M2 Breeze AI

Breeze AI 1.10.1

• Banner images come from the design, never from a made-up address. A section build could fill banner pictures with addresses the model invented (example.com/image1.png), and the page showed broken images with no warning. Images now arrive only as a region of the reference design, which is cropped; any address that is not a stored crop, the placeholder or one written word for word in your brief is dropped, the image is filled like any other, and the build notes which section it was. Crops are also aimed at the card's own picture, so a product card no longer gets a crop of its flat background.
• Cards side by side inside a tab panel are laid out as a grid. When the model stacked a tab's cards one under another, the build now reads where they sit in the design and puts them into the matching Grid or Columns layout of that tab, with a warning naming the section. Cards crammed into a single cell of a layout are spread out the same way.
• Picture cards are built as banners, not as a product grid. A row of cards with a photo, a title and a line of copy but no price was sometimes built as a Products block listing random catalog products. The layout of a tab panel is now asked for before its blocks, and Breeze Content Builder 1.2.11 tells the model that Products shows live catalog items with price and cart button.
• Text under a stacked banner is readable. A banner whose copy sits below its picture kept Content Builder's default white text on a light section. Its text colour is now set to dark or light, whichever contrasts better with the section colour measured on the design.
• The model now sees what each option of a select draws, e.g. Overlay (text over image) or Stack (text below/above image), not only its value.
• Works best with Breeze Content Builder 1.2.11.

Breeze AI 1.10.0

• Build from a reference design now plans the page first and asks you to confirm it. The AI chat panel reads the screenshot and lists one row per section: what the band does for the visitor (hero, benefits, category strip, product carousel, newsletter…), the element it will be built with, and its texts. You can drop a section or change its element before anything is built. The plan's section heights and dark/light backgrounds are measured on the screenshot itself instead of being guessed by the model. Behaviour change: a new design is only built from a confirmed plan, and the separate "Build from screenshot" modal is gone — the chat panel replaces it.
• Each section is built on its own. With a confirmed plan every section gets its own AI call and the bands are stacked in plan order, so a long page no longer comes back with half-filled containers; each slot of a Columns, Grid or Tabs is asked for by its own key. In our measurements on the same design and model the section score rose from 84.8 to 93.8.
• A hero slider is created from the design. When a section is an EasySlide slider, the build creates a new slider with the design's slide picture, heading, copy and button instead of pointing at one of the store's existing sliders — edit its slides and give the button a link under Swissup > EasySlide. The slide copy is kept readable over the photo and the button is sized to its label.
• A category strip renders as category tiles. An Easy Catalog Images widget now gets its category count, columns and "show image" from the design, instead of the widget defaults that rendered the whole category tree as text links.
• Cropped images no longer carry the design's own text. A crop taken from the reference design kept the heading, copy or form baked into it, and the real text was then drawn on top. When an image model is configured under Generated Images, every crop with text has the text painted out and keeps the photo (about $0.02 per image at low quality, within the Max images limit).
• Tabs and grids hold the right blocks. A tab panel keeps its blocks in its own tab instead of spreading them one per tab, and a panel of one large card beside smaller ones is built as a grid of banners. The ten-slot gallery-2x5 Grid preset of Breeze Content Builder 1.2.7 is supported, a gallery is no longer left with empty cells, and a section title is no longer printed twice.
• Wide or tall screenshots are placed on the right scale. A page whose bands ran past the end of the grid is rescaled, and when the crops may still sit on the wrong vertical scale the build says so. A band no widget matches no longer gets a widget hint.
• Developer: test sets under dev/test-sets/ and bin/magento breezeai:testset:score, which scores a build or a plan section by section (element, width, texts, images, data, defects); --model-id runs a billed build, --document / --plan-file score saved ones for free.
• Works best with Breeze Content Builder 1.2.10, whose Banner and Grid descriptions the planner relies on.

