Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent.
Read an MCP tool answer that arrived as structuredContent (#619)
kevinf8ec80ad3064
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 86f8058e..3f1d5915 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -6,10 +6,8 @@ afterwards, and does not when only the code moved.Newest first. `Unreleased` is what is on `main` and not yet tagged.-## Unreleased-### A tool cannot be granted for an app this deployment has not addedGranting a Bot a connector's tool checked only that the person asking was an administrator, so a@@ -20,7 +18,6 @@ the trail saying so. The grant is now refused, naming the app. Taking a grant awa dead row an administrator can see is still one they can remove, and a tool a connector has stoppedadvertising can still be granted: what a vendor lists today is not what somebody decided yesterday.-### A Google Drive shortcut is read as the file it points atSearch and recent files return shortcuts as ordinary hits, and reading one by that id was refused@@ -29,6 +26,12 @@ shortcut — the usual way a shared drive file is kept at hand — could be nameconnector now follows the target once and reads that file the same way it would have if search hadreturned it directly. A shortcut that names nothing, or another shortcut, is still declined.+### An MCP tool that answers in structuredContent is no longer read as empty++A tool that declares an output schema often puts the answer in `structuredContent` and leaves the+content list empty. That empty list was reported as "nothing was found", so the model filled the+gap from memory while the vendor had answered. The structured object is now read when the content+list had nothing to say. A tool that already sent text is unchanged.### A coworker can be a file of its own@@ -42,7 +45,6 @@ stop the server and both files are named, rather than one quietly winning on thewas listed in. The directory is in the package checksum, so a coworker added or edited there is apackage change a running deployment notices.-### Ten more example coworkers, each doing one jobThe example package shipped three coworkers, which is enough to prove the format and not enough to@@ -57,7 +59,6 @@ themes it can cite. Each says what the job is, what the coworker must not do, anit cannot find something. They grant nothing: a coworker names skills, a skill names tools, and whatit may call is what an administrator has granted. Delete the ones you do not want.-### Built-in and Mastra Bots can run on Anthropic API keysThe example built-in Bots and the Mastra Bot now use the selected model provider instead of@@ -65,7 +66,6 @@ assuming OpenAI. A deployment with `BOT_PROVIDER=anthropic`, `BOT_MODEL=claude-sAnthropic key routes built-in model calls, tool selection and Mastra runs through Anthropic's nativeAPI. Source startup no longer waits for the unused OpenAI-only sample when Anthropic is selected.-### The Agno Bot answers on an OpenAI keyPicked with an OpenAI key, the Agno Bot failed every run before reaching OpenAI. Agno sends a@@ -74,7 +74,6 @@ Compose passes when the setup screen names none, so the run ended in `Unsupporteparameter the model does not take is now dropped instead, the way the LlamaIndex Bot already doesit. The Anthropic key and an OpenAI-compatible endpoint behave as before.-### The Pydantic AI Bot starts on an Ollama model named with its tagOllama names every model with a tag after a colon, as in `llama3.1:8b`, and that is the name the@@ -83,7 +82,6 @@ model's name to mean the name already carried a provider, so Pydantic AI read `lrefused it as unknown, and the Bot never started. A model's name now carries a provider only when itbegins with the chosen provider's own, as in `anthropic:claude-sonnet-4-5`.-### The Langroid Bot starts on an Anthropic keyPicked with an Anthropic key, the Langroid Bot exited on startup asking for an OpenAI key. It names@@ -92,7 +90,6 @@ builds an OpenAI client for such a model all the same, from the `OPENAI_API_KEY`empty when the choice was not OpenAI. The image now installs Langroid's litellm extra and an emptyOpenAI key is treated as none, so the Bot starts and answers with the Anthropic model chosen.-### The AG2 Bot answers on an Anthropic keyThe AG2 Bot built an OpenAI client whatever the setup screen chose, and read `BOT_MODEL` but never@@ -101,7 +98,6 @@ Compose writes that one empty when the choice was Anthropic. It now reaches Anthown Anthropic client when that is the provider chosen, and OpenAI or an OpenAI-compatible endpointas before otherwise.-### The Microsoft Agent Framework Bot starts on an Anthropic keyThe Microsoft Agent Framework Bot built an OpenAI client whatever the setup screen chose, and read@@ -110,10 +106,8 @@ an OpenAI key, because Compose writes that one empty when the choice was AnthropAnthropic through Agent Framework's own Anthropic client when that is the provider chosen, andOpenAI or an OpenAI-compatible endpoint as before otherwise.-## 0.0.13-### Fresh desktop setup installs its runtime before sign-inDesktop setup installs local software before asking for an AI connection. Fresh Mac Podman@@ -121,7 +115,6 @@ installations no longer duplicate stack port bindings. The shared installer downprebuilt Bun binaries without developer tools; Linux setup also installs its native Podman runtimeand certificate bundle. (#604)-### Ctrl+B shows and hides the sidebar on a layout that does not write Latin lettersThe sidebar toggle's tooltip names Ctrl+B, or ⌘B on a Mac, and the shortcut was recognised by the@@ -129,7 +122,6 @@ character the key writes. On a Russian or Greek layout the B key writes "и" ordid nothing there. It is now read the way the app's own shortcuts read it since Shift+N was fixedfor the same layouts: from the physical key when the layout writes a character outside ASCII there.-### Scrolling a Bot's browser no longer scrolls or zooms the page around itWhile somebody drives a Bot's browser, a turn of the mouse wheel over its screen is sent to it, and@@ -138,7 +130,6 @@ as a passive listener, which a browser does not allow to do that, so the wheel sholding the Bot's screen along with the Bot's page, and Ctrl with the wheel zoomed the app. Thewheel is now handled by a listener that can hold it, so it reaches only the Bot's browser.-### Typing into a Bot's browser no longer triggers the app's own shortcutsWhile somebody drives a Bot's browser, every keystroke is sent to it. The app's shortcuts listen for@@ -147,7 +138,6 @@ keystrokes too, and they heard each one first, so typing a capital N into the Boalso showed or hid the sidebar. A keystroke sent to the Bot's browser now reaches only the Bot'sbrowser. Escape still closes the view, and the paste shortcut still pastes.-### The Python LangGraph Bot on Anthropic answers after a skill was pickedA skill somebody picks reaches the Bot as a system message just ahead of their message, and it@@ -157,7 +147,6 @@ every run in that conversation failed from then on, before the model was asked.Python LangGraph Bot now puts every system message ahead of the conversation, in the order given,where the integration joins them into its one prompt. Runs on OpenAI and Gemini are unchanged.-### The LangGraph Bot answers on Anthropic and GeminiWith `BOT_PROVIDER=anthropic` or `BOT_PROVIDER=google`, the LangGraph Bot answered nothing: every run@@ -167,7 +156,6 @@ coworker's standing role, which the server puts at the head of every run. The printegration refused the second one. On those two providers the Bot now folds them into one systemprompt, in the same order, and a skill picked for a message joins it. Runs on OpenAI are unchanged.-### The proof-of-concept Bot reads a model name set with whitespace around itA `BOT_MODEL` carrying a leading space reached this Bot as it was written. Its startup check refuses@@ -179,7 +167,6 @@ provider for a model with no name. Both now fall back to the default, and the nait reached only deployments that set the variable directly, such as a Kubernetes manifest or`docker run -e`.-### Pasting into a Bot's browser works on a layout that does not write Latin lettersWhile somebody drives a Bot's browser, Ctrl+V or Cmd+V is left to the local page so its paste event@@ -189,7 +176,6 @@ browser instead and nothing was pasted. The V is now read the way the app's ownsince Shift+N was fixed for the same layouts: from the physical key when the layout writes acharacter outside ASCII there.-### Reading a large file from Google Drive no longer downloads all of it`read_file_content` shows a Bot at most the first 20,000 characters of a file, but it downloaded the@@ -199,7 +185,6 @@ reading once it has more than it can show and cancels the rest of the download.is unchanged, except that the note on a cut file no longer gives the file's full length, which is nolonger known.-### A Bot reads the text an MCP server returns as an embedded resourceA tool result can carry an embedded resource, and a text resource holds content, such as a file the@@ -209,7 +194,6 @@ resource reached the model as `[resource]` and its contents were dropped. GitHubfile. The text of an embedded resource is now passed on like a text part. A resource that carriesbytes is still named.-### The Google Drive connector reaches files in shared drivesDrive leaves shared drive items out of any `files.get` or `files.list` request that does not say it@@ -219,7 +203,6 @@ appeared in `search_files` or `list_recent_files`. Those requests now say they sand the listings ask for shared drive items. Listings keep Drive's default `user` scope rather thansearching every shared drive.-### Google Drive search and recent files leave out what is in the trashDrive's `files.list` returns trashed files unless the query excludes them, and neither `search_files`@@ -228,7 +211,6 @@ as a recently changed file, with nothing in its line to say it was in the trash,answer from it as though it were current. Both now ask Drive to leave the trash out. Reading a file by itsid is unchanged.-### New conversations are still named once some older ones could not beEvery pass of the job that names conversations offered at most twenty of those still without a name.@@ -239,7 +221,6 @@ up, a new conversation could miss out on every pass and keep showing its plain nA pass now skips any conversation that the job already holds work for, so new ones get a place. Oneit could not name is still tried again later, as before.-### An offboarding one app refuses still records the apps that answeredRemoving somebody withdraws each brokered account they connected. When the broker refused one of@@ -251,7 +232,6 @@ got. Each app is now asked, recorded with the answer it actually gave and its rothe apps that were refused are left standing for the retry. The act still fails and still answers500, so a refusal is as loud as it was.-### Paging the audit trail no longer skips rows written in the same millisecond`GET /api/admin/audit-events` hands out a `nextCursor` built from the last row's timestamp, which the@@ -261,7 +241,6 @@ the same millisecond, were on no page at all. Anything that walked the trail pagthem without any sign of it. The cursor now carries the row's full timestamp. A cursor issued beforethis change still reads.-### The New chat shortcut works on a Russian or Greek keyboard layoutSettings lists New chat as Shift+N, and the app matched the character the keystroke wrote. A layout@@ -269,7 +248,6 @@ that writes another script has no key that writes an N: Shift and the N key writ"Ν" on Greek, so the shortcut never fired there. When the character is not ASCII, the app now readsthe physical key instead. A layout that writes Latin letters, such as Dvorak, still goes by the letter.-### Enter that confirms a typed character no longer saves a name, a rule or a wizard stepJapanese, Chinese and Korean are typed through an input method, where Enter confirms the character@@ -278,7 +256,6 @@ the character still unconfirmed, the new-coworker wizard moved on to its next strule was saved into the policy in force. Those fields now wait for the character, the way the chatcomposer already does, and an ordinary Enter works as before.-### A built-in coworker can be edited where the deployment's own Bot is on localhostA coworker created as Built in is stored pointing at the managed Bot's address. Editing its name,@@ -289,7 +266,6 @@ failed with "That address is inside this deployment's own network, so an agent mThe dialog now leaves a built-in coworker's address where it is stored. A coworker somebody hostsstill sends its own endpoint, as before.