Changelog

Studio and the CLI/TUI get a real Git surface — merge, rebase, resolve conflicts, and restore points with undo. Run it from Studio's UI or from the terminal with /git ...; AI agents can propose changes through the same surface but are never allowed to force-push, reset, or discard on their own.

Studio's new Infrastructure surface lets you choose where tests actually run, per run:

• Provision test machines in your own AWS, Azure, or GCP account, or connect BrowserStack or TestMu AI.
• Credentials only ever land in your OS keychain — they never round-trip to the UI — and a run opens a single port to your IP that closes on every exit path; there's no standing inbound rule.
• Cloud mode now shows the platform's real navigation (Repos, Test Runs, Monitoring) instead of local-only surfaces, and the AI assistant (Memory, Automations, Artifacts) works from Local mode too, not just Cloud.

Infrastructure requires a Pro plan. Studio itself now admits Pro and Enterprise orgs — if yours is still on Growth, you'll need to upgrade before Studio unlocks again.

Studio's browser MCP is bumped to Playwright MCP 0.0.79, adding browser_generate_locator so an AI agent can synthesize a selector instead of guessing one, four browser_verify_* assertion tools for checking page state, and browser_find under the Core capability.

Studio's auto-discovery doesn't catch every setup — monorepos, Python repos, non-standard script names. You can now add those servers yourself, picked from every package.json script in the repo up to two levels deep:

• Custom servers get the same boot signals, reachability panel, and logs as auto-discovered ones, and you can edit or remove them later.
• A new working-directory field lets a monorepo subfolder (like client/) run correctly instead of failing from the repo root.
• "Use as app" now lets you pick which running server your tests and emulator actually target, remembered per repo.

Augur is now a live workbench surface in Studio, alongside Browser, Mobile, and Evals:

• Embeds each run's own report.html verbatim — verdict banner, banded pHash/SSIM/pixel checks, five evidence viewers (side-by-side, slider, flicker, overlay, diff mask), region annotations, and plain-language root causes.
• Run history spans both real suite runs and augur demo runs, with demo runs labeled so a demo NO-SHIP is never mistaken for the suite's own result.
• Verdict chips (SHIP / NEEDS REVIEW / NO-SHIP) match the report's own banner.
• A one-click demo streams augur demo live — 5 scenes, 9 real GenAI failure patterns, no API keys — into an activity feed and opens the run when it finishes.
• The report's Apply and reject buttons now work directly in the pane.

Studio's workbench rail adds an Evals surface, next to Browser, Mobile, and Augur:

• Scans the open repo for DeepEval test-run JSON (.deepeval/, deepeval-results/, or a DEEPEVAL_RESULTS_FOLDER export) — no upload required.
• Groups metric scores by DeepEval's own taxonomy — RAG, Agentic, Multi-Turn, Safety & Compliance, Multimodal, and Custom (G-Eval/DAG) — each scored against its threshold with the judge's reason, evaluation model, and cost.
• Every case links to the agent conversation (single-turn or multi-turn) and the tool calls behind it, so a failing score can be read against what the agent actually said and did.
• Opens on a hero band: pass-rate donut, per-category health bars, and a pass-rate trend across the repo's run history, before drilling into individual cases.

Studio's Live Device pane now mirrors a real Android device connected over USB the same way it mirrors an emulator — plug one in and the pane picks it up automatically, no separate setup. For machines without an emulator yet, testrelic device setup takes a bare machine to a booted, hardened AVD with no JDK or Android Studio required — packages come from dl.google.com, and the AVD it creates is tuned from real usage (4 GB RAM, Wi-Fi off, 4 cores). Studio's Install Android tooling button drives the same steps as a live checklist, then boots the device and starts serving WebDriver.

augur run can now generate candidates from several models in one pass and grade them against the same baseline, so you can compare backends like Gemini, Flux, and OpenAI Images without maintaining separate suites. Attribution grounding (opt-in via --attribution) shows which part of your prompt each region of the image is honoring.