refactor(core): simplify request preparation

.changeset/share-session-model-request.md

@@ -0,0 +1,5 @@
+---
+"@opencode-ai/core": patch
+---
+
+Apply shared Session model-request preparation to transient generation.

packages/core/src/session/generate-node.ts

@@ -1,85 +1,54 @@
 export * as SessionGenerateNode from "./generate-node.js"
 
-import { LLM, LLMClient, Message, SystemPart } from "@opencode-ai/ai"
+import { LLMClient, Message } from "@opencode-ai/ai"
 import { Effect, Layer } from "effect"
 import { Database } from "../database/database.js"
 import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
-import { App } from "../app.js"
 import { llmClient } from "../effect/app-node-platform.js"
-import { PluginHooks } from "../plugin/hooks.js"
 import { SessionContext } from "./context.js"
 import { SessionGenerate } from "./generate.js"
 import { SessionHistory } from "./history.js"
-import { SessionModelHeaders } from "./model-headers.js"
-import { SessionPromptCacheKey } from "./prompt-cache-key.js"
+import { SessionModelRequest } from "./model-request.js"
 import { SessionRunnerModel } from "./runner/model.js"
-import { SessionSystemPrompt } from "./system-prompt.js"
-import { toLLMMessages } from "./runner/to-llm-message.js"
 
 export const layer = Layer.effect(
   SessionGenerate.Service,
   Effect.gen(function* () {
     const context = yield* SessionContext.Service
     const database = yield* Database.Service
-    const hooks = yield* PluginHooks.Service
     const llm = yield* LLMClient.Service
     const models = yield* SessionRunnerModel.Service
-    const app = yield* App.Metadata
+    const modelRequests = yield* SessionModelRequest.Service
 
     return SessionGenerate.Service.of({
       generate: Effect.fn("SessionGenerate.generate")(function* (input) {
         const selection = yield* context.select(input.sessionID)
         const model = yield* models.resolve(selection.session)
         const history = yield* SessionHistory.preview(database.db, selection.session.id, selection.instructions)
-        const providerMetadataKey = model.model.route.providerMetadataKey ?? model.model.provider
-        const tools = selection.tools
-        const toolDefinitions = tools.definitions
-        const toolsByName = new Map(toolDefinitions.map((tool) => [tool.name, tool]))
-        const contextEvent = yield* hooks.trigger("session", "context", {
-          sessionID: selection.session.id,
-          agent: selection.agent.id,
-          model: model.ref,
-          system: [
-            selection.agent.info.system
-              ? selection.agent.info.system
-              : SessionSystemPrompt.make(toolDefinitions.map((tool) => tool.name)),
-            history.initial,
-          ]
-            .filter((part) => part.length > 0)
-            .map(SystemPart.make),
-          messages: [
-            ...toLLMMessages(history.messages, model.ref, providerMetadataKey),
-            ...(history.instructionUpdate ? [Message.system(history.instructionUpdate)] : []),
-            Message.user(input.prompt),
-          ],
-          tools: Object.fromEntries(
-            toolDefinitions.map((tool) => [
-              tool.name,
-              { description: tool.description, input: { ...tool.inputSchema } },
-            ]),
-          ),
+        const transcript = SessionModelRequest.baseTranscript({
+          agent: selection.agent.info,
+          model,
+          tools: selection.tools,
+          initial: history.initial,
+          messages: history.messages,
         })
-        const hookedTools = Object.entries(contextEvent.tools).flatMap(([name, tool]) => {
-          const registered = toolsByName.get(name)
-          return registered
-            ? [Object.assign({}, registered, { description: tool.description, inputSchema: tool.input })]
-            : []
+        const prepared = yield* modelRequests.prepare({
+          scope: { session: selection.session, agentID: selection.agent.id, model, tools: selection.tools },
+          transcript: {
+            system: transcript.system,
+            messages: [
+              ...transcript.messages,
+              ...(history.instructionUpdate ? [Message.system(history.instructionUpdate)] : []),
+              Message.user(input.prompt),
+            ],
+          },
         })
         yield* Effect.logInfo("sending session generation request", {
           sessionID: selection.session.id,
           providerID: model.ref.providerID,
           modelID: model.ref.id,
         })
-        const response = yield* llm.generate(
-          LLM.request({
-            model: model.model,
-            http: { headers: SessionModelHeaders.make(selection.session, app) },
-            promptCacheKey: SessionPromptCacheKey.make(selection.session.id),
-            system: contextEvent.system,
-            messages: contextEvent.messages,
-            tools: hookedTools,
-          }),
-        )
+        const response = yield* llm.generate(prepared.request, prepared.options)
         yield* Effect.logInfo("session generation usage diagnostic", { usage: response.usage })
         return response.text
       }),
@@ -90,5 +59,5 @@ export const layer = Layer.effect(
 export const node = makeLocationNode({
   service: SessionGenerate.Service,
   layer,
-  deps: [SessionContext.node, Database.node, PluginHooks.node, SessionRunnerModel.node, App.node, llmClient],
+  deps: [SessionContext.node, Database.node, SessionModelRequest.node, SessionRunnerModel.node, llmClient],
 })

