dsh-provider-hub 0.1.1 → 0.3.0

package/README.md

@@ -4,7 +4,7 @@
 
 为 DSH 扩大可用的服务商：**选一个服务商，粘贴 API Key，点启用**——插件把官方 `llm-pi-ai` 路由与凭据一次性写好，立即生效、无需重启、无需手改 `settings.yaml`。
 
-内置 23 家预设（OpenAI 兼容协议 + Anthropic Messages 协议），另有「自定义服务商」入口可接入任意 OpenAI 兼容端点；DSH 已内置的服务商默认折叠、不重复展示。
+内置 **40 家服务商预设**（OpenAI 兼容 / Anthropic Messages / OpenAI Responses 三种协议），覆盖国际与国内的普通 API 与 **订阅渠道**（百炼 Token Plan、智谱 Coding、Kimi Coding、MiniMax Coding Plan、小米 MiMo、OpenCode Zen、**Command Code GOAT** 等），另有「自定义服务商」入口可接入任意兼容端点；DSH 已内置的服务商默认折叠、不重复展示。
 
 ## 安装
 
@@ -42,9 +42,54 @@ dsh plugin --profile web add link:/path/to/dsh-provider-hub
 
 ## 预设清单
 
-OpenAI、OpenRouter、Anthropic、Google Gemini（OpenAI 兼容）、xAI Grok、Mistral、Groq、Together、Fireworks、DeepInfra、Novita、硅基流动、月之暗面 Kimi（国内/国际）、智谱 GLM（国内）、Z.ai GLM（国际）、MiniMax、阿里云百炼 Qwen、百度千帆、腾讯混元、火山方舟（豆包）、**StepFun Step Plan（国内/国际）**。
+OpenAI、OpenRouter、Anthropic、Google Gemini、xAI Grok、Mistral、Groq、Together、Fireworks、DeepInfra、Novita、Cerebras、NVIDIA NIM、Baseten、HuggingFace Router、**Vercel AI Gateway**、**OpenCode Zen / Zen Go**、硅基流动、月之暗面 Kimi（国内/国际）、**Kimi Coding Plan**、智谱 GLM（国内）、**智谱 Coding Plan**、Z.ai GLM、**MiniMax Coding Plan（国际/国内）**、MiniMax（国内）、**阿里云百炼 Token Plan（国内/国际）**、阿里云百炼 Qwen、百度千帆、腾讯混元、火山方舟（豆包）、**小米 MiMo**、**Ant Ling**、**LongCat（美团）**、**StepFun Step Plan（国内/国际）**。
 
