Foundry Studio release log ========================== Foundry Studio (https://app.foundrystudio.dev) runs AI agents and agent work for teams: hosted agents and threads, workers on your own computers, the `studio` CLI and the Athena terminal app. This is the public log of what changed, for agents and people using the platform. It is newest first: one entry per release, "## RELEASE (date)", a one-line headline, then what changed and how to use it. Read from the top until you reach a release you already know. "Off by default" means a workspace manager or the Studio's host has to turn the feature on. Enterprise Studios may run an older release; their workspace managers see an upgrade bar when they are behind. For the full guide run `studio explain`, or read https://www.foundrystudio.dev/llms.txt. Update the CLI with `studio update`, or install it with: curl -fsSL https://app.foundrystudio.dev/install.sh | sh ## 2026.11.22 (2026-11-22) An agent can work with your own notes folder in Dropbox, and Studio can ring your phone. - A Studio's host can connect one person's Dropbox notes folder (/Brain) to the agents they choose. Those agents get brain_list and brain_read (anything in the folder) and brain_drop, which only ever adds a new note to the folder's inbox: nothing is edited, replaced, moved or deleted. Off by default; other agents, and other people's conversations, are refused. - Studio can ring your own phone and put the call in a thread: the thread's agent talks with you in its call voice, and what's said goes into the thread like any call. Set the number it rings with `studio phone set +447700900123` (only you can), then `studio call THREAD --phone`. Off by default: the Studio's host turns phone calls on. - Agents given the phone tools can make a short check-in call (call_me), once: no redial. If you don't answer, they send one text instead and note it; if you do, they note what you said. - `studio schedules pause ID`, `resume ID` and `run ID` (fire one now, as a test). CLI 0.35.0. ## 2026.11.20 (2026-11-20) Reports from a connected platform's assistant say what they are, even with no free text. - A platform that may only send ids once a chat has read records (connected platforms with strict privacy rules) can now send, with report_problem: kind (bug, feature_request or question), conversation_ref (the conversation's id and its https link on the platform) and record_refs (up to 20 records the chat read, as model name and uuid). Studio checks each as an id or a fixed choice and refuses anything else, including free text. - The item shows a link to open the conversation on the platform, where the platform's own access control applies, and the records it read. A feature request becomes a feature and a question a task labelled "question"; only bugs are matched against Sentry and joined to an open bug. - When the summary is the generic "Problem reported on ", the item is titled from the reporter's own first line instead, and repeats are matched on that title. - Items a link files say "via link , reporter " instead of naming the person who set up the link, on the item page and in studio sdlc item show. - Managers can list the platform's web hosts on a link (app_hosts); a conversation link on another host is left out and named in the answer's "ignored". The report tool's schema is version 2; any argument it doesn't list is refused. ## 2026.11.19.2 (2026-11-19) Fewer false decisions: report-only work items finish, and the counter counts only yours. - A work item whose agent rightly changed no code (a daily test run, a report, a check) used to go on to the merge step, fail there because it had no branch, retry six times and then ask its owner to decide. It now finishes when the merge step finds no branch for it. Items that pushed commits, or that fail to merge for any other reason, behave as before. - The Decide counter beside the logo shows only decisions waiting on you: work items you own and questions in your own threads. What waits on other people in workspaces you can see is named in its tooltip and still opens from there. The Decide page itself lists everything, yours first, as before. ## 2026.11.19.1 (2026-11-19) The workspace assistant can queue work for workers, see what they do and ask your Orchestrator. - Ask the workspace assistant (the "Studio" agent each workspace starts with) for something to be built, fixed or looked into by a worker. It checks the work queue for the same request, then adds an item and gives you its number (W-12). - Ask it what the workers are doing: it reads the queue, the workers and any one item, and can pass your answer to a worker's question. It pauses, stops or retries an item only when you ask. - For anything more (pointing an item at one worker, starting a worker, looking at a worker's checkout) it asks your Orchestrator, in your Orchestrator conversation, and reads the answer back to you. That conversation is also in your thread list, and in the terminal as `athena orchestrate`. - It can leave a message in a person's or a worker's inbox, as `studio tell` does. Inbox messages wake nobody. - A new conversation with it already knows which project it is in and that project's repository, so "the codebase" needs no explaining and it no longer asks which one. - It acts as you, with your access, and only in the thread's workspace. Filing an item in a project's delivery tracker still works as before. - Studio hosts: run `manage.py studio_default_agents --apply` after migrating, so existing built-in assistants get the new tools and instructions. Customised assistants keep theirs. ## 2026.11.19 (2026-11-19) Project code access: a "Can read code" switch per member, and the workspace assistant reads code. - Project settings, members: each member has a "Can read code" switch. Workspace owners and admins can always read a project's code; other members only when it is on. People already in a project keep access; anyone added from now on starts with it off. A member who creates a project can read its code. - The switch covers every agent that reads a project's code, including Code questions. - The workspace assistant can now list and read files in the project's repository itself, and hands searches to the Code questions agent. For a question that needs real digging it offers a work item for a worker. Without access it says so and who can switch it on. - Studio hosts: run `manage.py studio_default_agents --apply` after migrating, so existing built-in assistants get the new tools and instructions. Customised assistants keep theirs. ## 2026.11.18 (2026-11-18) Connected platforms: a grid of who can do what, one link per box, and why a call was refused. - Workspace settings, Integrations, Connected platforms: each link now shows a grid. Rows are what the other platform's assistant can do, columns are that platform's roles, and each cell is a tick box. Existing links keep their id, secret and access. - "Report bugs" can be given to a role on its own. It covers a person's report whichever report tool the platform uses; "Report on its own" is separate. - A new link starts from safe defaults: admins everything, other roles little or nothing. - A link has a name you can change, an environment (production, pilot, dev, test) and a box. "Duplicate" makes a link with the same grid and its own new secret. - A new secret is shown once as STUDIO_LINK_URL, STUDIO_LINK_ID and STUDIO_LINK_SECRET, each with a copy button, and "Copy as .env". "Test link" signs a call that does nothing. - Recent use is a table: when, who, role, capability, result and the reason a call was refused (role not allowed, daily limit, bad signature, clock skew), with the work item a call filed. Filter by result and capability; see each person's writes today. - Each link says where reports land: the project's backlog, and Linear when it is synced. ## 2026.11.17 (2026-11-17) Fair-use limits for Studio Support and for new sign-ups. - Studio Support (`studio support`) answers at most 20 questions an hour and 60 a day for each person; people whose workspaces are all on the Free plan get 5 an hour and 20 a day. When very many people ask at once it says it's busy: try again in an hour. - Sign-up with Google can pause for the rest of the hour when a great many accounts are made at once. Signing in is never paused. ## 2026.11.14.1 (2026-11-14) CLI 0.34.0: workers are ready for the doorbell. - Workers on CLI 0.34.0 wait on a Studio's doorbell when the Studio offers one, and otherwise ask for work every 20 seconds exactly as before. No Studio offers it yet. - Update with `studio update`, or `studio worker update` for running workers. Older CLIs keep working. ## 2026.11.14 (2026-11-14) Groundwork for workers that wait quietly instead of asking every 20 seconds. - Nothing changes yet: workers ask for work and check in as before. - A Studio's host can now let chosen workers wait on a "doorbell" instead, so they hear about new work within a second and ask far less often. It is off by default and starts with a few named workers. Workers fall back to asking every 20 seconds by themselves whenever the doorbell isn't there. - Update the CLI with `studio update` when the next CLI release is out; older CLIs keep working as they do now. ## 2026.11.12.2 (2026-11-12) Connections shows one connection at a time. - Settings → Connections lists the connections by name across the top. Pick one (Slack, Google, GitHub, ...) to see only its setup; nothing else is shown until you do. - A link to one connection opens it directly: Slack's /studio link, `studio connect` links, and coming back from connecting GitHub, Linear or Google. ## 2026.11.12.1 (2026-11-12) Links into a workspace's Connections open Connections again. - A link to a workspace's Connections page (Slack's /studio link, `studio connect` links, returns from installing an app) opened the workspace's home instead. It now opens Connections, so "Link my Slack" from /studio works. ## 2026.11.12 (2026-11-12) See at a glance how many decisions wait on you, and get them in your own Slack DM. - Beside the logo in the web app, a counter shows how many decisions wait on you across your workspaces. It lights up when there are any; click it to open Decide and go through them one at a time. With decisions in several workspaces it lists them; when one is done, Decide offers the next. - A question waiting on you now goes to your own Slack DM (once you've linked your Slack), not the workspace's group channel. Failures, landings and deploys still go to the channel. ## 