minimal-agent 0.18.13 → 0.18.14

package/src/memory/retrieval.js

@@ -5,6 +5,23 @@ const snapshotIndexes = new WeakMap();
 export function memoryTerms(query) {
     return searchTerms(query);
 }
+/**
+ * Stable personal facts keep their force however old they are. A preference
+ * like "always answer in Chinese" does not expire by calendar: age used to
+ * relabel it "待复核，不能当作当前已确认事实" after the review window, and
+ * restating it never refreshed the clock, so a standing preference silently
+ * stopped binding after six months. Project/decision/topic records change with
+ * the work, so for them age is a real signal. Only the injected label and the
+ * ranking penalty use this; `isReviewDue` still feeds the user's review queue.
+ */
+export function agingDowngrades(record) {
+    return record.category !== 'preference' && record.category !== 'profile';
+}
+/** The label and ranking view of staleness: explicit marks always count, age only where it is evidence. */
+export function isUnconfirmed(record, settings, now = Date.now()) {
+    return record.status === 'needs_review'
+        || (agingDowngrades(record) && isReviewDue(record, settings, now));
+}
 export function isReviewDue(record, settings, now = Date.now()) {
     const days = settings.reviewAfterDays?.[record.category] ??
         DEFAULT_REVIEW_WINDOWS[record.category];
@@ -13,14 +30,18 @@ export function isReviewDue(record, settings, now = Date.now()) {
 export function scoreMemory(record, query, settings) {
     if (record.status === 'superseded' || (record.validUntil !== undefined && record.validUntil <= Date.now()))
         return Number.NEGATIVE_INFINITY;
+    const stalePenalty = isUnconfirmed(record, settings) ? 8 : 0;
+    const { terms, phrase } = substringParts(record, query);
+    return terms + phrase - stalePenalty + record.lastConfirmedAt / 1e13;
+}
+/** Substring evidence: 10 per matching query term, 25 for the whole phrase. */
+function substringParts(record, query) {
     const haystack = `${record.category} ${record.subject} ${record.topicKey} ${record.aliases.join(' ')} ${record.statement}`.normalize('NFKC').toLocaleLowerCase();
-    const terms = memoryTerms(query);
-    const lexical = terms.reduce((sum, term) => sum + (haystack.includes(term) ? 10 : 0), 0);
+    const lexical = memoryTerms(query).reduce((sum, term) => sum + (haystack.includes(term) ? 10 : 0), 0);
     const phrase = query.trim() && haystack.includes(query.trim().normalize('NFKC').toLocaleLowerCase())
         ? 25
         : 0;
-    const stalePenalty = isReviewDue(record, settings) || record.status === 'needs_review' ? 8 : 0;
-    return lexical + phrase - stalePenalty + record.lastConfirmedAt / 1e13;
+    return { terms: lexical, phrase };
 }
 export function recallFromSnapshot(snapshot, query, settings, limit) {
     const eligible = snapshot.records.filter((record) => record.status !== 'superseded' && (record.validUntil === undefined || record.validUntil > Date.now()));
@@ -53,18 +74,26 @@ export function recallFromSnapshot(snapshot, query, settings, limit) {
         }
         recordSearchStrategy('bm25', eligible.length);
         const bm25 = index.search(query);
-        const normalized = query.trim().normalize('NFKC').toLocaleLowerCase();
+        // BM25 only knows whole tokens, so `postgres` never reaches a record that
+        // says `PostgreSQL`; the small-corpus path finds it by substring. Growing
+        // past the threshold must not make a findable record vanish, so substring
+        // hits stay candidates. A token hit keeps its original score (BM25's IDF
+        // is what puts a rare word above two common ones, and the flat per-term
+        // substring bonus would flatten it); substring evidence only scores the
+        // records BM25 could not see.
         return eligible
             .map((record) => {
-            const haystack = `${record.category} ${record.subject} ${record.topicKey} ${record.aliases.join(' ')} ${record.statement}`.normalize('NFKC').toLocaleLowerCase();
-            const phrase = normalized && haystack.includes(normalized) ? 25 : 0;
-            const stalePenalty = isReviewDue(record, settings) || record.status === 'needs_review' ? 8 : 0;
+            const { terms, phrase } = substringParts(record, query);
+            const tokens = bm25.get(record.id) ?? 0;
+            const stalePenalty = isUnconfirmed(record, settings) ? 8 : 0;
             return {
                 record,
-                score: (bm25.get(record.id) ?? 0) * 10 + phrase - stalePenalty + record.lastConfirmedAt / 1e13,
+                hit: tokens > 0 || terms + phrase >= 2,
+                score: (tokens > 0 ? tokens * 10 + phrase : terms + phrase)
+                    - stalePenalty + record.lastConfirmedAt / 1e13,
             };
         })
-            .filter((item) => (bm25.get(item.record.id) ?? 0) > 0)
+            .filter((item) => item.hit)
             .sort((a, b) => b.score - a.score)
             .slice(0, Math.min(100, Math.max(1, limit)))
             .map((item) => item.record);

package/src/memory/retrieval.ts

@@ -21,6 +21,29 @@ export function memoryTerms(query: string): string[] {
   return searchTerms(query);
 }
 
