@opendash-ai/cli 0.1.0-beta.1

package/src/provider/transform.ts

@@ -0,0 +1,1376 @@
+import type { ModelMessage } from "ai"
+import { mergeDeep, unique } from "remeda"
+import type { JSONSchema7 } from "@ai-sdk/provider"
+import type { JSONSchema } from "zod/v4/core"
+import type * as Provider from "./provider"
+import type * as ModelsDev from "./models"
+import { iife } from "@/util/iife"
+import { Flag } from "@/flag/flag"
+
+type Modality = NonNullable<ModelsDev.Model["modalities"]>["input"][number]
+
+function mimeToModality(mime: string): Modality | undefined {
+  if (mime.startsWith("image/")) return "image"
+  if (mime.startsWith("audio/")) return "audio"
+  if (mime.startsWith("video/")) return "video"
+  if (mime === "application/pdf") return "pdf"
+  return undefined
+}
+
+// MiMo vision support isn't reflected in models.dev modality data, so the
+// generic capability check would strip images before they reach the model.
+// mimo-auto and mimo-v2.5 accept images; mimo-v2.5-pro is text-only.
+function supportsImageInput(model: Provider.Model): boolean {
+  if (model.providerID === "mimo" || model.providerID === "xiaomi") {
+    const id = model.id.toLowerCase()
+    if (id.includes("v2.5-pro")) return false
+    if (id === "mimo-auto" || id.includes("v2.5")) return true
+  }
+  // Claude and GPT models are all multimodal regardless of catalog data.
+  const id = model.id.toLowerCase()
+  const apiID = model.api.id.toLowerCase()
+  if (id.includes("claude") || apiID.includes("claude") || model.providerID === "anthropic") return true
+  if (id.includes("gpt") || apiID.includes("gpt")) return true
+  return model.capabilities.input.image
+}
+
+export const OUTPUT_TOKEN_MAX = Flag.OPENDASH_EXPERIMENTAL_OUTPUT_TOKEN_MAX || 32_000
+const MIMO_OUTPUT_TOKEN_MAX = 128_000
+
+// Maps npm package to the key the AI SDK expects for providerOptions
+function sdkKey(npm: string): string | undefined {
+  switch (npm) {
+    case "@ai-sdk/github-copilot":
+      return "copilot"
+    case "@ai-sdk/azure":
+      return "azure"
+    case "@ai-sdk/openai":
+      return "openai"
+    case "@ai-sdk/amazon-bedrock":
+      return "bedrock"
+    case "@ai-sdk/anthropic":
+    case "@ai-sdk/google-vertex/anthropic":
+      return "anthropic"
+    case "@ai-sdk/google-vertex":
+      return "vertex"
+    case "@ai-sdk/google":
+      return "google"
+    case "@ai-sdk/gateway":
+      return "gateway"
+    case "@openrouter/ai-sdk-provider":
+      return "openrouter"
+  }
+  return undefined
+}
+
+function normalizeMessages(
+  msgs: ModelMessage[],
+  model: Provider.Model,
+  _options: Record<string, unknown>,
+): ModelMessage[] {
+  // Anthropic rejects messages with empty content - filter out empty string messages
+  // and remove empty text/reasoning parts from array content
+  if (model.api.npm === "@ai-sdk/anthropic" || model.api.npm === "@ai-sdk/amazon-bedrock") {
+    msgs = msgs
+      .map((msg) => {
+        if (typeof msg.content === "string") {
+          if (msg.content === "") return undefined
+          return msg
+        }
+        if (!Array.isArray(msg.content)) return msg
+        const filtered = msg.content.filter((part) => {
+          if (part.type === "text" || part.type === "reasoning") {
+            return part.text !== ""
+          }
+          return true
+        })
+        if (filtered.length === 0) return undefined
+        return { ...msg, content: filtered }
+      })
+      .filter((msg): msg is ModelMessage => msg !== undefined && msg.content !== "")
+  }
+
+  if (model.api.id.includes("claude")) {
+    const scrub = (id: string) => id.replace(/[^a-zA-Z0-9_-]/g, "_")
+    msgs = msgs.map((msg) => {
+      if (msg.role === "assistant" && Array.isArray(msg.content)) {
+        return {
+          ...msg,
+          content: msg.content.map((part) => {
+            if (part.type === "tool-call" || part.type === "tool-result") {
+              return { ...part, toolCallId: scrub(part.toolCallId) }
+            }
+            return part
+          }),
+        }
+      }
+      if (msg.role === "tool" && Array.isArray(msg.content)) {
+        return {
+          ...msg,
+          content: msg.content.map((part) => {
+            if (part.type === "tool-result") {
+              return { ...part, toolCallId: scrub(part.toolCallId) }
+            }
+            return part
+          }),
+        }
+      }
+      return msg
+    })
+  }
+  if (["@ai-sdk/anthropic", "@ai-sdk/google-vertex/anthropic"].includes(model.api.npm)) {
+    // Anthropic rejects assistant turns where tool_use blocks are followed by non-tool
+    // content, e.g. [tool_use, tool_use, text], with:
+    // `tool_use` ids were found without `tool_result` blocks immediately after...
+    //
+    // Reorder that invalid shape into [text] + [tool_use, tool_use]. Consecutive
+    // assistant messages are later merged by the provider/SDK, so preserving the
+    // original [tool_use...] then [text] order still produces the invalid payload.
+    //
+    // The root cause appears to be somewhere upstream where the stream is originally
+    // processed. We were unable to locate an exact narrower reproduction elsewhere,
+    // so we keep this transform in place for the time being.
