From a2d76cc03420573b2fd1ebd8acd487078c6bb85f Mon Sep 17 00:00:00 2001
From: Sen Huang <48879559+ployts@users.noreply.github.com>
Date: Thu, 4 Jun 2026 19:56:50 +0800
Subject: [PATCH] refactor(auto_dream): improve recall workflow and
 documentation (#272)

* feat(file_store): add concurrency protection to LocalFileStore.dump()

* feat(file_catalog): replace file_store with file_catalog in DreamStep

* feat: add ChannelSink for Claude Code channel notifications

* feat(auto-memory): add transcript_path support and enhance metadata
---
 docs4/old/auto_cognition_design.md            |   2 +-
 docs4/old/auto_consolidate_design.md          |  29 +-
 docs4/old/auto_dream_design.md                |  89 ++-
 docs4/old/structure.md                        |  32 +-
 reme4/components/application_context.py       |   8 +-
 reme4/components/job/background_job.py        |   6 +-
 reme4/components/service/mcp_service.py       | 106 ++-
 reme4/config/default.yaml                     |  34 +-
 reme4/steps/__init__.py                       |  18 +-
 reme4/steps/base_step.py                      |  14 +-
 reme4/steps/evolve/auto_dream.py              | 684 ++++--------------
 reme4/steps/evolve/auto_memory.py             |  55 +-
 reme4/steps/evolve/dream.py                   | 499 +++++++++++++
 .../evolve/{auto_dream.yaml => dream.yaml}    | 227 ++++--
 reme4/steps/file_io/daily_list.py             |   1 +
 reme4/steps/file_io/daily_reindex.py          |   5 +
 reme4/steps/file_io/delete.py                 |  11 +
 reme4/steps/file_io/frontmatter_delete.py     |   4 +
 reme4/steps/file_io/frontmatter_read.py       |   4 +
 reme4/steps/file_io/frontmatter_update.py     |   2 +
 reme4/steps/file_io/move.py                   |   7 +
 reme4/steps/file_io/stat.py                   |   5 +
 reme4/steps/index/channel_notify.py           |  67 ++
 reme4/steps/index/claim_channel.py            |  56 ++
 reme4/steps/index/node_search.py              | 158 ++++
 reme4/steps/index/scan_changes.py             | 141 +++-
 reme4/steps/index/traverse.py                 |  20 +-
 reme4/steps/index/watch_changes.py            |  35 +-
 reme4/steps/transfer/download.py              |   5 +
 reme4/steps/transfer/ingest.py                |   6 +
 reme4/steps/transfer/upload.py                |   6 +
 reme4/utils/service_utils.py                  |   1 +
 reme4/utils/transcript.py                     | 195 +++++
 tests4/integration/test_dreamer_cli.sh        |  68 --
 tests4/integration/test_dreamer_inproc.py     |  67 +-
 tests4/unit/test_auto_dream.py                | 360 +++++++++
 tests4/unit/test_background_steps.py          |  18 +-
 tests4/unit/test_channel_notify.py            | 104 +++
 tests4/unit/test_channel_sink.py              | 100 +++
 39 files changed, 2459 insertions(+), 790 deletions(-)
 create mode 100644 reme4/steps/evolve/dream.py
 rename reme4/steps/evolve/{auto_dream.yaml => dream.yaml} (77%)
 create mode 100644 reme4/steps/index/channel_notify.py
 create mode 100644 reme4/steps/index/claim_channel.py
 create mode 100644 reme4/steps/index/node_search.py
 create mode 100644 reme4/utils/transcript.py
 delete mode 100755 tests4/integration/test_dreamer_cli.sh
 create mode 100644 tests4/unit/test_auto_dream.py
 create mode 100644 tests4/unit/test_channel_notify.py
 create mode 100644 tests4/unit/test_channel_sink.py

diff --git a/docs4/old/auto_cognition_design.md b/docs4/old/auto_cognition_design.md
index 5daa3eb6..185631a7 100644
--- a/docs4/old/auto_cognition_design.md
+++ b/docs4/old/auto_cognition_design.md
@@ -318,7 +318,7 @@ auto-memory 写入的 daily event 节点也是图的一部分(承载 daily → d
 ## 11. 演进 / 待补
 
 **当前实现状态**:
-- ✅ Stage 1 dream 已实现并跑通(`reme4/steps/evolve/auto_dream.py` + `auto_dream.yaml`)
+- ✅ Stage 1 dream 已实现并跑通(`reme4/steps/evolve/dream.py` + `dream.yaml`)
 - ⏳ Stage 2 consolidate split 部分将实现;dups / community / decay / archived 待实现
 - ⏳ Stage 3 recall 增强未实现(当前 search.py 已有 vector + keyword + RRF + 一跳 expand)
 
diff --git a/docs4/old/auto_consolidate_design.md b/docs4/old/auto_consolidate_design.md
index e9d0d170..8061c75d 100644
--- a/docs4/old/auto_consolidate_design.md
+++ b/docs4/old/auto_consolidate_design.md
@@ -29,18 +29,20 @@ dream 写入是单点视角,有三类视野局限:**写入瞬间没有跨节点
 | 4 | **腐败** | 长期不激活的痕迹 | 旧节点过时 / 半年没人读 / 内容已被矛盾 | archive |
 | 5 | **抽象缺位** | 跨多 instance 缺 schema | vault 只有原子节点,没有"主题层"视角承接全局问 | abstract |
 
-### 0.1 五大动作 + 优先级
+### 0.1 四大动作 + 优先级
 
 | 优先级 | 动作 | 解决问题 | 触发节奏 | 改 vault | 风险 | 收益 |
 |---|---|---|---|---|---|---|
-| **P0** | **community detection** | (基础设施) | weekly batch | 否 | 0(只产 meta) | 基础(其它三个都靠它)|
+| **P0** | **community detection** | (基础设施) | weekly batch | 否 | 0(只产 meta) | 基础(下游动作的依据)|
 | **P1** | **abstract** | 抽象缺位 | weekly batch(基于 P0) | 是(新建 summary) | 低(additive) | **最高**(GraphRAG 核心) |
 | **P2** | **merge** | 冗余 | weekly batch(基于 P0) | 是(合并 + retarget) | 高(lossy) | 中(消除可见冗余) |
-| **P3** | **reinforce** | 稀疏 | weekly batch(基于 P0) | 是(additive 加 wikilink) | 低 | 低(retrieve multi-hop 已部分弥补)|
 | **(独立)** | **split** | 过载 | inline 写后(D3) | 是(拆 parent + children) | 低 | 中 |
 | **(独立)** | **archive** | 腐败 | daily batch | 软(meta 标记) | 0 | 中 |
+| ~~P3 reinforce~~ | **已并入 dream synapse** | 稀疏 wikilink | 由 dream Phase 2 step 4 织突触承担 | (不在 consolidate 范围内) | — | — |
 
-**关键论断**:**P1 比 P2 优先** —— abstract additive 失败可逆且回报最大;merge lossy 失败要回滚 inbound,价值是消除冗余(必要但不增能力)。
+**关键论断**:
+- **P1 比 P2 优先** —— abstract additive 失败可逆且回报最大;merge lossy 失败要回滚 inbound,价值是消除冗余(必要但不增能力)。
+- **reinforce 已取消**(2026-06-02)—— 详 §4 标作废说明;wikilink 稀疏的解决方案是 dream Phase 2 在写入瞬间多召回 + 织突触(详 `auto_dream_design.md` §4.2.2),不再由 consolidate 周期补救。
 
 ### 0.2 实施路径
 
@@ -51,9 +53,10 @@ M0:  P0 community detection (基础设施)
 
 M1.1: P1 abstract (additive,最低风险开始改 vault)
 M1.2: P2 merge   (lossy,高门槛 + 多数票)
-M1.3: P3 reinforce (additive,价值最低,可缓做)
 
