JetBrains announced the new Junie /demo mode in the Junie CLI tool, aimed at automating the interactive part of verifying code changes. Instead of merely running tests, a developer can describe the required scenario, and Junie will build and run the application, interact with its interface, and then produce an HTML report containing the result, screenshots, and a video of the execution.
How does the new mode work?
Junie analyzes the project and suggests a build and run plan, then creates the configuration files after the user approves. Changes can be selected from the branch, session, working tree, or latest commit before entering a request specifying the actions and expected results. In the example presented by JetBrains, the agent identifies two issues in a task-tracking tool, changes their status to Done, verifies that the counters have changed, and then reloads the page to confirm that the result persists.
The user can follow the execution live as the agent navigates the interface. The video includes introductory slides for each scenario and a closing results slide, along with explanatory subtitles that can be enabled or disabled. JetBrains explains that a model analyzes screenshots after recording to prepare these slides, while voice-over may be added in a later update.
From manual demonstration to auditable review
The HTML report combines the original request, result, video, screenshots, and executed steps, while clarifying which tests passed, failed, or remained incomplete. This makes the output reviewable by a developer, quality assurance engineer, or team member who wants to see the feature’s behavior instead of merely reading the code change.
certi.news sees the practical value here not as replacing automated tests, but as adding visual evidence of interface behavior. The judgment about whether a change is ready remains the reviewer’s responsibility; an agent completing its execution does not automatically mean that the check passed. For this reason, JetBrains’ examples ask Junie to issue an explicit verdict, and pass the result only when PASS is written, while values such as FAIL and PARTIAL, or a missing result, cause the result check to fail.
Integration with GitHub Actions and cost
JetBrains used the mode internally with GitHub Actions on more than 1,500 unique pull requests and created more than 2,100 demo videos. It also used it in 22 scenarios for smoke tests on release branches, totaling more than 1,300 internal tests. The company provides two examples of YAML files that can be adapted: one for checking changes that merit a demo and adding evidence links to the pull request, and the other for release tests triggered on push or manually.
The examples rely on GitHub Artifacts, so they do not require a separate service to host the video. JetBrains states that running 22 cases in an internal measurement cost $19.94 using GPT-5.6 SOL under the subscription conversion it uses, excluding the cost of the CI execution environment. These are internal figures affected by the application, build steps, and request wording, and are not a general estimate for every project.
Environment and operational limitations
The mode runs inside a Docker container based on Debian Bookworm, including Chromium, Node.js, Xvfb, a window manager, and tools such as xdotool and ffmpeg. It uses a model that supports Computer Use to control clicks, the keyboard, and screenshots. Complex repositories can define multiple virtual machine templates, with separate settings for services and interfaces.
Execution does not start without a supported model. Junie uses the active model if it is available and supports Computer Use; otherwise, it selects from a list including GPT-5.6 SOL, GPT-6 Astra, and GPT-5.5, followed by GPT-5.4, with High reasoning effort enabled in /demo mode. Execution also takes minutes, and its greatest benefit remains moving repetitive interactive verification into an auditable workflow, not guaranteeing the correctness of the result without human review.