feat(agent-core-v2): anchor compaction resumption on the latest user message

.changeset/compaction-resume-anchor.md

@@ -0,0 +1,5 @@
+---
+"@moonshot-ai/kimi-code": patch
+---
+
+Fix the agent resuming the wrong request after automatic context compaction in long sessions.

apps/vis/server/src/lib/context-projector.ts

@@ -346,7 +346,9 @@ export function projectContext(
           },
         };
         const legacyTail = rec.legacyTail === true || rec.keptUserMessageCount === undefined;
-        const summaryIndex = legacyTail ? 0 : shape.messages.length - 1;
+        const summaryIndex = legacyTail
+          ? 0
+          : shape.messages.findIndex((message) => message.origin?.kind === 'compaction_summary');
         const modelSummaryBubble: ProjectedMessage = {
           ...summaryBubble,
           message: modelFacingMessage(shape.messages[summaryIndex] ?? summaryBubble.message),

apps/vis/server/test/lib/context-projector.test.ts

@@ -1,6 +1,7 @@
 // apps/vis/server/test/lib/context-projector.test.ts
 import { describe, it, expect, afterEach } from 'vitest';
 import { estimateTokensForMessages } from '@moonshot-ai/agent-core-v2/llm-adapter/contract/tokens';
+import { buildCompactionContinuationText } from '@moonshot-ai/agent-core-v2/agent/contextMemory/compactionHandoff';
 import { buildSessionFixture } from '../fixtures/build';
 import { projectContext } from '../../src/lib/context-projector';
 import { readAgentWire } from '../../src/lib/wire-reader';
@@ -440,16 +441,23 @@ describe('context-projector', () => {
           keptUserMessageCount: 2 }, raw: {} },
     ];
     const proj = projectContext(entries as any);
-    // [m0, m1, summary] — real user prompts are kept verbatim, the assistant
-    // tail is dropped.
-    expect(proj.messages).toHaveLength(3);
+    // [m0, m1, summary, anchor] — real user prompts are kept verbatim, the
+    // assistant tail is dropped, and the continuation anchor follows the summary.
+    expect(proj.messages).toHaveLength(4);
     expect(proj.messages.map((m) => m.source)).toEqual([
-      'append_message', 'append_message', 'compaction_summary',
+      'append_message', 'append_message', 'compaction_summary', 'append_message',
     ]);
     expect(proj.messages[0]!.message.content[0]).toMatchObject({ text: 'm0' });
     expect(proj.messages[1]!.message.content[0]).toMatchObject({ text: 'm1' });
     expect(proj.messages[2]!.compaction).toEqual({ compactedCount: 3, tokensBefore: 100, tokensAfter: 10 });
     expect(proj.messages[2]!.message.content[0]).toMatchObject({ text: 'sum' });
+    expect(proj.messages[3]!.message.origin).toEqual({
+      kind: 'injection',
+      variant: 'compaction_continuation',
+    });
+    expect(proj.messages[3]!.message.content[0]).toMatchObject({
+      text: buildCompactionContinuationText(),
+    });
   });
 
   it('apply_compaction mirrors the legacy verbatim tail for records without keptUserMessageCount (model)', () => {
@@ -499,9 +507,10 @@ describe('context-projector', () => {
     ];
 
     const proj = projectContext(entries as any);
-    // [FIRST, head slice of middle, marker, tail slice of middle, LAST, summary]
-    // — mirrors the engine's selectCompactionUserMessages + elision marker.
-    expect(proj.messages).toHaveLength(6);
+    // [FIRST, head slice of middle, marker, tail slice of middle, LAST, summary, anchor]
+    // — mirrors the engine's selectCompactionUserMessages + elision marker, with
+    // the continuation anchor after the summary.
+    expect(proj.messages).toHaveLength(7);
     const texts = proj.messages.map((m) =>
       m.message.content.map((p: any) => (p.type === 'text' ? p.text : '')).join(''),
     );
@@ -517,9 +526,14 @@ describe('context-projector', () => {
     expect(middle.endsWith(texts[3]!)).toBe(true);
     expect(texts[4]).toBe(last);
     expect(proj.messages[5]!.source).toBe('compaction_summary');
+    expect(proj.messages[6]!.message.origin).toEqual({
+      kind: 'injection',
+      variant: 'compaction_continuation',
+    });
+    expect(texts[6]).toBe(buildCompactionContinuationText());
     // Synthesized entries (the head slice of the same message that anchors the
     // tail, and the marker) get fractional lineNos so keys stay unique.
-    expect(new Set(proj.messages.map((m) => m.lineNo)).size).toBe(6);
+    expect(new Set(proj.messages.map((m) => m.lineNo)).size).toBe(7);
   });
 
