Changelog
What's new in TellWang. We ship continuously; this is the plain-language record of the changes that reach you. Dates are when the change went live.
September 7, 2026 — Unapplied Fixes open from their full row
Every Fix card in a Site's Fixes tab now opens its retained detail, including a Fix that is still waiting for approval. The shared history table also keeps its row action on each actual Fix instead of its header. Pointer selection, Enter, and Space open the same durable Fix address without applying it.
September 7, 2026 — AI visibility evidence can become content Fixes
A Search and AI visibility run now checks every approved buyer question against the Site's current relevant pages, its mention and citation evidence, and exact titles cited for tracked competitors. It proposes an update when a suitable page is incomplete, or a new page when the Site has no suitable answer. Competitor pages can guide topical coverage and answer structure, but Lollipop writes an original outline and does not copy wording, branding, or claims. Facts that only the owner knows appear as questions inside the Fix; the related change waits for the saved answer. An approved new page remains a Content draft for review and is never published automatically.
September 7, 2026 — Current plan pricing joins the Lollipop homepage
The Lollipop homepage now places a pricing chart between its product comparison and FAQ. Free, Pro, Scale, Scale + HA, and Enterprise prices and allowances come from the public plan catalog rather than duplicated page copy. The chart shows managed Sites, parallel optimizations, monthly model allowances, database failover, and signup mode. If the catalog cannot be read, the page shows an unavailable state and retry instead of a stale price.
September 7, 2026 — Audits hand off to fresh Optimize runs
A Site audit that has no Fixes now points to Optimize, where its evidence can become scheduled Fix recommendations. Run now and recurring Optimize checks refresh Site audit evidence older than 24 hours before work begins, using the owner's approved buyer questions and disclosed repeat-audit allowance. A Site that has never completed question review stops there instead of inventing questions. The Share report preview also lists every distinct buyer question in an accordion instead of showing only the first question. Expanding one reveals every frozen answer excerpt and its retained links under Answers and citation sources.
September 4, 2026 — Copied audit links keep their messaging preview
Opening a public audit no longer leaves its metadata-free application address in the browser. The address bar and the report's Copy link action use the same versioned share URL that carries the score thumbnail, audit title, and quick verdicts for messaging apps. Previously copied /report/… links now redirect to that metadata page at the edge, so messaging crawlers do not need to run the app first.
September 4, 2026 — Site path redirects return real 301 responses
Redirect rules added in Lollipop now route the exact old Preview path to its destination with a real 301 response and preserve its query string. A staged release checks both the status and destination instead of accepting an unrelated homepage response. Existing static Sites with redirect rules receive the corrected routing configuration when the service starts.
September 4, 2026 — Receive payment is separate from Billing
Each Project now has a Receive payment page for its customer revenue, Stripe readiness, and payment connections. An Owner, Admin, or Billing member can create a fresh single-use Stripe setup link there whenever onboarding was interrupted or the old link expired. Lollipop plan charges, prepaid funds, usage, invoices, and receipts remain under Billing.
September 3, 2026 — Shared audits keep their citation links
A newly shared Lollipop audit now lists the retained citation sources below each sampled answer, with the source title, domain, and clickable public link. Answers whose source returned no citation say so. Unsafe, local, duplicate, credential-bearing, and platform-hosted links are not exposed.
September 3, 2026 — Member re-adds and DeepSeek answers are more reliable
Removing and re-adding the same member now works regardless of email-letter casing and keeps the existing account identity. New DeepSeek AI Answers treat the saved market as authoritative, answer without asking which same-named place the buyer meant, and retain exact public citation links returned in the answer without counting uncited search results.
September 2, 2026 — Naver visibility and complete competitive mentions
AI Answers now checks native Naver AI Briefing and DataForSEO Naver Organic results alongside its current sources. Search evidence remains separate from AI-answer rates. The Competitive mentions section lists every occurrence of an explicitly tracked competitor, with the buyer question, source, domain, and evidence excerpt, without guessing that every external citation is a competitor.
