@koda-sl/baker-cli 0.182.1-dev.c8185c1b5 → 0.185.0

package/README.md

@@ -1753,7 +1753,6 @@ Each external source is its own subcommand. Pick the verb that matches the sourc
 | `baker images stock <q> [--type photo\|vector\|psd]` | Magnific (Freepik's dev API) — photos, vectors, illustrations, PSDs (~250M assets) | off |
 | `baker images google <q>` | Google Images via the official Custom Search JSON API | off |
 | `baker images pinterest <q>` | Pinterest reference imagery via ScrapeCreators (photo-real mood boards) | off |
-| `baker images generate <prompt>` | AI image generation via OpenRouter image models (cost-tracked) | **always on** (the bytes are the artifact) |
 | `baker images logo <domain>` | Brand logo via Brandfetch CDN | **on** (top 1) |
 | `baker images icon <name>` | Iconify (200+ icon sets, no API key) | off (CDN URL is stable) |
 | `baker images gif <q>` | Reaction GIFs / memes via Giphy (paid-social creative) | off |
@@ -1854,7 +1853,7 @@ baker images stock "office reaction" --people only
 
 Free tier exists but watermarks previews — pass `--license freemium` to filter to clean free assets explicitly.
 
-An empty result comes back with `hints` naming the filters that narrowed the search plus the fallbacks (`baker images find --sources …`, `baker images generate`); a provider failure returns an `error.fix` pointing at the same exits. `baker images google` does the same, minus the cross-provider fallback (it is the last-resort provider).
+An empty result comes back with `hints` naming the filters that narrowed the search plus the fallbacks (`baker images find --sources …`, `baker studio generate`); a provider failure returns an `error.fix` pointing at the same exits. `baker images google` does the same, minus the cross-provider fallback (it is the last-resort provider).
 
 **Flags:**
 
@@ -1916,53 +1915,9 @@ baker images pinterest "cozy coffee shop interior" --auto-ingest 2 --context "Mo
 
 Endpoint `api.scrapecreators.com/v1/pinterest/search` (`SCRAPE_CREATORS_API_KEY`), $0.00188/request, 1-day cache.
 
-### `baker images generate <prompt>`
+### AI image generation
 
-Generate an image with AI (OpenRouter image models) and ingest it into the library. The generated bytes **are** the artifact, so the result is always auto-ingested (described + embedded) — the next `baker images library` query finds it. Cost is tracked per request via OpenRouter's reported usage. Models and defaults mirror the Baker canvas.
-
-```bash
-baker images generate "a friendly golden retriever in a bright modern living room" --aspect-ratio 16:9
-baker images generate "hero shot of a matte black water bottle on wet marble, studio light" \
-  --model google/gemini-3-pro-image-preview --image-size 2K --aspect-ratio 4:5
-baker images generate "home office hero, warm minimalist, natural light, 35mm photo" \
-  --reference "https://i.pinimg.com/…/pin1.jpg,https://…/brand-product.png"
-baker images generate "put this product on a marble countertop in soft daylight" \
-  --reference "./src/brand/logos/product.png,./refs/kitchen-mood.jpg"   # local sandbox files
-baker images generate "flat geometric mascot, brand palette" \
-  --model recraft/recraft-v4.1-pro-vector --rgb-colors "[[10,10,10],[255,80,0]]" --bg-rgb "[255,255,255]"
-```
-
-**Models** (`--model`, default `google/gemini-3.1-flash-image-preview`):
-
-| Model | Best for | Aspect ratios | Sizes |
-|---|---|---|---|
-| `google/gemini-3.1-flash-image-preview` **(default)** | Fast iteration, reference editing, extreme aspect ratios (Nano Banana flash) | standard **+** `1:4` `4:1` `1:8` `8:1` | `0.5K`–`4K` |
-| `google/gemini-3.1-flash-lite-image` | Cheapest tier (Nano Banana 2 Lite) — half the flash price per output token. High-volume variants and any 1K-sufficient frame, portrait ad formats included | standard **+** `1:4` `4:1` `1:8` `8:1` | `1K` **only** |
-| `google/gemini-3-pro-image-preview` | Highest fidelity (Nano Banana Pro) — strong prompt adherence and reference grounding | standard set | `1K` `2K` `4K` |
-| `openai/gpt-image-2` | Photoreal + cleanest in-image text — ad/landing reproduction | standard set **minus** `4:5` `5:4` | *(none — derived from the ratio)* |
-| `openai/gpt-5.4-image-2` | **Legacy** — accepted so canvases authored before the gpt-image-2 switch still re-run. Don't pick it for new work. | standard set | `1K` `2K` `4K` (ignored upstream) |
-| `recraft/recraft-v4.1-pro-vector` | Vector/flat/SVG-style with palette control | standard set | `1K` `2K` `4K` |
-
-Standard aspect ratios: `1:1` `2:3` `3:2` `3:4` `4:3` `4:5` `5:4` `9:16` `16:9` `21:9`.
-
-`openai/gpt-image-2` accepts neither `image_size` nor the `4:5` / `5:4` portrait ratios — OpenAI picks the pixel dimensions from the ratio. To ship a 4:5 Meta feed ad from it, render the hero at `3:4` and adapt with `image_aspect_adapt`; the scaffolders do this for you.
-
-`google/gemini-3.1-flash-lite-image` renders at `1K` and nothing else — `2K` / `4K` is a provider 400 (`resolution: not supported. Accepted: 1K`), not a soft clamp. It costs half of `google/gemini-3.1-flash-image-preview` per output image token ($0.00003 vs $0.00006), so it is the cheap-iteration tier for images. Keep hero stills and reference sheets that need `2K`+ on the flash or pro tier.
-
-**Flags:**
-
-| Flag | Description |
-|---|---|
-| `--model` | Model id (default `google/gemini-3.1-flash-image-preview`) |
-| `--aspect-ratio` | Output aspect ratio (default `1:1`) |
-| `--image-size` | Resolution: `1K` (default) `2K` `4K` (Gemini flash also `0.5K`; Gemini flash-lite is `1K` only) |
-| `--reference` | Comma-separated visual references, each either a **public image URL** (Pinterest / stock / library `imageUrl`) **or a local file path** (a sandbox image — brand logo, product shot, cropped photo, screenshot). Local files are downscaled (≤1536px) and inlined automatically — no manual upload. Applied in order; the biggest quality lever for photographed, on-brand output. Split is on `,`, so a URL containing a literal comma in its query string would be torn in two (rare for image CDNs — pass it alone if it occurs); a single `data:` URL is taken whole. |
-| `--strength` | Recraft only: vectorization strength 0–1 |
-| `--rgb-colors` | Recraft only: JSON palette `[[r,g,b],…]` |
-| `--bg-rgb` | Recraft only: JSON background `[r,g,b]` |
-| `--context` | Describe-hint override for the ingested row (defaults to the prompt) |
-
-Returns `{ images: [{ imageId, imageUrl, deduped, width, height }], model, costUsd }`. `imageUrl` is library-owned and ready to place. Requires `OPENROUTER_API_KEY` on the Convex deployment. No `seed` (OpenRouter image_config has no seed slot). Identical re-generations dedup by content hash.
+Its own family — see [`baker studio`](#baker-studio--make-new-imagery) below. `baker images` sources what already exists; `baker studio` makes what doesn't.
 