Breeze AI 1.9.0

• A page built from a reference design now reproduces it far more reliably. The page document the AI returns is checked by the provider against a schema instead of being written by hand inside a text field, which is where most failed builds used to end — on the cheaper models a build that came back with nothing was the common case rather than the rare one. In measurement on the same reference design and the same model, this took a usable page from 2 builds in 9 to 18 in 18.
• A band of blocks side by side is laid out with the right container. Columns offers up to five, so a row of six category tiles used to be squeezed back to two and four of the tiles thrown away. Six or more blocks, or blocks of unequal size, now go to a Grid with the preset whose slot count matches — six across is gallery, five uneven is bento — and a band under a row of tab labels becomes a Tabs with content behind every label rather than one open panel and two blanks.
• The builder no longer rebuilds your theme's own header and footer. Content Builder content sits inside a page your theme already wraps, so a header or footer in the reference design is skipped instead of becoming a section of its own. A bare strip of colour at the very top or bottom of a screenshot — usually where the image was cropped — is skipped for the same reason, rather than being turned into a section the AI invents a name for.
• A section that came out wrong now says so. A band with nothing in it is reported by position, a container holding more blocks than it has cells is reported instead of silently stacking two of them into one, and a page that came back with nothing in any section is refused outright rather than replacing the page you had already built with blank space.
• A build that fails now tells you which of three things happened. Each used to arrive as the same sentence about invalid JSON. They are worded apart now: the reply ran past the output limit, so raise Max Tokens on the model and try again; the provider declined the request, so the brief is the thing to change; or the reply missed the shape that was asked for, which is worth one more try on the same brief. The same three apply to Plan sections. Authorization and rate-limit errors from OpenAI are recognised as themselves, and a rate-limited bulk run is retried rather than failed.
• Not every model can build a page from a written brief alone. A cheaper model can be worse at a whole task rather than slightly worse at every task: in our testing GPT-4o mini wrote bulk attribute copy well and built a page from a reference screenshot, but asked for a page from a brief with no screenshot it reported the right number of sections and then returned every one of them empty. The Model field now says so, and a build that comes back empty names both things that help — attaching a reference design, or choosing a stronger model under Swissup > Breeze AI > Models.
• The AI chat panel can be closed. There is a close button in the panel's own header, the button that opens it now shows whether it is open, and closing from the panel returns the keyboard to that button instead of dropping focus at the top of the page.
• Debug logs may include provider response content, so enable them only while troubleshooting and disable them afterward.

Breeze AI 1.8.22

• Fixed: setup:di:compile failed on Magento 2.4.7 with "Incompatible argument type: Required type: \Exception. Actual type: \Throwable" for StructuredResponseException, introduced in 1.8.21. The exception now accepts \Exception as its cause, matching LocalizedException.

Breeze AI 1.8.21

• Security: Medium — strengthened authorization and input validation for bulk AI operations.
• Security: Low — improved validation of model connection settings.
• Security: Low — improved validation of the MCP endpoint settings.
• Security: Low — improved validation of links in generated page drafts.
• Developer API: AI providers (Claude, Gemini, OpenAI) can now return structured output validated against a schema (ProviderInterface).
• Developer API: a curated, per-store-view cached vocabulary of filterable attributes and categories, groundwork for AI product search. No visible change yet.

Upgrade note: permission requirements have changed; review administrative roles that use bulk AI operations.

Breeze AI 1.8.20

• The AI chat panel in the Content Builder editor completes a request for the first time. Pressing Send — or, since 1.8.14, Plan sections — in the editor's chat panel never finished: the loader sat there and the built page never reached the screen, even though the build had already been sent to the provider and billed for in full. The panel has behaved this way since it appeared in 1.3.0. Both buttons now complete normally, the result is applied to the page, and the reply appears in the panel. The build dialog on the CMS page form was never affected, which is why this went so long without being caught. If you tried the editor's chat panel, got nothing back and went back to the page form, it is worth another try.

• Attaching a design now asks what you want done with it. Attaching a screenshot in the chat panel used to leave you to work out for yourself that there were two ways forward. The panel now asks, and both answers are one click: Build it now, which is exactly the behaviour of the previous release, or Plan it with me first, which brings back the list of sections to approve before the build is paid for. When a build lands, the panel speaks first — how many sections it built, anything it could not place, and where the pictures on the page came from — and then refining by chat continues as before. Neither the question nor the summary costs anything: no extra call to the provider is made for either. Two long-standing annoyances in the same panel are fixed alongside: long notes now wrap instead of being clipped at the edge of the dock, and Send stays reachable on a laptop screen after a five-section plan rather than being pushed out of view.

• Correction to the 1.8.14 release note: leaving Section-Plan Model empty did not use the model the build runs on. That note said it did. In fact the planning call went to the installation's default model row, so a plan could be billed to a model nobody had chosen, and the cost sentence shown beside the button was priced against a third reading again. On one measured setup — default row claude-sonnet-5, the editor's own build model gpt-4o-mini — the sentence said planning would add about $0.0002 and it added $0.0065, roughly thirty times the quoted increment. The field, the call and the sentence now all name the same model row. If you left Section-Plan Model empty on the strength of that note, your planning calls were going somewhere else until this release.

