Skip to content

Open-source AI teammates in Slack, Teams, Discord, Telegram and Mattermost. Self-host it, with any agent harness and any model.

346 KB
Raw Blame

Changelog

All notable changes to Switch are documented in this file.

The format is based on Keep a Changelog, and this project aims to adhere to Semantic Versioning.

Switch ships several independently-versioned artifacts, each with its own section below. Every artifact carries three-part semver (MAJOR.MINOR.PATCH) and a changelog, without exception.

A version says where an artifact is — which release you are running. It says nothing about what that release can talk to. Compatibility is carried separately by the contract revisions in artifacts.yaml, which move independently: a release that changes nothing on the wire bumps its version and leaves its contracts alone.

  • switch-core — the backend service (core/). Released by tagging switch-v<version>; the version lives in core/pyproject.toml. See RELEASING.md.
  • switch-console — the desktop app (dash/). Released by tagging switch-console-v<version>; the version lives in dash/apps/switch-console-desktop/package.json.
  • agent-runtime — the Switch protocol client and MCP runtime (dash/packages/switch-agent-runtime/), published to a package registry.
  • sidecar — the remote runtime Switch Console deploys to an agent host (dash/apps/switch-console-desktop/src/sidecar/). Versioned in sidecar-version.ts and deployed by Switch Console, not published separately.
  • switch-connector / switch-connector-codex / switch-connector-opencode — the three connectors (connectors/), versioned in their respective manifests. The first two are installed from the plugin marketplace; the OpenCode one is written by Switch Console, which has no marketplace to install from, so its version identifies the source rather than an install.

Three things are not separately versioned, and ship under the switch-core release so a single tag pins the whole stack: the operator dashboard (gateway/), the Helm chart, and the standalone compose artifact. Their versions are stamped at package time from the switch-core version; do not add a version of their own to them without also giving them a release of their own.


switch-core

[Unreleased]

Added

  • A one-screen Datadog board and alerts that say what to do. deploy/observability/signal-board.json shows the four golden signals (latency, traffic, errors, saturation) plus whether agents are connected and messaging apps are up, grey while normal and red only past a limit. The monitors are rewritten to what a user would notice: each says what users see and what to check, posts when it fires, every hour while it stays firing and once on recovery, and error alerts count failures instead of taking a percentage, so low traffic doesn't page on noise. Causes such as slow queries stay on the dashboards. deploy/observability/apply.sh creates or updates them in Datadog by name.
  • switch-core reports which bridge crashed. switch.bridges.crashed is 1 per crashed collaboration bridge, tagged with its platform and display name, and 0 once it runs again, so the bridge alert can name it.
  • An agent is told when Switch Trust changed or flagged what it sent. The result of post_message and send_targeted_message, the agent HTTP message endpoint, the status endpoint, and the media upload endpoint (for a caption and the files) now carries a trust field when Switch Trust did something: {"outcome": "redacted", "delivered": "<what the room received>"} when parts were redacted, or alerted / errored with the note posted under the message. A redaction that came with a flag or an incomplete check, say a caption redacted beside a file that was flagged, also says alert: alerted or errored. A message Switch Trust left alone has no trust field, so its result is as before. Until now an agent whose words were redacted had no way to know.
  • How many users and agents each workspace has. GET /gateway/tenants/counts answers for every workspace the caller may enter, counting people as the workspace's members list shows them: its members, and its organization's admins who are not one. The dashboard's new workspace switcher reads it.
  • A dev-only role switcher for the dashboard. just seed-dev-roles creates local test accounts for an org admin, a workspace admin and a member, plus a second workspace only the org admin can open, and just gateway-dev, on a server with gateway_v2 on, shows a Role: control in the top bar that signs in as any of them, or as the operator from .env, so a page can be checked as each role. Switching is an ordinary sign-in, so every check is the server's own. A production build leaves the switcher out, and npm run build now fails if it gets in.

Changed

  • The dashboard no longer claims the system is healthy. The "System healthy" indicator in its top bar was hard-coded and never checked anything, so it is gone.
  • GATEWAY_ADMIN_PASSWORD is optional when password login is off. With GATEWAY_PASSWORD_LOGIN_ENABLED=false the admin account is seeded without a password, and the first OIDC sign-in with GATEWAY_ADMIN_EMAIL links to it. A password set anyway is logged as unused at startup. While password login is on it is still required, and the server now says so at startup instead of failing on a missing setting. The Helm chart follows: with switchCore.oidc.passwordLoginEnabled: false (and Mattermost off), secrets.gatewayAdminPassword may be left empty, and the Secret key becomes optional.
  • Switch Trust notices about a person's own message reach only that person. When someone's message is blocked, redacted, flagged or not fully checked, the notice used to be posted in the channel, where everyone could see what they had tried to say. Now it goes to the sender alone: an ephemeral message on Slack and Mattermost, and a direct message from the bot on Discord and Telegram. A direct message only works for someone who accepts one: on Discord, unless they turned off messages from server members; on Telegram, only once they have started a chat with the bot, which most members of a group never have, so there most notices are still posted in the group. Microsoft Teams has no way for a bot to do either unprompted, so there the notice is still posted in the channel, as it is for a post by an app or by Telegram's anonymous group admins. switch-core logs a warning each time a notice goes to the channel for want of a private route. Notices about an agent's message are still posted for the room.

Security

  • Only the people in a direct or group conversation decide who is in it. The room routes and the agent tools added agents and people to one from outside it, though its platform fixes its members; a DM the bridge adopted is writable by every member, so anyone could put their own agent in someone else's DM and have it receive their messages. The room routes now refuse adding or removing with a 409, as do invite_agent_to_room, add_users_to_room and remove_agent_from_room, and a template whose new agent joins one is refused before anything is made. Rooms carry can_change_members, false for one, and Switch Console no longer offers it. The people in the conversation still invite and remove agents with !invite-agent and !remove-agent, and a Switch Console chat still changes its own through its chat routes.

  • Adding to a package takes being able to read what is added. Only the package owner's reading was checked, and a package lists what it holds to everyone who may read it, so someone who could edit a publicly writable package could add a private reference or document by id and then read it.

  • A refused agent transfer names the services it uses to its owner only. An admin who tried to transfer someone's agent was told which of the owner's connections it was granted, which the owner's service access otherwise shows nobody.

  • Sign out of all devices. Signing out used to clear only that browser's cookie, so a copied cookie or a browser left signed in elsewhere stayed valid until it expired. The account menu now has Sign out of all devices, which ends every session the account holds (POST /auth/sessions/revoke), and the server administrator can do the same for anyone (POST /users/{id}/sessions/revoke, recorded in the server audit log). Changing a password, including a reset made directly in the database, and losing the server-admin role also end every session; the device that changed its own password stays signed in. Sessions from before this release stay valid.

  • Dashboard sessions end 14 days after sign-in. A session could be refreshed forever, so a browser that stayed active never had to sign in again. Refreshing or switching workspace now keeps the original sign-in time, and 14 days after it the session ends and the person signs in again, however active they are. A session still ends after 24 hours without a refresh. Sessions from before this release count from when their cookie was issued.

  • Registration keys expire. A registration key made under API keys worked until someone deleted it. A new key now expires after 90 days by default; when creating it you can choose 1 year instead, or Never, which warns that the key keeps working if it leaks. POST /api-keys takes expires_in_days (1 to 365, or "never"; 90 when left out), the key list shows each key's expiry or No expiry, and an expired key is refused with 401 by the agent API, the MCP server and agent registration. Its owner can still reveal it under API keys, to see which key it was. Keys made before this release, agents' own keys and the deployment's bootstrap key do not expire.

  • Messages with files attached are checked by Switch Trust. A message with a file used to skip the check entirely, typed text included, both from a person on a chat platform and from an agent sending files. Now the text is checked as any message is, and so is the text inside each file: plain text, JSON, YAML, CSV, Markdown, source code and anything else that reads as text whatever type or name it is sent with, Word documents (.docx) and PDFs. All of it is checked before anything reaches the room. If the text or any file is blocked, nothing from the message is delivered and the sender is told why. A file whose text Switch Trust would redact blocks the message too, since a file cannot be redacted safely. Files with no text to read, such as images, archives and other binaries, are delivered without a check or a note, since Switch never reads their kind. Scanned or encrypted PDFs and files that cannot be opened are delivered with a note on the message saying Switch Trust did not check them. Only the first 20,000 characters of a file are checked (and at most 200 pages of a PDF): anything past that reaches agents unchecked, and the message carries a note saying the rest was not checked. switch-core now depends on pypdf to read PDFs.

  • Logs no longer carry email addresses. Signing up, creating a user, creating or deleting an API key, seeding the admin account and finding two accounts with the same address each logged the person's email. They now log the user id. A test fails if a log call passes an email again.

  • Counts and replies no longer name private rooms either. Beyond the room listings below, a reference, document or package's attached-rooms count took in every room it was attached to, an agent's detail named every room it is in, and removing an item from a package named every room that lost it. Each now names, and counts, only rooms the viewer may read; workspace admins still see every room. Whoever may manage a resource is still told how many rooms they cannot open it is in, without their names, and deleting it warns about those rooms too.

  • A refused request is no longer repeated back. When the gateway or the agent API refused a request as invalid, the 422 repeated what was sent: a password too short to accept, or the whole body, password included, when another field was missing. Registering or editing a Slack, Discord or Teams connection with a config that did not fit quoted part of its bot token or secret, and on an edit that could be the stored token, which nothing else returns. Anything that records error responses (APM, proxies, browser HAR captures) could end up holding passwords and tokens. Each error now says only where and what is wrong (type, loc and msg), and the agent API no longer writes the request body to its log. A script that read input or ctx from a 422 will find them gone. A validation rule that raised its own error on the agent API now answers 422 rather than 500.

  • Agents are no longer handed a key that never expires. Datadog connects with an API key and an application key that do not expire and that only Datadog can revoke, and an agent granted Datadog was handed both. Switch now refuses: Datadog can no longer be connected or granted, and an agent with an existing Datadog grant is told why when it asks. A Datadog connection made before can still be disconnected; revoke the keys in Datadog if they were ever granted to an agent.

  • Service connections start off in a new workspace. A workspace created from now on has every service connection switched off until a workspace admin switches each on under Connectors. Workspaces that already had members keep every service on by default, as before, along with every switch an admin already set.

  • A Google Workspace token is never handed out with more than an hour left, and one Google issues for longer is refused.

  • Withdrawing access says what it could not take back. Removing a grant, disconnecting, or switching a service off for the workspace cannot revoke an Atlassian or Google Workspace token already given to an agent, since that token is the owner's own. The change now marks those tokens withdrawn and its warning names the service and the time until which they may keep working: up to an hour for Google Workspace, and up to eight hours for Atlassian, whose tokens cannot be asked for shorter.

  • A template's room params name only rooms its runner can see. A room param checked its value against every room on the server, and $first or a list of candidates could settle on someone else's private room. That told the runner the room existed and could write its name into their new room. Now a room the runner cannot read reads as missing.

  • Registering an agent under a parent requires owning the parent. The agent API's register-known and register-known-bulk accepted any agent as parent_agent_id, copying its name and settings to the new child. They now refuse a parent the registering user does not own, as the dashboard's own subagent registration already did.

  • Private rooms stay out of room listings for people who cannot read them. The room link graph, the rooms a document, package or reference is attached to, and an agent's list_all_rooms now show only rooms the caller may read (an agent also sees the rooms it is in). Linking a room to another now needs read access to the target, whose name the link shows, including links an agent's create_room makes. A room's links, outbound and inbound, leave out rooms the caller cannot read, for people and agents alike. An agent can list a room's participants only from inside the room.

  • Only a workspace admin manages room groups. Groups are shared by the whole workspace and have no owner, so any member could rename, move or delete one. Creating, changing and deleting a group, from the dashboard, the API or an agent's create_room_group, now needs a workspace admin (for an agent, its owner). Group room counts, and the rooms get_room_group_detail lists, now leave out private rooms the caller cannot read.

  • Only a workspace admin puts a room in a room group. A room's group can decide who may address an agent, since an addressing policy can admit everyone in a group's rooms, so any member who filed a room they could write into a group could address that group's agents. Putting a room in a group, whether when creating it, from the room's page, from the groups page or through an agent's create_room, now needs a workspace admin (for an agent, its owner). Running a template that makes a room group now needs a workspace admin too, and an agent cannot run one, since it runs templates without its owner's admin reach. Taking a room out of its group still only needs write access to the room.

Added

  • The dashboard's new layout, behind the gateway_v2 flag. Overview, Templates and Connectors on the rail; Workspace and Organization settings as areas of their own, each with a menu of only the pages you may open, reopening the page you last used; and the operator's pages under Platform. Today's pages move into it, and their old URLs lead there. It is still being built, so turn it on for development and staging only. With it off, the dashboard is laid out as before, and the new URLs lead back to today's pages.

  • The new layout's Overview counts the workspace's agents, users and rooms. Each card opens its list, and View graph opens the ecosystem graph, which leaves the new rail. A workspace's admins see its totals; a member sees the agents they own and the rooms they take part in, marked "Yours", and every user. A count that cannot be read shows a dash and why.

  • The gateway says what the caller may do with every object it shows, so no page works a permission out for itself. Alongside the room, template and agent flags already there:

    • references, reference types, documents and packages carry can_edit and can_manage, documents and rooms can_transfer, and rooms can_change_members;
    • agents carry can_edit_options, can_rename, can_transfer with transfer_blocked_reason, and, on their detail, can_change_manage_agents and can_grant_services; an agent's sub-agents carry them too;
    • workspace and organization members carry can_change_role and can_remove, workspace members can_receive_transfer, and organization members can_grant_org_admin;
    • GET /tenants/current carries can_delete and can_manage_data_protection (setting messages to be deleted, listing and erasing people: its admins by their own membership), and GET /organizations/current administers and can_manage_members;
    • API keys carry can_reveal and can_delete, machines can_edit and can_remove, managed agents can_edit and can_delete;
    • messaging apps carry can_manage and can_delete, installs can_disconnect, claims on a chat account can_unlink, connections can_revoke, and the operator's list of accounts can_delete. Each is decided by the same rule its route enforces, and a test holds every flag to its route for each kind of caller. A flag says whether the caller may, not whether there is anything left to do: an ended install or a revoked machine says that by its own status.
  • The server's administrator can delete someone's account. The Users page has a Delete action (DELETE /users/{id}) that removes the account in every workspace, with the same refusals as deleting your own. Administrator accounts and your own account are refused, and every attempt is recorded in the server audit log with its outcome.

  • Transfer ownership. The owner, or a workspace admin, can hand an agent, room, package, reference, reference type, library document or template to another member of the workspace (POST /ownership/transfer, and a Transfer ownership action on the detail pages). Each transfer is audited. An agent that uses its owner's service connections, is managed on its owner's machine, can manage agents, has sub-agents or signs in with its own OAuth client is refused; any other agent gets a new API key that only the new owner can see, so it has to be reconnected. Chats and room documents are not transferable.

  • Workspaces keep messages for a chosen time. A workspace admin sets a message-retention window on the Workspace page; once an hour, messages older than it are deleted with their attachments, and stored files no message carries any more are swept. Lapsed invitations, settled approval requests and expired install links are deleted 30 days after they end.

  • A workspace admin can erase a person. The People list on the Workspace page erases everything a person sent in the workspace, in every room: their messages, the files only those carried, the copies kept for agents, and their name on approval answers. It works for people on a chat platform, people from a chat app that has since been disconnected, and members who wrote from Switch Console. A chat-platform person's identity goes too; a Switch Console member keeps their account and can go on writing. The admin confirms by typing the person's name, and the request and its outcome are recorded in the audit log.

  • A deleted account's messages read "Former member". Deleting an account keeps the messages it sent, so conversations stay whole, but takes the person's name off them, off the copies kept for agents and commands, and out of the names of the Switch Console chats they were in. Invitations addressed to their email are deleted in every workspace, whether or not they were still open. Erasing the messages themselves is the workspace admin's erase action above.

  • A deleted workspace's audit history goes after two years. It is kept while the workspace exists and for two years after it is deleted, then deleted by the hourly retention pass. The server's own record of the deletion stays. Deleting audit rows needs the schema owner's connection (DB_OWNER_USER and DB_OWNER_PASSWORD); without it, switch-core logs an error at startup saying the history will not be purged.

  • Workspace admins can delete an empty workspace. Workspace settings has a Delete workspace action that asks for the workspace's name (DELETE /tenants/{id}). It is refused, with what is left by kind and count, until the workspace has no rooms (archived ones included), agents, packages, references, reference types, documents, templates, room groups, messaging or service connections, machines, connectors, skills or other members. Its audit history is kept, and the deletion is recorded in the server audit log. The server's default workspace can't be deleted. Deleting your account now also deletes a workspace you leave empty.

  • The gateway says what the caller may change in a room. Rooms, in the list (GET /rooms) and in a room's detail, carry can_write, can_manage and can_unlink. Each room a reference, document or package is attached to carries can_unlink, and a room's documents, listed or fetched one at a time, carry can_remove. A reference, document or package carries hidden_rooms_count: for whoever may manage it, how many of the rooms it is in they may not read. All are decided by the same rules the routes enforce, so a client offers a change only where the server would allow it.

  • An agent controller can send its agents' operations on its socket. Every operation a controller's agents call was an HTTP request of its own, parsed, routed and authenticated with the controller's token, though the controller already holds an authenticated WebSocket to the server. The socket now takes them as call frames and answers each with a call.result frame carrying the status and body the HTTP route (POST /agents/{id}/ops/{operation}) would have answered with. A call is authorized as that request is: the controller checked against its row, and the agent's binding to it. The socket's first frame offers calls (calls.max_in_flight, 64); a call over that is answered 429 capacity_exceeded without running, so the controller sends it over HTTP.

  • Organizations can follow WorkOS (experimental, off by default). With ORGANIZATION_DIRECTORY=workos, an organization an operator links to a WorkOS organization (switch-link-workos) takes its members, and who is its Org Admin, from there: at sign-in, from WorkOS's events every 30 seconds, and in a full check once a day. A change from WorkOS is applied even when it breaks a rule Switch keeps (the last Org Admin, agents left without an owner who can act for them), and the Organization page tells its admins what is left to fix. With a members service configured (MEMBERS_SERVICE_URL, WORKOS_CLIENT_ID, WORKOS_REDIRECT_URL), an Org Admin's role changes and removals on the Organization page are made in WorkOS too, with the admin's own WorkOS token, which the dashboard gets by sending them through the organization's sign-in when needed. Without one, making someone Org Admin is done in WorkOS, and taking it away in Switch lasts only while WorkOS agrees. Adding someone new is done in WorkOS either way. Setup and limits: docs/old/workos-directory.md.

  • A reconnect storm is paced instead of taking the pool. When a pod goes away, every agent and controller it held opens a connection on the pods left within seconds. On staging (two pods, 1,000 agents, one pod deleted) those opens took the whole database pool: requests and the process lease timed out with them, and the surviving pod dropped agents of its own. Each process now opens at most AGENT_OPEN_LIMIT (8) agent and controller connections at once. An open that finds no slot within 2 s, which keeps it well inside the timeout of the proxy in front, is refused before authentication with 503 and retry_after_s, its place in the line at the pace opens are taking in that process, so the storm comes back as a queue spread over what the process can take. A caller that comes back early keeps its time. switch.agent.opens_deferred counts the refusals, by route.

  • The Helm chart can put switch-core on chosen nodes. switchCore.nodeSelector and switchCore.tolerations go onto the switch-core pods as written, so they can run on a node group of their own, such as one kept for load tests. Both are empty by default, which places the pods as before.

  • Organizations set up their own connector apps and switch services for their workspaces, behind organization_connector_apps, off by default. With the flag on (and service_connections), the dashboard's Organization page gains two sections for organization admins:

    • Connectors. An organization saves its own GitHub App, Google client or Airtable integration, checked with the vendor as the operator's are. Every workspace in the organization then signs in, is issued and is refreshed through it rather than the server's. An organization's app that stops working makes the service unavailable in its workspaces, with the reason, and never falls back to the server's. A workspace's service listing says whose app it goes through (app_source).
    • Services. An organization switches each service on or off for its workspaces. Unlocked, its setting applies where a workspace admin hasn't set the service themselves; locked, it applies everywhere and workspace admins can't change it. Wherever a change leaves a service off, the tokens issued for it there are revoked; any a change could not reach, the periodic token upkeep revokes within five minutes. A workspace admin can drop their own switch (DELETE /gateway/workspace/services/{service}) to take the organization's setting.

    Every change is recorded in the organization's audit log. The platform operator has no say in an organization's apps or services unless they are one of its admins. With the flag off, nothing changes: the routes aren't mounted and no workspace consults its organization. The migration adds two empty tables either way.

  • Google Workspace without Google's yearly security assessment. A server can set SERVICE_ACCESS_PROFILES='{"google-workspace": "drive-file"}' so sign-ins ask for neither drive nor drive.readonly. Google counts those two as restricted: an External app needs its strictest review before anyone outside the organization can grant them, with a security assessment renewed every year. With drive-file, agents still read and change any Doc, Sheet or Slides deck the person can open, by its link, and Calendar as before; in Drive they reach only the files and folders they create, so they cannot search the person's Drive or read their PDFs and other uploads. Agents are given a skill that says so. The setting covers every Google sign-in on the server, an organization's own Google client included, so an organization with its own Internal client also loses Drive search. Changing the setting makes everyone connected before the change reconnect: until they do, their connection shows as needing reconnection and their agents get no Google Workspace token. With Helm, set switchCore.serviceAccessProfiles (and switchCore.disabledServices for DISABLED_SERVICES). Without the setting nothing changes.

  • The [CONN] log lines say when a client fell back to the event stream. /events takes optional fallback and fallback_close_code parameters, and the [CONN] opened and [CONN] reattached lines end with fallback=<reason>:<close code> only when they are present, so a new agent runtime whose WebSocket never opened is told apart from a client built before the socket. Unknown reasons log as other and anything but a short numeric close code is dropped.

  • Templates carry packages. A room in a template takes packages:, a list of packages that already exist, by id or by name, attached when the room is created. The person running the template must be able to attach each one by hand; a package that is missing, ambiguous or not theirs stops the run before the room exists. A template never creates or changes a package. Capturing a room (GET /rooms/{id}/yaml) now lists its packages by id, so a room built again from its capture gets them back (#910).

  • An agent can read its own addressing policy. get_agent_detail now returns addressing: the kind of policy (owner_only, owner_and_owner_agents, anyone or custom) and, when an agent looks itself up, its rules with the people, agents and rooms they name. So an agent can tell someone who may answer its questions and approvals instead of guessing. Another agent sees only the kind.

  • Console hears when your managed agents or machines change. A signed-in user's Console opens one WebSocket per workspace (/gateway/changes/ws), and switch-core tells it when one of their managed agents or machines changed, so Console reads it again instead of polling. A notice names a kind and an id and carries nothing else. Notices travel between replicas on the signal bus, so a change made through one pod reaches a Console connected to another (#717).

  • The Airtable connection. A person connects Airtable from Switch Console, choosing at Airtable's consent screen which bases and workspaces to share, and a granted agent reads and updates records and comments in them, as that person, through Airtable's MCP server. Each server's operator registers and publishes an Airtable OAuth integration and saves it under Connector apps, or names it in AIRTABLE_CLIENT_CONFIG_PATH (Helm: switchCore.airtableClientSecret); see docs/old/airtable-setup.md. Without it, Airtable is listed as not set up. Like Google Workspace, it needs GATEWAY_PUBLIC_URL: Airtable sends people back to the server's own callback. On a server without one, both are listed as unavailable (needs_public_url), saying why, rather than failing when someone connects. Airtable offers no read-only sign-in, so every connection can write; changing tables and fields is not asked for, and Airtable refuses it. Airtable cannot revoke a sign-in, and an integration reaches the bases of every grant a person gave it, so a sign-in Switch stops using still works and widens their next connection. Disconnecting or reconnecting says so and where in Airtable to remove it, as any OAuth service without revocation now does rather than reporting a failed revoke.

  • Connector apps set up Airtable. An app's entry in GET /connector-apps gains redirect_uris (every address to register), needs_public_url and scopes. The dashboard's Connectors page walks Airtable's Builder Hub, with its redirect URL to copy and the scopes to tick, and saves only once the operator confirms the integration's support details are set: no check can see whether it is published, so a passing one says so. Saving an app answers with its checks too, so the page can show them. A passing Google Workspace check now says it can't see the redirect URI either: Google compares it only when someone signs in.

Changed

  • The dashboard, and what Switch tells people, use Switch's own words. A computer that runs agents is a device, not a machine; an outside service is a connector, and a chat app a messaging app, not a connection; a person is a user, with "member" kept for the role; and nothing calls Switch a server. Dashboard labels are in sentence case. Register Agent is now Add agent, Register messaging app is Add messaging app, and the API keys page lists Keys for adding agents and Agent keys. The same words reach the errors, refusals and status the server sends to the dashboard, Switch Console and chat, and messages that sent people to "Settings, Connections" in Switch Console now point to Connectors, where that page is. Text written for agents still names their tools, such as list_machines, and Switch Console's own cloud-machine messages are unchanged for now. A dashboard test fails on the old words, naming what to write instead.
  • Admins step down and leave from their own row. The workspace and organization members pages offer changing your own role and leaving, while another admin remains; the last one is refused, as before. Leaving the workspace your session is on moves the session to your other workspace, or to none, as deleting a workspace does, rather than leaving it on one you left.
  • A room's own document is changed in its room. PATCH and DELETE /documents/{id} refuse one with a 409, and its library flags say no.
  • Revoking an invitation that has already ended is refused with a 409, rather than rewriting when, and by whom, it ended.
  • An addressing rule that can never let anyone in is refused when it is saved, with a 422 naming the rule and why: one that names no rooms or no room groups, or lets in no people, no agents, not the owner and not the platform. Next to other rules it changed nothing while looking as if it did, and a policy of only such rules shut the agent off, its owner included. Policies already saved are left as they are.
  • Removing a member hands what they own to an admin. Their rooms, packages, references, reference types, library documents and templates in that workspace stayed theirs, though they could no longer reach them, and were only cleared when their account was deleted. They now pass to the admin removing them, or, when the removal comes from the organization's directory or from an operator who is not a member, to the workspace's longest-standing other admin. The same happens in a workspace an organization admin reached only through that role, when they lose it. Each is audited as an ownership transfer. Chats, a room's own documents and a template whose name the admin already uses, in any workspace, stay as they were, and a member who owns agents still can't be removed until those are transferred or deleted.
  • One role runs a workspace: owner merges into admin. Every workspace owner becomes an admin, and every admin now manages every member, other admins included, as well as everything an owner could. A workspace always keeps at least one admin: demoting or removing the last one is refused (409), and so is deleting the account of the last admin of a workspace others are in. Two admins demoting each other at once can't both succeed. A client that still asks for owner, in an invitation or a role change, gets admin. The organization issue workspace_without_owner is now workspace_without_admin. The migration is not expand-only: until every replica runs this release, an older one creating a workspace writes owner, which the new check refuses, and an older one sees former owners as plain admins with no last-admin rule. Roll every replica together.
  • Only the owner or a workspace admin can change a reference's, reference type's, package's or library document's visibility. This tightens permissions: before, anyone with write access could, so on a publicly writable resource any member could make it private and lock everyone else out of it and the rooms that use it. Editing its other fields still needs only write access, as it does for rooms.
  • Unlinking a resource from a room is for those taking part in it. It took write access to the room, which on a publicly writable room, the default, is everyone's. It is now for the room's owner, people in it (in Switch Console or through a claimed platform account), owners of agents in it, and workspace admins, whether or not they may write to it. The dashboard shows it as an "Unlink" button rather than a bin, since nothing is deleted. Attaching still takes write access.
  • A room's own documents are deleted by their creator, the room's owner or an admin. Anyone who could write to the room could delete them. A document an agent made counts as made by the agent's owner.
  • Usage telemetry tells people from bots. Slack, Discord and Mattermost posts from an integration (an app, a bot account, a webhook that isn't Switch's own) were counted as people. The bridge now marks them, and the daily snapshot reports person_identity_active_*, bot_identity_active_* and account_active_* beside the old chat_identity_active_* and user_active_*, which are still sent for one more stable release. Room and agent message events report such posts as sender_kind = bot. An app's post on Discord or Mattermost also no longer answers a request card or wakes an agent following the conversation, as already on Slack.
  • Usage telemetry follows Switch Cloud. Every event now carries deployment_kind (switch_cloud, managed_by_console or self_hosted), so Cloud can be charted on its own or left out. New events cover its funnel: account_created (sign-up, a first single sign-on sign-in, or an admin), signup_refused with the reason, workspace_created, and a cloud machine being requested, becoming ready (with how long a new one took), failing (with the reason), being woken, and being stopped. Machine events are sent only once their change is committed.
  • Usage telemetry counts each finished day, week and month. The daily snapshot's *_7d counts are a trailing week, so a chart adding them up over a week counted a room once for every day it was active. A new usage_period event is sent once per finished UTC day, ISO week and calendar month, with distinct active rooms, people and accounts for exactly that period, dated to its middle. Summed at the matching interval it is the real figure. A server that was off sends the periods it missed, up to 35 days back, when it next runs.
  • Agent telemetry says connection where it meant one. agent_connected, agent_disconnected and the first_agent_connected milestone are sent alongside agent_session_started, agent_session_ended and first_session_started, which measured an agent coming online, not a coding-agent session. The old names stay for one more stable release. The new first_agent_session milestone is the first session a host reports. A deployment that already passed either milestone claims it at startup without reporting it.
  • Workspaces belong to organizations. Upgrade by replacing every pod, and downgrade the database before rolling back. Every existing workspace gets an organization of its own, and its owners become the organization's admins. An organization admin administers every workspace in it without being added to each. A release from before organizations doesn't know about organization admins: if old and new pods serve together during a rolling update, or after a rollback that keeps the new database, an owner removed from a workspace by an old pod is still the organization's admin and can come back into it. So for this release, replace every switch-core pod at once rather than rolling them, and to roll back, first downgrade the database to the revision before 238c68b7306e (alembic downgrade 238c68b7306e-1), then start the older release.
  • Messages no longer queue for the database's NOTIFY lock. Postgres lets one transaction that sends a NOTIFY commit at a time, and the triggers on messages and agent_events sent one from every message's transaction, so the whole deployment stopped near 300 messages a second however many pods or vCPUs it had. The router now writes without them and announces its rows once they have committed, through one announcer per process that sends everything queued since its last send in a single transaction (announcer.py). Same channels and payloads, so pods of the previous release and this one hear each other during a rolling update; the migration is expand only. A reader re-checks its rooms, and the agent event listener forgets the heads it was told, every 20 to 30 seconds, so an announcement lost with a process that died right after its commit costs a delay, not a message. Signals go through the same announcer.
  • The statements every message write costs run on asyncpg directly. Taking the room's lock, reading its head, storing the agents' own events and reading the room's agent members again under the lock go straight to asyncpg on the transaction's connection (db/raw.py), as the delivery reads do, rather than through SQLAlchemy's execution.
  • The two reads behind every delivery run on asyncpg directly. A stream reads an agent's events, and counts its unread messages, for every event it delivers, and under load SQLAlchemy's work around each statement cost more than Postgres took to run it. db/raw.py runs a statement SQLAlchemy compiled once on the session's own asyncpg connection, in its transaction and under its tenant, with SQLAlchemy's type conversions and timed into db.query.duration as every other statement is.
  • Pods start in seconds again with many rooms. Every process checks at startup that each room has its clients, under a lock that lets one process at a time through, and read each room on its own to do it. With tens of thousands of rooms that took about 100 seconds, so processes starting together were restarted while they waited. Rooms are now read a thousand at a time.
  • Idle agents behind a controller cost the database far less. A controller stream's pass, including the one after every idle interval, read each of its agents' events in a session of its own; it now reads all their heads in one query and reads only the agents with something new. And each process's 30-second re-check of controllers held by another process read every agent's rooms with a query of its own, which grew with processes times agents; it is now one query per controller.
  • A delivery to an agent behind a controller takes two database transactions instead of four. The controller's stream read the heads of all its agents to find the work, again after delivering, and again before waiting. It now reads them once a pass, and an event that arrives after that read wakes the stream for the next pass.
  • A controller keeps its socket when a delivery pass cannot get a database connection. When a pod was too busy to hand its streams a connection within the pool timeout, the controller's stream ended and its socket closed, and every agent on it reconnected, which needs more of what was short. The pass is now tried again after a short wait, doubling to 5 seconds, and the socket closes only if it is still failing after a minute.
  • Switch's own reader is woken only by commands. The process that holds a workspace's admin client read every row of every room to find the !command rows it answers. A new row's announcement now names its type, and that reader is woken for command rows only and reads only those, from a partial index on them. Its warnings for tags nobody can answer are already posted by the process that writes the message.
  • The heads-up for tagging an agent that isn't in the room comes from the process that wrote the message, and costs nothing when every name tagged is one of the room's agents. The admin's consumer, which reads every message in a workspace on one process, checked each message with an @: two database sessions and a lookup for every name tagged. Under load that one process fell behind the rest. The check now runs once a message has committed, on whichever process wrote it, and only when the message tags a name that is none of the room's agents (a role, a person, an agent not in the room). The warnings read as before.
  • The server spends less CPU on its event loop and on finding the agent sending a message. It now runs on uvloop wherever uvloop is installed, which is every platform but Windows. Finding the actor of the agent sending a message no longer looks through every client's actor, and compiled mention patterns are kept per name. In a CPU profile under message load the event loop's own work was about 15% of a pod's CPU and the actor lookup about 3%.
  • Creating a room or an agent no longer looks through every running client. Every process keeps an actor for every client, and finding a workspace's admin, or the actor of an agent registered in another process a moment before, read all of them. Each is now found from an index of the running clients by type and by agent. While 10,000 agents were being created, these lookups took about 11% of a process's CPU in a profile.
  • Each database transaction costs the server a little less. Binding the workspace to a transaction built and compiled the same statement every time; it is now sent to the driver as it is. And timing who holds a pooled connection worked out the borrower's name and a short stack on every checkout; the name is now kept per function, and the stack is written out only for a hold long enough to log.
  • Usage analytics read the database far less per message in a large workspace. With usage reporting on, each message looked up its room's platform and member counts, its sender and the agent's runtime, each kept for five minutes in caches of at most 10,000 entries. Quiet rooms missed every time, and with thousands of agents the caches evicted what they were about to be asked again. A room's facts are now kept for fifteen minutes, a sender and an agent's runtime, which do not change, for an hour, and up to 50,000 of each.
  • A post's first checks are kept until a signal says they changed. Every post checks that its agent is in the room and that no budget covering it has run out, and a controller relaying it checks the membership again: two transactions per post. Each process now keeps those answers as it keeps a controller token's rows. Triggers on room_agents, rooms and usage_budgets signal every change, so an answer is kept while the signal connection listens and for AGENT_AUTH_CACHE_TTL_SECONDS otherwise. A refusal is never kept. ROOM_ACCESS_CACHE_MAX_ENTRIES (20,000) bounds what is kept.
  • An agent's stream reads in one database session per pass, and pooled connections are no longer pinged before use. A pass that delivered a message opened a session for each of its reads, six pool checkouts and six transactions, and the pre-ping added a transaction to every checkout. The reads of a pass now share one session, for an agent on its own socket and for each agent an agent controller serves. DB_POOL_PRE_PING now defaults to false: after a database restart the first query on each pod fails, along with any query running at the time, and the pool then replaces every connection it held. Set it to true behind anything that drops idle connections. In a load test of 300 agents at 10 messages a second on 2 pods, CPU per message fell by a quarter and database transactions per message by more than two thirds. At most half of a pod's pool goes to streams' passes at once, so requests still get a connection when the pod's CPU is saturated.
  • A cloud machine can move to another instance type and keep its disk. The hosted controller now launches every new instance of a machine as the instance type in its gateway.json, at the same moment it already moves the machine onto the configured image, and upgrade --gateway-config moves a stopped machine onto both. Changing instance_type therefore resizes existing machines as well as new ones; a machine that only stops and starts keeps its instance and type. allowed_instance_types is no longer bound to the controller's state database, so it can change while machines exist, and status and list report instance_type_allowed for each machine. A database bound before this release is accepted and rebound on first start, as long as allowed_instance_types has not changed in the same deploy.
  • The Helm chart runs switch-core as a ServiceAccount of its own. It is <release>-switch-core (switchCore.serviceAccount) where it used to be the namespace's default, which every pod that names no account shares, so anything trusting switch-core's identity, such as the distributed Teams app's federated credential, trusts switch-core alone. create: false with a name runs it as an account made elsewhere. A Teams app whose federated credential trusts default stops getting tokens after this upgrade: add a credential for system:serviceaccount:<namespace>:<release>-switch-core first, and remove the old one once the new pod is up.
  • Agents and controllers are pinged every 10 seconds instead of every 2. A connection is now declared dead after 30 seconds without a beat, up from 6, so a client that goes away holds its room slot and shows online for up to 30 seconds. The interval is advertised on every open, so clients follow it without an update. The agent auth cache now gives each entry a random 80 to 120% of AGENT_AUTH_CACHE_TTL_SECONDS, so entries filled together after a restart do not all expire together, and that setting must stay below 25 seconds (the 30 second heartbeat TTL, less the jitter) or startup refuses it.
  • Following the conversation is described the same way everywhere, and the chart can turn it on. The feature added in 0.31.0 is off unless follow_the_conversation is among the enabled feature flags (FEATURE_FLAGS_ENABLED, or switchCore.featureFlags.enabled in the Helm chart). The Switch skill and the agent runtime's instructions no longer tell agents that a message without their name can never reach them: a person's follow-up can, while other agents' chatter still does not. The chart lists the flag and, while it is on, sets FOLLOW_THE_CONVERSATION_WINDOW_SECONDS from switchCore.followTheConversation.windowSeconds (default 300).
  • Agents name room groups by path. create_room and create_room_group take a group as a path such as Fraud/Incidents (the top-level group, then each child), so two groups called Incidents under different parents no longer break every agent request that names one. A bare name still works when only one group has it; when several do, the error lists each match by its path so the agent can retry with the right one, and an unknown path names the first part that was not found. A group id is accepted too.
  • Room group names are unique among siblings. Creating, renaming or moving a group next to one with the same name (ignoring case) is refused with an error naming the existing group, from the dashboard, the API and agents alike. The same name under different parents is still fine. Deleting a group whose subgroup would move up next to a namesake is refused too, and the dashboard now shows why. Names can no longer contain / or be blank. Existing duplicates are left as they are and can be reached by id.
  • Delivering an event to an agent behind a controller no longer reads the agent's head first. The notification of a new agent event already carries its number, and the controller's stream now keeps it as the agent's head, so a delivery costs the agent's own read alone, where it read the heads of all the controller's agents first. The heads are still read after each idle interval, when an agent is attached, and after the listener reconnects.
  • Delivering an event to an agent costs three statements, where it cost seven. The rooms dropped after the agent's cursor are read with its events, and the unread count reads the agent's baseline, the room's window and the agent's user id in the statement that counts.
  • Writing a message no longer costs four statements for each agent it is for. The agents' own events are stored together, in two statements and one savepoint, where each agent's event took a savepoint, a numbering, an insert and a release of its own.
  • A controller's stream waits on three events, where it waited on one per agent and two more. Every attached agent's doorbell now rings the stream's one bell.
  • Reading an agent's own events no longer builds its query each time. Each shape of the read is built once and run with bind parameters.
  • A message addressed to a connected agent no longer reads the agent's membership after it commits. That read stops an agent removed from the room in the meantime from saying anything there, and a connected agent has nothing to say about a message addressed to it, so the read is now made only before a refusal or an offline notice is posted, or a machine woken.
  • Message telemetry reads a workspace's rooms, clients and agents at once. The first lookup that misses reads all of them, where each room, sender and agent was read on its own, so in a large workspace most messages no longer cost a lookup.
  • Registering, renaming or deleting an agent no longer makes every process read every client and agent. The change now names the clients it touched, and each process reads only those, where it read every client of the workspace and, with a messaging app connected, every agent as well. The 30-second re-check reads every client only when a fingerprint of the workspace's clients and agents has changed since its last full read, where it read them all every time.
  • The 30-second re-check of which controller runs each agent reads one query, and nothing when no placement changed. Every process read every placed agent's definition, document included, then each controller with a query of its own. It now reads the columns a binding is made of in one query per workspace, and only when a fingerprint of the placements has changed since its last full read.
  • The 30-second re-check of who holds each controller's socket no longer reads every controller each time. It compares a fingerprint of their connections and of the live processes first, and when it reads, reads only the connection columns, where it read every row with its status and platform, and the rooms of every agent on a controller another process held.
  • Writing to a room for the first time through a process no longer reads the room. A process's clients now share the room ids they look up, where each kept its own, and the first one a process misses reads every open room of the workspace at once, so with many agents behind a load balancer most sends no longer start with a read of the room.
  • Checking a message against Switch Trust no longer reads its settings each time. Each process keeps the settings it read until they change: saving or clearing them in the gateway tells every process, so a change still takes effect on the next message. Settings are read again after 30 seconds at most, for a change whose signal was lost.
  • Reading an agent's budget standings no longer builds their query each time. The standings are read twice for every message an agent posts, and building the query cost more than running it; it is now built once.
  • Reading a room's context costs one transaction. The room's newest message number, the membership check and the history read share one session, where each opened its own.
  • A controller's requests no longer read its row and its agent's row each time. They were kept for 5 seconds, and a controller's requests, spread over several processes, rarely came back to one within that. They are now kept until a signal says they changed, which triggers on agents and agent_controllers send, and a miss reads all of the controller's agents at once. CONTROLLER_AUTH_CACHE_MAX_AGENTS (20,000) bounds the agent rows kept.
  • Reading an agent's rooms uses an index. room_agents is indexed on agent_id; reads by agent scanned the whole table, 2.9 ms each at 10,000 agents.
  • Recording a message's usage builds its statement once. The insert was built for every message an agent sent; it is built once and run with bind parameters.
  • A controller's access token is verified once. Each process keeps the tokens that verified until they expire, where every request verified its token again; revocation is still checked on every request.