-关于内置：DSH 自带 DeepSeek 官方适配器（`llm-deepseek`），其目录里的 `stepfun` 指向普通付费端点 `api.stepfun.com/v1`——**Step Plan 订阅端点（`step_plan/v1`）没有内置条目，因此本插件提供该预设**；其余与本插件重合的服务商（openai、openrouter 等 13 个）由 DSH 原生目录覆盖，默认折叠在「显示 DSH 已内置」开关后。已配置过的 Key（包括来自环境变量的）会被识别，留空输入框即可直接启用。
+关于内置：DSH 自带 DeepSeek 官方适配器（`llm-deepseek`），其目录里的 `stepfun` 指向普通付费端点 `api.stepfun.com/v1`——**Step Plan 订阅端点（`step_plan/v1`）没有内置条目，因此本插件提供该预设**；其余与本插件重合的服务商由 DSH 原生目录覆盖的那些默认折叠在「显示 DSH 已内置」开关后。已配置过的 Key（包括来自环境变量的）会被识别，留空输入框即可直接启用。
+
+**Command Code GOAT** 预设收录了 GOAT 计划的 58 个模型（GPT-6 Luna、MiMo V2.6 全系、Kimi K3、Grok 4.7、Qwen 3.8、GLM-5.3、DeepSeek V4 系列、Gemini 3.8 Flash 等，含 `poolside/laguna-s-2.1-free`、`inclusionai/ling-3.0-flash-sante:free` 两个免费模型）；需 Pro/Max 计划的高级模型（GPT-6 Astra/Sol、Claude 系列等）不在其中。
+
+## WorkBuddy（腾讯）适配
+
+同一份预设目录也能写入 WorkBuddy 的本地自定义模型文件（`~/.workbuddy/models.json`，明文 JSON，应用启动时读取）：
+
+```sh
+node tools/workbuddy.mjs list                    # 查看预设与 WorkBuddy 里的已装状态
+node tools/workbuddy.mjs add-all                 # 目录内全部预设：无 Key 的先装上，用时在 WorkBuddy 里补
+node tools/workbuddy.mjs add stepfun-step-plan --key-from-dsh STEPFUN_API_KEY
+node tools/workbuddy.mjs add openrouter --key <key> --models deepseek/deepseek-chat
+node tools/workbuddy.mjs remove stepfun-step-plan
+node tools/workbuddy.mjs add-custom --model gpt-5.1 --url https://api.openai.com/v1 --key sk-...
+```
+
+- 条目按 WorkBuddy 的校验规则（`isValidLocalCustomModel`）生成，写入前本地校验、临时文件+rename 原子替换
+- **写入后需重启 WorkBuddy**（CLI 会在应用运行中时提示）
+- 应用会把本地模型标记为 `custom-local:` 前缀 + `custom` tag，实际请求前还原原始模型 ID
+- `useCustomProtocol: false`（默认）＝应用自动在 URL 后追加 `/chat/completions`；需要完整 URL 时加 `--full-url`
+- Key 来源三选一：`--key`、`--key-env NAME`、`--key-from-dsh REF`（复用 DSH 凭据服务里的 Key）
+- WorkBuddy 的模型按 ID 全局唯一：国内/国际预设共用模型 ID 时，**目录里靠前的一家生效**；带 Key 的 `add` 是显式安装、必定覆盖；`add-all` 不会覆盖已配置 Key 的条目。想换另一家先 `remove <另一家>` 再 `add <目标家> --key ...`
+
+## OpenCode 适配
+
+同一份预设目录也能写入 OpenCode 的配置（`~/.config/opencode/opencode.json` 或 `.jsonc`，自动备份）：
+
+```sh
+node tools/opencode.mjs list                          # 查看预设与 OpenCode 里的已装状态
+node tools/opencode.mjs add commandcode --key-from-dsh COMMANDCODE_API_KEY
+node tools/opencode.mjs add openrouter --key <key> --models deepseek/deepseek-chat
+node tools/opencode.mjs add-all [--prune]             # 全部预设（无 Key 的先用占位）
+node tools/opencode.mjs remove commandcode
+node tools/opencode.mjs add-custom --route myapi --model gpt-5.1 --url https://api.openai.com/v1 --key sk-... [--reasoning] [--model-name "GPT-5.1"]
+```
+
+- Key 存进 provider 的 `options.apiKey`；模型带 OpenCode 需要的元数据（`name` / `limit` / `reasoning` / `tool_call` / `attachment`）
+- OpenCode 的 models.dev 目录里没有的路由自动补上正确的 npm SDK（依据本地 `~/.cache/opencode/models.json` 判定协议）
+- 改完配置**重启 OpenCode 应用**（命令行 `opencode run` 每次启动即读）
+- Key 来源三选一：`--key`、`--key-env NAME`、`--key-from-dsh REF`
+
+## 目录维护
+
+- 模型清单与逐模型兼容开关可从任意权威目录刷新：`node tools/refresh-catalog.mjs --catalog <catalog.json>`（先 `--dry` 审阅）；路由级只保留全部模型一致的开关，不一致的由模型自己携带 `compat`
+- WorkBuddy 端用 `node tools/workbuddy.mjs add-all --prune` 同步：装入全部预设并清掉目录里已不存在的旧条目
 
 ## 注意
 
@@ -58,7 +103,11 @@ OpenAI、OpenRouter、Anthropic、Google Gemini（OpenAI 兼容）、xAI Grok、
 
 Expand the providers DSH can use: **pick a provider, paste the API key, press Enable**. The plugin writes the official `llm-pi-ai` route and its credential in one step - effective immediately, no restart, no hand-editing `settings.yaml`.
 
-23 built-in presets (OpenAI-compatible and Anthropic Messages protocols) plus a custom-endpoint form for any OpenAI-compatible gateway. Presets your target app already ships natively are collapsed by default, never deleted.
+40 built-in presets (OpenAI-compatible, Anthropic Messages and OpenAI Responses protocols) plus a custom-endpoint form for any compatible gateway. Presets your target app already ships natively are collapsed by default, never deleted.
+
+The **Command Code GOAT** preset carries all 58 models of the GOAT plan (GPT-6 Luna, the MiMo V2.6 family, Kimi K3, Grok 4.7, Qwen 3.8, GLM-5.3, DeepSeek V4 series, Gemini 3.8 Flash, plus the free `poolside/laguna-s-2.1-free` and `inclusionai/ling-3.0-flash-sante:free`). Pro/Max-only models (GPT-6 Astra/Sol, the Claude family) are excluded.
+
+The same catalog ships two CLIs: `provider-hub-workbuddy` writes WorkBuddy's local `models.json`, `provider-hub-opencode` writes OpenCode's config; both accept `--key-from-dsh <REF>` to reuse a key already stored in DSH.
 
