fix(llm): preserve native provider options

2026-05-23 15:42:28 -05:00
parent 7fe7b9f258
commit 6aaaac27d0
12 changed files with 629 additions and 106 deletions
@@ -20,6 +20,7 @@ import {
 import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared"
 import * as Cache from "./utils/cache"
 import { Lifecycle } from "./utils/lifecycle"
+import { ProviderOptions } from "./utils/provider-options"
 import { ToolStream } from "./utils/tool-stream"

 const ADAPTER = "anthropic-messages"
@@ -136,6 +137,7 @@ const AnthropicTool = Schema.Struct({
  description: Schema.String,
  input_schema: JsonObject,
  cache_control: Schema.optional(AnthropicCacheControl),
+  eager_input_streaming: Schema.optional(Schema.Boolean),
 })
 type AnthropicTool = Schema.Schema.Type<typeof AnthropicTool>

@@ -144,10 +146,10 @@ const AnthropicToolChoice = Schema.Union([
  Schema.Struct({ type: Schema.tag("tool"), name: Schema.String }),
 ])

-const AnthropicThinking = Schema.Struct({
-  type: Schema.tag("enabled"),
-  budget_tokens: Schema.Number,
-})
+// Anthropic accepts several `thinking` shapes (enabled with `budget_tokens`,
+// adaptive with optional `display`, and disabled). The body schema permits the
+// full union so explicit lowering can pick the correct fields per model.
+const AnthropicThinkingBody = Schema.Record(Schema.String, Schema.Unknown)

 const AnthropicBodyFields = {
  model: Schema.String,
@@ -161,9 +163,12 @@ const AnthropicBodyFields = {
  top_p: Schema.optional(Schema.Number),
  top_k: Schema.optional(Schema.Number),
  stop_sequences: optionalArray(Schema.String),
-  thinking: Schema.optional(AnthropicThinking),
+  thinking: Schema.optional(AnthropicThinkingBody),
 }
-const AnthropicMessagesBody = Schema.Struct(AnthropicBodyFields)
+// Unknown provider options pass through verbatim with top-level keys snake-cased.
+const AnthropicMessagesBody = Schema.StructWithRest(Schema.Struct(AnthropicBodyFields), [
+  Schema.Record(Schema.String, Schema.Any),
+])
 export type AnthropicMessagesBody = Schema.Schema.Type<typeof AnthropicMessagesBody>

 const AnthropicUsage = Schema.Struct({
@@ -254,11 +259,16 @@ const signatureFromMetadata = (metadata: ProviderMetadata | undefined): string |
  return typeof anthropic.signature === "string" ? anthropic.signature : undefined
 }

-const lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition): AnthropicTool => ({
+const lowerTool = (
+  breakpoints: Cache.Breakpoints,
+  tool: ToolDefinition,
+  eagerInputStreaming: boolean | undefined,
+): AnthropicTool => ({
  name: tool.name,
  description: tool.description,
  input_schema: tool.inputSchema,
  cache_control: cacheControl(breakpoints, tool.cache),
+  eager_input_streaming: eagerInputStreaming ? true : undefined,
 })

 const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
@@ -413,24 +423,46 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
  return messages
 })

-const anthropicOptions = (request: LLMRequest) => request.providerOptions?.anthropic
+// Typed AI SDK Anthropic options. Mirrors the subset of
+// `anthropicLanguageModelOptions` that opencode's provider transform actually
+// emits today (see `packages/opencode/src/provider/transform.ts`). Unknown
+// keys flow through the index signature and pass through to the wire body
+// with their top-level key snake-cased.
+type AnthropicEffort = "low" | "medium" | "high" | "xhigh" | "max"
+type AnthropicThinking =
+  | { readonly type: "enabled"; readonly budgetTokens?: number; readonly budget_tokens?: number }
+  | { readonly type: "adaptive"; readonly display?: "omitted" | "summarized" }
+  | { readonly type: "disabled" }

-const lowerThinking = Effect.fn("AnthropicMessages.lowerThinking")(function* (request: LLMRequest) {
-  const thinking = anthropicOptions(request)?.thinking
-  if (!ProviderShared.isRecord(thinking) || thinking.type !== "enabled") return undefined
-  const budget =
-    typeof thinking.budgetTokens === "number"
-      ? thinking.budgetTokens
-      : typeof thinking.budget_tokens === "number"
-        ? thinking.budget_tokens
-        : undefined
-  if (budget === undefined) return yield* invalid("Anthropic thinking provider option requires budgetTokens")
+interface AnthropicOptions {
+  readonly thinking?: AnthropicThinking
+  readonly effort?: AnthropicEffort
+  readonly toolStreaming?: boolean
+  readonly [extra: string]: unknown
+}
+
+const ANTHROPIC_KNOWN_KEYS: ReadonlySet<string> = new Set(["thinking", "effort", "toolStreaming"])
+
+const lowerThinking = (thinking: AnthropicOptions["thinking"]) => {
+  if (thinking === undefined) return undefined
+  if (thinking.type === "disabled") return undefined
+  if (thinking.type === "adaptive") {
+    return { type: "adaptive" as const, ...(thinking.display ? { display: thinking.display } : {}) }
+  }
+  const budget = thinking.budgetTokens ?? thinking.budget_tokens
+  if (budget === undefined) return undefined
  return { type: "enabled" as const, budget_tokens: budget }
-})
+}