2026.11.11.8 (2026-11-11) Tidier messages: one email when an agent writes to you, and work stays off the bus. - When an agent emails you from a thread, you get one email, not two if your notification emails are on. - Agent work no longer posts lines to the project bus. You hear of work in your inbox (and Slack, if connected), and `studio work events` lists it. ## 2026.11.11.7 (2026-11-11) Decide shows everything on the Board's "Needs you", not only questions. - Items flagged for a look, paused, or failing again and again come after the questions, one card each, with plain choices: carry on, resume, try again or stop. Something else leaves a note on the item for its next worker. - The Board's "One at a time" button opens Decide whenever anything needs you. ## 2026.11.11.6 (2026-11-11) Link your Slack to Studio from Slack. - Type /studio in Slack, or send the Studio app a message: it replies, only to you, with a link. Open it signed in to Studio and press Link my Slack. /studio unlink undoes it. ## 2026.11.11.5 (2026-11-11) Decide can check whether a question still matters, and close the ones that don't. - "Still relevant?" (key R) asks the workspace's Orchestrator to look at the current state and say whether the question still needs an answer. Its reply appears beside the card, and you can talk to it there. - "Close…" dismisses an agent's question (and can archive or delete its thread), or stops a worker's item. Old questions say how old they are. ## 2026.11.11.4 (2026-11-11) Workspace links open again. - Opening a workspace page from a link, a new tab or a reload could say "This workspace is unavailable". Fixed. ## 2026.11.11.3 (2026-11-11) CLI 0.33.2 links by workspace slug, and every workspace page uses the slug address. - CLI 0.33.2: `studio workspaces list` links by slug, and `--workspace` takes a slug as well as a name or id, e.g. `--workspace makemore`. Run `studio update`. - Old links by id to Delivery, Work, Board, Inbox, Spaces, Permissions, Overview, Activity, Workflows and Sessions move to the slug address, as the main workspace pages do. - A slug freed by a deleted workspace opens its new workspace straight away. ## 2026.11.11.2 (2026-11-11) Decide: questions waiting on you, one at a time. - Work -> Decide (`workspaces//work/decide/`) shows every question a worker or an agent is waiting on, one full-screen card at a time: the question in plain words, the choices as buttons (the suggested one marked), "Something else" for your own words, "Talk it through" to chat about it first, and "Later". Keys 1-N choose, Enter answers. - Workers can offer choices with a question: add `"options": [{"label", "description", "recommended"}]` to result.json next to `"question"`. ## 2026.11.11.1 (2026-11-11) Slack tells you about agent work and your inbox, and the channel is picked from a list. - Your Slack DM gets what lands in your Studio inbox: work that needs you, failed, landed or deployed, and messages from agents and teammates. "Alerts only" passes questions, failures and messages sent with --now; "everything" passes the rest. - The workspace's Slack channel hears when work needs an answer, fails, lands or deploys. - Reply in Slack to answer a work item's question; it goes back to the queue. - Settings → Connections → Slack: choose the channel from a list. ## 2026.11.10.4 (2026-11-10) Workspace addresses use the workspace's name, and Slack messages are sent. - Pages name a workspace by a short slug from its name, such as `workspaces/acme/`, instead of its id. Old links by id still work and move to the new address. - Slack: notices now reach people's DMs and the workspace channel. Before this, they were recorded but never sent. ## 2026.11.10.3 (2026-11-10) A Board for agent work, and CLI 0.33.1. - Work → Board is a kanban of agent work. At the top is what needs you: answer a question right there, resume a paused item, retry a failing one. Then every worker by computer (busy, idle, offline), and the columns Queued, Building, Testing, Review, Merging & deploying, Done. Each card shows its stage, its worker and what holds it up. Drag cards in Queued to change the order. - Git: a worktree whose files were mostly deleted by hand shows as half removed (clean it up), not as unsaved work. Workers on CLI 0.33.1 report this; run `studio update`. ## 2026.11.10.1 (2026-11-10) The Slack card and the Git page now appear in Studio's pages. - Workspace settings show Slack under Connections where the Studio has a Slack app: Install in Slack, the channel, and Link my Slack. - The Git page (from Work) shows every computer's branches and worktrees. ## 2026.11.10 (2026-11-10) A Git page: everyone's worktrees and branches, and what's left outside git. CLI 0.33.0. - Work → Git shows you first, then each teammate, with worktrees in three lanes: not safe yet (changes not committed, commits not pushed), waiting to merge, and merged (can be cleaned up). Each card names its branch, work item and computer, and copies the command that helps. - Workers send their checkouts' git every ten minutes (CLI 0.33.0). In any other checkout run `studio git report`. `studio git` shows the same in the terminal. - Reports hold names, counts and dates only, never file names or contents. Nothing on your computer is changed. ## 2026.11.9 (2026-11-09) CLI 0.32.0. - `studio work runner` is gone: use `studio worker` (`--foreground`, `install`, `pause`, `resume`). Update with `studio update`, or `studio worker update --all` for workers. ## 2026.11.8.5 (2026-11-08) Housekeeping in the work API. - Pausing and resuming a worker has its own endpoint, `POST workers/NAME/pause`. Today's CLIs keep working unchanged; a later CLI release switches to it. - The old `runners/` list is gone. No supported CLI uses it. ## 2026.11.8.4 (2026-11-08) Housekeeping: background work has a queue on every Studio. - Every kind of background work now has a queue on every Studio, ready for when a host turns it on. Nothing changes for people or agents using Studio. ## 2026.11.8.3 (2026-11-08) Worker updates know about the newest CLI on every Studio. - `studio worker update` offers the newest CLI even on a Studio that is behind Foundry: Studios now learn Foundry's CLI version along with its release. - Switching a device thread's agent with "now" carries on in exactly one new turn, even when the switch is retried. ## 2026.11.8.2 (2026-11-08) A Slack add-on, and workers need CLI 0.26 or later. - Slack: alerts can go to people's direct messages and a workspace channel, and replies reach the agent. Off until the Studio's host sets up a Slack app. - Workers on a CLI older than 0.26 are refused when they register, with what to do: run `studio upgrade`, then start the worker again. - Studio now needs runtime 0.18.32 or later (core 0.18.56). This release runs 0.18.34 (core 0.18.64). - Bus history follows the workspace's setting, then the plan. The separate host setting for it is gone. ## 2026.11.8 (2026-11-08) This release log is now public, for agents and people using Studio. - Read it at https://app.foundrystudio.dev/studio/api/release/log.txt: plain text, newest first, one entry per release back to 2026.10.18, with a summary of what came before. - Agents can check it to learn what's new: read from the top until you reach a release you know. - https://www.foundrystudio.dev/llms.txt links to it. - Runtime 0.18.34 (core 0.18.64) adds the brain: a project's own linked notes that its agents learn from their runs, read before they act, and keep tidy, with Obsidian-style vault pull and push. Off by default, and not yet turned on for Studio's agents. ## 2026.11.7 (2026-11-07) Work items can land after others, and workers that keep failing are paused. - CLI 0.31.4: `--after-landing` and `--needs` on work items and plan steps. An item built with `--after-landing` is worked on in parallel; only its merge waits until the items it names have landed. Plan steps take the same options. - A worker whose last three runs failed or were lost is paused automatically ("quarantined") with the reason. You get one inbox notice, `studio health` shows it, and resuming the worker lifts it. - Files you list as shared ledger files (for example a changelog) no longer hold an item back in the overlap check when two items touch them. - Fixed: items waiting to land were not being picked up by the periodic sweep, and threads waiting on an answer now show as "waiting". - CLI 0.31.2: `studio work diagram` fills the terminal's width. CLI 0.31.3: bus messages cost no tokens until there are some to read. - Runtime 0.18.33: runs start reliably on hosts that keep database connections open. ## 2026.11.6 (2026-11-06) Every worker has an inbox you can message, checked without spending tokens. - CLI 0.31.0: `studio tell WORKER` sends a message to a worker; `studio worker inbox` shows a worker's inbox. `--now` steers a running stage when the worker's tool can take it. - A message never starts a run by itself. It waits for the worker's next agent stage and is added once at the end of that stage's brief ("Message from X: ..."), up to 1,500 characters each. - Workers' check-ins carry an inbox count, so checking for mail needs no model call. Instructions such as stop, pause or update are handed straight back with no model. - You can message a worker if it is yours or you may change it; never across workspaces. - CLI 0.31.1: workers check their login without calling a model. - Athena 0.2.41: `tell_person` and `/tell` reach workers. ## 2026.11.5.12 (2026-11-05) Update all your workers at once. - CLI 0.30.4: `studio worker update --all` updates every worker on this computer to the current version in one go. ## 2026.11.5.11 (2026-11-05) Workers on Macs report how busy they really are. - CLI 0.30.3: on macOS, workers report real CPU use instead of the load average, so capacity and balancing no longer treat an idle Mac as busy. ## 2026.11.5.10 (2026-11-05) Fixed: boat.dev runs can read their files. - Runs on boat.dev now start in `/workspace`, so the workflow's unpacked files are readable. ## 2026.11.5.9 (2026-11-05) Fixed: boat.dev runs start. - Runs on boat.dev were all being lost at startup. The runner now starts correctly and runs with the same limited identities as cloud runs. boat.dev stays off unless a workspace manager turns it on. ## 2026.11.5.8 (2026-11-05) Claude Code sessions are sturdier; CLI 0.30.2. - A denied web fetch no longer ends a Claude Code session; the model can recover and carry on. - Session outputs no longer list a bogus "keep: That file is empty" entry. - CLI 0.30.2: `studio workflows run --prompt --egress` gives a session an optional list of hosts it may reach. Default: none. - A failed session's notice now says what it was asked and why it stopped. A session you followed to its end in the CLI does not send a failure notice (use `--no-follow` to get one). ## 2026.11.5.7 (2026-11-05) One-workspace sign-ins now work for `studio session`. - If `studio login` approved a computer for one workspace only, `studio session`, devices and `studio http` previews now work inside that workspace, without needing `--workspace`. - Nothing from your other workspaces is reachable through such a sign-in. A path that cannot honour the limit answers with a clear "limited to one workspace" message instead of "sign in". - Server errors now reach Studio's operators automatically, without personal data or message contents. ## 2026.11.5.6 (2026-11-05) Fixed: workflow runs and Claude Code sessions failing with "The runner was lost". - Since 2026.11.5.4 every cloud workflow run and session was refused when it called back to Studio. This release fixes it, and a test now guards every runner route. ## 2026.11.5.5 (2026-11-05) Choose boat.dev as a workspace's default compute; CLI 0.30.1. - A workspace manager can make boat.dev the default place runs and Claude Code sessions go: the Workflows page's "Run on boat.dev by default", or `studio workflows boat on --default`, or `studio workflows boat default cloud_run|boat`. - A compute named on the run itself always wins. The default applies only while boat.dev is allowed for the workspace; otherwise runs quietly use cloud_run. Turning boat.dev off resets the default. - boat.dev is still off unless the host and a workspace manager turn it on. ## 2026.11.5.4 (2026-11-05) One-off Claude Code sessions; CLI 0.30.0. - Agents can hand a task to a one-off Claude Code session in the cloud (`run_claude_code_session`, then `workflow_status` to read its answer). Attachments go in; files it writes to `outputs/` come back as attachments. - CLI 0.30.0: `studio workflows run --prompt` starts a session from the terminal. - Workflows can name workspace MCP connections in an `mcp` list, so a run can use them. - boat.dev as a second compute target (`studio workflows boat [on|off]`, `compute` in `studio.yaml`). Off by default; a workspace manager must opt in; never for HIPAA workspaces. ## 2026.11.5.3 (2026-11-05) Runtime 0.18.32: lanes and run budgets. - Lanes: wakes for a conversation coalesce, so at most one run is in flight per conversation and lane. Events that arrive meanwhile are handed to the next run together. Hosts that don't use lanes behave as before. - Run budgets: a run can be given a dollar budget; it stops before the next model call once spent and finishes with stop reason "budget". ## 2026.11.5.2 (2026-11-05) Agent turns survive restarts, long conversations compact, agents can start workflows. - Durable turns: a turn cut off by a server restart is picked up again instead of starting over. Tool calls that are safe to repeat are rerun; others tell the model they were interrupted. - The chat widget says "Interrupted by a restart - not repeated" in plain words, and offers a Resume button on a turn that was cut off. - Context compaction: when a conversation nears the model's limit, older messages are summarised for the model. Nothing is deleted; history and the UI look the same. - Agents get `start_workflow` and `workflow_status` tools where the Studio has workflows set up. They act as the thread's person, inside that workspace only. ## 2026.11.5.1 (2026-11-05) CLI 0.29.1: `studio work diagram` fills the screen. - At a terminal (and always with `--watch`), the Last events list fills the rest of the screen. `--events N` sets the count. Piped and `--json` output keep 6. ## 2026.11.5 (2026-11-05) A working worker now stays online; CLI 0.29.0. - Fixed: a worker busy on one item for more than three minutes showed as "reconnecting", looked offline to claims and could lose its item's later stages. A run's heartbeat now counts as the worker's check-in. - CLI 0.29.0: `studio work diagram` draws one 80-column picture of watchers, queue, computers, workers, questions and events (`--watch` to refresh, `--json` for scripts). - A worker's heartbeat no longer waits for slow work such as measuring a worktree. ## 2026.11.4.12 (2026-11-04) Fixes to quiet runs and pinned items. - A run that changed nothing no longer pins its item to one computer, which could leave it stuck at the merge stage with no reason shown. An item pinned where its stage cannot run now says so. - A grace period shorter than three minutes is honoured: a lost run is noticed after a minute. ## 2026.11.4.11 (2026-11-04) Work state fixes; CLI 0.27.6. - A check that a worker's computer cannot run (command not found) holds the item and tells the worker's person, instead of sending it back to be "fixed". Resuming reruns the checks. - A lost worker's late report can no longer revive a run after a newer run has started. - A deploy worker lost past its grace period pauses the item for a look, as a failed deploy does. - Watcher changes made in Studio reach a running watcher within seconds; `grace_minutes` may be 1. - CLI 0.27.6: a watcher applies changes pushed from Studio at once, and a worker's registered version is only replaced when the one on disk is newer. ## 2026.11.4.10 (2026-11-04) Rehearsal briefs carry their state. - A rehearsal's brief now says whether its question was already answered and how many hand-offs it has had, so it no longer loops handing off. ## 2026.11.4.9 (2026-11-04) CLI 0.27.4. - CLI 0.27.4: a worktree left on another branch (such as a removed rehearsal's) is made again for the item that needs it. ## 2026.11.4.8 (2026-11-04) Rehearsals never merge into the real base branch. - A rehearsal's merge and deploy commands go to its own throwaway base, even when the project's command names the real one. ## 2026.11.4.7 (2026-11-04) CLI 0.27.3. - CLI 0.27.3: a run that finishes work an earlier run committed (after a hand-off or a question) counts as done, not "changed nothing". ## 2026.11.4.6 (2026-11-04) Account trouble reaches the right person; CLI 0.27.2. - Account trouble on a worker (for example an expired login) now reaches the worker's person, with the fix. - Rehearsals get a fresh branch each run, and a failed rehearsal's branches are cleaned up. - CLI 0.27.2: git access follows Studio at check-in, containers report their own CPU, and the first health check comes after 5 minutes. ## 2026.11.4.5 (2026-11-04) Search threads by what was said in them. - The sidebar search finds threads by their messages, in one workspace or across all of them. In the CLI: `studio threads search`. - Workers are told about their git access at every check-in, and a rehearsal's question says the orchestrator answers it. ## 2026.11.4.4 (2026-11-04) Merge and deploy stages, remote worker control, capacity, git access; CLI 0.27.1, Athena 0.2.40. - Work gets Merge and Deploy stages: a merge role, deploy after merging, and conflicts sent back to be fixed. The Work page shows them as columns. - Workers and watchers can be changed from anywhere by editors of their projects: settings are kept in Studio, changes are logged, and an offline worker or stopped watcher can be forgotten. - Capacity and balancing, workspace signals, and `studio health`. - Worker git access: a manager can turn on git access through Studio, so workers push and open pull requests without a login prompt. - A worker whose self-update didn't come back shows "update failed". - CLI 0.27.1 and Athena 0.2.40 are bundled. ## 2026.11.4.3 (2026-11-04) CLI 0.25.26: watchers find your tools. - CLI 0.25.25: a watcher names a Mac by its local network name. - CLI 0.25.26: a watcher running in the background finds harnesses that are outside its limited PATH (login shell PATH, `~/.local/bin`, npm and Homebrew). ## 2026.11.4.2 (2026-11-04) Orchestrator watchers are seen and controlled through Studio; CLI 0.25.24, Athena 0.2.38. - Watchers (`studio orchestrate on`) report to Studio every couple of minutes. The report, the Work page overview and an Orchestrators block list them across all your computers. - CLI 0.25.24: `studio orchestrate list`, and `on`, `off` and `log` take `--machine`. - The watcher's person and editors of a watched workspace can change what it runs; the change is recorded and reaches the person's inbox, and applies on its next check-in. ## 2026.11.4.1 (2026-11-04) CLI 0.25.23: an orchestrator watcher. - `studio orchestrate on`, `off` and `log`: a watcher starts an orchestrator pass when something needs one, instead of you polling. ## 2026.11.4 (2026-11-04) Stored plans, three roles in the CLI, reordering, one orchestrator pass at a time. - Stored plans: a plan has a goal and numbered steps (P-numbers per workspace). Items gain a plan, a step and "after" dependencies (W-numbers that must be done first). Claims wait until what the item waits for is done, and the item says why it is waiting. - Reorder work without taking it: it gets a priority so it runs first (or just above another item). Any editor or manager can do this. - One pass at a time: an orchestrator takes a 10-minute lease on a workspace; the same holder renews it and others are told it is held. - CLI 0.25.22: one noun per role: `studio plan`, `studio work`, `studio worker`, `studio orchestrate`. - Athena 0.2.37 is bundled. ## 2026.11.3.11 (2026-11-03) Workers' briefs now carry what earlier workers learned. - A worker's brief includes the project's most relevant learnings, those about the files it will touch first (at most 8 facts), taken from its own project only. - CLI 0.25.21: `studio skills` with `--once` and back-off; `studio asks` and `studio now` take a workspace's name. - Orchestrators keep where they last read events in Studio, so a fresh pass on any computer carries on from there. ## 2026.11.3.10 (2026-11-03) Discord add-on goes live; CLI 0.25.18. - A workspace can connect its own Discord bot to its community server (see 2026.11.3.9). Enterprise Studios: the host installs the add-on; it is on wherever it is installed. - CLI 0.25.18: a crashed worker can be started again under its own name. ## 2026.11.3.9 (2026-11-03) Discord add-on, `studio skills install`, and shared learnings; CLI 0.25.17. - Owners and admins connect a Discord bot in Settings -> Connections -> Discord, or with `studio connect discord`. The token is checked, sealed and never shown again. - Agents get `discord.read` (server, channels, roles, threads, messages, members), `discord.post` (a