@tanstack/ai-byteplus 0.0.0

package/src/model-meta.ts

@@ -0,0 +1,954 @@
+/**
+ * BytePlus ModelArk model metadata.
+ *
+ * Every Ark model id in this file — chat, video and image — was verified live
+ * against `https://ark.ap-southeast.bytepluses.com/api/v3` on 2026-07-31. The
+ * two Seed Speech ids are the exception: they live on the voice host, which
+ * needs a separate key that was not available, so they are docs-derived.
+ * Capability metadata is a mix of probed and docs-derived facts; anything not
+ * confirmed against the live API is annotated as such at its declaration.
+ * BytePlus
+ * deactivates model ids aggressively (the whole `seedance-1-0-lite-*` family,
+ * `seed-1-6-lite-*`, `seedream-3-0-*`, and the `doubao-`/`skylark-` names are
+ * all 404s internationally), so only dated, probe-confirmed ids are shipped.
+ *
+ * Prefix rules, also probe-confirmed:
+ * - `dola-seed-2-1-turbo-260628` and `dola-seedream-5-0-pro-260628` are the
+ *   canonical ids; the bare forms resolve as aliases but the API echoes the
+ *   prefixed id back.
+ * - The Seedance 2.0 family *requires* the `dreamina-` prefix.
+ * - Older models reject the `dola-` prefix outright.
+ */
+import type { DurationOptions } from '@tanstack/ai/adapters'
+import type { BytePlusTextProviderOptions } from './text/text-provider-options'
+
+/**
+ * BytePlus exposes no server-side provider tools (no hosted web search, code
+ * interpreter, …) on the international Ark endpoint, so every chat model
+ * advertises an empty tool set. Typing it as `never` makes passing another
+ * provider's `ProviderTool` to a BytePlus adapter a compile-time error.
+ */
+export type BytePlusProviderToolKind = never
+
+/**
+ * Internal metadata structure describing a BytePlus model.
+ */
+interface ModelMeta {
+  name: string
+  supports: {
+    input: ReadonlyArray<'text' | 'image' | 'audio' | 'video' | 'document'>
+    output: ReadonlyArray<'text' | 'image' | 'audio' | 'video'>
+    capabilities?: ReadonlyArray<
+      'reasoning' | 'tool_calling' | 'structured_outputs'
+    >
+    tools?: ReadonlyArray<BytePlusProviderToolKind>
+  }
+  context_window?: number
+  max_input_tokens?: number
+  max_output_tokens?: number
+}
+
+// ============================================================================
+// Chat models (Seed / GLM / DeepSeek / gpt-oss on Ark)
+// ============================================================================
+
+const DOLA_SEED_2_1_TURBO = {
+  name: 'dola-seed-2-1-turbo-260628',
+  context_window: 256_000,
+  max_input_tokens: 256_000,
+  max_output_tokens: 256_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_2_0_LITE_260428 = {
+  name: 'seed-2-0-lite-260428',
+  context_window: 256_000,
+  max_input_tokens: 256_000,
+  max_output_tokens: 128_000,
+  supports: {
+    input: ['text', 'image', 'video', 'audio'],
+    output: ['text'],
+    // Live-probed 2026-07-31: rejects both json_schema and json_object.
+    capabilities: ['reasoning', 'tool_calling'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_2_0_MINI_260428 = {
+  name: 'seed-2-0-mini-260428',
+  context_window: 256_000,
+  max_input_tokens: 256_000,
+  max_output_tokens: 128_000,
+  supports: {
+    input: ['text', 'image', 'video', 'audio'],
+    output: ['text'],
+    // Live-probed 2026-07-31: rejects both json_schema and json_object.
+    capabilities: ['reasoning', 'tool_calling'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_2_0_PRO_260328 = {
+  name: 'seed-2-0-pro-260328',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 128_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    // Live-probed 2026-07-31: accepts json_schema, despite the docs table
+    // saying otherwise.