September 2, 2026 — Buy domains and manage their DNS in Lollipop
The Domains page now separates domains purchased through Lollipop from externally registered hostnames. Check availability, review the exact charge today and current yearly renewal estimate, choose automatic renewal, and confirm before the Account balance is charged. Purchased rows show their registrar expiry date and renewal state. Active purchased domains include DNS record list, add, edit, and confirmed delete controls; external hostnames remain in the separate Site connection and verification flow.
September 2, 2026 — Active Preview releases no longer idle out
A Git deployment or browser check now keeps its Preview Wok awake for the complete operation and refreshes the idle clock before it begins. The hourly staging cleanup rechecks both immediately before stopping containers. A Preview that was selected from an older idle snapshot can therefore remain online when new work has started, instead of returning a 502 seconds after a successful check.
September 2, 2026 — Email attachments and inbound routing fail visibly and recover safely
Outbound email now carries Base64 attachments from the public or per-Wok send request through its durable queue to Resend or SES. Both content_type and contentType are accepted, while malformed or oversized files are rejected instead of disappearing behind a successful response. TellWang also keeps mx.tellwang.com pointed directly at the mail origin with DNS proxying off and reports a semantic diagnostic if that shared inbound route drifts.
September 1, 2026 — DeepSeek joins AI Answers
Each Site can now sample its approved buyer questions through DeepSeek as well as ChatGPT and Google AI Overview. DeepSeek performs a live web search and keeps its answer excerpt, mention result, citation result, and provider receipt separate. Ordinary successful question checks cost $0.01 each; pre-account and first-Site audits remain included.
September 1, 2026 — A moved Site opens in its destination Project
Moving a Site inside an account now selects the destination Project, keeps the Site open, and replaces an older account-only address with the exact Account → Project → Site link. The Site's Fixes, activity, leads, AI-answer questions and evidence, audit-question drafts, creative work, and Site-bound optimizations move in the same transaction, so the source Project cannot retain a partial history.
September 1, 2026 — Every Project has its own canonical link
Switching Projects now changes the address even when both Projects live in the same account workspace. Refresh, browser back and forward, bookmarks, and shared links restore the exact Project and reject a Site that belongs to another Project. Existing account-only links remain valid and upgrade to the canonical Account → Project path after loading.
September 1, 2026 — Move this site stays available before the first audit
Site management no longer disappears while a Site is waiting for its first audit. Owners and admins can open Move this site from Overview in every audit state. The destination list matches reachable Projects: native Projects move inside the account, while eligible managed Sites can move to a legacy Project in another workspace the same person controls.
September 1, 2026 — First Preview builds remain visible after you leave
The no-site Review now shows every frozen section, marks unresolved questions and permission choices, and uses a second explicit build confirmation. A running build retains six truthful stages and owner requests across navigation. Success shows the verified Preview summary, checks, and creative brief without publishing; failure retains its evidence and offers either a same-brief retry or a new reviewed revision tied to that exact Site, even when several Site builds have stopped.
September 1, 2026 — Design styles can be previewed, compared, and refined
The no-site flow now shows why one pinned style is recommended and renders the owner's real brief at desktop and mobile widths. Owners can search and filter the verified local catalog, inspect production guidance and warnings, select or undo a style, and request an allowlisted refinement with an exact before/after token-group record. Catalog recovery never executes a third-party site or silently changes an existing selection.
September 1, 2026 — Look & feel keeps owners in control
The no-site flow now accepts a written direction, optional mood signals, up to three preferred colours, or “Let Lollipop choose.” Inspiration images can be uploaded or dropped, remain local until “Extract colors & mood,” and show filename, size, private-analysis limits, exact recovery actions, and a keep-or-remove-palette choice. Results use five semantic roles with provenance, locks, drag reordering, contrast checks, reversible accessibility fixes, and a private written-brief fallback when no image is used.
