feat(tui): project picker with footer typewriter flash

.opencode/plugins/tui/discovery-smoke.ts

@@ -2,15 +2,12 @@ import type { Context } from "../../../packages/plugin/src/tui/context"
 
 export default {
   id: "test.tui-discovery-smoke",
-  setup(context: Context) {
-    const timer = setTimeout(() => {
-      context.ui.toast.show({
-        title: "TUI plugin discovery works",
-        message: "Loaded .opencode/plugins/tui/discovery-smoke.ts",
-        variant: "success",
-        duration: 30_000,
-      })
-    }, 1_000)
-    return () => clearTimeout(timer)
+  setup(_context: Context) {
+    // context.ui.toast.show({
+    //   title: "TUI plugin discovery works",
+    //   message: "Loaded .opencode/plugins/tui/discovery-smoke.ts",
+    //   variant: "success",
+    //   duration: 30_000,
+    // })
   },
 }

bun.lock

@@ -600,6 +600,7 @@
         "zod": "catalog:",
       },
       "devDependencies": {
+        "@opencode-ai/theme": "workspace:*",
         "@opentui/core": "catalog:",
         "@opentui/keymap": "catalog:",
         "@opentui/solid": "catalog:",
@@ -611,12 +612,14 @@
         "typescript": "catalog:",
       },
       "peerDependencies": {
+        "@opencode-ai/theme": "workspace:*",
         "@opentui/core": ">=0.4.5",
         "@opentui/keymap": ">=0.4.5",
         "@opentui/solid": ">=0.4.5",
         "solid-js": ">=1.9.0",
       },
       "optionalPeers": [
+        "@opencode-ai/theme",
         "@opentui/core",
         "@opentui/keymap",
         "@opentui/solid",

docs/design/v1-v2-database-migration.md

@@ -0,0 +1,118 @@
+# V1 to V2 Database Migration
+
+## Approach
+
+- Use the `dev` branch database schema and migration registry as the V1 baseline.
+- Remove migrations that exist only on the V2 branch.
+- Generate one canonical migration from the `dev` schema to the final V2 schema.
+- Add explicit data operations to that migration where generated DDL is insufficient.
+- Test the migration against a populated database at the exact `dev` schema.
+
+## Preserve
+
+The canonical V1 data remains in its existing tables. In particular, preserve `session`, `message`, and `part` rows.
+
+Preserve `workspace` rows and existing `session.workspace_id` values unchanged. The migration must not clear or rebuild
+workspace relationships.
+
+Keep the `todo` table and its data unchanged. V2 does not currently migrate todos into another representation, and the
+generated migration must not drop the table.
+
+## Truncate
+
+Truncate these pre-launch V2 tables before applying schema changes:
+
+- `event`
+- `event_sequence`
+- `session_message`
+
+These rows are not canonical V1 data. Truncating `event` before adding the required `event.created` column means the
+column needs neither a backfill nor a default. After truncation, rebuild `session_message` from canonical V1 `message`
+and `part` rows rather than retaining its pre-launch V2 contents.
+
+## Message Backfill
+
+Backfill canonical V1 history from `message` and `part` into `session_message`. This is the main data transformation in
+the migration. Preserving the V1 tables alone keeps the data safe but does not make existing history visible through the
+V2 session APIs, which read `session_message`.
+
+Reuse each V1 `message.id` as the corresponding `session_message.id`. Stable IDs keep the migration deterministic and
+avoid rewriting other persisted state that may refer to a message.
+
+Within each session, order V1 messages by `time_created` and then `id`, matching the existing V1 message index. Assign
+contiguous `session_message.seq` values starting at `0`.
+
+Map ordinary V1 messages one-to-one by role. Each ordinary V1 user message becomes one V2 `user` row, and each ordinary
+V1 assistant message becomes one V2 `assistant` row. Fold the source message's ordered V1 parts into that row's V2
+payload.
+
+Handle semantic marker parts before applying the ordinary mapping. In particular, a V1 user message containing a
+`compaction` part and its paired assistant summary represent one compaction operation, not two ordinary messages. Special
+part mappings must be decided explicitly before implementing the backfill.
+
+V1 synthetic content is represented by user text parts with `synthetic: true`, not by a separate message role. A V1 user
+message whose visible text parts are all synthetic should become a V2 `synthetic` message. If a V1 user message mixes
+ordinary and synthetic content, preserve the ordinary content in the V2 `user` row and emit the synthetic content as an
+adjacent V2 `synthetic` row. Ignore text parts marked `ignored`, matching V1 model-history behavior.
+
+Use the V1 compaction user message ID as the ID of the collapsed V2 compaction message. This matches V2's use of the
+admitted compaction input ID and preserves references to the initiating message.
+
+For a completed compaction, create one V2 `compaction` row with `status: "completed"`. Set `reason` from the V1
+compaction part's `auto` flag, join the paired summary assistant's nonempty text parts with blank lines for `summary`, and
+serialize the retained V1 tail beginning at `tail_start_id` for `recent`. Use an empty `recent` value when no tail was
+retained, and use the compaction user message creation time. Do not emit the paired summary assistant as a separate V2
+assistant row.
+
+After rebuilding `session_message`, seed `event_sequence` with one row per migrated session. Set its watermark to that
+session's maximum backfilled `session_message.seq`. This prevents new V2 events from reusing sequence numbers or sorting
+before migrated history. The `event` table remains empty.
+
+## Drop
+
+Drop these pre-launch V2 tables without preserving or transforming their rows:
+
+- `session_input`
+- `session_context_epoch`
+
+Do not transfer `session_input` rows into `session_pending`.
+
+## Create Empty
+
+Let the generated migration create these tables empty:
+
+- `instruction_blob`
+- `instruction_entry`
+- `instruction_state`
+- `session_pending`
+- `kv`
+
+V1 has no canonical data to backfill into these tables. V2 initializes their state as it runs.
+
+## Fork Storage
+
+V1 has no fork-boundary state to backfill. New V2 forks use a required message boundary and persist it in
+`session.fork_boundary`. The durable fork event contains no parent sequence. Its resolved boundary is one of:
+
+- `before`: copy messages before the identified message.
+- `through`: copy messages through the identified message.
+
+Forking an empty session is not supported. `session.fork_seq` and `session.fork_message_id` are not part of the final V2
+schema.
+
+New nullable session columns, including `fork_session_id`, `fork_boundary`, and `time_suspended`, require no explicit
+backfill. Existing rows naturally receive `NULL` when the generated migration adds the columns.
+
+## Verification
+
+The canonical migration test should seed representative V1 sessions, messages, parts, todos, projects, accounts,
+credentials, permissions, shares, and workspaces. After migration, it should verify:
+
+- Preserved rows and encoded values remain unchanged.
+- Todo rows remain available in the unchanged `todo` table.
+- `event` is empty, and stale pre-launch rows are absent from the rebuilt projections.
+- Backfilled `session_message` rows represent the canonical V1 `message` and `part` history.
+- Each migrated session's `event_sequence` watermark matches its maximum backfilled message sequence.
+- Dropped tables no longer exist.
+- New tables exist and are empty.
+- The final schema has no ungenerated changes.

packages/ai/AGENTS.md

@@ -46,7 +46,7 @@ const response = yield * LLMClient.generate(request)
 
 `LLM.request(...)` builds an `LLMRequest`. `LLMClient.generate(...)` reads the executable route carried by `request.model.route`, builds the provider-native body, asks the route's transport for a real `HttpClientRequest.HttpClientRequest`, sends it through `RequestExecutor.Service`, parses the provider stream into common `LLMEvent`s, and finally returns an `LLMResponse`.
 
-Use `LLMClient.stream(request)` when callers want incremental `LLMEvent`s. Use `LLMClient.generate(request)` when callers want those same events collected into an `LLMResponse`. Use `LLMClient.prepare<Body>(request)` to compile a request through the route pipeline without sending it — the optional `Body` type argument narrows `.body` to the route's native shape (e.g. `prepare<OpenAIChatBody>(...)` returns a `PreparedRequestOf<OpenAIChatBody>`). The runtime body is identical; the generic is a type-level assertion.
+Use `LLMClient.stream(request)` when callers want incremental `LLMEvent`s. Use `LLMClient.generate(request)` when callers want those same events collected into an `LLMResponse`.
 
 Filter or narrow `LLMEvent` streams with `LLMEvent.is.*` (camelCase guards, e.g. `events.filter(LLMEvent.is.toolCall)`). The kebab-case `LLMEvent.guards["tool-call"]` form also works but prefer `is.*` in new code.
 
@@ -138,13 +138,13 @@ packages/ai/src/
     ids.ts                  branded IDs, literal types, ProviderMetadata
     options.ts              Generation/Provider/Http options, Limits, Model, cache policy
     messages.ts             content parts, Message, ToolDefinition, LLMRequest
-    events.ts               Usage, individual events, LLMEvent, PreparedRequest, LLMResponse
+    events.ts               Usage, individual events, LLMEvent, LLMResponse
     errors.ts               error reasons, LLMError, ToolFailure
     index.ts                barrel
   llm.ts                    request constructors and convenience helpers
   route/
     index.ts                @opencode-ai/ai/route advanced barrel
-    client.ts               Route.make + LLMClient.prepare/stream/generate
+    client.ts               Route.make + LLMClient.stream/generate
     executor.ts             RequestExecutor service + transport error mapping
     protocol.ts             Protocol type + Protocol.make
     endpoint.ts             Endpoint type + Endpoint.path

packages/ai/README.md

@@ -196,7 +196,6 @@ The hosted result is represented as a provider-executed tool call and tool resul
 - **`LLM.generate` / `LLM.stream`** — re-exported from `LLMClient` for one-import use.
 - **`Message.user(...)` / `Message.assistant(...)` / `Message.tool(...)`** — message constructors from the canonical schema model.
 - **`Model.make(...)` / `ToolCallPart.make(...)` / `ToolResultPart.make(...)` / `ToolDefinition.make(...)`** — model and tool-related constructors from the canonical schema model.
-- **`LLMClient.prepare(request)`** — compile a request through protocol body construction, validation, and HTTP preparation without sending. Useful for inspection and testing.
 - **`LLMEvent.is.*`** — typed guards (`is.textDelta`, `is.toolCall`, `is.finish`, …) for filtering streams.
 - **`Image.generate({...})`** — generate images through a provider-neutral image request and response model.
 - **`ImageClient`** — Effect service and layer for image execution, parallel to `LLMClient`.

packages/ai/example/call-sites.md

@@ -568,7 +568,7 @@ App boundary = explicit durable-config -> typed-provider call
       calling `.model(...)`.
 - [x] Remove request-shaping defaults from `Model`; selected models now carry only
       id, provider, and configured route while defaults live on routes or requests.
-- [x] Rework `LLMClient.prepare` / `stream` / `generate` to read
+- [x] Rework `LLMClient.stream` / `generate` to read
       `request.model.route` directly instead of calling `registeredRoute(...)`.
 - [x] Remove `Route.make(...)` global registration from the normal execution
       path; keep route ids only as diagnostics/provider API labels.

packages/ai/example/tutorial.ts

@@ -50,18 +50,6 @@ const request = LLM.request({
   },
 })
 
-// `http` is intentionally not needed for normal calls. This shows the shape for
-// newly released provider fields before they deserve a typed provider option.
-const rawOverlayExample = LLM.request({
-  model,
-  prompt: "Show the final HTTP overlay shape.",
-  http: {
-    body: { metadata: { example: "tutorial" } },
-    headers: { "x-opencode-tutorial": "1" },
-    query: { debug: "1" },
-  },
-})
-
 // 3. `generate` sends the request and collects the event stream into one
 // response object. `response.text` is the collected text output.
 const generateOnce = Effect.gen(function* () {
@@ -222,33 +210,15 @@ const FakeEcho = {
   }),
 }
 
-// `LLMClient.prepare` is the lower-level inspection hook: it compiles through
-// body conversion, validation, endpoint, auth, and HTTP construction without
-// sending anything over the network.
-const inspectFakeProvider = Effect.gen(function* () {
-  const prepared = yield* LLMClient.prepare(
-    LLM.request({
-      model: FakeEcho.configure().model("tiny-echo"),
-      prompt: "Show me the provider pipeline.",
-    }),
-  )
-
-  console.log("\n== fake provider prepare ==")
-  console.log("route:", prepared.route)
-  console.log("body:", Formatter.formatJson(prepared.body, { space: 2 }))
-})
-
 // Provide the LLM runtime and the HTTP request executor once. Keep one path
-// enabled at a time so the tutorial can demonstrate generate, prepare, stream,
-// or tool-loop behavior without spending tokens on every example.
+// enabled at a time so the tutorial can demonstrate generate, stream, or
+// tool-loop behavior without spending tokens on every example.
 const requestExecutorLayer = RequestExecutor.fetchLayer
 const llmDeps = Layer.mergeAll(requestExecutorLayer, WebSocketExecutor.layer)
 const llmClientLayer = LLMClient.layer.pipe(Layer.provide(llmDeps))
 
 const program = Effect.gen(function* () {
   // yield* generateOnce
-  // yield* inspectFakeProvider
-  // yield* LLMClient.prepare(rawOverlayExample).pipe(Effect.andThen((prepared) => Effect.sync(() => console.log(prepared.body))))
   // yield* streamText
   // yield* generateStructuredObject
   // yield* generateDynamicObject.pipe(Effect.andThen((response) => Effect.sync(() => console.log(response.object))))

packages/ai/src/llm.ts

@@ -9,25 +9,28 @@ import {
   LLMRequest,
   LLMResponse,
   Message,
+  Model,
   SystemPart,
   ToolChoice,
   ToolDefinition,
   type ContentPart,
+  type ModelProviderOptions,
 } from "./schema"
 import { make as makeTool, toDefinitions, type ToolSchema } from "./tool"
 
 /** Input accepted by `LLM.request`, normalized into the canonical `LLMRequest` class. */
-export type RequestInput = Omit<
+export type RequestInput<SelectedModel extends Model = Model> = Omit<
   ConstructorParameters<typeof LLMRequest>[0],
-  "system" | "messages" | "tools" | "toolChoice" | "generation" | "http" | "providerOptions"
+  "model" | "system" | "messages" | "tools" | "toolChoice" | "generation" | "http" | "providerOptions"
 > & {
+  readonly model: SelectedModel
   readonly system?: string | SystemPart | ReadonlyArray<SystemPart>
   readonly prompt?: string | ContentPart | ReadonlyArray<ContentPart>
   readonly messages?: ReadonlyArray<Message | Message.Input>
   readonly tools?: ReadonlyArray<ToolDefinition.Input>
   readonly toolChoice?: ToolChoice.Input
   readonly generation?: GenerationOptions.Input
-  readonly providerOptions?: ConstructorParameters<typeof LLMRequest>[0]["providerOptions"]
+  readonly providerOptions?: NoInfer<ModelProviderOptions<SelectedModel>>
   readonly http?: HttpOptions.Input
 }
 
@@ -35,7 +38,7 @@ export const generate = LLMClient.generate
 
 export const stream = LLMClient.stream
 
-export const request = (input: RequestInput) => {
+export const request = <const SelectedModel extends Model>(input: RequestInput<SelectedModel>) => {
   const {
     system: requestSystem,
     prompt,
@@ -63,7 +66,7 @@ const GENERATE_OBJECT_TOOL_NAME = "generate_object"
 
 const GENERATE_OBJECT_TOOL_DESCRIPTION = "Return the structured result by calling this tool."
 
-type GenerateObjectBase = Omit<RequestInput, "tools" | "toolChoice">
+type GenerateObjectBase<SelectedModel extends Model = Model> = Omit<RequestInput<SelectedModel>, "tools" | "toolChoice">
 
 export class GenerateObjectResponse<T> {
   constructor(
@@ -80,11 +83,13 @@ export class GenerateObjectResponse<T> {
   }
 }
 
-export interface GenerateObjectOptions<S extends ToolSchema<any>> extends GenerateObjectBase {
+export interface GenerateObjectOptions<S extends ToolSchema<any>, SelectedModel extends Model = Model>
+  extends GenerateObjectBase<SelectedModel> {
   readonly schema: S
 }
 
-export interface GenerateObjectDynamicOptions extends GenerateObjectBase {
+export interface GenerateObjectDynamicOptions<SelectedModel extends Model = Model>
+  extends GenerateObjectBase<SelectedModel> {
   /** Raw JSON Schema object describing the expected output shape. */
   readonly jsonSchema: JsonSchema.JsonSchema
 }
@@ -137,11 +142,11 @@ const runGenerateObject = Effect.fn("LLM.generateObject")(function* (
  * 2. `jsonSchema: JsonSchema.JsonSchema` — `.object` is `unknown`. Use when
  *    the schema is only available at runtime (MCP, plugin manifests). Caller validates.
  */
-export function generateObject<S extends ToolSchema<any>>(
-  options: GenerateObjectOptions<S>,
+export function generateObject<const SelectedModel extends Model, S extends ToolSchema<any>>(
+  options: GenerateObjectOptions<S, SelectedModel>,
 ): Effect.Effect<GenerateObjectResponse<Schema.Schema.Type<S>>, LLMError>
-export function generateObject(
-  options: GenerateObjectDynamicOptions,
+export function generateObject<const SelectedModel extends Model>(
+  options: GenerateObjectDynamicOptions<SelectedModel>,
 ): Effect.Effect<GenerateObjectResponse<unknown>, LLMError>
 export function generateObject(options: GenerateObjectOptions<ToolSchema<any>> | GenerateObjectDynamicOptions) {
   if ("schema" in options) {

packages/ai/src/provider-package.ts

@@ -1,4 +1,4 @@
-import type { Model } from "./schema"
+import type { Model, ProviderOptions } from "./schema"
 
 export interface Settings extends Readonly<Record<string, unknown>> {
   readonly headers?: Readonly<Record<string, string>>
@@ -9,8 +9,11 @@ export interface Settings extends Readonly<Record<string, unknown>> {
   }
 }
 
-export interface Definition<ProviderSettings extends Settings = Settings> {
-  readonly model: (modelID: string, settings: ProviderSettings) => Model
+export interface Definition<
+  ProviderSettings extends Settings = Settings,
+  Options extends ProviderOptions = ProviderOptions,
+> {
+  readonly model: (modelID: string, settings: ProviderSettings) => Model<Options>
 }
 
 export * as ProviderPackage from "./provider-package"

packages/ai/src/providers/anthropic-compatible.ts

@@ -47,7 +47,7 @@ export const configure = (input: Config) => {
   })
   return {
     id: ProviderID.make(provider),
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<AnthropicMessages.ProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -57,7 +57,10 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, AnthropicMessages.ProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined && settings.authToken !== undefined)
     throw new Error("Anthropic-compatible apiKey cannot be combined with authToken")
   return configure({

packages/ai/src/providers/anthropic.ts

@@ -52,7 +52,10 @@ export const configure = (input: Config = {}) => {
 }
 
 export const provider = configure()
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, AnthropicMessages.ProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined && settings.authToken !== undefined)
     throw new Error("Anthropic apiKey cannot be combined with authToken")
   return configure({

packages/ai/src/providers/azure.ts

@@ -99,10 +99,14 @@ export const configure = (input: Config) => {
   const modelDefaults = defaults(input)
 
   const responses = (modelID: string | ModelID) =>
-    configuredResponsesRoute.with(withOpenAIOptions(modelID, modelDefaults)).model({ id: modelID })
+    configuredResponsesRoute
+      .with(withOpenAIOptions(modelID, modelDefaults))
+      .model<OpenAIProviderOptionsInput>({ id: modelID })
 
   const chat = (modelID: string | ModelID) =>
-    configuredChatRoute.with(withOpenAIOptions(modelID, modelDefaults)).model({ id: modelID })
+    configuredChatRoute
+      .with(withOpenAIOptions(modelID, modelDefaults))
+      .model<OpenAIProviderOptionsInput>({ id: modelID })
 
   return {
     id,
@@ -133,8 +137,12 @@ const config = (settings: Settings): Config => {
   throw new Error("Azure requires resourceName or baseURL")
 }
 