Fixed

  • A room's detail says whether the caller takes part in it, and whether it is a Switch Console chat. The gateway's GET /rooms/{id}, and the room it returns from a create or change, always answered viewer_is_member: false and is_console_chat: false, so Console's room settings named a Console chat's messaging app "None". Both now say what the room list says.
  • !list-documents and !list-references answer again. Since they began listing what a room's packages bring, both failed on the admin client that runs them, and the channel got no answer at all.
  • The dashboard's addressing policy editor judges rules as the server does. A rule letting in only the agent's owner, or its owner's agents, was taken for one that lets in nobody, and Save was held; the editor now shows both choices, and says why a rule can never let anyone in in the server's own words. A restricted policy with no rules lets anyone address the agent, as documented, though the editor said it let in nobody: choosing Restricted now starts with an owner-only rule, the last rule cannot be removed, and a policy saved with no rules shows as Open.
  • A dashboard page someone may not open says so instead of loading. Typing the address of a page they may not open, or one the server does not run, loaded it into a column of refusals; it now says why. A member could open the operator's Users page; it is refused now.
  • The dashboard offers each action only where the server would allow it. Every action follows the server's own answer for that thing, rather than the dashboard's guess from someone's role or ownership. An organization admin was offered Delete workspace in a workspace they are not an admin member of, and was refused. Anyone may edit a publicly writable reference, document or package, but only its owner and admins were offered the fields. An agent that cannot be handed over now says why beside a disabled Transfer.
  • A bridged room's description can be edited. It was locked, though nothing sets it from the channel; whoever may write to the room can now edit it, as on any other room.
  • The dashboard reads the server's feature flags instead of guessing. Agent management, connectors and organization connector apps were taken to be off when a page answered "not found"; they are now read from the server's flags, and if those cannot be read the dashboard says so rather than hiding pages.
  • Two organization admins demoting or removing each other at once could leave the refused one without their keys and connections in the workspaces they reached as an organization admin: those were taken away before the last-admin check that refused the change. The check is now made, under a lock held until the change is saved, before anything is taken away.
  • Deleting an API key an agent still signs in with failed with a 500. It is refused with a 409 that says to delete the agent; once the agent is gone, its leftover key can be deleted.
  • Attaching a document or package that is already attached, to a room or a package, failed with a 500. It now changes nothing, as for a reference, and attaching anything a room already has is no longer reported as a new attachment.
  • A package in a room's list said the caller may open it only on that list; every route now works it out from the package's read rule.
  • Stable Switch Console can link chat accounts again. Since links have to be confirmed from the account with a one-time code, a claim from a Console that cannot show the code was refused with "Update Switch Console", and no stable release could show it, so nobody on stable could link a new account. Switch now issues the code anyway and returns it as the 409's text, which stable Switch Console shows in the link dialog: what to post from the account and that the code expires in 30 minutes. An account already confirmed links normally. An admin linking someone else from an older client is still refused while the link needs confirming, since the code would be shown to the admin.
  • A file that names nobody now follows the conversation like text does. A post with one or more files and no mention never reached the room's assigned agent, the agent whose thread it was posted in, or the agent the sender was just talking to, even with follow-ups on; the same words as a text message did. A post is now decided by the same rules, read from its first file's caption and thread, and the agent gets the post with every file on it, in that thread.
  • The dashboard offers only the room changes the server would allow. On a publicly writable room, the default, anyone could change the Access setting and press Save, only for the server to refuse it. Someone who could only read a room was still offered adding and removing its agents and users, attaching and detaching its resources, linking it to other rooms and deleting its room documents. A document, reference or package's page offered attaching it to, and detaching it from, rooms the viewer could not change. The rooms list offered archiving, and the groups page filing, rooms the viewer could not change; a bulk archive skipped those without saying so. Each of these was refused when tried. Access is now read-only for everyone but the room's owner and admins, and its description says so. Everything else is offered only to those who may write to the room, and only rooms they may write to are offered for attaching. A reader still sees each agent's "Notify on join" setting, but cannot change it. A publicly writable room's other settings stay editable by everyone, as before.
  • Those taking part in a private room can read its resources. The room itself could be read by its participants, but its references, documents, packages and linked rooms were refused to anyone but its owner and admins.
  • Deleting a room document from the room page asks first. It cannot be undone, and its own page already asked.
  • Reconnecting Google Workspace no longer ends the new connection. On a reconnect, Switch revoked the sign-in it replaced. Google ends every sign-in of the account through the client when one is revoked, so the new connection also stopped working, and the next token refresh asked the person to reconnect again. A reconnect to the same Google account through the same client no longer revokes the old sign-in. A reconnect to another account still revokes it, and so does a disconnect. A sign-in that Switch does not keep, because the workspace switched Google Workspace off during it, no longer ends the existing connection either.
  • A client in more than 32,767 rooms starts reading them. The reads a client makes as it starts matched its rooms with one parameter each, and Postgres takes at most 32,767 per statement. Switch's own client is in every room of its workspace, archived ones included, so past that many it failed to start, retried, and failed again while the process ran, taking up rooms one at a time in between. The rooms now go as one array.
  • On Slack, the 👀 and the activity card no longer arrive after the agent's reply. Every running turn's card was redrawn every 5 seconds to tick its clock. Slack closes a streamed card after 10 minutes, so on a longer turn each tick was a full message edit, and Slack allows about 50 of those a minute for the whole workspace. A few long turns used that up, Slack made the bridge wait, and new turns' cards and 👀 waited behind the clock redraws. The clock now moves once a minute, and a running card shows whole minutes ("Working… 3m"). Clock redraws stop while Slack is making the bridge wait, and new turns go ahead of them. The 👀 goes on as soon as a turn starts rather than once its card has posted. A card is redrawn as soon as Slack closes its stream, instead of showing still-running steps as failed until the next change. A turn whose first draw into its stream fails keeps that card rather than posting a second one under it.
  • An agent is told why a machine's CLI cannot be used. When create_agent or update_agent_detail was refused because a provider was not logged in, it said the owner needed to log in, even when the machine had reported a login its agents could not use, which logging in does not fix (a Codex login in the OS keyring, say). The refusal now gives what the machine said instead, and list_machines shows it as each provider's problem.
  • A Teams app on a federated credential can deploy from the Helm chart. switch-core's migrate init container and the pre-upgrade migration Job were given the Teams app's settings but not its projected token, so TEAMS_APP_FEDERATED_TOKEN_FILE named a missing file and both refused to load their configuration: with switchCore.teamsApp.credential: federated, no release could migrate or start. The two that migrate are no longer given the Teams app at all.
  • The agent HTTP API answers a refused or malformed request with a 4xx, not a 500. Inviting an agent to a room the caller may not change answered 500 instead of 403, and so did looking up a room the agent is not in. A room history request whose since or before is not an ISO 8601 timestamp answered 500 instead of 400; it now says which parameter was wrong, and so does read_context (#1045). One whose offset moves it outside the years 1 to 9999 in UTC is refused the same way instead of answering 404, and a time in the last millisecond of 9999 is no longer refused at all.
  • Creating a user refuses a role the server does not understand. An account is a user or an admin, and only the exact role admin grants operator rights. POST /users accepted any role string, so Admin or owner created an account that was shown with that role, was stored with it, and had no operator rights. Anything other than user or admin is now a 422 and creates nothing. The Add User dialog offers only those two, so only scripts calling the API were affected. Accounts already stored with another role are left as they are, still listed with that role and still treated as ordinary users; SELECT id, email, role FROM users WHERE role NOT IN ('user', 'admin') finds them.
  • Every way of setting a password now applies the same rules. Signup already required 8 characters and at most 72 bytes. Changing your own password checked only the 8 characters, and an admin creating a user checked nothing, so a short or empty password went through, and a long passphrase was saved as just its first 72 bytes, the most bcrypt reads. Both now refuse what signup refuses, saying why. Creating a user also trims, lowercases and checks the email the way signup does, so the same address with a stray space no longer makes a second account, and an address with no dot in its domain is refused. A script that creates users through POST /users gets a 422 for either. Passwords already saved still work (#1035).
  • Full garbage collections no longer stall the server for hundreds of milliseconds. Once boot has built what lasts for the life of the process (modules, the ORM's mappers, the app's routes), the server freezes it out of garbage collection, so a full pass walks only what was made since. The clients and bridges started after that stay collectable, since replicas hand them to one another. The dashboard's pause chart now shows the average pause next to p99, as Datadog keeps the average for longer.
  • Agents no longer read offline while their replica shuts down. With several replicas, a stopping replica released its lease before its 5 s drain, so for up to about 6 s the other replicas counted its connected agents as gone: @role mentions did not reach the holder, rooms got false "offline" notices, and stop and reset commands were refused. It now drains first and closes its sockets as it releases.
  • A request card posted in a thread now mentions the person who can answer it. Approval and question cards go into the thread of the turn that raised them, and on Slack they used to mention nobody there, so an owner who was not in the thread was never told and the agent waited. A card in a thread now mentions the person who asked when the agent's addressing policy lets them answer, otherwise the agent's owner if they may answer, otherwise nobody. It never mentions more than one person, even when the policy lets anyone answer. Cards at the top of a channel mention whoever asked, as before.
  • !list-references and !list-documents include what packages bring. They listed only what was attached to the room directly, so a room whose references all came from a package answered "No references in this room." while its agents were sent them. Items from attached packages now follow the direct ones, each marked with the package it comes from (for example (github, via Payments service context)); an item attached both ways is listed once, as direct (#911).
  • The template guide agents read now covers a room's roles:, references: and docs:. All three already worked in a room resource and appear when a room is captured, but the guide stopped at aliases:, so an agent writing a template did not know they existed or what shape they take. It now gives each one's fields, a sentence on what it does, and a short example.
  • switch-buddy no longer says the Mattermost switch waits for a server restart. Turning it on starts Mattermost straight away on a running local server, but switch-buddy's instructions still said it took effect at the next start, which sent people to restart the server for nothing (#1021).
  • The Teams command menu fits Teams' limit again. Switch's own Teams app listed thirteen commands, one more than Teams allows, once /remove-agent joined it. Both Teams apps now offer the same twelve: /remove-agent is in, and /room-url leaves the menu but still works when typed. A self-hosted Teams app shows the change once its package is rebuilt with --bump and uploaded again.

Removed

  • Breaking for HTTP clients: the /agents/{id}/moderation/… routes are removed and answer 404. They predate the operations door, which runs the same operations as MCP. Call POST /agents/{id}/ops/<operation> with the operation's arguments as the JSON body instead: POST moderation/rooms is create_room, POST moderation/rooms/{room_id}/invite is invite_agent_to_room, GET moderation/rooms is list_all_rooms, GET moderation/rooms/{room_id} is get_room_detail, GET moderation/agents is list_agents, and GET moderation/bridges is list_bridges. The reply is {"result": …} in the operation's own shape, so list_all_rooms gives room_id where the old route gave id.
  • !run-cmd, and the terminal command in offline replies. A CLI started with that command got none of the Switch tools, so it never brought the agent online. The reply for an agent that is offline or not live now says what's wrong and who has to act, with no start command to paste. Where the owner starts it from Switch Console, the reply names the room to connect it to and the provider's sign-in command, for when Console says it is signed out. An always_on agent says its connector is down, whatever provider it registered through. Slack apps that list /run-cmd in their manifest keep showing it, answering "Unknown command", until an admin deletes it; Discord and Telegram drop it from their command lists by themselves (#730).
  • The agent event stream and POST /connection/beat. Agents connect over the WebSocket at /agents/{id}/connection/ws, which Switch Console has used since 0.39.0. A client still asking for the Server-Sent Events stream is refused with 410 transport_removed and told to update, and agent-protocol accepts rises to 8, so Consoles 0.38.2 and older can no longer connect their agents (#673).
  • Deleting your account now also clears workspaces you were removed from. Removing a member leaves what they own in that workspace, and deleting their account afterwards failed on it. Those rows are now deleted or handed over as they are in the workspaces you are still in.

[0.31.0] - 2026-10-20

Added

  • One address per Switch server. A deployment is reached at a single address instead of separate gateway and API hosts (#753).
  • Agents can follow a conversation without being @-mentioned. A person's untagged reply in a thread, their next message in an exchange with an agent, or any message in a room with an assigned agent reaches that agent. It is off unless follow_the_conversation is among the enabled feature flags (switchCore.featureFlags.enabled); people control it per room with !follow-ups on|off, !assign-agent and !unassign-agent (#869).
  • Template version 2. A template can start with version: 2 and describe its agents and rooms as resources:, each with a type:, that refer to each other by key. An entry can be made conditional with if:, and {key.name} fills in another resource's name. Templates without version: 2 keep working as before (#851).
  • Official templates. Switch ships templates every workspace sees, listed first, with ids starting official:. They can be run but not edited or deleted. The Switch expert, which used to be bundled in Switch Console, is now one of them (#851).
  • Event-loop stall probe. switch-core checks for event-loop stalls every 100 ms and reports them, so a blocked loop shows up (#713).
  • The server's operator sets up connector apps from the dashboard. A Connectors page lets you set up your GitHub App and Google Workspace's OAuth client with no restart. Changes are recorded in a new server-wide audit log (server_audit_events). To use this you need the service_connections flag on.
  • A connection knows the app it signed in through. A connection made through an app the operator has since replaced shows as needing reconnection, and an agent asking for its token is told to reconnect rather than handed the vendor's refusal. A sign-in that started before the app was replaced is asked to start again. Connections made before this release are unaffected.
  • Confluence through the Atlassian connection. Connecting Atlassian now also asks for Confluence's read, search and write scopes, so a granted agent reaches Confluence pages as well as Jira. People who connected before must connect again to get Confluence.
  • Operators can keep names away from outside avatar services. An agent with no icon of its own is drawn by api.dicebear.com with its name as the seed. With THIRD_PARTY_AVATARS_ENABLED=false (Helm switchCore.thirdPartyAvatars.enabled), no names are sent to outside avatar services and each platform shows its own default icon.
  • The Helm chart can take the invitation SMTP username from its Secret. Some relays issue usernames that are credentials in their own right (an Amazon SES SMTP username is an access key id). With switchCore.smtp.usernameFromSecret: true, the chart reads it from secrets.gatewaySmtpUsername, or from GATEWAY_SMTP_USERNAME in secrets.existingSecret, instead of switchCore.smtp.username. Setting both is refused at render time.
  • Agent controllers update themselves to the release Switch names. Set switchCore.controllerRelease (version, URL and SHA-256, all three or none) and any controller running another version downloads that release, checks it and installs it once it is idle. Every Switch release now publishes a matching controller package, release candidates included, so controllers follow the server they belong to. To roll back, point the values at the previous release. Switch Console's own controller still updates with Console.
  • Workspace admins can turn a service off for their workspace. With the service_connections flag on, an admin can switch off a service such as GitHub or Google Workspace. While it is off, nobody in the workspace can connect it, grant it to an agent or get a token for it, and running sessions lose it within half a minute (GitHub tokens are revoked straight away). Grants are kept, so turning it back on needs no re-grant. Services are on by default, so an upgraded server behaves as before.

Changed

  • Upgrading from 0.30 or earlier: two things now work the other way round.

    • switch-core can run as multiple replicas. Before 0.31.0 the chart refused any switchCore.replicaCount above 1. Now any replica can serve any agent, each bridge and connector is held by one replica at a time, and if a replica goes away another picks up its work. With more than one replica an upgrade rolls a pod at a time instead of taking the server down. Agent events now live in Postgres rather than one process's memory, so a restart no longer loses what was queued (#886, #815, #858).
    • Running a template creates agents. Before 0.31.0, run_template created rooms only, and agents= filled every agent slot with an agent you already had. Now a run creates each agent the template describes, as a managed agent on the machine the entry names (this needs agent management and the "can manage agents" permission). agents= now means "use this existing agent instead of creating one", and only for an entry that says allow_existing: true. Every agent entry also needs provider, machine (or location) and directory, so a template written for 0.30 gets warnings and won't save until you add them or save with bypass_warnings=true (#851).
  • A caller repeating a refused request is answered with 429 from memory, rather than re-doing the refused work each time (#750).

  • Console links work on chat platforms. A switchdash://open/... link an agent posts reaches Slack, Discord, Telegram and Teams as a clickable https link to the new public GET /deeplink/open/{page}, which opens that page in Switch Console. A server with no public address (GATEWAY_PUBLIC_URL) sends the link's label and the page's name instead. Mattermost servers that allow the switchdash scheme get the link unchanged.

  • Deleting a room can retire its channel. DELETE /gateway/rooms/{id} and bulk delete take retire_channel (default false). When it is set, Switch archives the room's Slack or Mattermost channel, deletes its Discord channel, or has the bot leave its Telegram chat; Teams channels and direct messages stay. Without it, the channel is left as it is. The delete responses say what happened (channel), and room detail says what retiring would do (channel_on_delete). A channel that could not be retired is reported, and the room is still deleted. Disconnecting a messaging app leaves channels as they are.

  • The Switch expert is offered only where switch-buddy is not. A server with agent management and switch-buddy on serves switch-buddy alone; every other server keeps the official Switch expert. Agents made from it are unchanged.

  • Usage telemetry goes to telemetry.switchagents.ai by default. The default TELEMETRY_ENDPOINT (Helm switchCore.telemetry.endpoint) was https://telemetry.flintai.dev, an older name for the same relay that is being retired. A deployment that sets the endpoint explicitly is unaffected; one relying on the default moves to the new host on upgrade. Nothing is sent unless TELEMETRY_ENABLED=true.

  • Documentation links point at docs.switchagents.ai. Links in the docs and the synced docs/official pages used docs.flintai.dev, the older name for the same site. Old links still redirect.

  • A sign-in that returns through the server goes straight back to Switch Console. Google Workspace and GitHub send the browser back to the server rather than to the computer it left from. The server used to stop there on a page naming the Switch account and asking the person to continue; it now sends the browser on to Switch Console at once. The page remains only when the return cannot be verified or the consent was cancelled. What ties a sign-in to the browser and the Console that started it is unchanged.

  • Signing in to a service connects it without a confirm step. Once the browser returns, the server exchanges the code, reads the account and keeps the connection in one step, and Switch Console shows the result. It used to hold the sign-in until the person confirmed the account in Console. The account is chosen at the vendor, and Disconnect undoes a wrong one. The flow's complete route now answers with the account and any warning, its status reads connected or failed with the reason, and the confirm route is gone, for GitHub's flow as well; a Switch Console from before this release cannot finish a sign-in against this server, so update both together. Migration bf016e18bc16, expand-only.

  • New agents answer their owner and the owner's own agents. Every path that creates an agent without an addressing choice (registration, create_agent, template runs, switch-buddy's template) now gives it the "me and my agents" policy instead of owner-only, so the agents one person creates can address each other out of the box. A template agent with no addressing gets owner-agents. Other people and their agents stay out, an explicit owner-only choice is kept, and existing agents keep their policy. The "others can use these grants" warning no longer counts the owner's own agents.

Removed

  • The /admin command. It did nothing and is no longer offered in rooms or registered with Slack, Discord or Telegram; a command without a handler is never registered.
  • Breaking (agent-facing): the task protocol is removed. The delegate_task, accept_task, update_task, finalise_task, cancel_task and list_tasks tools are gone from MCP and the operations door, the /agents/{id}/tasks… HTTP routes answer 404, and no task events are delivered. The documentation has said not to use them since 0.9.x. Agents coordinate with post_message and send_targeted_message, and the instructions they receive no longer mention tasks. connect_to_room drops each participant's can_delegate and can_accept, and neither !status nor the dashboard shows them. Stored integration profiles that still carry task_protocol load as before, with the key ignored. Existing rows in the tasks table are left in place; no migration.

Security

  • An agent icon URL must spell its host plainly. A link with a backslash, or a host with percent-encoding or non-ASCII characters, is refused: browsers and chat clients read such a host differently from Switch, so its checks could not tell where the image would be loaded from.

Fixed

  • A controller Switch Console runs on an SSH host is no longer shown as behind. It enrolls like a standalone controller, so Core named it the release and the Machines page marked it behind, though it runs Console's bundle and cannot follow Core. Console now starts it with --updated-by-console; it follows no release, reports controller.updated_by: "console" in its status, and Core names it none, as for Console's own controller. One started from an older Console bundle shows behind until Console replaces that bundle.
  • Usage telemetry no longer carries the ENVIRONMENT or SERVICE_NAME settings. From 0.28.0 through 0.30.0, with TELEMETRY_ENABLED=true, every event sent to the Flint relay carried both, as deployment.environment, service.name and the OTLP scope name. They are free text an operator sets for their own logs (Helm switchCore.logging.environment and serviceName), so a value such as a company name left the deployment despite the promise of no free text. Usage events now always say switch-core and omit the environment; flint_env still says which Amplitude project they belong in. The operational OTLP export to your own collector is unchanged.