September 1, 2026 — Each Site's AI Answers leads back to its audit
AI Answers keeps separate buyer questions, samples, mention rates, and citation rates for every Site. Before the included first audit or a paid repeat samples an answer engine, Lollipop infers one to six Site-specific questions and asks the owner to edit and confirm them. A name copied from the question into an answer is not counted as a mention; matching source titles and domains remain separate citations. Google AI Overview uses a canonical DataForSEO location and retains partial-provider receipts instead of hiding a missing engine. The selected Site remains in the page address across refreshes and bookmarks. It also has a direct audit action that opens question review, live progress, or recovery as appropriate. A watch-only Site exposes Overview and Fix review, while managed Site, Content, and Redirect controls remain unavailable until import creates a Preview.
September 1, 2026 — New Sites keep an owner-approved page plan
The “I don’t have a site yet” funnel now keeps purpose planning separate from visual design. Custom Sites, visitor actions, and functions require a plain-language description; Calendar, Gallery, file upload, and the other promised function types are available. After private analysis, the owner can rename, add, remove, and reorder proposed pages, answer blocking questions, save them, and explicitly continue to Look & feel. At least one named page and path must remain in the frozen brief before Preview work can start.
September 1, 2026 — First login points to the exact Site and audit
After adding a free audit, Lollipop now stays on Sites, identifies the exact Site card, and offers separate actions for the Site and its Audit report. When the Site already exists, the review names findings that match an open Fix; confirmation links those Fixes instead of creating duplicates. Audit links retain a focused report URL and the full Project → Sites → Site → Audit report breadcrumb. Choosing “Not now” keeps “Add your free audit” in the account menu until the claim expires. If the one-time Site pointer is dismissed after attachment, Help → “Where is my audit?” restores it without replaying or rerunning the audit.
New-account signup now verifies email and phone before asking the person to create or choose a Project, so an unverified signup cannot create workspace records.
September 1, 2026 — Removing a watched Site no longer resets its first audit
A normalized watched domain now has one durable identity across the Account. Removing and re-adding it restores the same Site, audit attempts, history, Fixes, events, and included-audit ledger in the selected Project. Older removed Sites are repaired from their retained audit snapshot when possible, so protocol, www, paths, Project moves, and removal cannot mint another included audit.
September 1, 2026 — Included and paid audits share one usage history
The public audit now keeps “Included · $0.00” visible while it runs, if it stops, and on the completed result. Billing → Usage can be filtered to Audits or Email. Audit rows name the Site and completion date. A claimed pre-account audit and the first completed audit for every Site show their reason and $0.00; repeat audits show the dollar charge for successful answer checks plus their model-funding terms.
August 31, 2026 — Wang drafts replies only after a credit estimate
Open an Inbox message, describe what the reply should achieve, and get a zero-credit estimate before confirming one metered model request. Wang uses only that message and approved Project knowledge, treats email instructions as untrusted, and returns an editable draft without sending it.
August 31, 2026 — Billing can open directly on Email usage
A verified publish now opens a durable outcome window for every newly published Fix in the release. The Fix retains its expected metric, direction, horizon, pre-release Site/audit reference, and first check time. A Site-scoped Optimize report can explicitly attach measured or inconclusive evidence, and can link its first new draft Fix as “Review next opportunity”; the server rejects unrelated or cross-Site links and never executes that next Fix automatically.
Delivery health now opens Billing with an Email filter. All Sites also shows a failure-first Site breakdown with current receiving, forwarding, sending, automation, last-success, pending/failed, and monthly-unit evidence. Each row opens the exact Site recovery screen, partial failures retain the other Site results, and a role without Billing access sees unavailable totals instead of a false zero. The ledger separates received, sent, and forwarded units by Site, quantity, charge state, and receipt; it explicitly keeps email delivery and model-credit usage apart.
August 31, 2026 — Sent replies stay with their Inbox thread
Replying from the Email workspace now returns to the source Inbox message and appends the sent copy, delivery state, and receipt. The sent copy is stored in the same reservation transaction as the idempotent email delivery, so retries cannot duplicate it and provider-payload cleanup cannot erase the thread.
