What's new

What's new#

What each release changed, newest first. Ask the Assistant "what changed in the last update?" and it reads this page.

0.3.11.17 - 2026-11-08#

  • A phone's task page is the view's height, its message box at the foot on every kind
  • The demo is a connected world on an invented machine
  • On a phone the chat box's From chip sits on a line of its own over the words
  • Invites, new meetings and a dashboard's share are set up like an email: a card and one press

0.3.11.16 - 2026-11-08#

  • The agent's own turn picks the lit button; the bar drops Reply and Hand it to a person
  • No Send is lit over a card whose channel is not connected; a phone's chat box stays at its foot
  • Search reaches the agent chat: the owner's lines, the agent's answers and its drafts
  • A video tool's card signs in with one press
  • A turn whose words equal its summary's What I did draws its words

0.3.11.15 - 2026-11-08#

  • An agent's email keeps its CC, and a withdrawn email can be written again
  • An email the agent set up waits on you: store.waiting_on_owner
  • A finished run's row says "agent finished", on the rail and in the header alike
  • A chat turn a restart cut keeps what it said, with one line that it was cut
  • A Send over a blank asks once - "This still has [ROOM]. Send anyway?"
  • A #task link opens its task over setup, and setup has a quiet "Set up later"
  • The demo's weekly dashboard runs on a clock the scheduler reads
  • The agent's next turn hears the owner's hand-off
  • A goal slipped by a late step says that step's day
  • The agent's next turn hears every press the owner made since its last turn
  • Opening a closed task shows it and leaves the Tasks band alone
  • A closed task looked up by any road carries closed
  • Each press the owner makes on a task is their line in its chat
  • A turn is drawn once: the agent's words, and only the cards it set up under them
  • #settings=audit off Developer mode says why it is not there
  • A send card whose channel is not connected says so, and its move is Connect
  • Every kind opens on its chat with Came from folded thin, unless it has no chat
  • Send and close is filled on every reply card unless a move drawn beside it is lit
  • A chat agent whose CLI died while it sat idle is noticed at the next sweep
  • Log says why a quick answer failed; expected outcomes are no warnings
  • An agent's browser closes with its session, its view answering or not
  • A task's agent is Agent everywhere; prompt machinery stays out of comment readers
  • An answer's raw file path is its file card, not a line of its own
  • Copy: one word for one thing, no internal words on screen
  • New's kinds wrap instead of scrolling out of sight
  • UI tests read the agent as Agent; Send anyway's label comes from buttons.js
  • Trust follows whose words are in a task, and who wrote a tracker item
  • A task bound to a teammate chat runs in safe mode by default
  • A stranger's session may not name the CLI's own sign-in variables
  • An unverified From on a verifying mailbox starts nothing by itself
  • A read-level database card runs one SELECT, checked before it connects
  • The browser gate judges an open that carries a token, and never reads a password box
  • A signed build runs only an installer signed by its own signer
  • The cloud-fix workflow pins its actions and CLI, and push keeps no token on disk
  • A task-bound agent session reads only its own task, and no other pane's screen
  • A run that finished with a post or act waiting reads as that press, not agent done
  • A reply with no mailbox leads with Connect email under the draft
  • One email card for the reply and the agent's email: To, CC, BCC, Subject, files
  • New: a general agent's job is typed into the agent chat's own box
  • The Assistant reports what the agent needs; no name yet is ? on the avatar
  • Said once: the ask that is its message, New playbook, a single page's back crumb
  • A held line wears the owner's head and stands under everything the turn drew
  • A phone's task view fits its chat body; a reply's card shows its subject
  • Tests follow the shared CC/BCC row and the New box; the send's envelope edits are one helper

0.3.11.14 - 2026-11-08#

  • Video tools for the agents: Descript, InVideo, Vidova, OpusClip, video-use, Kling and Runway video

0.3.11.13 - 2026-11-07#

  • Taskuary's own reports are never a new workstream
  • Dashboard charts draw again: an orphaned CSS line had dropped the palette
  • The token count is in the glance's top row, and its AI tiles sit up with the charts
  • White cards on a light page for the dashboard, and the End of day checkup laid out as a brief
  • An agent can show a table, chart or checklist in its chat
  • A table, chart or checklist goes through show, not a file
  • A shown view is not repeated in words
  • A chat agent's question is kept with the task, and a turn that asked nothing is done for now
  • On a chat task nothing tells the owner to press a Continue it does not have
  • Each turn keeps the file as it made it; the newest turn shows the file itself
  • Every line the owner sends is kept the moment it is sent, with the time it was typed
  • A line held behind a running turn takes the first line's road, and never says "reload the page"
  • After a restart, unsent lines go back into the owner's box, never to the agent
  • An agent's view draws every block as its own card, with tiles side by side and a chart beside its table