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_2_0_LITE_260228 = {
+  name: 'seed-2-0-lite-260228',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 128_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_2_0_MINI_260215 = {
+  name: 'seed-2-0-mini-260215',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 128_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_2_0_CODE_PREVIEW_260328 = {
+  name: 'seed-2-0-code-preview-260328',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 128_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_1_8_251228 = {
+  name: 'seed-1-8-251228',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 64_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_1_6_250915 = {
+  name: 'seed-1-6-250915',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 32_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_1_6_250615 = {
+  name: 'seed-1-6-250615',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 32_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_1_6_FLASH_250715 = {
+  name: 'seed-1-6-flash-250715',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 32_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const SEED_1_6_FLASH_250615 = {
+  name: 'seed-1-6-flash-250615',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 32_000,
+  supports: {
+    input: ['text', 'image', 'video'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const GLM_5_2_260617 = {
+  name: 'glm-5-2-260617',
+  context_window: 1_024_000,
+  max_input_tokens: 1_024_000,
+  max_output_tokens: 128_000,
+  supports: {
+    input: ['text'],
+    output: ['text'],
+    // Live-probed 2026-07-31: accepts json_schema, despite the docs table
+    // saying otherwise.
+    capabilities: ['reasoning', 'tool_calling', 'structured_outputs'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const GLM_4_7_251222 = {
+  name: 'glm-4-7-251222',
+  context_window: 256_000,
+  max_input_tokens: 224_000,
+  max_output_tokens: 128_000,
+  supports: {
+    input: ['text'],
+    output: ['text'],
+    // Adherence-probed 2026-07-31: ACCEPTS a json_schema with 200 but ignores
+    // it and answers in prose, so it is not a structured-output model. A
+    // status-code-only probe reads this as support — see the note on
+    // BYTEPLUS_STRUCTURED_OUTPUT_CHAT_MODELS.
+    capabilities: ['reasoning', 'tool_calling'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const DEEPSEEK_V4_PRO_260425 = {
+  name: 'deepseek-v4-pro-260425',
+  context_window: 1_024_000,
+  max_input_tokens: 1_024_000,
+  max_output_tokens: 384_000,
+  supports: {
+    input: ['text'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+const DEEPSEEK_V4_FLASH_260425 = {
+  name: 'deepseek-v4-flash-260425',
+  context_window: 1_024_000,
+  max_input_tokens: 1_024_000,
+  max_output_tokens: 384_000,
+  supports: {
+    input: ['text'],
+    output: ['text'],
+    capabilities: ['reasoning', 'tool_calling'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+// The one model on Ark that defaults to `thinking: disabled`.
+const DEEPSEEK_V3_2_251201 = {
+  name: 'deepseek-v3-2-251201',
+  context_window: 128_000,
+  max_input_tokens: 128_000,
+  max_output_tokens: 32_000,
+  supports: {
+    input: ['text'],
+    output: ['text'],
+    // Live-probed 2026-07-31: rejects both json_schema and json_object.
+    capabilities: ['reasoning', 'tool_calling'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+// The only model accepting `thinking: {type: 'auto'}`. Tool calling is
+// undocumented on Ark and unverified, so it is not advertised.
+const GPT_OSS_120B_250805 = {
+  name: 'gpt-oss-120b-250805',
+  context_window: 128_000,
+  max_input_tokens: 96_000,
+  max_output_tokens: 64_000,
+  supports: {
+    input: ['text'],
+    output: ['text'],
+    capabilities: ['reasoning'],
+    tools: [] as const,
+  },
+} as const satisfies ModelMeta
+
+/**
+ * All supported BytePlus chat model identifiers.
+ */
+export const BYTEPLUS_CHAT_MODELS = [
+  DOLA_SEED_2_1_TURBO.name,
+  SEED_2_0_LITE_260428.name,
+  SEED_2_0_MINI_260428.name,
+  SEED_2_0_PRO_260328.name,
+  SEED_2_0_LITE_260228.name,
+  SEED_2_0_MINI_260215.name,
+  SEED_2_0_CODE_PREVIEW_260328.name,
+  SEED_1_8_251228.name,
+  SEED_1_6_250915.name,
+  SEED_1_6_250615.name,
+  SEED_1_6_FLASH_250715.name,
+  SEED_1_6_FLASH_250615.name,
+  GLM_5_2_260617.name,
+  GLM_4_7_251222.name,
+  DEEPSEEK_V4_PRO_260425.name,
+  DEEPSEEK_V4_FLASH_260425.name,
+  DEEPSEEK_V3_2_251201.name,
+  GPT_OSS_120B_250805.name,
+] as const
+
+/**
+ * Union of all supported BytePlus chat model names.