-export const responsesModel: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
-  configure(config(settings)).responses(modelID)
-export const chatModel: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
-  configure(config(settings)).chat(modelID)
+export const responsesModel: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => configure(config(settings)).responses(modelID)
+export const chatModel: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => configure(config(settings)).chat(modelID)
 export const model = responsesModel

packages/ai/src/providers/cloudflare.ts

@@ -4,6 +4,7 @@ import { Auth } from "../route/auth"
 import { AuthOptions, type AtLeastOne, type ProviderAuthOption } from "../route/auth-options"
 import type { RouteDefaultsInput } from "../route/client"
 import { ProviderID, type ModelID } from "../schema"
+import type { OpenAIProviderOptionsInput } from "./openai-options"
 
 export const aiGatewayID = ProviderID.make("cloudflare-ai-gateway")
 export const workersAIID = ProviderID.make("cloudflare-workers-ai")
@@ -20,10 +21,11 @@ type GatewayURL = AtLeastOne<{
 }
 
 export type AIGatewayOptions = GatewayURL &
-  RouteDefaultsInput &
+  Omit<RouteDefaultsInput, "providerOptions"> &
   ProviderAuthOption<"optional"> & {
     /** Cloudflare AI Gateway authentication token. Sent as `cf-aig-authorization`. */
     readonly gatewayApiKey?: CloudflareSecret
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 type WorkersAIURL = AtLeastOne<{
@@ -31,7 +33,11 @@ type WorkersAIURL = AtLeastOne<{
   readonly baseURL: string
 }>
 
-export type WorkersAIOptions = WorkersAIURL & RouteDefaultsInput & ProviderAuthOption<"optional">
+export type WorkersAIOptions = WorkersAIURL &
+  Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly providerOptions?: OpenAIProviderOptionsInput
+  }
 
 export const aiGatewayBaseURL = (input: GatewayURL) => {
   if (input.baseURL) return input.baseURL
@@ -98,7 +104,7 @@ const configureAIGateway = (options: AIGatewayOptions) => {
   })
   return {
     id: aiGatewayID,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenAIProviderOptionsInput>({ id: modelID }),
     configure: configureAIGateway,
   }
 }
@@ -111,7 +117,7 @@ const configureWorkersAI = (options: WorkersAIOptions) => {
   })
   return {
     id: workersAIID,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenAIProviderOptionsInput>({ id: modelID }),
     configure: configureWorkersAI,
   }
 }

packages/ai/src/providers/github-copilot.ts

@@ -50,9 +50,11 @@ export const configure = (options: ModelOptions) => {
   const responsesRoute = configuredResponsesRoute(options)
   const chatRoute = configuredChatRoute(options)
   const responses = (modelID: string | ModelID) =>
-    responsesRoute.with(withOpenAIOptions(modelID, defaults(options))).model({ id: modelID })
+    responsesRoute
+      .with(withOpenAIOptions(modelID, defaults(options)))
+      .model<OpenAIProviderOptionsInput>({ id: modelID })
   const chat = (modelID: string | ModelID) =>
-    chatRoute.with(withOpenAIOptions(modelID, defaults(options))).model({ id: modelID })
+    chatRoute.with(withOpenAIOptions(modelID, defaults(options))).model<OpenAIProviderOptionsInput>({ id: modelID })
   return {
     id,
     model: (modelID: string | ModelID) =>

packages/ai/src/providers/google-vertex-chat.ts

@@ -1,8 +1,9 @@
 import type { ProviderPackage } from "../provider-package"
 import { OpenAICompatibleChat } from "../protocols/openai-compatible-chat"
 import type { RouteDefaultsInput } from "../route/client"
-import { ProviderID, type ModelID, type ProviderOptions } from "../schema"
+import { ProviderID, type ModelID } from "../schema"
 import { GoogleVertexShared } from "./google-vertex-shared"
+import type { OpenAIProviderOptionsInput } from "./openai-options"
 
 export const id = ProviderID.make("google-vertex")
 
@@ -11,6 +12,7 @@ export type Config = RouteDefaultsInput &
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
@@ -19,7 +21,7 @@ export interface Settings extends ProviderPackage.Settings {
   readonly baseURL?: string
   readonly location?: string
   readonly project?: string
-  readonly providerOptions?: ProviderOptions
+  readonly providerOptions?: OpenAIProviderOptionsInput
 }
 
 const route = OpenAICompatibleChat.route.with({
@@ -56,7 +58,7 @@ export const configure = (input: Config = {}) => {
   const route = configuredRoute(input)
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenAIProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -66,7 +68,7 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) => {
   if (settings.apiKey !== undefined) throw new Error("Google Vertex Chat does not support API keys")
   return configure({
     accessToken: settings.accessToken,

packages/ai/src/providers/google-vertex-messages.ts

@@ -91,7 +91,7 @@ export const configure = (input: Config = {}) => {
   const route = configuredRoute(input)
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<AnthropicMessages.ProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -101,7 +101,10 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, AnthropicMessages.ProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined) throw new Error("Google Vertex Messages does not support API keys")
   return configure({
     accessToken: settings.accessToken,

packages/ai/src/providers/google-vertex-responses.ts

@@ -1,8 +1,9 @@
 import type { ProviderPackage } from "../provider-package"
 import { OpenAICompatibleResponses } from "../protocols/openai-compatible-responses"
 import type { RouteDefaultsInput } from "../route/client"
-import { ProviderID, type ModelID, type ProviderOptions } from "../schema"
+import { ProviderID, type ModelID } from "../schema"
 import { GoogleVertexShared } from "./google-vertex-shared"
+import type { OpenResponsesProviderOptionsInput } from "./open-responses-options"
 
 export const id = ProviderID.make("google-vertex")
 
@@ -11,6 +12,7 @@ export type Config = RouteDefaultsInput &
     readonly baseURL?: string
     readonly location?: string
     readonly project?: string
+    readonly providerOptions?: OpenResponsesProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
@@ -19,7 +21,7 @@ export interface Settings extends ProviderPackage.Settings {
   readonly baseURL?: string
   readonly location?: string
   readonly project?: string
-  readonly providerOptions?: ProviderOptions
+  readonly providerOptions?: OpenResponsesProviderOptionsInput
 }
 
 const route = OpenAICompatibleResponses.route.with({
@@ -58,7 +60,7 @@ export const configure = (input: Config = {}) => {
   const route = configuredRoute(input)
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenResponsesProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -68,7 +70,10 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, OpenResponsesProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined) throw new Error("Google Vertex Responses does not support API keys")
   return configure({
     accessToken: settings.accessToken,

packages/ai/src/providers/google-vertex.ts

@@ -77,7 +77,8 @@ const configuredRoute = (input: Config, modelID: string | ModelID) => {
 export const configure = (input: Config = {}) => {
   return {
     id,
-    model: (modelID: string | ModelID) => configuredRoute(input, modelID).model({ id: modelID }),
+    model: (modelID: string | ModelID) =>
+      configuredRoute(input, modelID).model<Gemini.ProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -86,7 +87,10 @@ export const provider = {
   id,
   configure,
 }
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, Gemini.ProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => {
   if (settings.apiKey !== undefined && settings.accessToken !== undefined)
     throw new Error("Google Vertex apiKey cannot be combined with accessToken or auth")
   return configure({

packages/ai/src/providers/google.ts

@@ -50,14 +50,14 @@ export const configure = (input: Config = {}) => {
     })
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<Gemini.ProviderOptionsInput>({ id: modelID }),
     image,
     configure,
   }
 }
 
 export const provider = configure()
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
+export const model: ProviderPackage.Definition<Settings, Gemini.ProviderOptionsInput>["model"] = (modelID, settings) =>
   configure({
     apiKey: settings.apiKey,
     baseURL: settings.baseURL,

packages/ai/src/providers/openai-compatible-responses.ts

@@ -36,7 +36,7 @@ export const configure = (input: Config) => {
   })
   return {
     id: ProviderID.make(provider),
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenResponsesProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }
@@ -46,7 +46,10 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
+export const model: ProviderPackage.Definition<Settings, OpenResponsesProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) =>
   configure({
     apiKey: settings.apiKey,
     baseURL: settings.baseURL,

packages/ai/src/providers/openai-compatible.ts

@@ -4,13 +4,15 @@ import type { RouteDefaultsInput } from "../route/client"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options"
 import type { ProviderPackage } from "../provider-package"
 import { profiles, type OpenAICompatibleProfile } from "./openai-compatible-profile"
+import type { OpenAIProviderOptionsInput } from "./openai-options"
 
 export const id = ProviderID.make("openai-compatible")
 
-type GenericModelOptions = RouteDefaultsInput &
+type GenericModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
   ProviderAuthOption<"optional"> & {
     readonly provider?: string
     readonly baseURL: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 export interface Settings extends ProviderPackage.Settings {
@@ -19,9 +21,10 @@ export interface Settings extends ProviderPackage.Settings {
   readonly provider?: string
 }
 
-export type FamilyModelOptions = RouteDefaultsInput &
+export type FamilyModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
   ProviderAuthOption<"optional"> & {
     readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 export const routes = [OpenAICompatibleChat.route]
@@ -37,7 +40,8 @@ export const configure = (input: GenericModelOptions) => {
   })
   return {
     id: ProviderID.make(provider),
-    model: (modelID: string | ModelID) => route.model({ id: modelID, provider: ProviderID.make(provider) }),
+    model: (modelID: string | ModelID) =>
+      route.model<OpenAIProviderOptionsInput>({ id: modelID, provider: ProviderID.make(provider) }),
     configure,
   }
 }
@@ -63,7 +67,7 @@ export const provider = {
   configure,
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) =>
   configure({
     apiKey: settings.apiKey,
     baseURL: settings.baseURL,

packages/ai/src/providers/openai.ts

@@ -86,10 +86,15 @@ export const configure = (input: Config = {}) => {
   const chatRoute = configuredRoute(OpenAIChat.route, input)
   const modelDefaults = defaults(input)
   const responses = (id: string | ModelID) =>
-    responsesRoute.with(withOpenAIOptions(id, modelDefaults, { textVerbosity: true })).model({ id })
+    responsesRoute
+      .with(withOpenAIOptions(id, modelDefaults, { textVerbosity: true }))
+      .model<OpenAIProviderOptionsInput>({ id })
   const responsesWebSocket = (id: string | ModelID) =>
-    responsesWebSocketRoute.with(withOpenAIOptions(id, modelDefaults, { textVerbosity: true })).model({ id })
-  const chat = (id: string | ModelID) => chatRoute.with(withOpenAIOptions(id, modelDefaults)).model({ id })
+    responsesWebSocketRoute
+      .with(withOpenAIOptions(id, modelDefaults, { textVerbosity: true }))
+      .model<OpenAIProviderOptionsInput>({ id })
+  const chat = (id: string | ModelID) =>
+    chatRoute.with(withOpenAIOptions(id, modelDefaults)).model<OpenAIProviderOptionsInput>({ id })
   const image = (modelID: string | ModelID) =>
     OpenAIImages.model({
       id: modelID,
@@ -132,15 +137,17 @@ const config = (settings: Settings): Config => {
   }
 }
 
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+export const model: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (modelID, settings) => {
   const configured = configure(config(settings))
   if (settings.transport === undefined || settings.transport === "http") return configured.responses(modelID)
   if (settings.transport === "websocket") return configured.responsesWebSocket(modelID)
   throw new Error(`Unsupported OpenAI Responses transport: ${String(settings.transport)}`)
 }
 
-export const chatModel: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
-  configure(config(settings)).chat(modelID)
+export const chatModel: ProviderPackage.Definition<Settings, OpenAIProviderOptionsInput>["model"] = (
+  modelID,
+  settings,
+) => configure(config(settings)).chat(modelID)
 export const responses = provider.responses
 export const responsesWebSocket = provider.responsesWebSocket
 export const chat = provider.chat

packages/ai/src/providers/openrouter.ts

@@ -107,7 +107,7 @@ export const configure = (input: ModelOptions = {}) => {
   const route = configuredRoute(input)
   return {
     id,
-    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    model: (modelID: string | ModelID) => route.model<OpenRouterProviderOptionsInput>({ id: modelID }),
     configure,
   }
 }

packages/ai/src/providers/xai.ts

@@ -5,12 +5,14 @@ import * as OpenAICompatibleProfiles from "./openai-compatible-profile"
 import * as OpenAICompatibleChat from "../protocols/openai-compatible-chat"
 import * as OpenAIResponses from "../protocols/openai-responses"
 import { XAIImages } from "../protocols/xai-images"
+import type { OpenAIProviderOptionsInput } from "./openai-options"
 
 export const id = ProviderID.make("xai")
 
-export type ModelOptions = RouteDefaultsInput &
+export type ModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
   ProviderAuthOption<"optional"> & {
     readonly baseURL?: string
+    readonly providerOptions?: OpenAIProviderOptionsInput
   }
 
 export type { XAIImageOptions } from "../protocols/xai-images"
@@ -42,8 +44,8 @@ const configuredChatRoute = (input: ModelOptions) => {
 export const configure = (input: ModelOptions = {}) => {
   const responsesRoute = configuredResponsesRoute(input)
   const chatRoute = configuredChatRoute(input)
-  const responses = (modelID: string | ModelID) => responsesRoute.model({ id: modelID })
-  const chat = (modelID: string | ModelID) => chatRoute.model({ id: modelID })
+  const responses = (modelID: string | ModelID) => responsesRoute.model<OpenAIProviderOptionsInput>({ id: modelID })
+  const chat = (modelID: string | ModelID) => chatRoute.model<OpenAIProviderOptionsInput>({ id: modelID })
   const image = (modelID: string | ModelID) =>
     XAIImages.model({
       id: modelID,

packages/ai/src/route/client.ts

@@ -10,7 +10,7 @@ import { WebSocketExecutor } from "./transport"
 import type { Protocol } from "./protocol"
 import { applyCachePolicy } from "../cache-policy"
 import * as ProviderShared from "../protocols/shared"
-import type { LLMError, PreparedRequestOf, ProtocolID, ProviderOptions } from "../schema"
+import type { LLMError, ProtocolID, ProviderOptions } from "../schema"
 import {
   GenerationOptions,
   HttpOptions,
@@ -20,7 +20,6 @@ import {
   ModelLimits,
   LLMError as LLMErrorClass,
   LLMEvent,
-  PreparedRequest,
   ProviderID,
   mergeGenerationOptions,
   mergeHttpOptions,
@@ -46,7 +45,7 @@ export interface Route<Body, Prepared = unknown> {
   readonly defaults: RouteDefaults
   readonly body: RouteBody<Body>
   readonly with: (patch: RoutePatch<Body, Prepared>) => Route<Body, Prepared>
-  readonly model: (input: RouteMappedModelInput) => Model
+  readonly model: <Options extends ProviderOptions = ProviderOptions>(input: RouteMappedModelInput) => Model<Options>
   readonly prepareTransport: (body: Body, request: LLMRequest) => Effect.Effect<Prepared, LLMError>
   readonly streamPrepared: (
     prepared: Prepared,
@@ -93,12 +92,12 @@ export interface RoutePatch<Body, Prepared> extends RouteDefaultsInput {
 
 type RouteMappedModelInput = RouteModelInput | RouteRoutedModelInput
 
-const makeRouteModel = (route: AnyRoute, mapped: RouteMappedModelInput) => {
+const makeRouteModel = <Options extends ProviderOptions = ProviderOptions>(route: AnyRoute, mapped: RouteMappedModelInput) => {
   const provider = route.provider ?? ("provider" in mapped ? mapped.provider : undefined)
   if (!provider) throw new Error(`Route.model(${route.id}) requires a provider`)
   if (!endpointBaseURL(route.endpoint))
     throw new Error(`Route.model(${route.id}) requires an endpoint baseURL — configure it on the route first`)
-  return Model.make({
+  return Model.make<Options>({
     ...mapped,
     provider,
     route,
@@ -142,17 +141,6 @@ export const httpOptions = (input: HttpOptionsInput | undefined) => {
 }
 
 export interface Interface {
-  /**
-   * Compile a request through protocol body construction, validation, and HTTP
-   * preparation without sending it. Returns the prepared request including the
-   * provider-native body.
-   *
-   * Pass a `Body` type argument to statically expose the route's body
-   * shape (e.g. `prepare<OpenAIChatBody>(...)`) — the runtime body is
-   * identical, so this is a type-level assertion the caller makes about which
-   * route the request will resolve to.
-   */
-  readonly prepare: <Body = unknown>(request: LLMRequest) => Effect.Effect<PreparedRequestOf<Body>, LLMError>
   readonly stream: StreamMethod
   readonly generate: GenerateMethod
 }
@@ -296,7 +284,8 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
           defaults: mergeRouteDefaults(route.defaults, defaults),
         })
       },
-      model: (input) => makeRouteModel(route, input),
+      model: <Options extends ProviderOptions = ProviderOptions>(input: RouteMappedModelInput) =>
+        makeRouteModel<Options>(route, input),
       prepareTransport: (body, request) =>
         routeInput.transport.prepare({
           body,
@@ -370,9 +359,6 @@ export function make<Body, Prepared, Frame, Event, State>(
   })
 }
 
-// `compile` is the important boundary: it turns a common `LLMRequest` into a
-// validated provider body plus transport-private prepared data, but does not
-// execute transport.
 const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest) {
   const resolved = applyCachePolicy(resolveRequestOptions(request))
   const route = resolved.model.route
@@ -390,17 +376,17 @@ const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest) {
   }
 })
 
-const prepareWith = Effect.fn("LLMClient.prepare")(function* (request: LLMRequest) {
+/** @internal Test-only projection of the execution compiler; not exported from package barrels. */
+export const compileRequest = Effect.fn("LLM.compileRequest")(function* (request: LLMRequest) {
   const compiled = yield* compile(request)
-
-  return new PreparedRequest({
+  return {
     id: compiled.request.id ?? "request",
     route: compiled.route.id,
     protocol: compiled.route.protocol,
     model: compiled.request.model,
     body: compiled.body,
     metadata: { transport: compiled.route.transport.id },
-  })
+  }
 })
 
 const streamRequestWith = (runtime: TransportRuntime) => (request: LLMRequest) =>
@@ -422,9 +408,6 @@ const generateWith = (stream: Interface["stream"]) =>
     )
   })
 
-export const prepare = <Body = unknown>(request: LLMRequest) =>
-  prepareWith(request) as Effect.Effect<PreparedRequestOf<Body>, LLMError>
-
 export function stream(request: LLMRequest): Stream.Stream<LLMEvent, LLMError> {
   return Stream.unwrap(
     Effect.gen(function* () {
@@ -453,7 +436,7 @@ export const layer: Layer.Layer<Service, never, RequestExecutor.Service> = Layer
       http: yield* RequestExecutor.Service,
       webSocket: Option.getOrUndefined(yield* Effect.serviceOption(WebSocketExecutor.Service)),
     })
-    return Service.of({ prepare: prepareWith as Interface["prepare"], stream, generate: generateWith(stream) })
+    return Service.of({ stream, generate: generateWith(stream) })
   }),
 )
 
@@ -462,7 +445,6 @@ export const Route = { make } as const
 export const LLMClient = {
   Service,
   layer,
-  prepare,
   stream,
   generate,
 } as const

packages/ai/src/schema/events.ts

@@ -1,6 +1,5 @@
 import { Schema } from "effect"
-import { ContentBlockID, FinishReason, ProtocolID, ProviderMetadata, RouteID, ToolCallID } from "./ids"
-import { ModelSchema } from "./options"
+import { ContentBlockID, FinishReason, ProviderMetadata, ToolCallID } from "./ids"
 import { Message, ToolCallPart, ToolOutput, ToolResultPart, ToolResultValue, type ContentPart } from "./messages"
 import { ProviderFailureClassification } from "./errors"
 
@@ -314,29 +313,6 @@ export const LLMEvent = Object.assign(llmEventTagged, {
 })
 export type LLMEvent = Schema.Schema.Type<typeof llmEventTagged>
 
