OpenAI /v1/chat/completions: tool-call wire-format support (parity with /v1/messages)

[pekkah]

Background

`ChatCompletionRequest` has no `tools` / `tool_choice` field today, so the OpenAI endpoint silently drops any tool definitions the client sends and the model never sees them. `/v1/messages` (Anthropic) has full tool support; `/v1/chat/completions` does not.

OpenAI wire shape:

```json
{
"model": "...",
"messages": [...],
"tools": [{"type":"function","function":{"name":"get_weather","description":"...","parameters":{...}}}],
"tool_choice": "auto"|"none"|{"type":"function","function":{"name":"..."}}
}
```

Response:
```json
{
"choices":[{
"message":{
"role":"assistant",
"content": null,
"tool_calls":[{"id":"call_abc","type":"function","function":{"name":"...","arguments":"{...}"}}]
},
"finish_reason":"tool_calls"
}]
}
```

Streaming variant emits incremental `tool_calls` deltas (each `delta.tool_calls[i]` is an array element addressed by index).

History side: a `role:"tool"` message with `{tool_call_id, content}` echoes a prior result.

Scope

1. Add `Tools`, `ToolChoice` to `ChatCompletionRequest`. Plus matching `tool_calls` field on `OaiAssistantMessage` and a new `ToolCall` record in the response shape. Register types in `SharpInferenceJsonContext`.
2. Render tool definitions into the Jinja context (`tools` key) on the prompt side, same path the Anthropic endpoint already uses via `ChatTemplateRenderer.Format(dict)`.
3. Use the new tool-call adapter (issue Pluggable tool-call translation layer: Anthropic wire format ↔ model-specific syntax #96) to parse model output → emit OpenAI's `tool_calls` array shape and set `finish_reason: "tool_calls"` when calls were produced.
4. Accept history-side `role:"tool"` messages in `BuildMessageList` and pass them through to the template.
5. Streaming: emit `delta.tool_calls` increments per chunk, matching OpenAI's spec (incremental name + JSON-fragment arguments).
6. Tests: parallel coverage to the Anthropic suite (`ChatCompletion_WithTools_NonStreaming`, streaming variant, history-echo).

Acceptance

• OpenAI-protocol tool-calling client (e.g. `openai` Python SDK, LiteLLM proxy) can do a tool-use roundtrip against `/v1/chat/completions`.
• Streaming variant emits incremental `tool_calls` deltas in the documented shape.
• Shares the parser layer from Pluggable tool-call translation layer: Anthropic wire format ↔ model-specific syntax #96 — no duplicate Qwen3 / Qwen3-Coder / Llama branching here.
• Existing OpenAI tests + Anthropic tool tests stay green.

Related

• Pluggable tool-call translation layer: Anthropic wire format ↔ model-specific syntax #96 — pluggable tool-call adapter (this depends on the parsing side of that)
• `src/SharpInference.Server/Endpoints/OpenAiEndpoints.cs` — request/response shape, endpoint flow
• `src/SharpInference.Server/Endpoints/AnthropicEndpoints.cs` — reference impl for the Anthropic side
• OpenAI tool-calling spec: https://platform.openai.com/docs/guides/function-calling

[pekkah]

Closed by 6fa096d — OpenAiEndpoints gained Tools / ToolChoice on ChatCompletionRequest, tool_calls on responses, and parses model output via the shared adapter layer from #96. Streaming emits incremental tool_calls deltas per the OpenAI spec.