← Back to Updates

Every fix, listed.

A detailed, customer-readable record of 282 fixes in v0.3.177 to v0.3.255. Updated 19 September 2026.

The current version is v4.0.21, released 5 November 2026.The Updates page has a short card for each release: what is new, what improved, and the headline changes worth knowing about, newest first. Updates →

Release

Fixes in v0.3.255

282 fixes across 10 categories.

v0.3.25519 September 2026

Telephone calls

20 fixes
  • v0.3.185 Player One can answer and place telephone calls, with the whole conversation kept in the chat.
  • v0.3.185 A live banner sits above the composer while a call runs, with a button to hang up.
  • v0.3.185 Only the part of a reply meant to be spoken is read aloud, and the rest stays in the chat.
  • v0.3.185 A caller's words appear as an ordinary message, marked as spoken.
  • v0.3.185 A phone line keeps its credentials in the vault, never in a visible record.
  • v0.3.185 A call no longer breaks off in silence when the caller interrupts.
  • v0.3.185 The call card reads all six ways a call can end.
  • v0.3.186 A caller reaches a chat that already exists for them, so the agent remembers the last call.
  • v0.3.186 An abandoned call leaves no half-finished record behind.
  • v0.3.186 A call the agent ends politely is recorded as a goodbye, not a refusal.
  • v0.3.186 A phone line can be changed in place without losing its number.
  • v0.3.187 An agent on a call speaks instead of typing.
  • v0.3.187 A sandboxed agent can speak on a live call.
  • v0.3.187 A call started from the owner's chat runs with that chat's permissions.
  • v0.3.187 Numbers are read in international format, so a caller is matched however they dial.
  • v0.3.187 The Phone lines admin page lists your lines with their recent calls beneath.
  • v0.3.187 A call that refuses to end is retried and then closed.
  • v0.3.187 The ended card says why a call finished, in the provider's own words.
  • v0.3.187 Messaging and Webhooks moved into a Communications group in Admin.
  • v0.3.192 A phone call uses the same model and thinking level as any other chat.

Your own computer

39 fixes
  • v0.3.196 Player One works on your own computer through Sidekick.
  • v0.3.196 Commands run on your computer in a shell that stays where you left it.
  • v0.3.197 You can sign in to Sidekick through your browser.
  • v0.3.197 Player One can list the computers you have paired, and forget one.
  • v0.3.197 Output from your computer is shown, kept and paged under three separate limits.
  • v0.3.198 Player One can capture your screen, and see your displays and windows.
  • v0.3.198 Player One can act on your computer, clicking, typing and dragging.
  • v0.3.198 Player One can work with the files on your computer.
  • v0.3.198 A change of access mode appears in the chat as a single note.
  • v0.3.198 The chat says when a paired computer disconnects, and names it when something fails.
  • v0.3.198 A long-running command tells you the chat carries on by itself.
  • v0.3.198 A request with no computer attached refuses by name instead of failing obscurely.
  • v0.3.199 A computer that cannot do something says so, rather than appearing to be absent.
  • v0.3.205 Four access modes decide how much a paired computer allows.
  • v0.3.205 The chat shows what a computer is allowed to do beneath its mode.
  • v0.3.205 A refusal names the mode you actually see in the picker.
  • v0.3.222 Sidekick has a side panel on your desktop, with its own chat.
  • v0.3.222 A preview opens over the side panel instead of sending you to a browser.
  • v0.3.222 The side panel's header shows the agents in the chat, and drags the window.
  • v0.3.223 A signed-in Sidekick stays signed in.
  • v0.3.229 A refusal from your computer keeps its own reason, and an installation problem is named as one.
  • v0.3.231 The side panel lists your chats from every project, in folders, loading more as you scroll.
  • v0.3.231 The side panel stays on its own page, wherever the site navigates.
  • v0.3.231 The panel's header carries the other agents, and a close button.
  • v0.3.231 The panel's composer is the full desktop composer, at desktop size.
  • v0.3.231 Menus and dialogs inside the panel read at the panel's own size.
  • v0.3.234 A page preview opens a viewer window in the side panel, so a browser session is visible at all.
  • v0.3.237 The panel's chat list is the same list as the sidebar's.
  • v0.3.237 Menus, submenus and popovers stay inside the panel's column.
  • v0.3.237 The panel's composer takes the desktop controls in a flat frame.
  • v0.3.237 A mark in the panel's header starts a new chat, and says why one waits under All projects.
  • v0.3.238 The panel signs in through the extension that frames it.
  • v0.3.238 The parent avatar in the panel opens the parent chat.
  • v0.3.242 A change to what a paired computer allows takes effect while it stays connected.
  • v0.3.245 A face in the agent strip opens that chat inside the Sidekick panel.
  • v0.3.248 An action on your computer checks itself and reports the new state in one step.
  • v0.3.251 A command on a paired computer reads as finished when it finishes.
  • v0.3.251 Typed text keeps its edges, and a refusal no longer suggests the thing it just refused.
  • v0.3.255 A text area on a Mac can be reached by its role alone.

