[NativeTool] Phase 4: Streaming tool_use (deferred)

[bug-ops]

Summary

chat_with_tools() is currently non-streaming. The entire model response must complete before tool calls are parsed and dispatched. For long-running model reasoning, the user sees no output until completion.

Claude API supports streaming with tool_use (SSE content_block_start with type: tool_use). OpenAI also supports streaming function calls.

Proposed Approach

New trait method

fn chat_with_tools_stream(
    &self,
    messages: Vec<Message>,
    tools: Vec<ToolDefinition>,
    options: &ChatOptions,
) -> Pin<Box<dyn Stream<Item = Result<ChatStreamEvent>> + Send + '_>>;

ChatStreamEvent enum (new)

pub enum ChatStreamEvent {
    TextChunk(String),
    ToolCallStart { id: String, tool_id: String },
    ToolCallDelta { id: String, input_delta: String },
    ToolCallEnd { id: String },
}

Agent loop changes

• Process stream events in real-time
• Display TextChunk events immediately to the channel
• Accumulate ToolCall* events into complete tool calls
• After stream ends, dispatch accumulated tool calls

Acceptance Criteria

• Text chunks displayed in real-time during model reasoning
• Tool calls extracted correctly from stream
• Multi-turn tool loops work with streaming
• Fallback to non-streaming when provider does not support streaming tool_use

Trade-offs

• (+) Much better UX — user sees model "thinking" text before tool calls
• (-) Significant implementation complexity (different SSE formats per provider)
• (-) Ollama tool streaming may not be supported by ollama_rs
• (-) Increases test surface substantially

Status

Deferred to post-1.0. Implement only after Phases 1-3 are stable and if UX feedback demands it.

References

• Epic: Epic: [NativeTool] Improve native model tool_use integration #940
• Plan: .local/plan/native-tools-improvement.md (Phase 4)