[0.30.0] - 2026-10-12

Added

  • Multi-tenancy. Several workspaces per person, with sign-up, invitations, workspace roles, per-workspace usage budgets, and Switch Cloud sign-in (#627).
  • Distributed Telegram app. One Telegram app any workspace connects, like the other platforms (#583).
  • Distributed Teams app. A Teams app any organisation installs for itself (#632).
  • Switch Trust guardrails. Every message is checked before it is sent; blocked, redacted and alerted/errored outcomes are handled, blocked senders are named and redacted senders notified. Configurable from a Switch Trust settings panel backed by a gateway API (#679, #739, #747).
  • Per-message telemetry and activity tiers. switch-core reports per-message events and rolls activity into tiers for the dashboards, on top of an events refactor and dashboard-accuracy fixes (#719, #610, #558).
  • Room-wide mention. A mention that pages a room's people on its platform and wakes no agent (#542).
  • Windows support for managed agents and the controller (#765).
  • Self-contained integration points. Adding a messaging platform (#720) or an agent CLI (#763) is now a single plugin entry.
  • Build provenance and usage reporting. switch-core reports the commit it was built from with its metrics and logs (#710), how long garbage collection pauses the process by generation (#711), and the count of active staff accounts in the usage snapshot (#724).
  • Deploy-time feature flags. Capabilities are gated by a feature-flag setting (FEATURE_FLAGS_ENABLED, Helm switchCore.featureFlags.enabled) rather than per-capability environment variables. Known flags: agent_management, hosted_agents, switch_cloud, and ecosystem.show_owners.
  • An agent or controller that keeps repeating a refused request is answered 429 from memory. After 8 identical refusals (401, 403, 404 or 426 for the same bearer token, or the same controller credential, on the same method and path of an agent or controller route) within two minutes, Switch answers that request 429 with Retry-After for a minute without looking up its credential or touching the database. One request goes through when the wait is over; refused again, the wait doubles, up to 15 minutes, and any other answer clears it. It protects the server from old clients that retry a refusal forever. New metric switch.http.refusals_throttled, with a panel on the dashboard.
  • Switch cloud machines that run the agents controller sleep when idle, and wake on a message. Such a machine is stopped once its status reports say no session has run for HOSTED_IDLE_STOP_MINUTES (released, as before, when no agent is placed on it). A message addressed to one of its agents starts it again; the room is told once that the machine is starting, and Switch keeps the message for the agent until it is back. Placing an agent on a sleeping machine starts it too, and is judged by its last report. A machine its owner stopped stays stopped, and says so.
  • Switch cloud machines can run the agents controller. With HOSTED_MACHINE_RUNTIME=controller (Helm switchCore.hostedMachineRuntime) and agent management on, a cloud machine claimed from then on enrolls the standard switch-agent-controller with a one-time code Core hands the hosted controller at prepare (the same code on every retry of a revision, valid 30 minutes), instead of running the hosted worker. The controller it enrolls as becomes the machine's (hosted_machines.controller_id, kind ec2, "Switch cloud"); its status reports make the machine ready; the managed agents placed on it are the machine's agents, and keep it from being retained. Launches are refused in that mode, and such machines are not stopped when idle yet. Machines claimed before keep the worker runtime. Migration e7a2c4b9d013.
  • Provider logins sealed to a machine. A controller registers an X25519 public key (at enrollment, or once afterwards); its owner gives the machine a provider login sealed to that key, which Switch stores and relays as ciphertext it cannot open (PUT/GET/DELETE /gateway/management/controllers/{id}/provider-logins/{provider}, and GET /v1/management/controllers/{id}/provider-credentials/{provider} for the controller). Giving one queues a provider.login operation the machine takes it up with; revoking the machine deletes its logins. Migration d41c7a9e2b58 adds sealed_provider_logins.
  • The service_connections feature flag shows or hides service connections. Off, which is the default, the routes for connections, sign-ins, keys and grants are not mounted: agent hosts read that as no grants, and Switch Console and the dashboard offer GitHub's link alone, as before. Turn it on with FEATURE_FLAGS_ENABLED=service_connections (Helm: switchCore.featureFlags.enabled) for everything below.
  • Agents use their owner's GitHub through grants. A person connects their GitHub account once; an agent's owner then grants it repositories, to read or to read and push, from the agent's page in the gateway or in Switch Console. The agent's sessions get installation tokens for exactly those repositories, valid for at most an hour, each issue recorded. They are revoked when the grant narrows or goes, the owner disconnects GitHub or re-links another account, the agent is deleted, or its owner leaves the workspace. Agents with a runtime of their own cannot be granted anything yet. New routes: GET /agents/{id}/service-grants and POST /agents/{id}/service-tokens/{service} for agents and their controllers, and /gateway/service-connections and /gateway/agents/{id}/service-grants for people. Pushes, pull requests and comments made this way appear as the GitHub App.
  • Agents can use their owner's Jira, through the Atlassian connection. A person connects Atlassian from Switch Console; nothing needs configuring on the server, which registers its own OAuth client with Atlassian on the first connect. An agent's owner then turns Atlassian on for it: on or off, at the level the person consented to when connecting (read, or read and write). The agent acts as its owner in Jira, through Atlassian's MCP tools, which its sessions call without the token ever reaching the coding tool. Only Jira's scopes are asked for. Connecting from the web dashboard is not possible yet.
  • Services beyond GitHub from the catalog. A catalog entry can name the generic OAuth/MCP adapter: its sign-in, the vendor's MCP servers, and whether its tokens are minted per issue or the owner's own, passed through with a lifetime cap. A service whose OAuth client the operator registers reads it from the JSON file <PREFIX>_CLIENT_CONFIG_PATH names; unset, the service is shown as not set up on this server, and set but unusable, the server does not start. Connecting goes through /gateway/service-connections/{service}/flows.
  • Agents can use their owner's Google Drive, Docs, Sheets, Slides and Calendar, on self-hosted servers. The server's operator registers an internal Google app and names its client in GOOGLE_WORKSPACE_CLIENT_CONFIG_PATH (see docs/old/google-workspace-setup.md); without it, Google Workspace is listed as not set up. A person connects from Switch Console, signing in to Google through the server's own callback; their agents' sessions then run Google's gws command-line tool with the person's access, on/off like Atlassian. Gmail is not offered. Disconnecting revokes the sign-in at Google.
  • A catalog service's tools can be the vendor's command-line tool. An oauth entry (the generic adapter, formerly oauth-mcp) names either the vendor's MCP servers or a cli block: the binary and its pinned builds with their SHA-256s, the commands and flags allowed and refused, which arguments are local files, an output cap, a timeout, and how a run says the vendor refused its token. auth.oauth.authorization_params sends extra sign-in parameters a vendor needs, from a fixed allowlist. The grants answer gains cli_tools, additively.
  • Services people connect with a key, starting with Datadog. A catalog entry with adapter: api_key names the fields a person pastes (one to three secret keys) or picks (a site or region, never free text), a call that checks them, and header or environment templates that put them into the vendor's MCP requests or CLI runs. Core checks a key with its vendor before storing it, encrypted, through PUT /gateway/service-connections/{service}/key, and hands it to sessions an hour at a time like any pass-through token; the coding tool never sees it. Keys are not refreshed, and disconnecting forgets one but cannot revoke it at the vendor, which the answer says. The grants answer's MCP servers carry their headers templates and a CLI its env ones (both additive). Datadog needs MCP access turned on for the organization and an application key with mcp_read.
  • DISABLED_SERVICES switches catalog services off on a server, as JSON mapping each to the reason people are shown (an empty one reads "Switched off on this server."). A switched-off service is listed with its reason and cannot be connected, granted or issued a token; disconnecting still works.
  • The service token answer gains expires_in and use_until: a holder asks again by use_until, at most an hour away, whatever the vendor's token lifetime. The grants answer gains each service's mcp_servers. Both are additive.
  • A grant change reaches running sessions within a minute. The grants answer carries each grant's revision and the token answer the grant_revision it was issued under (both additive). An agent host holding a token reads the grants again at most every 30 seconds and asks for a new token once its grant has moved on, so a widened grant, one turned off, or one removed and made again applies without restarting the session. A service first granted after a session started reaches it when it next starts, as its tools are set up then.
  • The dashboard's Service access shows services such as Atlassian as on/off switches beside GitHub's grant, and says why one cannot be turned on: not connected yet, or switched off or not set up on this server.
  • GITHUB_APP_CONFIG_PATH and GITHUB_APP_PRIVATE_KEY_PATH configure the GitHub App (both or neither; the key path absolute). The Helm chart sets them from switchCore.githubConnectionsSecret and the new switchCore.githubAppPrivateKeyPath, the signing key's path in a Secret the chart mounts. SERVICE_TOKEN_RETENTION_DAYS (default 30) sets how long issue records are kept; tokens' values are dropped as soon as they expire.
  • The Helm chart turns agent management on from its values. switchCore.agentManagement.enabled and secrets.controllerTokenSecret (or CONTROLLER_TOKEN_SECRET in secrets.existingSecret) set AGENT_MANAGEMENT_ENABLED, CONTROLLER_TOKEN_SECRET and CONTROLLER_STATUS_INTERVAL_SECONDS on switch-core and its migration Job. Until now they had to be added with kubectl set env after each upgrade, rolling the pods twice. Rendering fails when the secret is missing or shorter than 32 characters. Remove variables set by hand before the first upgrade that renders them (see the chart README).
  • global.preStopSleepSeconds (default 5): the gateway waits that long in a preStop hook before SIGTERM, so a roll stops sending it requests before it shuts down while its replacement is already serving. Rendered only on Kubernetes 1.30+ (the native sleep action); 0 turns it off. switch-core and Mattermost roll with Recreate, so they get no hook.

Changed

  • The published Helm chart pins its images by digest. The release writes each image's registry digest into the chart's values before packaging, so switchCore.image, gateway.image and setup.image default to <image>:<version>@sha256:…, and global.imageRegistry defaults to the registry the release published to. Pinning the chart version now pins the exact images; a deployment no longer needs to set the three image values.
  • Breaking for some values files: global.imageRegistry now has a default. An image value that names a registry (ghcr.io/acme/switch-core:1.0, localhost/…, host:5000/…) is used as written, as before. One that names a Docker Hub namespace without a registry (acme/switch-core:1.0) now gets the default registry prepended; write it as docker.io/acme/switch-core:1.0.
  • imagePullPolicy for switch-core, gateway and setup defaults to "": IfNotPresent for a digest-pinned image, Always for anything else, so an install that opts out of the pin with a mutable tag keeps pulling it. An explicit policy still wins.
  • Every merge to main publishes a dev build of the images and chart, under ghcr.io/<owner>/dev/ and <next patch>-dev.<run>.g<sha>: never latest, apart from the release packages, and pruned after 30 days beyond the newest
    1. Each publishing run records every digest in its summary and in a release-pins artifact, and a version already in the registry is never published again.
  • The hosted controller is published with every release and dev build: its image as switch-hosted-controller and its chart as charts/switch-hosted-controller, versioned with switch-core and pinned to its image by digest (image.repository, image.digest). The dev charts also carry a main tag that follows the newest build of main.

Changed

  • The agent connection runs over a single WebSocket between the server and the runtime, replacing the previous connection (#652, #687).
  • Room access-permission changes are restricted to owners/admins (#638).
  • The bundled Mattermost no longer calls Mattermost Inc. (#733).
  • Agent icons switched to DiceBear gaze (#685).
  • Faster startup and fewer requests. Agents and rooms are read in bulk at startup (#743), the room and agent list queries are batched so they don't grow with the data (#709), and Console makes fewer managed-agent requests with one query for their machines (#702).
  • Agent management moved behind the agent_management feature flag. The old AGENT_MANAGEMENT_ENABLED setting is gone: switch-core refuses to start when it is still set to true, and the Helm chart refuses to render with switchCore.agentManagement.enabled: true. Enable the agent_management flag instead.
  • Hosted launches require agent_management. HOSTED_LAUNCH_CAPACITY above 0 now needs the agent_management flag enabled.
  • GitHub moves onto service connections when the new release first starts, once per workspace: each GitHub connection. A connection that cannot be read is skipped and logged, and its owner connects GitHub again. So is a second person linked to the same GitHub account as someone else: they must connect a different GitHub account, or the first must disconnect theirs. A workspace whose move fails is logged and tried again at the next start; the others still move. The old table is left as it was, but nothing is written back to it: after rolling back to the previous release, people connect GitHub again (GitHub renews sign-ins by replacing them, so the copies there stop working within hours), and disconnects and removed grants made on this release do not carry back. Rolling back means downgrading this release's migrations, which drops the service connection, grant and token record tables and everything made in them on this release. Tokens already handed out then stay valid until they expire, within the hour, as nothing is left to revoke them.

Removed

  • The feature_flags table is dropped. The migration logs which flags were on at upgrade so they can be re-expressed through the feature-flag setting.
  • A new migration adds service_oauth_clients, where the server keeps the OAuth clients it registered with vendors such as Atlassian, encrypted with the server keys and re-encrypted when they rotate. Downgrading it drops the table; the server registers again on the next connect.

Deprecated

  • HOSTED_GITHUB_CONFIG_PATH still configures the GitHub App's sign-in, with a warning, for this release. It names no signing key, so GitHub can then be connected but not granted.

Security

  • Multi-tenant security hardening across the server and gateway (#641).

Fixed

  • A silent Mattermost no longer freezes all of switch-core (#660).
  • Teams no longer adds users to the whole team when a private channel is mistaken for public (#591).
  • A client that fails on the database at start is retried instead of being dropped (#742).
  • The clock no longer redraws turns for offline agents (#715).
  • The Helm chart's Ingress now routes every agent API path to switch-core. /agent-sessions (session starts, turn activity, approval requests and their answers), /version and /deeplink were missing from ingress.agentApiPaths and from the sample Ingress, and so is /v1, where agents controllers enroll and connect. Behind a path-routed ingress they reached the gateway, which answered with its web page. Agent sessions then failed mid-turn with Unexpected token '<', and activity and approval cards never reached messaging platforms. An install that overrides ingress.agentApiPaths, or routes with its own Ingress, must add these paths itself. A test now fails when the agent bridge serves a path the chart does not route.

[0.29.0] - 2026-09-29

Added

  • Agents can find, run and save workspace templates (#554, CHOO-2923). New agent tools — get_template_guide, list_templates, get_template, run_template, save_template, update_template, delete_template — let an agent discover the shared and owned templates available to it, read a template's document and its params, create the rooms a template describes by filling each slot with an existing agent, and save, update or delete its own templates (private or shared). Backed by template runs, agent-run tracking, and refusal reasons that explain why an action was declined.

[0.28.1] - 2026-09-26

Added

  • Observability: a peak db.pool.in_use gauge and a counter for database connection-pool checkout timeouts (#560).

[0.28.0] - 2026-09-25

Added

  • Sessions live with the host; switch-core only relays (#543). Session ownership moves to Console and its sidecar — switch-core relays the agent connection rather than owning session lifecycle. The agent bridge API is reworked accordingly (activity routes replace the old session routes, with the operation and command surfaces updated to match).
  • Publish, edit, share and search room templates (#501). Server-side template visibility, a room-YAML export endpoint, and parameterized export (values replaced with {key} placeholders), backing the Console's capture-as-template flow and the gateway template management UI.
  • Bound workspace creation per person (#458, CHOO-2722). GATEWAY_MAX_WORKSPACES_PER_USER (default 3) caps self-service workspace creation; the check, creation and count run under a NOWAIT lock on the caller so concurrent requests cannot slip past, and the bound counts workspaces created (users.workspaces_created) so handing one over does not free the allowance. 0 closes self-service; deployment operators are exempt.
  • Richer connector-event telemetry, plus DB and bridge timing (#488). Connector events gain a failure reason and a duration_ms, and server-side DB/bridge lifecycle timing is recorded.
  • Multi-tenancy phase 1 — tenant isolation enforced in the database. Every request and every unit of background work now binds a tenant onto its session, the schema is scoped to a tenant, and row-level security enforces the boundary under a restricted database role whose only exemption is the tenant-resolution lookups. Tenant admin is split from the deployment-operator capability, and a session with several memberships can select which tenant it acts as.
  • Distributed messaging-install protocol. A collaboration app can be installed into a workspace and later ended through a first-class lifecycle (messaging_installs, install/end routes, and the tenant-routing lookup), inbound platform events are routed to the installing tenant and handled exactly once, and MESSAGING_PUBLIC_URL is configurable separately from the gateway's public origin.
  • Workspaces, invitations and members. Gateway routes to manage a workspace, its invitations and its members, backed by an invitations table and the lookup that resolves an invitation to a tenant.
  • Datadog observability (#491). OpenTelemetry metrics, a liveness/readiness split (only the database gates readiness), log export, and a shipped Datadog dashboard and monitor set. Tracing is not yet wired.
  • Product-usage telemetry for switch-core (#487).
  • Template registry (#421) — server-side storage and API for room templates.
  • Index on rooms.group_id (#404).

Fixed

  • Removing a bridge or agent now deletes its clients and their room memberships in one transaction, so client rows are no longer orphaned when a client had ever joined a room (#525, CHOO-1492).
  • Removing an agent from a room no longer stops message delivery to the rest of the room (#412).
  • A failed bridge read is no longer mistaken for an empty bridge list (#515).
  • The gateway refuses to delete a bridge an install created, and refuses to remove a member who still owns agents in the workspace.
  • The gateway requires GATEWAY_OIDC_SCOPES to include openid and no longer 500s on a userinfo hiccup (#418).
  • Migration lock waits are bounded, and the deploy job runs through boot's path.
  • A log context token is never restored in a context that did not create it (#428).
  • Dependency: pyjwt 2.12.1 → 2.13.0 (#508).

Security

  • Room message bodies are delivered to an agent inside BEGIN/END SWITCH MESSAGE markers carrying a per-message nonce. Any room participant writes the body, and it previously sat as plain text directly under the [Switch] … addressed you … header, so it could spell out a convincing header of its own — or its own trailing read_context notice — and be read as something Switch said. The nonce cannot be predicted, so the frame cannot be closed early or reopened; sender display names are collapsed to a single line for the same reason (#527).

[0.26.0] - 2026-09-09

Added

  • Multi-tenancy groundwork (phase 0). Tenant scoping across the backend, and the tenant id plus request context are now stamped on every log line.
  • Typed params block in rooms YAML with {var} interpolation, so room definitions can be parameterised (#402).
  • An agent can read any room it is a member of without connecting to it (CHOO-2628) — a new agent operation. (Reaches agents once the runtime and connector plugins that expose it are released; the server side ships here.)
  • Users can change their own password in the gateway (#323).

Fixed

  • OIDC account linking hardened. A verified OIDC identity is linked to its account by email; multi-issuer and null-issuer collisions are handled rather than guessed, duplicate emails no longer resolve silently, legacy identities are backfilled and matched case-insensitively, and the race is retried with savepoints (CHOO-2624).
  • Agent bootstrap-key and registration hardening — resilient bootstrap-key seeding across admin changes and legacy migrations, retired key hashes are revoked and keys retired rather than deleted, a config fault is no longer misreported as a bad token, and the deployment-wide registration token is scoped to a non-admin owner (CHOO-2624).
  • A Postgres server that fails over without closing the connection is now noticed by the notify listener (CHOO-2622).
  • The setup job stays authenticated to the gateway (CHOO-2622); Helm applies the secret and CA bundle before the migration that needs them.

[0.25.0] - 2026-09-06

Added

  • GATEWAY_OIDC_REQUIRE_EMAIL_VERIFIED (default true) makes the OIDC email_verified check configurable, so a deployment whose IdP addresses are authoritative (e.g. Okta directory-provisioned users that never report the claim as true) can opt out. The relaxed check logs a warning on every login naming the claim value, so it is never silent (#387).

Removed

  • Matrix is gone. Following the 0.24.0 move to a Postgres message store, switch-core no longer runs or talks to a Matrix homeserver: the Tuwunel service, the matrix-nio dependency, and all MATRIX_* configuration are removed, and the Helm chart drops the tuwunel deployment, service and PVC. The standalone compose no longer ships a Tuwunel service. Ships a migration (drop_matrix_columns) that drops the now-unused Matrix columns (#380).

Changed

  • stack-compose contract → speaks 2 (accepts 1). The published standalone compose no longer carries the tuwunel service or the MATRIX_* environment variables; Switch Console's local-server mode drives the new shape (#380).

[0.24.1] - 2026-09-06

Changed

  • The one-time history backfill now carries a bounded amount of each room's history — 50 messages per room, newest first, instead of walking every room to its start (an unbounded migration that could never finish on a busy, long-lived channel and ran at boot with switch-core waiting on it). Override with --messages-per-room (0 restores the old full walk). A capped room is marked backfilled and reported as capped, not incomplete (#379).

[0.24.0] - 2026-09-06

Security

  • Mattermost connections verify TLS by default; disabling it is now an explicit opt-in (#294).
  • OIDC logins are bound to the identity provider's (iss, sub) pair rather than email, so a changed or reused email address can no longer take over an account (#291).
  • Room roster changes are gated on room membership, not on a resource's write-visibility (#292).
  • Room-scoped resource endpoints now require access to the room (#293).
  • Agent registration and re-registration are confined to the owner's tenant, and re-registration cannot overwrite an agent owned by someone else (#290).

Changed

  • Internal message transport moved to Postgres LISTEN/NOTIFY. The message/notify path and event cursors are served from Postgres instead of Matrix, which is kept alongside for now. Ships a database migration (message_notify_and_cursors) (#363).

Added

  • Agents render under their human display name on every collaboration bridge, defused so an agent still never sees its own echo (#330).

Fixed

  • !command now resolves @mentions and runs code-wrapped commands, matching /command behaviour — both had broken once agents gained autocomplete user groups (#313).

[0.23.0] - 2026-09-03

Performance

  • Matrix I/O no longer runs inside a database transaction. add_agents_to_room, remove_agents_from_room and delete_room each held a connection-pool slot open across a fan-out of Matrix invites/kicks (up to 10s per call), a major contributor to pool exhaustion under load. The three now read short, commit, then do the Matrix work holding nothing — ordered so Matrix membership can never exceed what the database records (a crash leaves an under-privileged, repairable agent, never one silently reading a room Switch has no record of) (#354).
  • Per-client sync-cursor writes are staggered. With one Matrix client per participant (hundreds in a single process) and the throttle interval equal to the sync long-poll, every client persisted its cursor in the same instant on an idle instance — a periodic burst of concurrent one-row transactions that starved the pool and cost heartbeats their connection. A sync that carried nothing durable no longer earns a write, and the writes that remain no longer align (#357).

Changed

  • On the collaboration bridges, a turn that ends on an error now renders as the stall it is — the reason is shown as why the turn stopped, not streamed as an in-progress step under a spinner (#356).

[0.22.1] - 2026-09-02

Performance

  • Agent Bridge no longer holds a database connection while doing other work. Under load the connection pool could be exhausted by slots checked out and left idle in transaction, which surfaced as QueuePool limit ... connection timed out and, because heartbeats share the same pool, as fleet-wide agent loss. Three fixes: a request never takes a second pool checkout while holding one (targeted messages to owner-only agents were the common trigger); the message path collapses six checkouts to one on an internal room and two on a bridged one; and authentication is taken off the heartbeat path via a single outer join plus a narrow in-process cache of resolved credentials (#351).

Added

  • New operator settings for the above: AGENT_AUTH_CACHE_TTL_SECONDS and AGENT_AUTH_CACHE_MAX_ENTRIES tune the credential cache (set the TTL to 0 to disable it), and the opt-in DB_IDLE_IN_TRANSACTION_SESSION_TIMEOUT (default off) turns a leaked idle-in-transaction slot into a loud error. All three are exposed as Helm chart values (#351).

Fixed

  • Rejoin the two Alembic migration heads left by the parallel 0.22.0 merges, so alembic upgrade head is unambiguous again — deploys from main were failing on the two heads (#351).

[0.22.0] - 2026-09-02

Added

  • Agents can carry an optional human display name, stored on the agent and surfaced over the agent protocol (agent detail) and the gateway (#326).
  • User-defined external reference types. Reference types are no longer a closed set of four. Any signed-in user can define one — slug, display name, agent-facing instructions, value hint and a read/write visibility pair, exactly like a Reference — from the new Reference types tab under Resources, or over POST /references/reference-types. The four built-ins (Google Drive, Confluence, GitHub, Jira) stay in code, cannot be edited or deleted, and win any slug collision. Agents see the types their owner can read, via list_reference_types; each entry now also carries value_hint and origin (builtin or user). A type's visibility gates who may pick it, never delivery of its instructions for a reference already attached to a room — the attachment is the grant.
  • list_all_references lets an agent discover references across the whole instance. Until now an agent only saw the references attached to its current room. The result is scoped to what the agent's owner may read. Three optional filters narrow it — name_contains, type and owner_name — and they are ANDed. Each row carries the reference's metadata but never its value. When the caller is connected to exactly one room, each row also reports attached_to_current_room. An agent can therefore find a reference and attach it in the same turn.

Changed

  • A reference value must now carry at least one URL. An empty urls list was already rejected, but an absent urls key silently validated to an empty list and was stored. It is now a 400. Existing rows still read fine; the failure appears the first time someone saves one — including a no-op save from the reference detail page. Find them before upgrading and give each a real URL:

    SELECT id, type FROM "references"
    WHERE value->'urls' = '[]'::jsonb OR NOT (value ? 'urls');
    
  • GET /references/reference-types now requires a session and returns 401 without one. It was the last unauthenticated route on the references router; the set it returns is per-caller now, so it cannot be answered anonymously.

Fixed

  • Discord: create the room on the first command, not only on the first message (#341).
  • Deeplinks: replaced the handoff page with a branded two-state design (#339).

Performance

  • Agent bridge: take the sequential scan off the heartbeat path and size the DB pool for it, backed by a new index on agents.api_key_id (#344).

[0.21.1] - 2026-09-01

Fixed

  • Slack: normalize typed command mentions so commands addressed with a typed @agent are recognized (#316).
  • Slack: resolve agent mentions that crossed a workspace boundary (#327).

[0.21.0] - 2026-08-25

Added

  • The gateway's registration keys page can now search agent keys, and its table fills the page height rather than being capped short (#177).
  • Switch's commands appear in Teams' / autocomplete picker. The shipped manifest moves to schema v1.29 and declares supportsTargetedMessages plus a second command list with triggers: ["slash"]. The two surfaces disagree about the slash — the picker prints it and inserts the bare name, while the mention menu inserts a title verbatim — so the lists carry the same commands spelled for their own surface. A command picked from the picker is sent privately to Switch and needs no @-mention.

Fixed

  • A Slack turn triggered by another app's post — a workflow, a scheduled summary — is handled as an ordinary request: bot_id was read as "nobody asked", so the 👀 went on the thread root (which in such a channel need not be a message) and Slack answered message_not_found. The message that asked is now remembered whoever sent it; a mark Slack cannot place is recorded as unmarkable rather than retried every progress report; and the warning names the message, not just the channel. A workflow-triggered turn still gets the status message and mark but no card, since Slack will not open a session addressed to an app (#288).
  • A Teams message that merely began with a slash is no longer swallowed as a command. Pasting /Users/ada/notes.md came back "unknown command" instead of reaching an agent. / is not Switch's prefix — it opens paths, dates and fractions — so an unknown /name is now read as text. ! is Switch's own, so !list-agent still answers that it is a typo.
  • An agent no longer answers Switch's own notices. A system message — a command's result, "Added X to this room", the guidance shown when someone tags the app itself — is output, not a request for a reply, and the marker saying so was honoured by the admin client but not by agents. Two ways in: a direct room addresses its agent with every message, so running /list-agents in a 1:1 chat had the agent start a session to respond to its own roster; and any notice that lists the agents present writes each @name, which tagged all of them. Such a message now arrives as ordinary unaddressed context on every platform.

Added

  • just teams-app-package builds the Teams app package an operator uploads. Every route into a tenant wants a .zip — the Developer Portal's Import app, Upload a custom app, the admin centre — and none of them takes a bare manifest.json, so the guide's "now zip these three files" was a step people had to get right by hand. It also writes the app id into the three places it has to match, checks the limits Teams enforces without explaining them, and exits non-zero naming any placeholder left in, so a null app id cannot ship quietly.

Changed

  • The shipped Teams app manifest declares supportsChannelFeatures: tier1. Teams requires it of any manifest at schema v1.25 or later whose bot takes the team scope and refuses the upload without it — a validator rule, not a schema one, so the package validated cleanly against the published schema and was still rejected. The package builder now checks the same rule.
  • The shipped Teams app manifest moves to schema v1.27. It was written at v1.19 on the reasoning that older schemas upload most reliably; the tenant we actually run on is on v1.27, which settles that better than the reasoning did. Nothing is removed between the two and the required fields are identical, so it is a version bump rather than a rewrite — and v1.25 raised a command list's cap from ten entries to twelve, so the menu gains /list-switch-agents and /room-url.

Fixed

  • A Teams posts channel no longer fills with "This message has been deleted." Teams substitutes that for any deleted message and keeps it in the post, so an agent's status line — posted, moved as the conversation went on, and removed at the end — left one behind per move and per turn. Talking to an agent produced a column of them. The status is now posted once, stays where it is, and is edited to ✓ Done · <elapsed> when the turn ends, as it already was on Mattermost for the same reason. Threads-layout channels delete cleanly and are unchanged.
  • An agent addressing a Teams user by name produces a real mention again, rather than flat text that notifies nobody. Teams offers a display name on an inbound activity, and that was taken as the person's handle — so someone was filed under Louis Amaudruz, space and all, while the matcher that turns @name back into a mention stops at the first space. It looked for Louis, the directory held louis amaudruz, and the two could never meet: anyone whose handle contained a space was permanently unmentionable. A display name is now traded for the one-word principal name the directory holds, and the matcher recognises the names it knows however many words they run to, so people already stored the old way work without being renamed.
  • Concurrent writes to a bridge's learned connection_config no longer lose each other. The Teams bridge records a channel's team the first time it sees that channel, each from its own session, so a burst of new channels was a burst of read-modify-writes on one JSONB value: measured, eight concurrent writes landed two. A dropped entry is exactly the state that makes Graph refuse that channel's subscription after the next restart. The row is now locked for the update, which also stops the serviceUrl writer clobbering it.
  • The Graph notification endpoint compares its clientState in constant time. It is the only control that authenticates a notification's origin, and a plain != on a secret leaks its prefix through timing.
  • Repairing a placeholder username costs one lookup per person rather than one per message. It runs on every inbound message on every bridge, and the common case — a stored name that is already fine — was still opening a session and querying.
  • A Graph read of a Teams channel that fails is retried later instead of being believed forever. The name and layout of a channel are read together and cached, and a failure was cached with them — so one blip, or a permission granted a minute after the bridge started, left that channel nameless and threaded as a posts channel for the life of the process. Learning a channel's team now also discards anything read against the previous one, since a read made against the wrong team is exactly the read Graph refuses.
  • The rejection an inbound Teams activity gets when it is addressed to the wrong app id names the audience the token carries — and now reads that back out of a signature-verified decode rather than an unverified one. The signature had in fact already been checked by the time this ran, so nothing was exploitable, but quoting an unverified claim in the message explaining why a caller was rejected is a habit worth not having.
  • Teams channel capture no longer dies at a restart for any channel outside the connection's configured team. A Graph subscription names the team as well as the channel, and the team is only ever carried on an inbound activity — held in memory it was lost on restart, and those channels fell back to the configured team, where Graph answers "Channel is not present in the team". Nothing recovered it: refilling the map needs an activity from the channel, and without capture Teams delivers only messages mentioning the bot, so addressing an agent — the normal way to use the room — never arrived. The mapping is now persisted the way the Bot Connector endpoint already was.

Added

  • Teams accepts /command as well as !command, and the setup guide carries the app-manifest snippet that puts the commands above the compose box. Teams has no server-registered slash commands: a manifest command list types the text and the bot parses it, so the two prefixes are the same thing by the time Switch sees them — but people arriving from Slack try / first, and until now got nothing.

  • The Helm chart can publish and route the Microsoft Teams bridge listener. The Teams adapter serves Bot Framework activities and Graph change notifications from its own HTTP server on port 3978, separate from the API on 8000, and nothing in the chart exposed it — so a Teams bridge could create channels and post while receiving nothing back, looking healthy throughout. Set switchCore.teamsBridge.enabled=true to publish the container and Service port, and switchCore.teamsBridge.ingress.mode to say how the two callback paths are routed: dedicated renders a second Ingress carrying only those paths on their own host and certificate, so Microsoft can reach them while the gateway stays internal; shared adds them to the chart's managed Ingress; external leaves the routing to you (see samples/ingress.example.yaml). Off by default, because bridges are created at runtime and the chart cannot detect one.

  • The Teams Ingress supports sitting behind a CDN or reverse proxy: leave its host empty and TLS off, and it renders a host-less rule that answers whatever Host arrives. That is the shape needed when the public name belongs to something upstream, and it is the only option that needs no domain of your own — the setup doc carries the CloudFront recipe, including the two settings (forward query strings, do not cache) without which Graph's subscription handshake fails. Empty host with TLS enabled is still rejected, since a host-less rule cannot carry a certificate.

  • The chart now refuses to render a Teams bridge that nothing can reach — enabled with no routing mode, shared without a managed Ingress, or either managed mode without a hostname. Publishing the port without a route is the silent half-failure the rest of this work exists to prevent, so it is a render-time error rather than something found days later.

  • helm test checks the Teams listener, running Graph's own validation handshake against the Service. It proves the listener is bound and the port published; it runs in-cluster, so it deliberately does not claim anything about public reachability, and says so.

  • Post-install notes print the exact public_base_url to paste into the bridge, and warn when its Ingress has TLS disabled — Graph only calls HTTPS it trusts.

  • The Helm chart has a README.md, covering which of Switch's three network surfaces have to be reachable from where. Only the Teams listener needs the public internet; every other bridge connects outbound.

  • Registering a bridge that would contend with an existing one for a host resource is refused, naming what clashed. Teams is the case that has one: its inbound listener needs a TCP port, and a second bridge on the same port failed to bind inside a background task, was dropped from the running set, and never retried — surfacing much later as Bridge not running: <id> with no stated cause. Adapters declare this by implementing exclusive_resource; those that only dial out declare nothing and are unaffected. The same check runs at startup, so rows predating it get a clear error instead of a bind failure.

  • Bridge credentials are verified before a bridge is saved. Adapters start in a background task whose exceptions are logged and swallowed, so credentials that the platform rejects used to be stored, reported as success, and surface hours later as an unrelated-looking failure. Registering a Teams bridge now asks Azure for both tokens it will need first, and a refusal comes back as a 400 carrying Microsoft's own explanation — which names the mistake outright, including the client secret value vs secret ID mix-up. Adapters opt in by implementing verify_credentials; the default still accepts anything, because an adapter that cannot check cheaply should not pretend to.

Changed

  • A Teams bridge is five fields instead of nine. The clientState shared secret and the Graph encryption certificate, public key and private key are generated when the bridge is created rather than asked for: they are values the operator invents rather than gets from Microsoft, and asking meant an invented secret, an openssl invocation, and three PEM fields whose only symptom when pasted wrong is channel capture that silently never decrypts. Supplying your own through the API still wins, all three or none. Existing bridges are untouched, including any left without encryption material.
  • The Teams Graph private key is no longer typed into a plaintext form field. The gateway decides which inputs to mask by matching the field name against token|password|secret|api_key, which encryption_private_key does not match; generating it removes the field from the form entirely.
  • docs/bridges/TEAMS_SETUP.md rewritten. It now walks through Azure setup value by value and says where each one comes from, documents the deployment and public-ingress requirements (previously absent), and adds verification and troubleshooting sections keyed to the errors these failures actually produce — starting with the client secret value vs secret ID mix-up, which surfaces only as an opaque AADSTS7000215 at the first Graph call.

