Changelog

New

• The agent now hunts feature by feature, fully checking a few features at a time instead of chasing every suspicion first. Screen scans start much earlier in a run.

Improved

• Sandbox Docker stacks now have readable names in docker ps.

Fixed

• Published reports no longer show "Something went wrong". Re-export to fix older ones.
• Parallel lanes no longer erase each other's Playwright results, which made passing tests fail.
• A screen scan that recorded nothing is now retried instead of counting as done.
• Stopping a run no longer loses the work in progress.
• The link crawl now works on apps whose navigation is rendered client-side.
• The screen scan no longer misses the first screens of long walks.
• A new session on an already-built sandbox reuses the same stack instead of starting a broken duplicate.
• Map paths like /<slug>/projects, /{slug}/projects and /[slug]/projects are now all read correctly.

New

• New four-page onboarding that walks you through how Heal works

Improved

• The attack scan now runs logged in, so it reaches pages behind the login and can test whether one account can access another's data.
• The Defects screen now shows how many suspicions are on record, even when none have been reproduced yet.

Fixed

• The attack-surface scan reads as Scanning the attack surface in the agent's output list, and sits with the rest of the testing. It was the one kind of work the list had no name for, so it showed as scan surface under its own task heading.

New

• Heal 2.0 is rebuilt around a new agent. Its mission is to find and reproduce frontend, behavior and runtime security defects against your apps thanks to the Heal sandbox. We improved the sandbox to be deeper, built faster and agent first. The user experience to interact with the agent was improved, no more plan, no more AI slop, only sharp questions for an accurate sandbox.

New

• You can run the agent on your own Claude subscription instead of Heal's. The toggle is in Settings → General, under LLM routing, and inline in the account menu so you can flip it without leaving what you're doing.
• Settings has an Updates tab, with the version you're on and updater status.

Improved

• Plans are now much more compact and readable. Also, the agent now asks for input on plans only when really needed.
• Then run feed now shows everything that happened in the workspace, instead of just the current conversation. You can filter it for recent items.
• Plans are kept in the conversation after you've answered instead of disappearing.

Fixed

• Task headers, plans, and the end-of-run summary used to disappear when you closed Studio or switched sessions. Reopening a conversation brings the whole thing back.
• Fixed a bug where "Restart now" got stuck when an update was available.
• Failed update checks now show a warning instead of failing silently.

Improved

• The agent writes better tests and uses Playwright best practices (assertions, locators, timeouts) more rigorously.
• Improved the user interaction in planning: you can now ask for clarification on agent questions and iterate from there.

New

• Introducing the live view: see exactly what the agent is working on. If it's working on a test, you'll see that test.
• Ask the agent anything. "Fix this bug and re-run the tests", "the code changed, update the tests", "tidy up this folder" now work. Heal works in your workspace only, and asks for permission before editing your source code.
• When the agent finishes its tasks, it waits for you instead of closing and forgetting the session. Use /quit or ctrl-c to end a conversation when you're done.
• Talk to the agent while it's working! Instead of ignoring your messages, it now reacts to input while running: "what are you doing?" or "stop, do this instead".

Improved

• Tasks that exit now explain why clearly.
• The agent panel can now be resized.
• In git repositories, the agent now ignores your own staged work and only tracks its own changes.

Fixed

• Fixed a bug where slow background jobs caused the agent to lose its results.
• Fixed a bug where answering the agent mid-run wasn't immediately taken into account.
• Quitting Studio cleans up after itself. A run still in flight gets its sandbox torn down before the app exits, and the shells behind your console tabs are closed with them, so a stray container or a background process from the last session isn't there waiting when you reopen.
• Fixed a bug where updates needed to be accepted twice. "Restart now" now updates the app on the first click.

Fixed

• The "update ready, restart now" prompt shows on the sign-in screen too. A signed-out window used to sit on an update it had already downloaded without ever offering to restart.

Fixed

• Signing in brings Studio back to the front by itself, and the browser tab closes behind it. No more finishing in the browser and wondering which window to go find.
• A sign-in that fails now says why, right under the button — an expired code, a timeout, a backend that didn't answer. It used to fail in total silence, which looked like the button doing nothing.
• Fixed a race where finishing sign-in too quickly — already logged in to Heal in your browser, so the round trip is instant — left the app waiting the full two minutes before giving up on a code it had already received.

Fixed

• Fixed a bug that prevented the app from updating automatically.

New

• The agent chat has a status bar that summarizes what the agent is doing in a few words. It also links to the feature, spec, or test that the agent is working on. Know what's happening at a glance.
• The status bar is also available in the CLI. This can be disabled with --no-monitor.
• The product feedback modal accepts screenshots, in case you need to add one in your feedback about Heal. Paste, drag and drop, or browse.
• Test generation is now ~2.7x faster, in exchange of a longer initial sandbox setup time. The agent ceates a more complete sandbox initially, and spends less iterations updating it next.

Improved

• Improved test names. Heal names tests more concisely and descriptively.
• Tests have fewer comments. Previously, comments were too verbose.

