Skip to content

Enhanced ChatGPT Clone: Features Agents, MCP, Skills, DeepSeek, Anthropic, AWS, OpenAI, Responses API, Azure, Groq, o1, GPT-5, Mistral, OpenRouter, Vertex AI, Gemini, Artifacts, AI model switching, message search, Code Interpreter, langchain, DALL-E-3, OpenAPI Actions, Functions, Secure Multi-User Auth, Presets, open-source for self-hosting. Active

✂️ fix: Keep Parts After a Summary in the Summary-Bounded History Read (#16434)

Marco Beretta0a85ff5c5955
diff --git a/api/app/clients/BaseClient.js b/api/app/clients/BaseClient.js
index 4a4f6975bd..643fd796ec 100644
--- a/api/app/clients/BaseClient.js
+++ b/api/app/clients/BaseClient.js
@@ -25,6 +25,7 @@ const {
withBalanceReservations,
findCheckpointSummaryPart,
getSummaryPartText,
+ resolveCheckpointMessage,
runAfterSeed,
saveTurnConversation,
seedTurnConversation,
@@ -1419,6 +1420,7 @@ class BaseClient {
* - The message's 'role' is set to 'system'.
* - The message's 'text' is set to its 'summary'.
* - If the message has a 'summaryTokenCount', the message's 'tokenCount' is set to 'summaryTokenCount'.
+ * - A message with a summary content block keeps its content from that block on, for the SDK formatter to promote.
* The traversal stops at the message with the 'summary' property.
*
* Each message object should have an 'id' or 'messageId' property and may have a 'parentMessageId' property.
@@ -1468,35 +1470,13 @@ class BaseClient {
break;
}
- let resolved = message;
- let hasSummary = false;
- if (summary) {
- const summaryBlock = findCheckpointSummaryPart(message.content);
- if (summaryBlock) {
- const summaryText = getSummaryPartText(summaryBlock);
- resolved = {
- ...message,
- role: 'system',
- content: [{ type: ContentTypes.TEXT, text: summaryText }],
- tokenCount: summaryBlock.tokenCount,
- };
- hasSummary = true;
- } else if (message.summary) {
- resolved = {
- ...message,
- role: 'system',
- content: [{ type: ContentTypes.TEXT, text: message.summary }],
- tokenCount: message.summaryTokenCount ?? message.tokenCount,
- };
- hasSummary = true;
- }
- }
-
+ const checkpoint = summary ? resolveCheckpointMessage(message) : null;
+ const resolved = checkpoint ?? message;
const shouldMap = mapMethod != null && (mapCondition != null ? mapCondition(resolved) : true);
const processedMessage = shouldMap ? mapMethod(resolved) : resolved;
orderedMessages.push(processedMessage);
- if (hasSummary) {
+ if (checkpoint) {
break;
}
diff --git a/api/app/clients/specs/BaseClient.test.js b/api/app/clients/specs/BaseClient.test.js
index 2c675aa5dc..375f5f2d27 100644
--- a/api/app/clients/specs/BaseClient.test.js
+++ b/api/app/clients/specs/BaseClient.test.js
@@ -548,9 +548,10 @@ describe('BaseClient', () => {
summary: true,
});
expect(result).toHaveLength(2);
- expect(result[0].role).toBe('system');
- expect(result[0].content).toEqual([{ type: 'text', text: 'Content block summary' }]);
- expect(result[0].tokenCount).toBe(42);
+ expect(result[0].role).toBeUndefined();
+ expect(result[0].content).toEqual([
+ { type: 'summary', text: 'Content block summary', tokenCount: 42 },
+ ]);
});
it('should prefer content block summary over legacy summary field', () => {
@@ -571,8 +572,9 @@ describe('BaseClient', () => {
summary: true,
});
expect(result).toHaveLength(2);
- expect(result[0].content).toEqual([{ type: 'text', text: 'Content block summary' }]);
- expect(result[0].tokenCount).toBe(20);
+ expect(result[0].content).toEqual([
+ { type: 'summary', text: 'Content block summary', tokenCount: 20 },
+ ]);
});
it('should fallback to legacy summary when no content block exists', () => {