August 31, 2026 — Receiving and sending domain health are separate
Email domains now show independent Receiving and Sending states: not started, DNS pending, verified, or error. View DNS separates the required inbound MX from sending-provider records and shows what Lollipop has detected. Safe recovery guidance never exposes a stored provider error, and repeated verification does not consume model credits.
August 31, 2026 — Forwarding failures explain the next retry
Pending and failed forwarding deliveries now open into a safe detail view with the recipient, attempt count, documented retry limit, next retry time, stable error code, and the next action. Automatic retry state remains visible even after forwarding is stopped. Message bodies and raw provider errors are never exposed in the diagnostic.
August 31, 2026 — Inbox read state and filters persist
Inbox rows now distinguish unread and read messages with durable server state. Opening a message marks it read, Mark unread reverses that state, and the current Inbox can be narrowed by unread state, Site, recipient domain, date, or safe metadata search. Clear filters restores the selected Site or All Sites scope. None of these controls sends message content to a model or consumes model credits.
August 31, 2026 — Redirect mistakes can be removed safely
Redirects now have a confirmed Remove action. The rule is removed from Preview as a new staged release, while Published remains untouched until the normal review and publish flow. This completes the add, inspect, verify, and recover lifecycle without requiring a release rollback.
August 31, 2026 — Inbox attachments and original messages are downloadable
An opened Lollipop Inbox message now provides an authenticated download for each decoded attachment and, when retained, the exact original .eml. Downloads stay private, do not put the account key in a URL, do not cache, and force untrusted files to download rather than execute in the dashboard. Missing or malformed raw mail fails closed instead of returning partial bytes.
August 31, 2026 — Lollipop completes safe Published checks after cutover
Publishing a verified Lollipop Preview now carries its release through the checks the product can perform safely: it loads the Published homepage and calls the non-mutating health contract on protected form functions. The release remains bound to the exact reviewed Preview receipt, and arbitrary customer functions are never invoked automatically. This removes the indefinite “Publishing…” state for sites whose generated forms were already verified in Preview.
August 31, 2026 — Lollipop Email gains Project scope and delivery-focused navigation
Email now opens with the six customer workflows named directly: Inbox, Compose, Forwarding, Automations, Domains & setup, and Delivery health. All Sites aggregates readable Inbox metadata across the current Project, while every delivery-changing action requires a specific Site. The last scope is remembered, changing Site clears unfinished recipients and settings, forwarding has an explicit stop confirmation, and health cards lead to the exact setup or recovery surface.
August 31, 2026 — Optimize begins with business outcomes
Optimize now asks what the customer wants to improve before exposing implementation details. It offers leads or bookings, search and AI visibility, clarity, speed or accessibility, competitor monitoring, and a custom outcome; marks one audit-backed recommendation; scopes the work to one Site; asks which evidence it may use; explains cadence and action boundaries; and ends on a plain-language review with advanced details available on demand. The wizard saves each person's answers to their account, confirms before replacing a dependent outcome, and lets them edit an exact answer before returning to Review. A new setup or reviewed edit can be closed for later or explicitly discarded; discarding an edit does not change the active plan. “Ask Wang to help me choose” discloses an unmetered estimate first, then uses approved Project evidence only after a separate confirmation; its recommendation includes a reason and confidence and is ordinary metered Optimize work, never part of the first-audit exemption. Starting freezes plan version 1 and immediately queues its first durable run. Reviewed edits append a new version. Daily, weekly, fortnightly, monthly, and verified after-publish recurrence runs automatically under the human owner's current role; Run now, Pause, Resume, and retained-history Remove remain available. A run that needs an answer or permission reappears under Waiting for you with Send reply, Allow once, and Reject. Optimize never publishes.
August 15, 2026 — Aggregate audience analytics and the private Live Wok Data Room
Analytics now opens on cookieless Audience analytics with aggregate country and region, device, language, campaign and source, privacy-safe clicked-element labels, conversion outcomes, pages, and daily traffic. Session replay remains available as the adjacent submenu. Raw IP addresses are discarded after local geography lookup; age and gender are not collected or inferred.
