DocSign Leaves Beta, the My Tasks Dashboard & Statuses by Type
DocSign leaves beta for everyone: every signed contract now carries the evidence a signature needs to hold up in the US and Brazil, with a rebuilt editor, a rebuilt signing page, an assistant that answers the signer's questions, and contracts generated from board rows. My Tasks gains a dashboard you build yourself and a source picker that browses every board. Status columns can offer different statuses for different kinds of work. And AI agents now keep following a board, an inbox or a conversation after their turn ends, take board rows as tasks, build automations for you, and ask before a browser sends anything.
New Features
DocSign — Out of Beta, for Everyone
DocSign is no longer a beta feature: it is on for every workspace, with the evidence, the screens and the intelligence a contract product needs (#4066).
- Evidence that holds up — every envelope freezes the original document the moment it is sent, with its own fingerprint, so signing never depends on the file in Drive again. Everything that happens afterwards — opened, consented, each field, signed, declined, voided, expired, reminded, completed, downloaded — is written to a log that can only be added to, with the server's time, the address it came from and how the signer proved who they were.
- Consent recorded properly — every party, approvers included, consents to signing electronically in the language they chose, and a stricter consumer-disclosure mode is available where the law asks for it.
- A signing link that is really yours — the emailed link is exchanged for a session tied to that one envelope and that one party, so two contracts or two signers in the same browser can never act for each other. An email code is asked by default for people outside your workspace.
- A certificate and audit pages on the final document — the completed PDF carries a certificate of completion and audit pages naming every party, the sender, each consent and the acts that make the record. Its fingerprint is shown on a public verification page, where anyone holding the file can check it matches.
- Sealed, or honestly unsealed — the final document is sealed with a timestamp, and a document that could not be sealed is never presented as if it were: the sender is told, and it is sealed as soon as that is possible again.
- Everyone gets their copy — the sealed document is mailed to every party after sealing, retried if a delivery fails, and an address that bounces is recorded as undeliverable. Parties keep their link, their copy and the verification page even after the sender hides the contract from their own list, and even after the workspace is closed.
- Retention you choose — signed contracts are kept for a number of years you set (five by default, five as the floor), in a vault that survives workspace deletion, a Drive purge and a member leaving. Existing contracts never get a shorter date than they already had.
- Word files become PDFs properly — a Word document is converted with real fonts, and tracked changes, comments or substituted fonts are reported to you as warnings. When conversion is unavailable the wizard says so in a sentence and falls back to printing from Office.
- A rebuilt editor — field chips in each signer's colour, a recipient-first palette, drag and resize with snapping and guides, undo and redo, zoom, a page rail, rotated pages, nine field types and auto-placement suggestions. A field you place just before leaving is saved.
- A rebuilt signing page — one scroll with a bottom bar on phones, a guided reading order, fields that open on a click or a tap and behave as their type, a signature you type, draw, upload or reuse, decline, download, print and language controls, saved progress, and an end state for every outcome. It follows your system's light or dark preference.
- An assistant for the signer — the signer can ask questions about the contract in writing or by voice, and get answers with citations into the document, from the document and the sender's own private notes. It never sees anything else in your workspace, never quotes the sender's notes back, and never affects signing: questions can be limited, signing cannot.
- Agents can prepare, never sign — an agent can prepare an envelope or draft a contract, and both come back as an approval card naming the title, parties, message and field count. Sending always asks. No agent can sign anything.
- Contracts from board rows — a board automation sends a contract from a template, mapping board columns to the contract's variables and roles. Money uses the board column's currency, a value that does not fit is refused rather than silently cut, one contract is sent per row and template version, and the outcome is written back to the row. Sent, signed, declined and expired each start automations of their own.
DocSign — Sender and Signer Screens Rebuilt to the Approved Design
On top of leaving beta, both sides of DocSign were rebuilt to the approved design (#4086).
- A draft opens straight into the editor — no second header above it: the editor is the page, with Delete and Save as template in its menu and any warning as a banner under the header.
