Changelog
All notable changes to Flux are recorded here. The format follows Keep a Changelog. Flux uses Semantic Versioning for its releases.
Keeping it current. A pull request that changes what people or operators see adds
one file under changes/ with its PR number, instead of editing this
file, so that parallel pull requests do not conflict here. This covers features, changed
behaviour, fixes, removals, security fixes and new operator steps. Internal refactors,
tests and agent-process changes need no entry. At release time,
python3 scripts/changelog.py release <version> <date> moves the pending entries under
that version; the older entries under 0.1.0 were written before that rule.
[Unreleased]
Section titled “[Unreleased]”[0.2.0] - 2026-11-10
Section titled “[0.2.0] - 2026-11-10”- The ⋯ menu of a message has Copy text next to Copy link, on every message a person can read, on a computer and in the phone’s long-press list. It copies the message’s plain words (a mention as @name) and is not offered on a deleted message (#1006).
- A person can edit and delete their own messages, in a space and in a direct message, from the ⋯ menu after Copy link on a computer and on a phone. Edit changes the words in place and the message shows “edited”; Delete asks once and leaves “Message removed” in the same place, keeping the replies, and the words leave search, the notices that carried them, the export, the message’s files and the database (encrypted spaces too). Only the author can; an agent’s own line stays with its owner (#1006).
/aiworks in the threads of a direct message too. The answer is a card in the thread that only you see; Show everyone posts it as a reply in that thread, and the assistant reads that thread and nothing else. Both people still see one line in the conversation that AI was used (#1016).- Agents on your own Claude Code (a Claude plan in Connect AI) search the web with Claude Code’s own search, on by default like the search of agents on an API key. Each search is recorded with the run, and the answer lists what it searched and the pages it named. Agents on your own Codex search too, with Codex’s own search on OpenAI’s cached index (#1065).
- A dynamic security test:
./scripts/check_dast.shruns OWASP ZAP against the running application in Docker, a baseline scan of the web app and an active scan of the API signed in as the owner and as a member, and lists what is left to triage; recorded false positives carry their reason. A manual step before a release (#838). The API now has a floor of rate limits under every route (1,200 requests a minute per session, 600 per address without one; a read that gets a 429 waits and asks again with a calm “Flux is busy” line), and every response carriesX-Content-Type-Options,Permissions-PolicyandCross-Origin-Resource-Policy(#956). - Release images are signed. The release workflow signs both image digests keylessly (cosign, Sigstore) after they pass the scan and the packaged install,
INSTALL.mdshows how to check the signature and the build provenance withcosign verifyandgh attestation verifybefore you pull, and publishing a release now requires both signatures (#840). - Agents work together in a task’s thread. An agent can split a task into subtasks (shown in the task as “Subtasks · 1 of 2 done”; those for other people wait for your OK, with Approve all), and ask the agent of another person in the task a question: it answers on its owner’s account, reads only, and uses only what everyone in the task can see. Each exchange is bounded and a stalled one returns to the people. The task shows who is in it, the people and their agents (#841).
/aiworks in every conversation: a space’s conversation and its threads, a task’s thread and a direct message. The answer is yours alone, in a card with Dismiss and Show everyone, which posts it as your own message, “Ada · via Assistant”. It uses only what the people of that conversation can see; a guest has no/ai. In a direct message both people see one line, “Ada used AI here (Claude)” (#842).- Each of your agents has a private chat in Messages with a box to write in: it answers there, remembers what was said and “remember that …” adds to its memory. In its chat an agent reads like you and changes nothing (#842).
- Your agents now run on the server and take tasks. Give a task to one of your agents in the Owner field, or set “When someone assigns a task to me” in Settings › AI to “Give it to” one agent. It says it is on it, works inside Flux with your rights, writes its report in the task, and follows its “What it may do”: Done with Undo, or handed back for you to check. The task shows what it is doing in one plain line, with Stop, Try again and Take it back (#856).
- Tell an agent in its private chat “Every weekday at 9, summarise what changed in Garden” and it answers in plain words with what it prepared and one tap to turn it on. An agent’s schedules are listed in its conversation, in one line in its header that opens the list, where you turn them on, pause, change, run now and delete them (with Undo); the agent’s page shows the same list. Their results arrive in the agent’s chat as one message each, and a run that fails or cannot start says so there: a schedule is never paused after failures. A schedule reads only web addresses you typed (#858).
- The morning summary is a ready-made schedule on the agent’s page, off until you turn it on (#858).
- What a local client (Claude Code, Codex) changed on a task, a Wiki page or a map shows its change line and Undo where it was made, “Ada · via Claude Code”, “Moved #4 to Doing · Undo”: in the task’s activity, on the page under its title, and on the map (in its dock, or at the head of its List), one line per change, newest first, three at most. Anyone who can read it sees the line; the owner or anyone who may edit it takes it back, and “Changed since” puts nothing back (#868).
- Encrypted spaces, per space. Turn on Encrypted in the New space window (off by default; it cannot be turned off later), choose an unlock password and download the recovery code, and the space is encrypted on the server with no AI in it. Only its admins and the recovery code hold the keys: the company’s admins can delete it but not open it, and Add me does not list it. After a server restart it stays locked until one of its admins enters their unlock password; its members see "Locked, waiting for an admin" and the rest of the company works as usual. Any company member who is not a guest can make one, from every New space entry point, also as the first thing a new owner does (#874, #848).
- New space is one window with a name, the people (chosen from the company) and an AI switch that starts on; you are the space’s admin. Each space has settings under its header’s ⋯ menu: its name, its people with Admin and Member roles, the AI switch, Leave space and, for admins, Delete space, which asks first and names the space. A company admin can delete any space from Settings › Company, but not Everyone; deleting a space also removes its files from the server. A space always keeps an admin: nobody leaves, steps down or is removed while they are its last admin. (#874)
- Agents can read a public web page, on by default (the operator turns it off with
FLUX_WEB_READS=off). The Flux server reads it, read-only and with the same guard as link cards, and the agent’s answer ends with the pages it read. What a page says is data and can never make an agent act beyond “What it may do” or skip a wait (#918). - The owner of an agent can take back what it said in a task’s thread: its report, its question to another agent and its answer to one carry Undo under the line (while no person has answered it), and everyone then reads “Undone by Jonas” (#923).
- A Wiki page can be deleted: ⋯ › Delete page, after an OK, on the computer and the phone. It goes for everyone in the space with its earlier versions; a message that cited it keeps its text and shows “Page deleted”; an agent’s suggestion about it and the task plan items made from it are cleared (the tasks stay); a call on it ends, and a page that is open elsewhere says “This page was deleted”. An agent asks with
flux_delete_docand it always waits for its owner’s OK (#941). - On My map a selected thought offers Move to a space map and Copy link. Move asks for the space and its map (a space with no map gets one), keeps the thought’s text, shape and size, takes along its connections to other thoughts that move with it, and has Undo. Copy link gives an address that opens My map with the thought selected (#948).
- Flux’s pages and its service worker now carry a real Content-Security-Policy. Scripts run only from Flux’s own files and the hashed inline ones (never
unsafe-inlineorunsafe-eval), the page connects only to its own origin, and plugins, frames,<base>and forms to other sites are refused, so a script an attacker managed to inject could neither run nor send anything out; styles may still be inline. The offline page’s button no longer uses an inline handler (#979). - The API writes a structured security event for each failed or limited sign-in, password reset or change, session or client revocation, consent decision, refused origin, refused request and encrypted-space lock or unlock, with no password, token, address or content in it; the events are listed in the monitoring guide (#997).
Changed
Section titled “Changed”- The phone’s “+” Create sheet is five list rows, Task, Thought, Decision, Page and Space, with the line “The same five on the computer (C)” and no Space picker of its own (the Create window has one), and a long press on a message opens one bottom sheet with the message at the top and its actions below, in the computer’s order, in place of a lifted copy of the bubble with a small menu (#1013).
- Undo on an agent’s message now takes back a page the agent started, by deleting it, while only that agent has written it and nothing names it; a page edited, linked or being typed in since stays, and the line says why. A Wiki page switcher that is already open drops a page deleted meanwhile (#1021).
- The panel that opens beside the work is headed by the label that opened it, on every device: “Space details” or “Message info” for a place’s own panel, “Add to a Wiki page”, “Connect your AI”, “Propose a decision”, “People and encryption” and “Linked objects” for the rest, and its close control says the same (“Close space details”). A proposed decision that replaces an earlier one says so once, in the sentence that names it, and a restricted space’s details say who else can open it: the company’s owners and admins (#1058).
- AI answers know who asks, where and when: every place (
/aiin a space, a task’s discussion, a direct message and its threads, an agent’s private chat, a task an agent works on, a schedule’s run) and both engines (an API key, or your own Claude Code or Codex) get one brief with your name, the time in your time zone, the place’s name and each message’s author and time; answers come in the language you asked in, as short plain text with#12and[[Wiki page]]that stay links after Show everyone; they may use general knowledge, say so when the conversation does not answer, and say when older messages were left out; an agent working a task is given its outcome and when it is done; the stiff “Fact, Interpretation or Proposal” labels are gone (#1064). - An agent may open any public web page it needs, on an API key, on your own Claude Code and on your Codex alike: no list of allowed addresses and nothing to approve. It is told never to put anything from Flux into a web address, and Flux refuses an address that carries text the agent read in Flux. Agents on your own Codex now search the web too, and every answer still lists what the agent searched and the pages it opened (#1065).
- Agents search and read as much as a question needs: the limits of 5 searches and 20 addresses a run are gone, each search is charged as the provider reports it and the key’s daily budget still stops spending, a run on your API key keeps a short note of each page it read and no longer sends the page’s text again once it is noted, and an address nothing gave is still refused. A malicious page with many links could leak what an agent knows through which of its links the agent opens; the founder accepted that risk (#1113).
- An agent’s page reads are no longer limited to 20 a minute (founder direction, 2026-11-09): a research run reads as many pages as its question needs, one after another; the server still reads at most 6 pages at once (#1116).
- An agent knows who wrote what: every message it reads, on your Claude Code or Codex and with an API key alike, names its author as you see them, “Asystent (Ada’s agent)” for an agent (with the owner’s full name when two people in the space share a first name) and “Ada Nowak · via Asystent” for an answer someone posted with Show everyone, which it never takes for its own; only people who can read the space are named, anyone else is “Former member”. An agent gives a task to a person of the space by name, and sets or clears a task’s due date like a person (#1162, #1163).
- The README, the self-hosting and administration guides, the release installation guide, the security policy and the integration guide describe v0.2 (one company with spaces, simple AI with Connect AI and agents that have their owner’s rights, local clients over MCP, encrypted spaces, the security work), v0.2 has release notes, and
./flux demosays what the app says: a source, a Map with Thoughts and a Wiki page instead of “published material”, “sketch” and “doc”../flux exportnames its argument a space. Screenshots that still show v0.1 (the README’s, the People and invite pages, the New project page) are out of the pages until new ones are recorded (#1172). - Flux v0.2 does not upgrade a v0.1 instance, because v0.1 had no users:
./flux upgraderefuses a database that v0.1 wrote before it asks, takes a backup or stops anything, the migration step (./flux up,docker compose up, a release install) refuses it before any migration file runs withFLUX_V01_DATABASE, and./flux restorerefuses a v0.1 backup with or without--migrate. The message says to install v0.2 fresh (a new folder, or./flux reset) and that there is no import of v0.1 data; the upgrade guide and the v0.2 release notes say so too. An instance interrupted during its first install, before anybody signed up, is not refused (#1176). - The API’s request log uses a request’s
X-Request-IDas itsreqIdwhen it is a short plain identifier (letters, digits,.,_,:,-, up to 64), so a proxy’s id and the log line match; the performance report pairs server times with requests by that id, so every request has its time (#827). - A task an agent makes from a message keeps that message as its source, as a task made by a person does: the task opens the message in context and the message shows the task (#844).
