@yeaft/webchat-agent 0.1.972 → 0.1.973

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "0.1.972",
+  "version": "0.1.973",
   "description": "Remote agent for Yeaft WebChat — connects worker machines to the central server",
   "main": "index.js",
   "type": "module",

package/yeaft/dream/apply.js

@@ -43,8 +43,8 @@ function rawResponseSnippet(raw) {
 
 function applySystem(language) {
   return String(language || '').toLowerCase().startsWith('zh')
-    ? '你是梦境流水线的 Apply 阶段。你会根据最近的群组对话重写单个 scope 的 memory.md 和 summary.md。请只回复严格 JSON，不要输出说明文字或 markdown fence。memory_md 和 summary_md 的自然语言内容必须使用中文；JSON key、scope、schema 字段和代码标识符保持英文。'
-    : 'You are the Apply stage of a dream pipeline. You rewrite a single scope\'s memory.md and summary.md based on recent group conversations. Reply with strict JSON only — no prose, no fences.';
+    ? '你是梦境流水线的 Apply 阶段。你会根据最近的 session 对话重写单个 scope 的 memory.md 和 summary.md。请只回复严格 JSON，不要输出说明文字或 markdown fence。memory_md 和 summary_md 的自然语言内容必须使用中文；JSON key、scope、schema 字段和代码标识符保持英文。'
+    : 'You are the Apply stage of a dream pipeline. You rewrite a single scope\'s memory.md and summary.md based on recent session conversations. Reply with strict JSON only — no prose, no fences.';
 }
 
 /**

package/yeaft/dream/prompts/create.md

@@ -1,14 +1,31 @@
+<!-- lang:en -->
 You are creating a new memory scope from scratch.
 
 Scope path: {{target}}   (must be ≤2 levels)
 
-Source conversations:
+Source session conversations:
 {{sources}}
 
 {{siblingsBlock}}
 Task:
-1. Write memory.md from scratch with reasonable section structure.
+1. Write memory.md from scratch with a reasonable section structure.
 2. Write summary.md (1–3 sentences).
 
 Reply with strict JSON of the shape:
 { "memory_md": "...", "summary_md": "..." }
+
+<!-- lang:zh -->
+你正在从零创建一个新的 memory scope。
+
+Scope path: {{target}}   (must be ≤2 levels)
+
+来源 session 对话：
+{{sources}}
+
+{{siblingsBlock}}
+任务：
+1. 从零编写 memory.md，并使用合理的章节结构。
+2. 编写 summary.md（1–3 句）。
+
+只回复严格 JSON，结构如下：
+{ "memory_md": "...", "summary_md": "..." }

package/yeaft/dream/prompts/extract-session.md

@@ -1,57 +1,59 @@
+<!-- lang:en -->
 # Dream Extract — Session Scope
 
-You are extracting **memory segments** from a conversation between the
-user and a Yeaft AI companion. This pass focuses on a specific
-**`sessions/<id>` scope**: long-lived facts about one collaboration session
-(a project team, a study cohort, a working set of people/agents).
+You are extracting **memory segments** from a session conversation. This pass focuses on a specific **`sessions/<id>` scope**: durable facts about one collaboration session and its shared context.
 
 The target session id is provided as `{{sessionId}}`.
 
 ## What to extract for `sessions/<id>` scope
 
-- **purpose** — what this session exists to do, its charter / mission
-- **members** — people, VPs, and roles in the session, and what each is
-  responsible for
-- **conventions** — how the session works (rituals, cadences, naming,
-  channels, languages used)
-- **shared decisions** — durable agreements ("we ship on Fridays",
-  "all PRs need two reviewers")
-- **shared context** — domain knowledge the whole session relies on
-- **relations** — other sessions, features, or topics this session owns or
-  depends on
-- **lessons** — collective takeaways ("we tried X in Q1, it didn't
-  scale, switched to Y")
+- **purpose** — what this session exists to do, its charter or mission
+- **members** — users, VPs, roles, and what each is responsible for
+- **conventions** — how the session works: rituals, naming, languages, review flow
+- **shared decisions** — durable agreements the session should remember
+- **shared context** — domain or project knowledge the whole session relies on
+- **current state** — recent PRs, reviews, blockers, next steps, or release status that matter now
+- **lessons** — corrections or decisions that should guide future turns in this session
 