- Calm recipient rows — one row on one grid with number, colour, name and email on a line, a quiet role control and a menu for language, email code and remove. The add row takes a pasted list and puts the cursor back on the name.
- Review and send, as a dialog — a centred dialog with its own scroll and a sticky footer, full screen on a phone, listing recipients with their role and field count.
- It tells you what is in the way — instead of a greyed-out button with no reason, pressing Review and send opens a short "Before you send" note when a signer has no fields or nobody can sign, with the fix one click away.
- Preview as the signer — open the real signing page as any party, from the editor or the send dialog, and walk it exactly as they will.
- A sent contract opens in the same shell — read-only, with the state as a quiet pill, one primary action, the document on the same canvas and a panel of recipients, details and activity that becomes a slide-over on narrow windows. Everything the old page could do is still there.
- A clearer list — the third box is now Finished, and each card in it says how it ended (completed, declined, voided, expired) with the date. A party's run of filled fields reads as one line in the activity log.
- Pages turn from the bar — previous and next page buttons, plus PageUp and PageDown while the canvas has focus.
My Tasks — A Dashboard You Build Yourself
The stats header on My Tasks is replaced by a dashboard that is yours to arrange (#4093).
- Five kinds of card on a grid you drag, resize and reorder: breakdowns, numbers, donuts, task lists and a timeline. Up to twelve, saved per person, reset whenever you like.
- Picks that narrow the list — click a bar, a slice, a day or a due bucket and the task list below follows it, shown as one removable chip. The dashboard filter and the card you picked combine, each on its own.
- A quiet top row — Filter with a chip per value, Customize, and a fold control. Folded, it asks for nothing and still remembers your filter.
- It says when done tasks are hidden — if your filter asks for done tasks while Hide done is on, the list says so and offers one action to show them.
- A project column of your own choosing — table owners can name which column My Tasks reads as the project, picked from Dropdown, Labels and linked-row columns, kept apart from the Status and Priority slots.
- A start page preference — Preferences gains a Start page section: carry on where you left off, or open My Tasks.
My Tasks — Browse Every Board and Pick What Shows
The Sources panel was one flat, unsorted list of checkboxes with no search (#4032, #4038).
- A two-pane dialog — every board on the left with a box for the whole board, the chosen board's tables on the right, and a switch between "Everything" and "Only what I pick".
- Search across boards and tables, with one action to select every match.
- It names the column that blocks a table — a table My Tasks cannot read used to say only "a column My Tasks needs is hidden from you". Now each row names the column, the role it plays, and whether you can lift the restriction yourself or need to ask a board admin. Where the column name itself is sensitive, the role is named instead.
- Tables sorted by board name, then table order, with a link to open any table.
Boards — Statuses by Type
A Status column can now offer different statuses depending on the kind of work a row is (#4093).
- Turn it on per Status column — pick the Dropdown column that says what type a row is, then give each type the statuses it uses, with a tab per type plus All statuses.
- Every picker follows it — the list, the grid, the touch popover, bulk edit, the add-row form and the row dialog each offer only the statuses the row's type uses, keeping the row's current one.
- Kanban respects it — dragging a card onto a lane its type does not use is refused like a blocked workflow move, with a plain reason. Lanes no selected type uses are hidden; lanes holding cards always stay.
- Clear refusals — a refused status says it is not a status of that type, naming both, and after a type change a toast says the status moved to the first one that type allows.
- A status that exists elsewhere — searching for a status another type uses no longer creates a duplicate: the picker says it exists for other types and where to add yours.
- Automations know the types — status options show their type labels in the builder, a publish warning fires when a status does not apply to every type, and a run step that was skipped for that reason says so plainly instead of reading as an error.
- Duplicating a column carries it, and both the public interface and the agent tools can read and set it.
AI Agents — They Keep Working After the Conversation Ends
An agent used to go quiet the moment its turn ended. Now it can follow things, be handed work, and pick it up again later.