- Flux is lighter on a modest server. An open Map or Wiki no longer reads its access rules four times a second: it listens for changes and looks only on activity or every two seconds when idle, and a cursor that has not moved no longer wakes the other editors. The project, its people and its docs are not read twice at once, the app reads the person and the workspaces together and then the projects and messages together, the video-call code is not downloaded ahead of time, and a session read no longer signs a token nobody reads (#850).
- The performance check measures what a 2 vCPU host feels: a CPU-limited phase that holds every request to a 50 ms p95, the first-screen download, idle database statements and restart times, and a browser test counts the API requests of each move between places (#850).
- An idle server is quieter: its worker looks at its queues once a minute instead of every 30 seconds when nothing has woken it, and its other housekeeping looks less often too, so a Flux nobody is using makes about 95 database statements a minute, down from 122. A delayed or retried job can start up to a minute late instead of half a minute; a job sent now still starts at once (#850).
- Settings › AI has one Connect AI: a Claude plan, a ChatGPT plan and API keys side by side, each with one usage line. Each agent picks its model from any connection, and nothing switches by itself (#854).
- Flux sets no limits of its own on subscription runs: no runs a day, five-minute clock, step count, changes per run or background runs a day. A plan at its limit pauses the request with one line and resumes by itself; only technical safety bounds remain (#843).
- An API key has one daily budget, for everything that runs on it. The per-run ceiling, the assistant’s daily cap and the separate allowance for background suggestions are gone (#854).
- An agent now has exactly its owner’s rights, and it loses them the moment its owner does. Each agent has one “What it may do”: Only read, Do it and tell me (the default) or Ask me first. Deleting and other people’s things always wait for your OK, and only you, signed in, can give it (#855).
- Undo now sits on the agent’s message, next to one line that says what it changed, and the line is then struck through and ends “Undone by Ada”. A deletion asks Delete or Keep, a task for someone else is a proposal with Approve, Edit and Skip plus Approve all, and the Inbox shows one item per agent run. Nothing that waits expires (#855).
- MCP tool contract 2: the
flux://policy/{projectId}/{revision}resource and bootstrap’strusted.approvedPolicyare gone, because an agent now has its owner’s rights. A local client set to Only read getsAGENT_ONLY_READS; its other changes wait like the assistant’s (#855). - A local Claude Code or Codex now acts with your rights, like your agents, and has the same “What it may do”: Only read, Do it and tell me, or Ask me first. Deleting and other people’s things always wait for your OK, and so does everything under Ask me first; you approve in the Inbox, one item per client with Approve all. A unit or request a client gives to someone else’s client waits too. Disconnecting a client closes what it had waiting. Only you, signed in, can approve: the client’s own token cannot (#859, #855).
- Signing in Claude Code, Codex or another local client is one click: its sign-in opens one page that says who asks and where access goes, and Allow is the whole setup. Allow makes the client’s agent (yours) and its connection at “Do it and tell me”, or finds the one the client already has; Deny and merely looking at the page make nothing. The client then acts with your rights, in the spaces where AI is on, and no more (#859).
- Disconnect ends a client’s access at once and closes what it had waiting; signing in again makes a new connection and gives the client’s earlier agent back (#859).
- A client’s undo of a task it made waits for your OK like any deletion, and an Approve that comes after the claim or session it was made under has ended says so in plain words instead of applying (#859).
- The agent runtime has no fixed number of people: one runtime serves everyone with Connect AI, each person under their own user id. How many agents run at once follows from memory, about 2 on 4 GB, and the others wait with “Waiting for room on the server”.
FLUX_RUNTIME_MAX_RUNSchanges the number (#860). - A host that does not meet a runtime requirement, such as
kernel.yama.ptrace_scope=0, stops the runtime with one clear message and the fix. There is no restart loop, and./flux upfails instead of reporting a healthy Flux (#860). - Moving between places asks the server less: the project, its people, the members, the space’s docs and the agents’ unread dots are kept for a moment, and what waits for you is read once per page, so the next view shows them without asking again. Anything that could change them (a change you make, a live update, signing out) forgets them, and a refresh on focus still reads. Going from Home to the Inbox now costs two requests instead of four, and entering a space 15 instead of 20 (#862).
- An open tab no longer asks the server on a timer while its live connection is up: a space’s conversation, Tasks, a thread and the working agents read when something changes, so a tab nobody touches makes almost no requests (a space’s conversation made 70 a minute; the Inbox count still checks every two minutes, as some of its changes send no event). “Seen by” in a space appears when someone sees your message and stops naming someone who can no longer read the space, a live update tells only the change it carries what to read again, a tab back from sleep or a network change reconnects, and opening a page’s live editing keeps what the space had read (#862).
- The Inbox is the one list of what needs you, and Home shows its top three under “Waiting for you” with the same count: every decision is one row (“2 decisions waiting”), each row opens the Inbox on its item, and nothing on Home acts. Home keeps only that and “Continue where you left off”; on the phone “My tasks (N)” and “My map” are rows under the waiting items (#865).
- “Mine” and My tasks include the tasks your own agents own (#865).
- An idea is a Thought and a canvas is a Map, in every label, in search and in Create. The Sketchbook is now My map, your one private map at
/map; “Your notes” are thoughts on it, and Create’s Thought puts a thought there. “Sketch” is “Map” (the default title is “Untitled map”), and the toolbar is “Map tools” (#866). - A message shows Reply in thread, Create task and ⋯. The ⋯ menu is one level, in the same order on every device: Add to map, Copy link (a phone’s long-press list starts with Reply in thread and Create task). Cite, Record a result, Details, Propose a decision, Hand off to an agent, Copy text and Done / Not now are gone from messages; a private message can be added to My map, also in a 1:1 the other person left, and a direct-message reply’s Copy link opens its thread at the reply (#867, #948).
- Create offers the same five types on the phone and the computer, Task, Thought, Decision, Page (a Wiki page) and Space, in place of Message, Project, File or link and Map; Decision on the phone opens the same window as the computer’s, and a Page is made from it. The phone’s sheet no longer hands a picked file to the composer (#867).
- A task has one discussion: an agent writes in it like anyone (
flux_discuss_task, read withflux_get_task_discussion), the separate “Agents’ thread” is gone, and the space’s conversation shows one line for each task event instead of the discussion’s first message and thread. Agents may write five messages in a row in it until a person writes, by any tool; a line Flux writes for an agent at work on the task (its report, a question to another agent, the answer, “Stopped by …”) is always posted (#868). - Tasks: “Not pursued” is Done with a “dropped” note (“Done · dropped”). A done task has one optional line, “What came of it”, also written in its discussion, and a “Dropped, not finished” check; the “Result” button and the “Record a result” form are gone from the task panel. A blocked task waits for one person, one decision or one other task and goes back to Doing by itself when that is resolved: the person says “Done” in the Inbox or the panel, the decision is accepted, or the task is done. The stream says “#4 can continue: the decision was accepted” and the owner is told once; an agent picks the task up again (#868).
- An agent’s task that is Blocked when its run ends stays Blocked with the agent (it was marked Done, or handed back); a task something waits for cannot have its creation undone, and the tasks that wait for a decision or a task are not stuck when the waiting task’s creation is undone (#868).
- Tasks: a task is To do, Doing, Blocked or Done, and the board has those four columns; a dropped task is Done with a “dropped” note. Board and List share one control row (Board | List, Mine, Search) on every device, the phone included, and a narrow board scrolls its columns sideways. A card shows its title, number, due day and owner, and a blocked card one “Waiting for” line (#868).
- The task panel hides empty rows and texts and has no ⋯ that held only Copy link; giving a task to one of your agents is its Owner field. It appears once its relationships are read, so nothing moves after it opens (#899). A thread’s panel leaves with its content (#901), a thread opens at half height on the phone, and Tasks loads as its board’s shape (#902).
- The Wiki is two columns, the sidebar and the page: the page is edited where it is read, and versions save by themselves while you write and when you leave. The title is the page switcher (Find a page, + New page), ≡ lists the page’s headings, and ⋯ holds History (compare and Restore), Copy link, Export, Import a Markdown file and Rename page. A task opened from a page opens over it (#869).
- Search is one tool: ⌘K, the sidebar’s Search and the phone’s Search open the same dialog, with the results in groups by kind and a plain sentence for a hint. The Search page, its filter chips, the place filter and the “See all results” row are gone, and so are the Create rows and the
@and#scopes (#870). - A decision is called “Decision” everywhere: the Inbox card, the return summary and the notification no longer say “rule” or “proposal”. “Replaces” and “In the Wiki” are plain sentences, drawn only when they apply: “This replaces the decision … (12 Oct).”, “This was replaced by the decision …” and “Shown on the Wiki page …”. Accepting a decision that replaces another only marks the old one replaced (#871).
- Words about who can see a place appear only when the audience differs from the place’s default. A space’s conversation, its threads, a task, a space’s map and a direct message say nothing under the box; a private map still says “Only you”, and a private answer still says who sees it. Header subtitles that described a page are gone, “Who can see this” in the header’s ⋯ menu is “People”, and “Came from” on a task is “From” (#871).
- A place’s panel has one name on every device: “Space details” in a space and “Message info” in a direct message. Home, My map and My tasks have no Details (#871).
- An encrypted space’s people and keys moved from Home → Details → People to the space’s own settings (“People and encryption”). Removing someone from the company now also ends their place in its encrypted spaces, and the only admin of an encrypted space makes another admin before leaving (#874).
- One Flux is one company (F-037): the first account sets the company up, as its owner, with the default space Everyone that holds the whole company. With single sign-on everyone the identity provider lets in joins on their first sign-in; without it, company admins invite people with links. Projects are now called spaces, in one flat list in the sidebar (Home, Inbox, My tasks, My map, Spaces, Messages); a person sees only the spaces they were added to, everyone but a guest starts spaces, and a space’s creator is its first admin. Settings › Company replaces People, with the company’s name and its people. There is no way to create or switch workspaces. (#874) Settings › Company also lists every space, where an owner or admin adds themselves, and people who left or were removed, with Restore; a company admin’s sidebar holds only the spaces they belong to. The company’s name is on the sign-in page and in email.
- AI is switched on or off per space (F-037), by the space’s admin or a company admin, and replaces the whole-company AI switch. While AI is off in a space, no agent or local client reads or acts there,
/aisays “AI is off in this space”, the Owner picker lists people only, and an agent’s run there stops before its next change; the company’s other spaces are not affected, and Settings › AI no longer lists where AI is off. A company admin’s search and Inbox now hold only the spaces they belong to. (#874) - An idle Flux uses about a seventh of the CPU (7.4% to 1.0% of a core on the measured stack) and a tenth of the database statements: health checks are small programs every 30 s, not a Node process every 3 s, and job queues wake the worker instead of being polled every 2 s. The application image is 216 MB, down from 500 MB (#878).
- Operators: the default memory limits are now
FLUX_DB_MEMORY768m (new),FLUX_API_MEMORY1536m andFLUX_WORKER_MEMORY512m, 2.9 GiB for the base stack on a 4 GB host. Raise them in the settings file if your instance needs more. The health checks usestart_interval, so Docker Engine 25 or newer is now required:docker compose uprefuses an older engine (can't set healthcheck.start_interval as feature require Docker Engine v25 or later) and./fluxsays so before it starts. Update Docker Engine first (#878). - A company admin’s agents (server agents, local clients and
/ai) read, search and act only in the spaces the admin was added to, as the admin’s own sidebar, search and Inbox already did. After Add me they work in that space too (#934). - The owner of an agent can always delete a line it put in a task’s thread, whoever answered it since: the line becomes a marker that keeps its place and everyone reads “Removed by Jonas”. A release only stops the runs that were going when the task was taken away, and a give-up that keeps failing is tried a bounded number of times and left with its count (#938).