+/**
+ * Stable personal facts keep their force however old they are. A preference
+ * like "always answer in Chinese" does not expire by calendar: age used to
+ * relabel it "待复核，不能当作当前已确认事实" after the review window, and
+ * restating it never refreshed the clock, so a standing preference silently
+ * stopped binding after six months. Project/decision/topic records change with
+ * the work, so for them age is a real signal. Only the injected label and the
+ * ranking penalty use this; `isReviewDue` still feeds the user's review queue.
+ */
+export function agingDowngrades(record: MemoryRecordV2): boolean {
+  return record.category !== 'preference' && record.category !== 'profile';
+}
+
+/** The label and ranking view of staleness: explicit marks always count, age only where it is evidence. */
+export function isUnconfirmed(
+  record: MemoryRecordV2,
+  settings: MemorySettings,
+  now = Date.now(),
+): boolean {
+  return record.status === 'needs_review'
+    || (agingDowngrades(record) && isReviewDue(record, settings, now));
+}
+
 export function isReviewDue(
   record: MemoryRecordV2,
   settings: MemorySettings,
@@ -38,10 +61,16 @@ export function scoreMemory(
   settings: MemorySettings,
 ): number {
   if (record.status === 'superseded' || (record.validUntil !== undefined && record.validUntil <= Date.now())) return Number.NEGATIVE_INFINITY;
+  const stalePenalty = isUnconfirmed(record, settings) ? 8 : 0;
+  const { terms, phrase } = substringParts(record, query);
+  return terms + phrase - stalePenalty + record.lastConfirmedAt / 1e13;
+}
+
+/** Substring evidence: 10 per matching query term, 25 for the whole phrase. */
+function substringParts(record: MemoryRecordV2, query: string): { terms: number; phrase: number } {
   const haystack =
     `${record.category} ${record.subject} ${record.topicKey} ${record.aliases.join(' ')} ${record.statement}`.normalize('NFKC').toLocaleLowerCase();
-  const terms = memoryTerms(query);
-  const lexical = terms.reduce(
+  const lexical = memoryTerms(query).reduce(
     (sum, term) => sum + (haystack.includes(term) ? 10 : 0),
     0,
   );
@@ -49,9 +78,7 @@ export function scoreMemory(
     query.trim() && haystack.includes(query.trim().normalize('NFKC').toLocaleLowerCase())
       ? 25
       : 0;
-  const stalePenalty =
-    isReviewDue(record, settings) || record.status === 'needs_review' ? 8 : 0;
-  return lexical + phrase - stalePenalty + record.lastConfirmedAt / 1e13;
+  return { terms: lexical, phrase };
 }
 
 export function recallFromSnapshot(
@@ -93,18 +120,26 @@ export function recallFromSnapshot(
     }
     recordSearchStrategy('bm25', eligible.length);
     const bm25 = index.search(query);
-    const normalized = query.trim().normalize('NFKC').toLocaleLowerCase();
+    // BM25 only knows whole tokens, so `postgres` never reaches a record that
+    // says `PostgreSQL`; the small-corpus path finds it by substring. Growing
+    // past the threshold must not make a findable record vanish, so substring
+    // hits stay candidates. A token hit keeps its original score (BM25's IDF
+    // is what puts a rare word above two common ones, and the flat per-term
+    // substring bonus would flatten it); substring evidence only scores the
+    // records BM25 could not see.
     return eligible
       .map((record) => {
-        const haystack = `${record.category} ${record.subject} ${record.topicKey} ${record.aliases.join(' ')} ${record.statement}`.normalize('NFKC').toLocaleLowerCase();
-        const phrase = normalized && haystack.includes(normalized) ? 25 : 0;
-        const stalePenalty = isReviewDue(record, settings) || record.status === 'needs_review' ? 8 : 0;
+        const { terms, phrase } = substringParts(record, query);
+        const tokens = bm25.get(record.id) ?? 0;
+        const stalePenalty = isUnconfirmed(record, settings) ? 8 : 0;
         return {
           record,
-          score: (bm25.get(record.id) ?? 0) * 10 + phrase - stalePenalty + record.lastConfirmedAt / 1e13,
+          hit: tokens > 0 || terms + phrase >= 2,
+          score: (tokens > 0 ? tokens * 10 + phrase : terms + phrase)
+            - stalePenalty + record.lastConfirmedAt / 1e13,
         };
       })
-      .filter((item) => (bm25.get(item.record.id) ?? 0) > 0)
+      .filter((item) => item.hit)
       .sort((a, b) => b.score - a.score)
       .slice(0, Math.min(100, Math.max(1, limit)))
       .map((item) => item.record);