-### A long page's text is cut between characters, not through oneA navigation hands the Bot the first 6000 UTF-16 code units of the page's readable text. When that@@ -297,7 +273,6 @@ limit fell between the two halves of an emoji, the Bot was handed text ending onwhich reads as U+FFFD: a broken character that is not on the page. It now stops one code unit shortin that case, which is what a control's name and value in a page snapshot already did.-### Tool selection reads a skill choice the model wrapped in a code fenceBefore a run, the deployment's model picks which of a Bot's skills the message needs, so a Bot holding@@ -307,7 +282,6 @@ or lead with a sentence. Every such answer read as no answer, so the Bot was offholds and the audit row said `unavailable`. The object is now read out of the answer, the way therouter already reads its own. A bare JSON answer is read as before.-### A long message reaches the coworker it is for, and is recordedA message over 10,000 characters, such as a pasted email thread or log, was refused by the router@@ -317,7 +291,6 @@ so the message went to the default coworker rather than the one it is for, and amessage's opening, and the whole message still goes to the coworker. Shorter messages route asbefore.-### Removing somebody is recorded even when retiring what they owned failsRemoving somebody denies their access and ends their sessions, then retires the credentials and@@ -327,7 +300,6 @@ and removing them a second time reported success without retrying it, leaving thstanding. The removal is now recorded as soon as it takes effect, and removing somebody alreadyremoved finishes the retirement that failed.-### Removing a connector takes its grants with itA grant naming a connector's tool outlived the connector. Removing an app revoked every credential@@ -340,7 +312,6 @@ been made. An app's grants are now removed in the same step as the app, the remogrants it released and from which Bots, and a migration drops the grants earlier removals leftbehind. Grants for other connectors, and skill grants, are untouched.-### A vendor that broke no longer reads as a refusal to a Bot running its own loopWhen a Bot that calls tools back from its own process, such as the LangGraph Bots, called a tool@@ -349,7 +320,6 @@ holding. The conversation drew it as blocked and the model read it as not allowetrail recorded a failed call. Only a refusal is marked now. A vendor that broke reads "That tool couldnot be called: …", the way it already did for a Bot running here, and a refusal reads as before.-### A Bot running its own loop is told a vendor's error is the vendor'sA vendor that says no by answering with an error, the way an MCP server refuses a call, reached a Bot@@ -360,7 +330,6 @@ ordinary result, and could tell the person they had no access rather than that tBoth kinds of Bot are now told the same thing. A result that is not an error, and this deployment'sown refusals, read as before.-### A long Composio result or failure is cut between characters, not through an emojiA Composio action's answer over 20,000 characters, and a failure sentence as long, were cut by UTF-16@@ -370,10 +339,8 @@ Plane, the text handed to the model ended on half of it: a lone surrogate that Jthe way the MCP and built-in transports' cuts already do. Anything that fits is untouched, and thenote saying the answer was cut reads as before.-## 0.0.12-### A deployment can broker its Bots into a few hundred apps through ComposioComposio holds a person's connections to a few hundred SaaS apps behind one account. A deployment@@ -383,7 +350,6 @@ through the gateway like the rest. Unset, there is nothing to connect, nothing tComposio tool for a Bot to call, and the Plugins page says so under **More apps** rather thanpretending otherwise. See [docs/plugins/composio.md](docs/plugins/composio.md).-### A skill's grants are removed when it is uninstalledUninstalling a skill deleted the skill but left its tool grants, which are keyed by its slug. A new@@ -391,7 +357,6 @@ skill created under the same slug then inherited them, and was offered on the Bobeen granted to with no grant ever made for it. Uninstalling now removes a skill's grants along withit, in one transaction, and an upgrade drops any grants already left orphaned this way.-### Malformed requests are refused instead of coerced, and a fail-open is closedA pass across the write and query surface answers a malformed request with a 400 that names the bad@@ -406,7 +371,6 @@ One of these closed a hole rather than tightening an edge: a skill installed witin its `tools` list had that entry silently dropped, so the skill declared nothing and installed as asuccess. It is refused now.-### Naming a conversation asks the endpoint OPENAI_BASE_URL names, not OpenAIThe job that names a conversation sent its request to api.openai.com whatever `OPENAI_BASE_URL` said,@@ -416,7 +380,6 @@ conversation, to OpenAI; OpenAI refused the key, so no conversation was ever namgoes to the same endpoint as every other model call. A deployment that never set `OPENAI_BASE_URL`behaves as before.-### A Bot's question to a person survives a route that failsWho "a person" is, is a seam a deployment fills in with its own on-call rota or duty desk. If that@@ -427,7 +390,6 @@ told, in a sentence it can say, that nobody could be asked and that it must notan `agent.escalation_failed` row goes down carrying what the route actually threw. Deployments usingthe shipped in-conversation route are unaffected: it cannot fail.-### Resetting a Bot's computer while the Bot is acting signs it outOn a deployment with one shared computer, which is what the published image and the Helm chart run by@@ -438,7 +400,6 @@ and the audit trail recorded the saved state as deleted. A Bot's browser is no lit is being closed: the action waits for the reset to finish and starts signed out. Computers thesupervisor makes per Bot were not affected.-### A long control name or value in a page snapshot is cut between charactersThe computer's page snapshot keeps the first 200 UTF-16 code units of each control's accessible@@ -447,7 +408,6 @@ ending on half a character, which reads as U+FFFD: a broken character that is noat the end of a message the Bot had just typed into a text box. The cut now stops one code unitshort in that case, the same rule tool results and relayed answers already follow.-### A Bot's shell can no longer read the deployment's keys from a neighbouring processIn the all-in-one image the API and the browser ran under one account, and a Bot's shell — a child@@ -459,10 +419,8 @@ the variables it needs, so its own environment carries none of those keys; and t`/run/s6/container_environment` are closed to the shell. A deployment that runs each Bot in its ownsandboxed computer, as the documentation asks for, was never exposed to this.-## 0.0.11-### The LlamaIndex Bot answers with the model the setup screen choseThe LlamaIndex Bot built an OpenAI client from `BOT_MODEL` and ignored `BOT_PROVIDER`, so it only@@ -475,7 +433,6 @@ way the Agno Bot does, so all three choices answer. A model LiteLLM does not knoto call tools, which the AG-UI workflow requires, and parameters a model does not accept, such asthe temperature LlamaIndex sends to a reasoning model, are dropped rather than refused.-### The Audit page says "not enforced" only under a dry-run refusal that went aheadOn a deployment in `dry-run`, the Audit page printed "dry-run: recorded, not enforced" under every@@ -486,7 +443,6 @@ which is the one case it describes. A tool call refused for carrying credentialrecorded with `carriedOut: false` in every mode. Rows already written keep their old value, and thepage reads them correctly either way.-### The Routines page says when nothing is there to run themA routine needs a second process to fire it, and a deployment that never started one looked exactly@@ -497,7 +453,6 @@ sweeping is told so — that no worker has ever checked in, or when the last onebeing shown a page that looks correct. The window is the fifteen minutes a routine's own schedulealready has as its floor, so a gap longer than that is one no routine could have wanted.-### A long tool result or relayed answer is cut between characters, not through an emojiA tool result over 20,000 characters, and a Bot's answer over 12,000 relayed back through a handoff,@@ -507,7 +462,6 @@ JSON carries as a bare `\ud83d` and UTF-8 turns into a replacement character. Thunit short in that case, the way an attached text file's already did. Anything that fits isuntouched, and the note saying the result was cut reads as before.-### A boundary rule can ask what started a run, not only whose authority it carriesA routine's turn goes through exactly the path a person's chat turn does, as the routine's owner:@@ -538,7 +492,6 @@ rule about routines into a deployment that refused every ordinary click.Replaying a rule against history reads the initiator off the row when it is there and treats a rowthat predates the field, or carries a shape this version does not recognise, as a person.-### The Python LangGraph Bot tells its model why the deployment refused a tool callWhen the deployment would not run a tool call from the Python LangGraph Bot — a token it no longer@@ -548,7 +501,6 @@ could say a call was refused but not why, and could not correct a call the deplomalformed. The reason now follows the status, the way the TypeScript LangGraph Bot already passes iton. A refusal with no readable reason reads exactly as before.-### A tool argument that starts with "Basic" or "Bearer" is no longer refused as a credentialContent inspection read any MCP tool argument whose first word was "basic" or "bearer", in any case,@@ -560,10 +512,8 @@ punctuation or mixed case after `Bearer`. A shorter bearer token, or a Basic valdecode to `user:password`, is no longer caught by this pattern; the same value under an`authorization` field is still refused by name.-## 0.0.10-### The LangGraph Bot says a refused tool call was refused, not that it found nothingWhen the deployment would not run a tool call from the LangGraph Bot — a token it no longer accepts,@@ -573,7 +523,6 @@ tells its model the call was refused, with the status and the deployment's reasoPython LangGraph Bot already does, and the transcript draws it as a refusal. A tool that answered ispassed on exactly as before.-### Revoking a grant with a blank ref or Bot is refused instead of reported as done`DELETE /api/plugins/grants` checked its query params with truthiness, and a query param is@@ -583,7 +532,6 @@ match, still wrote a `plugin_revoked` audit row naming whitespace, and answeredon the trimmed values, so a blank ref or Bot is a 400 with the same message, no delete, and noaudit row.-### The Bot in the box and the LangGraph Bot read a message that has a file attachedA message with a file attached reached both Bots as `[object Object],[object Object]`, in place of@@ -591,7 +539,6 @@ what the person typed and the file both: they read a message as a string, and ona list of parts. Each part now reaches the model as what it is — the words, the text of an attachedfile, an attached image — and a part neither can read is named rather than dropped.-### A fractional or out-of-range computer setting takes the fallback instead of breaking the boot`numberFromEnv` accepted anything `Number` called finite and positive, so `PORT=80.5` bound@@ -602,7 +549,6 @@ so only whole numbers on sight are values now and anything else takes the documethe port additionally keeps its 1–65535 range, the way the supervisor's own port parser alreadydoes. Zero semantics are unchanged: `COMPUTER_BROWSER_IDLE_MS=0` still keeps browsers resident.-### A malformed page size is refused instead of silently coerced`GET /channels` and `GET /api/admin/people` read `?limit=` with `Number.parseInt`, which@@ -611,7 +557,6 @@ wrong page. Both now share one strict parser with the audit list's rule: absentthe store default alone, a run of digits is clamped into range against the same ceiling the storeenforces, and anything else is a 400 naming the parameter, before the database is reached.-### A skill written with a non-string slug or summary is refused instead of failing the insert`POST /api/plugins/skills` checked presence with truthiness and then ran the slug regex, which@@ -621,7 +566,6 @@ caller error. A slug, a title and instructions must be non-empty strings now, thtested only after that, and a summary must be absent or a string; anything else is a 400 namingthe field, before any refusal check, store write, or audit row.