+ */
+export type BytePlusChatModel = (typeof BYTEPLUS_CHAT_MODELS)[number]
+
+/**
+ * Chat models that emit a `encrypted_content` blob alongside
+ * `reasoning_content` when thinking is enabled ("thinking summary" models).
+ *
+ * The blob is an opaque signature over the reasoning trace: when it is
+ * present it must be echoed back verbatim on the assistant message in the
+ * next turn. Live probing showed omitting it did *not* fail a simple tool
+ * round-trip, so adapters preserve and replay it when present but must never
+ * treat its absence as an error.
+ */
+export const BYTEPLUS_THINKING_SUMMARY_MODELS = [
+  DOLA_SEED_2_1_TURBO.name,
+  SEED_2_0_LITE_260428.name,
+  SEED_2_0_MINI_260428.name,
+  SEED_2_0_PRO_260328.name,
+] as const
+
+/**
+ * Union of chat models that emit `encrypted_content`.
+ */
+export type BytePlusThinkingSummaryModel =
+  (typeof BYTEPLUS_THINKING_SUMMARY_MODELS)[number]
+
+const THINKING_SUMMARY_MODEL_SET: ReadonlySet<string> = new Set(
+  BYTEPLUS_THINKING_SUMMARY_MODELS,
+)
+
+/**
+ * True when the model emits `encrypted_content` that should be round-tripped
+ * on subsequent turns.
+ */
+export function emitsEncryptedContent(model: string): boolean {
+  return THINKING_SUMMARY_MODEL_SET.has(model)
+}
+
+/**
+ * Chat models that accept `response_format: {type: 'json_schema'}`.
+ *
+ * Live-probed against all 18 chat models on 2026-07-31, not docs-derived — the
+ * BytePlus capability tables are wrong here in both directions.
+ *
+ * Membership needs TWO things, because the API has both failure modes:
+ * 1. The request is accepted. Seven models (`seed-2-0-lite-260428`,
+ *    `seed-2-0-mini-260428`, `seed-2-0-code-preview-260328`, both
+ *    `deepseek-v4-*`, `deepseek-v3-2-251201`, `gpt-oss-120b-250805`) answer a
+ *    JSON schema with 400 InvalidParameter — and reject
+ *    `{type: 'json_object'}` too, so there is no JSON-mode fallback.
+ * 2. The schema is actually honoured. `glm-4-7-251222` accepts the request
+ *    with 200 and then ignores the schema, answering in prose (reproduced
+ *    twice by the adherence probe). A status-code-only probe wrongly reads
+ *    that as support, so it is excluded.
+ *
+ * Models that fail either check need tool-shaped extraction instead.
+ *
+ * Note that the default chat model `seed-2-0-lite-260428` is one of the
+ * rejecting models: structured-output work needs `seed-2-0-lite-260228` or
+ * `dola-seed-2-1-turbo-260628`.
+ */
+export const BYTEPLUS_STRUCTURED_OUTPUT_CHAT_MODELS = [
+  DOLA_SEED_2_1_TURBO.name,
+  SEED_2_0_PRO_260328.name,
+  SEED_2_0_LITE_260228.name,
+  SEED_2_0_MINI_260215.name,
+  SEED_1_8_251228.name,
+  SEED_1_6_250915.name,
+  SEED_1_6_250615.name,
+  SEED_1_6_FLASH_250715.name,
+  SEED_1_6_FLASH_250615.name,
+  GLM_5_2_260617.name,
+] as const
+
+const STRUCTURED_OUTPUT_MODEL_SET: ReadonlySet<string> = new Set(
+  BYTEPLUS_STRUCTURED_OUTPUT_CHAT_MODELS,
+)
+
+/**
+ * True when the model supports native JSON-schema structured output.
+ */
+export function supportsStructuredOutput(model: string): boolean {
+  return STRUCTURED_OUTPUT_MODEL_SET.has(model)
+}
+
+/**
+ * Type-only map from chat model name to whether it supports native
+ * JSON-schema structured output.
+ */
+export type BytePlusChatModelStructuredOutputByName = {
+  [K in BytePlusChatModel]: K extends BytePlusStructuredOutputChatModel
+    ? true
+    : false
+}
+
+/**
+ * Union of chat models supporting native JSON-schema structured output.
+ */
+export type BytePlusStructuredOutputChatModel =
+  (typeof BYTEPLUS_STRUCTURED_OUTPUT_CHAT_MODELS)[number]
+
+/**
+ * Type-only map from chat model name to its supported input modalities.