+    msgs = msgs.flatMap((msg) => {
+      if (msg.role !== "assistant" || !Array.isArray(msg.content)) return [msg]
+
+      const parts = msg.content
+      const first = parts.findIndex((part) => part.type === "tool-call")
+      if (first === -1) return [msg]
+      if (!parts.slice(first).some((part) => part.type !== "tool-call")) return [msg]
+      return [
+        { ...msg, content: parts.filter((part) => part.type !== "tool-call") },
+        { ...msg, content: parts.filter((part) => part.type === "tool-call") },
+      ]
+    })
+  }
+  if (
+    model.providerID === "mistral" ||
+    model.api.id.toLowerCase().includes("mistral") ||
+    model.api.id.toLocaleLowerCase().includes("devstral")
+  ) {
+    const scrub = (id: string) => {
+      return id
+        .replace(/[^a-zA-Z0-9]/g, "") // Remove non-alphanumeric characters
+        .substring(0, 9) // Take first 9 characters
+        .padEnd(9, "0") // Pad with zeros if less than 9 characters
+    }
+    const result: ModelMessage[] = []
+    for (let i = 0; i < msgs.length; i++) {
+      const msg = msgs[i]
+      const nextMsg = msgs[i + 1]
+
+      if (msg.role === "assistant" && Array.isArray(msg.content)) {
+        msg.content = msg.content.map((part) => {
+          if (part.type === "tool-call" || part.type === "tool-result") {
+            return { ...part, toolCallId: scrub(part.toolCallId) }
+          }
+          return part
+        })
+      }
+      if (msg.role === "tool" && Array.isArray(msg.content)) {
+        msg.content = msg.content.map((part) => {
+          if (part.type === "tool-result") {
+            return { ...part, toolCallId: scrub(part.toolCallId) }
+          }
+          return part
+        })
+      }
+      result.push(msg)
+
+      // Fix message sequence: tool messages cannot be followed by user messages
+      if (msg.role === "tool" && nextMsg?.role === "user") {
+        result.push({
+          role: "assistant",
+          content: [
+            {
+              type: "text",
+              text: "Done.",
+            },
+          ],
+        })
+      }
+    }
+    return result
+  }
+
+  if (typeof model.capabilities.interleaved === "object" && model.capabilities.interleaved.field) {
+    const field = model.capabilities.interleaved.field
+    return msgs.map((msg) => {
+      if (msg.role === "assistant" && Array.isArray(msg.content)) {
+        const reasoningParts = msg.content.filter((part: any) => part.type === "reasoning")
+        const reasoningText = reasoningParts.map((part: any) => part.text).join("")
+
+        // Filter out reasoning parts from content
+        const filteredContent = msg.content.filter((part: any) => part.type !== "reasoning")
+
+        // Include reasoning_content | reasoning_details directly on the message for all assistant messages
+        if (reasoningText) {
+          return {
+            ...msg,
+            content: filteredContent,
+            providerOptions: {
+              ...msg.providerOptions,
+              openaiCompatible: {
+                ...msg.providerOptions?.openaiCompatible,
+                [field]: reasoningText,
+              },
+            },
+          }
+        }
+
+        return {
+          ...msg,
+          content: filteredContent,
+        }
+      }
+
+      return msg
+    })
+  }
+
+  return msgs
+}
+
+// Determines whether a model's provider respects inline cache_control / cachePoint
+// markers. Pure name matching (model.api.id.includes("claude")) is fragile — a Claude
+// model behind an OpenAI-compatible proxy gets matched but markers are silently dropped.
+// See docs/cache-policy.md and upstream opendash#26786.
+function supportsCacheMarkers(model: Provider.Model): boolean {
+  // Anthropic-only SDKs — always support inline markers
+  if (model.api.npm === "@ai-sdk/anthropic" || model.api.npm === "@ai-sdk/google-vertex/anthropic") return true
+  if (model.providerID === "anthropic" || model.providerID === "google-vertex-anthropic") return true
+  // Bedrock cachePoint is a Converse API feature, works across model families
+  if (model.api.npm === "@ai-sdk/amazon-bedrock") return true
+  // Multi-model providers: only Anthropic/Claude models support cache markers
+  if (
+    model.api.npm === "@openrouter/ai-sdk-provider" ||
+    model.api.npm === "@ai-sdk/github-copilot" ||
+    model.api.npm === "@ai-sdk/alibaba"
+  ) {
+    return (
+      model.api.id.includes("claude") ||
+      model.api.id.includes("anthropic") ||
+      model.id.includes("claude") ||
+      model.id.includes("anthropic")
+    )
+  }
+  return false
+}
+
+// The cache-control marker shape differs per provider/SDK. This is the single
+// source of truth, keyed by the SDK provider-options namespace. `applyCaching`
+// attaches the whole object (keyed by stored providerID) and lets `message()`
+// remap the active provider's namespace to its SDK key; `tools()` (which
+// bypasses that remap) resolves a single namespace up front via `cacheMarkerFor`.
+// Only Anthropic and OpenRouter expose a TTL in their AI SDK — the others ignore
+// an unknown `ttl`, so we thread it only there.
+function cacheMarkerOptions(model: Provider.Model) {
+  const ttl = model.cachePromptTTL === "1h" ? { ttl: "1h" as const } : {}
+  return {
+    anthropic: { cacheControl: { type: "ephemeral", ...ttl } },
+    openrouter: { cacheControl: { type: "ephemeral", ...ttl } },
+    bedrock: { cachePoint: { type: "default" } },
+    openaiCompatible: { cache_control: { type: "ephemeral" } },
+    copilot: { copilot_cache_control: { type: "ephemeral" } },
+    alibaba: { cacheControl: { type: "ephemeral" } },
+  }
+}
+
+// Resolve the marker for a single model, already keyed under the SDK namespace
+// the AI SDK expects — i.e. the remap that `message()` performs for messages,
+// done up front. Used by `tools()`, whose tools never pass through `message()`.
+// Returns undefined for providers that don't take inline markers (callers gate
+// on `supportsCacheMarkers` first, so this is just a type-safety fallback).
+function cacheMarkerFor(model: Provider.Model): Record<string, unknown> | undefined {
+  const shapes = cacheMarkerOptions(model)
+  const ns: keyof typeof shapes | undefined =
+    model.api.npm === "@ai-sdk/anthropic" || model.api.npm === "@ai-sdk/google-vertex/anthropic"
+      ? "anthropic"
+      : model.api.npm === "@openrouter/ai-sdk-provider"
+        ? "openrouter"
+        : model.api.npm === "@ai-sdk/amazon-bedrock"
+          ? "bedrock"
+          : model.api.npm === "@ai-sdk/github-copilot"
+            ? "copilot"
+            : model.api.npm === "@ai-sdk/alibaba"
+              ? "alibaba"
+              : undefined
+  if (!ns) return undefined
+  return { [ns]: shapes[ns] }
+}
+
+function applyCaching(msgs: ModelMessage[], model: Provider.Model): ModelMessage[] {
+  const providerOptions = cacheMarkerOptions(model)
+
+  // Strategy: prefix caching is longest-common-prefix based with a backward
+  // lookback window (Anthropic walks back ~20 blocks from a breakpoint to find
+  // a prior write). The markers that grow the cached prefix are pinned to the
+  // *tail* of the request. We place up to three stable breakpoints (Anthropic
+  // allows max 4):
+  // 1. Last system message — the immutable prompt prefix.