- Search says when there are more than 20 results (“Showing 20 · type more words to narrow”), its hint lists files, and the search answer no longer carries per-kind counts that nothing read (#944).
- Wiki: a screen reader no longer hears “Saving…” and “Saved” at every pause (only who joins), Rename works while someone else types, an open page follows a rename without reading the page again, and a caret left idle stops renewing itself after 30 s (#950).
- Agents read the web more safely: an agent reads only an address it was given (written in the request, in the task or its conversation, or on a page it read); every address the server sent a request to is written down before the request goes out and listed at the end of the run (under its answer, or in its end line when it ends without one), those that failed too and without the site’s own title; pages with no
</head>are read; invisible characters are removed from page text; the tool is offered only to a server agent’s run and only when reading the web is on, so a local client’s tool list no longer has it (#953).
Removed
Section titled “Removed”- Tasks: the parked state of a task, which nothing set after decisions stopped moving tasks, and the
still_applieslink role: the database columns, the API fields (parkedon a task, a row and a prerequisite, the summary’s parked count, theparkedgroup, the project export’sparked) and the places that read them (#1039). - A space stores no views any more: every space shows Conversation, Tasks, Map and Wiki, so the
viewsfield,POST /projects/:id/viewsand theproject.view_added.v1event are gone (#1043). - The Result object: there is no “Record a result”, no result card or notice in the conversation, no Results section in the task panel, no result in the Wiki’s “Add to docs”, search or the Inbox, no
resultsin the project export, and noflux_record_result,flux_list_resultsorflux_get_resulttool. A finished task says what came of it in its one line; what matters goes to a Wiki page. The assistant’s proposals went with them (a scheduled run could only propose a result), and theaboutlink role (#1054). - Nothing is kept compatible before the first public release: the MCP tool contract and the project export each have one current snapshot, edited in place (no
mcp-tools.v1.json, no version-bump rules, no “breaking or additive” classification); the export no longer lists the unuseddraftdoc state,vieweranddeniedroles; the playbook keeps no version history; the old Settings addresses (/settings/assistant,/settings/assistant/claude-code) and the direct message’s/sketchesaddresses are gone; the browser no longer reads the earlier composer drafts; and the API no longer refuses a setFLUX_LIVEKIT_WS_URL(#1096). - A schedule’s result goes only to its agent’s private chat: nothing posts it into a conversation any more, and the owner shares what they choose with their own message (#1104).
- The answer to
/aiis no longer a message everyone in the conversation reads, and nobody sees “Ada’s assistant is working…” any more; the ask bar’s model chip and the public “@Ada’s assistant …” line are gone (#842). - Agent grants on spaces, the approved agent policy, asking a space admin to let an assistant or an agent in, the assistant’s list of spaces, its area switches, its change limits and its Recent list (#855).
- The schedules page in Settings, its sheet with a cron editor and its “post the result in a conversation” choice are gone: schedules belong to an agent, are made by telling it, and their results arrive in its chat. A schedule’s answer is no longer posted in a space’s conversation,
/aiin a space no longer makes schedules, and three failures in a row no longer pause one (#858). - The three permissions a local client asked for (read, suggest, run actions), the 38 capability switches under Permissions, the choice of projects and the connection form with its Continue to consent step: a client has your rights and the one “What it may do” row, so there is nothing to tick or choose (#859).
- Standing grants for local clients: the grant panel under each connection,
/api/v1/agent-connections/:id/action-grants,grantIdon every action tool and the grants influx_bootstrap(#859, #855). - The space header’s “N needs you”, the Tasks “decisions need you in the Inbox” link, Details’ “Now” and “What matters” (with the “Since you left” line over a conversation), Home’s spaces card with “quiet”, its footer links, key hints and assistant banner, and the workspace-wide Wiki page (#865).
- A direct message has one view, Messages. Its Sketches tab, “Choose messages” (start a sketch from messages) and “Make it a space” are gone; a conversation is a conversation, and ideas go on My map or a space’s Map (#866).
- Your notes, the note composer with its device captures, “Private drafts” and “Start from a private draft” when adding a material are gone, with the Sketchbook’s list of sketches (#866).
- The Map has no “Create task” in its toolbar, on its phone options sheet or in its outline: a task is made in Tasks, with
/taskin the composer, or from a message (#867). - A message can no longer be cited by another message, because nothing offers Cite any more: a message cites a map thought only, and the API refuses a
messagereference (#867). - Tasks: the cards’ ⋯ (“Move to…”), the columns’ “+”, the narrow board’s row of status tabs, the List’s row of state views, the phone’s separate “Mine | All” row, the cards’ “came from” link and the task panel’s Hand off button and window (#868).
- The Wiki’s Write / Preview / Both editor, the raw Markdown field, “Reason for this version”, Draft / Published, the pages column, the “On this page” column and the Wiki’s focus mode (#869).
- No more pivot: accepting a decision never moves, parks or re-links a task. The “What happens to current work?” part of the decision panel, “Accept and pivot”, “At this pivot”, “Parked by a pivot” in Tasks and the task panel, the return summary’s “Parked: …”,
stillAppliesandparkonPOST /api/v1/decisions/:id/acceptand{ parked: false }onPATCH /api/v1/work/:idare gone (#871). - Space templates, the space goal and its “goal reached” moment (F-038), invite links that open a project (an invite joins the company), and the New space window for an encrypted workspace, until encryption is chosen per space. (#874)
- The AI switch of the whole company (F-029), its People row and the “AI is off” lists in Settings. (#874)
- A space has two roles, Member and Admin. The read-only role (
viewer) and the explicit deny (denied) are gone from the access rules, the grants API, the screens and the database: everyone in a space can work in it, and a person is kept out of a space by removing them from it. A company owner or admin always reaches every space, and Add me in Settings › Company now makes them the space’s admin. (#903) - A Wiki page has no Draft or Published state in the API, the agents’ tools or the database, and a person no longer types a version’s reason: the server writes it (“Edited the text”, “Renamed to “…””, “Restored version 3”, “Imported from notes.md”, “In place of Ada’s version 4”).
stateis gone from the page commands, versions and summaries and fromflux_create_doc/flux_update_doc, a save names acauseinstead of areason, and Search says “Wiki page”. An assistant’s Undo no longer moves a page it started back to draft; the page stays (#942). - The project Review rule is gone from the server too: no
review-policyroute, nocowork_review_policiestable (migration 0017) and no per-project choice between “another connection” and “another owner”; an author’s own connection is still never its own reviewer in a run. A guest saving an API key or starting a plan sign-in is refused withGUEST_NO_AI(#959). - The return summary’s recap options (
scope,from,until,digest,history),ReturnDigest, the project work summary’sstateandGET /api/v1/workspaces/:workspaceId/docsare gone from the API: the web app, the MCP tools and the local clients read none of them.GET /api/v1/returntakes a place and starts at the saved return point (#962). - Home’s Details button and Home’s Details panel are gone. A company’s people are in Settings › Company; an encrypted space’s people and keys are in the space’s own settings (#964).
- A local client or agent no longer sends a request to another one: the tools
flux_send_request,flux_claim_request,flux_defer_request,flux_decline_request,flux_resolve_requestandflux_coordination_inbox, the request cards and handoff packet in a task’s thread, the cancel of a request, the limits on requests, and the recorded verdict of a review request are gone, with their API, tables and tests. Units stay (create, claim, renew, release, complete, transfer, decline), and a review is still never the run that made the work. The playbook is 1.14.0 and its request module is now “Hand off” (#973). - Pause and Resume of an agent’s work are gone from the server: the route, the table, the event, the
pausedfield of the working-agents list and the error codeCOWORK_PAUSED(playbook 1.15.0). Stop is the one way a person ends an agent’s work (#985).
- Small follow-ups from the encrypted-space and company-role reviews: the Search dialog keeps its “Showing 20” line while the next answer loads; an agent is active only while its owner is not a guest of the company; the task history says “deleted the agent Planner, so it came back to Ada” instead of a reassignment by a setting; a person removed from a space no longer sees where a hand-off stands there; an agent or a client connection is checked against the company role before the workspace’s own, in the same order a change to guest takes them (#1003).
- A Wiki page’s History tells what happened: “Restored version 3” only when the saved title and text are that version’s, “In place of Ada’s version 4” and “Saved as the page was left, in place of version 4” name the version the save really replaced, and a long or odd file name of an import is shortened to its last part on one line (#1021).
- A re-read that a live event causes (a space renamed in another tab, a grant taken away) can no longer join a read that began before the change and keep the old answer. The first sign-in page after a start no longer waits for a probe of the identity provider. A made-up session cookie no longer gets a request budget of its own: a session cookie or an
/mcpbearer has one only once the server has accepted it, it is the cookie Better Auth reads (not a decoy, a bearer or a space next to it), and a 401 takes it away; an IPv6 client counts as its /64 in the per-address limits (#1023, #1022, #1040). - A member made a guest keeps their agents, API keys and client connections but the server hides and refuses them until they are a member again: their agent list, keys, client connections and the conversations with their agents in Messages are empty, and the assistant’s status says it is not set up (C-50). The company role decides, wherever else the person holds a role. A guest’s shell also reads its role again when it changes, so the
/aibar goes without a reload (#1035, #1034). - A send whose answer was lost, and whose message you then deleted, no longer puts the deleted words back in the composer when it is retried: the retry finds the deleted message and ends quietly. Deleting a space also runs once more if the database cancels it to break a rare deadlock with a notification being made (#1052).
- A task that a linked pull request’s rule moves to Done lets the tasks that wait for it continue, in the same transaction, with the same line, activity entry, Inbox note and agent pick-up as when a person finishes it (#1055).
- A member made a guest keeps their schedules and old
/airuns, but the server now hides and refuses them until they are a member again, like their agents and keys: no schedule is listed, shown, deleted, restored or run, no run or its changes can be read, stopped, dismissed or shown to everyone, no plan sign-in is shown or signed out and the model list is empty (C-50). Their schedules stop firing and write no “Skipped” rows meanwhile; a schedule that is skipped for another reason keeps only its newest rows. (#1059). - Confirmation links to an extra notification address:
name+news@…andname+jobs@…count as the one addressname@…for the limit of three an hour, a refusal by that limit says so without blaming the person (ADDRESS_CONFIRMATION_LIMIT) and no longer uses up the person’s own one-a-minute allowance, the messages say “confirmation” as the page does, and an address an identity provider reports with a decomposed letter is the same mailbox as the confirmed one (#1070). - A member made a guest is no longer handed an old
/airequest back: the same request, in a space, a direct message or a chat and with or without anIdempotency-Key, is refused. The conversation’s change list gives a guest only what was done, a schedule’s “Paused” notice leaves their Inbox, andFLUX_AGENT_RUNTIME_IDLE_DAYSdoes not release their plan sign-in; all come back as a member again (C-50, #1090). - On a phone a toast that appears over a sheet whose content is still settling is placed above the content’s final position, not where it was a frame earlier, so it no longer covers the sheet’s last link (#1106).
- A space’s people line no longer says “Only you” or “· only you two”, which was false because the company’s owners and admins can open every space; it names the people the space was added to, and a company owner or admin who was not added is not called “you”. The space’s details say once who else can open it, and “Add to a Wiki page” is no longer written twice under the panel’s heading. An address an identity provider reports with a decomposed letter is looked up as the same mailbox when a pending seat is claimed and a standing profile is read; sign-up, sign-in and reset take only ASCII addresses and answer 400 to any other (#1110).