   it('apply_compaction drops shell/local-command/background messages in model mode only', () => {
@@ -543,10 +557,10 @@ describe('context-projector', () => {
 
     const model = projectContext(entries as any);
     expect(model.messages.map((m) => m.source)).toEqual([
-      'append_message', 'compaction_summary', 'append_message',
+      'append_message', 'compaction_summary', 'append_message', 'append_message',
     ]);
     expect(model.messages.map((m) => m.message.content[0])).toMatchObject([
-      { text: 'real user' }, { text: 'sum' }, { text: 'new' },
+      { text: 'real user' }, { text: 'sum' }, { text: buildCompactionContinuationText() }, { text: 'new' },
     ]);
 
     const full = projectContext(entries as any, 'full');
@@ -591,12 +605,14 @@ describe('context-projector', () => {
           keptUserMessageCount: 3 }, raw: {} },
     ];
     const proj = projectContext(entries as any);
-    // Correct: [u1, u3, u4, summary]. The marker is gone, all real prompts kept.
+    // Correct: [u1, u3, u4, summary, anchor]. The marker is gone, all real
+    // prompts kept, and the continuation anchor follows the summary.
     expect(proj.messages.map((m) => m.source)).toEqual([
-      'append_message', 'append_message', 'append_message', 'compaction_summary',
+      'append_message', 'append_message', 'append_message', 'compaction_summary', 'append_message',
     ]);
     expect(proj.messages.map((m) => m.message.content[0])).toMatchObject([
       { text: 'u1' }, { text: 'u3' }, { text: 'u4' }, { text: 'sum' },
+      { text: buildCompactionContinuationText() },
     ]);
   });
 
@@ -994,10 +1010,10 @@ describe('context-projector', () => {
           keptUserMessageCount: 2 }, raw: {} },
     ];
     // No 2nd arg → 'model' default: the real user prompts are kept verbatim and
-    // the summary is appended after them.
+    // the summary is appended after them, followed by the continuation anchor.
     const proj = projectContext(entries as any);
     expect(proj.messages.map((m) => m.source)).toEqual([
-      'append_message', 'append_message', 'compaction_summary',
+      'append_message', 'append_message', 'compaction_summary', 'append_message',
     ]);
     expect(proj.messages[0]!.message.content[0]).toMatchObject({ text: 'm0' });
     expect(proj.messages[1]!.message.content[0]).toMatchObject({ text: 'm1' });

apps/vis/server/test/routes/context.test.ts

@@ -1,4 +1,5 @@
 import { describe, it, expect, afterEach } from 'vitest';
+import { buildCompactionContinuationText } from '@moonshot-ai/agent-core-v2/agent/contextMemory/compactionHandoff';
 import { buildSessionFixture } from '../fixtures/build';
 import { contextRoute } from '../../src/routes/context';
 
@@ -77,10 +78,11 @@ describe('context route', () => {
       messages: { source: string; message: { content: { type: string; text?: string }[] } }[];
     };
     expect(modelBody.messages.map((m) => m.source)).toEqual([
-      'append_message', 'compaction_summary', 'append_message',
+      'append_message', 'compaction_summary', 'append_message', 'append_message',
     ]);
     expect(modelBody.messages[0]!.message.content[0]).toMatchObject({ text: 'before compaction' });
-    expect(modelBody.messages[2]!.message.content[0]).toMatchObject({ text: 'after compaction' });
+    expect(modelBody.messages[2]!.message.content[0]).toMatchObject({ text: buildCompactionContinuationText() });
+    expect(modelBody.messages[3]!.message.content[0]).toMatchObject({ text: 'after compaction' });
 