TellWang platform operators also have a private Live Wok Data Room covering every active production Wok, whether or not it has a connected domain. It combines optional-domain health and aggregate audience breakdowns with cumulative and latest Wang/Agent Cell build time and tokens. Cross-customer reads remain platform-admin gated and audit-logged.
Coming next — The chat composer shows its send shortcut
A compact key cue now sits beside Send: Ctrl ↵ on Windows and Linux, or ⌘ ↵ on macOS. The existing keyboard behavior is unchanged, and plain Enter still creates a new line.
Coming next — Managed DNS propagation no longer looks like an abandoned run
When TellWang has already published a managed domain's records and the required one-shot verification returns a future retry time, Agent Cell records that result as an expected external wait. lollipop finish-run gives the Cell an explicit terminal action: after the server validates that wait, completion cannot force another turn or verification call. The command is deliberately refused for external DNS or ordinary unfinished work. If the rest of the request is complete, Wang's truthful pending-activation answer is accepted and chat does not show “stopped before finishing.” Raw tool output never enters completion evidence; only revalidated operational labels do. Domain verification now checks an independent public resolver alongside the host resolver, and managed DNS reconnects repair an existing orange-cloud CNAME to the required DNS-only mode.
August 12, 2026 — Agent Cell reasoning and public text are separated
Agent Cell chat routes OpenCode reasoning and trusted internal continuation guidance into a collapsed Thinking accordion and shows public model text as ordinary Wang messages outside it. OpenCode's echo of a customer or harness prompt is never public narration. Both streams survive refreshes and stopped runs, while the working stopwatch remains visible. Existing conversations from before typed streams keep their historical mixed rows inside Thinking because TellWang cannot reliably split old content after the fact. The independently accepted final answer remains a separate immutable message.
August 11, 2026 — Agent Cell answers have a hard publication boundary
Agent Cell chat publishes only the final answer that passes TellWang's independent completion check. Publication and status=done commit together behind an explicit marker; older mixed rows without that trustworthy boundary are hidden wholesale. Safe status, file, release, verification, queue, and question progress remains visible while Wang works. When an owner explicitly requires a named tool or tellwang command, Wang must record a successful tool call; pasting the command cannot count as execution. Static Wok guidance maps frontend/index.html to / and nested index files to their matching URL routes, so two valid pages are not mistaken for duplicate checkouts.
Coming next — Wang stops inspection loops early
A coding turn that keeps rereading files without editing, testing, or naming a real blocker is interrupted after a bounded number of calls. Wang gets one focused continuation in the retained app checkout; if it repeats the loop, the run stops visibly while the chat, checkout, and completed work remain available for the next request. Missing base-app files are treated as absent instead of searched repeatedly. Agent Cell model credentials also converge across TellWang control-plane replicas, so starting another chat cannot invalidate a key already mounted in a healthy worker.
Coming next — App chats keep their coding workspace
This release candidate makes choosing an app return to its latest focused chat instead of silently creating another conversation. The associated coding workspace, cloned repository, branch and unfinished edits remain attached to that chat across later messages, refreshes, worker replacement and idle scale-to-zero. Organization chat also keeps a run-level app focus: a named app resolves to its Preview, while a short follow-up such as “go ahead” inherits that focus without permanently binding the channel. TellWang snapshots and prepares the retained Preview checkout before Wang starts, so it no longer has to search an empty workspace or guess how to fetch the app. Tool progress now preserves safe target details and marks invalid or unavailable tool calls as failures instead of successful ticks. The completion check receives the recent user goal, so an acknowledgement cannot erase the work it authorized.
August 9, 2026 — Recordings no longer collapse to a white frame
A browser can briefly report a zero-width viewport while a page becomes visible. Those recordings still contain the complete page, but the replay frame previously inherited that zero and looked blank. Playback now reconstructs older affected recordings at a safe viewport, and capture repairs an invalid dimension before upload so new recordings keep a usable viewport.
