fix(core): estimate semantic request context

packages/core/src/session/compaction.ts

@@ -80,7 +80,60 @@ export interface Interface {
 
 export class Service extends Context.Service<Service, Interface>()("@opencode/v2/SessionCompaction") {}
 
-const estimate = (value: unknown) => Token.estimate(JSON.stringify(value))
+const serializeString = (value: unknown) => {
+  try {
+    return String(value)
+  } catch {
+    return "[unserializable]"
+  }
+}
+
+const serializeJson = (value: unknown) => {
+  try {
+    return JSON.stringify(value) ?? serializeString(value)
+  } catch {
+    return serializeString(value)
+  }
+}
+
+const serializeError = (value: unknown) => {
+  try {
+    const prototype =
+      typeof value === "object" && value !== null && !Array.isArray(value) && Object.getPrototypeOf(value)
+    const structured = Array.isArray(value) || prototype === Object.prototype || prototype === null
+    return structured ? serializeJson(value) : serializeString(value)
+  } catch {
+    return serializeString(value)
+  }
+}
+
+const serializeContent = (part: LLMRequest["messages"][number]["content"][number]) => {
+  if (part.type === "text" || part.type === "reasoning") return part.text
+  if (part.type === "media") return ""
+  if (part.type === "tool-call") return `${part.name}\n${serializeJson(part.input)}`
+  // OpenAI replays hosted image generations by item reference; the opaque JSON result contains the image bytes.
+  if (
+    part.providerExecuted &&
+    part.name === "image_generation" &&
+    part.result.type === "json" &&
+    typeof part.providerMetadata?.openai?.itemId === "string"
+  )
+    return part.name
+  if (part.result.type === "content")
+    return [part.name, ...part.result.value.flatMap((item) => (item.type === "text" ? [item.text] : []))].join("\n")
+  if (part.result.type === "text") return `${part.name}\n${serializeString(part.result.value)}`
+  if (part.result.type === "error") return `${part.name}\n${serializeError(part.result.value)}`
+  return `${part.name}\n${serializeJson(part.result.value)}`
+}
+
+const estimate = (request: LLMRequest) =>
+  Token.estimate(
+    [
+      ...request.system.map((part) => part.text),
+      serializeJson(request.tools),
+      ...request.messages.flatMap((message) => message.content.map(serializeContent).filter(Boolean)),
+    ].join("\n"),
+  )
 
 const truncate = (value: string) =>
   value.length <= TOOL_OUTPUT_MAX_CHARS ? value : `${value.slice(0, TOOL_OUTPUT_MAX_CHARS)}\n[truncated]`
@@ -285,11 +338,7 @@ const make = (dependencies: Dependencies) => {
     const context = input.request.model.route.defaults.limits?.context
     if (context === undefined || context <= 0) return false
     const output = input.request.generation?.maxTokens ?? input.request.model.route.defaults.limits?.output ?? 0
-    if (
-      estimate({ system: input.request.system, messages: input.request.messages, tools: input.request.tools }) <=
-      context - Math.max(output, config.buffer)
-    )
-      return false
+    if (estimate(input.request) <= context - Math.max(output, config.buffer)) return false
     return yield* compactAfterOverflow(input)
   })
   return {

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

@@ -1,6 +1,16 @@
 import { expect, test } from "bun:test"
-import { LLMClient, LLMEvent, Model, type LLMRequest } from "@opencode-ai/llm"
-import { OpenAIChat } from "@opencode-ai/llm/protocols"
+import {
+  LLM,
+  LLMClient,
+  LLMEvent,
+  Message,
+  Model,
+  ToolCallPart,
+  ToolResultPart,
+  type LLMRequest,
+} from "@opencode-ai/llm"
+import { OpenAIChat, OpenAIResponses } from "@opencode-ai/llm/protocols"
+import { Base64, FileAttachment } from "@opencode-ai/schema/prompt"
 import { Config } from "@opencode-ai/core/config"
 import { Database } from "@opencode-ai/core/database/database"
 import { AppNodeBuilder } from "@opencode-ai/core/effect/app-node-builder"
@@ -12,12 +22,15 @@ import { SessionCompaction } from "@opencode-ai/core/session/compaction"
 import { SessionEvent } from "@opencode-ai/core/session/event"
 import { SessionMessage } from "@opencode-ai/core/session/message"
 import { SessionProjector } from "@opencode-ai/core/session/projector"
+import { toLLMMessages } from "@opencode-ai/core/session/runner/to-llm-message"
 import { SessionRunnerModel } from "@opencode-ai/core/session/runner/model"
 import { SessionTable } from "@opencode-ai/core/session/sql"
 import { SessionStore } from "@opencode-ai/core/session/store"
 import { SessionV2 } from "@opencode-ai/core/session"
 import { Project } from "@opencode-ai/core/project"
 import { ProjectTable } from "@opencode-ai/core/project/sql"
+import { ModelV2 } from "@opencode-ai/core/model"
+import { ProviderV2 } from "@opencode-ai/core/provider"
 import { AbsolutePath } from "@opencode-ai/core/schema"
 import { DateTime, Effect, Fiber, Layer, Stream } from "effect"
 import { asc, eq } from "drizzle-orm"
@@ -68,6 +81,138 @@ test("compaction describes tool media without embedding base64", () => {
   expect(serialized).not.toContain(base64)
 })
 