 ### Install
 

package/core/adapters/opencode.js

@@ -0,0 +1,137 @@
+// dsh-provider-hub - OpenCode adapter core (pure).
+//
+// OpenCode (sst/opencode, desktop app + CLI) reads `~/.config/opencode/opencode.json`
+// (or `.jsonc`) and keeps provider keys either in `options.apiKey` or in its
+// auth store (`~/.local/share/opencode/auth.json`, keyed by provider id).
+// A custom provider entry is:
+//
+//   "provider": {
+//     "<route>": {
+//       "npm": "@ai-sdk/openai-compatible",   // only when OpenCode does not
+//       "api": "https://…",                    // know the route from models.dev
+//       "options": { "apiKey": "…" },
+//       "models": {
+//         "<model id>": { "name", "attachment", "reasoning", "tool_call",
+//                         "temperature", "limit": { context, input, output },
+//                         "cost", "interleaved", "modalities", "variants" }
+//       }
+//     }
+//   }
+//
+// The builders here are pure; the CLI decides `npm` (from the local models.dev
+// cache) and performs the file I/O.
+
+/** Validate the pieces OpenCode needs from a provider entry. */
+export function providerError(entry) {
+  if (!entry || typeof entry !== 'object') return 'provider entry 必须是对象'
+  if (typeof entry.api !== 'string' || entry.api.trim() === '') return 'api（Base URL）不能为空'
+  try {
+    const url = new URL(entry.api)
+    if (url.protocol !== 'https:' && url.protocol !== 'http:') return 'api 必须是 http(s)'
+  } catch {
+    return 'api 不是合法 URL'
+  }
+  if (!entry.models || typeof entry.models !== 'object' || Object.keys(entry.models).length === 0) return '至少需要一个模型'
+  return undefined
+}
+
+/**
+ * Build the OpenCode provider entry for one preset.
+ * @param preset - catalog preset (`id`, `name`, `baseURL`, `models`).
+ * @param options - `{ key?, apiKey?, npm?, modelIds? }`; `npm` is set by the
+ *   caller when OpenCode does not ship the route in its models.dev catalog.
+ * @returns a `{ [route]: provider }` object ready to merge into the config.
+ */
+export function buildOpencodeProvider(preset, options = {}) {
+  const apiKey = typeof options.key === 'string' && options.key.trim() !== '' ? options.key.trim() : options.apiKey
+  const wanted = Array.isArray(options.modelIds) && options.modelIds.length > 0 ? new Set(options.modelIds) : undefined
+  const models = {}
+  for (const model of preset.models) {
+    if (wanted !== undefined && !wanted.has(model.id)) continue
+    const config = {
+      name: model.name ?? model.id,
+      attachment: Array.isArray(model.input) && model.input.includes('image'),
+      reasoning: options.reasoning ?? Boolean(model.reasoningEfforts),
+      temperature: true,
+      tool_call: true,
+    }
+    if (typeof model.contextWindow === 'number') {
+      config.limit = { context: model.contextWindow, input: model.contextWindow }
+      if (typeof model.maxTokens === 'number') config.limit.output = model.maxTokens
+    }
+    models[model.id] = config
+  }
+  const provider = { name: preset.name }
+  if (options.npm) provider.npm = options.npm
+  provider.api = preset.baseURL
+  if (apiKey !== undefined) provider.options = { apiKey }
+  provider.models = models
+  return { [preset.id]: provider }
+}
+
+/**
+ * Merge provider entries into an OpenCode config, replacing same-route models
+ * in place while keeping every other key (including user edits on fields this
+ * hub does not own).
+ * @param config - parsed OpenCode config.
+ * @param entries - `{ [route]: provider }` from {@link buildOpencodeProvider}.
+ * @returns a new config object.
+ */
+export function mergeOpencodeConfig(config, entries) {
+  const next = config && typeof config === 'object' ? JSON.parse(JSON.stringify(config)) : {}
+  next.provider = next.provider && typeof next.provider === 'object' ? next.provider : {}
+  for (const [route, provider] of Object.entries(entries)) {
+    const existing = next.provider[route] && typeof next.provider[route] === 'object' ? next.provider[route] : {}
+    const merged = { ...existing, ...provider }
+    merged.options = { ...(existing.options ?? {}), ...(provider.options ?? {}) }
+    if (existing.models || provider.models) {
+      merged.models = { ...(existing.models ?? {}) }
+      for (const [id, model] of Object.entries(provider.models ?? {})) {
+        merged.models[id] = { ...(existing.models?.[id] ?? {}), ...model }
+      }
+    }
+    next.provider[route] = merged
+  }
+  return next
+}
+
+/** Remove one provider route; returns a new config. */
+export function removeOpencodeProvider(config, route) {
+  const next = config && typeof config === 'object' ? JSON.parse(JSON.stringify(config)) : {}
+  if (next.provider && typeof next.provider === 'object') delete next.provider[route]
+  return next
+}
+
+/**
+ * Drop model ids that left the catalog for every preset the config already
+ * has (matched by api URL), keeping routes and models this hub does not own.
+ * @returns `{ config, pruned }`.
+ */
+export function pruneOpencodeConfig(config, presets) {
+  const next = config && typeof config === 'object' ? JSON.parse(JSON.stringify(config)) : {}
+  let pruned = 0
+  for (const preset of presets) {
+    const provider = next.provider?.[preset.id]
+    if (!provider || provider.api !== preset.baseURL || !provider.models) continue
+    const ids = new Set(preset.models.map((model) => model.id))
+    for (const id of Object.keys(provider.models)) {
+      if (!ids.has(id)) {
+        delete provider.models[id]
+        pruned += 1
+      }
+    }
+  }
+  return { config: next, pruned }
+}
+
+/** Which catalog presets already exist in an OpenCode config (by route id). */
+export function installedPresets(config, presets) {
+  const routes = new Map()
+  for (const preset of presets) {
+    const provider = config?.provider?.[preset.id]
+    if (provider && typeof provider === 'object') {
+      routes.set(preset.id, Object.keys(provider.models ?? {}).length)
+    }
+  }
+  return routes
+}