Fixed

  • Following an "Open in Switch Console" link no longer strands the browser tab. The gateway answered the click with a bare 302 to the switchdash:// deeplink; the browser hands that to the desktop app without loading a page, so the tab kept whatever it last rendered — on Teams, Defender's Safe Links interstitial, still saying "Verifying link . . ." after Switch Console had opened, which reads as a link that hung. It now serves a page that opens the app, tries to close itself, and — since browsers only let a script close a window a script opened — otherwise says the handover worked and the tab can be closed.
  • A Graph permission granted while the bridge is running takes effect on the next call instead of within the hour. An app's roles are stamped into its access token when it is issued, so consent granted afterwards is invisible for the token's lifetime — Graph keeps answering Authorization_RequestDenied on a permission the operator can see plainly granted in Azure, which sends them back to look at it again. A 401 or 403 from Graph now discards the cached token and retries once with a fresh one. Once, and only when the token was old enough to have missed the grant, so a genuine denial costs one extra round trip rather than looping.
  • A person recorded under a Teams id gets their name back on their next message. Switch files someone under the name it is first given, and Teams often gives none — a 1:1 activity carries no sender name — so its own id went in and then read as that person's name in room titles, on their Matrix account and in every agent reply addressing them. Resolving names properly helped only people met afterwards; this repairs the records already written, without a migration. One-way by design: an id becomes a name, never the reverse, and one name never replaces another.
  • A Teams channel with no name in its activity is looked up in Graph rather than falling back to its id, so an auto-created room is called something a person recognises instead of 19:7641f9de326b4…. Cached, including the "asked and could not say" case.
  • An agent's first message in a Teams channel opens a post and everything it says next joins that post. It used to open a fresh one each time, so an agent introducing itself produced a column of one-line posts rather than a conversation.
  • An agent answering in a Teams channel replies inside the post it was asked in. A channel is a list of posts rather than a stream, so posting at the root opens a new conversation — the question sat in one post and the answer appeared as a fresh one below it. A reply that names no thread now goes to the post the bridge last saw that channel speak in. Per channel, so two conversations at once in the same channel can still cross; Teams offers no better signal on an untied reply, and the wrong post beats a new one every time. Chats and group chats are unaffected.
  • Teams messages are rendered for Teams. Shared bridge code emitted Slack's <@id> mention form on every platform, so a Teams user was told they had tagged <@28:11111111-…>; an @name was inert text, because Teams needs <at> markup and a matching entity and the adapter emitted neither, so an agent could never notify its owner; and single line breaks vanished, because an Adaptive Card text block follows Markdown in treating one newline as whitespace. Outbound bodies were also being rendered twice — callers of send_message render first by contract, and Slack and Teams rendered again, which Slack's real conversion has been quietly suffering.
  • A Teams sender with no name attached is looked up rather than filed under their id. Teams omits the sender's name from some activities — 1:1 chats especially — and the raw 29:… id became that person's name in the room title, on their Matrix account and in every reply addressing them. The directory and the message paths now also agree on one handle (the local part of the principal name) instead of storing a principal name from one and a display name from the other. Existing records keep the name they were created with.
  • Directory search matches on user principal name, so confirming a claim finds the person who was just picked from the list. It searched display name and mail only, then re-searched by principal name to confirm — a 404 for anyone whose mail differs from their UPN or who has none. User.ReadBasic.All is also now documented; without it the search fails and nobody can link an account at all.
  • The "Open in Switch Console" link no longer renders as empty brackets on Teams. Teams strips a link whose scheme is not http(s) — label included — and the console deeplink is switchdash://. Teams now declares that, so the URL is rewritten to the gateway redirect as it already was for Discord and Telegram, and a deployment missing GATEWAY_PUBLIC_URL is told at startup.
  • Adding a room to a bridge no longer answers 500 when the platform refuses to create the channel. Graph says exactly what is wrong — which permission is missing, which value it will not take — and that sentence went to the log while the operator got an unhandled exception. Platform refusals are now BridgeOperationError, the counterpart to BridgeCredentialError for everything after the credentials are accepted, and room creation returns a 502 quoting the platform. Registering a bridge only proves it will issue a token, so this class of failure always surfaces after setup has reported success. Only the Teams adapter raises it so far.
  • A Teams channel subscription that fails is retried instead of leaving capture dead until someone restarts the process. Graph validates a notification URL by calling it, and the bridge asks for its subscriptions seconds after binding its port — behind a load balancer that is precisely when nothing answers yet, so the first attempt after every restart was refused and the restart that broke capture was also the only thing that could fix it. The retry backs off to a few minutes and says so when capture recovers; repeats of an unchanged failure are quiet, but a change of reason is not.
  • Binding a room to a collaboration channel that already exists now establishes message capture for it. Capture was only ever set up as a side effect of creating a channel, so a room pointed at an existing one — on creation, on linking a bridge to an internal room, or on moving a room between bridges — heard nothing but the messages that @mentioned it. Invisible on bridges whose capture is one bridge-wide stream, and on Teams it survived only until the next restart happened to reconcile it.
  • A Teams activity rejected for the wrong audience says which audience it carried and which one the bridge expects. The rejection is almost always the Azure Bot resource and the registered app registration having different app ids, and the message it replaced — Audience doesn't match — named neither, so the one fact needed to fix it was the one fact not logged.

[0.20.0] - 2026-08-24

Added

  • Each agent gets a mentionable Discord role, so its @name autocompletes in the composer and arrives as a real pill rather than plain text that only looks like one. The roles are minted empty and permissionless — mentioning one notifies nobody — and an inbound <@&…> resolves back to the agent's name before the addressing layer sees it. A Discord role carries no metadata, so an agent's role is the one named exactly after it: a role made by hand is adopted rather than duplicated, and one that has members is never deleted. Off with agent_roles: false, and it turns itself off, saying why, on a server missing Manage Roles or at Discord's 250-role cap (CHOO-2316).
  • Discord puts 👀 on the message an agent is answering for as long as its turn lasts. The posted status says an agent is busy; the reaction is what says which message it is busy with, and an agent answering two people at once marks both and clears both together. Needs the bot's Add Reactions permission; without it the bridge warns and posts no reaction rather than faking one (CHOO-2316).

Fixed

  • A Discord command argument naming someone picked from the composer's @ menu is resolved to their name instead of being refused. Both the slash and typed ! forms branch off before the translation an ordinary message gets, so an agent or person chosen from the menu arrived as raw <@&…> / <@…> markup and was rejected as "not a single name" — quoting a string the invoker never typed. Newly reachable because an agent's name now autocompletes, but the typed form was wrong before that too (CHOO-2316).
  • Disconnecting a messaging app no longer fails with a 500 when a room member has already left. Deleting a bridge kicks every client of every room it had, and one that had already left answered a 403; "already not a member" is now treated as done, while every other failure — permission denials included — still raises (CHOO-2344).
  • An addressed auto-session agent that is offline now asks its owner to open Switch Console — the person who can actually bring it online — instead of telling the whole room to run a terminal command. The reply threads off the triggering message, names the asker as the reason (dropped when they are the owner), and the connect command names the agent it was generated for, so a directory holding several agents answers select_agent on its own (CHOO-2344).

[0.19.0] - 2026-08-22

Added

  • Telegram marks the message an agent is working on with 👀, and clears it when the turn ends — in groups, channels and 1:1 chats alike, with no administrator rights needed. Outside forum topics Telegram has no threads, so the mark goes on the last thing a person said, which is what the agent is answering. A chat with reactions switched off loses the mark, not the turn (CHOO-2314).
  • Telegram setup docs now spell out that BotFather shows the bot token once, and what to do if it was not saved (CHOO-2314).

Known limitations

  • Telegram cannot autocomplete an agent's name. Slack's per-agent user groups have no Telegram equivalent: @ autocomplete offers only real chat members, there is no user-group concept, and every agent posts through one bot. Real per-agent autocomplete would need a Telegram bot account per agent (CHOO-2314).
  • Mattermost marks the message an agent is working on with 👀, added when the turn opens and removed when it ends. Inside a thread the mark goes on the reply that asked rather than the root it hangs off, and an agent handling two messages at once clears both together instead of leaving the first marked for good. The reaction is added by the agent's own bot, so two agents on one message read as two marks naming who is on it. Documented alongside it: Mattermost has no equivalent of Slack's native progress card, and Switch does not approximate one (CHOO-2317).

Changed

  • The gateway's Docs button opens the published documentation at docs.flintai.dev rather than the GitHub repository (CHOO-2313).

Fixed

  • Slack's native agent-session card could vanish from a workspace and never come back until the pod was replaced. A burst of Slack rate limits was being counted as the app being unfit to host sessions and tripped the permanent give-up after three in a row; throttling now makes the bridge stand down for the Retry-After Slack sent rather than counting against it, and a give-up over an unrecognised error expires after ten minutes instead of lasting the life of the process (#276).
  • A stale or hand-deleted Slack agent card no longer turns cards off for the whole workspace. A status update arriving just after a card is torn down — or a user deleting the card by hand — wrote to a message that was gone, and three in a row tripped the same permanent give-up. Such a card is now forgotten: the turn falls back to the posted status message and the next turn opens a fresh card. A step that failed to land is also no longer recorded as the card's current step (#283).

[0.18.0] - 2026-08-21

Added

  • Each agent gets a Slack user group, so its @name autocompletes in the composer and arrives as a structured mention rather than plain text that only looks like one. A single app still fronts every agent; the groups are minted empty (mentioning one notifies nobody) and exist only to be mentionable. Inbound, a <!subteam^…> mention resolves back to the agent's name before the addressing layer sees it — including a name folded into a legal handle — and outbound mentions render as the same pill. Slack owns the objects, so the id↔name maps are rebuilt by reading them back, and only groups carrying Switch's marker are adopted (CHOO-2277).
  • An agent's turn is mirrored onto Slack's native agent-session status — the streamed "in progress / complete" indicator Slack renders for an app — pinned to the message being worked on. It needs the app's assistant:write scope; without it the bridge says so rather than failing silently (CHOO-2277).

[0.17.3] - 2026-08-19

Fixed

  • The Mattermost bridge now bridges the channels a new bot is already a member of. A bot's websocket opens only after Mattermost has added it to the team's default channels, so it never witnessed its own joins — the first agent on an instance bridged nothing, and creating a second adopted Town Square and Off-Topic late and attributed to the wrong action. Current membership is now reconciled on connect rather than relying on live join broadcasts alone (CHOO-2203).

[0.17.2] - 2026-08-18

Fixed

  • The collaboration-bridge runtime status is delivered where the reply is. When an agent was addressed at channel level its reply opened a thread on the triggering message while the status line stayed at the channel root, so the two landed in different places. Adapters can now opt into anchoring the status on the triggering message; Mattermost takes it — where a thread is a side panel, a status left at the root is the one nobody opens. The typing indicator is reported against the trigger's actual location (channel vs thread) rather than the status's, so it shows where the person who addressed the agent is watching (CHOO-2173).

[0.17.1] - 2026-08-16

Fixed

  • Agent avatars no longer show a bright square on Slack's dark message list. A generated avatar is a transparent DiceBear PNG and Slack composites it onto white; newly generated icon URLs now name Slack's resting dark surface as the background instead. Only newly generated URLs change — an icon already stored keeps its URL (#239).

[0.17.0] - 2026-08-16

Added

  • Agents can carry a custom icon. Switch stores a per-agent icon_url — a link, never image bytes — accepted at registration and set, changed or cleared through a dedicated set_agent_icon operation and PUT /agents/{id}/icon; it is reported on the agent summary and detail. The Mattermost bridge fetches the icon and re-uploads it as the bot avatar. Because Switch itself dereferences the URL, it is validated to an absolute https address with no embedded credentials and refused when it names a local, private, loopback or link-local host (CHOO-2171).

Fixed

  • An agent joining a room greets it again. The greeting fires when an agent notices its own arrival, and since 0.9.0 it never did: auto-accepting the invite began recording the room as joined a moment before the join arrived over sync, so the arrival was read as one already handled. Every room an agent is invited to went ungreeted, on every connection, regardless of the per-connection greeting toggle. Whether the arrival has been announced is now tracked separately from membership, which the bridge needs recorded as early as possible. Departing a room resets it, so being added back greets again.
  • just test-integration runs again. Its wiring had drifted behind RoomService, ProtocolService and AgentClient, so every test errored during setup — unnoticed, because the suite is excluded from the default run and from CI. It is the suite that would have caught the greeting regression.
  • The bundled-stack setup image links the seeded Mattermost admin to its owner, so a freshly seeded managed deployment comes up already knowing which chat account the owner is — owner-only addressing then works without the owner linking an account by hand first (CHOO-2172).

[0.16.0] - 2026-08-15

Added

  • A Switch user can claim their messaging-app account as their own, linking a Slack / Mattermost / Teams / Discord identity to their Switch login (external_user_claims). Claiming searches the platform's own user directory, so someone can be recognised before they have ever posted — previously Switch only learned of a person when they first spoke, which left a freshly connected workspace with nobody to pick. Each platform is searched in its own way: Slack filters a paged listing, Mattermost and Discord have server-side search, and Teams queries the directory through Graph. Telegram has no directory a bot may search at all, so there the list falls back to the accounts Switch has already seen and says why it is narrower — a smaller answer rather than a refusal, since refusing would leave owner-only addressing unusable on the platform for people Switch already knows. Claims are not exclusive: several Switch users may claim the same account, so nobody can keep the real person from being recognised by claiming it first (CHOO-2137).
  • An addressing rule can name the agent's owner, or any agent that owner runs, rather than a list of identities. Both resolve when the message arrives, so they survive connecting a new workspace, recreating a bridge, registering another agent, or the agent changing hands. The second is what lets one person's manager agent keep dispatching their own workers under an owner-scoped policy — that is the owner acting through a program, where someone else's agent is not (CHOO-2137).
  • GET /agents reports each agent's addressing_policy, which until now was only on GET /agents/{id}. So "which of these agents answers only its owner" is one list read rather than a read per agent — the shape a client needs to ask that at all, instead of a request storm (CHOO-2137).

Changed

  • A messaging app declares whether its user directory can be searched (directory_search_supported, beside channel_creation_supported). False for Telegram, where a bot can only name people who have messaged it. Read from the adapter class, so it is answerable before a connection of that type exists — which is when a client has to decide whether asking someone to pick themselves out of a directory is a question worth putting (CHOO-2137).
  • Disconnecting a messaging app now removes the identities Switch made for it — the app's own Matrix client and the puppet behind every person Switch saw on it. They were left behind, and the app's own client is not merely untidy: its Matrix name was derived from the app's type and display name, so disconnecting an app and connecting one named the same failed outright on the leftover (duplicate key value violates unique constraint "clients_matrix_user_id_key"). The name now carries a random tail as well, because deleting the row is necessary but not sufficient: the homeserver has no call for removing an account, so the old Matrix user is still there, and adopting it would mean logging in with a password Switch no longer holds — shared-secret registration reports an existing user as success without applying the new one, which reads as a working connection that can never connect. Abandoned homeserver accounts are logged on removal rather than passed over in silence (CHOO-2137).
  • An agent pings its owner, not a handle typed into its config. The per-agent notify_user option is gone; who to @-mention when an agent needs input is now the agent's owner, resolved through the messaging account that person has claimed on the platform the room is bridged to. A handle only ever means something on one platform — the same person is one name on Slack and another on Telegram — so a single configured string was at best right in one room, and on Discord and Teams it went out as plain text that notified nobody while looking like it had. An agent with no owner, or an owner who has linked no account there, now says so in the nudge instead of posting a line with the mention silently missing: a ping that reaches no one and an agent that never asked look identical otherwise. Existing notify_user values are ignored rather than migrated, and drop out of an agent's options the next time they are written (CHOO-2137).
  • Newly created agents are owner-only by default: only their owner may address them, from any room, and no agent unless the owner says so. The default is applied where every registration path converges, so it holds for Switch Console, the gateway, the agent bridge, the configure skill and bulk subagent registration alike. Existing agents are left open — the default is not applied retroactively, since that would mute every agent whose owner has not yet claimed an identity. Server-side connector agents opt out: they are services a deployment offers everyone, not one person's assistant (CHOO-2137).
  • In-room commands are subject to the addressing policy, not just messages — !reset, !interrupt and !compact drive an agent as surely as a mention does, and previously reached a restricted agent from anyone in the room. A command naming the agent draws the same one-line refusal; a room-wide command is declined quietly rather than producing a refusal from every restricted agent present (CHOO-2137).
  • send_targeted_message reports a target whose policy forbids the sender as not_permitted, rather than as "live" for a message that will never move it. The message is still sent and the target still declines it in the room: a refusal belongs where the request was made, not only in the sender's account of it, and the same request should not succeed or fail depending on which tool carried it — a plain @name was never blocked. delegate_task remains the exception and fails at the sender, because a task is a row somebody is expected to work rather than something a room can decline (CHOO-2137).
  • An agent that refuses a sender because its owner cannot be identified says so, and points at linking the account in Switch Console — rather than giving the owner the generic refusal from their own agent (CHOO-2137).

[0.15.0] - 2026-08-14

Added

  • Telegram collaboration bridge, at parity with Slack, Mattermost and Discord (CHOO-1686). A single bot backs every agent — Telegram has no per-message identity override — so an agent is named at the head of its messages, beside a stable colour derived from that name, which is the closest thing to a per-agent avatar Telegram allows. Inbound arrives by long polling, so no public ingress is needed. Chats cannot be created by a bot, so a Telegram connection declares that up front and rooms are provisioned when the bot is added to a chat. Multi-file messages post as a single album, and outbound Markdown is converted to Telegram's HTML subset rather than its stricter MarkdownV2. See docs/bridges/TELEGRAM_SETUP.md.

  • One-click install links for a bridge, offered on the operator dashboard and built by the adapter (install_links); empty for platforms installed through their own admin UI. Telegram supplies one, for groups: pick the group, confirm, done. It asks for no permissions at all — a bot posts and deletes its own messages in a group as an ordinary member — so adding the bot no longer means promoting it by hand. Setting a Telegram bridge up is now one BotFather setting (Group Privacy off, once per bot, before the bot is added anywhere) and then a link per group. Channels are added from their Administrators screen, and get no link: that would need Telegram's admin= parameter, which not every client implements, and the ones that do not open a chat with the bot instead — a link that works for some people is worse than a documented step that works for all of them. The dashboard's Add to a chat dialog says so: alongside the links, a bridge can supply an install_note naming the chats no link reaches and the route they take instead, so a missing button is never all the operator sees.

  • A bridged chat is told what the bridge can see in it, and told again whenever that changes — so promoting the bot confirms itself and retracts the earlier warning, and a demotion does not pass unmentioned. Visibility is settled per chat rather than inferred from the bot's global privacy setting, which is what made an administrator bot report a fault it did not have. A chat the bridge can only see mentions in — no admin, privacy mode on — is a supported way to run and is disclosed as one: a notice in the chat, and a warning naming each such chat at startup. Where the answer rests on the global setting rather than on admin status it is taken on trust and logged as such: Telegram reads that setting when the bot joins and no API call reports which value a given chat got, so a bot that predates the change is still filtered and only a re-add or a promotion fixes it.

  • Whether a messaging-app connection may create channels is now declared rather than discovered by failing, in two parts: what the platform can do at all (Telegram cannot — the Bot API has no call for it) and what an operator permits this connection to do, set when the app is registered and changeable afterwards. The second can only narrow the first. A deployment can therefore withhold channel creation from Slack or Teams as well, for a bot that holds no such permission or an organisation that would rather channels were made by people. The answer reaches everyone who acts on it: the room forms in the dashboard and Switch Console stop offering the option and say which reason applies, list_bridges carries can_create_channels so an agent can choose a bridge that works, and a refused create is a 400 naming what to do instead.

Fixed

  • Creating a room on a bridge that cannot make channels returned "Internal Server Error". The adapter's explanation — make the chat on the platform, add the app, Switch adopts it — was thrown away at four of the six doors into room creation, because NotImplementedError subclasses RuntimeError and every caller catches only ValueError. It is now a ValueError subclass, so the message survives to the caller as a 4xx. Opening a DM is governed the same way.
  • Telegram admin notices reached the chat as Markdown source, ** and backticks included. They are written in Switch Markdown like every other body but were handed to the API without the conversion, and everything goes out with parse_mode HTML. Two notices were also written with single-asterisk emphasis, which no platform's converter recognises — Slack bolded them by accident and Telegram italicised them.
  • The "Open in Switch Console" link on Telegram no longer disappears, and can no longer take its message with it. It is a switchdash:// URL and Telegram renders only http(s)/tg:: the client dropped the link and kept the label, or the API rejected the whole message. An unsupported scheme is now posted as tap-to-copy text, and the bridge warns at startup when a platform that only renders web links is running without GATEWAY_PUBLIC_URL — which is what makes the link real, in the Telegram app and on Telegram Web alike.
  • A Telegram message Telegram rejects is retried unformatted whatever the reason. The retry only fired when the error contained "parse", so a rejection worded any other way — "unsupported URL protocol" among them — lost the entire message to a single log line.
  • A Telegram command that needs an argument is usable from the / menu. Telegram sends a command the instant it is tapped, with no chance to type one and no way for a bot to declare that one is wanted, so /invite_agent always arrived bare — and answering with its usage line left retyping the whole command by hand as the only way through, which is what the menu was for. The bot now asks for what is missing, with Telegram's own reply prompt so the composer opens ready, and runs the command when you answer. Giving the argument up front skips it; the prompt is one-shot, so talk that continues under it stays ordinary chat.
  • The no-agents notice advertises /invite_agent on Telegram, the only spelling Telegram will register and therefore the only one its command menu offers. It named the hyphenated form, which the client will not autocomplete. All three forms are now listed.
  • The Telegram brand icon renders correctly in Switch Console's dark theme. The plane was a hole punched through the disc rather than a white shape, so it showed the background through it and came out black.
  • Bridge row action icons line up down the column again. A row carries one icon or two depending on whether its platform offers an install link, and the pair was not anchored to the same edge as the single, so the delete buttons sat at different positions on different rows.
  • Telegram links are built from the username the Bot API reports rather than the one in the bridge's config. A configured name that is not the bot's resolves to whichever account does own it, so every link opened a chat with a stranger — and the mismatch was already detected and then ignored, with the wrong value used anyway. The bot's own name wins and the correction is logged.
  • The "add to a Telegram group" link is withheld from a bot BotFather has barred from groups. Telegram answers such a link by opening a chat with the bot, which from the outside is indistinguishable from a link that does nothing; the reason is logged with the setting to change instead.
  • The group install link no longer asks to be an administrator, which made it unusable for the groups most people have. Telegram builds that picker from groups the person already administers, excluding basic groups — promoting a bot in one converts it to a supergroup — so for anyone whose groups are all basic it offered nothing to pick and Telegram opened a chat with the bot instead, reading as a link that did nothing. Admin was only ever bought for the privacy-mode exemption, and turning Group Privacy off in BotFather buys the same thing once per bot rather than once per group. Promotion remains documented as the repair for a single chat that was added before that setting was changed, and the bot asks for it in the chat when it applies.
  • A room follows its Telegram chat when the chat is reissued a new id, which Telegram does silently whenever a group becomes a supergroup. The room was left bound to the old id, so nothing anyone typed reached Switch again while sends kept working — Telegram forwards those — leaving a bridge that looked alive and was deaf. Adapters now report the change (CollaborationAdapter gains set_channel_migration_handler) and the room is re-pointed, or the refusal is logged with the id to re-point it at by hand.

[0.14.0] - 2026-08-14

Added

  • opencode is a registerable known agent type, alongside claude-code and codex. It carries its own option set — auto-session, working directory and notify-user — rather than borrowing another type's, and the start-session command it shows in a room passes the prompt as a flag, since OpenCode reads its first positional argument as a directory to open.

Fixed

  • The OpenCode server-side connector no longer hangs when the OpenCode server raises a tool permission request. The connector's event loop ignored permission.updated, so no reply was ever sent, the session never went idle, and the response stream blocked forever. Permission requests are now answered automatically — this connector reports tool calls after the fact and performs no pre-invocation mediation, so there is nothing for a prompt to gate.
  • The Mattermost bridge no longer litters a channel with "(message deleted)" placeholders while an agent works. Mattermost's web client shows that placeholder for any message removed while it is on screen — a permanent delete looks no different to it, and no server setting turns it off — so the agent's status line is now retired by editing it into a "✓ Done · 2m14s" marker rather than being deleted, and it no longer moves down the channel to follow the conversation (each move was a delete of its own). The marker carries nothing else: it stays in the channel permanently, so it is kept to the fact that the turn finished and how long it took.

[0.13.2] - 2026-08-12

Fixed

  • read_context's own documentation no longer tells agents to page by passing oldest_timestamp back as before. The first is epoch milliseconds and the second is parsed as ISO-8601, so following it raised instead of paging.

[0.13.1] - 2026-08-11

Fixed

  • The first message bridged from a channel member is no longer dropped when that member's puppet had joined the room in an earlier run (#199). wait_joined only resolved on a join it observed live, so a puppet already in the room from a previous process never satisfied the wait and the send was skipped. It now checks current membership up front and returns immediately when the puppet is already joined.

[0.13.0] - 2026-08-10

Added

  • The Discord collaboration bridge registers native slash commands (/…) that map onto the same in-room command dispatcher as typed ! commands, giving Discord the interaction surface Slack already had. One implementation, two entry points: a slash invocation is reassembled into the same positional @token string the ! handlers already parse. Argument shapes come from a new shared Command.args_spec in the command registry — validated up front, so a malformed spec names the offending command and fails in CI rather than taking the bridge down at startup — and registration is guild-scoped. Requires the applications.commands OAuth2 scope: a bot invited without it runs but answers only !-prefixed commands.

Changed

  • Breaking (agent-facing): read_context returns {threads, truncated, oldest_timestamp} rather than a bare list of thread groups, and every entry carries a kind. Both connector skills are updated.

Fixed

  • read_context seeks to a before window via the homeserver's timestamp_to_event rather than paging over everything newer to reach it, and falls back to a scan with its own page budget when the server cannot answer. Previously the walk shared the read budget, so a window deep in a busy room returned empty however small a limit was asked for.
  • read_context now pages the homeserver instead of reading a single /messages page, so agents see the whole room rather than the last slice of it. before pages backwards into older history rather than filtering one page, and the response reports truncated and oldest_timestamp so a shortened read can never pass as a complete one. Room joins now appear in the timeline as kind: "room_join" entries. The GET /agents/{id}/rooms/{id}/history endpoint, which read the result as a flat message list when it has always been thread groups and so returned no events at all, is fixed with it.
  • Discord no longer silently drops outbound messages over its 2,000-character limit. Both send paths caught the 400, logged it, and returned None, so the message never reached the channel; text is now split on line (or word) boundaries, reopening a fenced code block across the break so a torn fence doesn't render the remainder as plain text. help, at 2,063 characters, had been failing to post as !help for as long as the command list has been this long.
  • The "no agents in this channel" notice now gives each platform its own invite syntax instead of one Slack-shaped /invite-agent @agent-name line. Discord names the argument as a field (/invite-agent agent:agent-name), and Mattermost and Teams — which register no slash commands — no longer point at a command that does not exist.
  • Outbound @name mentions on the Slack bridge resolve to real <@U…> mentions from the external_users table, primed at startup and topped up as puppets are created. The name→id map was previously filled only as a side effect of resolving an inbound sender, so it was empty after every restart and whether a person was actually notified depended on whether they happened to have posted since — otherwise the mention went out as plain text. Matching is now case-insensitive, and app/bot rows, whose B… id cannot form a valid user mention, are skipped.
  • The first message from a channel member the room does not yet know is no longer dropped. Provisioning the sender's puppet only invited it; the join landed asynchronously from the puppet's own sync loop while the triggering message was relayed immediately, and a send issued before the join is rejected by the homeserver — a failure that was then swallowed. _ensure_user_in_matrix_room now blocks on a new ClientBase.wait_joined() (bounded by a 30s timeout) after inviting, and returns None rather than relaying into a room the sender is not in; inbound relay failures are logged as errors instead of silently skipped. This hit app/bot senders every time — they key on bot_id and are never pre-warmed by a join.

[0.12.4] - 2026-08-09

Added

  • Cross-artifact version compatibility, part 1 — every artifact declares what it is and what it speaks. artifacts.yaml is the one authored registry of contract revisions; per-language modules are generated from it and CI fails if they are stale or hand-edited. switch-core reads its own release version, discloses it and its ranges on authenticated surfaces only (the connection_state frame, the 409 refusal body, the gateway session response, and an authenticated credential-scoped GET /version), and records what each client declared. Nothing acts on the declarations yet — a client that declares nothing connects exactly as before.
  • The authentication surface is now frozen: never versioned, permanently backward-compatible, and excluded from gateway-api, so a client can always authenticate far enough to be told what is wrong.

Fixed

  • The agent-protocol check no longer defaults a silent client to the server's own version. It defaulted to agreement, and since no shipped client ever sent ?protocol=, the check had never once fired. Absent now records as unknown — and still connects.
  • A protocol refusal names which side is behind. It always said "update the Switch agent runtime", which for a client ahead of the server sent the user to downgrade the side that was already right.
  • The compatibility check is a range overlap rather than exact equality. The two numbers are equal today, so nothing changes yet.
  • SWITCH_VERSION is required by the standalone compose instead of defaulting to latest, which silently floated the whole stack to whatever had most recently been published.
  • The Helm chart no longer claims a version it has not been at for months. The real one is stamped at package time; the file now says 0.0.0-dev.
  • Refreshed the stale switch-core version in uv.lock, which the release procedure never re-locked.

Security

  • Bump core Python dependencies (cryptography, aiohttp, starlette, mcp, pydantic-settings, joserfc, python-multipart) to current releases.

[0.12.3] - 2026-08-07

Added

  • Collaboration bridges expose a workspace/home deeplink so a client can open the messaging app's workspace from the gateway.

Security

  • Admin-gate every collaboration-bridge write: updating and deleting a bridge were ungated, so any authenticated user could toggle agent greetings or delete a bridge (which cascades into deleting every room on it) — now admin-only .

[0.12.2] - 2026-08-07

Changed

  • The collaboration-bridge runtime indicator now follows the agent: it moves to the foot of the conversation and into the thread the triggering message belongs to (instead of staying where the turn opened), and the agent — not the bridge watching traffic — decides when it moves (new anchor_event_id on the runtime-state protocol), so it no longer jumps below a human's message before the agent has been handed it.

Fixed

  • Runtime-indicator robustness: serialise refresh vs repositioning, and don't strand the indicator when a turn ends mid-move.
  • Discord: delete an agent's messages through the same webhook that posted them .
  • Agent bridge: clamp a resumed cursor to the in-memory buffer head on heartbeat, so a connection no longer silently skips events up to a stale cursor after a switch-core restart.
  • Agent bridge: write the legacy room-binding row only for connectionless (MCP-transport) callers, so a connection-backed connect_to_room no longer leaves an unread, never-cleaned binding row behind.

[0.12.1] - 2026-08-05

Security

  • Retire the unauthenticated /collab bridge-admin API (#120).
  • Re-scan and bump dependency CVEs (#123).
  • Local stacks are secure by default: published ports bind to 127.0.0.1 (SWITCH_BIND_ADDR=0.0.0.0 is now an explicit opt-in for network exposure), .env.example ships every secret blank, and just init-env generates strong random credentials — no more admin/admin reachable on 0.0.0.0 (#122).

Fixed

  • Enforce one session of an agent per room: connect_to_room now refuses with HTTP 409 when another session of the same agent already holds the room, and a takeover evicts and reports the displaced session — instead of silently stranding a session in a room whose events go elsewhere (#109).

[0.12.0] - 2026-08-04

Added

  • Codex registered as its own gateway known-agent (connector_type "Codex") with a Codex-flavored "no live session" command, instead of falling back to Claude's builder (#91).
  • cancel_task agent operation — a requester can abandon a task it delegated, recording the reason and notifying the room; served on both the MCP server and the HTTP front door (#79).

Changed

  • list_participants now includes each participant's status and alias (both null when unset) (#79).

[0.11.0] - 2026-08-03

Added

  • Agent bridge push transport (design + Stages A & B): a sequenced, non-destructive per-agent event buffer plus a push connection with SSE delivery (GET /agents/{id}/events, one POST /connection/beat for liveness and cursor), replacing long-poll's destroy-on-read queue; presence becomes a union of heartbeat and live connection, and connection status is derived from what a connection observes rather than a declared connection_model. Backward compatible — polling keeps working and old/new state are read together (#100).

[0.10.0] - 2026-07-30

Added

  • Full attachment support — any file type and multiple files per message, in both directions across Slack, Mattermost, and Discord. Non-image files are relayed instead of silently dropped, multi-file messages are coalesced into a single platform post, and the 20MB size cap is now enforced inbound too; oversize or failed files are disclosed in the room rather than dropped (#93).

[0.9.0] - 2026-07-29

Added

  • Every room now gets a collaboration bridge by default: a bridge can be marked default and is used when a room names none, with internal_only as the explicit opt-out; a bundled Mattermost is provisioned across compose and Helm (#84).

Changed

  • Restyle the Switch Gateway web UI onto the Hoot design system — new light/dark theme, icon-rail shell, and consistent formatting conventions; nav destinations unchanged (#87).

[0.8.1] - 2026-07-24

Fixed

  • Prevent a duplicate Switch room from being created when provisioning a Slack channel (race on bridge channel provisioning).

[0.8.0] - 2026-07-24

Added

  • Scoped agent-addressing permissions — control which agents and rooms an agent may address, enforced server-side.

[0.7.0] - 2026-07-24

Added

  • Microsoft Teams collaboration bridge — rooms can bridge to Teams channels, joining Slack, Mattermost, and Discord as supported platforms.
  • Discord bridge polish: deeplink redirect, room icon, and outbound image relay .

Changed

  • Relicensed to Apache-2.0 + Commons Clause, with a CLA gate for contributions .

Fixed

  • Detach dependent rooms before removing a bridge, so bridge removal no longer leaves rooms in a broken state (#35).

[0.6.0] - 2026-07-20

Added

  • Discord collaboration bridge — rooms can bridge to Discord channels, joining Slack and Mattermost as supported platforms.

[0.5.0] - 2026-07-20

Added

  • New POST /gateway/auth/refresh endpoint: re-mints the switch_auth cookie from a still-valid session, enabling clients (switchdash) to silently renew sessions before the 24h expiry.
  • Standalone compose: bind address and host ports are configurable via env vars (SWITCH_BIND_ADDR, POSTGRES_HOST_PORT, API_HOST_PORT, …) — consumed by switchdash's local-server mode.

