@modusensus/dsh-mneme 0.8.2 → 0.8.4

package/src/inject.js

@@ -1,253 +1,299 @@
-import { createScopeResolver } from "./scope.js";
-import { createHotMemory } from "./hot-memory.js";
-import { STR, langOf } from "./lang.js";
-
-// Best-effort extraction of the current user's latest message text from the
-// live session, for semantic-first injection (Bug4). The system-prompt
-// interpolator renders synchronously, so this walks the already-materialized
-// session event log (same event shape summarize.js consumes) and returns the
-// most recent human message. Any failure degrades to "" — the injector then
-// falls back to the legacy rule-based pick, never breaking the render.
-function lastUserQuery(ctx) {
-  try {
-    const session = ctx?.agent?.session;
-    // DSH ≥0.1.2-rc only exposes events via snapshotEvents(); older builds
-    // still have the .events property, so fall back to it.
-    const events = session?.snapshotEvents?.() ?? session?.events;
-    if (!Array.isArray(events) || events.length === 0) return "";
-    for (let i = events.length - 1; i >= 0; i--) {
-      const event = events[i];
-      if (event?.type !== "user/message") continue;
-      const kind = event.data?.source?.kind;
-      if (kind !== undefined && kind !== "user") continue;
-      const parts = event.data?.content;
-      if (!Array.isArray(parts) || parts.length === 0) continue;
-      return parts
-        .map((p) => (typeof p === "string" ? p : p?.text ?? ""))
-        .filter(Boolean)
-        .join("\n")
-        .slice(0, 500);
-    }
-  } catch { /* session internals unavailable: degrade to no query */ }
-  return "";
-}
-
-// Hot-memory round extraction (v0.5.0 1.3): pairs each user/message with the
-// next assistant reply from the materialized session log. Tolerates shapes
-// where assistant events carry a different type tag — anything whose payload
-// has content parts and is not a user message counts as a reply. Best-effort:
-// returns [] on any failure, and the hot block simply does not render.
-function extractRounds(ctx, maxRounds) {
-  try {
-    const session = ctx?.agent?.session;
-    const events = session?.snapshotEvents?.() ?? session?.events;
-    if (!Array.isArray(events) || events.length === 0) return [];
-    const rounds = [];
-    let pendingQuery = null;
-    const textOf = (event) => {
-      // user/message 的正文就在 data 上，而 assistant/message（以及 tool/result）的消息体
-      // 嵌在 data.message 下 —— DSH 侧由 @deepseek-ai/dsh-session 的
-      // assertMessageEventShape 一行写死：const message = type === "user/message" ? record : record?.["message"]。
-      // 只读 data.content 会让每一轮的 response 恒为空（Issue #129）。
-      // 这里用 ?? 兜底而不是直接改成 data.message.content：user 消息与旧的扁平形状都照旧可用。
-      const parts = event?.data?.content ?? event?.data?.message?.content;
-      if (!Array.isArray(parts)) return "";
-      // 只取正文 part（type === "text" 或字符串），跳过 reasoning（issue #162）：
-      // reasoning 里出现 Go template / Vue / Handlebars / 正则 / 日志原文的概率远高于
-      // 正文，把它带进 hot memory 会让 `{{...}}` 模板片段注入 prompt。旧的扁平形状
-      // （字符串 part）照旧取用，无 type 字段的兜底视为正文。
-      return parts
-        .map((p) => {
-          if (typeof p === "string") return p;
-          if (p?.type && p.type !== "text") return "";
-          return p?.text ?? "";
-        })
-        .filter(Boolean)
-        .join("\n")
-        .trim();
-    };
-    for (const event of events) {
-      const kind = event?.data?.source?.kind;
-      const isUser = event?.type === "user/message" && (kind === undefined || kind === "user");
-      if (isUser) {
-        if (pendingQuery) rounds.push({ query: pendingQuery, response: "" });
-        pendingQuery = textOf(event).slice(0, 500);
-        continue;
-      }
-      // Only assistant-originated events close a round; tool/system events
-      // carrying text must not be mistaken for the model's reply.