 ### `baker images logo <domain>`
 
@@ -2920,6 +2875,155 @@ baker schema images.search      # Get image search schema
 
 Not every command registers a schema. On `NOT_FOUND` the error carries a structured `fix` pointing at `--help` and the family's agent doc (`__tooling__/docs/tools/baker/<family>.md` in sandboxes). Every top-level family's `--help` also ends with a `Full guide:` line pointing at that doc; `packages/cli/__tests__/doc-coverage.test.ts` fails CI if a mounted command is missing from its family doc.
 
+## `baker studio` — make new imagery
+
+The generation lane. `baker images` **sources** what already exists; `baker studio` **makes** what doesn't. Every call opens a *batch* — one brief, 1-8 takes, rendered in parallel — recorded against the calling chat and shown in the dashboard Studio next to the batches the client ran themselves.
+
+| Command | Purpose |
+|---|---|
+| `baker studio generate <prompt>` | Render 1-8 takes of one still. Each take is ingested into the media library as it lands |
+| `baker studio animate <prompt> --image <…>` | Turn a still into a clip. The prompt describes movement, not the look |
+| `baker studio get <generationId>` | Read one batch: takes, urls, library state, per-take failure reason |
+| `baker studio list` | Recent batches for the calling chat (`--all` widens to the company) |
+| `baker studio keep <generationId> --slot N` | Mark the keeper. Stars an image; **ingests** a clip into the video library |
+| `baker studio models` | Per model: shapes, resolutions, clip lengths, prompt cap, reference limit, every knob. `--kind image\|video`, `--model <id>` |
+| `baker studio skills` | Craft directions accepted by `generate --skill <id>` |
+| `baker studio improve <prompt>` | Sharpen a rough brief — the same rewrite the client's prompt-bar wand performs. Returns the rewrite and the original; generates nothing |
+
+**Every per-model limit is enforced before anything is billed.** Prompt character caps (Kling 2,500 · Seedance 4,000 · Recraft 10,000 · gpt-image 32,000), reference-image counts (Recraft 1 · Gemini 14 · gpt-image 16 · Seedance 9 in reference mode), the uint32 seed range, RGB channel bounds, and `width:height` shapes are all checked against the chosen model up front — the request is refused with the number to fix rather than opening a batch that fails a take at a time. A knob the model does not have is refused too; only the *shape* snaps to the nearest it renders. `baker studio models` reads all of this from the same contract table the backend validates against, so it cannot drift.
+
+Two of those checks run in the CLI itself, before the request is even sent. The **reference-image count** is one: `--reference` / `--image` accept file paths and URLs, which are uploaded to the media library first, so checking the count only at the backend means paying to ingest and describe every image in a set that was always going to be refused. The **first-frame rule** is the other — a clip started from one exact frame takes exactly one image whatever the model's reference slot allows, and reference mode is refused outright on Veo and Kling, which cannot read references at all.
+
+**References must be JPG, PNG or WebP.** The models read nothing else, and the library legitimately holds SVGs because Recraft's vector model writes them — so "render the logo I just made, on a kitchen counter" is a shape the trap fits exactly. It is refused on the way in, naming the format.
+
+Starting a batch returns 202 + a `generationId`; the CLI polls until the takes settle. `--wait false` / `--max-wait 0` returns the id immediately, and a wait that runs out is **not** a failure — the id comes back with `status: running` and a hint to collect it with `baker studio get`.
+
+### `baker studio generate <prompt>`
+
+```bash
+baker studio generate "a friendly golden retriever in a bright modern living room" --aspect-ratio 16:9
+baker studio generate "hero shot of a matte black water bottle on wet marble, studio light" \
+  --model google/gemini-3-pro-image-preview --image-size 2K --aspect-ratio 4:5 --count 3
+baker studio generate "home office hero, warm minimalist, natural light, 35mm photo" \
+  --reference "https://i.pinimg.com/…/pin1.jpg,./src/brand/logos/product.png,j57abc123def456ghi789"
+baker studio generate "founder-style selfie, kitchen background" --skill ugc-selfie-hook
+baker studio generate "flat geometric mascot, brand palette" \
+  --model recraft/recraft-v4.1-pro-vector --rgb-colors "[[10,10,10],[255,80,0]]" --bg-rgb "[255,255,255]"
+```
+
+**Models** (`--model`, default `google/gemini-3.1-flash-image-preview`):
+
+| Model | Best for | Aspect ratios | Sizes |
+|---|---|---|---|