Fixed

• "What the agent knows" content now wraps instead of overflowing. We're investigating how to catch that with Heal next!
• On error, feedback submission recommends to download and send the feedback bundle. This can happen when the feedback is very large because of a large sandbox or conversation.

Improved

• Heal is now between 2.4x and 3.1x faster and 2.4x-3.2x cheaper on benchmarks for sandbox generation and test generation.

Fixed

• Fixed a race condition in which the auto-updater would not execute the update after quitting the app in some cases.

Fixed

• Fixed an issue where the agent failed with "ENOTDIR: not a directory, opendir".

New

New

• We overhauled the test generation methodology so that Heal generates deeper tests, faster.

Improved

• Improved performance of displaying long lists.
• The code view in a run's detail handles large files smoothly; it only renders what's on screen.
• Long agent sessions restore faster and stay lighter on memory.
• Traces reopen quicker: we reuse what we've already parsed instead of re-reading every time.
• Trace galleries load their screenshots as you scroll to them.

Fixed

• Fixed an issue where the tests filter dropdowns stopped working when clicked several times in a row.

New

• There's a new mode in the agent to ask heal anything. This comes on top of the add sandbox and run test coverage mode.

Fixed

• Fixed an issue where the infra would appear as errored even if it was up.

New

• Experimental: Run & fix from the Bugs screen, launch an agent that runs your suite, diagnoses each failing test against your changes, and applies a fix to your code or to your tests.
• Added Zed to the 'Open in..." menu, thanks FBienefelt for suggesting.

Improved

• Heal can create tests as out of scope when it needs a more complete sandbox to implement them.
• Heal checks for updates while it's running, so you're prompted to update without restarting the app.
• On Mac, Heal bounces for attention when it needs input.
• Valid mermaid charts now render.

Fixed

• The agent's latest output stays pinned to the bottom as it streams. Scroll up to read back, and a jump-to-latest button brings you home.
• Fixed the agent chat input. The agent prompt grows as you type. Add multi-line input with Shift+Enter.
• Markdown tables in the agent's replies render properly.
• Fixed an issue where the agent input looked too much like a pop-up.
• Notification sounds play only when Heal is in the background, not while you're looking at it.
• Fixed a bug that prevented test sources from appearing in the test view.

Misc

• The agent's waiting messages cycle a little more often.

New

• New agent version: better adheres to instructions and spins up tests faster.
• Bring your sandbox infra up or down straight from the app.
• Run your tests without leaving Heal. Kick off a single test, a feature's tests, or the whole suite, and watch the output stream live — no terminal required.

Improved

• Heal asks logged-out users to log in before starting a session. Fixes an issue where the agent would hang without a clear error for logged-out users.
• You can now copy bugs from the Bugs view to clipboard.
• While the sandbox builds, the status and every build phase are clickable.
• Improved navigation in the sandbox overview. Click on any step to view its details view.
• Cleaner agent panel header: New and History are now icon buttons with hover tooltips.

Fixed

• Fixed an issue where the app would freeze when submitting many screenshots for Visual Defects analysis.
• Visual defect screenshots show up again.
• The back and close buttons on the Settings screen are clickable again.
• A single-action step no longer shows an expand arrow that had nothing to expand.
• Buttons and cards across the agent and sandbox screens show the right cursor on hover.

New

• Heal catches visual defects across your tests.
• Watch your run history in the new Runs screen.
• Caffeinate mode keeps your Mac awake while the agent works.
• Opt in to crash diagnostics so we can fix issues faster.
• Log in from the app to see your token quota and plan.
• Open a new window with a keyboard shortcut.

Improved

• Heal remembers where you were when you reload.
• The layout adapts when you open the agent panel.
• Improved the feature pass and fail tracker to be easier to read.
• You can now open a run's Playwright trace.

Fixed

• Fixed an issue where Heal crashed on quit.
• The Features screen survives missing data.

Removed

• Removed the Copy seed button from the Golden Seed screen.

New

• The Bugs tab lists every bug Heal finds.
• Watch the agent's status, and get a nudge when it needs you.

Improved

• Every view updates live as the agent works.
• You can collapse the architecture legend.

Fixed

• The terminal keeps focus when you expect it to.
• The start command runs reliably.

New

• Open Heal in as many windows as you like.
• Added sort by Kind, Features, and Status to the Tests view.
• Added support for several workspaces.

Improved

• Polished the look across the app.
• Columns no longer cut off their content.

New

• Added a built-in terminal with tabs that stay open as you switch workspaces.
• Added a trace viewer with a multi-tab editor, syntax highlighting, and a file tree scoped to your trace.
• Added a command palette (cmd+p / cmd+shift+p).
• A chat panel to work with the agent: review its plan, approve its actions, and see diffs full screen.

Improved

• Agent sessions retry on their own and pick up where they left off.
• New terminals open in the workspace you're working in.

Fixed

• Removing a repository leaves Heal in a clean state.
• Fixed a range of terminal and chat issues.