 const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (request: LLMRequest) {
  const toolChoice = request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined
  const generation = request.generation
+  const options = ProviderOptions.merge(request, ["anthropic"]) as AnthropicOptions
+  // AI SDK's `toolStreaming` controls per-tool `eager_input_streaming`. opencode
+  // sets `toolStreaming: false` for non-Claude models routed through
+  // `@ai-sdk/anthropic`; otherwise the field is left unset so the provider
+  // applies its own default.
+  const eagerInputStreaming = options.toolStreaming === true
  // Allocate the 4-breakpoint budget in invalidation order: tools → system →
  // messages. Tools live highest in the cache hierarchy, so when callers
  // over-mark we keep their tool hints and shed the message-tail ones first.
@@ -438,7 +470,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
  const tools =
    request.tools.length === 0 || request.toolChoice?.type === "none"
      ? undefined
-      : request.tools.map((tool) => lowerTool(breakpoints, tool))
+      : request.tools.map((tool) => lowerTool(breakpoints, tool, eagerInputStreaming))
  const system =
    request.system.length === 0
      ? undefined
@@ -454,6 +486,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
    )
  }
  return {
+    ...ProviderOptions.passthrough(options, ANTHROPIC_KNOWN_KEYS),
    model: request.model.id,
    system,
    messages,
@@ -465,7 +498,8 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
    top_p: generation?.topP,
    top_k: generation?.topK,
    stop_sequences: generation?.stop,
-    thinking: yield* lowerThinking(request),
+    thinking: lowerThinking(options.thinking),
+    ...(options.effort !== undefined ? { effort: options.effort } : {}),
  }
 })

@@ -15,6 +15,7 @@ import {
 } from "../schema"
 import { isRecord, JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared"
 import { OpenAIOptions } from "./utils/openai-options"
+import { ProviderOptions } from "./utils/provider-options"
 import { Lifecycle } from "./utils/lifecycle"
 import { ToolStream } from "./utils/tool-stream"

@@ -50,6 +51,10 @@ const OpenAIChatAssistantToolCall = Schema.Struct({
 })
 type OpenAIChatAssistantToolCall = Schema.Schema.Type<typeof OpenAIChatAssistantToolCall>

+// `reasoning_content` is a plain string per DeepSeek/OpenAI-compatible spec.
+// `reasoning_details` is an OpenRouter-style array of typed reasoning objects
+// (summary / encrypted / text). We accept the structured payload as-is so it
+// round-trips verbatim to the provider on continuation requests.
 const OpenAIChatMessage = Schema.Union([
  Schema.Struct({ role: Schema.Literal("system"), content: Schema.String }),
  Schema.Struct({ role: Schema.Literal("user"), content: Schema.String }),
@@ -58,6 +63,7 @@ const OpenAIChatMessage = Schema.Union([
    content: Schema.NullOr(Schema.String),
    tool_calls: optionalArray(OpenAIChatAssistantToolCall),
    reasoning_content: Schema.optional(Schema.String),
+    reasoning_details: Schema.optional(Schema.Array(Schema.Record(Schema.String, Schema.Unknown))),
  }),
  Schema.Struct({ role: Schema.Literal("tool"), tool_call_id: Schema.String, content: Schema.String }),
 ]).pipe(Schema.toTaggedUnion("role"))
@@ -79,7 +85,7 @@ export const bodyFields = {
  stream: Schema.Literal(true),
  stream_options: Schema.optional(Schema.Struct({ include_usage: Schema.Boolean })),
  store: Schema.optional(Schema.Boolean),
-  reasoning_effort: Schema.optional(OpenAIOptions.OpenAIReasoningEffort),
+  reasoning_effort: Schema.optional(Schema.String),
  max_tokens: Schema.optional(Schema.Number),
  temperature: Schema.optional(Schema.Number),
  top_p: Schema.optional(Schema.Number),
@@ -88,7 +94,7 @@ export const bodyFields = {
  seed: Schema.optional(Schema.Number),
  stop: optionalArray(Schema.String),
 }
-const OpenAIChatBody = Schema.Struct(bodyFields)
+const OpenAIChatBody = Schema.StructWithRest(Schema.Struct(bodyFields), [Schema.Record(Schema.String, Schema.Any)])
 export type OpenAIChatBody = Schema.Schema.Type<typeof OpenAIChatBody>

 // =============================================================================
@@ -125,9 +131,16 @@ const OpenAIChatToolCallDelta = Schema.Struct({
 })
 type OpenAIChatToolCallDelta = Schema.Schema.Type<typeof OpenAIChatToolCallDelta>

+// Streaming reasoning fields. `reasoning_content` (DeepSeek) and `reasoning`
+// (AI SDK fallback) are strings; `reasoning_details` (OpenRouter) is an array
+// of typed reasoning detail objects. We surface their plaintext via reasoning
+// deltas and preserve the structured array for downstream round-trip.
+const OpenAIChatReasoningDetail = Schema.Record(Schema.String, Schema.Unknown)
 const OpenAIChatDelta = Schema.Struct({
  content: optionalNull(Schema.String),
  reasoning_content: optionalNull(Schema.String),
+  reasoning: optionalNull(Schema.String),
+  reasoning_details: optionalNull(Schema.Array(OpenAIChatReasoningDetail)),
  tool_calls: optionalNull(Schema.Array(OpenAIChatToolCallDelta)),
 })

@@ -188,6 +201,16 @@ const lowerToolCall = (part: ToolCallPart): OpenAIChatAssistantToolCall => ({
 const openAICompatibleReasoningContent = (native: unknown) =>
  isRecord(native) && typeof native.reasoning_content === "string" ? native.reasoning_content : undefined

+// `reasoning_details` rounds-trips the OpenRouter structured array. Accept the
+// array shape as canonical; tolerate a string for legacy callers that already
+// flattened it.
+const openAICompatibleReasoningDetails = (native: unknown) => {
+  if (!isRecord(native)) return undefined
+  const value = native.reasoning_details
+  if (Array.isArray(value)) return value as ReadonlyArray<Record<string, unknown>>
+  return undefined
+}
+
 const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (message: OpenAIChatRequestMessage) {
  const content: TextPart[] = []
  for (const part of message.content) {
@@ -220,6 +243,7 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
    content: content.length === 0 ? null : ProviderShared.joinText(content),
    tool_calls: toolCalls.length === 0 ? undefined : toolCalls,
    reasoning_content: openAICompatibleReasoningContent(message.native?.openaiCompatible),
+    reasoning_details: openAICompatibleReasoningDetails(message.native?.openaiCompatible),
  }
 })

