Atlas Changelog

About this changelog

This page lists product changes a customer or operator could actually notice — new features, UX changes, fixes, and access or permission changes — not every commit or internal refactor. Legacy date groups have been assigned sequential v0.x.0 versions, beginning with v0.1.0; each heading retains its release date.

Scope. Atlas is pre-release and iterating quickly on staging before each change is promoted to main. This changelog only records changes once they reach main; staging-only work is not listed here yet.

November 2026

v0.55.0 — November 3, 2026

• Home leads your day. Atlas opens on a Home page at /today with everything that matters in one scroll — today's schedule, your open tasks, recent activity, this month's metrics and the latest completed work. The Work panel now shows only Home and Tasks (for managers too), Workbench is renamed Tasks and lives at /today/tasks with old links redirected, and Business overview, Sales, Accounts receivable, My performance and Team overview leave the navigation for now — their pages stay where they were.
• Tasks opens fast. The Tasks page now reads reply details for the whole list together instead of one row at a time, stops asking the server for the same lists several times while opening, and loads the Inbox sorting and linking counts only after the task rows are in. Each task sits on one line in its section — Require attention, Replies, Reach out, Scheduled and Tidy up — and one click opens it, with Atlas's draft already prepared before you arrive.
• The Inbox is for reading. Sort Filters now sit under Deal stages and Non-customer emails, with Other on its own below them. All inboxes and every connected mailbox are listed at the bottom of the rail with their unread counts, replacing the mailbox dropdown. The mail list is one scrolling list — no pages, no select-all; tick boxes and stars appear on hover, and Cmd or Ctrl-click selects several. In the reader, the star sits on each email beside Reply, Archive, Spam and Delete stay at the top for the whole conversation, quoted mail folds behind a small ••• button, and every folder and filter still shows its count.
• Sorting is decided by twelve email types. New filters cover After-sales, Supplier, Promotions and Job application, with wider Inquiry and Logistics coverage; a short reply takes its thread's topic; a Notices topic holds sign-in codes, security alerts and account messages. To link keeps customer-related email: confident Supplier, Promotions, Job application and Notices mail from unknown senders stays in the Inbox instead. Mail Atlas cannot place waits in To sort with its reply task kept, and only a person chooses Other.
• Mail connects, files and reads more reliably. IMAP setup works with 263 and QQ expectations and imports existing messages when a provider omits optional mailbox metadata; new emails send from the selected mailbox without being blocked by an unrelated task lookup. A connected mailbox can be expanded to all Inbox history without re-entering credentials, slow mailboxes get up to two minutes per import session, and personal mailboxes from another domain work too. Email sent from Atlas also files into the mailbox's own Sent folder (English and Chinese folder names recognised), a customer's reply arrives on the next sync instead of waiting behind a long history import, and members can connect a private mailbox the team cannot read. HTML email keeps its tables, emphasis and spacing in a safe reading view, HTML-only mail shows as readable text, and placeholder-only names fall back to their addresses. Mailbox switching, list paging and sending are noticeably snappier.
• The new web client, atlas.v2.forth.ai, reaches v1 capability. The same screens, styles and translations as the desktop app, under existing accounts and permissions, with the corrected CRM hierarchy — Home, Contacts, Companies, Deals and Tasks. The five core pages are complete: Inbox (compose window, folders, topic filters and the sort/link queues with counts), the customer profile (figures header, AI-extraction review, labels and conversation logging), the Contacts directory (whole-directory sorting and New contact), Deals (board totals, list view and a full deal workspace with stage history, milestones, quotations and documents) and Knowledge (My/Company/Customer scopes, upload and the paginated reader). The v1 web client stays independent.
• Assistants set themselves up and talk. Name an assistant and choose an avatar; Atlas infers its job or asks one short question, confirms creation in a single action, and steers its skills at task time under existing controls. A created assistant's conversation is live on web v2 — send, replies and history replace the unavailable notice, PDFs are read through the governed pipeline, and the composer offers browser dictation where the engine exists.
• Deal and quotation writes are safer. Editing a customer, contact or deal that another member changed since your read is refused with a side-by-side comparison of the server record and your draft, and applies only after you review and save again. Deal and quotation creation is replay-safe — a timed-out or restarted submit reads back the saved record instead of creating a duplicate — and replaying a lost create request as a colleague who cannot see the original refuses without exposing any record data.
• Your workspace keeps its shape. Personal saved views remember each member's search, filters, sort, columns and board/table mode across sessions; Home shows your own open work grouped by due day, and completing, rescheduling and reopening keep one canonical task everywhere; the sales results report explains pipeline, closed outcomes and overdue follow-ups with clickable record lists that reconcile to the counted totals; and a cited knowledge answer can be saved as a durable evidence reference that re-opens under your current access.
• A new app icon and steadier screens. The new brand mark on a white background serves as the browser favicon and home-screen icon on both web clients. Customer creation keeps its pinned header over scrolling content, WhatsApp setup shows its channel guidance after saving, and the 390px phone layouts keep Home, Tasks and the Inbox usable with a single-pane reader.

October 2026

v0.54.0 — October 24, 2026

October 25 follow-up to this assigned release. Its original October 24 release date is retained.

• October 25 follow-up: customer recycling keeps the latest qualifying contact when duplicate records merge, so an older timestamp cannot hide a recent call. Warning delivery, grace periods and protection remain required; the Public customer guide shows manual release, claim and assignment with English and Chinese screenshots.
• October 25 follow-up: spreadsheet import recovery keeps the current file and choices. Replacing a file or resetting the form retires older preview results and errors; confirmation still controls when customers are created.
• Your Inbox is rebuilt around conversations. Pages of 20, one thread at a time, star, read/unread, Archive, Spam and a recoverable Trash act on the selected threads with Undo. Every folder and sort filter shows its count, Spam counts the automated mail a mailbox received, a chosen topic shows its “Sorted as…” banner immediately, and rows waiting to be sorted no longer carry an amber edge.
• Writing email gains review and safety. Compose, Reply, Reply all and Forward open editable drafts with reviewed Cc/Bcc (Bcc stays private), an explicit send confirmation that lists every recipient and attachment, clear missing-field messages and safe discard with Undo. They all take attachments from your computer (program files refused; only you can send what you uploaded), a promised document in a draft is satisfied by attaching the file, each person has per-language plain-text signatures, and replying again to an answered thread opens its own reviewable task. Mail sent from a connected mailbox no longer names the approver as Reply-To, so the customer's answer returns to the same conversation. The October 25 update checks the current account, organisation and task again before sending; unconfirmed send outcomes stay locked for review. The floating Assistant no longer covers recovery controls after Discard.
• Duplicate customers can be merged. Select them, choose the surviving profile, and resolve conflicting fields with the latest copy or your own choice. Related records stay traceable, the original profiles remain as linked read-only records, you keep working while the merge runs in the background, and completion offers to open the result; merge history lives in the customer's property menu.
• The Workbench explains itself. Today at a glance opens four working numbers (Overdue, Urgent, Due today, Done today); an Overdue label says which condition applies — the customer asked by a date that passed, or no reply within your window — quoting the customer's own words. Sections sit in separate cards with an ⓘ that explains how their tasks are created, pages show five rows at a time, and quiet or never-contacted customers are raised across sweeps instead of being capped. Task windows open with why the task exists and what is left, keep the draft in one column with only its own email under it, offer Mark as spam for suspicious mail, and a locked Send always says why; reopening a task by its link restores the saved draft.
• Email feeds the customer record. Linking a message files its supported attachments into your My Docs and the customer's documents, the sender can be added as a contact in the same flow, and a thread from an already-saved contact links itself. A sorted email can suggest the deal-stage change to confirm or dismiss, and a proven-sender purchase order opens the pending Won check with the source file — only a person confirms a change or a win. Every reply-needed email gets one owned task, linked or not, and a reply missing its required source citation gets one bounded retry before it stays blocked. The October 25 update also explains running, failed, no-facts and approval states; Check analysis status reads saved progress without starting another AI analysis.
• Scheduled greetings and reminders. Admins configure holiday greetings and recurring reminders; each rep then receives one batch task with per-customer drafts to review, send or skip, sent from their own authorized mailbox after confirmation. Scheduled instructions accept up to 2,000 characters, and a new or changed task can take up to an hour to appear.
• A faster, cleaner Assistant and clearer email setup. The Assistant answers simple daily-priority questions by reading your current work immediately, suppresses echoed internal markup in streamed and saved answers, and settles attachment questions honestly with the file's own lines as evidence. Setting up a mailbox starts from the address — Alibaba Enterprise Mail (including newer qiye.aliyun.com servers), Aliyun Mail and 263 are recognized, the app password is entered once for both receiving and sending, and screens explain when an administrator must help.
• Navigation, accounts and trust. Desktop and narrow-screen navigation keep Inbox, account and Workbench destinations usable with focus returning to the right place, the obsolete Open work list is gone, and you can switch between previously authenticated accounts with isolated workspace state and an unsaved-work confirmation. Sign-in no longer claims a code was sent when the transport has not confirmed it, and organisation owners can review scoped, redacted assistant diagnostics under live access checks.
• Operator verification keeps uncertainty visible. Health and backup reports retain open incidents, role-scoped diagnostics keep validated page history after Retry, and local release, Workbench, intent and documentation checks report missing or invalid evidence. Offline intent results remain synthetic; documentation findings do not imply that the whole manual passes.

v0.53.0 — October 19, 2026

• Customer location is easier to identify. Customer forms, detail editing and list filters now say Country/Region, with 国家/地区 in Chinese. The import template uses Country/Region, while existing Country and 国家 column headers remain accepted. On phones, customer creation has full-width guidance, labels above fields, larger controls, a pinned action bar and an Add customer button outside the scrolling customer tabs.
• Customer imports explain existing-customer matches and let you review them or skip all unresolved matches in one action. A separate confirmation imports the new customers, with decisions kept visible while scrolling. The ? icon opens an explanation of all four choices; Leave for now replaces Merge later and does not schedule an automatic merge.

v0.52.0 — October 13, 2026

• The Email plugin responds immediately when you connect, open or disconnect an account. Disconnecting updates the page at once instead of re-reading every account first, and the Plugins “Connected” view no longer lists disconnected accounts by default. Opening a connection’s page loads that one account directly instead of waiting for the whole organisation’s account list. A newly connected mailbox starts fetching its mail immediately, and the Inbox says “Fetching emails…” while the first download is under way instead of looking empty; mail appears as it arrives. Operators also gain a documented single-connection REST read and a per-account sync state on the Inbox accounts read.

v0.51.0 — October 12, 2026

• Your Inbox sorts itself and links mail to customers. New mail is sorted in the background into topics such as inquiry, quotation, sample, purchase order and invoice, grouped by conversation, with a line that says how much is sorted. What Atlas cannot sort waits under Needs you to sort; a conversation suggests its customer and deal, and one tap links it or creates the customer. The reading pane puts Atlas's suggestion and your decision first, with the evidence below.
• Send from Atlas. Reply to an email from the Inbox, send a task's draft reply, or ask the Assistant to draft an email and send it through your connected mailbox after you approve it; the sent email joins the conversation.
• Connecting a mailbox starts from your address. Type your email address and Atlas identifies the provider, including 263 企业邮箱, Alibaba, Tencent Exmail and NetEase, and walks you through the authorization code; no server settings unless you ask for them.
• One Workbench page. Needs you, Atlas working and Done replace the old tabs and notification list. Tasks come from your organisation's own settings: a reply owed to a customer, a deal missing its next step, a customer gone quiet, missing customer details and files to review each open in a focused window that shows why the task exists, the Atlas draft and what closes it.
• Document review is faster and complete. Uploaded files open a compact review beside the original; extraction progress, the proposed buyer, product lines, prices and a new deal are reviewed in one form; unsaved reviews stay as visible drafts; files filed before the review page existed open the same review, and a file whose original is gone can be retired.
• The Assistant answers as it writes and keeps its place. Answers appear sentence by sentence, a dropped connection resumes, every conversation keeps a server-side record, enabled skills load on demand, and organisations can write their own text skills.
• Knowledge answers cover every document type. Semantic search and Ask Knowledge recall from PDF, image, Word, PowerPoint and video documents, answer sooner, cite the place in the document and no longer dead-end on a spreadsheet question.
• Billing and Settings are clearer. Invoice status uses the complete payment history, usage recovery times are stated accurately, Settings is easier to find by task with clearer role labels, and phone usage entries keep their request, token counts and provider cost together.
• The Email connector page and the Inbox count the same accounts. A connector page now lists every account you can see, disconnected ones included, so its count matches the Inbox sidebar; “Connected & paused” and “Disconnected” remain available as filters. A shared account you manage but cannot read is marked “No inbox access”, which is why it does not appear in the Inbox.
• Pages load several seconds faster. Atlas now does its server work next to its database, so the sign-in check every page makes first, the Plugins page and its connected email accounts, and other lists and API reads no longer wait seconds for data.
• Manage mailboxes from the Email plugin. Superpower → Plugins → Email lists the ways to connect, Gmail and other IMAP/SMTP email, followed by your connected accounts. Select an account to reconnect it or disconnect it; disconnected accounts are folded away.
• Open any skill to see what it does. Each installed skill on Superpower → Skills now opens its own page with its description, the teammates and Assistant it applies to, the file types and phrases that trigger it, its inputs, outputs and constraints, and your organisation’s usage over the last 30 days. Organisation skills also show their instructions and keywords, and taught methods name the teammate they follow.

v0.50.0 — October 7, 2026

• PowerPoint previews preserve the deck’s appearance. Original view shows supported slide backgrounds, font sizes and colors, alignment and picture placement, scaling together on smaller screens. Existing uploads benefit without re-uploading. Supported draft slides can be previewed before Confirm without starting processing. After processing, unsupported slide structures keep a readable extracted view, and speaker notes and chart values remain available. The original download remains available.
• Large originals can be saved with clear reading limits. Library, My Docs, Assistant uploads and customer documents accept supported files up to 1 GiB. Files over 50 MiB are saved without automatic reading or AI processing; smaller files still follow their reader limits. Permissions and storage quotas still apply, and processing supported videos over 50 MiB requires selecting the original file.
• Assistant progress is easier to follow. Customer-intake requests record the skill used in English and Chinese. Revised plans use readable action names, activity counts include every batched action, PDF reading shows its running step, and answer controls remain usable on a phone.
• My Docs and file associations keep their context. Customer uploads remain visible even when your personal Library is empty. On desktop, unresolved associations survive reload; saved associations show readable customer names, and Knowledge pickers show the selected document’s correct status.
• Sales pages keep the current view and recover from failed reads. Contacts, deals, quotations and new-work-item searches ignore obsolete responses; in-progress work can retry missing customer names. Customer pickers read beyond the first directory page, quarterly targets follow the selected period, and failed task lookups or Pipeline board reads offer recovery.
• Communication review follows your latest selection. The current review item, mailbox and queue filter stay in sync. Late reads or action completions cannot replace the selected item, and refresh uses the current selection.
• AI-team and routine runs finish on their own. Account Manager, Sales Assistant and routine runs complete on a durable workflow instead of waiting for the next five-minute sweep, and a routine run is billed to the person who pressed Run now. A supervisor can hand a customer's RFQ to the Account Manager with one confirmation, and an AI-team member can be configured to run when a linked inbound email arrives, proposing reviewable profile updates and deal signals from the email's own words.
• The Assistant answers as it writes and remembers everywhere. Answers appear sentence by sentence while they are still being written; after a dropped connection or reload the full answer is recovered. Conversation memory is kept on the server, so a cleared or second browser continues the same conversation, and a branch keeps its policy answers on every device. Long answers stay within budget and re-read only the passages they need, without claiming a full read it did not make.
• Every uploaded document gets one honest review page. Values say where they came from, an amount the document never stated stays blank, and Save files one document at a time. My Docs keeps a Needs-you filter, a daily task groups the files still waiting on you, and every document opens in the same viewer with View and Download side by side.
• Organisations write their own text skills. Under Settings, AI team, Skills, owners and admins can create, publish, edit and retire text-only skills, and the Assistant loads the skills your organisation has enabled, naming them on the Skills used line. Taught methods appear in the same library as read-only rows.
• Financial readings keep their printed sources, and originals are retained. Delivery-note references and labelled product quantities and prices keep their citations back to the document, and an original Atlas cannot process is retained with its limits explained. The staged Superpower workspace groups existing capabilities under Home, Knowledge, Plugins and Skills with guided email setup; existing accounts, permissions and work remain intact.

v0.49.0 — October 6, 2026

• Knowledge reads product decks that carry one very large picture. A PowerPoint file whose single picture is larger than 8 MB no longer fails with a wrong “cannot be read as Word, Excel or PowerPoint” message; the deck extracts with its slide text, tables and charts, and a picture too large to describe visually is skipped and noted instead.

v0.48.0 — October 5, 2026

• Settings navigation and page layout. Settings keeps you in the right place. Team review, Adoption and My issues retain Settings navigation, while all Settings pages use consistent widths, margins and compact headings. Labels follow the same heading hierarchy. Existing permissions, bookmarks, browser history and return navigation work on desktop and phone.
• Read-only email Inbox is available across organisations, with existing account permissions preserved. Search loaded messages beside the selected inbox heading, read message contents on desktop or phone, and keep your place when returning to the list. A connected account without messages has a distinct empty state; unavailable or discarded messages reveal no Inbox content. Instant messaging is outside this rollout’s verified scope and will be covered separately. The existing Inbox switch also governs IM; it does not hide existing IM entries. Reply, Reply All, Forward, Delete, Pin, unread tracking and account management remain outside this release.
• Everyday pages answer faster with nothing changed about their answers. The customers list loads its health badges in one batched read, a customer's detail and engagement panels fetch their parts in parallel, knowledge search resolves what you may see in a single pass, and the Assistant reads a named customer's owner, contacts, open tasks and receivables at the same time instead of one after another.