+ * Used for type inference when constructing multimodal messages.
+ */
+export type BytePlusModelInputModalitiesByName = {
+  [DOLA_SEED_2_1_TURBO.name]: typeof DOLA_SEED_2_1_TURBO.supports.input
+  [SEED_2_0_LITE_260428.name]: typeof SEED_2_0_LITE_260428.supports.input
+  [SEED_2_0_MINI_260428.name]: typeof SEED_2_0_MINI_260428.supports.input
+  [SEED_2_0_PRO_260328.name]: typeof SEED_2_0_PRO_260328.supports.input
+  [SEED_2_0_LITE_260228.name]: typeof SEED_2_0_LITE_260228.supports.input
+  [SEED_2_0_MINI_260215.name]: typeof SEED_2_0_MINI_260215.supports.input
+  [SEED_2_0_CODE_PREVIEW_260328.name]: typeof SEED_2_0_CODE_PREVIEW_260328.supports.input
+  [SEED_1_8_251228.name]: typeof SEED_1_8_251228.supports.input
+  [SEED_1_6_250915.name]: typeof SEED_1_6_250915.supports.input
+  [SEED_1_6_250615.name]: typeof SEED_1_6_250615.supports.input
+  [SEED_1_6_FLASH_250715.name]: typeof SEED_1_6_FLASH_250715.supports.input
+  [SEED_1_6_FLASH_250615.name]: typeof SEED_1_6_FLASH_250615.supports.input
+  [GLM_5_2_260617.name]: typeof GLM_5_2_260617.supports.input
+  [GLM_4_7_251222.name]: typeof GLM_4_7_251222.supports.input
+  [DEEPSEEK_V4_PRO_260425.name]: typeof DEEPSEEK_V4_PRO_260425.supports.input
+  [DEEPSEEK_V4_FLASH_260425.name]: typeof DEEPSEEK_V4_FLASH_260425.supports.input
+  [DEEPSEEK_V3_2_251201.name]: typeof DEEPSEEK_V3_2_251201.supports.input
+  [GPT_OSS_120B_250805.name]: typeof GPT_OSS_120B_250805.supports.input
+}
+
+/**
+ * Type-only map from chat model name to its supported provider tools.
+ * BytePlus exposes no provider-tool factories, so every model gets an empty
+ * tuple — passing another provider's tool is then a compile-time error.
+ */
+export type BytePlusChatModelToolCapabilitiesByName = {
+  [K in BytePlusChatModel]: ReadonlyArray<BytePlusProviderToolKind>
+}
+
+/**
+ * Type-only map from chat model name to its provider options type.
+ */
+export type BytePlusChatModelProviderOptionsByName = {
+  [K in BytePlusChatModel]: BytePlusTextProviderOptions
+}
+
+// ============================================================================
+// Video models (Seedance, async task API)
+// ============================================================================
+
+/**
+ * Output aspect ratios accepted by the Seedance task API. `adaptive` is only
+ * meaningful for image-to-video, where the ratio follows the input frame.
+ */
+export type BytePlusVideoRatio =
+  | '16:9'
+  | '9:16'
+  | '4:3'
+  | '3:4'
+  | '1:1'
+  | '21:9'
+  | 'adaptive'
+
+/**
+ * Resolution tiers accepted by the Seedance task API.
+ *
+ * All four are probe-verified per model (2026-07-31). Two findings contradict
+ * the BytePlus prose docs: there is **no 2K tier on any Seedance model** —
+ * `2k`/`2K` is rejected everywhere, including on the 2.0 flagship documented
+ * as reaching 4K — and `4k` exists only on `dreamina-seedance-2-0-260128`.
+ *
+ * The API matches this field case-insensitively (`4K`, `4k` and `1080P` are
+ * all accepted), so this package standardizes on the lowercase spelling.
+ */
+export type BytePlusVideoResolution = '480p' | '720p' | '1080p' | '4k'
+
+/**
+ * Generic `size` template for Seedance models: either a bare aspect ratio
+ * ("16:9") or `ratio_resolution` ("16:9_720p"). The Seedance API takes the
+ * two as separate `ratio` / `resolution` fields; the adapter splits this
+ * template back apart.