0.3.11.12 - 2026-11-07#

  • A plain to-do is the one chat too: no Close out section; it still opens on what came in (WIP, bundles not rebuilt)
  • The stage model has two sections, Came from and the chat: close_out is to_send, and the Advisor names no Close out section
  • An FYI handful says "6 messages", and the dock's files ride its box's own line
  • A line the owner sends is in the chat once, and its files leave the box with the send
  • A follow-up line their reply ended is marked done once, not on every rail build
  • Triage names no repository: it says a job is coding and nothing about where
  • The coding agent picks its checkout: use_repo and find_repos
  • No repository chosen is no hold on the page: the agent picks, the owner can still change it
  • A rail build waiting on the membership worker wakes it instead of waiting out its tick
  • Every task is Came from and its one chat: the Close out column, its tab, its pairing and its notice are gone
  • The settle wait looks again in milliseconds once it has woken the worker
  • Items lay no hidden Close out section: what registers their row's verbs is ItemSections' 'moves'
  • What the Assistant, the agents, the phone and the demo say: the chat, never Close out
  • AI used and 'Talk to this agent from…' are quiet buttons in the task's top-right corner, each a popover
  • Docs: a task is Came from and its one chat - no Close out section, stage or step names
  • One prompt box everywhere: every box that talks to an agent is the same component
  • Next turns blue as soon as the move saves: the walk asks for the light answer and reads the rail behind it
  • AI routing: an optional AI per level for the agents' work
  • The mail a task came in on stays in its chat after the owner's first line
  • Coding sessions over ACP: a task's claude or codex opens as one ACP connection in its checkout, the pty its fallback
  • Coding chat over ACP: the plan is a checklist, Stop ends the turn, the Terminal shows once a command printed
  • ACP coding: a file it touches is named from the checkout's long name
  • A stranger's task never runs over ACP, on either road, whatever a caller hands the gate
  • Docs: ACP step 2 - coding sessions on claude and codex, the registry they share, the stranger floor, the live check
  • Regenerate the assistant tools docs page after the Close out wording change
  • Install the agent connector beside Claude Code and Codex, and run only the pinned one
  • Task button row: four sections, one button size, the lit one alone filled
  • Hold every relay road when a task turns a stranger's under a full session
  • Setup's bottom bar takes the step's width, centred under its cards
  • Assistant's own icons back in the chat head on the home screen
  • One chat picker for the Assistant chat and the agent channels
  • Every dialog's actions are the task row's one ButtonBar: no band, left, one button tall
  • Playbooks back on the Reports page; About you keeps You, Docs and Memory
  • About you is the avatar in the top bar's corner, not a sidebar line
  • ACP coding: steer a running turn, idle out quietly, pin the plan
  • Setup: the job and what you work on are one step, About your work
  • Tell the agent when a task is the owner's own ask, and offer it no reply tools
  • Read a coding task's whole work so far as one diff
  • Write a brain's route row on its first call, not when it is built
  • Name the connector's press on a CLI card: Install connector, Update connector
  • Docs: the connector's presses are Install connector and Update connector
  • A failing membership tick waits out its backoff; a reader's kick no longer cuts it short (CI)
  • Remove AdvisorNote.jsx: nothing draws the task's Advisor section any more
  • Coding chat: no Terminal button, rows in arrival order, no spinner after the turn ends
  • Add a quiet Changes button to the coding task's header
  • Route a general session by tier unless the owner picked its model
  • Count each ACP turn's tokens; match model ids to the agent's spelling
  • Latest message wins: each turn answers the newest message, and a draft belongs to its turn
  • A task works in a checkout whatever its kind; a helper's calls name the call they run under
  • A helper's steps nest under its row; a call still running beside finished ones says it runs in parallel
  • One chat for every task's agent: coding and general share one component, one mapper, three detail levels
  • A stale draft stays under the turn that wrote it, read off the chat payload's drafts
  • The one chat keeps the agent's words at Summary and reads no general conversation on a coding task
  • An owner-opened own-ask session still hears it is the owner's to end
  • The owner's Allow or Deny on a permit goes through on a stranger's task
  • Latest-wins facts ride in the turn, so a CLI chat keeps one process
  • A resumed chat turn carries latest-wins only when a draft waits or a reply went

