A program that writes programs: deterministic harness (tree, scheduler, verification) with a tiny LLM as one computing layer for semantic fills.
See docs/PARADIGM.md and docs/VERIFICATION.md.
Requires uv.
uv sync --all-extras
cp ../research-watch/.env .env # or copy OPENAI_API_KEY onlyuv run unfolder init
uv run unfolder plan "A CLI that prints hello" --heuristic
uv run unfolder statusLayer-bound swarm architecture (docs/V2_ARCHITECTURE.md,
docs/V2_IMPLEMENTATION.md):
spec swarm → blueprint swarm (tree.yaml) → deterministic materialization →
implementation swarms (upward invalidation, sibling locking) → acceptance swarm.
uv run unfolder v2 build "A tiny expense tracker library" --project /tmp/app # full pipeline
uv run unfolder v2 spec "..." # phase 1 only (interactive elicitation)
uv run unfolder v2 status --project /tmp/appAdd --headless to let the spec swarm decide instead of asking questions.
The harness receives only a spec and acceptance tests — no src/. It must generate the full application.
uv run unfolder build run --target-id ledger-cli --heuristic --dest /tmp/ledger --keepWith OpenAI (harness + LLM only, no reference finalize):
uv run unfolder build run --target-id ledger-cli --dest /tmp/ledger-llmTarget layout: examples/verify-target/ contains SPEC.md, tests/, pyproject.toml — no src/.
./scripts/e2e.sh # full quality gate (recommended before handoff)
uv run pytest tests/ -q
uv run ruff check src tests
uv run unfolder viz --open # requires viz extraSee HANDOFF.md for the complete checklist.