Also in this release: security hardening — sessions refuse to verify when a deployment secret is missing instead of weakening to a public development key, evaluation-only request annotations are refused for ordinary users, and API responses through the app gateway carry x-content-type-options: nosniff; two additive database indexes back the customers list and customer task lookups.

v0.47.0 — October 4, 2026

• Ask Knowledge answers sooner and refuses less. Document questions in a library that also holds spreadsheets no longer wait on a spreadsheet check, a question that asks no calculation is answered from the documents instead of being refused in spreadsheet wording, and the documents found appear within about a second while the answer is checked. A calculation over a spreadsheet is still answered by the spreadsheet engine.
• Part numbers, orders and standards are answered from the right place. A part's MOQ, price or tolerance is answered from its row in a catalog table or price list with the page or sheet cited, a question that names a document is answered from that document first, and references written with a space or a version — PO 8841, ISO 9001:2015 — are answered only when a cited document names that exact reference.
• The assistant reads the whole passage it cites. When it finds the right page it can read the full passage with the passages before and after it, or a whole short document, before answering; it searches at most twice and then says plainly when the documents do not state the answer.
• Library uploads confirm at once and show their progress. Text files, and Word and PowerPoint files without images, are read in the background after Confirm, with the row showing Queued, Reading or Indexing; long documents are read much further; and every file's detail page shows whether semantic search covers it, with Retry for administrators when it failed.
• Knowledge files now open in one consistent page. Images, PDFs and spreadsheets from the Library or a customer show the same navigation, viewer and file details, with available previews and extracted content for each format. An unavailable customer document offers recovery without unusable file actions.
• Inbox is easier to scan and read across email and messaging. Search opens from an icon beside the selected inbox name; the reader shows a clear subject, sender, recipients and date without extra review panels. On a phone, search returns to the message list and Back keeps your place. When no email or messaging accounts are available, Inbox points you to Extensions. This release remains read-only.

Also in this release: a passage read for agents and the command line — GET /knowledge/library/files/:id/passages, the MCP tool ekb_library_passages_read and atlas library read; an opt-in two-step Ask Knowledge response (Accept: application/x-ndjson); a route stage in Ask Knowledge Server-Timing; an administrator retry for semantic search (POST /knowledge/library/files/:id/embedding/retry); and an answer speed and quality baseline tool for operators.

v0.46.0 — October 3, 2026

• Atlas opens from a smaller chat bubble into a compact input, then a conversation when you send. Prompt me and right-panel expansion remain, with New chat at the top right and quieter replies without the skill record. Choose Library files, customer files or whole folders in a separate modal with a compact selection review and existing attachment limits; local file selection keeps its existing intake path.
• Assistant drafts and attachments now follow you between the floating chat, side panel and full page. The Add menu offers Add from local and Add from library. Sending waits for My Docs saving; if saving fails or takes too long, keep your draft and choose Retry or send without a library link. Stop, New chat and account changes retire earlier replies safely, Chinese composition no longer sends early, and shared startup reads avoid duplicate requests.
• Ask the Assistant how to do something in Atlas or where to find a feature. It answers from the User Manual with numbered footnotes that open the exact manual section, and says when it is not sure instead of guessing. When it is unsure, or when you ask it to report a problem or request a feature, it fills in the Send feedback form for you; you can change anything, and nothing is sent until you press Send.
• Knowledge search and Ask Knowledge now find documents by meaning across every file type — PDFs, images, Word, PowerPoint and video as well as text and spreadsheets — and beyond a long document's opening pages; existing libraries are indexed in the background. The Assistant's library search uses the same keyword-plus-meaning recall, keeps part numbers whole and cites the page, sheet or slide. When a workbook in your selection cannot be read, a question that needs no calculation is answered from the other documents, and Ask Knowledge declines rather than answer about a part number or reference that no cited document names.

v0.45.0 — October 2, 2026

• Deals move through six simpler stages: Inquiry, Needs confirmed, Evaluation, Negotiation, Won and Lost. A deal's stepper shows only the stages it actually entered — a skipped stage is drawn as skipped — and marking Lost asks for a reason that is stored once and shown on the deal, in its Stage history, on the customer's timeline and in the monthly review. Quoted, Sample sent and Sample evaluation come back as milestones on a panel that traces each to its own record — a sent quotation, a sample record, a linked purchase-order file. A Customer needs section summarises what the customer has said with a source for every fact, Atlas moves a deal forward when the evidence says so — marked · Atlas, naming the rule and records, with a one-click Undo — and a purchase order lays itself beside the deal and asks a person to confirm the win while Today lists what is waiting to be verified.
• Workbench becomes the shared home for your work. Every member starts there: Overview separates decisions, agent activity and important signals with the selected record and its existing actions beside the list, and Planning brings real commitments, owners and due dates into the same workspace — old task links keep working and the separate My Tasks navigation entry is retired. Its customer figures — current, active, at-risk and open pipeline — match the scope you actually carry: the company for owners and admins, your named team for managers, your own customers for reps, with contributors inspectable and pipeline separated by currency.
• Settings can manage named human teams. Owners and admins create and rename teams, assign each member to at most one team, give a team several managers, and archive empty teams — current permissions, observed revisions and conflict recovery protect every change, and existing tokens receive no additional grants automatically.
• The workspace reads calmer and explains itself where you are. Customer records and imports, agent teams, retrospectives, Knowledge, insight and team views, feedback, handover, review and settings share one compact heading with real ancestor breadcrumbs; explanatory copy moves into a contextual Help button beside the title — exactly one per page, fitting narrow phone screens even beside long headings — while scope, dates, identity and counts stay beside their records. Existing actions, return navigation, bilingual wording and phone layouts are retained.
• Knowledge and Settings get focused homes. Document review opens a pending-decision queue with separate history and source tools; Knowledge Q&A keeps its chosen scope visible beside a wider conversation; customer documents get a dedicated library page with customer and customer-type filters. Settings opens with a searchable overview that finds permitted pages and folded sections, and Integrations is now Extensions (扩展) everywhere it is named — the page lives at /settings/extensions, every old /settings/integrations address redirects permanently, connected accounts and connection health come before the service catalogue, and the pilot scaffolding (attention list, readiness card, self-test and MCP-server cards, hint lines) is gone.
• Every document answers which customer and which deal it belongs to — and is allowed to answer neither. A customer file shows a Deal column: the deal's name as a link, "Not a deal", or "A deal before Atlas" with the reference you typed; files without an answer show "Needs an answer" and are answered in place, and an upload refuses to land until its question is answered — nothing is guessed, and no file is ever assigned on its own. Files dropped in the assistant are asked the same question, and a Library file that names a company gets the same Customer and Deal fields, pre-filled and marked read, while staying in exactly one place — never copied between the Library and a customer's page. A customer file that has waited an hour without an answer raises one task — "Which deal is <file name> for?" — and the answer, or deleting the file, closes it by itself.
• A dedicated Inbox sits in the navigation rail above Assistant. Your linked Email, WhatsApp and WeCom accounts live in the sidebar's Inbox panel — an "All inboxes" aggregate first, each account with its waiting-review count — the message list gathers every account newest-first, and the reading pane shows participants, time, channel, parsed body and attachment names in a full-height layout whose panes scroll independently. Items awaiting review link straight into the existing review; search covers loaded subjects, senders and summaries; on a phone the account picker collapses and a message opens full-width with Back returning to your row. Read-only for now, and shown only where the platform Inbox switch is on.
• The assistant's document replies become one clear decision. The reply has two lines — what the document states, then your answer — and when it cannot tell which deal or customer applies it asks with the real options, the suggested one tagged "Recommended" and starting selected, and exactly one button whose label says what will happen: Review the form, File under this customer, Save to My Docs, File as a transaction on <deal>. "Review the form" opens the new-customer form carrying what the document said — even on a conversation reopened later — and a file dropped on the floating assistant is now genuinely read and filed. When a turn carries no document, the assistant says so instead of naming an organisation boundary that has nothing to do with it.
• The Money page is now called Accounts receivable (应收账目) everywhere it is named — the sidebar entry, the business overview's More detail chip, the page title and breadcrumb, and the pin label. Routes, links and earlier pins keep working.
• Routines arrive: recurring assistant tasks on a schedule. Settings gains a Routines page for owners and admins with two kinds — Customer status update, which re-evaluates a bounded batch of customers against one label set and proposes keep-or-change with a reason per customer, and Customers needing attention, which lands every customer that needs attention as exactly one attributed task on its owner's workbench, due 09:00 the next morning. Proposals apply in one action, or automatically only where the new "Customer status update by routine" permission is set to By itself; applied changes carry assistant attribution, appear in Done-for-you and can be undone; re-runs never duplicate a task, uncertain interrupted runs are not replayed automatically, and scheduled model calls meter credits through the governed gateway.
• WeCom (企业微信) and WhatsApp become connectable channels. Submitting the credentials verifies them with the provider before anything is created — a refused credential names the field to correct, and an unreachable provider creates nothing — and after connecting, Atlas shows the exact callback URL to register in the provider's console. Verified messages enter the same review queue as email (incoming only, no sending), connected accounts behave exactly like email accounts — list, health, pause, disconnect, reconnect — and a wrong signature, wrong token or unknown callback address is refused and reads like a route that never existed.
• The account popover carries the Support list again, and the issues page reads 工单. The profile menu at the bottom of the sidebar has its Support group back with one entry — Issues — opening the page where your submitted issues and their statuses live (also still under Settings → Account); on the issues page the Chinese title, breadcrumb and menu labels now read 工单 instead of 历史问题/我的问题, and English labels are unchanged.
• Navigation and work pages get a round of polish: Settings-first tools, calmer loading, one width. The Business overview keeps its figures but no longer shows an AI insight panel, a deal's page drops its Back link, "More tools" and "My issues" live in Settings (with Settings above View as in the account menu), feeds keep their skeleton while loading instead of flashing an empty state, the Workbench page gains a breadcrumb, and every Work secondary page shares one page width.
• Quieter reliability fixes: a late response from an older session can no longer replace a completed organisation switch, verified sign-in or sign-out — the confirmed identity stays; and the assistant's conversation list loads in a fixed three queries no matter how many conversations exist, which takes roughly a second off every page in the app..

Also in this release: a staging-built retrieval benchmark for the Knowledge library (passage BM25, a context-enriched dense hybrid and an ontology-grounded entity graph over a 24-file bilingual OEM/ODM corpus with a ~120-query gold set) reports per-lane quality, latency and cost against today's search, which stays the recommended path; the report and its frozen decision rules live in the operator documentation. Tenant isolation is proven with zero cross-organisation results in both directions. No shipped search behaviour changes.

September 2026

v0.44.0 — September 30, 2026

• Dropping a purchase order, invoice or quotation on the assistant now comes back as one sentence. The reply names the document, the customer, the amount with its currency, and the delivery date when the document states them, with just Confirm and Change. A value the document does not state is left out rather than shown as zero, and a currency is never assumed. When a document names two different companies in the same role, Atlas says so and writes nothing until you choose. One decision lives in one place: the triage card owns the reply, the question is never put twice, and the seeded new-customer form stays behind “Review the form”.
• My Docs lists every file you uploaded. Knowledge → Library → My Docs now includes files uploaded straight onto a customer’s page, which were previously invisible there. Each row shows the document type, the customer and the deal as links, so the Library answers “which files fed this deal?” — and a deal name appears only when the deal belongs to that document’s own customer.
• A customer’s industry is stored once, in the organisation’s Industry set. Every write — creating, editing, importing or the assistant filing a new customer — resolves to one value of that set, and a typed “Other” industry with custom text becomes a hidden value of the set: nothing is lost, and nothing new appears in anyone’s picker until a manager unhides it. The Customers column, the profile’s About panel and exports all read the carried value with today’s field shapes; saved industry filter links keep their meaning; free-text search still finds customers by industry. The customer profile’s Labels section stops listing Industry — industry keeps its own field. The REST manuals record the write forms and the industry filter’s semantics.
• Duplicate label values can be merged or moved, and the Customers filter narrows to an exact value. Labels settings gains two cleanup actions on each value: “Merge into…” moves every customer from duplicate values onto one value — naming beforehand how many customers each value carries — then retires the duplicates in one atomic step; “Move to…” relocates a value to another category (within its set or to another set of the same kind) without touching any customer’s labels, except that a value cannot leave the protected Industry set. The Customers filter panel now filters to exact label values — Region = Europe, not just “has a Region label” — with the chosen values kept in the link. The REST API records the two new endpoints, the label_value filter and per-value customer counts.

