Kiwi MCP

# Phase 8.1: Extend http_client with Streaming **Estimated Time:** 3-4 days **Dependencies:** None **Status:** 📋 Planning ## Overview Extend the `http_client` primitive to support Server-Sent Events (SSE) streaming with configurable destination sinks. This is the foundation for all streaming features. ## What This Phase Accomplishes - Adds streaming support to `http_client` primitive - Implements SSE parsing and event fan-out - Creates data-driven sink architecture (file_sink, null_sink, websocket_sink) - Adds built-in `return` sink for buffering - Enables recursive body templating for parameter substitution ## Files to Create/Modify - `kiwi_mcp/primitives/http_client.py` (extend - add ReturnSink class + _execute_stream) - `.ai/tools/sinks/file_sink.py` (new - Python-only, no YAML) - `.ai/tools/sinks/null_sink.py` (new - Python-only, no YAML) - `.ai/tools/sinks/websocket_sink.py` (new - Python-only, no YAML) - `tests/primitives/test_http_streaming.py` (new) ## Architecture Notes **Key Principle:** `http_client` is a pure primitive. It receives pre-instantiated sink objects from the tool executor (during chain resolution). Sinks are passed via `__sinks` parameter. **Sink Types:** - `return` (built-in): Buffers events in memory for inclusion in result - `file_sink` (data-driven): Appends events to file in JSONL format - `null_sink` (data-driven): Discards events (fire-and-forget) - `websocket_sink` (data-driven): Forwards events to WebSocket endpoint **Sink Instantiation:** Happens in tool executor during chain resolution, BEFORE http_client execution. See doc section "Sink Architecture (Data-Driven)" for details. **Tool Pattern:** Sink tools are Python-only files (no YAML sidecars). Metadata is declared at the top of the Python file using module-level variables: `__tool_type__`, `__version__`, `__executor_id__`, `__category__`. Dependencies are declared as `DEPENDENCIES = [...]` if needed. ## Task Order 1. Add StreamConfig and StreamDestination dataclasses to http_client.py 2. Add ReturnSink class to http_client.py (built-in primitive) 3. Implement _execute_stream method with SSE parsing 4. Add recursive body templating 5. Write tests for http_client streaming 6. Create file_sink tool (Python-only) - data-driven tool 7. Create null_sink tool (Python-only) - data-driven tool 8. Create websocket_sink tool (Python-only) - data-driven tool ## Success Criteria - [ ] http_client supports `mode="stream"` parameter - [ ] SSE events are parsed correctly - [ ] Events fan-out to all configured sinks - [ ] Built-in `return` sink buffers events - [ ] All three data-driven sinks work correctly - [ ] Body templating works recursively (dict/list/str) - [ ] Tests cover sync, stream, and error cases ## Related Sections - Doc lines 262-504: Layer 1 (Extended http_client Primitive) - Doc lines 5775-6125: Appendix A.7 (Sink Architecture)