Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

feat(chat): say when a turn is summarizing earlier conversation

Sanil40c64bf99ee0
diff --git a/app/src/features/conversations/aui/AgentRunningStatus.test.tsx b/app/src/features/conversations/aui/AgentRunningStatus.test.tsx
index 6e04f8359e..6cef2abe59 100644
--- a/app/src/features/conversations/aui/AgentRunningStatus.test.tsx
+++ b/app/src/features/conversations/aui/AgentRunningStatus.test.tsx
@@ -105,6 +105,30 @@ describe('AgentRunningStatus waiting states', () => {
});
});
+describe('AgentRunningStatus while compacting', () => {
+ it('says it is summarizing earlier conversation instead of thinking', () => {
+ renderStatus({
+ inferenceStatusByThread: {
+ [THREAD_ID]: { phase: 'compacting', iteration: 1, maxIterations: 0 },
+ },
+ });
+ expect(screen.getByTestId('agent-running-status-thinking')).toHaveTextContent(
+ 'Summarizing earlier conversation'
+ );
+ });
+
+ it('thinks again once the next model call starts', () => {
+ renderStatus({
+ inferenceStatusByThread: {
+ [THREAD_ID]: { phase: 'thinking', iteration: 1, maxIterations: 0 },
+ },
+ });
+ expect(screen.getByTestId('agent-running-status-thinking')).not.toHaveTextContent(
+ 'Summarizing'
+ );
+ });
+});
+
describe('waitingOnUser', () => {
it('ranks approval over review over answer', () => {
expect(waitingOnUser({ approval: true, review: true, answer: true })).toBe('approval');
diff --git a/app/src/features/conversations/aui/AgentRunningStatus.tsx b/app/src/features/conversations/aui/AgentRunningStatus.tsx
index f89c5ef3e5..84f30d66aa 100644
--- a/app/src/features/conversations/aui/AgentRunningStatus.tsx
+++ b/app/src/features/conversations/aui/AgentRunningStatus.tsx
@@ -27,6 +27,9 @@
* 3. **Thinking.** The vendored `GenerationLoader`, with "Thinking · {elapsed}"
* once a second has passed. The clock starts when this slot mounts, which is
* when the turn starts running: the runtime keeps no turn-start timestamp.
+ * While the turn summarizes earlier conversation before its next model call
+ * (`inferenceStatusByThread.phase === 'compacting'`), the same loader says
+ * so instead: that can take minutes with nothing else to show.
*/
import { useAuiState } from '@assistant-ui/react';
import { useEffect, useState } from 'react';
@@ -140,6 +143,14 @@ function useParentWaitsOnSubagents(): boolean {
);
}
+/** Whether the turn is summarizing earlier conversation before its next call. */
+function useCompacting(): boolean {
+ const threadId = useAuiThreadId();
+ return useAppSelector(state =>
+ threadId ? state.chatRuntime.inferenceStatusByThread?.[threadId]?.phase === 'compacting' : false
+ );
+}
+
function WaitingLine({ kind }: { kind: WaitingOnUser }) {
const { t } = useT();
return (
@@ -163,6 +174,7 @@ export function AgentRunningStatus() {
const strings = useAgentStatusStrings();
const waiting = useWaitingOnUser();
const parentWaits = useParentWaitsOnSubagents();
+ const compacting = useCompacting();
const [tick, setTick] = useState(0);
const [startedAt] = useState(() => Date.now());
const turnTasks = useAuiState(s => turnHasTasks(s.thread.tasks, s.thread.messages));
@@ -178,9 +190,14 @@ export function AgentRunningStatus() {
if (waiting !== null) return <WaitingLine kind={waiting} />;
if (!turnTasks) {
- const label =
- elapsedMs !== undefined && elapsedMs >= 1000
- ? t('chat.status.thinkingElapsed').replace('{elapsed}', formatElapsed(elapsedMs))
+ const elapsed =
+ elapsedMs !== undefined && elapsedMs >= 1000 ? formatElapsed(elapsedMs) : undefined;
+ const label = compacting
+ ? elapsed
+ ? t('chat.status.compactingElapsed').replace('{elapsed}', elapsed)
+ : t('chat.status.compacting')
+ : elapsed
+ ? t('chat.status.thinkingElapsed').replace('{elapsed}', elapsed)
: t('chat.thinkingDots');
return (
<GenerationLoader
diff --git a/app/src/lib/i18n/ar.ts b/app/src/lib/i18n/ar.ts
index 1e05070734..9c2d6d26aa 100644
--- a/app/src/lib/i18n/ar.ts
+++ b/app/src/lib/i18n/ar.ts
@@ -1271,6 +1271,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'فشل تحميل الرسائل',
'chat.thinkingDots': 'جارٍ التفكير...',
'chat.status.thinkingElapsed': 'جارٍ التفكير · {elapsed}',
+ 'chat.status.compacting': 'جارٍ تلخيص المحادثة السابقة…',
+ 'chat.status.compactingElapsed': 'جارٍ تلخيص المحادثة السابقة · {elapsed}',
'chat.status.waitingApproval': 'في انتظار موافقتك',
'chat.status.waitingReview': 'في انتظار مراجعتك',
'chat.status.waitingAnswer': 'في انتظار إجابتك',