+  // 2+3. The last TWO messages — a "rolling double buffer". Each turn marks
+  //      messages[-2] and messages[-1]; next turn the old [-1] is now [-2] and
+  //      still carries its marker, so the lookback gets a cache READ hit, while
+  //      the new [-1] is the WRITE for the turn after.
+  //
+  //      Why two and not one: the second (next-to-last) marker is the safety
+  //      net for the tail boundary. When the last message is removed — a
+  //      tool-call retry, a Ctrl-C, or the user editing/deleting their latest
+  //      message — a lone tail marker disappears with it, and how much of the
+  //      surrounding prefix the provider then evicts depends on the upstream
+  //      (Anthropic) KV-cache implementation. The next-to-last marker is a
+  //      still-present, further-back write the next lookback can land on, so the
+  //      worst case degrades to "recompute only the removed message" instead of
+  //      "recompute the whole history". It also covers turns that append >20
+  //      blocks (tool spam pushes the prior write outside the lookback window).
+  //      Cost is ~equal to a single marker: the two adjacent breakpoints write
+  //      roughly the same incremental bytes as one, split in two, and a hit
+  //      never rewrites. A third marker would write a segment never read
+  //      independently, so two is the minimum that covers the boundary.
+  // We deliberately do NOT mark a drifting midpoint or a fixed before-last-user
+  // INDEX: those shift every turn without tracking the tail.
+  const targets: ModelMessage[] = []
+
+  const systemMsgs = msgs.filter((msg) => msg.role === "system")
+  if (systemMsgs.length > 0) targets.push(systemMsgs[systemMsgs.length - 1])
+
+  const nonSystem = msgs.filter((msg) => msg.role !== "system")
+  for (const msg of nonSystem.slice(-2)) targets.push(msg)
+
+  for (const msg of unique(targets)) {
+    const useMessageLevelOptions =
+      model.providerID === "anthropic" ||
+      model.providerID.includes("bedrock") ||
+      model.api.npm === "@ai-sdk/amazon-bedrock"
+    const shouldUseContentOptions = !useMessageLevelOptions && Array.isArray(msg.content) && msg.content.length > 0
+
+    if (shouldUseContentOptions) {
+      const lastContent = msg.content[msg.content.length - 1]
+      if (
+        lastContent &&
+        typeof lastContent === "object" &&
+        lastContent.type !== "tool-approval-request" &&
+        lastContent.type !== "tool-approval-response"
+      ) {
+        lastContent.providerOptions = mergeDeep(lastContent.providerOptions ?? {}, providerOptions)
+        continue
+      }
+    }
+
+    msg.providerOptions = mergeDeep(msg.providerOptions ?? {}, providerOptions)
+  }
+
+  return msgs
+}
+
+function unsupportedParts(msgs: ModelMessage[], model: Provider.Model): ModelMessage[] {
+  return msgs.map((msg) => {
+    if (msg.role !== "user" || !Array.isArray(msg.content)) return msg
+
+    const filtered = msg.content.map((part) => {
+      if (part.type !== "file" && part.type !== "image") return part
+
+      // Check for empty base64 image data
+      if (part.type === "image") {
+        const imageStr = String(part.image)
+        if (imageStr.startsWith("data:")) {
+          const match = imageStr.match(/^data:([^;]+);base64,(.*)$/)
+          if (match && (!match[2] || match[2].length === 0)) {
+            return {
+              type: "text" as const,
+              text: "ERROR: Image file is empty or corrupted. Please provide a valid image.",
+            }
+          }
+        }
+      }
+
+      const mime = part.type === "image" ? String(part.image).split(";")[0].replace("data:", "") : part.mediaType
+      const filename = part.type === "file" ? part.filename : undefined
+      const modality = mimeToModality(mime)
+      if (!modality) return part
+      const supported = modality === "image" ? supportsImageInput(model) : model.capabilities.input[modality]
+      if (supported) return part
+
+      const name = filename ? `"${filename}"` : modality
+      return {
+        type: "text" as const,
+        text: `ERROR: Cannot read ${name} (this model does not support ${modality} input). Inform the user.`,
+      }
+    })
+
+    return { ...msg, content: filtered }
+  })
+}
+
+// Returns the decoded byte size of a base64 data URL, or undefined for inputs
+// that aren't data URLs (remote URLs, raw binary) and therefore can't be sized.
+function imageByteSize(image: string): number | undefined {
+  if (!image.startsWith("data:")) return undefined
+  const base64 = image.slice(image.indexOf(",") + 1)
+  if (!base64) return 0
+  const padding = base64.endsWith("==") ? 2 : base64.endsWith("=") ? 1 : 0
+  return Math.floor((base64.length * 3) / 4) - padding
+}
+
+function limitImages(msgs: ModelMessage[]): ModelMessage[] {
+  const maxImages = Flag.OPENDASH_MAX_PROMPT_IMAGES
+  const maxSize = Flag.OPENDASH_MAX_PROMPT_IMAGE_SIZE
+  if (maxImages === undefined && maxSize === undefined) return msgs
+
+  const total = msgs.reduce(
+    (sum, msg) =>
+      msg.role === "user" && Array.isArray(msg.content)
+        ? sum + msg.content.filter((part) => part.type === "image").length
+        : sum,
+    0,
+  )
+  // Drop the oldest excess images so the most recent ones reach the model.