@@ -596,6 +598,34 @@ describe('BaseClient', () => {
expect(result[0].tokenCount).toBe(15);
});
+ it('keeps the parts a response produced after its summary (summary mode)', () => {
+ const trailingText = { type: 'text', text: 'Answer after summarizing' };
+ const messagesWithTrailingParts = [
+ { id: '1', parentMessageId: null, text: 'Message 1' },
+ {
+ id: '2',
+ parentMessageId: '1',
+ text: 'Before summarizing',
+ content: [
+ { type: 'text', text: 'Before summarizing' },
+ { type: 'summary', text: 'Earlier context', tokenCount: 6 },
+ trailingText,
+ ],
+ },
+ { id: '3', parentMessageId: '2', text: 'Message 3' },
+ ];
+ const result = TestClient.constructor.getMessagesForConversation({
+ messages: messagesWithTrailingParts,
+ parentMessageId: '3',
+ summary: true,
+ });
+ expect(result.map((message) => message.id)).toEqual(['2', '3']);
+ expect(result[0].content).toEqual([
+ { type: 'summary', text: 'Earlier context', tokenCount: 6 },
+ trailingText,
+ ]);
+ });
+
it('should not stop traversal at a failed summary, keeping the prior history', () => {
/** A summarize round that errored keeps the deltas it streamed, so its
* text is a truncated prefix; treating it as the checkpoint would send
diff --git a/api/server/controllers/agents/__tests__/client.retainedAnswers.spec.js b/api/server/controllers/agents/__tests__/client.retainedAnswers.spec.js
index aa0f035dbf..27a308a303 100644
--- a/api/server/controllers/agents/__tests__/client.retainedAnswers.spec.js
+++ b/api/server/controllers/agents/__tests__/client.retainedAnswers.spec.js
@@ -222,6 +222,11 @@ describe('AgentClient retained answers', () => {
client.user = 'user-123';
client.processMemory = jest.fn();
const rows = compactedBranch();
+ /** The stored count covers the whole response, including what preceded its summary. */
+ const storedResponseTokens = 5000;
+ const response = rows.find((row) => row.messageId === 'a2');
+ response.tokenCount = storedResponseTokens;
+ response.content.unshift({ type: ContentTypes.TEXT, text: 'Checking the pipeline.' });
getMessages.mockResolvedValue(rows);
/** The real loader: one read of the conversation, then the summary-bounded walk. */
const cut = await client.loadHistory('convo-123', 'u3');
@@ -239,6 +244,10 @@ describe('AgentClient retained answers', () => {
expect(text.indexOf(ANSWER_LINE)).toBeLessThan(text.indexOf(LATEST_TEXT));
expect(text.endsWith(LATEST_TEXT)).toBe(true);
expect(client.options.agent.additional_instructions ?? '').not.toContain(ANSWER_LINE);
+ expect(text).toContain('Deployed.');
+ expect(text).not.toContain('Checking the pipeline.');
+ expect(tokenCountMap.a2).toBeGreaterThan(0);
+ expect(tokenCountMap.a2).toBeLessThan(storedResponseTokens);
expect(cut[1].text).toBe(LATEST_TEXT);
expect(cut[1].content).toBeUndefined();
expect(counts[prompt.length - 1]).toBe(tokenCountMap.u3);
diff --git a/e2e/specs/mock/scenarios/failed-summary-history.spec.ts b/e2e/specs/mock/scenarios/failed-summary-history.spec.ts
index 50765dac37..fbe1da623e 100644
--- a/e2e/specs/mock/scenarios/failed-summary-history.spec.ts
+++ b/e2e/specs/mock/scenarios/failed-summary-history.spec.ts
@@ -26,12 +26,12 @@ async function startConversation(page: Page) {
}
/**
- * Appends a turn whose only content is a summary part, cloning the endpoint and
+ * Appends a turn whose content is the given parts, cloning the endpoint and
* model of the turn it hangs off. A summarize round that streamed deltas and
* then errored cannot be produced through the composer, and the behavior under
* test is what the *next* turn sends once such a turn is persisted.
