feat: add context management and compaction features to responses API

Author: caozhiyuan

src/lib/config.ts

@@ -9,6 +9,7 @@ export interface AppConfig {
   }
   extraPrompts?: Record<string, string>
   smallModel?: string
+  responsesApiContextManagementModels?: Array<string>
   modelReasoningEfforts?: Record<
     string,
     "none" | "minimal" | "low" | "medium" | "high" | "xhigh"
@@ -54,6 +55,7 @@ const defaultConfig: AppConfig = {
     "gpt-5.4": gpt5CommentaryPrompt,
   },
   smallModel: "gpt-5-mini",
+  responsesApiContextManagementModels: ["gpt-5.4", "gpt-5.3-codex"],
   modelReasoningEfforts: {
     "gpt-5-mini": "low",
     "gpt-5.3-codex": "xhigh",
@@ -179,6 +181,19 @@ export function getSmallModel(): string {
   return config.smallModel ?? "gpt-5-mini"
 }
 
+export function getResponsesApiContextManagementModels(): Array<string> {
+  const config = getConfig()
+  return (
+    config.responsesApiContextManagementModels
+    ?? defaultConfig.responsesApiContextManagementModels
+    ?? []
+  )
+}
+
+export function isResponsesApiContextManagementModel(model: string): boolean {
+  return getResponsesApiContextManagementModels().includes(model)
+}
+
 export function getReasoningEffortForModel(
   model: string,
 ): "none" | "minimal" | "low" | "medium" | "high" | "xhigh" {

src/routes/messages/handler.ts

@@ -22,7 +22,11 @@ import {
   translateAnthropicMessagesToResponsesPayload,
   translateResponsesResultToAnthropic,
 } from "~/routes/messages/responses-translation"
-import { getResponsesRequestOptions } from "~/routes/responses/utils"
+import {
+  applyResponsesApiContextManagement,
+  compactInputByLatestCompaction,
+  getResponsesRequestOptions,
+} from "~/routes/responses/utils"
 import {
   createChatCompletions,
   type ChatCompletionChunk,
@@ -108,7 +112,10 @@ export async function handleCompletion(c: Context) {
   }
 
   if (shouldUseResponsesApi(selectedModel)) {
-    return await handleWithResponsesApi(c, anthropicPayload, initiatorOverride)
+    return await handleWithResponsesApi(c, anthropicPayload, {
+      initiatorOverride,
+      selectedModel,
+    })
   }
 
   return await handleWithChatCompletions(c, anthropicPayload, initiatorOverride)
@@ -181,10 +188,23 @@ const handleWithChatCompletions = async (
 const handleWithResponsesApi = async (
   c: Context,
   anthropicPayload: AnthropicMessagesPayload,
-  initiatorOverride?: "agent" | "user",
+  options?: {
+    initiatorOverride?: "agent" | "user"
+    selectedModel?: Model
+  },
 ) => {
+  const { initiatorOverride, selectedModel } = options ?? {}
+
   const responsesPayload =
     translateAnthropicMessagesToResponsesPayload(anthropicPayload)
+
+  applyResponsesApiContextManagement(
+    responsesPayload,
+    selectedModel?.capabilities.limits.max_prompt_tokens,
+  )
+
+  compactInputByLatestCompaction(responsesPayload)
+
   logger.debug(
     "Translated Responses payload:",
     JSON.stringify(responsesPayload),

src/routes/messages/responses-stream-translation.ts

@@ -19,6 +19,7 @@ import {
 import { type AnthropicStreamEventData } from "./anthropic-types"
 import {
   THINKING_TEXT,
+  encodeCompactionCarrierSignature,
   translateResponsesResultToAnthropic,
 } from "./responses-translation"
 