-      const isAssistant = typeof event?.type === "string" && event.type.includes("assistant")
-        || kind === "assistant";
-      const body = isAssistant ? textOf(event) : "";
-      if (!body || !pendingQuery) continue;
-      rounds.push({ query: pendingQuery, response: body.slice(0, 800) });
-      pendingQuery = null;
-    }
-    if (pendingQuery) rounds.push({ query: pendingQuery, response: "" });
-    return rounds.slice(-maxRounds);
-  } catch {
-    return [];
-  }
-}
-
-export function createInjector(ctx, service, settings, config) {
-  const language = langOf(config);
-  const maxItems = config.maxInjectedItems ?? 5;
-  const threshold = config.importanceThreshold ?? 3;
-  // v0.8.0 A3（issue #17）：注入路径的会话 scope 解析——strictScope 硬过滤
-  // 需要。渲染 ctx 与工具 exec 同形（agent.session），解析器直接复用。
-  // logger 透传：registry 反查失败时 warnOnce 才有出口。
-  const resolveSessionScope = createScopeResolver({ ctx, config, logger: ctx.logger });
-
-  // Bug6: bound the injected memory block. Each entry's content is truncated to
-  // MAX_CONTENT chars (trailing `…`); the whole block gets a MAX_BLOCK budget
-  // and an entry that would exceed it collapses to its title only, so a long
-  // memory can never push the injected context past a few thousand chars.
-  const MAX_CONTENT = 300;
-  const MAX_BLOCK = 1500;
-
-  // Compressed injection (v0.5.0 2.1): a sleep-demoted row already carries its
-  // summary in `content` with the original parked in `_full_content` — inject
-  // the summary verbatim instead of re-truncating the (already short) text.
-  // Regular long rows keep the hard truncate.
-  function injectMemory(m, maxLength = MAX_CONTENT) {
-    if (m?._full_content) return String(m.content ?? "");
-    const text = String(m?.content ?? "");
-    return text.length <= maxLength ? text : `${text.slice(0, maxLength)}…`;
-  }
-
-  // Prompt-variable brace escaping (issue #162, restored from v0.7.4 #40):
-  // DSH's interpolate() scans every injected section for `{{...}}` and throws
-  // unless the variable name matches /^[a-z][a-z0-9_]*$/ — memory/profile
-  // content carrying legal template syntax (docker `--format "{{.Server.Version}}"`,
-  // Obsidian-style `{{hl|}}`, `{{挖空}}`, `{{关键词}}`) would hit an illegal
-  // variable name and throw, crashing the whole turn. At the injection boundary
-  // we escape every run of 2+ consecutive braces, inserting a `\` between each
-  // pair, so no `{{`/`}}` substring survives into the prompt: `{{a}}` → `{\{a\}\}`,
-  // and odd runs like `{{{a}}}` (which pair-wise escaping would leave with a
-  // literal `{{`) are handled too. The text keeps its readable template form,
-  // the transform is idempotent (escaped braces are single + `\`, never two
-  // adjacent), and single braces pass through untouched — interpolate only
-  // scans `{{`. Off via config.escapePromptVariables=false (default true).
-  function escapePromptVars(text) {
-    if (config.escapePromptVariables === false) return String(text);
-    return String(text).replace(/[{}]{2,}/g, (run) => run.split("").join("\\"));
-  }
-
-  // Hot memory (v0.5.0 1.3): the latest rounds of THIS session, rebuilt from
-  // the materialized event log on every render — stateless, so it survives
-  // session switches and never persists anywhere.