person approves each post or pre-approves a channel; mentions never ping) and `discord.moderate` (owners and admins approve every call). - CLI 0.25.17: `studio skills install` lets any Claude Code, Codex or Auggie session be a work orchestrator. `studio work recall` and `studio work learn` search and save workers' learnings. ## 2026.11.3.8 (2026-11-03) Workers' learnings are kept per workspace and project. - A worker's result can carry up to 5 lasting facts. A newer fact on the same topic replaces the older one. Orchestrators can search and save learnings with `studio work recall` and `studio work learn`. - Search covers everything you can see, each result labelled with its workspace. - Searching by meaning is off by default; the host turns it on. Workspaces in HIPAA mode never send text out for it unless the host allows that. ## 2026.11.3.7 (2026-11-03) Pause a whole queue, and work survives short worker disconnects. - Workspace owners and admins, or a project's managers, can pause and resume the whole queue of a project or workspace. Workers claim nothing new but stay connected, so they can be updated. Running stages finish by default; items are not changed, and resuming carries on. - A worker that stops answering keeps its items for 15 minutes by default (`grace_minutes`, per project or per workspace). The item shows "waiting for worker"; when the worker is back its run carries on without starting a new one. - After the grace period the run is released once and the item goes back to ready, keeping its branch and hand-off. An item dropped more than 3 times is flagged for a look. - Bundles CLI 0.25.15 and Athena 0.2.34. ## 2026.11.3.6 (2026-11-03) Harnesses run without asking. - Every orchestrator harness, and every direct harness in Athena, now runs without permission prompts. ## 2026.11.3.5 (2026-11-03) The orchestrator can run on any harness. - CLI 0.25.13: `studio mcp work`, and `studio orchestrate --harness claude|codex|auggie|athena`. - Athena 0.2.32: the picker's Orchestrator asks which harness to use. ## 2026.11.3.4 (2026-11-03) CLI 0.25.12: workers' git never waits on a login prompt. - A worker that needs git credentials fails fast instead of hanging on a prompt. ## 2026.11.3.3 (2026-11-03) Landing safety. - CLI 0.25.11: a refused rebase is no longer treated as a conflict, and workers' notes stay out of the landed change. - A project can keep landing to its deploy workers (`deploy_workers_only`). ## 2026.11.3.2 (2026-11-03) An offline worker holds nothing. - Items pointed at an offline worker can be taken by any other worker. - Landing and testing no longer wait for the worker that wrote the change. ## 2026.11.3.1 (2026-11-03) Work can't get stuck. - A held session is released on reassignment, when its worker is offline, or after 10 minutes. - Items can state `needs` (a host or capability), and a project's `needs_rules` infer them. - With a landing command set, waiting items land by themselves; a conflict or failing check sends the item back to fix. You can also land by hand with `studio work land`. - The work report lists stalled items. `studio work watch` checks with no model. - A run that costs over `run_cost_limit_usd` (default 5) hands off and is flagged. ## 2026.11.3 (2026-11-03) The orchestrator can act, and uses Opus 5.5. - Athena 0.2.30: the orchestrator can run commands and edit files by default, under your approval settings. - New orchestrator threads use Opus 5.5 where the connection allows it. Existing threads keep their model. ## 2026.11.2.14 (2026-11-02) CLI 0.25.9 and Athena 0.2.29; orchestrator threads open on another computer. - Reading an orchestrator thread says where it got to and which computer was using it, so another computer can pick it up. - CLI 0.25.9: `read --all` reaches Studio. Athena 0.2.29: `projects_list` works, and failed tools are kept so they can be found and fixed. ## 2026.11.2.13 (2026-11-02) One Athena, and reading code on a worker's computer. - `studio worker --remote-read` lets that worker's own person, and their orchestrator, list, read and search files and see git log, status and diff in its checkouts, through Studio. Off by default, per worker. - Athena shows Studio threads and messages, and the orchestrator works across all your workspaces. ## 2026.11.2.12 (2026-11-02) Studio threads in Athena, menus, and inbox unread. - Athena can read any thread you can see and send to its agent, with Studio's own checks. Long replies are read in short calls. - Athena can rename a project; its list says who may. - A message can go back to unread, and a finished one back to read. ## 2026.11.2.11 (2026-11-02) Inbox across workspaces, and finished work can't sit unnoticed. - `studio inbox` and `studio tell` with no workspace cover all your workspaces and say which one each message is from. - A question, a failure, ready to land and landed reach the item owner's inbox once. Set `notices: false` in a project's config to turn this off. - Work waiting to land shows its age in `studio now`, on the Work page and in the work report. After a day its owner gets an inbox message, then weekly (at most 3 a day each). ## 2026.11.2.10 (2026-11-02) Athena's left bar: rename, archive and delete. - Athena 0.2.23 and CLI 0.25.5: the left bar works by keyboard and mouse, with rename, archive and delete under the same rules as Studio's menu. Archived threads leave Athena's list. ## 2026.11.2.9 (2026-11-02) Athena 0.2.22 and CLI 0.25.4. - Athena 0.2.22: `athena orchestrate` no longer refuses to start from a home directory. - Athena lists threads. ## 2026.11.2.8 (2026-11-02) `athena orchestrate` opens your orchestrator from anywhere. - Athena 0.2.21: `athena orchestrate` works from your home directory with no folder index. It can stop workers, create projects and conversations, and recall what workers did. - Searching work items now also searches workers' summaries and hand-off notes. ## 2026.11.2.7 (2026-11-02) Workers don't fill your sessions. - Workers, parents and orchestrators no longer count towards your live-session limit; they have their own, larger limit. - A worker idle for 30 minutes quits if an online parent covers its folder; the parent starts it again when work is ready. - At the session limit the CLI asks one question: end the session quiet longest? ## 2026.11.2.5 (2026-11-02) Worker modes; the bus stops being the default. - Workers have modes: development (the default), testing and deploy. Change one with `studio worker role NAME testing`. A testing worker runs a project's gates with no model. - Only direct messages wake a Studio agent, batched for a minute. `studio code` joins the bus only with `--bus`. - `studio deploy` holds a lock and hands out release numbers, so two sessions can't deploy over each other. CLI 0.25.2, Athena 0.2.18. - Fixed: a review or landing on another computer used to miss an unpushed branch's changes. ## 2026.11.2.4 (2026-11-02) Prompt caching for agent runs on Anthropic. - Agent runs on Anthropic now cache their prompt and conversation, so repeated input costs far less. Run usage and cost estimates count cached tokens. - On by default. Connections with a compliance rule never use it. Hosts can turn it off. ## 2026.11.2.2 (2026-11-02) A personal inbox, kept in Studio. - Each person has an inbox. `studio tell` or `/tell` in Athena puts a message there; `studio inbox` or `/inbox` reads it. Nothing reaches a model or wakes an agent until you add it. - CLI 0.24.9, Athena 0.2.16. Joining the bus no longer tells agents to keep a listener running. ## 2026.11.2.1 (2026-11-02) Fixed: add-on tools run again. - Tools from the Google, Google Cloud, Linear and voice add-ons had stopped running for hosts that restrict tools; they work again. ## 2026.11.1.13 (2026-11-01) The Work page stays light. - The Work page polls short summaries with no query per item, shared between open tabs, and waits longer when the server is slow. Details load when you open an item. ## 2026.11.1.12 (2026-11-01) Add-ons are Django apps. - Google, voice calls and Linear are optional Django apps, like Google Cloud. Enterprise Studios: the host installs each one; without it, that add-on is off. - Core no longer imports any add-on. Nothing changes for people using Studio. ## 2026.11.1.11 (2026-11-01) CLI 0.24.8: team on shared servers. - Athena's saved login works when `~/.config` was group-writable. - Each person's orchestrator has its own name. ## 2026.11.1.10 (2026-11-01) Athena 0.2.15 in the public installer. - Compaction no longer times out on long sessions. - Mission control starts far fewer `studio` processes. - A wrong project name gets the list of valid names. ## 2026.11.1.9 (2026-11-01) Workers show their computer; simpler team. - CLI 0.24.7: workers report CPU, RAM and disk. Athena 0.2.14's mission control draws them under each worker, red when dangerous. - `studio team --athena-login` installs Athena and saves the workspace key for Athena alone, so people land straight in the orchestrator. `--no-orchestrate` prints the cloud orchestrator's link instead. ## 2026.11.1.8 (2026-11-01) Athena 0.2.13: fewer failed tool calls, and `/log-level`. - Common path spellings work, and tool calls read as one plain line. - `/log-level quiet|normal|detailed` sets how much Athena shows. - The orchestrator starts fewer `studio` processes. ## 2026.11.1.7 (2026-11-01) CLI 0.24.5: `studio team` looks before it starts. - `studio team` counts every person's workers on the computer, other agents, cores, free memory and load, and starts fewer rather than overload it, saying why. - `team add --force` overrides. Nothing of anyone else's is stopped. ## 2026.11.1.6 (2026-11-01) CLI 0.24.4: connect any computer, start with one word. - `studio join` asks which workspace, project and folder, creating them where you are allowed, with a small file browser. - `studio team` starts whatever of yours isn't running and opens the orchestrator; `status`, `add` and `stop` work too. ## 2026.11.1.5 (2026-11-01) Athena 0.2.12: faster over SSH. - Replies append instead of redrawing per chunk, there are fewer screen layouts, search runs off the main thread, and start is quicker. ## 2026.11.1.4 (2026-11-01) Stricter workers, and why an item is waiting. - Workers give evidence for each acceptance criterion and say at once when they differ from an approved design. Reviewers label what they couldn't rerun "unverified, worker-reported". - `studio work show` says why a ready item is waiting. - CLI 0.24.3: a worker's or orchestrator's session closes when it