-### More endpoints refuse a malformed request instead of coercing it or failing openThe same treatment reached the rest of the write and query surface: the agent tool-call endpoint's@@ -636,7 +580,6 @@ was filtered to the strings in it, so `{"functions":[123,null,{}]}` became an empermission check ran over nothing, and the answer came back `allowed` for functions the caller had infact named. The verdict is about the functions named now, or a 400; an absent list still means none.-### Generated interfaces, tables and formsGenerative UI is enabled by default; set `OPENBOT_GENERATIVE_UI=false` or `0` to disable it.@@ -647,7 +590,6 @@ interfaces correctly.The playground rejects invalid JSON before saving or publishing, confirms successful saves, andshows published custom components in the administrator's gallery.-### A Bot's computer is rebuilt when it holds a token the deployment has stopped usingA computer checks every caller against the `COMPUTER_TOKEN` it was created with, and holds that one@@ -668,10 +610,8 @@ already replaces one built from an older image, keeping the profile and workspaccomes back with its logins and its files. A deployment that sets no token is left alone, because acomputer with no door on it is a choice the environment made rather than a mismatch to act on.-## 0.0.9-### The People screen keeps a person's last sign-in when their sessions go away`Last signed in` was `max(sessions.created_at)`, so it disappeared whenever the session rows behind@@ -681,7 +621,6 @@ The moment of each sign-in is now recorded on the person, so the answer survivesremoved person's row shows when they were last here instead of leaving it out. Existing deploymentsare backfilled from whatever sessions they still hold.-### A message can carry filesPick them, drag them onto the composer or paste them in: up to eight files on one message, images up@@ -701,7 +640,6 @@ to seven other files, and the part sent says where it was cut so the model is noabout a file it read only part of. A megabyte of text is therefore stored whole and read as itsfirst eighth or so, and the person who attached it is not yet told that.-### Unsent attachments have to be swept, or somebody who stages 32 can attach nothing againA file is stored when it is picked, not when it is sent, so every abandoned draft leaves bytes@@ -718,14 +656,12 @@ behalf: where nothing runs it, the files are never cleared and anybody who reachnothing, in any channel, from then on. At 8 MiB a file, 32 staged files per person is the 256 MiBto size storage against.-### Fresh desktop installs pin the latest published deploymentThe desktop app resolves GitHub's latest published release on first setup and downloads that exacttag's source and image manifest. It records the version after both downloads finish and reuses iton subsequent starts, so new installs no longer stay tied to the app's old v0.0.8 default.-### The Bot computer refuses a malformed scroll or live input before the browser sees itA non-finite wheel delta travelled into Playwright and came back as a 502 that read as a broken@@ -733,14 +669,12 @@ computer, and any JSON object on the live-screen socket fell through to `Input.iforwarded wrong-typed coordinates to CDP. Scroll deltas must be finite numbers now, and liveinput must match its mouse, wheel, key, or text shape; anything else is a 400 naming the field.-### A non-string plugin grant or tool call is refused before it reaches the store`POST /api/plugins/grants` and `POST /api/plugins/call` checked presence, not shape, so a JSONnumber, object, or whitespace string passed and failed inside the store as a 500. Refs and Botids must be non-empty strings now, and anything else is a 400 naming what is required.-### A fractional or infinite snapshot id is refused as malformed, not staleA `snapshotId` of `1.5` or `Infinity` passed the acting routes and never matched the stored@@ -748,14 +682,12 @@ integer, so the answer was a 409 stale snapshot and the caller retried a requestmalformed. Non-integer ids are refused with a 400 naming the ref and its snapshot before anydecision or audit row.-### A `DATABASE_URL` with a port of zero is refused at start-up`postgres://…:0/…` parsed and booted, and every query then failed against a port nothing listenson. Ports outside 1-65535 are refused with a sentence naming `DATABASE_URL` before a socket isever opened.-### Setup installs the container engine, instead of telling somebody to go and get oneSetup ended at "Install Podman Desktop or Docker Desktop first" on any machine that had neither,@@ -771,7 +703,6 @@ container themselves and previously named an obstacle with no way past it. Everya resolved path rather than trusting the PATH this process was started with, so an engine installed aminute ago can be used by the run that installed it.-### Setup ends with a question the Bot has to answerEvery step before the last one proves that something started, which is not the same as proving the@@ -783,7 +714,6 @@ rather than an empty answer, because the framework catches its own 401 and logsof a refusal in the container's log and nowhere else. The sentence on screen is OpenBot's own andnames the choice to change, with the harness's log behind a disclosure for the developer half.-### A subscription picks the Bot that can spend itA plan is not a key, and only one Bot speaks each vendor's subscription. Signing in to a Claude plan@@ -799,7 +729,6 @@ which would give a Bot that works in the morning and fails after lunch. The windmenu now, so the shortcut works on the screen whose own instruction is "paste the code it shows you";macOS routes the clipboard through the menu bar, and a window without one has no Paste.-### Signing in to CopilotKit from the window worksThe sign-in that creates a key for somebody had never been run end to end, and it failed four times@@ -815,7 +744,6 @@ white rectangles, because the tile rule overrode the background to white and notsomebody to choose between six empty boxes. The sign-in address is kept on screen the way the plansign-ins keep theirs, for the machine whose browser is not the one in front of the person.-### An endpoint that needs no key can be connectedThe compatible row names Ollama and vLLM in its own summary and then refused to continue without an@@ -832,7 +760,6 @@ credential store, because it is not a credential, and a key somebody actually tyunchanged. The model name reaches the bundled Bot too, which reads `AGENT_BOT_MODEL` and had beenleft on a pin chosen for OpenAI's own catalogue.-### A model name no longer outlives the answer that chose itThe compatible row is the only one that names a model, and switching away from it kept the name.@@ -842,7 +769,6 @@ cannot use the model that was chosen" about a model this run never chose. The nathan emptied, so the compose default applies, and taken out of the file as well, because the writerkeeps the lines it did not write and that is what let it survive.-### Credentials go to the machine's own credential store, not the `.env`The `.env` is a settings file, and a settings file is something somebody opens, reads out to support@@ -861,7 +787,6 @@ so a secret arrives at exactly the services that declare it and is written downearlier version already wrote in plaintext is moved and then purged, or the change would have boughtnothing for anybody who already had OpenBot.-### The credential store is asked once per run, not once per screenFour Keychain dialogs every time the setup screen mounted, each needing a click before the window@@ -874,7 +799,6 @@ mount for something that was never there. Writes go through the same memory andso the two cannot disagree. This does not remove the prompts on a first run, and nothing in thisprocess can: that decision belongs to the operating system and to the signature.-### Stop stops the Bot that was picked, and the next Start no longer refuses because of itCompose only acts on a profiled service when the profile is named, so Stop left the one container the@@ -884,7 +808,6 @@ OpenBot itself had started, which the person never saw and could not find, and tforward from that screen. A port this deployment already publishes is not a stranger on the port, sothe check reclaims our own and keeps its teeth for somebody else's.-### Stop stops the host processes on WindowsMeasured on Windows Server 2022: Stop took the containers down, reported success, and left the server@@ -899,7 +822,6 @@ byte-order mark in front of `package.json`, which `Set-Content -Encoding UTF8` wthe manifest unreadable and was reported as "the deployment is older than this version of OpenBot",sending somebody looking for a newer installer over three bytes.-### The installed app is served without a development server"Show OpenBot" did nothing on a machine where the stack was up. The window said OpenBot was running,@@ -916,7 +838,6 @@ button also shows what it was told: the call behind it already answered "OpenBotport 3010 yet, so there is nothing to show", and the click handler threw that sentence away, which iswhy a dead process looked like a dead button.-### A conversation whose history this deployment cannot reach says soClicking a conversation in the rail drew the coworker's name and then nothing at all. The rail comes@@ -931,7 +852,6 @@ been spoken in and comes back empty has a history this deployment cannot reach.in the notice slot beside the existing explanations for a deleted coworker and for turns that couldnot be parsed.-### `bun run dev` no longer starts a routines worker that cannot start`bun run dev` fanned out across every workspace, and one of them is the routines worker. That worker@@ -942,7 +862,6 @@ successfully. The command now starts the app and the server, which is what `READ`docs/development.md` already say it does. Routines are unaffected: `scripts/start.sh` starts theworker exactly as before, and on Kubernetes the CronJob does.-### Double-clicking works while a person is driving a Bot's browserA double click was sent to the Bot's browser as two separate first clicks, because every press said@@ -953,7 +872,6 @@ the wheel simply could not do, with nothing on screen to say why — the clickscounted as the first. The count the person's own browser worked out is now the one that is sent, so adouble click is a double click and a single one is unchanged.-### Pressing Enter works while a person is driving a Bot's browserTaking the wheel of a Bot's browser is mostly for the sign-in it cannot do itself, and Enter is how a@@ -967,7 +885,6 @@ what makes Chrome carry out what the key means. Measured against Chromium 151: ekey — Backspace, Delete, Tab, Home, End and the arrows — already did what it meant and is unchanged,and a single-line field still holds exactly what was typed into it.-### A half-ticked box is no longer described to a Bot as tickedThe snapshot a Bot reads before it acts on a page says whether each box is ticked, and Playwright@@ -979,7 +896,6 @@ chosen when most of them were not. `mixed` is now reported as not ticked, whichhalf of a yes-or-no answer and the one that gets the right action: clicking a half-ticked box ticksit. An ordinary tick and an ordinary empty box are unchanged.-### A Bot cannot end its turn by asking a person nothing`ask_person` is how a Bot stops and puts something to a person instead of guessing, and a call with@@ -992,7 +908,6 @@ duty desk rather than the person already in the conversation, it is a page to soquestion on it. A blank question is now refused with the sentence that was already written for it,and a question typed with room around it is recorded as the question rather than as the spacing.-### A routine scheduled for Sunday says Sundays, whichever number it was written withCrontab has always let Sunday be either 0 or 7, the scheduler here takes both, and a routine written@@ -1002,7 +917,6 @@ that page as `0 9 * * 7` while its neighbour said "Sundays at 09:00" — the samtwo ways, with the raw one looking like something had gone wrong. Both spellings now read as Sunday,and a list that names the day under both of its numbers says it once.-### A tenant package's theme may carry a commentA package's `theme.css` is checked at start-up against what it is allowed to define: the `:root` and@@ -1016,7 +930,6 @@ over a comment, saying nothing about comments. Comments are now taken out beforedefinitions, which also closes a comment wedged into the middle of `url(` as a way past the ruleabove it.-### Test connection stops reading once it has seen the agent answerThe button that checks an agent before it is registered sends it a real run and reads what comes@@ -1029,7 +942,6 @@ agent that had answered correctly in its first two events. It now reads the openthe connection, and answers in the time the agent took to start rather than the time it took tofinish.