AthenaFlow runs a real browser, maps interaction paths, writes a human-readable spec first, then implements Playwright tests from that spec. The spec step is intentional, it's reviewable before any code runs, and every generated test traces back to a TC-ID. Self-healing works by resolving semantic identifiers rather than selectors, so when the DOM shifts, the identifier survives.

Three repos make up the stack:

athena-flow-cli is the workflow runtime. It hooks into Claude Code's event system, receives runtime events over a Unix domain socket as NDJSON, persists sessions to SQLite, and renders a live terminal UI. Sessions are fully resumable. In CI, exec mode emits JSONL with clean exit codes for every failure mode. https://github.com/lespaceman/athena-flow-cli

agent-web-interface is an MCP server that exposes semantic page snapshots to the model instead of raw DOM or full accessibility trees. Keeps context window usage low and action resolution stable across layout changes. https://github.com/lespaceman/agent-web-interface

athena-workflow-marketplace holds the QA domain knowledge as a Claude plugin -- composable skills for codebase analysis, coverage planning, browser exploration, spec generation, and test implementation, wired together as a multi-phase workflow with checkpointed progress. https://github.com/lespaceman/athena-workflow-marketplace

  npm install -g @athenaflow/cli
  https://athenaflow.in/