v0.43.0 — September 29, 2026

• One directory again: a prospect is a customer whose type says so. Every customer, whatever its stage, sits in the one Customers directory — the type filter offers Prospect, and the separate Prospects page is gone (its old links open the directory filtered to Prospect). Creating a prospect still needs only a company name, a customer drafted from a purchase order or invoice arrives with its type proposed from the file’s evidence, and a change of type is recorded on the customer’s history.
• One coherent customer-labelling system. Settings → Sales & CRM → Labels becomes the single management home for the four label concepts — customer types, customer classes, Industry, and label sets (organisation and personal) — each with a one-line hint; class values gain the same colour picker type values have, and deleting a label value customers still carry asks first. A customer's page reads one classification row — type, class, engagement — with both pickers showing each value's colour, and the class ladder's bottom rung is renamed Unranked / 未定级 (it was Lead / 潜在客户, the same word as the Prospect type). The Customers filter panel takes its type options from the organisation's own catalogue with their current names, grouped to match the profile page; saved links keep working.
• The AI Assistant finds customers by their labels. The assistant's customer search, customer detail and portfolio tools now filter by customer type, class and label set and name each result's type and class — the only filter it had before was an invisible “tier” field stuck on one value, so key and strategic searches always came back empty and Today's stale-customer panel stayed blank. Today now reads customer type, so Key and Strategic customers finally surface. The REST manuals record the removal of the customers-list tier filter and the tier patch field.
• Opportunities are now called Deals. Everywhere you read it — screens, buttons, filters, messages and the assistant’s replies — “opportunity” is now “deal”, and “商机” is now “业务”. The assistant still understands the old words, and nothing about how deals work changes.
• The Sales Pipeline gains a List view. Board stays as it was; the list shows every deal as a table with search, owner, filters and sortable columns. The separate Opportunities page is gone — its old links open the list with the same filters.
• Customers gains a board of its own. A Table / Board switch joins the customers list: the board shows one column per customer type (or per grade, via Group by), you choose which columns appear, and dragging a customer to another column changes its type or grade.
• The customer page reads as one page. The tab strip holds exactly the five tabs — Overview, Deals, Communications, Sales history and Files — the AI-extraction review entry moved into the overflow menu beside them, and Ownership and recycling protection collapsed into a rail card that opens only when you need it. About holds the company’s facts, Contacts holds the people, and profile gaps sit on the field they are missing.
• Insight grows Team and Money pages. Sales, Team and Money each open their own page beside the business overview — Team for managers and admins, Money for owners. The AI insight panel and the AI team and all-at-risk chips are gone from those pages, and the overview’s detail row lists Sales, Team and Money.
• Files reach the assistant from anywhere. A purchase order or invoice dropped anywhere on a workbench page opens the assistant rail with the file attached and read, and a file picked through the customer page’s rail is read too. A buyer written slightly differently on the document still matches the customer on file; when nobody matches, Atlas moves to the New Customer form prefilled from the file — every value citing its page, the evidence showing beside each field — and a second staged customer replaces the open draft or asks first.
• The assistant’s quality scorecard moves into Settings. The quality scorecard and alerts now live in Settings → AI & agents → Agent quality, above the alert policies; the old Performance address opens that page. The rail’s workspace formerly “Ask Atlas” now reads “AI Assistant”.
• The Knowledge Library’s menus widen, and customer docs grow up. The row and ⋯ menus are wider so long actions no longer wrap, and every document and folder gains Copy link; a Company docs folder’s menu adds New folder inside and Upload files here. In My Docs, an owner, admin or manager can move their own document into Company docs. The Customers Docs list now follows the Customers directory, with a customer-type filter, each row and header showing the customer’s type, and a calmer header. Files uploaded on a customer’s own page now appear there too, under “Files from the customer profile”, ready to preview or download and counted in the customer’s document total.
• Customer types can carry a colour. Admins and managers give each customer type a colour in Settings → Sales & CRM → Labels — each type’s row carries a visible Colour button — and the colour shows on the type chip across the customer directory, the customer profile and the Library.
• A round of smaller fixes lands with them. A blocked customer create raises the standard notice with a shortcut to the field, and red field errors follow your language the moment you switch it; timeline system entries read in your language; the Customers owner filter is type-to-search; the four summary cards and the Public customer rules button leave the Customers page; a transaction or a draft quotation can be deleted from its Sales history row; and a quotation the customer has already seen is voided instead — it leaves the working lists, the record and its history stay, and it can be brought back.
• Creating a customer now asks for the core profile up front. The new-customer form requires the company's name, customer type, country/region, main market, industry, lead source and interest/market, plus the primary contact's name, title, phone, email and sales owner; a missing field is marked in place and nothing is sent until it is filled. The interest/market you enter is now saved on the customer as a confirmed profile fact instead of being dropped, so it no longer shows as a profile gap. Choosing type Prospect marks only the company name as required and clears any “Required” messages it no longer needs, and a field’s label stays level with its box when a message appears under it.
• Library keywords are whole words, not character fragments. Chinese documents now earn keywords that are complete terms — a document about the 知识库 gets 知识库, not overlapping pairs like 知识 and 识库.
• Preparing a customer draft from an attached card or file no longer dies on an odd model answer. When the extraction model returns the source attribution in an unusual shape, the prepare understands that shape — the recognized values become one medium-confidence, fully-cited candidate — and the reviewable draft stages as usual. The instruction now teaches the model only the correct candidate shape, and genuinely malformed answers still fail with the honest retry message instead of a false “temporarily unavailable”.
• The new-customer page keeps Create and Cancel pinned at the top while the form scrolls, uses the full width with the header Cancel as the single way out, and always starts fresh. A company created on an earlier visit no longer reappears in the tab strip or its counter, unfinished drafts still resume from the Customers list's banner, and Tab moves between the real fields instead of stopping on the ⓘ field-help icons.
• When the assistant needs a decision — for example, a customer with the same name already exists — it shows the options as buttons in place of the message box, recommended option first. Pick one by click, number key or arrows and Enter, type your own under Other, or skip.
• Dropping a file on the assistant now triages it before reading it in full. On an exact match Atlas identifies the buyer or supplier, files the document under that customer and opens the Documents review with a five-line summary; on a close match it asks in chat first; with no match it offers to create a new customer. A multi-file turn, or a spreadsheet that looks like a customer list, is asked whether it is several customers' paperwork or a list to import. Nothing is created or recorded without a person's review.
• Document pages calm down and show more. The header is just breadcrumb and actions with Re-extract and Download living in the ⋯ menu, the Keywords card expands in place with inline add, edit and delete on hover, and the Extracted view lists everything Atlas read as labelled rows with a true Previous/Next pager for multi-page files. Image understanding now also shows the image's basic facts — format, size, dimensions, language, capture date and device — and its distinct visible objects, each with its own factual caption passage beside the whole-image description.
• CRM lists and the pipeline gain selection and right-click actions. The customers list has the same leading checkbox column as opportunities (select-all, per-row, Shift and Ctrl/⌘-click ranges) feeding the bottom bulk bar, and right-clicking a customer, contact or opportunity row — or a pipeline card — opens Atlas's own shortcut menu with open, start-a-task, copy email, advance stage, archive and delete, confirming before anything destructive.
• Real phone photos land in the Knowledge Library instead of failing. JPEGs that carry a producer footer after their end marker are accepted, camera-resolution photos are no longer refused by a pixel budget, and the ⋯ menu's Retry now succeeds for these files. Truncated or corrupt images still fail with their reason.
• The floating assistant reads as one chat again. The owner-only suggestion strip no longer stacks above the conversation, so every persona sees the normal chat widget — header, quick chips and one composer.
• Turning Enterprise Memory off now takes the module off. The Library navigation and its pages disappear for everyone — a platform superadmin included — Ask Knowledge withdraws its policy and procedure checkboxes, and the assistant answers policy questions with an honest “turned off for this organisation” instead of answering. An organisation lifted to Preview keeps everything, and a read outage never silently disables a working module.
• Record a transaction straight from an uploaded invoice or PO (API). Uploading through the Documents tab can now ask Atlas to read a purchase order, invoice or receipt once and propose a transaction with every value tied back to the exact line it came from — a currency is never guessed, only shown when the document states it. Nothing is recorded until the proposal is reviewed and accepted, and the same document can only ever record one transaction. (API only in this release — no Documents-tab UI yet.)
• The UAT fixture reload shows its work (staging UAT org only). The confirm dialog stays open through the rebuild with a spinner and step-by-step progress, then reports the counts, and the rebuilt demo org reads as a functioning OEM/ODM business — eleven customers, opportunities in every stage, timelines, labels, tasks, quotations, purchase orders, agreements and a sixteen-file bilingual knowledge library.
• Every work page now opens with a slim job strip under the breadcrumb: the job's type and status, “Plan N of M done”, the next step that needs you (or “Nothing waiting on you”), an “N running” link while background work runs, and — for task jobs — the due day with an Open task control. On a phone the strip stays at the top with an Assistant button that opens the job's conversation with its plan, and the record's title wraps by words instead of clipping.
• On a job's work page you can ask for background work in words: the assistant starts it as a read-only background job under the current one, replies with its running card and keeps talking. The result lands in the same conversation when it finishes, and the job nests under its parent in In progress. With the permission set to Never nothing starts and the assistant says so; outside a job it offers to make the chat a work item.
• Two assistant permissions now act on the inbox. “Log an email or inquiry to the customer it matches” at By itself auto-logs a mail whose match has exactly one winner and lists it in Done for you with Undo that restores the review and the customer's last contact; “Update an opportunity from a customer's inquiry” waits in Decide with the email quoted at With permission, applies itself with Undo at By itself, and builds nothing at Never. A new customer from an unknown sender always waits for your confirmation.
• The Work side panel gains an “Open work” list. Every job you picked up and have not finished, newest activity first, each with its status dot and one click from its work page. A task you picked up now reads “Waiting on you” on its work page, in In progress and in every count, so the same job carries one status everywhere; in Chinese the Working status has its own word, no longer the same as the In progress page name.
• Assistant answers read more naturally. The Workbench AI insight states each fact once in an ordinary sentence and shows its time as “today 15:16” in your language instead of a machine string, and records the assistant cites — customers, deals, timeline entries — appear as small named chips that open the record instead of bracketed ids, in Ask Atlas, the floating assistant and the work-page assistant alike.
• A job without a plan now says so and offers “Plan it with me”. When the assistant proposes a checklist in the job's conversation, it becomes the job's plan — each step marked By itself, Needs you or Never — the “Plan N of M done” count moves as steps finish, and a finished answer leaves no step of its checklist unticked.
• Every “Review & send” or “Open the draft” — on the Workbench, a customer's Open work, In progress or a work page — opens the same review page: who it goes to, the editable text, exactly what will leave Atlas with your edits marked, the sources and the history. Internal account ids and field names are gone; Send appears only after Approve, and Back returns to where you came from.
• An Ask Atlas answer about exactly one deal can become deal work in one tap: the “Make this a work item?” card appears with Opportunity already chosen and the deal's customer named, and Make it opens the deal's work page keeping the conversation. Answers spanning several customers offer nothing.
• Reps and managers, who can read but not change assistant permissions, now see each action's setting as a plain label — By itself, With permission, Never — instead of greyed-out radio buttons, and the page footer links to Decisions without design-frame references. Owners and admins keep the editable grid.
• One permission, “Change a customer's status when its rule is met”, now governs both of a customer's rule-driven statuses — the contact state and the warmth label. By itself applies rule changes and lists them in Done for you with an Undo button; With permission waits in Decide for Accept or Reject; Never changes nothing. Undo, or rejecting a warmth change, keeps the previous status until the customer's next contact, and the customer list, its filters and the customer page always show the same status.
• Library fixes from internal testing. A My Docs row's menu now offers the same create-folder action as the right-click menu, searching in Ask Knowledge's document picker no longer misses Draft or Failed documents, reviewers see every sandbox feedback report their organisation's members leave, and retrying a failed Company copy says the retry started.
• Team overview opens for managers and admins under Insight. The team strip says whether the team is on track, each seller's conversion shows against the team average, at-risk customers surface with Assign follow-up, opening a person explains their gap from the records, a Coaching panel supports the monthly conversation, and the monthly team review reads like a review. The manager's Insight menu holds Team overview beside My performance, and Team review moves under More tools.
• Knowledge's Policies section is now Enterprise memory. One home whose tabs gather Policies, Procedures, Standards and Built-in permissions — with the assistant permission table moved in from Settings — plus Learned lessons, My memory, Human feedback and, for admins and managers, Agent logs, each the page you already used. Opening a run in Agent logs shows its evidence at a Knowledge address, with Back returning to the list. Who can see what does not change, and old addresses such as the retired Settings screens land on the tab that now holds the page.
• Document review and My suggestions now sit beside the documents they govern: admins and managers open Document review from the Library menu with the same waiting badge, and reps open My suggestions with their own list and withdraw action — the same pages Settings had, now at /knowledge/review and /knowledge/suggestions.
• Settings now shows exactly the drawn menu — Account · Personal; Organisation · Organisation, Integrations, Billing; AI & agents · Model provider, Agent quality, Roles & skills, Assistant permissions; Sales · Customer management, Sales targets, Quotation rules — and every folded page, including bookmarked agent-run links, opens the section or tab that now holds it, keeping query strings and run ids.

v0.42.0 — September 28, 2026

• Settings now opens on the menu the owner journeys draw. Account, Organisation, AI & agents and Sales, in that order — Billing sits inside Organisation instead of its own group, and each role sees only the pages it can act on or needs to read, so a rep's menu no longer lists read-only organisation pages. The folded-in pages have arrived at their new homes while their old addresses keep working: Billing ends with the Subscription record, Personal shows your sessions, Customer management carries Labels, Public customer rules and Engagement strategy as tabs, Integrations › Email holds the sorting questions, and Roles & skills gains a Schedules tab with the two assistant switches.
• Manager team figures (API). The manager's Insight menu gains Team overview, opening the existing Team view for now, and two new reads give owners, admins and managers every seller's figures — per-person customers, deals, pipeline, won amounts and activity (GET /crm/team/members), and where each seller's deals get stuck with the counts behind each rate (GET /crm/team/conversion) — over any window of up to 400 days. Reps are refused before anything is read.
• Everywhere now agrees on the at-risk and stalled counts. Work lists used to describe never-contacted follow-ups as stalled deals while the dashboards correctly said none — those rows still appear, with their actions unchanged, but now say what they are: a customer nobody has contacted yet, with the deal waiting on a first touch. The overview's At-risk block names the real total beside its top rows, the Today strip's monthly wins figure now says it is your own, and the assistant answers at-risk and stalled questions with the standard definition the tiles use.
• Policies can now be released with enforcement modes, evidence gates and rollback. A policy releases as enforce or warn-only — a would-be block is delivered as a warning and recorded until the release's enforcement date arrives — a future effective date keeps a release out of evaluation until it starts, and any release can be rolled back to an earlier one, which re-enforces its content as a new release. Scenario cases run against the whole policy set before publishing: a failing case refuses the release until it is fixed, and every admission decision is logged immutably, with weekly per-policy metrics and an organisation-level spike flag for owners and admins (GET /agent/assistant/policy-metrics).
• Draft policies from a company document — with sources you can check, in English and Chinese. Atlas can turn a company document into a draft policy where every generated section and rule cites a verbatim passage of the source and uncheckable rules are flagged; accepting a proposal creates a draft only, never a release. A rule may state a permitted alternative of up to 300 characters, and sections can carry Chinese headings and bodies — shown verbatim in evaluation reasons and on a rep's block card, and omitted entirely when absent.

v0.41.0 — September 27, 2026

• Approving a rep's document suggestion now makes it answerable in one action. The Company copy is added and processed in the same request, reaching Ready without anyone opening it. If processing fails, the approving reviewer sees a count bubble leading to a Processing failed list with Retry, and the rep's My suggestions card says where each copy stands — processing, ready or failed; videos say they wait for Library processing.
• The side rail now keeps its place when you follow a shared link to a Knowledge page your role cannot open. The access message shows beside the Knowledge menu instead of the rail jumping to Work. Permissions and the messages themselves are unchanged.
• A draft in Ask Knowledge now says what you can do about it. In the document picker, a draft you cannot use yet no longer points you to a control only library managers have: library managers keep the Improve-understanding wording on their drafts, your own draft tells you to open it in My docs, press Confirm and ask once it is ready, and a draft you cannot act on — such as a company draft as a rep — simply says it is still being prepared and will be made available soon.
• The owner menu now leads with the new owner pages. Business overview, Decisions, Team view and the view pages have replaced the older owner destinations: the menu no longer offers Business direction, Quote decisions or Team review. Old addresses keep working — /full-view opens the Business overview, /owner-decisions opens Decisions (both keeping any query or section of the address), and the retired task calendar and month table return you to My Tasks, which is now the one task list.
• Long checks can run in the background inside a job. On a work item's page, a new "Run in the background" control hands a checking or comparison instruction to the assistant while you move on. The job runs as a child item under the one you opened — shown on its card while it works and listed under it in In progress — and when it finishes, its result is posted back into the job's conversation. If the organisation has set background work to Never, or the assistant's permissions cannot be read, the control says so and nothing runs.
• Turn a chat answer into a work item. When an Ask Atlas answer is about exactly one customer, a small card under the reply offers "Make this a work item". Confirming keeps the conversation with the job — it opens on the job's page with the plan docked beside it and waits in In progress — and "Not now" keeps the chat as it is. The offer never appears for answers that span several customers or name none.
• Knowledge QnA now answers spreadsheet questions in one plain sentence, whichever scope you ask in. A which-parts question over a workbook names the matching rows and their values — for example AX-100 (Power: 65W) — both for a picked document and under All knowledge, instead of refusing or answering in raw data. One workbook with a stale extraction no longer spoils the whole ask: a mixed scope answers from the workbooks that are ready and falls back to documents, and only an all-stale workbook selection still says so honestly, pointing you to Ask all knowledge instead. Spreadsheet keywords and passages also stopped showing internal version identifiers.
• Library fixes and polish. Selecting several of your files in My Docs and pressing Delete now removes all of them after one confirmation that names the count — cancelling keeps everything. Suggesting one of your files for Company docs shows its success message, a My suggestions line with the pending count appears while suggestions wait, and suggesting the same file again explains it is already suggested with a link to your suggestions. The Knowledge QnA scope line names the active scope instead of a zero count, a citation chip no longer claims its document left the Library while the source is still loading, English figures pluralize per their count, the file select control's label reads Select, and the Chinese stalled-deals tooltip is localized. Deleting a document from a list row's menu now asks the same confirmation the document page asks before removing anything. In Customers Docs, a document listed under a customer opens from its row — keyboard included — the panel links to the customer's page, and each customer row shows its owner so same-named customers tell apart.