@@ -193,11 +194,45 @@ const handleOutputItemDone = (
   const events = new Array<AnthropicStreamEventData>()
   const item = rawEvent.item
   const itemType = item.type
+  const outputIndex = rawEvent.output_index
+
+  if (itemType === "compaction") {
+    if (!item.id || !item.encrypted_content) {
+      return events
+    }
+
+    const blockIndex = openThinkingBlockIfNeeded(state, outputIndex, events)
+
+    if (!state.blockHasDelta.has(blockIndex)) {
+      events.push({
+        type: "content_block_delta",
+        index: blockIndex,
+        delta: {
+          type: "thinking_delta",
+          thinking: THINKING_TEXT,
+        },
+      })
+    }
+
+    events.push({
+      type: "content_block_delta",
+      index: blockIndex,
+      delta: {
+        type: "signature_delta",
+        signature: encodeCompactionCarrierSignature({
+          id: item.id,
+          encrypted_content: item.encrypted_content,
+        }),
+      },
+    })
+    state.blockHasDelta.add(blockIndex)
+    return events
+  }
+
   if (itemType !== "reasoning") {
     return events
   }
 
-  const outputIndex = rawEvent.output_index
   const blockIndex = openThinkingBlockIfNeeded(state, outputIndex, events)
   const signature = (item.encrypted_content ?? "") + "@" + item.id
   if (signature) {

src/routes/messages/responses-translation.ts

@@ -6,6 +6,7 @@ import {
 } from "~/lib/config"
 import {
   type ResponsesPayload,
+  type ResponseInputCompaction,
   type ResponseInputContent,
   type ResponseInputImage,
   type ResponseInputItem,
@@ -14,6 +15,7 @@ import {
   type ResponseInputText,
   type ResponsesResult,
   type ResponseOutputContentBlock,
+  type ResponseOutputCompaction,
   type ResponseOutputFunctionCall,
   type ResponseOutputItem,
   type ResponseOutputReasoning,
@@ -44,6 +46,8 @@ import {
 } from "./anthropic-types"
 
 const MESSAGE_TYPE = "message"
+const COMPACTION_SIGNATURE_PREFIX = "cm1#"
+const COMPACTION_SIGNATURE_SEPARATOR = "@"
 
 export const THINKING_TEXT = "Thinking..."
 
@@ -89,6 +93,44 @@ export const translateAnthropicMessagesToResponsesPayload = (
   return responsesPayload
 }
 
+type CompactionCarrier = {
+  id: string
+  encrypted_content: string
+}
+
+export const encodeCompactionCarrierSignature = (
+  compaction: CompactionCarrier,
+): string => {
+  return `${COMPACTION_SIGNATURE_PREFIX}${compaction.encrypted_content}${COMPACTION_SIGNATURE_SEPARATOR}${compaction.id}`
+}
+
+export const decodeCompactionCarrierSignature = (
+  signature: string,
+): CompactionCarrier | undefined => {
+  if (signature.startsWith(COMPACTION_SIGNATURE_PREFIX)) {
+    const raw = signature.slice(COMPACTION_SIGNATURE_PREFIX.length)
+    const separatorIndex = raw.indexOf(COMPACTION_SIGNATURE_SEPARATOR)
+
+    if (separatorIndex <= 0 || separatorIndex === raw.length - 1) {
+      return undefined
+    }
+
+    const encrypted_content = raw.slice(0, separatorIndex)
+    const id = raw.slice(separatorIndex + 1)
+
+    if (!encrypted_content) {
+      return undefined
+    }
+
+    return {
+      id,
+      encrypted_content,
+    }
+  }
+
+  return undefined
+}
+
 const translateMessage = (
   message: AnthropicMessage,
   model: string,
@@ -165,17 +207,25 @@ const translateAssistantMessage = (
       continue
     }
 
-    if (
-      block.type === "thinking"
-      && block.signature
-      && block.signature.includes("@")
-    ) {
-      flushPendingContent(pendingContent, items, {
-        role: "assistant",
-        phase: assistantPhase,
-      })
-      items.push(createReasoningContent(block))
-      continue
+    if (block.type === "thinking" && block.signature) {
+      const compactionContent = createCompactionContent(block)
+      if (compactionContent) {
+        flushPendingContent(pendingContent, items, {
+          role: "assistant",
+          phase: assistantPhase,
+        })
+        items.push(compactionContent)
+        continue
+      }
+
+      if (block.signature.includes("@")) {
+        flushPendingContent(pendingContent, items, {
+          role: "assistant",
+          phase: assistantPhase,
+        })
+        items.push(createReasoningContent(block))
+        continue
+      }
     }
 