+| `google/gemini-3.1-flash-image-preview` **(default)** | Fast iteration, reference editing, extreme aspect ratios (Nano Banana flash) | standard **+** `1:4` `4:1` `1:8` `8:1` | `0.5K`–`4K` |
+| `google/gemini-3.1-flash-lite-image` | Cheapest tier (Nano Banana 2 Lite) — half the flash price per output token. High-volume variants and any 1K-sufficient frame, portrait ad formats included | standard **+** `1:4` `4:1` `1:8` `8:1` | `1K` **only** |
+| `google/gemini-3-pro-image-preview` | Highest fidelity (Nano Banana Pro) — strong prompt adherence and reference grounding | standard set | `1K` `2K` `4K` |
+| `openai/gpt-image-2` | Photoreal + cleanest in-image text — ad/landing reproduction | standard set **minus** `4:5` `5:4` | *(none — derived from the ratio)* |
+| `openai/gpt-5.4-image-2` | **Legacy** — accepted so canvases authored before the gpt-image-2 switch still re-run. Don't pick it for new work. | standard set | `1K` `2K` `4K` (ignored upstream) |
+| `recraft/recraft-v4.1-pro-vector` | Vector/flat/SVG-style with palette control. Writes **SVG**, so its output cannot be fed back as a reference | `1:1` `4:3` `3:4` `16:9` `9:16` **only** | *(none — refuses `image_size`)* |
+
+Standard aspect ratios: `1:1` `2:3` `3:2` `3:4` `4:3` `4:5` `5:4` `9:16` `16:9` `21:9`.
+
+`openai/gpt-image-2` accepts neither `image_size` nor the `4:5` / `5:4` portrait ratios — OpenAI picks the pixel dimensions from the ratio. To ship a 4:5 Meta feed ad from it, render the hero at `3:4` and adapt with `image_aspect_adapt`; the scaffolders do this for you.
+
+`google/gemini-3.1-flash-lite-image` renders at `1K` and nothing else — `2K` / `4K` is a provider 400 (`resolution: not supported. Accepted: 1K`), not a soft clamp. It costs half of `google/gemini-3.1-flash-image-preview` per output image token ($0.00003 vs $0.00006), so it is the cheap-iteration tier for images. Keep hero stills and reference sheets that need `2K`+ on the flash or pro tier.
+
+`recraft/recraft-v4.1-pro-vector` is the narrowest of the roster in three ways, all read off its live endpoint schema: it renders **five** ratios (no `4:5` / `5:4` / `2:3` / `3:2` / `21:9`, so a portrait ad shape snaps to `3:4`), it has **no `image_size`** — the provider accepts one and ignores it, which is why it is refused here — and it takes **exactly one** reference image against 14-16 elsewhere. Its output is an SVG, and the image models read only JPG, PNG and WebP, so a Recraft take cannot be used as a reference for the next render.
+
+A knob the chosen model does not advertise is **refused, not silently dropped** — the error names the model and what it accepts, and carries an `error.fix` telling the agent to drop the flag rather than abandon the job.
+
+**Flags:**
+
+| Flag | Description |
+|---|---|
+| `--model` | Model id (default `google/gemini-3.1-flash-image-preview`) |
+| `--aspect-ratio` | Output aspect ratio (default `1:1`) |
+| `--count` | Takes to render, 1–8 (default 1) |
+| `--reference` | Comma-separated visual references, in order: **library image ids**, **local file paths**, and/or **image URLs**. Paths and URLs are added to the media library first and passed as ids — the studio takes library ids only, so the client can see what grounded the render and the reference stays reusable. The biggest quality lever there is. |
+| `--skill` | Craft direction id (`baker studio skills`) — art direction plus the model/shape/count it wants |
+| `--brand` | Append the company palette and visual style. **Off by default** for the agent (the dashboard defaults it on) |
+| `--image-size` | Resolution: `0.5K` (Gemini flash only) `1K` `2K` `4K`; Gemini flash-lite is `1K` only. **No default** — left unset the provider picks the pixels from the ratio. `gpt-image-2` has no resolution setting and **refuses** this flag |
+| `--quality` | `low` \| `medium` \| `high` \| `auto`. Only on models that honour it — OpenRouter lists it for gpt-image alone, so every Gemini model and Recraft **refuse** it |
+| `--strength` | Recraft only: vectorization strength 0–1 |
+| `--rgb-colors` | Recraft only: JSON palette `[[r,g,b],…]` |
+| `--bg-rgb` | Recraft only: JSON background `[r,g,b]` |
+| `--wait` / `--max-wait` | Wait for the takes (default on, 240000 ms). `--wait false` / `--max-wait 0` returns the batch id |
+| `--full` | Include settings, references, chat attribution, `createdAt` |
+
+Requires `OPENROUTER_API_KEY` on the Convex deployment. Identical re-generations dedup by content hash.
+
+### `baker studio improve <prompt>`
+
+Sharpen a rough brief into directed art direction, using the same model and prompt as the wand in the dashboard's Studio bar.
+
+```
+baker studio improve "make it look premium"