package/core/adapters/workbuddy.js

@@ -0,0 +1,179 @@
+// dsh-provider-hub - WorkBuddy adapter core (pure).
+//
+// Tencent WorkBuddy (com.tencent.workbuddy.mac) reads local custom models from
+// `<configDir>/models.json` -- a plain JSON array, no field encryption, only a
+// file lock and atomic write around it. The app validates every entry with the
+// equivalent of `isValidLocalCustomModel` and DROPS invalid entries with a
+// warning, so the builders here emit exactly that shape:
+//
+//   * `id` is required (the model id sent to the endpoint);
+//   * `name` / `vendor` / `url` / `apiKey` are strings;
+//   * `maxInputTokens` / `maxOutputTokens` / `temperature` are numbers;
+//   * `supportsToolCall` / `supportsImages` / `supportsReasoning` /
+//     `onlyReasoning` / `useCustomProtocol` are booleans;
+//   * `reasoning` is an object (`{ defaultEffort, supportedEfforts, canDisableThinking }`).
+//
+// `useCustomProtocol: false` means the app appends `/chat/completions` to the
+// URL (the normal OpenAI-compatible case); `true` sends the URL verbatim.
+// Entries are addressed by their URL for install/removal bookkeeping: the app
+// keys by id (one entry per id globally), so a preset owns exactly the entries
+// whose `url` equals its base URL.
+
+/** The separator between a preset's display name and a model name in `name`. */
+const NAME_SEPARATOR = ' · '
+
+/**
+ * Mirror of the app's `isValidLocalCustomModel`: would WorkBuddy keep this entry?
+ * @param entry - candidate entry.
+ * @returns true when the app would load it.
+ */
+export function isValidWorkbuddyEntry(entry) {
+  if (!entry || typeof entry !== 'object' || Array.isArray(entry)) return false
+  if (typeof entry.id !== 'string' || entry.id.trim() === '') return false
+  const stringOrAbsent = (value) => value === undefined || typeof value === 'string'
+  const numberOrAbsent = (value) => {
+    if (value === undefined) return true
+    if (typeof value === 'number') return Number.isFinite(value)
+    // The app coerces numeric strings (`coerceOptionalNumber`), so a numeric
+    // string is accepted here too -- kept for validator fidelity only.
+    if (typeof value === 'string' && value.trim() !== '') return Number.isFinite(Number(value))
+    return false
+  }
+  const booleanOrAbsent = (value) => value === undefined || typeof value === 'boolean'
+  if (!['name', 'vendor', 'url', 'apiKey'].every((field) => stringOrAbsent(entry[field]))) return false
+  if (!['maxInputTokens', 'maxOutputTokens', 'temperature'].every((field) => numberOrAbsent(entry[field]))) return false
+  if (!['supportsToolCall', 'supportsImages', 'supportsReasoning', 'onlyReasoning', 'useCustomProtocol'].every((field) => booleanOrAbsent(entry[field]))) return false
+  if (entry.reasoning !== undefined && (typeof entry.reasoning !== 'object' || entry.reasoning === null || Array.isArray(entry.reasoning))) return false
+  return true
+}
+
+/** Normalize one http(s) base URL the way the app's connectivity test does. */
+export function baseUrlError(value) {
+  if (typeof value !== 'string' || value.trim() === '') return 'url 不能为空'
+  let url
+  try {
+    url = new URL(value.trim())
+  } catch {
+    return 'url 格式不合法'
+  }
+  if (url.protocol !== 'https:' && url.protocol !== 'http:') return 'url 仅支持 HTTP 或 HTTPS'
+  return undefined
+}
+
+/**
+ * Build the WorkBuddy entry for one catalog model.
+ * @param preset - catalog preset (`id`, `name`, `baseURL`, `models`).
+ * @param model - one model entry from the preset.
+ * @param options - `{ key?, apiKey? }`.
+ * @returns a models.json entry.
+ */
+export function buildWorkbuddyEntry(preset, model, options = {}) {
+  const apiKey = typeof options.key === 'string' && options.key.trim() !== '' ? options.key.trim() : options.apiKey
+  const efforts = model.reasoningEfforts && typeof model.reasoningEfforts === 'object' ? Object.keys(model.reasoningEfforts) : []
+  const reasoning = {}
+  if (efforts.length > 0) {
+    reasoning.supportedEfforts = efforts
+    if (efforts.includes('low')) reasoning.defaultEffort = 'low'