- Watches (#4023, #4024, #4028, #4027, #4030) — an agent can follow a conversation, a board row, a whole table, a whole board, a customer line's new conversations, an email thread or a shared inbox, each with its own instructions. It is woken when something changes, and stops when the job is done, when you tell it to, when a person answers the customer themselves, or when its limits run out. Its own actions never wake it.
- A Watching panel — a card beside the chat listing what the agent follows, what needs a person and what ended recently, with a Stop on each. What you stopped stays stopped, whatever a slow answer says afterwards.
- A watch keeps your words — the wake quotes what you wrote when you asked for it, word for word, and says your words decide. A board, table or row wake says who made the change and what the cells held before and after.
- A board row assigned to an agent becomes its task (#4031) — put an agent in a row's people cell and it reads the whole row, does the work, puts the result back on the row, sets a status, comments what it did and reports in its channel. A comment on the row wakes it again; unassigning it ends that. Only the member the agent works for, and the people in its channel, can hand it work — for anyone else the agent asks in its channel first.
- Agents show up wherever people do — a row's people pickers gain an Agents group, and an assigned agent is shown by name and avatar in the grid, row drawer, kanban, list, calendar, timeline, Gantt, workload, to-do, linked rows, lookups and filters. Sharing, approvers, voting and mentions are untouched: agents are a separate list.
- A turn survives a server restart (#4050) — a deploy used to cut a long turn in half and run it again from the beginning, doing twice what it had already done. A turn now keeps a record of what it has done as it goes, hands itself over at shutdown, and the next server carries on from the last finished step and posts one answer.
- It watches a shared inbox (#4054) — each new mail a person writes to a watched inbox wakes the agent once and may be answered in that thread, from that inbox, to the people already on it. Automatic mail — auto-replies, list mail, delivery reports, no-reply senders — is left alone. A thread the agent had finished with is picked up again when a person writes in it.
- It answers a mention in a team channel (#4057) — the agent's creator can add it to a text channel, where a mention of it by name wakes a turn that reads the recent messages, answers the person who mentioned it, and reads nothing it should not. Broadcasts, team mentions and forwarded messages do not wake it.
- It sets its own follow-up (#4091) — asked to "follow up with him tomorrow morning", the agent can now schedule itself: once, daily, on weekdays, weekly or every so many minutes, on the same routine machinery your Routines settings already show. It tells you the exact day, time and zone it resolved, and refuses a time that has passed or is more than 90 days out.
- Routines run on the new runtime (#4017) — the Routines settings page has the form to add and change a routine again, instead of sending you to ask the agent in its channel.
AI Agents — Files, Pictures, Notes and Knowledge
- It opens what you send it (#4019) — a document comes back as text, part by part or searched; an image as a description answering what the agent needs to know; a recording as its transcript. The conversation line now says what was sent, so it can be opened.
- It makes pictures and posts them in the conversation as its own.
- It keeps notes between conversations — the notes your agent's settings page already lists and lets you edit. A note only reaches a conversation where it tells nobody something they were not there for.
- It moves files where the work is (#4029) — a file from the conversation, a picture it made, a Drive file or a row's file can be put on a row, saved to Drive, or sent as an email attachment. It can read a file sitting in Drive or on a row too.
- It answers from the knowledge your team gave it (#4021) — the facts, texts, documents and pages under Settings → Brain → Knowledge never reached the new runtime, so the agent answered business questions from whatever the model happened to know. It now looks them up, and a conversation a customer or a guest reads searches only what customers may be told.
- It builds automations for you (#4035) — from a plain request the agent drafts an automation with the builder's own AI, checks it, saves it and turns it on, with your access deciding what it can make. A guessed draft or one with blocking issues is never saved, and an automation drafted after the agent read something from outside your team is left as a draft for you to publish.
- Ultra-fast mode for a native agent (#4069) — a Speed card on the agent's Brain page (Standard or Ultra-fast) and a chip in its channel header with one-click off.
AI Agents — The Browser Asks Before It Sends
The agent's browser can now finish a job that ends in a Send, instead of handing it back to you (#4062, #4025, #4026).