-export class PreparedRequest extends Schema.Class<PreparedRequest>("LLM.PreparedRequest")({
-  id: Schema.String,
-  route: RouteID,
-  protocol: ProtocolID,
-  model: ModelSchema,
-  body: Schema.Unknown,
-  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
-}) {}
-
-/**
- * A `PreparedRequest` whose `body` is typed as `Body`. Use with the generic
- * on `LLMClient.prepare<Body>(...)` when the caller knows which route their
- * request will resolve to and wants its native shape statically exposed
- * (debug UIs, request previews, plan rendering).
- *
- * The runtime body is identical — the route still emits `body: unknown` — so
- * this is a type-level assertion the caller makes about what they expect to
- * find. The prepare runtime does not validate the assertion.
- */
-export type PreparedRequestOf<Body> = Omit<PreparedRequest, "body"> & {
-  readonly body: Body
-}
-
 const responseText = (events: ReadonlyArray<LLMEvent>) =>
   events
     .filter(LLMEvent.is.textDelta)

packages/ai/src/schema/options.ts

@@ -178,7 +178,8 @@ export namespace ModelCompatibility {
   export const make = (input: Input) => (input instanceof ModelCompatibility ? input : new ModelCompatibility(input))
 }
 
-export class Model {
+export class Model<Options extends ProviderOptions = ProviderOptions> {
+  declare protected readonly _ProviderOptions: Options
   readonly id: ModelID
   readonly provider: ProviderID
   readonly route: AnyRoute
@@ -193,8 +194,8 @@ export class Model {
     this.compatibility = input.compatibility
   }
 
-  static make(input: Model.Input) {
-    return new Model({
+  static make<Options extends ProviderOptions = ProviderOptions>(input: Model.Input) {
+    return new Model<Options>({
       id: ModelID.make(input.id),
       provider: ProviderID.make(input.provider),
       route: input.route,
@@ -203,7 +204,7 @@ export class Model {
     })
   }
 
-  static input(model: Model): Model.ConstructorInput {
+  static input<Options extends ProviderOptions>(model: Model<Options>): Model.ConstructorInput {
     return {
       id: model.id,
       provider: model.provider,
@@ -213,9 +214,9 @@ export class Model {
     }
   }
 
-  static update(model: Model, patch: Partial<Model.Input>) {
+  static update<Options extends ProviderOptions>(model: Model<Options>, patch: Partial<Model.Input>) {
     if (Object.keys(patch).length === 0) return model
-    return Model.make({
+    return Model.make<Options>({
       ...Model.input(model),
       ...patch,
     })
@@ -241,6 +242,8 @@ export namespace Model {
 
 export type ModelInput = Model.Input
 
+export type ModelProviderOptions<SelectedModel> = SelectedModel extends Model<infer Options> ? Options : never
+
 export const ModelSchema = Schema.declare((value): value is Model => value instanceof Model, { expected: "LLM.Model" })
 
 export class CacheHint extends Schema.Class<CacheHint>("LLM.CacheHint")({

packages/ai/src/testing.ts

@@ -99,7 +99,6 @@ export const layer = (options: LayerOptions = {}) =>
         )
       }) as LLMClientShape["stream"]
       const client = LLMClient.Service.of({
-        prepare: () => Effect.die("TestLLM does not prepare provider-native requests"),
         stream,
         generate: (request) =>
           stream(request).pipe(

packages/ai/test/adapter.test.ts

@@ -2,6 +2,7 @@ import { describe, expect } from "bun:test"
 import { Effect, Schema, Stream } from "effect"
 import { LLM, LLMRequest, LLMResponse } from "../src"
 import { Route, Endpoint, LLMClient, Protocol, type FramingDef } from "../src/route"
+import { compileRequest } from "../src/route/client"
 import { Model } from "../src/schema"
 import { testEffect } from "./lib/effect"
 import { dynamicResponse } from "./lib/http"
@@ -139,8 +140,7 @@ describe("llm route", () => {
 
   it.effect("selects routes by model route value", () =>
     Effect.gen(function* () {
-      const llm = yield* LLMClient.Service
-      const prepared = yield* llm.prepare(
+      const prepared = yield* compileRequest(
         LLMRequest.update(request, { model: updateModel(request.model, { route: configuredGemini }) }),
       )
 
@@ -173,7 +173,7 @@ describe("llm route", () => {
         framing: fakeFraming,
       })
 
-      const prepared = yield* (yield* LLMClient.Service).prepare(
+      const prepared = yield* compileRequest(
         LLMRequest.update(request, { model: updateModel(request.model, { route: duplicate }) }),
       )
 

packages/ai/test/cache-policy.test.ts

@@ -1,7 +1,8 @@
 import { describe, expect, test } from "bun:test"
 import { Effect } from "effect"
 import { CacheHint, LLM, Message } from "../src"
-import { Auth, LLMClient } from "../src/route"
+import { Auth } from "../src/route"
+import { compileRequest } from "../src/route/client"
 import { AmazonBedrock } from "../src/providers"
 import * as AnthropicMessages from "../src/protocols/anthropic-messages"
 import * as Gemini from "../src/protocols/gemini"
@@ -31,7 +32,7 @@ const geminiModel = Gemini.route
 describe("applyCachePolicy", () => {
   it.effect("undefined cache resolves to 'auto' (the recommended default)", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: "You are concise.",
@@ -50,7 +51,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'auto' marks the last tool, first and last system parts, and final message boundary on Anthropic", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: [
@@ -87,7 +88,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'auto' is a no-op on OpenAI (implicit caching protocol)", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: openaiModel,
           system: "Sys",
@@ -106,7 +107,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'auto' is a no-op on Gemini (out-of-band caching protocol)", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: geminiModel,
           system: "Sys",
@@ -123,7 +124,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'auto' on Bedrock emits cachePoint markers in the right places", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: bedrockModel,
           system: [
@@ -157,7 +158,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'none' disables auto placement even when manual hints exist", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: "Sys",
@@ -176,7 +177,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("granular object form: tools-only marks just tools", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: "Sys",
@@ -195,7 +196,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("auto policy preserves manual CacheHints on other parts", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: [
@@ -241,7 +242,7 @@ describe("applyCachePolicy", () => {
       expect("cache" in tail ? tail.cache : undefined).toBeUndefined()
       expect(applyCachePolicy(applied)).toBe(applied)
 
-      const prepared = yield* LLMClient.prepare(request)
+      const prepared = yield* compileRequest(request)
 
       const body = prepared.body as {
         tools: Array<{ cache_control?: unknown }>
@@ -261,7 +262,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("ttlSeconds in the policy flows through to wire markers", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           system: "Sys",
@@ -278,7 +279,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("messages: { tail: 2 } marks the last 2 message boundaries", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           messages: [Message.user("u1"), Message.assistant("a1"), Message.user("u2"), Message.assistant("a2")],
@@ -296,7 +297,7 @@ describe("applyCachePolicy", () => {
 
   it.effect("'latest-assistant' marks the last assistant message", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: anthropicModel,
           messages: [Message.user("u1"), Message.assistant("a1"), Message.user("u2")],

packages/ai/test/compile.test.ts

@@ -4,6 +4,7 @@ import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, mergeProviderOptions } from "../src"
 import { AnthropicMessages, OpenAIChat } from "../src/protocols"
 import { Auth, LLMClient } from "../src/route"
+import { compileRequest } from "../src/route/client"
 import { it } from "./lib/effect"
 import { dynamicResponse } from "./lib/http"
 import { deltaChunk } from "./lib/openai-chunks"
@@ -44,7 +45,7 @@ describe("request option precedence", () => {
     })
   })
 
-  it.effect("prepares bodies with route defaults, model defaults, and call options in order", () =>
+  it.effect("compiles bodies with route defaults, model defaults, and call options in order", () =>
     Effect.gen(function* () {
       const route = OpenAIChat.route.with({
         endpoint: { baseURL: "https://api.openai.test/v1/" },
@@ -59,7 +60,7 @@ describe("request option precedence", () => {
           providerOptions: { openai: { reasoningEffort: "medium" } },
         },
       })
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           prompt: "Say hello.",
@@ -141,7 +142,7 @@ describe("request option precedence", () => {
       const model = OpenAIChat.route
         .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("test") })
         .model({ id: "gpt-4o-mini" })
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model,
           prompt: "Say hello.",
@@ -164,10 +165,8 @@ describe("request option precedence", () => {
         limits: { output: 128 },
       })
       const model = route.model({ id: "claude-sonnet-4-5", defaults: { limits: { output: 64 } } })
-      const withoutMaxTokens = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
-        LLM.request({ model, prompt: "Say hello.", cache: "none" }),
-      )
-      const withMaxTokens = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const withoutMaxTokens = yield* compileRequest(LLM.request({ model, prompt: "Say hello.", cache: "none" }))
+      const withMaxTokens = yield* compileRequest(
         LLM.request({ model, prompt: "Say hello.", cache: "none", generation: { maxTokens: 32 } }),
       )
 

packages/ai/test/llm-option-types.types.ts

@@ -0,0 +1,47 @@
+import { Schema } from "effect"
+import { LLM, type Model, type ModelProviderOptions, type ProviderOptions } from "../src"
+import { OpenAIChat } from "../src/protocols"
+
+interface ExampleOptions {
+  readonly [key: string]: unknown
+  readonly mode?: "fast" | "thorough"
+}
+
+type ExampleProviderOptions = ProviderOptions & {
+  readonly example?: ExampleOptions
+}
+
+const model = OpenAIChat.route
+  .with({ endpoint: { baseURL: "https://example.com/v1" } })
+  .model<ExampleProviderOptions>({ id: "example" })
+
+LLM.request({ model, prompt: "Hello", providerOptions: { example: { mode: "fast" } } })
+LLM.request({ model, prompt: "Hello", providerOptions: { future: { option: true } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Known provider options preserve their value types.
+  providerOptions: { example: { mode: "slow" } },
+})
+
+LLM.generateObject({
+  model,
+  prompt: "Hello",
+  schema: Schema.Struct({ answer: Schema.String }),
+  providerOptions: { example: { mode: "thorough" } },
+})
+
+LLM.generateObject({
+  model,
+  prompt: "Hello",
+  jsonSchema: { type: "object" },
+  // @ts-expect-error Dynamic object generation uses the selected model's provider options.
+  providerOptions: { example: { mode: false } },
+})
+
+declare const generic: Model
+LLM.request({ model: generic, prompt: "Hello", providerOptions: { arbitrary: { option: true } } })
+
+const options: ModelProviderOptions<typeof model> = { example: { mode: "fast" } }
+void options

packages/ai/test/provider-options/anthropic-compatible.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { AnthropicCompatible } from "../../src/providers"
+
+const model = AnthropicCompatible.configure({ baseURL: "https://example.com" }).model("claude")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { anthropic: { effort: "high" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Anthropic effort must be a string.
+  providerOptions: { anthropic: { effort: 1 } },
+})

packages/ai/test/provider-options/anthropic.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { Anthropic } from "../../src/providers"
+
+const model = Anthropic.provider.model("claude-sonnet-4-5")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { anthropic: { thinking: { type: "adaptive" } } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Anthropic thinking modes are a fixed union.
+  providerOptions: { anthropic: { thinking: { type: "automatic" } } },
+})

packages/ai/test/provider-options/azure.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { Azure } from "../../src/providers"
+
+const model = Azure.configure({ resourceName: "example" }).responses("deployment")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { store: false } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Azure OpenAI store must be boolean.
+  providerOptions: { openai: { store: "false" } },
+})

packages/ai/test/provider-options/cloudflare.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { CloudflareWorkersAI } from "../../src/providers"
+
+const model = CloudflareWorkersAI.configure({ accountId: "account", apiKey: "test" }).model("model")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { promptCacheKey: "cache" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Cloudflare's OpenAI-compatible prompt cache key must be a string.
+  providerOptions: { openai: { promptCacheKey: 1 } },
+})

packages/ai/test/provider-options/github-copilot.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { GitHubCopilot } from "../../src/providers"
+
+const model = GitHubCopilot.configure({ baseURL: "https://example.com" }).model("gpt-5")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { reasoningSummary: "auto" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Copilot reasoning summaries use the OpenAI union.
+  providerOptions: { openai: { reasoningSummary: "full" } },
+})

packages/ai/test/provider-options/google-vertex-chat.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { GoogleVertexChat } from "../../src/providers"
+
+const model = GoogleVertexChat.configure({ accessToken: "test", project: "project" }).model("gemini")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { serviceTier: "priority" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Vertex OpenAI-compatible service tiers use the OpenAI union.
+  providerOptions: { openai: { serviceTier: "premium" } },
+})

packages/ai/test/provider-options/google-vertex-messages.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { GoogleVertexMessages } from "../../src/providers"
+
+const model = GoogleVertexMessages.configure({ accessToken: "test", project: "project" }).model("claude")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { anthropic: { effort: "medium" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Vertex Anthropic effort must be a string.
+  providerOptions: { anthropic: { effort: false } },
+})

packages/ai/test/provider-options/google-vertex-responses.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { GoogleVertexResponses } from "../../src/providers"
+
+const model = GoogleVertexResponses.configure({ accessToken: "test", project: "project" }).model("gemini")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openresponses: { textVerbosity: "high" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Vertex Responses verbosity uses the Open Responses union.
+  providerOptions: { openresponses: { textVerbosity: "verbose" } },
+})

packages/ai/test/provider-options/google-vertex.types.ts

@@ -0,0 +1,17 @@
+import { LLM } from "../../src"
+import { GoogleVertex } from "../../src/providers"
+
+const model = GoogleVertex.provider.configure({ apiKey: "test" }).model("gemini-2.5-pro")
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  providerOptions: { gemini: { thinkingConfig: { includeThoughts: true } } },
+})
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Vertex Gemini includeThoughts must be boolean.
+  providerOptions: { gemini: { thinkingConfig: { includeThoughts: "yes" } } },
+})

packages/ai/test/provider-options/google.types.ts

@@ -0,0 +1,17 @@
+import { LLM } from "../../src"
+import { Google } from "../../src/providers"
+
+const model = Google.provider.model("gemini-2.5-pro")
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  providerOptions: { gemini: { thinkingConfig: { thinkingBudget: 1024 } } },
+})
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Gemini thinking budgets must be numeric.
+  providerOptions: { gemini: { thinkingConfig: { thinkingBudget: "large" } } },
+})

packages/ai/test/provider-options/openai-compatible-responses.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { OpenAICompatibleResponses } from "../../src/providers"
+
+const model = OpenAICompatibleResponses.configure({ baseURL: "https://example.com" }).model("model")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openresponses: { reasoningSummary: "detailed" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Open Responses reasoning summaries use a fixed union.
+  providerOptions: { openresponses: { reasoningSummary: "full" } },
+})

packages/ai/test/provider-options/openai-compatible.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { OpenAICompatible } from "../../src/providers"
+
+const model = OpenAICompatible.deepseek.model("deepseek-chat")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { store: false } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error OpenAI-compatible store must be boolean.
+  providerOptions: { openai: { store: "false" } },
+})

packages/ai/test/provider-options/openai.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { OpenAI } from "../../src/providers"
+
+const model = OpenAI.responses("gpt-5")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { reasoningEffort: "high" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error OpenAI reasoning effort must be a string.
+  providerOptions: { openai: { reasoningEffort: 1 } },
+})

packages/ai/test/provider-options/openrouter.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { OpenRouter } from "../../src/providers"
+
+const model = OpenRouter.provider.model("anthropic/claude-sonnet-4.5")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openrouter: { usage: true } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error OpenRouter usage must be boolean or an option record.
+  providerOptions: { openrouter: { usage: "yes" } },
+})

packages/ai/test/provider-options/xai.types.ts

@@ -0,0 +1,13 @@
+import { LLM } from "../../src"
+import { XAI } from "../../src/providers"
+
+const model = XAI.provider.model("grok-4")
+
+LLM.request({ model, prompt: "Hello", providerOptions: { openai: { reasoningEffort: "high" } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error xAI's OpenAI-compatible reasoning effort must be a string.
+  providerOptions: { openai: { reasoningEffort: true } },
+})

packages/ai/test/provider/anthropic-messages.test.ts

@@ -3,6 +3,7 @@ import { Effect } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { CacheHint, LLM, LLMError, LLMRequest, Message, ToolCallPart, ToolDefinition, Usage } from "../../src"
 import { Auth, LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import * as AnthropicMessages from "../../src/protocols/anthropic-messages"
 import { continuationRequest, nativeAnthropicMessagesContinuation } from "../continuation-scenarios"
 import { it } from "../lib/effect"
@@ -44,7 +45,7 @@ const expectToolResult = (body: AnthropicMessages.AnthropicMessagesBody): Anthro
 describe("Anthropic Messages route", () => {
   it.effect("prepares Anthropic Messages target", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(request)
+      const prepared = yield* compileRequest(request)
 
       expect(prepared.body).toEqual({
         model: "claude-sonnet-4-5",
@@ -59,7 +60,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("lowers adaptive thinking settings with effort", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLMRequest.update(request, {
           providerOptions: {
             anthropic: { thinking: { type: "adaptive", display: "summarized" }, effort: "low" },
@@ -76,17 +77,17 @@ describe("Anthropic Messages route", () => {
 
   it.effect("normalizes enabled and disabled thinking settings", () =>
     Effect.gen(function* () {
-      const enabled = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const enabled = yield* compileRequest(
         LLMRequest.update(request, {
           providerOptions: { anthropic: { thinking: { type: "enabled", budgetTokens: 1_024 } } },
         }),
       )
-      const legacy = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const legacy = yield* compileRequest(
         LLMRequest.update(request, {
           providerOptions: { anthropic: { thinking: { type: "enabled", budget_tokens: 2_048 } } },
         }),
       )
-      const disabled = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const disabled = yield* compileRequest(
         LLMRequest.update(request, {
           providerOptions: { anthropic: { thinking: { type: "disabled" } } },
         }),
@@ -100,7 +101,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("rejects enabled thinking without a budget", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLMRequest.update(request, {
           providerOptions: { anthropic: { thinking: { type: "enabled" } } },
         }),
@@ -112,7 +113,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("lowers chronological system updates natively for Claude Opus 4.8 with cache hints", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: opus48,
           messages: [
@@ -137,7 +138,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("lowers chronological system updates to wrapped user text for unsupported Anthropic models", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -164,7 +165,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("rejects non-text chronological system update content before send", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model: opus48,
           messages: [
@@ -181,7 +182,7 @@ describe("Anthropic Messages route", () => {
   it.effect("falls back for unsupported native chronological system update placement", () =>
     Effect.gen(function* () {
       expect(
-        (yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+        (yield* compileRequest(
           LLM.request({
             model: opus48,
             messages: [Message.assistant("Plain."), Message.system("After plain assistant.")],
@@ -196,12 +197,11 @@ describe("Anthropic Messages route", () => {
         },
       ])
       expect(
-        (yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
-          LLM.request({ model: opus48, messages: [Message.system("First.")], cache: "none" }),
-        )).body.messages,
+        (yield* compileRequest(LLM.request({ model: opus48, messages: [Message.system("First.")], cache: "none" })))
+          .body.messages,
       ).toEqual([{ role: "user", content: [{ type: "text", text: "<system-update>\nFirst.\n</system-update>" }] }])
       expect(
-        (yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+        (yield* compileRequest(
           LLM.request({
             model: opus48,
             messages: [Message.user("Before."), Message.system("One."), Message.system("Two.")],
@@ -223,7 +223,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("rejects a system update between a local tool call and its result", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model: opus48,
           messages: [
@@ -242,7 +242,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("prepares tool call and tool result messages", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result",
           model,
@@ -273,7 +273,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("keeps tools and sends tool_choice none", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_choice_none",
           model,
@@ -303,7 +303,7 @@ describe("Anthropic Messages route", () => {
   // not JSON-stringified into `tool_result.content`.
   it.effect("lowers media tool-result content as structured blocks", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_image",
           model,
@@ -335,7 +335,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("lowers single-image tool-result content as a structured image block", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_image_only",
           model,
@@ -360,7 +360,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("rejects unsupported media in tool-result content with a clear error", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_unsupported_media",
           model,
@@ -384,7 +384,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("prepares the composed native continuation request", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         continuationRequest({
           id: "req_native_continuation_anthropic",
           model,
@@ -428,7 +428,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("lowers preserved Anthropic reasoning signature metadata", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -447,7 +447,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("round-trips redacted thinking as redacted_thinking blocks", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -628,9 +628,7 @@ describe("Anthropic Messages route", () => {
         { type: "reasoning", text: "", providerMetadata: { anthropic: { signature: "sig_1" } } },
       ])
 