package/src/memory/turnJournal.js

@@ -30,6 +30,7 @@ import { scheduleMemoryMaintenance } from './learning.js';
 import { enqueueFactTurn, factLearningStatus, scheduleFactLearning } from './factLearning.js';
 import { captureDependencies } from './dependencies.js';
 import { getWorkingDir } from '../bootstrap/workingDir.js';
+import { FENCE_CLOSE, FENCE_OPEN } from '../prompts/visualPresentation.js';
 import { isReadOnlyShellCommand } from '../tools/bash/semantics.js';
 function messageText(message) {
     if (typeof message.content === 'string')
@@ -48,6 +49,31 @@ function lastAssistantMessage(messages, start) {
     }
     return undefined;
 }
+/**
+ * The turn's answer as the user read it. When a live answer paused before a
+ * text-drawn diagram, the part already shown was committed as its own
+ * text-only progress message and the model continued in the final one; both
+ * halves are the answer. Only contiguous text-only assistant messages right
+ * before the last one join it (ephemeral notices in between are skipped); tool
+ * rounds and user messages end the run.
+ */
+function turnAnswerText(messages, start) {
+    let index = messages.length - 1;
+    while (index >= start && messages[index]?.role !== 'assistant')
+        index -= 1;
+    if (index < start)
+        return '';
+    const parts = [messageText(messages[index])];
+    for (let cursor = index - 1; cursor >= start; cursor -= 1) {
+        const message = messages[cursor];
+        if (message.role === 'system' && message._ephemeral)
+            continue;
+        if (message.role !== 'assistant' || message.tool_calls?.length || message.presentation !== 'progress')
+            break;
+        parts.unshift(messageText(message));
+    }
+    return parts.join('');
+}
 const PURE_CONTINUATION = /^(?:继续|接着(?:来|做)?|好的?|行|嗯+|ok(?:ay)?|go on|continue|next|下一步|开始吧?|谢谢|thanks?)[。.!！~\s]*$/i;
 /** Both current events and legacy previews use the same minimal evidence fields. */
 function evidenceArguments(raw, toolName) {
@@ -84,18 +110,56 @@ function evidenceArguments(raw, toolName) {
  *  - 斜杠命令：是给宿主的指令，不是陈述；
  *  - 纯续接词（「继续」「好的」）：提示词明确要求不展开成事实。明确确认（「是的」
  *    「按这个」）不在此列——它们命中下面的旧信号，是确认前文方案的唯一入口；
- *  - 长粘贴且无明确信号：粘进来的邮件/文档里的第一人称句子能通过逐字引用闸门，
+ *  - 长输入且无明确信号：粘进来的邮件/文档里的第一人称句子能通过逐字引用闸门，
  *    「引述他人不算」只靠提示词挡。这条是为准确性，不是为省调用。
+ *    曾经整条丢弃，于是 2300 字的消息末尾那句「以后回答简洁点」永远进不了队列。
+ *    现在只交出**用户自己写的部分**：去掉围栏代码块和 `>` 引用行（那是贴进来的资料，
+ *    不是陈述）；仍超长就取末尾——长消息里的诉求几乎总在最后。不新增任何模型判断，
+ *    逐字证据闸门不变，且交出去的文本就是证据来源，引用只能出自这一段。
  */
-function shouldExtractFacts(userInput, explicitSignal) {
+const FACT_TEXT_MAX_CHARS = 2_000;
+export function factTurnText(userInput, explicitSignal) {
     const value = userInput.trim();
     if (!value || value.startsWith('/'))
-        return false;
+        return undefined;
     if (explicitSignal)
-        return true;
+        return userInput;
     if (PURE_CONTINUATION.test(value))
-        return false;
-    return value.length <= 2_000;
+        return undefined;
+    if (value.length <= FACT_TEXT_MAX_CHARS)
+        return userInput;
+    const written = userWrittenLines(value).join('\n').trim();
+    if (!written)
+        return undefined;
+    if (written.length <= FACT_TEXT_MAX_CHARS)
+        return written;
+    const tail = written.slice(-FACT_TEXT_MAX_CHARS);
+    // Start at a line boundary so the first sentence is not cut mid-way.
+    const boundary = tail.indexOf('\n');
+    return boundary >= 0 && boundary < tail.length - 200 ? tail.slice(boundary + 1) : tail;
+}
+/** Lines outside fenced blocks and `>` quotes: what the user actually wrote. */
+function userWrittenLines(text) {
+    const kept = [];
+    let marker;
+    for (const raw of text.split('\n')) {
+        const line = raw.replace(/\r$/, '');
+        if (marker) {
+            const close = FENCE_CLOSE.exec(line);
+            if (close && close[1][0] === marker[0] && close[1].length >= marker.length)
+                marker = undefined;
+            continue;
+        }
+        const open = FENCE_OPEN.exec(line);
+        if (open) {
+            marker = open[1];
+            continue;
+        }
+        if (/^\s*>/.test(line))
+            continue;
+        kept.push(line);
+    }
+    return kept;
 }
 /**
  * 会话原文的保留清理：节流后台执行，绝不阻塞本回合。
@@ -278,8 +342,7 @@ export async function startTurnJournal(input) {
                 return;
             finished = true;
             structuredInputs.clear();
-            const assistant = lastAssistantMessage(history, startIndex);
-            const content = assistant ? messageText(assistant) : '';
+            const content = turnAnswerText(history, startIndex);
             if (content)
                 await episodicMemoryService.append({
                     ...target,
@@ -314,11 +377,13 @@ export async function startTurnJournal(input) {
             // "记住我喜欢简洁回答" answered with "好的，已记住" but no tool call landed
             // as claims_unverified and the fact was silently dropped — the loop broke
             // in exactly the case memory exists for.
-            if (settings.factLearning && !userInterrupted && shouldExtractFacts(input.userInput, userSignalsChange)) {
+            const factText = settings.factLearning && !userInterrupted
+                ? factTurnText(input.userInput, userSignalsChange) : undefined;
+            if (factText) {
                 // Separate stores, separate failures: a corrupt fact queue must not
                 // also cost this turn its experience source below.
                 const queued = await enqueueFactTurn(target.workspaceId, {
-                    id: turnId, sessionId: target.sessionId, text: input.userInput, createdAt,
+                    id: turnId, sessionId: target.sessionId, text: factText, createdAt,
                     handledQuotes: [...handledQuotes],
                     handledTopics: [...handledTopics.values()],
                     confirmationContext: priorAssistant,

package/src/memory/turnJournal.ts

@@ -37,6 +37,7 @@ import type { Message } from '../types.js';
 import { enqueueFactTurn, factLearningStatus, scheduleFactLearning } from './factLearning.js';
 import { captureDependencies } from './dependencies.js';
 import { getWorkingDir } from '../bootstrap/workingDir.js';
+import { FENCE_CLOSE, FENCE_OPEN } from '../prompts/visualPresentation.js';
 import { isReadOnlyShellCommand } from '../tools/bash/semantics.js';
 
 export interface TurnJournalInput {
@@ -79,6 +80,28 @@ function lastAssistantMessage(
   return undefined;
 }
 