     // Full history: every pre-compaction message (user prompt + assistant reply)
     // is KEPT, then the summary marker, then the post-compaction tail.

packages/agent-core-v2/src/agent/contextMemory/compaction-summary-prefix.md

@@ -1 +1 @@
-The conversation so far has been compacted to free up context. What follows is your own working summary of this task — use it to continue your train of thought rather than starting over. Treat it as notes, not proof: where it says a step was done, tests passed, or a fix worked, verify that yourself before relying on it. Any user messages earlier in this context are preserved verbatim from the compacted conversation; where a system-reminder note among them marks an omitted middle section, the user messages it replaced are covered by this summary.
+The conversation so far has been compacted to free up context. What follows is your own working summary of this task — use it to continue your train of thought rather than starting over. Treat it as notes, not proof: where it says a step was done, tests passed, or a fix worked, verify that yourself before relying on it. Any user messages earlier in this context are preserved verbatim from the compacted conversation; where a system-reminder note among them marks an omitted middle section, the user messages it replaced are covered by this summary. The summary records which earlier requests were already addressed.

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

@@ -8,6 +8,7 @@ export const COMPACTION_SUMMARY_PREFIX = summaryPrefixTemplate.trimEnd();
 export const COMPACT_USER_MESSAGE_MAX_TOKENS = 20_000;
 export const COMPACT_USER_MESSAGE_HEAD_TOKENS = 2_000;
 export const COMPACTION_ELISION_VARIANT = 'compaction_elision';
+export const COMPACTION_CONTINUATION_VARIANT = 'compaction_continuation';
 
 type MessageLike = ContextMessage;
 
@@ -94,11 +95,12 @@ export function buildContextCompactionShape(
     ? [...selection.head, ...selection.tail]
     : [...selection.head, elisionMessage, ...selection.tail];
   const contextSummary = input.contextSummary ?? input.summary;
+  const continuationMessage = createCompactionContinuationMessage();
   const tokensAfter =
     input.tokensAfter ??
     (input.requestOverheadTokens ?? 0) +
       (input.summaryOutputTokens ?? estimate.text(contextSummary)) +
-      estimate.messages(keptMessages);
+      estimate.messages([...keptMessages, continuationMessage]);
   const keptUserMessageCount =
     input.keptUserMessageCount ?? selection.head.length + selection.tail.length;
   const keptHeadUserMessageCount =
@@ -113,7 +115,11 @@ export function buildContextCompactionShape(
     keptUserMessageCount,
     keptHeadUserMessageCount,
     droppedCount: input.droppedCount,
-    messages: [...keptMessages, createCompactionSummaryMessage(contextSummary)],
+    messages: [
+      ...keptMessages,
+      createCompactionSummaryMessage(contextSummary),
+      continuationMessage,
+    ],
   };
 }
 
@@ -146,6 +152,21 @@ export function buildCompactionElisionText(omittedTokens: number): string {
   );
 }
 
+export function createCompactionContinuationMessage(): ContextMessage {
+  return {
+    role: 'user',
+    content: [{ type: 'text', text: buildCompactionContinuationText() }],
+    toolCalls: [],
+    origin: { kind: 'injection', variant: COMPACTION_CONTINUATION_VARIANT },
+  };
+}
+
+export function buildCompactionContinuationText(): string {
+  return wrapSystemReminder(
+    'Context compaction is complete — continue the work that was in progress when it began.',
+  );
+}
+
 export function collectCompactableUserMessages<T extends MessageLike>(messages: readonly T[]): T[] {
   return messages.filter(
     (message) => isRealUserInput(message) && !isCompactionSummaryMessage(message),

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

@@ -190,7 +190,7 @@ function recoverFoldedLength(
   const keptHeadUserMessageCount = readNumber(record, 'keptHeadUserMessageCount');
   const compactedCount = readNumber(record, 'compactedCount');
   if (keptUserMessageCount !== undefined) {
-    return keptUserMessageCount + (keptHeadUserMessageCount === undefined ? 1 : 2);
+    return keptUserMessageCount + (keptHeadUserMessageCount === undefined ? 2 : 3);
   }
   if (compactedCount !== undefined && compactedCount < foldedLength) {
     return 1 + (foldedLength - compactedCount);

packages/agent-core-v2/src/agent/fullCompaction/compaction-instruction.md

@@ -1,24 +1,20 @@
-You are about to run out of context. Write a first-person handoff note to
-yourself so you can seamlessly continue this task after the earlier
-conversation is cleared.
+You are about to run out of context. Create a handoff summary for the
+model that will resume this task after the earlier conversation is cleared.
 