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
-        LLM.request({ model, messages: [response.message], cache: "none" }),
-      )
+      const prepared = yield* compileRequest(LLM.request({ model, messages: [response.message], cache: "none" }))
       expect(prepared.body.messages).toEqual([
         { role: "assistant", content: [{ type: "thinking", thinking: "", signature: "sig_1" }] },
       ])
@@ -773,7 +771,7 @@ describe("Anthropic Messages route", () => {
           ),
         ),
       )
-      const prepared = yield* LLMClient.prepare<AnthropicMessages.AnthropicMessagesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -1133,7 +1131,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("round-trips provider-executed assistant content into server tool blocks", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_round_trip",
           model,
@@ -1184,7 +1182,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("rejects round-trip for unknown server tool names", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_unknown_server_tool",
           model,
@@ -1261,7 +1259,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("maps ttlSeconds >= 3600 to cache_control ttl: '1h'", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           system: { type: "text", text: "system", cache: new CacheHint({ type: "ephemeral", ttlSeconds: 3600 }) },
@@ -1277,7 +1275,7 @@ describe("Anthropic Messages route", () => {
 
   it.effect("emits cache_control on tool definitions and tool-result blocks", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           tools: [
@@ -1318,7 +1316,7 @@ describe("Anthropic Messages route", () => {
   it.effect("drops cache_control breakpoints past the 4-per-request cap", () =>
     Effect.gen(function* () {
       const hint = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           system: [
@@ -1344,7 +1342,7 @@ describe("Anthropic Messages route", () => {
   it.effect("spends breakpoint budget on tools before system before messages", () =>
     Effect.gen(function* () {
       const hint = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           tools: [

packages/ai/test/provider/bedrock-converse.test.ts

@@ -13,6 +13,7 @@ import {
   ToolDefinition,
 } from "../../src"
 import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { AmazonBedrock } from "../../src/providers"
 import * as BedrockConverse from "../../src/protocols/bedrock-converse"
 import { it } from "../lib/effect"
@@ -101,7 +102,7 @@ const baseRequest = LLM.request({
 describe("Bedrock Converse route", () => {
   it.effect("prepares Converse target with system, inference config, and messages", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(baseRequest)
+      const prepared = yield* compileRequest(baseRequest)
 
       expect(prepared.body).toEqual({
         modelId: "anthropic.claude-3-5-sonnet-20240620-v1:0",
@@ -114,7 +115,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("passes topK through additionalModelRequestFields as top_k", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLMRequest.update(baseRequest, {
           generation: GenerationOptions.make({ maxTokens: 64, temperature: 0, topK: 40 }),
         }),
@@ -129,14 +130,14 @@ describe("Bedrock Converse route", () => {
 
   it.effect("omits additionalModelRequestFields when topK is unset", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(baseRequest)
+      const prepared = yield* compileRequest(baseRequest)
       expect(prepared.body.additionalModelRequestFields).toBeUndefined()
     }),
   )
 
   it.effect("lowers chronological system updates to wrapped user text in order", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [Message.user("Before."), Message.system("Update."), Message.assistant("After.")],
@@ -153,7 +154,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("prepares tool config with toolSpec and toolChoice", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLMRequest.update(baseRequest, {
           tools: [
             ToolDefinition.make({
@@ -187,7 +188,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("keeps tools and omits the unsupported choice when tool choice is none", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLMRequest.update(baseRequest, {
           tools: [
             ToolDefinition.make({
@@ -217,7 +218,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("lowers assistant tool-call + tool-result message history", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_history",
           model,
@@ -256,7 +257,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("lowers image content in tool-result messages", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_image",
           model,
@@ -491,7 +492,7 @@ describe("Bedrock Converse route", () => {
         providerMetadata: { bedrock: { signature: "sig_1" } },
       })
 
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -546,9 +547,7 @@ describe("Bedrock Converse route", () => {
         },
       ])
 
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
-        LLM.request({ model, messages: [response.message], cache: "none" }),
-      )
+      const prepared = yield* compileRequest(LLM.request({ model, messages: [response.message], cache: "none" }))
       expect(prepared.body.messages).toEqual([
         {
           role: "assistant",
@@ -639,7 +638,7 @@ describe("Bedrock Converse route", () => {
         text: "",
         providerMetadata: { bedrock: { redactedData } },
       })
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -754,7 +753,7 @@ describe("Bedrock Converse route", () => {
           secretAccessKey: "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY",
         },
       }).model("anthropic.claude-3-5-sonnet-20240620-v1:0")
-      const prepared = yield* LLMClient.prepare(LLMRequest.update(baseRequest, { model: signed }))
+      const prepared = yield* compileRequest(LLMRequest.update(baseRequest, { model: signed }))
 
       expect(prepared.route).toBe("bedrock-converse")
       expect(prepared.model).toBe(signed)
@@ -764,7 +763,7 @@ describe("Bedrock Converse route", () => {
   it.effect("emits cachePoint markers after system, user-text, and assistant-text with cache hints", () =>
     Effect.gen(function* () {
       const cache = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_cache",
           model,
@@ -796,7 +795,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("does not emit cachePoint when no cache hint is set", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(baseRequest)
+      const prepared = yield* compileRequest(baseRequest)
       expect(prepared.body).toMatchObject({
         system: [{ text: "You are concise." }],
         messages: [{ role: "user", content: [{ text: "Say hello." }] }],
@@ -806,7 +805,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("lowers image media into Bedrock image blocks", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_image",
           model,
@@ -843,7 +842,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("base64-encodes Uint8Array image bytes", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_image_bytes",
           model,
@@ -865,7 +864,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("lowers document media into Bedrock document blocks with format and name", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_doc",
           model,
@@ -897,7 +896,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("requires names for document media", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model,
           messages: [Message.user({ type: "media", mediaType: "application/pdf", data: "UERGREFUQQ==" })],
@@ -910,7 +909,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("passes named document-only messages through for provider validation", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           cache: "none",
@@ -936,7 +935,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("lowers document media in tool results", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<BedrockConverse.BedrockConverseBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           cache: "none",
@@ -988,7 +987,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("rejects unsupported image media types", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_bad_image",
           model,
@@ -1002,7 +1001,7 @@ describe("Bedrock Converse route", () => {
 
   it.effect("rejects unsupported document media types", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_bad_doc",
           model,
@@ -1017,7 +1016,7 @@ describe("Bedrock Converse route", () => {
   it.effect("maps ttlSeconds >= 3600 to cachePoint ttl: '1h'", () =>
     Effect.gen(function* () {
       const cache = new CacheHint({ type: "ephemeral", ttlSeconds: 3600 })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           system: [{ type: "text", text: "system", cache }],
@@ -1034,7 +1033,7 @@ describe("Bedrock Converse route", () => {
   it.effect("appends cachePoint after marked tool definitions and tool-result blocks", () =>
     Effect.gen(function* () {
       const cache = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           tools: [{ name: "lookup", description: "lookup", inputSchema: { type: "object", properties: {} }, cache }],
@@ -1066,7 +1065,7 @@ describe("Bedrock Converse route", () => {
   it.effect("drops cachePoint markers past the 4-per-request cap", () =>
     Effect.gen(function* () {
       const cache = new CacheHint({ type: "ephemeral" })
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           system: [

packages/ai/test/provider/cloudflare.test.ts

@@ -3,7 +3,7 @@ import { ConfigProvider, Effect, Schema } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, LLMEvent } from "../../src"
 import { CloudflareAIGateway, CloudflareWorkersAI } from "../../src/providers/cloudflare"
-import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { it } from "../lib/effect"
 import { dynamicResponse } from "../lib/http"
 import { sseEvents } from "../lib/sse"
@@ -34,7 +34,7 @@ describe("Cloudflare", () => {
       })
       expect(model.route.endpoint.baseURL).toBe("https://gateway.ai.cloudflare.com/v1/test-account/test-gateway/compat")
 
-      const prepared = yield* LLMClient.prepare(LLM.request({ model, prompt: "Say hello." }))
+      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Say hello." }))
 
       expect(prepared.route).toBe("cloudflare-ai-gateway")
       expect(prepared.body).toMatchObject({
@@ -129,7 +129,7 @@ describe("Cloudflare", () => {
         openai: { reasoningField: "reasoning", reasoningDetails: merged },
       })
 
-      const replay = yield* LLMClient.prepare(LLM.request({ model, messages: [response.message] }))
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([
         { role: "assistant", content: "Hello", reasoning: "Thinking", reasoning_details: merged },
       ])
@@ -180,7 +180,7 @@ describe("Cloudflare", () => {
 
   it.effect("allows a fully configured baseURL override", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: CloudflareAIGateway.configure({
             baseURL: "https://gateway.proxy.test/v1/custom/compat",
@@ -208,7 +208,7 @@ describe("Cloudflare", () => {
       })
       expect(model.route.endpoint.baseURL).toBe("https://api.cloudflare.com/client/v4/accounts/test-account/ai/v1")
 
-      const prepared = yield* LLMClient.prepare(LLM.request({ model, prompt: "Say hello." }))
+      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Say hello." }))
 
       expect(prepared.route).toBe("cloudflare-workers-ai")
       expect(prepared.body).toMatchObject({

packages/ai/test/provider/gemini.test.ts

@@ -2,6 +2,7 @@ import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { LLM, LLMError, LLMRequest, Message, ToolCallPart, ToolDefinition, Usage } from "../../src"
 import { Auth, LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import * as Gemini from "../../src/protocols/gemini"
 import { ProviderShared } from "../../src/protocols/shared"
 import { it } from "../lib/effect"
@@ -26,7 +27,7 @@ const request = LLM.request({
 describe("Gemini route", () => {
   it.effect("prepares Gemini target", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(request)
+      const prepared = yield* compileRequest(request)
 
       expect(prepared.body).toEqual({
         contents: [{ role: "user", parts: [{ text: "Say hello." }] }],
@@ -38,12 +39,12 @@ describe("Gemini route", () => {
 
   it.effect("normalizes Gemini thinking options", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const prepared = yield* compileRequest(
         LLMRequest.update(request, {
           providerOptions: { gemini: { thinkingConfig: { thinkingBudget: 0, includeThoughts: false } } },
         }),
       )
-      const filtered = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const filtered = yield* compileRequest(
         LLMRequest.update(request, {
           providerOptions: { gemini: { thinkingConfig: { thinkingBudget: "invalid", includeThoughts: false } } },
         }),
@@ -59,7 +60,7 @@ describe("Gemini route", () => {
 
   it.effect("lowers chronological system updates to wrapped user text in order", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [Message.user("Before."), Message.system("Update."), Message.assistant("After.")],
@@ -75,7 +76,7 @@ describe("Gemini route", () => {
 
   it.effect("prepares multimodal user input and tool history", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result",
           model,
@@ -143,7 +144,7 @@ describe("Gemini route", () => {
 
   it.effect("continues media tool results as inline model input without base64 text", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -188,7 +189,7 @@ describe("Gemini route", () => {
 
   it.effect("strips matching data URLs to raw base64 inlineData", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -229,7 +230,7 @@ describe("Gemini route", () => {
   ] as const)
     it.effect(`rejects ${name}`, () =>
       Effect.gen(function* () {
-        const error = yield* LLMClient.prepare(
+        const error = yield* compileRequest(
           LLM.request({ model, messages: [Message.user({ type: "media", ...media })] }),
         ).pipe(Effect.flip)
         expect(error.message).toMatch(/does not support|does not match|valid base64/)
@@ -238,7 +239,7 @@ describe("Gemini route", () => {
 
   it.effect("rejects oversized image input", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -256,7 +257,7 @@ describe("Gemini route", () => {
 
   it.effect("keeps tools and sends function calling mode NONE", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_choice_none",
           model,
@@ -276,7 +277,7 @@ describe("Gemini route", () => {
 
   it.effect("sanitizes integer enums, dangling required, untyped arrays, and scalar object keys", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_schema_patch",
           model,
@@ -457,7 +458,7 @@ describe("Gemini route", () => {
         response.events.findIndex((event) => event.type === "tool-call"),
       )
 
-      const prepared = yield* LLMClient.prepare<Gemini.GeminiBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -691,7 +692,7 @@ describe("Gemini route", () => {
 
   it.effect("rejects unsupported assistant media content", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_media",
           model,

packages/ai/test/provider/google-vertex.test.ts

@@ -4,6 +4,7 @@ import { HttpClientRequest } from "effect/unstable/http"
 import { LLM } from "../../src"
 import { GoogleVertex, GoogleVertexChat, GoogleVertexMessages, GoogleVertexResponses } from "../../src/providers"
 import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { it } from "../lib/effect"
 import { dynamicResponse } from "../lib/http"
 import { deltaChunk, finishChunk } from "../lib/openai-chunks"
@@ -182,7 +183,7 @@ describe("Google Vertex providers", () => {
 
   it.effect("protects the Vertex Messages API version from body overlays", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model: GoogleVertexMessages.configure({
             accessToken: "vertex-token",

packages/ai/test/provider/openai-chat-reasoning.recorded.test.ts

@@ -5,6 +5,7 @@ import { OpenAIChat } from "../../src/protocols/openai-chat"
 import * as OpenAICompatible from "../../src/providers/openai-compatible"
 import * as OpenRouter from "../../src/providers/openrouter"
 import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { recordedTests } from "../recorded-test"
 import { expectWeatherToolLoop, goldenWeatherToolLoopRequest, runWeatherToolLoop } from "../recorded-scenarios"
 
@@ -84,9 +85,7 @@ for (const item of cases) {
             ),
           ).toBe(true)
 
-          const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-            LLM.request({ model: item.model, messages: [response.message] }),
-          )
+          const replay = yield* compileRequest(LLM.request({ model: item.model, messages: [response.message] }))
           expect(replay.body.messages).toMatchObject([
             { role: "assistant", content: response.text, reasoning: response.reasoning },
           ])

packages/ai/test/provider/openai-chat.test.ts

@@ -18,6 +18,7 @@ import * as OpenAI from "../../src/providers/openai"
 import * as OpenAIChat from "../../src/protocols/openai-chat"
 import { ProviderShared } from "../../src/protocols/shared"
 import { Auth, LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { it } from "../lib/effect"
 import { dynamicResponse, fixedResponse, truncatedStream } from "../lib/http"
 import { deltaChunk, usageChunk } from "../lib/openai-chunks"
@@ -42,11 +43,7 @@ const request = LLM.request({
 describe("OpenAI Chat route", () => {
   it.effect("prepares OpenAI Chat payload", () =>
     Effect.gen(function* () {
-      // Pass the OpenAIChat payload type so `prepared.body` is statically
-      // typed to the route's native shape — the assertions below read field
-      // names without `unknown` casts.
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(request)
-      const _typed: { readonly model: string; readonly stream: true } = prepared.body
+      const prepared = yield* compileRequest(request)
 
       expect(prepared.body).toEqual({
         model: "gpt-4o-mini",
@@ -64,7 +61,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("lowers chronological system updates to escaped user wrappers in order", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -87,7 +84,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("replays canonical reasoning as OpenAI-compatible reasoning_content", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -105,7 +102,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("writes reasoning to a configured custom field on every assistant message", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: Model.update(model, { compatibility: { reasoningField: "vendor_reasoning" } }),
           messages: [
@@ -131,7 +128,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("rejects reasoning fields that conflict with assistant message fields", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model: Model.update(model, { compatibility: { reasoningField: "content" } }),
           messages: [Message.assistant([{ type: "reasoning", text: "thinking" }])],
@@ -144,7 +141,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("maps OpenAI provider options to Chat options", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenAI.configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).chat("gpt-4o-mini"),
           prompt: "think",
@@ -159,7 +156,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("passes through custom OpenAI-compatible reasoning effort strings", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           prompt: "think",
@@ -253,7 +250,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("prepares assistant tool-call and tool-result messages", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result",
           model,
@@ -291,7 +288,7 @@ describe("OpenAI Chat route", () => {
   it.effect("preserves structured tool errors for the model", () =>
     Effect.gen(function* () {
       const error = { error: { type: "unknown", message: "Tool execution interrupted" } }
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -311,7 +308,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("continues image tool results as vision input without base64 text", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -355,7 +352,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("orders parallel tool responses before one aggregated vision message", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -405,7 +402,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("aggregates consecutive tool images with a following system update", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -446,7 +443,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("appends system updates without replacing multipart user content", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -474,7 +471,7 @@ describe("OpenAI Chat route", () => {
   ] as const)
     it.effect(`rejects ${name}`, () =>
       Effect.gen(function* () {
-        const error = yield* LLMClient.prepare(
+        const error = yield* compileRequest(
           LLM.request({ model, messages: [Message.user({ type: "media", ...media })] }),
         ).pipe(Effect.flip)
         expect(error.message).toMatch(/does not support|does not match|valid base64/)
@@ -483,7 +480,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("rejects oversized image input", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -501,7 +498,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("prepares raw and data URL image media as vision input", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_media",
           model,
@@ -528,7 +525,7 @@ describe("OpenAI Chat route", () => {
 
   it.effect("lowers reasoning-only assistant history", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_reasoning",
           model,
@@ -619,9 +616,7 @@ describe("OpenAI Chat route", () => {
           openai: { reasoningField: field },
         })
 
-        const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-          LLM.request({ model, messages: [response.message] }),
-        )
+        const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
         expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", [field]: "thinking" }])
       }
     }),
@@ -647,9 +642,7 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningField: "vendor_reasoning" },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model: custom, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model: custom, messages: [response.message] }))
       expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", vendor_reasoning: "thinking" }])
     }),
   )
@@ -692,9 +685,7 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningField: "reasoning", reasoningDetails: details },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([
         {
           role: "assistant",
@@ -737,9 +728,7 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningDetails: details },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", reasoning_details: details }])
     }),
   )
@@ -764,9 +753,7 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningField: "reasoning", reasoningDetails: details },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([
         { role: "assistant", content: "Hello", reasoning: "thinking", reasoning_details: details },
       ])
@@ -839,9 +826,7 @@ describe("OpenAI Chat route", () => {
         openai: { reasoningDetails: [] },
       })
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([{ role: "assistant", content: "Hello", reasoning_details: [] }])
     }),
   )
@@ -889,9 +874,7 @@ describe("OpenAI Chat route", () => {
         response.events.findIndex(LLMEvent.is.textStart),
       )
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([
         { role: "assistant", content: "Hello", reasoning: "thinking", reasoning_details: merged },
       ])
@@ -918,9 +901,7 @@ describe("OpenAI Chat route", () => {
       expect(response.events.filter(LLMEvent.is.reasoningStart)).toHaveLength(1)
       expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
 
