feat(agent-core): rework compaction to keep only user prompts and summary

.changeset/compaction-internals-cleanup.md

@@ -0,0 +1,5 @@
+---
+"@moonshot-ai/kimi-code": patch
+---
+
+Tighten compaction bookkeeping so compacted history stays consistent across retries and resume.

.changeset/rework-compaction-strategy.md

@@ -0,0 +1,5 @@
+---
+"@moonshot-ai/kimi-code": minor
+---
+
+Rework conversation compaction to keep only real user prompts followed by a user-role summary, dropping assistant and tool messages.

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

@@ -1,3 +1,9 @@
+import {
+  COMPACT_USER_MESSAGE_MAX_TOKENS,
+  collectCompactableUserMessages,
+  isRealUserInput,
+  selectRecentUserMessages,
+} from '@moonshot-ai/agent-core';
 import type {
   ContentPart,
   ContextMessage,
@@ -234,19 +240,21 @@ export function projectContext(
         break;
       case 'context.apply_compaction': {
         openSteps = new Map();
-        // Mirror agent-core's actual `applyCompaction` behaviour
-        // (`packages/agent-core/src/agent/context/index.ts`): history becomes
-        // `[summaryBubble, ...history.slice(compactedCount)]`. The summary is
-        // an *assistant* message tagged `origin.kind = 'compaction_summary'`
-        // (using 'system' would skew role counts and any downstream diff
-        // against agent-core history). The post-compaction tail is preserved
-        // rather than dropped, so messages still in context stay visible.
+        // Mirror agent-core's `applyCompaction`
+        // (`packages/agent-core/src/agent/context/index.ts`): the live history
+        // becomes the most recent real user messages (verbatim, within a token
+        // budget) followed by a single user-role summary tagged
+        // `origin.kind = 'compaction_summary'`. Assistant messages, tool calls,
+        // and tool results are dropped. The selection rule
+        // (`selectRecentUserMessages` / `collectCompactableUserMessages`) is the
+        // same helper agent-core's `ContextMemory` and the web transcript
+        // reducer apply, so all three views stay in sync.
         const summaryBubble: ProjectedMessage = {
           lineNo: entry.lineNo,
           time: rec.time,
           source: 'compaction_summary',
           message: {
-            role: 'assistant',
+            role: 'user',
             content: [{ type: 'text', text: rec.summary }],
             toolCalls: [],
             origin: { kind: 'compaction_summary' },
@@ -258,34 +266,49 @@ export function projectContext(
             tokensAfter: rec.tokensAfter,
           },
         };
+        const modelSummaryBubble: ProjectedMessage =
+          rec.contextSummary === undefined
+            ? summaryBubble
+            : {
+                ...summaryBubble,
+                message: {
+                  ...summaryBubble.message,
+                  content: [{ type: 'text', text: rec.contextSummary }],
+                } as ContextMessage,
+              };
         if (mode === 'model') {
-          // Drop the first `rec.compactedCount` HISTORY entries (NOT array
-          // entries): agent-core's `compactedCount` indexes into `_history`,
-          // which never contains our synthetic 'undo'/'clear' markers. Walk the
-          // array counting only history entries (`isHistoryEntry`) until
-          // `compactedCount` are passed, then slice there — any UI-only markers
-          // in the dropped region go with it (correct: they precede the
-          // compaction). With no markers this is exactly `slice(compactedCount)`.
-          let sliceAt = messages.length;
-          let passed = 0;
-          for (let i = 0; i < messages.length; i++) {
-            if (passed >= rec.compactedCount) {
-              sliceAt = i;
-              break;
-            }
-            if (isHistoryEntry(messages[i]!)) passed++;
-          }
-          if (passed < rec.compactedCount) sliceAt = messages.length;
-          messages = [summaryBubble, ...messages.slice(sliceAt)];
+          // Rebuild the model's-eye view as the kept user messages + summary.
+          // `realUserEntries` is filtered with the exact
+          // `collectCompactableUserMessages` predicate so it stays aligned with
+          // the selection below (genuine user input only — no injections,
+          // system triggers, or prior summaries). `selectRecentUserMessages`
+          // keeps a contiguous suffix of that subsequence, with only the oldest
+          // kept message possibly truncated, so each kept message maps back onto
+          // its original ProjectedMessage wrapper (preserving line/time); we swap
+          // in the (possibly truncated) message object.
+          const historyEntries = messages.filter(isHistoryEntry);
+          const realUserEntries = historyEntries.filter(
+            (pm) => collectCompactableUserMessages([pm.message]).length === 1,
+          );
+          const keptUserMessages = selectRecentUserMessages(
+            realUserEntries.map((pm) => pm.message),
+            COMPACT_USER_MESSAGE_MAX_TOKENS,
+          );
+          const suffixStart = realUserEntries.length - keptUserMessages.length;
+          const keptEntries: ProjectedMessage[] = keptUserMessages.map((message, i) => {
+            const original = realUserEntries[suffixStart + i]!;
+            return original.message === message ? original : { ...original, message };
+          });
+          messages = [...keptEntries, modelSummaryBubble];
         } else {
           // Full history: keep ALL preceding messages, just append the summary
           // marker inline so the compacted prefix stays visible.
           messages.push(summaryBubble);
         }
         // Mirror agent-core applyCompaction() → microCompaction.reset() (cutoff
-        // → 0): the message list is rebuilt as [summary, ...tail], so the old
-        // index-based cutoff no longer points at the same messages. (In full
-        // mode the blanking pass does not run, so this is a no-op there.)
+        // → 0): the message list is rebuilt, so the old index-based cutoff no
+        // longer points at the same messages. (In full mode the blanking pass
+        // does not run, so this is a no-op there.)
         microCutoff = 0;
         // Mirror agent-core applyCompaction() → _tokenCount = result.tokensAfter:
         // the live context-window fill is now the post-compaction count. Derived
@@ -577,16 +600,6 @@ function isHistoryEntry(pm: ProjectedMessage): boolean {
   return pm.source !== 'undo' && pm.source !== 'clear';
 }
 