- A card before the send — at a step that sends, posts or buys, the agent raises a card naming the page, the action and the exact text, and waits. The text is shown only to the person who asked.
- It presses in the same tab — an approved Send is pressed in the tab the run paused in, with the approved text, instead of opening a new tab and typing everything again.
- One "Send now" is one send (#4085) — a message you already sent by hand is never sent again, an approval covers only the message box its card named, and a press that did not reach the page is reported as unconfirmed rather than as sent.
- After one allowed send on a site, typing the next message there does not ask (#4088) — a request for three messages to three people used to raise five to seven cards. Every Send still asks; the typing in between no longer does.
- Allow a site once and for all — your always-allow rule for sending or posting on a site also covers a guarded run, so later sends there do not ask, and a rule for opening an address can be removed.
- Searching works (#4082) — searching Google, Google Maps or Bing is reading, and never raises a card or a false stop. The words that make a field a search field are recognised in all six of the product's languages.
- The Browser page is back in a native agent's settings: your connected browser, your site lists, the agent's Browser level and limits, and the history of its runs.
- A failed run says why (#4041) — a run whose AI call failed used to tell nobody the cause: the agent said "failed and gave no result" and the page said "Something went wrong". It now names the kind of failure, and says plainly that it is not your Chrome.
Omni — Shifts, and Agents That Own Their Conversations
- Shifts (#4056) — different people answer an inbox at different times of day, and an AI agent can take a window of its own. Shifts are laid out as compact blocks, whole-day shifts included, and an AI shift requires an agent that can actually take turns.
- A shift picks up what is waiting (#4067) — when a shift starts, conversations with a customer waiting move to it, in one pass per shift start, and one conversation that fails does not hold the rest back.
- A native agent as first responder (#4089) — it answers inbound customers and never leaves one waiting: a turn that failed, ran out of budget, could not start or ended with nothing to say hands the customer to a person, once. A burst of messages gets one answer. A paused or deleted agent, and the inbox's AI switch, stop a turn that is already working, not only the next message.
- A person who answered keeps the conversation — someone who typed a reply during an AI shift keeps that conversation for the shift, and whoever replies to a conversation nobody owns becomes its owner.
- An agent that reaches out keeps what it started (#4090) — one inbox can have a first responder and other agents that reach out to customers through it. When a customer answers, the agent that reached out continues that conversation, at any hour. A thread the first responder is in the middle of is refused with words that say who is talking to the customer.
- The inbox says which agent has a conversation (#4094) — the thread header names the agent ("Nina is answering", "Nina is paused", "Nina can't answer now"), and "Hand back to ..." names the agent it actually goes to.
- A hand-over says why (#4095) — the agent gives a short reason in its own words, written as one internal note with the hand-over, for your team only, in the inbox owner's language.
- A hand-back the agent cannot take goes back to a person — a conversation handed back to a paused, removed or out-of-budget agent used to sit with nobody while the customer waited.
- The agent writes to customers as the business (#4097) — asked whether it is a person or a machine, it never says or suggests a person is writing: it says what your instructions say, or that it is your business's virtual assistant, and offers a person. Unprompted, it introduces itself as nothing at all. What a customer asks of your team — a call, a visit, a discount, a time — is passed on, not agreed to on your behalf.
Omni — Mark Unread, Voice Notes and Line Breaks
- Mark a conversation unread from the header or the list (#4051).
- Transcribe an inbound voice note on demand, and the AI agent reuses the transcript rather than paying for it twice (#4051, #4053).
- Message line breaks are preserved, and stray artifacts no longer appear in message text (#4051).
Contacts — Field Kinds, Renaming and List Options
Contact field settings were rebuilt (#4080).
- Field kinds and inline editing — set what kind of value a field holds and edit it in place.
- Rename or hide the default fields — the fields Copera ships with are no longer fixed.
- List options with a delete that asks — removing a choice tells you first, and the values already on contacts are cleaned up properly.
- Drafts survive — a choice field's draft options are kept when focus moves, and a finished create never clears a newer draft.