v0.40.0 — September 26, 2026

• The Assistant now works from the company's own guidance, visibly. Outreach playbooks carry their how-to-sell-to-this-segment instructions — tone, opening, what to avoid, when to escalate — beside their cadence and themes; editing saves a draft that changes nothing until a manager deliberately publishes the next immutable version, and withdrawing removes a playbook from matching without touching its history. When Atlas proposes an engagement plan it records which playbook version matched and why, and a touch draft follows the instructions of exactly that recorded version — saying so on the draft — instead of whatever is newest; republishing never silently changes drafts under an accepted plan, withdrawn guidance produces no new drafts, and the plan offers an explicit refresh. Policies cover every assistant action: an admin writes rules in plain language over typed facts (the actor's role, the customer's class), and before any assistant action is confirmed — a task, a draft, a send, an external call, an AI-team approval — the action is checked: a warning shows with its cited section, a block shows the reason and goes ahead only with a recorded override reason from a person, and every override is listed under Settings → Knowledge → Agent logs. Scheduled runs and anything the policy check cannot verify are refused, never waved through. Your own send is covered the same way: pressing Send or Retry send on an email draft the assistant prepared runs the same check first — a blocked send goes ahead only with a recorded override reason, listed like any other. The Policies list also says truthfully whether the current published release already has rules in effect, beside its editable library copy.
• Knowledge now reads the way the company talks about it. The menu splits into Library — My Docs, Company Docs, Customers Docs and Knowledge QnA — and Policies, holding Policies, Procedures and Standards; the Chinese header and its first entry both read 制度, and every role works from one Work rail with My work and My Tasks under a Workbench header and Business overview and Business direction under an Insight header. Reviewers' pages — Document review, Learned lessons, Human feedback and the governed sources — moved into a new Settings → Knowledge group for exactly the people who could open them before, a small count bubble leads reviewers to suggestions waiting for them, and empty document folders say "No documents yet" with one Add button.
• Every document in the Library opens as one full page. A viewer on the left shows the original and switches to Extracted — what Atlas stored, each passage with its page or location, content kind and keyword chips, a Text/Images/Tables filter and an honest "N of M" footer; spreadsheets show sheets and cell ranges, Word files their body and headers, text files their starting line, and every passage says where it came from. The right rail holds the summary, keywords with their own page, the processing pipeline for PDFs and images or workbook coverage for spreadsheets, and Status. Company Docs and My Docs share one Create button that uploads straight from your file picker without a dialog, reps can suggest one of their own files for Company Docs from its row menu, and earlier regressions are fixed: confirming a document no longer reports a false conflict, text files no longer hang at Processing, and spreadsheets answer again.
• The Ask Knowledge page opens on the approved canvas. Before the first question it leads with the ask-your-documents heading and the check-before-you-quote caveat, suggested questions and the scope panel standing on the right; once a conversation starts, the transcript owns the page with the composer pinned to the bottom, and a header control shows the current scope and opens the scope panel as a side sheet at any width — one panel for search, the folder/document tree, the include-checks, and the selected-count line. Answers read as plain text with source chips, a use-in-response link and Helpful / Needs improvement, and the composer is a bordered card whose Send control lights up only when there is a question to send.
• The sales area reads honestly and opens faster. Every card on the Sales Pipeline board shows the real deal amount, all eight stages are first-class everywhere — list filter, board, forms — and a stage strip above the board shows every stage with its count and value; the Opportunities list opens a deal with one click (bulk selection moved to a checkbox column), the deal page opens with a five-tile band — amount in 万, probability, days in stage, expected close, priority — above its history, and owners and admins start on All owners so the whole company's deals are the default view. The new-opportunity form keeps its buttons reachable and offers every stage and priority even on short screens.
• A customer's page starts with your work and their decisions. Opening a customer leads with "Open work for you" — your own Today lines for them, each with one plain reason and the same button it has on Today; the About panel reads "Not on file" instead of a dash and offers to fill a missing field on the spot, and a Profile gaps panel counts what is still missing. Contacts joined the right-hand rail — every contact edits from a card visible on any tab — and Overview leads with Recent communications and AI insights. When the assistant drafts a new customer from a pasted note, it fills in a best-guess customer type, always marked "Verify" so you confirm or change it before saving.
• Reviewing and finishing work on Today keeps its place. Clicking Review & send on a draft row opens a full page — customer, purpose, state and version, recipient, subject and message with your edits highlighted, the exact preview of what will leave Atlas, and sources and history beside it; while a sent message waits for its receipt the page locks editing and offers Check send status. Saving draft edits refreshes the sender check and restores Approve & send when ready, without closing and reopening the review or approving anything automatically. Marking a task done in the drawer keeps the drawer on that row with time and an Undo, invites a completion note that lands on the customer's timeline and one suggested follow-up with dates — only created when you save one; asking Atlas to check in on a stalled deal explains what it will draft and that nothing sends until you approve it; and End of day opens the four groups as a drawer from the right, counts what remains, and says nothing is carried over or dropped without you. The review now also shows earlier sends and failures instead of saying they were never attempted — with a confirmed retry for a failed connected-mailbox send — and rechecks the current recipient and sender after you save an edit, before Approve & send unlocks.
• Owners get their own day: Work and Insight navigation with owner destinations. Work and Insight now lead each role to Workbench, In progress, My Tasks and My performance, while owners and admins also see Business overview and owner-only destinations under Insight. Addresses for pages that do not exist yet open an honest “This page is being built” state instead of a dead end, the Money address stays owner-only, and reusable owner-view cards and loading states give the coming owner pages one consistent bilingual layout.
• The owner decides what policy holds — and the rep always finds out. When company policy holds an assistant plan for the owner's decision, the plan waits up to seven days and only the owner or an admin can decide it with a reason. A rejection — or the plan expiring undecided — reaches the requesting rep both ways: sending the same unchanged message again shows “Rejected by owner” with the reason or “Expired after 7 days”, and a task carrying that reason lands on the rep's Workbench and My Tasks the same day, exactly once no matter how often they retry; asking again means changing the message. Every decision — who, when, what, why — joins one decision log.
• Owner business foundations (API). Owner and manager business figures use one documented source and set of definitions for won revenue, won deals, customers, forecast and risk; owners and admins can store six quarterly KPI targets per seller, with unset targets absent and period targets prorated by calendar days. Atlas can suggest one of nine owner views from a plain question, generate short fact-linked owner insight text with a safe fallback, and assemble an upcoming-events feed from recorded tasks and meetings — services that feed the owner pages as those views arrive.
• Assistant permission tiers, private work items and system-created member tasks (API). Organisation owners and admins can store the assistant action tier for their organisation — by itself, with permission, or never — and members can read the choices; stored tiers do not yet change execution authority. Each member's assistant work items, progress and plans remain private, and the assistant can create an attributed open task for a named member through the existing signed Tasks boundary — how owner decisions notify the requesting representative.
• The Business overview answers the owner's three questions on one page. Are we on track, what needs me, and who is on it: the page opens with the period and how fresh the numbers are, one AI-written insight over recorded facts, three goal tiles — Revenue won with a pace label and time-gone tick, New customers, Deals won — each saying what closes its gap, and a four-figure strip (open pipeline, weighted forecast, win rate year to date, stalled deals) from the same documented definitions used everywhere else. Edit goals changes the three targets in place and the tiles update without reloading. Below: the waiting decisions with a one-time setup row when quote approval is off, the at-risk changes — never-contacted customers, sends whose outcome is unknown, failing AI runs, failing setup — each with an Assign follow-up that names who is on it and when, then what's coming up and where to look next.
• Decisions are made one at a time, with the reason recorded. The Decisions page works the owner's queue one decision at a time — the item's recorded facts, gaps reading "None on file", quote approvals carrying their full evidence pack with the threshold chip — and requires a reason the requester will see before Approve or Reject. Every decision joins the log with who, when, what and why; "Change the plan" opens the item as a work item instead of deciding.
• Work moves one item at a time. The Workbench focuses on the one thing that needs the member — what it is, why now, and six fixed actions (accept, mark done, open the draft, dive in, later, reject) each visibly unavailable with a reason when they don't apply — with the next few ready below. My Tasks gathers everything waiting into one list grouped now, this week, later, no date, and any row opens a preview sheet of the whole item with its evidence, stepping through with ↑ ↓, where a prepared email can be approved and sent through the same guarded path as the review page. Dive in opens the record's own page with the assistant beside it already on that job's conversation, and a plan card above the chat shows each step — may do alone, needs permission, never allowed — with Done and Skip on every open step. A new In progress page lists everything picked up, with a waiting count in the menu.
• Team, Coming up and Org health views for the owner. Team shows every person and the AI helpers on one page — a Health label with its stated basis, a "who needs help this week" card grounded in records, per-person actions and a handover link. Coming up opens the dated feed to a full week, month or quarter. Org health gathers accounts needing attention, pilot readiness, seats and the last 30 days of AI usage, each section independent; every view can be pinned to the overview.
• Sales targets and Assistant permissions arrive in Settings — and the stored permissions now bite. Sales targets sets each seller's six quarterly targets in whole yuan with team totals and edits the stage win chances through the same pipeline setting forecasts use; only the owner or an admin edits targets. Assistant permissions shows the nine actions with their tiers — the owner can change the unlocked ones, the three fixed rows say why not — and those tiers now govern every assistant path: a Never action is refused, a With-permission action needs a person even from a scheduled run, the check precedes the policy gate so an override cannot lift a Never, and an unreadable state refuses rather than assumes. The Customers at-risk card and a new Opportunities stalled tile follow the overview's shared rule with a "how counted" link, and Knowledge opens on Company Docs for owners and admins.
• My performance figures (API). GET /crm/analytics/performance serves each seller's six KPIs with prorated targets, team means and forecasts (whole yuan, own-probability-first odds) for week, month, quarter, year-to-date and annual periods; owner and admin read any member, a manager their roster, everyone else only themself.
• Every seller now has a My performance page under Insight. Switch between Week, Month, Quarter, YTD and Annual and see six KPI tiles with your number against the target, the team average and the forecast, a revenue forecast strip that shows how the forecast is built, and a note that only reads back the recorded figures.
• The owner's floating assistant can open a view straight away. Suggestion chips (How is sales doing?, Who needs help, Key clients, What's coming) jump to their views with no waiting, and a typed question opens the matching view instead of a chat reply when Atlas recognises it. Everyone else keeps the assistant unchanged.
• A Money page under Business overview answers who owes us, how old it is and who is on it. The yuan total open and the customers owing it, four age bands plus money with no age on record, the top five debtors' share, other currencies in their own currency, and one row per customer with Assign follow-up on aged balances — owner-only, like the money figures it restates.
• Done for you lists the assistant's own changes. The Workbench's Done for you page lists everything the assistant decided by itself today — what it did, for which customer, and which permission allowed it — each undoable with the undo recorded. Nothing there ever reached a customer; sending always needs a person.
• Later now works on every Workbench item. A put-off item leaves the list, the Next column and every count for three days, then comes back by itself.

v0.39.0 — September 23, 2026

• Today is one dense list. Everything waiting for you — prepared messages, overdue and due tasks, planned follow-ups, stalled deals and quiet customers, promised profile details and the AI team's overnight picks — reads as a single list grouped Now, Today and This week, each line saying who it is for, what to do next and why in one plain sentence, with one button that does exactly what that kind of work always did. Clicking a line opens it in a drawer with its facts and the same buttons; Previous and Next step through every line, and a switch at the top shows the same work as four counted columns. Your month, the day's first reminder, an end-of-day Work summary and what the AI team did overnight with an Undo on each item all live in the same place, and overdue work is counted from one source so Today and Tasks always agree.
• Each role opens Atlas in its own workspace. Reps start with Work, managers with Team, and owners and admins with Insights; Customers, Knowledge, Ask Atlas and every authorized specialist destination remain reachable with their existing permissions, with explicit return links and saved owner landing choices preserved. A display-only session field keeps custom roles in the approved Rep presentation without ever granting access.
• Insights opens with one business overview. Owners and admins land on a compact Business overview with Customers and Sales summaries and descriptive links into the existing customer-activity and sales detail pages, each with breadcrumbs and a Back control that returns focus to the link you left from; the owner-only Business direction page leads with goal pace and customer attention before the member inventory, and following an exception opens the existing destination while saying plainly that it is unfiltered.
• Customer pages lead with decisions, and preferences are explicit. A customer's identity, confirmed facts, contacts, open opportunities and tasks now precede AI insights and specialist operations; successfully empty sections collapse without hiding loading or failed reads, and targeted links reveal their section with navigation state preserved. An authorized editor can explicitly confirm a customer's Email or Phone contact preference from real evidence, review what exists, and reject an unwanted fact separately — conflicting facts stay visible and a failed save keeps your choice.
• The Assistant's prepared work is safer to accept. New next-step recommendations use confirmed communication preferences and real contact availability, missing or conflicting evidence requires explicit review, and before you accept a recommendation for a focus customer, Today shows that customer's existing open tasks and pending or approved-unsent prepared work in one expandable block — matching titles are never merged, and a failed read says coverage is unknown instead of claiming there is nothing.
• The Assistant is more dependable. The floating launcher moved to the bottom-left of the navigation panel so it no longer covers page actions; it gets out of the way entirely — from clicks, keyboard and screen readers — while a modal dialog is open, and returns exactly as it was. A customer-response source becomes a readable link to the timeline entry it came from when the reading really was the one the handoff asked for; plan narration now matches the read the turn actually performed; and an edited portfolio Sales Assistant can be reactivated again because its configuration preview checks a bounded sample instead of the whole book of customers, while a run interrupted mid-way records what it completed instead of losing the result.
• A checked knowledge answer can be carried into a customer reply. Ask Knowledge can hand an answer straight into a customer response draft after re-reading and re-asking exactly the documents the answer rests on, so what reaches the draft is what those documents say right now; it only offers this for an answer it can stand behind, asks which customer the reply is for from customers you can already see, and never replaces unsent text you typed without asking. The page's title now matches its navigation entry — Ask Knowledge / 「询问知识库」.
• Settings → Integrations was rebuilt around its connectors. The page opens onto one catalogue of channel and system cards — Gmail, Email, WeCom, WhatsApp, MCP server — each with its own add or chat entry and live account counts, and below it the accounts that need attention with the one action that restores each. The Email card opens a short connect dialog that hands over a one-time connection key with its copy control and setup steps, and Gmail starts Google sign-in directly. Every connector has its page with breadcrumb, search and status filter; the account page keeps Receive and Send as separate verdicts beside the Google-revoked reconnect callout, moves all account actions into one bottom row, and disconnecting now names the consequences and asks for the address to be typed before the destructive button arms. Per-member access becomes two clear switches per person with an unsaved-changes bar, the activity log reads as a redacted table with event and period filters, and admins see separately scoped pilot-readiness evidence. Same endpoints, permissions and health states as before.
• Creating customers and quotations behaves. The New customer dialog no longer fails silently when a contact detail is typed without a name — it says why, jumps you to the field, and catches a contact email missing an @ or domain before save; and a new quotation automatically joins the customer's current negotiation when no other quotation or document is already under negotiation, exactly as if you had associated it yourself.
• Release news arrives quietly. New release information appears as a small dismissible notice inside the workspace instead of automatically opening the release-history dialog; the full history stays one click away in the account menu, and Today actions, keyboard focus and unsent Assistant drafts are never disturbed.
• The in-app user manual was rewritten against the product as it is today, in English and Simplified Chinese: navigation paths name the rail group each role starts in, and new sections cover the Priority entry and end-of-day Work summary on Today, the Business overview page and its drilldowns, and Enterprise Memory with Operating guidance. Ask Knowledge is named consistently, and sections that need a platform-operator account say plainly that they were not re-observed.
• Runs and budgets recover on their own. A portfolio run that meets an unresponsive billing write stops waiting, keeps the customers it already reviewed and names the one that did not complete, instead of spending the run in silence; and a monthly credit hold a run leaves behind settles within about half an hour at exactly the credits its own recorded usage shows — or zero — so a stranded reservation no longer consumes the organisation's allowance.

v0.38.0 — September 21, 2026

• Every aggregate now names whose numbers they are, the period they cover, and how fresh they are. Your month labels itself “Your work” while the dashboard KPI cards and the owner's Full view say “Organisation”, each beside its period, the server's own update stamp where one exists, and “Checked at” — the moment your browser actually read the data. A customer insight generated on an earlier day now reads “Insight generated on <date>” instead of a relative age, so its stored words stay in their day, and the customer's actual last contact appears as its own current line. A failed refresh keeps the last successful view visibly dated beside its retry path, missing values read as unavailable rather than zero, and zero reads as zero.
• Retrospectives now show how you handled your focus customers. The weekly rep and monthly team retrospectives gain a read-only Focus-customer handling section beside the counted evidence: every customer a day's brief recommended during the period, with the recorded decision — edited then accepted, accepted, or dismissed with its reason — the accepted work linked to real task history, and carry-over marked when a recommended customer moved into the next period. The actor who accepted is named separately from the actor who completed, a private task stays private to its owner, and absence of evidence is never shown as ignored. Coverage is stated honestly — complete, partial with reasons, or unavailable — and older retrospectives read on unchanged.
• Plan adherence reads a revised plan once. When a three-touch plan or an accepted schedule was revised during the period, the retrospective now counts each obligation under the plan version that governed it — a rescheduled touch no longer appears twice, and a superseded date stops showing as still due. History older than the lookback window is labelled truncated instead of silently dropped.
• One safe review path per outbound send state. Uncertain outbound work now has one safe review path. A draft that has attempted a send shows exactly one state-appropriate action: a never-attempted draft keeps the ordinary approve-and-send; an uncertain, refused or failed result offers only the existing duplicate-risk confirmation, which re-reads the live send state the moment it is armed and stands down if anything changed underneath; an in-flight send offers only wait, and an accepted send shows its receipt. When the send state itself cannot be read, the row says 「Send status unavailable」 — visibly different from a draft that has never been sent — and offers only re-check and operator help, never a guessed send. The review dialog leads with the same attempt state beside the exact message and sender, an incomplete draft names its missing subject or body instead of offering the send, and a known 【请补充】 placeholder is called out in review.

