What problem does it solve?
It makes it practical to validate Conductor workflow path behavior end-to-end without relying on real LLM calls, real Azure DevOps state, or nondeterministic external systems—so regressions are caught early in CI.
Core Features & Use Cases
- Workflow path coverage for scenarios: Runs scenarios defined under tests/harness/scenarios/<name>/scenario.yaml and asserts the workflow reaches the expected terminal node.
- Deterministic LLM boundary via FakeProvider: Replays scripted agent outputs in the exact call order the workflow expects, including both structured JSON and regex-matched unstructured text.
- Deterministic script interception via .NET shim: Matches script node calls (scripted commands like polyphony/twig/gh with argv prefix rules) and returns scenario-defined stdout/exit codes.
- Gate-seam simplification for repeatability: Uses --skip-gates to auto-select the first human gate option, enabling stable scenario routing while gate API work is pending.
Quick Start
Ask the AI to “run and debug scenario close_out_happy_path in the polyphony workflow harness using the Python driver with verbose output so I can see why the workflow did not reach workflow_completed.”