+ */
+export type BytePlusVideoSize<
+  TResolution extends BytePlusVideoResolution = BytePlusVideoResolution,
+> = BytePlusVideoRatio | `${BytePlusVideoRatio}_${TResolution}`
+
+// The Seedance 2.0 family's `audio` input modality is docs-derived, not
+// live-probed: the docs' multimodal-reference caps list `reference_audio`
+// parts (up to 3, never sent without a visual reference). Every 2.0 model id
+// below is itself probe-verified live; only the audio-reference capability
+// rests on the docs.
+const DREAMINA_SEEDANCE_2_0 = {
+  name: 'dreamina-seedance-2-0-260128',
+  supports: {
+    input: ['text', 'image', 'video', 'audio'],
+    output: ['video', 'audio'],
+  },
+} as const satisfies ModelMeta
+
+const DREAMINA_SEEDANCE_2_0_FAST = {
+  name: 'dreamina-seedance-2-0-fast-260128',
+  supports: {
+    input: ['text', 'image', 'video', 'audio'],
+    output: ['video', 'audio'],
+  },
+} as const satisfies ModelMeta
+
+const DREAMINA_SEEDANCE_2_0_MINI = {
+  name: 'dreamina-seedance-2-0-mini-260615',
+  supports: {
+    input: ['text', 'image', 'video', 'audio'],
+    output: ['video', 'audio'],
+  },
+} as const satisfies ModelMeta
+
+const SEEDANCE_1_5_PRO = {
+  name: 'seedance-1-5-pro-251215',
+  supports: {
+    input: ['text', 'image'],
+    output: ['video', 'audio'],
+  },
+} as const satisfies ModelMeta
+
+const SEEDANCE_1_0_PRO = {
+  name: 'seedance-1-0-pro-250528',
+  supports: {
+    input: ['text', 'image'],
+    output: ['video'],
+  },
+} as const satisfies ModelMeta
+
+const SEEDANCE_1_0_PRO_FAST = {
+  name: 'seedance-1-0-pro-fast-251015',
+  supports: {
+    input: ['text', 'image'],
+    output: ['video'],
+  },
+} as const satisfies ModelMeta
+
+/**
+ * All supported Seedance video model identifiers.
+ */
+export const BYTEPLUS_VIDEO_MODELS = [
+  DREAMINA_SEEDANCE_2_0.name,
+  DREAMINA_SEEDANCE_2_0_FAST.name,
+  DREAMINA_SEEDANCE_2_0_MINI.name,
+  SEEDANCE_1_5_PRO.name,
+  SEEDANCE_1_0_PRO.name,
+  SEEDANCE_1_0_PRO_FAST.name,
+] as const
+
+/**
+ * Union of all supported Seedance video model names.
+ */
+export type BytePlusVideoModel = (typeof BYTEPLUS_VIDEO_MODELS)[number]
+
+/**
+ * Type-only map from video model name to the non-text prompt modalities it
+ * accepts. The Seedance 2.0 family takes multimodal references (start/end
+ * frames, reference images, reference video and audio); the 1.x models take
+ * start/end frames only.
+ */
+export type BytePlusVideoModelInputModalitiesByName = {
+  [DREAMINA_SEEDANCE_2_0.name]: readonly ['image', 'video', 'audio']
+  [DREAMINA_SEEDANCE_2_0_FAST.name]: readonly ['image', 'video', 'audio']
+  [DREAMINA_SEEDANCE_2_0_MINI.name]: readonly ['image', 'video', 'audio']
+  [SEEDANCE_1_5_PRO.name]: readonly ['image']
+  [SEEDANCE_1_0_PRO.name]: readonly ['image']
+  [SEEDANCE_1_0_PRO_FAST.name]: readonly ['image']
+}
+
+/**
+ * Type-only map from video model name to the resolutions it accepts.
+ *
+ * Probe-verified per model on 2026-07-31. Note `seedance-1-0-pro-fast-251015`
+ * does accept `1080p`, despite the BytePlus docs listing it as 480p/720p.
+ */
+export type BytePlusVideoModelResolutionByName = {
+  [DREAMINA_SEEDANCE_2_0.name]: '480p' | '720p' | '1080p' | '4k'
+  [DREAMINA_SEEDANCE_2_0_FAST.name]: '480p' | '720p'
+  [DREAMINA_SEEDANCE_2_0_MINI.name]: '480p' | '720p'
+  [SEEDANCE_1_5_PRO.name]: '480p' | '720p' | '1080p'
+  [SEEDANCE_1_0_PRO.name]: '480p' | '720p' | '1080p'
+  [SEEDANCE_1_0_PRO_FAST.name]: '480p' | '720p' | '1080p'
+}
+
+/**
+ * Type-only map from video model name to its accepted `size` strings.