+/**
+ * The turn's answer as the user read it. When a live answer paused before a
+ * text-drawn diagram, the part already shown was committed as its own
+ * text-only progress message and the model continued in the final one; both
+ * halves are the answer. Only contiguous text-only assistant messages right
+ * before the last one join it (ephemeral notices in between are skipped); tool
+ * rounds and user messages end the run.
+ */
+function turnAnswerText(messages: readonly Message[], start: number): string {
+  let index = messages.length - 1;
+  while (index >= start && messages[index]?.role !== 'assistant') index -= 1;
+  if (index < start) return '';
+  const parts = [messageText(messages[index]!)];
+  for (let cursor = index - 1; cursor >= start; cursor -= 1) {
+    const message = messages[cursor]!;
+    if (message.role === 'system' && message._ephemeral) continue;
+    if (message.role !== 'assistant' || message.tool_calls?.length || message.presentation !== 'progress') break;
+    parts.unshift(messageText(message));
+  }
+  return parts.join('');
+}
+
 const PURE_CONTINUATION =
   /^(?:继续|接着(?:来|做)?|好的?|行|嗯+|ok(?:ay)?|go on|continue|next|下一步|开始吧?|谢谢|thanks?)[。.!！~\s]*$/i;
 
@@ -114,15 +137,47 @@ function evidenceArguments(raw: unknown, toolName: string): Record<string, unkno
  *  - 斜杠命令：是给宿主的指令，不是陈述；
  *  - 纯续接词（「继续」「好的」）：提示词明确要求不展开成事实。明确确认（「是的」
  *    「按这个」）不在此列——它们命中下面的旧信号，是确认前文方案的唯一入口；
- *  - 长粘贴且无明确信号：粘进来的邮件/文档里的第一人称句子能通过逐字引用闸门，
+ *  - 长输入且无明确信号：粘进来的邮件/文档里的第一人称句子能通过逐字引用闸门，
  *    「引述他人不算」只靠提示词挡。这条是为准确性，不是为省调用。
+ *    曾经整条丢弃，于是 2300 字的消息末尾那句「以后回答简洁点」永远进不了队列。
+ *    现在只交出**用户自己写的部分**：去掉围栏代码块和 `>` 引用行（那是贴进来的资料，
+ *    不是陈述）；仍超长就取末尾——长消息里的诉求几乎总在最后。不新增任何模型判断，
+ *    逐字证据闸门不变，且交出去的文本就是证据来源，引用只能出自这一段。
  */
-function shouldExtractFacts(userInput: string, explicitSignal: boolean): boolean {
+const FACT_TEXT_MAX_CHARS = 2_000;
+
+export function factTurnText(userInput: string, explicitSignal: boolean): string | undefined {
   const value = userInput.trim();
-  if (!value || value.startsWith('/')) return false;
-  if (explicitSignal) return true;
-  if (PURE_CONTINUATION.test(value)) return false;
-  return value.length <= 2_000;
+  if (!value || value.startsWith('/')) return undefined;
+  if (explicitSignal) return userInput;
+  if (PURE_CONTINUATION.test(value)) return undefined;
+  if (value.length <= FACT_TEXT_MAX_CHARS) return userInput;
+  const written = userWrittenLines(value).join('\n').trim();
+  if (!written) return undefined;
+  if (written.length <= FACT_TEXT_MAX_CHARS) return written;
+  const tail = written.slice(-FACT_TEXT_MAX_CHARS);
+  // Start at a line boundary so the first sentence is not cut mid-way.
+  const boundary = tail.indexOf('\n');
+  return boundary >= 0 && boundary < tail.length - 200 ? tail.slice(boundary + 1) : tail;
+}
+
+/** Lines outside fenced blocks and `>` quotes: what the user actually wrote. */
+function userWrittenLines(text: string): string[] {
+  const kept: string[] = [];
+  let marker: string | undefined;
+  for (const raw of text.split('\n')) {
+    const line = raw.replace(/\r$/, '');
+    if (marker) {
+      const close = FENCE_CLOSE.exec(line);
+      if (close && close[1]![0] === marker[0] && close[1]!.length >= marker.length) marker = undefined;
+      continue;
+    }
+    const open = FENCE_OPEN.exec(line);
+    if (open) { marker = open[1]!; continue; }
+    if (/^\s*>/.test(line)) continue;
+    kept.push(line);
+  }
+  return kept;
 }
 
 /**
@@ -284,8 +339,7 @@ export async function startTurnJournal(
       if (finished) return;
       finished = true;
       structuredInputs.clear();
-      const assistant = lastAssistantMessage(history, startIndex);
-      const content = assistant ? messageText(assistant) : '';
+      const content = turnAnswerText(history, startIndex);
       if (content) await episodicMemoryService.append({
         ...target,
         role: 'assistant',
@@ -315,11 +369,13 @@ export async function startTurnJournal(
       // "记住我喜欢简洁回答" answered with "好的，已记住" but no tool call landed
       // as claims_unverified and the fact was silently dropped — the loop broke
       // in exactly the case memory exists for.
-      if (settings.factLearning && !userInterrupted && shouldExtractFacts(input.userInput, userSignalsChange)) {
+      const factText = settings.factLearning && !userInterrupted
+        ? factTurnText(input.userInput, userSignalsChange) : undefined;
+      if (factText) {
         // Separate stores, separate failures: a corrupt fact queue must not
         // also cost this turn its experience source below.
         const queued = await enqueueFactTurn(target.workspaceId, {
-          id: turnId, sessionId: target.sessionId, text: input.userInput, createdAt,
+          id: turnId, sessionId: target.sessionId, text: factText, createdAt,
           handledQuotes: [...handledQuotes],
           handledTopics: [...handledTopics.values()],
           confirmationContext: priorAssistant,

