basemode-loom

Persistent branching exploration for LLM continuations.

basemode-loom lets you generate multiple continuations, navigate a tree of alternatives, and keep everything in a local SQLite store so you can resume later.

Install

pip install basemode-loom

Quickstart

# Create a new tree with 3 branches
basemode-loom run "The ship rounded the headland and" -n 3 -m gpt-4o-mini

# Open the interactive explorer
basemode-loom view

# Continue from selected branch
basemode-loom continue -b 2 -n 3

Core Commands

basemode-loom --help
basemode-loom view --help
basemode-loom run --help
basemode-loom continue --help
basemode-loom stats --help
basemode-loom serve --help

Useful commands:

• basemode-loom view: interactive TUI tree explorer
• basemode-loom run: create a new tree from a prompt
• basemode-loom continue: branch from current/selected node
• basemode-loom nodes|active|show|children: inspect stored trees
• basemode-loom stats: analyze tree depth/branching/model usage
• basemode-loom export|import: move trees in/out as JSON/Markdown
• basemode-loom serve: run REST/WebSocket API for frontend usage

Tree Search

The TUI tree picker (Tab) supports live metadata filtering, category/domain/ source/model facets, ID lookup, FTS5 keyword search, and semantic search when the selected database contains a compatible vector index.

Guardian Angel corpora use the optional MLX search dependencies:

pip install 'basemode-loom[embed-mlx]'
basemode-loom view --db /path/to/corpus.sqlite

Press / in the tree picker, enter a query, and press Enter to rank matching trees. Build or update an embedding index directly with:

basemode-loom embed --db /path/to/corpus.sqlite --model mlx
basemode-loom embed --db /path/to/corpus.sqlite --model mlx --incremental

The MLX model is downloaded lazily on first use. Embedding dependencies remain isolated in the optional embed and embed-mlx installation extras.

Model Selection

• TUI model picker is available via m.
• Picker can consume verified model metadata from basemode when available.
• Session state supports model-plan metadata for multi-model generation workflows.
• Force OpenRouter passthrough for new/unknown IDs with -m "or:vendor/model" (or openrouter:vendor/model), e.g. -m "or:moonshotai/kimi-k2.6".

Storage

By default, data is stored in a local SQLite DB under your user data directory. Use --db /path/to/file.sqlite to choose a custom location.

Docs

Project docs live in docs/ (MkDocs):

make docs
make docs-serve

Then open http://localhost:8001.