-/** Mirrors agent-core `isRealUserPrompt` (`agent/context/index.ts`): a message
- *  counts toward an undo only if it is a genuine user prompt. */
-function isRealUserPrompt(message: ContextMessage): boolean {
-  if (message.role !== 'user') return false;
-  const origin = message.origin;
-  if (origin === undefined || origin.kind === 'user') return true;
-  if (origin.kind === 'skill_activation') return origin.trigger === 'user-slash';
-  return false;
-}
-
 /** Single source of truth for the `context.undo` backward walk, shared by both
  *  projection modes. Mirrors agent-core `undo` (`agent/context/index.ts`): walk
  *  from the end, skip `origin.kind === 'injection'` (those are KEPT even when
@@ -612,7 +625,7 @@ function computeUndoCutoff(
     if (origin?.kind === 'compaction_summary') break; // stop
     removedMessageCount++;
     cutoff = i;
-    if (isRealUserPrompt(messages[i]!.message) && ++removedUserCount >= count) break;
+    if (isRealUserInput(messages[i]!.message) && ++removedUserCount >= count) break;
   }
   return { cutoff, removedMessageCount };
 }

apps/vis/server/test/fixtures/sessions/sample-compaction/agents/main/wire.jsonl

@@ -1,5 +1,6 @@
 {"type":"metadata","protocol_version":"1.1","created_at":1779256791085}
 {"type":"config.update","cwd":"/tmp/work","profileName":"agent","systemPrompt":"You are Kimi.","time":1779256791100}
 {"type":"context.append_message","message":{"role":"user","content":[{"type":"text","text":"before compaction"}],"toolCalls":[]},"time":1779256800001}
-{"type":"context.apply_compaction","summary":"compacted summary","compactedCount":1,"tokensBefore":100,"tokensAfter":30,"time":1779256800500}
+{"type":"context.append_message","message":{"role":"assistant","content":[{"type":"text","text":"assistant reply"}],"toolCalls":[]},"time":1779256800200}
+{"type":"context.apply_compaction","summary":"compacted summary","compactedCount":2,"tokensBefore":100,"tokensAfter":30,"time":1779256800500}
 {"type":"context.append_message","message":{"role":"user","content":[{"type":"text","text":"after compaction"}],"toolCalls":[]},"time":1779256801000}