Alynivo – Live-Chat & KI-Chatbot für den Kundenservice

Änderungsprotokoll

1.0.0

• First stable release. Nothing was removed and nothing behaves differently — the version number simply catches up with the plugin. Alynivo has been running on live sites since July, the agent inbox and the iPhone app are in place, and the setup, settings and data controls have settled.
• From here on, releases are collected rather than shipped one change at a time.

0.22.41

• The assistant title and welcome message now show the built-in wording as a placeholder instead of a filled-in value. On a site without a shop that wording was always the neutral one your visitors saw („Assistant“), but the settings screen showed the shop version („Product advisor“) as if it were active.
• Setup suggests your site name as the assistant title, so the chat greets visitors with your own name.
• Hints that assume a shop — the teaser bubble example and the „next step“ hint — now only mention products when WooCommerce is actually installed.

0.22.40

• Setup now reads your home page and services page once and offers a first draft for the assistant’s knowledge, so you do not start with empty fields. It only ever fills what is empty, shows you the text before it is used, and saves nothing until you finish the setup.
• Prices found on your pages are deliberately left out of that draft.

Older releases are listed in changelog.txt, which ships with the plugin.

0.22.39

• The assistant can now read the page the visitor is looking at, so it answers about that page instead of only knowing its title. Up to 1,200 characters, never stored, and switchable off in the settings.
• Fixes the keyboard focus ring, which sat at 1.5:1 contrast where 3:1 is required.
• Screen readers now announce a new visitor message in the agent inbox.
• Corrects the privacy statement: it claimed the managed service never stores message content, which is not true once Live Chat is on.

0.22.38

• The knowledge context and persona fields no longer use a monospace font; they hold prose, not code.
• The overlap hint now only appears for topics that are genuinely hard to tell apart, instead of every pair sharing two words.

0.22.37

• The settings now point out which of your own topics are too similar to tell apart, naming the words they share. Overlapping topics are why the assistant sometimes has to ask back instead of answering, and until now nothing showed you that.

0.22.36

• With Managed AI active, a question that matches several of your own topics equally is now answered by the assistant instead of being met with a „which one do you mean?“ menu. The free version keeps the menu, where there is nothing better to offer.

0.22.35

• The launcher now recognises on its own whether your logo carries its own background, and lets it fill the button instead of sitting inside a white ring. You can still override the decision either way.

0.22.34

• Adds a launcher background setting: automatic (unchanged), brand colour, white, or a colour of your own.
• Adds a „logo fills the launcher“ option for logos that carry their own background, so the mark no longer sits as a smaller tile inside a white button.
• Gives the closed button a light edge instead of the soft shadow when its background is dark, where the shadow was invisible.

0.22.33

• Reworks the chat sheet on phones: the page stays visible behind a dimmed backdrop instead of a full-screen takeover, and a larger gap at the top makes clear the chat is a layer over the site.
• Adds two more ways to close the chat on a phone — pull the sheet down by its handle, or tap the strip above it — alongside the existing close button.
• Softens the close button so it no longer sits in a grey circle, while keeping its full touch target.
• Stops a host theme’s section spacing from leaking into the chat panel, which left an empty white block above the header on themes that indent every section.
• Keeps a deliberately disabled enquiry form disabled when the setup assistant is run again.

0.22.30

• Simplifies the Managed AI and Live Chat settings copy so site owners understand what each hosted feature does before enabling it.

0.22.29

• Fixes the self-service WordPress connection flow so the one-time code form never falls through to the raw WordPress settings screen.
• Returns admins directly to the Managed AI & Live Chat screen after a connection attempt.

0.22.28

• Added a subtle setup restart action directly to the Alynivo dashboard setup card.

0.22.27

• Added branded HTML account activation e-mails with a clear clickable button while keeping the plain-text fallback.
• Made the WordPress one-time connection code field visible directly in the Managed AI & Live Chat settings screen.
• Removed duplicate single-site labels in the Agent Inbox header.