0.3.11.11 - 2026-11-06#

  • A plain to-do is the one chat too: no Close out section; it still opens on what came in (WIP, bundles not rebuilt)
  • The stage model has two sections, Came from and the chat: close_out is to_send, and the Advisor names no Close out section
  • An FYI handful says "6 messages", and the dock's files ride its box's own line
  • A line the owner sends is in the chat once, and its files leave the box with the send
  • A follow-up line their reply ended is marked done once, not on every rail build
  • Triage names no repository: it says a job is coding and nothing about where
  • The coding agent picks its checkout: use_repo and find_repos
  • No repository chosen is no hold on the page: the agent picks, the owner can still change it
  • A rail build waiting on the membership worker wakes it instead of waiting out its tick
  • Every task is Came from and its one chat: the Close out column, its tab, its pairing and its notice are gone
  • The settle wait looks again in milliseconds once it has woken the worker
  • Items lay no hidden Close out section: what registers their row's verbs is ItemSections' 'moves'
  • What the Assistant, the agents, the phone and the demo say: the chat, never Close out
  • AI used and 'Talk to this agent from…' are quiet buttons in the task's top-right corner, each a popover
  • Docs: a task is Came from and its one chat - no Close out section, stage or step names
  • One prompt box everywhere: every box that talks to an agent is the same component
  • Next turns blue as soon as the move saves: the walk asks for the light answer and reads the rail behind it
  • AI routing: an optional AI per level for the agents' work
  • The mail a task came in on stays in its chat after the owner's first line
  • Coding sessions over ACP: a task's claude or codex opens as one ACP connection in its checkout, the pty its fallback
  • Coding chat over ACP: the plan is a checklist, Stop ends the turn, the Terminal shows once a command printed
  • ACP coding: a file it touches is named from the checkout's long name
  • A stranger's task never runs over ACP, on either road, whatever a caller hands the gate
  • Docs: ACP step 2 - coding sessions on claude and codex, the registry they share, the stranger floor, the live check
  • Regenerate the assistant tools docs page after the Close out wording change

0.3.11.10 - 2026-11-06#

  • AI routing: Auto by default, effort for one-model brains, plain words in Settings
  • One clock on a running turn; the task's agent says Agent, the corner pill is the Assistant
  • AI routing: API cards with one model think less or harder too
  • A routing rule's sentence never repeats its job
  • One status cycle for a chat task: working, waiting on you, agent done
  • Check-backs: the owner can cancel one, and a chip says when it wakes
  • A nothing-new turn is one line: the agent judges it, the chat draws only the line
  • The day's opener names an Advisor idea by its own words
  • Attachments: one chip everywhere, a few in the row and then +N more
  • The rail plays a short confetti burst when the walk reaches All done
  • The rail's clear-burst count updates only when the count changes
  • The task's title bar shows its check-back chip beside the status
  • Items open on one box with their Came-from chip; an FYI handful gets the single one's box
  • Opening a closed task from Search only opens it; Reopening… never outlives its press
  • Every non-coding task is one chat with one box; the task's buttons are one row in one place
  • HIPAA AI card takes the effort setting
  • AI used: every call keeps its model, tokens and time; a dashboard and a task endpoint read them
  • The general agent can ask a stronger model itself when it is stuck
  • TaskButtons: the order lives in actionRow.taskButtons, and the row folds around Mark done and the lit move
  • The task shows what AI it used, behind one quiet "AI used" line

0.3.11.9 - 2026-11-06#

  • A reply block glued to the agent's words is its own paragraph, shown once
  • A chat reply's To line names the app: "in Teams", not "in teams"
  • The open chat reads a turn the server woke: new mail, a check-back
  • A check-back's turn opens on what it checks; the server's start keeps the clock
  • A report made a task: its row keeps its ref, never "Mo…"

0.3.11.8 - 2026-11-05#

  • One message everywhere: Taskuary is your AI work partner - your work, in the right order
  • An input bubble is its icon, who and when - no "sent to Taskuary"
  • A post card's picture is a preview, capped at 180px
  • No stop icon on a chat task; the chat box stops any turn at work
  • The card holds still while the walk brings the next one
  • One prompt box at the pane's foot, one attachment chip in one line

0.3.11.7 - 2026-11-05#

  • A mail read in mid-turn stands after that turn, over the re-think it woke
  • A turn that ended draws as ended: its own run length, no step still spinning
  • Copy at the corner of an answer and a code block
  • No "Taskuary answered" line over a chat
  • Send and close keeps the sent reply in place and says it went
  • "Starting…" is a quiet line like "Working on this…", with an honest clock
  • A burst from one sender reads as one; a summary keeps its breath under prose

0.3.11.6 - 2026-11-05#

  • An agent's lesson goes to LEARNED.md, named for its kind of job - never a card on the playbook
  • A lesson is only what changes how the next job like it is done: a lasting fact, a rule about the owner, a route that works - never a log of clicks
  • The general agent reads what LEARNED.md has learned about the owner: the trusted lines nearest its task, small, the rest on ask
  • A message that changes nothing gets one line from the agent - nothing new, the plan stands - not a re-run
  • The check before a turn shows only when it found new messages, and never calls a WhatsApp thread mail

0.3.11.5 - 2026-11-05#

  • An agent can check back on its own task later, on the reminders' clock
  • The agent searches what earlier tasks settled before asking the owner
  • Playbooks are a menu the agent reads, and learn from its turns