Drive — Folders First, in a Familiar Order
- Folders come first inside a disk, in both list and grid, with a natural name order that reads numbers as numbers — and they stay first in the sorts you choose yourself (#4071).
- Your own last open is shown on file cards, and listings still work when the recents lookup fails.
Computers — Hiring in Three Steps, With Our Own Size Names
Hiring a computer agent is now Agent, Computer, Review, with a summary and one final button (#4087).
- Our own size names — Standard, Pro and Ultra instead of supplier tier codes, and modes that read Desktop and Terminal only. Invoices and the billing portal use the same names.
- A clear wait — the computer page shows a setup timeline with the up-to-two-hours notice, and a brand-new agent waits with its composer closed until its computer is ready.
- You are told when it is ready — a computer bought on its own sends its buyer a notification and a push when it is first ready, and another if its setup gives up so you can try again.
- Honest side paths — a size that lost its price says so, a declined card keeps its reason, a lost purchase recovers, a retired size reads as retired rather than blaming the catalog, and members who cannot buy computers are told who can.
- Updates for older software — where you are allowed to, an update action appears in agent settings for a computer running older software, with guidance in your language.
Account Security — You Are Told When Something Changes
- An email to the account owner when a password, authenticator, passkey, recovery code or email address changes (#4077).
- Two-factor recovery codes (#4073) — codes you can save when you set up an authenticator, and sign in with if you lose it.
- Sessions end on a password change (#4070) — changing your password signs out the other sessions, and open documents reconnect on their own.
- Proof before changing a sign-in method — adding a passkey, enrolling an authenticator, adding an address or setting a password asks you to confirm it is you, in one shared dialog that keeps your place if you cancel.
- The second factor is asked on every road (#4076) — a password reset, a Google or Apple sign-in and the email link now all ask for it.
- Google or Apple sign-in takes over an unconfirmed password account cleanly (#4074), and the account owner is told it happened.
- Administrators cannot reset the workspace owner's two-factor (#4051).
Mentions — Lists That Load as You Scroll
Read and unread mention lists now load a page at a time as you scroll, with the counts for badges and tabs coming from the server instead of being guessed from a capped list.
Bug Fixes
- A deleted agent locked its computer — the agent's own records were removed in the same breath as the request to clear its files, so the machine could never confirm the clearing: its place was held for good and every later action on that computer was refused. The place is now freed properly, a machine that was offline at the time is retried, and a reinstall, mode change or deletion asked right afterwards is accepted instead of being refused for half an hour (#4048, #4055).
- A computer holding a deleted agent could not be reinstalled, changed or deleted — the page counted the agents it could list against the places in use, found none against one, and disabled everything with a tooltip that always claimed the machine was not ready. Deleted agents still holding a place are now listed with their own wording, each action says why it is unavailable, and the dialogs say what happens to that place (#4048).
- Computer agents stopped working in workspaces that were never enrolled in Workforce — adding one was refused with no reason, and existing ones could not be woken or resumed at all. A setting nobody ever set now allows, and only one deliberately turned off refuses, with a message in your language (#4063).
- "Add an agent" failed with a generic error — the list called a computer selectable and the request then refused it without saying why. One rule decides for both now, and each refusal says what to do: owned by someone else while reserved for you, agent software starting, not responding, busy, or records that need a repair. Every refusal of an agent's creation — a name already taken, a role that may not create agents, a budget above your role's cap, a guest among the people — is said in your own language instead of in English or not at all (#4075).
- An agent computer with no agent read as stuck — right after a reinstall the page said the console was unavailable and setup might be incomplete. It now says "This computer has no agent yet" and offers to add one (#4058).
- Renaming an agent failed on the avatar the settings page never sent (#4064).
- Reopening a channel landed on an old message, with a gap — a channel first opened on an unread message kept going back to it, and anything that arrived meanwhile appeared after a hole in the history. Reopening now lands on the present, and a stored place positions one open, once (#4060).
- "Invalid data" about eight seconds after sending a chat message — the list asked for newer messages starting from the message still being sent. It now asks from the newest confirmed one, in the main chat, threads and meeting chat alike (#4059).