@@ -246,13 +270,14 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
 })

 const lowerOptions = Effect.fn("OpenAIChat.lowerOptions")(function* (request: LLMRequest) {
-  const store = OpenAIOptions.store(request)
-  const reasoningEffort = OpenAIOptions.reasoningEffort(request)
-  if (reasoningEffort && !OpenAIOptions.isReasoningEffort(reasoningEffort))
-    return yield* invalid(`OpenAI Chat does not support reasoning effort ${reasoningEffort}`)
+  const options = OpenAIOptions.options(request)
+  const effort = options.reasoningEffort
+  if (effort !== undefined && !OpenAIOptions.isReasoningEffort(effort))
+    return yield* invalid(`OpenAI Chat does not support reasoning effort ${effort}`)
  return {
-    ...(store !== undefined ? { store } : {}),
-    ...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}),
+    ...ProviderOptions.passthrough(options, OpenAIOptions.KNOWN_KEYS),
+    ...(options.store !== undefined ? { store: options.store } : {}),
+    ...(effort !== undefined ? { reasoning_effort: effort } : {}),
  }
 })

@@ -325,8 +350,15 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>

    let lifecycle = state.lifecycle

-    if (delta?.reasoning_content)
-      lifecycle = Lifecycle.reasoningDelta(lifecycle, events, "reasoning-0", delta.reasoning_content)
+    // OpenRouter-style `reasoning_details` ships an array of typed reasoning
+    // objects (summary / text / encrypted). Concatenate the plaintext fields
+    // into the reasoning delta stream; the structured array is preserved on
+    // the assistant message for round-trip via `providerMetadata`.
+    const detailText = (delta?.reasoning_details ?? [])
+      .map((detail) => (typeof detail.text === "string" ? detail.text : typeof detail.summary === "string" ? detail.summary : ""))
+      .join("")
+    const reasoning = delta?.reasoning_content ?? delta?.reasoning ?? (detailText.length > 0 ? detailText : undefined)
+    if (reasoning) lifecycle = Lifecycle.reasoningDelta(lifecycle, events, "reasoning-0", reasoning)

    if (delta?.content) lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)

@@ -1,22 +1,84 @@
+import { Effect, Schema } from "effect"
 import { Route, type RouteRoutedModelInput } from "../route/client"
 import { Endpoint } from "../route/endpoint"
 import { Framing } from "../route/framing"
+import { Protocol } from "../route/protocol"
+import type { LLMRequest } from "../schema"
+import { ProviderOptions } from "./utils/provider-options"
 import * as OpenAIChat from "./openai-chat"

 const ADAPTER = "openai-compatible-chat"

 export type OpenAICompatibleChatModelInput = RouteRoutedModelInput

+const OpenAICompatibleChatBody = Schema.StructWithRest(
+  Schema.Struct({ ...OpenAIChat.bodyFields, reasoning_effort: Schema.optional(Schema.String) }),
+  [Schema.Record(Schema.String, Schema.Any)],
+)
+export type OpenAICompatibleChatBody = Schema.Schema.Type<typeof OpenAICompatibleChatBody>
+
+// Typed AI SDK `@ai-sdk/openai-compatible` options. Known keys are lowered
+// explicitly; everything else passes through to the wire body with its
+// top-level key snake-cased.
+interface CompatibleOptions {
+  readonly user?: string
+  readonly reasoningEffort?: string
+  readonly textVerbosity?: string
+  readonly strictJsonSchema?: boolean
+  readonly [extra: string]: unknown
+}
+const COMPATIBLE_KNOWN_KEYS: ReadonlySet<string> = new Set([
+  "user",
+  "reasoningEffort",
+  "textVerbosity",
+  "strictJsonSchema",
+])
+
+// Match AI SDK `@ai-sdk/openai-compatible` option resolution: the deprecated
+// `openai-compatible` alias, the canonical `openaiCompatible` key, the raw
+// provider name (dot-split so e.g. `opencode.internal` matches `opencode`),
+// and its camelCase variant. Later sources override earlier ones.
+const bodyOptions = (request: LLMRequest) => {
+  const provider = String(request.model.provider).split(".")[0]
+  const camel = provider.replace(/[_-]([a-z])/g, (_, value: string) => value.toUpperCase())
+  const options = ProviderOptions.merge(request, [
+    "openai-compatible",
+    "openaiCompatible",
+    provider,
+    camel,
+  ]) as CompatibleOptions
+  return {
+    ...ProviderOptions.passthrough(options, COMPATIBLE_KNOWN_KEYS),
+    ...(options.user !== undefined ? { user: options.user } : {}),
+    ...(options.reasoningEffort !== undefined ? { reasoning_effort: options.reasoningEffort } : {}),
+    ...(options.textVerbosity !== undefined ? { verbosity: options.textVerbosity } : {}),
+  }
+}
+
+export const protocol = Protocol.make({
+  id: ADAPTER,
+  body: {
+    schema: OpenAICompatibleChatBody,
+    // Drop providerOptions before delegating so OpenAI Chat's OpenAI-only
+    // option validation does not reject compatible-route requests whose
+    // provider id happens to be `openai` or use extended reasoning efforts.
+    from: (request) =>
+      OpenAIChat.protocol.body
+        .from({ ...request, providerOptions: undefined })
+        .pipe(Effect.map((body) => ({ ...body, ...bodyOptions(request) }))),
+  },
+  stream: OpenAIChat.protocol.stream,
+})
+
 /**
 * Route for non-OpenAI providers that expose an OpenAI Chat-compatible
- * `/chat/completions` endpoint. Reuses `OpenAIChat.protocol` end-to-end and
- * overrides only the route id so providers can be resolved per-family without
- * colliding with native OpenAI. Provider helpers configure the route endpoint
- * before model selection.
+ * `/chat/completions` endpoint. Reuses OpenAI Chat streaming behavior while
+ * allowing compatible providers to pass through additional request-body
+ * options such as `enable_thinking` and extended reasoning efforts.
 */
 export const route = Route.make({
  id: ADAPTER,
-  protocol: OpenAIChat.protocol,
+  protocol,
  endpoint: Endpoint.path("/chat/completions"),
  framing: Framing.sse,
 })
@@ -19,6 +19,7 @@ import {
 } from "../schema"
 import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared"
 import { OpenAIOptions } from "./utils/openai-options"
+import { ProviderOptions } from "./utils/provider-options"
 import { Lifecycle } from "./utils/lifecycle"
 import { ToolStream } from "./utils/tool-stream"