[0.4.0] - 2026-07-19

Added

  • The standalone Docker Compose file is published to GHCR as a versioned OCI artifact (standalone-compose:<version>, plus latest) on every release, pinned to the same version as the images and chart — switchdash's local-server mode consumes it.

[0.3.0] - 2026-07-17

Added

  • Agents can send images from rooms, relayed out through the Slack and Mattermost collaboration bridges.

[0.2.1]

Fixed

  • switch-core Deployment now uses strategy: Recreate so upgrades don't briefly run two pods — a second replica collided with the singleton's in-memory Matrix sync / collaboration-bridge sessions and could leave a bridge stuck until a manual restart.

[0.2.0]

Added

  • Helm chart: global.imagePullSecrets for pulling the images from a private registry (e.g. a private GHCR namespace); defaults to empty so public-image installs are unchanged.

[0.1.0]

Added

  • Packaging metadata for switch-core (authors, URLs, classifiers, license placeholder) and a switch-core console entry point.
  • Image-publishing CI: multi-arch (amd64/arm64) switch-core, gateway, and setup images published to GHCR, plus the Helm chart published as an OCI artifact, on switch-v* tags.
  • Documented release procedure (this file + RELEASING.md).

Changed

  • Distribution endpoints (electron-updater target, plugin-marketplace source, image registry/namespace) centralized so the eventual public-repo move is a configuration flip.

  • Initial internal version. History predating the changelog lives in the git log.


switch-console

[Unreleased]

Added

  • Replace a messaging app's token without disconnecting it. When a bot token is rotated or leaked, a workspace admin can paste the new one from the connection's menu on the server page. The connection restarts with it and keeps its rooms. Offered on connections to your own bot, not on ones that came from installing Switch's app.
  • Airtable's logo on the Connections grid.
  • Admin → Connector apps sets up Airtable. Its dialog walks Airtable's Builder Hub with the redirect URL to copy and the scopes to tick, and asks the operator to confirm the integration's support details are set, which no check can see. A server that does not send redirect_uris, needs_public_url and scopes is read as before.
  • Saving a connector app shows what its check said. After Check and save or an updated secret, the app's page opens with the vendor's answer and any note, as Check it works shows them.
  • Connectors says why Airtable or Google Workspace can't be connected on a server without a public address: the operator is told to set GATEWAY_PUBLIC_URL, and everyone else whom to ask.
  • Codex sessions can use an API key kept in the OS keyring. With cli_auth_credentials_store = "keyring" or "auto", the execution machine's Codex keeps its login in the keyring, where a session's own Codex home cannot find it, so every session failed with "Sign in on the execution machine with codex login" while codex login status said it was signed in (#1048). When the machine has no auth.json to copy, a session now asks its Codex for the API key as it starts and keeps it in the memory of the session's own Codex process: it is not written to a file, put in the environment or copied into another keyring entry, and a changed key is picked up at the next start. A ChatGPT login in the keyring cannot be handed over, and the session says so with the file-storage workaround; it also refuses to hand over the key when an administrator's Codex configuration would have it saved elsewhere. A session that still has a working login of its own, such as one it copied while the machine had an auth.json, starts as before; a copy that no longer works is removed.
  • A Claude Code agent's session can continue a conversation started in the Claude Code CLI. The new-session dialog lists the conversations started in the agent's folder, on this computer or its SSH host. The session continues a copy of the one picked, so the terminal's own is never written to, and its transcript starts with what was said in it. A conversation still open in a terminal can't be picked. No earlier prompt is sent again.

Changed

  • The Switch skill tells agents what to do when their message was redacted. When a send's result says Switch Trust redacted it, the agent reads what the room received and, if that no longer makes sense, says what it meant in words that pass instead of reposting the original.
  • The Switch Trust settings say what is checked in attached files. The text inside a file is checked up to its first 20,000 characters; files with no text to read, such as images and archives, are sent without being checked.
  • No Owner role. Inviting people and changing roles offer Admin and Member, and the copy says "an admin of this workspace" where it said "an owner or admin". A server from before the merge that still reports owner shows as Admin, including for workspaces stored before the update.
  • Usage data follows an install to a working agent. Three new activation_milestone steps (server_ready, agent_ready, first_agent_session) are each sent once per install, and only by installs made on this version or later, with path saying whether the install started switch-buddy or skipped it. Sessions of agents on the agent controller Console runs on this computer, which never reached session_started, are sent as managed_session_started. Updates send update_download_started and, on the first launch after one, update_installed with the version it came from. Adding a server by address, starting a server Console runs, and creating a room now say why they failed where they said error, and Switch Cloud is its own server_kind. An install once signed in with a staff account stays labelled internal after it signs out. agent_reset, which was listed but never sent, is gone from the list. docs/TELEMETRY.md has every field.
  • Managed agents and machines update when the server says they changed. Against a server that pushes change notices, Console reads those lists again when told to, and otherwise only every five minutes instead of every 15 to 30 seconds; against one that does not, it polls as before (#717).
  • One name for the controller. Console, the Gateway and the docs say "agent controller" everywhere, matching the switch-agent-controller package, instead of a mix of "agent controller" and "agents controller". The delete dialog on a server's page now calls it "this Switch server that Console runs" rather than a "managed server", which read like "managed agent"; it was the only place Console showed that name.

Fixed

  • Changing your password no longer signs Console out without a word. The server ends every session on a password change and hands the one that made it a new session, which Console now keeps, so it stays signed in. The account page says "Password changed successfully" in place of a toast. And whenever a server ends Console's session (it expired, or was signed out from elsewhere), Console re-checks its sign-in and shows the server as signed out with a sign-in prompt, instead of "Chats offline" and a warning that its rooms may be out of date.
  • Direct and group conversations no longer offer adding or removing agents, in a room's side column, the sidebar or an agent's "Add to rooms", where the server refuses it now that their platform decides who is in them. People in the conversation still use !invite-agent and !remove-agent.
  • Antigravity runs Google's own agy CLI, on Windows too. Switch Console finds agy (including %LOCALAPPDATA%\agy\bin on Windows), installs it with Google's installers, runs sessions with agy acp, and checks the sign-in you made with agy acp --login or a GEMINI_API_KEY, instead of a separate Antigravity ACP profile that a Google sign-in never reached. Antigravity's questions now reach you, too, including typed answers. A machine that only has the older antigravity-acp runtime keeps using it until agy is installed, and Update installs agy there. An Antigravity sign-in Switch holds for the older runtime cannot sign agy in: a session given one now says how to sign in on the machine instead of starting signed out.
  • Console no longer opens to a blank window on a database a newer build has changed. Stable and Canary share one database, so after Canary added a migration, an older build found tables it was not written for, applied nothing, and failed its first query with nothing on screen. Console now refuses a database carrying a migration it does not know, before changing anything, and says why: open the build that last used it, download a newer release (a build that cannot start cannot update itself), or move the data folder aside to start empty. Migrations are tracked one by one, so a migration applied out of order is applied rather than skipped. Any other startup failure now shows the error and a Reload button instead of a blank window. drizzle-kit now defaults to the dev data folder's database rather than the installed app's.
  • Giving a machine a Vertex AI login no longer pre-fills a project. The Google Cloud project field started out filled with one organisation's own project, so anyone else who kept it got a sign-in failure with nothing to say why. It now starts empty, with a placeholder; the region still starts at global.
  • Two agent hosts starting at once on Windows no longer crash each other. A host reading another host's ownership ticket while that host replaced or deleted it got EPERM from Windows and exited. It now tries the read again for a few seconds, as it already did for its own writes.
  • Codex is no longer shown as ready with a login sessions cannot use. A ChatGPT login kept in the OS keyring passed Console's sign-in check, so Codex showed as signed in while every session failed: a session takes that login only from auth.json. The check now reports a login sessions cannot reach as not signed in, and says why: a ChatGPT login or a Bedrock key Codex keeps in the keyring, an auth.json Codex ignores because it is set to use the keyring, and an API key in the keyring that Codex will not hand over or that an administrator's Codex configuration stops Switch handing to sessions. It gives the switch to file storage where that helps and names the Codex home it checked (from ~ when it is under your home folder), and when it cannot tell, it says so and why. Otherwise an API key in the keyring still counts as signed in, since sessions are handed it. The provider status card and the warning beside an agent now show what the check said when a provider is not signed in, instead of a generic line.
  • Console stops offering a sign-in that cannot help. When a provider's check says signing in would not fix it (an administrator pinning where Codex keeps logins, or a check that could not finish, say), the status card, Settings → Agent providers and switch-buddy say it can't be used yet and why, instead of "Not signed in" with the login command or a Sign in button.
  • A relative CODEX_HOME is refused. Codex reads one from whichever directory it starts in, which differed between Console's check and Switch's sessions, so the two could look at different Codex homes. Both now say to set an absolute path, and a session will not start with one.
  • Settings → Agent providers keeps this computer's report current. It prefers the report from the server Console is on, rereads it while open and after the controller reports again, and the install section says so when Console's own check disagrees with the controller's.
  • The local Switch server starts on Intel Macs without Homebrew. cryptography publishes no wheel for Intel Macs, so the Intel build of Console compiled its own against the build machine's Homebrew OpenSSL, and on a Mac without Homebrew the server failed to start with Library not loaded: /usr/local/opt/openssl@3/lib/libssl.3.dylib. The bundled cryptography now carries its own OpenSSL, as cryptography's wheels for Apple silicon do, and a macOS build fails if anything in the local server links a library that is neither macOS's nor shipped with it.
  • Setting up a server notices Docker once it starts. "Set up a server on this computer" checked for Docker only when the step opened, so installing or starting Docker meanwhile left the setup skipping Mattermost until you went back. It now checks again every few seconds, and when Console regains focus, while Docker isn't answering.
  • Starting Docker with the server page open now enables the Mattermost switch. The server page checked for Docker only when it opened, so Docker Desktop started afterwards left the switch greyed out until you left the page and came back. It now checks again the way the setup step does. A running server whose Mattermost was left out or failed to start, because Docker was off, shows the switch off, so turning it on brings Mattermost up without a restart (#1021).
  • A local Mattermost shows up as soon as it is ready. When the Mattermost beside a local server finished setting up, pages already open kept showing no Mattermost and asked you to link an account until the server was restarted. They now refresh the messaging apps and linked accounts the moment it is up (#1021).
  • Agents say who can answer their questions and approvals. The Switch skill said a request card needed the owner's linked identity and room membership. Its addressing policy decides, and agents now read theirs with get_agent_detail and name the people who can answer.
  • Giving this computer's Codex login to a machine explains a keyring login. When Codex keeps its login in the OS keyring (cli_auth_credentials_store = "keyring" or "auto"), it said "This computer is not signed in to codex". It now says the login is in the keyring, which cannot be given to another machine, and gives the workaround for a machine whose organization does not enforce the keyring: switch Codex to file storage, which keeps the login in plaintext, and sign in again (#1048).

Removed

  • The fallback to the agent event stream. The agent runtime connects only over the WebSocket, so it can no longer reach a Switch server older than the socket (agent-protocol 7 and older), and it stops sending its own heartbeat requests (#673).

[0.40.1] - 2026-10-21

Added

  • Managed agents and template-made rooms are reported, and managed servers point at the current relay (#963).

Fixed

  • Search under agent management shows managed agents, not legacy ones (#975).
  • Deleting a managed agent removes the Console agent it was moved from (#974).

[0.40.0] - 2026-10-20

Added

  • Chat-first Console backed by Switch rooms. Console leads with agents and rooms and a chat view, with a room summary shown in place and quieter tool activity in chats (#716, #934).
  • Guided first-run onboarding with switch-buddy. A fresh install walks through onboarding and ends in a chat with switch-buddy — an assistant that knows its context and can hand you Console links — with an intro animation and page entrances (#879, #915, #942, #932, #930).
  • Run a local Switch server without Docker (#915).
  • Delete my account from Console (#915).
  • Stable ships with onboarding and the managed server's feature flags on (Switch Cloud off). A stable build now shows the first-run onboarding, and the Switch server Console manages starts with agent management, agent chats, switch-buddy, follow-the-conversation and service connections enabled. Switch Cloud stays off in stable.
  • The server's operator sets up connector apps in Console. For the operator, the Admin page has a Connector apps section under Server, below the workspace's switches: the GitHub App and Google Workspace's OAuth client as tiles, the ones needing attention first, each with its state and the one thing to do next. A tile opens the app's page, with its client, its secret's last four characters or its key's fingerprint, who uses it and a check with the vendor; updating a secret, switching to a different app and removing it are in its menu. Setting one up walks through creating it at the vendor, with the values to paste and copy buttons, then checks and saves it in one step, marking the value the vendor refused. It is the dashboard's Connectors page, which stays for now. Needs a server from this release.
  • Console follows a server that keeps names away from avatar services. On a server with THIRD_PARTY_AVATARS_ENABLED=false, Console no longer loads DiceBear or ui-avatars.com images for that server's agents and shows their initials instead. Its icon picker offers only an image link and refuses one to either service, new agents are created without a generated icon, and it no longer writes generated icons onto existing agents. Until it has heard from a server, Console shows initials rather than send a name, and it reads the setting again every few minutes, so a server restarted with it changed needs no Console restart. A setting it cannot read counts as off: the icon picker says so and offers to ask again, and a new agent is still created, without a generated icon. Agents on servers that leave the setting on look as before. Older versions of Console load those images whatever the server says.

Changed

  • Link your account right after adding a messaging app. Every way into connecting an app, including a Console link from switch-buddy, now ends by offering to link your own account on it (Telegram explains how instead). The server's Home page keeps a Link your … account prompt until you do.
  • The delete-room dialog asks about the channel. Deleting a room keeps its channel unless you tick Also archive the … channel (or delete it on Discord, or have the bot leave on Telegram); the dialog says what will happen either way, and shows what to clean up by hand when the server could not retire it. Teams channels always stay. Needs a server from this release; an older one gets "the channel may remain".
  • A new agent starts on "Me and my agents". "Who can talk to your agent" now opens on Me and my agents (default): you and any agent you own. Only me is still one choice away, and agents that already exist keep their setting. A template agent with no addressing is created the same way.
  • Usage data goes to telemetry.switchagents.ai. Console reported to telemetry.flintai.dev, an older name for the same relay that is being retired. Only where it is sent changes; what is sent, and the opt-out, do not.
  • A sign-in comes back to Console on its own. When the browser returns from a service's sign-in, Console comes to the front with the next step showing, and the browser's page says the sign-in is done and the tab can be closed. It used to stay behind the browser, with a page telling the person to return to it and confirm. Services whose sign-in returns through the server (Google Workspace, GitHub) also no longer stop on the server's "Continue in Switch Console" page, with a server from the same release.
  • A sign-in connects without a confirm step. Once the browser returns, the server keeps the connection and Console shows the result: a service whose agents are added from its page opens that page with the search ready, and any other shows "connected" in the dialog. Console no longer asks to confirm the account first; the account is chosen at the vendor, and Disconnect undoes a wrong one. A sign-in the server refuses, such as an account already linked to someone else in the workspace, shows the reason in the dialog, and the browser's page says to look there. GitHub's sign-in works the same way. Needs a server from this release.
  • A connector's agents are managed on its page, the way a room's participants are. The connector's page lists the agents using it as cards, with each agent's avatar and what runs it, under an "Agents" heading with the count. "Add agent" opens a search of your other agents under the cards, and choosing one turns the service on for it at once. Connecting a service now lands on its page with that search open, instead of a checklist in the connect dialog. An agent others can address is offered with a note that it would act as you. A cross on a card, shown on hover, turns the service off for that agent again, so one agent's access can be taken away without disconnecting from every agent.
  • GitHub is connected, and handed to agents, from its page like any other connector. Its page has a Repositories section: one card per account the App is installed on, with its count, opening to a directory of names in columns that is searched rather than scrolled once it is long, re-read as the window regains focus. The section's "Choose repositories", "Change repositories" once there are some, goes straight to GitHub. The connect dialog ends with Done once GitHub is connected, the page behind already current; it no longer offers Back or Disconnect, which the page has. "Add agent" is offered for GitHub too: what every added agent gets is chosen above the search, and the page says so while nothing is chosen, since no agent can be added until something is. "Choose repositories" opens the choice in place: the account where there are several, the repositories as chips, added one by one from a search or all at once and removed with a cross, and whether the agent may push. An agent's own choice is changed on its Connections tab, whose form is the same choice in place of the account dropdown and checkbox list.
  • A workspace's service switches are on an Admin page. They were a "Workspace" tab on the Connectors page, beside Yours and Discover. An Admin entry in the sidebar, shown to the workspace's admins and the server's operator where service connections are on, opens a page laid out like Settings: its sections down the side, grouped under what they reach ("Workspace" and the workspace's name). Its Connectors section shows the switches as tiles like Discover's: each service with its category, how many members have connected it and agents use it, and who last changed it, with its switch. A tile keeps its place when it flips, and switching one off still asks first when anyone uses it. The Connectors page has Yours and Discover again, with "Manage in Admin" in its header for an admin, and its search now fills the row.
  • Connector tools ask in approval-required and auto-accept-edits mode. Calls to a granted service's tools (Jira and Confluence through Atlassian, Datadog, Google Workspace) used to run without asking in every mode. They now follow the agent's mode, as its other tools do: an approval card asks first, and declining tells the agent the call was refused. Full access still allows them, and Switch's own room tools never ask. The first such card in a session comes with a notice saying why. Claude Code, Codex and OpenCode cards name the server and tool and show what the call will send, so the text of a comment can be read before it is posted; a long input is cut and the card says how much is not shown. Codex cards can be approved for the rest of the session. A session started before the update keeps allowing them until it restarts.
  • A Console session in "approval required" no longer inherits your own Claude Code default mode. Console now starts a Claude Code session in that mode with Claude Code's default permission mode. It used to leave the mode unset, so a defaultMode such as auto in your own Claude Code settings applied, and tools, connector tools among them, could run without the approval card the mode promises.
  • A session says which services it lost or gained. When a running session restarts because its agent's services changed, it shows a notice naming them, for example "Datadog was removed from this agent's tools: it was switched off for the workspace, or its grant was removed." The agent is told the same as it resumes, so it does not look for tools it no longer has.

Fixed

  • Switch buddy stops waiting on a Claude Code it cannot check. A Claude Code from before claude auth status left the AI provider screen on "Checking" until its timeout; it now asks to sign in and says the CLI may be too old. A login that is there but cannot answer a test prompt shows Claude's own error next to Sign in.
  • Settings → Agent providers finds a CLI installed since Console started. Console now looks in ~/.local/bin, where the Claude Code and OpenCode installers put their CLI, at start and on every refresh, as this computer's agents controller already did. Before, a Claude Code installed by hand showed "CLI not installed" there while Switch buddy was using it.
  • Settings → Agent providers and Switch buddy agree on this computer's providers. When this computer's agents controller has reported to a signed-in server, Settings shows that report, the one buddy and Switch go by, and its refresh asks the controller to check again. Console's own check is used only when there is no such report. Install and update still run through Console's own check.
  • A failed agent install says why. Installing Claude Code, Codex or another provider from Switch buddy showed only "Install command failed." It now says when the disk is full, and otherwise adds the line of the install output that says what failed.
  • Signing in to a service through Core works on a server whose gateway and API are on different hosts. Core sends the browser through its own authorize step at its public address, the API's host, where the vendor returns it; Console only accepted the gateway's host and refused the sign-in as an invalid URL. It now accepts either of the server's own origins, over HTTPS. Google Workspace connects this way.
  • The usage-data note says what a "yes" does on a shared server. The consent dialog and Settings said the answer applies to a server Console runs for you. On a server other people also run from their own Console, a "no" turns sharing off for everyone at your next start, while a "yes" leaves it as it is — on stays on, off stays off — unless none of them has used the server in the last 14 days. The note now says so.

[0.39.1] - 2026-10-13

Fixed

  • Support the gateway and API on different hosts when connecting a service. Some deployments serve the Switch dashboard and its API from separate addresses. Console now accepts either of your server's own addresses, over HTTPS.

[0.39.0] - 2026-10-12

Added

  • Service connections follow the server's service_connections flag, apart from cloud machines. With it off, Your Agents offers no Connections, no agent shows Service access, and setting up Switch Cloud goes straight from signing in to the first cloud agent; turning it on reaches a running Console without a restart.
  • Agents / Rooms sidebar and a reworked workspace menu. On a server with agent management, Console leads with Agents and Rooms above the legacy Sessions, behind the server's feature flags, and the workspace menu is restyled (#751).
  • Managed server status and a Server page. A managed server shows status dots, has its own Server page, and surfaces a frozen-workspace state (#762).
  • Switch Trust settings. Guardrails are configured from a settings panel in Console, backed by the gateway's Switch Trust API (#747).
  • Windows support for managed agents and the controller (#765).
  • Canary has its own switch-core pin. The canary build can bundle and run a different switch-core release for the servers it manages than stable does — a pre-release can be tried on canary without moving stable. The canary pin can never be older than the stable one. Canary and stable share their managed servers on one computer, so once canary has updated one, stable will not start it again until stable's pin catches up (#764).
  • Switch cloud on the agents controller. On a server whose cloud machines run the agents controller, choosing Switch cloud in the New agent form starts your cloud machine and waits for it to come online (its first start takes a few minutes), then creates the agent as a managed agent on it, like on any machine. Once online, the machine is listed as "Switch cloud" among your machines. A server whose machines run the hosted worker works as before.
  • Give a machine a provider login. In the New agent form, a machine with a provider installed but not signed in offers to give it a login: a Claude setup token or API key, a Codex or Cursor API key, or this computer's own Codex, OpenCode or Antigravity sign-in. Console seals it to the machine's own key before it leaves, so the server only relays it, and shows whether the provider signs in with it on the machine.
  • Agents granted GitHub can use it in every session. Git's credential helper and a gh wrapper are set up for https://github.com in the session alone, and ask the agent host for a short-lived token; SSH remotes and other hosts keep your own setup. Nothing is set up for an agent without a GitHub grant, or on Windows yet. A token GitHub refuses is replaced on the next command. Tokens are scrubbed from session logs, transcripts, activity shown on messaging platforms and anything the agent posts to a room.
  • Service access in an agent's settings and on a managed agent's page, Switch cloud agents included: its grants, the GitHub repositories they reach, and a warning, with "Make owner-only", when anyone can address the agent.
  • The skills of an agent's granted services reach each session, and a session is told when its agent's grants could not be loaded.
  • A server's Connections show a connection that needs reconnecting, and why a service cannot be granted on that server.
  • Connecting GitHub says when an organization's single sign-on hides its installation of the App, and what to do, instead of waiting on access that cannot appear.
  • Connect Atlassian, and any other service the server offers, from Connections. Signing in happens in your browser and comes back to Switch Console, which confirms the account with you before the server keeps it. Atlassian takes a sign-in back only on a few fixed ports on your computer; if all are in use, Switch Console says which to free.
  • Agents granted a service whose tools are a command-line tool, such as Google Workspace, get that tool in every session. The session host runs the vendor's pinned build itself, never through a shell, after checking each command against the service's catalog entry, and keeps every file it reads or writes inside the session's folder. Each run works in a folder of its own, with your token in its environment for that run alone; the coding tool never sees the token, and it is scrubbed from the output. The build is downloaded on first use and checked against its pinned SHA-256; the agents controller reports it among its tools.
  • Agents granted a service such as Atlassian get its tools in every session, as an MCP server of their own beside Switch's. The session host calls the service with your token, which never reaches the coding tool, and scrubs it from everything the session records. A service that is slow or failing shows as one failed server, and Switch's tools keep working. Only text from a service's tools is passed on; anything else is replaced by a line saying what was left out.
  • Service access in an agent's settings shows services such as Atlassian as an on/off switch, saying what on means and how soon off takes effect.
  • Connect a service with a key, such as Datadog, from Connections. You paste its keys and pick its site in a form; Switch checks the key with the service before keeping it, and never shows it again. Agents granted the service use it with your key, which the session host puts in each request or run as the service's catalog entry says, and scrubs each key from everything the session records.

Changed

  • The agents controller reports git and gh as found, missing, or not supported on Windows.
  • An idle session now parks after a day rather than 30 minutes. Set SWITCH_SESSION_PARK_AFTER_MS on the session host to choose another timeout, or off to never park.

Fixed

  • A session whose provider offers hundreds of models no longer dies on its next message. OpenCode signed in to many providers lists more models than a session event can hold, so every update to the session was refused and the host stopped when the next message arrived, leaving the room unanswered. The catalog is now cut to fit (without the models' options first, then without the models past what fits, the selected one kept), and the session says so. A host that fails also records its failure without the catalog, rather than failing again on it.
  • A machine whose controller Switch refuses for good says so instead of restarting it forever. When the server needs a newer agents controller than Console carries, "This computer as a machine" shows that Console needs an update and stops restarting the controller. When the server no longer knows its credential, Console forgets it as it does for a removed machine, and enrolls the machine again on the next check. SSH hosts set up by Console stop restarting on either and enroll again on the second.
  • A session's helper endpoint for GitHub no longer hands out the token of any other service the agent was granted.
  • A Claude Code session is no longer parked while its background subagents are still working. The idle timer only looked at turns, so a session whose turn had ended with subagents still running in the background was stopped after the idle timeout, taking the subagents with it. Those subagents now count as activity, and the idle wait starts again when the last one stops.

[0.38.1] - 2026-10-01

Fixed

  • Clean recovery after the machine sleeps. On wake, a watcher no longer takes itself over, a silently-dropped SSH connection is detected rather than trusted until it says it closed, and an agent shows a single status row instead of duplicates (#621).

[0.38.0] - 2026-09-30

Changed

  • Updating the sidecars on a host with many agents takes a fraction of the SSH round trips. Each agent's bring-up used to check the host's shared bundle for itself — half a dozen commands asking the same question — and then stop an old sidecar, write its flags, stage its configuration and launch it in as many more. The bundle is now checked, and uploaded if needed, once per host, and each agent is brought up in one command after its configuration is staged: about two round trips instead of thirteen. Agents are still brought up one after another, so their reconnects to Switch stay spread out.
  • Every agent starts a session when it is addressed; the "Auto-create a session on notify" setting is gone. Its switch is removed from the agent page and the create form, and new agents register as auto_session.
  • Opening an agent's page only reads. Reading the auto-session setting used to re-launch the agent's controller as a side effect, which on a remote host uploaded the sidecar bundle and replaced an outdated sidecar, and checking a provider's sign-in or models uploaded the bundle too. The provider check now uses the bundle already on the host, and says so when there is none. The bundle is uploaded only when an agent is started, restarted or updated.

Fixed

  • A remote agent's status can no longer go stale. Console used to learn whether a remote agent was connected only from what its sidecar pushed down a long-lived connection, and trusted that connection until it said it had closed. One that died without saying so, as when every sidecar is replaced while its host's SSH connection is rebuilt, left the last thing it carried on screen for good: agents that were answering in Slack showed "Connection failed" and "The agent's room watcher is not running". Console now reads each host every five seconds, in one command for every agent on it, and the sidebar and the agent's page are both built from that one read.
  • Agents come up after Console starts without their page being opened. Console now brings each host's agents up side by side, so a slow or wedged SSH connection holds back only that host, never local agents or other hosts, and it keeps retrying an agent whose controller could not start, backing off from 30 seconds to 5 minutes. Before, an agent that failed once stayed off until its host reconnected or its page was opened.

[0.37.3] - 2026-09-29

Changed

  • The bundled Switch skill now documents the workspace-template tools an agent can use — finding, reading, running and saving templates (#554).
  • The bundled local Switch server moves to switch-core 0.29.0, so local-server mode ships the workspace-template feature those tools rely on.

[0.37.2] - 2026-09-27

Changed

  • Checking a remote agent's watcher reads only that agent's own state. The watcher panel, session start and the sidecar connection used to read every agent's watcher configuration on the host to find one, so any other agent's broken or duplicated state could fail them. The agent's directory is named from its identity, so it is now read directly; the others are only looked at when it has none, to adopt a watcher saved under an earlier name.

Fixed

  • A remote agent's room watcher no longer fails with "Competing saved watchers require explicit cleanup" or an unreadable-config error after a launch was interrupted. Each launch staged the agent's configuration in a scratch directory beside every agent's watcher state, and the watcher status check, session start and session listing all read that directory in full. A scratch directory left behind by a dropped SSH connection held either a half-written file, which broke the check for every agent on the host, or a complete copy of one agent's configuration, which made that agent look like it had two watchers. Launches now stage in a separate sdk-launch directory, the readers skip any scratch directory an earlier version left, and each launch removes staged configurations abandoned more than an hour ago, since they carry the agent's credentials.
  • Agent discovery failures no longer dump a raw error into the sidebar. A failed discovery now shows a compact discovery-failure indicator instead of the raw error text (#565).

[0.37.1] - 2026-09-26

Fixed

  • An agent's instructions and settings can no longer be blanked or reverted by the Console itself. The agent's config file (.switch/config/<name>.json) is now the only record: opening an agent's page never writes anything, Claude Code sessions get their agent definition from the config file directly instead of from .claude/agents/<name>.md, and that file is no longer written or read back. A missing config file is reported as an error rather than treated as an agent with no instructions. A one-time migration gives every existing agent a config file, taking over its description and any hand edits from the old definition file; attaching to an agent, or adopting a subagent, creates one when it is missing. Deleting an agent's provisioned files now removes its config file too, so an agent re-created under the same name starts clean.

[0.37.0] - 2026-09-25

Added

  • Codex sessions get the Switch skill (and the agent's own instructions) as developer instructions, like Claude/Cursor/Antigravity — no skill file is written, so a session no longer needs shell approval to read it, and a failed turn now shows its reason (#547).
  • Providers report per-turn token usage — prompt/completion plus cache reads and writes — on the session activity stream, feeding server-side usage metering.

Changed

  • Sidecar session handling refined: shared-agent-runtime initial prompt, session-command and stop-shared-session paths, and transcript handling.

Supersedes the never-tagged 0.36.0; that release's changes (below) ship to users for the first time here.

[0.36.0] - 2026-09-25

Added

  • Sessions live with the host. Console and its sidecar now own agent sessions and their lifecycle; switch-core only relays the connection (#543).
  • Capture, publish, share and search room templates. Capture a live room as a shareable, parameterized template (values become {key} placeholders), and publish, edit, share and search templates from the Console (#501).
  • Templates. Create a new agent from a template, with a dedicated create-from-template view (#405).
  • Load and manage agents on a remote host (#364).

Changed

  • Connector telemetry events now carry a failure reason and a duration_ms (#488).
  • Managed servers behind the app's switch-core pin are upgraded automatically, and sessions wait for it. The local server is upgraded at startup and a remote managed server when its host is reconciled (at startup and whenever the host becomes reachable); a stopped one is upgraded at its next start rather than started at the old version. Every database is backed up with pg_dumpall into a timestamped backups/ directory next to the stack first, and an upgrade.json journal resumes an interrupted upgrade or refuses it loudly. Sessions and room watchers for agents on the server wait while it updates and are refused with the error if it fails. The server page shows the update's progress, or its error and a Retry button, in place of the optional "Restart to update" banner.
  • Usage telemetry is now opt-out. The setting defaults to on and the first-run dialog is a notice rather than a question — it still states what is shared, is not dismissible, and carries the off switch (pre-set to on). Consent is re-read before every event, so turning it off stops the next event. A migration preserves anyone who had already opted out (#490). The managed server honours the same single telemetry answer (#529).
  • Watchers are console-parented locally and sidecar-parented remotely, so a watcher's lifecycle follows the process that should own it (#507).
  • Local-server mode now bundles switch-core 0.28.0 (was 0.25.0), and the refreshed agent-runtime 0.7.0 and sidecar 1.9.9.

Removed

  • Dead UI and RPC surface removed from the Console (#526, CHOO-2323).
  • The bundled connector plugins are retired: every session now gets the Switch tools from its own host and the Switch skill pushed by Console. The plugin files remain only so Consoles that have not updated can still install and update the plugin they expect (see connectors/DEPRECATED.md).

Fixed

  • An unusable remote working directory is reported before an identity is minted, rather than failing later (#373).
  • The managed stack provisions the database roles it now needs and records why it sets no messaging-app variables; the bundled standalone compose is re-synced.

Security

  • Untrusted room-message bodies are fenced in the prompt delivered to a session, matching the switch-core hardening (#527).

[0.34.0] - 2026-09-06

Changed

  • Local-server mode drops Matrix. The bundled standalone stack no longer runs a Tuwunel service or sets MATRIX_* environment, matching switch-core 0.25.0; an existing local stack is migrated off Matrix on start. Bundles switch-core 0.25.0 (was 0.24.x) (#380).

[0.33.0] - 2026-09-06

Added

  • Local control API for programmatic agent management. A localhost-only, token-gated HTTP server in the main process (127.0.0.1 on an OS-assigned port, credentials in a 0600 file in the app data dir) exposes the same agent-management operations as the UI — list/read agents, list an agent's sessions, and start a session (#362).

Changed

  • Agent names are auto-lowercased as you type in the New Agent dialog, matching the identifier rules (rather than rejecting after the fact).
  • Local-server mode now bundles switch-core 0.24.0 (was 0.23.0), and picks up the internal Postgres LISTEN/NOTIFY transport client-side (#363).
  • Refreshes the connectors so installs pick up the display-name skill updates (Claude Code 0.9.13, Codex 0.3.14, OpenCode 0.1.9).

