skill-tuple-space

Fuzzy Tuple Space for Agent Skills.

Inspired by David Gelernter's Linda coordination language and JavaSpaces, skill-tuple-space indexes your agent skill git repos into a local SQLite store and lets you query them using two complementary mechanisms:

• Fuzzy layer — membership-function scoring on structured metadata (skill_class, crud_verb, topic, requires_role). Pure Python, no heavy deps.
• Semantic layer (optional) — cosine similarity on SKILL.md descriptions via sentence-transformers. Opt in only if you want it.

These are two distinct things: the semantic layer finds what is similar, the fuzzy layer reasons about how well something fits your current taxonomy and context.

Install

Default — fuzzy matching only, lightweight, no PyTorch:

pip install skill-tuple-space

With semantic search — adds sentence-transformers + PyTorch (~200MB CPU / ~3GB CUDA):

pip install "skill-tuple-space[semantic]"

Quickstart

# 1. Copy and edit config
mkdir -p ~/.skill-space
cp config.example.toml ~/.skill-space/config.toml

# 2. Index your repos
skill-space index run

# 3. Semantic + fuzzy search
skill-space search query "create something visual with 3d"

# 4. Linda-style template matching
skill-space search read "skill_class=Process, crud_verb=create, topic=*"

# 5. Learning journal
skill-space learn claim create-openscad-from-construction-image-en
skill-space learn done create-openscad-from-construction-image-en --level 4
skill-space learn next --topic agent-skills

# 6. Space health
skill-space space stats
skill-space space drift

Architecture

src/skill_space/
├── cli.py        # typer CLI entry point
├── indexer.py    # git clone + SKILL.md parse + embed
├── embedder.py   # sentence-transformers (all-MiniLM-L6-v2, 384-dim)
├── store.py      # SQLite + sqlite-vec persistence
├── matcher.py    # fuzzy membership functions + cosine combiner
├── journal.py    # learning events CRUD
├── predictor.py  # readiness scoring + next-skill suggestion
└── display.py    # rich terminal output

License

MIT