diff --git a/app/src/lib/i18n/bn.ts b/app/src/lib/i18n/bn.ts
index 99ce16245c..853554fb7e 100644
--- a/app/src/lib/i18n/bn.ts
+++ b/app/src/lib/i18n/bn.ts
@@ -1290,6 +1290,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'বার্তা লোড করতে ব্যর্থ',
'chat.thinkingDots': 'ভাবছে...',
'chat.status.thinkingElapsed': 'ভাবছে · {elapsed}',
+ 'chat.status.compacting': 'আগের কথোপকথনের সারাংশ তৈরি হচ্ছে…',
+ 'chat.status.compactingElapsed': 'আগের কথোপকথনের সারাংশ তৈরি হচ্ছে · {elapsed}',
'chat.status.waitingApproval': 'আপনার অনুমোদনের অপেক্ষায়',
'chat.status.waitingReview': 'আপনার পর্যালোচনার অপেক্ষায়',
'chat.status.waitingAnswer': 'আপনার উত্তরের অপেক্ষায়',
diff --git a/app/src/lib/i18n/de.ts b/app/src/lib/i18n/de.ts
index 70b0976f88..02a418c80b 100644
--- a/app/src/lib/i18n/de.ts
+++ b/app/src/lib/i18n/de.ts
@@ -1331,6 +1331,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'Nachrichten konnten nicht geladen werden',
'chat.thinkingDots': 'Denken...',
'chat.status.thinkingElapsed': 'Denkt nach · {elapsed}',
+ 'chat.status.compacting': 'Fasse den bisherigen Verlauf zusammen…',
+ 'chat.status.compactingElapsed': 'Fasse den bisherigen Verlauf zusammen · {elapsed}',
'chat.status.waitingApproval': 'Wartet auf deine Freigabe',
'chat.status.waitingReview': 'Wartet auf deine Prüfung',
'chat.status.waitingAnswer': 'Wartet auf deine Antwort',
diff --git a/app/src/lib/i18n/en.ts b/app/src/lib/i18n/en.ts
index aecf1dd492..2a8d76d251 100644
--- a/app/src/lib/i18n/en.ts
+++ b/app/src/lib/i18n/en.ts
@@ -1488,6 +1488,8 @@ const en: TranslationMap = {
'chat.failedToLoadMessages': 'Failed to load messages',
'chat.thinkingDots': 'Thinking...',
'chat.status.thinkingElapsed': 'Thinking · {elapsed}',
+ 'chat.status.compacting': 'Summarizing earlier conversation…',
+ 'chat.status.compactingElapsed': 'Summarizing earlier conversation · {elapsed}',
'chat.status.waitingApproval': 'Waiting for your approval',
'chat.status.waitingReview': 'Waiting for your review',
'chat.status.waitingAnswer': 'Waiting for your answer',
diff --git a/app/src/lib/i18n/es.ts b/app/src/lib/i18n/es.ts
index d44075ce46..f59e3901a5 100644
--- a/app/src/lib/i18n/es.ts
+++ b/app/src/lib/i18n/es.ts
@@ -1322,6 +1322,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'No se pudieron cargar los mensajes',
'chat.thinkingDots': 'Pensando...',
'chat.status.thinkingElapsed': 'Pensando · {elapsed}',
+ 'chat.status.compacting': 'Resumiendo la conversación anterior…',
+ 'chat.status.compactingElapsed': 'Resumiendo la conversación anterior · {elapsed}',
'chat.status.waitingApproval': 'Esperando tu aprobación',
'chat.status.waitingReview': 'Esperando tu revisión',
'chat.status.waitingAnswer': 'Esperando tu respuesta',
diff --git a/app/src/lib/i18n/fr.ts b/app/src/lib/i18n/fr.ts
index 8ec019a85f..9da53a8290 100644
--- a/app/src/lib/i18n/fr.ts
+++ b/app/src/lib/i18n/fr.ts
@@ -1330,6 +1330,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'Échec du chargement des messages',
'chat.thinkingDots': 'Réflexion…',
'chat.status.thinkingElapsed': 'Réflexion · {elapsed}',
+ 'chat.status.compacting': 'Résumé de la conversation précédente…',
+ 'chat.status.compactingElapsed': 'Résumé de la conversation précédente · {elapsed}',
'chat.status.waitingApproval': 'En attente de votre approbation',
'chat.status.waitingReview': 'En attente de votre relecture',
'chat.status.waitingAnswer': 'En attente de votre réponse',
diff --git a/app/src/lib/i18n/hi.ts b/app/src/lib/i18n/hi.ts
index f079c1c5e7..397f08e6a0 100644
--- a/app/src/lib/i18n/hi.ts
+++ b/app/src/lib/i18n/hi.ts
@@ -1298,6 +1298,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'मैसेज लोड नहीं हो पाए',
'chat.thinkingDots': 'सोच रहा है...',
'chat.status.thinkingElapsed': 'सोच रहा है · {elapsed}',
+ 'chat.status.compacting': 'पिछली बातचीत का सारांश बना रहा है…',
+ 'chat.status.compactingElapsed': 'पिछली बातचीत का सारांश बना रहा है · {elapsed}',
'chat.status.waitingApproval': 'आपकी स्वीकृति की प्रतीक्षा',
'chat.status.waitingReview': 'आपकी समीक्षा की प्रतीक्षा',
'chat.status.waitingAnswer': 'आपके उत्तर की प्रतीक्षा',
diff --git a/app/src/lib/i18n/id.ts b/app/src/lib/i18n/id.ts
index 75f639880b..15ced87273 100644
--- a/app/src/lib/i18n/id.ts
+++ b/app/src/lib/i18n/id.ts
@@ -1308,6 +1308,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'Gagal memuat pesan',
'chat.thinkingDots': 'Berpikir...',
'chat.status.thinkingElapsed': 'Berpikir · {elapsed}',