-M2+: 多层 abstract (L2 super-community) / delete / typed predicate reinforce
+M2+: 多层 abstract (L2 super-community) / delete
+
+reinforce: 不再排期 —— 已由 dream Phase 2 synapse 织突触承担
 ```
 
 ### 0.3 显式排除
@@ -365,7 +368,19 @@ audit 记录 + cooldown 设置 (winner 进 cooldown 2 weeks)
 
 ---
 
-## 4. reinforce(P3,关系强化:补 dream 漏的 wikilink)
+## 4. ~~reinforce~~(**已作废,2026-06-02**)
+
+> ⚠️ **本节作废,reinforce 已并入 dream Phase 2 synapse 织突触**(详 `auto_dream_design.md` §4.2.2)。理由:
+> - reinforce 的本质 = "找语义相关但 wikilink 缺失的节点对,补 wikilink"
+> - 但 dream Phase 2 在写入新节点瞬间已经在做同样的事(多召回 + 内化判 related + 织 `[[Y.md]]`)
+> - 让 consolidate 周期事后补 wikilink = dream RECALL 不充分的兜底,与其兜底不如把 dream 召回做强
+> - F-2 自然守住:dream 只动新节点 body(自己的 subject),不需要 consolidate 改 leaf body 这种 F-2 破例
+>
+> **新立场**:wikilink 的稀疏由 dream Phase 2 在写入瞬间一次性解决,vault 不维护"事后周期补 wikilink"的通道(`auto_cognition_design.md` §9.2 立场:关系建立在写入瞬间)。详 `hierarchical_summary.md` §13.2 Q4。
+>
+> 以下保留原 reinforce 设计内容作为历史快照,**不实施**。
+
+**(以下内容已作废,仅作历史快照)**
 
 **类比**:NREM 突触强化 LTP —— 反复共激活的连接被强化。
 
diff --git a/docs4/old/auto_dream_design.md b/docs4/old/auto_dream_design.md
index b9924fac..83b8b8d7 100644
--- a/docs4/old/auto_dream_design.md
+++ b/docs4/old/auto_dream_design.md
@@ -34,7 +34,7 @@ dream 设计回答四个问题:**桶**怎么布局 / **节点**长什么样 / **
 |---|---|
 | **物理几何** | `digest/<bucket>/<slug>.md`;**浅桶一层**(顶多两层),桶内 flat |
 | **bucket 角色** | **仅承担物理归档 + OS-level 浏览锚点**;不承担语义本体角色 —— 主题由图中节点表达 |
-| **bucket 集合** | **代码内 hard-coded**(`reme4/steps/evolve/auto_dream.py` 的 `BUCKETS` 常量),不通过配置外置,不由 dreamer / maintainer 动态生成 —— 三桶设定是 dream 模型本身的一部分(Phase 2 prompt 按 bucket 专化),不是可调参数 |
+| **bucket 集合** | **代码内 hard-coded**(`reme4/steps/evolve/dream.py` 的 `BUCKETS` 常量),不通过配置外置,不由 dreamer / maintainer 动态生成 —— 三桶设定是 dream 模型本身的一部分(Phase 2 prompt 按 bucket 专化),不是可调参数 |
 | **集合视图** | 桶名内嵌在 prompt 中(extract 阶段三桶判别启发 + 三份独立 integrate prompt);不再生成独立 `_buckets.md` 视图 |
 | **初始化** | opinionated **三桶**,按"答什么问 + 谁在问"划分:`procedure`(答"怎么做 X" —— 步骤 / 方法 / runbook)/ `personal`(答"X 是谁 / 喜欢什么 / 不要做什么" —— 用户 / 团队 specific 身份 + 偏好)/ `wiki`(答"X 是什么 / 发生了什么 / 决策依据是什么" —— 通用知识 / 定义 / 原则 / 观察 / 决策先例;**也是默认兜底**) |
 | **bucket 主页** | 不强制存在;split 累积出层级时 parent 节点天然成为浏览主页(中心性涌现,非架构必需) |
@@ -179,11 +179,83 @@ agent 上报 IntegrateOutcome {action, target_path}
 
 **Phase 2 的 bucket 专化**:三桶各有独立 system prompt,因为各桶的 body 形态、决策偏置不同 —— `procedure` 节点是 runbook 风(触发 / 步骤 / 前置 / 失败模式),`personal` 节点是规则风(rule + Why + How to apply),`wiki` 节点是百科风(定义 + 性质 + 关系)。共用一份通用 prompt 会让"应该写成什么样"的指导被稀释,bucket 信号靠一段 if-this-then-that 散文承载,效果劣于让每桶自带专属 prompt。
 
-#### 4.2.2 召回二段
+#### 4.2.2 召回 → 内化分类 → 决策 → 织突触(ReAct agent 一体完成)
 
-**RECALL = search + traverse**:search 给关键词 + 向量 RRF 命中;只要 search 在 `digest/` 下返回任何 hit,就对 top hit 跑 `traverse depth=2 direction=both`。理由是 search 关键词导向,会漏掉用不同术语归档的语义相邻抽象,那些常常一跳之外。search 在 `digest/` 下完全无命中 → 无 traverse 起点 → 候选集为空 → 直接 CREATE。
+Phase 2 是单个 ReAct agent 在一个 loop 内完成 4 件事 —— **不拆 stage,不引入外部机械步骤**,只通过 prompt 引导 agent 把 dedup 与 synapse 这两类判断都做透。当前默认 `search(limit=5)` 不够,prompt 已显式引导更深召回。
 
-**HIT = frontmatter_read + read**:渐进披露 —— 先 `frontmatter_read` 读 `name + description` 廉价 triage 淘汰明显无关候选,剩下的再 `read` 整 body。**不可仅凭 chunk 片段或 frontmatter 决定 UPDATE**,body 才是判定依据。
+**4 步流程**(整段由 ReAct agent 自主组织调用):
+
+| # | 步 | 关键动作 |
+|---|---|---|
+| 1 | **召回 —— 多角度宽召** | 显式 `limit=20-30` × 两轮 search(一次 hybrid,一次 `vector_weight=1.0` 纯语义)+ `traverse depth=2` 拓扑补充 |
+| 2 | **内化分类** | `frontmatter_read` triage + 必要时 `read` body;对每个候选**内化打 label**(只在思考中分类,不输出):`same_abstraction` / `related` / `unrelated` |
+| 3 | **决策** | 0 个 `same_abstraction` → CREATE;1 个 → UPDATE(选 flavor) |
+| 4 | **织突触** | CREATE 或 UPDATE 都把所有 `related` 候选织入 body 作 `[[Y.md]]`;CREATE 一次性织全;UPDATE additive 加 wikilink |
+
+**两类内化判断的本质**:
+
+| 判断 | 服务 | 输出形态 |
+|---|---|---|
+| **同抽象?**(dedup)| 决定 CREATE / UPDATE | 0/1 个 target(决策面排他) |
+| **相关?**(synapse)| 决定织哪些 wikilink | N 个 related 候选(决策面累加) |
+
+两者是同一个 ReAct agent 在看完 candidates 后的**两层独立判断**,共享同一批召回结果,**不需要分两轮 LLM 调用**。
+
+**召回**(对应 prompt step 1):dream 用专属的 `node_search`(`reme4/steps/index/node_search.py`),**不**用通用 `search`,**也不用 `traverse`** —— 详 §4.2.2.1(traverse 是 retrieve-time 子图挖掘工具,跟 dream 写入场景错位)。
+
+| 调用 | 找什么 |
+|---|---|
+| `node_search(query=<...>, limit=20-30)` | digest 内节点级 hybrid 召回(vector + BM25 RRF),返回 path + frontmatter |
+
+**召回结果服务两类判断**:dedup(`same_abstraction` label,是否同抽象 → CREATE / UPDATE)和 synapse(`related` label,是否相关 → 织 wikilink)是 LLM 在**同一批 candidates** 上的两类内化 label。原"两轮 search(hybrid + vector_only)"是设计冗余 —— 同一批候选 LLM 自己能判 same/related/unrelated,模式切换无意义。**调用次数由 agent 自决**:一次通常够;若 unit 跨多个概念维度,agent 可发起多次不同 query 的召回,prompt 不强约束。
+
+**HIT = `node_search` 返回 + read**:`node_search` 已内嵌返回每个 hit 的 frontmatter(`name + description`),agent 直接据此 triage,**不需要额外调 `frontmatter_read` 批量取 metadata**;仅对需要看 body 的少数候选用 `read`。**不可仅凭 frontmatter 决定 UPDATE**,body 才是判定依据。
+
+##### 4.2.2.1 node_search vs 通用 search 的差别 + 为什么 dream 不用 traverse
+
+**node_search vs 通用 search**:dream 的召回需求跟外部 agent 的 RAG 检索**结构性不同**,因此用专属 step 而非复用 `search`:
+
+| 维度 | 通用 `search`(外部 agent)| `node_search`(dream Phase 2) |
+|---|---|---|
+| 用户 | 用户/外部 agent 的自然语言 query | dream 内部生成的 unit.summary |
+| 结果粒度 | **chunk 级**(可能同一 node 多个 chunk)| **node 级**(同 path 聚合 max score)|
+| 返回信息 | 完整 chunk text + scores | **path + name + description**(frontmatter 内嵌,无 body)|
+| 范围 | 全 vault(daily / resource / digest) | **digest-only**(dream 永远只在 digest 找候选)|
+| expand_links | 默认 `True`(给 agent 更多上下文)| **永远 `False`**(synapse 找的就是未 link 的)|
+| 默认 limit | 5 | **20**(dream 需要宽召覆盖 synapse)|
+
+复用通用 `search` 会让 dream 拿到的候选**既粒度不对**(chunk 级,同 node 多次出现)**又信息冗余**(chunk text 不必要)**又被噪声污染**(daily / resource hits 永远不是 dream 的 UPDATE 候选)**又召回偏窄**(expand_links 把已 link 的拖回来,挤掉真正未 link 的 synapse 候选)。所以 dream 需要自己的 `node_search`。
+
+**为什么 dream toolkit 不包含 traverse(或 dream_traverse)** —— traverse 是 **retrieve-time 子图挖掘工具**,跟 dream 写入场景**结构性错位**:
+
+| 维度 | traverse 的本性(retrieve / RAG)| dream 的真实需求(写入)|
+|---|---|---|
+| 方向 | 从已知中心向外扩散 | 从外部新材料找 vault 内相关候选 |
+| 输入 | 已知种子节点 | 新材料的 unit.summary |
+| 输出语义 | "X 的子图"(给读者上下文) | "X 应该 link 到哪些 Y" |
+| 图遍历的角色 | 主操作 | 召回兜底(可有可无) |
+
+dream 写新节点要回答"vault 中谁跟我相关",这是**召回**问题(给 query 找相关),不是**遍历**问题(给中心找邻居)。**召回工具 = node_search;遍历工具 = traverse(留给 retrieve / 外部 agent 用)。dream 不需要遍历**。
+
+(早期曾实现 `dream_traverse` 准备作为 dream toolkit 一员,后撤销 —— 实测拓扑遍历 vs vector 召回重叠率 ~95%,真正独特贡献 < 2%,且引入 LLM 调用 / 上下文 / 复杂度成本。详 git log。)
+
+**node_search 参数极简**(`query / limit` 两个):**mode 不需要**(同一批候选服务双判断);**exclude_paths 不需要**(self 由 LLM 自己识别,frontmatter 内嵌让 agent 一眼看出"这就是我");**min_score 不需要**(RRF 分数范围 0~0.025,跟 cosine 0~1 量纲完全不同,召回深度由 `limit` 控制就够)。**调用次数 agent 自决**:prompt 不约束"必须一次",unit 跨多个概念维度时 agent 可多次召回。
+
+**node_search 召回算法:weighted node-level RRF**(vector + BM25 hybrid):
+
+- vector + BM25 各自独立召回 → 各自得到 chunk list(按各自 score 排序)
+- 同 path 多 chunk 合并:取该 path 在两个 list 中的 max chunk score 位置作为 node rank
+- RRF 融合:`score(path) = vector_weight × 1/(60 + rank_v) + (1-vector_weight) × 1/(60 + rank_k)`
+- `vector_weight=0.7`(默认),vector 主导,BM25 作为兜底(覆盖专有名词 / 缩写等 embedding 可能 struggle 的字面 case)
+- 输出 score 是 RRF 分(0~0.025 量级,不是 cosine);LLM 不依赖具体分数,内化判 same/related/unrelated
+
+**reinforce 并入立场**(对照 `auto_consolidate_design.md` §4 标作废):reinforce 不再是独立的 consolidate 动作 —— 它就是 step 4 的"织突触"。新节点写入瞬间一次性建立关系,vault 不维护"事后周期 batch 补 wikilink"的通道。F-2 自然守住 —— dream 只动新节点 body,不动其它节点。
+
+**关键约束**(诚实承认):
+- **写入即定型** —— 今天没织的 wikilink 以后没机会再织;vault 单调演化
+- **一次性 commit,无事后兜底** —— prompt 明示"宁可多织"(false positive 一眼能否决;false negative 永远沉默)
+- **召回深度取决于 prompt 引导 + agent 配合** —— 不引入外部机械召回 step;prompt 已明示 `limit=20-30 × 两轮`,但仍是 ReAct agent 的开放执行
+- **dedup 与 synapse 在一次 LLM 调用内完成** —— 不拆独立 stage,共享召回结果,内化分类是免费的
 
 #### 4.2.3 UPDATE 三种 flavor
 
@@ -207,6 +279,8 @@ agent 上报 IntegrateOutcome {action, target_path}
 - **0 出边节点合法**(没识别到合适邻居),后续 dream 进入时其它节点可以反向链回来 —— 不强求 LLM 一次性给全
 - **dream 漏判去重**(同概念建成新节点)→ 不主动兜底,接受重复;若 vault 累积明显重复,由 auto-consolidate 的 dups 检测周期 batch 产报告(`auto_consolidate_design.md` §3)
 - **召回不做 bucket 粗筛** —— LLM 拥有完整跨桶视野,可识别"概念跨桶同抽象"(例如同一原则在 wiki 已有节点而 Phase 1 把新材料归入 personal,此时 UPDATE wiki 节点而非新建 personal 节点)
+- **reinforce 已并入 dream synapse recall** —— 不存在独立的 reinforce 动作或周期 batch;突触构建(原 `auto_consolidate_design.md` §4 reinforce 的职责)在 dream Phase 2 synapse recall 阶段完成,新节点写入瞬间织全(详 §4.2.2)
+- **vault 不维护事后补 wikilink 通道** —— 上一条的直接推论;cognition §9.2 立场("关系建立在写入瞬间")在此自然守住
 
 **provenance 写出**:
 - 行文中自然带:"... 该模式最早出现在 [[daily/2026/05/15.md]] 的实践中"
@@ -266,10 +340,13 @@ agent 上报 IntegrateOutcome {action, target_path}
 
 本文档覆盖 dream 模型(桶 / 节点 / 边 / 演化)。组织端实现清单(M split / D 检测 / CAS 框架)见 `auto_consolidate_design.md` §10。
 
-- ✅ **dream step 实现** —— Phase 1 extract(识别抽象 + 分配 bucket)+ Phase 2 integrate(per sub-unit,**bucket-specific prompt 分发**;`reme4/steps/evolve/auto_dream.py` + `auto_dream.yaml`,与 `auto_memory` 同级同形)
-- ✅ **三桶 hard-coded** —— `procedure / personal / wiki`,`BUCKETS` 常量在 `dreamer.py` 顶部,Phase 1 通过 `MemoryUnit.bucket: Literal[...]` 由 Pydantic 强制约束
+- ✅ **dream step 实现** —— Phase 1 extract(识别抽象 + 分配 bucket)+ Phase 2 integrate(per sub-unit,**bucket-specific prompt 分发**;`reme4/steps/evolve/dream.py` + `dream.yaml`,与 `auto_memory` 同级同形)
+- ✅ **三桶 hard-coded** —— `procedure / personal / wiki`,`BUCKETS` 常量在 `dream.py` 顶部,Phase 1 通过 `MemoryUnit.bucket: Literal[...]` 由 Pydantic 强制约束
 - ✅ **provenance prompt 规范** —— `derived_from:: [[daily/...]]` / `[[resource/...]]` 强制(三桶 prompt 各自重申)
 - ❌ ~~**边守恒校验工具**~~ —— 早期 `digest_edit` 子类的 outbound diff 校验已随子类一并移除(切到 canonical `edit`);E-1 现由 prompt 自律,详 §4.4
 - ❌ ~~**bucket 集合配置外置**~~ —— 撤销:三桶是 dream 模型本身的一部分,不做配置参数(`vault.yaml` 不再承载 `digest.buckets`,`_buckets.md` 视图也不再生成)
+- 🆕 **Phase 2 召回拆 dedup / synapse**(2026-06-02 沉淀,详 §4.2.2)—— 当前 prompt 共用一次 `search(limit=5)`,既不够 dedup 精度也不够 synapse 覆盖;落地:`dream.yaml` 6 处(en + zh × 3 buckets)Recall 段改写,加 synapse 模式说明 + 写入即定型纪律
+- 🆕 **`file_store.default.embedding_model` 启用**(blocker)—— `default.yaml` 当前 `""`,synapse recall 用 vector_weight=1.0 模式必须开启;否则 `search` 退化为纯 BM25,dedup 也劣化
+- 🆕 **reinforce 并入立场写入**(详 §4.2.2)—— 与 `auto_consolidate_design.md` §4 标作废同步;`hierarchical_summary.md` §13.2 Q4 标解决
 
 实现进入 `reme4/steps/evolve/` 时,本文档与 `auto_memory_design.md` / `auto_consolidate_design.md` / `auto_cognition_design.md` 共同作为契约依据。
diff --git a/docs4/old/structure.md b/docs4/old/structure.md
index 56bb89dc..8cdb0443 100644
--- a/docs4/old/structure.md
+++ b/docs4/old/structure.md
@@ -173,7 +173,7 @@ Service 与 Runtime 是同一份 vault 上的两个进程角色:
 
 | 不变量 | 内容 | 来源 |
 |---|---|---|
-| **I-1** | agent 不直接写 digest(digest 写权只属 digester / maintainer) | §2.4 |
+| **I-1** | agent 不直接写 digest(digest 写权只属 dreamer / maintainer) | §2.4 |
 | **I-2** | daily folder 单作者(同 folder 不并发改) | §2.4 |
 | **I-3** | resource 内容不可变,只允许 metadata appendable | §2.4 |
 | **I-4** | 三层共用同一套 wikilink 索引,跨层引用全靠 wikilink | §2.4 |
@@ -215,10 +215,10 @@ Service 与 Runtime 是同一份 vault 上的两个进程角色:
 | 维度 | resource/ | daily/ | digest/ |
 |---|---|---|---|
 | **组织主轴** | 时间(`<date>/<name>`) | 时间 + 任务(`<date>/<slug>/`) | 语义(`<slug>/<subslug>/...`,任意嵌套) |
-| **写权归属** | 入流通道唯一(webhook / upload / pull) | agent(写入任务过程) | digester / maintainer(无 agent 直写) |
+| **写权归属** | 入流通道唯一(webhook / upload / pull) | agent(写入任务过程) | dreamer / maintainer(无 agent 直写) |
 | **可变性** | 不可变,只追加新文件 | folder 内可反复更新 | 单节点可演化,可被合并/拆分/移动 |
 | **不变量** | 写入即冻结,原文永不变 | folder 名 = summary note 名(可移动单元);同 slug 同日只一份 | slug 全局唯一;每 folder 有 canonical entry;wikilink 全路径 |
-| **谁在用** | agent(查原文)、digester(双源输入之一) | agent(自己的工作记录)、digester(双源输入之一) | agent(召回主目标)、maintainer(自维护对象) |
+| **谁在用** | agent(查原文)、dreamer(双源输入之一) | agent(自己的工作记录)、dreamer(双源输入之一) | agent(召回主目标)、maintainer(自维护对象) |
 
 ### 2.3 写入纪律:并行写入 + 双源合流
 
@@ -239,12 +239,12 @@ Service 与 Runtime 是同一份 vault 上的两个进程角色:
                     ┌──────────────┐  ◄──╮
                     │   digest/    │     │ maintain
                     │  可重组      │ ────╯ (in-place,
-                    │ (digester +  │       fold-only)
+                    │ (dreamer +  │       fold-only)
                     │  maintainer) │
                     └──────────────┘
 ```
 
-写权按这个**两层并行 → 单层合流**的拓扑分配:resource 写权专属 ingester(外部入流通道),daily 写权专属 agent(sync 落入,可响应 notify 或自身任务驱动),digest 写权专属 digester + maintainer。**resource 与 daily 之间互不写入**(agent 不动 resource,ingester 不动 daily);任何一层都不能反向改写它的上游。这是整个架构的脊梁。
+写权按这个**两层并行 → 单层合流**的拓扑分配:resource 写权专属 ingester(外部入流通道),daily 写权专属 agent(sync 落入,可响应 notify 或自身任务驱动),digest 写权专属 dreamer + maintainer。**resource 与 daily 之间互不写入**(agent 不动 resource,ingester 不动 daily);任何一层都不能反向改写它的上游。这是整个架构的脊梁。
 
 ### 2.4 不变量(永远成立)
 
@@ -310,12 +310,12 @@ Service 与 Runtime 是同一份 vault 上的两个进程角色:
 
 | 维度 | 内容 |
 |---|---|
-| 主体 | Reme Runtime(`digester` 模块,LLM-driven) |
+| 主体 | Reme Runtime(`dreamer` 模块,LLM-driven) |
 | 输入 | 一组待蒸馏的 daily folder + 相关 resource(双源合流;通常以 daily 任务为线索,顺着 wikilink / 同主题搜索拉入相关 resource 原文) |
 | 产出 | digest 中 0~N 个新节点 或 已有节点的更新;新节点必须用 wikilink 反指至少一个上游来源 |
 | 不变量 | resource / daily 正文 0 修改;digest 新节点必须 wikilink 反指上游(provenance);digest 节点遵守第 2.4 节列的不变量 |
 | Provenance | digest → daily / resource 双源链条可达(资料源是 resource 时直接反指,任务过程是 daily 时反指 daily 进而可达 resource) |
-| 反例 | digester 改写 resource;digester 改写 daily 正文 |
+| 反例 | dreamer 改写 resource;dreamer 改写 daily 正文 |
 
 ### 3.6 maintain:digest → digest 折叠
 
@@ -605,12 +605,12 @@ L2 只关心"vault 当前是什么样",**无业务语义** —— 不知道 dail
 | **ingester** | ingest | External push / pull | × | 原样落 resource + 抽 frontmatter + 入索引 |
 | **notifier** | notify | Reme background(cron + L2 资源自治状态阈值) | × | 从 L2 资源自治状态选候选 → 写 L2 推送队列;Service MCP 拿走 |
 | **synchronizer** | sync | Agent on-demand | ✓ | 把当下事件织入 daily 工作叙事 |
-| **digester** | digest | Reme background | ✓ | resource + daily 双源合流成 digest 长期条目 |
+| **dreamer** | digest | Reme background | ✓ | resource + daily 双源合流成 digest 长期条目 |
 | **maintainer** | maintain | Reme background | ✓ | digest topic tree 的**密度折叠**(fold-only) |
 