-### A key pasted with a line break in it is now refused, instead of reported as an unreachable agentThe box that holds an agent's key takes whatever is pasted into it, and what comes off a clipboard is@@ -1043,7 +955,6 @@ every turn that Bot took afterwards failed on a value nothing on screen said anyplaces now check the value before accepting it and say which kind of character is in the way. Thecharacter is named; the key never is.-### A deployment directory pasted with a stray space goes where it saysThe desktop setup screen asks where OpenBot should live, enables Start once that box is not blank@@ -1058,7 +969,6 @@ separator — it stopped being absolute, and the deployment was laid out relativwindow happened to be running from. The path is trimmed at both ends now. Spaces inside it are partof a directory's name and are left alone.-### The desktop app notices a busy port whichever loopback holds itThe shell refuses to start when something already holds port 3001 or 3010, because otherwise the@@ -1070,7 +980,6 @@ held on `::1` alone was reported free and the start went ahead into it. Both addnow, so the two agree on what "in use" means and the person is told which port is taken and whatOpenBot wanted it for.-### A request for a secret no longer follows a Bot into tomorrow's conversationsAn unanswered ask to take the wheel stops being shown after ten minutes, because control belongs to a@@ -1083,7 +992,6 @@ so a value typed into a box left open in an old tab is refused rather than senthas ended. A request inside the window is unchanged, and a person actually holding the wheel is stillnever timed out.-### `COMPUTER_BROWSER_IDLE_MS=0` now keeps browsers resident, as it says it doesZero is the documented way to switch off the sweep that closes a Bot's browser after it has sat@@ -1096,7 +1004,6 @@ setting. A blank variable, which is what an unset variable declared in a composestill not zero: it means "not set" and takes the default, as do a negative and anything that is not anumber.-### A flag or a family emoji in a channel preview is no longer cut in halfThe one line a roster draws is cut to a cap, and the cut walked code points -- right for a plain@@ -1108,7 +1015,6 @@ generated channel title, and in the excerpt the titler is shown. The cut is nowgrapheme clusters, so what a person sees as one character is kept or dropped whole. Plain text iscut in exactly the same place as before.-### A scroll with an unusable `deltaY` is refused, rather than scrolling some other distance`POST /computers/:botId/scroll` and `POST /computers/:botId/human/scroll` accepted any JSON number@@ -1118,7 +1024,6 @@ its own default distance. The caller was answered 200 for a scroll it had not asthat is not a finite number now answers 400 and the page is not touched, the way the timeout on`exec` and the coordinates behind a person's click already did.-### An MCP call carrying `x-api-key` is stopped the same as one carrying `api-key`The check that keeps credentials out of MCP tool arguments compared each argument name against a@@ -1131,7 +1036,6 @@ adds the spellings of names already on the list that were missing from it: `passa credential: an argument named `x_axis`, `token_count`, `max_tokens` or `secretary` is passed asbefore.-### One command to stop what `start.sh` startedStopping the local stack meant four commands read off the end of a successful start, and the one@@ -1143,7 +1047,6 @@ OpenBot, so an unrelated process on 3010 is named and left alone rather than kildeleted: the database, the Bots' files and their browser profiles are volumes. `--keep-computers`leaves the browsers signed in.-### The trail says when an identity provider was added, not only when one was taken awayWhoever holds an identity provider decides who can sign in at all, and the audit trail recorded only@@ -1154,7 +1057,6 @@ through the library's endpoint rather than the screen was unrecorded for the samnow written where the deployment already stands in front of those routes to check that the personasking is an administrator, so a provider appearing or disappearing names itself and whoever did it.-### Two workers on one machine can no longer fire the same routine twiceEvery process that claims work from the shared queue named itself after its hostname, and the queue@@ -1167,7 +1069,6 @@ always carries a random suffix, so a second process is a different claimant. Ithostname, so a stuck claim still traces back to the machine holding it, and a blank `HOSTNAME` is nolonger read as a name — which had made every replica in a deployment share one.-### The desktop app writes its `.env` readable only by its ownerThe desktop `.env` holds `KEY_ENCRYPTION_KEY` and every minted token, and those are now long-lived:@@ -1176,7 +1077,6 @@ umask (`0644`), so on a shared macOS or Linux machine another local user could rdisk. The file is now narrowed to `0600` after it is written. Windows has no equivalent mode and itssingle-user desktop profile is already the boundary, so the change is Unix-only.-### The desktop app stops adding a banner to `.env` on every start`env::write` keeps the lines it did not write, and its own header comment is one of them, so each@@ -1185,7 +1085,6 @@ copies of "Written by OpenBot Desktop" and fifty blank lines stacked above its sis now recognised and replaced rather than kept, and comments somebody else put in the file are leftalone exactly as before.-### Starting the desktop app again keeps the secrets the first start generatedThe shell generated a fresh set of secrets every time Start was pressed, including the@@ -1197,7 +1096,6 @@ already carries are now kept, and only generated when there is nothing usable topublished in this repository does not count, and neither does a `KEY_ENCRYPTION_KEY` the server wouldrefuse to start on.-### The desktop app refuses a deployment download that writes outside its own directoryThe shell fetches the release tarball and lays it out under the directory it manages. The check that@@ -1207,7 +1105,6 @@ started with the root and still landed outside it. An entry has to begin with adeployment wants, which `app` does, so the file filter did not stop it either. Every component of apath inside the tree is now required to be an ordinary name, and the traversal is refused by name.-### A credential pasted with a stray space into the desktop setup screen now worksThe setup screen enables its button on `value.trim() !== ""` and then sends the untrimmed string, so@@ -1217,19 +1114,16 @@ entered on the same screen were not, so Compose passed the space through, the prcredential, and the failure the person saw named neither the space nor the field. All four are nowtrimmed the same way.-### Example LangGraph and Mastra Bots no longer bind an ephemeral port on empty `PORT=`An empty `PORT=` in compose or `.env` used to become `NaN` for those two example processes, so they listened on a random port while docs still named 4300/4400. They now use the same `listenPort` helper as `agent-bot`: empty is the documented default, and a prefix typo refuses to start.-### An empty app port is the default, not a random one`APP_PORT=` and `SERVER_PORT=` in a compose file or leftover `.env` used to become `NaN` for the Vitedev and preview servers, so the UI bound an ephemeral port while the proxy target was `http://localhost:`.Both empty values now mean the documented defaults (3010 and 3001), and a non-numeric value refuses to start.-### The trail says what started a run, not only whose authority it hadA routine runs as the person who set it up, and a Bot handing work to another Bot runs as the person@@ -1254,7 +1148,6 @@ by itself.Nothing about existing rows changes. Every row already written, and every row a person's own clickwrites from now on, reads as a person, because that is what it was.-### The live screen and live channel updates work again, and one request can no longer end the appThe app opened its two WebSockets against its own address, so they travelled through Vite's `/api`@@ -1268,10 +1161,8 @@ through the proxy and the proxy no longer offers to carry one. In production theapp itself and answers the upgrade on the same origin, so the sockets stay on the browser's ownhost, which is what an ingress terminating TLS on 443 requires and a fixed server port would break.-## 0.0.8-### A desktop shell that installs OpenBot and then becomes itOne window, on macOS, Windows and Linux. From a machine with nothing on it, one click fetches the@@ -1284,7 +1175,6 @@ quitting stops what starting started. Windows is told what it needs before anythits kernel, virtualisation and administrator rights are each named with the command that fixes them,because none of them is something OpenBot can fix on somebody's behalf.-### A title or name that is only spaces is refused rather than storedSandboxed components, skills, custom servers and the component catalogue each accepted a title, slug@@ -1292,7 +1182,6 @@ or id made entirely of whitespace, and stored it. What came back was a row nobodin the catalogue's case, entries that were empty strings. Each endpoint now refuses those the way itrefuses a missing field, and the catalogue drops empty entries instead of keeping them.-### A malformed socket message no longer takes the screen downThe live screen and the channel socket both assumed every payload they received was an object of the@@ -1300,7 +1189,6 @@ shape they expected. Anything else threw inside the handler, which in a browserstops updating with nothing on screen to say why. Both check the shape first and ignore what does notmatch.-### The worker refuses a bad server URL when it starts, not when it first needs it`SERVER_INTERNAL_URL` was read as a string and used as one. A value that was not an http or https URL@@ -1308,43 +1196,36 @@ failed later, inside whichever call happened first, with an error about that calthe setting. The worker validates its environment up front now and normalises the URL, so a typo stopsit immediately and says which variable is wrong.-### A sandbox template that is not valid JSON is reported as a sandbox errorA malformed template surfaced as a raw parse error from whatever tried to read it. It is a`SandboxError` now, which is what the caller already handles and what the surface already knows how toshow.-### A duration written in capitals is understood`durationMs` accepted `30s` and refused `30S`. Units are read case-insensitively now.-### The supervisor refuses a `HostPort` that is not a portThe value went through a bare `parseInt`, so `80abc` became 80 and an out-of-range number was used asgiven. Digits and range are checked, and anything else refuses to start rather than binding somewherenobody asked for.-### The audit endpoint refuses a bad date or limit instead of guessingAn unparseable `from` or `to`, or a `limit` that was not a whole number, was coerced and the query ranagainst whatever came out. All three are parsed strictly now and a bad one answers 400.-### Chat content that does not parse is dropped rather than thrownOne malformed entry in a user message threw and took the whole message with it. The malformed part isdropped and the rest is delivered.-### The example package is read by a path the host acceptsIt was addressed by a URL pathname, which is not a path on Windows. It is read by a real path now.-### The app can be served from a built bundle, and listens on both loopbacksThe app's only start script was `dev`, which runs a Vite dev server. That meant `NODE_ENV=development`@@ -1355,7 +1236,6 @@ and a machine with Bun and no Node exits 127 without saying why. Vite is also tosince Node resolves `localhost` to `::1` and Bun to `127.0.0.1`, so binding one of them left whoeverasked for the other looking at nothing.-### A person's Stop is recorded as a stop, not as a failed actionPressing Stop mid-action aborts the request, and the gateway wrote that outcome beside the decision@@ -1366,7 +1246,6 @@ type, the natural way to watch for outages, read every Stop as one. A stop now wit with the other did-not-happen outcomes, and a policy dry-run skips it the way it skips a failure,since both sit beside a decision row that is already scored.-### A malformed `DATABASE_URL` is refused without printing the password`DATABASE_URL` is taken apart before it reaches Bun, and the string most likely to fail that parse is@@ -1375,7 +1254,6 @@ string back to name the fault, which wrote the database password into the log liIt now names the variable and the shape it expects, the way the other refusals beside it already do,and never echoes the value.