v1.17.0

Added

• Live response streaming to IDE clients: the chibi stdio --tui transport can now stream the model's answer as it is generated instead of delivering it only at the end. IDE clients opt in by declaring capabilities.streaming: true in the initialize handshake (protocol v1 unchanged — additive, no version bump); they then receive delta frames ({"type": "delta", "request_id", "thread_id", "text"}) carrying partial answer text, followed by the authoritative final result frame as before, so existing renderers can simply overwrite the partial text. Deltas are coalesced by the transport (flushed at most every ~200 ms or 512 chars) to keep frame rates sane; deltas strictly precede the terminal frame on the wire; cancelling a request or hitting an error drops any buffered deltas — a delta frame never follows a terminal one. Deltas are in-memory only: they are never persisted to thread history. Clients that do not declare the capability see no protocol change at all (VS Code extension and other v1 consumers are unaffected; thinking/reasoning tokens are never streamed as body text).
• New LLM provider: Entrim (https://api.entrim.ai/v1) — an OpenAI-compatible endpoint serving DeepSeek, GLM and other open models, with defaults deepseek-ai/DeepSeek-V4.1-Flash for chat and zai-org/GLM-5.3-Flash for vision and moderation; activated by setting the ENTRIM_API_KEY environment variable.