Chat and conversations

54 fixes
  • v0.3.177 The compaction notice offers to start a new chat.
  • v0.3.183 Opening the side preview no longer moves you earlier in the thread.
  • v0.3.183 A Preview control opens the workspace pane.
  • v0.3.186 A new chat opens on a redesigned welcome screen.
  • v0.3.186 The compaction marker sits at the end of the transcript.
  • v0.3.187 Your own passthrough chats stay in the sidebar and in Recent chats.
  • v0.3.189 A live progress line replaces the activity boxes.
  • v0.3.189 The progress line says when the connection drops.
  • v0.3.189 A prompt card reads the project you are in, not one selected elsewhere.
  • v0.3.192 A light rule separates the messages of one reply.
  • v0.3.192 A finished answer beneath a run still going reads as in progress.
  • v0.3.194 The mobile Photo button keeps the photo it was given.
  • v0.3.195 No sideways scrollbar when the window is too narrow for avatars.
  • v0.3.196 The progress line opens into a log of what is happening.
  • v0.3.196 Steps are coloured by kind, so tools and skills are told apart at a glance.
  • v0.3.197 A yielded command no longer splits the reply into separate cards.
  • v0.3.197 The progress line names what a tool is acting on.
  • v0.3.197 Queued messages arrive one at a time instead of as a single block.
  • v0.3.217 A shortened command result says exactly where to continue reading.
  • v0.3.219 A helper waiting on your approval tells the agent it works for straight away.
  • v0.3.224 Agent faces sit over the composer, replacing the cards that were above it.
  • v0.3.224 A chevron separates the parent chat from the agents it started.
  • v0.3.226 A keyboard shortcut starts and stops a voice note.
  • v0.3.226 Voice options open in a wide panel, and recording pauses while it is open.
  • v0.3.227 The composer's effort chip shows the level the next turn would actually run at.
  • v0.3.227 A narrow window keeps the input focused after you send.
  • v0.3.229 Fold and email notices read as lines in the chat rather than messages from an agent.
  • v0.3.230 A timer's report reaches the channel it was set from.
  • v0.3.231 An agent working in the background wears a turning ring on its face.
  • v0.3.231 A chevron marks the parent chat in the row of agent faces.
  • v0.3.231 The agent faces above the composer sit better, and lift as you point at them.
  • v0.3.231 A voice note shows a moving level meter while you record.
  • v0.3.236 The chats an agent started are grouped under it in the sidebar, as a tree you can open.
  • v0.3.236 A cog on each chat carries its options in one place.
  • v0.3.236 The sidebar filters by what a chat is, so scheduled and flow runs can be listed on their own.
  • v0.3.238 Pickers open as bottom sheets on a phone, and in the side panel.
  • v0.3.238 The model, effort and agent pickers show a card as you point at a choice, and the model picker ticks the one in use.
  • v0.3.238 Keyboard focus shows a soft ring instead of the browser's black outline.
  • v0.3.238 The row of agent faces no longer reserves empty space above the composer.
  • v0.3.206 An expired approval says so, and a late click still gets an answer.
  • v0.3.207 Messages sent while a chat is opening wait for it instead of being lost.
  • v0.3.209 An approval card's command scrolls on a phone instead of being cut off.
  • v0.3.209 An approval card hides values that look like secrets, and leaves the rest readable.
  • v0.3.239 A reply that folded several model calls says so, and names its last prompt.
  • v0.3.240 Compaction reads as one card, from the request through to the result.
  • v0.3.240 The stop button is disabled while a conversation is being compacted.
  • v0.3.241 A new chat's effort chip reads the route rather than a default.
  • v0.3.243 Tool result panes show their size.
  • v0.3.246 The notice shown while an installation is away says what it usually means.
  • v0.3.248 A screenshot already shown stops re-appearing as a long reply grows.
  • v0.3.248 A screenshot step keeps its tool call and its thinking once the picture is attached.
  • v0.3.248 Streamed text keeps pace with a fast model instead of landing all at once.
  • v0.3.254 Menu and popup text stays sharp, with ClearType kept on Windows.
  • v0.3.254 A selected helper in the sidebar keeps its whole level in view.