package/src/memory/types.ts

@@ -172,7 +172,7 @@ export interface MemoryContextBundle {
   includedIds: string[];
   estimatedTokens: number;
   sections: Array<{
-    id: 'persona' | 'preferences' | 'relevant' | 'projectKnowledge' | 'experiences' | 'health';
+    id: 'persona' | 'preferences' | 'profile' | 'relevant' | 'projectKnowledge' | 'experiences' | 'index' | 'health';
     content: string;
     includedIds: string[];
     estimatedTokens: number;
@@ -189,6 +189,12 @@ export interface ConversationEpisode {
   content: string;
   createdAt: number;
   contentHash: string;
+  /**
+   * Set when one message was stored as several consecutive episodes sharing a
+   * `messageId` (each ≤ the per-episode cap). `omittedChars` marks a message so
+   * long that its middle was dropped, so a reader knows the text is not whole.
+   */
+  part?: { index: number; total: number; omittedChars?: number };
 }
 
 export interface ConversationSearchOptions {

package/src/prompts/delegation.js

@@ -1,3 +1,4 @@
+import { CONFIG_LIMITS } from '../config/numericLimits.js';
 /** One delegation contract shared by the main prompt, tool discovery and runtime. */
 export const RESEARCH_DELEGATION_POLICY = `# 大范围阅读与并行调研
 - 先读项目说明、目录、入口与基线，建立问题地图，不先串行读完整仓库。批量阅读涉及多个独立模块、各需多轮探索时，优先同轮派发多个 Agent(agent_type="explore")；数量取决于独立工作量和宿主额度，不按文件数量机械切块。
@@ -11,5 +12,5 @@ export function delegationRuntimeGuidance(input) {
     const remaining = Math.max(0, input.maxCalls - input.callsUsed);
     if (!remaining)
         return '本轮子任务派发次数已用完：收口已有子任务，其余调查由主 Agent 使用当前工具完成，不重复尝试派发。';
-    return `本轮可使用 Agent 进行隔离委派，剩余派发次数 ${remaining}，配置并发上限 ${Math.max(1, Math.min(8, input.maxConcurrent))}（实际槽位由调度器协调）。大型批量阅读先建立问题地图，再优先并行派发独立 explore 调查；不要把所有模块留在主上下文串行阅读。少量已知文件仍直接批量读。`;
+    return `本轮可使用 Agent 进行隔离委派，剩余派发次数 ${remaining}，配置并发上限 ${Math.max(1, Math.min(CONFIG_LIMITS.subagentMaxConcurrent.max, input.maxConcurrent))}（实际槽位由调度器协调）。大型批量阅读先建立问题地图，再优先并行派发独立 explore 调查；不要把所有模块留在主上下文串行阅读。少量已知文件仍直接批量读。`;
 }

package/src/prompts/delegation.ts

@@ -1,3 +1,5 @@
+import { CONFIG_LIMITS } from '../config/numericLimits.js';
+
 /** One delegation contract shared by the main prompt, tool discovery and runtime. */
 export const RESEARCH_DELEGATION_POLICY = `# 大范围阅读与并行调研
 - 先读项目说明、目录、入口与基线，建立问题地图，不先串行读完整仓库。批量阅读涉及多个独立模块、各需多轮探索时，优先同轮派发多个 Agent(agent_type="explore")；数量取决于独立工作量和宿主额度，不按文件数量机械切块。
@@ -16,5 +18,5 @@ export function delegationRuntimeGuidance(input: {
   if (!input.available) return '';
   const remaining = Math.max(0, input.maxCalls - input.callsUsed);
   if (!remaining) return '本轮子任务派发次数已用完：收口已有子任务，其余调查由主 Agent 使用当前工具完成，不重复尝试派发。';
-  return `本轮可使用 Agent 进行隔离委派，剩余派发次数 ${remaining}，配置并发上限 ${Math.max(1, Math.min(8, input.maxConcurrent))}（实际槽位由调度器协调）。大型批量阅读先建立问题地图，再优先并行派发独立 explore 调查；不要把所有模块留在主上下文串行阅读。少量已知文件仍直接批量读。`;
+  return `本轮可使用 Agent 进行隔离委派，剩余派发次数 ${remaining}，配置并发上限 ${Math.max(1, Math.min(CONFIG_LIMITS.subagentMaxConcurrent.max, input.maxConcurrent))}（实际槽位由调度器协调）。大型批量阅读先建立问题地图，再优先并行派发独立 explore 调查；不要把所有模块留在主上下文串行阅读。少量已知文件仍直接批量读。`;
 }