0.22.26

• Added self-service website pairing for Alynivo accounts: owners can create a website, generate a short-lived WordPress connection code, and pair the site without emailing permanent credentials.
• Added the WordPress-side pairing flow for Starter and Pro connections.
• Improved the Agent Inbox mobile reply composer with a shorter placeholder and more usable typing area.

0.22.25

• The widget now appears in the shop even when Alynivo was activated before WooCommerce. Activating wrote the page visibility into the database verbatim — including the four zeroes the default sets while WooCommerce is absent, and a stored value always beats a default. Anyone who tried the chat first and installed the shop afterwards had no widget on any shop, product, cart or checkout page, permanently, with the dashboard reporting „Live“. Activation no longer freezes those settings, and existing installations are corrected once — but only where the settings still look exactly as activation left them.
• The dashboard says „Not in the shop“ instead of „Live“ when the widget is switched on but hidden on every shop page.

0.22.24

• „C’è un umano?“ now reaches a person. The Italian for „is there a human?“ was deliberately left out because it shares two words with „Sei un umano?“ — „are you a human?“ — which asks the assistant about itself and deserves a different answer. The two are now told apart, so both get the right one. An explicit request still wins when a message contains both.

0.22.23

• Your own topics can now be written in every widget language. Until now only the frame around the click-through was translated: on a French page the heading said „En quoi puis-je vous aider ?“ and the topic buttons under it stayed in the language you typed them in. Each topic now has an optional field per language, right underneath it. Leave a language empty and the original text is used; a language is only taken when both the question and the answer are filled in, because a translated question with an untranslated answer sends the visitor somewhere they cannot read.
• A mirrored reply now reaches the Inbox in the language the visitor saw. The mirroring call runs through a REST request where a multilingual plugin often cannot tell which page the visitor was on, so the site looked the text up in its own language. The Inbox showed the German answer under the French question.
• A colleague who is suspended, or limited to fewer websites, loses their open live connection. Removing someone already did this; changing their permissions did not, so their open tab kept receiving conversation headers it should no longer see.
• The route that mirrors your own answers can no longer be used to flood a conversation. At most five assistant lines in a row; the longest real sequence is three.

0.22.22

• A greeting is answered as a greeting again. Site owners write their topics as full sentences, and „Hello, how do I book an appointment?“ is a perfectly normal line. Anyone who then typed just „Hello“ got the phone number from that topic instead of a welcome — on the most common opening message there is.
• The click-through no longer ends after an enquiry. Sending the enquiry form left the visitor in front of an empty window: no topics, no „Show questions“. Asking for a person a second time did the same, because the form is offered only once per visit. Both now lead back into the menu.
• Typing mistakes find their answer again. A fix against word mix-ups („kosten“ / „konten“) also threw away the most common typo of all — a finger on the neighbouring key. Measured over 26 everyday business words, 52 of 383 mistyped spellings were still recognised; now 292 are, and both mix-ups stay apart.
• Asking for a human works in every widget language. French, Italian, Spanish and Turkish knew two to four phrases each; over 43 real sentences, one of them matched. Now 42 do. Plain nouns like „customer service“ no longer hijack a pricing question about that same thing.
• French no longer switches from „vous“ to „tu“ mid-conversation, Turkish no longer switches from „sen“ to „siz“ in the enquiry form, and the Turkish greeting bubble asks about questions instead of asking whether something is broken.
• Nine interface texts that were missing from the translation catalogue — among them the whole click-through — now have their German version.
• Bulk actions in the shared inbox only ever touch what is on screen. With a search term active, selecting conversations and then deleting could remove ones the search had just hidden.
• The rate limit no longer lets a burst through at the moment its counting window rolls over.

0.22.21

• The „currently on the site“ panel no longer eats half the screen on a phone. With an empty conversation list it took 52 % of the height for three lines of text, squeezing the list into what was left.

