Release notes — September 2026 (1–20 September)

Summary

Studio now explains your test results instead of just listing them. A failed run shows its causes, a passing run is checked for whether it would have caught a regression, and a new Project Assistant answers questions from your own project's data and can carry out changes you approve. Requirements can be extracted from many documents at once, Excel workbooks are read sheet by sheet, and end-to-end workflows can build themselves from your requirements and API specs. Postman imports are considerably more faithful, and a broad performance and reliability pass makes large projects and large test packs noticeably faster and lighter on memory.

Improvements

• Run triage — a failed run shows its causes, not its rows. Failures are grouped by what went wrong, each cause carries a verdict (test problem, data problem, environment problem, or a suspected product defect), and a model is asked about each cause once rather than once per test. Failures are judged on what the server actually said, not just the status code: a firewall or sign-in page, a rate limit, a missing record, and an "[object Object]" sent in place of a real value are each recognised for what they are. For a cause Studio cannot explain, you can add a hint and ask for a second look.
• Triage reads your passing tests too. It flags a green test that would not have noticed a regression, tells you whether a requirement was actually verified (as opposed to merely "all its tests passed"), and compares each response with the documentation. Suggested strengthening only ever adds checks; nothing is applied without you, and a suspected product defect is never "fixed" by loosening a test.
• A smarter Fix button. Fix now reads the specification, the requirement, the test's own intent and what the run actually did, then either fixes the test or tells you why it will not. It refuses a change that would not alter the request, and it will not turn a negative test into one that cannot fail.
• Project Assistant. Ask questions about your project and get answers drawn from its own requirements, specifications, tests and run history, with citations that open the item in the project tree. It knows which project and which screen you are on, reads what you attach, answers questions about the product from its own documentation, and works across the projects you can already open.
• The assistant can also propose changes, and carries them out only when you accept: run a test, pack or workflow; analyse a run; generate tests for an endpoint, a feature or a whole project; edit a test; decide requirements; link a requirement to its endpoint and fill gaps; shape packs, tags, environments, mocks, scheduled reports and alerts; and create projects, features, endpoints, tests, requirements, documents, datasets and workflows. Every change goes through the same route as the equivalent button, so permissions, licensing and run limits apply exactly as they would if you clicked it. It never deletes anything, handles credentials, publishes, or changes what production points at. Fixes it applies are verified, and workflow edits can be undone.
• The assistant learns from being corrected. Give an answer a thumbs up or down, or simply tell it what was wrong; it remembers the correction for that project only, and you can review and retire what it has learned.
• Workflows can build themselves. Studio finds the end-to-end flows a project wants from use-case requirements, your API specifications, your existing tests and the order calls actually happen in, then offers to build them with each step's data wired to the earlier step that produces it. Where a step has no test or no endpoint, it says which, and can generate the missing tests and build the workflow in one accept. A workflow now counts as coverage for the requirements it exercises, and a workflow run can verify one.
• Workflow Doctor. In the workflow editor, Studio diagnoses a workflow's structure, data connections, run history and source requirement, and offers the change that fixes each finding. A step whose test was deleted can be given a new one.
• Many requirement documents at once. Upload several documents and they are read one after another in the background; every requirement records the document, section, row or page it came from, and review spans all documents. A rule that appears in more than one document is counted once, and documents that disagree wait for a person to decide, never for the system to pick a side.
• Excel workbooks are read sheet by sheet. Requirement sheets are read row by row, test-data sheets become datasets, and glossaries become project terms rather than requirements. You can correct how a sheet was read.
• Document versions. Replace a document with a new version of itself: unchanged requirements keep their review decision and test links, changed ones return to review with their tests flagged, and dropped ones are listed for you to retire or keep. Deleting a document now tells you what it will remove first.
• Postman imports are far more complete. A collection imports as one test per request, reports any requests it left out, and maps more assertions. Checks the translator could not map are compiled from the collection's own script, variable captures are imported, and a variable at the start of a URL is treated as the base URL whatever it is called. Approving a detected collection change now imports its tests, and an unchanged collection no longer asks for review.
• Convert imported Postman checks that came in as placeholders — from Fix, from the assistant, or in bulk — using the collection's own script, with no AI involved where the script is enough.
• Clearer assertion results. Each result row says which assertion it is, by number and name; you can state the check you want in words and have it built; custom scripts can write log lines that appear on the result; and a custom check says what it looked for and what it found. In the Test Workbench you can copy what a results tab shows.
• Dashboards. TV Mode and Schedule now work, scheduled reports and alert rules now run on their own, widgets can be resized and rearranged, and Focus Project narrows the multi-project dashboards. The environment matrix draws and the CIO radar is measured. Every time range in the toolbar, including 14 days and Last 24 hours, now applies as labelled.
• Activity feed switch. An administrator can now switch the activity feed on or off from Settings.
• Pack runs and reports. Pack runs open their summary when they finish, reports show the counters first with run causes folded below, and search survives a reload.
• Performance and memory. The reporting summary list, project open, the test table, the test editor, the sidebar tree, the requirements page and the workflow canvas are all faster. Large test packs run with less overhead per test, reuse TLS connections, and no longer time out the browser that started them. Long-running installations no longer accumulate memory from caches that never released. The application ships considerably smaller.

