A factory that maps multiple AI provider API formats (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages) to a single ProviderAdapter interface emitting normalized chunks. The trio of adapters (chat-completions.ts, responses.ts, anthropic-messages.ts) plus a common types.ts define request building, streaming consumption, and tool-call abstraction. This is directly extractable into any AI application needing multi-provider support without vendor lock-in.

A hand-rolled SSE parser that consumes OpenAI Chat Completions streaming responses and emits typed StreamDelta events (text, reasoning, tool calls, usage) in a vendor-neutral shape. It handles raw chunk parsing, [DONE] framing, and error propagation. Together with anthropic-messages-sse.ts and openai-responses-sse.ts, this forms a complete normalized streaming layer reusable in any chat client.

A custom StreamableHTTPClientTransport implementation for the MCP SDK that works inside the Tauri/browser sandbox where the official transport may not function. It negotiates SSE sessions, maintains the read/write stream lifecycle, and handles JSON-RPC batching and reconnection. This is non-obvious integration work that any MCP-enabling app could reuse.

A self-contained module for compressing long AI conversations: it estimates token usage (with distinct heuristics for English vs CJK text), decides when compression is needed, calls an LLM to produce a summary, and preserves the summary for subsequent API calls. No internal imports beyond shared types — it can be dropped into any LLM chat application that needs context-window management.

Parses <file path="..."> tags from AI model output and writes content to a workspace directory, with explicit path-traversal guards that reject absolute paths and ../ escapes, plus overwrite confirmation. This is a security-conscious implementation of the 'model writes files' pattern increasingly needed by agentic chat tools.