     const converted = translateAssistantContentBlock(block)
@@ -302,15 +352,43 @@ const createReasoningContent = (
   // align with vscode-copilot-chat extractThinkingData, should add id, otherwise it will cause miss cache occasionally —— the usage input cached tokens to be 0
   // https://github.com/microsoft/vscode-copilot-chat/blob/main/src/platform/endpoint/node/responsesApi.ts#L162
   // when use in codex cli, reasoning id is empty, so it will cause miss cache occasionally
-  const array = block.signature.split("@")
-  const signature = array[0]
-  const id = array[1]
+  const { encryptedContent, id } = parseReasoningSignature(block.signature)
   const thinking = block.thinking === THINKING_TEXT ? "" : block.thinking
   return {
     id,
     type: "reasoning",
     summary: thinking ? [{ type: "summary_text", text: thinking }] : [],
-    encrypted_content: signature,
+    encrypted_content: encryptedContent,
+  }
+}
+
+const createCompactionContent = (
+  block: AnthropicThinkingBlock,
+): ResponseInputCompaction | undefined => {
+  const compaction = decodeCompactionCarrierSignature(block.signature)
+  if (!compaction) {
+    return undefined
+  }
+
+  return {
+    id: compaction.id,
+    type: "compaction",
+    encrypted_content: compaction.encrypted_content,
+  }
+}
+
+const parseReasoningSignature = (
+  signature: string,
+): { encryptedContent: string; id: string } => {
+  const splitIndex = signature.lastIndexOf("@")
+
+  if (splitIndex <= 0 || splitIndex === signature.length - 1) {
+    return { encryptedContent: signature, id: "" }
+  }
+
+  return {
+    encryptedContent: signature.slice(0, splitIndex),
+    id: signature.slice(splitIndex + 1),
   }
 }
 
@@ -456,6 +534,13 @@ const mapOutputToAnthropicContent = (
         }
         break
       }
+      case "compaction": {
+        const compactionBlock = createCompactionThinkingBlock(item)
+        if (compactionBlock) {
+          contentBlocks.push(compactionBlock)
+        }
+        break
+      }
       default: {
         // Future compatibility for unrecognized output item types.
         const combinedText = combineMessageTextContent(
@@ -549,6 +634,23 @@ const createToolUseContentBlock = (
   }
 }
 
+const createCompactionThinkingBlock = (
+  item: ResponseOutputCompaction,
+): AnthropicAssistantContentBlock | null => {
+  if (!item.id || !item.encrypted_content) {
+    return null
+  }
+
+  return {
+    type: "thinking",
+    thinking: THINKING_TEXT,
+    signature: encodeCompactionCarrierSignature({
+      id: item.id,
+      encrypted_content: item.encrypted_content,
+    }),
+  }
+}
+
 const parseFunctionCallArguments = (
   rawArguments: string,
 ): Record<string, unknown> => {

src/routes/responses/handler.ts

@@ -14,7 +14,11 @@ import {
 } from "~/services/copilot/create-responses"
 
 import { createStreamIdTracker, fixStreamIds } from "./stream-id-sync"
-import { getResponsesRequestOptions } from "./utils"
+import {
+  applyResponsesApiContextManagement,
+  compactInputByLatestCompaction,
+  getResponsesRequestOptions,
+} from "./utils"
 
 const logger = createHandlerLogger("responses-handler")
 
@@ -31,6 +35,8 @@ export const handleResponses = async (c: Context) => {
   // Remove web_search tool as it's not supported by GitHub Copilot
   removeWebSearchTool(payload)
 
+  compactInputByLatestCompaction(payload)
+
   const selectedModel = state.models?.data.find(
     (model) => model.id === payload.model,
   )
@@ -50,6 +56,13 @@ export const handleResponses = async (c: Context) => {
     )
   }
 
+  applyResponsesApiContextManagement(
+    payload,
+    selectedModel?.capabilities.limits.max_prompt_tokens,
+  )
+
+  logger.debug("Translated Responses payload:", JSON.stringify(payload))
+
   const { vision, initiator } = getResponsesRequestOptions(payload)
 
   if (state.manualApprove) {