-### Wiping a computer is recorded even if clearing its stored state failsResetting a computer destroys the profile first and wrote the audit row last, after two Postgres@@ -1385,7 +1263,6 @@ login gone, no undo -- with nothing on the trail to say who wiped it or when. Thas soon as the profile is gone, which is the point after which nothing can be put back. A failure ineither delete is still reported to the caller.-### Pressing Stop is recorded as a stop, not as a computer that is not runningThe computer transport answered a Stop correctly only when it arrived before the request left. The@@ -1396,7 +1273,6 @@ was not running, and the gateway wrote that sentence into the action's audit rowso a deliberate stop read back as an outage. A genuinely unreachable computer and a timeout stillsay what they said.-### A component the server refused is no longer drawn anywayA sandboxed component asks the server at call time whether the Bot may still use it, and a refusal@@ -1407,7 +1283,6 @@ found, and the component rendered as though it had been allowed -- so revoking aBot did not take effect on screen until the five-second grant poll caught up, and a failed decisionrequest showed nothing at all.-### A component whose name has a stray space is the same componentThe catalogue announcement asked whether each component's `name`, `title`, `kind` and `description`@@ -1418,7 +1293,6 @@ second catalogue row beside `weatherPanel`: published, ungranted by anybody, anda Bot back from under the name people use. The four fields are now stored as the strings the guardapproved.-### A password with a `%` in it says so, instead of failing as `URI error``DATABASE_URL` is taken apart before it reaches Bun, and each part is percent-decoded. A part@@ -1430,12 +1304,10 @@ wrong, and that a literal `%` must be written `%25`.A correctly encoded password is unaffected.-### An empty Bot `PORT` is unset, so NaN never reaches Bun.serve`PORT=` on `agent-bot` and `agent-langgraph` used to parse as `NaN` (`??` does not treat empty as absent) and `Bun.serve` bound an ephemeral port while compose still published 4200/4201. A prefix typo (`42o0`) started on 42. Empty now means the shipped default; anything that is not a whole port number refuses to start.-### The server connects to Postgres on Windows, and `localhost` is no longer a coin tossTwo separate faults, both of which stop a deployment reaching its own database and neither of which@@ -1456,14 +1328,12 @@ to the other, so the same configuration worked on one machine and failed on thenothing in the error mentions. Every loopback port is now published on both addresses. Both areloopback, so nothing became reachable from another host.-### A bad `COMPUTER_MEMORY_BYTES` refuses to start the supervisor, instead of capping a computer at 512 bytes`COMPUTER_MEMORY_BYTES=512m` used to parse as `512` via `parseInt`, which Docker accepts as a memorycap Chromium cannot live in. Empty `COMPUTER_MEMORY_BYTES=` is still unset (no cap). A value that isnot a whole number of bytes now exits before any computer is created.-### An IPv6 address in `AGENT_ENDPOINT_ALLOWED_HOSTS` now matches however it is writtenThe endpoint check compares the list against the address as the URL parser spells it, compressed@@ -1475,7 +1345,6 @@ round. Bracketed entries are now stored in the parser's spelling, with the portcompared with their brackets on; an entry the parser does not read as an address is refused at boot,naming the entry, as a URL or a wildcard already was. Names and IPv4 entries are unaffected.-### A Bot's own decline is only recorded against a Bot the caller may reachA Bot reports that it declined a request through the person's session, and the audit row says@@ -1485,7 +1354,6 @@ against any coworker, one they cannot see included, and an administrator readingit for something the Bot said. The route now asks the store first, as every other route on a Botdoes, and answers not found for a Bot the caller cannot reach, writing nothing.-### The engine socket the supervisor is given can be pointed somewhere elseCompose mounted `/var/run/docker.sock` into the supervisor as a fixed path. That is correct for@@ -1498,7 +1366,6 @@ with a message about not reaching Docker.The mount source is now `ENGINE_SOCKET`, defaulting to `/var/run/docker.sock`, so nothing changesunless it is set. On rootless Podman on Linux, set it to `$XDG_RUNTIME_DIR/podman/podman.sock`.-### A Bot's computer is waited for properly on Podman, and the supervisor can reach the engine thereTwo things stopped OpenBot running on Podman, which nothing had tried before.@@ -1517,10 +1384,8 @@ arrived at a port nothing was listening on and came back as a computer that is nsupervisor now states the healthcheck when it creates a computer instead of inheriting it, soreadiness no longer depends on how the image was built. On Docker the behaviour is unchanged.-## 0.0.7-### The published service images are zstd rather than gzipAn image pull was already a compressed transfer, so this is not compression where there was none; it@@ -1534,7 +1399,6 @@ This applies to the five `ghcr.io/copilotkit/openbot-<service>` images and not tcurrent containerd do; the single image is pulled by deployments running whatever they have, so itstays gzip.-### A release publishes every service's image, not just the one`ghcr.io/copilotkit/openbot` was the only image a release produced, so anything running the Compose@@ -1556,7 +1420,6 @@ CI now builds those five Dockerfiles too. Nothing did before, because they are b`docker compose up --build`, which only the smoke journey runs and which cannot run in CI, so abroken one surfaced during a release after the first image had already been pushed.-### A skill can be written in the conversation instead of retyped into a formA skill is four fields and an instruction a Bot follows, and the instruction is the one that decides@@ -1578,7 +1441,6 @@ The tools run in the browser as the signed-in person, over the same `POST /api/pperson uses, so who may take a slug is answered the same way and the `configuration.changed` auditrow is written the same way.-### Resetting a computer clears the activity paneResetting a Bot's computer deletes the browser profile it worked on, but the activity pane beside the@@ -1587,7 +1449,6 @@ Bot did afterwards with nothing to tell the two apart, and only reloading the papane now forgets a Bot's history when that Bot's computer is reset. Stopping a browser is unchanged:the profile survives a stop and is meant to resume, so what it did is still true of the machine.-### A person can set standing instructions that every coworker followsSettings now has a box for standing instructions: one piece of text per person, saved once and@@ -1605,7 +1466,6 @@ as before.This adds migration `0026_user_instructions`, which creates one table.-### A coworker can be made in the conversation, and it starts able to reach nothingA coworker without an endpoint runs on its role description, which becomes the standing instruction@@ -1626,7 +1486,6 @@ Like the skill tools above, these run in the browser as the signed-in person oveperson uses, so who may create a coworker is answered the same way and `bot.created` carries theactor.-### A conversation has a name of its ownA channel's name was only the names of the Bots in it, so asking one Bot about six unrelated things@@ -1644,7 +1503,6 @@ It runs on the work queue rather than as a headless turn, so naming a conversatiIntelligence thread lock and cannot refuse somebody's own next message with a 409. A deployment withno model key names nothing and carries on.-### The trail says who a coworker was opened toMaking a coworker public admits every signed-in person to it, and being admitted to a coworker is@@ -1655,7 +1513,6 @@ the whole deployment was byte-identical on the audit page to one that correctednow carry the visibility, on every edit rather than only the edit that moved it, so reading the trailforward says who could reach each coworker at any point.-### Duplicating a Bot in the box keeps its instructionsA coworker that runs on this deployment's own Bot has no endpoint — it has a prompt, which is the@@ -1669,7 +1526,6 @@ box, which also means it can still be granted the right to hand work on — writcoworker it could never hold that grant, however the original was set up — and copying one no longerneeds a managed Bot to fall back to.-### Hiding a coworker no longer hides the grants pointing at itHiding a coworker is a preference about your own roster — one row per person — and the grants saying@@ -1682,7 +1538,6 @@ A coworker you have hidden now appears in that list when a grant already pointshidden from your roster, so it can be switched off. One you have hidden with nothing granted to itstays hidden.-### A tool call that failed no longer reads as one that workedThe audit page draws a row it does not recognise as `Allowed`, which is right for the many rows that@@ -1695,7 +1550,6 @@ this path every time somebody's token expires, so this was the most common failuand the one the trail was quietest about. Both now read as `Did not happen`, and both are in thatsaved view. Neither is filed as a refusal: nothing was forbidden on either row.-### A blank agentId on a channel activity says which field was wrong`POST /api/channels/:id/activity` accepted an `agentId` of only spaces, trimmed it to nothing, then@@ -1703,7 +1557,6 @@ looked that up and answered `404 Agent not found`. The field was malformed rathemissing, so the answer sent whoever was integrating to look for a coworker that was never named. Itis now a `400` naming the field, which is what the same endpoint already did for malformed text.-### Audit payloads are redacted by the store as well as by its callerRedaction of secrets out of audit payloads happened in `recordAuditEvent`, and every caller in the@@ -1712,7 +1565,6 @@ was handed, so a future direct caller would have written secrets to the audit ta`insert` now redacts too. Redaction is idempotent, so nothing about the existing path changes; thisis the floor under it rather than a fix to it.-### A stray space in NODE_ENV no longer lets the public example key throughA deployment that never changed `KEY_ENCRYPTION_KEY` encrypts its credential vault with the key@@ -1723,7 +1575,6 @@ invisible: Docker's `env_file` preserves it and so does every hosting dashboard`NODE_ENV=production ` tripped one refusal, slipped past the other, and started the deployment on thepublic key with only a warning at boot. Both gates now ask the same question the same way.-### A tool result from an MCP server with an empty part no longer crashes the turnReading a tool result cast every part to an object and asked for its type. A `null` or missing entry,@@ -1732,10 +1583,8 @@ tool failed. Such a part is now named `[unknown]`, which is the same naming-rathsurrounding code already does for parts it does not recognise, so the rest of the result stillreaches the Bot.-## 0.0.6-### Setting up needs one Intelligence credential, not two`COPILOTKIT_LICENSE_TOKEN` is no longer required. Managed Intelligence derives entitlement from the@@ -1749,7 +1598,6 @@ The Helm chart failed harder than the docs did: it *required* `secrets.licenseTomanaged-Intelligence install was refused at `helm install` rather than merely misdocumented. Thatvalue is now optional.-### Duplicating a coworker keeps the endpoint it was copied fromDuplicate used to point every copy at this deployment's own managed Bot, whatever the coworker being@@ -1764,7 +1612,6 @@ mean rotating either one's key silently changed the other's, so a copy starts wideployment with no managed Bot configured, duplicating a coworker that brought its own endpoint nowworks instead of being refused with advice to give it an endpoint it already had.-### Connecting an account survives a vendor that is downFinishing a connection used to end on a blank server error if the vendor could not be reached at the@@ -1785,7 +1632,6 @@ difference lives. Three lines to look for: `oauth-token-endpoint-unreachable` an`oauth-token-endpoint-unusable`, means the fault is this deployment's catalogue rather than thevendor.-### A custom MCP server token is sent with the scheme it namesEvery token stored against a custom MCP server went out as `Bearer`, whatever the vendor asked for.@@ -1797,7 +1643,6 @@ A token that begins with `Basic ` or `Bearer ` is now sent as written, so pastevendor gives you, scheme and all. A bare token is still sent as `Bearer`, so nothing alreadyworking needs to change.