When a saved login session expires mid-run, whether it's a password-based profile or a magic-link/OTP one, Simz now catches it, flips the account to "needs re-auth" in Settings, and posts an actionable notice with a direct link on the GitHub PR check, instead of leaving you to guess why a run failed.
Manually verifying a login by pasting an OTP or magic link no longer times out on you mid-paste: the resolve window widened from 2 minutes to 5.
If a deploy causes a brief sync hiccup, an affected run now fails as "errored" right away instead of getting stuck at "Queued" indefinitely.
The dashboard no longer shows a leftover "LOCAL" label in the top bar, including in production.
Sign-up verification emails now look like Simz instead of a generic default template: they adapt to light and dark mode, with an amber dashed-border VERIFY button, the Simz logo, and matching typography.
A "Forgot password?" link on the sign-in screen sends a branded reset email, and never reveals whether a given address actually has an account.
The dashboard now shows which account you're signed in as: your email appears next to the Sign out button, truncated with a tooltip for long addresses.
One "Continue" flow now handles both signing in and creating an account instead of separate tabs, email verification is detected automatically after you click the link with no extra step, and the page carries Simz's real logo.
Recorded steps that type a literal value into a field, not just a %variable%, can now be cached and replayed on future runs instead of re-exploring the page every time. Anything that looks like a secret (passwords, tokens, API keys, and similar) is automatically excluded, and literal values are only ever cached per test, never shared across tests.
Run reports now show a "Cache:" line with per-test hits, shared hits, re-explores, and misses, and two new Settings checkboxes, "Shared step cache" and "Verify replay fingerprint", give direct control over step-cache behavior (both off by default).
Settings is now split into three tabs (Run defaults, Authentication, and Display) instead of one long stacked page, with per-target login credentials collapsed into an accordion. Tabs are deep-linkable, and the Authentication tab shows an amber dot when a target has a login challenge waiting on your input.
Run logs no longer show a spurious "prompt injection risk" warning on every single AI-driven step, while genuine warnings still surface normally.
The Simz wordmark in the top navigation is no longer crowded against the edge.
Targets that use a magic link or a one-time email code instead of a password can now be set up and verified: pick the auth method per target, and when sign-in needs a human to click a link or type a code, the dashboard shows a paste box right where you'd expect it. A stalled verification times out to a clean failure instead of hanging forever, and the one-time code or link is never stored in readable form.
Once a target's login is verified, later test and discovery runs reuse that signed-in session instead of repeating the full sign-in flow, including SSO, every single run. This also fixed screenshots that would occasionally 404 after a run finished.
The dashboard now checks for updates only as often as actually useful. The runs list replaces its always-on auto-refresh with a manual Refresh button (with a live elapsed-time ticker while a run is in flight), and live run reports and the tests list poll less frequently in the background.
Costs across the dashboard can now show as simple "credits" instead of raw dollar amounts, making run costs easier to read at a glance. A Settings toggle switches every cost surface between credits and real dollars, remembered per browser, and defaults to credits.
A new Integrations tab replaces the old GitHub-only page: GitHub shows as connected, alongside Webhook and Jira as "coming soon" cards. The old GitHub page redirects here automatically.
Rows in the runs list now display the test's name instead of just an opaque ID, making run history much easier to scan. Multi-test runs show the joined names, and a deleted test simply falls back to no label.
Fixed the Simz logo floating off-center on the login screen, and the browser tab icon now automatically matches light or dark mode.
Fixed the budget number inputs in Settings visually overflowing into neighboring fields.
Every step in a run report can now show why it executed the way it did: replayed from the shared cache, re-explored after a stale cached selector, skipped a cached entry because the page changed underneath it, or escalated to a smarter (and slower) execution tier. These show as small tags next to the tier badge, so a cold run and a re-healed stale replay no longer look identical.
GitHub pull request checks now run against a real, always-on cloud service instead of a local machine, so the full webhook-to-check-result pipeline works end-to-end in production. Engine and browser failures are now logged and surfaced on the run, instead of silently producing a blank "errored" result with no explanation.
The dashboard is no longer wide open: it now sits behind a shared access key, with a lock screen prompting for the key before any data loads. The key is entered once and cached in your browser after that.