 `retrieve` 不构成独立 L4 模块,理由见 §7.4。
 
-**三个 reme 自治模块**:notifier(机械)、digester(LLM)、maintainer(LLM)。三者都由 scheduler 触发,都消费 L2 自治状态,但只有 notifier 是机械的 —— 候选选择不需要 LLM,LLM 决策在 agent 侧的 sync。
+**三个 reme 自治模块**:notifier(机械)、dreamer(LLM)、maintainer(LLM)。三者都由 scheduler 触发,都消费 L2 自治状态,但只有 notifier 是机械的 —— 候选选择不需要 LLM,LLM 决策在 agent 侧的 sync。
 
 ### 7.2 对称结构
 
@@ -621,7 +621,7 @@ L2 只关心"vault 当前是什么样",**无业务语义** —— 不知道 dail
   Inbound:                                        ingester
   Attention:                                      notifier
   Working:           synchronizer
-  Sink:              digester
+  Sink:              dreamer
   Organization:      maintainer (fold-only)
 ```
 
@@ -632,7 +632,7 @@ L2 只关心"vault 当前是什么样",**无业务语义** —— 不知道 dail
 | ingester | 外部异构格式 → vault 统一文件 |
 | notifier | L2 资源自治状态 → agent 注意力(`notify` 推送) |
 | synchronizer | agent 事件流 → 工作过程叙事(写 hot) |
-| digester | 工作过程 + 原始资料 → 长期知识(双源合流,写 cold) |
+| dreamer | 工作过程 + 原始资料 → 长期知识(双源合流,写 cold) |
 | maintainer | 散乱叶子 → 有层次的 topic tree(组织 cold) |
 
 ingester 和 notifier 是机械(确定性阈值/流水线);其它三个是 LLM 决策模块,各自跨越一层语义鸿沟。
@@ -728,11 +728,11 @@ Schema(资料的 frontmatter / wikilink / 章节约定)是横跨三层、各写
 
 | # | 反例 | 违反的不变量 |
 |---|---|---|
-| ✗-1 | Agent 通过任意 verb 直接写 digest | I-1(digest 写权只属 digester / maintainer) |
+| ✗-1 | Agent 通过任意 verb 直接写 digest | I-1(digest 写权只属 dreamer / maintainer) |
 | ✗-2 | 多 agent 并发改同一个 daily folder | I-2(daily 单作者) |
 | ✗-3 | 任何动作改写 resource 的原文 | I-3(resource immutable) |
 | ✗-4 | 跨层引用引入第二套机制(hash-id / external ref / SQL) | I-4(wikilink 是唯一跨层载体) |
-| ✗-5 | digester 改写 daily 正文 | `digest` 不变量(§3.5) |
+| ✗-5 | dreamer 改写 daily 正文 | `digest` 不变量(§3.5) |
 | ✗-6 | maintain 改写 daily / resource 的语义内容 | `maintain` 不变量(§3.6) |
 | ✗-7 | `notify` 维护 resource 上的 `referenced_by` 反指 | `notify` 完全单向(§3.3) |
 | ✗-8 | 把 state / semantic / topological 合并成单一 read verb | R-1 |
@@ -764,13 +764,13 @@ Schema(资料的 frontmatter / wikilink / 章节约定)是横跨三层、各写
 | **L5 Service** | 服务 agent 请求的进程(HTTP / MCP);执行栈最上层;也是 notify 的 MCP transport |
 | **L5 Runtime** | 自治维护 vault 的进程;scheduler 在其中按 L2 自治状态阈值触发 background Action |
 | **L4 Action** | 6 类动作语义:ingest / notify / sync / retrieve / digest / maintain |
-| **L4 模块** | 实现 Action 的架构角色;五个:ingester / notifier / synchronizer / digester / maintainer(retrieve 不构成独立模块) |
+| **L4 模块** | 实现 Action 的架构角色;五个:ingester / notifier / synchronizer / dreamer / maintainer(retrieve 不构成独立模块) |
 | **ingester** | L4 模块,机械:外部源原样落 resource + 抽 frontmatter + 入索引 |
 | **notifier** | L4 模块,机械:从 L2 资源自治状态选 notify 候选 → 写 L2 推送队列;Service MCP 拿走推给 agent |
 | **synchronizer** | L4 模块,LLM-driven:agent 事件织入 daily 工作叙事;响应 notify 的也走这里 |
-| **digester** | L4 模块,LLM-driven:resource + daily 双源合流为 digest 长期条目 |
+| **dreamer** | L4 模块,LLM-driven:resource + daily 双源合流为 digest 长期条目 |
 | **maintainer** | L4 模块,LLM-driven,**fold-only**:digest topic tree 的密度折叠 |
-| **scheduler** | L5 Runtime 内部触发器:按 cron + L2 自治状态阈值拉起 background L4 模块(notifier / digester / maintainer) |
+| **scheduler** | L5 Runtime 内部触发器:按 cron + L2 自治状态阈值拉起 background L4 模块(notifier / dreamer / maintainer) |
 | **Topic tree** | digest/ 的心智模型:文件夹 = 中间节点,文件 = 叶子 |
 | **Fold(密度折叠)** | maintainer 唯一操作:把过密叶子归簇到新子中间节点 + 写高密度摘要 |
 | **L3 原子工具** | 基础(create/append/edit/read/write/move/delete/list/stat,直 fs)+ 高级(search/traverse/frontmatter,走 L2)两组 |
diff --git a/reme4/components/application_context.py b/reme4/components/application_context.py
index 4b482b7b..611c188a 100644
--- a/reme4/components/application_context.py
+++ b/reme4/components/application_context.py
@@ -1,7 +1,7 @@
 """Application context: shared state container for components, jobs, and service."""
 
 from concurrent.futures import ThreadPoolExecutor
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Any
 
 from ..enumeration import ComponentEnum
 from ..schema import ApplicationConfig
@@ -29,3 +29,9 @@ class ApplicationContext:
         self.components: dict[ComponentEnum, dict[str, "BaseComponent"]] = {}
         self.jobs: dict[str, "BaseJob"] = {}
         self.thread_pool: ThreadPoolExecutor | None = None
+        # Side-channel for service/transport-specific objects that don't fit
+        # the shared component/job model — e.g. MCPService publishes a
+        # ChannelSink under "channel_sink" so MCP-specific steps
+        # (claim_channel, channel_notify) can find it. Keep keys narrow:
+        # if a value is needed across services, promote it to a typed field.
+        self.metadata: dict[str, Any] = {}
diff --git a/reme4/components/job/background_job.py b/reme4/components/job/background_job.py
index fa623d78..fc62662a 100644
--- a/reme4/components/job/background_job.py
+++ b/reme4/components/job/background_job.py
@@ -36,11 +36,13 @@ class BackgroundJob(BaseJob):
         backoff_cap: float = 60.0,
         close_timeout: float = 5.0,
         attempt_reset_after: float = 60.0,
-        enable_serve: bool = False,
         use_thread_pool: bool = False,
         **kwargs,
     ):
-        super().__init__(enable_serve=enable_serve, **kwargs)
+        # Background jobs are long-running loops, not request-shaped callables —
+        # forced off so they never get registered as service tools.
+        kwargs.pop("enable_serve", None)
+        super().__init__(enable_serve=False, **kwargs)
         self.supervisor: bool = supervisor
         self.backoff_base: float = backoff_base
         self.backoff_cap: float = backoff_cap
diff --git a/reme4/components/service/mcp_service.py b/reme4/components/service/mcp_service.py
index 9203aff9..99287ccd 100644
--- a/reme4/components/service/mcp_service.py
+++ b/reme4/components/service/mcp_service.py
@@ -1,23 +1,117 @@
-"""MCP (Model Context Protocol) service: exposes jobs as MCP tools."""
+"""MCP service: expose jobs as MCP tools.
 
+Channel binding (the `<channel source="reme" kind="vault_change" ...>`
+push from background steps to a specific Claude Code window) is uniform
+across transports: a single `ChannelSink` lives on
+`ApplicationContext.metadata["channel_sink"]`, unbound at startup, and any
+client calling the `claim_channel` MCP tool binds itself as the recipient
+via `fastmcp.server.dependencies.get_context().session`. Last-claim-wins.
+
+Under stdio (one client per server process) the client should claim once
+after init; until then channel events drop silently. Under shared
+streamable-http / sse the human picks which window receives events.
+
+``ChannelSink`` is colocated here because it is the runtime mechanism
+behind this service's channel feature — pushes ``notifications/claude/channel``
+frames to the bound MCP session. Lossy by design: not bound → no-op;
+``send_message`` raises → log warning, swallow (failed notifications must
+not surface as ingest failures). Uses ``ServerSession.send_message``
+(low-level raw frame) instead of ``send_notification`` because the latter
+validates against a closed ``ServerNotification`` RootModel union that
+does not include ``notifications/claude/channel`` — Pydantic rejects
+custom methods. Meta keys are filtered to ``[A-Za-z0-9_]+``: Claude Code
+silently drops keys with hyphens / other chars when projecting onto
+``<channel>`` attrs.
+"""
+
+import re
 from typing import TYPE_CHECKING
 
 from fastmcp import FastMCP
 from fastmcp.server.server import Transport
 from fastmcp.tools import FunctionTool
+from mcp.shared.message import SessionMessage
+from mcp.types import JSONRPCMessage, JSONRPCNotification
 
 from .base_service import BaseService
 from ..component_registry import R
 from ..job import BaseJob, StreamJob
 from ...constants import REME_DEFAULT_HOST, REME_DEFAULT_PORT
+from ...utils import get_logger
 
 if TYPE_CHECKING:
+    from mcp.server.session import ServerSession
+
     from ...application import Application
 
 
+_IDENT_RE = re.compile(r"^[A-Za-z0-9_]+$")
+_CHANNEL_METHOD = "notifications/claude/channel"
+
+
+class ChannelSink:
+    """Hold a bound MCP ``ServerSession`` and forward channel events to it."""
+
+    def __init__(self) -> None:
+        self._session: "ServerSession | None" = None
+        self._logger = get_logger()
+
+    def bind(self, session: "ServerSession") -> None:
+        """Set ``session`` as the recipient for subsequent ``emit`` calls (last-claim-wins)."""
+        self._session = session
+
+    def unbind(self) -> None:
+        """Drop the bound session; future ``emit`` calls become no-ops until rebind."""
+        self._session = None
+
+    async def emit(self, content: str, meta: dict[str, str] | None = None) -> None:
+        """Send one channel notification; no-op if unbound, log+swallow on transport failure."""
+        session = self._session
+        if session is None:
+            return
+
+        clean_meta = {k: str(v) for k, v in (meta or {}).items() if _IDENT_RE.match(k)}
+        message = SessionMessage(
+            JSONRPCMessage(
+                JSONRPCNotification(
+                    jsonrpc="2.0",
+                    method=_CHANNEL_METHOD,
+                    params={"content": content, "meta": clean_meta},
+                ),
+            ),
+        )
+
+        try:
+            await session.send_message(message)
+        except Exception as exc:
+            self._logger.warning(f"ChannelSink: send_message failed ({type(exc).__name__}: {exc})")
+
+
+_CHANNEL_INSTRUCTIONS = (
+    "Events from the reme channel arrive as\n"
+    '  <channel source="reme" kind="vault_change" count="N">\n'
+    "    added|modified|deleted: <vault-relative path>\n"
+    "    ...\n"
+    "  </channel>\n"
+    "The vault watcher fires one event per debounced batch of live changes "
+    "under daily/, digest/, and resource/ (initial-scan diffs at startup are "
+    "intentionally NOT replayed).\n"
+    "\n"
+    "Events are delivered ONLY to the MCP session that called the "
+    "`claim_channel` tool last. Call it once per Claude Code window that "
+    "should receive vault-change notifications.\n"
+    "\n"
+    "When new files appear under daily/ or resource/, treat it as a suggestion "
+    "to run `/dream <path>` on each new path -- unless the user is mid-task and "
+    "would be interrupted, in which case acknowledge in one line and continue. "
+    "For changes under digest/ (which /dream itself writes), just acknowledge; "
+    "do not re-dream them. For deletes or modifies elsewhere, just acknowledge."
+)
+
+
 @R.register("mcp")
 class MCPService(BaseService):
-    """Expose non-stream jobs as MCP tools over stdio, SSE, or other supported transports."""
+    """Expose non-stream jobs as MCP tools over stdio, SSE, or streamable-http."""
 
     def __init__(
         self,
@@ -34,14 +128,16 @@ class MCPService(BaseService):
     # ----- BaseService contract ------------------------------------------
 
     def build_service(self, app: "Application") -> None:
-        """Create the FastMCP server with an app-managed lifespan."""
+        """Construct the FastMCP server and publish an unbound ChannelSink."""
+        app.context.metadata["channel_sink"] = ChannelSink()
         self.service = FastMCP(
             name=app.config.app_name,
+            instructions=_CHANNEL_INSTRUCTIONS,
             lifespan=self._lifespan(app, self.host, self.port),
         )
 
     def add_job(self, job: BaseJob) -> None:
-        """Register a non-stream job as an MCP tool; StreamJobs are skipped (not supported)."""
+        """Register a non-stream job as an MCP tool; StreamJobs are unsupported."""
         if isinstance(job, StreamJob):
             return
 
@@ -59,7 +155,7 @@ class MCPService(BaseService):
         )
 
     def start_service(self, app: "Application") -> None:
-        """Run the MCP server; bind host/port only when the transport is network-based."""
+        """Run the MCP server; bind host/port only for network transports."""
         transport_kwargs: dict = {}
         if self.transport != "stdio":
             transport_kwargs["host"] = self.host
diff --git a/reme4/config/default.yaml b/reme4/config/default.yaml
index 837b9a53..782d5bcb 100644
--- a/reme4/config/default.yaml
+++ b/reme4/config/default.yaml
@@ -13,7 +13,7 @@ jobs:
     watch_paths: [ "daily", "digest" ]
     suffix_filters: [ "md" ]
     steps:
-      - backend: scan_changes_step
+      - backend: scan_store_changes_step
       - backend: update_index_step
         persist: true
       - backend: watch_changes_step
@@ -108,6 +108,26 @@ jobs:
         expand_links: true
         max_links_per_direction: 10
 
+  node_search:
+    backend: base
+    description: "Digest node recall — given a candidate abstraction's name+description, surface existing digest nodes similar enough to either dedup against or link to as related."
+    parameters:
+      type: object
+      properties:
+        query:
+          type: string
+          description: "search query"
+        limit:
+          type: integer
+          description: "max digest nodes to return"
+          default: 20
+      required:
+        - query
+    steps:
+      - backend: node_search_step
+        vector_weight: 0.7
+        candidate_multiplier: 5.0
+
   daily_create:
     backend: base
     description: "Provision a session note under a daily folder: daily/<date>/<session_id>.md or daily/<date>.md"
@@ -385,7 +405,7 @@ jobs:
       required:
         - path
     steps:
-      - backend: dreamer_step
+      - backend: dream_step
 
   auto-dream:
     backend: base
@@ -402,7 +422,7 @@ jobs:
           description: "caller guidance passed through to each per-file dream"
           default: ""
     steps:
-      - backend: cron_dreamer_step
+      - backend: auto_dream_step
 
   auto_memory:
     backend: base
@@ -456,19 +476,21 @@ components:
       stream: true
       context_size: 200000
       max_retries: 3
-      retry_delay: 1.0
       credential:
         api_key: ${LLM_API_KEY:-}
         base_url: ${LLM_BASE_URL:-https://dashscope.aliyuncs.com/apps/anthropic}
       parameters:
-        max_tokens: 100000
+        max_tokens: 65536
         thinking_enable: true
-        thinking_budget: 38000
 
   file_graph:
     default:
       backend: local
 
+  file_catalog:
+    default:
+      backend: local
+
   file_parser:
     linked:
       backend: linked
diff --git a/reme4/steps/__init__.py b/reme4/steps/__init__.py
index 82228673..fc492938 100644
--- a/reme4/steps/__init__.py
+++ b/reme4/steps/__init__.py
@@ -7,8 +7,9 @@ from .common.help import HelpStep
 from .common.llm_demo import LLMDemoStep
 from .common.stream_demo import StreamDemoStep1, StreamDemoStep2
 from .common.version import VersionStep
+from .evolve.auto_dream import AutoDreamStep
 from .evolve.auto_memory import AutoMemoryStep
-from .evolve.auto_dream import CronDreamer, Dreamer
+from .evolve.dream import DreamStep
 from .file_io.daily_create import DailyCreateStep
 from .file_io.daily_list import DailyListStep
 from .file_io.daily_reindex import DailyReindexStep
@@ -23,8 +24,11 @@ from .file_io.read import ReadStep
 from .file_io.read_image import ReadImageStep
 from .file_io.stat import StatStep
 from .file_io.write import WriteStep
+from .index.channel_notify import ChannelNotifyStep