Jobs, tasks and automations

25 fixes
  • v0.3.177 A coordinator wakes as soon as a worker finishes.
  • v0.3.177 Creating a job refuses an invalid workspace straight away.
  • v0.3.183 Stop, resume and close a job from the job page.
  • v0.3.183 The job board's third column says Finished.
  • v0.3.183 Stopping a job names the person who stopped it.
  • v0.3.183 Scheduled tasks pause and resume, rather than being enabled and disabled.
  • v0.3.183 Flow runs survive a restart instead of being discarded.
  • v0.3.183 A resumed flow that makes no progress stops instead of spinning.
  • v0.3.189 A throwaway flow's last build can be removed.
  • v0.3.205 A scheduled run takes its permissions from the task's owner.
  • v0.3.206 A stopped worker cannot overwrite its own record.
  • v0.3.207 A schedule that cannot happen is refused as you enter it.
  • v0.3.207 An impossible schedule reads as never, not as every two seconds.
  • v0.3.217 A scheduled task keeps the thinking level you chose for its chat.
  • v0.3.217 A scheduled task that names a model your installation does not allow fails by name.
  • v0.3.207 A scheduled run does its work in the project its task belongs to.
  • v0.3.209 Removing a schedule takes its project membership with it.
  • v0.3.221 A job in which no task was ever done no longer reads as complete.
  • v0.3.221 A failed task parks its job instead of leaving it adrift.
  • v0.3.221 Finishing a run names the runs still keeping a job open.
  • v0.3.228 A job with live work or an open review is not cut short by the resume cap.
  • v0.3.230 A refused worker request is recorded rather than passing in silence.
  • v0.3.231 A job can be given its own limit on how many times it retries.
  • v0.3.231 A parked job is flagged once, not on every check.
  • v0.3.241 A deleted job's queued announcements go with it.

Agents, skills and tools

21 fixes
  • v0.3.177 Changing a chat's project reloads the skills for that project.
  • v0.3.183 A skill's notes are shown on its detail page.
  • v0.3.183 The Skills page says where a skill came from.
  • v0.3.183 The Skills page no longer claims a control it does not have.
  • v0.3.185 Searching skills stops re-reading the whole workspace each time.
  • v0.3.186 Ten shipped skills whose details never loaded now work.
  • v0.3.187 A stale tool can be removed from one agent, or from every agent at once.
  • v0.3.187 An agent's card says when it is sandboxed.
  • v0.3.187 Deleting an agent offers to delete its Memory Space with it.
  • v0.3.187 A skill's emoji shows in search results.
  • v0.3.189 A permission entry that matches nothing explains why.
  • v0.3.194 A turn that resumes after an approval keeps its tools.
  • v0.3.196 A project skill is created with the visibility you asked for.
  • v0.3.216 Every agent can read the model catalogue and switch its own session.
  • v0.3.216 A worker runs on the model named by whoever started it.
  • v0.3.244 A map of the helpers working on a chat, opened from the row of agent faces.
  • v0.3.244 The helper tree shows every level beneath the chat you are in.
  • v0.3.244 A helper that is working wears a turning ring on the map, as it does in the strip.
  • v0.3.244 Helper popups and sidebar rows lead with the chat's own name.
  • v0.3.245 The sidebar keeps the parent's row and its helper list in view while a helper is open.
  • v0.3.245 A helper opens like any other chat.

Memory, knowledge and search

13 fixes
  • v0.3.183 A healthy search index is no longer reported as failed.
  • v0.3.183 Search never drops a source you asked for without saying so.
  • v0.3.185 An agent searches its own memory and only its own.
  • v0.3.185 Two files are two results, however similar their openings.
  • v0.3.186 This installation's own Fold projects can be connected as knowledge sources.
  • v0.3.186 A knowledge source on our own service needs no key of your own.
  • v0.3.187 A knowledge source whose project has gone is marked, skipped and removable.
  • v0.3.187 Removing a Memory Space asks whether its search index goes too.
  • v0.3.187 A scan finds a Memory Space wherever it lives.
  • v0.3.226 A knowledge source is checked with the credential that can actually see it.
  • v0.3.229 Searching for a skill and loading it now agree on what exists.
  • v0.3.230 A run that ends unverified parks only its own Memory Space.
  • v0.3.231 A Memory Space still being indexed reads as indexing, not as failed.