0.3.11.4 - 2026-11-05#

  • Fix built-in Windows dictation recovery and add voice selection card
  • A held press is the owner's own click on the real page, not an Approve card
  • Each connector declares its act's card; one card draws it, with per-connection defaults - LinkedIn audience and post now or draft first
  • A Slack line is answered in Slack: in its thread, in Slack's own markup, through the card that read it
  • Every hand-over on the page is watched; a desktop press is a line in the chat
  • The agent keeps going when a way fails, and a payment's last press is the owner's on the real page
  • An Advisor idea reads as its finding and what to do in one line, with no Whose move row; Based on stays
  • The Advisor's looking-out pass runs each morning and proposes up to four ideas across families
  • The demo world gets a general agent's finished working day
  • The static demo is re-recorded with the agent's day, and carries the files it made
  • README and docs show a task as one chat, reshot against the demo world

0.3.11.3 - 2026-11-05#

  • A trusted repo's own steps go through: a push, a pull request or a workflow run on a repo whose card may act no longer asks
  • A card's field saves only its own keys, merged over the config as stored; Remove takes an added card off the list
  • Every field on a connection card saves itself, and Remove asks once and takes the card away
  • A poll, a Test or a sign-in merges only its own keys over the card, so it never rolls back a field the owner saved meanwhile
  • The chat reads like a chat: local times, every input shown before the turn it woke, a turn's draft kept read-only in its history

0.3.11.2 - 2026-11-05#

  • The agent never closes a chat task: only the owner's Mark done, Send and close or Close with a note does
  • A saved sign-in goes straight in and raises no ask, so no browser is shown for it
  • Typing in a chat task is the way on: it reopens a closed one and carries on, and Continue and Run another are gone
  • A chat task draws the agent's browser only when the owner has a part on it, beside the chat

0.3.11.1 - 2026-11-05#

  • A chat task draws no Advisor line: its agent is the advisor
  • Reports, FYIs, dashboards and Advisor items have no Close out, and their box reads as a chat; an Advisor item answers questions too
  • An item's ask box sits under what it opens on, never in the Agent tab of an item no agent was sent to
  • The agent says its reply waits under Next steps, never in a Close out a chat task no longer has
  • A quick turn ends in the same What this is about / What I did as an agent turn
  • A burst of mail on a thread is one re-think, and a line is never lost to a session that just ended
  • The chat box takes a line while the agent works, shows no Suggested reply before one exists, and stays at the card's foot
  • One playbook proposal per task while it waits, never one per session ending
  • "Suggested reply" heads a reply in the chat, never a closed task's Mark done line
  • The agent's question is said once: never again as an On you step under it
  • Send and close closes a chat task whose agent is parked: the press stops it, saved, and closes
  • A chat task's chat has no agent picker, model box or Full screen: only the task's own actions
  • A chat reply's To names the person, never their phone or chat id, and no room for a one-to-one chat
  • When the thread moves, the agent says plainly what changed first: "I see you answered Ray there"
  • The owner's own reply on a thread wakes every open task on it, not only the one it is filed on

0.3.11.0 - 2026-11-05#

  • Triage sees every open item and what the owner's search finds, so a matter joins its task from any source
  • A row clicked lights Back and Next at once, a watched row included
  • Mark done on a watched row takes it off the rail at the press and keeps it off
  • The rail's age gutter splits the first half hour into < 5m, < 15m and < 30m
  • Triage no longer picks a worker for a task: every general task runs on the one assistant; routines still name theirs
  • A general task whose reply writer says NEEDS starts the assistant on it, as a reply task's does
  • A login the thread asks the owner for, kept in their own files, goes in the drafted reply and never in the conversation
  • A drafted reply's body stays out of the agent's pane, live trace and kept steps
  • A research task handed on by the reply writer's NEEDS keeps its kind, and counts as opening until its start lets go
  • A drafted reply's body stays off the CLI live line and an ACP call named by its kind, and a broken start lets its hold go
  • A Claude hook on the draft call names it by what it is: the worker's tool line and the approval text carry no body
  • A stranger session's permission card names the draft call, never its body
  • The general agent's one goal is to get the owner to done, starting from triage's reading
  • A stranger's task keeps the one goal but is never steered toward a login
  • The login rule lives only in the owner's own goal line, never in a document a stranger's task reads
  • Each agent turn ends with what it is about, what it did and what is on the owner
  • A summary block cut short never reaches the screen, and the owner's On you steps survive ticks, edits and empty summaries
  • The agent's chat says Working on this…, and ends each turn in three parts
  • An opened Working on this… stays open, a failure is never folded, and a pending Approve is not Nothing left
  • Profiles are for routines only: the roster and the "Can be given new work" switch go
  • The assistant-tools page is the catalogue as it stands
  • GOAL_LINE is defined once, and the danger check asks the verdict schema without a profile
  • The general agent's goal is a principle: take the work off the owner's hands
  • A PDF clicked in Came from or in the agent's answer is read in place, Download beside it
  • An agent's browser starts at its first navigate, never on an empty tab, and a hand-over with no page is a question in words
  • Only a browser on a page takes room in the agent's canvas: an empty tab is no view and no chip
  • An agent fills and signs a form with the signature the owner saved once
  • Settings has a Your signature card: draw it or upload it once, see it, clear it
  • A task whose home is the chat shows Came from and the agent, and opens on the chat
  • A live general agent's strip carries Mark done where the chat is home, and a coding task never flashes two sections
  • A browser test: an agent started on the open task says Starting… within 1.5 s
  • A start on the open task says Starting… at once, and a queued one Waiting its turn
  • A correction typed in the chat is learned the way the draft box's words were
  • Drafts read how the owner works, inside the same budget
  • The Advisor reads the agent's last summary, so Nothing left to do lights Mark done
  • The agent's process stays warm while its task is on screen (cold 7 s, warm 3 s)
  • Any change on the thread wakes the task's assistant, which re-plans the whole next step
  • The warm-keeping heartbeat route is the owner's: a session never keeps its own process alive
  • A reply draft is never left empty without saying so, and a rewrite says whether it changed
  • Close out shows a draft filed under the open card, says an empty one plainly, and the Assistant says Rewritten only when it changed
  • An FYI or a report says what it says and what matters to you
  • A question on a report, FYI or dashboard is answered from it - no task, no agent
  • The Assistant says an act is done only when the store shows it: every receipt from the result, never the plan
  • Draft cards, emails owed and On you steps live in the agent's summary; Close out is gone on chat tasks
  • A redraft that says NO REPLY marks the waiting draft Not needed, with the writer's WHY, instead of leaving it ready to send
  • Run now's notice sits under the Reports toolbar, not squeezed beside its buttons
  • Removing the WhatsApp card unpairs its bridge, and the open card stops scrolling to its top
  • CI's Python suite is green again: stale stand-ins, a leaked session, and two sentences that typed a place
  • Fix WhatsApp wrapped messages and first LID delivery (#273)
  • A pull-request row is shown to triage by what it holds and who sent it, so a chat line about that PR joins it
  • The agent-button tests open Send to agent's box only where it is not already up, and the walk says why Draft reply goes unpressed on chat tasks
  • A fresh question on a task its agent already worked reaches that agent's chat, and the demo lets an open task keep its agent warm
  • The Assistant says an agent it handed work to is on it, because the hand-off now says the agent started
  • A chat task's Agent section is one chat: the ask, each turn, a line for each new mail, and the box always under it
  • Came from no longer repeats triage's reading as "What's going on", and the source-text tests follow the one-chat pane