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(replay.body.messages).toEqual([{ role: "assistant", content: null, reasoning_details: details }])
     }),
   )
@@ -950,7 +931,7 @@ describe("OpenAI Chat route", () => {
     Effect.gen(function* () {
       const first = { type: "reasoning.text", text: "first", signature: "signed-0", index: 0 }
       const second = { type: "reasoning.text", text: "second", signature: "signed-1", index: 1 }
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const replay = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -979,7 +960,7 @@ describe("OpenAI Chat route", () => {
   it.effect("retains scalar replay for mixed structured reasoning parts", () =>
     Effect.gen(function* () {
       const detail = { type: "reasoning.encrypted", data: "opaque", index: 0 }
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const replay = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -1004,7 +985,7 @@ describe("OpenAI Chat route", () => {
   it.effect("replays native scalar reasoning alongside native details", () =>
     Effect.gen(function* () {
       const details = [{ type: "reasoning.encrypted", data: "opaque", index: 0 }]
-      const replay = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const replay = yield* compileRequest(
         LLM.request({
           model,
           messages: [

packages/ai/test/provider/openai-compatible-chat.test.ts

@@ -3,6 +3,7 @@ import { Effect, Schema } from "effect"
 import { HttpClientRequest } from "effect/unstable/http"
 import { LLM, LLMRequest, Message, ToolCallPart, ToolChoice, ToolDefinition } from "../../src"
 import { Auth, LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import * as OpenAICompatible from "../../src/providers/openai-compatible"
 import * as OpenAICompatibleChat from "../../src/protocols/openai-compatible-chat"
 import { it } from "../lib/effect"
@@ -52,7 +53,7 @@ const providerFamilies = [
 describe("OpenAI-compatible Chat route", () => {
   it.effect("prepares generic Chat target", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLMRequest.update(request, {
           tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
           toolChoice: ToolChoice.make({ type: "required" }),
@@ -127,7 +128,7 @@ describe("OpenAI-compatible Chat route", () => {
 
   it.effect("matches AI SDK compatible basic request body fixture", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(request)
+      const prepared = yield* compileRequest(request)
 
       expect(prepared.body).toEqual({
         model: "deepseek-chat",
@@ -145,7 +146,7 @@ describe("OpenAI-compatible Chat route", () => {
 
   it.effect("matches AI SDK compatible tool request body fixture", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_parity",
           model,

packages/ai/test/provider/openai-compatible-responses.test.ts

@@ -7,6 +7,7 @@ import { OpenResponses } from "../../src/protocols/open-responses"
 import { OpenAICompatibleResponses } from "../../src/protocols/openai-compatible-responses"
 import { OpenAIResponses } from "../../src/protocols/openai-responses"
 import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import { it } from "../lib/effect"
 import { fixedResponse } from "../lib/http"
 import { sseEvents } from "../lib/sse"
@@ -23,7 +24,7 @@ describe("Open Responses-compatible route", () => {
         baseURL: "https://responses.example.test/v1",
         provider: "example",
       }).model("example-model")
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           system: "You are concise.",
@@ -61,7 +62,7 @@ describe("Open Responses-compatible route", () => {
         apiKey: "test-key",
         baseURL: "https://responses.example.test/v1",
       }).model("example-model")
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({ model, prompt: "Draw.", tools: [OpenAI.imageGeneration()] }),
       ).pipe(Effect.flip)
 
@@ -76,7 +77,7 @@ describe("Open Responses-compatible route", () => {
         apiKey: "test-key",
         baseURL: "https://responses.example.test/v1",
       }).model("example-model")
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -102,7 +103,7 @@ describe("Open Responses-compatible route", () => {
         baseURL: "https://responses.example.test/v1",
         providerOptions: { openresponses: { reasoningEffort: "low", store: true } },
       }).model("example-model")
-      const prepared = yield* LLMClient.prepare(LLM.request({ model, prompt: "Think." }))
+      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Think." }))
 
       expect(prepared.body).toMatchObject({
         reasoning: { effort: "low" },

packages/ai/test/provider/openai-responses.test.ts

@@ -14,6 +14,7 @@ import {
   Usage,
 } from "../../src"
 import { Auth, LLMClient, RequestExecutor, WebSocketExecutor } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import * as Azure from "../../src/providers/azure"
 import * as OpenAI from "../../src/providers/openai"
 import * as XAI from "../../src/providers/xai"
@@ -56,7 +57,7 @@ const expectToolOutput = (body: OpenAIResponses.OpenAIResponsesBody): OpenAITool
 describe("OpenAI Responses route", () => {
   it.effect("prepares OpenAI Responses target", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(request)
+      const prepared = yield* compileRequest(request)
 
       expect(prepared.body).toEqual({
         model: "gpt-4.1-mini",
@@ -74,7 +75,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("lowers the hosted OpenAI image generation tool", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           prompt: "Show me a rooftop garden.",
@@ -92,7 +93,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("rejects invalid hosted image generation options locally", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           model,
           prompt: "Show me a rooftop garden.",
@@ -109,7 +110,7 @@ describe("OpenAI Responses route", () => {
     Effect.gen(function* () {
       const input = LLMRequest.update(request, { providerOptions: { openai: { serviceTier: "priority" } } })
       expect(input.providerOptions).toEqual({ openai: { serviceTier: "priority" } })
-      const prepared = yield* LLMClient.prepare(input)
+      const prepared = yield* compileRequest(input)
 
       expect(prepared.body).toMatchObject({ service_tier: "priority" })
       expect(prepared.body).not.toHaveProperty("serviceTier")
@@ -118,7 +119,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("passes through custom OpenAI reasoning effort strings", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLMRequest.update(request, { providerOptions: { openai: { reasoningEffort: "experimental" } } }),
       )
 
@@ -128,7 +129,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("omits unsupported semantic service tiers", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLMRequest.update(request, { providerOptions: { openai: { serviceTier: "unsupported" } } }),
       )
 
@@ -138,7 +139,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("flattens top-level object unions in function schemas", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLMRequest.update(request, {
           tools: [
             ToolDefinition.make({
@@ -191,7 +192,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("lowers chronological system updates to escaped user wrappers in order", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -217,7 +218,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("prepares OpenAI Responses WebSocket target", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLMRequest.update(request, {
           model: OpenAIResponses.webSocketRoute
             .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("test") })
@@ -395,7 +396,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("prepares function call and function output input items", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result",
           model,
@@ -432,7 +433,7 @@ describe("OpenAI Responses route", () => {
         content: [],
         structured: {},
       }
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -453,7 +454,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("keeps primitive tool errors as plain text", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -469,7 +470,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("keeps non-JSON tool errors as plain text", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -487,7 +488,7 @@ describe("OpenAI Responses route", () => {
   // image data is not JSON-stringified into `function_call_output.output`.
   it.effect("lowers image tool-result content as structured input_image items", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_image",
           model,
@@ -516,7 +517,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("lowers single-image tool-result content as structured input_image array", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_image_only",
           model,
@@ -540,7 +541,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("lowers PDF tool-result content as structured input_file array", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_pdf",
           model,
@@ -575,7 +576,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("uses xAI inline file encoding for PDF tool results", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: xaiModel,
           messages: [
@@ -610,7 +611,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("rejects unsupported media in tool-result content with a clear error", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_tool_result_unsupported_media",
           model,
@@ -633,7 +634,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("prepares the composed native continuation request", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         continuationRequest({
           id: "req_native_continuation_openai",
           model,
@@ -675,7 +676,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("maps OpenAI provider options to Responses options", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenAI.configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).model("gpt-5.2"),
           prompt: "think",
@@ -700,7 +701,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("accepts the full ResponseIncludable union", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           prompt: "hi",
@@ -722,7 +723,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("filters unknown includable values out of the include array", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           prompt: "hi",
@@ -739,7 +740,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("treats an explicit empty include as no include at all", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({ model, prompt: "hi", providerOptions: { openai: { include: [] } } }),
       )
 
@@ -749,7 +750,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("treats an all-invalid include as no include at all", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({ model, prompt: "hi", providerOptions: { openai: { include: ["bogus.thing"] } } }),
       )
 
@@ -759,7 +760,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("omits include when no include is set", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({ model, prompt: "hi", providerOptions: { openai: { store: false } } }),
       )
 
@@ -773,7 +774,7 @@ describe("OpenAI Responses route", () => {
       // reasoningSummary: "auto" by default. Without `include`, a follow-up
       // turn cannot replay reasoning state, so the facade also opts into
       // `reasoning.encrypted_content` automatically.
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenAI.configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).responses("gpt-5.2"),
           prompt: "hi",
@@ -788,7 +789,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("lets callers opt out of the GPT-5 default include", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenAI.configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).responses("gpt-5.2"),
           prompt: "hi",
@@ -802,7 +803,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("request OpenAI provider options override route defaults", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenAI.configure({
             baseURL: "https://api.openai.test/v1/",
@@ -934,9 +935,7 @@ describe("OpenAI Responses route", () => {
         },
       ])
 
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
-        LLM.request({ model, messages: [response.message] }),
-      )
+      const prepared = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
       expect(prepared.body.input).toEqual([
         {
           role: "assistant",
@@ -1270,7 +1269,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("preserves assistant content order around reasoning items", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_reasoning_order",
           model,
@@ -1308,7 +1307,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("references stored reasoning items by id", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -1330,7 +1329,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("references stored provider-executed hosted tool results by id", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -1367,7 +1366,7 @@ describe("OpenAI Responses route", () => {
   it.effect("continues stateless hosted image generation with the generated image", () =>
     Effect.gen(function* () {
       const imageTool = OpenAI.imageGeneration({ action: "edit" })
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           messages: [
@@ -1408,7 +1407,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("joins streamed summary blocks into one continuation reasoning item", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_multi_summary_continuation",
           model,
@@ -1445,7 +1444,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("skips non-persisted reasoning ids without encrypted state", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_reasoning_without_encrypted_state",
           model,
@@ -1762,7 +1761,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("lowers user image and PDF content", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           id: "req_media",
           model,
@@ -1793,7 +1792,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("uses xAI inline file encoding for user PDFs", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenAIResponses.OpenAIResponsesBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: xaiModel,
           messages: [
@@ -1825,7 +1824,7 @@ describe("OpenAI Responses route", () => {
 
   it.effect("rejects unsupported user media content", () =>
     Effect.gen(function* () {
-      const error = yield* LLMClient.prepare(
+      const error = yield* compileRequest(
         LLM.request({
           id: "req_media",
           model,

packages/ai/test/provider/openrouter.test.ts

@@ -2,6 +2,7 @@ import { describe, expect } from "bun:test"
 import { Effect } from "effect"
 import { LLM, Message } from "../../src"
 import { LLMClient } from "../../src/route"
+import { compileRequest } from "../../src/route/client"
 import * as OpenRouter from "../../src/providers/openrouter"
 import { it } from "../lib/effect"
 import { fixedResponse } from "../lib/http"
@@ -19,7 +20,7 @@ describe("OpenRouter", () => {
       })
       expect(model.route.endpoint.baseURL).toBe("https://openrouter.ai/api/v1")
 
-      const prepared = yield* LLMClient.prepare(LLM.request({ model, prompt: "Say hello." }))
+      const prepared = yield* compileRequest(LLM.request({ model, prompt: "Say hello." }))
 
       expect(prepared.route).toBe("openrouter")
       expect(prepared.body).toMatchObject({
@@ -32,7 +33,7 @@ describe("OpenRouter", () => {
 
   it.effect("applies OpenRouter payload options from the model helper", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenRouter.configure({
             apiKey: "test-key",
@@ -100,7 +101,7 @@ describe("OpenRouter", () => {
         { type: "reasoning.text", signature: "signed", format: "anthropic-claude-v1", index: 0 },
         { type: "reasoning.encrypted", data: "opaque", format: "openai-responses-v1", index: 1 },
       ]
-      const prepared = yield* LLMClient.prepare<OpenRouter.OpenRouterBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenRouter.configure({ apiKey: "test-key" }).model("anthropic/claude-sonnet-4.6"),
           messages: [
@@ -133,7 +134,7 @@ describe("OpenRouter", () => {
         { type: "reasoning.encrypted", id: "state", data: "opaque" },
         { type: "reasoning.encrypted", id: "state", data: "opaque" },
       ]
-      const prepared = yield* LLMClient.prepare<OpenRouter.OpenRouterBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenRouter.configure({ apiKey: "test-key" }).model("anthropic/claude-sonnet-4.6"),
           messages: [
@@ -158,7 +159,7 @@ describe("OpenRouter", () => {
         { type: "reasoning.text", id: "first", index: 0, text: "A", opaque: "first" },
         { type: "reasoning.text", id: "second", index: 1, text: "B", opaque: "second" },
       ]
-      const prepared = yield* LLMClient.prepare<OpenRouter.OpenRouterBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenRouter.configure({ apiKey: "test-key" }).model("anthropic/claude-sonnet-4.6"),
           messages: [
@@ -179,7 +180,7 @@ describe("OpenRouter", () => {
 
   it.effect("omits scalar reasoning without continuation details", () =>
     Effect.gen(function* () {
-      const prepared = yield* LLMClient.prepare<OpenRouter.OpenRouterBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model: OpenRouter.configure({ apiKey: "test-key" }).model("anthropic/claude-sonnet-4.6"),
           messages: [Message.assistant({ type: "reasoning", text: "Thinking" })],

packages/ai/test/tool-schema-projection.test.ts

@@ -3,7 +3,8 @@ import { Effect } from "effect"
 import { LLM } from "../src"
 import { OpenAIChat } from "../src/protocols"
 import { ToolSchemaProjection } from "../src/protocols/utils/tool-schema"
-import { Auth, LLMClient } from "../src/route"
+import { Auth } from "../src/route"
+import { compileRequest } from "../src/route/client"
 import { it } from "./lib/effect"
 
 describe("tool schema projections", () => {
@@ -79,7 +80,7 @@ describe("tool schema projections", () => {
       const model = OpenAIChat.route
         .with({ endpoint: { baseURL: "https://api.openai.test/v1/" }, auth: Auth.bearer("test") })
         .model({ id: "kimi-k2", compatibility: { toolSchema: "moonshot" } })
-      const prepared = yield* LLMClient.prepare<OpenAIChat.OpenAIChatBody>(
+      const prepared = yield* compileRequest(
         LLM.request({
           model,
           prompt: "Use the tool.",

packages/app/src/components/dialog-fork.tsx

@@ -69,7 +69,7 @@ export const DialogFork: Component = () => {
     const dir = base64Encode(sdk().directory)
 
     sdk()
-      .api.session.fork({ sessionID, messageID: item.id })
+      .api.session.fork({ sessionID, boundary: { type: "before", messageID: item.id } })
       .then((forked) => {
         dialog.close()
         prompt.set(restored, undefined, { dir, id: forked.id })

packages/app/src/utils/server-compat.test.ts

@@ -26,7 +26,6 @@ function setup(
         return new Response(undefined, { status: 204 })
       if (request.method === "POST" && request.url.endsWith("/prompt")) {
         return Response.json({
-          admittedSeq: 1,
           id: "msg_1",
           sessionID: "ses_1",
           timeCreated: 1,

packages/app/src/utils/server-compat.ts

@@ -229,7 +229,6 @@ function createV1Api(input: CompatibleInput): CompatibleApi {
           ],
         })
         return {
-          admittedSeq: 0,
           id: value.id ?? "",
           sessionID: value.sessionID,
           timeCreated: Date.now(),
@@ -255,7 +254,6 @@ function createV1Api(input: CompatibleInput): CompatibleApi {
           })),
         })
         return {
-          admittedSeq: 0,
           id: value.id ?? "",
           sessionID: value.sessionID,
           timeCreated: Date.now(),
@@ -280,7 +278,6 @@ function createV1Api(input: CompatibleInput): CompatibleApi {
           modelID: value.model.modelID,
         })
         return {
-          admittedSeq: 0,
           id: value.id ?? "",
           sessionID: value.sessionID,
           timeCreated: Date.now(),

packages/cli/src/acp/service.ts

@@ -244,7 +244,10 @@ export function make(input: { readonly client: OpenCodeClient; readonly connecti
       return {}
     },
     forkSession: async (params) => {
-      const forked = await input.client.session.fork({ sessionID: params.sessionId })
+      const forked = await input.client.session.fork({
+        sessionID: params.sessionId,
+        boundary: { type: "through" },
+      })
       const state = await attach(forked, forked.location.directory, params.mcpServers ?? [])
       await replay(state)
       return { sessionId: state.id, configOptions: configOptions(state) }

packages/cli/src/session-target.ts

@@ -105,7 +105,7 @@ async function selectSession(input: {
     return {
       session: input.fork
         ? await input.client.session
-            .fork({ sessionID: explicit.id }, ...requestOptions(input.signal))
+            .fork({ sessionID: explicit.id, boundary: { type: "through" } }, ...requestOptions(input.signal))
             .catch((error) => {
               throw new SessionTargetMutationError(error)
             })
@@ -118,9 +118,11 @@ async function selectSession(input: {
   if (!selected) return { session: undefined, location }
   return {
     session: input.fork
-      ? await input.client.session.fork({ sessionID: selected.id }, ...requestOptions(input.signal)).catch((error) => {
-          throw new SessionTargetMutationError(error)
-        })
+      ? await input.client.session
+          .fork({ sessionID: selected.id, boundary: { type: "through" } }, ...requestOptions(input.signal))
+          .catch((error) => {
+            throw new SessionTargetMutationError(error)
+          })
       : selected,
   }
 }

packages/cli/test/acp/service-lifecycle.test.ts

@@ -136,7 +136,7 @@ describe("acp service lifecycle", () => {
       method: "POST",
       path: "/api/session/ses_loaded/fork",
       query: {},
-      body: {},
+      body: { boundary: { type: "through" } },
     })
   })
 

packages/cli/test/run/noninteractive.test.ts

@@ -267,7 +267,7 @@ async function run(input: {
     values.push(...input.turn(messageID))
     wake?.()
     wake = undefined
-    return ok({ admittedSeq: 1, id: messageID, sessionID: "ses_1", timeCreated: 1 }) as never
+    return ok({ id: messageID, sessionID: "ses_1", timeCreated: 1 }) as never
   })
   await runNonInteractivePrompt({
     client: sdk,

packages/client/src/effect/api/api.ts

@@ -136,7 +136,7 @@ export type Endpoint5_4Input = { readonly sessionID: Session.ID }
 export type Endpoint5_4Output = void
 export type SessionRemoveOperation<E = never> = (input: Endpoint5_4Input) => Effect.Effect<Endpoint5_4Output, E>
 
-export type Endpoint5_5Input = { readonly sessionID: Session.ID; readonly messageID?: SessionMessage.ID | undefined }
+export type Endpoint5_5Input = { readonly sessionID: Session.ID; readonly boundary: Session.ForkRequestBoundary }
 export type Endpoint5_5Output = Session.Info
 export type SessionForkOperation<E = never> = (input: Endpoint5_5Input) => Effect.Effect<Endpoint5_5Output, E>
 
@@ -342,8 +342,10 @@ export type Endpoint5_26Output =
           readonly data: {
             readonly sessionID: Session.ID
             readonly parentID: Session.ID
-            readonly parentSeq: number
-            readonly from?: SessionMessage.ID | undefined
+            readonly boundary: Session.ForkBoundary
+            readonly instructions?:
+              | { readonly [x: string & Brand.Brand<"Instruction.Key">]: string & Brand.Brand<"Instruction.Hash"> }
+              | undefined
           }
         }
       | {

packages/client/src/effect/generated/client.ts

@@ -329,7 +329,7 @@ const Endpoint5_4 = (raw: RawClient["server.session"]) => (input: Endpoint5_4Inp
 
 const Endpoint5_5 = (raw: RawClient["server.session"]) => (input: Endpoint5_5Input) =>
   preserveEffect<Endpoint5_5Output>()(
-    raw["session.fork"]({ params: { sessionID: input["sessionID"] }, payload: { messageID: input["messageID"] } }).pipe(
+    raw["session.fork"]({ params: { sessionID: input["sessionID"] }, payload: { boundary: input["boundary"] } }).pipe(
       Effect.mapError(mapClientError),
       Effect.map((value) => value.data),
     ),

packages/client/src/promise/generated/client.ts

@@ -510,7 +510,7 @@ export function make(options: ClientOptions) {
           {
             method: "POST",
             path: `/api/session/${encodeURIComponent(input.sessionID)}/fork`,
-            body: { messageID: input["messageID"] },
+            body: { boundary: input["boundary"] },
             successStatus: 200,
             declaredStatuses: [404, 400, 401],
             empty: false,

packages/client/src/promise/generated/types.ts

@@ -14,6 +14,8 @@ export type PermissionEffect = "allow" | "deny" | "ask"
 
 export type PluginInfo = { id: string }
 
+export type SessionForkBoundary = { type: "before"; messageID: string } | { type: "through"; messageID: string }
+
 export type MoneyUSD = number
 
 export type TokenUsageInfo = {
@@ -43,13 +45,7 @@ export type PromptMention = { start: number; end: number; text: string }
 
 export type SessionPendingSyntheticData = { text: string; description?: string; metadata?: { [x: string]: JsonValue } }
 