+from .index.claim_channel import ClaimChannelStep
 from .index.clear_and_scan import ClearAndScanStep
-from .index.scan_changes import ScanChangesStep
+from .index.node_search import NodeSearchStep
+from .index.scan_changes import ScanCatalogChangesStep, ScanStoreChangesStep
 from .index.search import SearchStep
 from .index.traverse import TraverseStep
 from .index.update_catalog import UpdateCatalogStep
@@ -65,16 +69,20 @@ __all__ = [
     "FrontmatterReadStep",
     "FrontmatterUpdateStep",
     # index
+    "ChannelNotifyStep",
+    "ClaimChannelStep",
     "ClearAndScanStep",
-    "ScanChangesStep",
+    "NodeSearchStep",
+    "ScanCatalogChangesStep",
+    "ScanStoreChangesStep",
     "SearchStep",
     "TraverseStep",
     "UpdateCatalogStep",
     "UpdateIndexStep",
     "WatchChangesStep",
     # evolve (dream)
-    "CronDreamer",
-    "Dreamer",
+    "AutoDreamStep",
+    "DreamStep",
     # transfer
     "DownloadStep",
     "IngestStep",
diff --git a/reme4/steps/base_step.py b/reme4/steps/base_step.py
index 17587e51..ec507114 100644
--- a/reme4/steps/base_step.py
+++ b/reme4/steps/base_step.py
@@ -5,8 +5,9 @@ from abc import abstractmethod, ABC
 from pathlib import Path
 from typing import TypeVar, TYPE_CHECKING
 
+from agentscope.message import TextBlock
 from agentscope.model import ChatModelBase
-from agentscope.tool import Toolkit, FunctionTool
+from agentscope.tool import Toolkit, FunctionTool, ToolChunk
 
 from ..components.base_component import ComponentMixin
 from ..components.file_parser import BaseFileParser
@@ -132,8 +133,8 @@ class BaseStep(ComponentMixin, ABC):
 
         # Load class-level prompts first, then overlay caller-provided overrides.
         # Walk MRO in reverse so most-derived class wins; subclasses without their
-        # own YAML inherit prompts from their parent (e.g. CronDreamer inherits
-        # dreamer.yaml from Dreamer).
+        # own YAML inherit prompts from their parent (e.g. AutoDreamStep inherits
+        # dream.yaml from DreamStep).
         self.prompt = PromptHandler(language=self.language)
         for cls in reversed(self.__class__.__mro__):
             self.prompt.load_prompt_by_class(cls)
@@ -217,9 +218,12 @@ class BaseStep(ComponentMixin, ABC):
         if job is None:
             raise RuntimeError(f"Job {job_name} not found")
 
-        async def run_job(**_kwargs) -> str:
+        async def run_job(**_kwargs) -> ToolChunk:
             response = await job(**{**_kwargs, **kwargs})
-            return response.answer
+            return ToolChunk(
+                content=[TextBlock(text=str(response.answer))],
+                state="success" if response.success else "error",
+            )
 
         tool = FunctionTool(
             func=run_job,
diff --git a/reme4/steps/evolve/auto_dream.py b/reme4/steps/evolve/auto_dream.py
index 09511bcf..d2efd6e3 100644
--- a/reme4/steps/evolve/auto_dream.py
+++ b/reme4/steps/evolve/auto_dream.py
@@ -1,547 +1,154 @@
-"""Dreamer — auto-dream's create_or_update step.
+"""AutoDreamStep — daily-tick wrapper around :class:`DreamStep`.
 
-Reads one daily-event note or resource file at the given vault-relative
-``path``, identifies the ABSTRACTIONS the material teaches in Phase 1
-(each tagged with one of the three buckets), then in Phase 2 makes
-ONE cognitive write decision (CREATE or one of the three UPDATE
-flavors: CORROBORATE / REFINE / CORRECT) per abstraction using a
-**bucket-specific** integrate prompt.
+Each tick scans today's two surfaces under ``<daily_dir>/``:
 
-**Digest is the abstract memory layer** — raw details stay in the
-material; digest holds the principle, pattern, or precedent worth
-recalling once the specifics fade. Provenance wikilinks
-(``derived_from::``) let readers drill back down to the source.
+* ``<daily_dir>/<today>.md`` — the day-index file (auto-rebuilt rollup
+  of today's notes); included first so day-level abstractions land
+  before per-event details.
+* ``<daily_dir>/<today>/**/*.md`` — event notes for the day.
 
-Pipeline (external loop in Python, two distinct ReAct agent invocations,
-**light Phase 1 / heavy Phase 2**):
+The diff vs ``file_catalog`` follows the same shape as
+:class:`ScanCatalogChangesStep`: build ``existing`` (on-disk ``rel → mtime``)
+and ``indexed`` (catalog ``rel → mtime``, restricted to today's
+prefix so we never disturb entries from other days), then:
 
-    execute():
-        units, _ = _extract(material_blob)  # 1× ReAct: identify abstractions
-                                            #   agent emits ExtractedUnits
-                                            #   ({units: [{name, bucket, summary}, ...]})
-        for unit in units:                  # Python loop, K iterations
-            _integrate_unit(unit)           # 1× ReAct per abstraction, dispatched
-                                            #   to integrate_system_prompt_<bucket>;
-                                            #   recalls cross-bucket, decides write,
-                                            #   uses canonical write/edit/frontmatter_update tools.
+* ``existing`` keys not in ``indexed``      → **added**, dream
+* mtime mismatch                              → **modified**, dream
+* ``indexed`` keys not in ``existing``       → **deleted**, drop from catalog
+* mtime match                                 → **unchanged**, skip
 
-The bucket vocabulary is hard-coded (:data:`BUCKETS`) — three buckets,
-each with a dedicated Phase 2 prompt:
+After dreaming, successful (and Phase 1 vacuously-skipped) files
+upsert their current ``st_mtime`` so the next tick re-dreams only
+what actually changed. Failures leave the catalog untouched and will
+be retried on the next tick.
 
-* ``procedure`` — how-to-do-X: steps, methods, recipes, workflows.
-* ``personal``  — user/team specific: identity, preferences,
-  conventions, things they avoid.
-* ``wiki``      — general knowledge: definitions, principles,
-  observations, decisions-as-precedent. Default catch-all.
+The two writers (``WatchDreamStep`` event-driven, ``AutoDreamStep``
+catch-up scan) share the protocol: upsert ``(path, st_mtime)`` on
+success. Last-writer-wins is fine — same path + same content yields
+the same mtime, so they cannot disagree on what's "done".
 
-There is no SKIP outcome in Phase 2: Phase 1 is the gate for "not
-worth memorizing"; anything reaching Phase 2 warrants a write.
+Cron scheduling itself is out of scope; this step is the unit of
+work — invoke it from a system cron, ``reme auto-dream date=...``,
+or any other catch-up trigger when ``auto_dream_loop`` missed a file
+(e.g. process crashed before the watcher fired).
 
-Phase 2 uses the **canonical** ``write`` / ``edit`` jobs (no
-constrained variants). Bucket placement and edge conservation are
-prompt-level discipline; the tools themselves perform no path-shape
-or conservation validation.
+Inputs (RuntimeContext):
+    date (str, optional): YYYY-MM-DD to scan. Defaults to today
+        in the dreamer's timezone.
+    hint (str, optional): passed through to each per-file dream.
 
-Invocation form (CLI / MCP):
-    reme dream path=daily/2026-05-28/auth-refactor/auth-refactor.md
-    reme dream path=resource/2026-05-28/spec.pdf hint="focus on auth"
+Step kwargs (from yaml ``backend: auto_dream_step``):
+    persist (bool, default True): when True, ``file_catalog.dump()``
+        is called after the batch so progress survives a restart.
+
+The outer loop reuses ``DreamStep``'s prompt mounting via MRO — no
+separate YAML; ``dream.yaml`` is found through the parent class.
 """
 
-import datetime
-import zoneinfo
 from pathlib import Path
-from typing import Literal
 
-from agentscope.agent import Agent
-from agentscope.message import Msg, TextBlock
-from agentscope.permission import PermissionContext, PermissionMode
-from agentscope.state import AgentState
-from agentscope.tool import Toolkit
 from pydantic import BaseModel, Field
 
-from ..base_step import BaseStep
+from .dream import DreamStep, DreamResult
+from ..base_step import Ref
 from ...components import R
+from ...components.file_catalog import BaseFileCatalog
+from ...enumeration import ComponentEnum
+from ...schema import FileNode
 
 
-# Hard-coded bucket vocabulary. Phase 1 classifies each sub-unit into
-# one of these; Phase 2 dispatches to the bucket-specific prompt.
-# Order matters for prompt rendering — keep procedure/personal/wiki.
-BUCKETS: tuple[str, ...] = ("procedure", "personal", "wiki")
-
-# Bucket = Literal of BUCKETS. Pydantic Literal must be a static type;
-# update both BUCKETS and Bucket together if the vocabulary changes.
-Bucket = Literal["procedure", "personal", "wiki"]
-
-
-_EXTRACT_TOOLS: tuple[str, ...] = ("read",)
-
-_INTEGRATE_TOOLS: tuple[str, ...] = (
-    # read
-    "search",
-    "traverse",
-    "read",
-    "frontmatter_read",
-    # write
-    "write",
-    "edit",
-    "frontmatter_update",
-)
-
-
-def _pack_material(file_store, path: str) -> str:
-    """Render one daily-event note or resource file into a prompt block."""
-    try:
-        absolute = (Path(file_store.vault_path or ".") / path).resolve()
-    except Exception as e:
-        return f"### {path}\n(error resolving path: {type(e).__name__}: {e})\n"
-
-    if not absolute.is_file():
-        return f"### {path}\n(file not found)\n"
-
-    try:
-        return f"### {path}\n{absolute.read_text(encoding='utf-8')}\n"
-    except Exception as e:
-        return f"### {path}\n(error reading: {type(e).__name__}: {e})\n"
-
-
-class MemoryUnit(BaseModel):
-    """One memory sub-unit identified by Phase 1's structured output."""
-
-    name: str = Field(
-        description=(
-            "Short kebab-case identifier for the abstraction "
-            "(e.g. 'jwt-rotation-decision', 'pr-size-pref'). "
-            "Agent-internal handle — NOT the eventual digest slug; "
-            "Phase 2 picks the actual filing path."
-        ),
-    )
-    bucket: Bucket = Field(
-        description=(
-            "Which bucket this abstraction belongs in — Phase 2 dispatches "
-            "to a bucket-specific prompt based on this. Pick exactly one: "
-            "`procedure` (how-to-do-X — steps, methods, recipes, workflows), "
-            "`personal` (user/team-specific — identity, preferences, "
-            "conventions, things they avoid), `wiki` (general knowledge — "
-            "definitions, principles, observations, decisions-as-precedent; "
-            "default catch-all when nothing else fits)."
-        ),
-    )
-    summary: str = Field(
-        description=(
-            "1-2 sentences naming the abstraction AND pointing at where "
-            "in the material the supporting evidence lives "
-            "(e.g. 'short-credential compliance drives auth cadence; "
-            "illustrated by the 30→24h decision in the 'Decision' section "
-            "+ the SOC2 CC6.1 criticism in the 'Observation' section')."
-        ),
-    )
-
-
-class ExtractedUnits(BaseModel):
-    """Structured output emitted by Phase 1's extract agent."""
-
-    units: list[MemoryUnit] = Field(
-        default_factory=list,
-        description=(
-            "Memory sub-units identified in the material — orthogonal "
-            "abstractions (principles / patterns / precedents) worth "
-            "lifting into long-term memory. Each is tagged with its "
-            "bucket. Empty list = nothing worth lifting (Phase 2 is skipped)."
-        ),
-    )
-
-
-def _render_outcome_line(unit_name: str, bucket: str, o: "IntegrateOutcome") -> str:
-    """Format one IntegrateOutcome as a one-line summary entry."""
-    body = f"{o.action} {o.target_path}"
-    if o.note:
-        body += f" — {o.note}"
-    return f"[{unit_name}/{bucket}] {body}"
-
-
-class IntegrateOutcome(BaseModel):
-    """Structured outcome reported by Phase 2 for one sub-unit."""
-
-    action: Literal["CREATE", "CORROBORATE", "REFINE", "CORRECT"] = Field(
-        description=(
-            "Outcome of the write decision for this sub-unit. Phase 1 already "
-            "filtered out non-abstractions, so every sub-unit reaching you "
-            "warrants a write — pick the matching fine-grained action: "
-            "`CREATE` — brand-new digest node (recall returned no node "
-            "covering this abstraction); even thin first-encounter seeds go "
-            "here, they grow via CORROBORATE / REFINE on later passes. "
-            "`CORROBORATE` (most common when a covering node exists) — "
-            "provenance append + optional wording strengthening; the "
-            "abstraction already covers this material. `REFINE` — covering "
-            "node exists but the material reveals nuance, scope, or edge "
-            "cases the abstraction under-specified. `CORRECT` — covering "
-            "node exists but the material contradicts it; tighten the "
-            "abstraction or annotate the contradiction inline."
-        ),
-    )
-    target_path: str = Field(
-        description=("The digest path you wrote to — must match what your `write` / " "`edit` call(s) targeted."),
-    )
-    note: str = Field(
-        default="",
-        description=(
-            "Optional ONE short line, ≤ 200 chars, no newlines, summarizing "
-            "what landed (e.g. 'extended scope to also cover X'). Do NOT "
-            "dump recall summaries, search results, internal reasoning, or "
-            "transcripts here — those belong in the ReAct trace, not the "
-            "outcome note."
-        ),
-    )
-
-
-class DreamResult(BaseModel):
-    """Outcome of one dreamer invocation.
-
-    Per-tool audit lives in the toolkit layer (not exposed back to the
-    orchestrator). Structured outcome here is the input path the call
-    processed, the memory sub-units the agent declared in Phase 1, and
-    what got created / updated in Phase 2.
-    """
-
-    used_llm: bool = False
-    skipped: bool = False
-    path: str = ""
-    units: list[dict] = Field(default_factory=list)
-    nodes_created: list[str] = Field(default_factory=list)
-    nodes_updated: list[str] = Field(default_factory=list)
-    summary: str = ""
-    error: str = ""
-
-
-@R.register("dreamer_step")
-class Dreamer(BaseStep):
-    """auto-dream create_or_update step.
-
-    Inputs (from RuntimeContext):
-        path       (str, required): vault-relative path of one
-            daily-event note or resource file to dream over. Pass
-            empty string to no-op.
-        hint       (str, optional): caller guidance to the LLM
-            (e.g. "focus on the auth-related decisions").
-
-    Output (written to context.response.answer):
-        ``DreamResult`` JSON in ``metadata``; LLM summary in ``answer``.
-
-    CLI / MCP form:
-        reme dream path=daily/2026-05-28/auth-refactor/auth-refactor.md
-    """
-
-    def __init__(
-        self,
-        toolkit: Toolkit | None = None,
-        timezone: str | None = None,
-        **kwargs,
-    ):
-        super().__init__(**kwargs)
-        self.toolkit = toolkit
-        self.timezone = timezone
-
-    def _now(self) -> datetime.datetime:
-        if self.timezone:
-            try:
-                return datetime.datetime.now(zoneinfo.ZoneInfo(self.timezone))
-            except Exception as e:
-                self.logger.error(f"Invalid timezone: {self.timezone}, error={e}")
-        return datetime.datetime.now()
-
-    def _vault_dir(self) -> Path:
-        vr = getattr(self.file_store, "vault_path", None)
-        return Path(vr).resolve() if vr else Path.cwd().resolve()
-
-    def _llm_available(self) -> bool:
-        try:
-            return self.llm is not None
-        except Exception:
-            return False
-
-    def _build_extract_toolkit(self) -> Toolkit:
-        """Read-only toolkit for the extract agent. Sub-units come back via
-        :class:`ExtractedUnits` structured output, not via a tool call."""
-        toolkit = Toolkit()
-        for job_name in _EXTRACT_TOOLS:
-            self.add_as_tool(toolkit, job_name)
-        return toolkit
-
-    def _build_integrate_toolkit(self) -> Toolkit:
-        """Full read + canonical write/edit/frontmatter_update toolkit for
-        the integrate agent. All tools are registered via :meth:`add_as_tool`
-        — same as every other step in this codebase. Outcome tracking is
-        driven by the agent's :class:`IntegrateOutcome` structured emission,
-        not by per-tool callbacks."""
-        toolkit = self.toolkit or Toolkit()
-        for job_name in _INTEGRATE_TOOLS:
-            self.add_as_tool(toolkit, job_name)
-        return toolkit
-
-    async def _extract(self, material_blob: str, hint: str, vault_dir: Path) -> tuple[list[dict], str]:
-        """Phase 1: one ReAct invocation — read material + emit ExtractedUnits.
-
-        Returns ``(units, llm_summary)`` where ``units`` is the cleaned
-        sub-unit list (each entry has ``name`` / ``bucket`` / ``summary``)
-        and ``llm_summary`` is whatever free-form text the agent produced
-        alongside its structured emission.
-        """
-        toolkit = self._build_extract_toolkit()
-        agent = Agent(
-            name="reme_dreamer_extract",
-            model=self.llm,
-            system_prompt=self.prompt_format(
-                "extract_system_prompt",
-                vault_dir=str(vault_dir),
-                buckets=", ".join(BUCKETS),
-            ),
-            toolkit=toolkit,
-            state=AgentState(
-                permission_context=PermissionContext(
-                    mode=PermissionMode.BYPASS,
-                ),
-            ),
-        )
-        user_message = self.prompt_format(
-            "extract_user_message",
-            today=self._now().strftime("%Y-%m-%d"),
-            hint=hint or "(none)",
-            material_blob=material_blob,
-        )
-        msg = await agent.reply(
-            Msg(name="reme", role="user", content=[TextBlock(text=user_message)]),
-        )
-
-        structured_resp = await self.llm.generate_structured_output(
-            agent.state.context,
-            structured_model=ExtractedUnits,
-        )
-        meta = structured_resp.content if isinstance(structured_resp.content, dict) else {}
-        cleaned: list[dict] = []
-        for raw in meta.get("units") or []:
-            if not isinstance(raw, dict):
-                continue
-            name = str(raw.get("name") or "").strip()
-            summary = str(raw.get("summary") or "").strip()
-            bucket = str(raw.get("bucket") or "").strip()
-            if not name or not summary:
-                continue
-            if bucket not in BUCKETS:
-                # Defensive: structured_model should already reject this,
-                # but if it slips through we route to wiki (the catch-all).
-                self.logger.warning(
-                    f"[{self.name}] unit {name!r} emitted bucket {bucket!r} "
-                    f"not in {list(BUCKETS)}; routing to 'wiki'",
-                )
-                bucket = "wiki"
-            cleaned.append({"name": name, "summary": summary, "bucket": bucket})
-        return cleaned, (msg.get_text_content() or "").strip()
-
-    async def _integrate_unit(self, unit: dict, material_blob: str, hint: str, vault_dir: Path) -> IntegrateOutcome:
-        """One ReAct invocation per memory sub-unit, dispatched to the
-        bucket-specific system prompt. Returns the parsed
-        :class:`IntegrateOutcome` reported by the agent — that's the
-        single source of truth for what got written (action +
-        target_path)."""
-        bucket = unit.get("bucket") or "wiki"
-        toolkit = self._build_integrate_toolkit()
-        digest_dir = getattr(self.app_context.app_config, "digest_dir", "")
-        agent = Agent(
-            name=f"reme_dreamer_integrate_{unit.get('name', 'unit')}",
-            model=self.llm,
-            system_prompt=self.prompt_format(
-                f"integrate_system_prompt_{bucket}",
-                vault_dir=str(vault_dir),
-                digest_dir=digest_dir,
-                bucket=bucket,
-            ),
-            toolkit=toolkit,
-            state=AgentState(
-                permission_context=PermissionContext(
-                    mode=PermissionMode.BYPASS,
-                ),
-            ),
-        )
-        user_message = self.prompt_format(
-            "integrate_user_message",
-            hint=hint or "(none)",
-            unit_name=unit.get("name", ""),
-            unit_bucket=bucket,
-            unit_summary=unit.get("summary", ""),
-            material_blob=material_blob,
-        )
-        await agent.reply(
-            Msg(name="reme", role="user", content=[TextBlock(text=user_message)]),
-        )
-        structured_resp = await self.llm.generate_structured_output(
-            agent.state.context,
-            structured_model=IntegrateOutcome,
-        )
-        return IntegrateOutcome.model_validate(structured_resp.content)
-
-    async def dream_one(self, path: str, hint: str = "") -> DreamResult:
-        """Run the full extract + integrate pipeline on one vault-relative
-        material path. Returns a structured :class:`DreamResult`. Safe to
-        call repeatedly on the same instance — per-invocation trackers are
-        reset at the start of each call. Used both by :meth:`execute`
-        (single file from context) and by :class:`CronDreamer` (loop over
-        today's materials).
-        """
-        path = (path or "").strip()
-        hint = (hint or "").strip()
-
-        if not path:
-            return DreamResult(used_llm=False, skipped=True)
-
-        if not self._llm_available():
-            return DreamResult(
-                used_llm=False,
-                skipped=True,
-                path=path,
-                error="no llm configured; dreaming requires an LLM",
-            )
-
-        material_blob = _pack_material(self.file_store, path)
-
-        vault_dir = self._vault_dir()
-
-        # Phase 1 — extract (light). Agent emits ExtractedUnits structured output to commit the
-        # memory sub-units worth lifting. Each unit carries its own bucket.
-        self.logger.info(f"[{self.name}] extract phase: path={path!r}")
-        units, extract_summary = await self._extract(material_blob, hint, vault_dir)
-
-        if not units:
-            return DreamResult(
-                used_llm=True,
-                path=path,
-                summary=extract_summary or "no memory sub-units declared",
-                skipped=True,
-            )
-
-        unit_handles = ", ".join(f"{u['name']}/{u['bucket']}" for u in units)
-        self.logger.info(f"[{self.name}] integrate phase: {len(units)} sub-unit(s): {unit_handles}")
-
-        # Phase 2 — integrate, one fresh ReAct per sub-unit, dispatched to
-        # the bucket-specific system prompt. Python-level loop, not agent
-        # loop. Each session emits a structured IntegrateOutcome whose
-        # action + target_path are the source of truth for what landed.
-        nodes_created: list[str] = []
-        nodes_updated: list[str] = []
-        per_unit_lines: list[str] = []
-        for i, unit in enumerate(units, start=1):
-            name = unit.get("name", "?")
-            bucket = unit.get("bucket", "?")
-            try:
-                outcome = await self._integrate_unit(unit, material_blob, hint, vault_dir)
-            except Exception as e:
-                self.logger.error(
-                    f"[{self.name}] integrate {i}/{len(units)} "
-                    f"(unit={name}, bucket={bucket}) failed: {type(e).__name__}: {e}",
-                )
-                per_unit_lines.append(f"[{name}/{bucket}] FAILED: {type(e).__name__}: {e}")
-                continue
-            if outcome.action == "CREATE":
-                nodes_created.append(outcome.target_path)
-            else:
-                nodes_updated.append(outcome.target_path)
-            per_unit_lines.append(_render_outcome_line(name, bucket, outcome))
-
-        per_unit_block = "\n".join(per_unit_lines)
-        summary = (
-            f"Declared {len(units)} sub-unit(s) ({unit_handles}); "
-            f"created {len(nodes_created)}, updated {len(nodes_updated)}.\n"
-            f"{per_unit_block}"
-        )
-
-        return DreamResult(
-            used_llm=True,
-            path=path,
-            units=units,
-            nodes_created=nodes_created,
-            nodes_updated=nodes_updated,
-            summary=summary,
-            skipped=False,
-        )
-
-    async def execute(self):
-        assert self.context is not None
-        path: str = (self.context.get("path", "") or "").strip()
-        hint: str = (self.context.get("hint", "") or "").strip()
-
-        result = await self.dream_one(path, hint)
-
-        if not path:
-            self.context.response.success = True
-            self.context.response.answer = "Skipped: no path supplied"
-        elif result.error:
-            self.context.response.success = False
-            self.context.response.answer = f"Error: {result.error}"
-        elif result.skipped:
-            self.context.response.success = True
-            self.context.response.answer = result.summary or "Skipped: no memory sub-units declared"
-        else:
-            self.context.response.success = True
-            self.context.response.answer = result.summary
-        self.context.response.metadata.update(result.model_dump())
-
-
-# ============================================================
-# CronDreamer — daily-tick wrapper around Dreamer.
-#
-# Inherits the per-file pipeline from Dreamer.dream_one and adds the
-# outer loop over today's daily/ + resource/ files. Cron scheduling
-# itself is out of scope; this step is just the unit of work.
-#
-# Inputs (RuntimeContext):
-#     date  (str, optional): YYYY-MM-DD to scan. Defaults to today
-#         in the dreamer's timezone.
-#     hint  (str, optional): passed through to each per-file dream.
-# ============================================================
-
-
-class CronDreamResult(BaseModel):
-    """Aggregated outcome of one cron tick."""
+class AutoDreamResult(BaseModel):
+    """Aggregated outcome of one auto-dream tick."""
 
     date: str = ""
     files_scanned: int = 0
+    files_unchanged: int = 0
     files_dreamed: int = 0
     files_skipped: int = 0
     files_failed: int = 0
+    files_deleted: int = 0
     per_file: list[DreamResult] = Field(default_factory=list)
     summary: str = ""
 
 
-@R.register("cron_dreamer_step")
-class CronDreamer(Dreamer):
-    """Loop ``daily/<today>/`` + ``resource/<today>/`` and dream each file.
+@R.register("auto_dream_step")
+class AutoDreamStep(DreamStep):
+    """Scan ``daily/<today>.md`` + ``daily/<today>/`` and dream each file
+    whose ``st_mtime`` doesn't already match its ``file_catalog`` entry."""
 
-    Inherits :data:`auto_dream.yaml` from :class:`Dreamer` (no separate
-    yaml — there's no extra prompt for the outer loop).
-    """
+    file_catalog: BaseFileCatalog = Ref(BaseFileCatalog, ComponentEnum.FILE_CATALOG)
+
+    def __init__(self, persist: bool = True, **kwargs):
+        super().__init__(**kwargs)
+        self.persist: bool = persist
 
     async def execute(self):
         assert self.context is not None
         date_input: str = (self.context.get("date", "") or "").strip()
         hint: str = (self.context.get("hint", "") or "").strip()
 
-        # daily_dir / resource_dir come from app config — NOT tool params.
-        # Same convention as daily_create / daily_list / daily_reindex.
-        # resource_dir may be empty (default) — that just skips the resource scan.
+        # daily_dir comes from app config — NOT a tool param. Same convention
+        # as daily_create / daily_list / daily_reindex.
         cfg = self.app_context.app_config if self.app_context is not None else None
         daily_dir = (cfg.daily_dir if cfg else "") or "daily"
-        resource_dir = cfg.resource_dir if cfg else ""
 
         today = date_input or self._now().strftime("%Y-%m-%d")
         vault = self._vault_dir()
-        files = _scan_today_files(vault, today, daily_dir, resource_dir)
+        files = _scan_today_files(vault, today, daily_dir)
 
-        result = CronDreamResult(date=today, files_scanned=len(files))
+        # existing: today's on-disk paths → st_mtime. Insertion order = scan
+        # order (date.md first, then sorted event notes); preserved through
+        # the diff so the day-index file is dreamed before per-event notes.
+        existing: dict[str, float] = {}
+        for rel in files:
+            try:
+                existing[rel] = (vault / rel).stat().st_mtime
+            except OSError as e:
+                self.logger.error(f"[{self.name}] stat failed on {rel}: {e}")
+
+        # indexed: catalog entries restricted to today's prefix. Restriction
+        # is critical — get_nodes() returns all days, but we must only
+        # consider deletions within today's scan scope.
+        today_md = f"{daily_dir}/{today}.md"
+        today_dir = f"{daily_dir}/{today}/"
+        all_nodes = await self.file_catalog.get_nodes()
+        indexed: dict[str, float] = {
+            n.path: n.st_mtime for n in all_nodes if n.path == today_md or n.path.startswith(today_dir)
+        }
+
+        # Diff — same vocabulary as scan_*_changes_step (added/modified/deleted).
+        to_dream: list[tuple[str, float]] = [(rel, mt) for rel, mt in existing.items() if indexed.get(rel) != mt]
+        unchanged: list[str] = [rel for rel, mt in existing.items() if indexed.get(rel) == mt]
+        to_delete: list[str] = sorted(indexed.keys() - existing.keys())
+
+        result = AutoDreamResult(
+            date=today,
+            files_scanned=len(existing),
+            files_unchanged=len(unchanged),
+            files_deleted=len(to_delete),
+        )
         self.logger.info(
-            f"[{self.name}] cron tick date={today} scanned={len(files)} file(s) under "
-            f"{daily_dir}/{today}/ + {resource_dir}/{today}/",
+            f"[{self.name}] auto-dream tick date={today} scanned={len(existing)} "