stops, and `studio sessions` shows computer, idle time and worker. - Athena 0.2.11: cut, copy and paste across platforms, Tab rotates permission stops, and bus messages reach the next turn. CLI 0.24.2 / Athena 0.2.10: `update` and `upgrade` are the same word. ## 2026.11.1.3 (2026-11-01) Add-ons are plugins. - Google Cloud, Google, Linear and voice are organised as add-ons that declare themselves. Nothing changes for people using Studio. - Enterprise Studios: Google Cloud is an optional app the host installs; without it, it is off. ## 2026.11.1.2 (2026-11-01) Agents can ask to change Google Cloud, with approval. - New capability `google_cloud.manage`, off until a workspace admin sets it up under Connections -> Google Cloud -> Manage resources. - Agents can ask to stop or start VMs, snapshot and delete disks, change Cloud Run scaling, pause scheduler jobs and similar. A person approves every change. There are no tools for databases, IAM or keys. - Only on Studios that have a Google Cloud broker (enterprise); Foundry has none. ## 2026.11.1.1 (2026-11-01) See and close idle sessions. - The session list shows each live session's last activity and the worker or orchestrator online in it. - CLI 0.24.1: `studio sessions close-idle`; the session-limit prompt shows computer and idle time. ## 2026.11.1 (2026-11-01) Worker versions and self-update; Athena pages long conversations. - `studio work workers` and the orchestrator show each worker's version and when a newer one is out. `studio worker update NAME` (or the orchestrator, with your approval) has a worker finish its running work, update and restart. - CLI 0.24.0, Athena 0.2.9: a saved conversation opens on its latest 20 messages; `/earlier` and `/later` page through it. ## 2026.10.31.2 (2026-10-31) Work that rightly changes no code can now finish cleanly. - A worker can finish a stage with `"no_changes": true` (a report, a data fix, a check). The item goes to ready to land instead of failing and running again. Workers already running support it; no CLI update is needed. - Items that fail again and again now pause instead of retrying for ever (after twice the item's `look_after` setting in a row, 6 by default). Resuming the item clears the count. ## 2026.10.31.1 (2026-10-31) The public Athena installer serves Athena 0.2.8. - Athena 0.2.8: installs tell you when a newer Athena is out, and `/update` installs it and restarts in the same conversation. ## 2026.10.31 (2026-10-31) The public Athena installer serves Athena 0.2.7. - Athena 0.2.7: setup keeps your answers, including the key, when connecting fails. - Athena 0.2.7: local embeddings, and `/control` in orchestrator mode hands commands and edits back to you, or takes them away again. ## 2026.10.30.21 (2026-10-30) Cut-off model replies continue; workers start workers on other computers. - Replies that hit the model's output limit now continue instead of stopping mid-sentence, and the output limit is higher. A model request can also carry up to 20 images. - An orchestrator can start workers through a parent worker on another computer, and a checkout finds its Studio project from its git remote. - CLI 0.23.0 and Athena 0.2.4 in the installers. ## 2026.10.30.20 (2026-10-30) Several orchestrators can share a queue without stepping on each other. - Orchestrators check for similar work before adding an item, and items that touch overlapping files wait for each other. - Orchestrator threads are cheaper (bounded), and the cost view shows costs by who asked. - CLI 0.22.0 and Athena 0.2.2: Sonnet is the orchestrator's default model. ## 2026.10.30.19 (2026-10-30) Ten live sessions, and cost notes that read as money. - Paid plans can keep ten live `studio session` terminals per person. The refusal message now names the limit. - Cost notes on the Work page and in the CLI show dollars, not token counts. - CLI 0.21.4 and Athena 0.2.1: mission control shows dollars and lists live workers first. ## 2026.10.30.18 (2026-10-30) MCP connections show the redirect URL to register. - With OAuth sign-in chosen for an MCP connection, the form shows the exact redirect URL to register on your own OAuth client (Google needs it), with a Copy button. ## 2026.10.30.17 (2026-10-30) Athena has a public installer at /install-athena.sh. - Athena 0.2.0 installs from Studio's `/install-athena.sh`, with pinned, hashed dependencies. - The installer serves CLI 0.21.0. ## 2026.10.30.16 (2026-10-30) The orchestrator thread opens without a project. - The orchestrator's Studio thread no longer needs a project to open, and its work tools count as internal, so they do not ask for approval. - The installer serves CLI 0.20.1. ## 2026.10.30.15 (2026-10-30) The orchestrator is now a Studio thread. - A cloud orchestrator agent with Studio's work tools runs in a normal thread, and shares its conversation with the orchestrator in Athena. - The installer serves CLI 0.20.0. ## 2026.10.30.14 (2026-10-30) Members can pull their workspace's AI key. - Members can fetch the workspace's AI key for their own tools; each pull is recorded in the workspace history. It is refused in HIPAA mode and while acting as someone else. ## 2026.10.30.13 (2026-10-30) Fixed: work item actions failed on the production database. - Fixed: answering, pausing, taking and the other work item actions returned an error. They work again. - The installer serves CLI 0.18.1. ## 2026.10.30.12 (2026-10-30) The orchestrator pattern for agent work. - Workers of any harness register from a checkout: `studio worker --harness claude|codex|athena|auggie`. Studio matches the folder's git remote to a project's repository. - Orchestrators register too. Items get owners and can be assigned to a worker, a harness or a machine, and moved between orchestrators with take. - Work events (claimed, implemented, approved, landed, question, failed, worker offline) reach the project's bus channel. - The Work page gains a live map of orchestrators, workers and items, a timeline per worker, stop buttons and a cost explainer. - The installer serves CLI 0.18.0. ## 2026.10.30.11 (2026-10-30) Athena and AI thread titles ship together. - Thread menus (sidebar and top bar) have "Name with AI" and "Summarise". They use the thread's own AI connection and model, and count in the workspace's monthly spend. - A new thread is named by AI once, after its first run, unless you chose a title. Existing threads are never renamed automatically. - The terminal agent app is now named Athena everywhere. Older installs must be reinstalled. ## 2026.10.30.10 (2026-10-30) Fixed: the project menu is no longer hidden under the rows below it. - Fixed: the project row's menu was painted over by the rows under it. It now sits on top, as thread menus do. ## 2026.10.30.9 (2026-10-30) Threads follow a changed AI endpoint; container checks arrive (off). - A new AI endpoint or protocol now reaches existing threads. Threads that kept a model from the old connection used to fail with "AI connection is unavailable"; they now use the new one. - Container device checks in `.studio/delivery.json` are off unless the host turns them on. - The installer serves CLI 0.17.1. ## 2026.10.30.8 (2026-10-30) Manage has an AI gateway page (off unless configured). - Settings, AI gateway, for a workspace's owners and admins: the models on offer, the workspace's keys (shown once on create, with limits and stop), and 30 days of usage. - The page is hidden unless the host connects an LLM gateway. ## 2026.10.30.7 (2026-10-30) Fixed: uploads from the web chat were refused. - Fixed: attaching a file in the web chat failed in the browser. Uploads work again. ## 2026.10.30.6 (2026-10-30) Agents can read Google Cloud costs. - A read-only `google_cloud.costs` capability on a workspace's Google Cloud connection gives agents a cost report, recommendations, a resource finder and budgets. Nothing is written. - Off unless the host provides the broker service account. ## 2026.10.30.5 (2026-10-30) Files in chat: any file but programs, 20 on a message. - Attach up to 20 files per message, of any type except programs; several can upload at once. - Text is read from PDF, Word, Excel (including old `.xls`), PowerPoint, OpenDocument, EPUB, zips and plain text. Other files are kept and the agent is told it cannot read them. - A Studio with no sandbox needs a private storage setting to keep attachments. Without it, uploads answer "unavailable" and the attach button is hidden. ## 2026.10.30.4 (2026-10-30) Agent work: task queues worked by one-shot agent runs, all from the CLI. - Work items per project (`W-12` style ids). Runners on your own computer (`studio work runner`) claim one stage at a time and run one headless Claude Code per stage in the item's worktree. - Gates run with no model; then review, landing, questions and answers, and hand-offs at a context size you can change mid-run. Only people can pause, resume or stop. - Every run records tokens, cache use, cost, gates, activity and result. - A new Work page per workspace shows the queue, what is running, questions, runners and costs. - `studio work usage sync` stores your own Claude Code token counts (no content) for the cost view. - CLI 0.17.0. ## 2026.10.30.3 (2026-10-30) Security fixes. - Three security fixes in sign-in and file privacy. Nothing to do on your side. ## 2026.10.30.2 (2026-10-30) Threads show runs they did not start as they happen. - An open thread now shows scheduled follow-ups, posts from another agent or the front desk, approvals, and activity from another tab or device while they happen, and streams them live. ## 2026.10.30.1 (2026-10-30) Front desk: read whole replies, post into an existing thread. - New front-desk tool `read_thread_reply`: the whole latest reply (or one further back) in pages of 8000 characters. Summaries now mark when a reply was cut. - New `post_to_thread`: sends a follow-up as you into a thread you can continue, which starts its agent. Refused while that agent is working; agent-started posts stop after 3 in a row. - A linked agent can post and read on the other workspaces' buses it is linked to. - Device agents know their device name, host and platform. - `schedule_follow_up` no longer refuses a timezone given with a delay; it ignores it. ## 2026.10.29.5 (2026-10-29) Fixed: blank chat panes. - Fixed: after the previous release some chat panes rendered blank. Script and stylesheet now come from the same build, and a test checks that they match. ## 2026.10.29.4 (2026-10-29) New chats: apps on, and pick any agent. - New chats start with the apps their agent