0.22.20

• The click-through menu now speaks every widget language. On a French, Turkish, Italian or Spanish site the title, greeting and topic buttons were translated, but the six lines that carry the whole free tier were not: „What can I help you with?“, „More topics“, „Back“, „Something else“, „Show questions“. On the free plan that menu is the interface.
• The Inbox inside wp-admin no longer freezes when it stays open. WordPress renews its security token after a while; every request then failed silently and the stale list just sat there with no explanation. It now says what happened and offers a reload.
• The shared inbox is easier to scan and search. A wider desktop list, conversation search across site, message, page and language, and clearer status pills put requests that need a reply first without inventing an unread state.
• The empty inbox now helps instead of repeating itself. The main area shows open and waiting totals; the duplicate message in the context column is gone.
• Workspace and website now read as one scope control, account destinations use clear chevrons, and the profile menu reliably covers conversation action buttons instead of letting their three dots shine through.
• Login help and privacy text have stronger contrast in light and dark mode. The shared responsive Inbox improvements ship in both the browser Agent App and WordPress admin.

0.22.19

• File names of attachments stay readable. „mein foto.png“ arrived as „mein20foto.png“, and anything with an umlaut turned into gibberish: the name was cleaned while still URL-encoded, which strips the percent signs.
• A damaged settings value no longer takes the whole website down. If the stored options were ever anything other than an array — a broken serialisation, an import, a filter — every single request died with a fatal error, front end and wp-admin alike, with no way back except the database.
• Running the setup wizard a second time no longer resets where enquiries are sent. The field always showed the WordPress admin address, and the wizard saves what the form contains — so a business that had pointed enquiries at sales@ got them sent back to the admin address, silently.

0.22.18

• The assistant no longer answers the wrong question with full confidence. Two rules that help the free tier when a visitor words things differently were treating unrelated German words as the same: „Versicherung“ got the shipping answer, „Übersetzung“ the bank details, „Bescheinigung“ the complaints answer, „Bestätigung“ the order answer. Thirteen such mix-ups out of twenty word pairs tested. A shared prefix like ver- or an- no longer counts as a shared word stem — while „Öffnungszeiten“/“geöffnet“, the case the rule exists for, still matches.
• A typo is a slipped key, not a different word. „kosten“ and „konten“ are one letter apart, so are „anmelden“ and „abmelden“ — and the pricing question was answered with the account details. Words of the same length that differ in the middle are no longer treated as typos. Forgotten letters, swapped neighbours and word endings („kostet“/“kosten“) still match.

0.22.17

• Saying „hi“ no longer empties the chat. Every message you send clears the topic buttons above it, and the friendly replies — „hi“, „thanks“, „ok“, „are you a bot?“ — were the one path that did not bring them back. The most common opening line left the visitor with a greeting and nothing to click. The buttons now return after every one of them.
• An enquiry with a mistyped e-mail address is no longer lost. „maria@gmail“ without the ending was rejected by the server, but the widget showed „That did not work, please try again in a moment“ — a fault message. People tried the same address again, got the same line and gave up. The form now says which field needs fixing, in all six languages, and it checks for a real address instead of just an „@“ somewhere.
• Long AI answers reach the Inbox again. Two separate faults kept them out: the proof of authorship was computed over a different text than the one sent back — any answer containing „<“ or a link with an umlaut in it broke it — and the mirror rejected anything over 1000 characters while the service allows ten times that. Exactly the detailed answers an operator most needs to see.
• Turkish, Greek, Cyrillic and emoji get the full message length. The limit counted bytes, so 600 Turkish characters were refused while 600 German ones went through.
• „Please enter a shorter message“ no longer appears for an empty message.
• Long live chats are no longer cut off. The request limit reset its own clock on every request, so a chat open for more than two and a half minutes stopped updating for a minute at a time.