packages/core/src/session/model-request.ts

@@ -12,15 +12,16 @@ import { Permission } from "../permission.js"
 import { PluginHooks } from "../plugin/hooks.js"
 import { QuestionTool } from "../tool/plugin/question.js"
 import { Tool } from "../tool.js"
-import { SessionContext } from "./context.js"
 import { SessionModelHeaders } from "./model-headers.js"
 import { SessionModelHttp } from "./model-http.js"
 import { SessionModelTransport } from "./model-transport.js"
 import { SessionPromptCacheKey } from "./prompt-cache-key.js"
-import { PromptCacheDiagnostics } from "./prompt-cache-diagnostics.js"
-import { MAX_STEPS_PROMPT } from "./runner/max-steps.js"
+import { SessionRunnerModel } from "./runner/model.js"
+import { SessionSchema } from "./schema.js"
 import { SessionSystemPrompt } from "./system-prompt.js"
 import { toLLMMessages } from "./runner/to-llm-message.js"
+import type { SessionMessage } from "./message.js"
+import type { Agent } from "../agent.js"
 
 const IMAGE_BYTES_TRIGGER = 25 * 1024 * 1024 // 25 MiB
 const IMAGE_BYTES_TARGET = 15 * 1024 * 1024 // 15 MiB
@@ -47,20 +48,49 @@ const declineDefect = (cause: Cause.Cause<Tool.Error>) => {
 interface Prepared {
   readonly request: LLMRequest
   readonly options: StreamOptions
-  /** False when Session HTTP hooks require the request to remain on HTTP. */
-  readonly webSocketEligible: boolean
   /**
-   * One request-scoped execution operation. Unknown, hook-removed, and
-   * step-limit-violating calls fail individually through the same seam.
+   * One request-scoped execution operation. Unknown and hook-removed calls
+   * fail individually through the same seam.
    */
   readonly executeTool: (input: Parameters<Tool.Snapshot["execute"]>[0]) => Effect.Effect<Tool.Result, ExecuteError>
-  /** True when this request is the final Step; violating calls are rejected and no continuation follows. */
-  readonly stepLimitReached: boolean
 }
 