-### A conversation is no longer stuck after a tool call went unansweredA tool that runs in the browser can be torn down while its call is still open, most often because@@ -1818,7 +1663,6 @@ it. Ids are never rewritten and the stored thread is untouched, so the transcriphappened and a call waiting on a resume still gets its result. The same filter also covers a Botanswering a relayed question, whose seeded conversation kept every tool call and no tool result.-### Reopening a channel no longer hides the end of the last conversationOpening a channel joins the realtime gateway, and the snapshot the join returns can lag the durable@@ -1830,7 +1674,6 @@ The stored thread now wins whenever it holds more than the channel does and holdchannel already shows. A message typed while history is still loading is not in the store yet, anda run still streaming has messages the store has not seen, so neither is rolled back.-### The transcript stays with the question when an answer starts arrivingSending a message carried it up to the top of the view, correctly, and then the Bot's first token@@ -1840,7 +1683,6 @@ scrolled, so every answer began with a scroll back down to find it. The transcriquestion in place for the whole turn, and the scroll that puts it there is animated rather than ajump — unless the reader has asked their system for reduced motion.-### A conversation started from the sidebar is recorded like one started from the home screenThe trail had a `channel.routed` row for every conversation begun in the home composer — the@@ -1848,7 +1690,6 @@ coworker it went to and why, whether inferred or named with `@` — and nothingfrom the sidebar's +, a coworker's card or its profile, which read exactly like a row that failedto write. Picking a coworker in that To: field is now recorded the same way an `@` is.-### A browser clock that runs ahead no longer hides what a routine saidThe roster line and unread dot for a channel are moved by the last report that arrived, and only@@ -1857,7 +1698,6 @@ report from a correct clock — a routine's reply, a relayed handoff answer, ano— was dropped without a word until the real time caught up: the reply was in the thread, and theroster never said so. A reported time is now capped at the server's own clock.-### An empty supervisor PORT is unset, not an ephemeral bind`PORT=` left blank in compose or a `.env` used to reach `Bun.serve` as `NaN`, so the supervisor@@ -1865,7 +1705,6 @@ bound a random port while the published mapping still pointed at 4300. Empty now4300; a non-numeric or out-of-range value refuses to start instead of binding port 30 from a typolike `30o0`.-### Tool arguments and results are redacted in the audit trail whatever their spellingThe sensitive-key list redacted `tool_result` and `tool_arguments` but not `toolResult` and@@ -1875,7 +1714,6 @@ carried both spellings. New rows are redacted; rows already written are not rewrdeployment that has been running MCP or computer tools still holds unredacted arguments and resultsin its existing trail.-### An IPv6 address listed in `AGENT_ENDPOINT_ALLOWED_HOSTS` is now actually allowedA bracketed IPv6 host was normalised one way when the list was read and another way when an endpoint@@ -1883,7 +1721,6 @@ was checked, so the two could never match: `[::1]:8080` became `::1]:8080` on on`::1:8080` on the other. Registering a Bot at a listed IPv6 address was refused as a privateaddress anyway. IPv4 was unaffected.-### An empty `PORT` no longer starts the server on a port nobody asked for`PORT` and `SERVER_PORT` name one number, and either is meant to move the server. A `PORT` that was@@ -1894,14 +1731,12 @@ that polls `SERVER_PORT` reported it had never started. An empty value now countevery other setting already treats it, and a value that is not a whole port number (`30o1` used tostart the server on port 30) refuses to start instead.-### `TRUSTED_ORIGINS` falls back to the port the app is actually served onUnset, it fell back to `http://localhost:3000`, while the app is served on 3010 everywhere else inthis repository. `appUrl` reads the first trusted origin, so a deployment that left the variableblank built OAuth redirects against a port nothing was listening on.-### Coworkers are made in a wizard and managed in a dialogCreating a coworker is now a three-step wizard — who it is, who may see it, then where it runs,@@ -1912,7 +1747,6 @@ hide/duplicate/delete verbs — usable on a phone, where the old side panel hidpanel beside a conversation slims down to who-you-are-talking-to plus two buttons: start a newchannel, or open that dialog.-### Routines live on each coworkerThe sidebar's global Routines entry is gone; a coworker's routines are a section of its own dialog,@@ -1921,7 +1755,6 @@ Each routine row now wears its state as chips — the channel it posts to, how teither the next run, **Paused**, or **Due** when a firing is waiting on the sweep (which used torender as "Next 5 hours ago").-### A built-in coworker is no longer asked for credentials it will never needA coworker running on the deployment's own Bot was nagged for a callback token and shown an@@ -1930,7 +1763,6 @@ to authenticate. In the same spirit, the handoff panel explains once when a cowowork on (it runs as its own agent, outside this deployment's loop) instead of offering switches theserver can only refuse; its existing grants stay visible so they can still be revoked.-### A conversation is given a name of its ownA channel was labelled with the names of the Bots in it, so every conversation with the same@@ -1940,7 +1772,6 @@ question that prompted it. Once a conversation has an opening exchange, the deplis asked for a few words naming what it is about, and the roster draws those words in place of thepreview. A deployment with no model key configured names nothing and looks exactly as it did before.-### A channel stops showing a working indicator once its turn has endedSending a message and then opening another channel before the reply arrived left the first channel@@ -1953,7 +1784,6 @@ working. The report now belongs to the turn instead of to the screen, so the charunning is the channel told when it stops. Opening a channel also no longer announces that it isidle twice before anything has run in it.-### A hop the boundary refused now names the Bot that was refusedThe audit page renders its Bot column from `payload.bot` and nothing else. `agent.handoff_offered`@@ -1963,7 +1793,6 @@ Bot belongs. Those are the rows somebody actually opens the trail for: a hop thain the transcript anyway, and a refused or lost one is visible nowhere else. All four now name theasking Bot, exactly as the accepted pair and `agent.escalated` already did.-### A failed tool refresh no longer leaves a connector offering nothingRefreshing a connector's tools replaced the list with a delete and then an insert, as two separate@@ -1975,7 +1804,6 @@ none of that vendor's tools. Nothing brought them back until somebody read the epage and pressed Refresh. The two statements are now one, so a bad refresh is recorded and the toolsalready held are left alone, which is what the code always claimed to do.-### A rule tried in dry-run now says what it would have refused a Bot's tools`dry-run` exists so a boundary can be measured against live traffic before it starts refusing@@ -1987,7 +1815,6 @@ refusing Bots with nothing in the trail to have warned anybody. A refused tool cwhatever the mode does with it, carrying `carriedOut` so a reader can tell a call this deploymentstopped from one dry-run recorded and let past. Enforcing deployments behave exactly as before.-### A policy dry-run no longer counts a failed action twice, or invents a change it did not makeTesting a boundary against recent history replayed three kinds of audit row, and one of them is a@@ -1998,7 +1825,6 @@ decision row saying "refused" and the failure row reading as "allowed", so a canrefused it identically was reported as a new refusal it never introduced. The replay now scores eachaction once, from the row that recorded its decision.-### A message no longer routes to a specialist because a longer word contained a connector's nameWhen the intent router falls back — it is unreachable, or it declines — and exactly one coworker can@@ -2008,7 +1834,6 @@ giraffe) matched **jira**: a message naming neither system was pinned, for the lto a specialist that could not answer it. A connector's name now has to appear on a word boundary,so a system named on its own still routes and one buried inside another word does not.-### The audit page no longer says "Allowed" about six kinds of refusalA hop one Bot was not allowed to make, an endpoint this deployment would not dial, a rotation the@@ -2019,7 +1844,6 @@ page recognised six refusal types and the six added since were never added to itas refusals, the two saved views are built from the same lists the rows are labelled from, and arefusal added later is added in one place or in none.-### A conversation deleted while a server was reconnecting no longer lingers on the screenAnnouncements between servers travel as Postgres notifications, which reach whoever is subscribed at@@ -2034,7 +1858,6 @@ showing a conversation that had been deleted until the page was reloaded.A server now tells the browsers it is holding to refetch when its subscription is re-established.Nothing to configure, and no change for a deployment whose database connection never drops.-### A Bot can answer with a picture instead of describing oneAsk for a chart and a Bot replied in prose, or handed back a fenced block of HTML for somebody to@@ -2042,7 +1865,6 @@ read instead of look at. Set `OPENBOT_GENERATIVE_UI=true` and it may answer withwrites itself, drawn in the transcript. Off unless asked for, and deliberately so: it runs code amodel wrote, so a deployment acquires the capability by choosing it rather than by upgrading.-### The sidebar collapses, and the roster is reachable on a phoneThe sidebar could always collapse, but nothing in the app ever drew the trigger. The only@@ -2054,7 +1876,6 @@ already draws, with Cmd/Ctrl+B on it, and the collapsed choice survives a reloadwritten to a cookie that nothing ever read back. Fifteen screens that previously drew no header bargain 40px above their heading, because the toggle has to sit at the pane's edge.-### A button drawn as a link answers the keyboard and announces itselfSix controls navigate rather than submit, so they render a router link through the shared button:@@ -2064,7 +1885,6 @@ button that draws on five routes. Base UI was told each was a native `<button>`,needs: the `role="button"` that tells a screen reader what the control is, and Space-keyactivation, which a `<button>` gets from the browser and an anchor does not.-### A Bot's answer comes back to the conversation that asked**This reverses what 0.0.5 shipped.** The 0.0.5 notes below say the asking Bot does not relay text@@ -2075,7 +1895,6 @@ relays what came back — attributed by name — into the conversation you are wis the asking Bot's account of the answer rather than the answer verbatim; very long answers areclipped to keep the relay itself from failing.-### Channels say when a Bot is working in themA channel whose Bot is mid-turn shows a working indicator on its roster avatar — including turns no@@ -2083,20 +1902,17 @@ browser started, such as a handoff running on the server or a routine. An open cpicks up turns that arrived while nobody here streamed them, so a relayed answer appears withoutleaving and coming back.-### First sign-in gets an onboarding wizardA new person lands in a short welcome wizard before the app; everyone who signed in before thisupgrade is stamped as already onboarded by the migration and sees nothing.-### Shift+N starts a new chat from anywhereBound across the signed-in app, shown under **Settings → Keyboard shortcuts**, and inert while youare typing in a field. Handoff work is also picked up the moment it is queued rather than at thenext poll, so an answer's round trip no longer pays up to two seconds per leg.-### A Bot's shell can no longer reach the embedded database without a passwordIn the all-in-one image the cluster was `trust`-auth on loopback, and the Bot's shell runs in the@@ -2106,14 +1922,12 @@ the policy store, and the credential vault as the instance owner. The cluster noover the container environment. The shell has no way to learn it, so the connection is refused. Anexternal `DATABASE_URL` deployment is unaffected.-### A live screen that ends says so, instead of freezing the last frameWhen a Bot's live screen ended — the computer stopped, or the socket failed — the message explainingwhy was drawn only by a component the take-the-wheel view does not mount, so the screen sat frozen onits last frame with nothing said. The reason is now shown where the live screen is.