+it.effect("does not count image attachments as text context", () =>
+  Effect.gen(function* () {
+    requests = []
+    const compaction = yield* SessionCompaction.Service
+    const text = "context ".repeat(4_000)
+    const data = Base64.make(Buffer.alloc(64 * 1024).toString("base64"))
+    const image = FileAttachment.make({
+      data,
+      mime: "image/png",
+      source: { type: "inline" },
+      name: "screenshot.png",
+    })
+    const inputModel = Model.make({
+      id: "media-model",
+      provider: "test",
+      route: OpenAIResponses.route.with({ limits: { context: 30_000, output: 1_000 } }),
+    })
+    const inputModelRef = ModelV2.Ref.make({
+      id: ModelV2.ID.make(inputModel.id),
+      providerID: ProviderV2.ID.make(inputModel.provider),
+    })
+    const messages = [
+      SessionMessage.User.make({
+        id: SessionMessage.ID.create(),
+        type: "user",
+        text,
+        time: { created: DateTime.makeUnsafe(0) },
+      }),
+      SessionMessage.User.make({
+        id: SessionMessage.ID.create(),
+        type: "user",
+        text: "Inspect this image",
+        files: [image],
+        time: { created: DateTime.makeUnsafe(1) },
+      }),
+    ]
+    const request = LLM.request({
+      model: inputModel,
+      messages: [
+        ...toLLMMessages(messages, inputModelRef),
+        Message.assistant(
+          ToolResultPart.make({
+            id: "image_generation_1",
+            name: "image_generation",
+            result: { type: "image_generation_call", output: Buffer.alloc(64 * 1024).toString("base64") },
+            providerExecuted: true,
+            providerMetadata: { openai: { itemId: "image_generation_1" } },
+          }),
+        ),
+      ],
+    })
+
+    expect(request.messages.flatMap((message) => message.content)).toContainEqual({
+      type: "media",
+      mediaType: "image/png",
+      data,
+      filename: "screenshot.png",
+    })
+
+    expect(
+      yield* compaction.compactIfNeeded({
+        sessionID: SessionV2.ID.make("ses_media_compaction"),
+        messages,
+        request,
+      }),
+    ).toBe(false)
+    expect(requests).toHaveLength(0)
+  }),
+)
+
+it.effect("counts tool-call inputs as text context", () =>
+  Effect.gen(function* () {
+    requests = []
+    const db = (yield* Database.Service).db
+    const compaction = yield* SessionCompaction.Service
+    const text = "context ".repeat(4_500)
+    const sessionID = SessionV2.ID.make("ses_tool_input_compaction")
+    const inputModel = Model.make({
+      id: "tool-input-model",
+      provider: "test",
+      route: OpenAIChat.route.with({ limits: { context: 30_000, output: 1_000 } }),
+    })
+    const messages = [
+      SessionMessage.User.make({
+        id: SessionMessage.ID.create(),
+        type: "user",
+        text,
+        time: { created: DateTime.makeUnsafe(0) },
+      }),
+      SessionMessage.User.make({
+        id: SessionMessage.ID.create(),
+        type: "user",
+        text: "Continue",
+        time: { created: DateTime.makeUnsafe(1) },
+      }),
+    ]
+    yield* db
+      .insert(ProjectTable)
+      .values({ id: Project.ID.global, worktree: AbsolutePath.make("/project"), sandboxes: [] })
+      .onConflictDoNothing()
+      .run()
+      .pipe(Effect.orDie)
+    yield* db
+      .insert(SessionTable)
+      .values({
+        id: sessionID,
+        project_id: Project.ID.global,
+        slug: "tool-input-compaction",
+        directory: "/project",
+        title: "Tool input compaction",
+        version: "test",
+      })
+      .run()
+      .pipe(Effect.orDie)
+
+    expect(
+      yield* compaction.compactIfNeeded({
+        sessionID,
+        messages,
+        request: LLM.request({
+          model: inputModel,
+          messages: [
+            Message.user(text),
+            Message.assistant(ToolCallPart.make({ id: "call_read", name: "read", input: { path: "x".repeat(8_000) } })),
+          ],
+        }),
+      }),
+    ).toBe(true)
+    expect(requests).toHaveLength(1)
+  }),
+)
+
 test("compaction prompt requires the checkpoint headings in order", () => {
   const prompt = SessionCompaction.buildPrompt({ context: ["Conversation history"] })
   expect(prompt.match(/^#{2,3} .+$/gm)).toEqual([