@@ -111,12 +112,23 @@ const OpenAIResponsesCoreFields = {
  tools: optionalArray(OpenAIResponsesTool),
  tool_choice: Schema.optional(OpenAIResponsesToolChoice),
  store: Schema.optional(Schema.Boolean),
+  conversation: Schema.optional(Schema.String),
+  max_tool_calls: Schema.optional(Schema.Number),
+  metadata: Schema.optional(JsonObject),
+  parallel_tool_calls: Schema.optional(Schema.Boolean),
+  previous_response_id: Schema.optional(Schema.String),
  prompt_cache_key: Schema.optional(Schema.String),
-  include: optionalArray(Schema.Literal("reasoning.encrypted_content")),
+  prompt_cache_retention: Schema.optional(Schema.String),
+  safety_identifier: Schema.optional(Schema.String),
+  service_tier: Schema.optional(Schema.String),
+  top_logprobs: Schema.optional(Schema.Number),
+  truncation: Schema.optional(Schema.String),
+  user: Schema.optional(Schema.String),
+  include: optionalArray(Schema.String),
  reasoning: Schema.optional(
    Schema.Struct({
      effort: Schema.optional(OpenAIOptions.OpenAIReasoningEffort),
-      summary: Schema.optional(Schema.Literal("auto")),
+      summary: Schema.optional(Schema.String),
    }),
  ),
  text: Schema.optional(
@@ -129,10 +141,15 @@ const OpenAIResponsesCoreFields = {
  top_p: Schema.optional(Schema.Number),
 }

-const OpenAIResponsesBody = Schema.Struct({
-  ...OpenAIResponsesCoreFields,
-  stream: Schema.Literal(true),
-})
+// Unknown provider options are passed through verbatim with their top-level
+// key snake-cased; the rest record validates them against any JSON value.
+const OpenAIResponsesBody = Schema.StructWithRest(
+  Schema.Struct({
+    ...OpenAIResponsesCoreFields,
+    stream: Schema.Literal(true),
+  }),
+  [Schema.Record(Schema.String, Schema.Any)],
+)
 export type OpenAIResponsesBody = Schema.Schema.Type<typeof OpenAIResponsesBody>

 const OpenAIResponsesWebSocketMessage = Schema.StructWithRest(
@@ -293,14 +310,15 @@ const lowerToolResultOutput = Effect.fn("OpenAIResponses.lowerToolResultOutput")
  // Text/json/error results are encoded as a plain string for backward
  // compatibility with existing cassettes and provider expectations.
  if (part.result.type !== "content") return ProviderShared.toolResultText(part)
-  return yield* Effect.forEach(part.result.value, lowerToolResultContentItem)
+  const items: ReadonlyArray<ToolResultContentPart> = part.result.value
+  return yield* Effect.forEach(items, lowerToolResultContentItem)
 })

 const lowerMessages = Effect.fn("OpenAIResponses.lowerMessages")(function* (request: LLMRequest) {
  const system: OpenAIResponsesInputItem[] =
    request.system.length === 0 ? [] : [{ role: "system", content: ProviderShared.joinText(request.system) }]
  const input: OpenAIResponsesInputItem[] = [...system]
-  const store = OpenAIOptions.store(request)
+  const store = OpenAIOptions.options(request).store

  for (const message of request.messages) {
    if (message.role === "user") {
@@ -355,25 +373,47 @@ const lowerMessages = Effect.fn("OpenAIResponses.lowerMessages")(function* (requ
  return input
 })

