feat: add native OpenAI Responses compaction

.changeset/preserve-openai-compaction-state.md

@@ -0,0 +1,10 @@
+---
+"@moonshot-ai/kimi-code": patch
+"@moonshot-ai/agent-core": patch
+"@moonshot-ai/agent-core-v2": patch
+"@moonshot-ai/kosong": patch
+---
+
+Preserve opaque OpenAI Responses compaction state across turns and automatically
+use `/responses/compact` when the active provider exposes that capability,
+falling back to Kimi's existing local summarizer when it does not.

apps/kimi-code/src/tui/utils/message-replay.ts

@@ -171,6 +171,7 @@ export function collectReplayMessageContent(
         break;
       case 'audio_url':
       case 'image_url':
+      case 'openai_compaction':
       case 'video_url':
         break;
     }
@@ -285,6 +286,8 @@ function contentPartToText(part: ContentPart): string {
       return mediaUrlPartToText('video', part.videoUrl.url);
     case 'audio_url':
       return mediaUrlPartToText('audio', part.audioUrl.url);
+    case 'openai_compaction':
+      return '';
   }
 }
 

apps/vscode/src/runtime/replay-adapter.ts

@@ -511,6 +511,8 @@ function toLegacyContent(content: readonly ContentPart[]): LegacyContentPart[] {
       case "video_url":
         result.push({ type: "video_url", video_url: { ...part.videoUrl } });
         break;
+      case "openai_compaction":
+        break;
     }
   }
   return result;

apps/vscode/src/utils/session-context.ts

@@ -196,6 +196,7 @@ function formatPartMarkdown(part: ContentPart): string {
     case "image_url": return "[image]";
     case "audio_url": return "[audio]";
     case "video_url": return "[video]";
+    case "openai_compaction": return "";
   }
 }
 
@@ -205,7 +206,8 @@ function stringifyParts(parts: readonly ContentPart[]): string {
     if (part.type === "think") return part.think.trim() ? `<thinking>\n${part.think}\n</thinking>` : "";
     if (part.type === "image_url") return "[image]";
     if (part.type === "audio_url") return "[audio]";
-    return "[video]";
+    if (part.type === "video_url") return "[video]";
+    return "";
   }).filter(Boolean).join("\n");
 }
 

docs/en/guides/sessions.md

@@ -68,6 +68,12 @@ You can manage sessions without leaving the terminal. The following slash comman
 
 As a conversation grows, Kimi Code CLI automatically compresses the message history when the context approaches the window limit, freeing up token space. You can also trigger compression manually at any time:
 
+For OpenAI Responses-compatible providers, Kimi Code automatically tries the
+provider's native `/responses/compact` endpoint and preserves its opaque
+replacement state. If that capability is unavailable, or when `/compact`
+includes a custom instruction, it uses Kimi Code's existing summary compaction
+instead. No separate setting is required.
+
 ```
 /compact
 ```

docs/zh/guides/sessions.md

@@ -68,6 +68,10 @@ kimi --session
 
 对话变长时，Kimi Code CLI 会在上下文接近窗口上限时自动压缩历史消息，释放 token 空间。也可以随时手动触发：
 
+对于兼容 OpenAI Responses 的提供商，Kimi Code 会自动尝试其原生
+`/responses/compact` 接口，并保留接口返回的不透明替换状态。如果该能力不可用，
+或 `/compact` 带有自定义指引，则回退到 Kimi Code 原有的摘要压缩。无需额外配置。
+
 ```
 /compact
 ```

packages/agent-core-v2/docs/wire-manifest.d.ts

@@ -114,7 +114,7 @@ interface ContextAppendMessagePayload {
   message: {
     role: 'system' | 'user' | 'assistant' | 'tool';
     name?: string;
-    content: ('text' | 'think' | 'image_url' | 'audio_url' | 'video_url')[];
+    content: ('text' | 'think' | 'image_url' | 'audio_url' | 'video_url' | 'openai_compaction')[];
     toolCalls: {
       type: 'function';
       id: string;