0.3.10.29 - 2026-11-04#

  • Triage sees every open item and what the owner's search finds, so a matter joins its task from any source
  • A row clicked lights Back and Next at once, a watched row included
  • Mark done on a watched row takes it off the rail at the press and keeps it off
  • The rail's age gutter splits the first half hour into < 5m, < 15m and < 30m
  • Triage no longer picks a worker for a task: every general task runs on the one assistant; routines still name theirs
  • A general task whose reply writer says NEEDS starts the assistant on it, as a reply task's does
  • A login the thread asks the owner for, kept in their own files, goes in the drafted reply and never in the conversation
  • A drafted reply's body stays out of the agent's pane, live trace and kept steps
  • A research task handed on by the reply writer's NEEDS keeps its kind, and counts as opening until its start lets go
  • A drafted reply's body stays off the CLI live line and an ACP call named by its kind, and a broken start lets its hold go
  • A Claude hook on the draft call names it by what it is: the worker's tool line and the approval text carry no body
  • A stranger session's permission card names the draft call, never its body
  • The general agent's one goal is to get the owner to done, starting from triage's reading
  • A stranger's task keeps the one goal but is never steered toward a login
  • The login rule lives only in the owner's own goal line, never in a document a stranger's task reads
  • Each agent turn ends with what it is about, what it did and what is on the owner
  • A summary block cut short never reaches the screen, and the owner's On you steps survive ticks, edits and empty summaries
  • The agent's chat says Working on this…, and ends each turn in three parts
  • An opened Working on this… stays open, a failure is never folded, and a pending Approve is not Nothing left
  • Profiles are for routines only: the roster and the "Can be given new work" switch go
  • The assistant-tools page is the catalogue as it stands
  • GOAL_LINE is defined once, and the danger check asks the verdict schema without a profile

0.3.10.28 - 2026-11-04#

  • A helpdesk's answer on its own thread is shown the request it answers, so a watched ask and its ticket confirmation are one row
  • Settings drops "Your mail goes to": triage picks no worker for a task, so the row chose nothing
  • The role walk no longer asks where the owner's mail goes, and applying a role moves none of it
  • A file the agent hands over that cannot ride is kept on the reply, by name and why
  • Close out says which file did not attach and why, and Attach another names a refused file
  • The agent is told which file its reply does not carry, and what to offer instead
  • The admin approval link goes through the sign-in page, which takes the app's desktop redirect
  • Help page: why an old approval link said AADSTS900971, and the Entra admin center as a last resort