+  let toDrop = maxImages === undefined ? 0 : Math.max(0, total - maxImages)
+
+  return msgs.map((msg) => {
+    if (msg.role !== "user" || !Array.isArray(msg.content)) return msg
+    const content = msg.content.map((part) => {
+      if (part.type !== "image") return part
+      if (toDrop > 0) {
+        toDrop--
+        return { type: "text" as const, text: `[Image omitted: exceeds the configured limit of ${maxImages} prompt image(s).]` }
+      }
+      if (maxSize !== undefined) {
+        const size = imageByteSize(String(part.image))
+        if (size !== undefined && size > maxSize) {
+          return { type: "text" as const, text: `[Image omitted: exceeds the configured ${maxSize}-byte prompt image size limit.]` }
+        }
+      }
+      return part
+    })
+    return { ...msg, content }
+  })
+}
+
+export function message(msgs: ModelMessage[], model: Provider.Model, options: Record<string, unknown>) {
+  msgs = unsupportedParts(msgs, model)
+  msgs = limitImages(msgs)
+  msgs = normalizeMessages(msgs, model, options)
+  if (supportsCacheMarkers(model)) {
+    msgs = applyCaching(msgs, model)
+  }
+
+  // Remap providerOptions keys from stored providerID to expected SDK key
+  const key = sdkKey(model.api.npm)
+  if (key && key !== model.providerID) {
+    const remap = (opts: Record<string, any> | undefined) => {
+      if (!opts) return opts
+      if (!(model.providerID in opts)) return opts
+      const result = { ...opts }
+      result[key] = result[model.providerID]
+      delete result[model.providerID]
+      return result
+    }
+
+    msgs = msgs.map((msg) => {
+      if (!Array.isArray(msg.content)) return { ...msg, providerOptions: remap(msg.providerOptions) }
+      return {
+        ...msg,
+        providerOptions: remap(msg.providerOptions),
+        content: msg.content.map((part) => {
+          if (part.type === "tool-approval-request" || part.type === "tool-approval-response") {
+            return { ...part }
+          }
+          return { ...part, providerOptions: remap(part.providerOptions) }
+        }),
+      } as typeof msg
+    })
+  }
+
+  return msgs
+}
+
+// Place a cache breakpoint on the tool definitions. The cache hierarchy is
+// `tools` → `system` → `messages`, so marking the LAST tool caches the entire
+// tool-schema block (often several KB) as a stable prefix that sits in front of
+// the system + message caches. Tools are passed to the SDK separately from
+// `message()` and never go through its providerID→SDK-key remap, so we resolve
+// the SDK-keyed marker via `cacheMarkerFor`. Tool registration order is stable
+// (insertion order of the tools record), so "last tool" is deterministic.
+export function tools<T extends Record<string, any>>(tools: T, model: Provider.Model): T {
+  if (!supportsCacheMarkers(model)) return tools
+  const marker = cacheMarkerFor(model)
+  if (!marker) return tools
+  const names = Object.keys(tools)
+  if (names.length === 0) return tools
+
+  const last = tools[names[names.length - 1]]
+  last.providerOptions = mergeDeep(last.providerOptions ?? {}, marker)
+  return tools
+}
+
+export function temperature(model: Provider.Model) {
+  const id = model.id.toLowerCase()
+  if (id.includes("qwen")) return 0.55
+  if (id.includes("claude")) return undefined
+  if (id.includes("gemini")) return 1.0
+  if (id.includes("glm-4.6")) return 1.0
+  if (id.includes("glm-4.7")) return 1.0
+  if (id.includes("minimax-m2")) return 1.0
+  if (id.includes("mimo")) return 1.0
+  if (id.includes("kimi-k2")) {
+    // kimi-k2-thinking & kimi-k2.5 && kimi-k2p5 && kimi-k2-5
+    if (["thinking", "k2.", "k2p", "k2-5"].some((s) => id.includes(s))) {
+      return 1.0
+    }
+    return 0.6
+  }
+  return undefined
+}
+
+export function topP(model: Provider.Model) {
+  const id = model.id.toLowerCase()
+  if (id.includes("qwen")) return 1
+  if (["minimax-m2", "gemini", "kimi-k2.5", "kimi-k2p5", "kimi-k2-5"].some((s) => id.includes(s))) {
+    return 0.95
+  }
+  return undefined
+}
+
+export function topK(model: Provider.Model) {
+  const id = model.id.toLowerCase()
+  if (id.includes("minimax-m2")) {
+    if (["m2.", "m25", "m21"].some((s) => id.includes(s))) return 40
+    return 20
+  }
+  if (id.includes("gemini")) return 64
+  return undefined
+}
+
+const WIDELY_SUPPORTED_EFFORTS = ["low", "medium", "high"]
+const OPENAI_EFFORTS = ["none", "minimal", ...WIDELY_SUPPORTED_EFFORTS, "xhigh"]
+
+function anthropicAdaptiveEfforts(apiId: string): string[] | null {
+  if (["opus-4-7", "opus-4.7", "opus-4-8", "opus-4.8"].some((v) => apiId.includes(v))) {
+    return ["low", "medium", "high", "xhigh", "max"]
+  }
+  if (["opus-4-6", "opus-4.6", "sonnet-4-6", "sonnet-4.6"].some((v) => apiId.includes(v))) {
+    return ["low", "medium", "high", "max"]
+  }
+  return null
+}
+
+export function variants(model: Provider.Model): Record<string, Record<string, any>> {
+  if (!model.capabilities.reasoning) return {}
+
+  const id = model.id.toLowerCase()
+  const adaptiveEfforts = anthropicAdaptiveEfforts(model.api.id)
+  if (
+    id.includes("deepseek") ||
+    id.includes("minimax") ||
+    id.includes("glm") ||
+    id.includes("mistral") ||
+    id.includes("kimi") ||
+    id.includes("k2p5") ||
+    id.includes("qwen") ||
+    id.includes("big-pickle")
+  )
+    return {}
+
+  // see: https://docs.x.ai/docs/guides/reasoning#control-how-hard-the-model-thinks
+  if (id.includes("grok") && id.includes("grok-3-mini")) {
+    if (model.api.npm === "@openrouter/ai-sdk-provider") {
+      return {
+        low: { reasoning: { effort: "low" } },
+        high: { reasoning: { effort: "high" } },
+      }
+    }
+    return {
+      low: { reasoningEffort: "low" },
+      high: { reasoningEffort: "high" },
+    }
+  }
+  if (id.includes("grok")) return {}
+
+  switch (model.api.npm) {
+    case "@openrouter/ai-sdk-provider":
+      if (!model.id.includes("gpt") && !model.id.includes("gemini-3") && !model.id.includes("claude")) return {}
+      return Object.fromEntries(OPENAI_EFFORTS.map((effort) => [effort, { reasoning: { effort } }]))
+
+    case "@ai-sdk/gateway":
+      if (model.id.includes("anthropic")) {
+        if (adaptiveEfforts) {
+          return Object.fromEntries(
+            adaptiveEfforts.map((effort) => [
+              effort,
+              {
+                thinking: {
+                  type: "adaptive",
+                },
+                effort,
+              },
+            ]),
+          )
+        }
+        return {
+          high: {
+            thinking: {
+              type: "enabled",
+              budgetTokens: 16000,
+            },