+            f"unchanged={len(unchanged)} todo={len(to_dream)} deleted={len(to_delete)} "
+            f"under {daily_dir}/{today}{{.md,/}}",
         )
 
-        for rel_path in files:
+        # Drop catalog entries for files no longer on disk first. Cheap, no
+        # LLM, and keeps the catalog consistent even if the dream pass below
+        # errors.
+        if to_delete:
+            try:
+                await self.file_catalog.delete(to_delete)
+            except Exception as e:  # noqa: BLE001
+                self.logger.exception(
+                    f"[{self.name}] file_catalog.delete failed: {type(e).__name__}: {e}",
+                )
+
+        # Dream + upsert per-file. Single-file granularity means an LLM
+        # failure on file N doesn't block files N+1..K from advancing their
+        # catalog mtime.
+        upsert_nodes: list[FileNode] = []
+        for rel_path, mtime in to_dream:
             try:
                 dr = await self.dream_one(rel_path, hint)
             except Exception as e:  # pylint: disable=broad-except
@@ -554,63 +161,74 @@ class CronDreamer(Dreamer):
                 )
             result.per_file.append(dr)
             if dr.error:
+                # Failures leave the catalog untouched — next tick retries.
                 result.files_failed += 1
-            elif dr.skipped:
+                continue
+            if dr.skipped:
+                # Phase 1 said "nothing to extract" — still mark as seen so
+                # Phase 1 doesn't re-run on every tick.
                 result.files_skipped += 1
             else:
                 result.files_dreamed += 1
+            upsert_nodes.append(FileNode(path=rel_path, st_mtime=mtime))
 
-        result.summary = _render_cron_summary(result)
+        if upsert_nodes:
+            try:
+                await self.file_catalog.upsert(upsert_nodes)
+            except Exception as e:  # noqa: BLE001
+                self.logger.exception(
+                    f"[{self.name}] file_catalog.upsert failed: {type(e).__name__}: {e}",
+                )
+
+        if self.persist and (upsert_nodes or to_delete):
+            try:
+                await self.file_catalog.dump()
+            except Exception as e:  # noqa: BLE001
+                self.logger.exception(
+                    f"[{self.name}] file_catalog.dump failed: {type(e).__name__}: {e}",
+                )
+
+        result.summary = _render_auto_dream_summary(result)
         self.context.response.success = result.files_failed == 0
         self.context.response.answer = result.summary
         self.context.response.metadata.update(result.model_dump())
+        return self.context.response
 
 
-def _scan_today_files(
-    vault: Path,
-    today: str,
-    daily_dir: str,
-    resource_dir: str,
-) -> list[str]:
-    """Return vault-relative paths of today's daily notes + resource files.
+def _scan_today_files(vault: Path, today: str, daily_dir: str) -> list[str]:
+    """Return vault-relative paths of today's day-index + event notes.
 
     * ``<daily_dir>/<today>.md`` — the day-index file (auto-rebuilt
       rollup of all of today's notes). Included first so its day-level
       abstractions land before the per-event details.
     * ``<daily_dir>/<today>/**/*.md`` — event notes for the day,
-      sorted by path.
-    * ``<resource_dir>/<today>/**/*`` — any file type ingested under
-      today's resource folder. Skipped when ``resource_dir`` is empty.
-
-    Results are sorted for deterministic processing order within each
-    group; the day-index file leads.
+      sorted by path for deterministic processing order.
     """
     out: list[str] = []
+    if not daily_dir:
+        return out
 
-    if daily_dir:
-        day_index = vault / daily_dir / f"{today}.md"
-        if day_index.is_file():
-            out.append(str(day_index.relative_to(vault)))
-        daily_root = vault / daily_dir / today
-        if daily_root.is_dir():
-            for md in sorted(daily_root.rglob("*.md")):
-                if md.is_file():
-                    out.append(str(md.relative_to(vault)))
+    day_index = vault / daily_dir / f"{today}.md"
+    if day_index.is_file():
+        out.append(str(day_index.relative_to(vault)))
 
-    if resource_dir:
-        resource_root = vault / resource_dir / today
-        if resource_root.is_dir():
-            for f in sorted(p for p in resource_root.rglob("*") if p.is_file()):
-                out.append(str(f.relative_to(vault)))
+    daily_root = vault / daily_dir / today
+    if daily_root.is_dir():
+        for md in sorted(daily_root.rglob("*.md")):
+            if md.is_file():
+                out.append(str(md.relative_to(vault)))
 
     return out
 
 
-def _render_cron_summary(r: CronDreamResult) -> str:
-    """One-line header + one line per file with its outcome."""
+def _render_auto_dream_summary(r: AutoDreamResult) -> str:
+    """One-line header + one line per dreamed file with its outcome.
+    Unchanged + deleted files are counted in the header only."""
     lines = [
-        f"[CronDreamer] date={r.date} scanned={r.files_scanned} "
-        f"dreamed={r.files_dreamed} skipped={r.files_skipped} failed={r.files_failed}",
+        f"[AutoDreamStep] date={r.date} scanned={r.files_scanned} "
+        f"unchanged={r.files_unchanged} dreamed={r.files_dreamed} "
+        f"skipped={r.files_skipped} failed={r.files_failed} "
+        f"deleted={r.files_deleted}",
     ]
     for dr in r.per_file:
         if dr.error:
diff --git a/reme4/steps/evolve/auto_memory.py b/reme4/steps/evolve/auto_memory.py
index 4926cec6..7806f885 100644
--- a/reme4/steps/evolve/auto_memory.py
+++ b/reme4/steps/evolve/auto_memory.py
@@ -6,7 +6,17 @@ decides what to preserve, and writes the note via ``read`` / ``edit``
 / ``frontmatter_update`` / ``write`` tools.
 
 Inputs (from RuntimeContext):
-    messages (list[Msg], required): conversation slice to inspect.
+    messages (list[Msg], optional): conversation slice to inspect.
+        Mutually exclusive with ``transcript_path``; if both are
+        provided, ``messages`` wins.
+    transcript_path (str, optional): absolute path to a Claude Code
+        transcript JSONL file. When provided (and ``messages`` is
+        empty), the step parses the file via
+        :func:`reme4.utils.transcript.load_messages_from_transcript`
+        and proceeds as if those were the messages. This is what the
+        ``reme-service`` plugin's PreCompact / SessionEnd hooks pass
+        in directly via ``type: mcp_tool``, replacing the temporary
+        spawn-subagent bridge.
     session_id (str, optional): passed to daily_create to determine
         the note path.
     memory_hint (str, optional): caller-supplied hint for the agent.
@@ -14,7 +24,7 @@ Inputs (from RuntimeContext):
 
 Output (written to context.response):
     answer: one-line summary from the agent.
-    metadata: {path, created}.
+    metadata: {path, created, n_messages, transcript_path?}.
 """
 
 from agentscope.agent import Agent
@@ -26,6 +36,7 @@ from agentscope.tool import Toolkit
 from ._evolve import format_history, now
 from ..base_step import BaseStep
 from ...components import R
+from ...utils.transcript import load_messages_from_transcript
 
 
 @R.register("auto_memory_step")
@@ -46,24 +57,50 @@ class AutoMemoryStep(BaseStep):
 
     async def execute(self):
         assert self.context is not None
-        messages: list[Msg] = [self._to_msg(item) for item in self.context.get("messages", [])]
+        raw_messages = self.context.get("messages") or []
+        transcript_path: str = self.context.get("transcript_path", "") or ""
         session_id: str = self.context.get("session_id", "")
         memory_hint: str = self.context.get("memory_hint", "")
         current = now(self.context.get("timezone"))
 
+        # If caller passed transcript_path (the canonical Claude Code hook
+        # input) instead of messages, parse it here so the rest of the step
+        # stays unchanged.
+        if not raw_messages and transcript_path:
+            raw_messages = load_messages_from_transcript(transcript_path)
+            self.logger.info(
+                f"[{self.name}] loaded {len(raw_messages)} messages from transcript_path={transcript_path}",
+            )
+
+        messages: list[Msg] = [self._to_msg(item) for item in raw_messages]
+
         if not messages:
             self.context.response.success = True
-            self.context.response.answer = "Skipped: no messages supplied"
+            reason = (
+                f"Skipped: no messages in transcript_path={transcript_path}"
+                if transcript_path
+                else "Skipped: no messages supplied"
+            )
+            self.context.response.answer = reason
+            self.context.response.metadata.update(
+                {"n_messages": 0, "transcript_path": transcript_path},
+            )
+            self.logger.info(f"[{self.name}] skipped: {reason} session_id={session_id!r}")
             return
 
         create_response = await self.run_job("daily_create", session_id=session_id)
         if not create_response.success:
             self.context.response.success = False
             self.context.response.answer = f"daily_create failed: {create_response.answer}"
+            self.logger.info(f"[{self.name}] daily_create failed session_id={session_id!r}")
             return
 
         note_path: str = create_response.metadata["path"]
         created: bool = create_response.metadata["created"]
+        self.logger.info(
+            f"[{self.name}] note_path={note_path} created={created} "
+            f"messages={len(messages)} hint={'yes' if memory_hint else 'no'}",
+        )
 
         toolkit = Toolkit()
         for job_name in self.agent_tools:
@@ -95,4 +132,12 @@ class AutoMemoryStep(BaseStep):
 
         self.context.response.success = True
         self.context.response.answer = (final_msg.get_text_content() or "").strip()
-        self.context.response.metadata.update({"path": note_path, "created": created})
+        self.context.response.metadata.update(
+            {
+                "path": note_path,
+                "created": created,
+                "n_messages": len(messages),
+                "transcript_path": transcript_path,
+            },
+        )
+        self.logger.info(f"[{self.name}] done note_path={note_path}")
diff --git a/reme4/steps/evolve/dream.py b/reme4/steps/evolve/dream.py
new file mode 100644
index 00000000..f4218c41
--- /dev/null
+++ b/reme4/steps/evolve/dream.py
@@ -0,0 +1,499 @@
+"""DreamStep — single-file digest step (auto-dream's create_or_update primitive).
+
+Reads one daily-event note or resource file at the given vault-relative
+``path``, identifies the ABSTRACTIONS the material teaches in Phase 1
+(each tagged with one of the three buckets), then in Phase 2 makes
+ONE cognitive write decision (CREATE or one of the three UPDATE
+flavors: CORROBORATE / REFINE / CORRECT) per abstraction using a
+**bucket-specific** integrate prompt.
+
+**Digest is the abstract memory layer** — raw details stay in the
+material; digest holds the principle, pattern, or precedent worth
+recalling once the specifics fade. Provenance wikilinks
+(``derived_from::``) let readers drill back down to the source.
+
+Pipeline (external loop in Python, two distinct ReAct agent invocations,
+**light Phase 1 / heavy Phase 2**):
+
+    execute():
+        units, _ = _extract(material_blob)  # 1× ReAct: identify abstractions
+                                            #   agent emits ExtractedUnits
+                                            #   ({units: [{name, bucket, summary}, ...]})
+        for unit in units:                  # Python loop, K iterations
+            _integrate_unit(unit)           # 1× ReAct per abstraction, dispatched
+                                            #   to integrate_system_prompt_<bucket>;
+                                            #   recalls cross-bucket, decides write,
+                                            #   uses canonical write/edit/frontmatter_update tools.
+
+The bucket vocabulary is hard-coded (:data:`BUCKETS`) — three buckets,
+each with a dedicated Phase 2 prompt:
+
+* ``procedure`` — how-to-do-X: steps, methods, recipes, workflows.
+* ``personal``  — user/team specific: identity, preferences,
+  conventions, things they avoid.
+* ``wiki``      — general knowledge: definitions, principles,
+  observations, decisions-as-precedent. Default catch-all.
+
+There is no SKIP outcome in Phase 2: Phase 1 is the gate for "not
+worth memorizing"; anything reaching Phase 2 warrants a write.
+
+Phase 2 uses the **canonical** ``write`` / ``edit`` jobs (no
+constrained variants). Bucket placement and edge conservation are
+prompt-level discipline; the tools themselves perform no path-shape
+or conservation validation.
+
+Invocation form (CLI / MCP):
+    reme dream path=daily/2026-05-28/auth-refactor/auth-refactor.md
+    reme dream path=resource/2026-05-28/spec.pdf hint="focus on auth"
+"""
+
+import datetime
+import zoneinfo
+from pathlib import Path
+from typing import Literal
+
+from agentscope.agent import Agent
+from agentscope.message import Msg, TextBlock
+from agentscope.permission import PermissionContext, PermissionMode
+from agentscope.state import AgentState
+from agentscope.tool import Toolkit
+from pydantic import BaseModel, Field
+
+from ..base_step import BaseStep
+from ...components import R
+
+
+# Hard-coded bucket vocabulary. Phase 1 classifies each sub-unit into
+# one of these; Phase 2 dispatches to the bucket-specific prompt.
+# Order matters for prompt rendering — keep procedure/personal/wiki.
+BUCKETS: tuple[str, ...] = ("procedure", "personal", "wiki")
+
+# Bucket = Literal of BUCKETS. Pydantic Literal must be a static type;
+# update both BUCKETS and Bucket together if the vocabulary changes.
+Bucket = Literal["procedure", "personal", "wiki"]
+
+
+_EXTRACT_TOOLS: tuple[str, ...] = ("read",)
+
+_INTEGRATE_TOOLS: tuple[str, ...] = (
+    # read — dream uses its own node-level digest search (NOT the
+    # general chunk-level `search`), specialized for dedup + synapse
+    # recall. See reme4/steps/index/node_search.py for the rationale.
+    # NO traverse here: traverse is a retrieve-time subgraph mining
+    # tool (used by external retrieval agents); dream is a write-time
+    # candidate recall operation, structurally a different problem.
+    "node_search",
+    "read",
+    "frontmatter_read",
+    # write
+    "write",
+    "edit",
+    "frontmatter_update",
+)
+
+
+# ============================================================
+# Schema (Pydantic models for structured ReAct output) + helpers
+# ============================================================
+
+
+def _pack_material(file_store, path: str) -> str:
+    """Render one daily-event note or resource file into a prompt block."""
+    try:
+        absolute = (Path(file_store.vault_path or ".") / path).resolve()
+    except Exception as e:
+        return f"### {path}\n(error resolving path: {type(e).__name__}: {e})\n"