0.3.10.27 - 2026-11-04#

  • An agent run is never killed for running long, and a turn cut by a time limit says so and offers Continue
  • A draft keeps which mail it answers apart from the newest mail it saw
  • The agent's reply answers the mail that holds the ask, and can be pointed at another
  • A re-saved reply addressed from the mail that holds the ask, and a plain re-save that keeps what was pinned
  • The drafter answers the mail that holds the ask and leaves the owner one line on why
  • A draft with no thread reading keeps its From and Cc, and a WHY that wraps is taken off whole
  • A WHY is taken off only at the very end, set off by a blank line, so a Why heading in the reply stays
  • A thanks on the thread leaves the draft where it was, and Send goes to the mail that asked
  • A reply sent with a merge goes to the mail that asked, and an aside opens no draft and moves no unthreaded reply
  • A draft pinned to a thanks is not moved by the older ask it answers, and the store finds the newest line past the asides

0.3.10.26 - 2026-11-04#

  • The owner's exact words go under YOU; what the Assistant adds is Taskuary's, folded
  • A sign-in form that also asks for a third box says so, instead of calling the saved password wrong
  • The vault keeps the other boxes a sign-in form asks for, and fills them by their label
  • An approved click the page refuses tells the agent and ends the held act, instead of a raw refusal under Approve
  • A sign-in answered by Taskuary itself clears from the open task without a reload
  • Two poll tests follow the read made while the agent waits on the page; the new browser test joins the list

0.3.10.25 - 2026-11-03#

  • A run cut off while asking the owner reads as still asking, not stopped
  • Continue on a run that ended asking reads Answer it and shows the question
  • An agent a restart cut off while it asked the owner is back at its question
  • Triage is told who answered the like of an ask before, when colleagues share the mail
  • Came from shows whom a mail went to
  • A chat read never puts an item's old walk line after the table's own
  • A press on the grown chat pill's middle opens it
  • Draft reply spins until its draft lands, and always drafts
  • Triage gets a fourth verdict, watch: a thread the owner cares about that someone else will handle
  • An FYI batch reads the thread under every row, not only the first

0.3.10.24 - 2026-11-03#

  • Send's hover says what it does where it is: Close out's Send closes the task, a Send anywhere else leaves it open
  • An Advisor idea sent to an agent is briefed as the owner's ask, not a proposal for the agent to grade
  • Advisor job ideas come daily, never repeat what was raised, and an Advisor idea needs a high bar to become a task
  • Advisor job ideas check that a person they name still works there, and that a process they name still runs
  • Stage moves stay on the spine at every desk width; the bar is Back and Next alone
  • The chat pop-up shuts when the owner moves to another task, and keeps its conversation
  • No AI, no start: every road that starts an agent refuses when no AI can run it
  • An Advisor idea's task is as trusted as the least trusted thing it cites
  • Security page: an Advisor idea's task is trusted through what it cites
  • Stage moves stay on the spine from 601px up, and a squeezed bar's batch Next says All read
  • A report from the owner's own systems is trusted, and so is a memory line in their own words
  • A report or dashboard source card has a "my own system" switch
  • Security and Reports pages: a report is as trusted as what it read
  • The Assistant's history search says when it cut the list
  • A sweep of the FYI leaves the Advisor's ideas
  • A page the Assistant opens is named in its step without the model's note
  • A bare web address in a chat answer is a link
  • Docs: the Reports page links Security by its slug, and the docs site is rebuilt
  • One walk move at a time: Next and Back grey while a move is out, nothing is held
  • Next from the last task into FYI draws the batch at its press
  • A press the table moved off while it ran does not read the Advisor again
  • Browser and walk tests follow no AI, no start and the squeezed bar's All read

0.3.10.23 - 2026-11-03#

  • A double or triple click is one press, and never presses what the first click drew
  • The demo's refusal of Hand it to a person is said as the demo, not in red
  • A walk press is never lost: Back and Next pressed fast, over a move or a typed ask, are held, not dropped
  • A Next pressed just after a typed "next" is the second move, not lost behind it
  • An Advisor idea's Put it on my list and Not ours run again
  • An Advisor idea's Put it on my list keeps the task it made on the table, as a report's does
  • Back held over a Back that drew the top row goes up, instead of snapping back down

0.3.10.22 - 2026-11-02#

  • Send something written by the Assistant opens on the channel asked, to the chat named
  • FYI fills the rail before reports, and a scrolling rail draws a screenful of it
  • Next on the last row reads it and ends the walk, instead of standing greyed
  • Digest lists only meetings still ahead, and is named by when it was written
  • FYI shows a fixed 6 rows on a big window and 4 on a small one

0.3.10.21 - 2026-11-02#

  • The Assistant gives a file an agent made by its link, never a path on disk
  • A stopped agent's Result names its file, not the path into the outputs folder
  • A folded Agent pane says the file's name, not '**C:\Users...' with its bold marks
  • The Advisor looks out for the owner's job, not only the app
  • Every download is the installer, and a single-file install updates into the installed app