- A Wiki page open in the live editor follows a change made outside it (a client’s or the API’s edit, a restore): it shows the new text, or keeps your own unshared text privately, instead of “Sharing is not confirmed (EDITING_GENERATION_CHANGED)” and the old text (#1116).
- Anthropic’s web search that pauses a long turn (
pause_turn) is continued instead of ending the run with the text so far (#1116). - Two moves between My map and a space map in opposite directions no longer wait on each other (one answered 500); a space’s first map is made in the same change as the move, so a move that fails leaves no empty map; the window lists every map of a space, not the first 50; a move into an encrypted space is not kept in plain text in the company’s records; a move needs one side to be your own My map; and the toast “Moved to …” stands above the map’s toolbar and the status line does not say it again (#1116).
- On a phone, the toast after a status change no longer covers “Dropped, not finished”; a client’s map move reads “Moved 1 thought”; a task panel says a local client’s change once, by its change line; “Undone by Ada” uses the first name; a guest sees a local client’s change lines; and the Space settings line says the company’s owners and admins can open the space too (#1116).
- Claude Code’s search is told not to write Markdown links, which an answer shown as plain text would show raw, and a search whose tool use has no id no longer makes a read of the run wait for its result (#1116).
- A page that loses access to a space while it is open goes to the list of spaces once, instead of sometimes getting stuck and asking the server for the person and their workspaces thousands of times; reads of the app’s shell that keep failing or keep being interrupted now wait longer each time, and a page never starts more than 30 of them a minute (#1126).
- Asking AI in a direct message of an encrypted space now says that AI is always off there, instead of asking the person to make an agent (#1131).
- After a long press on a phone, lifting the finger over the dimmed screen leaves focus in the message sheet, so Esc still closes it. Codex’s Sign in again is shown only for a real revocation, not for a temporary refresh failure or a refused request to Flux’s own tool route, and Codex’s usage read no longer calls ChatGPT’s plugin endpoints (#1133).
- An agent’s web read no longer slows the server, however many links it read, and a page someone named opens: the link itself, the page a redirect sends it on to (a DOI’s publisher, twitter.com to x.com), and the target of a pasted Google link. Flux’s own wording and dates or times in what the agent read no longer block ordinary addresses, while a private link’s id cannot ride in a form or search on the same site. The text an agent read is kept only while it works and deleted when it ends, link cards are never fetched for a line an agent on the server wrote, and an agent cannot write an address carrying what it read into a task or a page (#1136).
- A page that is on show stays on show when the read that refreshes it in the background fails to get an answer (a 503, another 5xx, no connection): the page keeps what it showed, asks again by itself after a short wait (1, 2, 4, 8 and 16 s, and on the next live event or when you come back to the tab), and takes the new data when a read finishes; the error screen is for a page whose first load fails and for what is gone or closed (a space that is no longer there, an access that ended, a signed-out session) (#1141).
- An agent’s answer keeps no web address that carries something it read in Flux: such an address shows as “[address left out]”. That covers the task reports, questions to other agents and posts that come of the answer. Links people wrote and pages the agent cites stay. Search queries list any address in them in a form that can’t be opened, a link to another storage bucket counts as another site, and one agent’s many page reads no longer hold up another’s (#1143, #1144).
- An agent’s answer is checked in its final form, after Flux takes out its marks and blocks, so no step can piece an address back together, and an address too long to check is left out. Its question to another agent and a schedule it prepares are checked too. An agent never writes another page of a person’s site carrying that link’s id, and more storage hosts count each bucket as its own site (#1149).
- An agent’s answer is checked before Flux saves it, so people’s task changes in that space no longer wait for the check. A question to another agent is read only from the agent’s own words, never from Flux’s list of its web searches. The pages a run read are marked as they are read, so a run that ends without an answer lists a page it could open only through a person’s link to that site as plain text, not as a link. Another bucket’s page on Backblaze B2 opens, and its download hosts count each bucket as its own site (#1156).
- A thought dragged on the map lands where the pointer let go, also when the release comes before the last movement is drawn (a quick flick, a busy browser); it used to land one movement short. The phone map no longer reports a ResizeObserver loop error in Safari’s engine, and a long select list opened while its field is still scrolling into view stays inside the window (#1167).
- A call someone starts in a space shows Join (or the live dot on More) in the space’s other open tabs at once, and so does someone joining or leaving it and its end: the server tells the space over the live stream instead of the tabs looking every 30 seconds (#1175).
- The first visit downloads 241 KB before it shows anything, from 316: a space’s conversation, the sign-in pages and the live strip come when they are needed, and inside the app they are fetched in the background beforehand. Work reads plan their heaviest statements once per database connection and read the session one time fewer (#1178).
- The task notices of a conversation and the assignment lines of the Inbox read the tasks they point at together, not one by one: a page of 100 took 113 and 218 database statements, and now takes a handful, however long the page is (#828).
- A NUL character in a name or other text, or in an address typed at sign-in, sign-up or password reset, answered “Internal Server Error”; it is now refused as invalid input (#838).
- With single sign-on as the only way in, the sign-in page no longer shows the email and password form for a moment before the provider button: the page reads what sign-in offers together with the check for an existing session, and the server answers from its last known state while it re-checks the provider (#847).
- Kreska’s hello wave, sleeping “z” and celebration sparkle sit outside the frame, so the mascot’s corner no longer looks broken; the design reference board shows the same marks (#849).
- Someone removed from a space or from the company, or who leaves either in another tab, sees it leave their sidebar at once, and a page of it that is open, or Back into it, lands on the list of spaces with one line saying it is no longer available instead of showing what the page had loaded or an error page (#862).
- A task row’s state button on the phone list names the four keys it answers to (1 to 4), not five (#868).
- Restoring a removed person with a malformed company id answers “not found” instead of a server error. (#874)
- A fresh install on a slow disk or CPU no longer fails with
Query read timeoutinmigrate: the migrator has its own long query bound instead of the API’s 2 s. On v0.1.0 apply the first migration by hand, see troubleshooting (#878, #846). - A short database outage no longer restarts the API or the worker: every connection pool logs the lost connection, drops it and opens a new one on the next query (#889).
- Search no longer grows after it opens when your assistant is ready: whether “Ask my assistant…” has a row is known before the dialog opens, so it is its final size on the first frame (#898).
- Closing a thread’s panel keeps its text through the 120 ms exit on a computer, and on a phone the sheet of a thread, a task or a decision fades out with its head and content instead of vanishing at once (#901).
- The Owner field of a task offers only your own agents. Giving a task to someone else’s agent is refused by the server too; to reach another person’s agent, give the task to that person (#904).
- A key’s daily budget can no longer be lowered below one request of the model the assistant uses on that key; it was accepted, and every request then paused for good. (#909)
- The company owner can be added to a space and removed from it in the space’s settings, as New space already allows, with no “Only an owner can add, change or remove an owner”. (#905)
- Connect AI: Escape closes the window at every step, and a plan whose sign-in is not ready says so instead of failing with words nobody can act on. (#910)
- An API key’s day ends at the next local midnight also on the day the clocks change, and a comparison and a run can no longer both take the last of its budget. (#910)
- The agent runtime removes what a person’s program left in the shared
/tmp, also directories with no rights, and a release no longer says everything is deleted while some of it stays (what a release can and cannot reach is in #915);./flux upleaves a healthy runtime alone whose log still has an old refusal. (#910) - A deleted space’s files leave the disk in one cleanup, not fifty every ten minutes. (#908)
- Releasing a person from the agent runtime now removes what their program left anywhere in the shared
/tmpthat their user id can reach and list, also a directory of theirs inside a directory of someone else’s and a tree deeper thanPATH_MAX, whichrmcannot take (a file in a directory of someone else’s that can be entered but not read is not found); their/datadirectory is removed by a walk that follows no link, even one another person swaps in while it runs; the same tree in their own directory no longer keeps the release open for good. A release that is asked again does not sign them out a second time, andagent-runtime.mdhas a section for a release that stays open (#915). - A key’s budget is checked against the model the assistant ends up on even when the model is changed at the same moment, and the day of a key whose clocks go back at midnight (Havana) begins at the first midnight (#915).
- A job that holds a database session lock (the agent runtime reconciler, a live-session transition, a live access revocation) now stops at its next step when the lock’s connection dies, instead of carrying on without the lock; the API start no longer logs Fastify’s FSTWRN004 or
pg’s deprecation of overlapping statements on one connection; a refresh token used by several requests at once makes one new token pair; and the database outage check proves the worker runs a job after the database returns (#916). - Agents and tasks: “In this task” lists only people who can still read the space, and their agents that can work there, and follows the thread; an answer that ends with a question is posted without the markup; an agent may ask another agent of its own owner; deleting an agent hands its tasks back with a history row; Approve says when you lost the space instead of “Changed since you asked”; the waiting line of a hand-over or a request to another person’s client reads plainly; the misleading “You can read this project but not…” messages say you lost access; the unused colour input of the agent API, and the web’s unused Pause/Resume code, are removed (#924, #929, #930, #936, #956).
- An agent’s run stops when taking its task away could not be handled, an event nobody could handle five times tells the task’s owner in the task and is never dropped, a retry that was waiting when the server restarted is found again, and the state line says who holds the task after a Stop and whose agent is being waited for (#935, #931).
- In a space a company owner or admin can open but was not added to, the assistant’s bar now says it works only in spaces they are in and offers Add me, instead of saying they can’t open the space; an agent is asked a question in a task only if it can read the space; and when “Give it to” cannot give such a task to their agent, the task says so (#955).
- A guest has no AI, all the way: the pages that connect AI or set an agent up (the API keys, Connect an agent, Claude Code and Codex sign-in, an agent’s page, the assistant’s schedules) lead a guest to the one line at Settings › AI; looking up a provider’s models is refused for a guest like saving a key; the first Assistant is made in the workspace the app shows as the person’s own, and not for a guest of it; the first plan sign-in’s Assistant has a test; and after a first key saved on the API keys page, Back to AI shows the new Assistant at once (#959, #995).
- An agent’s line in a task’s thread is its owner’s to delete whichever agent put it there (the assistant and a local client too), the marker reads “[Removed]” next to “Removed by Jonas”, another open task panel shows the deletion without a reload, nothing of the line’s text is kept in the change, and an owner who was taken out of the space can no longer delete it. A reconciliation or a live-access change that loses its database lock stops before its next write, and its writes go through the lock’s own connection; statements queued behind one that timed out fail with it at once (#961).
- An owner who was removed from a space can no longer undo a removal there (a task-thread line, a map, a result), as intended in #961, but the check no longer reaches the other changes: a hand-off or an edit is refused in the domain’s own words (“Changed since. Open it”) again, one by one and in the whole-message Undo (#961).
- With live editing off, a Wiki tab that closes during a save no longer loses its newest text: the tab leaves a copy on this device that the next visit restores, and a version saved over another person’s version says which one it replaced (#967).
- A guest is decided by the role in the company, not by every workspace a person belongs to (the server refuses a company guest’s API key and plan sign-in); a task refused by “Give it to” names the agent and says it once; a question to an admin’s agent that cannot work in the space says so; Add me is one request on a double press; the hand-off refusal reads plainly (#968).
- A guest of the company cannot make an agent or sign in a local client through a role in another workspace (an encrypted space they were added to), and a client they already had stops; a question to an agent that cannot work in the space no longer guesses why (#978).
- Deleting an encrypted space tells its people: an event so their lists drop it at once and one Inbox notice naming the space and who deleted it. The server’s check of a recovery code the browser made rejects the alphabet in order, backwards, every second character and one block repeated, not only a code of a few characters. The New space window tells the creator of an encrypted space in one line that company admins will see its name, never what is in it (#980).
- Wiki without live editing: unsaved text a tab kept is dropped when a later version of yours already holds it, so a save that went through as the tab closed no longer shows as a conflict with text you saved. Copies a closed tab left on a device go when another account signs in, and a person’s cursor leaves the page for the others as soon as their editor loses focus (#982).