+ */
+export type BytePlusVideoModelSizeByName = {
+  [K in BytePlusVideoModel]: BytePlusVideoSize<
+    BytePlusVideoModelResolutionByName[K]
+  >
+}
+
+/**
+ * A Seedance model id: one this package knows, or any other string.
+ *
+ * The open half is a deliberate escape hatch for models BytePlus ships between
+ * releases of this package. **Seedance 2.5 is the live example.** Its real id
+ * is `dreamina-seedance-2-5-260628` — note the June date suffix, which is why
+ * guessing ids around its 2026-07-31 announcement never landed. It is absent
+ * from the table below because its capability cells are unverified, not
+ * because it is unreachable: probing it returns 404 `ModelNotOpen` ("your
+ * account has not activated the model"), so no capability question can be
+ * answered until someone enables it in the Ark Console. Passing it through
+ * the escape hatch works today for an account that has.
+ *
+ * Adding a model here *narrows* it — the adapter's guards switch on and reject
+ * against this file's tables. For a model whose real limits are unknown that
+ * is strictly worse than the open path, which lets Ark judge. So an id lands
+ * here only once probed.
+ *
+ * Discovering ids: `GET /models` on the Ark data plane enumerates the catalog
+ * (id, `task_type`, `modalities`, `status`) and is how 2.5 was found. It is
+ * not exhaustive — `seedream-5-0-lite-260128` answers requests but is missing
+ * from the listing — so absence there is not evidence of absence. The ModelArk
+ * release notes (https://docs.byteplus.com/en/docs/ModelArk/1159178) are the
+ * other watch surface.
+ *
+ * To probe an id, POST `/contents/generations/tasks` with only
+ * `{"model": "<id>"}`. Three outcomes, all live-verified:
+ * - 400 `MissingParameter` (about `content`) — live and usable.
+ * - 404 `ModelNotOpen` — real, but not activated on this account.
+ * - 404 `InvalidEndpointOrModel.NotFound` — no such model.
+ *
+ * Unknown ids trade compile-time narrowing for reach: the full size surface is
+ * accepted, provider options are ungated, and the adapter's model-specific
+ * runtime guards stand down so a new model's legitimate request reaches Ark.
+ * Known ids keep their probe-verified narrowing.
+ */
+export type BytePlusVideoModelOrString = BytePlusVideoModel | (string & {})
+
+/**
+ * Resolve the `size` type for a video model: the model's probe-verified
+ * template union when known, otherwise the full template surface plus any
+ * string (a future model may bring ratios or resolution tiers that do not
+ * exist today).
+ */
+export type ResolveBytePlusVideoSize<TModel extends string> =
+  TModel extends BytePlusVideoModel
+    ? BytePlusVideoModelSizeByName[TModel]
+    : BytePlusVideoSize | (string & {})
+
+/**
+ * Resolve the accepted non-text prompt modalities for a video model. Unknown
+ * models accept all three rather than none, so a new model's reference media
+ * is not a compile error.
+ */
+export type ResolveBytePlusVideoInputModalities<TModel extends string> =
+  TModel extends BytePlusVideoModel
+    ? BytePlusVideoModelInputModalitiesByName[TModel]
+    : readonly ['image', 'video', 'audio']
+
+const VIDEO_MODEL_SET: ReadonlySet<string> = new Set(BYTEPLUS_VIDEO_MODELS)
+
+/**
+ * True when the id is one this package has probe-verified metadata for.
+ *
+ * The adapter uses this to decide whether its model-specific guards apply:
+ * see {@link BytePlusVideoModelOrString}.
+ */
+export function isKnownBytePlusVideoModel(
+  model: string,
+): model is BytePlusVideoModel {
+  return VIDEO_MODEL_SET.has(model)
+}
+
+/**
+ * Per-model duration type. Seedance accepts any integer second inside the
+ * model's range, so this is a continuous range expressed as `number` — a
+ * literal union cannot represent it. (The API also accepts `duration: -1` on
+ * Seedance 2.0 and 1.5-pro to let the model choose; that is reachable through
+ * provider options, not through the generic `duration`.)