0.22.16

• A message about suicide did not reach a person. The rule that pulls someone into the chat for a crisis only knew „umbringen“ written as one word — but in German the verb splits, and the sentence people actually write is „Ich bringe mich um“. „I want to kill myself“, „Ich will nicht mehr leben“ and every wording in the four other widget languages fell through as well. Measured against twenty-two real phrasings, the old rule caught three. All six languages are covered now, and twelve ordinary sentences containing „bringen“ or „leben“ still call nobody.
• „How many employees do you have?“ no longer calls an employee. The bare word was enough to trigger a handover, and since 0.22.13 that happens on the very first message — so an ordinary question notified the whole team. The same for phrases with words far apart: „How many staff do I need to run the chat?“ matched „staff chat“ because the words were only required to appear in order, anywhere in the sentence. Now they have to be close together. While fixing it, a third bug surfaced: „I would like to chat with an employee“ was not recognised at all.
• A link can no longer claim one address and lead to another. An answer containing [alynivo.de](https://alynivo.de.attacker.example/login) showed your domain and went somewhere else. The real domain is now named next to the link whenever the label claims a different one — the same rule every mail client uses. An honest label is left alone.
• Long answers no longer stall the chat. A message packed with „@“ characters took a third of a second per line to render, freezing the browser while it worked. Now it is immediate.
• The Agent Inbox inside wp-admin shows AI answers the way the visitor sees them: bold where it says bold, links as links, paragraphs as paragraphs — instead of raw asterisks and brackets.
• Line breaks in the chat window stay line breaks. A list of points used to collapse into one block of running text.

0.22.15

• Asking for a person is recognised far more reliably. The check only knew phrasings with a verb („speak to an employee“), so „I would like an employee on the chat“ fell through to your own answers — and on a site whose answers cover pricing, someone asking for help was handed a sales pitch. Measured against eleven real phrasings, nine were missed; now all of them are caught, and ten ordinary questions still call nobody.
• The same request now also comes before your own answers on sites without Live Chat. There nobody can take over, but the reply is honest and leads to the enquiry form instead of selling something.

0.22.14

• The enquiry form is no longer a dead end. Once it was open there was no way back: closing and reopening the chat did not help, because the form stays where it is — the only way out was to actually hand over your contact details. There is now a „Back to the questions“ button next to Send, and it brings the question menu back.
• The setting for the question menu is in German again on German sites. Four new labels had been added in 0.22.11 without their translation.

0.22.13

• When a rule brings a person into the chat, the visitor is now told. Complaints and sensitive topics have always been able to pull someone in without the visitor asking — but it happened silently: an operator was waiting in the Inbox while the assistant kept talking, and the visitor never found out. That mattered most in exactly the case the rule exists for.
• The line the assistant says about it reaches the Agent Inbox too, so your team reads what the visitor was promised.

0.22.12

• Everything from 0.22.11. The new self-check no longer reports the opening greeting as a missing answer: it is not a reply to anything, and at that point there is no conversation for it to be missing from. A warning that cries wolf on every visit is one nobody reads a week later.

0.22.11

• A question menu for the free version. The chat now opens with your own answers as buttons instead of an empty input box, so a visitor does not have to guess what it knows. It is fed entirely by the answers you already entered — there is no second list to maintain, and it works from the first minute without another setting. Visitors can still type freely at any time, and asking for a person always works. Six topics per page, and the menu comes back after every answer so nobody reaches a dead end. On by default; switch it off under Question menu.
• Common wording now finds your answers. „When are you open?“ reaches an entry stored as „Opening hours“ — around twenty topic groups covering prices, hours, appointments, directions, delivery, payment, cancellation, warranty and more. A topic match never answers on its own, it only ever suggests, so a broad word costs an unnecessary button rather than a confidently wrong answer.
• Typos no longer cost the answer. One edit from five characters, two from eight, and swapped neighbours count as one — „Presie“ finds „Preise“. The first letter has to match, so „Preis“ and „Kreis“ never become neighbours.
• All three only apply without Managed AI. With it, the model answers the question properly and in the visitor’s own language.

0.22.10

• Everything from 0.22.9, plus: without Managed AI, a question that nearly matches one of your own answers is now offered as a suggestion instead of being turned away. The free tier has nothing but your own answers, and a question phrased slightly differently used to end in „I cannot answer that reliably“ while the right answer sat one wording away. A weak match still never answers on its own — a single short word would hit the wrong entry — it asks, and the visitor decides. With Managed AI nothing changes: the model answers the question properly, in the visitor’s own language.
• If none of the suggestions fit, the enquiry form follows, so the conversation still has a way out.