v0.9.22

Changes

• Fix missing streaming output when tool-calling clients (such as OpenClaw) chat with cloud or third-party provider models through /v1/chat/completions: responses now stream incrementally over SSE instead of arriving in one chunk after the full completion (#59).

v0.9.21

Changes

• Show Xiaozhi's default administrator email and password in the app details panel.
• Add one-click copying of the local login credentials with localized guidance.

v0.9.20

Changes

• Add one-click Xiaozhi installation and lifecycle management with a storage-root-managed Docker Compose stack.
• Configure task-specific language, speech recognition, embedding, and image generation models before launch.
• Support Apple Silicon through Docker amd64 emulation and initialize a local administrator with a private Workspace.

v0.9.19

Changes

• Add one-click ZCode desktop installation and launch support for macOS, Linux, and Windows using the official domestic CDN.
• Detect existing user-managed ZCode installations without taking ownership or exposing uninstall actions.
• Configure the selected CSGLite model in ZCode before launch while preserving existing providers and settings.

v0.9.18

• Added retryable 32 MiB chunk uploads for large local model files, with offset validation and whole-file client compatibility.
• Added --openai-stream-default and CSGHUB_LITE_OPENAI_STREAM_DEFAULT to stream OpenAI chat responses when requests omit stream.

v0.9.17

• Fixed delayed OpenAI-compatible streaming through third-party providers such as DeepSeek by forwarding SSE chunks immediately.

v0.9.16

What's New

• AI Apps now detects Codex App installed in custom locations on Windows by reading registry Uninstall entries and scanning common install directories, instead of only the managed install path. (#55)
• Added a "Specify Install Location" field in the Codex App details drawer (backed by the new POST /api/apps/path endpoint) so you can point csghub-lite at any Codex App install that automatic detection misses.

v0.9.15

What's New

• Dashboard now shows GPU memory on Windows machines with AMD or Intel GPUs, including integrated GPUs such as the Radeon 780M. Previously the GPU card always showed N/A when no NVIDIA card was present. (#39)
• Dashboard on macOS now shows live GPU memory usage instead of the total amount only.

v0.9.14

• Fixed Windows embedding runtime readiness checks so broken worker imports are caught before startup.
• Added clearer guidance when Windows embedding imports fail due to missing Visual C++ runtime components.

v0.9.13

• Fixed model library loading status text so long progress messages no longer overlap the date and operation columns.
• Added a download queue that runs at most three model or dataset downloads at once and shows queued downloads clearly.
• Kept model library filters, run, details, and delete actions responsive while unrelated downloads are running.