Models and providers

28 fixes
  • v0.3.177 Routing explains the decisive reason it fell back to another model.
  • v0.3.177 A billing refusal is shown as billing, and keeps the provider's own remedy.
  • v0.3.182 Provider tabs follow the order you set.
  • v0.3.182 The status panel says when it last checked, and offers to look again.
  • v0.3.182 The status panel stops blinking, and reads at a normal size.
  • v0.3.186 Each model in a routing chain gets its own thinking level.
  • v0.3.186 A missing model is reported as missing, not as an unreadable error.
  • v0.3.187 The effort controls line up down a chain.
  • v0.3.189 A provider you switched off stops being offered.
  • v0.3.189 A connected xAI account shows its Grok Imagine models.
  • v0.3.216 New models this release: Gemini 3.8 Flash, GPT-6 Astra, GLM-5.3, GLM-5.3 Flash, Qwen3.8-Flash-Next and Fable 5.1.
  • v0.3.217 A provider failover does not overwrite the model you pinned.
  • v0.3.218 A model refusal returns as a clear refusal rather than an unexplained failed run.
  • v0.3.221 An unknown thinking level is refused plainly.
  • v0.3.227 New model this release: DeepSeek V4.1 Flash.
  • v0.3.227 The retired DeepSeek V3 names are gone from the catalogue.
  • v0.3.234 A run that falls back to another model uses the right thinking level for the model actually serving it.
  • v0.3.236 An Anthropic account on extra usage credits is no longer reported as out of usage.
  • v0.3.236 A refusal about your plan names the ChatGPT subscription rather than Codex.
  • v0.3.237 A model one account's plan lacks is tried on your other accounts.
  • v0.3.237 The signed-out notice names the provider and the account.
  • v0.3.238 Retired models no longer linger in the catalogue after an update.
  • v0.3.246 On Anthropic models the tools sent with every request sit in one stable cached block, so a long conversation costs less and starts faster.
  • v0.3.246 Loading a tool part-way through a conversation no longer rebuilds that cached block.
  • v0.3.240 A substituted model resolves its own thinking level.
  • v0.3.248 You confirm a provider's own data terms before that provider is connected.
  • v0.3.248 TypeSafe can be connected, with a vendor-neutral tool for judging output.
  • v0.3.248 A Grok chat is no longer refused when the judging tool is present.

Admin, settings and security

29 fixes
  • v0.3.183 Gateway is now called Runtime, and Instance is now called Installation, throughout the interface.
  • v0.3.183 The Settings page stops having three names for the same thing.
  • v0.3.183 You can test a connection from the page where you added it.
  • v0.3.183 The help panel is called Help.
  • v0.3.183 Plain words on the Backups import tab.
  • v0.3.183 API keys read the token the runtime actually sends.
  • v0.3.183 A hosted link is handed over with a privacy notice.
  • v0.3.185 An approval message never shows the raw command.
  • v0.3.187 A command recorded in the chat never shows its secrets.
  • v0.3.196 Memory pressure is measured against one ceiling, and notices name it.
  • v0.3.201 A vault refusal names the field and the shape it wanted.
  • v0.3.217 A local settings fault ends the run without wrongly blaming a provider.
  • v0.3.217 The default agent cannot be accidentally made sandboxed.
  • v0.3.206 Reading another program's memory needs a grant you give.
  • v0.3.206 Every command runs as the ordinary runtime user.
  • v0.3.207 Deleting a vault item names what still uses it, and one action clears them.
  • v0.3.209 A stored permission entry never keeps the raw command line.
  • v0.3.235 Prompt capture is a debug switch that writes outside your workspace, and is off unless you turn it on.
  • v0.3.235 A one-time sweep takes old prompt captures out of customers' workspaces.
  • v0.3.235 Prompt captures never cross a workspace sync, and never reach cloud storage.
  • v0.3.235 A Prompt capture switch sits behind Advanced in Settings.
  • v0.3.236 The tool catalogue reports what each permission profile grants.
  • v0.3.236 The deny picker offers every tool an agent can actually call.
  • v0.3.236 A tool configuration update keeps the overrides you set.
  • v0.3.238 The Admin menu opens with a Back to site row.
  • v0.3.238 A vault injection the runtime does not recognise is named rather than silently dropped.
  • v0.3.252 A command approval can be answered for five minutes rather than two.
  • v0.3.252 An Always allow answer sits on the approval card and in the operator alerts.
  • v0.3.253 An Always allow answer never grants an interpreter.