 interface PrepareInput {
-  readonly context: SessionContext.Loaded
-  readonly step: number
+  readonly scope: {
+    readonly session: SessionSchema.Info
+    readonly agentID: Agent.ID
+    readonly model: SessionRunnerModel.Resolved
+    readonly tools: Tool.Snapshot
+  }
+  readonly transcript: {
+    readonly system: Array<SystemPart>
+    readonly messages: Array<Message>
+  }
+  readonly toolChoice?: LLM.RequestInput["toolChoice"]
+  /** Stateful Session WebSocket channels require an explicit durable-runner opt-in. */
+  readonly webSocket?: "session"
+}
+
+export const baseTranscript = (input: {
+  readonly agent: Agent.Info
+  readonly model: SessionRunnerModel.Resolved
+  readonly tools: Tool.Snapshot
+  readonly initial: string
+  readonly messages: ReadonlyArray<SessionMessage.Info>
+}) => {
+  const providerMetadataKey = input.model.model.route.providerMetadataKey ?? input.model.model.provider
+  return {
+    providerMetadataKey,
+    system: [
+      input.agent.system
+        ? input.agent.system
+        : SessionSystemPrompt.make(input.tools.definitions.map((tool) => tool.name)),
+      input.initial,
+    ]
+      .filter((part) => part.length > 0)
+      .map(SystemPart.make),
+    messages: toLLMMessages(input.messages, input.model.ref, providerMetadataKey),
+  }
 }
 
 const mimeToModality = (mime: string) => {
@@ -174,30 +204,11 @@ export const layer = Layer.effect(
       Config.withDefault(false),
       Effect.orDie,
     )
-    const diagnostics = yield* Config.boolean("OPENCODE_PROMPT_CACHE_DIAGNOSTICS").pipe(
-      Config.withDefault(false),
-      Effect.orDie,
-    )
-    const promptCacheSnapshots = diagnostics ? new Map<string, PromptCacheDiagnostics.Snapshot>() : undefined
-
     const prepare = Effect.fn("SessionModelRequest.prepare")(function* (input: PrepareInput) {
-      const session = input.context.session
-      const agent = input.context.agent
-      const resolved = input.context.model
+      const session = input.scope.session
+      const resolved = input.scope.model
       const model = resolved.model
-      const providerMetadataKey = model.route.providerMetadataKey ?? model.provider
-      const stepLimitReached = agent.info.steps !== undefined && input.step >= agent.info.steps
-      // The final Step keeps definitions available to protocols with native "none",
-      // preserving their prompt cache prefix. Calls are still rejected at execution.
-      const tools = input.context.tools
-      const system = [
-        agent.info.system ? agent.info.system : SessionSystemPrompt.make(tools.definitions.map((tool) => tool.name)),
-        input.context.initial,
-      ]
-        .filter((part) => part.length > 0)
-        .map(SystemPart.make)
-      const history = toLLMMessages(input.context.messages, resolved.ref, providerMetadataKey)
-      const messages = stepLimitReached ? [...history, Message.assistant(MAX_STEPS_PROMPT)] : history
+      const tools = input.scope.tools
       const registry = new Map(tools.definitions.map((tool) => [tool.name, tool]))
       // The definition objects we hand to hooks, mapped back to their tools. Hooks rename a
       // tool by moving its definition to a new key; recognizing the object recovers the tool.