+baker studio improve "zoom in a bit" --kind video --has-reference
+```
+
+| Flag | Notes |
+| --- | --- |
+| `--kind` | `image` (default) sharpens a still brief; `video` sharpens a **motion** brief |
+| `--skill` | Sharpen it for one craft direction (`baker studio skills`) |
+| `--has-reference` | Say a reference image is wired, so the rewrite describes the **change** rather than the whole scene |
+
+Both `prompt` (the rewrite) and `original` come back, so you can show the client what changed instead of silently replacing their words. Nothing is generated — pass the result to `baker studio generate` yourself.
+
+Reach for it when you are relaying the **client's** own words. When you are writing the art direction yourself, just write it: you will do a better job than a flash model rewriting you.
+
+### `baker studio animate <prompt>`
+
+```bash
+baker studio animate "slow push in, model turns to camera and smiles" --image j57abc123def456ghi789
+baker studio animate "handheld drift right, steam rising from the cup" \
+  --image ./out/hero.png --duration 6 --quality 1080p
+baker studio animate "product rotates once on a turntable" --image j57abc…,j57def… --from references
+```
+
+The image fixes the look, so the prompt describes **movement**. Models: `bytedance/seedance-2.0` (default), `google/veo-3.1`, `google/veo-3.1-fast`, `kwaivgi/kling-v3.0-pro`.
+
+| Flag | Description |
+|---|---|
+| `--image` | **Required.** Comma-separated library image ids / local paths / image URLs; the first is the opening frame |
+| `--from` | `frame` (default — animate the first image) \| `references` (use the images as visual reference). Mutually exclusive at every provider |
+| `--model` | Video model id (default `bytedance/seedance-2.0`) |
+| `--aspect-ratio` | Default `9:16` |
+| `--duration` | Seconds; per-model set (default 5, snapped onto what the model renders) |
+| `--quality` | Resolution, e.g. `720p` (default) \| `1080p` |
+| `--count` | Takes, 1–3 (default 1). A clip costs ~45x an image |
+| `--sound` | Let the model generate audio. Off by default (billed on top) |
+| `--negative` / `--seed` / `--people` / `--enhance-prompt` / `--conditioning-scale` / `--cfg-scale` | Per-model knobs; refused when the chosen model doesn't advertise them |
+| `--wait` / `--max-wait` | Wait for the takes (default on, 600000 ms) |
+| `--full` | Include settings, references, chat attribution, `createdAt` |
+
+**A rendered clip is not usable anywhere until it is kept** — `keep` is what ingests it into the video library. Unkept takes are never ingested, which is what makes a rejected batch cheap.
+
+### `baker studio get` / `list` / `keep` / `skills`
+
+```bash
+baker studio get j57abc123def456ghi789 --full
+baker studio list --kind video --limit 5      # this chat's batches
+baker studio list --all                       # every batch for the company
+baker studio keep j57abc123def456ghi789 --slot 2
+baker studio keep j57abc123def456ghi789 --slot 2 --undo
+baker studio skills
+```
+
+Compact output per take is `{ slot, status, imageId | videoId, url, inLibrary, error }` alongside the batch's prompt, model, aspect ratio and status; `--full` adds the settings echo, the reference ids, the chat id and `createdAt`. `keep` stars an image (it was already in the library) and **ingests** a clip; `--undo` reverses either.
+
+**Craft presets.** `--skill <id>` appends ready-written art direction to your brief — camera, lens, light, plus a clean-plate rule (no text, logos or overlays in the picture; type is composited afterwards) and a preference for physical cause over adjective.
+
+| `--skill` | Model it picks | Shape | Takes | Wants a reference |
+|---|---|---|---|---|
+| `product-pack-shot` | `openai/gpt-image-2` | `1:1` | 3 | **yes** |
+| `ugc-selfie-hook` | `openai/gpt-image-2` | `9:16` | 3 | **yes** |
+| `lifestyle-scene` | `openai/gpt-image-2` | `4:5` | 3 | **yes** |
+| `cinematic-still` | `google/gemini-3-pro-image-preview` | `16:9` | 3 | no |
+| `brand-icon` | `recraft/recraft-v4.1-pro-vector` | `1:1` | 4 | no |
+| `background-plate` | `google/gemini-3.1-flash-image-preview` | `4:5` | 3 | no |
+
+A skill fills only what you did **not** pin — `--model`, `--aspect-ratio` and `--count` always win. Which means a skill can change the model, and the model decides what the rest of the flags may be: `--skill brand-icon` lands on Recraft, so it takes one reference image rather than fourteen, refuses `--image-size`, and writes SVG that cannot be fed back in as a reference. The CLI resolves the skill's model before checking the reference count, so an over-long `--reference` list is refused before anything is uploaded.
+
 ## Creative Canvas
 
