refactor(core): tune retained media budget

packages/core/src/session/compaction.ts

@@ -23,7 +23,7 @@ const DEFAULT_BUFFER = 20_000
 const DEFAULT_KEEP_TOKENS = 15_000
 const OUTPUT_TOKEN_MAX = 32_000
 const TOOL_OUTPUT_MAX_CHARS = 2_000
-const IMAGE_TOKEN_ESTIMATE = 2_000
+const MEDIA_TOKEN_ESTIMATE = 1_500
 const SUMMARY_TEMPLATE = `Output exactly the Markdown structure shown inside <template> and keep the section order unchanged. Do not include the <template> tags in your response.
 <template>
 ## Objective
@@ -127,9 +127,12 @@ export const serializeToolContent = (content: SessionMessage.ToolStateCompleted[
     )
     .join("\n")
 
-export const estimateImageTokens = (message: SessionMessage.Info) => {
+const isEstimatedMedia = (mime: string) =>
+  mime.toLowerCase().startsWith("image/") || mime.toLowerCase() === "application/pdf"
+
+export const estimateMediaTokens = (message: SessionMessage.Info) => {
   if (message.type === "user")
-    return (message.files?.filter((file) => file.mime.toLowerCase().startsWith("image/")).length ?? 0) * IMAGE_TOKEN_ESTIMATE
+    return (message.files?.filter((file) => isEstimatedMedia(file.mime)).length ?? 0) * MEDIA_TOKEN_ESTIMATE
   if (message.type !== "assistant") return 0
   return (
     message.content
@@ -138,8 +141,7 @@ export const estimateImageTokens = (message: SessionMessage.Info) => {
           ? (part.state.content ?? [])
           : [],
       )
-      .filter((content) => content.type === "file" && content.mime.toLowerCase().startsWith("image/")).length *
-    IMAGE_TOKEN_ESTIMATE
+      .filter((content) => content.type === "file" && isEstimatedMedia(content.mime)).length * MEDIA_TOKEN_ESTIMATE
   )
 }
 
@@ -205,7 +207,7 @@ const select = (
   let total = 0
   let split = conversation.length
   for (let index = conversation.length - 1; index >= 0; index--) {
-    const next = total + Token.estimate(conversation[index].text) + estimateImageTokens(conversation[index].message)
+    const next = total + Token.estimate(conversation[index].text) + estimateMediaTokens(conversation[index].message)
     if (split < conversation.length && next > tokens) break
     total = next
     split = index

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

@@ -114,7 +114,7 @@ test("compaction describes tool media without embedding base64", () => {
   expect(serialized).not.toContain(base64)
 })
 
-test("compaction estimates image context without counting base64", () => {
+test("compaction estimates media context without counting base64", () => {
   const image = FileAttachment.make({
     data: Base64.make("a".repeat(10_000)),
     mime: "image/png",
@@ -125,11 +125,11 @@ test("compaction estimates image context without counting base64", () => {
     id: SessionMessage.ID.create(),
     type: "user",
     text: "Compare these images.",
-    files: [image, image],
+    files: [image, image, FileAttachment.make({ ...image, mime: "application/pdf" })],
     time: { created: DateTime.makeUnsafe(0) },
   })
 
-  expect(SessionCompaction.estimateImageTokens(message)).toBe(4_000)
+  expect(SessionCompaction.estimateMediaTokens(message)).toBe(4_500)
 })
 
 test("compaction prompt requires the checkpoint headings in order", () => {
@@ -197,7 +197,7 @@ it.effect("auto compaction reserves a buffer below the prompt ceiling", () =>
   }),
 )
 
-it.effect("manual compaction summarizes short context instead of no-op", () =>
+it.effect("manual compaction preserves ordered media in the retained tail", () =>
   Effect.gen(function* () {
     requests = []
     const db = (yield* Database.Service).db
@@ -209,7 +209,7 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
     const userMessage = {
       id: SessionMessage.ID.create(),
       type: "user" as const,
-      text: "Manual compaction should include this short conversation.",
+      text: `Manual compaction should include this older conversation. ${"older context ".repeat(4_500)}`,
       time: { created: DateTime.makeUnsafe(0) },
     }
     const recentMessage = SessionMessage.User.make({
@@ -223,9 +223,21 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
           source: { type: "inline" },
           name: "prompt.pdf",
         }),
+        FileAttachment.make({
+          data: Base64.make("aW1hZ2U="),
+          mime: "image/png",
+          source: { type: "inline" },
+          name: "prompt.png",
+        }),
       ],
       time: { created: DateTime.makeUnsafe(1) },
     })
+    const latestMessage = SessionMessage.User.make({
+      id: SessionMessage.ID.create(),
+      type: "user",
+      text: "Newest text after the retained media.",
+      time: { created: DateTime.makeUnsafe(2) },
+    })
     yield* db
       .insert(ProjectTable)
       .values({ id: Project.ID.global, worktree: AbsolutePath.make("/project"), sandboxes: [] })
@@ -261,7 +273,7 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
     expect(
       yield* compaction.compactManual({
         session,
-        messages: [userMessage, recentMessage],
+        messages: [userMessage, recentMessage, latestMessage],
         inputID: SessionMessage.ID.make("msg_manual_compaction"),
       }),
     ).toEqual({ status: "completed" })
@@ -278,13 +290,15 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
       "x-opencode-client": "opencode",
     })
     expect(requests[0]?.generation).toBeUndefined()
-    expect(JSON.stringify(requests[0]?.messages)).toContain("Manual compaction should include this short conversation.")
+    expect(JSON.stringify(requests[0]?.messages)).toContain("Manual compaction should include this older conversation.")
     expect(yield* store.context(sessionID)).toMatchObject([
       {
         type: "compaction",
         reason: "manual",
         summary: "manual summary",
-        recent: expect.stringContaining("[Attached application/pdf: prompt.pdf]"),
+        recent: expect.stringMatching(
+          /\[User\]: Compare the retained media\.\n\[Attached application\/pdf: prompt\.pdf\]\n\[Attached image\/png: prompt\.png\]\n\n\[User\]: Newest text after the retained media\./,
+        ),
         media: [
           {
             type: "file",
@@ -292,6 +306,12 @@ it.effect("manual compaction summarizes short context instead of no-op", () =>
             mime: "application/pdf",
             name: "prompt.pdf",
           },
+          {
+            type: "file",
+            uri: "data:image/png;base64,aW1hZ2U=",
+            mime: "image/png",
+            name: "prompt.png",
+          },
         ],
       },
     ])

packages/www/content/docs/(Configure)/compaction.mdx

@@ -112,10 +112,10 @@ The newest serialized context up to `keep.tokens` is retained separately. This
 is not a byte-for-byte transcript: tool output is limited to 2000 characters,
 and non-media attachments become textual descriptors. Media in the retained
 context is attached to the checkpoint in the same order as its descriptors.
-Tail selection budgets 2000 additional tokens per image as a provider-neutral
-planning estimate; it does not count base64 request bytes as text tokens. On
-later compactions, V2 updates the previous summary and carries forward its
-retained recent context before selecting a new tail.
+Tail selection budgets 1500 additional tokens per image or PDF as a
+provider-neutral planning estimate; it does not count base64 request bytes as
+text tokens. On later compactions, V2 updates the previous summary and carries
+forward its retained recent context before selecting a new tail.
 
 The completed compaction is presented to the model as historical conversation
 context, explicitly not as new instructions. Running and failed compactions are