+
+    if not absolute.is_file():
+        return f"### {path}\n(file not found)\n"
+
+    try:
+        return f"### {path}\n{absolute.read_text(encoding='utf-8')}\n"
+    except Exception as e:
+        return f"### {path}\n(error reading: {type(e).__name__}: {e})\n"
+
+
+class MemoryUnit(BaseModel):
+    """One memory sub-unit identified by Phase 1's structured output."""
+
+    name: str = Field(
+        description=(
+            "Short kebab-case identifier for the abstraction "
+            "(e.g. 'jwt-rotation-decision', 'pr-size-pref'). "
+            "Agent-internal handle — NOT the eventual digest slug; "
+            "Phase 2 picks the actual filing path."
+        ),
+    )
+    bucket: Bucket = Field(
+        description=(
+            "Which bucket this abstraction belongs in — Phase 2 dispatches "
+            "to a bucket-specific prompt based on this. Pick exactly one: "
+            "`procedure` (how-to-do-X — steps, methods, recipes, workflows), "
+            "`personal` (user/team-specific — identity, preferences, "
+            "conventions, things they avoid), `wiki` (general knowledge — "
+            "definitions, principles, observations, decisions-as-precedent; "
+            "default catch-all when nothing else fits)."
+        ),
+    )
+    summary: str = Field(
+        description=(
+            "1-2 sentences naming the abstraction AND pointing at where "
+            "in the material the supporting evidence lives "
+            "(e.g. 'short-credential compliance drives auth cadence; "
+            "illustrated by the 30→24h decision in the 'Decision' section "
+            "+ the SOC2 CC6.1 criticism in the 'Observation' section')."
+        ),
+    )
+
+
+class ExtractedUnits(BaseModel):
+    """Structured output emitted by Phase 1's extract agent."""
+
+    units: list[MemoryUnit] = Field(
+        default_factory=list,
+        description=(
+            "Memory sub-units identified in the material — orthogonal "
+            "abstractions (principles / patterns / precedents) worth "
+            "lifting into long-term memory. Each is tagged with its "
+            "bucket. Empty list = nothing worth lifting (Phase 2 is skipped)."
+        ),
+    )
+
+
+def _render_outcome_line(unit_name: str, bucket: str, o: "IntegrateOutcome") -> str:
+    """Format one IntegrateOutcome as a one-line summary entry."""
+    body = f"{o.action} {o.target_path}"
+    if o.note:
+        body += f" — {o.note}"
+    return f"[{unit_name}/{bucket}] {body}"
+
+
+class IntegrateOutcome(BaseModel):
+    """Structured outcome reported by Phase 2 for one sub-unit."""
+
+    action: Literal["CREATE", "CORROBORATE", "REFINE", "CORRECT"] = Field(
+        description=(
+            "Outcome of the write decision for this sub-unit. Phase 1 already "
+            "filtered out non-abstractions, so every sub-unit reaching you "
+            "warrants a write — pick the matching fine-grained action: "
+            "`CREATE` — brand-new digest node (recall returned no node "
+            "covering this abstraction); even thin first-encounter seeds go "
+            "here, they grow via CORROBORATE / REFINE on later passes. "
+            "`CORROBORATE` (most common when a covering node exists) — "
+            "provenance append + optional wording strengthening; the "
+            "abstraction already covers this material. `REFINE` — covering "
+            "node exists but the material reveals nuance, scope, or edge "
+            "cases the abstraction under-specified. `CORRECT` — covering "
+            "node exists but the material contradicts it; tighten the "
+            "abstraction or annotate the contradiction inline."
+        ),
+    )
+    target_path: str = Field(
+        description=("The digest path you wrote to — must match what your `write` / " "`edit` call(s) targeted."),
+    )
+    note: str = Field(
+        default="",
+        description=(
+            "Optional ONE short line, ≤ 200 chars, no newlines, summarizing "
+            "what landed (e.g. 'extended scope to also cover X'). Do NOT "
+            "dump recall summaries, search results, internal reasoning, or "
+            "transcripts here — those belong in the ReAct trace, not the "
+            "outcome note."
+        ),
+    )
+
+
+class DreamResult(BaseModel):
+    """Outcome of one DreamStep invocation.
+
+    Per-tool audit lives in the toolkit layer (not exposed back to the
+    orchestrator). Structured outcome here is the input path the call
+    processed, the memory sub-units the agent declared in Phase 1, and
+    what got created / updated in Phase 2.
+    """
+
+    used_llm: bool = False
+    skipped: bool = False
+    path: str = ""
+    units: list[dict] = Field(default_factory=list)
+    nodes_created: list[str] = Field(default_factory=list)
+    nodes_updated: list[str] = Field(default_factory=list)
+    summary: str = ""
+    error: str = ""
+
+
+# ============================================================
+# DreamStep — the per-file create_or_update step.
+# ============================================================
+
+
+@R.register("dream_step")
+class DreamStep(BaseStep):
+    """auto-dream create_or_update step — one file per call.
+
+    Inputs (from RuntimeContext):
+        path       (str, required): vault-relative path of one
+            daily-event note or resource file to dream over. Pass
+            empty string to no-op.
+        hint       (str, optional): caller guidance to the LLM
+            (e.g. "focus on the auth-related decisions").
+
+    Output (written to context.response.answer):
+        ``DreamResult`` JSON in ``metadata``; LLM summary in ``answer``.
+
+    CLI / MCP form:
+        reme dream path=daily/2026-05-28/auth-refactor/auth-refactor.md
+    """
+
+    def __init__(
+        self,
+        toolkit: Toolkit | None = None,
+        timezone: str | None = None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.toolkit = toolkit
+        self.timezone = timezone
+
+    def _now(self) -> datetime.datetime:
+        if self.timezone:
+            try:
+                return datetime.datetime.now(zoneinfo.ZoneInfo(self.timezone))
+            except Exception as e:
+                self.logger.error(f"Invalid timezone: {self.timezone}, error={e}")
+        return datetime.datetime.now()
+
+    def _vault_dir(self) -> Path:
+        vr = getattr(self.file_store, "vault_path", None)
+        return Path(vr).resolve() if vr else Path.cwd().resolve()
+
+    def _llm_available(self) -> bool:
+        try:
+            return self.llm is not None
+        except Exception:
+            return False
+
+    def _build_extract_toolkit(self) -> Toolkit:
+        """Read-only toolkit for the extract agent. Sub-units come back via
+        :class:`ExtractedUnits` structured output, not via a tool call."""
+        toolkit = Toolkit()
+        for job_name in _EXTRACT_TOOLS:
+            self.add_as_tool(toolkit, job_name)
+        return toolkit
+
+    def _build_integrate_toolkit(self) -> Toolkit:
+        """Full read + canonical write/edit/frontmatter_update toolkit for
+        the integrate agent. All tools are registered via :meth:`add_as_tool`
+        — same as every other step in this codebase. Outcome tracking is
+        driven by the agent's :class:`IntegrateOutcome` structured emission,
+        not by per-tool callbacks."""
+        toolkit = self.toolkit or Toolkit()
+        for job_name in _INTEGRATE_TOOLS:
+            self.add_as_tool(toolkit, job_name)
+        return toolkit
+
+    async def _extract(self, material_blob: str, hint: str, vault_dir: Path) -> tuple[list[dict], str]:
+        """Phase 1: one ReAct invocation — read material + emit ExtractedUnits.
+
+        Returns ``(units, llm_summary)`` where ``units`` is the cleaned
+        sub-unit list (each entry has ``name`` / ``bucket`` / ``summary``)
+        and ``llm_summary`` is whatever free-form text the agent produced
+        alongside its structured emission.
+        """
+        toolkit = self._build_extract_toolkit()
+        agent = Agent(
+            name="reme_dreamer_extract",
+            model=self.llm,
+            system_prompt=self.prompt_format(
+                "extract_system_prompt",
+                vault_dir=str(vault_dir),
+                buckets=", ".join(BUCKETS),
+            ),
+            toolkit=toolkit,
+            state=AgentState(
+                permission_context=PermissionContext(
+                    mode=PermissionMode.BYPASS,
+                ),
+            ),
+        )
+        user_message = self.prompt_format(
+            "extract_user_message",
+            today=self._now().strftime("%Y-%m-%d"),
+            hint=hint or "(none)",
+            material_blob=material_blob,
+        )
+        msg = await agent.reply(
+            Msg(name="reme", role="user", content=[TextBlock(text=user_message)]),
+        )
+
+        structured_resp = await self.llm.generate_structured_output(
+            agent.state.context,
+            structured_model=ExtractedUnits,
+        )
+        meta = structured_resp.content if isinstance(structured_resp.content, dict) else {}
+        cleaned: list[dict] = []
+        for raw in meta.get("units") or []:
+            if not isinstance(raw, dict):
+                continue
+            name = str(raw.get("name") or "").strip()
+            summary = str(raw.get("summary") or "").strip()
+            bucket = str(raw.get("bucket") or "").strip()
+            if not name or not summary:
+                continue
+            if bucket not in BUCKETS:
+                # Defensive: structured_model should already reject this,
+                # but if it slips through we route to wiki (the catch-all).
+                self.logger.warning(
+                    f"[{self.name}] unit {name!r} emitted bucket {bucket!r} "
+                    f"not in {list(BUCKETS)}; routing to 'wiki'",
+                )
+                bucket = "wiki"
+            cleaned.append({"name": name, "summary": summary, "bucket": bucket})
+        return cleaned, (msg.get_text_content() or "").strip()
+
+    async def _integrate_unit(self, unit: dict, material_blob: str, hint: str, vault_dir: Path) -> IntegrateOutcome:
+        """One ReAct invocation per memory sub-unit, dispatched to the
+        bucket-specific system prompt. Returns the parsed
+        :class:`IntegrateOutcome` reported by the agent — that's the
+        single source of truth for what got written (action +
+        target_path)."""
+        bucket = unit.get("bucket") or "wiki"
+        toolkit = self._build_integrate_toolkit()
+        digest_dir = getattr(self.app_context.app_config, "digest_dir", "")
+        agent = Agent(
+            name=f"reme_dreamer_integrate_{unit.get('name', 'unit')}",
+            model=self.llm,
+            system_prompt=self.prompt_format(
+                f"integrate_system_prompt_{bucket}",
+                vault_dir=str(vault_dir),
+                digest_dir=digest_dir,
+                bucket=bucket,
+            ),
+            toolkit=toolkit,
+            state=AgentState(
+                permission_context=PermissionContext(
+                    mode=PermissionMode.BYPASS,
+                ),
+            ),
+        )
+        user_message = self.prompt_format(
+            "integrate_user_message",
+            hint=hint or "(none)",
+            unit_name=unit.get("name", ""),
+            unit_bucket=bucket,
+            unit_summary=unit.get("summary", ""),
+            material_blob=material_blob,
+        )
+        await agent.reply(
+            Msg(name="reme", role="user", content=[TextBlock(text=user_message)]),
+        )
+        structured_resp = await self.llm.generate_structured_output(
+            agent.state.context,
+            structured_model=IntegrateOutcome,
+        )
+        return IntegrateOutcome.model_validate(structured_resp.content)
+
+    async def dream_one(self, path: str, hint: str = "") -> DreamResult:
+        """Run the full extract + integrate pipeline on one vault-relative
+        material path. Returns a structured :class:`DreamResult`. Safe to
+        call repeatedly on the same instance — per-invocation trackers are
+        reset at the start of each call. Used by :meth:`execute` (single
+        file from context), :class:`AutoDreamStep` (loop over today's
+        materials), and :class:`WatchDreamStep` (loop over a change batch).
+        """
+        path = (path or "").strip()
+        hint = (hint or "").strip()
+
+        if not path:
+            return DreamResult(used_llm=False, skipped=True)
+
+        if not self._llm_available():
+            return DreamResult(
+                used_llm=False,
+                skipped=True,
+                path=path,
+                error="no llm configured; dreaming requires an LLM",
+            )
+
+        material_blob = _pack_material(self.file_store, path)
+
+        vault_dir = self._vault_dir()
+
+        # Phase 1 — extract (light). Agent emits ExtractedUnits structured output to commit the
+        # memory sub-units worth lifting. Each unit carries its own bucket.
+        self.logger.info(f"[{self.name}] extract phase: path={path!r}")
+        units, extract_summary = await self._extract(material_blob, hint, vault_dir)
+
+        if not units:
+            return DreamResult(
+                used_llm=True,
+                path=path,
+                summary=extract_summary or "no memory sub-units declared",
+                skipped=True,
+            )
+
+        unit_handles = ", ".join(f"{u['name']}/{u['bucket']}" for u in units)
+        self.logger.info(f"[{self.name}] integrate phase: {len(units)} sub-unit(s): {unit_handles}")
+
+        # Phase 2 — integrate, one fresh ReAct per sub-unit, dispatched to
+        # the bucket-specific system prompt. Python-level loop, not agent
+        # loop. Each session emits a structured IntegrateOutcome whose
+        # action + target_path are the source of truth for what landed.
+        nodes_created: list[str] = []
+        nodes_updated: list[str] = []
+        per_unit_lines: list[str] = []
+        for i, unit in enumerate(units, start=1):
+            name = unit.get("name", "?")
+            bucket = unit.get("bucket", "?")
+            try:
+                outcome = await self._integrate_unit(unit, material_blob, hint, vault_dir)
+            except Exception as e:
+                self.logger.error(
+                    f"[{self.name}] integrate {i}/{len(units)} "
+                    f"(unit={name}, bucket={bucket}) failed: {type(e).__name__}: {e}",
+                )
+                per_unit_lines.append(f"[{name}/{bucket}] FAILED: {type(e).__name__}: {e}")
+                continue
+            if outcome.action == "CREATE":
+                nodes_created.append(outcome.target_path)
+            else:
+                nodes_updated.append(outcome.target_path)
+            per_unit_lines.append(_render_outcome_line(name, bucket, outcome))
+
+        per_unit_block = "\n".join(per_unit_lines)