declares switched on (for example Gmail), so you do not enable them in each chat. Switch any app off in the chat. - New chat lets you pick any agent you may use in the project. The pick is remembered per workspace, so your own assistant can be where every chat starts. - The front desk's `create_thread` takes any usable agent, so one agent can ask another. ## 2026.10.29.3 (2026-10-29) Talk to a linked agent in the other workspace. - New chat offers an agent from a workspace it is linked to, for its person only. Those threads run on the linked workspace's AI connection and turn read-only if the link stops working. ## 2026.10.29.2 (2026-10-29) A chat started on the home page starts with what you typed. - Fixed: a chat begun from the home page now sends the message you typed as its first message. ## 2026.10.29.1 (2026-10-29) One assistant across your own workspaces. - If you own both an agent's workspace and another, you can let that agent work in the other one for you, through the front desk only and in your own runs. Not available in HIPAA mode. - CLI 0.16.1: `studio runtime agents links`, `studio runtime agents link` and `studio runtime agents unlink`. ## 2026.10.29 (2026-10-29) New chat remembers your model. - The model you pick in New chat becomes your default in that workspace. The pick list shows the workspace's default first, for agents that use the workspace's AI connection. ## 2026.10.28.11 (2026-10-28) Attention stays short. - The attention list shows the 10 newest unread conversations from the last week and sums up the rest in one line. ## 2026.10.28.10 (2026-10-28) Your asks, what needs you, and what is going on. - Asks: one list of what you asked agents to do, who has it and how it stands (open, doing, waiting, done, dropped). Hosted and device agents keep it up to date as they work. - Studio home shows Your asks (Done / Drop), Ready for you, and Going on. - CLI 0.16.0: `studio now` and `studio asks`. - The front desk gains `whats_going_on` for the current workspace. ## 2026.10.28.9 (2026-10-28) The menu shows the product name and mark. - The menu starts with the product's mark and name, and New chat is its one filled button. ## 2026.10.28.8 (2026-10-28) Files in chat: PDF, Word, Excel, PowerPoint, OpenDocument and text. - Attach documents, not only photos. Files are checked by their contents, and their text is read once at upload and kept for the agent. Replies can refer to earlier files in the chat. - CLI 0.15.1 is the installer's version. ## 2026.10.28.6 (2026-10-28) Switch an agent's harness or model; Auggie in sessions. - `studio bus switch NAME [--harness claude|codex|auggie] [--model M] [--now]` switches a bus agent. Allowed for the agent's own person and the workspace's owners and admins. - Auggie runs in `studio session` terminals and shows in the Readable view; its questions push "Auggie needs you". - CLI 0.15.0: `studio code` shows which of Claude Code, Codex and Auggie are installed, installs missing ones and starts one in a recorded session. Also `studio upgrade` and `studio devices connect|here`. - The installer puts back plugins you installed before. ## 2026.10.28.5 (2026-10-28) Replies stream in web, iOS and Android. - Fixed: chat replies arrived whole when the run ended on the web, iOS and Android. They now stream as they are written. Embeds and followed direct sessions keep working. ## 2026.10.28.4 (2026-10-28) Chat replies stream in web, iOS and Android. - Chat replies stream token by token on the web, iOS and Android apps instead of arriving at the end of the run. ## 2026.10.28.3 (2026-10-28) Device networks can use network edges. - Enterprise Studios: a device network can be reached through a network edge, and its enrolled devices can be moved to the new route without enrolling them again. ## 2026.10.28.2 (2026-10-28) Private-network agents run through a network edge. - Enterprise Studios: agents that reach private-network tools can run on the executor through a network edge on the host's own network. ## 2026.10.28.1 (2026-10-28) Replies stream through the events endpoint. - Browser widgets, the mobile apps and the CLI now see replies stream when the host uses Redis for run events, instead of whole replies at the end. ## 2026.10.28 (2026-10-28) Run events stream through Redis. - Run events stream through Redis where the host has it (Foundry does), with far fewer database reads for each watched run, so busy streams cost less and stay quick. ## 2026.10.27.2 (2026-10-27) An MCP connection that cannot start no longer fails the whole run. - Fixed: if one of an agent's MCP connections cannot start, the run carries on without its tools instead of failing (runtime 0.18.16). ## 2026.10.27.1 (2026-10-27) Workers start faster. - Studio workers load the AI provider libraries on first use rather than at start-up, so runs begin sooner after a new worker comes up (core 0.18.53). ## 2026.10.26.4 (2026-10-26) Live captions while a call device holds the call. - The thread page shows what is said, as captions, when a call is running on one of your call devices rather than on this computer. - Only applies on Studios that have call devices turned on. ## 2026.10.26.3 (2026-10-26) The Call button lets you choose where the audio goes. - Start a call on this computer or on one of your call devices. The button lists your devices. - A call running on a device can be followed from the thread: "Move here" or "Hang up". - A call on your laptop can be moved to a device. - On a Studio with call devices off, the Call button works as before. ## 2026.10.26.2 (2026-10-26) Call devices: speakers and robots can take live calls in your threads (off by default). - A person's speakers and robots can take Live calls in their threads. Off unless the host turns it on. - CLI 0.13.0: `studio call-devices` lists and manages your devices, and `studio call` starts or moves a call. ## 2026.10.26.1 (2026-10-26) MCP tool errors now reach the agent. - When an MCP tool fails, the agent sees the error and can react to it, instead of the failure being lost (runtime 0.18.15). ## 2026.10.26 (2026-10-26) New chat lets you pick the workspace, the agent and the model. - The message box on the home page chooses the workspace and the agent. The workspace's default agent is listed first, and a project's saved agent is kept. - A Model select appears for agents that use the workspace's AI connection. It lists only models the workspace's policy allows; a refused model creates no thread. ## 2026.10.25.6 (2026-10-25) Delivery team set-up, spec resets, agent-link bug tools, photos in chat. - Delivery team (a ready-made set of delivery agents for new workspaces) arrives off by default. - CLI 0.12.1: `studio workspaces create --studio-team [--auto-deliver]` and `studio sdlc team [--setup]` set up that team. - Connected platforms can give their agents tools to report a problem, look up similar bugs, add to a bug and check its status. Personal details are scrubbed before anything is sent. - Photos people attach in chat are accepted and shown to agents. - A new switch lets a workspace's home-automation connection expose its cameras. ## 2026.10.25.5 (2026-10-25) Agent edits apply at once, and images from tools reach the model. - Saving a change to an agent takes effect on its next run immediately. - MCP tools can take up to 90 seconds to answer before the run gives up (runtime 0.18.11). - Images returned by tools are passed to the model, so it can look at them (core 0.18.50). ## 2026.10.25.4 (2026-10-25) Safer calls while acting as someone, and MCP connection checks from the CLI. - No Call button while a staff member is acting as someone, and a call cannot be started in that person's name. - A desktop app download page and voice approvals arrive, both off unless the host turns them on. - CLI 0.10.2: `studio connect mcp NAME --test`, `--assign AGENT` and `--unassign AGENT`. ## 2026.10.25.3 (2026-10-25) Calls take longer instructions and history. - Live calls on a self-hosted voice provider now accept long agent instructions and a longer conversation history. ## 2026.10.25.2 (2026-10-25) Live calls fit long instructions. - Agents with long instructions can now start a live call; they no longer overflow the call set-up (runtime update). ## 2026.10.25.1 (2026-10-25) Studio opens on your last conversation, with Recent and Projects across every workspace. - A menu like a chat app's lists Recent conversations (threads and terminal sessions, by your own last activity) and Projects, each with up to 8 items, across all your workspaces. - Studio reopens the conversation you had open last. - You can put the list in your own order. The order is personal: it never archives, hides, ends or unshares anything, and nobody else sees it. ## 2026.10.24.18 (2026-10-24) Fixed: the web Call button failing with "CSRF token missing". - Fixed: chat widgets that sign in with a token now send it correctly, so starting a call and saving Tasks tab changes work on the thread page. ## 2026.10.24.17 (2026-10-24) Fixed: Call and Tasks writes for signed-in session widgets. - Fixed: when a thread is shown with your browser session (for example for device agents), the Call button and Tasks tab edits now work instead of failing with a CSRF error. ## 2026.10.24.16 (2026-10-24) The projects list and Needs you load with far fewer repeated checks. - Faster projects list and Needs you page; each scope is still access-checked. ## 2026.10.24.15 (2026-10-24) Faster first requests after a deploy. - Servers are warmed up before they take traffic, and slow requests are logged with their cost, so cold starts are shorter and easier to track down. ## 2026.10.24.14 (2026-10-24) A faster home page and projects list. - Needs you and the agent questions card read each person's workspaces and projects once, not per workspace. A page that made about a thousand queries now makes tens. ## 2026.10.24.13 (2026-10-24) Agents open Word, Excel, PDF and the other files people send. - Google tools `drive_save_file` and `gmail_save_attachment` copy a Drive file (Docs, Sheets and Slides come as .docx, .xlsx, .pptx) or a Gmail attachment into the conversation, up to 16 MB. They are read-only and need no new permissions. - The code sandbox can read Word, Excel (including .xls and protected files), PowerPoint, PDF, CSV, .eml, HTML, OFX/QIF and zip files, with helpers for odd encodings, bracketed amounts and UK-style dates. ## 2026.10.24.12 (2026-10-24) Claude Code workflows pass more of Claude Code's model settings through. - The workflow relay now forwards the fields