August 4, 2026 — Session replay across the workspace
Analytics is now a first-level workspace tab for teams with many Woks. Search the complete Published Wok list, choose one, and enable recording with an owner privacy acknowledgement. The list is filtered on the server to the last 24 hours, three days, or seven days, with pagination for busy Woks instead of silently stopping at 100 recordings.
Each session carries an entry page, duration, viewport, event count, completeness, and deterministic client-error, rage-click, and dead-click signals before you open the player. Wang can inspect the same summaries and change the owner-controlled setting without receiving raw DOM events. Each Wok remains off by default and respects Global Privacy Control; TellWang records every eligible human session after an owner/admin opts that Wok in. Automated browser checks are excluded, private lifecycle markers preserve idle reading time, and the retired Do Not Track signal does not block recording. Custom domains proxy the same recorder and ingest paths as TellWang Wok domains, and seven-day cleanup is enforced by the control plane.
July 30, 2026 — The workspace stays; only Wang chat changes
The dashboard keeps its familiar Feed, Email, Decisions, Apps, Documents, Automations, Replay, Published/Preview, and split-preview workspace. Feed is now the navigation label and default view. Inbound customer email can still be opened and replied to directly. The unreliable Wang conversation area is the only part replaced: it now shows connected coding clients, the Streamable HTTP MCP address, and editable, Preview-safe prompts for Claude Code, Codex, Cursor, or another MCP client. It does not start a Wang run. Existing Wang records and APIs were not deleted. The bottom-left account avatar opens a menu; signing out requires choosing the labeled Sign out action.
July 30, 2026 — App choices stay attached to their app
When you choose Edit app and select a Wok, the new conversation now belongs to that app instead of the general workspace. The chat shows the selected app before you send, and the same scope follows app-targeted GEO, domain, email, marketing email, and MCP server work. Opening an existing thread returns to the normal chat view, shows a loading message while its history arrives, and opens at the newest reply instead of leaving the app builder over a clipped transcript.
July 25, 2026 — A new TellWang workspace
The customer dashboard is now a purpose-built React workspace instead of features layered onto the old administration screen. Chat and Apps lead the experience. App chats contain focused threads; Feed and Decisions behave like conversations; Documents owns files, storage, generated assets and governed knowledge; Automations and privacy-masked session replay stay tied to the app they belong to. Apps open the published version first, with Preview available only when a staging version exists. The previous administration screen is no longer served; moved settings live in the React workspace, and remaining administrative work uses the documented API.
May 31, 2026 — Teach Wang your business
Wang can now learn the specifics of your business. Give it a knowledge base — your policies, product facts, the answers your team repeats all day — and Wang draws on them instead of guessing. Add standing instructions ("quote prices in CAD," "send any refund over $100 to a human first") and Wang follows them on every request, both in the dashboard chat and through the tools you've connected from Cursor or Claude Code. Search tolerates typos, so "refnd polcy" still finds the refund policy.
May 30, 2026 — Wang reads your incoming email
Mail sent to your Wok's address now lands somewhere Wang can see it. Ask Wang to summarize what came in overnight, pull the order numbers out of a thread, or draft a reply — your inbox is part of the same conversation, not a separate tab.
May 30, 2026 — A rebuilt dashboard
The dashboard is now organized around your Woks: a card for each, plan and credit usage along the top, and tabs for Wang, API keys, domains, email, and activity. Less hunting, more doing.
May 29, 2026 — Bring your own model key
The built-in model gateway can now point at your own provider account — OpenAI, Anthropic, Groq, and others. Your calls run on your key and bill to your account; TellWang doesn't meter them a second time. Keys are encrypted at rest and never shown back to you in full.
May 29, 2026 — Run code the moment your data changes
Functions can now fire on a database change. A new row lands and a function runs — to send the confirmation email, call a webhook, write an audit trail, whatever that moment needs. No polling loop to babysit.
May 29, 2026 — Receive email, not only send it
Your Wok can take delivery of email now, not just send it on your domain. Incoming messages are stored and queryable, ready for a function — or Wang — to act on.
Sizing this for your team, or want something that isn't here yet? Talk to the founders →