@shop-os/modelkit 0.1.7 → 0.1.9

package/CHANGELOG.md

@@ -0,0 +1,166 @@
+# Changelog
+
+All notable changes to `@shop-os/modelkit`. The format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Dates are npm publish dates.
+
+A release is a PR that touches only `packages/modelkit/**` and bumps `version`; merging it to
+`dev` publishes. Add an entry under **Unreleased** with the change, and move it under the version
+header in the release PR.
+
+## [Unreleased]
+
+## [0.1.9] - unreleased (publish pending)
+
+### Added
+- `display.sortOrder`: pickers show lower values first, then catalog order; set it to surface a
+  new model at the top.
+- `usage.slug` and `usage.providerLabel`: an app's historic usage-event names for a row, so
+  metering continuity no longer needs a per-model table in the app.
+- Text rows carry their stream tuning: `providerOptions` (Anthropic thinking / adaptive, OpenAI
+  reasoning summaries, DeepSeek thinking), `capabilities.reasoning`, and `defaults.temperature`
+  (Kimi rows).
+- `constraints.exactSizes` on the gpt-image-2 rows: the historic WxH per aspect ratio that Cowork
+  sends as `exactSize`.
+
+### Changed
+- `deepseek-flash` (DeepSeek V4.1 Flash) replaces the `deepseek-v4-flash` row; the old id is an
+  alias, so stored selections and plan lists keep resolving. DeepSeek retired V4 Flash upstream and
+  serves the old id from V4.1 Flash at Flash rates.
+- `resolveVideoOptions` enforces conditioning compatibility: a `sourceVideo` on a row without
+  `capabilities.acceptsSourceVideo`, or frames next to references on a row with
+  `framesExcludeReferences`, is an `unsupported_option` error, or is dropped with a warning
+  under `onUnsupported: "coerce"`.
+
+## [0.1.8] - 2026-09-11
+
+### Added
+- `exposure: "catalog" | "internal"` is required on every row. `catalog` rows are offered in
+  product pickers (Creative Studio, chat, Space Builder); `internal` rows are callable by id
+  only. 47 rows are `catalog`. A snapshot test pins the set.
+- `display.costTier` (1 to 3) and `display.recommended` on text rows, for the chat picker.
+- `constraints.qualityOptions` on rows that support `quality`; `constraints.durationPresets`
+  where the UI offers a short list over a duration range. `resolveOptions` ignores both.
+- `@shop-os/modelkit/catalog`: a browser-safe subpath with `Catalog`, `MODELS`, the per-type
+  arrays, `resolveOptions`, the error types and every type. No adapter, client, key handling
+  or provider SDK is reachable from it; a test walks the import graph.
+- Alias `meta-muse` on `muse-image-1.0`.
+
+### Changed
+- Labels: `gemini-3.1-flash-image-preview` is "Nano Banana 2" again and `gpt-image-2` is
+  "GPT-image-2" (the 0.1.7 labels read as different models in the picker).
+- `Catalog` rejects duplicate ids and aliases across rows.
+- `veo-3.1-lite` offers 720p and 1080p only; Google publishes no 4k rate for it.
+
+## [0.1.7] - 2026-09-10
+
+### Added
+- `display { provider, description, logo? }` and `capabilities` (`imageGenerate`,
+  `imageEdit`, `acceptsSourceVideo`, `framesExcludeReferences`, `vision`, `tools`,
+  `structuredOutput`, `contextWindow`) on every active row, so a UI can render a picker from
+  `kit.catalog.list()` without a hand-written catalog. The nine Cowork picker rows keep their
+  descriptions and logos verbatim.
+
+## [0.1.6] - 2026-09-10
+
+### Added
+- `MediaRef.label`: a text part before the media on Gemini image and Omni.
+- `ImageRequest.temperature` (Gemini image) and `ImageRequest.upscale` (Topaz enhance model,
+  subject detection, face enhancement with creativity and strength; face enhancement off by
+  default).
+- `VideoRequest.watermark` (Seedance) and `VideoRequest.safetyChecker` (fal MiniMax).
+- MiniMax base rows accept `lastFrame`; Seedream 4.5 gets the 4K tier and 9:21.
+
+### Changed
+- `onUnsupported: "error"` now covers aspect ratio and duration as well as size and
+  resolution. **Breaking** for callers that relied on silent snapping: pass `"coerce"`.
+- Aborts are never retried and are forwarded to OpenAI, Replicate, Prodia, Ark and reference
+  downloads. One retry policy across adapters.
+- Gateway token is required only for vendors that go through the gateway; the
+  enterprise-to-pro key fallback is logged.
+- Seedance poll keeps usage, duration, fps and resolution on `raw`; the token split lands on
+  the usage event. Grok sends `resolution` only when the caller set it.
+- The unused OpenRouter text route is removed.
+
+## [0.1.5] - 2026-09-10
+
+### Fixed
+- `size: "512"` reaches the adapter (Gemini 3.1 Flash requests were generated at 1K).
+- Every status-less SDK throw is a retryable transport error; five adapters had classified
+  them as permanent.
+- Grok `firstFrame` maps to the SDK's first frame (image-to-video); references capped at 7.
+- Omni: `delivery: "uri"` and `store: true`, `firstFrame` runs image-to-video, references run
+  reference-to-video, no `resolution` on 1.0, uploads deleted after create.
+
+### Changed
+- `constraints.supports` undefined means "supports nothing"; every row carries an explicit list.
+- `aspectRatio: "adaptive"` means "send no ratio" and is never coerced or defaulted.
+- Usage `unit` comes from the catalog row; adapters supply quantities.
+- Gemini `safetySettings` are per-call opt-in.
+- fal direct video is queue submit plus poll, only the submit is retried; the fal client takes
+  the kit's `fetch`. Cloudflare gateway slugs `anthropic/v1` and `grok/v1`.
+
+## [0.1.4] - 2026-09-10
+
+### Added
+- `topaz` row takes references and `scaleFactor` and returns `outputUrl` next to the bytes.
+- Gemini image usage events carry input and output token counts.
+- Aliases `claude-haiku-4` and `claude-sonnet-4-5-20250929`.
+
+## [0.1.3] - 2026-09-10
+
+### Fixed
+- The remaining medium findings from the adoption review.
+
+## [0.1.2] - 2026-09-10
+
+### Added
+- `gpt-image-2.5-sunburst` and `gpt-image-2.5-flare` with the `xhigh` and `max` quality tiers.
+- Aliases `gpt-5.1-mini` and `gpt-5.2-mini` on `gpt-5.4-mini`.
+
+### Changed
+- Google text goes through the gateway at `/google-ai-studio/v1beta`.
+- Reasoning defaults are dropped from the text catalog.
+- Seedance and Kling Omni keep last frames and seeds; Google transport failures retry.
+
+### Fixed
+- The seven blocking findings from the adoption review.
+
+## [0.1.1] - 2026-09-09
+
+### Added
+- `generateText`, `streamText`, `generateObject`, `streamObject` with catalog models. The
+  package re-exports every `ai` export, so `from "ai"` becomes `from "@shop-os/modelkit"`.
+- `generateAudio` (Lyria 3, Lyria 2, MMAudio) and `prepareMediaPart` for video understanding.
+- `generateImage` accepts the `upscale` and `bg-remove` catalog types, `n`, exact `WxH` sizes.
+- Catalog rows for every id the apps still sent to direct SDKs; retired Claude 4.0 ids alias to
+  Sonnet 4.6 and Opus 4.8; fal passthrough with a gateway-stored key.
+
+### Fixed
+- Text keys fall back per provider; dotted Claude aliases; native DeepSeek, Moonshot and
+  Perplexity providers; AI SDK-shaped `providerOptions`.
+- gpt-image-2 edit sent `input_fidelity`, which the API rejects; a filtered Gemini prompt maps
+  to `content_blocked`; `@google/genai` no longer switches to Vertex from the host env.
+
+## [0.1.0] - 2026-09-09
+
+First publish, renamed from `@shop-os/model-sdk`.
+
+- Catalog of 74 models. The id is the pro-backend pricing key; every historical id is an alias.
+- `resolveOptions` fills defaults, snaps the aspect ratio, errors or coerces on unsupported
+  size, resolution and duration, trims references.
+- Ten adapters by provider API: AI SDK text, Gemini image, gpt-image-2, Google GenAI (Veo,
+  Omni, Lyria 3, video understanding), Ark, fal, Replicate, Prodia, Higgsfield, xAI video.
+- Keys per tier (pro, enterprise with per-provider fallback); one retry classifier, one error
+  type, one usage event with the canonical id.
+
+[Unreleased]: https://github.com/Shop-OS/shopos-backend-ts/compare/modelkit-0.1.9...dev
+[0.1.9]: https://github.com/Shop-OS/shopos-backend-ts/compare/modelkit-0.1.8...dev
+[0.1.8]: https://github.com/Shop-OS/shopos-backend-ts/commit/9661a109d
+[0.1.7]: https://github.com/Shop-OS/shopos-backend-ts/commit/2d3d19ac2
+[0.1.6]: https://github.com/Shop-OS/shopos-backend-ts/commit/f67fee08f
+[0.1.5]: https://github.com/Shop-OS/shopos-backend-ts/commit/7f3f745ec
+[0.1.4]: https://github.com/Shop-OS/shopos-backend-ts/commit/a2f3ffcc7
+[0.1.3]: https://github.com/Shop-OS/shopos-backend-ts/commit/785960dbd
+[0.1.2]: https://github.com/Shop-OS/shopos-backend-ts/commit/4581615e6
+[0.1.1]: https://github.com/Shop-OS/shopos-backend-ts/commit/c159c97d3
+[0.1.0]: https://github.com/Shop-OS/shopos-backend-ts/commit/6b70c91cf

