fix(stream) + feat(api): emit Usage before Done; expose model capabilities in /v1/models

[shivam2014]

Two independent fixes:

1. Stream: emit Usage delta before Done

parse_openai_chunk currently pushes Done before Usage. This means format_deltas emits usage={0,0,0} in the chunk before [DONE], and the OpenAI SDK stops at that first [DONE] — the real token counts in the subsequent Usage chunk are never seen by the client.

Fix: reorder so Usage delta precedes Done. Includes a regression test.

2. API: expose model capabilities in /v1/models

The /v1/models endpoint currently returns bare {id, object, created, owned_by}. This adds max_context_length and max_output_tokens by pre-fetching ModelCapabilities from the admin for each route's target model.

{
  "id": "deepseek-v3",
  "object": "model",
  "owned_by": "Nyro",
  "max_context_length": 131072,
  "max_output_tokens": 8192
}

---

Files: stream.rs (+64, -5, 1 test), handler.rs (+41, -9)
Tests: cargo test --workspace --exclude nyro-desktop --exclude nyro-server — 153 passed, 0 failed

[shuaijinchao]

fetch nyro model context is virtual model, it include 1 to many providers real model, theri context length difference, so context length should not in /v1/models api.
what do you think ?

[shivam2014]

fetch nyro model context is virtual model, it include 1 to many providers real model, theri context length difference, so context length should not in /v1/models api. what do you think ?

Fair point. Instead of dropping the field entirely, what about reporting the minimum context_window across all routes for that virtual model? Deterministic, safe for every route, and still useful to clients.

[shuaijinchao]

good idea.
bro, you can add me wechat: xincps

[shivam2014]

good idea. bro, you can add me wechat: xincps

Appreciate the gesture. Unfortunately I don't have WeChat. Happy to connect on X if you prefer. Link on my GitHub bio. Also very happy with your router. Great job !