 ## What NOT to extract here
 
-- Personal preferences of the user — those go to `user` scope.
-- Single-VP traits — those go to that VP's `vp/<id>` scope.
-- Feature-specific implementation detail — those go to
-  `feature/<id>` scope.
-- Transient status updates — only durable session facts.
-
-## Segment shape
-
-Each segment is **self-contained** and **about one thing**. Detail is
-OK; one-line summaries are NOT — preserve specifics.
-
-## Output format
-
-Reply with a JSON array of segment objects:
-
-```json
-[
-  {
-    "kind": "decision",
-    "tags": ["process", "review"],
-    "sourceMessages": ["m_201"],
-    "body": "Session {{sessionId}} decided every PR touching the payments
-    module needs sign-off from both the security VP and the payments
-    feature owner before merge. Rationale: a near-miss in March."
-  }
-]
-```
-
-`kind` ∈ {`fact`, `preference`, `decision`, `lesson`, `relation`,
-`goal`, `context`}. `scope` is filled in by the runner — do not include
-it. If nothing session-scope is in this batch, return `[]`.
+- Private user preferences that should live in `user`.
+- Facts about one VP only; those belong in the session VP scope.
+- Topic-specific details that should live under a session topic scope.
+- One-off chatter with no future value.
+
+## Output
+
+Return a JSON array only. Each item must have `kind`, `body`, `tags`, `sourceMessages`, and `confidence`.
+
+Make the segment useful to a future session member who needs the big picture plus the latest actionable detail.
+
+<!-- lang:zh -->
+# Dream Extract — Session Scope
+
+你正在从一段 session 对话中抽取 **memory segments**。本轮关注特定 **`sessions/<id>` scope**：一个协作 session 的稳定事实和共享上下文。
+
+目标 session id 会以 `{{sessionId}}` 提供。
+
+## `sessions/<id>` scope 应抽取什么
+
+- **purpose** — 这个 session 存在的目的、charter 或 mission
+- **members** — 用户、VP、角色，以及各自负责什么
+- **conventions** — session 的协作方式：流程、命名、语言、review 规则
+- **shared decisions** — session 之后应该记住的稳定决策
+- **shared context** — 整个 session 依赖的领域或项目信息
+- **current state** — 当前重要的 PR、review、阻塞、下一步或 release 状态
+- **lessons** — 未来 turn 应继续遵守的纠正和决策
+
+## 不要在这里抽取什么
+
+- 应进入 `user` 的用户私人偏好。
+- 只关于某个 VP 的事实；这些应进入 session VP scope。
+- 应进入 session topic scope 的主题细节。
+- 没有未来价值的一次性闲聊。
+
+## 输出
+
+只返回 JSON array。每一项必须包含 `kind`、`body`、`tags`、`sourceMessages` 和 `confidence`。
+
+segment 应帮助未来的 session member 快速理解大背景和最新可执行细节。

package/yeaft/dream/prompts/extract-topic.md

@@ -1,59 +1,55 @@
+<!-- lang:en -->
 # Dream Extract — Topic Scope
 
-You are extracting **memory segments** from a conversation between the
-user and a Yeaft AI companion. This pass focuses on a specific
-**`topic/<id>` scope**: long-lived facts and viewpoints about a domain
-topic that recurs across conversations (e.g. `topic/lang/rust`,
-`topic/auth/jwt`, `topic/ml/transformers`).
+You are extracting **memory segments** from a session conversation. This pass focuses on a specific **session topic scope**: durable facts, decisions, and viewpoints about one recurring topic.
 
 The target topic id is provided as `{{topicId}}`.
 
-## What to extract for `topic/<id>` scope
+## What to extract for the topic scope
 
-- **core facts** — durable knowledge about the topic (how it works,
-  key terms, gotchas) that the user has either taught the assistant
-  or confirmed
-- **viewpoints / opinions** — formed views the user holds on this
-  topic ("I think GraphQL is overkill for internal tools")
-- **canonical references** — sources the user trusts on this topic
-  (specific docs, papers, people)
-- **patterns** — how the user typically applies this topic
-  ("when using JWT, always set short access-token TTL + refresh")
-- **lessons** — things that bit them, things that worked
-- **relations** — features, projects, or other topics tied to this one
+- **core facts** — durable knowledge about the topic that the user taught or confirmed
+- **viewpoints / opinions** — stable views the user or session formed on this topic
+- **canonical references** — docs, files, PRs, issues, papers, or people treated as authoritative
+- **patterns** — how this session usually applies the topic
+- **lessons** — gotchas, mistakes, corrections, or rules learned while working on the topic
+- **current state** — latest PR/review/todo/status that still matters for this topic
 