+ 'chat.status.compacting': 'Meringkas percakapan sebelumnya…',
+ 'chat.status.compactingElapsed': 'Meringkas percakapan sebelumnya · {elapsed}',
'chat.status.waitingApproval': 'Menunggu persetujuan Anda',
'chat.status.waitingReview': 'Menunggu tinjauan Anda',
'chat.status.waitingAnswer': 'Menunggu jawaban Anda',
diff --git a/app/src/lib/i18n/it.ts b/app/src/lib/i18n/it.ts
index 74225b8b78..983f43fcd2 100644
--- a/app/src/lib/i18n/it.ts
+++ b/app/src/lib/i18n/it.ts
@@ -1318,6 +1318,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'Impossibile caricare i messaggi',
'chat.thinkingDots': 'Sto pensando...',
'chat.status.thinkingElapsed': 'Sto pensando · {elapsed}',
+ 'chat.status.compacting': 'Riassumo la conversazione precedente…',
+ 'chat.status.compactingElapsed': 'Riassumo la conversazione precedente · {elapsed}',
'chat.status.waitingApproval': 'In attesa della tua approvazione',
'chat.status.waitingReview': 'In attesa della tua revisione',
'chat.status.waitingAnswer': 'In attesa della tua risposta',
diff --git a/app/src/lib/i18n/ja.ts b/app/src/lib/i18n/ja.ts
index 2a06d07cd3..552c63e898 100644
--- a/app/src/lib/i18n/ja.ts
+++ b/app/src/lib/i18n/ja.ts
@@ -1481,6 +1481,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'メッセージの読み込みに失敗しました',
'chat.thinkingDots': '思考中...',
'chat.status.thinkingElapsed': '思考中 · {elapsed}',
+ 'chat.status.compacting': 'これまでの会話を要約中…',
+ 'chat.status.compactingElapsed': 'これまでの会話を要約中 · {elapsed}',
'chat.status.waitingApproval': '承認を待っています',
'chat.status.waitingReview': 'レビューを待っています',
'chat.status.waitingAnswer': '回答を待っています',
diff --git a/app/src/lib/i18n/ko.ts b/app/src/lib/i18n/ko.ts
index ffbcd85798..790b79160b 100644
--- a/app/src/lib/i18n/ko.ts
+++ b/app/src/lib/i18n/ko.ts
@@ -1278,6 +1278,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': '메시지를 불러오지 못했습니다',
'chat.thinkingDots': '생각 중...',
'chat.status.thinkingElapsed': '생각 중 · {elapsed}',
+ 'chat.status.compacting': '이전 대화를 요약하는 중…',
+ 'chat.status.compactingElapsed': '이전 대화를 요약하는 중 · {elapsed}',
'chat.status.waitingApproval': '승인을 기다리는 중',
'chat.status.waitingReview': '검토를 기다리는 중',
'chat.status.waitingAnswer': '답변을 기다리는 중',
diff --git a/app/src/lib/i18n/pl.ts b/app/src/lib/i18n/pl.ts
index 76cc901c21..84e0ce5643 100644
--- a/app/src/lib/i18n/pl.ts
+++ b/app/src/lib/i18n/pl.ts
@@ -1310,6 +1310,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'Nie udało się wczytać wiadomości',
'chat.thinkingDots': 'Myślę...',
'chat.status.thinkingElapsed': 'Myślę · {elapsed}',
+ 'chat.status.compacting': 'Podsumowuję wcześniejszą rozmowę…',
+ 'chat.status.compactingElapsed': 'Podsumowuję wcześniejszą rozmowę · {elapsed}',
'chat.status.waitingApproval': 'Czeka na Twoją zgodę',
'chat.status.waitingReview': 'Czeka na Twoją ocenę',
'chat.status.waitingAnswer': 'Czeka na Twoją odpowiedź',
diff --git a/app/src/lib/i18n/pt.ts b/app/src/lib/i18n/pt.ts
index 3f58e7beaa..1502f956ee 100644
--- a/app/src/lib/i18n/pt.ts
+++ b/app/src/lib/i18n/pt.ts
@@ -1317,6 +1317,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'Falha ao carregar mensagens',
'chat.thinkingDots': 'Pensando...',
'chat.status.thinkingElapsed': 'Pensando · {elapsed}',
+ 'chat.status.compacting': 'Resumindo a conversa anterior…',
+ 'chat.status.compactingElapsed': 'Resumindo a conversa anterior · {elapsed}',
'chat.status.waitingApproval': 'Aguardando sua aprovação',
'chat.status.waitingReview': 'Aguardando sua revisão',
'chat.status.waitingAnswer': 'Aguardando sua resposta',
diff --git a/app/src/lib/i18n/ru.ts b/app/src/lib/i18n/ru.ts
index ea925b7a4a..e18439dbc4 100644
--- a/app/src/lib/i18n/ru.ts
+++ b/app/src/lib/i18n/ru.ts
@@ -1305,6 +1305,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': 'Не удалось загрузить сообщения',
'chat.thinkingDots': 'Думаю...',
'chat.status.thinkingElapsed': 'Думаю · {elapsed}',
+ 'chat.status.compacting': 'Резюмирую предыдущий разговор…',
+ 'chat.status.compactingElapsed': 'Резюмирую предыдущий разговор · {elapsed}',
'chat.status.waitingApproval': 'Ожидает вашего одобрения',
'chat.status.waitingReview': 'Ожидает вашей проверки',
'chat.status.waitingAnswer': 'Ожидает вашего ответа',
diff --git a/app/src/lib/i18n/tr.ts b/app/src/lib/i18n/tr.ts
index 7a333124cd..0dc6df510f 100644
--- a/app/src/lib/i18n/tr.ts
+++ b/app/src/lib/i18n/tr.ts
@@ -4208,6 +4208,8 @@ const messages: TranslationMap = {
'memoryPage.import.progress': '{imported}/{total} öğe içe aktarıldı',
// Chat status, history, and tool activity.