-const lowerOptions = Effect.fn("OpenAIResponses.lowerOptions")(function* (request: LLMRequest) {
-  const store = OpenAIOptions.store(request)
-  const promptCacheKey = OpenAIOptions.promptCacheKey(request)
-  const effort = OpenAIOptions.reasoningEffort(request)
-  if (effort && !OpenAIOptions.isReasoningEffort(effort))
-    return yield* invalid(`OpenAI Responses does not support reasoning effort ${effort}`)
-  const summary = OpenAIOptions.reasoningSummary(request)
-  const encryptedState = OpenAIOptions.encryptedReasoning(request)
-  const verbosity = OpenAIOptions.textVerbosity(request)
-  const instructions = OpenAIOptions.instructions(request)
+const lowerOptions = (request: LLMRequest) => {
+  const options = OpenAIOptions.options(request)
+  // OpenAI Responses does not accept the `max` reasoning effort variant.
+  const effort = OpenAIOptions.isReasoningEffort(options.reasoningEffort) ? options.reasoningEffort : undefined
+  const summary = options.reasoningSummary
+  const verbosity = options.textVerbosity
+  // `logprobs` is enabled only by `true` or a numeric top-N. `false` and
+  // `undefined` leave the request without the logprobs include + top_logprobs.
+  const logprobsEnabled = options.logprobs === true || typeof options.logprobs === "number"
+  const include = (() => {
+    const base = options.include ? [...options.include] : []
+    if (options.includeEncryptedReasoning && !base.includes("reasoning.encrypted_content")) {
+      base.push("reasoning.encrypted_content")
+    }
+    if (logprobsEnabled && !base.includes("message.output_text.logprobs")) {
+      base.push("message.output_text.logprobs")
+    }
+    return base.length > 0 ? base : undefined
+  })()
+  const topLogprobs = typeof options.logprobs === "number" ? options.logprobs : options.logprobs === true ? 20 : undefined
  return {
-    ...(instructions ? { instructions } : {}),
-    ...(store !== undefined ? { store } : {}),
-    ...(promptCacheKey ? { prompt_cache_key: promptCacheKey } : {}),
-    ...(encryptedState ? { include: ["reasoning.encrypted_content"] as const } : {}),
-    ...(effort || summary ? { reasoning: { effort, summary } } : {}),
-    ...(verbosity ? { text: { verbosity } } : {}),
+    ...ProviderOptions.passthrough(options, OpenAIOptions.KNOWN_KEYS),
+    ...(options.instructions !== undefined ? { instructions: options.instructions } : {}),
+    ...(options.store !== undefined ? { store: options.store } : {}),
+    ...(options.conversation !== undefined ? { conversation: options.conversation } : {}),
+    ...(options.maxToolCalls !== undefined ? { max_tool_calls: options.maxToolCalls } : {}),
+    ...(options.metadata !== undefined ? { metadata: options.metadata } : {}),
+    ...(options.parallelToolCalls !== undefined ? { parallel_tool_calls: options.parallelToolCalls } : {}),
+    ...(options.previousResponseId !== undefined ? { previous_response_id: options.previousResponseId } : {}),
+    ...(options.promptCacheKey !== undefined ? { prompt_cache_key: options.promptCacheKey } : {}),
+    ...(options.promptCacheRetention !== undefined ? { prompt_cache_retention: options.promptCacheRetention } : {}),
+    ...(options.safetyIdentifier !== undefined ? { safety_identifier: options.safetyIdentifier } : {}),
+    ...(options.serviceTier !== undefined ? { service_tier: options.serviceTier } : {}),
+    ...(topLogprobs !== undefined ? { top_logprobs: topLogprobs } : {}),
+    ...(options.truncation !== undefined ? { truncation: options.truncation } : {}),
+    ...(options.user !== undefined ? { user: options.user } : {}),
+    ...(include ? { include } : {}),
+    ...(effort !== undefined || summary !== undefined ? { reasoning: { effort, summary } } : {}),
+    ...(verbosity !== undefined ? { text: { verbosity } } : {}),
  }
-})
+}

 const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request: LLMRequest) {
  const generation = request.generation
@@ -386,7 +426,7 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
    max_output_tokens: generation?.maxTokens,
    temperature: generation?.temperature,
    top_p: generation?.topP,
-    ...(yield* lowerOptions(request)),
+    ...lowerOptions(request),
  }
 })

@@ -1,60 +1,73 @@
 import { Schema } from "effect"
-import type { LLMRequest, ReasoningEffort, TextVerbosity as TextVerbosityValue } from "../../schema"
-import { ReasoningEfforts, TextVerbosity } from "../../schema"
+import type { LLMRequest } from "../../schema"
+import { ReasoningEfforts, TextVerbosity, type ReasoningEffort } from "../../schema"
+import { ProviderOptions } from "./provider-options"

 export const OpenAIReasoningEfforts = ReasoningEfforts.filter(
  (effort): effort is Exclude<ReasoningEffort, "max"> => effort !== "max",
 )
 export type OpenAIReasoningEffort = (typeof OpenAIReasoningEfforts)[number]

-const REASONING_EFFORTS = new Set<string>(ReasoningEfforts)
 const OPENAI_REASONING_EFFORTS = new Set<string>(OpenAIReasoningEfforts)
-const TEXT_VERBOSITY = new Set<string>(["low", "medium", "high"])

 export const OpenAIReasoningEffort = Schema.Literals(OpenAIReasoningEfforts)
 export const OpenAITextVerbosity = TextVerbosity

-const isAnyReasoningEffort = (effort: unknown): effort is ReasoningEffort =>
-  typeof effort === "string" && REASONING_EFFORTS.has(effort)
-
 export const isReasoningEffort = (effort: unknown): effort is OpenAIReasoningEffort =>
  typeof effort === "string" && OPENAI_REASONING_EFFORTS.has(effort)

-const isTextVerbosity = (value: unknown): value is TextVerbosityValue =>
-  typeof value === "string" && TEXT_VERBOSITY.has(value)
-
-const options = (request: LLMRequest) => request.providerOptions?.openai
-
-export const store = (request: LLMRequest): boolean | undefined => {
-  const value = options(request)?.store
-  return typeof value === "boolean" ? value : undefined
+// Typed AI SDK OpenAI options. Mirrors the camelCase surface AI SDK accepts.
+// Known keys are typed; everything else passes through to the wire body with
+// its top-level key snake-cased.
+export interface Options {
+  readonly store?: boolean
+  readonly promptCacheKey?: string
+  readonly promptCacheRetention?: string
+  readonly reasoningEffort?: ReasoningEffort
+  readonly reasoningSummary?: string
+  readonly textVerbosity?: "low" | "medium" | "high"
+  readonly include?: ReadonlyArray<string>
+  readonly includeEncryptedReasoning?: boolean
+  readonly instructions?: string
+  readonly conversation?: string
+  readonly maxToolCalls?: number
+  readonly metadata?: Record<string, unknown>
+  readonly parallelToolCalls?: boolean
+  readonly previousResponseId?: string
+  readonly safetyIdentifier?: string
+  readonly serviceTier?: string
+  readonly logprobs?: boolean | number
+  readonly truncation?: string
+  readonly user?: string
+  readonly [extra: string]: unknown
 }