-### A Bot's browser drops the automation flags a person needs gone to sign inThe browser announced itself as automated (`navigator.webdriver`, the enable-automation switch),@@ -2122,7 +1936,6 @@ at the source — the browser flag, not a script that patches `navigator.webdrivother tells. A headless build still reports `HeadlessChrome` in its user agent, which only runningheaded under a virtual display removes; that heavier change is tracked separately.-### An empty model reply, or a run with no question, no longer ends in silenceTwo failures on strict OpenAI-compatible providers (z.ai GLM, Anthropic): a follow-up run that@@ -2131,7 +1944,6 @@ call ended the run with nothing on screen. A run with no human turn now carriescontinuation, and an empty reply ends on a visible line rather than in silence. OpenAI, whichtolerated both, is unchanged.-### Embedded PostgreSQL initialises on a platform volume, and says so when it cannot`EMBEDDED_POSTGRES=on` could not create its cluster on a platform whose persistent volume is an ext4@@ -2161,7 +1973,6 @@ cluster and is not empty is now refused with a sentence naming the mount to useReported by [@jerelvelarde](https://github.com/CopilotKit/OpenBot/issues/269) with the container logsfor both mount paths, which is what made the two failure modes separable.-### Turning on network policies no longer leaves the culler pod open`networkPolicy.enabled` rendered policies selecting the server and the computers. The culler@@ -2173,7 +1984,6 @@ delete Sandboxes. It has a third policy now, narrower than the other two, and`networkPolicy.cullerExtraEgress` narrows its database egress separately from `extraEgress`. Arender that leaves any pod unfenced is refused by the chart checks.-### A sign-in a site opens in a new window is shown, and can be clickedA Bot's browser was bound to the page it launched with, and to nothing the site opened afterwards.@@ -2190,7 +2000,6 @@ cannot act on the wrong document.Nothing to configure.-### Stopping a Bot's computer stops it, and the person watching is toldA computer somebody stopped came back up on its own about a second later, and reset did the same. The@@ -2215,10 +2024,8 @@ not yet show that message**: it arrives at the browser and is held in state theread, so a person still sees the last frame until they reopen the screen. That half is trackedseparately in #287.-## 0.0.5-### One Bot can hand work to another, and reach a person when no Bot will doA Bot asked something it is not the right Bot for can now put the question to one that is. The@@ -2249,7 +2056,6 @@ that took three attempts reads there as having been asked three times.No new tables: this uses the work queue that already fires the culler.-### A conversation that used a tool no longer stops answering for goodA channel could reach a state where every turn in it failed and the only thing it said was@@ -2267,7 +2073,6 @@ A result now only answers a call it follows. One that arrives early is moved tokeeping what it actually said, and a result whose call is nowhere in the thread is dropped. Affectedconversations start answering again on their own; there is nothing to run and nothing to reset.-### The framework Bot answers on 5.6-tier models, and can be told how hard to thinkPointing `BOT_MODEL` at a `gpt-5.6-*` model gave a Bot that started, reported healthy, and then said@@ -2282,7 +2087,6 @@ have, or one set where it cannot be sent — a provider that is not OpenAI, or aResponses API — stops the Bot at startup with a message naming what to change, rather than startingwith a setting that goes nowhere.-### A Bot can be asked to do something on a schedule"Every weekday at nine, post the standup notes here" is now something a Bot can be asked rather than@@ -2303,7 +2107,6 @@ than accepting one it cannot attribute. `scripts/start.sh` runs the worker localturns it on with `routines.enabled` and takes the secret as `secrets.workerSharedSecret`. No new portis opened for any of this — the worker only ever calls out to the server it already trusts.-### Turn screenshots are swept in every deployment, not oneA page a Bot opens is photographed and kept in `computer_page_frame`, so a conversation read back@@ -2328,7 +2131,6 @@ have been enforcing, but it has never been applied anywhere else, so it is worthupgrade rather than after. It is drained in batches, forty thousand rows an hour, rather than in onestatement.-### A channel a Bot has spoken in unseen shows a dotThe sidebar marks a channel when a Bot has said something since you last had it open: a dot beside@@ -2338,7 +2140,6 @@ the membership row like the pin — so one person reading does not clear anybodyThe deployment gains one nullable column, via migration `0019`.-### The API can reach Intelligence and sign-in when a NetworkPolicy is on`networkPolicy.enabled` wrote a rule for the API server that named DNS, the database and the Bots'@@ -2359,7 +2160,6 @@ destination and so permitted everything. That rule now covers the API server alo`networkPolicy.kubernetesApiCidr` narrows it to your cluster's service range; left empty it stays asit was, because a chart cannot know that range.-### Taking the wheel stops the Bot's shell, not just its clicksWhile a person held the wheel the Bot was refused on the page, and not in the shell. `/exec` and a@@ -2375,7 +2175,6 @@ to be able to say what it was doing.Nothing to configure. A Bot that acts during a takeover gets the refusal it already got for a click,and the trail records the attempt and the failure the same way.-### A computer that was suspended once suspends againScale-to-zero worked once per Bot. A computer suspended, resumed, used and then left alone again was@@ -2393,7 +2192,6 @@ the same clock the offer runs on, so a Bot cannot come back round as idle untilNothing to configure, and the sweep already runs on a schedule. A deployment where each Bot has itsown computer stops paying for browsers that were used once.-### A Bot's egress proxy is reachable on Kubernetes, or the install is refusedThe chart named no egress variable anywhere, so a Helm deployment read the per-Bot proxy settings@@ -2408,7 +2206,6 @@ addresses and nothing else, which is almost no proxies: they sit on a private ad`networkPolicy.computerExtraEgress`, rather than found later as a Bot that fails on every page.Nothing changes for a deployment that sets no proxy, or one that already opened a path to it.-### A finished turn shows the page it opened, not the one open nowReopening a conversation made every past turn fetch the screen as it is now, so an answer about@@ -2441,7 +2238,6 @@ wait there for ever. And opening one full size shows that same kept frame, withwheel: zooming a past turn used to mount the socket and offer Take control, so the one gesture forlooking closer at what a turn did replaced it with whatever the Bot has open now.-### A conversation keeps the browsing that produced its answersEvery turn in which a Bot used a tool was disappearing from the transcript on reload. The sentence@@ -2459,7 +2255,6 @@ The two spellings are now read as the same thing. The check stays for turns thatmalformed, and a mixed or unrecognised array is still refused rather than half-translated, because areader that rewrites what it does not recognise is worse than one that refuses it.-### Run this on KubernetesA Helm chart under `charts/openbot`, Bots and all, and the fixes that installing it for real turned@@ -2539,7 +2334,6 @@ shape it was written for, the check that stops a ref from a replaced computer relive one was silently absent. It now asks the supervisor when it does not know, by listing ratherthan by ensuring, so asking never starts a computer that had stopped.-### A routing trail says why a message was not routed, not only that it was notEvery untagged message writes a `channel.routed` row, and that row carried `fallback: true` for two@@ -2565,7 +2359,6 @@ the model. It is now reported as unparsed, which is what it is.Nothing changes about where a message goes. Every routing decision is the same decision it was.-### Notion joins the connector catalogueNotion is now a governed MCP connector, reached through Notion's own hosted server on the@@ -2585,7 +2378,6 @@ its tools are granted to it. Setup is enable at `/admin/plugins/notion`, connect`/admin/plugins/notion` grants a batch of tools to a batch of Bots in one pass, one grant and oneaudit row per Bot per tool. No migration.-### Refresh tokens rotate in place, and replicas take turns spending themA vendor that rotates refresh tokens invalidates the one it just handed out, so two replicas racing@@ -2595,7 +2387,6 @@ token now locks the credential's vault row for the length of the exchange, so arather than races, and the rotated token is written back in the same transaction that held the lock.Nothing to configure; a connection just stops going stale under concurrent traffic.-### An MCP token is spent only by its own server, and only at the address it was givenPointing a server at a credential is the one place this deployment takes a reference to a stored@@ -2635,7 +2426,6 @@ Curated servers keep working as they did. Their URL comes from the catalogue ratrequest, and a per-instance hostname is matched against the vendor's own anchored pattern beforeanything is stored, so re-adding one cannot point it at an address of the caller's choosing.-### A configured egress proxy reaches the browser that uses it`EGRESS_PROXY_DEFAULT` and `EGRESS_PROXY_<BOT>` were documented as the way to give a Bot a stable@@ -2658,7 +2448,6 @@ proxy URL can carry a password.**Move these two out of `.env` and into `egress.env`.** In `.env` they reach no process.-### Screens without a conversation stop polling for a Bot that does not existEvery surface asks which components its Bot holds, and asks again every few seconds so a revoked@@ -2674,7 +2463,6 @@ that matters is invisible.The grant queries now wait for a surface to declare a real Bot, and simply do not run while theplaceholder holds. Conversation surfaces — the Bot page, channels — declare one and are unchanged.-### A rule can be tested against history before it is savedA boundary was written blind: an administrator typed a CEL rule, saved it, and learned what it@@ -2691,7 +2479,6 @@ Replay, not simulation: the context is rebuilt from the audit row exactly as thedecision time, through the same helpers, so a rule behaves here as it will behave live. The scan isbounded and biased to recency, and the reply says how many rows it covered.-### A browser refusal names the element againEvery browser context carries a neutral all-empty `mcp` object, so a rule naming `mcp.effect`@@ -2704,7 +2491,6 @@ The branch now keys on the server and tool being named, which a real tool call aclick reads "“Submit order” on shop.example is blocked by the rule ..." again, which is what the Botrelays to the person asking.-### Knowledge searches instead of guessingA package can say which of its skills each coworker gets, and the fintech example gives Knowledge the@@ -2729,7 +2515,6 @@ A coworker naming a skill its package does not ship is refused at load rather thas a channel naming an agent that is not there: a typo that silently attaches nothing looks exactlylike working.-### A Bot's computer is no longer on the same network as the databaseCompose declared no networks, so every service shared one and reached the others by service name.@@ -2757,7 +2542,6 @@ This does not reach back in time. A deployment that has been running with the twshould assume a Bot could have read or written the database, and look at the trail with that inmind.-### A credential in an MCP server address is refused in the query and the fragment tooRefusing `https://user:[email protected]/mcp` closed the userinfo spelling of a credential in the@@ -2785,7 +2569,6 @@ it read as an ordinary vendor name. The long spelling was only ever refused inci`.internal` rule. Both are now named, so the address this check was written for is refused on purposerather than by luck.-### A curated MCP server is pointed at its own kind of credential tooAdding a server by URL was made to check which credential it is being pointed at. Adding one from the@@ -2815,7 +2598,6 @@ behind with nothing pointing at it and nothing to revoke it, and everybody who haccount was told the deployment has no client registered. A re-add that names no credential nowleaves the one that is there alone.-### Name the private addresses an agent may live atRefusing `AGENT_COMPUTER_ALLOW_PRIVATE_HOSTS` in production closed a hole and took something with@@ -2840,7 +2622,6 @@ It is narrow on purpose:Unset means none, which is what every deployment has today.