@@ -209,10 +220,10 @@ export const layer = Layer.effect(
       // Hooks mutate this record in place: edit descriptions and schemas, rename, or remove.
       const context = yield* hooks.trigger("session", "context", {
         sessionID: session.id,
-        agent: agent.id,
+        agent: input.scope.agentID,
         model: resolved.ref,
-        system,
-        messages,
+        system: input.transcript.system,
+        messages: input.transcript.messages,
         tools: Object.fromEntries(Array.from(given, ([definition, tool]) => [tool.name, definition])),
       })
       // Match each surviving entry back to its tool, by recognizing a moved definition or
@@ -235,7 +246,7 @@ export const layer = Layer.effect(
         system: context.system,
         messages: boundImages(unsupportedParts(context.messages, resolved.capabilities)),
         tools: Array.from(hooked, ([name, tool]) => ({ ...tool, name })),
-        toolChoice: stepLimitReached ? "none" : undefined,
+        toolChoice: input.toolChoice,
       })
       const webSocketEligible =
         !(yield* hooks.has("session", "http.request")) && !(yield* hooks.has("session", "http.response"))
@@ -243,37 +254,20 @@ export const layer = Layer.effect(
         ? undefined
         : SessionModelHttp.middleware(hooks, {
             sessionID: session.id,
-            agent: agent.id,
+            agent: input.scope.agentID,
             model: resolved.ref,
           })
       const options: StreamOptions = {
         ...(http ? { http } : {}),
-        ...(webSocket &&
+        ...(input.webSocket === "session" &&
+        webSocket &&
         webSocketEligible &&
         resolved.ref.providerID === Provider.ID.openai &&
         model.route.id === "openai-responses"
           ? { webSocket: transport.bind(session.id) }
           : {}),
       }
-      if (promptCacheSnapshots) {
-        const current = PromptCacheDiagnostics.snapshot(request)
-        const comparison = PromptCacheDiagnostics.compare(promptCacheSnapshots.get(session.id), current)
-        promptCacheSnapshots.delete(session.id)
-        promptCacheSnapshots.set(session.id, current)
-        const oldest = promptCacheSnapshots.keys().next().value
-        if (promptCacheSnapshots.size > 100 && oldest !== undefined) promptCacheSnapshots.delete(oldest)
-        yield* Effect.logInfo("prompt cache prefix").pipe(
-          Effect.annotateLogs({
-            sessionID: session.id,
-            toolCount: current.tools.length,
-            systemParts: current.system.length,
-            messageCount: current.messages.length,
-            ...comparison,
-          }),
-        )
-      }
       const executeTool: Prepared["executeTool"] = (input) => {
-        if (stepLimitReached) return new Tool.Error({ message: "Tools are disabled after the maximum agent steps" })
         const tool = hooked.get(input.call.name)
         // A registered tool absent from the hooked set was removed or renamed by a hook.
         if (!tool && registry.has(input.call.name))
@@ -285,9 +279,7 @@ export const layer = Layer.effect(
       return {
         request,
         options,
-        webSocketEligible,
         executeTool,
-        stepLimitReached,
       }
     })
 