[0.32.0] - 2026-09-03

Added

  • A session that dies on an API error now says so. A Claude Code turn that fails on expired credentials, quota or a provider outage fires StopFailure, which nothing listened for — the channel just went quiet and the operator had to open the TUI to find out why. It now maps onto the error status and carries the reason through, so the deeplink ping names what broke instead of a bare "needs your input" (#356).

Changed

  • Sidecar upgrades are taken under live sessions now, except a major. A sidecar whose build differed from the client's used to be left alone whenever it owned a session and upgraded only once idle, so a busy host could sit on a stale sidecar indefinitely. The upgrade (a few seconds of room injection) is now taken while sessions run; a major bump still waits for idle (#356).
  • Registering an agent with a name that isn't accepted now offers a valid slug instead of rejecting it, and shows the rejected-name message as an error banner (#325).
  • Bundles agent-runtime 0.4.1 (deleted-room / heartbeat backoff fixes), sidecar 1.9.7 (reaps dead tmux targets), and the updated connectors (Claude Code 0.9.12, Codex 0.3.13, OpenCode 0.1.8).
  • Local-server mode now bundles switch-core 0.23.0 (was 0.21.0) — the pinned images and standalone compose it pulls move up to the current release.

Fixed

  • An errored turn no longer reads as a fresh request for input: the stalled prompt that follows a failure is suppressed so it doesn't bury the reason, and on the Slack card the reason is marked as a stall rather than streamed under a spinner as if it were progress (#356).

[0.31.3] - 2026-09-01

Added

  • The disabled "Add agent" button now explains why on hover, splitting the not-ready tooltip into distinct checking and blocked messages (#314).

Changed

  • Bundles agent-runtime 0.3.4 (reaps orphaned agent runtimes at boot) and the updated connectors (Claude Code 0.9.11, Codex 0.3.12, OpenCode 0.1.7) and sidecar 1.9.6.

[0.31.2] - 2026-08-27

Added

  • Orphaned agent runtimes are reaped at boot: a runtime whose host has gone — left behind because a user updated the app but never revisited the connector — is cleared on startup, using the same "is this runtime abandoned?" predicate the runtime itself uses, so a live session is never touched (#309).
  • A one-shot catch-up brings already-installed Switch connectors up to date once, so an install that never revisits Settings → Agents stops drifting several runtime versions behind. It runs once, latches, and leaves the Update button as the normal path; it never installs a connector that isn't already there (#309).

Changed

  • The bundled sidecar (1.9.5) and the OpenCode connector (0.1.6) this build writes now run agent-runtime 0.3.3, which exits with its host instead of leaving stale processes behind.

[0.31.1] - 2026-08-26

Fixed

  • The setup checklist's ✕ now actually dismisses it. The button did nothing, so a checklist a user had finished with stayed on screen.

[0.31.0] - 2026-08-25

Added

  • Creating an agent, starting a session and adding a server now report whether they worked, and an enumerated reason when they did not. Previously only the ones that succeeded were counted, so nothing showed where people got stuck. A session start also records which button it came from, whether a room and an opening prompt were chosen, and whether a person started the session, the app started it on their behalf, or it was found already running on a remote host — never the room, the prompt or the host.
  • Removing an agent, resetting a remote agent, removing a server, attaching a remote session's terminal and retrying a session that failed to set up are all reported now, each with whether it worked. Removing an agent distinguishes a person doing it from a server teardown sweeping every agent up with it.
  • Installing, updating and removing an agent's CLI is reported — the app's main first-run wall, and previously invisible. So are creating and deleting rooms, signing in and out, connecting and disconnecting a messaging app, setting up a remote host, starting and stopping a managed server, and the update check, download and install.
  • Where people stop is now visible: the first-run checklist, the add-a-server wizard, which screens get opened and which commands get run. Opening a link from a message reports whether this copy of the app could find what the link pointed at.
  • Turning usage sharing on is reported. Turning it off is not, and cannot be: the check happens immediately before anything is sent, so the moment someone declines is the moment nothing more is sent. The opt-out rate stays unknown rather than being obtained that way.

Fixed

  • Retrying a session that failed to set up now says whether it worked. The retry reported failure by throwing, which nothing was listening for, so a second attempt that failed left the session sitting on "setting up" with no error and no explanation.
  • Deleting an agent left its sessions counted as still running. The sessions were removed by the database itself, so nothing in the app was told they had ended.
  • A session that is un-archived can end again. Its first ending was remembered for good, so the second one — and every one after it — went unrecorded.
  • A remote agent whose host refuses the connection now says so, rather than reporting the generic failure every other fault shares.
  • Installing an update no longer reports that it both worked and failed. The success was sent before handing over to the installer, so an install that then failed to take was counted twice, in both directions.

Changed

  • Local-server mode now bundles and pulls switch-core 0.21.0 (was 0.19.0): the bundle pin / COMPATIBLE_SWITCH_VERSION is raised to the current core release.
  • Third-party dependency updates (#193).

[0.30.0] - 2026-08-24

Changed

  • Windows in-app updates now verify the installer's Authenticode signature before installing it, against SandboxAQ's certificate. The check was held off until a signed release confirmed the certificate's subject verbatim; a mismatched name would have blocked updates for every existing install rather than warning. The release build now fails if the signature and the name it ships for the updater ever diverge (CHOO-1468).

Fixed

  • The first-run Share usage data prompt no longer opens with focus on the consent toggle. Base UI focused the dialog's first tabbable element, which drew a highlighted band around the row that read as a pre-selected answer and vanished on the first click; focus now goes to the dialog itself, so nothing is pre-selected (CHOO-2344).
  • Error toasts no longer pile up or repeat on every refresh. The agent-icon backfill is reported once per server keyed by outcome — a signed-out or too-old server that could not store icons was raising a toast every few seconds for one standing condition — and toast now replaces a toast it already shows rather than stacking another copy (CHOO-2344).

[0.29.0] - 2026-08-22

Added

  • Switch Console now sends a small, fixed set of product events — the app launching, an agent being created, a session starting and ending, a server being added, a connector being installed — so we can see which features get used and where the app runs into trouble. Nothing is sent unless you agree to it, and the first-run prompt now defaults to off, because each event carries a random id for your installation.
  • Events go to a relay we run, which forwards them on. The analytics services therefore never see your network address, and the app itself ships no credential — there is nothing in it to leak.

Fixed

  • The OpenCode connector the app writes carries its own copy of the room-workflow skill, so it picked up the same wrong claim that a Telegram DM is adopted like Mattermost's. Corrected alongside the connector directory (CHOO-2314).

Changed

  • Local-server mode now bundles and pulls switch-core 0.19.0 (was 0.18.0): the bundle pin / COMPATIBLE_SWITCH_VERSION is raised to the current core release.
  • Every Docs affordance opens the published documentation at docs.flintai.dev instead of the GitHub repository, deep-linked to the page that answers the question: the sidebar and app-menu Docs entries and the Settings Docs tab go to getting-started, the remote-hosts hint and the hosting cards to Host remotely, the messaging-app cards to Messaging apps, and the Rooms card to Rooms and agents. The welcome footer's GitHub link keeps pointing at the repository (CHOO-2313).
  • The "How to set up {Platform}" link on the connect-messaging-app form opens the platform's published setup page — Slack, Mattermost, Discord, Microsoft Teams or Telegram — rather than the repository's markdown, with the messaging-apps index as the fallback for an unrecognised platform (CHOO-2313, unblocking CHOO-2269).
  • The Share usage data prompt and its Settings row default to off, and say plainly that a random id for the installation is part of what is sent. The 0.27.2 entry below describes a payload of anonymous counters carrying no identifier of any kind; that was the shape of a feature that sent nothing, and this replaces it. Anyone who had already agreed under the old default keeps that answer — it is not silently reset to a refusal.

[0.28.0] - 2026-08-21

Added

  • Connecting a Slack app can enable per-agent user groups — the mentionable @name handles Switch mints for each agent — from the Connect-messaging-app dialog (CHOO-2277).

Changed

  • Local-server mode now bundles and pulls switch-core 0.18.0 (was 0.17.3): the bundle pin / COMPATIBLE_SWITCH_VERSION is raised to the current core release.

Fixed

  • The New-agent modal's Cancel / Add-agent footer no longer overlaps the "you have not linked a messaging account" callout at the bottom of the form, which had half-hidden its Open-Messaging-apps link (CHOO-2243).

[0.27.4] - 2026-08-20

Fixed

  • Search results show an agent's own icon rather than its provider logo, so two agents on the same provider are no longer the same picture; the provider mark now follows the "Agent type mark" toggle in both the sidebar and the palette (CHOO-2203).
  • A long instruction no longer takes over the agent page. Multi-line boxes cap at roughly a dozen lines and scroll inside themselves, and the agent's instructions field gains an Expand control for when you want the ceiling lifted (CHOO-2203).

[0.27.3] - 2026-08-19

Added

  • Set an agent's instructions — what it is for, addressed to the agent itself — as a first-class field edited from its page, applied to whichever mechanism its provider uses (a Claude repo-agent, a Codex or OpenCode launch profile) rather than being a per-provider setting only (CHOO-2228).
  • Build switch-core from your checkout for the managed local stack. A dev build launched from a Switch checkout can run switch/gateway/setup from the working tree (up -d --build) instead of this build's pinned GHCR images, via a toggle on the local server's page. Dev builds only, and only when an ancestor of the app's directory carries the Dockerfiles and core/pyproject.toml (#258).

Changed

  • Local-server mode now bundles and pulls switch-core 0.17.3 (was 0.17.2): the bundle pin / COMPATIBLE_SWITCH_VERSION is raised to the current core release.

Fixed

  • CLI first-run prompts that leave a spawned session alive but never answering are now cleared or detected before the session starts — Claude Code auto-trust (silently broken since a config-key rename), its first-run setup wizard, and the bypass-permissions warning (CHOO-2213).
  • @mentions in the embedded Mattermost view are readable in the light theme again — the link/mention colour pointed at the text-selection-highlight background and rendered at ~1.4:1 on the white channel (CHOO-2203).
  • A newly onboarded agent's default Mattermost rooms show in the sidebar without waiting for a background reconcile to happen to fire (CHOO-2203).
  • Sign-in details for a managed server's bundled chat now stays open. The dialog was rendered inside the messaging app's dropdown menu, so the same click that opened it closed the menu — and the menu took the dialog with it, before anyone could read a credential off it.
  • Deleting a Switch server no longer leaves the app pointed at it. The deleted server's id survived in the saved page state, so the app could return to a page for a server that no longer exists and keep asking its gateway for a sign-in configuration nothing could answer. The page now falls back to home and the dead id is discarded.

[0.27.2] - 2026-08-18

Added

  • Switch Console now asks, the first time you open it, whether you are happy to share anonymous usage data, and states plainly what that would and would not include. The choice is saved per install and can be changed at any time from Settings → General (CHOO-1955).
  • The app still sends nothing: there is no telemetry to consent to yet. What this adds is the control and the contract — a single gate that any future collection must ask before sending, which fails closed, so a fresh install that has not reached the prompt sends nothing regardless of how the toggle reads. Because the toggle defaults to on, console/AGENTS.md now pins what a payload may ever contain: anonymous counters, and no identifier of any kind.
  • A signed, notarized Intel (x86_64) macOS build. macOS now builds per-arch on native runners, so an Intel Mac has an installer again after the app went Apple-silicon-only; the auto-update channel manifest is written per arch so the two builds no longer overwrite each other's update feed (CHOO-2195).

Changed

  • Windows builds are now signed with Azure Trusted Signing, so the installer and app no longer trip SmartScreen's unknown-publisher warning (CHOO-1468).
  • Local-server mode now bundles and pulls switch-core 0.17.2 (was 0.17.1): the bundle pin / COMPATIBLE_SWITCH_VERSION is raised to the current core release.
  • The add-agent addressing control is renamed "Who can talk to your agent", and its Settings section now unfolds itself when the owner-only default would reach nobody, so the warning that the agent will answer no one on that messaging app is visible instead of hidden in the fold (CHOO-2173).

Fixed

  • A directory configured by the connector's configure skill is recognised as an agent again. Detection read the identity only out of .claude/settings.local.json, which the skill deliberately no longer writes into — writing two of the three SWITCH_* values there is what broke standalone sessions. It now falls back to the credentials store, and names no agent when the store holds several, since choosing between them is select_agent's job and not detection's.
  • An auto-started session now receives the message it was started for. Its room connection opens before the terminal is spawned, so the replayed trigger had nowhere to go and — unlike the other wait paths — never retried: the agent booted, greeted, and never answered. It now retries until the connection closes, and a pty accepts injected input only once the runtime has delivered the session's opening prompt, so the trigger can't be appended to that prompt and sent as one (CHOO-2173).
  • The agent tree no longer highlights every copy of an open room, and the titlebar breadcrumb shows the agent's own icon rather than its provider logo (CHOO-2173).
  • User-facing error messages across the app were audited and rewritten to say what happened and what to do, rather than surfacing raw internal errors (CHOO-2060).
  • Two Switch Console installs sharing a host no longer destroy each other's agents. Per-agent credentials live at .switch/agents/<name>.json, keyed by name alone, and each install's uniqueness check only sees its own database — so an agent created with a name another install had already used in that directory overwrote its credentials file. Because the write merged into what was there, the result was one file carrying the new agent's identity and the old one's remaining keys: the displaced agent's sessions authenticated as the new agent rather than failing, and its API token, issued once, was gone. Creating, provisioning and renaming an agent now refuse a name whose credentials in that directory belong to a different Switch server, and say which one. Refusal happens before the identity is minted, so nothing is left stranded on the server.

[0.27.1] - 2026-08-17

Changed

  • Local-server mode now bundles and pulls switch-core 0.17.1 (was 0.17.0): the bundle pin / COMPATIBLE_SWITCH_VERSION is raised to the current core release, so a managed stack picks up the Slack avatar-background fix (#239).

[0.27.0] - 2026-08-16

Added

  • Give an agent a custom icon. The add-agent dialog and an agent's settings carry an icon picker — set, change, or clear it; cleared means an icon generated from the agent's name. Icons render in the sidebar, the agent pages and dialogs, and existing agents are backfilled on first run (CHOO-2171).

Changed

  • Local-server mode now bundles and pulls switch-core 0.17.0 (was 0.16.0): the bundle pin / COMPATIBLE_SWITCH_VERSION is raised to the current core release.

Fixed

  • Collapse all collapses again. It cleared state the redesigned sidebar no longer reads, so nothing moved; it now names the rows to collapse. Clearing that state had also been dropping Next/Previous Session's list of visible sessions, which stops (CHOO-2173).
  • A sidebar session row's actions menu acts — Delete and Archive ran but the row's own click handler reopened the session on top of them, so they looked inert (right-click was unaffected because it is a sibling of the row). The guard now lives in the shared popup (CHOO-2173).

[0.26.0] - 2026-08-16

Changed

  • A visual design pass across the shell. The window is now one graded material surface with the sidebar sitting on it and the main panel floating above as an inset card with a hairline and a soft shadow, using the design spec's surface/material tokens and corner radii. Home is deliberately left for a later pass (CHOO-2164).
  • A server is now a workspace, and you switch between them. The list of servers that sat in the sidebar has been replaced by a switcher at the top, showing the one you are in and opening onto the rest — each with where it runs and whether it is reachable. Everything below it already belonged to that server; now the sidebar says so. Under the switcher is that server's Home, which is the page you used to reach by clicking the server in the list, and the titlebar names the path to it (CHOO-2158).
  • The session list is a section of its own, headed Sessions, with the grouping control labelled as what it is — group by By Agent or By Room — rather than sitting alone at the top of the tree where it read as the sidebar's whole navigation (CHOO-2158).
  • The welcome screen is reachable from the switcher as About Switch, as well as from the Switch mark at the bottom of the sidebar. With no servers added, the sidebar is just Add a server over the setup checklist (CHOO-2158).
  • A server's Home page reads as a dashboard rather than a stack of cards. It opens with the server, where it runs and who you are signed in as, then how many agents, rooms and messaging apps are on it, and gives the rest of the page over to the things you act on (CHOO-2158).
  • Messaging apps are a table. Which account is you, whether Switch may create channels there, and the app's actions are columns you can read down rather than facts to be opened one row at a time. Turning channel creation on or off is now in the row instead of behind its menu, and a platform that cannot create channels at all says so rather than showing the switch off. Unlinking stays in the menu — it is irreversible, and the row is where you press to change an account (CHOO-2158).
  • The managed stack has a section of its own with a Restart, and shows its live output, which until now was only visible while adding a server. An available switch-core update is announced at the top of the page rather than inside that section, and Reset sits alone at the bottom, away from Start and Stop (CHOO-2158).
  • The Mattermost sign-in details Switch Console generates open in a dialog from the app's row, rather than expanding underneath it (CHOO-2158).
  • A server's agents and rooms each have a page. Under the switcher, next to Home, Your Agents lists every agent registered on the server with the provider it runs, where that is, and how many of the server's rooms it is in; Your Rooms lists the rooms with the messaging app each is bridged to and how many agents are in it. Both start a session or add a member from the row, and carry the same actions the sidebar rows do. Until now the only way to see either as a list was to scroll the sidebar tree (CHOO-2158).
  • An agent's page opens with the agent. Its mark, its name, the server it is registered on and what it runs, with Create Session alongside — where the page used to begin at a list of sessions, with the agent named only in the titlebar. Sessions are framed as a list, each saying which room it is talking in, or that it is in none (CHOO-2158).
  • The titlebar says where you are, in full. A session now reads agent / room / session rather than naming its directory and itself, and a room is its own heading with its messaging app's mark. An agent's page carries the same connection status and actions menu its server's pages do (CHOO-2158).
  • Agents are described by their provider's name — Claude Code, Codex — rather than the claude and codex they are keyed by (CHOO-2158).
  • Every session in the sidebar has an actions button. Pin, rename, archive and delete were reachable only by right-clicking, which nothing in the interface said you could do; they are now on a menu in the row as well. Both menus are built from one list, so an action cannot arrive on one and be missing from the other (CHOO-2158).
  • Expanding a row in the sidebar is now the chevron's job alone, and the chevron has moved to the end of the row, appearing on hover with that row's other actions. Opening an agent or a room no longer unfolds it, so reading one thing does not rearrange the tree around it, and the agent's provider mark stays visible instead of turning into a chevron under the pointer. A row with nothing beneath it has no chevron at all (CHOO-2158).
  • A room in the sidebar no longer counts its agents. The pair of numbers on the row — agents this computer runs, agents it does not — sat where the expand control now goes and told you little. The sidebar lists the agents it can actually open; a room's full membership is a column on Your Rooms, which is the page for reading rooms as a list (CHOO-2158).
  • Anything you can click now says so under the pointer. Buttons, sidebar rows, menu entries and tabs show the hand cursor; Tailwind's reset had left them all on the arrow (CHOO-2158).

Fixed

  • Adding your first server left the app with no workspace selected. Nothing chose it, so the sidebar stayed on "Add a server" until the next launch (CHOO-2158).
  • Removing the server you were in could leave the app with no workspace at all — no switcher, no destinations, no sidebar tree — because the remembered server was still named as the active one. A remembered server that no longer exists now counts as no choice, and the app picks one (CHOO-2158).
  • Removing an agent from its own page removed the first agent in that directory, which is a different agent whenever a directory holds more than one (CHOO-2158).
  • The empty session list offered to "spawn a claude session" whatever the agent actually runs (CHOO-2158).

[0.25.0] - 2026-08-15

Added

  • You can tell Switch which messaging-app account is yours. A new dialog searches the workspace's own user directory — Slack, Mattermost — and links the account you pick to your Switch login. It searches the platform rather than Switch's record of who has spoken, so you can find yourself in a workspace you have never posted in, and it shows the display name, handle and email so you can tell two similar accounts apart. An account someone else has already linked says who — linking is not exclusive, so that is information about who else is recognised on it rather than a lock — and one you have linked yourself offers to unlink instead of linking it twice. A platform with no searchable directory says so — and says a message has to arrive first — instead of showing an empty list (CHOO-2137).
  • The dialog is offered as step 2 of connecting a messaging app, straight after the connection succeeds, because that is the one moment the workspace is on your mind — except on an app whose directory cannot be searched, where it is not offered at all. On Telegram, Switch can only name people who have messaged it, and nobody has messaged a connection made a second ago: the search was guaranteed to come back empty, which reads as "you are not in your own workspace" rather than "not yet". Linking waits for the server page, where the warning that you own an agent nobody can reach is what prompts it, and by then someone has messaged the app and there is a name to pick. The search prompt no longer promises a "workspace directory" on a platform that has neither. It is skippable, and linking lives on the server page for later: Messaging apps lists one row per app — the app, the account on it that is you, and what you can do with it — where your handle is the button that changes it and a Link button stands in when there is none, opening the dialog already pointed at that app. So you can see at a glance where you are recognised and where you are not, without a separate list to reconcile against. Linking is not an admin action: a member of the server sees the same card and can link on any app in it, and only the Connect button is withheld.
  • Each messaging app carries its own menu, holding the actions that are rare or destructive: unlinking your account, whether the app may create channels, and disconnecting it. The row itself is left saying one thing — which account here is you — in one control whether or not there is one. Unlinking in particular is off the row: it used to be a bare cross sitting beside the handle, one mis-click from the button you press to change an account, with no confirmation between. The Default badge also sits against the app's name now rather than drifting into the middle of a wide row.
  • The server page warns when an owner-only agent cannot recognise you, and only then. One line at the top of Messaging apps names the apps you have not linked — shown when you own an agent on that server whose rule admits its owner, and never otherwise. Link everywhere, or own no such agent, and there is nothing to see: a warning shown to everybody teaches people to ignore it.
  • A new agent now answers only its owner. Agents used to be created open to everyone in every room; the add-agent dialog now defaults to Only me. The rule names the owner rather than a list of identities, so it survives connecting a new workspace or the agent changing hands. Only me and my agents is one step away when an agent of yours has to hand this one work. Existing agents are untouched.
  • A messaging app can be disconnected from the server page, by an admin, on the same Messaging apps row that connects one. It is not a pause: the server deletes every Switch room on that app before removing it, so the rooms and their conversations go with it, and an app another admin has already removed says so rather than reporting a disconnect that never happened (CHOO-2137).

Changed

  • Who can send instructions to an agent is one question with four answers — Only me (default), Only me and my agents, Anyone, or Custom rules — rather than an open / restricted switch above a rule builder. The two owner-scoped answers differ in one thing: whether agents you own may hand this one work. Anyone drops the policy entirely; Custom rules opens the rule editor, seeded from whichever answer the agent was on, so nothing has to be rebuilt to add one exception. Rules built by hand are kept while the chooser is on a shortcut, and a policy too specific for any of them reads back as Custom rules rather than being flattened into one. The same control is on the add-agent dialog and on the agent's Settings tab, so a policy is changed the way it was set. The box shows the answer that was picked, rather than the word stored behind it.

  • The add-agent dialog offered "Advanced configuration" twice for Claude Code. A provider keeps its per-agent settings in one place — a repo-agent definition or a launch profile — and the agent's Settings tab picks between them by asking which; the creation form rendered both sections instead, and the launch-profile one reads a "the fields, from wherever they live" call that falls back to the definition fields. So the same settings appeared in two identical boxes writing to two different places. It now picks the same way the Settings tab does.

  • An agent's Settings tab no longer boxes Advanced configuration and Who can send instructions in borders of their own. Every section on the page is already a section; drawing two of them again was one frame too many.

  • You and your agents appear in the rule editor's own pickers, as Me in the Users list and My agents in the Agents list, rather than as a checkbox off to the side. They read as what they are — two more senders to admit — and they compose: "me and Alice" is one rule. Setting either list to Any or None clears the matching entry, so a rule cannot say "no humans" and "the owner may" at once.

  • An owner rule can only recognise an owner who has linked their messaging account, so the addressing editor warns — with a button into the linking dialog — when the signed-in user has linked none. A privacy control that silently admits nobody is the failure this exists to prevent.

  • The linking dialog is opened on one workspace rather than asking which. It is titled for the platform it is linking on — "Link your Discord user account" — and its search names the workspace, not the platform, since two workspaces on the same platform can be connected and only the name tells them apart.

Removed

  • The "Give Feedback" feature — modal, Help menu entry, command-palette command and its event — along with the hardcoded third-party Discord webhook it posted to. The webhook arrived with the initial emdash import and is already revoked upstream. Help → Report Issue remains the way to send us something; it opens an issue on this repository (CHOO-2040).

Fixed

  • A session on a remote agent now opens at the size of the pane it opens into, instead of a fraction of it that only corrected itself when the window was resized or the session was switched away from and back (CHOO-2066). A remote session opens its terminal over SSH, and the renderer mounts and measures its pane partway through that: the measurement arrived after the spawn size had been read and before there was a PTY to resize, so it was discarded and the session kept the 80x24 it was spawned with. Measured dimensions are now kept whether or not a PTY is live yet, and applied to the PTY when it registers.

[0.24.0] - 2026-08-14

Added

  • An OpenCode agent can be configured like a Codex one. Its model, reasoning variant, temperature, top-p, step limit, web search and instructions can be set when the agent is created and changed afterwards from its Settings tab, and a change applies to the next session — or to a running one with Restart, which resumes the conversation.

    OpenCode's settings are not Codex's, so the per-agent configuration an agent stores is now keyed by what its own provider offers rather than by a fixed list. Existing Codex agents keep their settings; nothing needs re-entering.

    Its instructions are added to OpenCode's own, the way an AGENTS.md is, rather than replacing them.

  • You can pick a model instead of remembering it. Switch Console asks the machine an OpenCode agent runs on which models it offers and lists them as you type, grouped by provider and annotated with the reasoning variants each accepts. Typing something not on the list still works — it is a shortcut, not a restriction.

  • The model and reasoning fields check themselves against the agent's own host. A model name that host does not have is flagged as you type, and the reasoning variant becomes a menu of what the chosen model actually accepts — greyed out, with a reason, for a model that has none, as local models generally do. Both were places OpenCode would otherwise accept a value and silently never apply it.

    The check warns rather than blocks, since the list is a snapshot and a model can appear a moment later. If the host cannot be reached, or OpenCode is not installed on it, the fields say so and go back to plain text rather than flagging everything as wrong.

    An OpenCode agent can also be pointed at a local model — define the provider once in your OpenCode config and set the agent's model to it. The new utility-model setting is worth setting too if the point is to keep everything on one machine: it is what OpenCode uses for background work like naming a conversation, which otherwise goes wherever your own config sends it.

  • Telegram brand icon, platform label and setup-guide link, so Telegram-bridged rooms show the "open channel" button and the attach form links the right guide (CHOO-1686). Telegram bot tokens are also redacted from the diagnostic logs.

Fixed

  • Opening a room from a deeplink expands the sidebar groups hiding it (CHOO-1686). The reveal ran once, before the room a session belongs to had loaded, and never again — so the view routed correctly to a row that stayed collapsed. Affects any bridge, not only Telegram.
  • Two Switch Console installs sharing an agent host no longer trade the sidecar back and forth when they are on the same release but carry different builds — the everyday case for dev builds, and the half of the shared-host problem that version ordering could not reach. Each install now mints a deployer identity and stamps it on whatever sidecar it deploys, so an install can tell its own build from another's without relying on the version string. When the versions are equal and the builds differ, the install that got there first keeps the sidecar.
  • The agent's Settings tab says so: the sidecar reads Another install's build, names who deployed the running one, and offers Restart — a deliberate takeover — instead of an Update that both sides would keep clicking at each other. Replacement is otherwise unchanged: an older sidecar is still replaced whoever deployed it, and a rebuild of your own is still picked up.
  • Search results no longer read as though agents, sessions and commands were servers. Those three had no heading of their own, so they were drawn after the last group on screen — usually "Servers" — and appeared to belong to it. Every result now sits under a heading that names it.
  • The server running on this computer shows a laptop everywhere — the sidebar and search as well as the dialog that adds it. The sidebar and search each chose the icon themselves; they now share one rule, so the next change lands in all three at once.
  • The local server's setup log now fills as the work happens rather than all at once at the end. Every line Docker printed was applied to the screen on its own, and a pull narrates faster than the UI can redraw — so the renderer spent the whole install rebuilding the list instead of painting it. Lines are applied in batches now. The same fix covers a remote host's setup.
  • Setting up the local server no longer looks stalled before Docker says anything. The output panel appeared only once the first line arrived, so the seconds Docker spends resolving the registry showed as an empty dialog; the panel is now there from the start and says what it is waiting for.
  • The command palette no longer opens onto a "Notifications" list of sessions and rooms. It is a search field; the first thing under it should be what you searched for.
  • Creating a room offers only the agents this Switch Console registered. The server answers with everyone on it, including agents belonging to another install — which this app cannot show under a room or drive.
  • Creating a room switches the sidebar to its Rooms list and selects the new room. It was already opened in the main panel, but the sidebar stayed on Agents, which does not list rooms — so the room you had just made was nowhere in it.
  • Installing a Switch connector now finishes the onboarding step it belongs to. The list of agent types you can onboard is cached separately from the agent's own status, and the install refreshed everything but that — so the step stayed unticked, and because the checklist locks each step behind the one above it, the rest of onboarding stayed greyed out with nothing explaining why.

Changed

  • An update to an agent's own CLI is no longer reported as though something were wrong. It never gated anything, but an amber badge ahead of "Installed" read like a fault on an agent that worked, and it was the nearest explanation for onboarding appearing stuck. A newer CLI is now mentioned only on the agent's own page, in plain text with the command that installs it.
  • The Switch connector's own updates keep their badge and gain a name — "Connector update" rather than "Update available". That one is worth acting on, and the two used to share a badge that could not say which was behind.

[0.23.0] - 2026-08-14

Added

  • OpenCode is now a supported Switch agent type, locally and on remote hosts. An OpenCode session can be onboarded as a Switch agent, join rooms, take injected prompts, and be reset, compacted or interrupted from a room. It reports working and completed rather than sitting on "awaiting input" for its whole life, and names the tool it is running on the bridged message.

    OpenCode has no plugin marketplace to install a connector from, so Switch Console writes one: the Switch MCP server is registered in OpenCode's global config as a local server, which lets the runtime inherit its credentials from the session environment rather than having a token written to disk, and the room-workflow skill is written to OpenCode's global skill directory beside it, so a session gets the instructions along with the tools. Install, update and uninstall are on the agent's card in Settings → Agents, and on a remote host in that host's setup, exactly as for the other agent types.

    OpenCode agents previously registered as Claude Code, so an operator asked to start one by hand was told to run claude. They now register as themselves.

  • A Codex agent's configuration can now be changed after the agent is created, from the Advanced configuration section in its Settings tab — the same section Claude agents already had. Model, reasoning effort and instructions were previously write-once in the add-agent dialog, with no way to edit them (CHOO-1985); verbosity, reasoning summary and web search are newly exposed, and reasoning effort gains none. Each option was checked against the Codex binary's own config validation rather than assumed.

  • Leaving one of those fields blank is not the same as choosing its default: blank omits the setting entirely, so your own ~/.codex/config.toml still decides. That is why the on/off settings offer Default/On/Off rather than a checkbox, which has no way to say "leave it alone".

  • Codex reads these values only when a session starts, so a save applies to the next session. A session already running is named in the section, with a Restart that resumes it on the new configuration. Clearing every field removes the agent's launch profile rather than leaving it orphaned.

Changed

  • Advanced configuration is one editor for every provider rather than one per storage mechanism. Providers keep these settings in different places — a repo-agent definition for Claude Code, the launch profile for Codex — and the main process now routes the read and the write, so the difference is no longer visible as the shape of the Settings page.
  • Remote hosts are a tab of Settings again rather than an entry in the left sidebar. Older saved navigation state, and the host page's way back, both land on the tab.
  • The setup checklist on the welcome screen can now be dismissed. It shares the sidebar checklist's setting, so it goes away in both places at once and comes back from Settings → General.
  • Settings descriptions for an agent's own options are cut back to a line, with the detail behind the ⓘ beside each one.
  • The agent providers list no longer separates out a "recommended" group; each filter is a single list.
  • The sidebar's + opens a menu offering both "Add an agent" and "Create a room", instead of the single action matching the list you were looking at.
  • The remote hosts page explains what a host is for, and links out to the cloud-hosting documentation.

Removed

  • A batch of interface settings, along with the toggles for creating a branch and worktree, including issue context, and OS notifications. Each was already on by default and is now simply always on, so nothing changes for anyone who had not turned one off. Several of them — the left-sidebar line-change and PR-status toggles, the context bar, and confirm-on-tab-close — turned out to control nothing at all: no code read them.
  • Keyboard shortcuts are no longer rebindable from Settings, and open-in tools can no longer be hidden. The shortcuts themselves are unchanged.

Fixed

  • Add Agent no longer stalls with an unset agent type once more than one Switch connector is installed. The form auto-selected a type only when exactly one was available, so a second connector left it blank — and since the directory scan, the onboard-existing list and every submit button are gated on a chosen type, the rest of the dialog stayed inert behind a control that did not read as required. It now selects the only usable type, or the configured default agent when that is usable on the machine being targeted, and otherwise leaves the choice explicit. The rule no longer depends on how many agent types exist.
  • The unread tally injected into a session no longer claims to count "since your last read_context". Nothing here can observe a session reading, so the tally is cleared per delivered line; it now says so, rather than inviting an agent to read its absence as proof it is caught up.
  • The OpenCode icon no longer dwarfs the other agent icons. Its artwork is full-bleed and 4:5 where the others are glyphs on a square, and it carried its own width/height — which an inline SVG applies in preference to the size of the box it sits in. It is now square, unsized, and inset to the margin the others have.
  • A file-based Switch connector reports the version its own directory declares rather than the app's. The OpenCode connector is versioned and released as its own artifact, so a card reading 0.22.0 beside two connectors reading 0.9.1 and 0.3.1 named nothing the connector declares. It also means an app release that leaves the connector alone no longer offers an update that would rewrite identical bytes.
  • A session that fails to be created no longer marks its agent with a failure badge that cannot be cleared. The failed session exists only in the app — the server rejected the create — and nothing removed it: reloading only adds sessions, deleting one asks the server first and restores it when that fails, and the sidebar never listed it in the first place, so the badge pointed at a session there was no way to reach. It is now dismissed by clicking the badge.