'chat.status.thinkingElapsed': 'Düşünüyor · {elapsed}',
+ 'chat.status.compacting': 'Önceki konuşma özetleniyor…',
+ 'chat.status.compactingElapsed': 'Önceki konuşma özetleniyor · {elapsed}',
'chat.status.waitingApproval': 'Onayınız bekleniyor',
'chat.status.waitingReview': 'İncelemeniz bekleniyor',
'chat.status.waitingAnswer': 'Yanıtınız bekleniyor',
diff --git a/app/src/lib/i18n/zh-CN.ts b/app/src/lib/i18n/zh-CN.ts
index c72b7850bf..0f201405aa 100644
--- a/app/src/lib/i18n/zh-CN.ts
+++ b/app/src/lib/i18n/zh-CN.ts
@@ -1212,6 +1212,8 @@ const messages: TranslationMap = {
'chat.failedToLoadMessages': '加载消息失败',
'chat.thinkingDots': '思考中...',
'chat.status.thinkingElapsed': '思考中 · {elapsed}',
+ 'chat.status.compacting': '正在总结之前的对话…',
+ 'chat.status.compactingElapsed': '正在总结之前的对话 · {elapsed}',
'chat.status.waitingApproval': '等待你的批准',
'chat.status.waitingReview': '等待你的审阅',
'chat.status.waitingAnswer': '等待你的回答',
diff --git a/app/src/providers/ChatRuntimeProvider.tsx b/app/src/providers/ChatRuntimeProvider.tsx
index 0f3df67f70..09b0775229 100644
--- a/app/src/providers/ChatRuntimeProvider.tsx
+++ b/app/src/providers/ChatRuntimeProvider.tsx
@@ -19,6 +19,7 @@ import {
type ChatApprovalDecidedEvent,
type ChatApprovalRequestEvent,
type ChatCancelledEvent,
+ type ChatContextCompactingEvent,
type ChatDoneEvent,
type ChatErrorEvent,
type ChatEventListeners,
@@ -852,6 +853,25 @@ const ChatRuntimeProvider = ({ children }: { children: React.ReactNode }) => {
})
);
},
+ onContextCompacting: (event: ChatContextCompactingEvent) => {
+ rtLog('context_compacting', { thread: event.thread_id, request: event.request_id });
+ // A parallel branch's compaction is its own; the thread's status
+ // line describes the primary turn.
+ if (store.getState().chatRuntime.parallelRequestThreads[event.request_id] !== undefined) {
+ return;
+ }
+ const prev = store.getState().chatRuntime.inferenceStatusByThread[event.thread_id];
+ dispatch(
+ setInferenceStatusForThread({
+ threadId: event.thread_id,
+ status: {
+ phase: 'compacting',
+ iteration: event.round,
+ maxIterations: prev?.maxIterations ?? 0,
+ },
+ })
+ );
+ },
onToolCall: (event: ChatToolCallEvent) => {
const prev = store.getState().chatRuntime.inferenceStatusByThread[event.thread_id];
dispatch(
diff --git a/app/src/services/__tests__/chatService.test.ts b/app/src/services/__tests__/chatService.test.ts
index 831225c636..7e75fd3410 100644
--- a/app/src/services/__tests__/chatService.test.ts
+++ b/app/src/services/__tests__/chatService.test.ts
@@ -117,6 +117,19 @@ describe('chatService.subscribeChatEvents', () => {
// listener interface unchanged. Without this coverage the parent
// thread's live subagent block silently goes blank if a future
// refactor renames a socket event.