-export const reasoningEffort = (request: LLMRequest): ReasoningEffort | undefined => {
-  const value = options(request)?.reasoningEffort
-  return isAnyReasoningEffort(value) ? value : undefined
-}
+export const KNOWN_KEYS: ReadonlySet<string> = new Set([
+  "store",
+  "promptCacheKey",
+  "promptCacheRetention",
+  "reasoningEffort",
+  "reasoningSummary",
+  "textVerbosity",
+  "include",
+  "includeEncryptedReasoning",
+  "instructions",
+  "conversation",
+  "maxToolCalls",
+  "metadata",
+  "parallelToolCalls",
+  "previousResponseId",
+  "safetyIdentifier",
+  "serviceTier",
+  "logprobs",
+  "truncation",
+  "user",
+])

-export const reasoningSummary = (request: LLMRequest): "auto" | undefined => {
-  return options(request)?.reasoningSummary === "auto" ? "auto" : undefined
-}
-
-export const encryptedReasoning = (request: LLMRequest) =>
-  options(request)?.includeEncryptedReasoning === true ? true : undefined
-
-export const promptCacheKey = (request: LLMRequest) => {
-  const value = options(request)?.promptCacheKey
-  return typeof value === "string" ? value : undefined
-}
-
-export const textVerbosity = (request: LLMRequest) => {
-  const value = options(request)?.textVerbosity
-  return isTextVerbosity(value) ? value : undefined
-}
-
-export const instructions = (request: LLMRequest) => {
-  const value = options(request)?.instructions
-  return typeof value === "string" ? value : undefined
-}
+// Read the merged `openai` provider option bag. Producers
+// (`packages/opencode/src/provider/transform.ts`) emit typed values; we widen
+// only the index signature so passthrough keys remain reachable. Invalid
+// shapes surface in the lowerer where they're consumed, not at decode time.
+export const options = (request: LLMRequest): Options => ProviderOptions.merge(request, ["openai"]) as Options

 export * as OpenAIOptions from "./openai-options"
@@ -0,0 +1,34 @@
+import type { LLMRequest } from "../../schema"
+
+const isRecord = (value: unknown): value is Record<string, unknown> =>
+  typeof value === "object" && value !== null && !Array.isArray(value)
+
+// Convert a single top-level option key from camelCase to snake_case. Values
+// are left verbatim — recursive conversion would mangle structured payloads
+// (IDs, nested provider-shaped objects) and provider APIs do not require it.
+// PascalCase (`FooBar`) becomes `foo_bar` without a leading underscore.
+export const snakeKey = (key: string) =>
+  key.replace(/([a-z0-9])([A-Z])/g, "$1_$2").toLowerCase()
+
+// Merge provider option namespaces using AI SDK precedence semantics: later
+// sources override earlier ones, missing namespaces are skipped. Used by every
+// native protocol that reads request-level provider options.
+export const merge = (request: LLMRequest, keys: ReadonlyArray<string>) => {
+  const sources = keys.map((key) => request.providerOptions?.[key]).filter(isRecord)
+  return Object.assign({}, ...sources) as Record<string, unknown>
+}
+
+// Spread the unknown remainder of a merged option bag onto a provider body.
+// `consumed` lists keys already lowered explicitly so they aren't duplicated
+// or echoed at the wrong shape.
+export const passthrough = (options: Record<string, unknown>, consumed: ReadonlySet<string>) => {
+  const result: Record<string, unknown> = {}
+  for (const [key, value] of Object.entries(options)) {
+    if (consumed.has(key)) continue
+    if (value === undefined) continue
+    result[snakeKey(key)] = value
+  }
+  return result
+}
+
+export * as ProviderOptions from "./provider-options"
@@ -209,6 +209,80 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("lowers Anthropic thinking provider option (enabled)", () =>
+    Effect.gen(function* () {
+      const prepared = yield* LLMClient.prepare(
+        LLM.request({
+          model,
+          prompt: "think",
+          providerOptions: { anthropic: { thinking: { type: "enabled", budgetTokens: 12345 } } },
+        }),
+      )
+      expect(prepared.body).toMatchObject({ thinking: { type: "enabled", budget_tokens: 12345 } })
+    }),
+  )
+
+  it.effect("lowers Anthropic adaptive thinking with effort sibling", () =>
+    Effect.gen(function* () {
+      const prepared = yield* LLMClient.prepare(
+        LLM.request({
+          model,
+          prompt: "think",
+          providerOptions: { anthropic: { thinking: { type: "adaptive", display: "summarized" }, effort: "max" } },
+        }),
+      )
+      expect(prepared.body).toMatchObject({
+        thinking: { type: "adaptive", display: "summarized" },
+        effort: "max",
+      })
+    }),
+  )
+
+  it.effect("sets per-tool eager_input_streaming only when toolStreaming is true", () =>
+    Effect.gen(function* () {
+      const off = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+        LLM.request({
+          model,
+          prompt: "use tool",
+          providerOptions: { anthropic: { toolStreaming: false } },
+          tools: [{ name: "lookup", description: "lookup", inputSchema: { type: "object" } }],
+        }),
+      )
+      expect(off.body.tools?.[0]?.eager_input_streaming).toBeUndefined()
+
+      const on = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+        LLM.request({
+          model,
+          prompt: "use tool",
+          providerOptions: { anthropic: { toolStreaming: true } },
+          tools: [{ name: "lookup", description: "lookup", inputSchema: { type: "object" } }],
+        }),
+      )
+      expect(on.body.tools?.[0]?.eager_input_streaming).toBe(true)
+    }),
+  )
+
+  it.effect("passes unknown Anthropic provider options through with snake-cased keys", () =>
+    Effect.gen(function* () {
+      const prepared = yield* LLMClient.prepare(
+        LLM.request({
+          model,
+          prompt: "go",
+          providerOptions: {
+            anthropic: {
+              anthropicBeta: ["claude-2024-07-15"],
+              customField: { keepCamelCase: true },
+            },
+          },
+        }),
+      )
+      expect(prepared.body).toMatchObject({
+        anthropic_beta: ["claude-2024-07-15"],
+        custom_field: { keepCamelCase: true },
+      })
+    }),
+  )
+
  it.effect("lowers preserved Anthropic reasoning signature metadata", () =>
    Effect.gen(function* () {
      const prepared = yield* LLMClient.prepare(
@@ -6,7 +6,7 @@ import { Auth, LLMClient } from "../../src/route"
 import * as OpenAICompatible from "../../src/providers/openai-compatible"
 import * as OpenAICompatibleChat from "../../src/protocols/openai-compatible-chat"
 import { it } from "../lib/effect"
-import { dynamicResponse } from "../lib/http"
+import { dynamicResponse, fixedResponse } from "../lib/http"
 import { sseEvents } from "../lib/sse"

 const Json = Schema.fromJsonString(Schema.Unknown)
@@ -199,6 +199,134 @@ describe("OpenAI-compatible Chat route", () => {
    }),
  )

