Changelog

When something goes wrong, Studio now tells you what happened and how to fix it.

• Every error comes with steps. A short explanation, then what to do, in order. The technical detail goes to a log file instead of filling the dialog.
• Every error has a page. Click More help in the message to jump straight to the matching section of the new Troubleshooting Studio page.
• Two common failures are named properly now. Antivirus software that inspects HTTPS traffic can break Studio's connection — Studio recognises that and retries on its own. And a mistyped server address used to be blamed on antivirus, which no antivirus setting could fix.

• A chat that reached its session token cap used to stop there permanently. It now compacts and carries on — the cap triggers compaction rather than ending the conversation.
• Token spend that was quietly under-reported is now counted in full, so usage reflects what was actually consumed.
• An organization can be given its own Ask AI session token cap and spend ceiling instead of sharing one platform-wide default. The two move as a pair, so a ceiling raised alone can't silently change which limit you hit.

Sending the same file twice for a run — a Playwright retry re-sending a screenshot, or a repeated upload — counted it as a second artifact. The run's evidence then rendered twice, and because usage totals the size of stored artifacts, the same bytes were counted twice.

Re-sending a file the run already holds now reuses the artifact it already has. This affected the CLI, the JavaScript reporter and the Python reporters alike.

• tr upload no longer registers the same artifact twice, and the approval error names a command that actually works instead of a flag the CLI rejects.
• An inline artifact card measures its own document before claiming there is more below it. A three-line table no longer gets a fade and a "Show the whole document" button that did nothing when pressed.
• The app settles on one theme rather than rendering two.

Released as CLI v0.102.1 and v0.102.2.

When a turn failed, the whole reply was discarded: your question sat in the transcript with no answer, no error and — after a reload — no sign that anything had been attempted, even when minutes of real tool work had gone into it.

A failed turn now keeps whatever the model had already produced, because a partial answer beats a blank, and records why it ended. Turns that finish against a ceiling rather than a failure — idle, wall-clock, token cap or a stuck loop — now record which one, instead of mentioning it only in the prose of the answer.

• A link to a page that doesn't exist now renders Page not found inside the app. Signed-in users were previously bounced to Ask AI, which turned every dead link in the product into what looked like a deliberate visit to Ask AI. Start-page preferences also drop three options that were never real destinations.
• Share, wall, invite and account-setup links carry a token in the URL that grants access on its own. Those tokens are now scrubbed from product analytics and from server logs, and session replay is off on those pages. Run identifiers and route names are deliberately kept, so run analytics and path queries keep working.

• Planner is re-navigated around one journey. A progress spine — Intake · Criteria · Build · Run · Sign-off — sits on every plan screen, and one status header carries the verdict counts plus a single next-action button that rewrites itself as you go (Build 3 tests → Run 12 → Sign off). Eight named screens collapse into one visible path; the verdict rules and the vocabulary are unchanged. See Planner.
• Playwright projects now get the SDK fixture wired automatically. Navigations, network requests and console logs are collected by the fixture your specs import — not by the reporter — so a project that was only reporter-wired streamed those three empty while Studio said they were flowing. A run now repairs the import the way it already repairs the config.
• Before a signed-in run uploads anything, Studio states what that framework's SDK actually captures, per framework rather than one generic sentence — Python Playwright, pytest, Appium, Maestro and evals each capture a different set, and each now says its own.
• A plan's share sheet previews the host you are actually signed in to. Until the link exists, the row says it is created on publish instead of offering a URL that leads nowhere.
• Cloud mode gets its Artifacts shelf back, and separates agent chats from plain chat.

Released as CLI v0.101.0 and v0.102.0.

• Studio's two tabs now name the work instead of the network. Code lists the agent sessions on this computer, Cloud the Ask AI chats on your account — so a chat you started on another device finally appears here. A Cloud chat can't run on its own; it hands the ask to Code with your draft carried over and seeded, never sent for you.
• Artifacts moved out of the single run and into the account: searchable, filterable, pinnable against a storage trim that no longer happens in silence, and exportable as PDF, PNG, Markdown, HTML or JSON. Drill-downs inside a document open Studio's own run and test screens — and images inside an artifact render at all now.
• The three places that talk about uploading stop contradicting each other. Uploading follows your sign-in, runs an agent started upload too, and a run card now distinguishes "uploading was declined" from "nobody promised either way" instead of offering to switch on something already on.
• The Reporter pane stops looping: its sign-in button no longer signs you out, the same sentence no longer appears twice in two colours, and the one thing that actually clears the failure is visible again.
• Studio notices when a newer reporter has been published than the one your project has installed, and stays quiet when the check itself fails rather than crying "update available" at a bad network.

Released as CLI v0.96.0 through v0.100.0.