0.3.10.20 - 2026-11-02#

  • Windows installer: TaskuarySetup.exe, built and signed with every release
  • The mic offers Windows voice typing only while its speech setting is on, and says plainly when it stops
  • Desktop Download saves into the Downloads folder and says so
  • Waiting room never types a note into the owner's half-typed line
  • Words typed to a general agent waiting on you reach it, and a send that fails stays on screen
  • Rows filed as Not ours can be marked read instead
  • The Assistant knows the agent job it started; a second press of it is one start
  • WhatsApp chats keep their names across a bridge restart, and a timed-out group roster is asked again
  • A conflict elsewhere in the store no longer refuses every mark-read
  • The WhatsApp card's switched-on chats show their last message under the name
  • Emails use the account's own signature when the owner has set none
  • Ship only the sent-mail signature: no Gmail settings permission yet
  • Regenerate the Assistant's tool reference for the Not ours wording

0.3.10.19 - 2026-11-02#

  • A log file that cannot open in the data folder no longer stops the desktop app
  • A link in the Assistant's reply to the app's own New or a message stays in the desktop window

0.3.10.18 - 2026-11-02#

  • A right-click on a misspelt word in the desktop window offers the fix again
  • A restarted WhatsApp bridge no longer loses what arrived before Taskuary's next poll
  • A message relayed to a working agent waits for it to stop instead of being typed into its turn
  • A memory line named by its own words is found, and a failed edit is told in plain words
  • A Codex agent mid-turn keeps its own queue for a relayed message; only a typed line waits in the waiting room
  • The Assistant's model says what went wrong after a failed look-up, never a sentence written in code
  • A task its agent closed stays in Tasks until it is read, not under Reports
  • A question put to someone else by name is theirs, even with the owner on the To line
  • Next takes a re-run report off the rail at the press, and no early rail puts it back
  • A one-name nudge after the owner answered is the named person's, and triage knows the owner spoke last
  • A task whose checklist owes an email nobody has drafted opens on the task, not an empty Close out
  • A closed task opened from Search opens reopened, without a press on Reopen task
  • Agents tick each to-do as they finish it, not all at the close
  • Close out's To is the owed email's own address, and the sender's line says it is the reply's
  • Back moves at its press, as Next does
  • An agent's reply goes to the person it names, and Send never rewrites it for someone else
  • Say --to in the agent's reply command itself, so the seed keeps its room for the message
  • An agent's draft is approved where the agent is: a pop-up over its terminal, a card under its chat

0.3.10.17 - 2026-11-02#

  • Watch it stays on the walk, means one thing on desk and phone, and a phone's small changes act at once
  • The AI doors leave the chat's view; a report's builder runs and pauses itself; a route names who gets it
  • Advisor: honest agreement numbers, a held action lit, the Assistant's context, a finish check on the latest word
  • The Assistant reaches the bell, Search and the Hub's Write one; a removed Hub post is under Recently deleted
  • The Assistant starts an agent when told to, and tells one at work

0.3.10.16 - 2026-11-02#

  • The task page's everyday asks join the audit, and three it refused or misled now land
  • The rail audit: 26 everyday asks about the work rail and its walk, and six it refused or misread now work
  • The everyday-asks audit covers the Connections page, and four things the Assistant got wrong there are fixed
  • A report's Run now and switch are on its row: the Assistant brings up the shelf, and reads each row by name
  • The Settings audit: 30 asks with Settings on screen, and four gaps closed
  • The Assistant on New: names the press each tab draws, sees every box, and fills a box a press draws
  • The Assistant works the Hub and the Board where they stand: Write one, a post by number, named menus
  • A phone asks yes or no only for what the row can do, and Watch it by phone is Watch it
  • A pip-installed or container Assistant reads the help pages and What's new too
  • The Assistant's self-knowledge audit: 24 asks, and the three gaps it found fixed
  • Regenerate the Assistant's tool page for the merged audits
  • What a real model wrote now lands: a task page's bare decision, a report by name, a CALL in tool markup
  • A report's miss still opens "No report by that name", and the exe check knows help is docs/site
  • A report act over the report on the table runs as it did: a row's handle is no report name

0.3.10.15 - 2026-11-01#

  • The everyday-asks audit covers the Assistant's newer powers, and three asks it refused now work

0.3.10.14 - 2026-11-01#

  • The assistant stops reading names and labels with word lists
  • The Assistant can read its own help without an embedder, and keeps what it found
  • The phone's menu words stay: with no AI, a choice's own label typed is that choice (the owner: button words are fine)
  • Say the sign-in's age and two counts through timewords, as the release gate asks

0.3.10.13 - 2026-11-01#

  • Take the mail words back off the mailbox card: which card a name means is the model's to read

0.3.10.12 - 2026-11-01#

  • A What's new page the Assistant reads, written by each release
  • The chat never answers a question with the story of the task under it

0.3.10.10 - 2026-11-01#

  • Next's card drawn at once stays mounted when its answer fills it
  • What came in and What triage made of it sit on white, as the mail itself reads
  • One Advisor row for all habits, each said plainly with its own answer