v0.37.0 — September 20, 2026

• The Knowledge Library now works like Google Drive. The navigation gains a Library section with My Docs — a private space whose documents only you can see and manage, whatever your role — beside Company Docs, which everyone reads and owner, admin and manager manage; representative uploads now land in their own My Docs. The document list behaves the way Drive taught you: click a row to select it (highlighted — no checkboxes), Shift-click to select a range, double-click to open a file or enter a folder, right-click for that row's actions, drag selected files onto a folder to move them, and “Create folder for selected files” builds a folder and moves the selection in one step. Large videos finally upload: MP4, MOV and WebM up to 1 GB go up in chunks with progress and cancel, while documents keep their existing limits, and duplicate detection, quota and retry behave exactly as before. Existing saved views keep working — rename one inline from the new pen icon or a double-click, and removing one asks first. Selection actions stay in a floating toolbar without shifting rows; Customers, Contacts and Opportunities use the same click, double-click and Shift-selection patterns, with sortable columns. Sandbox and Assistant share a searchable My Docs and Company Docs folder picker with a selected-file list; Sandbox suggests questions from the chosen documents without sending them. Representatives can request to add a private document to a Company folder: an authorised reviewer sees only the submitted snapshot, and approval creates an independent shared draft while keeping the personal original. Company Docs holds the review queue and source history in place of Manage sources.
• Retrospectives now show plan adherence. The weekly rep and monthly team retrospectives carry a read-only Plan adherence section beside the counted evidence: accepted planned touches and accepted schedules that fell due in the closed period, each set beside its linked execution evidence — a recorded completion for a planned touch, a provider receipt for a scheduled send — with the owner at acceptance named and the completing actor kept separate. Cancellations, pauses, invalidations and supersessions that happened strictly before a due day stay listed as explicit exclusion reasons, and a due day belongs to the plan version accepted before that day began — down to the exact local midnight, DST included. Coverage is stated honestly — complete, partial with its reasons, or unavailable — and nothing is invented: no totals without complete records, no completion percentage anywhere, planned touches and schedules never summed, and provider acceptance never called delivery.
• The Assistant can stage a new opportunity from chat. Asking the Assistant to create an opportunity for a customer no longer ends in “I don’t have a way to create an opportunity record from chat” plus a substitute follow-up task: the customer’s Opportunities tab comes forward with the existing New Opportunity dialog prefilled — the title, and wherever stated the stage, amount, priority, expected close date and next action — and your click on the dialog’s own create button saves the record through the same write the page button uses. Relative dates you say (明天, 下周三, 月底, end of this month, next Wednesday) are resolved against the real clock server-side, and a date it cannot honestly resolve stays blank for you to fill; a deal value stated in a currency the form cannot record is never silently converted — the amount stays empty and the stated figure shows as a note under the field. The Assistant itself still writes nothing.
• Today’s focus now weighs your organisation’s recorded customer activity. When the daily activity check has measured a customer, its recorded level — Active, Cooling, Silent, At risk or Lost — drives the contact urgency in place of raw days-since-contact, within the same ceiling, so the company’s own contact expectations decide who needs the morning’s attention; the sweep’s own pending activity proposals no longer add extra points on top. Each focus entry shows its evidence: the recorded level and its points, the standard version as last measured (with a warning when an older standard recorded it), the ranking version, and the raw contact age explicitly marked as covered by activity — or a plainly stated last-contact fallback when no usable record exists. A day’s brief never changes rules mid-day.
• The Integrations page now opens on connector cards, and every connector has its own page. Settings → Integrations opens straight onto the connector cards — no older list first, no waiting. Each card's top-right control arms from live data: a + when you can connect, a chat icon once a connector has a connected account, and the chat icon opens the Assistant chat window with that connector's icon and name at its top. The custom-email accounts workspace — search, status filter, connect — lives on the connector's own page, and Systems management lives on the MCP server card's page; legacy ?provider= links land on the matching connector page. All setup and authority checks are unchanged.
• Customer management settings got simpler, and say what they are. The Settings area formerly labelled Sales & CRM now reads 「Customer management」 in the navigation and the page header (客户管理设置 in Chinese), its pipeline section reads 「Sales stages」 (销售阶段), and the descriptive paragraphs under the headers are gone. The sub-modules — sales stages, customer classification, custom fields, quotations, activity standards, today's focus and retrospective cadence — are plain sections split by separator lines instead of boxed cards, at every width including a 390px phone.
• Pipeline stages can now be deleted outright, not only hidden. Every stage row carries a delete action behind its own confirmation. Deleting refuses while opportunities still sit on the stage — naming the count — and the last active Won and Lost stages can never be removed, so the pipeline always stays usable; an unused stage leaves the catalogue for good and the remaining stages renumber cleanly.
• Adding a customer class takes one name. The class creation dialog asks for a single name instead of separate English and Chinese labels — one name shows for both languages until each side is renamed, so an entry never blocks on a translation.
• Customer and contact lists refresh themselves after a delete. Deleting from a list page refetches the current page from the server once the batch finishes: counts and pagination settle immediately, and a page emptied by its own deletion steps back instead of reading as an empty directory.
• Deal amounts read at the same scale on every screen. An opportunity's deal value used to render 100× apart across the product — Today's focus card might say ¥45K while the Opportunities row for the same deal said ¥450.00 — because some screens still divided the stored whole-yuan figure by 100. The Opportunities rows, the opportunity detail page, the customer pages' opportunity lists, the account pickers (Cmd-K, @-mentions, command palette) and the Today work queue now read the stored figure directly, so every surface quotes one number for one deal.
• Today's rows and decision actions got clearer and fairer. A needs-attention due timestamp stored with second precision no longer parses as a 1970 date stuck forever overdue; large Today KPI figures switch to compact form so nothing clips, and multi-currency amounts print ¥ instead of the ICU code CN¥ in English locales; loading skeletons on the dashboard, customers and opportunity strips announce themselves to screen readers; rejecting a quotation, a prepared suggestion or a standard-change proposal — and dismissing a Today row — now carries a destructive outline so a refusal never reads as a neutral action beside its solid Accept; and the Enterprise Memory Agent-logs page heading matches its navigation entry.
• Workbench leads the compact navigation. The primary rail stays compact with the existing destination order; Settings remains in the account menu. Hovering or focusing a module previews its secondary menu and highlights it in light gray; leaving navigation restores the current page’s menu. The organisation switcher appears once in the primary rail, and the secondary header names the selected module. Existing landing preferences and old links keep working.
• Pages share a consistent layout. Lists use the same page spacing, header, toolbar and create controls. Breadcrumbs show the current location, including Library > My Docs > folder > document, with Help immediately alongside. Record editors and specialised Assistant conversations retain their existing interactions.
• Configuration lives in Settings, findable by search. Quotation routing rules, customer ownership (now 「Public customer rules」), engagement strategy, and roles and skills have canonical homes in Settings, and the Settings navigation carries a permission-aware search: typing filters the areas to what the name matches and to what you may open, with a plain no-match state and one click to clear. Help content and past issues, usage views and knowledge management moved into the portal with their old addresses redirecting, and account limits stay visible only to the account holder.
• Knowledge work and monitoring sit beside daily work. Source management and approved-content operations are reachable in the Knowledge area without detours into Settings, adoption and quality monitoring have working destinations, and the Knowledge Library's loaded-count label now reflects what is actually loaded instead of an optimistic number.
• AI Team opens on the roster. The AI Team area opens on the team roster with search retained, and its performance and configuration views are URL-addressable, so a performance view can be linked or bookmarked directly.
• Customer pages group around active business work. The customer detail page groups its content around the active business — overview, deals, follow-up work — and an initialized draft (a half-filled form or staged change) survives navigating between groups instead of being silently discarded.
• Clearer stage labels and per-currency totals. Opportunity stage names render from your configured catalogue instead of raw internal keys, and opportunity totals and averages are computed separately per currency, so a mixed-currency book no longer sums incompatible amounts into one misleading figure.
• Customer emails carry no internal record ids, and every proposed follow-up is acceptable. A payment-collection email used to tell the customer 「单据「rcv_…」尚有剩余款项」 — an internal database id as the document name — whenever the source transaction carried no reference; emails now name a basis a customer recognises (the transaction's date and amount, or the invoice filename), and the scheduled follow-up email's subject no longer repeats the customer name twice, with each source listed once in review. Engagement emails prepared with a follow-up intent — the scheduled sweeps' standard shape — promised a follow-up task no screen ever offered; the review dialog now shows every follow-up a draft proposes, with the same explicit Accept that creates exactly one task per proposal.
• The Assistant sees the money picture when it answers risk questions. Asked about outstanding payments, the Assistant used to answer 「no records」 even with a confirmed balance outstanding, because its customer read carried no receivable data. It now cites the confirmed open receivables — balance, currency, state, due date, overdue flag — alongside the customer's open tasks, and the dates it reads carry a server-computed relative marker so 「today」 is never misread as yesterday.
• Six Assistant sales-flow repairs. Deadline phrases such as 「下周五之前」 ("by next Friday") and 「月底前」 ("before month end") now stage real dates instead of blank ones; the staged-task card carries a fillable due-date input; committing a staged task refreshes the page's follow-up list; a staged task lookup and quotation-stage routing were corrected; and the staged card no longer bounces back after a restore.
• Four Assistant Actions-module repairs. A staged greeting composes from fact values instead of leaking internal fact-key labels; due-date labels render in the same working-day timezone the date parser uses (Asia/Shanghai) instead of shifting a day early; dates stated as ISO strings or weekday phrases stage as written, with a guard against echoing the instruction back verbatim; and contact routing for staged actions was corrected.

v0.36.0 — September 19, 2026

• Retrospectives now list the customer activity-level changes actually applied in the period. The weekly and monthly retrospectives gained an activity-changes section that reads the recorded activity-transition events verbatim: original from/to levels, the committed time, whether the daily sweep applied it automatically or a person confirmed it (finder and confirming actor named separately), and the standard version and reason inputs exactly as recorded at application time. Proposals that were never accepted are not transitions, A→B→A stays two separate transitions, and nothing is reclassified against today’s standard, owner or customer class. Where the records cannot support a claim the section says so: until transitions carry a recorded owner they appear as organisation-observed evidence without per-rep totals (a rep’s own view says attribution is not recorded rather than showing a zero), a period whose capture history is not proven complete and gap-free shows its coverage gap instead of totals, an unreadable source says the section is unavailable, and a fully covered period with no changes shows zero. Legacy retrospectives stored before this change are unchanged.
• An orphaned evaluation-campaign reservation can no longer deadlock staging deployments. A campaign reservation hold whose campaign record no longer exists — previously permanently unreconcilable, blocking every staging deployment behind a reserved-target refusal even for a platform superadmin — is now named as orphaned on the reservation status surface, can be reconciled by any platform operator with audited evidence, and is cleared automatically by the next deployment acquisition it would otherwise block. Platform operators can also read and conclude a campaign that holds the shared evaluation target, and every existing ownership guarantee is unchanged: elapsed time still never releases a live campaign's hold.
• A portfolio run's review stays readable while the customers it actually wrote about are unchanged. Opening a completed AI-team run's review no longer answers that the information is unavailable because a different customer in the same portfolio was touched after the run: a recorded review is withdrawn only when a customer the review itself cites has changed, so the review — and teaching anchored to it — survives routine activity elsewhere in the book. When content is genuinely withdrawn, the panel says the evidence changed instead of reporting an outage; and a team member that does not exist for your account says so plainly.
• The owner brief's money-drift section now reads the recorded money history. A customer whose recorded settlements — paid transactions and confirmed receivable settlements — have stopped arriving well past their own recorded rhythm is presented as drifting, each entry citing the exact records it derives from, with amounts, currencies and dates as recorded and never converted. Customers with too little recorded money history are shown as unknown rather than guessed, and the signals stay owner-only.
• The Assistant launcher no longer paints over open dialogs. The floating Assistant launcher stands down while registered dialogs are open, but six hand-built overlays — task create/edit, task detail, the month-view day popover, the customer note dialog, the meeting-summary prepare dialog and the timeline touchpoint recorder — never registered, so the launcher still covered their bottom-right corner, worst on a 390px screen. All six now join the same modal layer: four migrate onto the shared dialog shell (gaining focus trapping, Escape close and scroll lock) with the meeting-summary keeping its exact dismissal semantics, and the recorder registers directly with no behavior change. The launcher stands down while any of them is open and returns the moment they close.
• The owner brief's signals can now be acted on where the owner reads them. Every brief row that names a customer carries exactly two actions: create a follow-up task pre-filled with the signal's customer and the same evidence the row cites — saved only when confirmed, as the owner's own ordinary manual task — and, when that customer has a pending routed quote decision, a direct link that opens the decision surface focused on exactly that decision. Rows without a customer and unverifiable states (thin-history unknown lines, unknown-count aggregates) offer no action, and the brief still sends nothing, edits no status and changes no ownership.
• A task due today stays actionable all day. A date-only due date was stored as the start of that day, so a task due today read as overdue since midnight and fired the notification bell. A date-only due now means the end of that day, and the list still renders it as a bare date.
• Asking for an unsent draft's send state no longer logs console errors. An identified draft with no send attempt answers an explicit 「none」 state; a genuinely unknown draft still answers not-found.
• A cancelled or completed task stops nagging the notification bell. An overdue reminder that had already fired for a task stayed in the bell after the task was cancelled or completed, counting as overdue for work that no longer existed. Reminders for terminal tasks are no longer served; unseen reminders for open tasks are unchanged.
• Today's decision queue stops offering drafts for customers the reader can no longer act on. When a customer moves between teammates, AI-team outreach and engagement drafts prepared for the previous owner no longer fill that person's Today decision queue: a rep sees only drafts whose customers are still in their own book, while administrators, managers and the customer's current owner keep full visibility — an actionable 「Approve & send」 never points at a customer the reader cannot open. A hidden draft is not deleted; it reappears if the customer moves back.
• The money radar's empty state says what is actually absent — open exposure, not money as a whole. With a customer's settlements all recorded and settled, the Full-view money radar still announced 「No money recorded yet」 while the owner brief on the same page cited those exact settlements. The empty state now reads 「No open money exposure yet」: the panel's domain is the open-transaction remainder plus confirmed receivables, which can be empty while settled money exists.
• Deleting a customer no longer fails when a quotation was routed to the owner. A customer or opportunity whose quotation had ever been routed into the owner decision ledger could not be deleted: the delete answered a service failure because the ledger entry still referenced the quotation when the delete cascade committed. The ledger entry now goes with its quotation, so the delete completes and removes the routed decision record with it, and the staging UAT fixture reload converges again. No other cascade behavior changed.
• The bug-report form takes screenshots and its own diagnostics. Pasting a screenshot straight into the issue description attaches it like any uploaded file, under the same five-file and 20 MB limits. A "Detected context" panel shows the page the report is filed from plus the most recent console errors or alerts the browser recorded — collected passively into a short recent history — included by default, with a toggle to send the description alone. The attachment picker is now a proper button, and image attachments show thumbnails. English and 中文.
• The bug icon opens the feedback window in place, and Past issues moved into Settings. Clicking the rail-foot bug icon (Send feedback) no longer navigates to a separate form page: the floating feedback window opens over whatever you are reading, in the Assistant window’s corner, size and chrome, so reporting a problem never moves you off the record you are on. Escape or the close button returns you to the page with focus restored, and the form’s type, severity, description and attachment behavior is unchanged. Your submitted issues now live in Settings → Account as 「Past issues」, the sent confirmation links there, and the old /feedback/past address redirects to it; the deep-linkable full-page form at /feedback remains.
• Platform feature visibility picks the state in each row. On the Platform operations Feature visibility page, each feature row now carries its three states — Developers only, Preview, General — as a segmented control in the row itself: picking an option saves it immediately, a refused change reloads the authoritative state and says so, and the per-row Edit button and its dialog are gone. The control answers immediately so keyboard navigation can never skip to the wrong state, and registering a new feature keeps its dialog. English and 中文.