- Changing a task’s Owner in Details away from an agent now also ends the hold of a local client that had claimed it: its unit stops and its next call is told so, as after Stop. The unused Hand-off dialog and the list of a space’s connections behind it are gone (#985).
Security
Section titled “Security”- Production dependencies no longer carry a known HIGH or CRITICAL advisory: Better Auth 1.7.18, Fastify 5.12.8,
@fastify/static10, the MCP SDK 2.4 (one MCP request is limited to 1 MiB and 10,000 argument elements),markdown-it,zod,pg-boss12.38 and the other pinned packages are current (#877). - A password reset, a password change that signs out the other devices, and “Sign out all other devices” now disconnect the person’s local AI clients: their connections, runtime sessions and OAuth tokens end, so a client connected from a borrowed session no longer outlives the owner’s recovery. Each client connects again with one click on the consent page, and gets the same agent back. Better Auth’s own
/revoke-sessionsand/revoke-other-sessionsendpoints are off (#989). - Passwords need 12 to 128 characters, and a common or easy-to-guess password, or one built from the person’s name or email, is refused; the list ships with the server and nothing is sent anywhere. Sign-in limits are kept in PostgreSQL, so every API process counts together, and each account backs off after five wrong passwords (30 seconds, up to 15 minutes), whatever the client address, without keeping out a browser that has signed in to it before (a signed
flux.devicecookie), and the answer gives no exact wait; a reset clears it. Migration 0026 (#990). - An unexpected server error now answers
500with a code only in every space, and logs only its class and code: its message (a SQL or file-path text, or text it quotes from the request) no longer reaches the browser or the log. Fastify’s own 4xx answers keep their messages (#991). - Every response carries
X-Content-Type-Options: nosniffandReferrer-Policy: same-origin,/api/answers carryCache-Control: no-store, and an https public origin makes the API sendStrict-Transport-Security: max-age=31536000. The reverse-proxy guide says to pass it on (#992). FLUX_SESSION_IDLE(default 7 days) and the newFLUX_SESSION_MAX_AGE(default 30 days from sign-in, however much the session is used) set how long a session lasts;verifiedsign-up answers an address that has an account like a new one and mails its owner instead (#993).- The install guides create the owner account while the server is still on its loopback address, before it is public, and
./flux upand the API log warn while nobody has signed up on a public address, because the first account becomes the company’s owner (#994).
[0.1.0] - 2026-11-01
Section titled “[0.1.0] - 2026-11-01”The first release of Flux. It is published for linux/amd64; linux/arm64 follows in 0.1.1. There is no earlier
version, so there is nothing to upgrade from. The release notes say what
Flux is, how to install it and what is not in this version. Everything below was built since the repository was
created on 2026-09-26.
Run and operate Flux
Section titled “Run and operate Flux”- Operators can offer owners their own Claude Code inside Flux (off by default,
FLUX_AGENT_RUNTIME): a fixed pool of isolated runtime slots with no Docker socket,./flux runtime status|release|purge, and slot volumes kept out of backups (#303). - One Docker Compose application with the API, a separate worker, PostgreSQL and a durable job queue. Migrations are reviewed SQL with health checks (#34).
./flux up,./flux demoand./flux devstart, seed and develop Flux with only Docker installed (#102).- Backup, restore, project export and upgrade through
./flux backup,restore,exportandupgrade(#127). - Operator Compose files and versioned installation assets for released images (#159), plus a gated, explicitly triggered OCI release workflow (#104).
- Calls and screen sharing in a release install: each release ships
compose.live.yaml(the LiveKit media server, pinned by digest) andcompose.live.turn.yaml(TURN/TLS, added to that same service) to layer on itscompose.yaml, with the ports and firewall inINSTALL.md(#249). - An exact migration-ledger check at install and startup. A database that does not match the image’s migrations refuses to start (#130).
- Two public extension contracts with a written compatibility promise: MCP tool contract 1 and project export format 1. A versioned snapshot test catches a breaking change. The integration guide is for integrators and operators (#260).
- A tested HTTPS reverse-proxy guide (Caddy) that keeps one public origin for the UI, sign-in, MCP and live updates, and trusts only the measured proxy address (#366).
FLUX_BACKGROUND_COMPARISONS, an operator switch that registers the comparison runtime; off by default, with behaviour unchanged (#212).- Live media: a recorded encryption boundary (hop-by-hop, not end-to-end) (#284).
- A documentation site at https://coldphase.github.io/flux/, generated from the repository’s user-facing docs, and a README with the Kreska header, badges and highlights (#740).
- A self-hosting section in the documentation: requirements, a production installation on your own domain, tested Caddy, nginx and Traefik examples, an operator architecture diagram, operating encrypted spaces, health, monitoring and logs, capacity, hardening, troubleshooting and FAQ, and administration in the app, with a configuration reference generated from the code that the repository check keeps in step (#780).
./flux upasks once, on its first interactive run, whether people on this Flux may sign in with their own ChatGPT and Claude subscriptions, and the new./flux ai onasks the same on an existing instance. Enter means yes (Codex); Claude Code needs a typedyto Anthropic’s Commercial Terms, which writes today’s date as your statement. Hosts whose Docker engine cannot run the agent runtime (no Yamaptrace_scope, no internal networks, no limits, Docker Desktop) are not asked and get one line saying why; without a terminal nothing is asked or written (#494).
Encrypted workspaces
Section titled “Encrypted workspaces”- Operators: a new
keyringservice (encrypted workspaces, F-030, S1) runs on an internal network withapiandworker;./flux upgenerates itsFLUX_KEYRING_TOKEN.api,workerandkeyringnow have memory limits with swap limited to the same value (FLUX_API_MEMORY2g,FLUX_WORKER_MEMORY1g,FLUX_KEYRING_MEMORY128m) and core dumps off. (#655). - People can make an encrypted space from Home → Details → People → New space: a generated six-word unlock password or their own, a recovery code shown once to print, and no AI in it, for good. An encrypted space that waits to be unlocked shows as one row (“Waiting to be unlocked”) and as a locked panel at every address inside it; owners and admins unlock it there, and open tabs follow the lock and unlock without a reload. People → Encryption lets owners and admins change their unlock password, lock the space, let another admin set a first password and make a new recovery code, and says what removing an owner or admin changes (#655).
- Encrypted workspaces (F-030, S2b): search and live editing of docs and maps work in an encrypted workspace while it is unlocked. Each API process keeps an in-memory search index of an unlocked workspace (never written to disk, dropped when the workspace locks).
FLUX_SEALED_SEARCH_MAX_BYTES(default 256 MiB of estimated heap) bounds one workspace’s index andFLUX_SEALED_SEARCH_PROCESS_MAX_BYTES(default a quarter of the API container’s memory, at most 512 MiB) all of them in a process; a workspace that does not fit answerssealedUnavailableand idle indexes are evicted to make room. Live Wiki and map updates, snapshots and receipts are stored sealed, and locking a workspace closes its live rooms. (#655).
People and access
Section titled “People and access”- Email sign-up and sign-in with origin-checked sessions (#39).
- Workspaces, projects, grants and private drafts behind one authorization check (#42).
- An authorized live event stream; the worker rechecks access before it acts, and idempotent writes are safe to retry (#47).
- Optional single sign-on through one operator-configured OpenID Connect provider, next to email and password. Accounts are keyed by the provider’s subject, never by email, so an SSO identity cannot take over an existing account, and the provider’s groups or roles grant nothing in Flux (#240).
- Workspace People: owners and admins add an existing account by email, change roles and remove people, with project access controls and opt-in promotion grants (#198). An open project’s access line says accurately who can see it (#243).
- Settings → Account changes your name, email address (with your current password) and password, and lists where you are signed in with Sign out per device (#763).
- Settings → People shows each workspace’s people and roles in one place, also from the account menu; owners and admins can invite someone who has no account yet with an invite link (F-035) (#763).
- Someone who manages two workspaces chooses where a new project goes with “In … ▾” (#763).
- Owners and admins can invite someone with a one-time link from a workspace’s People: choose member or guest (admin for owners), optionally lock it to one email address, and pick how long it works (1, 7 or 30 days). The person opens the link, signs in or creates an account, and joins with that role. Pending links are listed with Copy and Revoke; a link works once, and Flux sends no email itself. In an encrypted workspace an admin who joins gets an unlock password only from someone who has one (F-035, #763).
Working together
Section titled “Working together”-
Search finds projects by name for the people who may open them; a result opens the project, and a Projects filter narrows the results (#465).
-
Installable PWA shell and Web Push notifications (#49).
-
The calm web app shell and design system (#55), later the project surface with view tabs and a state line (#122).
-
Project capture and conversations with replies and cited materials (#60).
-
Sketches: persistent maps of connected thoughts (#100).
-
Work items, decisions and results linked to the conversations they came from (#105). Native tasks gained criteria, prerequisites and plan intent (#171). A person who can only read a proposed decision sees who decides it: someone who can edit the project, never an agent or assistant (#259, O-009).
-
Canonical task discussions:
-
A return view: what changed since you left, its sources and one next step (#110). It also offers a private, sourced “What matters” recap (#140).
-
Direct messages that do not belong to a project (#111). A sketch can live inside a DM and be promoted to a project with an exact audience (#129).
-
Project docs and wiki with versions and links (#115).
-
Notifications: an inbox, per-reason preferences and email delivery (#120).
-
Privacy-safe search across everything you can read (#119).
-
Self-hosted live sessions (#109):
-
Phone task views. Your own messages sit on the right and other people’s on the left (#138).
-
A project can connect a read-only GitHub App repository and link tasks to verified pull requests, from signed, deduplicated webhook deliveries (#168).
-
“Let linked PRs move this task”: a person who can edit a task lets its required pull requests start, block and finish it, or mark it Ready to close. A manual status change pauses it until resumed (#270).
-
A task made in Tasks can be linked to a map thought from its Details: a “Linked thoughts” section and a “Link to a thought” picker for the project’s maps (#449).
-
Messages send instantly: the field empties at once, the message shows “Sending…” in place, and a failed send offers Retry or Remove. Offline, a line says so and queued messages resend automatically (#299).
-
A phone-first shell: on phones, Home and project views sit in a bottom bar, with a clear current place, motion and Settings (#267).
-
One chronological project conversation with no mandatory topics; replies sit with their message in a reply drawer (#195). Creating a task posts one compact announcement in the stream, and the task’s discussion root shows in the stream and Details (#224).
-
Project conversations and task discussions take durable file attachments, including file-only messages; uploads stay private until sent, and DMs refuse attachments (#225). A task’s unsent draft and staged files are shared across Conversation, Details and Agents (#229).
-
Tasks open as a Kanban board with drag, keyboard and menu moves; the grouped List stays one switch away (#194).
-
A two-pane Wiki with a page index, New page and Markdown import, and a top bar to edit, see history, export and share (#197).
-
Each map thought shows how many tasks link to it, and the count opens a chooser of those tasks (#196).
-
Home lists your tasks and first notes, and visiting Home acknowledges nothing (#211). Project sketches open inside their project, among other flow fixes from the Studio 11.6 audit (#210).
-
Choosing a project in the narrow sidebar (a tablet’s 64 px rail, or the rail a computer collapses to) moves the same quiet highlight to it at once, while the project loads, as the expanded list does (#155).
-
A task’s detail panel now shows who else is writing about the task, in the same quiet line and with the same rules as the task’s thread in Conversation and in Agents: one person, never yourself, ended by leaving the field or sending, and never stored (#155).