package/README.md

@@ -53,6 +53,18 @@ await clip.cleanup?.();
 
 Keys come from `process.env` by default: bare names for pro (`OPENAI_API_KEY`, `FAL_KEY`, `ARK_API_KEY` ...) and the `ENTERPRISE_` prefix for enterprise. Pass `keys: { pro, enterprise }` to override.
 
+## Browser-safe catalog
+
+```ts
+import { Catalog, MODELS, resolveOptions } from "@shop-os/modelkit/catalog";
+```
+
+The catalog subpath carries the model rows (`MODELS` plus the per-type
+arrays), `Catalog`, the option resolvers (`resolveOptions` and the per-type
+resolvers), `ModelKitError` and the catalog types. It includes no adapters, no
+client, no keys, no text helpers and no provider SDKs, and builds for
+`es2022` / neutral so it bundles straight into the browser.
+
 ## A superset of `ai`
 
 Every export of `ai` (the Vercel AI SDK) is re-exported here, so this import swap is
@@ -169,3 +181,7 @@ Prompt construction, agent tools, node schemas and UI, S3 upload, credit reserve
 ## Live smoke test
 
 `pnpm smoke [filter...]` runs one cheap real call per adapter (about $5 for the full matrix, 15 minutes, mostly video polling). Keys load from the `trigger-workflows`, `cowork`, and `vibe-triggers` env files next to this monorepo. Outputs land in `scripts/smoke-out/`. Run it before every release and after any adapter change. Never wire it into CI.
+
+## Changelog
+
+See [CHANGELOG.md](./CHANGELOG.md). Add an entry under Unreleased with every change; the release PR moves it under the version header.