Add tool-calling support with emulation for providers without native tool calling

[neuwcodebox]

Motivation

• Enable clients to send tools and tool_choice in chat completion requests and support tool-calling flows even when a provider lacks native tool calling.
• Provide a consistent emulation mode so providers can participate in tool workflows by exchanging special [tool-call]/[tool-result] messages.

Description

• Add tool-related types and validation to src/openai/schemas.ts, including tools, tool_choice, tool_calls, and a role: 'tool' message shape.
• Implement emulation logic in src/openai/tool-emulation.ts with functions to detect emulation (shouldEmulateTools), transform inbound messages (toProviderMessages), parse tool-call payloads, apply emulation responses (applyToolEmulationResponse), and convert streamed protocol JSON into structured tool_calls chunks (transformToolEmulationStream).
• Integrate emulation into the server by resolving provider capabilities and routing through toProviderMessages, transformToolEmulationStream for streams, and applyToolEmulationResponse for non-streamed responses in src/server.ts.
• Expose provider capability via supportsNativeToolCalling in src/providers/types.ts and mark the dummy provider as not supporting native tool calling while adding simple emulation-aware behavior in src/providers/dummy.ts.
• Update README.md with usage notes and an example of the emulation protocol.

Testing

• Added unit/integration tests in test/tool-emulation.test.ts and extended test/server.test.ts to cover schema validation, emulated tool call generation, plain-text passthrough, tool-result handling, and stream behavior, and executed the test suite with npm test (via vitest).
• All automated tests completed successfully.

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7fe19f3644

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Accept null assistant content when tool_calls are present

The assistant-message schema currently forces content to be a non-empty string/array, but this same change emits content: null when a tool call is detected (applyToolEmulationResponse). In a normal tool-calling loop, clients replay that assistant tool-call message on the next request; with this schema branch, the follow-up request is rejected with 400 before provider handling, so emulated multi-turn tool calling cannot continue. Please allow content: null for assistant messages that carry tool_calls.

Useful? React with 👍 / 👎.