[0.22.1] - 2026-08-12

Changed

  • Local-server mode now bundles and pulls switch-core 0.13.2 (was 0.13.1): the bundle pin / COMPATIBLE_SWITCH_VERSION is raised to the current core release.

Fixed

  • Windows builds now actually work — path handling, process spawning, the PTY, executable resolution, and Docker / managed-server invocation are corrected for Windows, so the x64 build introduced in 0.22.0 is functional rather than nominal (#207).
  • Settings is reachable again. A retired Settings tab left in the view registry could be selected and then render nothing, stranding the whole Settings page; the retired tab is now handled and navigation falls back to a valid tab (CHOO-2106).

[0.22.0] - 2026-08-12

Added

  • Linux arm64 desktop artifacts are now built and published alongside x64 — AppImage, deb, and rpm (#202).
  • Windows x64 releases are now built and published (unsigned) (CHOO-1468).

Changed

  • The Codex session runtime version is now derived from the artifact registry rather than a hand-maintained SWITCH_AGENT_RUNTIME_VERSION constant; the constant and its parity test are removed (#198).

[0.21.0] - 2026-08-11

Changed

  • The private-repo machinery is gone now that the repository and its packages are public (CHOO-2023). The updater no longer gates on a gh token or a private feed, so every user receives updates and the auth-required state is removed; managed-server image pulls drop docker login ghcr.io and the remote .ghcr-token forwarding; remote-host setup drops its interactive gh:auth step (a persisted plan still carrying a gh-auth step drops it on read); agent onboarding no longer withholds every agent type when gh cannot reach a registry; and Switch no longer installs, probes, or forwards credentials to gh on a remote host at all.
  • Local-server mode now bundles and pulls switch-core 0.13.1 (was 0.13.0): the bundle pin / COMPATIBLE_SWITCH_VERSION is raised to the current core release.
  • Codex sessions launched by Switch Console now run @sandboxaq/[email protected] (SWITCH_AGENT_RUNTIME_VERSION, was 0.1.6), matching the runtime's move to public npmjs.

[0.20.0] - 2026-08-10

Added

  • The bundled Mattermost row on the server page now has a "Sign-in details" disclosure — server URL, username and password, each with copy-to-clipboard — so signing in to a managed deployment's bundled chat no longer means finding the generated .env on disk. The password is masked until revealed, and can be copied without revealing it. Values are read from what the deployment actually runs (the host port Switch Console chose and the password from the encrypted app-secrets store), fetched only on expand so the secret does not cross into the renderer for everyone who opens the page. When a value cannot be read the card names which one is missing rather than showing a placeholder that could be mistaken for a real password.
  • A per-host cap on how many remote terminals stay attached at once (remote.maxAttachedSessionsPerHost, default 4), exposed as an Interface setting. Beyond the cap the least-recently-viewed session on that host is detached — its agent keeps running in its tmux pane and still reports status, so only the on-screen terminal goes away.

Changed

  • The app is now called Switch Console (formerly "switchdash") everywhere a user reads it: the window and menus, the home-screen wordmark, the server/agent dialogs, the auto-session / auto-approve / sidecar settings, the version-drift warning, the command palette, remote-host setup, and the "Open in Switch Console" deeplink the bridges post into Slack and Mattermost. Storage and OS identity deliberately keep the switchdash name (app id, userData directory, switchdash:// scheme, npm packages, SWITCHDASH_* env vars) so an existing install updates in place rather than being stranded. The source tree moved from dash/ to console/ and the release tag prefix is now switch-console-v* .
  • Local-server mode now bundles and pulls switch-core 0.13.0 (was 0.12.3): the app's bundle pin / COMPATIBLE_SWITCH_VERSION is raised so a managed local stack runs the current core release.
  • Bump the Codex session runtime pin (SWITCH_AGENT_RUNTIME_VERSION) to 0.1.6 so Switch Console-launched Codex sessions run the published runtime.
  • Remote terminals now attach on demand instead of on provision. Opening a session on a remote host is what brings up its terminal, driven by which session is in focus (debounced so arrowing through the sidebar doesn't ask a slow host for a terminal at every step), with least-recently-viewed eviction under the per-host cap. Previously every known session on a host wanted a terminal the moment it was provisioned — on one host that was 51 terminals opened at launch. The agent runs on its VM regardless; sidebar status, room badges and deep links are unaffected because they come from hook events and the session poll, not from a PTY.
  • Search no longer returns filesystem results, and the per-location file indexer behind them is removed. File hits were appended after the relevance cut and the 30-item slice, so up to 20 irrelevant rows survived regardless of ranking; the row navigated to the session, not a file. Removing search's only reader of the index also drops the crawl it ran on every location and the debounced re-crawl on every file-watch event; an existing database sheds the now-unused index tables.

Fixed

  • Clicking a session on a busy remote host no longer freezes, and opening a remote session no longer lands on an empty pane or a "Terminal not attached" message that Attach could not clear. On one host, 51 sessions sharing a single SSH transport re-attached in the same instant whenever the transport recovered, re-saturating the tunnel on a ~16-minute wedge cycle; separately, a runtime reaching the attachment path at provision time knew neither its session nor that it was attachable, so every attach quietly did nothing. Remote attachment is now serialised per host, sidecar event relays are shared one-per-sidecar instead of one-per-session, provisioned sessions are made attachable, and a failed attach throws a real error instead of leaving a silent empty pane. A detached remote session now shows that its agent is still running rather than an indefinite spinner.
  • The sidebar's staleness banner no longer warns about servers you are not looking at. Room state was loaded and reported across every server at once while the sidebar only ever shows one, so a healthy selected server could be buried under warnings about unrelated ones. Room state is now scoped to the server on screen (cross-server search still loads every server, but only when opened).
  • The sidebar no longer presents unknown room state as fact. A server that could not be read kept its last-known rooms with no hint they were stale; a room whose name had not loaded showed a short id styled like a name; an agent whose membership lookup failed vanished from its rooms. These are now disclosed distinctly ("asked and failed" vs "not connected, never asked"), an unloaded name renders as provisional, and a retry is offered. A room's member count, list and invite picker are now one read so they can't disagree.
  • A server you are not signed into now prompts you to sign in instead of reporting "couldn't reach" it with a Retry that cannot work. Signed-out, dormant, and genuinely-unreachable are now distinguished, and only a real failure offers Retry. Agent names are also read from the local database instead of fetched per-row from the gateway, which could silently render a plausible wrong name on failure.
  • Selecting a session, agent or room from outside the sidebar (keyboard, command palette, notifications, menu, modals, deep links, or the view restored at startup) now scrolls the highlighted row into view, expanding whatever group hides it. Previously only the deeplink handler scrolled, and only for session rows. It scrolls only when the row is off-screen, never while dragging, and keys on the selection so a reorder can't yank the viewport.
  • The per-room channel icon, the room pane's "Open in …" button and the Slack bridge card's Open button now open bridged channels even without the native desktop app installed. They handed the OS a native mattermost:// / slack:// deeplink, which nothing answers when no desktop app is registered, so the click did nothing. Links are now translated to the channel's web address, which works either way. Opens now also report a failure to the person who clicked instead of discarding it.
  • The bundled chat's sign-in prompt no longer offers to sign in "on your phone". The managed stack publishes onto 127.0.0.1 and the URL shown is a localhost one, which no other device can reach, so it now says "on this computer" .
  • The server page no longer sits on "Checking sign-in options…" forever after connectivity returns. Which login methods a server offers was read once when the page mounted, and a read that failed was never retried, so a blip left the sign-in form unrendered until you navigated away or restarted. It now recovers on the signals connection status already used — the page's Refresh button, returning to the app window, and a remote host becoming reachable again. The Refresh button in particular re-checked only the connection status, so the one control a user would reach for did nothing visible.
  • A server whose gateway cannot be reached now says so in one line and retries on its own, instead of stacking three surfaces that each described the same failure differently: a red banner quoting a raw internal error (Error invoking remote method 'switchServers.getAuthConfig': …), a "Not signed in" row, and a sign-in panel that claimed to be checking options it would never receive. There is nothing to sign into while the gateway is down, so the sign-in form no longer appears at all. The underlying error goes to the console.
  • The sidebar no longer offers "Sign in" on a server it cannot reach, which was an action that could not work. Unreachable is now modelled apart from signed-out, so the rooms list also stops reporting an unreachable server as merely needing sign-in. A server whose data is unavailable shows a red dot rather than amber, for both causes: neither is a transitional state .
  • A directory that already holds agents for one Switch server can now onboard its agents to another. "Already onboarded" was judged per directory rather than per server, so every candidate was filtered out as a duplicate and the modal offered an empty list — of agents that existed, on a server that did not have them .
  • The sidebar no longer draws a directory's agents under a server they do not belong to. A directory resolved to a single server — whichever of its agents was returned first — and then every agent in it was rendered under that one, so onboarding a single agent to a second server appeared to drag the rest across with it. Nothing had been onboarded; each agent is now drawn under its own server, and a shared directory shows under all of them.
  • Onboarding an agent whose credentials belong to another Switch server no longer overwrites those credentials. The import path looked the existing identity up on the target server, did not find it, and fell back to minting a fresh one — writing it back over the same .switch/agents/<name>.json and silently breaking the agent the other server was still running. It now refuses and names the server the identity belongs to. Such agents are listed in the Add Agent modal but not selectable, with the reason shown; a plain definition carrying no Switch identity is still adopted as before.
  • The Add Agent modal no longer offers a permanently disabled button over a list it will not submit. A directory carrying another server's leftover .claude/settings.local.json took the detected-agent branch, verified that foreign agent against the current server, failed, and disabled the only button on screen — with the onboardable agents listed above it. The onboard action now takes precedence, and a detected agent belonging to another server says so .
  • The Add Agent modal no longer shows the create form, or the existing-agents list, before an agent type is picked. Both depend on the type — it decides which agents can be brought in and how a new one runs — so the form asked for a name, a description and a config, then ended at a button that could not be pressed. Everything below the directory now waits for the type, and the modal asks for it instead.

[0.19.3] - 2026-08-09

Added

  • Declares its sidecar-control range in the sidecar's ready file, and reads back whatever a running sidecar declares. A sidecar that declares nothing records as unknown, never as agreement.

Fixed

  • A gateway call to a managed server whose stack is stopped now fails with the lifecycle state, naming the server, instead of timing out and reporting Could not reach http://localhost:<port> — a local address that was never the problem. It is logged as the modeled state it is rather than as an RPC error, and the proactive session renewal no longer warns about the same absence a second time.
  • A managed server whose deployed switch-core version cannot be read is now surfaced as such, in the banner and the sidebar. It previously reported "no drift" — the identical result a healthy, in-step stack gives — so a failed probe rendered as a green server and the user was told nothing.
  • Starting a stack whose version cannot be compared to this build's pin now says so loudly. It passed in complete silence, in the same branch as a clean match, while carrying the same risk as a downgrade.
  • CI verifies the bundled standalone compose is in step with the repo's copy. Nothing checked before — the two could drift with only a comment asking nicely, which is how the sync script once silently stopped running .
  • The host-unreachable panel now also shows on a location that mounted while its host was up, not only one that never mounted. Such a location stayed ready and kept every control live over a dead SSH transport — the sidebar showed the trouble icon, the main pane did not; it restores itself when the host returns .

Security

  • Bump bundled dependencies for two advisories — DOMPurify SAFE_FOR_TEMPLATES bypass in RETURN_DOM mode (GHSA-crv5-9vww-q3g8), and brace-expansion denial-of-service via exponential expansion (GHSA-3jxr-9vmj-r5cp).

[0.19.2] - 2026-08-07

Added

  • Connect a messaging app (Slack/Mattermost/Discord/Teams) to a server from within switchdash — per-platform setup-guide links and a Teams icon; open a messaging app's workspace from the server page.

Changed

  • New macOS app icon (max.tam's mark); a dark mark for non-release (canary/dev) channels; dropped the leftover emdash DMG art/wordmark.
  • Bump the bundled switch-core for local managed servers to 0.12.3 (COMPATIBLE_SWITCH_VERSION).

Fixed

  • Restore sidebar drag-to-reorder for agents and rooms.
  • Set GATEWAY_PUBLIC_URL so "Open in SwitchDash" links work; redact the credential key names configs actually use; vendor the real Teams logo .

[0.19.1] - 2026-08-07

Changed

  • switchdash sessions report their latest-message anchor so the collaboration-bridge runtime indicator can follow the agent to the foot of the conversation.
  • Bump the bundled switch-core for local managed servers to 0.12.2 (COMPATIBLE_SWITCH_VERSION), so a fresh local stack pulls the latest switch-core and existing stacks flag the drift for a one-click update.

Fixed

  • Fix an import cycle that could leave the view registry half-built — a renderer crash ("Cannot access 'remoteHostsView' before initialization"), deterministic in CI and load-order-dependent locally.

[0.19.0] - 2026-08-07

Added

  • Remote-host onboarding rewritten as per-host pages with a staged, one-click-per-prerequisite setup: install each prerequisite individually with live progress, inline GitHub CLI sign-in, and agent creation gated on host readiness.
  • Surface and act on available updates on a remote host — detect Codex/CLI updates remotely, re-check one dependency at a time, and show which version an update brings, flagged in the sidebar.
  • Discover and onboard already-configured agents on a shared remote host (#130).
  • Detect switch-core drift on a managed stack and update it; flag drift on the sidebar server rows.
  • Command palette now searches across your agents and rooms — not just navigation — labelling what each result is, jumping into another server's scope, and adding an "Add Switch Server" command (#140).

Changed

  • Split an agent type into separate CLI and connector rows; drop the servers-sidebar text actions.
  • Pin switch-core 0.12.1 for the managed stack.

Fixed

  • Remote-host setup hardening: a host without the GitHub CLI no longer reports itself Ready; the readiness probe no longer races itself; host vs agent-type readiness are judged separately; stale plans rebuild; a failed update is no longer treated as a broken dependency; only CLI updates switchdash can actually perform are offered; agent creation refuses an unchecked type; the agent type clears when the run location changes; remote-host pages scroll; and a remote install no longer hangs on an unanswerable prompt.
  • Dialogs opt out of the window drag region; the alert action no longer overlaps its text.
  • Migration safety: register migration 0046 in the drizzle journal, fail loud when a migration isn't registered, and assert the migration runner's timestamp precondition.
  • Command-palette matching: a hyphenated query is no longer split into separate terms, matches no longer land mid-word, results rank like the sidebar, and the search index no longer treats item_type as content (#140).

[0.18.3] - 2026-08-06

Added

  • The Codex connector plugin now ships the Switch MCP server itself (mcpServers + env_vars name-forwarding), so Codex can be used with Switch outside switchdash — previously the server was registered only by a switchdash-written profile, and a hand-run Codex session had the room-workflow skill but none of the tools it describes. The plugin also auto-approves the Switch tools, which no approval_policy setting could do: measured against codex-cli 0.146.0, that setting does not govern MCP tool calls at all .

    ⚠️ Codex upgrades a plugin only when a user clicks Update in Settings and caches each version separately, so an install still on an older connector has no Switch tools until it is upgraded.

Fixed

  • Bypass permissions is applied to sessions again: the toggle is stored on the agent but every launch read a copy frozen into the session at creation, so changing it never affected an existing session — on restart or resume — while the settings copy promised the opposite.
  • Codex runtime status reports tool calls ("Running tool …") instead of sitting on "Working on it…" for a whole turn; no tool hook was registered for Codex, so nothing could ever produce an activity update.
  • In-app release links resolve again: the sidebar update indicator and the Settings Update card pointed at .../releases/tag/v<version>, which is not a real tag (the app is tagged switch-console-v<version>), so the user hit a 404; the release-notes fetch had the same broken tag and now also authenticates with the gh CLI token so it can read the private release feed instead of silently returning nothing (#134).

Changed

  • The per-agent Codex profile carries only model, reasoning effort and instructions. An agent that sets none of them no longer gets a profile at all .

[0.18.2] - 2026-08-05

Fixed

  • Client side of one-session-per-room enforcement: a session whose room the server takes away (or refuses with HTTP 409 because another session of the agent holds it) is now reported as roomless and is not reconnected under a room it no longer attends — the remote-session reconciler and sidecar no longer strand two sessions in one room (#109).

[0.18.1] - 2026-08-05

Changed

  • Bundled Switch agent runtime bumped to 0.1.5 (#127).

Fixed

  • A session's Switch connection id is now derived deterministically, so a supervisor/sidecar restart no longer strands a remote session with a stale connection id (#125).

[0.18.0] - 2026-08-04

Added

  • Codex is now a supported agent provider. Add a Codex agent from the add-agent modal (with a dedicated Codex config), connected to Switch via a new Codex connector plugin registered in the marketplace, running locally or on a remote host. CODEX_SANDBOX_MODE / CODEX_APPROVAL_POLICY drive Codex's sandbox and approval (validated — an unknown value fails rather than silently widening the sandbox); a Codex agent defaults to a codex.<repo>.<user> identity; and switchdash registers the Switch MCP runtime for Codex itself via a per-agent Codex profile (#91, #79).

Fixed

  • Codex session correctness: per-agent Switch credentials are written to disk for providers without repo-agents (so Codex sessions actually receive SWITCH_*); Codex room tracking follows connect_to_room mid-session; and Codex hook payloads post correctly (local Codex sessions had been posting empty bodies to a portless URL) (#79).

[0.17.2] - 2026-08-04

Added

  • Status-aware "update available" UX: the sidebar indicator shows the target version when an update is available, live percentage + transfer rate while downloading, a restart prompt once ready, and a warning tint on failure — and opens a panel (current → new version, the right action, a link to the GitHub release) instead of just jumping to Settings. User-triggered failures are toasted and the real error message is shown (#107).

Fixed

  • The onboarding/home page can drag the window again: the empty home surface is now a drag region (action buttons opted out), so the window is no longer stuck when Home is the active view (#106).

[0.17.1] - 2026-08-04

Added

  • Linux x64 desktop builds are shipped again (AppImage, deb, rpm), with a stable desktop-entry name so the launcher, icon and pinning behave; INSTALL docs and release notes now cover Linux (#104).

Changed

  • Release builds macOS and Linux artifacts in parallel (the GitHub Release is created in its own job), cutting ~5 minutes off a release (#104).

Fixed

  • A dropped-events "gap" no longer wakes an agent and spends a turn: switchdash defers the warning onto the next event it was already going to surface, instead of injecting an addressed prompt (#105).

[0.17.0] - 2026-08-04

Added

  • In-app Switch room creation: create a room from switchdash (name, description, messaging app, agents, optional instructions) instead of the operator web app — the picker offers only running bridges and every created room is bridged (#103).
  • Room-centric sidebar: a room tree listing rooms by membership and ownership (not only rooms with a live session), showing each room's members, managing membership (add/remove agents), starting a session in a room with the agent pre-chosen, and opening the room in its messaging app when the gateway supplies a deeplink; rooms sort and filter on their own properties (#103).

Fixed

  • Sessions launched for a room now declare that room when their connection opens — both sidebar-started and auto-spawned (messaging-app-addressed) sessions appear under the correct room immediately instead of sitting under "Unassigned" until the agent calls connect_to_room (#103).

[0.16.1] - 2026-08-03

Added

  • Local GitHub-auth detection in Switch setup: a requirement row reporting gh missing / not logged in / missing read:packages, with an inline device-flow login that requests the scope. Sessions that can't fetch the runtime now toast the reason instead of starting silently broken (#102).

Fixed

  • GitHub auth changes now take effect without restarting: the updater no longer leaves a boot-time token in GH_TOKEN (which shadowed the keyring for gh), and the scope check judges the active account only (gh auth status --active) rather than any known account (#102).
  • A dialog's own terminal now receives keystrokes (the gh auth prompt was silenced inside the Agent Settings modal); agent usage — not just install — is gated on GitHub access; managed servers pinned to switch-core 0.11.0 (was 0.8.1) (#102).

[0.16.0] - 2026-08-03

Added

  • Move switchdash and the remote sidecar onto the agent-bridge push stream: one shared connection per session claims its room server-side instead of scraping it from a hook, backed by the new @sandbox-quantum/switch-agent-runtime shared protocol client used by both switchdash and the connector plugin (#100).

Removed

  • First-run welcome page and residual Emdash artwork; first run now lands on the existing home empty state with an "Add Switch agent" action (#96).

[0.15.3] - 2026-07-31

Fixed

  • Unbreak remote hosts with IdentitiesOnly or restricted agent forwarding: the IdentitiesOnly key-filter now extends ssh2's BaseAgent so it's no longer silently discarded (which had broken auth for keychain-only keys and failed connects under ForwardAgent yes), and a host that refuses agent forwarding now degrades to "everything except forwarding" — with a warning — instead of failing every command (#98).

[0.15.2] - 2026-07-30

Added

  • Logging & diagnostics foundation across main, renderer, and the remote sidecar: structured, session-scoped entries with a per-launch run id; file logging defaulting to info (was warn); PII redacted only on export (not in the local log) so local logs are debuggable while exported logs stay as safe as before; on-demand sidecar log capture folded into bug reports; and boot/exit + enumerated error markers to pinpoint crashes and launch failures (#83).

Fixed

  • Stop a destroyed managed server's agents from polling forever: resetting a managed server now deletes its agents (the cascade lives behind the reset in the main process), room connections are keyed per session with ON DELETE CASCADE so orphaned entries are unrepresentable, and the boot sweep refuses to restore sessions for orphaned agents instead of hammering a dead endpoint (#95).

[0.15.1] - 2026-07-30

Fixed

  • Surface non-image attachments to dash-managed sessions: switchdash's own session-notification builder now downloads and annotates any file type (not just images) and names failed downloads, so agents running under switchdash are actually told about files posted in a room (#94).

[0.15.0] - 2026-07-29

Added

  • Built-in chat: switchdash renders each room's default bridge conversation inline (Mattermost) with per-room embed resolution, a chromeless guest view, theming, and link/deeplink routing; Slack rooms get a deeplink and agent-only rooms a stated reason (#84).

Fixed

  • Gate remote managed servers on host reachability, so an unreachable host no longer shows a stale green "Running" card or a spinning sign-in — the server page shows a "host unreachable" state and disables actions until it recovers (#85).

[0.14.3] - 2026-07-28

Fixed

  • Model remote-host reachability as a first-class per-host state machine, so switchdash stops the unbounded reconnect/reconcile retries against an unreachable host (one bounded backoff probe per host), persists the state across restarts, and surfaces a clear "host unreachable — work paused" panel with a global Retry instead of raw SSH errors (#82).

[0.14.2] - 2026-07-28

Fixed

  • Key the remote session reconciler per-sidecar instead of per-location, so sessions from multiple agents sharing one working directory all get adopted into the sidebar (#81).

[0.14.1] - 2026-07-27

Fixed

  • Reap sidecars orphaned when their tmux session name changes across upgrades, instead of leaving duplicates running for the same agent directory.

[0.14.0] - 2026-07-27

Added

  • Remote sidecar now persists durable state and versioning, so running sessions survive sidecar restarts, upgrades, and token rotation.

Fixed

  • Client no longer deletes healthy sessions when it briefly can't reach the sidecar; fixes cross-build kill loops and deaf clients after a sidecar restart .

[0.13.1] - 2026-07-26

Fixed

  • Fix slow boot caused by the agent-storage migration, and restore sessions that vanished after the 0.13.0 upgrade.

[0.13.0] - 2026-07-26

Changed

  • Removed the subagent concept — agents are now flat and repository-defined; existing subagents migrate forward automatically on first launch.

Fixed

  • Honor the bypass-permissions toggle for remote agents and clear a stale session guard.
  • Skip the connection-status probe for a stopped managed server.

[0.12.0] - 2026-07-24

Added

  • Configure per-agent addressing policy from the app and gateway.

Fixed

  • Data-entry modals no longer dismiss on outside-click, so in-progress input isn't lost.

[0.11.4] - 2026-07-24

Added

  • Reset a remote agent — kill and reset all of its tmux sessions.

[0.11.3] - 2026-07-24

Changed

  • Sidebar and session UX polish.

[0.11.2] - 2026-07-24

Added

  • Delete and rename connected Switch servers.
  • Per-agent bypass-permissions setting — defaults off, on for remote agents (#57).

[0.11.1] - 2026-07-24

Added

  • Agent delete now tears down its credentials and optionally deletes the agent from Switch.
  • Agent error indicator with a retry button.

Changed

  • Discord bridge polish: deeplink redirect, room icon, and outbound image relay .
  • Relicensed to Apache-2.0 + Commons Clause, with a CLA gate for contributions .

[0.11.0] - 2026-07-20

Added

  • Managed Switch servers can now run on a remote host over SSH — the app provisions the Docker stack remotely, with port-forwarded access, alongside the existing local mode.

[0.10.1] - 2026-07-20

Added

  • Silent session token refresh — sessions renew before the 24h expiry instead of bouncing to sign-in; the managed local server is always-signed-in, and its gateway web page opens pre-authenticated in-app.
  • Editing a server's API URL cascades to its member agents' configs .

Fixed

  • Remote sessions recover their room connection after an app restart or machine sleep.

[0.10.0] - 2026-07-20

Added

  • Local-server mode: run a managed Switch stack via Docker straight from the app — pulls the versioned standalone compose artifact, provisions env/secrets/ports, and monitors health.

Fixed

  • Injected prompts are always bracketed-pasted, so prompts containing @ no longer swallow the submit.

[0.9.2] - 2026-07-19

Added

  • Sidebar agents are labelled by their registered Switch name.

[0.9.1] - 2026-07-19

Changed

  • Replaced the workspace/project abstraction with first-class Locations — agents attach directly to a location (working directory).
  • Collapsed sessions to a single conversation each and simplified session/conversation management throughout.

Existing databases migrate forward automatically on first launch (schema migrations 0031–0036, including a locations backfill).

[0.9.0] - 2026-07-17

Added

  • Newly created subagents default to auto-session on.

Removed

  • Removed the telemetry/analytics stack entirely — the app ships no tracking or phone-home behavior.

Fixed

  • tmux mouse scroll works again (set-option target parsing).
  • Switch setup re-points stale plugin-marketplace sources after the repo move and surfaces failed refreshes instead of silently skipping.

[0.8.8] - 2026-07-15

Changed

  • Updated Electron to 40.8.5 and drizzle-orm to 0.45.2.

Fixed

  • Switch settings writers no longer clobber settings.local.json when reading the existing file fails (local, remote-SSH, and plugin-fs providers).

Desktop-app releases predating this changelog live in the git log and in the per-release notes on their GitHub Releases (switch-console-v* tags).


agent-controller

The headless agents controller, switch-agent-controller (console/packages/agent-controller/), released on switch-agent-controller-v* tags; see RELEASING.md.

[Unreleased]

Added

  • Tells Switch Console when a session started. Once Switch has taken a session host's start report, the controller Console runs on its own computer passes the session's provider and how it started to Console over its IPC channel, for Console's usage data. A controller with no parent listening, on an SSH host or run as a service, sends nothing.

  • Waits its turn when Switch is busy opening connections. An open or an attach refused as busy comes with when to try again, the controller's place in the server's queue, and the controller now waits that long instead of its own backoff.

  • A provider report says when signing in would not help. A provider whose check says its login command would not fix the problem is reported with login_helps: false beside auth_problem, so Console does not offer the command. Absent otherwise, as before.

  • Provider logins given to the machine, on demand. The controller makes an X25519 keypair at enrollment (or on its next run, for one enrolled before) and registers the public half. When its owner gives the machine a login sealed to it, a provider.login operation has the controller open it, check the provider signs in with it, and report whether it does. The machine's own login still comes first; agents of a provider it has none for get the given one, run in a process of their own, and restart when it changes or is withdrawn. Status reports auth_source: "sealed" for it.

  • Each agent as a Linux user of its own. On Linux with systemd and polkit, sudo switch-agent-controller install-service --separate-users sets up, once, a pool of agent users, a unit template that runs an agent as one of them, a polkit rule that lets the controller start only those units, and the controller as a system service. Each agent then runs as its own user and sees only its own directory. It cannot see the home directories, the controller's data, the other agents or the cloud instance metadata, and the rest of the system is read-only to it. Agents get provider keys from the controller's --env-file, not its user's own logins. --agent-users sets how many agents can run (16 by default).

Changed

  • Points to the dashboard's Devices page. Its messages, installer and README sent people to the Machines page, which the dashboard now calls Devices.
  • Agents' operations go to Switch on the controller's socket. Every operation an agent host calls through the relay (POST /agents/{id}/ops/{operation}) is sent as a call on the socket the controller already holds, when Switch offers calls, instead of as an HTTP request of its own. Over HTTP as before against a server that does not offer them, while the socket is down, and when as many calls are in flight as Switch takes. A call the socket closes on is answered 502, as a request whose connection drops.

Fixed

  • A Claude Code login that is there but does not work is reported as not signed in. claude auth status only says a login or key is present, so a key left by an old install read as signed in while every session failed. When setting up, on a recheck, and whenever a login was not yet known to work, the controller now sends Claude Code one short test prompt, with the person's hooks off and stdin closed, and reports what it said if it fails, from its result, its errors or its stderr. A Claude Code too old to have auth status is reported as such instead of as a check that never finished.
  • A controller updates to the release Switch names while sessions sit idle. It waited until no session existed on the machine, and a session connected to a room counted even with nothing to do, so a machine with long-lived sessions stayed waiting_idle for good. It now waits only until no session is in the middle of a turn. Each agent host reports which of its sessions are busy (one that has not said yet counts as busy), and the controller looks again every few seconds while it waits rather than at the next status report. Idle sessions in the controller's own process are restarted with it and resume their conversations; agents running in a process of their own keep running through it. An agent host too old to report busy sessions still holds the update while it has any session.
  • Agents running in a process of their own move onto the release the controller updated to. Such an agent host (detached, or a systemd unit with separate users, as on Switch cloud machines) outlives the controller and kept the shared-host bundle it started with, so a fix in it reached the agent only when the agent was restarted by hand. Each agent host now records the bundle it runs (by SHA-256) in its health.json, and the controller restarts one on another bundle, or too old to say, once none of its sessions is mid-turn; idle sessions resume their conversations. A host that comes back still outdated is tried again after 10 minutes.
  • A refusal that cannot pass stops the controller instead of retrying it forever. The server answers protocol_unsupported when it no longer speaks the controller's protocol and invalid_credential when it knows no controller by its credential, and marks both not retryable. The controller retried them every few seconds for as long as it ran. It now stops its agents and exits with a new code, 5 (update the controller) or 6 (enroll the machine again), and the systemd unit and launchd job it installs do not restart it on either.

[0.1.1] - 2026-10-07

Fixed

  • Installer falls back to ~/.local when it can't write to a system prefix, and takes a --data-dir to place the controller's data directory explicitly.

[0.1.0] - 2026-10-07

Added

  • Installable on its own. Each release publishes one npm package (the CLI and the shared-host bundle, no dependencies) and an install.sh, which checks Node, installs the newest release, and can enroll the machine and install the service in one go.
  • install-service / uninstall-service: runs the controller as a systemd user unit on Linux or a launchd agent on macOS, restarted only after an error that may pass.
  • doctor: checks Node, enrollment, the credential, the server (including a proxy that does not route /v1 to switch-core), each provider CLI and its sign-in, the service, and updates.
  • update: installs the newest release and restarts the service. run logs when a newer one exists.
  • run --env-file and install-service --env-file: environment for the agents, such as provider API keys or Vertex AI and Bedrock settings. Sessions now inherit the Vertex and Bedrock variables Claude Code reads (ANTHROPIC_VERTEX_PROJECT_ID, CLOUD_ML_REGION, ANTHROPIC_BEDROCK_BASE_URL, …).
  • enroll --secret-store: keeps the controller credential in the macOS keychain (the default on a Mac) or the desktop keyring, as well as in owner-only files (the default elsewhere).

agent-runtime

The Switch protocol client and MCP runtime (console/packages/switch-agent-runtime/). Version lives in its package.json.

[Unreleased]

Added

  • send_attachment tells the agent what Switch Trust did. When Switch Trust redacted the caption, or flagged or could not fully check a file, the result has a Switch Trust: line with the same trust field post_message returns. It used to say only that the files were sent.
  • Waits its turn when Switch is busy opening connections. A server pacing a reconnect storm refuses an open as busy with when to open again, which is this agent's place in its queue. The stream now waits that long, even right after a restart, instead of retrying on its own schedule.
  • When the WebSocket never opens and the runtime asks for the event stream instead, the request carries fallback=socket_never_opened and the socket's close code as fallback_close_code, so the server's log can tell it from a client built before the socket. Opening the event stream straight away, against an old server, sends neither.