+ it('forwards context_compacting to its listener', () => {
+ const socket = createMockSocket();
+ bindMockSocket(socket);
+ const onContextCompacting = vi.fn();
+
+ subscribeChatEvents({ onContextCompacting });
+
+ expect(socket.on.mock.calls.map(call => call[0])).toEqual(['context_compacting']);
+ const event = { thread_id: 't', request_id: 'r', round: 1 };
+ socket.emit('context_compacting', event);
+ expect(onContextCompacting).toHaveBeenCalledWith(event);
+ });
+
it('subscribes and forwards live subagent events under canonical names', () => {
const socket = createMockSocket();
bindMockSocket(socket);
diff --git a/app/src/services/chatService.ts b/app/src/services/chatService.ts
index a1490ed903..1b4d97e82e 100644
--- a/app/src/services/chatService.ts
+++ b/app/src/services/chatService.ts
@@ -608,6 +608,17 @@ export interface ChatIterationStartEvent {
message: string;
}
+/**
+ * Emitted when the turn starts summarizing earlier conversation to fit the
+ * context window. The next `iteration_start` ends it.
+ */
+export interface ChatContextCompactingEvent {
+ thread_id: string;
+ request_id: string;
+ /** The iteration whose model call waits on the summary. */
+ round: number;
+}
+
/** Emitted when a sub-agent is spawned during tool execution. */
export interface ChatSubagentSpawnedEvent {
thread_id: string;
@@ -870,6 +881,7 @@ export interface ChatEventListeners {
onInferenceStart?: (event: ChatInferenceStartEvent) => void;
onInferenceHeartbeat?: (event: ChatInferenceHeartbeatEvent) => void;
onIterationStart?: (event: ChatIterationStartEvent) => void;
+ onContextCompacting?: (event: ChatContextCompactingEvent) => void;
onToolCall?: (event: ChatToolCallEvent) => void;
onToolResult?: (event: ChatToolResultEvent) => void;
onSubagentSpawned?: (event: ChatSubagentSpawnedEventV2) => void;
@@ -922,6 +934,7 @@ export function subscribeChatEvents(listeners: ChatEventListeners): () => void {
inferenceStart: 'inference_start',
inferenceHeartbeat: 'inference_heartbeat',
iterationStart: 'iteration_start',
+ contextCompacting: 'context_compacting',
toolCall: 'tool_call',
toolResult: 'tool_result',
subagentSpawned: 'subagent_spawned',
@@ -996,6 +1009,22 @@ export function subscribeChatEvents(listeners: ChatEventListeners): () => void {
handlers.push([EVENTS.iterationStart, cb]);
}
+ if (listeners.onContextCompacting) {
+ const cb = (payload: unknown) => {
+ const e = payload as ChatContextCompactingEvent;
+ chatLog(
+ '%s thread_id=%s request_id=%s round=%d',
+ EVENTS.contextCompacting,
+ e.thread_id,
+ e.request_id,
+ e.round
+ );
+ listeners.onContextCompacting?.(e);
+ };
+ socket.on(EVENTS.contextCompacting, cb);
+ handlers.push([EVENTS.contextCompacting, cb]);
+ }
+
if (listeners.onToolCall) {
const cb = (payload: unknown) => {
const e = payload as ChatToolCallEvent;
diff --git a/app/src/store/chatRuntimeSlice.ts b/app/src/store/chatRuntimeSlice.ts
index dcf385a32b..ca4b043d0e 100644
--- a/app/src/store/chatRuntimeSlice.ts
+++ b/app/src/store/chatRuntimeSlice.ts
@@ -88,7 +88,8 @@ export function selectSubagentChildrenByParentCallId(
/** Live progress of the running turn, as the socket handlers maintain it. */
export interface InferenceStatus {
- phase: 'thinking' | 'tool_use' | 'subagent';
+ /** `compacting`: summarizing earlier conversation before the next model call. */
+ phase: 'thinking' | 'compacting' | 'tool_use' | 'subagent';
iteration: number;
maxIterations: number;
activeTool?: string;
diff --git a/crates/openhuman-core/src/agent/tinyagents/context_wiring.rs b/crates/openhuman-core/src/agent/tinyagents/context_wiring.rs
index 2ee8e89f85..7689dc0eb0 100644
--- a/crates/openhuman-core/src/agent/tinyagents/context_wiring.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/context_wiring.rs
@@ -6,6 +6,7 @@
use std::sync::Arc;
+use async_trait::async_trait;
use tinyagents_harness::middleware::{
compression_policy, legacy_max_input_tokens, microcompact_for_window, split_input_allowance,
ArtifactIndexTocMiddleware, CapturedOutcomes, ContextCompressionMiddleware, ContextLadder,
@@ -13,9 +14,10 @@ use tinyagents_harness::middleware::{
};
use tinyagents_harness::runtime::AgentHarness;
use tinyagents_harness::summarization::{
- FaultTolerantCachingSummarizer, ModelSummarizer, SummarizationPolicy, TaskStateSummarizer,
- DEFAULT_TASK_STATE_CHUNK_TOKENS,
+ FaultTolerantCachingSummarizer, ModelSummarizer, SummarizationPolicy, Summarizer,
+ SummaryRecord, SummaryRequest, TaskStateSummarizer, DEFAULT_TASK_STATE_CHUNK_TOKENS,
};
+use tinyinference_llm::message::Message;
use crate::agent::tinyagents::host::OpenHumanRunContext;
use crate::agent::tinyagents::memory_summarizer::MemoryRecallSummarizer;
@@ -31,6 +33,51 @@ const TASK_STATE_KEEP_RECENT_TOKENS: u64 = 20_000;
const ARTIFACT_INDEX_STORE: &str =
crate::agent::harness::tool_result_artifacts::TINYAGENTS_TOOL_RESULT_ARTIFACT_STORE;
+/// Called when the turn starts summarizing earlier conversation.