*/
-async function appendSummaryTurn(conversationId: string, part: Record<string, unknown>) {
+async function appendSummaryTurn(conversationId: string, ...parts: Record<string, unknown>[]) {
await withMongo(async (db) => {
const rows = await db
.collection('messages')
@@ -53,7 +53,7 @@ async function appendSummaryTurn(conversationId: string, part: Record<string, un
parentMessageId: leaf.messageId,
isCreatedByUser: false,
text: '',
- content: [part],
+ content: parts,
createdAt: now,
updatedAt: now,
});
@@ -170,4 +170,32 @@ test.describe('failed summary history', () => {
messagesView(page).getByText(`E2E history assertion absent: ${token}`),
).toBeVisible({ timeout: 30000 });
});
+
+ /**
+ * A response that summarized mid-run keeps producing after the summary, and
+ * those parts are not covered by it. The text the turn wrote after its summary
+ * has to reach the model on the next turn.
+ */
+ test('a turn after a mid-response summary still sends what the response wrote after it @scenario:parts-after-summary-reach-next-turn', async ({
+ page,
+ }) => {
+ const { conversationId } = await startConversation(page);
+ conversationIds.push(conversationId);
+ const laterFact = `LATERFACT-${randomUUID().slice(0, 8)}`;
+ await appendSummaryTurn(
+ conversationId,
+ summaryPart('The user shared a passphrase and it was acknowledged.', {
+ boundary: completedBoundary,
+ }),
+ { type: 'text', text: `The deploy code is ${laterFact}.` },
+ );
+
+ await page.goto(`/c/${conversationId}`);
+ await expect(messagesView(page).getByText(laterFact)).toBeVisible();
+ await sendMessageAndWaitForCompletion(page, `E2E_ASSERT_HISTORY:${laterFact}`);
+
+ await expect(
+ messagesView(page).getByText(`E2E history assertion present: ${laterFact}`),
+ ).toBeVisible({ timeout: 30000 });
+ });
});
diff --git a/packages/api/src/agents/compaction.spec.ts b/packages/api/src/agents/compaction.spec.ts
index e5460ca7f7..a8910eeb3e 100644
--- a/packages/api/src/agents/compaction.spec.ts
+++ b/packages/api/src/agents/compaction.spec.ts
@@ -17,6 +17,7 @@ import {
getSummaryPartText,
markCompactionOutcome,
resolveFailedTurnContent,
+ resolveCheckpointMessage,
restoreCompactionSemanticIndex,
restoreCompactionSemanticIndexSnapshot,
stripUnusableSummaryParts,
@@ -246,6 +247,76 @@ describe('resolveFailedTurnContent', () => {
});
});
+describe('resolveCheckpointMessage', () => {
+ const summaryPart = {
+ type: ContentTypes.SUMMARY,
+ content: [{ type: ContentTypes.TEXT, text: 'Earlier context' }],
+ boundary: completedBoundary,
+ tokenCount: 7,
+ };
+
+ it('keeps the parts a response produced after its summary', () => {
+ const trailingText = { type: ContentTypes.TEXT, text: 'Answer after summarizing' };
+ const message = {
+ messageId: 'response',
+ tokenCount: 90,
+ content: [{ type: ContentTypes.TEXT, text: 'Before summarizing' }, summaryPart, trailingText],
+ };
+
+ const resolved = resolveCheckpointMessage(message);
+
+ expect(resolved).toEqual({ messageId: 'response', content: [summaryPart, trailingText] });
+ expect(resolved?.tokenCount).toBeUndefined();
+ });
+
+ it('starts at the last usable summary', () => {
+ const between = { type: ContentTypes.TEXT, text: 'Between summaries' };
+ const after = { type: ContentTypes.TEXT, text: 'After the latest' };
+ const latest = {
+ ...summaryPart,
+ content: [{ type: ContentTypes.TEXT, text: 'Latest context' }],
+ };
+