v0.35.0 — September 18, 2026

• Supervisors can publish an account-level learned correction to their AI team member without running an evaluation campaign. When an AI-team member's run gets something wrong, the reviewer's recorded rating on that run can anchor a corrected method instruction: the supervisor reviews the drafted correction and publishes it directly to that member's account-scoped guidance, and the member's next run locks and applies the published version — with the run's identity naming the version it applied. The published correction is scoped to the one member and its account, is withdrawn or replaced through the same version head as before, and organisation-wide lessons and every other kind of learned guidance still require the full evaluation campaign.
• Headless assistant runs honor the customer record they were aimed at. An assistant run started through the headless API with a page context now reads that page's customer identity for its work, so API-driven runs land on the customer they were aimed at instead of the organisation's default record.
• Mailbox-aware assistants are now enabled for production organisations. An authorised agent can answer from connected mailbox content — always citing the mailbox item and the account it came from. The source mailbox's authority follows everything derived from it: a colleague, an administrator or anyone without a current grant on the source mailbox sees nothing — not the content and not a hint the item exists — traces and exports carry only withheld metadata, and revoking a grant cuts off saved answers and further reads immediately, so retention never becomes access. Behaviour stays fail-closed: nothing is readable until your organisation connects mailboxes in Settings → Integrations and names its read-grant holders.
• The What's-new dialog now opens itself once for an unseen release — in the language you are reading Atlas in. Closing it is enough: it stays closed for that release, on reload and beyond, until the next one. The version pill in the account menu still opens it any time, and the entries now cover only changes a signed-in user can see for themselves.
• A refused mailbox grant migration now says so — and the migration works again. An organisation administrator running the mailbox grant migration gets the real answer: a refusal — for example from a platform-support session, which is never allowed to enumerate organisation membership — is reported as a refusal, while a genuine outage still reports as unavailable. The dry run lists the members it would grant, and the apply writes the named read grants with the ready marker set.
• Two people pressing the same send retry can no longer churn the approval. When the confirmed retry of the same failed or uncertain outbound message is pressed twice at once — a double-click racing a colleague — exactly one retry dispatches; every other attempt answers that a send is already in flight, without touching the draft's approval. The attempt ledger, the duplicate warning and the exactly-one-more-attempt guarantee are unchanged.
• A failing connected-system tool call now names its real cause. When an external tool call fails during an AI-team run, the run's notes say what actually went wrong and where — a remote server error, a refused redirect, a rejected credential, a timeout, an unreachable address — both for the run as a whole and in each customer's own section; the Assistant's refusal names the same reason. No connection secret ever appears in these messages.
• Engagement outreach now leaves from the organisation's own connected mailbox. When an organisation has connected its mailbox, an approved engagement message — a product recommendation, a promotion, a follow-up — leaves through that mailbox's own sending session: the customer sees it coming from the address the organisation actually uses, and Google-hosted inboxes — which silently discarded the previous Atlas-addressed mail — receive it. Replies still route to the sending person. Organisations without a connected mailbox keep the previous behaviour, and a refused or uncertain mailbox send is reported as exactly that instead of being substituted with another sender.
• A run whose trace carries more steps than a delivery can retain says so on the trace. The trace view already showed a run's governed steps, but a trace carrying more than forty stored steps silently lost its tail at the store. The trace now names that cut: a leading, counted marker states how many steps were kept and how many were dropped, so a long run's record is truncated visibly or not at all.

v0.34.0 — September 17, 2026

• Spreadsheets in the Knowledge Library are now fully indexed and calculable. Uploading or re-extracting an Excel workbook reads every occupied cell on every sheet — not just the first sheet's first window — and the file's detail shows honest coverage: complete cell, row and sheet counts with the source version, and semantic search coverage reported separately. Ask the Knowledge Sandbox or the Assistant in plain English or 中文: exact lookups, filters, counts, totals, averages, minima, maxima and two-group comparisons are computed deterministically over every matching row and cite the exact sheet and cell ranges that prove the answer — never by model arithmetic. What the workbook cannot answer is refused with the reason and the fix named: a formula without a cached result names its cell, duplicate or merged layouts ask to be fixed, oversize results ask to be narrowed, and a workbook among several selected files asks to be named instead of being answered from an arbitrary one. Older workbooks stay readable but are marked legacy until a manager runs the bounded, resumable backfill with honest counts; replacement, retry, re-extraction and deletion never mix evidence from two versions.
• Connected email accounts show honest diagnostics and can recover safely. Settings → Integrations now shows, for each account, whether incoming and outgoing checks pass and when they last ran, when the mailbox last synced or a message was last accepted, and exactly one plain-language cause and next step per problem — never a host, username, password or raw server line — with the account's activity log alongside and a metadata-only export. An authorised account manager can also ask an account to help itself: Run check tests the connection, and Sync now runs one explicit bounded sync through the same safe runner as the scheduled sync — it never resends, resets or duplicates anything, it refuses to overlap another running operation, and every request is recorded on the account's activity log.
• A portfolio run's trace is now complete and readable. The trace view shows the run's governed steps beside its model-call spans, the token usage it actually spent, and which customers' learned guidance it recalled — with the per-customer context for each lesson lookup. Oversize traces degrade in visible, named stages instead of being dropped whole, and a customer identifier in a step stays the identifier it is so the step can always say which customer it served; content-bearing arguments and results still never enter the trace.
• Engagement plans now say plainly what is happening, and a cancelled plan can start again. Accepting a purpose whose earlier plan was cancelled re-activates that plan with the newly reviewed dates — its history stays attached — and a genuine duplicate tells you what to do instead of dead-ending: edit the existing plan with Revise. When a due occurrence cannot prepare because no Sales Assistant is active, or the AI team is closed for the organisation, Today and the schedule control say exactly that, with the hire instruction for owners and admins.
• Platform operators can open the AI team for one organisation at a time. While the AI team stays developers-only, Platform → Feature visibility gains a per-organisation override: enabling it for one organisation (audited, version-guarded) lets that organisation's members — and only theirs — reach the AI-team surfaces. No other organisation sees any difference, and no tenant user can set or change an override.
• An AI team member reliably reaches the connected-system tool it was told to use. When a member is offered the read-only tools from your connected systems, it picks the exact one by a short label instead of typing out its full system-generated name, so a run — including one covering a whole portfolio of customers — actually gets the answer it was sent to fetch instead of the lookup being quietly refused. Only tools a role is enabled for are offered; an unrecognised label is refused before any system is contacted; anything that would change data still waits for a person's confirmation; and the run's record, trace and audit still name the real system and tool.
• The AI team page shows your organisation's run health Above the roster: over the last seven days, how many runs finished without a written review, how many failed, and what share of finished runs that is — with the number of runs started in the window beside it, and per-member counts on each affected team member's row. A rollup, not an alert: nobody is notified from it, trial runs are left out, and nothing about how a run behaves changes.
• Uploading to the Knowledge Library reports in the File table itself, not a pop-up dialog. The upload picker closes on submit and each file being uploaded appears as a semi-transparent row in the table — a spinner and the word Uploading in the Status column — until its own row lands, then returns to full opacity with the normal post-upload status. Several files upload side by side; navigating the Library never stops an upload; closing the tab mid-upload still asks first. A refused file, a duplicate whose bytes are already stored, or a document whose extraction failed says so, named, in the page's notice banner; a refused file is retried by uploading it again.
• Approved emails keep the exact sender identity the human approved. If the platform's sending address is later re-pointed or cleared, an approval that bound the verified Atlas address is refused with the same visible conflict instead of sending from an address the approver never saw; the Reply-To is always the approver recorded at approval time, even when a different member presses send, and the record still names who pressed it.
• Re-seeding the default pipeline stage catalogue follows the permission matrix — the Roles & permissions row 「Pipeline & activity standards」, the same grant as every other stage write, instead of the sales-CRM-administrator gate it borrowed. Still non-destructive: it only fills a missing catalogue from the shipped defaults and never overwrites, renames or hides customised stages, and the control appears only for viewers holding the grant.

v0.33.0 — September 16, 2026

• Approved email sends are now durable, bound and recorded. The review dialog shows exactly which mailbox account a message will send from and binds that account at the moment of approval; if the account is later paused, re-credentialed or replaced, or the draft's recipient or words change, the send is refused with a visible reason until a fresh approval binds the current sender. Each approved send executes as one recorded send — reserved before any connection is touched, at most one delivery attempt, finalised exactly once as refused, failed, accepted by the mail server (acceptance, not proof of receipt), or honestly uncertain. Repeating a send returns the recorded outcome instead of mailing again, racing sends collapse into one, a crash mid-send is never silently retried, and an explicit retry for a refused or uncertain send names the duplicate risk in its confirmation and records a fresh approval against the account as it stands now.
• Connected mailboxes stay private, and only explicitly chosen content is ever published. Every integration review surface — queue, discarded, quarantine, message detail, matching and confirm/discard — honours each mailbox's named permissions: without read access you see no hint a message exists, and revoking access takes effect on the very next read. An authorised reviewer can publish one synced message — its body and only the attachments they explicitly tick — onto a customer record they may write, after confirming the exact target; the mailbox itself and every unselected attachment stay private, the published copy answers to the customer's usual CRM permissions and survives later disconnects, and unpublishing removes copy and attachments completely with an audit entry. When the Assistant reads private mailbox content, the run inherits that privacy: its answers, saved history, prepared drafts and traces withhold the content, and the trace shows only that mail was used — never what it said.
• Read-only email sync recovers safely And shows what happened: interruptions, credential repairs and mailbox rebuilds no longer duplicate or skip messages — an expired run is fenced at storage level, a failed batch replays from the last durable cursor, a mailbox rebuild starts a fresh epoch without colliding with already-published mail, and an unavailable mailbox is skipped with a visible reason. Settings → Integrations shows each account's last successful sync and any skip reason in English and 中文. Scheduled sync itself now runs under the authority of the person who connected, resumed, reconnected or repaired the account, checked afresh on every run, so removing that person's access stops it at the next run; and when scheduled sync will not run at all, the account now says why — nobody has authorised it yet, or the recorded person no longer has permission to read it — rather than showing nothing. An account connected before this release needs one reconnect to authorise it.
• Assistant → AI team gains a full Skills page: an owner or administrator sees every active skill in the library, its purpose in their own language, which AI-team roles use it, and its 30-day attempts, outcomes and ratings — with an on/off switch per skill for the organisation. Turning a skill off stops new AI-team runs and Assistant turns from picking it up, while one already using it finishes unaffected.
• Learned lessons are now inspectable and revocable. A run's trace page and Today's review dialog show a Guidance used block — the learned guidance the run followed and the reviewed lessons it recalled, each with its version, scope and why it applied, with titles shown only to people who could already read that lesson. An administrator can withdraw the active learned-guidance version from the Teaching tab: the next run, draft and chat use the base skills only, the withdrawn version stays in history marked Withdrawn, and new guidance goes live only as a replacement published through teaching and its evaluation.
• An AI-team member's external lookups now work end to end. The lookup round carries the supervisor's own work instructions, so the hire knows when a lookup helps; and when a live model's write-up invents a citation it was never given, the review is no longer lost — invented citations are stripped, uncorroborated statements are dropped with a plain statement of the loss, and every surviving citation is one the server supplied. Replies inside code fences or behind a sentence of prose read correctly; a reply containing two different answers is still refused.
• The Assistant's skills gain follow-up drafting and business report composition, built only from confirmed facts. Following up on a customer produces a matched pair of drafts — English and 简体中文 together, never one without the other — written only from what you gave it and the customer's confirmed records, with the source named behind each claim; it never invents a price, discount, term or commitment, and it cannot send or confirm anything. Ask for a weekly, monthly or annual report and Atlas composes the narrative from your organisation's own analytics, naming the figure behind every statement; anything it could not read is named as a gap instead of being filled in, and wording that puts a figure against an unread metric is held for review rather than handed over clean.
• AI-team members use only the connected-system tools their role is enabled for — and never change data on their own. Before a member starts work, Atlas settles which tools it may use: only those an administrator enabled for the member's role, on a connection that is live. Whatever it may not use is counted into the member's record and named with reasons in the run's trace. Anything that would change data in a connected system is always held for a person's confirmation — exactly once — whatever the member's permission level; a tool the remote system changed since it was classified is excluded with nothing substituted, and credentials appear in no run, record or prompt.
• Owner quote decisions become contents-safe with a one-page evidence pack. When the owner opens a routed quote decision, a bilingual page assembles the customer's recorded facts — deal and quote history, payment record, confirmed open balance, last price actually paid — every section citing its source records, and anything without a recorded source shown as a named gap rather than an estimate. An approval now admits a send only while the quotation's version, total and currency still match the decided contents: a draft total edited after a hold is held afresh as a new decision, and decided rows can never admit contents the owner did not see.
• Pipeline and activity-standard editing follows the permission matrix. Creating, renaming, reordering and deactivating pipeline stages, and publishing or restoring the company activity standard, now follow the Roles & permissions row 「销售阶段与活动标准」 instead of hard-coded checks — managers and administrators keep today's access, an organisation can withhold stage editing from its managers, a custom-role grant can widen it, and the Sales & CRM settings entry follows the same grant.
• The AI team gets more honest about limits and steadier under failure. A platform-only account reaching the AI-team pages now sees a clear bilingual message that platform access alone does not provide this permission, instead of a misleading loss-of-access notice. An administrator's Today page loads the team's decisions and done-by-your-AI-team group again instead of failing, and no longer waits on per-lesson evaluation checks to do it. And a published method whose source evidence changed no longer stops the whole organisation: the team carries on with the most recent earlier version it can stand behind (or its built-in method), the run names which method was set aside, and the stale one is marked invalidated for an administrator to re-publish once the evidence settles.
• A fictional policy-admission evaluation pack (internal, operator-only): administrators running a diagnostic evaluation campaign on staging can select the new pack, which seeds fictional policy releases, drives the one protected follow-up-task action, and tries forged approvals, a signed-out user and a direct write — every attempt must be refused with nothing created; seeded policies are deleted after each trial, and the campaign report adds separately counted retrieval, policy, effect, latency and usage rates with a pass/fail check that no trial created an unauthorised effect.
• The Knowledge Sandbox answers from your documents and keeps its scope beside the question. A question now gets a direct answer written from the matched documents — each statement cited — instead of a quoted passage, and a question whose answer is a single value gets that value with its source. An “Ask from” panel beside the question lists All knowledge, each Library folder and “Pick documents by name”, so what every question draws on is chosen and shown in one place; the first question box is taller, with example questions beneath it. On a phone the panel sits under the examples, and after your first question a small scope button in the question box changes it.
• The round Assistant button steps aside while a dialog or side panel is open, so it no longer covers the panel's own buttons, and returns when the panel closes.

v0.32.0 — September 15, 2026

• The owner's Full view gains an Owner brief: the week's honest signals in one place. Customers gone quiet are judged against each customer's own recorded contact rhythm — never a global threshold — and every item cites the record it came from. Silence reads as “no verified contact for N days”, customers whose history is too thin are counted as unknown, money drift waits for recorded money history rather than being estimated, and a daily digest can be composed on demand from the same signals.
• The Money radar now shows drift and aging: each card states the customer's own baseline time to settle, groups open receivables into 30/60/90-day buckets, and marks balances drifting past that customer's normal rhythm — with per-record ages, confirmation dates, and an honest “no settlement history yet” when there is nothing to judge. It also shows the book's concentration per currency — the total exposure and the top five customers' exact share of it, never summed across currencies — and flags recorded data gaps: outstanding amounts without payment terms, only partially received, or explicitly recorded as zero received. Gaps are shown as gaps, never as zero.
• The AI team grows into a full catalog: a Pre-sale Assistant that briefs each key customer's open inquiry with cited evidence and named gaps; a Customer Success Manager whose health brief names the exact record behind every line — open receivables, order delivery, shipment progress and after-sales feedback; and one organisation-wide Business Analyst writing cited weekly and monthly analyses with up to five recommendations, each naming its customer. A key customer can now carry a whole team of independent hires (never a duplicate role), every customer's profile shows the AI team already deployed to that account, and every team member draws on the same versioned Skills library matched passively to their role — runs lock the exact skill versions they resolved and never switch mid-run. Two new governed skills join the library: inquiry qualification turns an inbound inquiry into structured requirements with named gaps and the next questions to ask, and solution match matches those requirements against approved product evidence, citing every match and naming honest gaps.
• AI team roles become a versioned catalog you govern: every role the hire wizard offers carries a version, a hire keeps running on the exact version it was hired on, and an owner or administrator can switch each role on or off for the organisation — a switched-off role leaves the hire wizard until it is switched back on. Every hire also carries a permission level: plan-only keeps its brief, next actions and drafts on the run page with nothing reaching Today; co-pilot prepares the work and waits for a person to decide before anything executes; auto-pilot executes only what its role allows — creating and updating tasks, saving drafts — while external sends and binding quotations are always excluded on every level. An auto-executed action lands with a receipt naming the team member that decided it and can be undone once from the work history with both records kept, and pausing a hire takes effect immediately, cancelling the run still in flight. Platform operators get an AI team roles tab listing every catalog role with its scope, current version, status and per-version history.
• Approved company milestone sharing: a manager or administrator can approve a specific piece of company knowledge — new equipment, entering a key account's supply chain, a technology breakthrough — together with its exact source version for external sharing. A salesperson can then prepare a customer-specific draft built only from that approval and the customer-fit evidence they are allowed to see; each statement cites its source, and nothing reaches a customer until a person reviews and sends the draft.
• Platform superadmins get a production SLA dashboard: current health of every required service, the monthly availability and error budget, daily backup evidence, incident history, and the delivery state of support notifications — checked every five minutes, with alerts to support when an incident opens, worsens or recovers.
• Customer ownership, recycling and handover keep their promises under pressure: profile edits and assignment retries cannot overwrite a customer after ownership changes; recycling re-checks every gate when releasing each customer and advances fairly across large books; each handover keeps its own completion receipt and needs a recorded next action with an explicit due date before it can complete; and handover recipient suggestions rest on current membership and authorised relationship evidence.
• Deciding a suggestion on Today now sticks: accept, reject or dismiss a CRM rule suggestion and the same suggestion for that customer stays quiet for seven days — even while the underlying gap is still open — instead of reappearing on the next rule scan a few minutes later. Suggestions nobody has acted on keep today's behaviour and still expire on their own schedule.
• Connect your own systems' tools over MCP And put them to governed use: an administrator connects an MCP tool server from Settings → Integrations, Atlas verifies it and records every tool it offers, and each tool stays inert until the administrator classifies it as a read or a change and names the AI roles allowed to use it — a tool the server later changes waits for re-classification. From then on a read runs when the Assistant needs it, while a change becomes a proposal naming the system, the tool and the values it would send, and only a person's confirmation makes the call. The same organisation-scoped tools appear in MCP clients such as Claude Code, Codex and Cursor and in the Atlas CLI, and AI team members — for one customer or across a rep's portfolio — can call up to three enabled read tools per run, with every result cited as evidence and every call traced.
• Knowledge works again for everyone signed in: the Library, the Sandbox and the other Knowledge pages load normally on production after the knowledge base moved to its own database, and a lesson learned from one customer can no longer appear as organisation-wide guidance in another customer's agent run.
• Engagement preparation and sending explain themselves: an accepted recurring schedule with no active Sales Assistant to prepare it — including while the organisation's AI team is not yet switched on — now says “Not preparing: no active Sales Assistant to claim this” on the customer's preparation list and schedule control, and prepares that occurrence exactly once when one is activated; a customer's schedules now show on the customer page even before an engagement plan exists; and a send the mail provider refuses or leaves uncertain shows the provider's own reason in Today's review dialog instead of a bare “Error”.
• Operators: the daily backup of the production database collects its export reliably — the backup workflow now polls the export every 3 seconds inside one step, picks up the briefly served download link as soon as it appears and restarts a dropped export job.