+ */
+export type BytePlusVideoModelDurationByName = {
+  [K in BytePlusVideoModel]: number
+}
+
+/**
+ * Runtime duration table backing `availableDurations()` / `snapDuration()`.
+ */
+export const BYTEPLUS_VIDEO_DURATIONS: {
+  readonly [TModel in BytePlusVideoModel]: DurationOptions<
+    BytePlusVideoModelDurationByName[TModel]
+  >
+} = {
+  'dreamina-seedance-2-0-260128': {
+    kind: 'range',
+    min: 4,
+    max: 15,
+    step: 1,
+    unit: 'seconds',
+  },
+  'dreamina-seedance-2-0-fast-260128': {
+    kind: 'range',
+    min: 4,
+    max: 15,
+    step: 1,
+    unit: 'seconds',
+  },
+  'dreamina-seedance-2-0-mini-260615': {
+    kind: 'range',
+    min: 4,
+    max: 15,
+    step: 1,
+    unit: 'seconds',
+  },
+  'seedance-1-5-pro-251215': {
+    kind: 'range',
+    min: 4,
+    max: 12,
+    step: 1,
+    unit: 'seconds',
+  },
+  'seedance-1-0-pro-250528': {
+    kind: 'range',
+    min: 2,
+    max: 12,
+    step: 1,
+    unit: 'seconds',
+  },
+  'seedance-1-0-pro-fast-251015': {
+    kind: 'range',
+    min: 2,
+    max: 12,
+    step: 1,
+    unit: 'seconds',
+  },
+}
+
+/**
+ * Duration hint for a model this package has no table for.
+ *
+ * Spans every range Seedance has shipped so far (2s on the 1.0 models through
+ * 15s on the 2.0 family) so `availableDurations()` can still drive a UI. It is
+ * a hint, not a contract: the adapter does **not** snap an unknown model's
+ * duration against it, because clamping a future model's legitimate 20-second
+ * request down to 15 would corrupt the request rather than protect it.
+ */
+export const BYTEPLUS_VIDEO_FALLBACK_DURATIONS: DurationOptions<number> = {
+  kind: 'range',
+  min: 2,
+  max: 15,
+  step: 1,
+  unit: 'seconds',
+}
+
+/**
+ * Look up the duration options for a Seedance video model, falling back to
+ * {@link BYTEPLUS_VIDEO_FALLBACK_DURATIONS} for an id this package does not
+ * know.
+ */
+export function getBytePlusVideoDurationOptions(
+  model: BytePlusVideoModelOrString,
+): DurationOptions<number> {
+  return isKnownBytePlusVideoModel(model)
+    ? BYTEPLUS_VIDEO_DURATIONS[model]
+    : BYTEPLUS_VIDEO_FALLBACK_DURATIONS
+}
+
+// ============================================================================
+// Image models (Seedream)
+// ============================================================================
+
+/**
+ * Shorthand size tokens accepted by `/images/generations`. A request uses
+ * either a token or an explicit `WxH` string — never both.
+ */
+export type BytePlusImageSizeToken = '1K' | '2K' | '4K'
+
+/**
+ * Accepted `size` values for Seedream models: a shorthand token or an
+ * explicit pixel size such as `2048x2048`.
+ */
+export type BytePlusImageSize = BytePlusImageSizeToken | `${number}x${number}`
+
+const DOLA_SEEDREAM_5_0_PRO = {
+  name: 'dola-seedream-5-0-pro-260628',
+  supports: {
+    input: ['text', 'image'],
+    output: ['image'],
+  },
+} as const satisfies ModelMeta
+
+const SEEDREAM_5_0 = {
+  name: 'seedream-5-0-260128',
+  supports: {
+    input: ['text', 'image'],
+    output: ['image'],
+  },
+} as const satisfies ModelMeta
+
+const SEEDREAM_5_0_LITE = {
+  name: 'seedream-5-0-lite-260128',
+  supports: {
+    input: ['text', 'image'],
+    output: ['image'],
+  },
+} as const satisfies ModelMeta
+
+const SEEDREAM_4_5 = {
+  name: 'seedream-4-5-251128',
+  supports: {
+    input: ['text', 'image'],
+    output: ['image'],
+  },
+} as const satisfies ModelMeta
+
+const SEEDREAM_4_0 = {
+  name: 'seedream-4-0-250828',
+  supports: {
+    input: ['text', 'image'],
+    output: ['image'],
+  },
+} as const satisfies ModelMeta
+
+/**
+ * All supported Seedream image model identifiers.
