Skip to content

v0.6.66b477

Choose a tag to compare

@tobocop2 tobocop2 released this 18 May 06:41
· 247 commits to main since this release
00db6ad

Agents

  • Any agent that speaks MCP can now read, change, and reset lilbee settings, with the same validation rules the TUI uses: nothing read-only can be touched, secrets stay write-only, dependent services reset when a provider or model changes.
  • lilbee_catalog_browse exposes the Hugging Face model catalog through MCP with the same filters as the TUI (task, size, sort, query).
  • lilbee_search falls back to the configured top_k when an agent omits it instead of erroring.

New features

  • Named lilbees. Give a project a short name with lilbee_name; the status bar shows the name instead of a long absolute path. Press F4 to flip between the name and the full path.
  • /rebuild command. A first-class command to force a full re-index of the current project from the chat screen, with a confirm step.
  • CUDA for NVIDIA. Linux and Windows ship a CUDA-built standalone binary alongside the default Vulkan/Metal one, and lilbee-cuda is on PyPI.

Overall improvements

  • Better defaults for documents AND code. Retrieval defaults were re-tuned. The old defaults were code-shaped and left long PDFs, manuals, and crawled docs thin. The new defaults ground prose-shaped questions without hurting code search.
  • Settings behave the same everywhere. Settings changed in the TUI, in the API, through MCP, or from Python all go through the same validated path.
  • Crawler stability. Matches the Chromium build it expects and retries cleanly when a binary is briefly unavailable.

Fixes

  • Invalid size or sort on /v1/models/catalog returns a clean client error instead of a server error.
  • Switching the LLM provider resets dependent services so old state can't leak.
  • Wiki tools moved to an experimental section so they don't crowd the default agent surface.
  • /reset and /rebuild use the standard confirm dialog.

What's Changed

  • demos: refresh the reel by @tobocop2 in #250
  • demos: pre-cached setup + mcp trims (re-apply) by @tobocop2 in #252
  • demos: tui-setup with pre-cached chat + live SmolLM download by @tobocop2 in #261
  • fix: block chat against a still-downloading model by @tobocop2 in #260
  • demos: move generation off main by @tobocop2 in #263
  • add cuda standalone binaries + lilbee-cuda package variant by @tobocop2 in #269
  • bug-fix: Crawler: match Chromium revision exactly + retry on missing binary by @tobocop2 in #268
  • Add .github/FUNDING.yml by @tobocop2 in #270
  • MCP-driven fine-tuning: expose lilbee settings to agents by @tobocop2 in #265

Full Changelog: v0.6.66b476...v0.6.66b477