Custom registry ignores advertised limit.output — providers reject requests with 400 "Invalid max_tokens"

[gfunc]

What version of Kimi Code is running?

0.42.0 (also verified present on latest main, commit b180725)

Which open platform/subscription were you using?

A custom registry (api.json, models.dev format) with OpenAI/Anthropic-compatible providers.

Which model were you using?

Any model imported from a custom registry whose advertised limit.output is below its context size (e.g. models with an 8k output cap).

What platform is your computer?

Linux x86_64 (platform-independent — the issue is in how registry data is imported).

What issue are you seeing?

When models are imported from a custom registry (api.json), each entry's advertised limit.output is ignored: the generated model alias never records maxOutputSize. The request's completion-token cap then falls back to the model's full context size and is sent as the protocol's output-token field — max_output_tokens for openai_responses, max_completion_tokens/max_tokens for openai chat-completions, and max_tokens for anthropic (which additionally falls back to a 128000 default for unrecognized model ids). Any registry model whose real output cap is smaller rejects the request, e.g.:

400 "Invalid max_output_tokens"

What steps can reproduce the bug?

1. Configure a custom registry whose model entry advertises a modest output limit, e.g. an openai_responses registry with this api.json entry:

   {
     "models": {
       "example-model": {
         "id": "example-model",
         "name": "Example Model",
         "limit": { "context": 1000000, "output": 8192 }
       }
     }
   }

2. Import that registry and select the model.
3. Send any prompt.
4. The request goes out with max_output_tokens ≈ 1000000 (the full advertised context size, since no output limit was recorded), and the provider responds 400 "Invalid max_output_tokens" because that exceeds the model's real 8192 output cap.

What is the expected behavior?

kimi-code records the registry's advertised limit.output as the model's max output size and sends a compliant output-token cap, the same way limit.context is already honored for maxContextSize.

Additional information

• Verified against the live models.dev dataset (~7700 models): every chat model publishes limit.output as a positive integer, so the field is reliable to consume.
• The models.dev integration path already maps this field to maxOutputSize; the custom-registry import path is the only one that drops it.
• Affects all custom-registry provider types (openai, openai_responses, anthropic, kimi) since the completion-budget cap is protocol-agnostic.
• I am willing to submit a PR for this bug fix myself (waiting for maintainer approval in this issue first, per the contributing policy).