+ */
+export const BYTEPLUS_IMAGE_MODELS = [
+  DOLA_SEEDREAM_5_0_PRO.name,
+  SEEDREAM_5_0.name,
+  SEEDREAM_5_0_LITE.name,
+  SEEDREAM_4_5.name,
+  SEEDREAM_4_0.name,
+] as const
+
+/**
+ * Union of all supported Seedream image model names.
+ */
+export type BytePlusImageModel = (typeof BYTEPLUS_IMAGE_MODELS)[number]
+
+/**
+ * Type-only map from image model name to its accepted `size` strings.
+ */
+export type BytePlusImageModelSizeByName = {
+  [K in BytePlusImageModel]: BytePlusImageSize
+}
+
+/**
+ * Maximum number of reference images accepted per editing request.
+ * Seedream 5.0 Pro caps at 10 references; the other editing-capable models
+ * accept up to 14.
+ *
+ * Docs-derived, not live-probed. The 14 for `seedream-5-0-260128` is weaker
+ * still — the docs never state a cap for that model, so it is inferred from
+ * the rest of the family.
+ */
+export const BYTEPLUS_IMAGE_MAX_REFERENCE_IMAGES: {
+  readonly [K in BytePlusImageModel]: number
+} = {
+  'dola-seedream-5-0-pro-260628': 10,
+  'seedream-5-0-260128': 14,
+  'seedream-5-0-lite-260128': 14,
+  'seedream-4-5-251128': 14,
+  'seedream-4-0-250828': 14,
+}
+
+// ============================================================================
+// Seed Speech models (voice host — separate product and API key)
+// ============================================================================
+
+const SEED_AUDIO_1_0 = {
+  name: 'seed-audio-1.0',
+  supports: {
+    input: ['text', 'audio'],
+    output: ['audio'],
+  },
+} as const satisfies ModelMeta
+
+// Seed Speech ASR is endpoint-addressed: `POST /api/v3/auc/bigmodel/recognize/
+// flash` selects the model through the `X-Api-Resource-Id` header
+// (`volc.seedasr.auc_turbo`) and takes no `model` field in the body. This
+// synthetic identifier satisfies the SDK's `TranscriptionOptions.model`
+// contract and gives logging and fixture matching a stable value.
+const SEED_ASR = {
+  name: 'seed-asr',
+  supports: {
+    input: ['audio'],
+    output: ['text'],
+  },
+} as const satisfies ModelMeta
+
+/**
+ * All supported Seed Speech TTS model identifiers.
+ *
+ * Note: TTS runs on `voice.ap-southeast-1.bytepluses.com` with an
+ * `X-Api-Key` header and a *different* API key from Ark.
+ */
+export const BYTEPLUS_TTS_MODELS = [SEED_AUDIO_1_0.name] as const
+
+/**
+ * All supported Seed Speech transcription model identifiers.
+ */
+export const BYTEPLUS_TRANSCRIPTION_MODELS = [SEED_ASR.name] as const
+
+/**
+ * Union of all supported Seed Speech TTS model names.
+ */
+export type BytePlusTTSModel = (typeof BYTEPLUS_TTS_MODELS)[number]
+
+/**
+ * Union of all supported Seed Speech transcription model names.
+ */
+export type BytePlusTranscriptionModel =
+  (typeof BYTEPLUS_TRANSCRIPTION_MODELS)[number]
+
+// ============================================================================
+// Type resolution helpers
+// ============================================================================
+
+/**
+ * Resolve provider options for a specific model. Models listed in the chat
+ * map get their explicit options; anything else falls back to the base chat
+ * options.
+ */
+export type ResolveProviderOptions<TModel extends string> =
+  TModel extends keyof BytePlusChatModelProviderOptionsByName
+    ? BytePlusChatModelProviderOptionsByName[TModel]
+    : BytePlusTextProviderOptions
+
+/**
+ * Resolve input modalities for a specific model. Models missing from the map
+ * are treated as text-only.
+ */
+export type ResolveInputModalities<TModel extends string> =
+  TModel extends keyof BytePlusModelInputModalitiesByName
+    ? BytePlusModelInputModalitiesByName[TModel]
+    : readonly ['text']