[0.8.0] - 2026-09-25

Added

  • Attaches as a hosted worker on agent-protocol 7: the event stream opens with the worker capability and host identity, hands the protocol-7 frames to the worker, stops for good on a terminal worker refusal or a superseded launch, and fences hosted up-calls by the connection and incarnation it holds.
  • States a retained volume's layout version on open and uploads its cutover manifest until Switch confirms it.

[0.7.2] - 2026-10-01

Fixed

  • Event-stream handling hardened against a self-takeover and a silently dropped stream when a connection resumes after the machine sleeps (#621).

[0.7.1] - 2026-09-30

Fixed

  • A connection no longer gives itself up to itself. A placements or room subscribe request sent under one incarnation and answered after this client had reopened its own stream was refused as "taken over", naming this client as the new holder, and the client stood down for good with nobody else anywhere near the connection. Such a refusal now only fails the request, which is stated again on the next change or reconnect; a real takeover is still caught by the next heartbeat.

[0.7.0] - 2026-09-25

Changed

  • Reworked for the host-owned session model: the runtime serves the Switch tool surface per session host while Console/sidecar own session lifecycle and switch-core relays (#543). No longer published on its own tag — ships inside Switch Console (0.36.0) and its sidecar.

[0.4.3] - 2026-09-23

Added

  • Support for reading any room the agent is a member of without connecting to it, and the create-from-template flow (#628, #405).

Changed

  • Watcher ownership follows the console/sidecar split (#507).

Security

  • Untrusted room-message bodies are fenced in the delivered prompt (#527).

[0.4.1] - 2026-09-03

Fixed

  • A deleted room no longer wedges the whole connection. When the client declared a room the server had dropped, the server refused the entire connection; the client reopened and re-declared the same room, so it was refused again — an unbreakable loop (measured at ~29 errors/min for over a day). A refusal that names one of our rooms is now terminal for that room: it is dropped, reported at error level, and the reduced room list is pushed over the same callback the server's own updates use, so the reopen is a strictly smaller request and cannot spin. Refusals that name no room keep their existing transport backoff (#353).
  • The heartbeat now backs off when the connection is rejected. A 404/409 on /connection/beat means we are not attached and only a reopen helps, but the loop treated "the server answered" as recovery and beat on at full cadence — one rejected request every two seconds against a connection that could never reopen. A rejection now counts as a beat that did not land: reopen, then double the interval to the 30s cap on the same curve as every other failure; the first beat that lands returns to base cadence (#353).

Added

  • Reap orphaned agent runtimes at boot: reapOrphanedRuntimes is now part of the package's public surface so the runtime and Switch Console share one definition of an abandoned runtime, and a session started on a dead host is cleaned up on the next start (#309).

[0.3.3] - 2026-08-27

Fixed

  • The runtime now exits when its host does, instead of running forever. Its intended shutdown path (via the transport's onclose) never fired for a host that was killed, crashed or force-quit, and the hook listener's TCP server plus the heartbeat and lease timers kept the process alive against a Switch that was already gone — leaving stale processes and loopback listeners piling up across sessions. Shutdown is now driven by four independent triggers — stdin end/close, termination signals, a stdout/stderr write error, and a parent-pid watchdog — so the process goes away whichever way its host vanishes (#307).

[0.3.2] - 2026-08-19

Fixed

  • An environment naming an agent but carrying no token now resolves against the local agent store instead of refusing to start. Any partial SWITCH_* environment was treated as a broken config, but that is the exact shape a host settings file produces when the credential is deliberately kept out of the working tree — and Claude Code exports its settings env block into the process, MCP subprocesses included. Switch Console writes exactly that shape for its own agents, so a session started by hand in a directory it set up degraded to switch_unavailable with a perfectly good store on disk beside it. The configure skill used to write it too; it no longer does, so this is a safety net for the Switch Console case rather than how the standalone path works.

    The agent id makes the lookup exact, so nothing is guessed: an id that matches no store entry, one belonging to a different server, or one claimed by two entries at once all still refuse. A token missing either of the others also still refuses — inferring where to send a credential is a different order of risk from inferring which one to send.

  • Refuses a partly expanded environment instead of filling the gap from disk. Some ${SWITCH_*} still literal while others resolved means a substitution step did not finish, which is not the same as a value deliberately omitted — and silently completing it from the store would authenticate as whatever is on disk without saying so. All three unexpanded is still the host's ordinary pre-expansion spawn.

  • normalizeEndpoint folds scheme and host case. It went from grouping endpoints for display to gating whether an identity binds, and a differently-cased host is the same server.

[0.3.1] - 2026-08-12

Changed

  • The MCP server instructions no longer tell the agent to read_context on every message event, or to connect before every call. Reading is now conditional on a signal that the agent is actually behind — an unread count above zero, a gap warning, an unfamiliar thread, a long silence — and the connection is described as holding for the session rather than as a precondition to re-establish per call.

[0.3.0] - 2026-08-11

Changed

  • Published to npmjs.com under the @sandboxaq scope (was @sandbox-quantum on GitHub Packages), so npx fetches it with no login and nothing on disk — anyone, inside or outside SandboxAQ, can run the connector plugins. GitHub Packages required a read:packages token even for public packages, which gh auth login does not grant, so external consumers hit a 404 naming neither the registry nor the missing credential.
  • Publishing now authenticates by npm trusted publishing (OIDC) and runs in the release environment, so a pushed tag queues the publish for approval; builds carry --provenance and no npm token is stored.

The package name (switch-agent-runtime) and the runtime's behavior are unchanged — only where and under what scope it ships. Version 0.2.x was skipped to 0.3.0 by request.

[0.2.0] - 2026-08-10

Added

  • Resolves its own Switch identity and credentials from the local agent store (./.switch/agents/), not only from the environment — so a connector-plugin session that is handed the server but no identity can start standalone instead of exiting before the handshake.
  • select_agent tool: when the store names several agents on one server, identity is left open and select_agent binds it; the event stream and heartbeat start on bind rather than at boot.
  • Degraded startup: on a resolution failure — no credentials, an ambiguous multi-server store, or an unreachable Switch — the runtime starts anyway and serves a single switch_unavailable tool whose result is the diagnostic, instead of vanishing so Switch tools read as silently missing.

Changed

  • An unexpanded ${SWITCH_*} placeholder now counts as absent rather than as a value, so a Claude connector session falls through to the store instead of dying on its own placeholder. Half a set of vars still refuses to start .

[0.1.6] - 2026-08-09

Added

  • Declares its agent-protocol range, artifact name and release version on the event stream it already opens, so switch-core can record what is connecting .
  • Logs the server's declaration from the connection_state frame, so which versions were actually talking to each other is answerable from a bug report rather than a guess.

Changed

  • Reports a failing event stream on a curve — first failure, then powers of two, plus a line when it recovers — instead of once per reconnect. The reconnect itself is unchanged; only the reporting is rationed, so a stack that is simply stopped costs a handful of log lines rather than one per session per 30s, forever.

[0.1.5]

Releases before this changelog existed are in the git log. They are not reconstructed here: an invented history reads exactly like a real one.


sidecar

The remote runtime Switch Console deploys to an agent host. Versioned in artifacts.yaml and deployed by Switch Console rather than published on its own.

[Unreleased]

Added

  • Serves the session-side half of service grants: granted skills in each session, the GitHub credential helper and gh wrapper, and scrubbing of issued tokens, as Switch Console does.

[1.9.12] - 2026-10-01

Changed

  • Picks up the agent-runtime 0.7.2 event-stream fix it bundles (#621).

[1.9.11] - 2026-09-30

Added

  • The room watcher records its connection to Switch in health.json beside its other state, each time it changes, so Console can read it from the host instead of relying on a live connection to the sidecar.

Fixed

  • A stuck process can no longer block replacing a watcher. Stopping a watcher or session host sent SIGTERM and waited; one waiting on a child that would never finish — a session host that hung up and stayed alive — never exited, so every update to that agent failed with "The SDK host has not stopped". A host asked to stop is now killed, with every process it started, if it has not gone after 10 s (a supervised host) or 20 s (one being replaced).
  • A session whose host hung up no longer blocks its room for good. After a reset, a session host could finish and close its link to the watcher while a provider process it had started kept it alive. The watcher still saw it as running, so every room message failed and it was "started again" every five seconds, forever, without anyone being told. Now the host exits within five seconds of finishing whatever it left running; the watcher treats a host that hung up as gone, and stops it and what it started if it has not exited fifteen seconds later; and a session that still will not take a message after five starts in a row is left alone with its messages queued, and the room is told.

Changed

  • Claude Code sessions accept their agent definition in the launch spec and pass it to the SDK, rather than reading .claude/agents/<name>.md on the host. Sessions saved by an earlier Console, which name the file, still relaunch.

[1.9.10] - 2026-09-25

Changed

  • Session handling refined: shared-agent-runtime initial prompt, session-command and stop-shared-session paths, and transcript handling. Ships with Switch Console 0.37.0.

[1.9.9] - 2026-09-25

Changed

  • Owns its agent sessions under the new host-owned session model; switch-core relays rather than owning session lifecycle (#543). Ships with Switch Console 0.36.0.

[1.9.8] - 2026-09-23

Changed

  • Watcher ownership follows the console/sidecar split — remote watchers are parented by the sidecar (#507). Rebuilt on agent-runtime 0.4.3.

[1.9.7] - 2026-09-03

Fixed

  • Reap dead tmux targets so a pane whose tmux session has gone is not treated as a live one (#336).

Changed

  • Rebuilt on agent-runtime 0.4.1 (a deleted room no longer wedges the whole connection, and the heartbeat backs off when the connection is rejected).

[1.9.6] - 2026-09-01

Changed

  • Rebuilt on agent-runtime 0.3.4, which reaps orphaned agent runtimes at boot.

[1.9.5]

Changed

  • Runs agent-runtime 0.3.3, which now exits when its host does instead of leaving stale runtime processes and loopback listeners behind (#307). The client↔sidecar wire (ready line, endpoints, on-disk layout) is unchanged, so the major stays 1.

[1.9.4]

Fixed

  • The remote session spawner clears the CLI first-run prompts that would otherwise leave a spawned session alive but never answering, matching the desktop-side fix (CHOO-2213). Behavior change only — the client↔sidecar wire (ready line, endpoints, on-disk layout) is unchanged, so the major stays 1.

[1.9.3]

Added

  • The ready file carries a deployer field: the identity of the Switch Console install that started this sidecar, echoed from the environment it was launched with. It is the one thing on the host that says whose build is running — the content hash says only which. Purely additive, so no contract revision moves: an older client ignores the field, and a client reading a sidecar that omits it must treat that as unknown rather than as its own.
  • Writes an agent's launch profile on the VM and completes the home-directory placeholder in its environment, so a provider that names a config file by absolute path (OpenCode's OPENCODE_CONFIG) launches correctly on a remote host. Additive — the client↔sidecar wire is unchanged, so the major stays 1.

[1.9.2]

Changed

  • The remote session spawner now installs a provider's hooks whether they are delivered as config files or as a dropped plugin (OpenCode), mirroring the desktop side. A plugin-delivered provider previously launched on the VM with nothing installed and never reported that it stopped (#203). Behavior change only — the client↔sidecar wire (ready line, endpoints, on-disk layout) is unchanged, so the major stays at 1.

[1.9.1]

Changed

  • Windows-compatible path and process handling in the session spawner, matching the desktop app's Windows fixes (#207). Behavior change only — the client↔sidecar wire (ready line, endpoints, on-disk layout) is unchanged, so the major stays at 1.

[1.9.0]

Changed

  • Dropped the npm-registry-auth machinery (npm-registry-auth.ts): now that the agent runtime is published to public npmjs (CHOO-2021, CHOO-2023), the sidecar no longer needs a token to npx it. Behavior change only — the client↔sidecar wire (ready line, endpoints, on-disk layout) is unchanged, so the major stays at 1.

[1.8.0]

Added

  • Declares its sidecar-control range in the ready file switchdash already reads.

Changed

  • Three-part semver: 1.7 becomes 1.8.0. The major stays at 1 — every switchdash already in the field judges compatibility on the major and parses two parts, so 1.7 and 1.8.0 order correctly and neither side replaces the other. 2.0.0 would have every existing install treat this sidecar as incompatible and replace it while a newer install replaces it back .
  • The version no longer carries compatibility. It says which release is running; what the sidecar can speak is the sidecar-control range it now declares in its ready file.
  • The version is declared in artifacts.yaml and generated into the code. Nothing else declares it — the sidecar is deployed by switchdash rather than published, so it has no packaging file of its own.

[1.7]

Earlier versions used a two-part x.y scheme in which the major was the compatibility signal. History for those is in the git log.


switch-connector (Claude Code)

connectors/claude-code-plugin/. Version lives in .claude-plugin/plugin.json.

[Unreleased]

[0.9.16] - 2026-09-23

Changed

  • Skill updated for the recent room-workflow changes (reading a room you belong to without connecting, templates). Pinned agent-runtime moves to 0.4.3; the plugin version bumps so installs re-download (#628, #527).

[0.9.14] - 2026-09-06

Changed

  • Skill updated: Matrix references removed to match the Postgres-backed message store (#380). Plugin version bumps so installs re-download.

[0.9.13] - 2026-09-06

Changed

  • Skills updated for agent display names: the switch and configure skills note that agents may render under a human display name on the bridges (#330). Plugin version bumps so installs re-download.

[0.9.12] - 2026-09-03

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.4) — a deleted room no longer wedges the whole connection, and the heartbeat backs off when the connection is rejected. Plugin version bumps so installs re-download.
  • Skill updated: documents user-defined external reference types and list_all_references for discovering references across the instance.

[0.9.11] - 2026-09-01

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.3) — picks up reaping of orphaned agent runtimes at boot.

[0.9.10] - 2026-08-27

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.2) — picks up the runtime's host-exit fix (#307), which stops stale runtime processes and loopback listeners piling up. Plugin version bumps so installs re-download.

[0.9.9] - 2026-08-26

Fixed

  • The configure skill's standalone feature list no longer claims the task protocol works — it was removed from the connector skills, so an agent reading the old list could try something unavailable. Plugin version bumped so installs re-download.

[0.9.8] - 2026-08-24

Changed

  • The configure skill's generated connect command now names the agent it was set up for, so a session started in a directory holding several agents answers select_agent on its own instead of asking the operator. Plugin version bumped so installs re-download (CHOO-2344).

[0.9.7] - 2026-08-22

Fixed

  • The room-workflow skill said a Telegram DM is adopted the way Mattermost's is, so an agent asked for a 1:1 on Telegram would tell the user to message the bot and wait for a room that never arrives. Telegram has no DM: a private chat with the bot is the lobby, which answers with setup guidance and provisions nothing. The skill now says so, and points at a group of two as the way to get DM-like behaviour (CHOO-2314).

[0.9.6] - 2026-08-19

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.1) — picks up the runtime's store-based credential resolution (CHOO-1862); the plugin version bumps so installs re-download.

Fixed

  • The configure skill no longer breaks the standalone path it exists to set up. It wrote SWITCH_API_ENDPOINT and SWITCH_AGENT_ID into .claude/settings.local.json, and Claude Code turns that env block into real environment variables for everything it spawns — so every session it configured started with two of the three values set. The runtime takes a complete environment or none at all and refuses anything in between, deliberately, so those sessions got no Switch tools whatsoever while the token sat unread in .switch/agents/<name>.json beside them. The skill now writes the store and nothing else, which is the mode the runtime already supports on every published version; Step 1 strips any SWITCH_* an earlier run left in either settings file, and Step 9 writes permissions.allow alone.
  • The hook no longer skips mediation on a standalone install. It read its credentials from the environment only, and the configure skill deliberately keeps the token out of there, so it reported itself unconfigured and returned without reporting or mediating anything — silently, on every tool call. It now falls back to the same .switch/agents/*.json store the runtime reads, keyed on the agent id the session recorded when it joined a room, and names the cause on stderr when it cannot resolve one instead of skipping quietly.
  • The runtime resolves an agent id carrying no token against the store rather than refusing outright. The skill no longer produces that state, but Switch Console writes the same two keys for its own agents, so a session started by hand in a directory Switch Console set up hit the identical dead end.
  • The configure skill read a 401 from the health probe as the wrong server and sent the user off to find a different URL. The bridge authenticates everything but a few public routes, so a path left on the end of an otherwise correct base URL answers 401 rather than 404 — the host was right and only needed stripping back.
  • The hook resolves an identity the same way the runtime does in three more cases, each of which let a session run while its governance quietly did not. It folds endpoint case, as the runtime does, so a differently-cased host no longer binds in one and matches nothing in the other. It refuses a partly unexpanded ${SWITCH_*} environment instead of completing it from disk. And it refuses a token that is missing either of the other two, rather than substituting a different credential from the store while the session itself refuses to start — a test had asserted that substitution as correct.
  • A failed mediation check says so. Any error reaching the bridge — a revoked token, an unreachable server, a timeout — was swallowed and the tool call proceeded with nothing printed, which is indistinguishable from approval. The call still proceeds, because a bridge outage must not wedge a session, but it now says on stderr that it was not checked.
  • The configure skill no longer overwrites a credentials file that belongs to another Switch setup. Its script writes .switch/agents/<name>.json by name alone and truncated whatever was there, which in a directory shared with a Switch Console install (or an earlier run against a different server) destroyed a token that is issued once and stored nowhere else. It now stops before registering if that file names a different Switch server, and reports which — the same guard the Codex skill and Switch Console's own write paths got. The subagent step gets it too, over .claude/switch-subagents/<name>.settings.json: that path is Claude-only, so it was outside that sweep, and it is checked across the whole batch before the bulk registration, since refusing afterwards would strand every agent in it.

Changed

  • The configure skill is rebuilt on the standalone shape the Codex connector uses: registering and writing the credentials are separate steps that still run as one command (the API key is returned once), the env-var expansion pitfall sits beside the request it applies to rather than after the script that trips on it, the heredoc wrapper is shown rather than described, and the server URL is probed before anything depends on it. It also writes permissions.allow for the connector's tools, which Switch Console always did and a skill-configured install went without — so every room action stopped for an approval prompt.
  • The per-project vs global scope choice is gone. Global wrote the identity machine-wide, but credentials are only ever read from the session's working directory, so it behaved as per-project with extra steps.
  • The skill no longer asks for or sends notify_user, which left the platform when owner-linked addressing arrived. It had survived here as a step collecting a value the server no longer reads.
  • Both room-workflow skills list the full set of reasons switch_unavailable can be the only tool, and both now point at the configure skill as the remedy for the ones it can fix.

[0.9.5] - 2026-08-18

Changed

  • The room-workflow skill no longer teaches the task protocol as an interaction mode: the lifecycle section is replaced by a short note that the protocol exists, is registered on the server, and must not be called yet, and the guidance to delegate rather than message is removed. The scoped-addressing material moves to its own section. Documentation only — the MCP/HTTP tool surface is unchanged (CHOO-1418).

[0.9.4] - 2026-08-15

Changed

  • The room-workflow skill describes owner-only addressing — the new default for agents created in Switch Console — and states that send_targeted_message now fails outright for a disallowed sender, and that in-room commands are covered by the policy too (CHOO-2137).

[0.9.3] - 2026-08-14

Changed

  • The room-workflow skill covers Telegram: attachments cross the bridge as real uploads, chats cannot be created by a bot at all (so create_room fails there for every channel type), forum topics thread natively, and formatting has no tables and a 4096-character cap (CHOO-1686). It also warns that a Telegram room may be mention-only, where unaddressed talk never reaches Switch at all and read_context cannot recover it — the one place the skill's "pull the rest with read_context" promise does not hold. The slash-command and attachment platform lists now also name Discord, which had been left out of both.
  • The room workflow adds a Mattermost-only rule: post at the root there unless you were asked in a thread. Mattermost shows a threaded reply as a reply count under the original post rather than in the channel, so a first-time reader can take it for no answer. Threading is unchanged everywhere else.

[0.9.2] - 2026-08-14

Changed

  • The room-workflow skill lists opencode alongside codex and claude-code wherever it enumerates known agent types — the list_agents filter and the per-type options of update_agent_detail (#203). The plugin version bumps so installs re-download.

[0.9.1] - 2026-08-12

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.0) — picks up the runtime's conditional-read MCP instructions; the plugin version bumps so installs re-download.
  • Skill: how to write in a room — answer first, plain words, a few sentences, one point per message — and post a one-line heads-up before going quiet for a long task, then return with the result in the same thread (#210).
  • Skill: document the room-document and room-admin tools it had never mentioned — load_internal_documents (without which an agent cannot read a document attached to its own room), list_references, the create/update/delete_room_document trio, add_users_to_room, and archive_room / unarchive_room.
  • The skill is state-aware: it loads once for a session instead of before every tool call, and no longer makes an agent reconnect and re-read the room to say one thing. The description still directs the agent to load the skill — a passive rewording stopped it loading at all in live runs — but says to load it once, and drops the 40-name tool inventory; a new "steady state" section says the connection holds for the session; and the unconditional "always read / always connect" instructions are replaced by triggers that fire on an actual signal.
  • Skill: correct the paging instruction. It told agents to pass oldest_timestamp straight back as before, but the first is epoch milliseconds and the second is parsed as an ISO-8601 string, so the call raised instead of paging.

Fixed

  • The channel's unread tally now resets when the agent reads the room. The PostToolUse matcher never included read_context, so the reset the hook already implemented was unreachable and the count only ever climbed from the moment a session connected.

[0.8.1] - 2026-08-11

Changed

  • Pin @sandboxaq/[email protected] — the runtime moved to public npmjs under the @sandboxaq scope (CHOO-2021), so npx no longer needs a GitHub token. The scope change had shipped without a plugin version bump and so never reached installs; this release carries it and bumps the plugin version so installs re-download.
  • Skill: document read_context's new response shape — truncated / oldest_timestamp and the per-entry kind — and tell the agent not to conclude anything from a truncated read.

Removed

  • Dropped the configure skill's Step 0, which installed the private-repo / gh setup machinery, now that the repository is public. The skill itself stays; an earlier version of this entry said it had been removed wholesale, which it had not.

[0.7.9] - 2026-08-09

Changed

  • Bump the pinned @sandbox-quantum/switch-agent-runtime to 0.1.6, so sessions pick up its rationed stream-failure reporting and version declaration . The plugin version bumps with the pin so installs re-download it.

[0.7.8]

Releases before this changelog existed are in the git log and in the plugin manifest history.


switch-connector-codex

connectors/codex-plugin/. Version lives in .codex-plugin/plugin.json.

[Unreleased]

[0.3.17] - 2026-09-23

Changed

  • Skill updated for the recent room-workflow changes (reading a room you belong to without connecting, templates). Pinned agent-runtime moves to 0.4.3; the plugin version bumps so installs re-download (#628, #527).

[0.3.15] - 2026-09-06

Changed

  • Skill updated: Matrix references removed to match the Postgres-backed message store (#380). Plugin version bumps so installs re-download.

[0.3.14] - 2026-09-06

Changed

  • Skills updated for agent display names: the switch and configure skills note that agents may render under a human display name on the bridges (#330). Plugin version bumps so installs re-download.

[0.3.13] - 2026-09-03

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.4) — a deleted room no longer wedges the whole connection, and the heartbeat backs off when the connection is rejected. Plugin version bumps so installs re-download.
  • Skills updated: document user-defined external reference types and list_all_references for discovering references across the instance.

[0.3.12] - 2026-09-01

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.3) — picks up reaping of orphaned agent runtimes at boot.

[0.3.11] - 2026-08-27

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.2) — picks up the runtime's host-exit fix (#307), which stops stale runtime processes and loopback listeners piling up. Plugin version bumps so installs re-download.

[0.3.10] - 2026-08-26

Fixed

  • The configure skill's standalone feature list no longer claims the task protocol works — it was removed from the connector skills, so an agent reading the old list could try something unavailable. Also corrects a mcp_servers → mcpServers reference. Plugin version bumped so installs re-download.

[0.3.9] - 2026-08-24

Changed

  • The configure skill's generated connect command now names the agent it was set up for, so a session started in a directory holding several agents answers select_agent on its own instead of asking the operator; the plugin README is updated to match. Plugin version bumped so installs re-download (CHOO-2344).

[0.3.8] - 2026-08-22

Fixed

  • The room-workflow skill said a Telegram DM is adopted the way Mattermost's is, so an agent asked for a 1:1 on Telegram would tell the user to message the bot and wait for a room that never arrives. Telegram has no DM: a private chat with the bot is the lobby, which answers with setup guidance and provisions nothing. The skill now says so, and points at a group of two as the way to get DM-like behaviour (CHOO-2314).

[0.3.7] - 2026-08-19

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.1) — picks up the runtime's store-based credential resolution (CHOO-1862); the plugin version bumps so installs re-download.
  • The room-workflow skill lists the full set of reasons switch_unavailable can be the only tool, and points at the configure skill as the remedy for the ones it can fix (CHOO-1862).

[0.3.6] - 2026-08-18

Changed

  • The room-workflow skill no longer teaches the task protocol as an interaction mode: the lifecycle section is replaced by a short note that the protocol exists, is registered on the server, and must not be called yet, and the guidance to delegate rather than message is removed. The scoped-addressing material moves to its own section. Documentation only — the MCP/HTTP tool surface is unchanged (CHOO-1418).

Fixed

  • The configure skill no longer overwrites a credentials file that belongs to another Switch setup. Its script writes .switch/agents/<name>.json by name alone and truncated whatever was there, which in a directory shared with a Switch Console install (or an earlier run against a different server) destroyed a token that is issued once and stored nowhere else. It now stops before registering if that file names a different Switch server, and reports which.

Changed

  • Skill: list the full set of reasons switch_unavailable can be the only tool. The runtime is shared, so the identity failures added there apply here too; three of the six were missing.
  • Skill: point at the configure skill as the remedy for the causes it can fix, rather than saying the state is unfixable from inside the session — the Codex configure skill shipped in 0.3.2, so the remedy now exists on both hosts.

[0.3.5] - 2026-08-15

Changed

  • The room-workflow skill describes owner-only addressing — the new default for agents created in Switch Console — and states that send_targeted_message now fails outright for a disallowed sender, and that in-room commands are covered by the policy too (CHOO-2137).

[0.3.4] - 2026-08-14

Changed

  • The room-workflow skill covers Telegram: attachments cross the bridge as real uploads, chats cannot be created by a bot at all (so create_room fails there for every channel type), forum topics thread natively, and formatting has no tables and a 4096-character cap (CHOO-1686). It also warns that a Telegram room may be mention-only, where unaddressed talk never reaches Switch at all and read_context cannot recover it — the one place the skill's "pull the rest with read_context" promise does not hold. The slash-command and attachment platform lists now also name Discord, which had been left out of both.
  • The room workflow adds a Mattermost-only rule: post at the root there unless you were asked in a thread. Mattermost shows a threaded reply as a reply count under the original post rather than in the channel, so a first-time reader can take it for no answer. Threading is unchanged everywhere else.

[0.3.3] - 2026-08-14

Changed

  • The room-workflow skill lists opencode alongside codex and claude-code wherever it enumerates known agent types — the list_agents filter and the per-type options of update_agent_detail (#203). The plugin version bumps so installs re-download.

[0.3.2] - 2026-08-12

Added

  • A configure skill: the standalone setup path. Registers this Codex instance as a Switch agent and writes .switch/agents/<name>.json in the working directory, so codex reaches Switch with no Switch Console involved (CHOO-1936). It writes no MCP config — the plugin's .mcp.json stays the single server definition and the runtime resolves its own identity from the store (switch-agent-runtime 0.2.0+).

[0.3.1] - 2026-08-12

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.0) — picks up the runtime's conditional-read MCP instructions; the plugin version bumps so installs re-download.
  • Skill: how to write in a room — answer first, plain words, a few sentences, one point per message — and post a one-line heads-up before going quiet for a long task, then return with the result in the same thread (#210).
  • Skill: document the room-document and room-admin tools it had never mentioned — load_internal_documents (without which an agent cannot read a document attached to its own room), list_references, the create/update/delete_room_document trio, add_users_to_room, and archive_room / unarchive_room.
  • The skill is state-aware: it loads once for a session instead of before every tool call, and no longer makes an agent reconnect and re-read the room to say one thing. The description is a trigger rather than a 40-name tool inventory — which also keeps it clear of Codex's skill-description budget, which silently truncates an overlong one; a new "steady state" section says the connection holds for the session; and the unconditional "always read / always connect" instructions are replaced by triggers that fire on an actual signal.
  • Skill: an unread count is documented as proof you are behind, but its absence is no longer documented as proof you are current — Switch Console resets the tally on every line it delivers, not when the session reads.
  • Skill: correct the paging instruction. It told agents to pass oldest_timestamp straight back as before, but the first is epoch milliseconds and the second is parsed as an ISO-8601 string, so the call raised instead of paging.

[0.2.3] - 2026-08-11

Changed

  • Pin @sandboxaq/[email protected] — the runtime moved to public npmjs under the @sandboxaq scope (CHOO-2021); the scope change had shipped without a plugin version bump and never reached installs, so this release carries it and bumps the plugin version to force re-download.
  • Drop SWITCHDASH_GITHUB_TOKEN and npm_config_userconfig from the forwarded env_vars in .mcp.json — npx no longer needs a GitHub token now that the runtime is on public npmjs (CHOO-2023).
  • Skill updated for read_context's new response shape — truncated / oldest_timestamp and the per-entry kind.

[0.2.1] - 2026-08-09

Changed

  • Bump the pinned @sandbox-quantum/switch-agent-runtime to 0.1.6. The plugin version bumps with the pin so installs re-download it.

[0.2.0]

Releases before this changelog existed are in the git log and in the plugin manifest history.


switch-connector-opencode

connectors/opencode-plugin/. Version lives in package.json.

Unlike the other two connectors, nothing installs this one from a marketplace — OpenCode has none. Switch Console writes its files directly, so the version is for humans reading a diff rather than for an installer, and an install reports the app version that wrote it rather than a version of its own.

[Unreleased]

[0.1.12] - 2026-09-23

Changed

  • Skill updated for the recent room-workflow changes (reading a room you belong to without connecting, templates). The embedded agent-runtime pin moves to 0.4.3; the version bumps so Switch Console re-writes the connector (#628, #527).

[0.1.10] - 2026-09-06

Changed

  • Skill updated: Matrix references removed to match the Postgres-backed message store (#380). (Delivered at the next Switch Console update, which rewrites the embedded connector.)

[0.1.9] - 2026-09-06

Changed

  • Skill updated for agent display names: the switch skill notes that agents may render under a human display name on the bridges (#330). (Delivered at the next Switch Console update, which rewrites the embedded connector.)

[0.1.8] - 2026-09-03

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.4) — a deleted room no longer wedges the whole connection, and the heartbeat backs off when the connection is rejected. (Delivered at the next Switch Console update, which rewrites the embedded connector.)
  • Skill updated: documents user-defined external reference types and list_all_references for discovering references across the instance.

[0.1.7] - 2026-09-01

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.3) — picks up reaping of orphaned agent runtimes at boot.

[0.1.6] - 2026-08-27

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.2) — picks up the runtime's host-exit fix (#307). Reaches users with the next Switch Console release, which writes this connector.

[0.1.5] - 2026-08-22

Fixed

  • The room-workflow skill said a Telegram DM is adopted the way Mattermost's is, so an agent asked for a 1:1 on Telegram would tell the user to message the bot and wait for a room that never arrives. Telegram has no DM: a private chat with the bot is the lobby, which answers with setup guidance and provisions nothing. The skill now says so, and points at a group of two as the way to get DM-like behaviour (CHOO-2314).

[0.1.4] - 2026-08-19

Changed

  • Pin @sandboxaq/[email protected] (was 0.3.1) — picks up the runtime's store-based credential resolution (CHOO-1862). The pin lives in distribution.ts and opencode.json; the version bumps so a re-written install is stamped fresh.

[0.1.3] - 2026-08-18

Changed

  • The room-workflow skill no longer teaches the task protocol as an interaction mode: the lifecycle section is replaced by a short note that the protocol exists, is registered on the server, and must not be called yet, and the guidance to delegate rather than message is removed. The scoped-addressing material moves to its own section. Documentation only — the MCP/HTTP tool surface is unchanged (CHOO-1418).

[0.1.2] - 2026-08-15

Changed

  • The room-workflow skill describes owner-only addressing — the new default for agents created in Switch Console — and states that send_targeted_message now fails outright for a disallowed sender, and that in-room commands are covered by the policy too (CHOO-2137).

[0.1.1] - 2026-08-14

Changed

  • The room-workflow skill covers Telegram: attachments cross the bridge as real uploads, chats cannot be created by a bot at all (so create_room fails there for every channel type), forum topics thread natively, and formatting has no tables and a 4096-character cap (CHOO-1686). It also warns that a Telegram room may be mention-only, where unaddressed talk never reaches Switch at all and read_context cannot recover it. The slash-command and attachment platform lists now also name Discord.
  • The room workflow adds a Mattermost-only rule: post at the root there unless you were asked in a thread. Mattermost shows a threaded reply as a reply count under the original post rather than in the channel, so a first-time reader can take it for no answer. Threading is unchanged everywhere else.

[0.1.0] - 2026-08-14

First release. Ships the Switch room-workflow skill, registers the Switch MCP server in OpenCode's global config as a local server — so the runtime takes its credentials from the session environment and none is written to disk — and carries a reporting plugin that gives OpenCode sessions real working and completed states, naming the tool a turn is currently running.