+    if (efforts.includes('medium') && reasoning.defaultEffort === undefined) reasoning.defaultEffort = 'medium'
+  }
+  const entry = {
+    id: model.id,
+    name: `${preset.name}${NAME_SEPARATOR}${model.name ?? model.id}`,
+    vendor: preset.name,
+    url: preset.baseURL,
+    apiKey,
+    supportsToolCall: true,
+    supportsImages: Array.isArray(model.input) && model.input.includes('image'),
+    supportsReasoning: Boolean(model.reasoningEfforts),
+    useCustomProtocol: false,
+  }
+  if (typeof model.contextWindow === 'number') entry.maxInputTokens = model.contextWindow
+  if (typeof model.maxTokens === 'number') entry.maxOutputTokens = model.maxTokens
+  if (Object.keys(reasoning).length > 0) entry.reasoning = reasoning
+  for (const key of Object.keys(entry)) if (entry[key] === undefined) delete entry[key]
+  return entry
+}
+
+/**
+ * Build every entry one preset contributes.
+ * @param preset - catalog preset.
+ * @param options - `{ key?, apiKey?, modelIds? }` (`modelIds` limits the set).
+ * @returns entries, in catalog order.
+ */
+export function buildWorkbuddyEntries(preset, options = {}) {
+  const wanted = Array.isArray(options.modelIds) && options.modelIds.length > 0 ? new Set(options.modelIds) : undefined
+  return preset.models
+    .filter((model) => wanted === undefined || wanted.has(model.id))
+    .map((model) => buildWorkbuddyEntry(preset, model, options))
+}
+
+/**
+ * Upsert entries into a WorkBuddy models list, by entry id (the app's own key).
+ * An existing entry for the same id is replaced -- exactly what the app's
+ * `configSaveLocalCustomModel` does -- so a re-add updates in place.
+ * @param list - current models.json array.
+ * @param entries - entries to write.
+ * @returns `{ models, added, updated }`.
+ */
+export function upsertWorkbuddyModels(list, entries) {
+  const models = Array.isArray(list) ? list.map((entry) => ({ ...entry })) : []
+  let added = 0
+  let updated = 0
+  let kept = 0
+  for (const entry of entries) {
+    if (!isValidWorkbuddyEntry(entry)) throw new Error(`拒绝写入非法条目：${JSON.stringify(entry).slice(0, 120)}`)
+    const index = models.findIndex((candidate) => candidate.id === entry.id)
+    if (index < 0) {
+      models.push(entry)
+      added += 1
+      continue
+    }
+    const existing = models[index]
+    const incomingHasKey = typeof entry.apiKey === 'string' && entry.apiKey !== ''
+    const existingHasKey = typeof existing.apiKey === 'string' && existing.apiKey !== ''
+    // WorkBuddy keys models by id globally, so a catalog pass can meet an id
+    // twice (CN vs global presets share model ids) or meet an entry the user
+    // already configured. Rules: a keyless add never displaces a configured
+    // entry; among keyless duplicates the first catalog entry wins; a keyed
+    // add always wins (that is an explicit per-provider install).
+    if (!incomingHasKey && existingHasKey) {
+      kept += 1
+      continue
+    }
+    if (!incomingHasKey && existing.url !== entry.url) {
+      kept += 1
+      continue
+    }
+    models[index] = { ...existing, ...entry }
+    updated += 1
+  }
+  return { models, added, updated, kept }
+}
+
+/**
+ * Remove every entry that a preset owns (matched by base URL).
+ * @param list - current models.json array.
+ * @param preset - catalog preset.
+ * @returns `{ models, removed }`.
+ */
+export function removeWorkbuddyModels(list, preset) {
+  const models = Array.isArray(list) ? list : []
+  const kept = models.filter((entry) => entry?.url !== preset.baseURL)
+  return { models: kept, removed: models.length - kept.length }
+}
+
+/**
+ * Which presets already have entries in the WorkBuddy models list.
+ * @param list - current models.json array.
+ * @param presets - catalog presets.
+ * @returns Map presetId -> model count.
+ */
+export function installedPresets(list, presets) {
+  const models = Array.isArray(list) ? list : []
+  const counts = new Map()
+  for (const preset of presets) {
+    const count = models.filter((entry) => entry?.url === preset.baseURL).length
+    if (count > 0) counts.set(preset.id, count)
+  }
+  return counts
+}