Files, backups and media

18 fixes
  • v0.3.177 Dropbox team folders that could not be reached now sync.
  • v0.3.183 The agent browser opens maximised.
  • v0.3.186 The browser has a memory limit and a time limit for page actions.
  • v0.3.187 Chats are included in backups, with their own tab and a chat history box.
  • v0.3.187 A chat in Trash is not treated as a clash when importing.
  • v0.3.187 A chat exported without secrets can be imported again.
  • v0.3.187 A move that meets an out-of-date record refuses instead of quietly doing nothing.
  • v0.3.189 An agent can delete a folder tree it created.
  • v0.3.189 A generated image names its type and size.
  • v0.3.191 A generated image keeps its location for the rest of the turn.
  • v0.3.192 The Drive connection is ready at start-up rather than on first use.
  • v0.3.225 Microsoft 365 mailboxes can be connected, and stay connected.
  • v0.3.228 Audio transcription no longer invents a prompt you did not ask for.
  • v0.3.230 Trash follows the scope in the header, and All projects can be emptied.
  • v0.3.201 A file error names your own path, not an internal one.
  • v0.3.204 A chat created and never used can be removed.
  • v0.3.237 Cloud storage syncing runs under a memory ceiling, so a large sync cannot crowd out everything else.
  • v0.3.238 A text file opened from a workspace keeps its accents and symbols.

Speed and reliability

35 fixes
  • v0.3.177 Dozens of refusals answer with a sentence instead of a bare code.
  • v0.3.181 Creating, updating and deleting agents works on a running installation.
  • v0.3.183 Start-up repairs its own settings instead of failing.
  • v0.3.184 Hundreds of messages that printed a meaningless value now print the real one.
  • v0.3.186 One event-loop monitor per installation, however many times it reloads.
  • v0.3.189 Health checks answer before anything else at start-up.
  • v0.3.189 Team chat attributes every teammate's message, not just the newest.
  • v0.3.194 A chat with no messages yet reads as empty rather than as an error.
  • v0.3.194 A busy database waits its turn instead of holding everything up.
  • v0.3.195 A chat run that failed reports failure.
  • v0.3.195 Compaction attributes teammate messages correctly.
  • v0.3.196 Every installation tells its clients which build it is running.
  • v0.3.198 Long conversations no longer stall the runtime while being read or written.
  • v0.3.198 A brief disconnection no longer removes a computer that has reconnected.
  • v0.3.198 The reconnecting screen's logo loads while the runtime is down.
  • v0.3.198 A teammate's reply reaches someone who already has messages waiting.
  • v0.3.219 A settings write-lock left by a stopped installation is reclaimed safely.
  • v0.3.220 Work the platform runs on your behalf is not blocked by a permission you never set.
  • v0.3.221 A run interrupted by a restart resumes on the next start-up.
  • v0.3.226 One faulty account no longer takes the whole channel status down.
  • v0.3.228 A background run that timed out says so, instead of reporting success.
  • v0.3.228 A cost refusal names its cause, its numbers and whose allowance it is.
  • v0.3.230 A failed import names its cause, with the sensitive part hidden.
  • v0.3.231 One channel that cannot describe itself no longer takes the whole status summary down.
  • v0.3.231 Health files are replaced cleanly, so a slow check cannot corrupt them.
  • v0.3.235 Long chat lists are read a page at a time instead of loading everything.
  • v0.3.235 Clearing out expired runs uses an index rather than reading every record.
  • v0.3.237 Dozens of operations stopped reading the whole chat store to change one thing.
  • v0.3.238 A connection that times out while starting says so plainly.
  • v0.3.238 A voice note added between turns is no longer treated as an interrupted reply.
  • v0.3.242 A background command's completion still reaches the chat after a restart.
  • v0.3.242 A chat gets at most six background-command notices an hour, and the rest as one summary.
  • v0.3.243 A search probe that times out no longer reports the index as down.
  • v0.3.244 The last whole-transcript reads leave the path every turn takes.
  • v0.3.245 Clearing out old sessions keeps to its own steady cadence.