 The creative canvas runs declarative JSON pipelines. You author a graph of nodes (text, image, video, audio, composition, control-flow), `baker canvas run` it, and every node's output drops to disk under `canvas/<run_id>/` so you can watch progress and consume the artifacts.
@@ -4976,32 +5080,6 @@ Without `--slug` the command behaves as before (one creative record per publishe
 
 ## Landing Quality (`baker landing`)
 
-Two tools that bracket a landing build: **`inspiration`** before you design (what does good look like here?) and **`critique`** after you build (did this settle into generic defaults?).
-
-### `baker landing inspiration`
-
-A searchable library of real landing-page sections. Each entry is one section of a real page: a screenshot, a standalone HTML+CSS reproduction with a measured fidelity score, a description of how it moves, and the transferable idea behind it — classified in the same vocabulary the landing skill teaches, so a hit is directly buildable.
-
-```bash
-baker landing inspiration search "pricing with a monthly/annual toggle" --scope all
-baker landing inspiration view <section id>     # full DNA + screenshots + motion filmstrip
-baker landing inspiration code <section id>     # the standalone bundle
-baker landing inspiration page <source id>      # a whole page as a section sequence
-baker landing inspiration add https://linear.app --note "client likes this density"
-baker landing inspiration favorites             # what this company has saved
-baker landing inspiration scrape <url> --out <dir>   # capture any page now, synchronously
-```
-
-- **Hybrid search over three signals** — keywords, meaning, and *appearance* (the screenshot is embedded, so a query like "dark developer hero with a terminal" can match a section whose text never says "terminal"). Filters: `--type --composition --register --interaction --motion --media --device --theme --max-rank --min-craft --min-fidelity --domain --similar-to --scope --limit`.
-- **Screenshots are downloaded to `.baker/inspiration/`** so an agent can actually look at a result rather than read a description of it.
-- **`fidelity` says how much to trust the code.** A measured 0–1 match between the live section and our standalone re-render, always reported. Above ~0.95 the markup renders like the original; below ~0.75 the section is scroll- or JS-driven and it does not.
-- **`used_on_pages`** appears when a section shows up on more than one of that site's pages — a nav or footer they ship everywhere, which is a stronger reference than a one-off.
-- **The corpus is shared across companies; favorites are per-company**, and search defaults to this company's saved sections.
-- **`add` studies in the background; `scrape` returns with the page on disk.** `add` grows the shared library and takes minutes, so it can never answer "build our page like this one" within the same turn. `scrape` runs the identical capture locally and blocks until it finishes, writing section screenshots, standalone markup, a whole-page reproduction and `report.html`. `--no-motion` roughly halves the runtime; `--no-mobile`, `--no-code` and `--no-report` skip further passes.
-- **Inspiration, never a clipboard.** Both `code` and `scrape` record what was consulted, and `critique`'s `originality` family blocks a publish that reuses a reference's copy verbatim.
-
-### `baker landing critique`
-
 Deterministic design-quality critic for landing pages — an offline, no-LLM detector for the well-known "AI slop" tells (gradient text, overused fonts like Inter, side-tab colored borders, cream/parchment grounds, purple-on-heading palettes, buzzword copy, broken images, and more). It scores a landing's `.astro` source and returns findings tiered **block / warn / advisory**.
 
 ```bash
@@ -5019,9 +5097,7 @@ baker landing critique spring-offer summer-offer claude  # score three landings
 
 - **Agent washing is a copy tell.** A page that claims autonomy ("fully autonomous", "while you sleep", "no human intervention") and never says anywhere what the buyer approves, overrides, or audits scores an `agent-washing` warn. One signal about oversight of the agent's actions — approval, review, override, undo/rollback, audit log, escalation — anywhere on the page clears it; the rule is page-scope precisely because the hero can defer the trust story to a section below it. Generic privacy boilerplate ("GDPR", "encrypted", "your data") deliberately does **not** clear it, since it appears in most footers and says nothing about who is in charge of the agent.
 
-- **The `originality` family compares your copy against inspiration sections you actually consulted.** It fires only on sections opened with `baker landing inspiration code`, so it flags facts rather than resemblances. Reusing a reference's headline, subhead, or CTA verbatim is block-tier.
-
-Output is the standard envelope `{ ok, data, hints }` with `data = { advisory, slug, overall, counts, dimensions, findings }`. `dimensions` scores eight design families (typography, color, borders_depth, motion, spacing, copy, integrity, originality) 0–1 (higher is better); `counts` is the block/warn/advisory tally.
+Output is the standard envelope `{ ok, data, hints }` with `data = { advisory, slug, overall, counts, dimensions, findings }`. `dimensions` scores seven design families (typography, color, borders_depth, motion, spacing, copy, integrity) 0–1 (higher is better); `counts` is the block/warn/advisory tally.
 