package/src/prompts/policyRegistry.js

@@ -118,7 +118,7 @@ const ROUTABLE_TOOLS = [
     'EnterWorktree', 'ExitWorktree', 'TerminalCapture', 'Snip', 'OptimizeConversation',
 ];
 const PARAMETERIZED_TOOLS = [
-    ...ROUTABLE_TOOLS, ...MEMORY_TOOLS, 'AskUserQuestion', 'Advisor', 'AdvisorDecision', 'Agent',
+    ...ROUTABLE_TOOLS, ...MEMORY_TOOLS, 'AskUserQuestion', 'Advisor', 'Agent',
     'PresentResult', 'GenerateImage', 'GenerateSpeech',
 ];
 const CAPABILITY_POLICIES = [
@@ -199,7 +199,7 @@ const CAPABILITY_POLICIES = [
         render: ({ surface, output, toolNames }) => [
             visualPresentation(output ?? surface),
             toolNames.has('Skill') || toolNames.has('Read')
-                ? '- 配图前加载可用的 dynamic-ui。'
+                ? '- 需要配图时从技能目录加载匹配的图形技能。'
                 : '',
         ].filter(Boolean).join('\n'),
     },
@@ -249,14 +249,11 @@ const CAPABILITY_POLICIES = [
     {
         id: 'capability.advisor', layer: 'capability', profiles: MAIN_PROFILE, source: 'native',
         requiresTools: ['Advisor'],
-        // 注意：文本内不得硬编码其他工具名（如 PresentResult），否则该工具缺席时这里
-        // 成为唯一残留的「幽灵引用」——capability 规则交叉引用必须在 toolNames 条件下拼接。
-        render: ({ toolNames }) => `- Advisor（顾问模型）使用规则：提供第二意见，由主 Agent 核实取舍；不授予权限、不替用户验收。
-  咨询时机：实质计划执行前（plan_review）、实质交付前（delivery_review）、同一问题两次或以上处理失败后（stalled）、用户明显不满意（user_dissatisfaction）、非常不确定或无法解决（uncertainty）、关键证据冲突无法消解（evidence_conflict）。换工具不清零失败次数，无关失败不合并。
-  文字调研、计算、比较和推荐也是交付；自信、没文件不豁免。自动咨询不再问用户；寒暄、进度、拼写修正无需咨询。
-  同阶段不重复问；成果、假设实质变化或新冲突才重开。questions/note 指明疑点与证据 ID，历史由宿主附带。缺材料先补查，不算验证通过。读完意见再操作；暂缓操作未执行，提醒不算咨询。
-  不设咨询配额，每次咨询只尝试一次；失败、空答或超时自行降级继续，不转成人工审批，也不宣称获得验证。
-  以代码、测试和用户要求核实意见，${toolNames.has('AdvisorDecision') ? '可用 AdvisorDecision 记录 accept/partial/reject；登记失败不阻止执行。' : '直接说明采纳判断。'}顾问不是事实来源，冲突以实测为准。`,
+        render: () => `- Advisor（顾问模型）使用规则：独立的第二意见，由主 Agent 核实取舍；不授予权限、不替用户验收。它能看到整段对话（含你的思考和全部工具结果），无需复述；没有工具，不能替你验证。
+  咨询时机：实质工作之前（先调查清楚，再在写、改、定方案之前问）；认为完成时（实质交付或下结论之前）；同一问题两次或以上处理失败后；想改变既定方案或方向前；非常不确定、关键证据冲突无法消解、用户明显不满意时。换工具不清零失败次数，无关失败不合并。
+  文字调研、计算、比较和推荐也是交付，自信、没文件不豁免；简短即时问答、寒暄、进度、拼写修正无需咨询；自动咨询不问用户。
+  宿主只会在动态上下文里提醒你还没问，不拦截工具也不替你调用；是否咨询由你判断。每次咨询只尝试一次，不设配额；失败、空答或超时即自行继续，不转成人工审批，也不宣称获得验证。note 只写想被挑战的疑点或关键原文；缺材料先补查，不算验证通过。读完意见再操作。
+  以代码、测试和用户要求核实意见：一手证据或用户明确要求否定才拒绝；自测通过不能证明意见错误，它可能没测到所指问题。已有证据与意见方向不同又无法判定时，不要默默改道，写明冲突再咨询一次。顾问不是事实来源。`,
     },
     {
         id: 'capability.subagent', layer: 'capability', profiles: MAIN_PROFILE, source: 'native',
@@ -271,7 +268,7 @@ const CAPABILITY_POLICIES = [
     {
         id: 'capability.agent-write-plan', layer: 'capability', profiles: MAIN_PROFILE, source: 'native',
         requiresTools: ['AgentResolve'],
-        render: ({ toolNames }) => `- 并行写任务先由主 Agent 铺路：确认需求和验收，检查当前改动，统一共享接口、文件归属、依赖与基线。${toolNames.has('Advisor') ? '复杂/高风险设计按顾问规则咨询 Advisor 并记录采纳判断；不可用或仅手动模式时自行核实，不等待顾问。' : '复杂设计由主 Agent 核实证据和取舍，不等待未提供的顾问能力。'}小型独立修改不强制走大流程。先完成前置接口再派发依赖实现，每份任务书携带 context.baseline/acceptance/interfaces/verification、决定及证据来源和 write_scopes。共享目录仅并行不重叠的局部编辑，构建与跨文件脚本由主 Agent 在子任务退出后统一执行；需独立构建用 worktree，但它不含未提交的铺路代码，不能为方便擅自提交用户改动。共享接口、锁文件、迁移与最终集成归主 Agent 单一负责。收到求助先决定接口或范围，必要时停止旧任务重新派发；不要让多个子 Agent 自行协调扩大范围。子任务结果是候选改动：检查差异、集成到目标工作区、处理冲突并验证组合行为后，才 AgentResolve accept/complete，写任务必须填写 writeReview 及主 Agent 的真实检查证据。接管 worktree 任务先检查其分支、未提交改动和后台进程，不把失败现场当主工作区已修改。最终只汇总设计决定、实际集成、验证证据、风险与未完成项，不拼接完整子历史。`,
+        render: ({ toolNames }) => `- 并行写任务先由主 Agent 铺路：确认需求和验收，检查当前改动，统一共享接口、文件归属、依赖与基线。${toolNames.has('Advisor') ? '复杂/高风险设计按顾问规则咨询 Advisor；不可用或仅手动模式时自行核实，不等待顾问。' : '复杂设计由主 Agent 核实证据和取舍，不等待未提供的顾问能力。'}小型独立修改不强制走大流程。先完成前置接口再派发依赖实现，每份任务书携带 context.baseline/acceptance/interfaces/verification、决定及证据来源和 write_scopes。共享目录仅并行不重叠的局部编辑，构建与跨文件脚本由主 Agent 在子任务退出后统一执行；需独立构建用 worktree，但它不含未提交的铺路代码，不能为方便擅自提交用户改动。共享接口、锁文件、迁移与最终集成归主 Agent 单一负责。收到求助先决定接口或范围，必要时停止旧任务重新派发；不要让多个子 Agent 自行协调扩大范围。子任务结果是候选改动：检查差异、集成到目标工作区、处理冲突并验证组合行为后，才 AgentResolve accept/complete，写任务必须填写 writeReview 及主 Agent 的真实检查证据。接管 worktree 任务先检查其分支、未提交改动和后台进程，不把失败现场当主工作区已修改。最终只汇总设计决定、实际集成、验证证据、风险与未完成项，不拼接完整子历史。`,
     },
     {
         id: 'capability.delivery', layer: 'capability', profiles: MAIN_PROFILE, source: '05',
@@ -308,7 +305,7 @@ const CAPABILITY_POLICIES = [
             if (toolNames.has('SearchConversations'))
                 rules.push('提及共享历史时用 SearchConversations 按内容关键词定位，片段不足才 read 和展开邻接消息；证据充分即答。无命中可换关键词，不能声称从未讨论；原文缺失须说明。');
             if (toolNames.has('MemoryMaintenance')) {
-                rules.push('项目经验须核对环境、失效条件和依赖；召回不等于采用，不为记忆重复任务。具体维护动作以 MemoryMaintenance 工具说明为准。', '有写权限且证据充分才用 resolve/resolveKnowledge 处理候选；后台失败不等于已保存，遗忘条目不等于删除原文。');
+                rules.push('项目经验须核对环境、失效条件和依赖；召回不等于采用，不为记忆重复任务。具体维护动作以 MemoryMaintenance 工具说明为准。', '事实候选经用户确认后 resolve 结清；UpdateMemory 改写记录即失效。后台失败不等于已保存，遗忘条目不等于删除原文。');
             }
             return rules.map((rule) => `- ${rule}`).join('\n');
         },