 ## What NOT to extract here
 
-- Facts about the user as a person — `user` scope.
-- Facts specific to one feature implementation — `feature/<id>`.
-- Group conventions — `sessions/<id>`.
-- Generic encyclopedia facts the assistant already knows — only the
-  user's *durable views and confirmed knowledge* about the topic.
-
-## Segment shape
-
-Each segment is **self-contained** and **about one thing**. Detail is
-OK; one-line summaries are NOT — capture rationale and specifics so
-the segment is useful next time without rehydrating the conversation.
-
-## Output format
-
-Reply with a JSON array of segment objects:
-
-```json
-[
-  {
-    "kind": "lesson",
-    "tags": ["jwt", "auth"],
-    "sourceMessages": ["m_330", "m_331"],
-    "body": "On topic {{topicId}}: user always pairs short-lived JWT
-    access tokens (≤15 min) with rotating refresh tokens stored as
-    HttpOnly cookies. Got burned by a 24h-token leak in 2024 — single
-    long-lived token is treated as a non-starter."
-  }
-]
-```
-
-`kind` ∈ {`fact`, `preference`, `decision`, `lesson`, `relation`,
-`goal`, `context`}. `scope` is filled in by the runner — do not include
-it. If nothing topic-scope is in this batch, return `[]`.
+- User biography or broad preferences unrelated to the topic.
+- Session-wide charter information that is not specific to this topic.
+- Details about one VP unless they matter to this topic.
+
+## Output
+
+Return a JSON array only. Each item must have `kind`, `body`, `tags`, `sourceMessages`, and `confidence`.
+
+Keep the topic memory sharp: enough background to orient the next turn, plus current details that should not be lost.
+
+<!-- lang:zh -->
+# Dream Extract — Topic Scope
+
+你正在从一段 session 对话中抽取 **memory segments**。本轮关注特定 **session topic scope**：关于某个反复出现主题的稳定事实、决策和观点。
+
+目标 topic id 会以 `{{topicId}}` 提供。
+
+## topic scope 应抽取什么
+
+- **core facts** — 用户教过或确认过的、关于该主题的稳定知识
+- **viewpoints / opinions** — 用户或 session 对该主题形成的稳定观点
+- **canonical references** — 被视为权威的文档、文件、PR、issue、论文或人员
+- **patterns** — 这个 session 通常如何应用该主题
+- **lessons** — 围绕该主题学到的坑、错误、纠正或规则
+- **current state** — 该主题当前仍重要的 PR、review、todo 或状态
+
+## 不要在这里抽取什么
+
+- 与主题无关的用户履历或广泛偏好。
+- 与该主题无关的 session-wide charter 信息。
+- 除非影响该主题，否则不要抽取单个 VP 的细节。
+
+## 输出
+
+只返回 JSON array。每一项必须包含 `kind`、`body`、`tags`、`sourceMessages` 和 `confidence`。
+
+保持 topic memory 精准：既要有下一轮需要的大背景，也要保留不该丢失的当前细节。

package/yeaft/dream/prompts/extract-user.md

@@ -1,56 +1,67 @@
+<!-- lang:en -->
 # Dream Extract — User Scope
 
-You are extracting **memory segments** from a conversation between the
-user and a Yeaft AI companion. This pass focuses on the **`user`
-scope**: long-lived facts about the user themselves.
+You are extracting **memory segments** from a session conversation. This pass focuses on the **`user` scope**: long-lived facts about the user themselves.
 
 ## What to extract for `user` scope
 
 Extract segments that describe the user as a person:
 
 - **identity** — name, role, location, languages spoken, time zone
-- **preferences** — tools they use (zsh, vim, JS over TS, …), code style,
-  communication style, what they value
-- **habits / workflow** — how they work, when they work, recurring
-  patterns
-- **goals (long-term)** — what they're building, where they want to go
-- **relations** — people / projects / orgs they regularly mention
-- **lessons / opinions** — formed views ("I don't trust X", "Y is the
-  right tool for Z")
+- **preferences** — tools they use, code style, communication style, what they value
+- **habits / workflow** — how they work, recurring patterns, process expectations
+- **goals (long-term)** — what they are building, where they want to go
+- **relations** — people, projects, or orgs they regularly mention
+- **lessons / opinions** — durable views the user has formed or confirmed
 