v0.30.0 — September 14, 2026

• Your Assistant conversations are kept by Atlas: past chats live on the server, so opening Atlas on another browser or device finds them there; deleting a conversation removes it everywhere, and the dialog says so.
• The Assistant speaks one language at a time and shows its reasoning honestly: English sessions no longer leak Chinese or raw internal tool names, customer counts come from one organisation-wide figure, task coaching text follows your display language, and a failed run or a refusing mailbox explains itself in one plain sentence instead of a raw error code.
• Customer records read right: the customer-type value reads “Ongoing” so it no longer collides with Status “Active”, the Contacts tab badge counts every contact, the label picker searches large taxonomies, the Contacts directory filters by customer and role, a foreign-currency amount stated in a note survives closing and reopening a draft, and an unfinished draft shows a Continue banner on the Customers page.
• Knowledge shows its confidence: a Library search that finds no literal wording falls back to the closest documents by meaning and says so, the Sandbox flags when cited documents state different figures for the same question, and clicking a citation opens the document in a new tab without losing the chat.
• Video knowledge stays under your control: MP4, MOV and WebM originals remain unprocessed until you explicitly start AI Readiness. Atlas then transcribes English or Chinese speech and describes sampled frames locally in your browser, warns you to keep the page open, and lets you cancel, stay or leave. Saved evidence is timestamped, searchable and linked to playback of the original, with honest coverage and device limits. Atlas does not use paid video inference or video embeddings.
• Next-year cooperation plans only quote confirmed cooperation goals as goals — internal notes, personal contact details and internal amounts are never listed or accepted.
• The AI team keeps its promises: an assistant keeps its confirmed credit budget through every reconfiguration and activation, the hire wizard picks approved knowledge through the Library's searchable folder picker, scheduled method-optimization drafts are bounded to one per month, and skills quality evidence fits small screens with readable currency amounts and version-fair comparisons. Organisation credit caps can meter usage instead of hard-failing calls.
• Connections tell the truth: self-test connections are no longer flagged as failing mailboxes, and each connection's card shows its own receive/send state with truthful synced counts. Concurrent account edits report a conflict instead of overwriting another change, and checks completed for old credentials cannot mark replacement credentials ready. Shared-account grant migration marks readiness only when access matches and configuration remains unchanged, preserving explicit grants. Integrations now centres on your accounts, with bookmarkable details, separate incoming and outgoing email setup, and repair that preserves history.
• Company guidance now reaches every Assistant chat, the Assistant can remember each user's own confirmed preferences (visible only to them), and reviewing an assistant-proposed task shows the company-policy verdict with its citations before anything is created.
• Unattended customers now have a way back into play: the organisation can run a shared Customer pool, with automatic recycling it configures itself. A Pool page under Customers lists every released customer with its country, industry, previous owner, pooling time and reason — searchable and paginated — and any teammate can claim a customer with one confirmed action; on the customer detail page a dedicated Ownership block shows the current state, lets an owner or manager release a customer to the pool with a stated reason (the confirmation names the open work left behind), lets managers and admins assign a new owner directly, and lets them protect a customer from automatic recycling for up to a year. In Settings → Organisation an admin previews exactly who would move under a proposed inactivity window before activating it, and every gate is honoured: inactivity counts real human communication only, an assignment grace applies, the owner and the managers receive a pool warning in the Today notification bell — customer, countdown and what to do — and nothing is ever released without a delivered warning. Every ownership change rides the same revision-checked, audited transition that refuses a stale view by naming the real conflict, and everything is bilingual and usable at phone widths. Profile edits and assignment retries cannot overwrite a customer after ownership changes. Recycling rechecks current follow-up, protection and policy when releasing each customer, and continues fairly across larger customer books.
• Say 记住… / “remember…” in a chat and the Assistant stages a memory card with exactly what it proposes to remember — nothing is saved until you click Save on the card, and the reply never claims it was. A confirmed memory applies from your next reply on, and the My-memory panel in the Assistant workspace lists, adds, edits and retires your memories. Everything stays private to your own account.
• Full view's Sales team table opens a continuity view per salesperson: pick anyone who owns customers — including someone who has already left — and see their whole customer book with what is recorded for each customer (history, evidence, open deals, next actions), then open one customer to check its completeness against the same profile standard used elsewhere, with the original author and date of each supporting record. What is missing is said plainly, what you are not allowed to see is marked as such and never counted as missing, and private tasks are never counted or named. Read-only and owner-only. Open-task counts belong to each customer: private work stays unknown, unavailable sources are named, and capped counts are shown as lower bounds.
• Handing customers over is now a completable workflow for the organisation owner: prepare one confirmed batch handover of many customers to a single active member, with a preview of current and proposed ownership, permitted open-work counts and the required transfer reason before anything changes; ask for recipient suggestions ranked from recorded, organisation-scoped evidence — portfolio size, open work and activity — each shown with its evidence and freshness, and gaps explained plainly when the evidence cannot distinguish anyone; and mark a handover complete only once a next action with an explicit due date is recorded for the customer. Every transfer rides the same revision-checked, audited transition as single ownership changes, each customer reports its own success, conflict or failure without undoing the rest, and retrying a partial batch returns the recorded outcomes instead of duplicating anything. Existing authorised next actions can satisfy completion without duplicate work. Each new transfer has its own completion check and keeps prior receipts; failed task creation and cleanup remain visible and retryable. Recipient suggestions use current membership and the selected customers’ authorised relationship evidence.
• Large quotes can now require the organisation owner's own decision before they are marked sent: the owner sets one threshold (amount and currency) for the organisation, and any send at or above it — or in a different currency than the threshold — is held at the source and appears on a new owner-only Quote decisions page with the quotation, customer, amount and requester on one row. Approving lets the requester mark the quote sent (its sent record names the decision); rejecting keeps it unsent; every decision is recorded permanently with who decided, when, and on which quotation version, and decided rows can never be changed. Routing is off until the owner turns it on, quotes at or below the threshold flow exactly as before, and if the routing setting cannot be read the send is refused rather than allowed past the gate.

v0.28.0 — September 13, 2026

• Connect the organisation's own mailbox in Settings → Integrations: a guided setup enters the mailbox account, app password, incoming and outgoing servers with port and security, the From address and the history window — each with a worked example — then signs in to the inbox and verifies the account may send before the app password is sealed, and port 25 or a security mismatch is refused up front with one plain sentence and one action. Once connected, approved outreach and replies leave from that own mailbox — customers see the organisation's From address and replies reach the sending rep — and Atlas keeps the connected inbox in sync on its own schedule so incoming customer messages match conversations without hand-importing; the sync is read-only, can be paused, and a refusing mailbox leaves the message unsent with a plain reason instead of silent retries.
• The Knowledge › Library list sorts straight from its column headers — click Name, Size, Status or Updated to sort, again to reverse, a third time for the default order — while the old sort dropdown and the search's file-type/updated filter dropdown go away (the folder picker and its chip stay), the page's help hint sits beside the title, and adding a view is a compact “+” that rides at the end of the views bar.
• Hand the assistant a customer list in CSV or TSV form — a file or pasted text, English or 中文 — and it works end to end: the delimiter is detected, text decodes as UTF-8 first, columns map onto the Atlas import fields with the company name required, the mapped rows show for review before anything is saved, and saving routes through the governed Customer import review.
• Platform → Feature visibility becomes easier to run: every feature opens a detail page at its own address showing its current visibility, who last changed it and which portal surfaces the switch controls; the edit dialog picks the state through one segmented three-option toggle group instead of three loose buttons; and the list rows slim down to the feature name and Edit.
• Accepted recurring industry & technology digests now really prepare their drafts: each due occurrence produces one reviewable draft — or a visible skip when no eligible attested source exists — knowledge ranking broadens to rank candidates rather than drop them, the overnight sweep's time budget covers the preparation step, and the digest still refuses external research.
• The AI team page gains a Skills usage view for owners and admins: each skill's last 30 days of attempts, outcomes, conversation results and 👍/👎 ratings, drawn from run receipts that are now correlated with the assistant reply that triggered them — so a rated reply is attributed to the skills that ran inside it. Ordinary members see none of this data. Skills views — the Platform catalog and the new usage table — switch between table and card grid from their top-right corner and remember your choice.
• Learned lessons gains a method-proposals tab: a shared, read-only view of each assistant's teaching candidates and the instruction versions currently active, with links into the AI team page where every action still happens.
• Recalled lessons have a real lifecycle: stewards explicitly mark which approved records may ever be recalled, and can correct, restrict or erase a recalled lesson, place and release legal holds, and run a retention sweep — every control takes effect on the very next read, and erased lessons leave only a tombstone.
• Assistant runs now compose steward-approved lessons as clearly labelled advisory context — kept strictly separate from authoritative customer facts and company policy — and record exactly which lesson and instruction versions they used.
• Customer imports read tab-separated text files (.tsv) as well as spreadsheets, recognising the delimiter automatically.
• The round help button on every page now sits immediately after the page title instead of at the far right of the header row — the placement the Knowledge pages already use — while page actions such as create, edit and refresh, and the Today notification bell, keep their places. On detail pages it hugs the customer, contact, opportunity or team-member name; the chat pane keeps its corner placement because it has no title.

v0.27.0 — September 12, 2026

• Team-plan accounts now carry real per-account usage limits: before every model call Atlas sums the signed-in account's usage over rolling 5-hour and 7-day windows and compares it with the recorded tier limits; a call at the limit is refused with a clear bilingual message naming when it frees up, Settings → Billing shows members their own window state while billing admins see every member's, and everything that is not a model call keeps working.
• Customer imports recognize 线索来源 (lead source) end to end — a spreadsheet's 来源 column maps automatically and assistant drafts propose the stated source for confirmation — and bringing data in got faster and steadier: files convert in the background, batch intake runs in parallel, customer creation stops waiting on serial writes, and transient OCR errors retry once.
• An Account Manager's evaluation campaign now launches on a reviewed paired case pack and produces control-versus-proposed pass counts, costs and per-trial records, while withholding every quality or cost claim unless execution identities exactly match the frozen launch; comparison reasons read as localized sentences with the raw code kept as a small detail line; teaching suggestions surface budget limits instead of generic failures and settle their credits correctly; a run that called no model says so plainly; and a hire's credit bounds are recorded at the exact revision confirmed.
• Teaching proposals can name where a lesson comes from: an anchor to the Sandbox or agent-run rating it is based on is verified against the real rating before anything is saved, and the agent's teaching tab opens directly by link.
• Information records become recallable only when a steward explicitly marks them: marking is one-way, reads default to deny, and marked records answer only while active, within validity and retention.
• The rep prepares an after-sales survey report from a customer's page: pick one coverage period, the recorded survey evidence and a recipient; the draft states only what the selected sources say, names missing months as partial coverage, and waits in Today for review before anything is sent.
• Approved emails now send from Atlas's verified sending mailbox with the sending rep kept on Reply-To, so approve-and-send completes even when members use other email domains.
• The workbench reads better at the edges: the About card finally shows the customer's primary contact, closing a dialog returns focus to its trigger, the small hours greet you honestly, creating a customer no longer writes the primary contact twice, task due labels follow the calendar, discarding prepared work demands a second confirming click, and an already-accepted follow-up says so instead of a generic error.
• Platform and operations hardening: authentication admits the organisation's current policy authority itself with recorded revisions, proposal approvals are bound to the exact policy and authority digests they were shown (refusing and recording nothing when policy has moved on), and the staging deploy script can no longer publish a stale web bundle.
• The Knowledge Sandbox reads as one conversation: the scope is a chip inside the question box, every in-place notice lands in one slot at the top of it, a first visit centres the purpose line, the question box and the example questions together, and the scope chooser shows only the active kind while keeping each kind's choices.
• Customer imports read any UTF-8 file correctly, an attached list keeps its stated deal amount and lead source, and a foreign-currency figure is never silently entered as CNY.
• Today is safer and clearer: sending an approved draft takes a confirming second click that names the recipient, decision reasons read in your language, amounts share one compact format, and the AI-team summary no longer shows an error it cannot act on.
• The customer list and Customer 360 keep your place: search, sort, page and the open tab live in the URL so Back returns where you were; the phone layout shows three readable columns; counts, AI insights, prepare buttons and the evidence picker all say what is true.
• Knowledge studios read as one product: workflow branch conditions are readable rules, policies and workflows share one status vocabulary, uploads count finished files, and help links open the manual of the environment and language you are in.
• Settings, Platform and AI team name things properly: organisations, customers and AI members appear by name, the platform feedback queue carries the submitting organisation's name, dialogs behave like dialogs, required fields are pointed out, and zh-CN copy uses Chinese punctuation.
• Platform roster accounts signed in to their home organisation can use the AI team page as administrators: review the team, propose, confirm or reject team changes, and run or resume hires. The Account Manager access list and evaluation environment rules are unchanged.
• When a staged customer's deal amount was stated in a foreign currency, the New Customer form now shows the stated figure and currency beside the amount field instead of leaving it blank, so the rep knows what to convert and enter.
• AI team hires explain a failed run: the Overview reports the most recent run with its time and a plain-language reason (for example, no model provider key is set up for the organisation), run history shows the same line, and features released in preview carry a Preview badge in the sidebar and page title.

v0.26.0 — September 11, 2026

• The Knowledge Library separates customer and company documents into two workspaces with personal saved views, and managers review an AI-proposed folder plan — including the moves it could not confidently place — before anything moves; a removed source document stays clearly labelled as unavailable instead of exposing an internal id.
• Extraction evidence is complete and readable: spreadsheet charts keep titles, categories and numeric series separate, governed image reading accepts the supported structured responses, and Word/PowerPoint results show full file metadata with explicit reading limitations in both languages.
• Atlas adds monitored daily long-retention database backups and documents a 99.9% monthly availability commitment with a recovery point of at most 24 hours.
• Platform operators can start a new organisation from an existing business template — its CRM defaults and labels copy over, Manufacturing stays the default when none is chosen, existing organisations are never re-seeded, and a failed seed or invitation keeps the pending organisation for retry instead of deleting it.
• Enterprise Memory becomes the home for governance workflows: the Policies and Procedures studios join it as one tabbed page, the steward's Information review queue joins as Learned lessons (approving a candidate stays a steward decision), and a new Human feedback workbench reviews Knowledge Sandbox and agent-run ratings together with per-source filters, honest availability counts and run evidence opening through its own trace.
• Every product page shows a “?” help icon beside its title that opens a short bilingual explainer of what the page does and how to use it — read-only, closing with the button, Escape, or a click outside.
• Industry pickers recognise 照明器具制造 (lighting fixture manufacturing) with one consistent Business services option; the Today needs-attention list shows readable stage names for sample-stage opportunities in both languages; and a website captured from a buyer's message now reaches the assistant's customer-entry form and the saved record instead of being silently dropped.
• Reps can prepare an after-sales survey report straight from a customer's page: pick one coverage period and the recorded survey or feedback evidence, optionally add open questions, and review the draft in Today before anything is sent — the report states only what the selected sources say, marks missing months as partial coverage, and never sums response counts a source does not itself provide. Approved emails now send from Atlas's verified sending mailbox with the sending rep kept on Reply-To.
• Account Manager evaluation campaigns can launch for real: a reviewed two-case diagnostic pack ships with its exact method binding so the control-versus-proposed comparison produces real evidence, and the credit-budget row is now written at the hire revision activation confirms — fixing the off-by-one that refused every launch after an activation.