-export type SessionPendingCompaction = {
-  admittedSeq: number
-  id: string
-  sessionID: string
-  timeCreated: number
-  type: "compaction"
-}
+export type SessionPendingCompaction = { id: string; sessionID: string; timeCreated: number; type: "compaction" }
 
 export type SessionMessageAgentSelected = {
   id: string
@@ -628,7 +624,7 @@ export type SessionForked = {
   type: "session.forked"
   durable: { aggregateID: string; seq: number; version: 2 }
   location?: LocationRef
-  data: { sessionID: string; parentID: string; parentSeq: number; from?: string }
+  data: { sessionID: string; parentID: string; boundary: SessionForkBoundary; instructions?: { [x: string]: string } }
 }
 
 export type SessionInputPromoted = {
@@ -1210,7 +1206,6 @@ export type PromptFileAttachment = {
 export type PromptAgentAttachment = { name: string; mention?: PromptMention }
 
 export type SessionPendingSynthetic = {
-  admittedSeq: number
   id: string
   sessionID: string
   timeCreated: number
@@ -1755,7 +1750,7 @@ export type PermissionRuleset = Array<PermissionRule>
 export type SessionInfo = {
   id: string
   parentID?: string
-  fork?: { sessionID: string; messageID?: string }
+  fork?: { sessionID: string; boundary: SessionForkBoundary }
   projectID: string
   agent?: string
   model?: ModelRef
@@ -1966,7 +1961,6 @@ export type AgentInfo = {
 export type SessionsResponse = { data: Array<SessionInfo>; cursor: { previous?: string | null; next?: string | null } }
 
 export type SessionPendingUser = {
-  admittedSeq: number
   id: string
   sessionID: string
   timeCreated: number
@@ -2702,7 +2696,9 @@ export type SessionRemoveOutput = void
 
 export type SessionForkInput = {
   readonly sessionID: { readonly sessionID: string }["sessionID"]
-  readonly messageID?: { readonly messageID?: string | undefined }["messageID"]
+  readonly boundary: {
+    readonly boundary: { readonly type: "before"; readonly messageID: string } | { readonly type: "through" }
+  }["boundary"]
 }
 
 export type SessionForkOutput = { data: SessionInfo }["data"]

packages/client/test/effect.test.ts

@@ -268,7 +268,6 @@ const session = {
 
 const admission = {
   data: {
-    admittedSeq: 0,
     id: "msg_test",
     sessionID: "ses_test",
     type: "user",
@@ -281,7 +280,6 @@ const admission = {
 const compactionAdmission = {
   data: {
     type: "compaction",
-    admittedSeq: 1,
     id: "msg_compaction",
     sessionID: "ses_test",
     timeCreated: 1_717_171_717_000,

packages/client/test/promise.test.ts

@@ -304,7 +304,6 @@ test("session.pending.list uses the public HTTP contract", async () => {
   const requests: Array<{ method: string; url: string }> = []
   const pending = [
     {
-      admittedSeq: 3,
       id: "msg_pending",
       sessionID: "ses_test",
       timeCreated: 1_717_171_717_000,
@@ -547,7 +546,6 @@ const session = {
 
 const admission = {
   data: {
-    admittedSeq: 0,
     id: "msg_test",
     sessionID: "ses_test",
     type: "user",
@@ -559,7 +557,6 @@ const admission = {
 
 const syntheticAdmission = {
   data: {
-    admittedSeq: 1,
     id: "msg_synthetic",
     sessionID: "ses_test",
     type: "synthetic",
@@ -572,7 +569,6 @@ const syntheticAdmission = {
 const compactionAdmission = {
   data: {
     type: "compaction",
-    admittedSeq: 1,
     id: "msg_compaction",
     sessionID: "ses_test",
     timeCreated: 1_717_171_717_000,

packages/core/schema.json

@@ -1,9 +1,9 @@
 {
   "version": "7",
   "dialect": "sqlite",
-  "id": "a4ba73b4-21bc-41ab-a415-94e2ca38d798",
+  "id": "db37a97f-9b5e-4c87-be8b-4feace35136c",
   "prevIds": [
-    "5f0a1db8-d4bf-42c3-becb-96b46fe66bed"
+    "a4ba73b4-21bc-41ab-a415-94e2ca38d798"
   ],
   "ddl": [
     {
@@ -1266,17 +1266,7 @@
       "autoincrement": false,
       "default": null,
       "generated": null,
-      "name": "fork_message_id",
-      "entityType": "columns",
-      "table": "session"
-    },
-    {
-      "type": "integer",
-      "notNull": false,
-      "autoincrement": false,
-      "default": null,
-      "generated": null,
-      "name": "fork_seq",
+      "name": "fork_boundary",
       "entityType": "columns",
       "table": "session"
     },

packages/core/src/codemode/catalog.ts

@@ -6,6 +6,7 @@ export const Entry = Schema.Struct({
   path: Schema.String,
   description: Schema.String,
   signature: Schema.String,
+  pinned: Schema.optionalKey(Schema.Boolean),
 })
 export type Entry = typeof Entry.Type
 
@@ -56,26 +57,39 @@ export function summarize(entries: ReadonlyArray<Entry>, budget = INLINE_BUDGET)
           if (left.path > right.path) return 1
           return 0
         })
+      const ranked = rankListings(listings)
+      const pinned = new Set(
+        namespaceEntries
+          .filter((entry) => entry.pinned)
+          .map((entry) => listings.find((listing) => listing.path === entry.path))
+          .filter((listing) => listing !== undefined),
+      )
       return {
         name,
         listings,
-        selectionOrder: rankListings(listings),
-        selectedListings: new Set<typeof Listing.Type>(),
+        selectionOrder: ranked.filter((candidate) => !pinned.has(candidate.listing)),
+        selectedListings: pinned,
+        selectionIndex: 0,
       }
     })
 
   const active = new Set(namespaces)
-  let remaining = budget
+  let remaining =
+    budget -
+    namespaces
+      .flatMap((namespace) => namespace.listings.filter((listing) => namespace.selectedListings.has(listing)))
+      .reduce((total, listing) => total + Math.round(listing.line.length / CHARACTERS_PER_TOKEN), 0)
   while (active.size > 0) {
     for (const namespace of active) {
-      const candidate = namespace.selectionOrder[namespace.selectedListings.size]
+      const candidate = namespace.selectionOrder[namespace.selectionIndex]
       if (!candidate || candidate.cost > remaining) {
         active.delete(namespace)
         continue
       }
       namespace.selectedListings.add(candidate.listing)
+      namespace.selectionIndex += 1
       remaining -= candidate.cost
-      if (namespace.selectedListings.size === namespace.selectionOrder.length) active.delete(namespace)
+      if (namespace.selectionIndex === namespace.selectionOrder.length) active.delete(namespace)
     }
   }
 

packages/core/src/codemode/instructions.ts

@@ -6,7 +6,9 @@ import { Instructions } from "../instructions/index"
 import { CodeModeCatalog } from "./catalog"
 
 // prettier-ignore
-const prompt = (hasMoreTools: boolean) => `The Code Mode tool catalog below is ${hasMoreTools ? "partial" : "complete"}.${hasMoreTools ? `
+const prompt = (hasMoreTools: boolean) => `The Code Mode tool catalog below is ${hasMoreTools ? "partial" : "complete"}.
+
+${hasMoreTools ? "The Code Mode catalog and `search` results are" : "This catalog is"} the complete set of tools available within Code Mode. Tools presented elsewhere are not available in this runtime.${hasMoreTools ? `
 
 ## Search
 

packages/core/src/codemode/tool.ts

@@ -33,7 +33,7 @@ type CollectedFiles = {
 
 // Invariant model-facing guidance; the changing tool catalog is delivered through Instructions.
 const description = [
-  "Run JavaScript to orchestrate tool calls and compose their results through `{ code }` in a confined Code Mode runtime.",
+  "Run JavaScript in a confined Code Mode runtime to orchestrate tool calls and compose their results.",
   "Imports, direct filesystem access, and timers are unavailable. Do not use `fetch`; all external access goes through `tools`.",
   "Within `{ code }`, the only callable tools are those explicitly listed in the Code Mode catalog instructions or returned by `search`. Inside `{ code }`, ignore tools shown outside the Code Mode catalog. They are not available in the Code Mode runtime.",
   'Call tools through `tools` using only exact paths and signatures from the catalog. Do not infer or normalize tool names; preserve bracket notation such as `tools.<namespace>["tool-name"](input)`.',
@@ -138,7 +138,14 @@ export const create = (
 }
 
 export const catalog = (registrations: ReadonlyMap<string, Info>) => {
-  return runtime(registrations, () => Effect.fail(toolError("Execute context is unavailable"))).catalog()
+  const pinned = new Set(
+    Array.from(registrations.values())
+      .filter((registration) => registration.options?.pinned === true)
+      .map(qualifiedName),
+  )
+  return runtime(registrations, () => Effect.fail(toolError("Execute context is unavailable")))
+    .catalog()
+    .map((entry) => ({ ...entry, pinned: pinned.has(entry.path) }))
 }
 
 function runtime(
@@ -149,11 +156,7 @@ function runtime(
   const tools: Record<string, Tool.Tool<never>> = {}
   for (const [name, registration] of registrations) {
     const child = definition(registration)
-    const normalized = registration.name.replace(/[^a-zA-Z0-9_-]/g, "_")
-    const path =
-      registration.options?.namespace === undefined
-        ? normalized
-        : `${registration.options.namespace}.${normalized}`
+    const path = qualifiedName(registration)
     tools[path] = Tool.make({
       description: child.description,
       input: child.inputSchema,
@@ -164,6 +167,12 @@ function runtime(
   return CodeMode.make<typeof tools>({ tools, ...hooks })
 }
 
+function qualifiedName(registration: Info) {
+  const normalized = registration.name.replace(/[^a-zA-Z0-9_-]/g, "_")
+  if (registration.options?.namespace === undefined) return normalized
+  return `${registration.options.namespace}.${normalized}`
+}
+
 // Tool inputs arrive as parsed JSON, so the JSON value cast is a boundary fact.
 function displayInput(input: unknown): Record<string, typeof Schema.Json.Type> | undefined {
   if (input === null || input === undefined) return

packages/core/src/config.ts

@@ -393,7 +393,8 @@ export const layer = (options?: Options) => Layer.effect(
         const key = JSON.stringify(target)
         if (watched.has(key)) continue
         watched.add(key)
-        yield* watcher.subscribe(target).pipe(
+        const stream = yield* watcher.subscribe(target)
+        yield* stream.pipe(
           Stream.runForEach((update) => PubSub.publish(updates, update)),
           Effect.forkScoped({ startImmediately: true }),
         )

packages/core/src/database/migration.gen.ts

@@ -57,5 +57,6 @@ export const migrations = (
     import("./migration/20260716020354_kv"),
     import("./migration/20260722011141_delete_tool_progress_events"),
     import("./migration/20260722170000_canonical_tool_results"),
+    import("./migration/20260729022634_session_fork_boundary"),
   ])
 ).map((module) => module.default) satisfies DatabaseMigration.Migration[]

packages/core/src/database/migration/20260729022634_session_fork_boundary.ts

@@ -0,0 +1,13 @@
+import { Effect } from "effect"
+import type { DatabaseMigration } from "../migration"
+
+export default {
+  id: "20260729022634_session_fork_boundary",
+  up(tx) {
+    return Effect.gen(function* () {
+      yield* tx.run(`ALTER TABLE \`session\` ADD \`fork_boundary\` text;`)
+      yield* tx.run(`ALTER TABLE \`session\` DROP COLUMN \`fork_message_id\`;`)
+      yield* tx.run(`ALTER TABLE \`session\` DROP COLUMN \`fork_seq\`;`)
+    })
+  },
+} satisfies DatabaseMigration.Migration

packages/core/src/database/schema.gen.ts

@@ -213,8 +213,7 @@ export default {
           \`workspace_id\` text,
           \`parent_id\` text,
           \`fork_session_id\` text,
-          \`fork_message_id\` text,
-          \`fork_seq\` integer,
+          \`fork_boundary\` text,
           \`slug\` text NOT NULL,
           \`directory\` text NOT NULL,
           \`path\` text,

packages/core/src/filesystem/location-watcher.ts

@@ -47,13 +47,12 @@ const layer = Layer.effect(
       const home = path.resolve(location.directory) === path.resolve(os.homedir())
 
       if (!home && location.vcs) {
-        yield* watcher
-          .subscribe({
-            path: location.directory,
-            type: "directory",
-            ignore: [...Ignore.PATTERNS, ...config, ...protecteds(location.directory)],
-          })
-          .pipe(Stream.runForEach(publish), Effect.forkScoped)
+        const updates = yield* watcher.subscribe({
+          path: location.directory,
+          type: "directory",
+          ignore: [...Ignore.PATTERNS, ...config, ...protecteds(location.directory)],
+        })
+        yield* updates.pipe(Stream.runForEach(publish), Effect.forkScoped)
       }
       if (home) {
         yield* Effect.logInfo("location watcher skipped home directory", { directory: location.directory })
@@ -68,9 +67,8 @@ const layer = Layer.effect(
           const ignore = (yield* fs.readDirectoryEntries(vcs).pipe(Effect.catch(() => Effect.succeed([])))).flatMap(
             (entry) => (entry.name === "HEAD" ? [] : [entry.name]),
           )
-          yield* watcher
-            .subscribe({ path: vcs, type: "directory", ignore })
-            .pipe(Stream.runForEach(publish), Effect.forkScoped)
+          const updates = yield* watcher.subscribe({ path: vcs, type: "directory", ignore })
+          yield* updates.pipe(Stream.runForEach(publish), Effect.forkScoped)
         }
       }
     }).pipe(

packages/core/src/filesystem/watcher.ts

@@ -59,7 +59,7 @@ export interface NativeInterface {
 export class Native extends Context.Service<Native, NativeInterface>()("@opencode/Watcher/Native") {}
 
 export interface Interface {
-  readonly subscribe: (input: WatchInput) => Stream.Stream<Update>
+  readonly subscribe: (input: WatchInput) => Effect.Effect<Stream.Stream<Update>>
 }
 
 export const Options = Schema.Struct({
@@ -83,7 +83,7 @@ export const layer = (options?: Options) =>
     Service,
     Effect.gen(function* () {
       if (options?.enabled === false) {
-        return Service.of({ subscribe: () => Stream.empty })
+        return Service.of({ subscribe: () => Effect.succeed(Stream.empty) })
       }
       const native = yield* Native
 
@@ -131,11 +131,19 @@ export const layer = (options?: Options) =>
       const subscribe = (input: WatchInput) => {
         const target = path.resolve(input.path)
         const ignore = [...new Set(input.type === "directory" ? (input.ignore ?? []) : [])].toSorted()
-        return Stream.unwrap(
-          RcMap.get(watchers, { type: input.type, target, ignore }).pipe(
-            Effect.map((pubsub) => Stream.fromPubSub(pubsub)),
-          ),
-        )
+        return Effect.gen(function* () {
+          yield* Effect.logInfo("watcher subscribe", {
+            path: target,
+            type: input.type,
+            ignores: ignore.length,
+          })
+          return Stream.unwrap(
+            Effect.gen(function* () {
+              const pubsub = yield* RcMap.get(watchers, { type: input.type, target, ignore })
+              return Stream.fromPubSub(pubsub)
+            }),
+          )
+        })
       }
 
       return Service.of({ subscribe })

packages/core/src/plugin/hooks.ts

@@ -2,6 +2,7 @@ export * as PluginHooks from "./hooks"
 
 import type { AISDKHooks } from "@opencode-ai/plugin/effect/aisdk"
 import type { SessionHooks } from "@opencode-ai/plugin/effect/session"
+import type { ShellHooks } from "@opencode-ai/plugin/effect/shell"
 import type { ToolHooks } from "@opencode-ai/plugin/effect/tool"
 import { Context, Effect, Layer, Scope } from "effect"
 import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
@@ -10,6 +11,7 @@ import { State } from "../state"
 export interface Domains {
   readonly aisdk: AISDKHooks
   readonly session: SessionHooks
+  readonly shell: ShellHooks
   readonly tool: ToolHooks
 }
 

packages/core/src/plugin/host.ts

@@ -296,6 +296,9 @@ export const make = Effect.fn("PluginHost.make")(function* (plugin: import("../p
           })
         }),
     },
+    shell: {
+      hook: (name, callback) => hooks.register("shell", name, callback),
+    },
     tool: {
       transform: (callback) =>
         tools

packages/core/src/plugin/promise.ts

@@ -326,6 +326,10 @@ export function fromPromise(plugin: Plugin) {
               ),
             interrupt: (input) => run(host.session.interrupt({ sessionID: Session.ID.make(input.sessionID) })),
           },
+          shell: {
+            hook: (name, callback) =>
+              register(host.shell.hook(name, (event) => Effect.promise(() => Promise.resolve(callback(event))))),
+          },
         }
 
         const cleanup = yield* Effect.promise(() => Promise.resolve(plugin.setup(context2)))

packages/core/src/plugin/supervisor.ts

@@ -253,7 +253,8 @@ const layer = Layer.effect(
         // inside), so don't watch what can't trigger anything.
         if (yield* fs.isDir(operation.target)) continue
         watched.add(operation.target)
-        yield* watcher.subscribe({ path: operation.target, type: "file" }).pipe(
+        const updates = yield* watcher.subscribe({ path: operation.target, type: "file" })
+        yield* updates.pipe(
           Stream.runForEach(() => PubSub.publish(configuredChanges, undefined)),
           Effect.catchCause((cause) =>
             Effect.logError("configured plugin watch failed", { target: operation.target, cause }),

packages/core/src/session.ts

@@ -28,11 +28,12 @@ import { fromRow } from "./session/info"
 import { SessionRunner } from "./session/runner/index"
 import { SessionStore } from "./session/store"
 import { SessionExecution } from "./session/execution"
-import { MessageDecodeError, NotFoundError } from "./session/error"
+import { ForkEmptyError, MessageDecodeError, NotFoundError } from "./session/error"
 import { makeGlobalNode } from "@opencode-ai/util/effect/app-node"
 import { LocationServiceMap } from "./location-service-map"
 import { SessionEvent } from "./session/event"
 import { SessionPending } from "./session/pending"
+import { InstructionState } from "./session/instruction-state"
 import { SessionGenerate } from "./session/generate"
 import { Snapshot } from "./snapshot"
 import { SessionRevert } from "./session/revert"
@@ -106,7 +107,7 @@ type CompactInput = {
 
 type ForkInput = {
   sessionID: SessionSchema.ID
-  messageID?: SessionMessage.ID
+  boundary: Session.ForkRequestBoundary
 }
 
 export class OperationUnavailableError extends Schema.TaggedErrorClass<OperationUnavailableError>()(
@@ -181,7 +182,9 @@ export interface Interface {
     readonly data: SessionSchema.Info[]
   }>
   readonly create: (input: CreateInput) => Effect.Effect<SessionSchema.Info, NotFoundError>
-  readonly fork: (input: ForkInput) => Effect.Effect<SessionSchema.Info, NotFoundError | MessageNotFoundError>
+  readonly fork: (
+    input: ForkInput,
+  ) => Effect.Effect<SessionSchema.Info, NotFoundError | MessageNotFoundError | ForkEmptyError>
   readonly get: (sessionID: SessionSchema.ID) => Effect.Effect<SessionSchema.Info, NotFoundError>
   readonly remove: (sessionID: SessionSchema.ID) => Effect.Effect<void, NotFoundError>
   readonly messages: (input: {
@@ -395,25 +398,33 @@ const layer = Layer.effect(
       }),
       fork: Effect.fn("Session.fork")(function* (input) {
         const parent = yield* result.get(input.sessionID)
-        const boundary = input.messageID
-          ? yield* db
-              .select({ seq: SessionMessageTable.seq })
-              .from(SessionMessageTable)
-              .where(
-                and(eq(SessionMessageTable.session_id, input.sessionID), eq(SessionMessageTable.id, input.messageID)),
-              )
-              .get()
-              .pipe(Effect.orDie)
-          : undefined
-        if (input.messageID && !boundary)
-          return yield* new MessageNotFoundError({ sessionID: input.sessionID, messageID: input.messageID })
+        const boundary = yield* db
+          .select({ id: SessionMessageTable.id, seq: SessionMessageTable.seq })
+          .from(SessionMessageTable)
+          .where(
+            and(
+              eq(SessionMessageTable.session_id, input.sessionID),
+              input.boundary.type === "before" ? eq(SessionMessageTable.id, input.boundary.messageID) : undefined,
+            ),
+          )
+          .orderBy(desc(SessionMessageTable.seq))
+          .limit(1)
+          .get()
+          .pipe(Effect.orDie)
+        if (!boundary && input.boundary.type === "before")
+          return yield* new MessageNotFoundError({
+            sessionID: input.sessionID,
+            messageID: input.boundary.messageID,
+          })
+        if (!boundary) return yield* new ForkEmptyError({ sessionID: input.sessionID })
         const sessionID = SessionSchema.ID.create()
-        const parentSeq = boundary ? boundary.seq - 1 : yield* Bus.latestSequence(db, parent.id)
+        const instructionThrough =
+          input.boundary.type === "before" ? boundary.seq - 1 : yield* Bus.latestSequence(db, parent.id)
         yield* bus.publish(SessionEvent.Forked, {
           sessionID,
           parentID: parent.id,
-          parentSeq,
-          from: input.messageID,
+          boundary: { ...input.boundary, messageID: boundary.id },
+          instructions: yield* InstructionState.valuesAt(db, parent.id, instructionThrough),
         })
         return yield* result.get(sessionID).pipe(Effect.orDie)
       }),

packages/core/src/session/error.ts

@@ -10,6 +10,14 @@ export class NotFoundError extends Schema.TaggedErrorClass<NotFoundError>()("Ses
   sessionID: SessionSchema.ID,
 }) {}
 