-  const hot = createHotMemory({
-    maxRounds: config.hotMemoryRounds ?? 5,
-    maxTokens: config.hotMemoryMaxTokens ?? 2000
-  });
-
-  function renderHotContext(ctx) {
-    if (config.hotMemoryEnabled === false) return "";
-    const rounds = extractRounds(ctx, config.hotMemoryRounds ?? 5);
-    if (!rounds.length) return "";
-    hot.clear();
-    for (const r of rounds) hot.add(r);
-    const body = hot.getContext();
-    if (!body) return "";
-    return escapePromptVars(STR.hotHeader[language](rounds.length, body));
-  }
-
-  function render(candidates) {
-    if (!candidates.length) return "";
-    const header = STR.memoryHeader[language];
-    const lines = [header];
-    let budget = MAX_BLOCK - header.length;
-    for (const m of candidates) {
-      // Epistemic trust (v0.4.5): when enabled, measured observations are
-      // flagged so the agent can weigh them above guesses/opinions.
-      const verified = config.trustEpistemicWeighting === true && m.epistemic_status === "observation"
-        ? STR.verified[language]
-        : "";
-      const title = STR.entryTitle[language](m.title, m.importance);
-      const content = injectMemory(m);
-      const full = STR.entryLine[language](m.type, verified, title, content);
-      if (budget - full.length >= 0) {
-        lines.push(full);
-        budget -= full.length;
-      } else {
-        lines.push(`- [${m.type}] ${verified}${title}`);
-      }
-    }
-    return escapePromptVars(lines.join("\n"));
-  }
-
-  // Bug4: the system-prompt render is synchronous, so the semantic query vector
-  // must be prefetched asynchronously and cached for the next assembly. The
-  // first render after a new user message may still fall back to the rule-based
-  // pick; later assemblies in the same session reuse the cached vector. Bounded
-  // cache (cap 8, drop oldest) so a long session never grows it unbounded.
-  const QUERY_VECTOR_CACHE_MAX = 8;
-  const queryVectorCache = new Map();
-  let lastPrefetched = "";
-
-  function prefetchQueryVector(query) {
-    if (!query || query === lastPrefetched || queryVectorCache.has(query)) return;
-    lastPrefetched = query;
-    service.embedQuery(query).then((vec) => {
-      if (Array.isArray(vec) && vec.length) {
-        queryVectorCache.set(query, vec);
-        if (queryVectorCache.size > QUERY_VECTOR_CACHE_MAX) {
-          queryVectorCache.delete(queryVectorCache.keys().next().value);
-        }
-      }
-    }).catch(() => { /* prefetch is best-effort */ });
-  }
-
-  // User profile + rules: injected ahead of the memory block because they are
-  // always-relevant instructions the agent should follow every turn.
-  function renderUserSettings() {
-    const profile = settings.getProfile().trim();
-    const rules = settings.getRules();
-    if (!profile && !rules.length) return "";
-    const lines = [STR.userSettingsHeader[language]];
-    if (profile) lines.push(STR.profileLine[language](profile));
-    for (const rule of rules) lines.push(STR.ruleLine[language](rule));
-    return escapePromptVars(lines.join("\n"));
-  }
-
-  const disposers = [
-    ctx.systemPrompt.context({
-      name: "memory",
-      order: 90,
-      text: (ctx) => {
-        // Bug4: pass the latest user query so injection prefers semantically
-        // relevant memories; lastUserQuery is best-effort (empty → legacy).
-        // The query vector is prefetched asynchronously (cached) because the
-        // render itself must stay synchronous.
-        const query = lastUserQuery(ctx);
-        if (query) prefetchQueryVector(query);
-        const queryVector = queryVectorCache.get(query);
-        // v0.8.0 A3：scope 随请求解析（strict 开启时 injectCandidates 内硬过滤；
-        // flag 关闭时解析器返回 null，注入行为不变）。
-        const scope = resolveSessionScope(ctx);
-        const candidates = service.injectCandidates({ query, queryVector, maxItems, threshold, scope });
-        // Hot memory (v0.5.0 1.3) leads the single memory block: the agent
-        // sees the short-term rounds first, then the cross-session recall —
-        // the documented injection order 1→2. Folding it here (instead of a
-        // separate context) keeps the prompt assembly stable at two blocks.