packages/agent-core-v2/src/agent/contextMemory/compactionHandoff.ts

@@ -24,6 +24,7 @@ export interface CompactionUserSelection<T> {
 
 export interface ContextCompactionShapeInput {
   readonly summary: string;
+  readonly replacementMessages?: readonly ContextMessage[];
   readonly legacySummaryMessage?: ContextMessage;
   readonly contextSummary?: string;
   readonly compactedCount: number;
@@ -51,6 +52,22 @@ export function buildContextCompactionShape(
   history: readonly ContextMessage[],
   input: ContextCompactionShapeInput,
 ): ContextCompactionShape {
+  if (input.replacementMessages !== undefined) {
+    const messages = [...input.replacementMessages];
+    const contextSummary = input.contextSummary ?? input.summary;
+    return {
+      summary: input.summary,
+      contextSummary,
+      compactedCount: input.compactedCount,
+      tokensBefore: input.tokensBefore,
+      tokensAfter: input.tokensAfter ?? estimateTokensForMessages(messages),
+      keptUserMessageCount:
+        input.keptUserMessageCount ?? messages.filter((message) => message.role === 'user').length,
+      keptHeadUserMessageCount: input.keptHeadUserMessageCount,
+      droppedCount: input.droppedCount,
+      messages,
+    };
+  }
   if (usesLegacyTailShape(input)) {
     const contextSummary = input.contextSummary ?? input.summary;
     const messages = [

packages/agent-core-v2/src/agent/contextMemory/contextMemory.ts

@@ -6,6 +6,8 @@ import type { ContextMessage } from './types';
 
 export interface ContextCompactionInput {
   readonly summary: string;
+  /** Provider-owned canonical replacement window, when native compaction is used. */
+  readonly replacementMessages?: readonly ContextMessage[];
   readonly contextSummary?: string;
   readonly compactedCount: number;
   readonly tokensBefore: number;

packages/agent-core-v2/src/agent/contextMemory/contextMemoryService.ts

@@ -113,6 +113,8 @@ export class AgentContextMemoryService extends Disposable implements IAgentConte
     this.wire.dispatch(
       contextApplyCompaction({
         summary: result.summary,
+        replacementMessages:
+          input.replacementMessages === undefined ? undefined : [...input.replacementMessages],
         contextSummary: result.contextSummary,
         compactedCount: result.compactedCount,
         tokensBefore: result.tokensBefore,

packages/agent-core-v2/src/agent/contextMemory/contextOps.ts

@@ -156,6 +156,7 @@ const contextCompactionBaseShape = {
   keptHeadUserMessageCount: z.number().optional(),
   droppedCount: z.number().optional(),
   legacyTail: z.boolean().optional(),
+  replacementMessages: z.array(contextMessageSchema).optional(),
 };
 
 const contextApplyCompactionSchema = z.union([
@@ -210,6 +211,7 @@ export function readContextCompactionShapeInput(
   const keptUserMessageCount = readOptionalNumber(fields, 'keptUserMessageCount');
   return {
     summary: readContextCompactionRawSummary(fields),
+    replacementMessages: readReplacementMessages(fields),
     legacySummaryMessage: readLegacySummaryMessage(fields),
     contextSummary: readOptionalString(fields, 'contextSummary'),
     compactedCount: readContextCompactedCount(fields),
@@ -222,6 +224,11 @@ export function readContextCompactionShapeInput(
   };
 }
 
+function readReplacementMessages(record: UnknownRecord): readonly ContextMessage[] | undefined {
+  const value = record['replacementMessages'];
+  return Array.isArray(value) ? (value as ContextMessage[]) : undefined;
+}
+
 export function readContextCompactedCount(record: ContextCompactionRecord): number {
   const fields = record as UnknownRecord;
   const compactedCount = fields['compactedCount'];

packages/agent-core-v2/src/agent/contextMemory/messageProjection.ts

@@ -36,7 +36,7 @@ function toProtocolRole(role: ContextMessage['role']): MessageRole {
   return role as MessageRole;
 }
 
-function mapContentPart(part: ContextMessage['content'][number]): MessageContent {
+function mapContentPart(part: ContextMessage['content'][number]): MessageContent | undefined {
   switch (part.type) {
     case 'text':
       return { type: 'text', text: part.text };
@@ -59,13 +59,22 @@ function mapContentPart(part: ContextMessage['content'][number]): MessageContent
         ? { type: 'video', source: { kind: 'file', file_id: ref.fileId } }
         : { type: 'video', source: { kind: 'url', url: part.videoUrl.url, id: part.videoUrl.id } };
     }
+    case 'openai_compaction':
+      return undefined;
   }
 }
 