+pub(super) type CompactionNotice = Arc<dyn Fn() + Send + Sync>;
+
+/// A summarizer that reports each summarization before it runs, so the client
+/// can say what a turn is doing while a long compaction holds its next model
+/// call. It sits under the cache, so a cached summary is not reported.
+struct AnnouncingSummarizer {
+ inner: Box<dyn Summarizer>,
+ notice: CompactionNotice,
+}
+
+#[async_trait]
+impl Summarizer for AnnouncingSummarizer {
+ async fn summarize(
+ &self,
+ messages: &[Message],
+ ) -> tinyagents_harness::error::Result<SummaryRecord> {
+ (self.notice)();
+ self.inner.summarize(messages).await
+ }
+
+ async fn summarize_request(
+ &self,
+ request: &SummaryRequest,
+ ) -> tinyagents_harness::error::Result<SummaryRecord> {
+ (self.notice)();
+ self.inner.summarize_request(request).await
+ }
+
+ async fn merge(
+ &self,
+ summaries: &[SummaryRecord],
+ ) -> tinyagents_harness::error::Result<SummaryRecord> {
+ self.inner.merge(summaries).await
+ }
+}
+
+/// `inner`, reporting each summarization to `notice` when there is one.
+fn announced(inner: Box<dyn Summarizer>, notice: Option<CompactionNotice>) -> Box<dyn Summarizer> {
+ match notice {
+ Some(notice) => Box::new(AnnouncingSummarizer { inner, notice }),
+ None => inner,
+ }
+}
+
/// Read side of the tool-outcome capture sink, for the wrap-up's restoration.
pub(super) struct OutcomeSinkSource(pub(super) ToolOutcomeSink);
@@ -60,6 +107,7 @@ pub(super) fn install_context_ladder(
wrap_up_at_cap: bool,
tool_outcome_sink: &ToolOutcomeSink,
memory_turn: Option<Arc<crate::memory::lifecycle::hooks::MemoryTurn>>,
+ compaction_notice: Option<CompactionNotice>,
) -> (
Option<Arc<ContextCompressionMiddleware>>,
Option<Arc<FinalCallWrapUpMiddleware>>,
@@ -76,6 +124,7 @@ pub(super) fn install_context_ladder(
summarizer_model,
model,
memory_turn,
+ compaction_notice,
))
});
// One allowance split between the contents list and wrap-up restoration.
@@ -126,6 +175,7 @@ fn compression_middleware(
summarizer_model: TurnChatModel,
model: &str,
memory_turn: Option<Arc<crate::memory::lifecycle::hooks::MemoryTurn>>,
+ compaction_notice: Option<CompactionNotice>,
) -> ContextCompressionMiddleware {
let window = context_window.filter(|w| *w > 0);
match strategy {
@@ -138,8 +188,12 @@ fn compression_middleware(
});
tracing::info!(model, keep_recent_tokens = keep_recent, max_chunk_tokens = chunk, context_window = ?window, "[context_compression] strategy=task_state");
let summarizer = FaultTolerantCachingSummarizer::new(
- Box::new(
- TaskStateSummarizer::new(summarizer_model, model).with_max_chunk_tokens(chunk),
+ announced(
+ Box::new(
+ TaskStateSummarizer::new(summarizer_model, model)
+ .with_max_chunk_tokens(chunk),
+ ),
+ compaction_notice,
),
&policy,
);
@@ -152,7 +206,10 @@ fn compression_middleware(
crate::config::CompactionStrategy::Summary => {
tracing::info!(model, "[context_compression] strategy=summary");
let summarizer = FaultTolerantCachingSummarizer::new(
- Box::new(ModelSummarizer::new(summarizer_model, model)),
+ announced(
+ Box::new(ModelSummarizer::new(summarizer_model, model)),
+ compaction_notice,
+ ),
&policy,
);
ContextCompressionMiddleware::with_summarizer(
diff --git a/crates/openhuman-core/src/agent/tinyagents/context_wiring_tests.rs b/crates/openhuman-core/src/agent/tinyagents/context_wiring_tests.rs
index 44f05b5174..533148f39d 100644
--- a/crates/openhuman-core/src/agent/tinyagents/context_wiring_tests.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/context_wiring_tests.rs
@@ -21,3 +21,54 @@ fn outcome_source_returns_captured_content_by_call_id() {
fn budget_notice_thresholds_stay_at_half_and_eighty_percent() {
assert_eq!(BUDGET_NOTICE_THRESHOLDS, [0.5, 0.8]);
}
+
+/// Records each call it gets, then fails: the tests read the order only.