 ## Brand Fonts (`baker brand fonts`)
 

package/dist/{chunk-UZ37VVP4.js → chunk-3TXZKSAR.js}

@@ -7,9 +7,9 @@ import {
   __toESM
 } from "./chunk-YL3HDEIJ.js";
 
-// ../../.pnpm-store/v10/links/@/safe-stable-stringify/2.5.0/810146e81bae4e3a061fe487864f2fde80c4b03b886877dc0f1fffbc6480b67e/node_modules/safe-stable-stringify/index.js
+// ../../node_modules/.pnpm/safe-stable-stringify@2.5.0/node_modules/safe-stable-stringify/index.js
 var require_safe_stable_stringify = __commonJS({
-  "../../.pnpm-store/v10/links/@/safe-stable-stringify/2.5.0/810146e81bae4e3a061fe487864f2fde80c4b03b886877dc0f1fffbc6480b67e/node_modules/safe-stable-stringify/index.js"(exports, module) {
+  "../../node_modules/.pnpm/safe-stable-stringify@2.5.0/node_modules/safe-stable-stringify/index.js"(exports, module) {
     "use strict";
     var { hasOwnProperty } = Object.prototype;
     var stringify = configure2();
@@ -1075,7 +1075,7 @@ function resolveAdaptFormats(params) {
   return params.formats ?? [];
 }
 
-// ../../.pnpm-store/v10/links/@/safe-stable-stringify/2.5.0/810146e81bae4e3a061fe487864f2fde80c4b03b886877dc0f1fffbc6480b67e/node_modules/safe-stable-stringify/esm/wrapper.js
+// ../../node_modules/.pnpm/safe-stable-stringify@2.5.0/node_modules/safe-stable-stringify/esm/wrapper.js
 var import__ = __toESM(require_safe_stable_stringify(), 1);
 var configure = import__.default.configure;
 var wrapper_default = import__.default;
@@ -1120,11 +1120,22 @@ function normalizeForCanonical(value) {
 var OPENROUTER_IMAGE_AR = ["1:1", "2:3", "3:2", "3:4", "4:3", "4:5", "5:4", "9:16", "16:9", "21:9"];
 var OPENROUTER_IMAGE_AR_EXTREME = [...OPENROUTER_IMAGE_AR, "1:4", "4:1", "1:8", "8:1"];
 var OPENROUTER_IMAGE_AR_GPT = ["1:1", "2:3", "3:2", "3:4", "4:3", "9:16", "16:9", "21:9"];
+var GPT_IMAGE_PROMPT_MAX = 32e3;
+var RECRAFT_PROMPT_MAX = 1e4;
+var SEEDANCE_PROMPT_MAX = 4e3;
+var KLING_PROMPT_MAX = 2500;
+var VEO_SEED_MAX = 4294967295;
+var GPT_IMAGE_MAX_REFERENCES = 16;
+var GEMINI_IMAGE_MAX_REFERENCES = 14;
+var RECRAFT_MAX_REFERENCES = 1;
+var SEEDANCE_MAX_REFERENCES = 9;
 var OPENROUTER_IMAGE_SIZE = ["1K", "2K", "4K"];
-var OPENROUTER_IMAGE_SIZE_EXTENDED = ["0.5K", ...OPENROUTER_IMAGE_SIZE];
+var OPENROUTER_IMAGE_SIZES = ["0.5K", ...OPENROUTER_IMAGE_SIZE];
+var OPENROUTER_IMAGE_SIZE_EXTENDED = OPENROUTER_IMAGE_SIZES;
 var GEMINI_LITE_IMAGE_SIZE = ["1K"];
 var OPENROUTER_IMAGE_QUALITY = ["auto", "low", "medium", "high"];
-var SEEDANCE_DURATIONS = [4, 5, 6, 8, 10, 12, 15];
+var RECRAFT_IMAGE_AR = ["1:1", "4:3", "3:4", "16:9", "9:16"];
+var SEEDANCE_DURATIONS = [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15];
 var KLING_DURATIONS = [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15];
 var GEMINI_OMNI_DURATIONS = [3, 4, 5, 6, 7, 8, 9, 10];
 var GEMINI_OMNI_DEFAULT_SECONDS = 10;
@@ -1153,6 +1164,7 @@ var IMAGE_GENERATE_MODELS = [
   "google/gemini-3-pro-image-preview",
   "recraft/recraft-v4.1-pro-vector"
 ];
+var DEFAULT_IMAGE_GENERATE_MODEL = "google/gemini-3.1-flash-image-preview";
 var VIDEO_GENERATE_MODELS = [
   "bytedance/seedance-2.0",
   "google/gemini-omni-flash",
@@ -1161,6 +1173,10 @@ var VIDEO_GENERATE_MODELS = [
   "kwaivgi/kling-v3.0-pro"
 ];
 var DEFAULT_VIDEO_GENERATE_MODEL = "google/gemini-omni-flash";
+var REFERENCE_TO_VIDEO_MODELS = ["bytedance/seedance-2.0", "google/gemini-omni-flash"];
+function supportsReferenceToVideo(model) {
+  return REFERENCE_TO_VIDEO_MODELS.includes(model);
+}
 function isGeminiOmniModel(model) {
   return model.startsWith("google/gemini-omni");
 }
@@ -1265,12 +1281,18 @@ var MODEL_REGISTRY = {
     // no `image_size` (OpenAI picks the pixel dimensions from the ratio, and a
     // `resolution` is silently ignored) and the narrower `..._AR_GPT` ratio set.
     "openai/gpt-image-2": {
+      // Its live endpoint also advertises `background` (auto|opaque) and
+      // `output_compression` (0-100), which we deliberately do not declare. Not an
+      // oversight: a registry param is a promise the whole system keeps — the
+      // canvas node schema, the dashboard control, the CLI flag and the cost model
+      // all have to carry it — and neither knob changes what the picture IS.
+      // Declaring them would widen four surfaces to expose a file-size dial.
       label: "OpenAI GPT Image 2",
       inputs: [],
-      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES }],
+      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES, max: GPT_IMAGE_MAX_REFERENCES }],
       required: ["prompt"],
       params: {
-        prompt: { kind: "string" },
+        prompt: { kind: "string", maxLength: GPT_IMAGE_PROMPT_MAX },
         aspect_ratio: { kind: "string", enum: OPENROUTER_IMAGE_AR_GPT },
         quality: { kind: "string", enum: OPENROUTER_IMAGE_QUALITY }
       }
@@ -1285,10 +1307,10 @@ var MODEL_REGISTRY = {
     "openai/gpt-5.4-image-2": {
       label: "OpenAI GPT-5.4 Image 2 (legacy)",
       inputs: [],
-      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES }],
+      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES, max: GPT_IMAGE_MAX_REFERENCES }],
       required: ["prompt"],
       params: {
-        prompt: { kind: "string" },
+        prompt: { kind: "string", maxLength: GPT_IMAGE_PROMPT_MAX },
