fix(consult): drop traces, budget checkpoints, silent summary inherit

BUG-729: dropped history rounds leave an omission marker in the model-visible summary slot. BUG-730: checkpoint threshold is 0.4 of historyBudgetChars (128k still 16,000). BUG-731: session_full new chat copies owned context_summary on the server; clients send only continued_from_session_id.
2026-09-16 07:38:36 +08:00
parent e61535f464
commit 149e1ec4c3
18 changed files with 451 additions and 31 deletions
@@ -59,6 +59,7 @@ export const chatSessionCreateSchema = z.object({
  messages: z.array(chatMessageSchema).max(0),
  session_type: z.enum(["consultation", "birth_time_rectification"]),
  rectification_case_id: z.string().uuid().nullable(),
+  continued_from_session_id: z.string().uuid().optional(),
  ...chartBindingSchema,
  updated_at: z.string().datetime().optional(),
 }).strict();
@@ -148,11 +149,41 @@ export type ChatSessionCreate = Readonly<{
  messages: readonly [];
  session_type: "consultation" | "birth_time_rectification";
  rectification_case_id: string | null;
+  continued_from_session_id?: string;
  chart_profile_id?: string | null;
  chart_profile_name?: string | null;
  chart_profile_role?: "self" | "other" | null;
 }>;

+export type ChatSessionCreateInsertValues = Readonly<{
+  title: string;
+  theme: ConsultationDomain;
+  model_id: string;
+  messages: readonly unknown[];
+  session_type: "consultation" | "birth_time_rectification";
+  rectification_case_id: string | null;
+  chart_profile_id?: string | null;
+  chart_profile_name?: string | null;
+  chart_profile_role?: "self" | "other" | null;
+}>;
+
+export function chatSessionCreateInsertRow(input: {
+  id: string;
+  userId: string;
+  values: ChatSessionCreateInsertValues;
+  inheritedSummary?: unknown;
+  updatedAt: string;
+}): Record<string, unknown> {
+  const row: Record<string, unknown> = {
+    id: input.id,
+    user_id: input.userId,
+    ...input.values,
+    updated_at: input.updatedAt,
+  };
+  if (input.inheritedSummary) row.context_summary = input.inheritedSummary;
+  return row;
+}
+
 export type ChatSessionWrite = Readonly<{
  title: string;
  theme: ConsultationDomain;
@@ -1,11 +1,14 @@
 export const CONSULTATION_HISTORY_LIMIT = 12;
 export const CONSULTATION_HISTORY_MESSAGE_CHARS = 12_000;
-export const CONSULTATION_HISTORY_TAIL_MAX_CHARS = 16_000;
 export const CONSULTATION_HISTORY_SYSTEM_RESERVE_TOKENS = 60_000;
 export const CONSULTATION_HISTORY_CHAR_PER_TOKEN = 1.5;
 export const CONSULTATION_HISTORY_BUDGET_MIN_CHARS = 4_000;
 export const CONSULTATION_HISTORY_BUDGET_MAX_CHARS = 40_000;
 export const DEFAULT_MODEL_CONTEXT_WINDOW = 128_000;
+// 0.4 of the history budget. At the default 128k window the budget is 40,000,
+// so the checkpoint fires at 16,000 — the previous hard-coded threshold.
+// Smaller windows then checkpoint before the tail can exceed the budget.
+export const CONSULTATION_HISTORY_CHECKPOINT_BUDGET_RATIO = 0.4;

 export const SESSION_CONTEXT_SUMMARY_HEADING = "【会话摘要（服务端维护）】";

@@ -51,6 +54,15 @@ export function historyBudgetChars(contextWindow: number | null | undefined): nu
  );
 }

+export function consultationHistoryCheckpointChars(contextWindow: number | null | undefined): number {
+  const budget = historyBudgetChars(contextWindow);
+  const threshold = Math.floor(budget * CONSULTATION_HISTORY_CHECKPOINT_BUDGET_RATIO);
+  if (!(threshold < budget)) {
+    throw new Error("consultation history checkpoint threshold must stay below the history budget");
+  }
+  return threshold;
+}
+
 export function parseSessionContextSummary(value: unknown): SessionContextSummaryV1 | null {
  if (!value || typeof value !== "object" || Array.isArray(value)) return null;
  const row = value as Record<string, unknown>;
@@ -74,6 +86,13 @@ export function omissionMarker(omittedChars: number): string {
  return `……（以下省略 ${omittedChars} 字，结论已并入会话摘要）`;
 }