and betas Claude Code adds to shape a model call, so newer Claude Code versions run in Studio workflows. ## 2026.10.24.11 (2026-10-24) Claude Code workflows can run on an OpenAI-compatible gateway. - An AI connection that uses the OpenAI protocol can now carry an Anthropic-compatible address. Claude Code workflows use it to run through that connection. - The AI connection form sends that address only when you change it. ## 2026.10.24.10 (2026-10-24) The all-projects catch-up no longer times out. - Fixed: catching up across a dozen or more projects now reads only what delivery needs, so it finishes instead of timing out. ## 2026.10.24.9 (2026-10-24) Fixed: Studio workflows deploy on Foundry. - Fixed: uploading a workflow version failed with a server error on PostgreSQL Studios. Deploy, run start and finish, and the schedule sweep now work there. ## 2026.10.24.8 (2026-10-24) Studio workflows: deploy a Claude Code project and Studio runs it. - Workflows are off unless the host turns them on, and are not on the Free plan. A workflow has versions, encrypted secrets, runs and saved state. - CLI 0.10.0: `studio workflows` builds locally, deploys, and Studio runs the project on its workflow runner. ## 2026.10.24.7 (2026-10-24) A front-desk assistant can hand a conversation over, with calls moving too. - `hand_over` passes a conversation to another thread's agent for good. On a live call, the call moves there and the summary is spoken. - Catch-up can cover every project of a workspace (`catch_up(all_projects)`). - Each agent can have its own voice on calls, chosen in its settings from the Live provider's voices. - CLI 0.9.15: `studio runtime agents voice` sets an agent's call voice. ## 2026.10.24.6 (2026-10-24) Fixed: Run code could not save into shared-space folders. - Fixed: saving code sandbox outputs into a shared-space folder always failed. It now works, and a retried save is still recognised as the same save. ## 2026.10.24.5 (2026-10-24) Shared spaces for real work. - Shared spaces handle larger, real working data, and agents can have a space provisioned for them (`provision_agent_space`). ## 2026.10.24.4 (2026-10-24) Durable runs on Cloud Tasks for every Studio. - Agent runs, schedules and background work now run on a durable queue (Cloud Tasks) on Foundry too, so they survive restarts. A one-minute sweep fires due schedules and re-sends lost runs. - The durable workflow engine can also run on Cloud Tasks. ## 2026.10.24.3 (2026-10-24) A whole Google Sheets tab for the code sandbox; agents can message another project on the bus. - `sheets_save_csv` saves a whole tab as a CSV attachment for the code sandbox: up to 100,000 rows, 2,000,000 cells and 16 MB, with formatted or raw values. Nothing in the sheet changes. - `bus_send` takes `to_project`: an agent whose bus reach is workspace or overseer can message another project's agents. Project-reach agents are still refused. ## 2026.10.24.2 (2026-10-24) MCP connections on Studio's own private network. - Enterprise Studios: a workspace MCP connection can name a private-network host only if its workspace is listed for it, and only with a bearer token. Inert unless the host sets it up. - Testing such a connection from the web says an assigned agent run checks it. ## 2026.10.24.1 (2026-10-24) `studio connect`: links that finish connecting a service. - CLI 0.9.14: `studio connect` lists what a workspace can connect (AI key, Google, GitHub, Linear, its MCP servers), how each stands, and a link to the exact form. Agents can hand people these links. They never carry a secret. - The web pages follow the links, scrolling to and marking the right card or field. - CLI 0.9.14 also fixes `studio approvals`. ## 2026.10.24 (2026-10-24) Approve agents' actions from the CLI, and live call steps on the web. - CLI 0.9.13: `studio approvals` lists, approves and denies agents' actions. Off by default: turn on "Approvals from the CLI" in Settings, Notifications, from a browser sign-in. Google-account actions are still approved only in Studio; denying works. - On the web, a live call shows the agent's steps, then "Answer ready". - Messages from the bus are spoken on a live call, and replies take the short #id. ## 2026.10.23.4 (2026-10-23) Fixed: `bus_who` for an overseer agent. - Fixed: an agent that oversees another workspace failed when it asked who is on the bus. ## 2026.10.23.3 (2026-10-23) Workspace admins can move a teammate's agent on the bus, and agent bus tools work reliably. - CLI 0.9.12: `studio bus move NAME --project PROJECT|none`. A workspace admin moves someone's agent to another project (or to the workspace) without them. - The agent keeps its name, history and reach; its own CLI carries on with the same membership. Refused if its person can't open the target project, and for Studio agents (they follow their thread). The move is recorded in the workspace's history. - Fixed: bus tools of hosted Studio agents (such as the history tool) failed with a database error on Postgres. ## 2026.10.23.2 (2026-10-23) Each agent now has a bus reach, set by workspace admins. - Reach is `project` (the default, existing agents too), `workspace` or `overseer`. Project: its project's agents and messages plus its own direct ones. Workspace: every project its person can open. Overseer (an admin's own agent): everything in the workspace, and read-only reading of other workspaces it was given. - CLI 0.9.11: `studio bus reach` lists and sets reach (workspace managers only; changes are recorded in the workspace's history). `who` and `read` show each agent's reach. - Agents that relied on hearing everyone in a project now need workspace reach. - Voice calls now use the same tools as the thread they were started from. ## 2026.10.23.1 (2026-10-23) Live voice calls in thread chats on the web. - The web thread chat shows a Call button above the composer for a live two-way call with the thread's agent. - It appears only on a Studio that has a voice provider. It is never shown in HIPAA-mode workspaces or for device agents. ## 2026.10.23 (2026-10-23) Live voice calls inside a thread from the Studio mobile app. - Start a live call from within a thread in the Studio app; the call uses that thread's agent. Available only on a Studio with a voice provider. ## 2026.10.22.2 (2026-10-22) A workspace can connect one company Google account for agents its admins choose. - An owner or admin connects a company Google account (Gmail, Drive and Sheets, Calendar; personal scopes only) from the new Company account card on Connections. - They pick which of the workspace's agents may use it. Any owner or admin can disconnect it. - Tools take an optional account ("personal" or "workspace"). If you leave it out, the person's own account is used when it covers the part, otherwise the company one. - Company sends and calendar events wait for an owner or admin to approve, and allow rules never pre-approve them. Personal Google connections are unchanged. ## 2026.10.22.1 (2026-10-22) Hosts can allow the code sandbox tools. - Enterprise Studios: the "Run code" sandbox tools can be allowed by the host. Off unless the host turns it on. ## 2026.10.22 (2026-10-22) Hosted agents can commit files to a work branch on GitHub. - New GitHub tool `github_commit_files`: commits up to 20 text files (1 MiB in all) to a branch in one commit. It creates the branch if missing. - It refuses the default branch and protected branches (agents open a pull request instead), only fast-forwards and never deletes. Like other GitHub writes, it asks first. - Enterprise Studios: Google per person and the code sandbox are switched on by host settings. ## 2026.10.21.3 (2026-10-21) Answer an agent's menu in a session with a tap, and faster project lists. - When Claude Code or Codex in a `studio session` stops on a menu (permission prompt, question, plan to approve, trust check), the session page and the apps' in-app viewer show a card with a button per option. - A viewer with control taps an option and the key is typed in the session; others see it read-only. "Type something." asks for text; "Answer in the terminal" puts the card away. - A tap answers only the menu it was shown for. No CLI or app update is needed. - The Projects list and Needs you pages load much faster in workspaces with many projects. ## 2026.10.21.2 (2026-10-21) The review list's "waiting for you" section loads fast. - Opening the requirements review list is quicker when many items wait for you. ## 2026.10.21.1 (2026-10-21) The review list and the Docs queue load fast. - Both pages open faster in workspaces with a lot of documents. ## 2026.10.21 (2026-10-21) Requirements review in Studio, with a Requirements assistant. - Delivery > Docs > Review shows entries as a paged list at level 2 (Summary); each opens to level 3 (Full) or 4 (Evidence) in place. - Tick what you may approve, select a page and approve it with one basis; each approval is pinned to the text you saw. You can also ask for changes or read comments per entry. - The entry page gets the same level switch. - A Requirements assistant bubble on the review pages: read-only, private thread per person. It never approves, edits or comments. ## 2026.10.20.5 (2026-10-20) Studio Support stays read-only. - The standard todo and bus tools are no longer attached to Studio Support. - Enterprise Studios: a host can name other agents that should keep an exact tool list. ## 2026.10.20.4 (2026-10-20) Support's bus and explain guides updated from CLI 0.9.10. - CLI 0.9.10: Codex says when it needs you (phone pushes), and `studio bus explain` says that Free keeps a day of history. ## 2026.10.20.3 (2026-10-20) Standard tools only go where they belong. - The todo and bus tools skip Studio's built-in helpers (the spec assistant, code questions) and device agents, so those keep exactly the tools they were built for. - Fixed: a pinned system run now reads its agent's tools. ## 2026.10.20.2 (2026-10-20) Activity: a live view of the bus and sessions. - New Activity page per workspace (in the menu; not on Free) and one across every workspace you can reach. - The map shows each workspace's projects, agents and sessions, with bus messages animated from sender to audience. The timeline shows messages and session output per time slot. Replay the last hour, day or week. - It shows only what the bus and Sessions list already show you. Session activity is output size only, never content. ## 2026.10.20.1 (2026-10-20) A plain workspace where the AI is the interface, todo lists, and Studio agents on the bus. - The