packages/core/src/session/runner/llm.ts

@@ -4,11 +4,12 @@ import {
   LLMClient,
   AIError,
   LLMEvent,
+  Message,
   isContextOverflowFailure,
   type ProviderErrorEvent,
   type ToolCall,
 } from "@opencode-ai/ai"
-import { Cause, Data, Effect, Exit, Fiber, FiberSet, Layer, Option, Pull, Schedule, Stream } from "effect"
+import { Cause, Config, Data, Effect, Exit, Fiber, FiberSet, Layer, Option, Pull, Schedule, Stream } from "effect"
 import { Database } from "../../database/database.js"
 import { Bus } from "../../bus.js"
 import { Permission } from "../../permission.js"
@@ -34,6 +35,9 @@ import { toSessionError } from "../to-session-error.js"
 import { SessionRunnerRetry } from "./retry.js"
 import { SessionUsage } from "../usage.js"
 import { ToolOutput } from "../../tool-output.js"
+import { Tool } from "../../tool.js"
+import { PromptCacheDiagnostics } from "../prompt-cache-diagnostics.js"
+import { MAX_STEPS_PROMPT } from "./max-steps.js"
 
 /** How one model call ended: settled, awaiting retry/recovery, or restarted by compaction. */
 type CallOutcome = Data.TaggedEnum<{
@@ -116,6 +120,32 @@ const layer = Layer.effect(
     const compaction = yield* SessionCompaction.Service
     const title = yield* SessionTitle.Service
     const toolOutput = yield* ToolOutput.Service
+    const diagnostics = yield* Config.boolean("OPENCODE_PROMPT_CACHE_DIAGNOSTICS").pipe(
+      Config.withDefault(false),
+      Effect.orDie,
+    )
+    const promptCacheSnapshots = diagnostics ? new Map<string, PromptCacheDiagnostics.Snapshot>() : undefined
+    const diagnosePromptCache = Effect.fn("SessionRunner.diagnosePromptCache")(function* (
+      sessionID: SessionSchema.ID,
+      request: Parameters<typeof PromptCacheDiagnostics.snapshot>[0],
+    ) {
+      if (!promptCacheSnapshots) return
+      const current = PromptCacheDiagnostics.snapshot(request)
+      const comparison = PromptCacheDiagnostics.compare(promptCacheSnapshots.get(sessionID), current)
+      promptCacheSnapshots.delete(sessionID)
+      promptCacheSnapshots.set(sessionID, current)
+      const oldest = promptCacheSnapshots.keys().next().value
+      if (promptCacheSnapshots.size > 100 && oldest !== undefined) promptCacheSnapshots.delete(oldest)
+      yield* Effect.logInfo("prompt cache prefix").pipe(
+        Effect.annotateLogs({
+          sessionID,
+          toolCount: current.tools.length,
+          systemParts: current.system.length,
+          messageCount: current.messages.length,
+          ...comparison,
+        }),
+      )
+    })
     // Title generation starts once input is visible and must not delay model execution.
     // The in-flight set coalesces overlapping prompts while title presence records success durably.
     const titlesRunning = new Set<SessionSchema.ID>()
@@ -278,10 +308,32 @@ const layer = Layer.effect(
           return CallOutcome.Restart({ step: currentStep, recoveredOverflow: false })
         return yield* new StepFailedError({ error: compacted.error })
       }
-      const prepared = yield* modelRequests.prepare({
-        context: loaded,
-        step: currentStep,
+      const stepLimitReached = agent.info.steps !== undefined && currentStep >= agent.info.steps
+      const transcript = SessionModelRequest.baseTranscript({
+        agent: agent.info,
+        model: resolved,
+        tools: loaded.tools,
+        initial: loaded.initial,
+        messages: loaded.messages,
       })
+      const prepared = yield* modelRequests.prepare({
+        scope: { session, agentID: agent.id, model: resolved, tools: loaded.tools },
+        transcript: {
+          system: transcript.system,
+          messages: stepLimitReached
+            ? [...transcript.messages, Message.assistant(MAX_STEPS_PROMPT)]
+            : transcript.messages,
+        },
+        // The final Step keeps definitions available to protocols with native "none",
+        // preserving their prompt cache prefix. Calls are still rejected at execution.
+        toolChoice: stepLimitReached ? "none" : undefined,
+        webSocket: "session",
+      })
+      yield* diagnosePromptCache(session.id, prepared.request)
+      const executeTool = (input: Parameters<typeof prepared.executeTool>[0]) => {
+        if (stepLimitReached) return new Tool.Error({ message: "Tools are disabled after the maximum agent steps" })
+        return prepared.executeTool(input)
+      }
       // Every local tool call forked here is owned until it reaches one durable settlement.
       const toolRuns: Array<{
         readonly call: ToolCall
@@ -295,7 +347,7 @@ const layer = Layer.effect(
         // The selected catalog identity, not model.id: route-level ids are provider API
         // model ids (for example gpt-5.5-fast resolves to api id gpt-5.5).
         model: resolved.ref,
-        providerMetadataKey: model.route.providerMetadataKey ?? model.provider,
+        providerMetadataKey: transcript.providerMetadataKey,
         snapshot: startSnapshot,
         assistantMessageID,
       })
@@ -356,7 +408,7 @@ const layer = Layer.effect(
               call: event,
               fiber: yield* Effect.uninterruptibleMask((restore) =>
                 restore(
-                  prepared.executeTool({
+                  executeTool({
                     sessionID: session.id,
                     agent: agent.id,
                     messageID: assistantMessageID,
@@ -504,7 +556,7 @@ const layer = Layer.effect(
             // A local call or malformed tool input requires another model step, unless
             // this step already exhausted the agent's allowance.
             needsContinuation:
-              !prepared.stepLimitReached &&
+              !stepLimitReached &&
               record.calls.some((call) => !call.providerExecuted && (call.called || call.settled)),
             step: currentStep,
           })