-        const hotText = renderHotContext(ctx);
-        const body = render(candidates);
-        if (!hotText) return body;
-        return body ? `${hotText}\n\n${body}` : hotText;
-      }
-    }),
-    ctx.systemPrompt.context({
-      name: "user-settings",
-      order: 85,
-      text: renderUserSettings
-    })
-  ];
-
-  return () => {
-    queryVectorCache.clear();
-    for (const dispose of disposers) {
-      if (typeof dispose === "function") dispose();
-    }
-  };
-}
+import { createScopeResolver } from "./scope.js";
+import { createHotMemory } from "./hot-memory.js";
+import { STR, langOf } from "./lang.js";
+
+// Best-effort extraction of the current user's latest message text from the
+// live session, for semantic-first injection (Bug4). The system-prompt
+// interpolator renders synchronously, so this walks the already-materialized
+// session event log (same event shape summarize.js consumes) and returns the
+// most recent human message. Any failure degrades to "" — the injector then
+// falls back to the legacy rule-based pick, never breaking the render.
+function lastUserQuery(ctx) {
+  try {
+    const session = ctx?.agent?.session;
+    // DSH ≥0.1.2-rc only exposes events via snapshotEvents(); older builds
+    // still have the .events property, so fall back to it.
+    const events = session?.snapshotEvents?.() ?? session?.events;
+    if (!Array.isArray(events) || events.length === 0) return "";
+    for (let i = events.length - 1; i >= 0; i--) {
+      const event = events[i];
+      if (event?.type !== "user/message") continue;
+      const kind = event.data?.source?.kind;
+      if (kind !== undefined && kind !== "user") continue;
+      const parts = event.data?.content;
+      if (!Array.isArray(parts) || parts.length === 0) continue;
+      return parts
+        .map((p) => (typeof p === "string" ? p : p?.text ?? ""))
+        .filter(Boolean)
+        .join("\n")
+        .slice(0, 500);
+    }
+  } catch { /* session internals unavailable: degrade to no query */ }
+  return "";
+}
+
+// Hot-memory round extraction (v0.5.0 1.3): pairs each user/message with the
+// next assistant reply from the materialized session log. Tolerates shapes
+// where assistant events carry a different type tag — anything whose payload
+// has content parts and is not a user message counts as a reply. Best-effort:
+// returns [] on any failure, and the hot block simply does not render.
+function extractRounds(ctx, maxRounds) {
+  try {
+    const session = ctx?.agent?.session;
+    const events = session?.snapshotEvents?.() ?? session?.events;
+    if (!Array.isArray(events) || events.length === 0) return [];
+    const rounds = [];
+    let pendingQuery = null;
+    const textOf = (event) => {
+      // user/message 的正文就在 data 上，而 assistant/message（以及 tool/result）的消息体
+      // 嵌在 data.message 下 —— DSH 侧由 @deepseek-ai/dsh-session 的
+      // assertMessageEventShape 一行写死：const message = type === "user/message" ? record : record?.["message"]。
+      // 只读 data.content 会让每一轮的 response 恒为空（Issue #129）。
+      // 这里用 ?? 兜底而不是直接改成 data.message.content：user 消息与旧的扁平形状都照旧可用。
+      const parts = event?.data?.content ?? event?.data?.message?.content;
+      if (!Array.isArray(parts)) return "";
+      // 只取正文 part（type === "text" 或字符串），跳过 reasoning（issue #162）：
+      // reasoning 里出现 Go template / Vue / Handlebars / 正则 / 日志原文的概率远高于
+      // 正文，把它带进 hot memory 会让 `{{...}}` 模板片段注入 prompt。旧的扁平形状
+      // （字符串 part）照旧取用，无 type 字段的兜底视为正文。
+      return parts
+        .map((p) => {
+          if (typeof p === "string") return p;
+          if (p?.type && p.type !== "text") return "";
+          return p?.text ?? "";
+        })
+        .filter(Boolean)
+        .join("\n")
+        .trim();
+    };
+    for (const event of events) {
+      const kind = event?.data?.source?.kind;
+      const isUser = event?.type === "user/message" && (kind === undefined || kind === "user");
+      if (isUser) {
+        if (pendingQuery) rounds.push({ query: pendingQuery, response: "" });
+        pendingQuery = textOf(event).slice(0, 500);
+        continue;
+      }
+      // Only assistant-originated events close a round; tool/system events
+      // carrying text must not be mistaken for the model's reply.
