but that'll change as part of #1036.

[JoseArroyave]

@guillaumegay13 I previously routed requests to openai/gpt-oss-120b and qwen/qwen3-235b-a22b-thinking-2507 through OpenRouter, but those models no longer appear in the available models list.

The issue I encountered is not that the models are missing from the pricing source itself, but that models accessible through OpenRouter are not being exposed in Local Mode when OpenRouter is configured as an active provider.

When I connected an OpenRouter API key and checked the /available-models endpoint, I noticed that the models returned by pricingCache.getAll() did not include all models available through OpenRouter, including the two mentioned above. The pricing sync currently derives the provider from the vendor prefix, so no entries exist with provider = OpenRouter, even though those models can be routed through OpenRouter.

Since getAvailableModels() filters models based on the agent’s active providers, if OpenRouter is the only active provider those models are filtered out because they are stored under their vendor provider instead.

Let me know if this aligns with the intended provider model, or if OpenRouter models are expected to be handled differently.

[guillaumegay13]

Thanks @JoseArroyave for the context. I'll reproduce & come back to you!

[guillaumegay13]

I've figured out that if models already exist in other providers than OpenRouter, they're not resolved anymore by OpenRouter. This is why openai/gpt-oss-120b and qwen/qwen3-235b-a22b-thinking-2507 don't appear anymore in the list (already exist in OpenAI & Alibaba providers).

I've proposed a fix to preserve all OpenRouter available models as long as they have text output (required for OpenClaw). PR #1121.

I've reproduced the issue & confirmed that with this fix, I can see a lot more models in OpenRouter providers and in particular openai/gpt-oss-120b and qwen/qwen3-235b-a22b-thinking-2507.

FYI @JoseArroyave @brunobuddy