+function mapContentParts(parts: ContextMessage['content']): MessageContent[] {
+  return parts.flatMap((part) => {
+    const mapped = mapContentPart(part);
+    return mapped === undefined ? [] : [mapped];
+  });
+}
+
 function buildProtocolContent(msg: ContextMessage): MessageContent[] {
   if (msg.role === 'tool') {
     if (msg.toolCallId === undefined) {
-      return msg.content.map((p) => mapContentPart(p));
+      return mapContentParts(msg.content);
     }
     const hasMediaPart = msg.content.some(
       (p) => p.type === 'image_url' || p.type === 'video_url' || p.type === 'audio_url',
@@ -89,7 +98,7 @@ function buildProtocolContent(msg: ContextMessage): MessageContent[] {
     return [part];
   }
 
-  const base = msg.content.map((p) => mapContentPart(p));
+  const base = mapContentParts(msg.content);
 
   if (msg.role === 'assistant' && msg.toolCalls.length > 0) {
     for (const call of msg.toolCalls) {

packages/agent-core-v2/src/agent/fullCompaction/fullCompactionService.ts

@@ -74,6 +74,7 @@ const OVERFLOW_CONTEXT_SAFETY_RATIO = 0.85;
 const OVERFLOW_STATUS_RECOVERY_RATIO = 0.5;
 const MAX_COMPACTION_OVERFLOW_SHRINK_ATTEMPTS = 3;
 const COMPACTION_OVERFLOW_SHRINK_RATIOS = [0.7, 0.5, 0.35] as const;
+const REMOTE_COMPACTION_SUMMARY = '[OpenAI server compaction checkpoint]';
 const EMPTY_TOOL_PARAMETERS: Record<string, unknown> = {
   type: 'object',
   properties: {},
@@ -115,6 +116,7 @@ export class AgentFullCompactionService extends Disposable implements IAgentFull
   private compactionCountInTurn = 0;
   private _compacting: ActiveCompaction | null = null;
   private readonly observedMaxContextTokensByModel = new Map<string, number>();
+  private readonly remoteCompactionUnavailableModels = new Set<string>();
   private lastCompactedTokenCount: number | null = null;
   private consecutiveOverflowCompactions = 0;
   private activeTurnId: number | undefined;
@@ -537,6 +539,71 @@ export class AgentFullCompactionService extends Disposable implements IAgentFull
 
       const resolvedModel = this.profile.resolveModelContext();
       thinkingEffort = resolvedModel.thinkingLevel;
+
+      const modelAlias = resolvedModel.modelAlias;
+      const hasCustomInstruction = (data.instruction?.trim().length ?? 0) > 0;
+      // Provider-owned compaction has no portable custom-instruction field.
+      if (!hasCustomInstruction && !this.remoteCompactionUnavailableModels.has(modelAlias)) {
+        try {
+          const remote = await this.llmRequester.compact?.(
+            {
+              messages: stripDynamicToolContext(originalHistory),
+              source: {
+                type: 'operation',
+                turnId: active.originTurnId,
+                requestKind: 'remote_compaction',
+              },
+            },
+            signal,
+          );
+          if (remote !== undefined) {
+            if (!historySafeToCompact(this.context.get(), originalHistory)) {
+              this.cancelActive(active);
+              throw compactionCancelledReason(active);
+            }
+            const result = this.context.applyCompaction({
+              summary: REMOTE_COMPACTION_SUMMARY,
+              contextSummary: REMOTE_COMPACTION_SUMMARY,
+              replacementMessages: remote.messages as readonly ContextMessage[],
+              compactedCount: originalHistory.length,
+              tokensBefore,
+            });
+            const properties: CompactionFinishedEvent = {
+              turn_id: active.originTurnId,
+              source: data.source,
+              tokens_before: result.tokensBefore,
+              tokens_after: result.tokensAfter,
+              duration_ms: Date.now() - startedAt,
+              compacted_count: result.compactedCount,
+              retry_count: 0,
+              round: 1,
+              thinking_effort: thinkingEffort,
+              ...usageTelemetry(remote.usage ?? null),
+            };
+            this.telemetry.track2('compaction_finished', properties);
+            return result;
+          }
+          this.remoteCompactionUnavailableModels.add(modelAlias);
+        } catch (error) {
+          if (isAbortError(error)) throw error;
+          if (
+            isError2(error) &&
+            (error.code === ErrorCodes.AUTH_LOGIN_REQUIRED ||
+              error.code === ErrorCodes.PROVIDER_AUTH_ERROR)
+          ) {
+            throw error;
+          }
+          const status = findAPIStatusError(error)?.statusCode;
+          if (status === 400 || status === 404 || status === 405 || status === 501) {
+            this.remoteCompactionUnavailableModels.add(modelAlias);
+          }
+          this.log.warn('remote compaction unavailable; falling back to local summary', {
+            model: modelAlias,
+            status,
+          });
+        }
+      }
+
       const maxContextTokens = resolvedModel.modelCapabilities.max_context_tokens;
       const defaultCompactionCap =
         maxContextTokens > 0