+export class ForkEmptyError extends Schema.TaggedErrorClass<ForkEmptyError>()("Session.ForkEmptyError", {
+  sessionID: SessionSchema.ID,
+}) {
+  override get message() {
+    return `Cannot fork empty session: ${this.sessionID}`
+  }
+}
+
 export class MessageDecodeError extends Schema.TaggedErrorClass<MessageDecodeError>()("Session.MessageDecodeError", {
   sessionID: SessionSchema.ID,
   messageID: SessionMessage.ID,

packages/core/src/session/info.ts

@@ -8,7 +8,6 @@ import { AbsolutePath, RelativePath } from "../schema"
 import { Workspace } from "../workspace"
 import { SessionSchema } from "./schema"
 import { SessionTable } from "./sql"
-import { SessionMessage } from "./message"
 import { PersistedRevert } from "@opencode-ai/schema/session-revert"
 import { Money } from "@opencode-ai/schema/money"
 
@@ -20,12 +19,13 @@ export function fromRow(row: typeof SessionTable.$inferSelect): SessionSchema.In
     projectID: Project.ID.make(row.project_id),
     title: row.title,
     parentID: row.parent_id ? SessionSchema.ID.make(row.parent_id) : undefined,
-    fork: row.fork_session_id
-      ? {
-          sessionID: SessionSchema.ID.make(row.fork_session_id),
-          messageID: row.fork_message_id ? SessionMessage.ID.make(row.fork_message_id) : undefined,
-        }
-      : undefined,
+    fork:
+      row.fork_session_id && row.fork_boundary
+        ? {
+            sessionID: SessionSchema.ID.make(row.fork_session_id),
+            boundary: row.fork_boundary,
+          }
+        : undefined,
     agent: row.agent ? Agent.ID.make(row.agent) : undefined,
     model: row.model
       ? {

packages/core/src/session/instruction-state.ts

@@ -10,11 +10,12 @@ import { SessionEvent } from "./event"
 import { SessionMessage } from "./message"
 import { Event } from "@opencode-ai/schema/event"
 import { SessionSchema } from "./schema"
-import { InstructionBlobTable, InstructionStateTable, SessionTable } from "./sql"
+import { InstructionBlobTable, InstructionStateTable } from "./sql"
 
 type DatabaseService = Database.Interface["db"]
 
 const decodeInstructionsUpdated = Schema.decodeUnknownSync(SessionEvent.InstructionsUpdated.data)
+const decodeForked = Schema.decodeUnknownSync(SessionEvent.Forked.data)
 
 export interface Observation extends Instructions.Admission {
   readonly sessionID: SessionSchema.ID
@@ -94,6 +95,26 @@ export const apply = Effect.fn("InstructionState.apply")(function* (
     .pipe(Effect.orDie)
 })
 
+export const initialize = Effect.fn("InstructionState.initialize")(function* (
+  db: DatabaseService,
+  sessionID: SessionSchema.ID,
+  seq: number,
+  values: Instructions.Values,
+) {
+  yield* db
+    .insert(InstructionStateTable)
+    .values({
+      session_id: sessionID,
+      epoch_start: seq,
+      through_seq: seq,
+      initial_values: values,
+      current_values: values,
+    })
+    .onConflictDoNothing()
+    .run()
+    .pipe(Effect.orDie)
+})
+
 export const advanceEpoch = Effect.fn("InstructionState.advanceEpoch")(function* (
   db: DatabaseService,
   sessionID: SessionSchema.ID,
@@ -255,6 +276,14 @@ const stateFromEvents = Effect.fnUntraced(function* (db: DatabaseService, sessio
   return folded ? foldedState(sessionID, folded) : undefined
 })
 
+export const valuesAt = Effect.fn("InstructionState.valuesAt")(function* (
+  db: DatabaseService,
+  sessionID: SessionSchema.ID,
+  through: number,
+) {
+  return fold(yield* instructionEvents(db, sessionID, through))?.current
+})
+
 const latestRelevantSequence = Effect.fnUntraced(function* (db: DatabaseService, sessionID: SessionSchema.ID) {
   return yield* db
     .select({ seq: EventTable.seq })
@@ -324,15 +353,23 @@ const revertedEventType = Bus.versionedType(
   SessionEvent.RevertEvent.Committed.type,
   SessionEvent.RevertEvent.Committed.durable.version,
 )
-const relevantEventTypes = [instructionEventType, compactionEventType, movedEventType, revertedEventType]
+const forkedEventType = Bus.versionedType(SessionEvent.Forked.type, SessionEvent.Forked.durable.version)
+const relevantEventTypes = [
+  forkedEventType,
+  instructionEventType,
+  compactionEventType,
+  movedEventType,
+  revertedEventType,
+]
 
 type InstructionEventRow = typeof EventTable.$inferSelect
 
 const instructionEvents = Effect.fnUntraced(function* (
   db: DatabaseService,
   sessionID: SessionSchema.ID,
+  through?: number,
 ): Effect.fn.Return<ReadonlyArray<InstructionEventRow>> {
-  return yield* eventRows(db, sessionID, relevantEventTypes)
+  return yield* eventRows(db, sessionID, relevantEventTypes, undefined, through)
 })
 
 const instructionUpdatesAfter = Effect.fnUntraced(function* (
@@ -348,48 +385,22 @@ const eventRows = Effect.fnUntraced(function* (
   sessionID: SessionSchema.ID,
   types: ReadonlyArray<string>,
   after?: number,
-): Effect.fn.Return<ReadonlyArray<InstructionEventRow>> {
-  const segments = (yield* lineage(db, sessionID)).filter(
-    (segment) => after === undefined || segment.through === undefined || segment.through > after,
-  )
-  return (yield* Effect.forEach(segments, (segment) =>
-    db
-      .select()
-      .from(EventTable)
-      .where(
-        and(
-          eq(EventTable.aggregate_id, segment.sessionID),
-          inArray(EventTable.type, types),
-          segment.through === undefined ? undefined : lte(EventTable.seq, segment.through),
-          after === undefined ? undefined : gt(EventTable.seq, after),
-        ),
-      )
-      .orderBy(asc(EventTable.seq))
-      .all()
-      .pipe(Effect.orDie),
-  )).flat()
-})
-
-const lineage = Effect.fnUntraced(function* (
-  db: DatabaseService,
-  sessionID: SessionSchema.ID,
   through?: number,
-): Effect.fn.Return<ReadonlyArray<{ readonly sessionID: SessionSchema.ID; readonly through?: number }>> {
-  const session = yield* db
-    .select({ parentID: SessionTable.fork_session_id, forkSeq: SessionTable.fork_seq })
-    .from(SessionTable)
-    .where(eq(SessionTable.id, sessionID))
-    .get()
+): Effect.fn.Return<ReadonlyArray<InstructionEventRow>> {
+  return yield* db
+    .select()
+    .from(EventTable)
+    .where(
+      and(
+        eq(EventTable.aggregate_id, sessionID),
+        inArray(EventTable.type, types),
+        after === undefined ? undefined : gt(EventTable.seq, after),
+        through === undefined ? undefined : lte(EventTable.seq, through),
+      ),
+    )
+    .orderBy(asc(EventTable.seq))
+    .all()
     .pipe(Effect.orDie)
-  const inherited =
-    session?.parentID && session.forkSeq !== null
-      ? yield* lineage(
-          db,
-          session.parentID,
-          through === undefined ? session.forkSeq : Math.min(session.forkSeq, through),
-        )
-      : []
-  return [...inherited, { sessionID, ...(through === undefined ? {} : { through }) }]
 })
 
 function fold(rows: ReadonlyArray<InstructionEventRow>) {
@@ -402,6 +413,12 @@ function fold(rows: ReadonlyArray<InstructionEventRow>) {
       }
     | undefined
   >((state, row) => {
+    if (row.type === forkedEventType) {
+      const instructions = decodeForked(row.data).instructions
+      return instructions
+        ? { epochStart: row.seq, throughSeq: row.seq, initial: instructions, current: instructions }
+        : undefined
+    }
     if (row.type === movedEventType || row.type === revertedEventType) return undefined
     if (row.type === compactionEventType)
       return state

packages/core/src/session/pending.ts

@@ -53,7 +53,6 @@ export class LifecycleConflict extends Schema.TaggedErrorClass<LifecycleConflict
 
 const fromRow = (row: typeof SessionPendingTable.$inferSelect): Info => {
   const base = {
-    admittedSeq: row.admitted_seq,
     id: SessionMessage.ID.make(row.id),
     sessionID: SessionSchema.ID.make(row.session_id),
     timeCreated: DateTime.makeUnsafe(row.time_created),
@@ -134,7 +133,6 @@ const promotedFromHistory = Effect.fn("SessionPending.promotedFromHistory")(func
     const decoded = decodeAdmittedEvent(row.data)
     if (decoded._tag !== "Some" || decoded.value.inputID !== id) continue
     const base = {
-      admittedSeq: row.seq,
       id,
       sessionID,
       timeCreated: DateTime.makeUnsafe(row.created),
@@ -172,10 +170,7 @@ export const admit = Effect.fn("SessionPending.admit")(function* (
     })
     .pipe(
       Effect.flatMap((event) => {
-        if (event.durable === undefined)
-          return Effect.die(new Error("Session input admission event is missing aggregate sequence"))
         const base = {
-          admittedSeq: event.durable.seq,
           id: request.id,
           sessionID: request.sessionID,
           timeCreated: event.created,

packages/core/src/session/projector.ts

@@ -1,6 +1,6 @@
 export * as SessionProjector from "./projector"
 
-import { and, asc, desc, eq, gt, gte, inArray, lt, sql } from "drizzle-orm"
+import { and, asc, desc, eq, gt, gte, inArray, lt, lte, sql } from "drizzle-orm"
 import { DateTime, Effect, Layer, Schema, Stream } from "effect"
 import { Database } from "../database/database"
 import { Bus } from "../bus"
@@ -174,25 +174,28 @@ const projectFork = Effect.fn("SessionProjector.projectFork")(function* (
     .get()
     .pipe(Effect.orDie)
   if (!parent) return yield* Effect.die(new Error(`Fork parent session not found: ${event.data.parentID}`))
-  const boundary = event.data.from
-    ? yield* db
-        .select({ seq: SessionMessageTable.seq })
-        .from(SessionMessageTable)
-        .where(
-          and(eq(SessionMessageTable.session_id, event.data.parentID), eq(SessionMessageTable.id, event.data.from)),
-        )
-        .get()
-        .pipe(Effect.orDie)
-    : undefined
-  if (event.data.from && !boundary)
-    return yield* Effect.die(new Error(`Fork boundary message not found: ${event.data.from}`))
+  const boundary = yield* db
+    .select({ seq: SessionMessageTable.seq })
+    .from(SessionMessageTable)
+    .where(
+      and(
+        eq(SessionMessageTable.session_id, event.data.parentID),
+        eq(SessionMessageTable.id, event.data.boundary.messageID),
+      ),
+    )
+    .get()
+    .pipe(Effect.orDie)
+  if (!boundary)
+    return yield* Effect.die(new Error(`Fork boundary message not found: ${event.data.boundary.messageID}`))
   const copied = yield* db
     .select({ seq: SessionMessageTable.seq })
     .from(SessionMessageTable)
     .where(
       and(
         eq(SessionMessageTable.session_id, event.data.parentID),
-        boundary === undefined ? undefined : lt(SessionMessageTable.seq, boundary.seq),
+        event.data.boundary.type === "before"
+          ? lt(SessionMessageTable.seq, boundary.seq)
+          : lte(SessionMessageTable.seq, boundary.seq),
       ),
     )
     .orderBy(desc(SessionMessageTable.seq))
@@ -207,8 +210,7 @@ const projectFork = Effect.fn("SessionProjector.projectFork")(function* (
       id: event.data.sessionID,
       parent_id: null,
       fork_session_id: event.data.parentID,
-      fork_message_id: event.data.from,
-      fork_seq: event.data.parentSeq,
+      fork_boundary: event.data.boundary,
       project_id: parent.project_id,
       workspace_id: parent.workspace_id,
       slug: Slug.create(),
@@ -314,8 +316,9 @@ const projectFork = Effect.fn("SessionProjector.projectFork")(function* (
 
     cursor = rows.at(-1)!.seq
   }
-  yield* Bus.reserveSequence(db, event.data.sessionID, event.data.parentSeq)
-  yield* InstructionState.rebuild(db, event.data.sessionID)
+  if (copiedSeq !== undefined) yield* Bus.reserveSequence(db, event.data.sessionID, copiedSeq)
+  if (event.data.instructions)
+    yield* InstructionState.initialize(db, event.data.sessionID, event.durable.seq, event.data.instructions)
 })
 
 function run(db: DatabaseService, event: MessageEvent) {

packages/core/src/session/sql.ts

@@ -32,8 +32,7 @@ export const SessionTable = sqliteTable(
     workspace_id: text().$type<Workspace.ID>(),
     parent_id: text().$type<SessionSchema.ID>(),
     fork_session_id: text().$type<SessionSchema.ID>(),
-    fork_message_id: text().$type<SessionMessage.ID>(),
-    fork_seq: integer(),
+    fork_boundary: text({ mode: "json" }).$type<Session.ForkBoundary>(),
     slug: text().notNull(),
     directory: directoryColumn().notNull(),
     path: pathColumn(),

packages/core/src/shell.ts

@@ -12,6 +12,8 @@ import { Bus } from "./bus"
 import { Location } from "./location"
 import { Global } from "@opencode-ai/util/global"
 import { ShellSelect } from "./shell/select"
+import type { ShellCreateBefore } from "@opencode-ai/plugin/effect/shell"
+import { PluginHooks } from "./plugin/hooks"
 
 export class NotFoundError extends Schema.TaggedErrorClass<NotFoundError>()("Shell.NotFoundError", {
   id: Shell.ID,
@@ -45,7 +47,10 @@ type Active = {
  */
 export interface Interface {
   readonly name: () => Effect.Effect<string>
-  readonly create: (input: Shell.CreateInput) => Effect.Effect<Shell.Info>
+  readonly create: <E = never, R = never>(
+    input: Shell.CreateInput,
+    before?: (input: ShellCreateBefore) => Effect.Effect<void, E, R>,
+  ) => Effect.Effect<Shell.Info, E, R>
   // Currently running commands only; exited shells are retained for get/output but excluded here.
   readonly list: () => Effect.Effect<Shell.Info[]>
   readonly get: (id: Shell.ID) => Effect.Effect<Shell.Info, NotFoundError>
@@ -68,6 +73,7 @@ export const layer = (options?: ShellSelect.Options) => Layer.effect(
     const config = yield* Config.Service
     const global = yield* Global.Service
     const appProcess = yield* AppProcess.Service
+    const hooks = yield* PluginHooks.Service
     const context = yield* Effect.context()
     const runFork = Effect.runForkWith(context)
     const sessions = new Map<string, Active>()
@@ -172,24 +178,34 @@ export const layer = (options?: ShellSelect.Options) => Layer.effect(
       }
     })
 
-    const create = Effect.fn("Shell.create")(function* (input: Shell.CreateInput) {
+    const create = Effect.fn("Shell.create")(function* <E = never, R = never>(
+      input: Shell.CreateInput,
+      before?: (input: ShellCreateBefore) => Effect.Effect<void, E, R>,
+    ) {
+      const invocation: ShellCreateBefore = {
+        command: input.command,
+        cwd: input.cwd ?? location.directory,
+        timeout: input.timeout,
+        shell: yield* resolve(),
+        env: {
+          ...process.env,
+          TERM: "xterm-256color",
+          OPENCODE_TERMINAL: "1",
+        },
+      }
+      yield* hooks.trigger("shell", "create.before", invocation)
+      if (before) yield* before(invocation)
+
       const id = Shell.ID.ascending()
-      const cwd = input.cwd ?? location.directory
-      const shell = yield* resolve()
-      const args = ShellSelect.args(shell, input.command)
+      const args = ShellSelect.args(invocation.shell, invocation.command)
       const file = path.join(outputDir, `${id}.out`)
-      const env = {
-        ...process.env,
-        TERM: "xterm-256color",
-        OPENCODE_TERMINAL: "1",
-      } as Record<string, string>
 
       const info: Info = {
         id,
         status: "running",
-        command: input.command,
-        cwd,
-        shell,
+        command: invocation.command,
+        cwd: invocation.cwd,
+        shell: invocation.shell,
         file,
         metadata: input.metadata ?? {},
         time: { started: Date.now() },
@@ -203,9 +219,9 @@ export const layer = (options?: ShellSelect.Options) => Layer.effect(
         Effect.scoped(
           Effect.gen(function* () {
             const handle = yield* appProcess.spawn(
-              ChildProcess.make(shell, args, {
-                cwd,
-                env,
+              ChildProcess.make(invocation.shell, args, {
+                cwd: invocation.cwd,
+                env: invocation.env,
                 stdin: "ignore",
                 detached: process.platform !== "win32",
                 forceKillAfter: Duration.seconds(3),
@@ -297,7 +313,7 @@ export const layer = (options?: ShellSelect.Options) => Layer.effect(
                 )
               })
 
-            yield* session.timeout(input.timeout)
+            yield* session.timeout(invocation.timeout)
 
             runFork(
               handle.exitCode.pipe(
@@ -327,7 +343,7 @@ export function configured(options?: ShellSelect.Options) {
   return makeLocationNode({
     service: Service,
     layer: layer(options),
-    deps: [Bus.node, Location.node, Config.node, Global.node, AppProcess.node],
+    deps: [Bus.node, Location.node, Config.node, Global.node, AppProcess.node, PluginHooks.node],
   })
 }
 

packages/core/src/shell/select.ts

@@ -173,7 +173,7 @@ export function args(file: string, command: string) {
   if (n === "nu" || n === "fish") return ["-c", command]
   if (n === "zsh" || n === "bash") return ["-c", command]
   if (n === "cmd") return ["/c", command]
-  if (ps(file)) return ["-NoProfile", "-Command", command]
+  if (ps(file)) return ["-NoLogo", "-NoProfile", "-NonInteractive", "-Command", command]
   return ["-c", command]
 }
 

packages/core/src/tool/plugin/shell.ts

@@ -1,6 +1,5 @@
 export * as ShellTool from "./shell"
 
-import path from "path"
 import { ToolFailure } from "@opencode-ai/ai"
 import type { Content } from "@opencode-ai/schema/tool"
 import type { Context as PluginContext } from "@opencode-ai/plugin/effect/plugin"
@@ -66,18 +65,14 @@ const Output = Schema.Struct({
   ...StructuredOutput.fields,
   output: Schema.String,
   status: Schema.optionalKey(Schema.Literals(["completed", "running"])),
-  warnings: Schema.optionalKey(Schema.Array(Schema.String)),
 })
 
 type Output = typeof Output.Type
 
 const modelOutput = (output: Output): string | undefined => {
-  const warnings = output.warnings?.length
-    ? `\n\nWarnings:\n${output.warnings.map((warning) => `- ${warning}`).join("\n")}`
-    : ""
-  if (output.status === "running") return `${warnings.trimStart()}${warnings ? "\n\n" : ""}${BACKGROUND_INSTRUCTION}`
-  if (output.timeout) return `${warnings.trimStart()}${warnings ? "\n\n" : ""}Command timed out before completion.`
-  return `${warnings.trimStart()}${warnings ? "\n\n" : ""}Command exited with code ${output.exit}.`
+  if (output.status === "running") return BACKGROUND_INSTRUCTION
+  if (output.timeout) return "Command timed out before completion."
+  return `Command exited with code ${output.exit}.`
 }
 
 /**
@@ -86,31 +81,12 @@ const modelOutput = (output: Output): string | undefined => {
  */
 // TODO: Port tree-sitter bash / PowerShell parser-based approval reduction.
 // TODO: Port BashArity reusable command-prefix approvals.
-// TODO: Replace token-based command-argument external-directory advisories with parser-based detection.
 // TODO: Add plugin shell.env environment augmentation once plugin hooks exist.
 // TODO: Persist job status and define restart recovery before exposing remote observation.
 // TODO: Add HTTP job observation only after durable status, restart recovery, and authorization are defined.
 // TODO: Revisit process-group cleanup and platform coverage with shell-specific tests if current AppProcess semantics do not fully cover it.
 // TODO: Revisit binary output handling if stdout/stderr decoding is text-only.
 
