REA 6.8.0 Turns Application Captures into Agent Test Environments, Enables Stateful Service by Default

The open-source reverse-engineering tool REA released rea-agents 6.8.0 on October 25, and the new version is also listed on npm. Through MCP, REA provides coding agents with code, references, and runtime observations. A key focus of this update is turning captured application behavior into a local environment that agents can operate. Release announcement, npm package

The new version can project network captures and Evidence files into environment descriptions. It can also create environments from browser interaction scenarios or prompts, then compare them with results from new application probes. environment-serve now infers and stores state by default, allowing eligible read tools to answer based on current data. From an engineering perspective, this gives agent tests a way to check data outcomes after operations, rather than only verifying call order; actual coverage still depends on the captured material. Stateful service implementation notes

Write capabilities have clear limits: the service uses recorded evidence to determine which effects it can apply, and does not enable a write when observations conflict. The new version also handles cases where a create operation's response omits fields that later reads supply, as well as differences in timestamp precision. These rules limit the behavior the environment can simulate, so they do not establish that it fully reproduces the original application's business logic. Write handling changes

When upgrading, workflows that depend on stateless record-and-replay should check whether they need --replay-only. Another compatibility change is that analyze_function now returns a condensed summary by default; MCP callers that need the previous full, embedded analysis must specify detail: "complete". CLI JSONL output now emits collection items one per line, followed by a metadata record, so parsers must recognize control records. Migration notes

Development teams should validate read/write consistency using their own captured data, check how unobserved inputs are handled, and verify that state is correct after an agent completes a task. What can currently be verified is the formal release, package listing, and implementation rules; that is not enough to infer reconstruction accuracy across applications or agent task success rates.