 ## What NOT to extract here
 
-- Anything specific to a single feature, project, or VP — those go to
-  `feature/*`, `vp/*`, or `topic/*` scopes (handled by other passes).
-- Transient state ("I'm tired today") — only durable facts.
-- Verbatim message copies — segments are *secondary processing*, not
-  transcripts.
-
-## Segment shape
-
-Each segment is **self-contained** and **about one thing**. A 30-turn
-conversation typically yields 1–3 user-scope segments, not 30. Detail
-is OK; one-line summaries are NOT — preserve the specifics that make
-the memory useful next time (e.g. "uses zsh with starship prompt" beats
-"uses a shell").
-
-## Output format
-
-Reply with a JSON array of segment objects:
-
-```json
-[
-  {
-    "kind": "preference",
-    "tags": ["shell", "terminal"],
-    "sourceMessages": ["m_142", "m_143"],
-    "body": "User uses zsh with the starship prompt and prefers
-    keyboard-only workflows. Wants suggestions to assume zsh + vim
-    keybindings unless told otherwise."
-  }
-]
-```
-
-`kind` ∈ {`fact`, `preference`, `decision`, `lesson`, `relation`,
-`goal`, `context`}. `scope` is filled in by the runner — do not include
-it. If nothing user-scope is in this batch, return `[]`.
+- Anything specific to one session, feature, project, VP, or topic; those belong in narrower scopes.
+- Temporary task status, transient debugging notes, or one-off implementation details.
+- Facts inferred only from the assistant's guesswork.
+
+## Output
+
+Return a JSON array only. Each item must have:
+
+- `kind`: short category such as `preference`, `workflow`, `goal`, `identity`, `opinion`
+- `body`: one concise factual sentence, grounded in the conversation
+- `tags`: short lowercase tags
+- `sourceMessages`: message ids that support the segment
+- `confidence`: number from 0 to 1
+
+Prefer fewer, higher-value segments. Do not invent facts.
+
+<!-- lang:zh -->
+# Dream Extract — User Scope
+
+你正在从一段 session 对话中抽取 **memory segments**。本轮只关注 **`user` scope**：关于用户本人的长期事实。
+
+## `user` scope 应抽取什么
+
+抽取描述用户本人的 segment：
+
+- **identity** — 姓名、角色、所在地、使用语言、时区
+- **preferences** — 使用的工具、代码风格、沟通风格、重视什么
+- **habits / workflow** — 工作方式、重复出现的习惯、流程期待
+- **goals (long-term)** — 用户长期在构建什么、想达到什么
+- **relations** — 经常提到的人、项目或组织
+- **lessons / opinions** — 用户已经形成或确认的稳定观点
+
+## 不要在这里抽取什么
+
+- 只属于某个 session、feature、project、VP 或 topic 的内容；这些应进入更窄的 scope。
+- 临时任务状态、短期调试记录、一次性实现细节。
+- 只来自模型猜测、没有对话证据支持的事实。
+
+## 输出
+
+只返回 JSON array。每一项必须包含：
+
+- `kind`：短类别，例如 `preference`、`workflow`、`goal`、`identity`、`opinion`
+- `body`：一句简洁事实，必须基于对话证据
+- `tags`：简短 lowercase tags
+- `sourceMessages`：支持该 segment 的 message ids
+- `confidence`：0 到 1 的数字
+
+宁可少抽，也要抽高价值内容。不要编造事实。

package/yeaft/dream/prompts/extract-vp.md

@@ -1,55 +1,57 @@
+<!-- lang:en -->
 # Dream Extract — VP Scope
 
-You are extracting **memory segments** from a conversation between the
-user and a Yeaft AI companion. This pass focuses on a specific
-**`vp/<id>` scope**: long-lived facts about one Virtual Person (a
-persona / sub-agent / role the user works with).
+You are extracting **memory segments** from a session conversation. This pass focuses on a specific **session VP scope**: durable facts about one Virtual Person in this session.
 
-The target VP id is provided as `{{vpId}}`. Only extract facts about
-*this* VP, not other VPs.
+The target VP id is provided as `{{vpId}}`. Only extract facts about this VP, not other VPs.
 