+  it.effect("passes through compatible options and prior reasoning for tool continuations", () =>
+    Effect.gen(function* () {
+      const prepared = yield* LLMClient.prepare(
+        LLM.request({
+          model,
+          providerOptions: {
+            deepseek: {
+              reasoningEffort: "max",
+              textVerbosity: "low",
+              promptCacheKey: "session_123",
+              strictJsonSchema: false,
+              enable_thinking: true,
+            },
+          },
+          messages: [
+            Message.user("Audit the site"),
+            Message.make({
+              role: "assistant",
+              native: { openaiCompatible: { reasoning_content: "I should inspect the page." } },
+              content: [ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "page" } })],
+            }),
+          ],
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        reasoning_effort: "max",
+        verbosity: "low",
+        prompt_cache_key: "session_123",
+        enable_thinking: true,
+        messages: [
+          { role: "user", content: "Audit the site" },
+          {
+            role: "assistant",
+            reasoning_content: "I should inspect the page.",
+            tool_calls: [
+              {
+                id: "call_1",
+                function: { name: "lookup", arguments: '{"query":"page"}' },
+              },
+            ],
+          },
+        ],
+      })
+      expect(prepared.body).not.toHaveProperty("strictJsonSchema")
+    }),
+  )
+
+  it.effect("preserves structured reasoning_details on compatible continuations", () =>
+    Effect.gen(function* () {
+      const details = [
+        { type: "reasoning.text", text: "Let me work through this.", format: "anthropic-claude-v1", index: 0 },
+        { type: "reasoning.encrypted", data: "sha256:abc123", format: "anthropic-claude-v1", index: 1 },
+      ]
+      const prepared = yield* LLMClient.prepare(
+        LLM.request({
+          model,
+          messages: [
+            Message.make({
+              role: "assistant",
+              native: { openaiCompatible: { reasoning_details: details } },
+              content: [ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })],
+            }),
+          ],
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        messages: [{ role: "assistant", reasoning_details: details }],
+      })
+    }),
+  )
+
+  it.effect("resolves dot-scoped compatible provider options", () =>
+    Effect.gen(function* () {
+      const prepared = yield* LLMClient.prepare(
+        LLM.request({
+          model: OpenAICompatibleChat.route
+            .with({ provider: "opencode.internal", endpoint: { baseURL: "https://api.example.test/v1" } })
+            .model({ id: "reasoning-model" }),
+          prompt: "Think.",
+          providerOptions: { opencode: { reasoningEffort: "max", enable_thinking: true } },
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({ reasoning_effort: "max", enable_thinking: true })
+    }),
+  )
+
+  it.effect("does not apply OpenAI effort limits to compatible providers", () =>
+    Effect.gen(function* () {
+      const prepared = yield* LLMClient.prepare(
+        LLM.request({
+          model: OpenAICompatibleChat.route
+            .with({ provider: "openai", endpoint: { baseURL: "https://compatible.example.test/v1" } })
+            .model({ id: "reasoning-model" }),
+          prompt: "Think.",
+          providerOptions: { openai: { reasoningEffort: "max" } },
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({ reasoning_effort: "max" })
+    }),
+  )
+
+  it.effect("parses compatible reasoning field variants", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              deltaChunk({ reasoning: "fallback" }),
+              deltaChunk({
+                reasoning_details: [
+                  { type: "reasoning.text", text: " text-detail", format: "anthropic-claude-v1", index: 0 },
+                  { type: "reasoning.summary", summary: " summary-detail", format: "anthropic-claude-v1", index: 1 },
+                ],
+              }),
+              deltaChunk({}, "stop"),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("fallback text-detail summary-detail")
+    }),
+  )
+
  it.effect("posts to the configured compatible endpoint and parses text usage", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(request).pipe(
@@ -407,6 +407,30 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("passes unknown OpenAI provider options through with snake-cased keys", () =>
+    Effect.gen(function* () {
+      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+        LLM.request({
+          model: OpenAI.configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).model("gpt-5.2"),
+          prompt: "passthrough",
+          providerOptions: {
+            openai: {
+              customCamelCaseField: "value",
+              already_snake_case: 42,
+              nested: { keepCamelCase: true },
+            },
+          },
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        custom_camel_case_field: "value",
+        already_snake_case: 42,
+        nested: { keepCamelCase: true },
+      })
+    }),
+  )
+
  it.effect("request OpenAI provider options override route defaults", () =>
    Effect.gen(function* () {
      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
@@ -110,7 +110,9 @@ const messages = (input: readonly ModelMessage[]) => {
      Message.make({
        role: message.role,
        content: content(message.content),
-        native: isRecord(message.providerOptions) ? { providerOptions: message.providerOptions } : undefined,
+        // Message provider options are already provider-native wire metadata
+        // (for example DeepSeek's reasoning_content continuation field).
+        native: isRecord(message.providerOptions) ? message.providerOptions : undefined,
      }),
    ]
  })
@@ -17,7 +17,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"gpt-5.2-codex\",\"input\":[{\"role\":\"system\",\"content\":\"Answer using tools when appropriate.\\nUse the get_weather tool exactly once to look up Paris, then reply with exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get the current weather for a city.\",\"parameters\":{\"$schema\":\"http://json-schema.org/draft-07/schema#\",\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}],\"store\":false,\"prompt_cache_key\":\"session-recorded-opencode-loop\",\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"max_output_tokens\":32000,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.2-codex\",\"input\":[{\"role\":\"system\",\"content\":\"Answer using tools when appropriate.\\nUse the get_weather tool exactly once to look up Paris, then reply with exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get the current weather for a city.\",\"parameters\":{\"$schema\":\"http://json-schema.org/draft-07/schema#\",\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}],\"store\":false,\"prompt_cache_key\":\"session-recorded-opencode-loop\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"max_output_tokens\":32000,\"stream\":true}"
      },
      "response": {
        "status": 200,
@@ -35,7 +35,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"gpt-5.2-codex\",\"input\":[{\"role\":\"system\",\"content\":\"Answer using tools when appropriate.\\nUse the get_weather tool exactly once to look up Paris, then reply with exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]},{\"type\":\"function_call\",\"call_id\":\"call_DfI0RwTrlaizfnQ9zkJC8rks\",\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":{}}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_DfI0RwTrlaizfnQ9zkJC8rks\",\"output\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get the current weather for a city.\",\"parameters\":{\"$schema\":\"http://json-schema.org/draft-07/schema#\",\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}],\"store\":false,\"prompt_cache_key\":\"session-recorded-opencode-loop\",\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"max_output_tokens\":32000,\"stream\":true}"
+        "body": "{\"model\":\"gpt-5.2-codex\",\"input\":[{\"role\":\"system\",\"content\":\"Answer using tools when appropriate.\\nUse the get_weather tool exactly once to look up Paris, then reply with exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":[{\"type\":\"input_text\",\"text\":\"What is the weather in Paris?\"}]},{\"type\":\"function_call\",\"call_id\":\"call_DfI0RwTrlaizfnQ9zkJC8rks\",\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":{}}\"},{\"type\":\"function_call_output\",\"call_id\":\"call_DfI0RwTrlaizfnQ9zkJC8rks\",\"output\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"name\":\"get_weather\",\"description\":\"Get the current weather for a city.\",\"parameters\":{\"$schema\":\"http://json-schema.org/draft-07/schema#\",\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}],\"store\":false,\"prompt_cache_key\":\"session-recorded-opencode-loop\",\"include\":[\"reasoning.encrypted_content\"],\"reasoning\":{\"effort\":\"medium\",\"summary\":\"auto\"},\"max_output_tokens\":32000,\"stream\":true}"
      },
      "response": {
        "status": 200,
@@ -6,6 +6,7 @@ import { Effect, Layer, Stream } from "effect"
 import { LLMNative } from "@/session/llm/native-request"
 import { LLMNativeRuntime } from "@/session/llm/native-runtime"
 import type { Provider } from "@/provider/provider"
+import { ProviderTransform } from "@/provider/transform"
 import { ModelID, ProviderID } from "@/provider/schema"
 import { OAUTH_DUMMY_KEY } from "@/auth"
 import { testEffect } from "../lib/effect"
@@ -70,6 +71,21 @@ const providerInfo: Provider.Info = {
  models: {},
 }