Fixes

• Execution: a request timeout is now reported once, in one unit, from every runner; a GET no longer sends a body, and the log shows what was actually sent; a pack cancelled while still queued no longer runs anyway; and pack duration is reported as wall-clock time.
• Scheduling: a pack saved for the current minute runs immediately, recurring packs keep their slot, and a past start time is refused.
• Postman and imports: an import creates only the tests the collection contains, and the Auto-Detect hint no longer suggests an API type Studio does not support.
• Test generation: a 204 response is no longer given assertions no correct server can pass; a SOAP negative test that XML cannot express is no longer generated; a header requirement such as Content-Type or SOAPAction can now produce a test; and Generation Setup no longer presents a default as the specification's own answer.
• Workflows: generated journeys now draw their connectors, cards no longer overlap on the canvas, wiring no longer fills in the field a negative test deliberately omits, and a workflow that belongs to no project is visible from every project. Deleting a workflow now removes it from the packs that run it instead of leaving them broken.
• Licensing: a blocked request no longer looks like a rejected key; a fresh install is licensed from first start and the Free licence runs six months; Deactivate now takes effect and the page reports the true state; and a routine check-in no longer erases an activation.
• Sign-in and access: password reset and change now report what actually happened, password rules tick as you type, the key-recovery screen supports dark mode, and a permission that was switched off is no longer granted back on the next load.
• Security: API endpoint definitions (URLs, headers and schemas) could previously be read without signing in; reading them now requires authentication, as do several diagnostic routes. Decrypted secrets are no longer kept in memory after use.
• Mocks: a second mock on the same path and method no longer returns 404 for every request.
• AI: a model reply that was cut off is now recognised as such and not treated as complete.
• Desktop Agent: an authentication profile that carries a secret field now authenticates on the Agent.
• Engine and installer: an Engine that cannot start now writes the reason to its log instead of crash-looping; Settings > Email loads and saves profiles again; a template secret is replaced during Engine updates; JSONPath works in the packaged Engine; and every network failure is no longer reported as "releases not available".
• Assistant: run cards no longer report "done" the instant a pack starts, a screen you are looking at is no longer reported as not existing, and citations open the item.
• Reporting: the execution summary no longer shows a missing duration, and workflow pack reports appear.
• Custom assertions: a custom check's script box now saves to the field the engine actually runs, so deleting a script really removes the check.

Known issues

• WebSocket-RPC tests run on the cloud Engine only; the desktop Agent does not yet execute them and will say so rather than misreporting a result.
• A Postman check that cannot be converted automatically stays as a flagged placeholder and fails on every run until it is converted or rewritten. Fix explains why it could not be converted.
• The desktop Agent must be updated to this release to receive the authentication and execution fixes above.
• SOAP tests generated before the August release may still carry a stored "[object Object]" request value until they are regenerated.

Related links

• Browse all release notes
• Download the latest build
• Help Center
• Product documentation
• Plans and editions