v0.2.0
v0.2.0
Smarter lazy tool loading: the model now activates only the tools a task needs, so each request carries just the schemas in use instead of the full toolset.
New
- On-demand tool loading: the model fetches a compact catalog of tool names and enables or disables tools as a task needs them.
- Calling a listed tool directly activates it too, saving a round trip.
- The active tool set is saved with the session, so resuming a chat restores the exact working set.
Improved
- Unknown tool names now get an error listing the real tool names, so models self-correct instead of retrying a bad name.
- Redesigned prompt input (CLI): a thin rounded-border prompt with a
>marker replaces the filled teal block.
Fixed
- OpenRouter models with mandatory reasoning (e.g. gpt-oss) no longer fail session-title generation with a 400.
- Turns where the model emits its whole response on the reasoning channel (seen with GLM and gpt-oss) no longer error as an empty response — the reasoning text becomes the answer.
- Tool-call names mangled by gpt-oss's harmony format are sanitized, so the calls resolve instead of failing as unknown tools.