         aspect_ratio: { kind: "string", enum: OPENROUTER_IMAGE_AR },
         image_size: { kind: "string", enum: OPENROUTER_IMAGE_SIZE },
         quality: { kind: "string", enum: OPENROUTER_IMAGE_QUALITY }
@@ -1297,7 +1319,7 @@ var MODEL_REGISTRY = {
     "google/gemini-3.1-flash-image-preview": {
       label: "Google Gemini 3.1 Flash Image (Preview)",
       inputs: [],
-      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES }],
+      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES, max: GEMINI_IMAGE_MAX_REFERENCES }],
       required: ["prompt"],
       params: {
         prompt: { kind: "string" },
@@ -1322,7 +1344,7 @@ var MODEL_REGISTRY = {
       // Source: GET /api/v1/images/models/google/gemini-3.1-flash-lite-image/endpoints.
       label: "Google Gemini 3.1 Flash Lite Image",
       inputs: [],
-      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES }],
+      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES, max: GEMINI_IMAGE_MAX_REFERENCES }],
       required: ["prompt"],
       params: {
         prompt: { kind: "string" },
@@ -1333,7 +1355,7 @@ var MODEL_REGISTRY = {
     "google/gemini-3-pro-image-preview": {
       label: "Google Gemini 3 Pro Image (Nano Banana Pro)",
       inputs: [],
-      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES }],
+      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES, max: GEMINI_IMAGE_MAX_REFERENCES }],
       required: ["prompt"],
       params: {
         prompt: { kind: "string" },
@@ -1345,14 +1367,24 @@ var MODEL_REGISTRY = {
       }
     },
     "recraft/recraft-v4.1-pro-vector": {
+      // The vector/logo tier, and the odd one out of this roster in three ways.
+      // All three are read off its live endpoint schema, and all three are the
+      // kind of fact that only shows up as a billed 400 if it is guessed:
+      //  - it renders FIVE ratios (no 4:5 / 5:4 / 2:3 / 3:2 / 21:9), so portrait
+      //    ad shapes snap to 3:4 rather than being sent as-is;
+      //  - it has NO resolution knob (`supported_parameters` lists none), so
+      //    `image_size` is refused here instead of accepted and ignored;
+      //  - it takes exactly ONE reference image, against 14-16 everywhere else.
+      // Its palette levers are real but arrive as provider passthrough
+      // (`controls`), which is why they are `json` rather than typed enums; the
+      // backend adapter builds that block in `buildImageGenBody`.
       label: "Recraft v4.1 Pro Vector",
       inputs: [],
-      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES }],
+      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES, max: RECRAFT_MAX_REFERENCES }],
       required: ["prompt"],
       params: {
-        prompt: { kind: "string" },
-        aspect_ratio: { kind: "string", enum: OPENROUTER_IMAGE_AR },
-        image_size: { kind: "string", enum: OPENROUTER_IMAGE_SIZE },
+        prompt: { kind: "string", maxLength: RECRAFT_PROMPT_MAX },
+        aspect_ratio: { kind: "string", enum: RECRAFT_IMAGE_AR },
         strength: { kind: "number", min: 0, max: 1 },
         rgb_colors: { kind: "json" },
         background_rgb_color: { kind: "json" }
@@ -1437,19 +1469,22 @@ var MODEL_REGISTRY = {
       // wrongly recommended and which sends you round the same loop.
       label: "ByteDance Seedance 2.0",
       inputs: [],
-      optional_inputs: [{ kind: "image", mimes: REPLICATE_IMAGE_MIMES }],
+      optional_inputs: [{ kind: "image", mimes: REPLICATE_IMAGE_MIMES, max: SEEDANCE_MAX_REFERENCES }],
       required: ["prompt"],
       params: {
         // Replicate's Seedance wrapper hard-caps the prompt at 4000 chars; gate
         // it here so an over-length prompt fails validate (free) not the billed call.
-        prompt: { kind: "string", maxLength: 4e3 },
+        prompt: { kind: "string", maxLength: SEEDANCE_PROMPT_MAX },
         aspect_ratio: {
           kind: "string",
           enum: ["1:1", "3:4", "9:16", "4:3", "16:9", "21:9", "9:21"]
         },
         resolution: { kind: "string", enum: ["480p", "720p", "1080p", "4k"] },
         duration: { kind: "number", enum: SEEDANCE_DURATIONS },
-        seed: { kind: "number" },
+        // Replicate types this `integer`, and a fractional seed is a schema
+        // rejection rather than a rounded one. No range is advertised, so none is
+        // invented — see `InputSpec.max` on why absent must stay ungated.
+        seed: { kind: "number", integer: true },
         generate_audio: { kind: "boolean" }
       }
     },
@@ -1505,7 +1540,8 @@ var MODEL_REGISTRY = {
         aspect_ratio: { kind: "string", enum: ["16:9", "9:16"] },
         resolution: { kind: "string", enum: ["720p", "1080p", "4K"] },
         duration: { kind: "number", enum: [4, 6, 8] },
-        seed: { kind: "number" },
+        // Vertex declares `seed` a uint32: "accepted range 0-4294967295".
+        seed: { kind: "number", integer: true, min: 0, max: VEO_SEED_MAX },
         generate_audio: { kind: "boolean" },
         person_generation: { kind: "string", enum: ["allow_all", "allow_adult"] },
         enhance_prompt: { kind: "boolean" },
@@ -1529,9 +1565,10 @@ var MODEL_REGISTRY = {
       required: ["prompt"],
       params: {
         // Kling caps the prompt shorter than Seedance; gate it here so an
