polygraph

{
  "name": "polygraph",
  "owner": {
    "name": "cognitive-fab",
    "url": "https://github.com/cognitive-fab"
  },
  "plugins": [
    {
      "name": "polygraph",
      "source": "./",
      "description": "A polygraph for your state machine. The agent does the hard part for you: it instruments your code, builds any test doubles needed to run it, and captures real execution traces — then has an LLM derive a transition-function spec from the source (default: a SAM v2 strict-profile module with named intents/schemas/domains, keyed acceptors, observable reject(reason), sealed model, next-state (prime) acceptors), replays the traces against it (conformance), model-checks it against invariants (finds bugs by exhaustive iteration), and can escalate the winning spec to TLC via mechanical TLA+ transpilation (--tla), surfacing every disagreement as a spec-error, code-finding, or contract-error. Also includes polygen (author NEW verifiable code from a feature description, self-repaired against reachable invariant violations before it ships) and polyvers (version a state machine with mechanical compatibility gates: classify the change into lanes, replay old-version stimuli and validate migrations against fleet snapshots, and model-check whether any live state can be driven to an invariant violation under the new rules — no API key), and polynv (elicit the invariants themselves: harvest candidates from the contract vocabulary, traces, and fleet snapshots, pre-check each against the machine, drive a plugin-led interview into an attributed intent ledger, and grade invariant-set strength by mutation — no API key except the optional --llm source). New in 7.0: the plugin now starts working at CODE-GENERATION time, before verification. The workflow skill builds a feature as a WORKFLOW, not an agentic loop: design stages along the data flow with the user (every failure either bends back to one named stage carrying its exact signal under a bounded budget, or escalates to a human), f