+      const isAssistant = typeof event?.type === "string" && event.type.includes("assistant")
+        || kind === "assistant";
+      const body = isAssistant ? textOf(event) : "";
+      if (!body || !pendingQuery) continue;
+      rounds.push({ query: pendingQuery, response: body.slice(0, 800) });
+      pendingQuery = null;
+    }
+    if (pendingQuery) rounds.push({ query: pendingQuery, response: "" });
+    return rounds.slice(-maxRounds);
+  } catch {
+    return [];
+  }
+}
+
+export function createInjector(ctx, service, settings, config) {
+  const language = langOf(config);
+  const maxItems = config.maxInjectedItems ?? 5;
+  const threshold = config.importanceThreshold ?? 3;
+  // Issue #205：注入位跨轮轮换。rotationTurns = 最近 N 个「不同用户查询」轮次
+  // 注入过的记忆本轮不再优先（0 = 关闭，保持既有行为）。历史按会话维护——
+  // 新会话从零开始；同一查询的多次渲染（工具调用轮）视为同一轮，不推进窗口。
+  const rotationTurns = Math.max(0, Math.floor(config.injectRotationTurns ?? 0));
+  const rotationHistory = new Map(); // sessionId -> [{ query, ids: Set }]
+  const ROTATION_HISTORY_MAX = 32; // 覆盖 max(20) 配置档仍有余量
+
+  function recentInjectedIds(sessionId, query) {
+    if (rotationTurns <= 0) return null;
+    const deque = rotationHistory.get(sessionId);
+    if (!deque || deque.length === 0) return null;
+    // 只看「之前的轮次」：排除与当前查询相同的条目——同一查询的重复渲染
+    // （工具调用轮）不该拿本轮自己的集合来转自己。
+    const prior = query ? deque.filter((e) => e.query !== query) : deque;
+    if (prior.length === 0) return null;
+    const recent = new Set();
+    for (const entry of prior.slice(-rotationTurns)) {
+      for (const id of entry.ids) recent.add(id);
+    }
+    return recent.size > 0 ? recent : null;
+  }
+
+  function recordInjection(sessionId, query, candidates) {
+    if (rotationTurns <= 0 || !query) return;
+    const deque = rotationHistory.get(sessionId) ?? [];
+    const ids = new Set(candidates.map((c) => c.id));
+    const last = deque[deque.length - 1];
+    if (last && last.query === query) {
+      last.ids = ids; // 同一查询的重复渲染：覆盖本轮内容，不推进窗口
+    } else {
+      deque.push({ query, ids });
+      if (deque.length > ROTATION_HISTORY_MAX) deque.shift();
+      rotationHistory.set(sessionId, deque);
+    }
+  }
+  // v0.8.0 A3（issue #17）：注入路径的会话 scope 解析——strictScope 硬过滤
+  // 需要。渲染 ctx 与工具 exec 同形（agent.session），解析器直接复用。
+  // logger 透传：registry 反查失败时 warnOnce 才有出口。
+  const resolveSessionScope = createScopeResolver({ ctx, config, logger: ctx.logger });
+
+  // Bug6: bound the injected memory block. Each entry's content is truncated to
+  // injectContentMaxChars chars (issue #164①: configurable, was a hardcoded
+  // 300); a truncated entry carries a tail hint (limit / original length /
+  // memory_get id) so the cut is never silent — BUDGET_EXCEEDED principle, the
+  // agent can always fetch the full text. The whole block gets a block budget
+  // (scales up with the per-entry cap so raising the cap is not defeated by a
+  // stale 1500) and an entry that would exceed it collapses to its title only.