+struct RecordingSummarizer(Arc<std::sync::Mutex<Vec<&'static str>>>);
+
+#[async_trait]
+impl Summarizer for RecordingSummarizer {
+ async fn summarize(&self, _: &[Message]) -> tinyagents_harness::error::Result<SummaryRecord> {
+ self.0.lock().unwrap().push("summarize");
+ Err(tinyagents_harness::error::TinyAgentsError::Model(
+ "stub".into(),
+ ))
+ }
+
+ async fn merge(&self, _: &[SummaryRecord]) -> tinyagents_harness::error::Result<SummaryRecord> {
+ self.0.lock().unwrap().push("merge");
+ Err(tinyagents_harness::error::TinyAgentsError::Model(
+ "stub".into(),
+ ))
+ }
+}
+
+fn recording() -> (
+ Arc<std::sync::Mutex<Vec<&'static str>>>,
+ Box<dyn Summarizer>,
+) {
+ let calls = Arc::new(std::sync::Mutex::new(Vec::new()));
+ let notice_calls = calls.clone();
+ let notice: CompactionNotice = Arc::new(move || notice_calls.lock().unwrap().push("notice"));
+ let summarizer = announced(Box::new(RecordingSummarizer(calls.clone())), Some(notice));
+ (calls, summarizer)
+}
+
+#[tokio::test]
+async fn a_summarization_is_announced_before_it_runs() {
+ let (calls, summarizer) = recording();
+ let _ = summarizer.summarize(&[]).await;
+ let _ = summarizer
+ .summarize_request(&SummaryRequest::new(Vec::new()))
+ .await;
+ assert_eq!(
+ *calls.lock().unwrap(),
+ ["notice", "summarize", "notice", "summarize"]
+ );
+}
+
+#[tokio::test]
+async fn merging_summaries_is_not_announced_again() {
+ let (calls, summarizer) = recording();
+ let _ = summarizer.merge(&[]).await;
+ assert_eq!(*calls.lock().unwrap(), ["merge"]);
+}
diff --git a/crates/openhuman-core/src/agent/tinyagents/observability/event_bridge.rs b/crates/openhuman-core/src/agent/tinyagents/observability/event_bridge.rs
index c67bb0fbf7..6180be2c52 100644
--- a/crates/openhuman-core/src/agent/tinyagents/observability/event_bridge.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/observability/event_bridge.rs
@@ -318,6 +318,14 @@ impl OpenhumanEventBridge {
}
}
+ /// Tell the client the turn is summarizing earlier conversation. Only a
+ /// top-level turn reports it: a subagent's compaction is not the turn's.
+ pub(crate) fn announce_compaction(&self) {
+ if self.scope.is_none() {
+ self.send(AgentProgress::ContextCompacting);
+ }
+ }
+
pub(super) fn iteration(&self) -> u32 {
self.cursor.load(Ordering::SeqCst)
}
diff --git a/crates/openhuman-core/src/agent/tinyagents/observability_tests.rs b/crates/openhuman-core/src/agent/tinyagents/observability_tests.rs
index 5998a425c2..d113168b98 100644
--- a/crates/openhuman-core/src/agent/tinyagents/observability_tests.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/observability_tests.rs
@@ -141,7 +141,11 @@ async fn the_top_level_run_announces_the_turn_once_and_first() {
});
}
assert_eq!(kinds.first(), Some(&"turn"), "{kinds:?}");
- assert_eq!(kinds.iter().filter(|k| **k == "turn").count(), 1, "{kinds:?}");
+ assert_eq!(
+ kinds.iter().filter(|k| **k == "turn").count(),
+ 1,
+ "{kinds:?}"
+ );
}
#[tokio::test]
@@ -743,3 +747,34 @@ async fn bridge_forwards_tool_progress_with_a_message() {
)]
);
}
+
+#[tokio::test]
+async fn only_a_top_level_turn_announces_its_compaction() {
+ let (tx, mut rx) = tokio::sync::mpsc::channel(8);
+ OpenhumanEventBridge::new(Some(tx), "mock-model", 10).announce_compaction();
+ assert!(matches!(
+ rx.try_recv(),
+ Ok(AgentProgress::ContextCompacting)
+ ));
+
+ let (tx, mut rx) = tokio::sync::mpsc::channel(8);
+ OpenhumanEventBridge::with_scope(
+ Some(tx),
+ "mock-model",
+ "custom",
+ 10,
+ Some(SubagentScope {
+ agent_id: "researcher".into(),
+ task_id: "t1".into(),
+ extended_policy: false,
+ }),
+ Arc::default(),
+ Arc::default(),
+ Arc::default(),
+ Arc::default(),
+ Vec::new(),
+ crate::core::runtime::DispatchCtx::unbound(),
+ )
+ .announce_compaction();
+ assert!(rx.try_recv().is_err());
+}
diff --git a/crates/openhuman-core/src/agent/tinyagents/turn_exec.rs b/crates/openhuman-core/src/agent/tinyagents/turn_exec.rs
index 0c9bf28448..20a7a57fd9 100644
--- a/crates/openhuman-core/src/agent/tinyagents/turn_exec.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/turn_exec.rs
@@ -575,6 +575,10 @@ async fn drive_turn(
.map(|parent| {
crate::storage::sessions::stores_or_ephemeral(&parent.workspace_dir).kv
}),
+ compaction_notice: Some({
+ let bridge = bridge.clone();
+ Arc::new(move || bridge.announce_compaction())
+ }),
},
);
if let Some(scope) = run_context.response_shape.clone() {
diff --git a/crates/openhuman-core/src/agent/tinyagents/turn_middleware.rs b/crates/openhuman-core/src/agent/tinyagents/turn_middleware.rs
index f55bb7d318..0730b4d88b 100644
--- a/crates/openhuman-core/src/agent/tinyagents/turn_middleware.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/turn_middleware.rs
@@ -96,6 +96,9 @@ pub(super) struct TurnMiddlewareInputs {
/// plan-order middleware read it. `None` for a turn with no workspace or
/// without the `todo` tool.
pub(super) todo_store: Option<Arc<dyn tinyagents_harness::store::Store>>,
+ /// Called when the turn starts summarizing earlier conversation, so the
+ /// client can show it. `None` for a turn nobody watches.