 --- This message is a direct task, not part of the above conversation ---
 
-Write the note as your own continuing train of thought — first person, present
-tense, the way you would reason through the next move. Do not write a
-third-party report about someone else's work, and do not impose rigid section
-headings; let the shape follow the task. Write the note in the same language the
-conversation has been using — do not switch to English just because these
-instructions happen to be in English.
+Do not impose rigid section headings; let the shape follow the task. Write it
+in the same language the conversation has been using — do not switch to English
+just because these instructions happen to be in English.
 
-Make the note self-sufficient: the next turn will see only your most recent user
-messages and this note — every assistant message, tool call, and tool result
-above will be gone. In your own words, preserve what you genuinely need to
-continue:
+Make the summary self-sufficient: the next turn will see only the preserved
+messages and this summary — every other assistant message, tool call, and tool
+result above will be gone. In your own words, preserve what you genuinely need
+to continue:
 
 - What the latest request is actually asking for: your reading of its intent and
   any ambiguity you have already resolved — not a re-transcription, since what
-  fits is kept verbatim in your most recent messages. But those kept messages are
+  fits is kept verbatim in the preserved messages. But those kept messages are
   size-capped, so a long request is truncated there: if the latest request is
   large (a big paste or file), preserve the parts at risk of being dropped —
   above all the actual ask. If several requests are in play, say which one governs
@@ -52,9 +48,9 @@ continue:
   here is one less thing the next turn must rediscover. Include any required
   format for the final answer.
 
-This conversation's event log stays on disk and a recovery pointer is appended below your note automatically, so you need not reproduce long outputs verbatim — keep exact identifiers, key values and error lines, and name anything the next turn should look up.
+This conversation's event log stays on disk and a recovery pointer is appended below this summary automatically, so you need not reproduce long outputs verbatim — keep exact identifiers, key values and error lines, and name anything the next turn should look up.
 
-Your TODO list is re-attached automatically below this note from its live
+Your TODO list is re-attached automatically below this summary from its live
 source, so do not transcribe it — copying it wastes space and can contradict the
 live version. What that list cannot hold is the reasoning between tasks — why one
 was reordered or dropped, or a decision on one that constrains another — so
@@ -65,9 +61,9 @@ was never verified (tests "passing", a fix "working", a file "created"), say so
 plainly and treat it as unverified rather than fact — re-check before relying
 on it.
 
-Be concise, and keep the note proportional to the task: a long multi-step task
-warrants detail, but a trivial or nearly finished exchange needs only a sentence
-or two — do not pad it out. Include the critical data, identifiers, and
+Be concise, and keep the summary proportional to the task: a long multi-step
+task warrants detail, but a trivial or nearly finished exchange needs only a
+sentence or two — do not pad it out. Include the critical data, identifiers, and
 references needed to continue, and omit anything that does not change the next
 move.
 

packages/agent-core-v2/test/agent/contextMemory/context.test.ts

@@ -788,8 +788,9 @@ describe('Agent context', () => {
       );
 
       expect(shape.tokensAfter).toBe(0);
-      expect(shape.messages.map((m) => m.role)).toEqual(['user', 'user']);
+      expect(shape.messages.map((m) => m.role)).toEqual(['user', 'user', 'user']);
       expect(shape.messages[1]?.origin?.kind).toBe('compaction_summary');
+      expect(shape.messages[2]?.origin).toEqual({ kind: 'injection', variant: 'compaction_continuation' });
     });
 
     it('prefers the measured summary output tokens over the text estimate', () => {

packages/agent-core-v2/test/agent/contextMemory/contextTranscript.test.ts

@@ -109,7 +109,7 @@ describe('reduceContextTranscript', () => {
       compaction('SUM', 3, 1),
       appendMessage(userMessage('u4')),
     ]);
-    expect(result.foldedLength).toBe(3);
+    expect(result.foldedLength).toBe(4);
   });
 