v0.25.0 — September 10, 2026

• Bringing customer data in from files got an end-to-end overhaul: spreadsheet imports now carry in every explicitly listed company and contact, including websites and email addresses; attaching an Excel or photo customer file in the Assistant prefills the customer-creation form again instead of failing; a file that is only partly readable keeps the details that were extracted and states exactly what is missing; and the import mapping step explains in plain language that each dropdown picks a column of the uploaded file.
• Sales teams can run the whole deal follow-up loop in Atlas: mark which exact quotation revisions are currently under negotiation (external quotations can be attached with their metadata and are never treated as sent or accepted), keep one confirmed tender brief per tender with every requirement citing the customer's own evidence, follow each order's delivery commitment from its PO or contract, record each sample's confirmed test result (pass, fail, inconclusive or retest required), ask for inventory and sales check-ins where quantity, period and amount stay three distinct measures, and chase the current sent quotation by number and revision. Every record keeps what the customer actually confirmed, versions each correction, and stops following up on its own once the deal moves on.
• The chat composer's paperclip became a “+” button with one menu: attach files or whole folders from your computer, or pick Knowledge Library documents in a side drawer — they attach exactly like uploaded files, and the per-turn Sources checklist lives in the same menu.
• Engagement sends gained double-send protection: when a provider call fails ambiguously (for example the connection drops after the provider may have accepted), the send is recorded as outcome unknown and the draft locks instead of offering a retry that could deliver twice; and two approved drafts for the same customer, purpose and record can no longer send while one is in flight or of unknown outcome.
• Sign-in now checks only the address's shape and lets the server's configured admission decide, so a newly admitted organisation's domain needs no web release before its people can log in. Platform superadmins reach every Platform page from ⌘K, new organisations start with only the generic customer labels (grading labels are added per organisation), and the Settings organisation page shows platform operators a link to the Platform console instead of a delete control their session could never use.
• Customer search accepts several clues at once and requires every clue to match the same visible customer — across company details, linked contacts, country or industry names, classifications and visible labels — in the Customers page and the Assistant alike.
• The platform Feedback workbench matured: every new issue gets a readable code such as FB-2026-0042 that platform search finds, cards drag between status lanes that scroll on their own, the issue page's Delete / Resolve / KIV buttons collapse into one status control, replies with files join the issue's attachments immediately, superadmins can edit or delete any submitted comment with an audit trail and an “edited” marker, and multi-select filters, inline editing and a calmer submission form speed up triage, with the form's secondary New issue / Past issues navigation and clear title restored. The navigation rail no longer duplicates the feedback entry — the bug icon at its foot is the single way in.
• Transactions record agreed payment terms (a 30% deposit, net-30 wording) next to their payment status plus the amount received so far, with outstanding totals counting only what is still unpaid per currency. Logged communication records can be corrected from the customer's Communications tab — method, time, duration, summary and details — keeping the original for audit with an “edited” mark. Industry pickers add 专业技术服务业 (Professional technical services), and choosing Other reveals an optional box to name the specific industry.
• Atlas prepares more cited, reviewable documents: an accepted annual-report schedule now produces the 年度合作报告 draft when it comes due; a recurring per-customer industry & technology digest drafts up to five passages from attested knowledge paired with the customer's confirmed facts; a promotion meeting's summary draft carries only what was explicitly confirmed, with at most three proposed follow-ups accepted exactly once; and next-year cooperation planning drafts an evidence-backed, non-binding proposal from the annual evidence summary and confirmed goals — suggestions stay clearly labelled, never quotations, orders or budgets.
• Organisation owners get a Full View (全景图) dashboard — company goals per period with record-backed pace, team performance per rep, customer status counts, and AI-suggested improvements and next actions that each cite their figures and become My Tasks in one click. My Tasks opens with a 待你处理 “Needs you” section counting everything waiting on the user, arranged by priority, type or time, with the chosen view kept in the URL.
• Long content behaves: review dialogs scroll with their actions pinned instead of clipping Save / Approve off-screen, the in-app What's-new dialog scrolls its notes and is keyboard-reachable, retrospective detail pages stop scrolling into blank space, and the shared dialog close button announces 关闭 in Chinese.
• Two governed reads opened to integrations: a cited engagement-context read returns one permission-checked customer context whose every item cites its source, content version, observation date and confirmation state (REST, MCP and the atlas CLI), and a read-only endpoint lists assistant-reply Like/Dislike feedback organisation-wide with role-scoped visibility and keyset pagination — both documented in the REST and MCP manuals.

v0.21.0 — September 9, 2026

• New Integrations page in Settings: one card per channel Atlas can connect to — Email, WeCom (企业微信) and WhatsApp — with a plain-language description, live status in words, and admin-only visibility. Connecting a channel is three guided steps — what Atlas will and will not do, only the fields that provider needs with copyable examples, and a one-sentence live verification; secrets are encrypted before storage and afterwards appear only as a masked hint. Platform superadmins can finish an organisation's channel setup, with every action audited twice and the customer's own page reading “set up by Atlas support” with the name and date.
• Pressing ⌘K opens a command palette that reaches every page the navigation shows — destinations appear above the familiar customer search, and typing filters both at once.
• Governed product promotion arrives in three pieces: admins save named, reviewable selections of approved product claims with live approval state per claim; from a customer page a manager drafts up to three tailored product recommendations bound to approved claims and the customer's confirmed needs, reviewed in Today before anything is sent; and a monthly new-product promotion schedule drafts only while every product keeps live customer-use approval, otherwise showing a visible skipped reason.
• Exhibitions become trackable and actionable: reps record which exhibition facts are verified — dates, venue, whose participation and booth — with the confirmer, time and source document on the record; a pre-exhibition promotion schedule counts back from a verified date and waits with a precise reason if the event moves or participation lapses; and invitation drafts are prepared from the verified facts — timing, venue and booth only when the company's own booth is confirmed — reviewed in Today.
• Receive desktop notifications for each new reminder while Atlas sits in a background tab, and click one to open the item; opt in from the notification bell (closed-browser delivery is not provided).
• Feedback refinements: superadmins can edit or delete submitted comments, with an edited marker for both sides and a full audit trail; and the navigation rail's duplicate feedback entry is gone — the bug icon at its foot is the single way in, with existing pages and links still working.
• A schedule change that accepts a follow-up without naming the customer record it attaches to is rejected immediately with a clear error, instead of consuming a review confirmation and failing afterwards.
• Navigation icons reveal their names at once on hover or keyboard focus, and the Platform console drops its back button in favour of an anchored “Return to Atlas” link with a calmer header.
• Skill review can compare a proposed method against a control, pairing diagnostic evidence with cost and latency context, and explains in plain language when a comparison or instruction is withheld.

v0.20.0 — September 8, 2026

• Platform Feedback triage gains filtering, search across titles and descriptions, multi-select with confirmed deletion and safe attachment cleanup.
• Organisation labels can be applied by numeric rules checked against reviewed customer evidence.
• Schedule task-start reminders, protected against stale concurrent edits.
• Knowledge sources can link to customers with capacity limits and deletion cleanup, and every customer-facing knowledge claim carries versioned, retractable source-use approvals.
• Knowledge Policies and Workflows publish as immutable releases — current-release reads, typed rules with example preview pinned into release history, and operational evidence scopes in the Sandbox and Assistant.
• Sales & CRM configuration returns to Settings, with retired links redirecting and all permissions unchanged.
• Receivable follow-ups and annual cooperation evidence record confirmed observations without combining currencies; transactions and quotations carry customer and internal part numbers; and totals stay separate per currency.
• Governed model attempts gain traceable identities and usage-cost reporting by purpose.
• The Account Manager reuses more of each day's context, communication review recovers more safely, and customer-draft corrections guide with explicit external-party attribution; a method-teaching suggestion is reserved before any model call.
• The Knowledge Sandbox conversation and composer are redesigned; policy pagination and Today opportunity amounts for sales reps are fixed; and the duplicated Knowledge feedback navigation entry is removed so Feedback has one home.

v0.19.0 — September 7, 2026

• Eligible leaders can safely preview the Manager and Sales representative roles, with the real signed-in identity and organisation preserved.
• Unreleased AI-team surfaces follow the organisation's Feature visibility setting.
• The Feedback issues workspace arrives with severity, status, dedicated navigation, detail pages, replies and status notes.
• Customer detail tabs, Information review, Knowledge Feedback and the Library get denser, easier-to-scan layouts, and communication review recovers more reliably.
• Prepaid wallet charges keep a defined order behind a no-negative-balance gate.
• Enterprise Memory navigation arrives, Outreach playbooks relocate, and Governance health reports bounded memory readiness.

v0.18.0 — September 6, 2026

• The notification bell warns before a follow-up SLA lapses, naming the rule, the customer and the hours left.
• Industry labels group collapsibly with clearer usage counts, and proposing an existing industry under another Chinese name no longer creates duplicates.
• The Library's Recent section becomes a compact chip list, document detail shows preview and extracted content side by side, and long keyword lists move to their own page.
• Every organisation can claim the Owner role when none exists, instead of only the originally seeded account.
• Pilot usage gains an Outreach outcomes section counting prepared, sent, replied, dismissed and expired outreach per trigger.
• Knowledge Policies and Workflows open read-only first, with an explicit Edit action.
• Fixes: Billing respects the roles matrix; platform operators see a console link instead of a dead delete option; the decision queue shows a real empty state; contact social accounts carry a clear limit; same-file concurrent uploads report a conflict instead of an error; and assistant tool actions have complete English and Chinese names.

v0.17.0 — September 5, 2026

• Today becomes a daily workbench: a Dashboard for revenue and pipeline analytics, a “Your month” goal tracker, and a “Needs your decision” queue for reviewing AI-prepared work.
• New organisations choose a starter template (manufacturing, consumer electronics trading, B2B distribution), and admins configure the industry list — add, rename, reorder, hide — while reps can propose new industries for review.
• The AI run/trace browser moves into Settings under “AI & agents”.
• Knowledge Sandbox questions can be scoped to specific documents or folders, conversations persist for later review, and repeating or changing an answer rating no longer creates duplicates.
• The Library gains clear searching/no-results/failed states, a Recent documents view, better keyword extraction with re-extract and “improve understanding”, sorting, load-more retry, multi-select bulk move/delete, drag-and-drop onto folders, and honest failures instead of silent empty results.
• Knowledge gains a Policies page where everyone reads the organisation's AI-answer policies and managers or admins publish them, plus a Workflows page of step-by-step procedures.
• Platform admins can multi-select and delete organisations (the protected home organisation excepted) and always see a Return to Atlas link.
• Customers gain a Partner type, website, background notes and social accounts; reps create personal labels alongside shared ones and see only their own customers while admins and managers see all; the customer-type filter returns; type and class categories become tenant-configurable; and label chips edit in place wherever they appear.
• Superadmins can reassign a feedback handler and a Bug Report type joins the form; Chinese, Japanese and Korean PDF extraction works through the CRM intake; failed Office previews no longer claim AI readiness; conversations show run status; and the import label for customer class is fixed.

v0.16.0 — September 4, 2026

• The Owner role arrives with a company-wide Full view dashboard — goals, per-salesperson performance and AI-generated suggestions.
• Organisation deletion requires typing the exact name, role permissions become editable, and Users and Permissions merge into one Organisation settings page with clearer Billing and Agent Quality pages.
• Feedback submissions take images, videos and documents plus a handler choice, a detail page shows the full thread with status controls, and status changes or summary waits no longer blank or hang the pages.
• The Knowledge Sandbox becomes conversational, with source citations showing which documents backed each answer and ratings for answer feedback.
• Word, Excel and PowerPoint files read directly into the Knowledge Library, and previews show an honest “ready for AI” message instead of garbled text.
• Settings becomes a full-screen workspace with an organisation-level AI model policy — Atlas-managed or your own model keys.
• Assistant messages render headings, tables, links and code properly, responses get faster with less thinking time, uploads finish before a message sends, and Chinese “business services” and “retail” industries are recognised.
• Customer import accepts Chinese country and industry names, and bulk delete can select every matching customer across pages.
• Platform and Settings move language and return links into the sidebar, “Billing and payments” becomes “Billing”, and superadmins can be assigned tasks again with real errors instead of silent failures.

v0.15.0 — September 3, 2026

• The AI assistant becomes a floating chat window that stays usable while other dialogs are open, with popups unified so they can appear above it.
• The Knowledge Library gains full-organisation document search showing each match's folder path.
• Customer import no longer flags valid Chinese industry and country names (such as 中国, 电子制造) as needing correction.

v0.14.0 — September 2, 2026

• Customer lists and profiles show contact-frequency status (Hot, Active, Cold) and a customer-type badge instead of the old class and lifecycle columns.
• Knowledge Library folders and documents get shareable, bookmarkable links with correct browser-back behavior, and mistranslated Chinese class labels (Lead, C, B, A) are fixed.

v0.13.0 — September 1, 2026

• The customer profile is redesigned — cleaner identity card, key metrics on every tab, simplified actions — and labels display in consistent colors with a redesigned Labels management page.
• New-customer creation becomes numbered sections with a live progress checklist, and import returns as a dialog over the list so filters and scroll position survive.
• The Knowledge Library unifies folders and files in one browsable list with clear AI-readiness status and direct open and delete.

August 2026

v0.12.0 — August 31, 2026

• Launched the Platform control tower for company admins — usage metering, billing, credit policies and financial reporting in one place.
• Added the Knowledge Library: browse your organisation's documents and folders from the sidebar.
• Organisations can create their own customer labels — tag categories and values — managed from a new Labels page and shown on each customer's profile.
• The assistant drafts personalised outreach emails for customers who haven't been contacted in a while, ready to review and send, and now streams long replies as they're generated instead of timing out.

v0.11.0 — August 28, 2026

• Creating multiple customers at once lets you remove unwanted entries, and “New customer” and “Import” become pop-up dialogs over the customer list instead of separate pages.
• Tasks unifies its List and Month views into one screen with shared search and filters.
• Contacts can have profile photos, and the Assistant panel works properly on mobile instead of squeezing the screen.

v0.10.0 — August 27, 2026

• Customers classify automatically (Lead, Class C, B, A) and upgrade when they win their first deal; the assistant fills in country and industry when creating a customer, and those pickers become searchable instead of long scrolling lists.
• Fixes: customer deletion reliably removes the customer and all related records in one step, and clicking a sidebar link while a new-customer draft is open now actually navigates instead of only changing the address.

v0.9.0 — August 26, 2026

• Create opportunities directly from the Opportunities list and Sales Pipeline pages.
• The Today page becomes a metrics and weekly-calendar workspace, adding a notification bell with a reminders popover and a 12-month revenue trend chart.

v0.8.0 — August 25, 2026

• My Tasks is redesigned into a simpler table with clearer filters, and clicking a task opens a read-only detail first, with a separate edit action.
• Primary contacts can no longer be deleted directly — another contact must be promoted first — and contact and customer info stays in sync between the two views.
• A read-only Subscription page shows plan, seat usage and contract term.
• Fixed customer deletion so it always removes all related records — contacts, deals, notes and more — instead of sometimes failing with an error.

v0.7.0 — August 24, 2026

• One account can belong to and switch between multiple organisations, and creating a new customer now requires only a company name — every other field is optional.
• Fixes: the “needing follow-up” count correctly includes customers in critical health regardless of type, and a customer's primary contact always shows on the company-wide Contacts page.

v0.6.0 — August 23, 2026

• The old Smart/Manual switch is retired: Actions leads the navigation with the Assistant opening directly, and duplicate dashboard and chat action buttons are removed.
• Starting a new chat always shows a clean conversation instead of old messages.

v0.5.0 — August 22, 2026

• Every customer shows a profile status (Captured, Actionable, or Enriched) with what information is still missing, plus a review flow to resolve missing-information alerts, including marking one as not needed; reminders now appear only when they actually block an open deal or an overdue follow-up.
• The daily “Today's Work” briefing is redesigned with a friendlier summary of what needs attention, and its summary text and empty states now match the real counts.
• Priority cards suggest the next action to take, and “Open Conversation” starts the chat with a relevant opening message already written.
• Customer pages show AI-generated insights with a refresh button and timestamp instead of static content, and documents preview in place instead of always downloading.
• Fixed a production blank-page issue affecting some visitors.

v0.4.0 — August 20, 2026

• A quick-search popup (Ctrl/Cmd+K) jumps straight to any customer account, and the Today page gains a “what's changed” activity feed of recent customer account activity.
• The assistant prepares a draft reply for review after a customer responds, and suggests a follow-up task when a deal's stage changes or a new opportunity appears.
• Approved draft replies can be sent to the customer by email with one click.

v0.3.0 — August 18, 2026

• Added the Opportunities page: every open deal across all customers in one table, with filters by stage, owner or tier and sortable columns.

v0.2.0 — August 10, 2026

• Added a “Continue with Google” option so people can sign in with their work Google account instead of an email code.

v0.1.0 — August 8, 2026

• Added the first sign-in flow — enter a work email, get a one-time code by email, and land inside the app.