- An agent in a team channel could reach its own private channel — a turn a mention woke could search chat messages and read threads under the agent's own channel. Those roads are closed, only the agent's creator can add it to a channel on every road, and such a turn is offered neither the watch tools nor a view of what the agent watches (#4061).
- A customer's turn held what your team had said in private — a turn a customer starts was given the agent's channel like any other, so the team's private messages sat in the prompt and a teammate's waiting request was read, and marked read, by it. Such a turn now answers its own notice alone (#4020).
- Browser runs failed in production before their first step — the model request carried a setting no host of the chosen models accepts, and every run was refused. Automation authoring had the same mistake and got the same fix. A router call now drops the settings its hosts refuse rather than being rejected whole (#4039, #4040).
- Every Ask Copera call on the documents and files lane was rejected — one tool's file-name rule was written in a form the provider could not read, so the whole call failed at step zero. Every tool's patterns are now sent in a form every provider accepts (#4010).
- An answer that had already been served could still fail — the check that reads over a generated document hit its own rate limit, failed closed, and the finished answer was thrown away. That check now tries a second reviewer, and an answer is kept when no verdict could be reached at all (#4010).
- Email domains stuck pending with every record green — the email provider stops looking for the signing records after 72 hours and never checks again, so a domain whose records went up late never activated and the every-minute check returned silently. Verification is now restarted once every record resolves, and a failed restart is retried in five minutes instead of an hour (#4046, #4047).
- A domain's verified marks could land on the wrong record — two checks running at once could write a tick against a record that never resolved. Marks are now matched to the record itself (#4047).
- Signing: an adopt in progress locked the signature sheet and was dropped if the field changed underneath it (#4092).
- The signing page title was unreadable in dark theme, and the sign-up button lost its inverse colours (#4053).
- Rich text in row lookup widgets showed as raw text, chained rich-text lookups included (#4051).
- Typing after a URL kept extending the link in the editor (#4051).
- Moving an all-day event, password values in table exports, and zipped folder names — passwords are masked in exports, zip entry names are sanitised in folder downloads, and an automation's run history explains an assignee it cannot show you (#4051).
- SLA elapsed time can be copied into a Duration column by an automation, and the run step payload label sits on its own line (#4051, #4053).
- Row activity errors were silent — they now show, with a retry (#4051).
- A voice channel kept showing a stale loudest speaker (#4051).
- Meeting AI actions could not be opened from every surface (#4051).
- A login loop after a failed start-up — a failed boot sent you between the sign-in page and the signed-in app forever (#4052).
- An automation's message to an agent arrived with its tokens unresolved — a "notify the agent" step handed over text with the row placeholders still spelled out in it instead of filled in, so the agent worked on the wrong row. Tokens are now filled in before the message is used, and the agent is told which board, table and row the automation ran for (#4034).
- An automation's private message never reached the agent — when the step may not show its text in the channel, the agent was woken with a note and no instruction. It now reads the message, and a turn configured by someone who cannot read the agent's channel keeps a narrow reach (#4033).
- A withheld report claimed nothing happened — a report replaced for privacy now says the agent did work on it and that the details stay private, and the line no longer names the automation, whose name can carry a client's (#4036).
- An agent reported in the wrong language — it used the row's language instead of the language you last wrote to it in (#4036).
- An agent whose user was deactivated could still be assigned rows, listed and woken (#4036, #4031).
- A teammate who mentioned an agent was told to "talk to it in its own channel" — a channel they cannot open. They are now told who the agent works for and to ask that person to add them (#4033).
- New My Tasks source reads raised an error toast every minute while the web ran ahead of the server. The page recovers on its own now (#4037).
- My Tasks due dates — a legacy text due date is counted in the stats and the dashboard, and editing a row no longer changes how its due date is stored (#4093).
- Private computers appeared in administrators' channel lists that they were not shared into, while administrators keep their view of the workspace's inventory (#4065).