-        // over-length prompt fails validate (free) not the billed call.
-        prompt: { kind: "string", maxLength: 2500 },
-        negative_prompt: { kind: "string" },
+        // over-length prompt fails validate (free) not the billed call. Its
+        // negative prompt is a SEPARATE 2,500 budget, not a shared one.
+        prompt: { kind: "string", maxLength: KLING_PROMPT_MAX },
+        negative_prompt: { kind: "string", maxLength: KLING_PROMPT_MAX },
         aspect_ratio: { kind: "string", enum: ["16:9", "9:16", "1:1"] },
         resolution: { kind: "string", enum: ["720p"] },
         duration: { kind: "number", enum: KLING_DURATIONS },
@@ -1552,7 +1589,7 @@ var MODEL_REGISTRY = {
         aspect_ratio: { kind: "string", enum: ["16:9", "9:16"] },
         resolution: { kind: "string", enum: ["720p", "1080p", "4K"] },
         duration: { kind: "number", enum: [4, 6, 8] },
-        seed: { kind: "number" },
+        seed: { kind: "number", integer: true, min: 0, max: VEO_SEED_MAX },
         generate_audio: { kind: "boolean" },
         // Image-to-video and EU/UK/CH/MENA regions cap this at `allow_adult`;
         // `allow_all` is text-to-video only. Allow both so an image-conditioned
@@ -1715,6 +1752,22 @@ var MODEL_REGISTRY = {
 function supportsParam(kind, model, param) {
   return MODEL_REGISTRY[kind]?.[model]?.params[param] !== void 0;
 }
+var DEFAULT_CLIP_DURATION_S = 5;
+function nearestClipDuration(model, preferred) {
+  const schema = MODEL_REGISTRY.video_generate[model]?.params.duration;
+  const allowed = schema?.kind === "number" ? schema.enum : void 0;
+  if (!allowed?.length || allowed.includes(preferred)) return preferred;
+  return [...allowed].sort((a, b) => Math.abs(a - preferred) - Math.abs(b - preferred) || a - b)[0];
+}
+function promptMaxLength(kind, model) {
+  const schema = MODEL_REGISTRY[kind]?.[model]?.params.prompt;
+  return schema?.kind === "string" ? schema.maxLength : void 0;
+}
+function maxInputReferences(kind, model) {
+  const spec = MODEL_REGISTRY[kind]?.[model];
+  const slot = spec?.optional_inputs?.find((input) => input.kind === "image") ?? spec?.inputs.find((input) => input.kind === "image");
+  return slot?.max;
+}
 function supportedAspectRatios(kind, model) {
   const schema = MODEL_REGISTRY[kind]?.[model]?.params.aspect_ratio;
   return schema?.kind === "string" ? schema.enum : void 0;
@@ -1859,6 +1912,9 @@ function validateNumberValue(key, value, schema, model) {
       message: `${key}: ${value} not allowed by "${model}" (allowed: ${schema.enum.join(", ")})`
     };
   }
+  if (schema.integer && !Number.isInteger(value)) {
+    return { code: "invalid_value", param: key, message: `${key}: ${value} must be a whole number` };
+  }
   if (schema.min !== void 0 && value < schema.min) {
     return { code: "invalid_value", param: key, message: `${key}: ${value} below min ${schema.min}` };
   }
@@ -7258,7 +7314,7 @@ var ImageGenerateParams = z17.object({
   model: imageModelEnum,
   prompt: z17.string().min(1),
   aspect_ratio: z17.enum(["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3", "4:5", "5:4", "21:9", "1:4", "4:1", "1:8", "8:1"]).optional(),
-  image_size: z17.enum(["0.5K", "1K", "2K", "4K"]).optional(),
+  image_size: z17.enum(OPENROUTER_IMAGE_SIZES).optional(),
   // Rendering quality. OpenRouter models without a quality knob ignore it; the
   // registry gates which models accept it (gpt-image, Gemini).
   quality: z17.enum(["auto", "low", "medium", "high"]).optional(),
@@ -7292,7 +7348,7 @@ var ImageAspectAdaptParams = z18.object({
   formats: z18.array(z18.enum(ASPECT_ADAPT_FORMATS)).min(1).max(6).refine((formats) => new Set(formats).size === formats.length, { message: "formats must be unique" }).optional(),
   platform: z18.enum(AD_FORMAT_PLATFORMS).optional(),
   guidance: z18.string().min(1).optional(),
-  image_size: z18.enum(["0.5K", "1K", "2K", "4K"]).optional()
+  image_size: z18.enum(OPENROUTER_IMAGE_SIZES).optional()
 }).strict().refine((p) => p.formats === void 0 !== (p.platform === void 0), {
   message: "provide exactly one of `formats` (explicit ratios) or `platform` (a placement preset)",
   path: ["formats"]
@@ -7378,7 +7434,7 @@ var ImageReferenceSheetParams = z21.object({
   style: z21.string().optional(),
   prompt_override: z21.string().min(1).optional(),
   aspect_ratio: z21.enum(["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3", "4:5", "5:4", "21:9", "1:4", "4:1", "1:8", "8:1"]).optional(),
-  image_size: z21.enum(["0.5K", "1K", "2K", "4K"]).optional()
+  image_size: z21.enum(OPENROUTER_IMAGE_SIZES).optional()
 }).strict();
 var imageReferenceSheetNode = delegated({
   id: "image_reference_sheet",
@@ -8148,12 +8204,20 @@ export {
   SEEDANCE_DURATIONS,
   ELEVENLABS_MAX_MUSIC_LENGTH_MS,
   IMAGE_GENERATE_MODELS,
+  DEFAULT_IMAGE_GENERATE_MODEL,
+  VIDEO_GENERATE_MODELS,
   DEFAULT_VIDEO_GENERATE_MODEL,
+  supportsReferenceToVideo,
   supportsLastFrame,
   MODEL_REGISTRY,
   supportsParam,
+  DEFAULT_CLIP_DURATION_S,
+  nearestClipDuration,
+  promptMaxLength,
+  maxInputReferences,
   nearestSupportedAspectRatio,
   nearestSupportedImageSize,
+  estimateVideoCredits,
   resolveConcurrency,
   ulid,
   isPersistedAssetRef,
@@ -8180,4 +8244,4 @@ export {
   defaultRegistry,
   createEngineFromEnv
 };
-//# sourceMappingURL=chunk-UZ37VVP4.js.map
+//# sourceMappingURL=chunk-3TXZKSAR.js.map