package/src/prompts/policyRegistry.ts

@@ -170,7 +170,7 @@ const ROUTABLE_TOOLS = [
 ];
 
 const PARAMETERIZED_TOOLS = [
-  ...ROUTABLE_TOOLS, ...MEMORY_TOOLS, 'AskUserQuestion', 'Advisor', 'AdvisorDecision', 'Agent',
+  ...ROUTABLE_TOOLS, ...MEMORY_TOOLS, 'AskUserQuestion', 'Advisor', 'Agent',
   'PresentResult', 'GenerateImage', 'GenerateSpeech',
 ];
 
@@ -248,7 +248,7 @@ const CAPABILITY_POLICIES: PromptPolicySpec[] = [
     render: ({ surface, output, toolNames }) => [
       visualPresentation(output ?? surface),
       toolNames.has('Skill') || toolNames.has('Read')
-        ? '- 配图前加载可用的 dynamic-ui。'
+        ? '- 需要配图时从技能目录加载匹配的图形技能。'
         : '',
     ].filter(Boolean).join('\n'),
   },
@@ -298,14 +298,11 @@ const CAPABILITY_POLICIES: PromptPolicySpec[] = [
   {
     id: 'capability.advisor', layer: 'capability', profiles: MAIN_PROFILE, source: 'native',
     requiresTools: ['Advisor'],
-    // 注意：文本内不得硬编码其他工具名（如 PresentResult），否则该工具缺席时这里
-    // 成为唯一残留的「幽灵引用」——capability 规则交叉引用必须在 toolNames 条件下拼接。
-    render: ({ toolNames }) => `- Advisor（顾问模型）使用规则：提供第二意见，由主 Agent 核实取舍；不授予权限、不替用户验收。
-  咨询时机：实质计划执行前（plan_review）、实质交付前（delivery_review）、同一问题两次或以上处理失败后（stalled）、用户明显不满意（user_dissatisfaction）、非常不确定或无法解决（uncertainty）、关键证据冲突无法消解（evidence_conflict）。换工具不清零失败次数，无关失败不合并。
-  文字调研、计算、比较和推荐也是交付；自信、没文件不豁免。自动咨询不再问用户；寒暄、进度、拼写修正无需咨询。
-  同阶段不重复问；成果、假设实质变化或新冲突才重开。questions/note 指明疑点与证据 ID，历史由宿主附带。缺材料先补查，不算验证通过。读完意见再操作；暂缓操作未执行，提醒不算咨询。
-  不设咨询配额，每次咨询只尝试一次；失败、空答或超时自行降级继续，不转成人工审批，也不宣称获得验证。
-  以代码、测试和用户要求核实意见，${toolNames.has('AdvisorDecision') ? '可用 AdvisorDecision 记录 accept/partial/reject；登记失败不阻止执行。' : '直接说明采纳判断。'}顾问不是事实来源，冲突以实测为准。`,
+    render: () => `- Advisor（顾问模型）使用规则：独立的第二意见，由主 Agent 核实取舍；不授予权限、不替用户验收。它能看到整段对话（含你的思考和全部工具结果），无需复述；没有工具，不能替你验证。
+  咨询时机：实质工作之前（先调查清楚，再在写、改、定方案之前问）；认为完成时（实质交付或下结论之前）；同一问题两次或以上处理失败后；想改变既定方案或方向前；非常不确定、关键证据冲突无法消解、用户明显不满意时。换工具不清零失败次数，无关失败不合并。
+  文字调研、计算、比较和推荐也是交付，自信、没文件不豁免；简短即时问答、寒暄、进度、拼写修正无需咨询；自动咨询不问用户。
+  宿主只会在动态上下文里提醒你还没问，不拦截工具也不替你调用；是否咨询由你判断。每次咨询只尝试一次，不设配额；失败、空答或超时即自行继续，不转成人工审批，也不宣称获得验证。note 只写想被挑战的疑点或关键原文；缺材料先补查，不算验证通过。读完意见再操作。
+  以代码、测试和用户要求核实意见：一手证据或用户明确要求否定才拒绝；自测通过不能证明意见错误，它可能没测到所指问题。已有证据与意见方向不同又无法判定时，不要默默改道，写明冲突再咨询一次。顾问不是事实来源。`,
   },
   {
     id: 'capability.subagent', layer: 'capability', profiles: MAIN_PROFILE, source: 'native',
@@ -320,7 +317,7 @@ const CAPABILITY_POLICIES: PromptPolicySpec[] = [
   {
     id: 'capability.agent-write-plan', layer: 'capability', profiles: MAIN_PROFILE, source: 'native',
     requiresTools: ['AgentResolve'],
-    render: ({ toolNames }) => `- 并行写任务先由主 Agent 铺路：确认需求和验收，检查当前改动，统一共享接口、文件归属、依赖与基线。${toolNames.has('Advisor') ? '复杂/高风险设计按顾问规则咨询 Advisor 并记录采纳判断；不可用或仅手动模式时自行核实，不等待顾问。' : '复杂设计由主 Agent 核实证据和取舍，不等待未提供的顾问能力。'}小型独立修改不强制走大流程。先完成前置接口再派发依赖实现，每份任务书携带 context.baseline/acceptance/interfaces/verification、决定及证据来源和 write_scopes。共享目录仅并行不重叠的局部编辑，构建与跨文件脚本由主 Agent 在子任务退出后统一执行；需独立构建用 worktree，但它不含未提交的铺路代码，不能为方便擅自提交用户改动。共享接口、锁文件、迁移与最终集成归主 Agent 单一负责。收到求助先决定接口或范围，必要时停止旧任务重新派发；不要让多个子 Agent 自行协调扩大范围。子任务结果是候选改动：检查差异、集成到目标工作区、处理冲突并验证组合行为后，才 AgentResolve accept/complete，写任务必须填写 writeReview 及主 Agent 的真实检查证据。接管 worktree 任务先检查其分支、未提交改动和后台进程，不把失败现场当主工作区已修改。最终只汇总设计决定、实际集成、验证证据、风险与未完成项，不拼接完整子历史。`,