+          },
+          max: {
+            thinking: {
+              type: "enabled",
+              budgetTokens: 31999,
+            },
+          },
+        }
+      }
+      if (model.id.includes("google")) {
+        if (id.includes("2.5")) {
+          return {
+            high: {
+              thinkingConfig: {
+                includeThoughts: true,
+                thinkingBudget: 16000,
+              },
+            },
+            max: {
+              thinkingConfig: {
+                includeThoughts: true,
+                thinkingBudget: 24576,
+              },
+            },
+          }
+        }
+        return Object.fromEntries(
+          ["low", "high"].map((effort) => [
+            effort,
+            {
+              includeThoughts: true,
+              thinkingLevel: effort,
+            },
+          ]),
+        )
+      }
+      return Object.fromEntries(OPENAI_EFFORTS.map((effort) => [effort, { reasoningEffort: effort }]))
+
+    case "@ai-sdk/github-copilot":
+      if (model.id.includes("gemini")) {
+        // currently github copilot only returns thinking
+        return {}
+      }
+      if (model.id.includes("claude")) {
+        return Object.fromEntries(WIDELY_SUPPORTED_EFFORTS.map((effort) => [effort, { reasoningEffort: effort }]))
+      }
+      const copilotEfforts = iife(() => {
+        if (id.includes("5.1-codex-max") || id.includes("5.2") || id.includes("5.3"))
+          return [...WIDELY_SUPPORTED_EFFORTS, "xhigh"]
+        const arr = [...WIDELY_SUPPORTED_EFFORTS]
+        if (id.includes("gpt-5") && model.release_date >= "2025-12-04") arr.push("xhigh")
+        return arr
+      })
+      return Object.fromEntries(
+        copilotEfforts.map((effort) => [
+          effort,
+          {
+            reasoningEffort: effort,
+            reasoningSummary: "auto",
+            include: ["reasoning.encrypted_content"],
+          },
+        ]),
+      )
+
+    case "@ai-sdk/cerebras":
+    // https://v5.ai-sdk.dev/providers/ai-sdk-providers/cerebras
+    case "@ai-sdk/togetherai":
+    // https://v5.ai-sdk.dev/providers/ai-sdk-providers/togetherai
+    case "@ai-sdk/xai":
+    // https://v5.ai-sdk.dev/providers/ai-sdk-providers/xai
+    case "@ai-sdk/deepinfra":
+    // https://v5.ai-sdk.dev/providers/ai-sdk-providers/deepinfra
+    case "venice-ai-sdk-provider":
+    // https://docs.venice.ai/overview/guides/reasoning-models#reasoning-effort
+    case "@ai-sdk/openai-compatible":
+      return Object.fromEntries(WIDELY_SUPPORTED_EFFORTS.map((effort) => [effort, { reasoningEffort: effort }]))
+
+    case "@ai-sdk/azure":
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/azure
+      if (id === "o1-mini") return {}
+      const azureEfforts = ["low", "medium", "high"]
+      if (id.includes("gpt-5-") || id === "gpt-5") {
+        azureEfforts.unshift("minimal")
+      }
+      return Object.fromEntries(
+        azureEfforts.map((effort) => [
+          effort,
+          {
+            reasoningEffort: effort,
+            reasoningSummary: "auto",
+            include: ["reasoning.encrypted_content"],
+          },
+        ]),
+      )
+    case "@ai-sdk/openai":
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/openai
+      if (id === "gpt-5-pro") return {}
+      const openaiEfforts = iife(() => {
+        if (id.includes("codex")) {
+          if (id.includes("5.2") || id.includes("5.3")) return [...WIDELY_SUPPORTED_EFFORTS, "xhigh"]
+          return WIDELY_SUPPORTED_EFFORTS
+        }
+        const arr = [...WIDELY_SUPPORTED_EFFORTS]
+        if (id.includes("gpt-5-") || id === "gpt-5") {
+          arr.unshift("minimal")
+        }
+        if (model.release_date >= "2025-11-13") {
+          arr.unshift("none")
+        }
+        if (model.release_date >= "2025-12-04") {
+          arr.push("xhigh")
+        }
+        return arr
+      })
+      return Object.fromEntries(
+        openaiEfforts.map((effort) => [
+          effort,
+          {
+            reasoningEffort: effort,
+            reasoningSummary: "auto",
+            include: ["reasoning.encrypted_content"],
+          },
+        ]),
+      )
+
+    case "@ai-sdk/anthropic":
+    // https://v5.ai-sdk.dev/providers/ai-sdk-providers/anthropic
+    case "@ai-sdk/google-vertex/anthropic":
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/google-vertex#anthropic-provider
+
+      if (model.providerID === "github-copilot") {
+        if (model.api.id.includes("opus-4.7")) {
+          return Object.fromEntries(["medium"].map((effort) => [effort, { reasoningEffort: effort }]))
+        }
+      }
+
+      if (adaptiveEfforts) {
+        return Object.fromEntries(
+          adaptiveEfforts.map((effort) => [
+            effort,
+            {
+              thinking: {
+                type: "adaptive",
+                ...(["opus-4-7", "opus-4.7", "opus-4-8", "opus-4.8"].some((v) => model.api.id.includes(v))
+                  ? { display: "summarized" }
+                  : {}),
+              },
+              effort,
+            },
+          ]),
+        )
+      }
+
+      return {
+        high: {
+          thinking: {
+            type: "enabled",
+            budgetTokens: Math.min(16_000, Math.floor(model.limit.output / 2 - 1)),
+          },
+        },
+        max: {
+          thinking: {
+            type: "enabled",
+            budgetTokens: Math.min(31_999, model.limit.output - 1),
+          },
+        },
+      }
+
+    case "@ai-sdk/amazon-bedrock":
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/amazon-bedrock
+      if (adaptiveEfforts) {
+        return Object.fromEntries(
+          adaptiveEfforts.map((effort) => [
+            effort,
+            {
+              reasoningConfig: {
+                type: "adaptive",
+                maxReasoningEffort: effort,
+                ...(["opus-4-7", "opus-4.7", "opus-4-8", "opus-4.8"].some((v) => model.api.id.includes(v))
+                  ? { display: "summarized" }
+                  : {}),
+              },
+            },
+          ]),
+        )
+      }
+      // For Anthropic models on Bedrock, use reasoningConfig with budgetTokens
+      if (model.api.id.includes("anthropic")) {
+        return {
+          high: {
+            reasoningConfig: {
+              type: "enabled",
+              budgetTokens: 16000,
+            },
+          },
+          max: {
+            reasoningConfig: {
+              type: "enabled",
+              budgetTokens: 31999,
+            },
+          },
+        }
+      }
+
+      // For Amazon Nova models, use reasoningConfig with maxReasoningEffort