+ pub(super) compaction_notice: Option<super::context_wiring::CompactionNotice>,
}
/// The handles the run loop reads after the drive returns.
@@ -147,6 +150,7 @@ pub(super) fn install_turn_middleware(
provider_usage_carry,
tool_outcome_sink,
todo_store,
+ compaction_notice,
} = inputs;
let super::host::HarnessHostPolicy {
stop_hooks,
@@ -451,6 +455,7 @@ pub(super) fn install_turn_middleware(
pause_at_cap && subagent_scope.is_none(),
&tool_outcome_sink,
memory_turn,
+ compaction_notice,
);
let is_subagent = subagent_scope.is_some();
let agent_id = tool_policy.as_ref().map(|p| p.agent_definition_id.as_str());
diff --git a/crates/openhuman-core/src/progress.rs b/crates/openhuman-core/src/progress.rs
index 9ed06a59a6..bae59ecf24 100644
--- a/crates/openhuman-core/src/progress.rs
+++ b/crates/openhuman-core/src/progress.rs
@@ -27,6 +27,11 @@ pub enum AgentProgress {
max_iterations: u32,
},
+ /// The turn is summarizing earlier conversation to fit the context window
+ /// before its next model call. That call starts with the next
+ /// [`Self::IterationStarted`].
+ ContextCompacting,
+
/// The LLM responded and the agent is about to execute a tool.
ToolCallStarted {
/// Provider-assigned (or synthesised) tool call id that ties
diff --git a/crates/openhuman-core/src/skills/run_log.rs b/crates/openhuman-core/src/skills/run_log.rs
index 6fcb68c4f0..5fa42dca53 100644
--- a/crates/openhuman-core/src/skills/run_log.rs
+++ b/crates/openhuman-core/src/skills/run_log.rs
@@ -117,6 +117,7 @@ pub fn format_event(ev: &AgentProgress) -> Option<String> {
iteration,
max_iterations,
} => format!("· iteration {iteration}/{max_iterations}"),
+ AgentProgress::ContextCompacting => "· summarizing earlier conversation".to_string(),
AgentProgress::ToolCallStarted {
tool_name,
arguments,
diff --git a/crates/openhuman-core/src/threads/turn_state/mirror/observe.rs b/crates/openhuman-core/src/threads/turn_state/mirror/observe.rs
index fd2f27e8af..42ef67973b 100644
--- a/crates/openhuman-core/src/threads/turn_state/mirror/observe.rs
+++ b/crates/openhuman-core/src/threads/turn_state/mirror/observe.rs
@@ -470,6 +470,12 @@ impl ObserveProgress for TurnStateMirror {
// running, so a flush per beat would buy nothing.
false
}
+ AgentProgress::ContextCompacting => {
+ // Live-only (`context_compacting` on the socket): the snapshot's
+ // phases come from tinyagents' `TurnPhase`, which has none for
+ // it, and the next `IterationStarted` sets the phase anyway.
+ false
+ }
}
}
}
diff --git a/crates/openhuman-core/src/web_chat/progress_bridge.rs b/crates/openhuman-core/src/web_chat/progress_bridge.rs
index f6f63f689e..9403f3ee5a 100644
--- a/crates/openhuman-core/src/web_chat/progress_bridge.rs
+++ b/crates/openhuman-core/src/web_chat/progress_bridge.rs
@@ -475,6 +475,20 @@ pub(crate) fn spawn_progress_bridge(
},
);
}
+ AgentProgress::ContextCompacting => {
+ publish_seq_stamped(
+ &dispatch_ctx,
+ &mut emit_seq,
+ WebChannelEvent {
+ event: "context_compacting".to_string(),
+ client_id: client_id.clone(),
+ thread_id: thread_id.clone(),
+ request_id: request_id.clone(),
+ round: Some(round),
+ ..Default::default()
+ },
+ );
+ }
AgentProgress::ToolCallStarted {
call_id,
tool_name,
diff --git a/crates/openhuman-core/src/web_chat/turn_activity.rs b/crates/openhuman-core/src/web_chat/turn_activity.rs
index 2ebf0dbb03..4ff77b94cc 100644
--- a/crates/openhuman-core/src/web_chat/turn_activity.rs
+++ b/crates/openhuman-core/src/web_chat/turn_activity.rs
@@ -125,6 +125,7 @@ fn step_of(event: &AgentProgress) -> Option<Cow<'static, str>> {
Some(match event {
AgentProgress::TurnStarted => "the turn started".into(),
AgentProgress::IterationStarted { .. } => "a model call started".into(),
+ AgentProgress::ContextCompacting => "it began summarizing earlier conversation".into(),
AgentProgress::ModelCallCompleted {
subagent_task_id: None,
..