-## What to extract for `vp/<id>` scope
+## What to extract for the VP scope
 
-- **identity** — who this VP is, role, persona name, charter
-- **voice / style** — how they speak, tone, formatting habits, language
-- **expertise** — what they know well, what they should defer on
-- **interaction patterns** — how the user and this VP work together,
-  what kinds of questions go to this VP, expected response shape
-- **boundaries** — what this VP will NOT do, or things to avoid
-- **relations** — other VPs, projects, or topics this VP is tied to
-- **decisions made by/about this VP** — durable choices ("VP X always
-  reviews schema migrations"), not single-turn outcomes
+- **identity / charter** — who this VP is in the session and what it is expected to own
+- **voice / style** — how this VP should think, speak, format, and decide
+- **expertise** — what the VP is good at and what it should defer
+- **interaction patterns** — what the user typically asks this VP to do and expected response shape
+- **boundaries** — things this VP must not do, or should avoid
+- **current state** — this VP's latest task, mistake, correction, review result, or handoff
+- **relations** — other VPs, topics, or project areas this VP works with
 
 ## What NOT to extract here
 
-- Facts about the user themselves — those go to `user` scope.
-- Facts about other VPs — those go to *their* `vp/<other>` scope.
-- Single-turn task state — only durable persona facts.
-- Verbatim message copies — segments are *secondary processing*.
-
-## Segment shape
-
-Each segment is **self-contained** and **about one thing**. Detail is
-OK; one-line summaries are NOT — preserve specifics that make the
-memory useful next time.
-
-## Output format
-
-Reply with a JSON array of segment objects:
-
-```json
-[
-  {
-    "kind": "preference",
-    "tags": ["voice", "style"],
-    "sourceMessages": ["m_88"],
-    "body": "VP {{vpId}} writes in concise bullet form, never uses
-    emoji, and always cites file paths with line numbers. Switches to
-    Chinese when the user does."
-  }
-]
-```
-
-`kind` ∈ {`fact`, `preference`, `decision`, `lesson`, `relation`,
-`goal`, `context`}. `scope` is filled in by the runner — do not include
-it. If nothing about this VP is in this batch, return `[]`.
+- Generic user preferences that belong in `user`.
+- Session-wide decisions that do not specifically involve this VP.
+- Facts about other VPs.
+
+## Output
+
+Return a JSON array only. Each item must have `kind`, `body`, `tags`, `sourceMessages`, and `confidence`.
+
+Prefer segments that make this VP behave better in the next turn.
+
+<!-- lang:zh -->
+# Dream Extract — VP Scope
+
+你正在从一段 session 对话中抽取 **memory segments**。本轮关注特定 **session VP scope**：这个 session 里某个 Virtual Person 的稳定事实。
+
+目标 VP id 会以 `{{vpId}}` 提供。只抽取关于这个 VP 的事实，不要抽取其他 VP。
+
+## VP scope 应抽取什么
+
+- **identity / charter** — 这个 VP 在 session 中是谁，用户期待它负责什么
+- **voice / style** — 这个 VP 应如何思考、表达、排版和判断
+- **expertise** — VP 擅长什么，哪些事情应交给别人
+- **interaction patterns** — 用户通常让这个 VP 做什么，期待什么输出形态
+- **boundaries** — 这个 VP 不能做或应避免的事情
+- **current state** — 这个 VP 最新任务、错误、纠正、review 结果或 handoff
+- **relations** — 这个 VP 关联的其他 VP、topic 或项目区域
+
+## 不要在这里抽取什么
+
+- 应进入 `user` 的通用用户偏好。
+- 与该 VP 无关的 session-wide 决策。
+- 关于其他 VP 的事实。
+
+## 输出
+
+只返回 JSON array。每一项必须包含 `kind`、`body`、`tags`、`sourceMessages` 和 `confidence`。
+
+优先抽取能让这个 VP 在下一轮表现更好的 segment。

