Evaluate free open-source models via OpenRouter

[ndoschek]

Make open-weight models hosted on OpenRouter usable in Theia AI. OpenRouter exposes an
OpenAI-compatible API, so it works today through the existing OpenAI custom endpoint preference
(ai-features.openAiCustom.customOpenAiModels) with no code changes. This sub-issue verifies that
path, tests the candidate models, and documents the setup. Less urgent than Hugging Face, since we
do not have an OpenRouter account yet.

Setup

1. Create an OpenRouter account (https://openrouter.ai/) and get an API key. Give it an expiration
   date and use it only for testing. Do not use sensitive data in the test cases.
2. Add an entry to the ai-features.openAiCustom.customOpenAiModels preference:

"ai-features.openAiCustom.customOpenAiModels": [
  {
    "model": "z-ai/glm-5.2",
    "url": "https://openrouter.ai/api/v1",
    "id": "openrouter-glm",
    "apiKey": "sk-or-...",
    "enableStreaming": true
  }
]

3. Configure an agent to use this model (by its id), then start a chat.

What to test

Look up model ids on the OpenRouter models list (https://openrouter.ai/models). Test the candidate
models and record for each:

• Does a basic chat response work?
• Does tool calling work (try with Coder on a prompt that triggers a tool)?
• Does it need supportsStructuredOutput: false?

Candidate models (confirm exact slugs):

• GLM (z-ai/glm-5.2)
• Gemma 4 (google/gemma-4-31b-it)
• A recent Qwen coding model (e.g. qwen/qwen3-coder, confirm the latest)
• A Kimi model (moonshotai/kimi-k2)

Deliverables

• A working OpenRouter configuration in Theia (the customOpenAiModels entry).
• An assessment of which models work for which agents, covering chat, tool calling, and structured
  output.
• Setup documentation so users can reproduce it.
• Findings reported in the parent issue.

[eneufeld]

I tested GLM5.2 using openrouter and openai api.
My overall feeling was very good especially if you take the cost into account.
I encountered some small errors with the llm trying to call tools and using invalid format. but that only happened 1 or 2 times.
The main thing missing for me was:

• search (which is only needed for tasks that include research)
• compaction
• memory

We would need to provide from Theia and not rely on built in tools. For Coding following a plan this probably would be more than enough. I will definitely try it out more