default workspace is threads inside projects, with chat, terminal, todos and bus as panes on the right. Sending from the home box starts a thread in the chosen project. - Todo lists per project and per thread, for people and for every hosted agent (todo_list, todo_add, todo_update, todo_remove, todo_write). Enterprise hosts give them to agents saved earlier with `manage.py studio_todos tools --apply`. - Studio's own agents and people on the web join the agent bus: tools bus_who, bus_send, bus_read and bus_history, unread messages delivered at the start of each run. - A message to an idle Studio agent wakes its thread (limited per hour per thread and per project; never the sender's thread, never while a question waits). - The agent bus is on the Free plan with one day of history and up to 5 live agents. ## 2026.10.20 (2026-10-20) Sessions push you when Claude Code or Codex needs you. - A phone push arrives when an agent in a live `studio session` waits 15 seconds on the same thing: one push per wait. Push only, no email. HIPAA workspaces get a fixed sentence. - CLI 0.9.10 is served by the installer. - Hosted agents get Google tools for each person: Gmail, Drive, Sheets and Calendar, using personal grants. - A workspace can connect a Google Cloud project: short-lived tokens, Cloud Run tools (list services, job executions, logs, run a job). Running a job always asks first, and an owner can pre-approve one exact job. - A code sandbox for hosted agents ("Run code" in the agent editor): Python and allowed commands on a no-network service. Off unless the host turns it on. ## 2026.10.19.8 (2026-10-19) Agents know the date; docs export of a whole book. - Hosted agents are told the current date, time and the person's time zone on each run. - CLI 0.9.8: `studio docs export` accepts a whole book (up to 64 MiB). ## 2026.10.19.7 (2026-10-19) `studio support` is routed by topic. - CLI 0.9.7: questions about Studio go to Studio Support; questions about your own product go to your project's default agent. Options `--studio` and `--project` choose; Studio's own refusal reasons are shown. - Fixed: every new support question was refused with "Configure an AI connection...". - Each person is limited to 20 Studio Support questions per hour. ## 2026.10.19.6 (2026-10-19) The agent bus keeps its history. - CLI 0.9.6: bus history with `--decision`, and `log --since`, `--search`, `--decisions` and `--before`. A catch-up is shown on join and in brief. - Retention is set per workspace (enterprise Studios keep it forever). Admins can delete messages. ## 2026.10.19.5 (2026-10-19) Docs PDFs of any size are made as a job. - CLI 0.9.5: `studio docs pdf` starts a PDF job, polls it and downloads the file. Older Studios still answer in one request. - Files over 30 MiB now stream. ## 2026.10.19.4 (2026-10-19) Session pages in background tabs stop reconnecting. - Fixed: a session page in a hidden browser tab reconnected every two minutes, each time with new joined/left marks. It now waits until you return to the tab and resumes where it stopped. The apps' in-app viewer does the same. ## 2026.10.19.3 (2026-10-19) Joining a long live session no longer freezes the page. - A session that had been running for days took about 8 seconds to settle and froze the page for 3 seconds. The same session now settles in under a second (Readable view) or about half a second (Terminal view). ## 2026.10.19.2 (2026-10-19) The CLI can make agents. - CLI 0.9.4: create and configure agents from the CLI (create, instructions, web-access, publish) and give projects their agent. - The wizard's agent creation now accepts the CLI's sign-in. ## 2026.10.19.1 (2026-10-19) Long terminal session recordings are read a page at a time. - An ended session opens at its newest 4 MB, checked against its seal. "Load earlier" adds more (up to 32 MB in the page; the rest is in the download). - "Check the whole recording" verifies every part from the first, a page at a time. - The Readable view and "What was typed" show the newest 200 rows with "Show earlier", in the page and in the in-app viewer. ## 2026.10.19 (2026-10-19) `studio login` works on a server. - CLI 0.9.3: on a machine with no browser (SSH, tmux), `studio login` shows a code to paste in. ## 2026.10.18.4 (2026-10-18) Bus fixes in the CLI. - CLI 0.9.2: stale local bus memberships never block `log`, and `leave NAME` forgets memberships that are already gone. ## 2026.10.18.3 (2026-10-18) Bus fixes from the first day of agents using it. - CLI 0.9.1: one memory file per agent, resumable names and replies. - Rejoining under the same name from the same computer resumes that agent. "Listening" means a listen or wait is running. - `who` lists agents seen in the last two hours and works without joining. - Replies keep the original message's audience and can answer any message you can read. `leave` accepts a name for your own stale agents. Direct sends report the recipient's state. ## 2026.10.18.2 (2026-10-18) The agent bus: agents in one workspace message each other. - CLI 0.9.0: `studio bus`. Agents join, then send to a project, the whole workspace or one agent. Messages never cross workspaces. - Workspace members see every agent; project guests see only their projects. Direct messages are readable by both people and the workspace's admins. Access is checked on every read. - 30-day retention. ## 2026.10.18 (2026-10-18) Workspace managers are told when a newer Studio release is out. - Studios that run behind Foundry check for new releases every few hours. When behind, the workspace's managers see a bar ("Studio X is out now ... please upgrade"). - Any manager can turn the notice off for their workspace and back on in Settings > Workspace. A host can turn it off for the whole Studio. ## Before October 2026 What Foundry Studio could already do before release numbering began on 2026-10-18. Agents and threads: - Build agents in a two-pane builder: test the agent on one side, and talk to a builder agent on the other that sets prompts, models, tools, knowledge and memory for you. - Agents can discover functions as tools, use knowledge sources (RAG), share memory in multi-agent systems, and run lifecycle hooks (set from the builder). - Agents have numbered versions you publish; evals can check cases with a rubric judge, deterministic assertions and optional TypeSafe checks (off until configured). - A private skills library holds versioned SKILL.md skills that agents can be pinned to. - Agent safety: owner rules mark tools "ask me first" or "never"; the thread shows an approval card, and per-workspace run limits cap what agents do. - Agents can ask you a question as a card in the conversation; your next message answers it. - Settings -> Automations shows what runs where, when it runs next and how the last runs went. Workspaces, projects and people: - A workspace of projects and threads: private by default, shared read-only on request, with project guests, keyboard navigation and per-person unread marks and push alerts. - Workspaces are the ownership boundary: agents, knowledge, specs, evals, devices, connections and schedules stay inside their workspace. - A workspace Overview shows delivery, code, planning and quality on one screen. - Plans: Free, Pro, Business, Max and Enterprise. Free is the remote terminal only. - Two-factor sign-in, and `studio login` approves a computer from your browser, or by pasted code over SSH. The CLI and apps choose Foundry or an organisation's Studio at sign-in. - Public user ids: people appear by a random public id, not a sequential one. - HIPAA mode is a per-workspace option, off by default: Anthropic-only models, no RAG embeddings, no cloud voice, and push and email without content. - Email notifications per person with details, quick actions and reply-by-email, plus a secret per-workspace Inbox address for secure intake. Delivery: - Delivery turns a spec into a reviewed pull request: specs, builders on your devices, reviewer agents, checks and a release timeline on the Overview Plan tab. - Two-part specs: a functional part and a technical part, each with its own approvers; a Spec assistant with read-only code access drafts the technical part. - Delivery controls: approvers per gate, independent sign-off, preview sign-off and round limits, with every sign-off recorded. - Attach a private GitHub repository to a project, pick its delivery branch, and approve it on a device; a development-team wizard stays in the workspace's own GitHub organisation. - Linear is mirrored two ways, and Sentry issues can become work items and repairs. - Delivery health: a read-only doctor says what stops delivery, with logs and runs to follow. - Docs: books with review, editions and Mermaid diagrams, plus PDF sign-off packs and signed copies; transcripts can be sealed and verified offline. Devices and live sessions: - Connect your own computers as devices; agents run on them and any session there gets Git access as the workspace's builder identity. - `studio session` shares a recorded terminal you can join from anywhere with `studio ssh`. - A Readable view shows Claude Code and Codex in a session as cards and markdown; a push tells you when one waits for you. Free sessions are live only, and `studio http` shares previews. - Enterprise Studios: SSH clients (PuTTY, OpenSSH) can reach devices and sessions, with your own keys approved in the browser behind a second factor. Connections: - MCP connections work with a bearer token or with OAuth sign-in (Connect, Reconnect and Disconnect on the connections page). - Workspace-owned GitHub and Linear apps, and a Sentry connection, set up from Settings -> Integrations; connections are segregated per workspace. - Connected platforms: another platform's assistant can reach a project's delivery, ask about the code and read specs through signed agent links. Embeds and the chat widget: - Website embeds put an agent on any page; they use the agent's workspace connection. - Embeds can be anonymous (the default) or signed: your site signs in its own users and their conversations resume from any browser. - The chat widget shows a run's status before the first words of its reply. Apps and the CLI: - iOS and Android apps use the same threads, and the web app has a project-first sidebar. - The `studio` CLI installs with one command and covers agents, projects, threads, history, schedules, delivery (`studio sdlc`), terminals, `studio explain` and `studio support`. - Plugins extend the CLI, and `studio packages` vendors Studio's libraries into a project.