Canceling a running test no longer risks taking down the entire test runner and stranding every other queued run. The browser now closes immediately and the rest of the queue keeps running.
API calls now route correctly, and refreshing any dashboard page works instead of 404ing.
Simz now watches your GitHub repo and automatically runs the relevant tests, chosen by which files the PR actually changed, on every pull request, posting a pass/fail check and a sticky summary comment right on the PR as soon as the run finishes. Quarantined and flaky tests still report in the summary but never fail the check.
A new "Connect GitHub" page links a GitHub repository to one of your targets, so PR checks know which tests belong to which repo. It lists active connections with their default branch and route-rule count, and lets you add, update, or delete a binding.
Crawls now produce a more complete, less cluttered map out of the box. Hover-menu exploration, which could add 700 to 1,700+ mostly-noise entries per crawl, is now off by default, while vision-assisted page scanning (which catches icon buttons and tabs that plain text scanning misses) is on by default.
Sidebar and top-nav links that appear on every page no longer clutter the map with a duplicate copy per page. Anything seen on 3+ pages is consolidated into a single "Global navigation & chrome" node instead of being duplicated across every page.
Simz gets a working web dashboard: create targets and tests, edit them, trigger a run, and watch it flip from queued to running to passed or failed live. Opening a run shows the full report, a per-step timeline, screenshots, and a cost/token breakdown. Repeat runs of the same test can now replay without needing the AI model at all, making them faster and cheaper.
Pick which Gemini model powers a test as a default, or override it for a specific run, right from the Run dialog. Four models are supported, each with its own pricing, and the pass/fail judge always uses the same model the run executed with so cost and results stay consistent.
Stop a run while it's still executing instead of waiting it out, and cap a run by total token usage in addition to cost and step count. A run stops cleanly between steps, and killing a still-queued run releases it instead of letting it start.
The automatic goal-check that grades a run now judges the outcome of the entire task, not just its last action, so it correctly catches cases where individual steps "worked" but the overall goal wasn't met. It also now rejects unverifiable comparative claims, like asserting something is "the cheapest one" without a way to check.
Every edit to a test is saved as a new version, and each run records exactly which version it executed. Past versions can be listed, viewed, restored, or deleted right from the test editor.
Optionally capture the agent's step-by-step reasoning during a run (toggle it per run or set an account-wide default), and replay a run's screenshots with a cursor that lands on the exact element that was clicked. Captured reasoning is scanned for secrets before it's stored.
Runs could previously come back marked passed even when the action silently failed. Simz now checks its own work: it derives a concrete success condition from the action and grades it against the real, final state of the page, failing closed on any uncertainty instead of trusting the agent's own claim of success. This verification now runs for the primary task goal too, not just follow-up actions, and the dashboard flags any failed step with a red badge and the reason why.
Every test run now captures visual proof: a screenshot at each step, a full browser trace, a recorded video, and the agent's own action history, so you can see exactly what happened instead of just a pass/fail label. Evidence is served via short-lived signed URLs rather than public links, a failed screenshot never fails the underlying test, and secrets are redacted from the action history before it's saved.
A new web dashboard lets you browse past test runs and drill into a single run's report: status, timing, cost, and a step-by-step timeline with inline screenshots. Each step also shows its execution tier, cache hits, masked inputs, and assertion results.
Runs kicked off manually, on a schedule, from a pull request, or from auto-generated tests now all flow through the same engine, so every run gets the same retry behavior, budget limits, and evidence capture regardless of how it started. PR-triggered runs prioritize tests tagged critical, a sharding option splits a large suite into balanced parallel runs, and infrastructure hiccups like a browser failing to launch retry with backoff before a run is marked errored.
A single command can log in and then perform a follow-up action with --then="<action>", and --profile=<name> saves a signed-in browser session by name so cookies and session state survive across later runs.
Typing a new value into a field that already has content no longer appends to it: the field is cleared first, so the value is replaced cleanly. New --set-field / --set-value flags add a deterministic fill step for cases where precision matters, auto-verified against the exact requested value.
A test kicked off from the command line now appears on the dashboard while it's running and after, instead of only living in the terminal that started it.
AI test agents plus real QA engineers, from $300/mo. New teams get 5 free test runs.