   it('accounts for the elision marker when the record kept a head segment', () => {
@@ -119,7 +119,7 @@ describe('reduceContextTranscript', () => {
       ...assistantStep('s1', 'a1'),
       compaction('SUM', 3, 2, 1),
     ]);
-    expect(result.foldedLength).toBe(4);
+    expect(result.foldedLength).toBe(5);
   });
 
   it('carries the originating wire record time per entry', () => {
@@ -159,7 +159,7 @@ describe('reduceContextTranscript', () => {
     ]);
     expect(texts(result)).toEqual(['message A', 'reply A', 'summary text']);
     expect(result.entries.map((m) => m.role)).toEqual(['user', 'assistant', 'user']);
-    expect(result.foldedLength).toBe(2);
+    expect(result.foldedLength).toBe(3);
   });
 
   it('undo without compaction keeps the earlier exchange intact', () => {
@@ -388,9 +388,10 @@ describe('live fold parity', () => {
     ];
     const live = foldLive(records);
     const transcript = reduceContextTranscript(records);
-    expect(live).toHaveLength(5);
+    expect(live).toHaveLength(6);
     expect(transcript.foldedLength).toBe(live.length);
     expect(live[2]!.origin).toEqual({ kind: 'compaction_summary' });
+    expect(live[3]!.origin).toEqual({ kind: 'injection', variant: 'compaction_continuation' });
   });
 
   it('settles a frame left open by a failed attempt when compaction lands mid-fold', () => {
@@ -403,7 +404,7 @@ describe('live fold parity', () => {
     ];
     const live = foldLive(records);
     const transcript = reduceContextTranscript(records);
-    expect(live.map((m) => m.role)).toEqual(['user', 'user', 'assistant']);
+    expect(live.map((m) => m.role)).toEqual(['user', 'user', 'user', 'assistant']);
     expect(texts(transcript)).toEqual(['u1', 'a1', 'SUM', 'a3']);
     expect(transcript.foldedLength).toBe(live.length);
   });

packages/agent-core-v2/test/agent/contextMemory/splice-replay.test.ts

@@ -15,6 +15,7 @@ import {
   ContextUndo,
 } from '#/agent/contextMemory/contextEvents';
 import { contextMemoryKey } from '#/agent/contextMemory/contextOps';
+import { buildCompactionContinuationText } from '#/agent/contextMemory/compactionHandoff';
 import type { ContextMessage } from '#/agent/contextMemory/types';
 import { ISessionTokenCountingService } from '#/session/tokenCounting/sessionTokenCounting';
 import { IEventBus } from '#/app/event/eventBus';
@@ -376,11 +377,19 @@ describe('AgentContextMemoryService (wire-backed)', () => {
     );
 
     const model = replay.agentState.get(contextMemoryKey);
-    expect(model.map((message) => message.role)).toEqual(['user', 'user', 'user']);
-    expect(model.map(textOf)).toEqual(['old user', 'recent user', 'model-facing summary']);
+    expect(model.map((message) => message.role)).toEqual(['user', 'user', 'user', 'user']);
+    expect(model.map(textOf)).toEqual([
+      'old user',
+      'recent user',
+      'model-facing summary',
+      buildCompactionContinuationText(),
+    ]);
     expect(model[2]).toMatchObject({
       origin: { kind: 'compaction_summary' },
     });
+    expect(model[3]).toMatchObject({
+      origin: { kind: 'injection', variant: 'compaction_continuation' },
+    });
   });
 
   it('replays pre-contextSummary kept-user records without adding a new prefix', async () => {
@@ -406,7 +415,12 @@ describe('AgentContextMemoryService (wire-backed)', () => {
     );
 
     const model = replay.agentState.get(contextMemoryKey);
-    expect(model.map(textOf)).toEqual(['old user', 'recent user', 'OLD SUMMARY']);
+    expect(model.map(textOf)).toEqual([
+      'old user',
+      'recent user',
+      'OLD SUMMARY',
+      buildCompactionContinuationText(),
+    ]);
     expect(model[2]).toMatchObject({
       role: 'user',
       origin: { kind: 'compaction_summary' },