• Pointing Section-Plan Model at different models in different store views now hides the cost sentence rather than quoting one of them. The panel has no page open when that sentence is written, so it cannot know which store view's model it would be pricing. An installation that sets this setting once, or never, sees the sentence exactly as before.

Breeze AI 1.8.14

• See which sections the builder intends to build — and edit that list — before a build is paid for. The Content Builder chat panel now has a Plan sections button beside Send. Attach a screenshot of the design, press it, and the list of sections comes back first: hero banner, three feature cards, quote strip, newsletter sign-up. Strike a line out and that section is genuinely absent from the page you get — on a four-section mockup, striking out the quote strip produced a three-section page with that band's copy nowhere in it. Where the design could honestly be read two ways, the plan asks which reading you meant instead of guessing. Planning is optional: building straight away is unchanged, still one click, and the build itself is still a single call over the whole page rather than a section-by-section loop. Planning adds one small call in front of that build — on the same four-section mockup it answered in 4.4 seconds and cost $0.0039 on gpt-4o-mini.

• The planning call can run on a cheaper model than the build. A new Section-Plan Model setting under Page Builder gives the planning call its own model row; leave it empty and it uses the same model the build runs on, so nothing changes until you set it. The planning call carries none of the component catalogue the build prompt needs, which is why a smaller model usually answers it just as well. The panel also shows what planning first adds to the cost before you choose it — read that as an estimate rather than a quote. It is counted from the size of the text sent, and some models charge for an attached screenshot at a near-fixed rate however short the prompt is, so on those the real difference is bigger than the figure shown: on gpt-4o-mini the measured difference was about 1.7× a plain build where the sentence showed about 1.1×.

Breeze AI 1.8.13

• The page builder's settings can now be set per website and per store view, and a build genuinely honours them. Everything under Page Builder and Page Builder: Generated Images — whether images are cut out of the reference screenshot and how many, which widgets the builder may place, whether pictures are generated at all, which image model draws them, its quality, the per-page limits on generated pictures and icons, the negotiated price per image and the timeout per image — used to be one value for the whole installation. Each of them now carries website and store view scope. Nothing changes for a merchant who sets nothing: a build reads the same installation-wide values it always did. Where a value is set per store, a build of a page that lives in exactly one store view reads that store's values — so one store view can cap a page at two generated pictures while another allows seven, or draw with a different image model at its own negotiated rate. That holds on every route into a build: the admin build dialog, queued bulk builds, the command line and the MCP server. The estimate shown before a build and the figures written to the usage log are read from the same store, so the two cannot disagree. One boundary worth knowing: a page assigned to several store views, or to All Store Views, has no single store to resolve and reads the installation-wide values on purpose — quietly picking one of the assigned stores and spending that store's image budget would be a surprise rather than a fix. Keep Unused Images For (days) also stays installation-wide, because the nightly cleanup sweeps every store in one pass.

Breeze AI 1.8.12

• The MCP Server settings can now be set per website and per store view. Enabled and GraphQL Endpoint Override were previously a single value for the whole installation; both now carry website and store view scope, so one store can run the MCP server against its own GraphQL endpoint while another keeps it switched off. A value set that way is genuinely honoured at request time — a call made to a store's own REST URL reads that store's setting rather than the installation-wide one. Everything else in the Breeze AI configuration section stays installation-wide on purpose: those settings are read by background jobs and cleanup tasks that run across every store at once, so a per-store value would look like it applied and never be used.

Breeze AI 1.8.11

• A billed call that failed is no longer invisible in the usage log. If a Claude reply ran out of its answer limit or came back unusable, the request had already been paid for in full — one measured case cost $0.028 for 844 tokens of output — but nothing showed up in the usage log to account for it, so total spend looked lower than it really was. Such a call is now logged and priced exactly like a successful one, so the log matches what you were actually charged.
• Choose where a build's images come from, per build, without touching store configuration. Both the page builder modal and the Content Builder chat panel now offer an Images choice — store default, extract from the source, use placeholders, or generate — that applies to that one build only; your store's own settings are left untouched. Picking Generate with no image model configured tells you so the moment you choose it, before the build runs and before anything is billed.