Quick start

SkillWeave is a runtime for composing LLM tasks from small, focused, testable micro-skills. This page gets the reference chain running in under a minute.

Install

git clone https://github.com/manojmallick/skillweave
cd skillweave
npm install

Requires Node.js 20+. The runtime has three dependencies (the Anthropic, Google, and OpenAI SDKs), used only by the boundary judge.

Check you're ready

npm run cli -- doctor

doctor confirms your Node version and the active judge provider — with no API key set it reports the offline heuristic, so the chain runs end to end without any configuration. See the CLI guide.

Run the chain

npm start

You'll see an execution summary:

SkillWeave — document-grounding v0.2.0
executor: heuristic (no LLM key)
────────────────────────────────────────────────────────────────────────
✓ parse-input       deterministic extracted 4 blocks                     1ms  $0.0000
✓ validate-coverage deterministic coverage 1 (sufficient)                0ms  $0.0000
✓ extract-highlights probabilistic selected 3 highlights                  1ms  $0.0000
  └─ judge: 1.00
✓ memory-update     deterministic 1 records, first run                   0ms  $0.0000
────────────────────────────────────────────────────────────────────────
STATUS: SUCCESS   total: 2ms   cost: $0.0000
trace: ./traces/run-…ndjson

With no provider key set, the boundary judge runs a deterministic offline heuristic, so the whole chain completes with zero network calls.

Watch the reliability layer

Each --inject mode demonstrates one behaviour of the reliability layer:

npm start -- --inject lowconf        # low-confidence highlight → confidence routing RETRIES → recovers
npm start -- --inject hallucination  # ungrounded highlight → judge RETRIES → recovers
npm start -- --inject persistent     # always ungrounded → retries exhausted → HALTS
npm start -- --inject coverage       # too-thin input → deterministic assertion HALTS

Run on your own document

npm start -- --doc ./path/to/document.md

Use a real model

Set any one provider key and the judge calls that model with structured output:

export ANTHROPIC_API_KEY=...   # or GEMINI_API_KEY / GOOGLE_API_KEY / OPENAI_API_KEY
npm start

See the Multi-LLM judge guide for the full provider matrix.

Tests

npm test          # node:test suite (9 tests)
npm run typecheck # tsc --noEmit

Next

• Architecture — how the runtime is wired
• Primitives — SKILL, PIPELINE, STATE, ASSERTION
• Reliability layer — confidence routing, auto-judge, retry