0.3.10.9 - 2026-11-01#

  • An Update is no longer held back by a second window onto the app
  • Personal is never learned from: a personal ask is an FYI, and nothing personal teaches or is raised
  • Readers get the item's workstream: the Advisor and agents read its own section, triage reads every one first
  • Memory hygiene: real people only, promises from sent words, bring-up from Done, lines expire, caps per workstream
  • The rail's task rows wear their workstream as a small label
  • Workstreams read from six months of mail: a setup confirm screen, a weekly refresh, and a new hat noticed
  • The Assistant drafts a bug report or a feature request when the owner asks
  • The Advisor raises "Report this?" when Taskuary itself keeps failing one way
  • Browser test: a wish, a bug and the Advisor's Report this? open the scrubbed draft
  • Signing out of the CLI in the Agent pane no longer ends the task's agent
  • The Advisor no longer calls a usage limit a broken report
  • A genuine sign-in link is not a scam: triage weighs the sender on facts, not on mail banners
  • Nothing closes a task automatically while work on it is still live
  • Killer ideas: kinds that earn their place, a second look before raising, sure habits adopted and unsure ones asked once
  • The second look asks join-or-restate, and a kind's record starts with it
  • The second look keeps what the owner asked to be brought up, promises, and a date that has passed
  • Yardsticks with more than one workstream, and a personal ask lights no work move
  • STYLE per workstream, as a measured option: a draft reads how the owner writes there, else the shared style
  • Unblock 0.3.10.9: the stranger-token stub takes sid, and the walk says why it skips the two report buttons

0.3.10.8 - 2026-11-01#

  • A live agent's earlier runs are one line under its terminal

0.3.10.7 - 2026-11-01#

  • Quiet unless sure: a rule's button goes dark below 80% on the owner's record
  • Trial: a task's open pane opens side by side with its partner
  • Workstreams: every task, message and idea says which body of work it is under
  • A task's workstream is the owner's word: PUT /api/tasks/{id}/workstream is owner-only
  • The day's opener puts the walk down and is the top of the chat

0.3.10.6 - 2026-11-01#

  • A section opened by hand stays open when the task's status catches up
  • An item opened again keeps the task view it has on the table

0.3.10.5 - 2026-11-01#

  • Agent Desktop: agents operate desktop apps through Vercel's agent-desktop, held like the browser
  • A report's moves ride its own pane; the bar keeps Back and Next
  • Agent Desktop runs only on a task marked "needs desktop", and the demo shows its card
  • An item's folded spines carry its stage moves, as a task's do
  • An Advisor idea comes in like everything else: in Came from, under Advisor
  • A close made through the walk wears its badge again, and All read, next keeps its own
  • Page: no press waits on itself, the fyi count holds between batches, a reopened item stays one line
  • A set that would not clear says why
  • Counts said with plural, and the cold-rail clock catches a held lock, not a busy desk
  • Close out stacks its rounds: the current one answers what reopened the task, earlier ones fold under it
  • The needs-desktop mark is the owner's press, and the Assistant's payload lists the Agent Desktop card
  • Next and Mark done draw the row below at the press again (rule 7)
  • A folded pane's stage move presses the live card, as the bar's does

0.3.10.4 - 2026-10-31#

  • railwalk: the walk has one owner - Next, Back, open and every close decided in one module
  • Page: the walk's moves go to railwalk and the page draws the answer
  • Deletion: the navigation reservation goes, New chat takes the walk's lock, the words follow the rules
  • Tests to the seven rules; opening a task reads it, Put it on my list keeps its place
  • Page: the walk beside the rail follows the phone; one close is one move; the clicked table fixed
  • Deletion: the last put_down and prepared_next remnants go
  • The day's opener fits a 13-inch laptop: two cards of one height, the digest under them to the fold
  • Page: every walk move is one POST to /api/walk/*, drawn by one function; the close's early draw goes
  • Server: the walk's last outside decisions go to railwalk; tests rewritten to the seven rules
  • Trial: the chat pill grows under the mouse and eases back after it leaves
  • railwalk: a walk move reads the rail it has, never builds it for its own writes
  • The hovered chat pill grows taller only, never wider
  • A report is never logged as failed because its own row would not settle

0.3.10.3 - 2026-10-30#

  • The evening report does not drop a connection for last session's error (#270)
  • A bug report's scrub ignores store strings with no letter or digit (#270)
  • A new walk on the server clears the rows the page left in the last one
  • An FYI batch put up is shown, not read: Back leaves it on the rail, Next reads it
  • A spine's words stop above its moves where they actually stand, and refit when they arrive

0.3.10.2 - 2026-10-30#

  • An agent's own --done lets go of a question its CLI is already past
  • The opener is your day and the tasks, then the digest's first lines with More

0.3.10.1 - 2026-10-30#

  • A walk move's dots stay on the canvas until the next item is in
  • Reply opens Close out on its box at the press, the spinner in the box while it drafts

Documentation for Taskuary 0.3.11.17 · updated 2026-11-08