+  const maxContent = config.injectContentMaxChars ?? 300;
+  const MAX_BLOCK = Math.max(1500, maxContent + 600);
+
+  // Compressed injection (v0.5.0 2.1): a sleep-demoted row already carries its
+  // summary in `content` with the original parked in `_full_content` — inject
+  // the summary verbatim instead of re-truncating the (already short) text.
+  // Regular long rows keep the hard truncate.
+  function injectMemory(m, maxLength = maxContent) {
+    if (m?._full_content) return String(m.content ?? "");
+    const text = String(m?.content ?? "");
+    if (text.length <= maxLength) return text;
+    return `${text.slice(0, maxLength)}…${STR.truncatedHint[language](maxLength, text.length, m.id)}`;
+  }
+
+  // Prompt-variable brace escaping (issue #162, restored from v0.7.4 #40):
+  // DSH's interpolate() scans every injected section for `{{...}}` and throws
+  // unless the variable name matches /^[a-z][a-z0-9_]*$/ — memory/profile
+  // content carrying legal template syntax (docker `--format "{{.Server.Version}}"`,
+  // Obsidian-style `{{hl|}}`, `{{挖空}}`, `{{关键词}}`) would hit an illegal
+  // variable name and throw, crashing the whole turn. At the injection boundary
+  // we escape every run of 2+ consecutive braces, inserting a `\` between each
+  // pair, so no `{{`/`}}` substring survives into the prompt: `{{a}}` → `{\{a\}\}`,
+  // and odd runs like `{{{a}}}` (which pair-wise escaping would leave with a
+  // literal `{{`) are handled too. The text keeps its readable template form,
+  // the transform is idempotent (escaped braces are single + `\`, never two
+  // adjacent), and single braces pass through untouched — interpolate only
+  // scans `{{`. Off via config.escapePromptVariables=false (default true).
+  function escapePromptVars(text) {
+    if (config.escapePromptVariables === false) return String(text);
+    return String(text).replace(/[{}]{2,}/g, (run) => run.split("").join("\\"));
+  }
+
+  // Hot memory (v0.5.0 1.3): the latest rounds of THIS session, rebuilt from
+  // the materialized event log on every render — stateless, so it survives
+  // session switches and never persists anywhere.
+  const hot = createHotMemory({
+    maxRounds: config.hotMemoryRounds ?? 5,
+    maxTokens: config.hotMemoryMaxTokens ?? 2000
+  });
+
+  function renderHotContext(ctx) {
+    if (config.hotMemoryEnabled === false) return "";
+    const rounds = extractRounds(ctx, config.hotMemoryRounds ?? 5);
+    if (!rounds.length) return "";
+    hot.clear();
+    for (const r of rounds) hot.add(r);
+    const body = hot.getContext();
+    if (!body) return "";
+    return escapePromptVars(STR.hotHeader[language](rounds.length, body));
+  }
+
+  function render(candidates) {
+    if (!candidates.length) return "";
+    const header = STR.memoryHeader[language];
+    const lines = [header];
+    let budget = MAX_BLOCK - header.length;
+    for (const m of candidates) {
+      // Epistemic trust (v0.4.5): when enabled, measured observations are
+      // flagged so the agent can weigh them above guesses/opinions.