+      return Object.fromEntries(
+        WIDELY_SUPPORTED_EFFORTS.map((effort) => [
+          effort,
+          {
+            reasoningConfig: {
+              type: "enabled",
+              maxReasoningEffort: effort,
+            },
+          },
+        ]),
+      )
+
+    case "@ai-sdk/google-vertex":
+    // https://v5.ai-sdk.dev/providers/ai-sdk-providers/google-vertex
+    case "@ai-sdk/google":
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/google-generative-ai
+      if (id.includes("2.5")) {
+        return {
+          high: {
+            thinkingConfig: {
+              includeThoughts: true,
+              thinkingBudget: 16000,
+            },
+          },
+          max: {
+            thinkingConfig: {
+              includeThoughts: true,
+              thinkingBudget: 24576,
+            },
+          },
+        }
+      }
+      let levels = ["low", "high"]
+      if (id.includes("3.1")) {
+        levels = ["low", "medium", "high"]
+      }
+
+      return Object.fromEntries(
+        levels.map((effort) => [
+          effort,
+          {
+            thinkingConfig: {
+              includeThoughts: true,
+              thinkingLevel: effort,
+            },
+          },
+        ]),
+      )
+
+    case "@ai-sdk/mistral":
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/mistral
+      return {}
+
+    case "@ai-sdk/cohere":
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/cohere
+      return {}
+
+    case "@ai-sdk/groq":
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/groq
+      const groqEffort = ["none", ...WIDELY_SUPPORTED_EFFORTS]
+      return Object.fromEntries(
+        groqEffort.map((effort) => [
+          effort,
+          {
+            reasoningEffort: effort,
+          },
+        ]),
+      )
+
+    case "@ai-sdk/perplexity":
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/perplexity
+      return {}
+
+    case "@jerome-benoit/sap-ai-provider-v2":
+      if (model.api.id.includes("anthropic")) {
+        if (adaptiveEfforts) {
+          return Object.fromEntries(
+            adaptiveEfforts.map((effort) => [
+              effort,
+              {
+                thinking: {
+                  type: "adaptive",
+                },
+                effort,
+              },
+            ]),
+          )
+        }
+        return {
+          high: {
+            thinking: {
+              type: "enabled",
+              budgetTokens: 16000,
+            },
+          },
+          max: {
+            thinking: {
+              type: "enabled",
+              budgetTokens: 31999,
+            },
+          },
+        }
+      }
+      if (model.api.id.includes("gemini") && id.includes("2.5")) {
+        return {
+          high: {
+            thinkingConfig: {
+              includeThoughts: true,
+              thinkingBudget: 16000,
+            },
+          },
+          max: {
+            thinkingConfig: {
+              includeThoughts: true,
+              thinkingBudget: 24576,
+            },
+          },
+        }
+      }
+      if (model.api.id.includes("gpt") || /\bo[1-9]/.test(model.api.id)) {
+        return Object.fromEntries(WIDELY_SUPPORTED_EFFORTS.map((effort) => [effort, { reasoningEffort: effort }]))
+      }
+      return {}
+  }
+  return {}
+}
+
+export function options(input: {
+  model: Provider.Model
+  sessionID: string
+  providerOptions?: Record<string, any>
+}): Record<string, any> {
+  const result: Record<string, any> = {}
+
+  // openai and providers using openai package should set store to false by default.
+  if (
+    input.model.providerID === "openai" ||
+    input.model.api.npm === "@ai-sdk/openai" ||
+    input.model.api.npm === "@ai-sdk/github-copilot"
+  ) {
+    result["store"] = false
+  }
+
+  if (input.model.api.npm === "@ai-sdk/azure") {
+    result["store"] = true
+    result["promptCacheKey"] = input.sessionID
+  }
+
+  if (input.model.api.npm === "@openrouter/ai-sdk-provider" || input.model.api.npm === "@llmgateway/ai-sdk-provider") {
+    result["usage"] = {
+      include: true,
+    }
+    if (input.model.api.id.includes("gemini-3")) {
+      result["reasoning"] = { effort: "high" }
+    }
+  }
+
+  if (
+    input.model.providerID === "baseten" ||
+    (input.model.providerID === "opendash" && ["kimi-k2-thinking", "glm-4.6"].includes(input.model.api.id))
+  ) {
+    result["chat_template_args"] = { enable_thinking: true }
+  }
+
+  if (
+    ["zai", "zhipuai"].some((id) => input.model.providerID.includes(id)) &&
+    input.model.api.npm === "@ai-sdk/openai-compatible"
+  ) {
+    result["thinking"] = {
+      type: "enabled",
+      clear_thinking: false,
+    }
+  }
+
+  if (input.model.providerID === "openai" || input.providerOptions?.setCacheKey) {
+    result["promptCacheKey"] = input.sessionID
+  }
+
+  if (input.model.api.npm === "@ai-sdk/google" || input.model.api.npm === "@ai-sdk/google-vertex") {
+    if (input.model.capabilities.reasoning) {
+      result["thinkingConfig"] = {
+        includeThoughts: true,
+      }
+      if (input.model.api.id.includes("gemini-3")) {
+        result["thinkingConfig"]["thinkingLevel"] = "high"
+      }
+    }
+  }
+
+  // Enable thinking by default for kimi-k2.5/k2p5 models using anthropic SDK
+  const modelId = input.model.api.id.toLowerCase()
+  if (
+    (input.model.api.npm === "@ai-sdk/anthropic" || input.model.api.npm === "@ai-sdk/google-vertex/anthropic") &&
+    (modelId.includes("k2p5") || modelId.includes("kimi-k2.5") || modelId.includes("kimi-k2p5"))
+  ) {
+    result["thinking"] = {
+      type: "enabled",
+      budgetTokens: Math.min(16_000, Math.floor(input.model.limit.output / 2 - 1)),
+    }
+  }
+
+  // Enable thinking for reasoning models on alibaba-cn (DashScope).
+  // DashScope's OpenAI-compatible API requires `enable_thinking: true` in the request body
+  // to return reasoning_content. Without it, models like kimi-k2.5, qwen-plus, qwen3, qwq,
+  // deepseek-r1, etc. never output thinking/reasoning tokens.
+  // Note: kimi-k2-thinking is excluded as it returns reasoning_content by default.
+  if (
+    input.model.providerID === "alibaba-cn" &&
+    input.model.capabilities.reasoning &&
+    input.model.api.npm === "@ai-sdk/openai-compatible" &&
+    !modelId.includes("kimi-k2-thinking")
+  ) {
+    result["enable_thinking"] = true
+  }
+
+  if (input.model.api.id.includes("gpt-5") && !input.model.api.id.includes("gpt-5-chat")) {
+    if (!input.model.api.id.includes("gpt-5-pro")) {
+      result["reasoningEffort"] = "medium"
+      // Only inject reasoningSummary for providers that support it natively.