+const compatibleModel: Provider.Model = {
+  ...baseModel,
+  id: ModelID.make("deepseek-v4-flash-free"),
+  providerID: ProviderID.make("opencode"),
+  api: {
+    id: "deepseek-v4-flash-free",
+    url: "https://ai.example.test/v1",
+    npm: "@ai-sdk/openai-compatible",
+  },
+  capabilities: {
+    ...baseModel.capabilities,
+    interleaved: { field: "reasoning_content" },
+  },
+}
+
 const it = testEffect(
  LLMClient.layer.pipe(Layer.provide(Layer.mergeAll(RequestExecutor.defaultLayer, WebSocketExecutor.layer))),
 )
@@ -326,6 +342,70 @@ describe("session.llm-native.request", () => {
    ])
  })

+  it.effect("preserves OpenAI-compatible reasoning continuation and provider options", () =>
+    Effect.gen(function* () {
+      const messages = ProviderTransform.message(
+        [
+          { role: "user", content: "Audit the site" },
+          {
+            role: "assistant",
+            content: [
+              { type: "reasoning", text: "I should inspect the page." },
+              {
+                type: "tool-call",
+                toolCallId: "call-1",
+                toolName: "devtools_new_page",
+                input: { url: "https://example.test" },
+              },
+            ],
+          },
+        ] as ModelMessage[],
+        compatibleModel,
+        {},
+      )
+
+      const prepared = yield* LLMClient.prepare(
+        LLMNative.request({
+          model: compatibleModel,
+          apiKey: "test-key",
+          messages,
+          providerOptions: ProviderTransform.providerOptions(compatibleModel, {
+            reasoningEffort: "max",
+            textVerbosity: "low",
+            promptCacheKey: "session-1",
+            enable_thinking: true,
+          }),
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        model: "deepseek-v4-flash-free",
+        reasoning_effort: "max",
+        verbosity: "low",
+        prompt_cache_key: "session-1",
+        enable_thinking: true,
+        messages: [
+          { role: "user", content: "Audit the site" },
+          {
+            role: "assistant",
+            content: null,
+            reasoning_content: "I should inspect the page.",
+            tool_calls: [
+              {
+                id: "call-1",
+                type: "function",
+                function: {
+                  name: "devtools_new_page",
+                  arguments: '{"url":"https://example.test"}',
+                },
+              },
+            ],
+          },
+        ],
+      })
+    }),
+  )
+
  test("selects native request routes for provider packages", () => {
    const openai = LLMNative.model({
      model: { ...baseModel, api: { ...baseModel.api, url: "", npm: "@ai-sdk/openai" } },