+      const verified = config.trustEpistemicWeighting === true && m.epistemic_status === "observation"
+        ? STR.verified[language]
+        : "";
+      const title = STR.entryTitle[language](m.title, m.importance);
+      const content = injectMemory(m);
+      const full = STR.entryLine[language](m.type, verified, title, content);
+      if (budget - full.length >= 0) {
+        lines.push(full);
+        budget -= full.length;
+      } else {
+        lines.push(`- [${m.type}] ${verified}${title}`);
+      }
+    }
+    return escapePromptVars(lines.join("\n"));
+  }
+
+  // Bug4: the system-prompt render is synchronous, so the semantic query vector
+  // must be prefetched asynchronously and cached for the next assembly. The
+  // first render after a new user message may still fall back to the rule-based
+  // pick; later assemblies in the same session reuse the cached vector. Bounded
+  // cache (cap 8, drop oldest) so a long session never grows it unbounded.
+  const QUERY_VECTOR_CACHE_MAX = 8;
+  const queryVectorCache = new Map();
+  let lastPrefetched = "";
+
+  function prefetchQueryVector(query) {
+    if (!query || query === lastPrefetched || queryVectorCache.has(query)) return;
+    lastPrefetched = query;
+    service.embedQuery(query).then((vec) => {
+      if (Array.isArray(vec) && vec.length) {
+        queryVectorCache.set(query, vec);
+        if (queryVectorCache.size > QUERY_VECTOR_CACHE_MAX) {
+          queryVectorCache.delete(queryVectorCache.keys().next().value);
+        }
+      }
+    }).catch(() => { /* prefetch is best-effort */ });
+  }
+
+  // User profile + rules: injected ahead of the memory block because they are
+  // always-relevant instructions the agent should follow every turn.
+  function renderUserSettings() {
+    const profile = settings.getProfile().trim();
+    const rules = settings.getRules();
+    if (!profile && !rules.length) return "";
+    const lines = [STR.userSettingsHeader[language]];
+    if (profile) lines.push(STR.profileLine[language](profile));
+    for (const rule of rules) lines.push(STR.ruleLine[language](rule));
+    return escapePromptVars(lines.join("\n"));
+  }
+
+  const disposers = [
+    ctx.systemPrompt.context({
+      name: "memory",
+      order: 90,
+      text: (ctx) => {
+        // Bug4: pass the latest user query so injection prefers semantically
+        // relevant memories; lastUserQuery is best-effort (empty → legacy).
+        // The query vector is prefetched asynchronously (cached) because the
+        // render itself must stay synchronous.
+        const query = lastUserQuery(ctx);
+        if (query) prefetchQueryVector(query);
+        const queryVector = queryVectorCache.get(query);
+        // v0.8.0 A3：scope 随请求解析（strict 开启时 injectCandidates 内硬过滤；
+        // flag 关闭时解析器返回 null，注入行为不变）。
+        const scope = resolveSessionScope(ctx);
+        // Issue #205：跨轮轮换——最近 N 个查询轮次注入过的 id 本轮不再优先。
+        const sessionId = ctx?.agent?.session?.id ?? "_";
+        const rotate = recentInjectedIds(sessionId, query);
+        // rotateWindow 随集合一起传入：候选池按 maxItems×(N+1) 扩容——窗口要的
+        // 牌比既有 maxItems×2 池多时，轮换才有新鲜牌可换（#205 补测）。
+        const candidates = service.injectCandidates({ query, queryVector, maxItems, threshold, scope, rotate, rotateWindow: rotationTurns });
+        recordInjection(sessionId, query, candidates);
+        // Hot memory (v0.5.0 1.3) leads the single memory block: the agent
+        // sees the short-term rounds first, then the cross-session recall —
+        // the documented injection order 1→2. Folding it here (instead of a
+        // separate context) keeps the prompt assembly stable at two blocks.
+        const hotText = renderHotContext(ctx);
+        const body = render(candidates);
+        if (!hotText) return body;
+        return body ? `${hotText}\n\n${body}` : hotText;
+      }
+    }),
+    ctx.systemPrompt.context({
+      name: "user-settings",
+      order: 85,
+      text: renderUserSettings
+    })
+  ];
+
+  return () => {
+    queryVectorCache.clear();
+    rotationHistory.clear();
+    for (const dispose of disposers) {
+      if (typeof dispose === "function") dispose();
+    }
+  };
+}