-const shellTokens = (command: string) => command.match(/(?:[^\s"']+|"[^"]*"|'[^']*')+/g) ?? []
-const unquote = (value: string) => value.replace(/^(['"])(.*)\1$/, "$2")
-const externalCommandDirectories = Effect.fn("ShellTool.externalCommandDirectories")(function* (
-  fs: FSUtil.Interface,
-  command: string,
-  cwd: string,
-) {
-  const directories = new Set<string>()
-  for (const token of shellTokens(command)) {
-    const value = unquote(token).replace(/[;,|&]+$/, "")
-    if (!path.isAbsolute(value)) continue
-    const resolved = yield* fs.resolve(value)
-    if (FSUtil.contains(cwd, resolved)) continue
-    directories.add(yield* fs.resolve(path.dirname(resolved)))
-  }
-  return [...directories]
-})
-
 export const Plugin = {
   id: "opencode.tool.shell",
   effect: Effect.fn("ShellTool.Plugin")(function* (ctx: PluginContext) {
@@ -170,38 +146,40 @@ export const Plugin = {
                   messageID: context.messageID,
                   callID: context.callID,
                 }
-                const target = yield* mutation.resolve({ path: input.workdir ?? ".", kind: "directory" })
-                const external = target.externalDirectory
-                if (external)
-                  yield* permission.assert({
-                    ...LocationMutation.externalDirectoryPermission(external),
-                    sessionID: context.sessionID,
-                    agent: context.agent,
-                    source,
-                  })
-                const warnings = (yield* externalCommandDirectories(fsUtil, input.command, target.canonical)).map(
-                  (directory) =>
-                    `Command argument references external directory ${path.join(directory, "*").replaceAll("\\", "/")}. Shell runs with host-user filesystem, process, and network authority; this scan is advisory only.`,
-                )
-                yield* permission.assert({
-                  action: name,
-                  resources: [input.command],
-                  save: [input.command],
-                  sessionID: context.sessionID,
-                  agent: context.agent,
-                  source,
-                })
-
-                if ((yield* fsUtil.stat(target.canonical)).type !== "Directory")
-                  return yield* Effect.fail(new Error(`Working directory is not a directory: ${target.canonical}`))
-
                 const timeout = input.background === true ? (input.timeout ?? 0) : (input.timeout ?? DEFAULT_TIMEOUT_MS)
-                const info = yield* shell.create({
-                  command: input.command,
-                  cwd: target.canonical,
-                  timeout,
-                  metadata: { sessionID: context.sessionID },
-                })
+                let finalTimeout = timeout
+                const info = yield* shell.create(
+                  {
+                    command: input.command,
+                    cwd: input.workdir,
+                    timeout,
+                    metadata: { sessionID: context.sessionID },
+                  },
+                  (invocation) =>
+                    Effect.gen(function* () {
+                      const target = yield* mutation.resolve({ path: invocation.cwd, kind: "directory" })
+                      invocation.cwd = target.canonical
+                      finalTimeout = invocation.timeout
+                      const external = target.externalDirectory
+                      if (external)
+                        yield* permission.assert({
+                          ...LocationMutation.externalDirectoryPermission(external),
+                          sessionID: context.sessionID,
+                          agent: context.agent,
+                          source,
+                        })
+                      yield* permission.assert({
+                        action: name,
+                        resources: [invocation.command],
+                        save: [invocation.command],
+                        sessionID: context.sessionID,
+                        agent: context.agent,
+                        source,
+                      })
+                      if ((yield* fsUtil.stat(target.canonical)).type !== "Directory")
+                        return yield* Effect.fail(new Error(`Working directory is not a directory: ${target.canonical}`))
+                    }),
+                )
                 yield* context.progress({ shellID: info.id })
 
                 const captureShell = Effect.fn("ShellTool.captureShell")(function* () {
@@ -220,20 +198,20 @@ export const Plugin = {
 
                 const settleShell = Effect.fn("ShellTool.settleShell")(function* () {
                   const final = yield* shell.wait(info.id)
+                  const capture = yield* captureShell()
 
                   // `exit` is optionalKey in the Output schema; a present-but-undefined key
                   // fails output encoding, so omit it when the process has no exit code.
                   if (final.status === "timeout") {
                     return {
                       ...(final.exit !== undefined ? { exit: final.exit } : {}),
-                      output: `Command exceeded timeout of ${timeout} ms. Retry with a larger timeout if the command is expected to take longer.`,
-                      truncated: false,
+                      output: `${capture.output}\n\nCommand exceeded timeout of ${finalTimeout} ms. Retry with a larger timeout if the command is expected to take longer.`,
+                      truncated: capture.truncated,
                       timeout: true,
                       status: "completed" as const,
                     }
                   }
 
-                  const capture = yield* captureShell()
                   return {
                     ...(final.exit !== undefined ? { exit: final.exit } : {}),
                     output: capture.output,
@@ -251,20 +229,19 @@ export const Plugin = {
                 const job = yield* runtime.job.start({
                   id: context.callID,
                   type: name,
-                  title: input.command,
+                  title: info.command,
                   metadata: { sessionID: context.sessionID, shellID: info.id },
                   run,
                 })
 
                 if (input.background === true) {
                   yield* runtime.job.background(job.id)
-                  yield* notifyWhenDone(context.sessionID, context.callID, input.command)
+                  yield* notifyWhenDone(context.sessionID, context.callID, info.command)
                   return {
                     output: BACKGROUND_STARTED,
                     shellID: info.id,
                     truncated: false,
                     status: "running" as const,
-                    ...(warnings.length ? { warnings } : {}),
                   }
                 }
 
@@ -273,20 +250,19 @@ export const Plugin = {
                 )
                 if (result?.type === "backgrounded") {
                   yield* shell.timeout(info.id, 0)
-                  yield* notifyWhenDone(context.sessionID, context.callID, input.command)
+                  yield* notifyWhenDone(context.sessionID, context.callID, info.command)
                   return {
                     output: BACKGROUND_STARTED,
                     shellID: info.id,
                     truncated: false,
                     status: "running" as const,
-                    ...(warnings.length ? { warnings } : {}),
                   }
                 }
                 if (result?.info.status === "error")
                   return yield* Effect.fail(new Error(result.info.error ?? "Command failed"))
                 if (result?.info.status === "cancelled") return yield* Effect.fail(new Error("Command cancelled"))
 
-                return { ...(yield* Deferred.await(settled)), ...(warnings.length ? { warnings } : {}) }
+                return yield* Deferred.await(settled)
               }).pipe(
                 Effect.map((output) => {
                   const content: Array<Content> = [{ type: "text", text: output.output }]

packages/core/test/aisdk.test.ts

@@ -1,9 +1,10 @@
-import type { LanguageModelV3, LanguageModelV3CallOptions, LanguageModelV3StreamPart } from "@ai-sdk/provider"
+import type { LanguageModelV3, LanguageModelV3StreamPart } from "@ai-sdk/provider"
 import { AISDK } from "@opencode-ai/core/aisdk"
 import { Model } from "@opencode-ai/core/model"
 import { Provider } from "@opencode-ai/core/provider"
 import { LLM, LLMError, LLMEvent, Message } from "@opencode-ai/ai"
 import { LLMClient, RequestExecutor } from "@opencode-ai/ai/route"
+import { compileRequest } from "@opencode-ai/ai/route/client"
 import { expect } from "bun:test"
 import { Effect, Layer } from "effect"
 import { testEffect } from "./lib/effect"
@@ -87,9 +88,7 @@ it.effect("projects request settings, headers, and body overlays", () =>
       headers: { "x-test": "header" },
       body: { safety_setting: "strict" },
     })
-    const prepared = yield* LLMClient.prepare<LanguageModelV3CallOptions>(
-      LLM.request({ model: resolved, prompt: "Hello" }),
-    )
+    const prepared = yield* compileRequest(LLM.request({ model: resolved, prompt: "Hello" }))
 
     expect(prepared.body.providerOptions).toEqual({
       google: { thinkingConfig: { thinkingBudget: 1024 } },
@@ -112,9 +111,7 @@ it.effect("maps pro reasoning bodies to AI SDK provider options", () =>
       ...model("@ai-sdk/openai"),
       body: { reasoning: { mode: "pro" } },
     })
-    const prepared = yield* LLMClient.prepare<LanguageModelV3CallOptions>(
-      LLM.request({ model: resolved, prompt: "Hello" }),
-    )
+    const prepared = yield* compileRequest(LLM.request({ model: resolved, prompt: "Hello" }))
 
     expect(body).toBeUndefined()
     expect(prepared.body.providerOptions).toEqual({
@@ -139,9 +136,7 @@ it.effect("maps package-specific AI SDK provider option keys", () =>
     ] as const
     for (const [packageName, key, settings] of cases) {
       const resolved = yield* aisdk.model(model(packageName, { reasoningEffort: "high" }))
-      const prepared = yield* LLMClient.prepare<LanguageModelV3CallOptions>(
-        LLM.request({ model: resolved, prompt: "Hello" }),
-      )
+      const prepared = yield* compileRequest(LLM.request({ model: resolved, prompt: "Hello" }))
       expect(prepared.body.providerOptions).toEqual({ [key]: settings })
     }
   }),
@@ -155,17 +150,13 @@ it.effect("forces reasoning and projects both Azure AI SDK namespaces", () =>
     })
 
     const openai = yield* aisdk.model(model("@ai-sdk/openai", { reasoningEffort: "high" }))
-    const openaiPrepared = yield* LLMClient.prepare<LanguageModelV3CallOptions>(
-      LLM.request({ model: openai, prompt: "Hello" }),
-    )
+    const openaiPrepared = yield* compileRequest(LLM.request({ model: openai, prompt: "Hello" }))
     expect(openaiPrepared.body.providerOptions).toEqual({
       openai: { reasoningEffort: "high", forceReasoning: true },
     })
 
     const azure = yield* aisdk.model(model("@ai-sdk/azure", { reasoningEffort: "high" }))
-    const azurePrepared = yield* LLMClient.prepare<LanguageModelV3CallOptions>(
-      LLM.request({ model: azure, prompt: "Hello" }),
-    )
+    const azurePrepared = yield* compileRequest(LLM.request({ model: azure, prompt: "Hello" }))
     expect(azurePrepared.body.providerOptions).toEqual({
       openai: { reasoningEffort: "high", forceReasoning: true },
       azure: { reasoningEffort: "high", forceReasoning: true },
@@ -187,9 +178,7 @@ it.effect("routes AI Gateway model options by upstream prefix", () =>
       }),
       modelID: Model.ID.make("anthropic/claude-sonnet-5"),
     })
-    const anthropicPrepared = yield* LLMClient.prepare<LanguageModelV3CallOptions>(
-      LLM.request({ model: anthropic, prompt: "Hello" }),
-    )
+    const anthropicPrepared = yield* compileRequest(LLM.request({ model: anthropic, prompt: "Hello" }))
     expect(anthropicPrepared.body.providerOptions).toEqual({
       gateway: { order: ["anthropic"] },
       anthropic: { thinking: { type: "adaptive" } },
@@ -199,9 +188,7 @@ it.effect("routes AI Gateway model options by upstream prefix", () =>
       ...model("@ai-sdk/gateway", { reasoningConfig: { type: "enabled" } }),
       modelID: Model.ID.make("amazon/nova-2-lite"),
     })
-    const bedrockPrepared = yield* LLMClient.prepare<LanguageModelV3CallOptions>(
-      LLM.request({ model: bedrock, prompt: "Hello" }),
-    )
+    const bedrockPrepared = yield* compileRequest(LLM.request({ model: bedrock, prompt: "Hello" }))
     expect(bedrockPrepared.body.providerOptions).toEqual({
       bedrock: { reasoningConfig: { type: "enabled" } },
     })
@@ -210,9 +197,7 @@ it.effect("routes AI Gateway model options by upstream prefix", () =>
       ...model("@ai-sdk/gateway", { reasoningEffort: "high" }),
       modelID: Model.ID.make("deepseek/deepseek-v4"),
     })
-    const fallbackPrepared = yield* LLMClient.prepare<LanguageModelV3CallOptions>(
-      LLM.request({ model: fallback, prompt: "Hello" }),
-    )
+    const fallbackPrepared = yield* compileRequest(LLM.request({ model: fallback, prompt: "Hello" }))
     expect(fallbackPrepared.body.providerOptions).toEqual({
       deepseek: { reasoningEffort: "high" },
     })
@@ -228,7 +213,7 @@ it.effect("projects replay metadata onto AI SDK prompt parts", () =>
 
     const resolved = yield* aisdk.model(model("@ai-sdk/anthropic"))
     expect(resolved.route.providerMetadataKey).toBe("anthropic")
-    const prepared = yield* LLMClient.prepare<LanguageModelV3CallOptions>(
+    const prepared = yield* compileRequest(
       LLM.request({
         model: resolved,
         messages: [

packages/core/test/codemode.test.ts

@@ -16,6 +16,7 @@ describe("CodeMode", () => {
             description: "Echo text",
             input: Schema.Struct({ text: Schema.String }),
             output: Schema.String,
+            options: { pinned: true },
             execute: ({ text }) => Effect.succeed({ output: text }),
         }),
       )
@@ -27,6 +28,7 @@ describe("CodeMode", () => {
           path: "echo",
           description: "Echo text",
           signature: "tools.echo(input: {\n  text: string,\n}): Promise<string>",
+          pinned: true,
         },
       ])
     }).pipe(

packages/core/test/codemode/catalog.test.ts

@@ -2,10 +2,11 @@ import { describe, expect, test } from "bun:test"
 import { CodeModeCatalog } from "@opencode-ai/core/codemode/catalog"
 import { CodeModeInstructions } from "@opencode-ai/core/codemode/instructions"
 
-const entry = (path: string, description: string, signature?: string): CodeModeCatalog.Entry => ({
+const entry = (path: string, description: string, signature?: string, pinned = false): CodeModeCatalog.Entry => ({
   path,
   description,
   signature: signature ?? `tools.${path}(input: {\n  q: string,\n}): Promise<string>`,
+  pinned,
 })
 
 const lookup = entry(
@@ -46,6 +47,30 @@ describe("CodeModeCatalog.summarize", () => {
     expect(catalog.namespaces.every((namespace) => namespace.entries.length === 0)).toBe(true)
   })
 
+  test("always retains pinned tools beyond the inline budget", () => {
+    const pinned = [
+      entry("alpha.first", "First", undefined, true),
+      entry("beta.second", "Second", undefined, true),
+    ]
+    const catalog = CodeModeCatalog.summarize([...pinned, entry("alpha.unpinned", "Unpinned")], 0)
+
+    expect(catalog.shown).toBe(2)
+    expect(catalog.namespaces.flatMap((namespace) => namespace.entries.map((item) => item.path))).toEqual([
+      "alpha.first",
+      "beta.second",
+    ])
+  })
+
+  test("spends the budget remaining after pinned tools on unpinned tools", () => {
+    const pinned = entry("alpha.pinned", "Pinned", undefined, true)
+    const unpinned = entry("beta.unpinned", "Unpinned")
+    const pinCost = Math.round(`  - ${pinned.signature} // Pinned`.length / 4)
+    const unpinnedCost = Math.round(`  - ${unpinned.signature} // Unpinned`.length / 4)
+
+    expect(CodeModeCatalog.summarize([pinned, unpinned], pinCost + unpinnedCost).shown).toBe(2)
+    expect(CodeModeCatalog.summarize([pinned, unpinned], pinCost + unpinnedCost - 1).shown).toBe(1)
+  })
+
   test("retains only the rendered portion of inline descriptions", () => {
     const catalog = CodeModeCatalog.summarize([entry("alpha.one", `Summary\n${"detail".repeat(10_000)}`)])
     expect(catalog.namespaces[0]?.entries[0]?.line).toEndWith("// Summary")
@@ -67,6 +92,7 @@ describe("CodeModeInstructions.render", () => {
     expect(instructions).toContain(`  - ${lookup.signature} // Look up an order by ID`)
     expect(instructions).not.toContain("## Search")
     expect(instructions).toContain("The Code Mode tool catalog below is complete.")
+    expect(instructions).toContain("This catalog is the complete set of tools available within Code Mode.")
     expect(instructions).not.toContain("surrounding top-level agent tools")
   })
 
@@ -76,6 +102,9 @@ describe("CodeModeInstructions.render", () => {
     expect(partial).toContain("- orders (1 tool, none shown)")
     expect(partial).toContain("## Search")
     expect(partial).toContain("The Code Mode tool catalog below is partial.")
+    expect(partial).toContain(
+      "The Code Mode catalog and `search` results are the complete set of tools available within Code Mode.",
+    )
     expect(partial).not.toContain("surrounding top-level agent tools")
     expect(partial).toContain("- search(input: {")
     expect(partial).toContain("  limit?: number,\n  offset?: number,")

packages/core/test/codemode/instructions.test.ts

@@ -44,6 +44,9 @@ describe("CodeModeInstructions", () => {
   it.effect("renders the initial catalog, semantic deltas, and removal", () =>
     Effect.gen(function* () {
       const initialized = yield* readInitial(CodeModeInstructions.make([echo]))
+      expect(initialized.text).toContain(
+        "This catalog is the complete set of tools available within Code Mode. Tools presented elsewhere are not available in this runtime.",
+      )
       expect(initialized.text).toContain("## Available tools")
       expect(initialized.text).not.toContain("## Search")
       expect(initialized.text).toContain(`  - ${echo.signature} // Echo text`)