-
Maps and the Wiki are edited together live with identified authors: a thought moves on a shared map while a person drags it, Wiki text is written by several people at once, and a map shows who is here with a named cursor, each person told apart by a shape and not by colour alone. An operator can turn it off, unsaved text is kept for recovery, and saved versions stay immutable (#389, #585, #613, #646, #648).
-
Link cards show the linked page’s own title and description in three lines instead of the address. The Flux server fetches the page, only when the operator sets
FLUX_LINK_PREVIEWS=on(off by default): publichttp/httpsaddresses on the default ports only, redirects re-checked, 3 s and 256 KiB limits, cached in memory, rate limited per person, and never for an encrypted workspace. No image is fetched (#469). -
A message can cite a map thought or another message. Cite in a message’s menu, and Cite on the map, put it in the composer as a tile you can take out; the sent message shows a card with the object’s glyph, its kind, its title and one line of context, and it opens the thought or the message. A target the reader may not open (another project’s, a direct message, a private sketch) is a muted “Not available to you” card with nothing of it (#469).
-
A file row names what the server read from its bytes: “PDF · 2 pages · 310 KB”, “Table · 48 rows · 4 KB”, and the playing time of a voice note (#469).
-
A task can have a due date (a calendar day, no time zone) and keeps its history: every change of its state, owner, due date or title is recorded with who made it and when, and
GET /api/v1/work/:workId/activityreturns it newest first (#488). -
In a private message, reply to a message in a private thread (the count stays under it), see who really saw your messages (“Seen by”), and copy a chosen part of a message into a project task after seeing the text, the project and who will see it (#534, #536, #537).
-
A project can have one optional goal, one sentence shared by everyone in it. People who can write set it and mark it reached from Details; the conversation then shows Kreska once with Write a wrap-up and Keep going, and the project goes on. Projects without a goal look as before, and agents can read the goal but not change it (#520).
-
A blocked task can name the one person it waits for (Details, the Create window,
flux_create_task/flux_update_taskaswaitingForUserId); they get an Inbox item and a push, and the conversation shows one line when a task is blocked or unblocked (#550). -
Following a reference (a wiki card, a source message, a search result, an Inbox item, a person, an agent link) now shows one quiet “‹ Conversation”, “‹ Search” or “‹ Inbox” control where you land, in the header row and in place of the phone’s back (an object opened inside the panel keeps the panel’s own “‹ Task #12” / “‹ Thread” step). Going back by it, browser Back, Alt+← or the phone’s back gesture returns to the same page, filters, open thread or panel, reading position (also when new messages arrived) and unsent draft. Following a reference adds one history entry; panels and threads keep opening and closing in place. A direct message and its thread come back to the message that was in view. A place you have lost access to (a project, a material, a direct message you left) has no control, and Back lands on the nearest place that opens, with one short line (#592).
-
A decision proposed or accepted in a thread, and a result recorded, now show as a line in that thread and in the project conversation, with the decision or result as a card that opens it and shows where it stands; other members see it and its state change at once (#694).
-
Flux can be installed from Settings → Appearance: the browser’s own offer where there is one, “Share, then Add to Home Screen” on iPhone and iPad Safari, and nothing once it runs installed. An installed Flux opens at its start page with your session, a link reopens its project, and a closed app receives Web Push that opens the item. A refused notification permission is said in Settings and the in-app Inbox still works. An unsent phone draft survives applying an update, and the update prompt is a row above the shell, never over controls (#756, #622). Operators: mobile and push.
AI that stays yours
Section titled “AI that stays yours”- Connect your own agent to Flux through scoped MCP and OAuth (#103).
- Several named connections per person, with consent bound to each request (#167).
- Standing-grant actions for tasks, results and decisions (#174), and for the shared map (#176).
- Standing-grant actions for project docs and the project conversation, with the agent named as the author (#193).
- A versioned co-work playbook delivered through Start/Resume prompts (#175).
- Co-work tools for connected agents: create a work unit, and claim or decline a request addressed to it (#254, #255, #261).
- Owners grant, narrow and revoke standing grants for each of their own connections on the Connect page (#262).
- Owner-only personal assistant runs. Only the owner can ask, stop or pay (#141, #142). Crashed runs recover without a second invocation (#161).
- Owner AI connections work with any provider and model on equal terms (#192).
- A project Agents view lists every connected agent, one entry per connection, with the task thread (#183).
- Owner background setup with paused rules and sourced comparisons. These are prepared, but production activation stays off (#124).
- In a project’s Agents view, workspace owners and admins edit and publish the agent policy (scope, priorities, review criteria, allowed work) (#214). A conflicting publish keeps your text and shows what the other manager changed; everyone else reads the policy with its managers named (#292).
- Kreska’s small moments: the startup splash, loading, empty Inbox, offline, no-results and agent-thinking faces, pull to refresh, and the “#N done” toast wink. Settings → Appearance → “Kreska in loading and empty screens” turns them all off for this device, leaving plain text, including the startup line and the sidebar and working-line status (#382).
- In a task’s agents’ thread, handoff, review and teammate requests appear as live cards at the end of the thread with their handoff packet (
handoff-packet.mddownload and Copy), Accept / Decline for the recipient, Cancel for the requester or a project manager while waiting, and a project review rule in Agents → Project policy (#153, #400). - Your assistant can run on your own Codex: sign in with ChatGPT on Settings → Agents and AI (Flux shows the one-time code with Copy and open ChatGPT; an OpenAI API key or access token are under Other ways to sign in). Codex runs with every local tool switched off and Flux as its only tool server; a run that shows any other tool is stopped with nothing posted, and a ChatGPT plan limit pauses it with the reset time when Codex reports one (#281).
- You can switch off single tools and projects of one of your MCP connections in Settings, within what the connection was granted. A connection none of whose selected projects is reachable any more is refused before any tool runs (#390).
- Undo for a task an agent or the assistant created and nobody has used yet: it keeps the task’s history, its sources and who made it, appends a record of the reversion, and leaves the task as read-only history. A task that has been used cannot be undone this way (#394).
- Owners sign in to Claude Code from a console in Flux (a terminal on the real CLI’s own sign-in, with no key or token field), and signing out or removing the runtime also deletes the keyless Console profile Claude Code keeps outside its configuration directory (#279).
- Your assistant can work on a schedule (Settings → Your assistant → On a schedule): a sentence, one project and a time in your own time zone, with Run now, pause, edit, delete with Undo and a history of the last 20 runs. A time Flux could not start within ten minutes is skipped, three failures in a row pause the schedule, and “Let your assistant write your morning summary” puts one line under your Home greeting with the same one push (#408).
- Signing in with Claude is now a card: Open claude.ai and Paste (Flux reads the clipboard and sends it to Claude Code’s own prompt), with the terminal under Show details and the Console and single sign-on methods under Other ways to sign in. Reloading the page, or a lost connection, while a Claude or ChatGPT sign-in is pending brings the same sign-in back in the same session, with its original deadline; another session sees nothing of it (#411).
- A connection can grant fewer permissions than an MCP client asks for. The consent page says what is and is not granted, the token carries exactly the granted permissions, tools outside them are not listed, a call to one says which permission is missing, and the Agents view names the permissions in plain words (#571).
- Your assistant’s work is easier to control: one message per run and a one-request queue, field-level Undo of its edits (done-when, prerequisites, waiting-for, leaving Blocked, a map’s name), and owner-only Apply, Apply all, Not now and “Ask me first” with the run’s authority named (#586, #597).
- You choose your assistant’s model when you turn it on, and change it from Settings or from the ask bar and the assistant’s panel in a conversation. It is one model everywhere: a change applies from your next request, a request already working keeps its model, and each answer’s footer names the model that answered. A model your plan or key does not offer is refused in plain words and the previous one stays (#596).
- An owner or admin can turn AI off for a whole workspace and back on (People, “AI in
”). While it is off, the server refuses every AI route, the workspace’s agents are inactive and their MCP tokens stop working, a run that was under way ends without posting, and the interface shows no assistant, Agents view, hand-off, agent mention or suggestion. Committed history stays. Only people see the change (#633, #634, #643, #653). - Undo of your assistant’s run now also takes back a map, a recorded result, a proposed decision and a message it made, as long as nobody has used it since (edited, decided, replied to, cited or copied); when someone has, the item says why and stays (#68).
- Co-work with your own agent: Connect and Agents offer Claude Code’s
/mcp__flux__start_workand/mcp__flux__resume_workand Codex’s./flux cowork start|resume, which resume the same project and thread from a saved checkpoint. A person who may stop an agent can also pause it; it pauses at its next safe point and can still save a checkpoint. Events of a linked GitHub pull request reach the connection whose agent holds the task, and an agent can link the pull request it produced to that task (#616, #649, #751, #760).
Changed
Section titled “Changed”-
Operators can re-key an account to its provider identity with
./flux identity link(audited, refused in SSO-only mode without--allow-sso), and people cannot remove their last way to sign in (#315). -
With one sign-on provider, operators can run
FLUX_SSO_MODE=prepareso existing password accounts link the provider in Settings before the cutover tosso, which closes password sign-in (#315). -
The Wiki reads as drawn: a “Pages” list beside the page, who edited it with their faces, an “On this page” outline, tasks with their state and decisions as cards in the text, and on the phone a row of page chips with 16px text and Edit at the bottom. The compact page count and action menu join the project header; on the phone the project title is the Wiki’s view menu, with no row of views, and New page and Import .md sit in the page’s menu. The phone reading area stays clear above Edit, including related links. Tables preserve sanitized content in scrollable regions, and unavailable references show as plain text. Inline Wiki-page references use a page-icon chip, and outline scrolling follows the person’s Reduce motion choice as well as the system setting (#374).
-
The phone has one floating tab capsule (Home, Projects, Inbox) with a round Search, one floating “+” that opens Create, and the project title as the view menu (Conversation, Tasks, Map, Wiki, Agents, then Details). Direct messages live inside Projects, there are no chip rows, and a conversation has no tab bar (#341).
-
The Inbox is one “Needs you” queue: proposed decisions you can accept, questions to you, your blocked tasks and mentions, with filters, Done (
E), Not now (S: tomorrow, next week or when a task is done) and Accept (A), each with Undo (Z). Home shows the top three, a “Continue where you left off” card and your projects; your private notes moved to the Sketchbook as “Your notes”. The Decisions view and its archive are gone; their links open the Inbox (#342). -
A new project takes one step: a name and a template (Build something, Research, Plan an event or Blank); Map, Wiki and Agents appear the first time a project needs them and can be added from More, existing projects keep every view, and sign-in and account pages follow the final design with the plain Kreska logo tile (#351).
-
The computer’s project header is one row: the name, the views as a segmented control, what needs you, the people and agents, and More, which holds Details.
Fturns on focus: the sidebar folds, and push and email wait on the server until the shown time; Notification settings shows the pause with Resume now.]no longer opens Details (#340). -
With a single sign-on provider, password sign-in, sign-up and reset are closed; a provider sign-in whose email another account holds is refused, or can claim the address only when that account is unverified and unlinked; and
FLUX_SIGNUP=verifiedwithout email closes sign-up (#313). -
When an agent app opens Flux’s sign-in page, people can now use the operator’s single sign-on there, then pick the connection and approve as before; a cancelled sign-in comes back to the same page, and the page says when the provider cannot be reached (#310).