package/core/profile.js

@@ -67,6 +67,10 @@ export function sanitizeModel(entry) {
     maxTokens: Number.isFinite(entry.maxTokens) && entry.maxTokens > 0 ? Math.floor(entry.maxTokens) : undefined,
     input: Array.isArray(entry.input) && entry.input.length > 0 ? [...entry.input] : undefined,
     reasoningEfforts: entry.reasoningEfforts,
+    // Per-model wire switches: when models on one route disagree (Kimi's k3 vs
+    // k2.x thinking formats), the model entry carries its own compat and wins
+    // over the route consensus.
+    compat: entry.compat && typeof entry.compat === 'object' && !Array.isArray(entry.compat) ? { ...entry.compat } : undefined,
   })
   return model
 }
@@ -133,12 +137,24 @@ export function buildProfile(input) {
       : route
   const env = preset ? preset.env : customEnvName(route)
 
+  // Route-level compat: the hub's safe defaults for OpenAI-compatible
+  // endpoints, overridden by the preset's own authoritative switches (learned
+  // from the served catalogs, e.g. thinkingFormat qwen/zai/deepseek/ant-ling),
+  // then by any explicit caller override. anthropic-messages routes carry their
+  // own switch set and no OpenAI defaults; openai-responses carries none here.
+  const presetCompat = preset && preset.compat && typeof preset.compat === 'object' ? preset.compat : undefined
+  const callerCompat = input && input.compat && typeof input.compat === 'object' ? input.compat : undefined
+  let compat
+  if (api === 'openai-completions') compat = { ...OPENAI_COMPAT, ...(presetCompat ?? {}), ...(callerCompat ?? {}) }
+  else if (api === 'anthropic-messages') compat = { ...(presetCompat ?? {}), ...(callerCompat ?? {}) }
+  if (compat !== undefined && Object.keys(compat).length === 0) compat = undefined
+
   return compact({
     displayName,
     api,
     baseURL,
     apiKeyEnv: env,
-    compat: api === 'openai-completions' ? { ...OPENAI_COMPAT, ...(input.compat ?? {}) } : undefined,
+    compat,
     models,
   })
 }