+    render: ({ toolNames }) => `- 并行写任务先由主 Agent 铺路：确认需求和验收，检查当前改动，统一共享接口、文件归属、依赖与基线。${toolNames.has('Advisor') ? '复杂/高风险设计按顾问规则咨询 Advisor；不可用或仅手动模式时自行核实，不等待顾问。' : '复杂设计由主 Agent 核实证据和取舍，不等待未提供的顾问能力。'}小型独立修改不强制走大流程。先完成前置接口再派发依赖实现，每份任务书携带 context.baseline/acceptance/interfaces/verification、决定及证据来源和 write_scopes。共享目录仅并行不重叠的局部编辑，构建与跨文件脚本由主 Agent 在子任务退出后统一执行；需独立构建用 worktree，但它不含未提交的铺路代码，不能为方便擅自提交用户改动。共享接口、锁文件、迁移与最终集成归主 Agent 单一负责。收到求助先决定接口或范围，必要时停止旧任务重新派发；不要让多个子 Agent 自行协调扩大范围。子任务结果是候选改动：检查差异、集成到目标工作区、处理冲突并验证组合行为后，才 AgentResolve accept/complete，写任务必须填写 writeReview 及主 Agent 的真实检查证据。接管 worktree 任务先检查其分支、未提交改动和后台进程，不把失败现场当主工作区已修改。最终只汇总设计决定、实际集成、验证证据、风险与未完成项，不拼接完整子历史。`,
   },
   {
     id: 'capability.delivery', layer: 'capability', profiles: MAIN_PROFILE, source: '05',
@@ -351,7 +348,7 @@ const CAPABILITY_POLICIES: PromptPolicySpec[] = [
       if (toolNames.has('MemoryMaintenance')) {
         rules.push(
           '项目经验须核对环境、失效条件和依赖；召回不等于采用，不为记忆重复任务。具体维护动作以 MemoryMaintenance 工具说明为准。',
-          '有写权限且证据充分才用 resolve/resolveKnowledge 处理候选；后台失败不等于已保存，遗忘条目不等于删除原文。',
+          '事实候选经用户确认后 resolve 结清；UpdateMemory 改写记录即失效。后台失败不等于已保存，遗忘条目不等于删除原文。',
         );
       }
       return rules.map((rule) => `- ${rule}`).join('\n');

package/src/prompts/promptAssembly.js

@@ -142,11 +142,7 @@ async function buildAdvisorStatusLine() {
         if (!advisor)
             return 'Advisor 状态：未配置。';
         const mode = advisor.mode === 'auto' ? '自动' : '仅手动';
-        const vision = advisor.supportsVision && advisor.shareSensitive
-            ? '允许图片'
-            : advisor.supportsVision
-                ? '支持图片但敏感共享关闭'
-                : '不支持图片';
+        const vision = advisor.supportsVision ? '支持图片' : '不支持图片';
         return `Advisor 状态：${advisor.enabled ? '已启用' : '已关闭'}，${advisor.provider.name}/${advisor.provider.model}，${mode}，${vision}，超时 ${Math.round(advisor.timeoutMs / 1000)} 秒。`;
     }
     catch {
@@ -219,6 +215,7 @@ async function buildMemoryTransient(runtime) {
         return memoryService.buildContextBundle(runtime.currentUserInput ?? '', runtime.workspaceId, {
             includePersona: runtime.inheritPersona !== false,
             budgetTokens: memoryBudgetTokens(runtime),
+            previousInput: runtime.previousUserInput,
         });
     }
     catch {
@@ -385,9 +382,13 @@ export async function buildTransientBundle(runtime, profile, tools, options = {}
         const priorities = {
             persona: 82,
             preferences: 78,
+            profile: 76,
             relevant: 70,
             projectKnowledge: 68,
             experiences: 65,
+            // The subject index only helps the model choose what to Recall; below
+            // skills (45) so a tight budget drops it before anything the model acts on.
+            index: 40,
             health: 10,
         };
         for (const section of memory.sections) {