-
People who sign in through the operator’s single sign-on now lose access once the provider has not confirmed them for
FLUX_OIDC_CONFIRMATION_MAX_AGE(default 7 days, 1 hour to 30 days): their browser sessions return to sign-in, their agents get a 401 and must sign in again, and the connection an agent held is chosen for it (#312). -
With single sign-on, Flux now checks every 15 minutes that the provider still honours each signed-in person: when an administrator disables someone there, their browser sessions end and their agents’ MCP requests get
401“Sign in again with” until the provider accepts them again, with nothing revoked. A provider that gives Flux no refresh token refuses the sign-in with a message; FLUX_OIDC_STANDING=offturns the check off, and backups leave the stored tokens out (#311). -
Flux now accepts OpenID Connect back-channel logout from the single sign-on provider: when someone signs out at the provider, their matching Flux browser sessions end and their agents are re-checked at once, and with the provider’s “revoke offline sessions” setting their agent connections must be authorized again. Register
<origin>/api/v1/identity/oidc/<provider id>/backchannel-logoutat the provider. Sign-in now visits the provider twice, the second time silently for offline access, so Keycloak keeps the person’s single sign-on session; signing in on another device no longer stops an earlier device’s logout from reaching Flux (#314). -
The agent runtime retains unconfirmed sign-out history, treats unreadable client files conservatively, refuses startup without the required kernel isolation settings, and caps both egress listeners per source and globally; remaining installation-history and supported-host observations stay open in the operator guide (#385).
-
A project export now lists the project’s “Let linked PRs move this task” rules as dormant, disabled entries; connected repositories are not exported, because their names need your own GitHub authorization (#74).
-
Changing a task’s status or blocker by hand now pauses its “Let linked PRs move this task” rule at once, with a history line, and the rule carries a revision that a task change or a rule command can pin so you override only the state you saw (#74).
-
The map fills the work pane with floating tools: dragging from a thought’s dot connects it to another thought, releasing on empty space adds a connected thought as a draft until you save it, the selection’s actions float above it, and Create task sits in the bottom toolbar; on the phone the map is for viewing and adding a thought and says connecting and arranging happen on a computer (#380).
-
Agents is one list with each agent’s colour, owner and what it is doing now, then Requests, Policy and “Connect your own agent” and one “Hand off a task” button; handing a task to an agent takes two steps (choose the agent, then see what its access lets it do), by button, from the task’s details or, on a computer, by dragging the task onto the agent (#347)
-
Settings → Appearance adds Text size (Small, Default, Large) on the computer, a Reduce motion switch that turns motion off on this device as the system setting does, and a phone text-size row that follows the phone (#452).
-
The Agents view shows what each connected agent holds now (its task and whether it is working, assigned or paused) and its open requests with sender, task and why they wait, and Claude Code users start Flux work from the built-in Start work prompt (#160).
-
A working external agent is always visible with Stop: on its row and in its panel in Agents, on the sidebar card and on Home when one of your own agents works, and in the phone conversation header. A project manager, the agent’s owner or the task’s creator ends its hold on the task and its co-work claim; the agent’s next call is told it was stopped, and Agents says “Stopped by
” (#347). -
An agent can ask a person a question with two to four ready answers (
flux_ask_question): it appears in the conversation as a card with one-tap answers and “Reply in your own words”, the person asked gets a question notification, and the answer is recorded once as their reply in the same thread (#347). -
Faster reads: the Inbox count no longer compiles query code on each visit, conversation and material lists use one statement instead of one or two per row, and conversation roots and comparison outcomes are read without locks (#300).
-
The app’s files are sent as Brotli or gzip, and large JSON answers are compressed (#269).
-
Settings is one calm place with Account, Appearance (Light, Dark or Match system, and a switch for Kreska’s small moments), Notifications, Agents and AI and Keyboard shortcuts. Notifications now default to Only “Needs you” (replies no longer push), with Everything or Nothing, quiet hours and an optional morning summary push (#370).
-
One neutral look from the final design: Soft volume surfaces in light, dark or Match system, Geist type served with Flux, pill buttons, and task states shown as shapes with words. The accent colours and their picker are gone (#353).
-
Kreska, a face drawn in a few lines, is the Flux logo, the app icon and every agent’s icon. Agents carry an “Agent” tag and, where you may know them, the person they work for; the personal assistant’s face shows what its run is doing (#356).
-
The conversation as drawn: every author has a 32 px face and full name on one left edge, the composer offers
@to mention and/for actions (/task,/decide,/handoff,/file) instead of Ask and Sources buttons, a message shows hover actions on a computer and swipe left plus long press on a phone, task notices say where the task came from and fold when several arrive together, a “Since you left” line jumps to the first unread item, and your assistant shows “is thinking…” with Stop (#343) -
Tasks as drawn: board columns and list rows show each task’s state as a glyph with its word, its
#number, title and owner, with Blocked as an inverted pill. One tap on the glyph or the keys1–5change the state, on the phone you can also swipe left then Done, and a toast offers Undo (orZ); the phone Tasks screen is a list with only Mine | All (#346). -
The computer sidebar as drawn: New (C) and Search at the top, Home, Inbox with its count and Sketchbook, projects as letter tiles, Messages, a card for your working assistant with Stop, and the account with Settings.
[folds it to a 64 px rail andGIopens the Inbox (#357). -
One Create window (a sheet on the phone) makes a task from New or C, Tasks, a message and Search, with the same fields everywhere, “Create task” as its only name, and the message kept as the task’s source; ⌘K now also creates, and
@finds people and agents and#finds tasks (#345). -
The Studio 11.6 interface: a sidebar with workspace and project names, a calmer header, Conversation · Map · Tasks · Wiki tabs and a reworked conversation (#184).
-
The application lives under
app/and the Docker inputs underdocker/(#165). -
The compiled, non-root runtime image no longer ships TypeScript tooling (#43).
-
A desktop map that is narrow, as beside docked Details on a small laptop, shows its thoughts in two readable columns and keeps the computer’s toolbar, zoom and Fit (#151).
-
Files, references and photos sit inside their message: file rows with a type icon and Download, voice notes with Play, task and decision cards, inline
#4and@personchips, link cards, a frameless photo grid with send state on the photo, composer thumbnails, drop to attach and an always-dark photo viewer (#369). -
A photo or file sent while it uploads shows its upload percentage, then “Sending” until the message is stored (#369).
-
On a phone, choosing several photos opens a numbered picker: tap them in the order they should send, tap again to take one out, and the rest renumber (#369).
-
A file row in a draft or a message that is still sending names its type with its size, as a stored file does (#369).
-
A link preview’s address line (its site and path, read from the link alone, with no page fetched) runs to three lines before it clamps; the page’s own description is not shown yet. A Wiki page cited by a message shows as a compact card inside that message’s bubble, and it opens the cited version (#369).
-
Mobile Details and Thread share a sheet with a grabber, half and full heights and a visible keyboard focus ring. Task details have label/value rows, a blocker card and Results; the title and chosen status save in place, with keyboard state shortcuts and truthful Undo (#344).
-
Settings → Agents and AI lists Sign in with Claude and Sign in with ChatGPT first, with API keys under “Other ways to connect”; when sign-in is off, one plain sentence says so (#277).
-
Your assistant’s suggestions show only in projects where it is a member, a request to your assistant now appears in the conversation as sent (“@Ada’s assistant …”), “Ask my assistant…” in the search window asks “In which conversation?” outside a conversation, and when this Flux does not allow signing in with a subscription, “Tell whoever runs Flux” sends the workspace’s owners one item (#407).
-
The assistant’s tool runs send shorter Flux tool definitions and may carry up to 32,000 input tokens per request (a request without tools keeps 16,000), so the default per-request limit is now $0.08; the daily cap is unchanged. Answers from runs on your own API key now have token-shaped strings removed, like Claude Code answers (#484).
-
A task, a decision, an agent and a thread open in the same right-hand panel (one sheet on the phone); the separate thread drawer, the agent and task-thread panes and the phone’s linked-tasks sheet are gone. Opening a thread from a task’s “Came from” replaces the panel’s content with “‹ Task #1” back, and the conversation marks the source message “Source of #1” for two seconds. The task panel is calm (rows for Status, Owner, Came from and Due, edited in place; the blocker as a card; one row of small actions; Activity with its authors and times; the composer at the bottom) (#488).
-
A signed-in OpenAI API key now reads “OpenAI · API key” on Connect AI, the same plain account labels as “Claude · your subscription”, “ChatGPT · your subscription” and “Anthropic Console · API billing” (#493).
-
Home, the Sketchbook and New project no longer show the “Overview · Map · Tasks · Wiki” bar; Home keeps its quiet links to All my tasks, the Sketchbook and the Wiki. The Sketchbook is titled “Sketchbook” with one privacy line, empty untitled sketches are not listed (in the Sketchbook, a project’s Map or Details), sketches show their thought count, and a note’s time has no “You · v1” and no leading zero (#508).
-
Settings show the same sections in the same order on the computer and the phone (Keyboard shortcuts only where there is a keyboard). Agents and AI starts with one calm assistant card, with costs, data and consent on a “How it works” page; your own API keys live on one AI connections page; and Background suggestions show one status line above your usage and project rules when the server does not run them (#513).
-
Search results are grouped (Wiki, Tasks, Map, Messages, Files, People and projects) with “Show N more”, snippets are one plain line without Markdown, tasks show their glyph and number, and the place filter is the last chip of the kind row, which on the phone fades at its edge to show that it scrolls (#514).
-
A whole project can now be deleted in one step: its action receipts, standing grants, task discussions, drafts and proactive comparison rows go with it (#551).
-
The Map is an infinite canvas: drag empty space or use Move around to pan, zoom around the pointer, and use the minimap and Fit; thoughts can sit at negative positions. The List is the map’s tree: a thought sits under the thought its connection was drawn from, with “also under …” for its other parents, and Tab / Shift+Tab move the connection. Create task on the Map opens the Create window with the selected thoughts as its sources (#554, #509, #510).
-
Light motion across the interface: dialogs and ⌘K, the detail panel and what it shows, menus, popovers and toasts fade and travel a few pixels in at most 180 ms and leave in 120 ms without delaying focus; switching a view or place crossfades the old content into the new in the content area only (and the detail panel’s old content fades out over the new, which slides in from the side of travel); a task’s state, a decision’s outcome, rows that move between groups, rows added and rows removed change in place instead of jumping; every pressable control takes the one press (scale .97, spring back in 250 ms); nothing animates on first load or while the page is hidden, and reduced motion (system setting or Settings → Appearance) shows the static end state (#593, PR #608).
-
Every dropdown is now Flux’s own select instead of the browser’s: a list in the app’s colours and in dark mode, a sheet from the bottom on the phone, and the same keys everywhere (arrows, Home and End, type a few letters, Enter, Esc). Opening one never moves the control (#695).
-
A decision in a Wiki page is one live card with its short reason, not a copied section: “Add to a Wiki page” inserts the card, the card opens the decision and follows its state without new page versions, and proposed decisions stay out until accepted (#774).
-
The database schema starts from one v0.1 baseline migration (
0001_v0_1_baseline.sql); v0.1.0 ships it with0002_first_time_user.sql(schema version 2); databases of pre-release development builds are refused with the instruction to remove them (./flux clean), and the pre-release upgrade shims (root.envmove, missing keyring secrets on upgrade, migration reversals) are gone (#779). -
./flux upwithFLUX_SIGNUP=offsays sign-up is closed and points to “Decide who can sign up” in the install guide instead of inviting an account (#779).
-
A new map draft receives keyboard focus after its editor becomes visible, without taking focus again when the editor is repositioned (#380).
-
Background comparisons resume within a few minutes after the worker is killed mid-run, instead of waiting an hour for the dead run to expire; the interrupted request stays counted as possible spending and is never retried (#58).
-
Task details hide empty relationship paging after a successful read and keep status/owner labels separated from their values with enlarged text (#282).
-
People and agents keep a full 32px author avatar in the same conversation column on phones and computers, including your own replies (#356).
-
Historical agent contributions can still name their currently visible owner after the agent loses project access, without restoring the agent or exposing hidden owner identities (#356).
-
Task board columns describe the tasks being loaded, and the Wiki link picker names its search while waiting for results (#363).