package/yeaft/dream/prompts/index.js

@@ -77,6 +77,24 @@ function load(name) {
   return txt;
 }
 
+function extractExactLangSection(content, language) {
+  const marker = `<!-- lang:${language} -->`;
+  const start = content.indexOf(marker);
+  if (start === -1) return null;
+  const bodyStart = start + marker.length;
+  const nextMarker = content.indexOf('<!-- lang:', bodyStart);
+  return content.slice(bodyStart, nextMarker === -1 ? undefined : nextMarker).trim();
+}
+
+function selectTemplateLanguage(content, language) {
+  const effectiveLang = normalizePromptLanguage(language);
+  const selected = extractExactLangSection(content, effectiveLang);
+  if (selected !== null) return selected;
+  const fallback = extractExactLangSection(content, 'en');
+  if (fallback !== null) return fallback;
+  return content.trim();
+}
+
 /**
  * Render a template with `{{name}}` substitution. Missing keys throw —
  * silent rendering of half-built prompts has bitten us before.
@@ -86,7 +104,7 @@ function load(name) {
  * @param {{ language?: string, includeLanguageDirective?: boolean }} [opts]
  */
 export function render(name, vars, opts = {}) {
-  const tpl = load(name);
+  const tpl = selectTemplateLanguage(load(name), opts.language || 'en');
   const body = tpl.replace(/\{\{(\w+)\}\}/g, (_m, key) => {
     if (!(key in vars)) throw new Error(`prompts.${name}: missing var ${key}`);
     return vars[key];

package/yeaft/dream/prompts/summarize-scope.md

@@ -1,40 +1,49 @@
+<!-- lang:en -->
 # Dream Summarize — Per-Scope Compression
 
-You are summarizing the **memory segments** of a single scope into a
-short, dense prose summary. This is NOT extraction — the segments
-already exist. Your job is to compress them into a paragraph the
-session can keep resident in working memory.
+You are summarizing the **memory segments** of a single scope into a short, dense prose summary. This is NOT extraction — the segments already exist. Your job is to compress them into a paragraph the session can keep resident in working memory.
 
-The target scope is `{{scope}}` and contains `{{segmentCount}}`
-segments listed below.
+The target scope is `{{scope}}` and contains `{{segmentCount}}` segments listed below.
 
 ## Goals
 
-- A reader (the assistant in a future turn) should grasp **the gist of
-  this scope** from your summary alone, without reading the segments.
-- Detail is OK — but compress. Drop redundancy, keep specifics that
-  matter (names, numbers, decisions, durable views).
-- Stay faithful: do not invent facts that are not in the segments. Do
-  not "soften" decisions or opinions.
-- Bilingual: write in the same language the segments are mostly in.
+- A future session member should grasp the gist of this scope from your summary alone, without reading all segments.
+- Keep specifics that matter: names, numbers, decisions, durable views, current blockers, and next steps.
+- Stay faithful: do not invent facts not present in the segments. Do not soften decisions or opinions.
+- Write in the same language the segments are mostly in, unless the surrounding language directive says otherwise.
 
 ## Length
 
-- Target **≤ {{tokenBudget}} tokens**.
-- One paragraph for small scopes; two or three short paragraphs grouped
-  by theme for larger scopes. No bullet lists, no headings.
+- Target ≤ `{{tokenBudget}}` tokens.
+- One paragraph for small scopes; two or three compact paragraphs for larger scopes.
+- Prefer dense prose over bullet lists unless bullets are clearly more compact.
 
-## What NOT to do
+Segments:
+{{segments}}
 
-- Don't list every segment one by one — that defeats compression.
-- Don't quote verbatim chunks of segment bodies.
-- Don't add meta-commentary ("the segments show that..."). Speak
-  directly about the subject.
+Return only the summary text.
 
-## Segments
+<!-- lang:zh -->
+# Dream Summarize — Per-Scope Compression
 
-{{segments}}
+你正在把单个 scope 的 **memory segments** 压缩成一段简短、密集的 prose summary。这不是抽取 —— segments 已经存在。你的任务是把它们压缩成 session 可以常驻在工作记忆里的摘要。
+
+目标 scope 是 `{{scope}}`，包含下面列出的 `{{segmentCount}}` 个 segments。
+
+## 目标
 
-## Output
+- 未来的 session member 只读 summary，也应能理解这个 scope 的核心意思。
+- 保留重要具体信息：名字、数字、决策、稳定观点、当前阻塞和下一步。
+- 忠于事实：不要编造 segments 中没有的事实。不要弱化已经明确的决策或观点。
+- 使用 segments 主要使用的语言；如果外层语言指令另有要求，以外层语言指令为准。
+
+## 长度
+
+- 目标 ≤ `{{tokenBudget}}` tokens。
+- 小 scope 用一段；较大 scope 用两到三个紧凑段落。
+- 除非 bullet 明显更紧凑，否则优先使用密集自然段。
+
+Segments:
+{{segments}}
 
-Reply with the prose summary only. No preamble, no JSON, no fences.
+只返回 summary 文本。