packages/agent-core-v2/src/agent/llmRequester/llmRequester.ts

@@ -1,5 +1,9 @@
 import { createDecorator } from '#/_base/di/instantiation';
-import type { FinishReason, ThinkingEffort } from '#/kosong/contract/provider';
+import type {
+  FinishReason,
+  ProviderCompactionResult,
+  ThinkingEffort,
+} from '#/kosong/contract/provider';
 import type { Message, StreamedMessagePart } from '#/kosong/contract/message';
 import type { Tool } from '#/kosong/contract/tool';
 import type { TokenUsage } from '#/kosong/contract/usage';
@@ -70,6 +74,11 @@ export interface IAgentLLMRequesterService {
     onPart?: AgentLLMRequestPartHandler,
     signal?: AbortSignal,
   ): AgentLLMRequestTask;
+
+  compact?(
+    overrides?: AgentLLMRequestOverrides,
+    signal?: AbortSignal,
+  ): Promise<ProviderCompactionResult | undefined>;
 }
 
 export const IAgentLLMRequesterService = createDecorator<IAgentLLMRequesterService>(

packages/agent-core-v2/src/agent/llmRequester/llmRequesterService.ts

@@ -55,7 +55,10 @@ import {
   isRetryableGenerateError,
 } from '#/kosong/contract/errors';
 import { type Message } from '#/kosong/contract/message';
-import { type ThinkingEffort } from '#/kosong/contract/provider';
+import type {
+  ProviderCompactionResult,
+  ThinkingEffort,
+} from '#/kosong/contract/provider';
 import { type Tool } from '#/kosong/contract/tool';
 import { emptyUsage, inputTotal, type TokenUsage } from '#/kosong/contract/usage';
 import { ILogService, type LogContext } from '#/_base/log/log';
@@ -194,6 +197,23 @@ export class AgentLLMRequesterService implements IAgentLLMRequesterService {
     };
   }
 
+  async compact(
+    overrides: AgentLLMRequestOverrides = {},
+    signal?: AbortSignal,
+  ): Promise<ProviderCompactionResult | undefined> {
+    const request = this.resolveRequest(overrides);
+    const input = {
+      systemPrompt: request.systemPrompt,
+      tools: request.tools,
+      messages: request.messages,
+    };
+    const result = await request.requester.compact?.(input, signal, request.params);
+    if (result?.usage !== undefined) {
+      this.usage.record(request.modelAlias, result.usage, request.source);
+    }
+    return result;
+  }
+
   private async requestWithTrace(
     trace: MutableLLMRequestTrace,
     overrides: AgentLLMRequestOverrides,

packages/agent-core-v2/src/agent/loop/loopService.ts

@@ -948,6 +948,7 @@ export class AgentLoopService extends Disposable implements IAgentLoopService {
         case 'image_url':
         case 'audio_url':
         case 'video_url':
+        case 'openai_compaction':
           return;
         case 'function': {
           onResponseEvent();

packages/agent-core-v2/src/agent/mcp/output.ts

@@ -246,7 +246,11 @@ function applyBinaryPartCap(parts: readonly ContentPart[]): {
   const out: ContentPart[] = [];
 
   for (const part of parts) {
-    if (part.type === 'text' || part.type === 'think') {
+    if (
+      part.type === 'text' ||
+      part.type === 'think' ||
+      part.type === 'openai_compaction'
+    ) {
       out.push(part);
       continue;
     }