packages/core/test/session-generate.test.ts

@@ -9,6 +9,7 @@ import {
   type LLMRequest,
 } from "@opencode-ai/ai"
 import { OpenAIChat } from "@opencode-ai/ai/protocols"
+import type { StreamOptions } from "@opencode-ai/ai/route"
 import { Agent } from "@opencode-ai/core/agent"
 import { Database } from "@opencode-ai/core/database/database"
 import { AppNodeBuilder } from "@opencode-ai/core/effect/app-node-builder"
@@ -51,15 +52,17 @@ import { Effect, Layer, Schema, Stream } from "effect"
 import { testEffect } from "./lib/effect"
 
 const requests: LLMRequest[] = []
+const options: Array<StreamOptions | undefined> = []
 let instruction: string | Instructions.Unavailable = "Initial context"
 const sessionID = SessionSchema.ID.make("ses_generate_test")
 
 const model = LanguageModel.make({ id: "generate-model", provider: "test", route: OpenAIChat.route })
 const client = Layer.mock(LLMClient.Service)({
   stream: () => Stream.die(new Error("unused")),
-  generate: (request) =>
+  generate: (request, requestOptions) =>
     Effect.sync(() => {
       requests.push(request)
+      options.push(requestOptions)
       const response = LLMResponse.fromEvents([
         LLMEvent.stepStart({ index: 0 }),
         LLMEvent.textStart({ id: "generate" }),
@@ -221,6 +224,7 @@ const setup = Effect.gen(function* () {
 it.effect("generates from fresh settled Session context without durable mutation", () =>
   Effect.gen(function* () {
     requests.length = 0
+    options.length = 0
     instruction = "Initial context"
     const { db, bus, instructions } = yield* setup
     yield* InstructionState.prepare(db, bus, instructions, sessionID)
@@ -292,6 +296,7 @@ it.effect("generates from fresh settled Session context without durable mutation
         if (event.tools.lookup) event.tools.lookup.description = "Hooked lookup"
       }),
     )
+    yield* hooks.register("session", "http.request", () => Effect.void)
 
     const generate = yield* SessionGenerate.Service
     const result = yield* generate.generate({ sessionID, prompt: "Summarize privately" })
@@ -321,6 +326,8 @@ it.effect("generates from fresh settled Session context without durable mutation
     ).toEqual(["Settled partial answer"])
     expect(requests[0]?.tools).toMatchObject([{ name: "lookup", description: "Hooked lookup" }])
     expect(requests[0]?.toolChoice).toBeUndefined()
+    expect(options[0]?.http).toBeFunction()
+    expect(options[0]?.webSocket).toBeUndefined()
     expect(yield* durableState(db, sessionID)).toEqual(before)
   }),
 )

packages/core/test/session-runner.test.ts

@@ -1004,9 +1004,16 @@ describe("SessionRunnerLLM", () => {
       const database = yield* Database.Service
       const bus = yield* Bus.Service
       yield* InstructionState.prepare(database.db, bus, selected.instructions, sessionID)
+      const loaded = yield* context.load(selected)
       const prepared = yield* modelRequests.prepare({
-        context: yield* context.load(selected),
-        step: 1,
+        scope: {
+          session: loaded.session,
+          agentID: loaded.agent.id,
+          model: loaded.model,
+          tools: loaded.tools,
+        },
+        transcript: { system: [], messages: [] },
+        webSocket: "session",
       })
       const http = prepared.options.http ?? (yield* Effect.die("Expected Session HTTP middleware"))
 
@@ -1015,7 +1022,7 @@ describe("SessionRunnerLLM", () => {
         return Effect.succeed(HttpClientResponse.fromWeb(request, new Response("network")))
       })
 
-      expect(prepared.webSocketEligible).toBe(false)
+      expect(prepared.options.webSocket).toBeUndefined()
       expect(response.headers["x-response-hook"]).toBe("active")
       expect(requestTriggers).toBe(1)
       expect(responseTriggers).toBe(1)