- Phantom errors from row saves and inbox labels — a batch of row saves cut short by its deadline, and an inbox label failure with no reason attached, both reported properly now (#4065).
- Design and manifest errors — oversized sources, invalid manifest paths and oversized push payloads are reported as what they are (#4052).
Improvements
- A browser job for several people finishes the job — the run used to end after the first press, and the agent then claimed all three messages had gone. A press is confirmed by the browser's own read, the rest of the task goes on, and nothing is ever sent twice (#4085).
- The agent is honest about what a send did — new outcomes with words of their own: typed but not sent, pressed but the page could not be read, sent once and the rest not done, and "not confirmed" when it genuinely cannot tell. It never says a message was sent when it was not (#4085, #4078).
- One browse at a time — two browser calls in one step used to refuse each other as "the browser is busy"; the second now waits and runs with what the first read (#4082).
- A failed web search is reported, not retried blindly — a refusal is told apart from a provider fault, a busy workspace gets words of its own, and you are no longer told your plan refused a search it did not (#4082).
- An agent's credential is made once — requests arriving at the same moment each created one, so an access switch you flipped right then could be written to the copy nobody reads again and an area you turned off stayed on (#4083).
- After a card's answer, the agent goes on with that request only — it used to take up unrelated older requests from the conversation by itself (#4078).
- Agents say when an automation is only a draft — a line from Copera under the agent's answer, in your language, with the link that opens it. A model saying "Set up" for something that was never turned on is no longer the last word (#4057).
- Agent habits corrected from real runs (#4022, #4029) — an empty calendar is an answer, not a failure; it does not volunteer what it has not re-checked; it does not ask about a choice it was already told how to make; it does not promise to tell you later when nothing can wake it; it looks on your boards before saying a person cannot be found; and it asks for everything missing at once.
- Faster hand-back checks — a browser run asking every two seconds whether someone wrote now costs one quick check instead of reading twenty messages each time (#4049).
- Messages reach a running browser job — typing while the browser works ends the run where it is, reports what it found, and the turn carries on from your message (#4049).
- A turn's "working" line outlasts a long step — it is refreshed every 45 seconds, so the row and its Stop button no longer disappear after two minutes (#4049).
- DocSign review and send reads the document again before it opens, so it no longer says recipients have no fields right after the editor saved them (#4086).
- An email change saved while another write is running waits instead of vanishing with the dialog still open and nothing said (#4086).
- A recipient with no address yet is never a duplicate of another — template roles that had no address yet all compared as the same empty address, so with three of them every save answered "This email has already been added" (#4086).
- DocSign evidence is offered only where it can be built — the download appears for completed contracts only, instead of failing in every other state (#4086).
- Nothing in the DocSign editor starts a text selection — a drag beside the paper used to highlight the sidebar and the header (#4086).
- Resize handles are exactly 20 px and never overlap their neighbours on a small field, and a field let go out of reach is brought back into view (#4086).
- A field's own controls never place the armed field — with a tool kept armed, pressing Required, Duplicate or a resize handle put one more field down and moved the selection from under the button (#4086).
- My Tasks dashboard cards never clip — a card body taller than its card scrolls inside it, bars, donut legends, lists and timeline labels included, and the Add card tile always comes after every card (#4093).
- Fewer phantom errors, again — another pass made expected situations stop reading as failures: a provider outage, a job that will retry itself, a worker shutting down mid-task, a brief database blip. A reported failure is now logged once, whoever is handed it (#4041, #4052, #4065).
- Status moves are free where the feature is off — the detection that notices a status moved by a type change costs nothing on a table without statuses by type, and a large paste stays linear (#4093).
- Computers' terminals work properly — engine tools and the programs you install are back on the path in a terminal, outbound requests from a terminal follow the computer's own rules, and the console desktop runs as the agent's own user with its home, tools and network. The browser was also given room for far more work at once.
- Code Review explains itself — a round that met the required score but was held back says which rule held it, findings get an acknowledgement that promises a check on the next review, replies are kept after the per-finding limit, and members of a private organisation with write access are recognised as collaborators rather than outsiders.