+      // @ai-sdk/openai-compatible proxies (e.g. LiteLLM) do not understand this
+      // parameter and return "Unknown parameter: 'reasoningSummary'".
+      if (
+        input.model.api.npm === "@ai-sdk/openai" ||
+        input.model.api.npm === "@ai-sdk/azure" ||
+        input.model.api.npm === "@ai-sdk/github-copilot"
+      ) {
+        result["reasoningSummary"] = "auto"
+      }
+    }
+
+    // Only set textVerbosity for non-chat gpt-5.x models
+    // Chat models (e.g. gpt-5.2-chat-latest) only support "medium" verbosity
+    if (
+      input.model.api.id.includes("gpt-5.") &&
+      !input.model.api.id.includes("codex") &&
+      !input.model.api.id.includes("-chat") &&
+      input.model.providerID !== "azure"
+    ) {
+      result["textVerbosity"] = "low"
+    }
+
+    if (input.model.providerID.startsWith("opendash")) {
+      result["promptCacheKey"] = input.sessionID
+      result["include"] = ["reasoning.encrypted_content"]
+      result["reasoningSummary"] = "auto"
+    }
+  }
+
+  if (input.model.providerID === "venice") {
+    result["promptCacheKey"] = input.sessionID
+  }
+
+  if (input.model.providerID === "openrouter") {
+    result["prompt_cache_key"] = input.sessionID
+  }
+  if (input.model.api.npm === "@ai-sdk/gateway") {
+    result["gateway"] = {
+      caching: "auto",
+    }
+  }
+
+  return result
+}
+
+export function smallOptions(model: Provider.Model) {
+  if (
+    model.providerID === "openai" ||
+    model.api.npm === "@ai-sdk/openai" ||
+    model.api.npm === "@ai-sdk/github-copilot"
+  ) {
+    if (model.api.id.includes("gpt-5")) {
+      if (model.api.id.includes("5.") || model.api.id.includes("5-mini")) {
+        return { store: false, reasoningEffort: "low" }
+      }
+      return { store: false, reasoningEffort: "minimal" }
+    }
+    return { store: false }
+  }
+  if (model.providerID === "google") {
+    // gemini-3 uses thinkingLevel, gemini-2.5 uses thinkingBudget
+    if (model.api.id.includes("gemini-3")) {
+      return { thinkingConfig: { thinkingLevel: "minimal" } }
+    }
+    return { thinkingConfig: { thinkingBudget: 0 } }
+  }
+  if (model.providerID === "openrouter" || model.providerID === "llmgateway") {
+    if (model.api.id.includes("google")) {
+      return { reasoning: { enabled: false } }
+    }
+    return { reasoningEffort: "minimal" }
+  }
+
+  if (model.providerID === "venice") {
+    return { veniceParameters: { disableThinking: true } }
+  }
+
+  return {}
+}
+
+// Maps model ID prefix to provider slug used in providerOptions.
+// Example: "amazon/nova-2-lite" → "bedrock"
+const SLUG_OVERRIDES: Record<string, string> = {
+  amazon: "bedrock",
+}
+
+export function providerOptions(model: Provider.Model, options: { [x: string]: any }) {
+  if (model.api.npm === "@ai-sdk/gateway") {
+    // Gateway providerOptions are split across two namespaces:
+    // - `gateway`: gateway-native routing/caching controls (order, only, byok, etc.)
+    // - `<upstream slug>`: provider-specific model options (anthropic/openai/...)
+    // We keep `gateway` as-is and route every other top-level option under the
+    // model-derived upstream slug.
+    const i = model.api.id.indexOf("/")
+    const rawSlug = i > 0 ? model.api.id.slice(0, i) : undefined
+    const slug = rawSlug ? (SLUG_OVERRIDES[rawSlug] ?? rawSlug) : undefined
+    const gateway = options.gateway
+    const rest = Object.fromEntries(Object.entries(options).filter(([k]) => k !== "gateway"))
+    const has = Object.keys(rest).length > 0
+
+    const result: Record<string, any> = {}
+    if (gateway !== undefined) result.gateway = gateway
+
+    if (has) {
+      if (slug) {
+        // Route model-specific options under the provider slug
+        result[slug] = rest
+      } else if (gateway && typeof gateway === "object" && !Array.isArray(gateway)) {
+        result.gateway = { ...gateway, ...rest }
+      } else {
+        result.gateway = rest
+      }
+    }
+
+    return result
+  }
+
+  const key = sdkKey(model.api.npm) ?? model.providerID
+  // @ai-sdk/azure delegates to OpenAIChatLanguageModel which reads from
+  // providerOptions["openai"], but OpenAIResponsesLanguageModel checks
+  // "azure" first. Pass both so model options work on either code path.
+  if (model.api.npm === "@ai-sdk/azure") {
+    return { openai: options, azure: options }
+  }
+  return { [key]: options }
+}
+
+export function maxOutputTokens(model: Provider.Model): number {
+  if (model.providerID === "mimo" || model.providerID === "xiaomi" || model.id.toLowerCase().includes("mimo")) {
+    return MIMO_OUTPUT_TOKEN_MAX
+  }
+  return Math.min(model.limit.output, OUTPUT_TOKEN_MAX) || OUTPUT_TOKEN_MAX
+}
+
+// Flatten a root-level `anyOf` / `oneOf` (typically from `z.discriminatedUnion`)
+// into a single `type: "object"` schema with all variant properties merged at
+// the root. The discriminator key becomes an `enum`, and per-variant required
+// fields are encoded as a textual hint on the discriminator's description.
+//
+// OpenAI's function-calling validator rejects oneOf/anyOf/allOf/enum/not at the
+// top level outright, so this is the only shape that gets through. We retain
+// `additionalProperties: false` to keep the model from inventing fields, and
+// rely on zod's runtime parse (still using the original discriminated union)
+// to enforce per-action required fields strictly.
+function flattenDiscriminatedUnion(schema: JSONSchema.BaseSchema | JSONSchema7): JSONSchema7 {
+  const root = schema as Record<string, any>
+  const variants = (root.anyOf ?? root.oneOf) as Array<Record<string, any>> | undefined
+  if (!variants?.length) return schema as JSONSchema7
+
+  // Find the discriminator: a property whose `const` differs across all variants.