-
When someone’s sign-in and extra address are one mailbox and the notification email Flux sent there fails for good, the other copy is now sent instead, once, after rechecking access and preferences (#334).
-
The shared files volume works on SELinux hosts (#67).
-
Check scripts remove the images they built, so test runs no longer fill the disk (#73).
-
Project status and conversation navigation are correct for readers (#169).
-
Blocked project work stays visible on narrow phones (#172).
-
On touch screens, message fields, Reply, map Fit and wiki Edit have 44 px tap areas; the map keeps its place when a phone rotates or changes layout; wiki prose keeps a readable line length; and closing a sheet on a tablet no longer leaves the page untappable (#265).
-
The return view tells you when a pivot parked a task, even if the task itself did not change since you left (#293).
-
Security: a runtime CLI can no longer read or alter the agent runtime supervisor; switching the runtime off no longer blocks vendor logout, and purge no longer reports an unconfirmed logout as successful (#359).
-
A task root’s chip in the conversation no longer grows when its task number arrives (#368).
-
In the production build, drawers, sheets and popovers move for their full duration instead of jumping in one frame (#286).
-
Web Push reaches public push services again: a network guard had rejected ordinary public IPv4 addresses (#233). Flux warns about VAPID contacts that Apple’s push service refuses (#263).
-
An installed app’s title bar follows the chosen Light or Dark theme (#242).
-
Search says “1 result” in the singular and adds “+” when counts are capped; GitHub settings shows a line when a chosen task has no linked pull requests (#241).
-
A database connection whose BEGIN or ROLLBACK failed is discarded instead of going back to the pool with an open transaction (#235).
-
The map keeps its complete keyboard hint when task Details is open (#227), and Fit keeps thoughts clear of the zoom controls (#226).
-
Retrying a first space, project or private draft after a lost response no longer creates a second one (#178).
-
Security: a signed-in account can no longer register or change OAuth clients, so a member cannot pass off a look-alike client with their own redirect. The agent consent page shows where access goes and warns when that is not this computer, and no Flux page can be framed by another site (#287, #294).
-
The agent connection page no longer asks for an HTTPS address on the quick start’s local
http://address, and the README shows how to connect your own agent (#328). -
In WebKit and Safari the sticky “Hand off a task” button on the Agents view no longer sits under the agent rows when they are tapped, and Wiki prose keeps to 62 characters of width so a line holds at most 90 characters in every browser. A short conversation now sits against its message field instead of at the top of a tall screen (#151).
-
On a notched phone the message field, Send and the map’s Add a thought stay above the home indicator, and on a phone on its side the rail’s controls no longer reach into the left or bottom safe area or below the window (Settings is in the account menu there) (#151).
-
A short private reply thread in a direct message and a short task thread in Agents now sit against their reply field on tall screens, instead of leaving up to 940 px of empty space above it (#151).
-
On the phone, the thoughts you can link a task to read at the phone body size (15 px) and still wrap, keep their 44 px targets and follow the system text size (#289).
-
Codex, which offers MCP protocol revision
2025-06-18, reaches Flux’s tools:/mcpadmits that revision beside2026-07-28with the same sign-in, grant and provenance checks, and refuses any other (#467). -
A release or purge of the agent runtime no longer reads a missing Claude Code or Codex as “nothing to sign out” when Flux’s own record shows a login may exist for it: the database keeps a durable per-binding history of started logins, cleared only by a confirmed CLI logout (#480).
-
Own project messages now appear on the right, with other people and agents on the left. Map connections have visible controls to add, edit or remove a label, preserving drafts when a save fails or conflicts with another edit. #482
-
Adding someone to a workspace keeps focus in the form, so the next email address can be typed at once; before, focus moved to the people list a frame later and could take the first keystrokes. Pages no longer start API requests while the browser is already leaving them, which WebKit reported as errors (#491, #471).
-
On a computer, the toolbar beside a message (reply in thread, create task, more) shows for one message at a time and never covers a neighbour: it follows the pointer or keyboard focus and a click leaves none behind. The room for it is part of the layout, so a thread never scrolls sideways (#591).
-
Choosing messages for a sketch keeps every message in its place (yours on the right), with an even soft fill, the check on the message’s first line and the bar on the conversation’s column. Swapping the composer for the bar keeps the place you were reading (#594).
-
Encrypted spaces: one account can no longer keep others’ spaces locked by flooding new spaces or new recovery codes (five new encrypted spaces per person in 15 minutes, one password check at a time per person). A new space leads to its first project; its recovery code step can’t be closed by accident and People says when nobody confirmed keeping the code; the generated words are offered to a password manager; the code reads in one order on a phone. The locked panel says who locked it or that the server restarted, and who can unlock it; People → Encryption unlocks in place after Lock now; a locked link on a phone keeps Back and the tab bar. Printing a page no longer comes out blank, and a former member’s messages say “Former member” (#655).
-
Task, result and decision cards under a message show their title and their whole state in a thread panel, a thread sheet and a phone’s conversation, as the Spokój one-row card (icon, number, title with an ellipsis, state at the right edge), instead of a few characters of the title with the state on top of the rest; the conversation column on a computer keeps the drawn card (#665, PR #680).
-
./flux cleanand./flux resetwork after a first./flux upthat stopped before the migration ran (#671). -
Codex signs in with a connection that grants fewer than the three actions it asks for (its authorization request is no longer refused), and an agent’s suggestion is a “Needs you” item for the person who connected the agent, with Got it and Decline (#673).
-
Stop, Retry, Accept, Apply and Undo of your assistant keep the keyboard where it was when the button goes away, typing
/aikey by key no longer leaves/ain the reply box, and a result opened from a thread button takes focus into Details instead of dropping it to the page (#68). -
A Wiki page no longer rebuilds itself a moment after it opens: a Share or Download card you had just opened, and the focus, stay where they are. The shared editor’s preview now keeps wide tables in their own scrollable, keyboard-reachable region, as the page does (#701).
-
Changing view on a phone (Conversation, Tasks and the rest, through the title menu) no longer holds the screen for more than a tenth of a second at a time (the worst gap between two frames fell from 133 to 83 milliseconds at the 95th percentile on an emulated phone four times slower than the test machine): the reads a view asks for as it opens now start in a task of their own, after the view has been drawn; each message’s time label shares one date formatter; the picture of the old view that fades away copies only the rows near the screen of a long list; and the conversation measures its spare height once instead of twice (#716).
-
Switching between a project’s Conversation, Map, Tasks and Wiki no longer flashes: the old view stays whole and at full strength until the new one has painted and then dissolves into it (it used to fade as see-through ink over the new view, or while the new view was still empty), in light and dark, on computer and phone, and with reduced motion it simply switches. The computer’s sidebar now folds to its rail and unfolds with a short size change instead of snapping, with the toggle staying where it is and keeping focus (#719, #720).
-
A long conversation on a phone opens with its newest message against the message field; the map opens at a readable size with Fit as the overview, and its first-run hint steps aside from thoughts; on a short window the rail’s places and projects scroll as one band with a soft edge (#725, #726, #727).
-
Another member’s message appears in an open conversation or thread when it is sent, not after the 15 second refresh (#741).
-
On a shared map, undoing a change no longer reports “The map change is not confirmed” when the undone thoughts were still selected, and a connection label whose sketch has become unavailable opens the not-found page with the unsaved label to recover (#749).
-
The map’s first-run hint (“Drag empty space to move around …”) now steps aside, with a short fade, when a thought lies under it, as it was meant to, instead of sitting on the card’s text at 1280×800 and 1440×900 (#757).
-
A page that is being left no longer starts the shell’s reads (working agents, “needs you”, notification preferences and the others) that it had asked for a moment before; WebKit reported those as uncaught page errors and failed otherwise passing tests (#758, #759).
-
Agents and Connect your agent read as one calm place. A paused agent says “Paused by you · #1” with its asleep face on its Agents row and in the sidebar card, as its panel does; the toast after Pause offers Resume; a task an agent holds without working on it reads “Has #1 · …”. Stop lives in the agent’s panel and the sidebar card, no longer on the row; the panel’s Message, Pause, Stop and Hand off keep their places and are dimmed when they can’t be used. The empty “Agent questions” row is gone and the hand-off drop zone is one line. Connecting your agent is now a Settings page (Settings → Agents and AI → Local co-work) with one tab per client (Claude Code, Codex, other) and the way back on the left; release-verification notes and checkout wording are gone from the page. During a client’s sign-in, “Continue to consent” also saves a new connection, its client label is preselected from the asking client’s name, and Revoke stays in Settings. Settings lists a connection as “Claude Code · for you · Desk laptop”, and the app has one main landmark (#763).
-
A proposed decision’s panel now starts with the same answer as its Inbox card: Needs you, Accept, Not now, Discuss and who else can accept, inside a phone’s half sheet; empty “Based on”, “Evidence”, “Reports on” and “At this pivot” sections are gone. Discuss opens the decision’s thread (or the conversation with the decision beside it) ready to write. “Add to a Wiki page” closes back to the decision with one toast and Open. “Hand off to an agent” with no agent explains that and offers “Connect your own agent” instead of making a plain task. A phone’s long press lifts the message above the blur and shows the five drawn actions under it; the computer’s More actions no longer repeats Reply in thread and says “Record a result”. A task made from a message is titled with its first clause, selected (#763).
-
Phones and tablets: a live session started from a task’s Details keeps Mute and Leave usable above the panel; the header Back after opening a shared
?open=link stays in Flux; Home’s navigation at 641–680 px (Galaxy Z Fold 5 unfolded) is on screen; Inbox actions fit or wrap inside the card on narrow phones; a Wiki decision card opens in place on the phone; the Inbox names each person who can accept once, keeps compact titles to two lines and Home’s list scrolls under the “+” instead of being cut off (#763). -
Quieter screens: the Wiki’s Edit is a soft 32 px toolbar button, the microphone-on state is a soft fill, the live strip keeps its place on Home, very wide windows align the Inbox filters and Settings with their page, and key hints on dark buttons are legible (#763).
-
A first project no longer asks a new person to name a “space”: their own workspace is made with it, and the page says “Only you can see it until you invite someone” (#763).
-
New (C) opens on Project while you have no project, and ⌘K offers New project instead of a task that cannot be made (#763).
-
The splash is dark in dark mode before the stylesheet arrives, with the logo visible (#763).
-
With in-app AI off on the server, Agents and AI shows local co-work and one status line each for the assistant and background suggestions, not set-up forms; Notifications offers only the channels the server has; settings switches and rows are 44 px on touch; mention chips use the same initials as avatars (#763).
-
Forgot password without email shows only the message and the way back; a field’s error replaces its hint; on sign-in Tab goes from the email to the password before “Forgot password?”; the account menu holds Settings and People (#763).
-
The Inbox opens only the selected item and shows the others as one line with their one action; a new person’s Home starts with “Start your first project”; a decision proposed from a message starts from a short title and offers Undo before it reaches anyone; connecting Claude Code needs one page with the connection named and the agent made behind it, and Continue says what is missing; an account made from an invite link joins as it is made; a hand-off can be undone (#763).
-
Consistent controls: the project views no longer overflow the header at 1024 px; a task’s state circle and the board’s “Move to…” are no longer cut; the header’s Together says why it is unavailable instead of sitting greyed out; toolbar and sidebar icon buttons share the 32 px control size with one icon size per row; the header’s agent face is no longer squashed; placeholders in every field have readable contrast in Safari; sign-in, sign-up and Create project buttons on the phone take the 44 px control height; settings titles use the type scale (#763).
-
Settings, Notifications: the per-kind table and the cards for the email address and muted places keep their padding again, instead of running to the card’s edge when the settings styles load after the notification ones (#810).
-
FLUX_SIGNUP(open,verifiedoroff) now reaches the API through both Compose files and the settings template, so operators can close or verify password sign-up (#780).