Summary

Implements E2 — a reusable SSE streaming parser in jcode-llm-core to replace per-provider ad-hoc SSE parsing.

What's Included

crates/jcode-llm-core/src/sse.rs (680 lines)

• SseEvent — parsed event with event, data, id, retry fields. Event-type interning (Cow::Borrowed) for common Anthropic/OpenAI events eliminates per-event String allocation.
• SseParser — streaming state machine. Supports:
  • CR, LF, CRLF line endings
  • UTF-8 BOM stripping (SSE spec)
  • Configurable per-event data cap (100 MB default, 10 MB buffer cap)
  • Zero-copy fast path when buffer is empty
  • feed() returns complete events; flush() handles stream-end
  • 15 unit tests covering all parse states
• SseStream<S> — futures::Stream wrapper converting Result<Vec<u8>, _> → Result<SseEvent, _>. Handles UTF-8 multi-byte characters split across chunk boundaries.

docs/pr-plans/E2-sse-parser.md (46 lines)

Plan file with research summary, gap analysis, and implementation decisions.

Why a Shared Module?

Previously, SSE parsing was embedded per-provider (Anthropic has inline parsing; OpenRouter too). This module:

• Eliminates duplicated parse logic
• Adds UTF-8 tail handling (partial multi-byte chars at chunk boundaries)
• Adds proper EventSource-compatible SseStream for ergonomic provider code
• Follows pi-agent-rust reference implementation (MIT license)

Verification

• cargo check -p jcode-llm-core — clean (0 new errors)
• All 15 SseParser tests pass
• SseStream verified by cargo check (stream boilerplate; parser tests cover all parsing paths)

Refs: #E2, docs/pr-plans/E2-sse-parser.md