-### A custom MCP server can only be pointed at its own tokenAdding an MCP server by URL takes a credential id alongside the address, and the add is what spends@@ -2863,7 +2644,6 @@ the admin screen, which mints a token and points at the one it just made. If a dcustom server pointing at a credential of another kind, adding it again will now be refused, and theanswer is to give the server its own token.-### A failed action is recorded the same way it was decidedAn action the policy allowed and the computer then failed is recorded twice, once for the decision@@ -2879,7 +2659,6 @@ unchanged.Both rows now carry the same subject. Nothing about the boundary moves: the policy decided on acomplete context before and after, and no action is permitted that was not permitted before.-### Upgrading**A deployment that sets `AGENT_COMPUTER_ALLOW_PRIVATE_HOSTS=true` with `NODE_ENV=production` no@@ -2911,7 +2690,6 @@ NAT64 spellings of them — were refused whatever this switch said, before and anot moved. What changed is that it is no longer the only thing left standing in a productiondeployment that copied the example.-### The supervisor answers on loopback, not on every address the host hasThe supervisor's port was published without an interface in front of it, so it bound every address@@ -2928,7 +2706,6 @@ host port.Nothing changes for a default deployment. `scripts/start.sh` already reached it on `localhost`.-### An agent's address is checked where it ends up, not only where it starts`checkAgentEndpoint` decides whether this deployment will dial an address, and the request was then@@ -2951,7 +2728,6 @@ Refusals are now on the audit trail as `agent.dial_refused`, with the address anrefused run already told the person what happened; nothing told the deployment, and an agent thathas quietly started redirecting somewhere it should not is worth being able to count.-### A connector says which of its granted tools it no longer offersA grant names `serverId/toolName`, and a Bot is told about a tool only when the grant and the vendor's@@ -2969,7 +2745,6 @@ stays a decision somebody makes rather than a side effect of a vendor's bad afteNothing changes for a connector whose grants all match its tool list, which is the normal case — thesection is not drawn and no row is written.-### Opening a new chat no longer logs a server errorA thread id is minted before the thread exists — the platform creates it on the first run — so reading@@ -2983,7 +2758,6 @@ A thread the platform does not have now reads as having no messages. A 404 and oa 500, because an outage answered with an empty history would tell the browser the conversation is goneand invite somebody to start it over.-### Reconnecting a live screen no longer stops the screen you just reconnectedA Bot's screen allows one viewer, and opening a second `/stream` replaces the first. The replaced@@ -2997,7 +2771,6 @@ without a word, because the input path looks for a viewer before it looks for anA close now stops casting only when the socket closing is the one that was casting.-### A sidebar channel row can be pinned or deletedRight-click on a channel in the sidebar and a menu opens with two entries: Pin channel and Delete@@ -3016,7 +2789,6 @@ control in the product.The deployment gains two nullable columns, via migration `0016`.-### An MCP server address that points inside the deployment is refused in three more spellingsAdding an MCP server by URL is checked before the address is stored, because that form is otherwise a@@ -3039,7 +2811,6 @@ points at the token field instead.A deployment that reaches its MCP servers by ordinary vendor hostnames sees no difference.-### One unreadable turn no longer takes a whole conversation downRestoring a thread cast whatever the history store held straight to messages and handed it to the@@ -3058,7 +2829,6 @@ nothing accounts for — a turn that silently disappears reads as one that was ncontent and every well-formed tool call are unaffected, and a history that cannot be read at allstill opens the composer rather than blocking it.-### Refreshing no longer flashes white before the theme arrivesA person with the dark theme selected saw a white frame on every reload. The stored preference was@@ -3073,7 +2843,6 @@ one they are, so those match too.No configuration changes and nothing is stored differently; a deployment that was already on thelight theme sees no difference at all.-### `start.sh` refuses a port that answers but is not OpenBotThe startup checks asked whether a port answered, and treated that as proof the port belonged to this@@ -3099,7 +2868,6 @@ are now present, next to each other, saying they have to agree.**A run may now stop where it used to continue.** That is the point: it stops at the port that iswrong, naming it, rather than several steps later on a parse error.-### `docker compose up -d` configures the same stack `scripts/start.sh` does`SUPERVISOR_TOKEN` and `COMPUTER_TOKEN` defaulted to the empty string in `docker-compose.yml`, so@@ -3114,10 +2882,8 @@ and every call between them was refused at the door.Both now carry the same defaults `start.sh` applies, as `COMPUTER_IMAGE` already did two lines down.A value set in `.env` still wins, and a deployment should set one.-## 0.0.4-### A click citing a ref this deployment cannot resolve is refusedA Bot acts on a page by citing a ref from a snapshot, and the server turns that ref back into the@@ -3142,7 +2908,6 @@ the decision row is written, so an action somebody tried to take still appears owere being carried out without the boundary seeing what they touched. A Bot that meets one takes afresh snapshot and continues.-### A package ships its skills, so tool selection works on a cloneTool selection narrows a Bot's tools to the ones its matching skills declare, and a deployment starts@@ -3167,7 +2932,6 @@ refusing to boot over.The example package ships four: `/find-a-document`, `/whats-changed`, `/who-owns-this` and`/check-a-claim`.-### The tools a skill needs can be picked where the skill is writtenA package's skills arrive with their tools declared. A skill somebody writes here could not: the@@ -3186,10 +2950,8 @@ than left out. A package ships skills declaring tools for connectors nobody hasskill outlives the server it was written against, so a screen that drew only what matched wasstating part of the declaration as though it were all of it.-## 0.0.3-### A Bot is offered the tools its message needs, not every tool it holdsA model chooses the right tool reliably out of about ten and unreliably out of thirty, and it fails@@ -3210,7 +2972,6 @@ were offered out of how many granted, and why: the skills chosen, or that nothinthat the choice could not be made. It answers "why did it call that", and the harder question, "whydid it not call anything at all" — which until now left no trace.-### The intent router works again behind a gateway`OPENAI_BASE_URL` is documented ending in `/v1`, and the router appended `/v1/chat/completions` to@@ -3220,10 +2981,8 @@ gateway, a proxy, a self-hosted model, which is the only reason to set it — unsilently stopped being routed and nothing said why. The version segment is now added only when theconfigured URL does not already carry one.-## 0.0.2-### Upgrading`AGENT_TOOL_TOKEN` is generated for you on a laptop. `scripts/start.sh` mints one and writes it to@@ -3296,7 +3055,6 @@ already refused the same way.Sessions survive and nobody signs in again.-### Changed- **This deployment does not search documents itself.** A Bot answers from a live system by calling@@ -3306,7 +3064,6 @@ Sessions survive and nobody signs in again.`document_acls` — and the connector that filled it have both been dropped. Retrieval overa copy of a customer's corpus is not a thing OpenBot does.-### Added- **A skill can say which tools it needs.** `POST /api/plugins/skills` takes a `tools` list of`serverId/toolName` references, stored against the skill and returned with it. This is the unit@@ -3428,7 +3185,6 @@ Sessions survive and nobody signs in again.ever been here, since removing them deletes the sessions that were the only evidence. A trail thatis unavailable never blocks a sign-in.-### Fixed- **A ref could resolve against a page from a computer that no longer existed.** The generation acomputer stamps on a snapshot is unique only within one run of it, so a replaced container counts@@ -3788,7 +3544,6 @@ Sessions survive and nobody signs in again.it, and no explanation, because a run that finishes carrying nothing is not an error. Every Botaction in the browser was affected: opening a page, filling a form, asking for help at a sign-in.-### Changed- **A retention policy for the audit trail.** `AUDIT_RETENTION_DAYS` removes rows older than the@@ -3825,7 +3580,6 @@ Sessions survive and nobody signs in again.addresses are refused whatever a provider says.- The container image runs as an unprivileged user rather than root.-## 0.0.1First tag.diff --git a/server/src/plugins/mcp.ts b/server/src/plugins/mcp.tsindex 76da2e9c..ba76e6ce 100644--- a/server/src/plugins/mcp.ts+++ b/server/src/plugins/mcp.ts@@ -44,12 +44,15 @@ export const MAX_RESULT_CHARS = 20_000;* from memory. For a knowledge connector that is precisely the failure the whole slice exists to* prevent — an answer with nothing behind it. So nothing is stated, in words.*/-export function resultText(content: unknown): {+export function resultText(+ content: unknown,+ structuredContent?: unknown,+): {text: string;truncated: boolean;} {const parts = Array.isArray(content) ? content : [];- const joined = parts+ let joined = parts.map((part) => {if (!part || typeof part !== "object") return "[unknown]";const item = part as {@@ -76,10 +79,25 @@ export function resultText(content: unknown): {// that sent one newline has said nothing, and which shape of nothing arrived should not change// what the model is told.if (joined.trim() === "") {- return {- text: "The tool returned no content. Nothing was found, so there is nothing here to answer from.",- truncated: false,- };+ /*+ * MCP tools with an output schema often put the answer in `structuredContent` and leave+ * `content` empty. Treating that as "nothing was found" is the same lie an empty string was:+ * the tool answered, and the model fills the gap from memory. Read only when `content` had+ * nothing; a text part already present is the representation the server chose to show.+ */+ const structured =+ structuredContent !== null &&+ structuredContent !== undefined &&+ typeof structuredContent === "object"+ ? JSON.stringify(structuredContent)+ : "";+ if (structured === "") {+ return {+ text: "The tool returned no content. Nothing was found, so there is nothing here to answer from.",+ truncated: false,+ };+ }+ joined = structured;}if (joined.length <= MAX_RESULT_CHARS) {@@ -372,7 +390,10 @@ export async function callTool({ timeout: CALL_TIMEOUT_MS },);- const { text, truncated } = resultText(result.content);+ const { text, truncated } = resultText(+ result.content,+ "structuredContent" in result ? result.structuredContent : undefined,+ );return { text, isError: result.isError === true, truncated };});}diff --git a/server/tests/mcp-result.test.ts b/server/tests/mcp-result.test.tsindex e18f494f..5e3cf02e 100644--- a/server/tests/mcp-result.test.ts+++ b/server/tests/mcp-result.test.ts@@ -33,6 +33,34 @@ describe("a result with nothing in it", () => {test("is not reported as truncated", () => {expect(resultText([]).truncated).toBe(false);});++ test("reads structuredContent when the content list was empty", () => {+ // Tools that declare an output schema often put the answer only in structuredContent.+ // An empty content list used to be reported as nothing found, so the model filled the+ // gap from memory while the vendor had answered.+ const { text, truncated } = resultText([], {+ title: "Expense policy",+ meals: "under $75 need no receipt",+ });+ expect(truncated).toBe(false);+ expect(text).toContain("Expense policy");+ expect(text).toContain("under $75 need no receipt");+ expect(text.toLowerCase()).not.toContain("no content");+ });++ test("does not replace a text part with structuredContent", () => {+ const { text } = resultText(+ [{ type: "text", text: "the prose the server chose" }],+ { title: "ignored" },+ );+ expect(text).toBe("the prose the server chose");+ expect(text).not.toContain("ignored");+ });++ test("empty content and empty structuredContent still say nothing was found", () => {+ expect(resultText([], null).text).toBe(resultText([]).text);+ expect(resultText([], undefined).text).toBe(resultText([]).text);+ });});describe("a result with something in it", () => {