+  const constByKey = new Map<string, Set<unknown>>()
+  for (const v of variants) {
+    if (!v.properties) continue
+    for (const [key, prop] of Object.entries(v.properties as Record<string, any>)) {
+      if (prop && typeof prop === "object" && "const" in prop) {
+        if (!constByKey.has(key)) constByKey.set(key, new Set())
+        constByKey.get(key)!.add(prop.const)
+      }
+    }
+  }
+  let discriminator: string | undefined
+  for (const [key, values] of constByKey) {
+    if (values.size === variants.length) {
+      discriminator = key
+      break
+    }
+  }
+
+  // Merge non-discriminator properties from every variant. Track which variants
+  // each property appeared in so the description can tell the model
+  // "(only when action='X'|'Y')" — flat schemas tempt some models (notably
+  // gpt-5.5) to fill every property regardless of which action they chose.
+  const properties: Record<string, any> = {}
+  const propertyOwners: Record<string, unknown[]> = {}
+  for (const v of variants) {
+    if (!v.properties) continue
+    const variantValue = discriminator
+      ? (v.properties as Record<string, any>)[discriminator]?.const
+      : undefined
+    for (const [key, prop] of Object.entries(v.properties as Record<string, any>)) {
+      if (key === discriminator) continue
+      if (!(key in properties)) properties[key] = prop
+      if (variantValue !== undefined) {
+        if (!propertyOwners[key]) propertyOwners[key] = []
+        propertyOwners[key].push(variantValue)
+      }
+    }
+  }
+  if (discriminator) {
+    for (const [key, owners] of Object.entries(propertyOwners)) {
+      if (owners.length === variants.length) continue // present in every variant — no annotation
+      const tag = `(only when ${discriminator}=${owners.map((o) => JSON.stringify(o)).join("|")})`
+      const original = (properties[key] as Record<string, any>).description as string | undefined
+      properties[key] = {
+        ...properties[key],
+        description: original ? `${tag} ${original}` : tag,
+      }
+    }
+  }
+
+  // Discriminator becomes an enum with per-variant required fields hinted in
+  // its description. Without this hint the model only sees a flat bag of
+  // optional fields and forgets what to provide for each action.
+  if (discriminator) {
+    const proto = (variants[0].properties as Record<string, any>)[discriminator]
+    const enumValues = variants.map((v) => (v.properties as Record<string, any>)[discriminator!]?.const)
+    const baseDescription = (proto?.description as string | undefined) ?? ""
+    const hints = variants
+      .map((v) => {
+        const value = (v.properties as Record<string, any>)[discriminator!]?.const
+        const required = ((v.required as string[] | undefined) ?? []).filter((r) => r !== discriminator)
+        return required.length > 0 ? `${value}: requires ${required.join(", ")}` : `${value}: no extra required fields`
+      })
+      .join("; ")
+    properties[discriminator] = {
+      type: "string",
+      enum: enumValues,
+      description: baseDescription ? `${baseDescription}\n\nPer-${discriminator}: ${hints}.` : `Per-${discriminator}: ${hints}.`,
+    }
+  }
+
+  return {
+    type: "object",
+    properties,
+    required: discriminator ? [discriminator] : [],
+    additionalProperties: false,
+  } as JSONSchema7
+}
+
+export function schema(model: Provider.Model, schema: JSONSchema.BaseSchema | JSONSchema7): JSONSchema7 {
+  /*
+  if (["openai", "azure"].includes(providerID)) {
+    if (schema.type === "object" && schema.properties) {
+      for (const [key, value] of Object.entries(schema.properties)) {
+        if (schema.required?.includes(key)) continue
+        schema.properties[key] = {
+          anyOf: [
+            value as JSONSchema.JSONSchema,
+            {
+              type: "null",
+            },
+          ],
+        }
+      }
+    }
+  }
+  */
+
+  // Many providers reject root-level `anyOf`/`oneOf` in tool schemas:
+  // - OpenAI/Azure: "schema must have type 'object' and not have 'oneOf'/'anyOf'"
+  // - Bedrock: "input_schema.type: Field required"
+  // - Anthropic proxies to Bedrock: same Bedrock error
+  // Flatten unconditionally — all providers accept a flat `type: "object"` schema,
+  // and zod's runtime parse still enforces per-variant required fields strictly.
+  schema = flattenDiscriminatedUnion(schema)
+
+  // Convert integer enums to string enums for Google/Gemini
+  if (model.providerID === "google" || model.api.id.includes("gemini")) {
+    const isPlainObject = (node: unknown): node is Record<string, any> =>
+      typeof node === "object" && node !== null && !Array.isArray(node)
+    const hasCombiner = (node: unknown) =>
+      isPlainObject(node) && (Array.isArray(node.anyOf) || Array.isArray(node.oneOf) || Array.isArray(node.allOf))
+    const hasSchemaIntent = (node: unknown) => {
+      if (!isPlainObject(node)) return false
+      if (hasCombiner(node)) return true
+      return [
+        "type",
+        "properties",
+        "items",
+        "prefixItems",
+        "enum",
+        "const",
+        "$ref",
+        "additionalProperties",
+        "patternProperties",
+        "required",
+        "not",
+        "if",
+        "then",
+        "else",
+      ].some((key) => key in node)
+    }
+
+    const sanitizeGemini = (obj: any): any => {
+      if (obj === null || typeof obj !== "object") {
+        return obj
+      }
+
+      if (Array.isArray(obj)) {
+        return obj.map(sanitizeGemini)
+      }
+
+      const result: any = {}
+      for (const [key, value] of Object.entries(obj)) {
+        if (key === "enum" && Array.isArray(value)) {
+          // Convert all enum values to strings
+          result[key] = value.map((v) => String(v))
+          // If we have integer type with enum, change type to string
+          if (result.type === "integer" || result.type === "number") {
+            result.type = "string"
+          }
+        } else if (typeof value === "object" && value !== null) {
+          result[key] = sanitizeGemini(value)
+        } else {
+          result[key] = value
+        }
+      }
+
+      // Filter required array to only include fields that exist in properties
+      if (result.type === "object" && result.properties && Array.isArray(result.required)) {
+        result.required = result.required.filter((field: any) => field in result.properties)
+      }
+
+      if (result.type === "array" && !hasCombiner(result)) {
+        if (result.items == null) {
+          result.items = {}
+        }
+        // Ensure items has a type only when it's still schema-empty.
+        if (isPlainObject(result.items) && !hasSchemaIntent(result.items)) {
+          result.items.type = "string"
+        }
+      }
+
+      // Remove properties/required from non-object types (Gemini rejects these)
+      if (result.type && result.type !== "object" && !hasCombiner(result)) {
+        delete result.properties
+        delete result.required
+      }
+
+      return result
+    }
+
+    schema = sanitizeGemini(schema)
+  }
+
+  return schema as JSONSchema7
+}