Support tool calling while streaming Open Responses

[ActuallyTaylor]

The streaming path sent the session's tools but its event loop only handled text deltas and completion, so function-call output items were dropped and tools never fired when streaming.

OpenResponsesStreamEvent now decodes response.output_item.added, response.function_call_arguments.delta, and response.output_item.done. Streamed items accumulate per output index, argument fragments append to the item the server already opened, and a call is surfaced only once its arguments parse as JSON. Assembled calls run through the shared resolveToolCalls; the turn's output items are echoed back into the input list the same way the non-streaming path echoes resp.output, tool outputs follow as .tool messages, and another request starts until a turn ends without calls.

Tool activity reaches the transcript while the stream is running, with the tool calls entry ahead of its outputs, and a .stop decision finishes the stream without executing anything. .failed still throws streamFailed.

Note

Not verified against a live endpoint — this suite needs OPEN_RESPONSES_API_KEY and OPEN_RESPONSES_BASE_URL, which weren't available. It is the one provider in this series still unproven end to end.

Based on #4; merge that first.

🤖 Generated with Claude Code

[ActuallyTaylor]

Removed from the stack (now stack huggingface#15) and left open as a standalone PR against main, to be picked up after the schema issue below is resolved.

Status: the streaming implementation here is not verified end to end. Testing it against Ollama's Responses API surfaced a blocking bug that is not in this PR's streaming code:

Tool parameters are sent with a $defs block. GenerationSchema.withResolvedRoot() inlines the root schema but leaves the now-unreferenced definitions behind. Replaying the exact request two ways against Ollama:

with $defs → ['reasoning', 'message']        ← no tool call
without    → ['reasoning', 'function_call']  ← tool called

The model is reliable — 5/5 tool calls once $defs is dropped — so the server silently declines to call a tool whose parameters carry it.

This is pre-existing and shared: withResolvedRoot() lives in GenerationSchema.swift, and OpenAI, Ollama, Gemini, MLX, Anthropic and System all use the same pattern. It goes unnoticed against OpenAI's own API, which tolerates $defs.

The fix belongs in withResolvedRoot() — after inlining the root, drop defs no longer reachable from it. Deleting $defs wholesale would break nested @Generable types that genuinely $ref into it.

Non-tool paths of this PR do pass against Ollama (basicResponse, streaming, streamingString, withInstructions).

🤖 Generated with Claude Code

[ActuallyTaylor]

Chilling. Same format as the rest of the implementations