+export function droppedRoundsMarker(droppedCount: number, hasSummary: boolean): string {
+  if (droppedCount <= 0) return "";
+  return hasSummary
+    ? `……（更早的 ${droppedCount} 轮问答已并入上面的会话摘要）`
+    : `……（更早的 ${droppedCount} 轮问答未能进入本轮上下文，结论尚未并入会话摘要）`;
+}
+
 export function clipConsultationHistoryText(text: string): string {
  if (text.length <= CONSULTATION_HISTORY_MESSAGE_CHARS) return text;
  const omitted = text.length - CONSULTATION_HISTORY_MESSAGE_CHARS;
@@ -165,16 +184,18 @@ export function consultationUserTurnContent(input: {
  instruction: string;
  extra?: string;
  summaryText?: string | null;
+  droppedCount?: number;
  question: string;
 }): string {
+  const summary = input.summaryText?.trim() ?? "";
+  const dropped = droppedRoundsMarker(input.droppedCount ?? 0, Boolean(summary));
  return [
    input.currentTime,
    input.name ? `用户称呼：${input.name}` : "",
    input.instruction,
    input.extra ?? "",
-    input.summaryText?.trim()
-      ? `${SESSION_CONTEXT_SUMMARY_HEADING}\n${input.summaryText.trim()}`
-      : "",
+    summary ? `${SESSION_CONTEXT_SUMMARY_HEADING}\n${summary}` : "",
+    dropped,
    input.question,
  ].filter(Boolean).join("\n");
 }
@@ -1,7 +1,7 @@
 import { Agent } from "@mastra/core/agent";

 import {
-  CONSULTATION_HISTORY_TAIL_MAX_CHARS,
+  consultationHistoryCheckpointChars,
  lastConsultationPair,
  parseSessionContextSummary,
  storedConsultationTurns,
@@ -98,7 +98,7 @@ export function sanitizeSessionContextSummary(raw: string): string | null {
 export function tailCharCount(
  messages: unknown,
  summary: SessionContextSummaryV1 | null,
-  options: { excludeRequestId?: string } = {},
+  options: { excludeRequestId?: string; contextWindow?: number | null } = {},
 ): number {
  const turns = storedConsultationTurns(messages, { excludeRequestId: options.excludeRequestId });
  const tail = summary
@@ -110,9 +110,10 @@ export function tailCharCount(
 export function shouldCheckpoint(
  messages: unknown,
  summary: SessionContextSummaryV1 | null,
-  options: { excludeRequestId?: string } = {},
+  options: { excludeRequestId?: string; contextWindow?: number | null } = {},
 ): boolean {
-  return tailCharCount(messages, summary, options) > CONSULTATION_HISTORY_TAIL_MAX_CHARS;
+  return tailCharCount(messages, summary, options)
+    > consultationHistoryCheckpointChars(options.contextWindow);
 }

 export function messagesForSummaryInput(
@@ -215,6 +216,19 @@ export function nextSessionContextSummary(
  };
 }

+export async function resolveInheritedContextSummary(input: {
+  continuedFromSessionId?: string | null;
+  loadOwnedSummary: (sessionId: string) => Promise<unknown>;
+}): Promise<SessionContextSummaryV1 | null> {
+  const sourceId = input.continuedFromSessionId?.trim();
+  if (!sourceId) return null;
+  try {
+    return parseSessionContextSummary(await input.loadOwnedSummary(sourceId));
+  } catch {
+    return null;
+  }
+}
+
 export async function writeSessionContextSummary(input: {
  seenUpdatedAt: string | null;
  summary: SessionContextSummaryV1;
@@ -228,13 +242,17 @@ export async function checkpointSessionContextSummary(input: {
  messages: unknown;
  summary: unknown;
  excludeRequestId?: string;
+  contextWindow?: number | null;
  now?: () => Date;
  generateText: (prompt: string, signal?: AbortSignal) => Promise<string>;
  update: (summary: SessionContextSummaryV1, seenUpdatedAt: string | null) => Promise<boolean>;
  timeoutMs?: number;
 }): Promise<"written" | "skipped" | "abandoned" | "failed"> {
  const previous = parseSessionContextSummary(input.summary);
-  if (!shouldCheckpoint(input.messages, previous, { excludeRequestId: input.excludeRequestId })) {
+  if (!shouldCheckpoint(input.messages, previous, {
+    excludeRequestId: input.excludeRequestId,
+    contextWindow: input.contextWindow,
+  })) {
    return "skipped";
  }
  try {