+ expect(resolveCheckpointMessage({ content: [summaryPart, between, latest, after] })).toEqual({
+ content: [latest, after],
+ });
+ });
+
+ it('keeps the parts after a completed summary when a later one failed', () => {
+ const after = { type: ContentTypes.TEXT, text: 'After the completed summary' };
+ const failed = { ...summaryPart, failed: true };
+
+ expect(resolveCheckpointMessage({ content: [summaryPart, after, failed] })).toEqual({
+ content: [summaryPart, after, failed],
+ });
+ });
+
+ it('returns the row itself when it already starts at its summary', () => {
+ const message = { content: [summaryPart], tokenCount: 12 };
+
+ expect(resolveCheckpointMessage(message)).toBe(message);
+ });
+
+ it('replaces the whole row for a legacy summary field', () => {
+ expect(
+ resolveCheckpointMessage({ summary: 'Legacy', summaryTokenCount: 4, tokenCount: 30 }),
+ ).toEqual({
+ summary: 'Legacy',
+ summaryTokenCount: 4,
+ role: 'system',
+ content: [{ type: ContentTypes.TEXT, text: 'Legacy' }],
+ tokenCount: 4,
+ });
+ });
+
+ it('is null for a row that is no checkpoint', () => {
+ expect(resolveCheckpointMessage({ content: [{ ...summaryPart, failed: true }] })).toBeNull();
+ expect(
+ resolveCheckpointMessage({ content: [{ type: ContentTypes.TEXT, text: 'x' }] }),
+ ).toBeNull();
+ });
+});
+
describe('findCheckpointSummaryPart', () => {
const legacySummary = { type: ContentTypes.SUMMARY, text: 'Summary of conversation' };
diff --git a/packages/api/src/agents/compaction.ts b/packages/api/src/agents/compaction.ts
index e89311178f..241c394965 100644
--- a/packages/api/src/agents/compaction.ts
+++ b/packages/api/src/agents/compaction.ts
@@ -96,6 +96,45 @@ export function findCheckpointSummaryPart(content: unknown): SummaryContentPart
return checkpoint;
}
+/** The fields a history row offers when it may be a summary checkpoint. */
+export interface CheckpointCandidate {
+ content?: unknown;
+ summary?: string | null;
+ summaryTokenCount?: number | null;
+ tokenCount?: number | null;
+}
+
+/**
+ * The row a history read stops at, as it enters the prompt, or null when the
+ * row is no checkpoint and the read continues past it. A content-block summary
+ * keeps the row from that summary on: a response that summarized mid-run keeps
+ * producing after it, and the SDK formatter (`applySummaryBoundary`) promotes
+ * the summary and keeps the parts that follow, exactly as it does when the full
+ * history is sent. A sliced row drops its stored token count, which still covers
+ * the parts before the summary, so the prompt is charged for what it sends. A
+ * legacy `summary` field replaces the whole row.
+ */
+export function resolveCheckpointMessage<T extends CheckpointCandidate>(
+ message: T,
+): (T & { role?: 'system' }) | null {
+ const summaryPart = findCheckpointSummaryPart(message.content);
+ if (summaryPart != null && Array.isArray(message.content)) {
+ const summaryIndex = message.content.lastIndexOf(summaryPart);
+ return summaryIndex === 0
+ ? message
+ : { ...message, content: message.content.slice(summaryIndex), tokenCount: undefined };
+ }
+ if (!message.summary) {
+ return null;
+ }
+ return {
+ ...message,
+ role: 'system',
+ content: [{ type: ContentTypes.TEXT, text: message.summary }],
+ tokenCount: message.summaryTokenCount ?? message.tokenCount,
+ };
+}
+
/**
* The summary a warm event-actor continuation carries forward: the last usable
* one in the run's content parts, stamped as actor state. A failed or