+        summary = (
+            f"Declared {len(units)} sub-unit(s) ({unit_handles}); "
+            f"created {len(nodes_created)}, updated {len(nodes_updated)}.\n"
+            f"{per_unit_block}"
+        )
+
+        return DreamResult(
+            used_llm=True,
+            path=path,
+            units=units,
+            nodes_created=nodes_created,
+            nodes_updated=nodes_updated,
+            summary=summary,
+            skipped=False,
+        )
+
+    async def execute(self):
+        assert self.context is not None
+        path: str = (self.context.get("path", "") or "").strip()
+        hint: str = (self.context.get("hint", "") or "").strip()
+
+        result = await self.dream_one(path, hint)
+
+        if not path:
+            self.context.response.success = True
+            self.context.response.answer = "Skipped: no path supplied"
+        elif result.error:
+            self.context.response.success = False
+            self.context.response.answer = f"Error: {result.error}"
+        elif result.skipped:
+            self.context.response.success = True
+            self.context.response.answer = result.summary or "Skipped: no memory sub-units declared"
+        else:
+            self.context.response.success = True
+            self.context.response.answer = result.summary
+        self.context.response.metadata.update(result.model_dump())
diff --git a/reme4/steps/evolve/auto_dream.yaml b/reme4/steps/evolve/dream.yaml
similarity index 77%
rename from reme4/steps/evolve/auto_dream.yaml
rename to reme4/steps/evolve/dream.yaml
index b91ee2f9..cfe72a18 100644
--- a/reme4/steps/evolve/auto_dream.yaml
+++ b/reme4/steps/evolve/dream.yaml
@@ -185,20 +185,30 @@ integrate_system_prompt_procedure: |
     Plain-prose provenance does NOT count (only wikilinks survive
     future updates).
 
-  ## Recall → decide → write
+  ## Recall → classify → decide → weave
 
-  1. **Recall** — `search` (include verb stems: rotate, migrate,
-     deploy…) + `traverse depth=2 direction=both` on any hit under
-     `{digest_dir}/`. Cross-bucket on purpose: an existing match
+  1. **Recall** — call `node_search` with whatever queries best
+     fit the unit (verb stems work well: rotate, migrate, deploy…).
+     Use `limit=20-30` for broader coverage; one call usually
+     suffices but feel free to issue more if the unit spans
+     multiple concept dimensions. Recall feeds BOTH the dedup
+     judgment (same_abstraction label) and the synapse judgment
+     (related label). Cross-bucket on purpose: an existing match
      filed elsewhere beats a duplicate.
-  2. **Hit** — `frontmatter_read` triage, then `read` body for
-     survivors. Same procedure = same trigger + substantially
-     overlapping steps. New step or one-line nuance is REFINE,
-     not "different procedure".
-  3. **Decide** (exactly one):
-     - empty hit set → **CREATE** at
+  2. **Classify (internal)** — `node_search` returns name +
+     description inline; triage from that directly. Use `read`
+     only for the few survivors needing body inspection. Internally
+     label each candidate (reasoning only, not emitted to output):
+     - **same_abstraction** — same trigger + substantially
+       overlapping steps → UPDATE target (new step / nuance is
+       REFINE, not "different procedure")
+     - **related** — adjacent procedure / sub-step / failure-mode
+       cross-ref → synapse wikilink in body
+     - **unrelated** — drop
+  3. **Decide** (exactly one same_abstraction action):
+     - no same_abstraction hit → **CREATE** at
        `{digest_dir}/procedure/<slug>.md`.
-     - hit non-empty → **UPDATE** the best match:
+     - same_abstraction hit → **UPDATE** the best match:
        - **CORROBORATE** — same procedure observed again; append
          `derived_from::`, optionally strengthen wording
          ("consistently used across N runs").
@@ -208,6 +218,11 @@ integrate_system_prompt_procedure: |
        - **CORRECT** — wrong order, missing critical step, bad
          outcome; tighten or annotate inline (`> note:
          contradicted by [[new-material]] — <one-line>`).
+  4. **Synapse weave** (both CREATE and UPDATE) — weave every
+     `related` candidate from step 2 into the body as `[[Y.md]]`.
+     CREATE: woven from the start. UPDATE: additive `edit`
+     (only-add, never drop existing wikilinks). Default to weaving
+     more, not less — this is the only chance.
 
   ## Discipline
 
@@ -280,21 +295,31 @@ integrate_system_prompt_personal: |
   preference** (not one big node per person) — that's the
   granularity downstream search will hit.
 
-  ## Recall → decide → write
+  ## Recall → classify → decide → weave
 
-  1. **Recall** — `search` (user/team name + rule keywords:
-     `user-X-pr-size-pref`, `team-no-friday-deploys`) +
-     `traverse depth=2 direction=both` on any hit under
-     `{digest_dir}/`. Personal nodes often link to each other and
-     to the user's identity node; don't skip traverse.
-  2. **Hit** — `frontmatter_read` triage; `read` body for
-     survivors. Same rule = same actor scope + same governing
-     principle. A new context where the rule applies is REFINE,
-     not "different rule".
-  3. **Decide** (exactly one):
-     - empty hit set → **CREATE** at
+  1. **Recall** — call `node_search` with whatever queries best
+     fit the unit (user/team name + rule keywords:
+     `user-X-pr-size-pref`, `team-no-friday-deploys`). Use
+     `limit=20-30` for broader coverage; issue more calls if the
+     rule has multiple scope dimensions. Recall feeds BOTH the
+     dedup judgment (same_abstraction label) and the synapse
+     judgment (related label). Personal nodes often link to each
+     other and to the user's identity node — vector similarity
+     surfaces those even when literal names differ.
+  2. **Classify (internal)** — `node_search` returns name +
+     description inline; triage from that directly. Use `read`
+     only for the few survivors needing body inspection. Internally
+     label each candidate (reasoning only, not emitted to output):
+     - **same_abstraction** — same actor scope + same governing
+       principle → UPDATE target (new applicable context is
+       REFINE, not "different rule")
+     - **related** — adjacent rule / contrasting preference /
+       identity node cross-ref → synapse wikilink in body
+     - **unrelated** — drop
+  3. **Decide** (exactly one same_abstraction action):
+     - no same_abstraction hit → **CREATE** at
        `{digest_dir}/personal/<slug>.md`.
-     - hit non-empty → **UPDATE** the best match:
+     - same_abstraction hit → **UPDATE** the best match:
        - **CORROBORATE** — rule reaffirmed; append
          `derived_from::`, possibly strengthen certainty
          ("observed across N independent contexts").
@@ -305,6 +330,11 @@ integrate_system_prompt_personal: |
          evidence support, OR annotate (`> note: contradicted by
          [[new-material]] — user now prefers Y`) without
          arbitrating.
+  4. **Synapse weave** (both CREATE and UPDATE) — weave every
+     `related` candidate from step 2 into the body as `[[Y.md]]`.
+     CREATE: woven from the start. UPDATE: additive `edit`
+     (only-add, never drop existing wikilinks). Default to weaving
+     more, not less — this is the only chance.
 
   ## Discipline
 
@@ -366,23 +396,32 @@ integrate_system_prompt_wiki: |
   - **`derived_from:: [[<material-path>]]`** — at least one.
     Plain-prose provenance does NOT count.
 
-  ## Recall → decide → write
+  ## Recall → classify → decide → weave
 
-  1. **Recall** — `search` (noun phrases + common synonyms) +
-     `traverse depth=2 direction=both` on any hit under
-     `{digest_dir}/`. Skipping `traverse` is the main failure mode
-     producing duplicate concept nodes filed under different
-     terminology — semantically close abstractions often live one
-     wikilink away from a noisy hit.
-  2. **Hit** — `frontmatter_read` triage; `read` body for
-     survivors. Same abstraction = same definition / principle in
-     the body, even if wording differs. Slightly different framing
-     of the same idea is REFINE; outright different concepts are
-     different nodes.
-  3. **Decide** (exactly one):
-     - empty hit set → **CREATE** at
+  1. **Recall** — call `node_search` with whatever queries best
+     fit the unit (noun phrases + common synonyms). Use
+     `limit=20-30` for broader coverage; issue more calls when
+     the abstraction has multiple aspects worth querying
+     separately. Recall feeds BOTH the dedup judgment
+     (same_abstraction label) and the synapse judgment
+     (related label). Vector similarity catches abstractions
+     filed under different terminology even when surface words
+     don't overlap.
+  2. **Classify (internal)** — `node_search` returns name +
+     description inline; triage from that directly. Use `read`
+     only for the few survivors needing body inspection. Internally
+     label each candidate (reasoning only, not emitted to output):
+     - **same_abstraction** — same definition / principle in body,
+       even if wording differs → UPDATE target (slightly different
+       framing is REFINE; outright different concepts are
+       different nodes)
+     - **related** — concept-adjacent / contrasts / sup-/sub-
+       concept / instance cross-ref → synapse wikilink in body
+     - **unrelated** — drop
+  3. **Decide** (exactly one same_abstraction action):
+     - no same_abstraction hit → **CREATE** at
        `{digest_dir}/wiki/<slug>.md`.
-     - hit non-empty → **UPDATE** the best match:
+     - same_abstraction hit → **UPDATE** the best match:
        - **CORROBORATE** — principle reaffirmed by new instance;
          append `derived_from::`, optionally strengthen wording
          ("consistently observed across N sources" / replace
@@ -396,6 +435,11 @@ integrate_system_prompt_wiki: |
          evidence support, or annotate inline (`> note:
          contradicted by [[new-material]] — <one-line>`) without
          arbitrating.
+  4. **Synapse weave** (both CREATE and UPDATE) — weave every
+     `related` candidate from step 2 into the body as `[[Y.md]]`.
+     CREATE: woven from the start. UPDATE: additive `edit`
+     (only-add, never drop existing wikilinks). Default to weaving
+     more, not less — this is the only chance.
 
   ## Discipline
 
@@ -590,19 +634,26 @@ integrate_system_prompt_procedure_zh: |
   - **`derived_from:: [[<material-path>]]`** —— 至少一条。纯
     散文形式 **不算**(下次 update 时会消失)。
 
-  ## 召回 → 决策 → 写入
+  ## 召回 → 内化分类 → 决策 → 织突触
 
-  1. **召回** —— `search`(带动词词根:rotate / migrate /
-     deploy…)+ 对 `{digest_dir}/` 下 **任何** 命中跑
-     `traverse depth=2 direction=both`。**跨 bucket** 是有意 ——
-     就地更新优于复制创建。
-  2. **命中** —— `frontmatter_read` 廉价 triage,幸存者用
-     `read` 读完整 body。"同一流程" = 同触发 + 步骤大幅重叠。
-     新增一步 / 细微差异是 REFINE,**不是** 另一个流程。
-  3. **决策**(恰好一种):
-     - 命中空 → **CREATE** 在
+  1. **召回** —— 调 `node_search`,query 由你自己根据 unit 内容
+     选(带动词词根效果好:rotate / migrate / deploy…)。`limit=
+     20-30` 取更宽覆盖;通常一次够,但若 unit 跨多个概念维度可
+     多调几次。召回结果同时服务 dedup(same_abstraction label)
+     和 synapse(related label)两类判断。**跨 bucket** 是有意
+     —— 就地更新优于复制创建。
+  2. **内化分类** —— `node_search` 已内嵌返回 name + description,
+     直接据此 triage。仅对需要看 body 的少数候选用 `read`。
+     对每个候选**内化打 label**(只在思考中分类,不输出):
+     - **same_abstraction** —— 同触发 + 步骤大幅重叠 → UPDATE
+       目标(新增一步 / 细微差异是 REFINE,**不是** 另一个流程)
+     - **related** —— 邻近流程 / 子步骤 / 失败模式互引 → 织成
+       body 内的 synapse wikilink
+     - **unrelated** —— 丢弃
+  3. **决策**(恰好一个 same_abstraction 动作):
+     - 无 same_abstraction 命中 → **CREATE** 在
        `{digest_dir}/procedure/<slug>.md`。
-     - 命中非空 → **UPDATE** 最匹配的:
+     - same_abstraction 命中 → **UPDATE** 最匹配的:
        - **CORROBORATE** —— 同流程再次出现;加 `derived_from::`,
          可选强化措辞("跨 N 次运行一致使用");步骤不动。
        - **REFINE** —— 新前置 / 边界 / 失败模式;扩展相关片段,
@@ -610,6 +661,10 @@ integrate_system_prompt_procedure_zh: |
        - **CORRECT** —— 顺序错 / 缺关键步 / 结果不对;收紧或
          内联标注(`> note: contradicted by [[new-material]] —
          <一句话>`)。
+  4. **织突触**(CREATE 与 UPDATE 都要做)—— 把第 2 步所有
+     `related` 候选织入 body 作 `[[Y.md]]`。CREATE:写入时一次性
+     织全。UPDATE:additive `edit`(只增不删,绝不丢已有 wikilink)。
+     宁可多织 —— **这是唯一机会**。
 
   ## 纪律
 
@@ -673,20 +728,28 @@ integrate_system_prompt_personal_zh: |
   同一人有多条偏好时,**一条偏好一个节点**(不是一个人一大
   节点) —— 这才是下游搜索的粒度。
 
-  ## 召回 → 决策 → 写入
+  ## 召回 → 内化分类 → 决策 → 织突触
 
-  1. **召回** —— `search`(user / team 名 + 规则关键词:
-     `user-X-pr-size-pref`、`team-no-friday-deploys`)+ 对
-     `{digest_dir}/` 下 **任何** 命中跑
-     `traverse depth=2 direction=both`。personal 节点常彼此互
-     链并指向用户身份节点;**别跳过 traverse**。
-  2. **命中** —— `frontmatter_read` triage,幸存者 `read` body。
-     "同一规则" = 同 actor 范围 + 同支配原则。新增"规则适用情
-     境"是 REFINE,**不是** 另一条规则。
-  3. **决策**(恰好一种):
-     - 命中空 → **CREATE** 在
+  1. **召回** —— 调 `node_search`,query 由你自己根据 unit 内容
+     选(user / team 名 + 规则关键词:`user-X-pr-size-pref`、
+     `team-no-friday-deploys`)。`limit=20-30` 取更宽覆盖;规则
+     若有多个 scope 维度可多调几次。召回结果同时服务 dedup
+     (same_abstraction label)和 synapse(related label)两类
+     判断。personal 节点常彼此互链并指向用户身份节点 —— vector
+     相似度能在字面名不同时也召回这些。
+  2. **内化分类** —— `node_search` 已内嵌返回 name + description,
+     直接据此 triage。仅对需要看 body 的少数候选用 `read`。
+     对每个候选**内化打 label**(只在思考中分类,不输出):
+     - **same_abstraction** —— 同 actor 范围 + 同支配原则 →
+       UPDATE 目标(新增"规则适用情境"是 REFINE,**不是** 另一
+       条规则)
+     - **related** —— 邻近规则 / 对比偏好 / 用户身份节点互引 →
+       织成 body 内的 synapse wikilink
+     - **unrelated** —— 丢弃
+  3. **决策**(恰好一个 same_abstraction 动作):
+     - 无 same_abstraction 命中 → **CREATE** 在
        `{digest_dir}/personal/<slug>.md`。
-     - 命中非空 → **UPDATE** 最匹配的:
+     - same_abstraction 命中 → **UPDATE** 最匹配的:
        - **CORROBORATE** —— 规则在新场景再次坐实;加
          `derived_from::`,可选强化确定性("跨 N 个独立情境
          观察")。
@@ -696,6 +759,10 @@ integrate_system_prompt_personal_zh: |
          新旧证据都支持的形式,或内联标注
          (`> note: contradicted by [[new-material]] — 用户现在
          偏好 Y`)不仲裁。
+  4. **织突触**(CREATE 与 UPDATE 都要做)—— 把第 2 步所有
+     `related` 候选织入 body 作 `[[Y.md]]`。CREATE:写入时一次性
+     织全。UPDATE:additive `edit`(只增不删,绝不丢已有 wikilink)。
+     宁可多织 —— **这是唯一机会**。
 
   ## 纪律
 
@@ -750,19 +817,27 @@ integrate_system_prompt_wiki_zh: |
   - **`derived_from:: [[<material-path>]]`** —— 至少一条;纯
     散文形式不算。
 
-  ## 召回 → 决策 → 写入
+  ## 召回 → 内化分类 → 决策 → 织突触
 
-  1. **召回** —— `search`(名词短语 + 常见同义词)+ 对
-     `{digest_dir}/` 下 **任何** 命中跑
-     `traverse depth=2 direction=both`。跳过 `traverse` 是产生
-     重复概念节点(不同 slug 同语义)的主要失败模式 —— 语义相
-     邻的抽象常常就在某个噪音命中的一跳之外。
-  2. **命中** —— `frontmatter_read` triage,幸存者 `read` body。
-     "同一抽象" = body 中的定义 / 原则相同(措辞可不同)。同
-     思想的略不同表述是 REFINE;真正不同的概念是不同节点。
-  3. **决策**(恰好一种):
-     - 命中空 → **CREATE** 在 `{digest_dir}/wiki/<slug>.md`。
-     - 命中非空 → **UPDATE** 最匹配的:
+  1. **召回** —— 调 `node_search`,query 由你自己根据 unit 内容
+     选(名词短语 + 常见同义词)。`limit=20-30` 取更宽覆盖;若
+     抽象本身有多个侧面,值得分别查询时可多调几次。召回结果
+     同时服务 dedup(same_abstraction label)和 synapse
+     (related label)两类判断。vector 相似度能捕获以不同术语
+     归档的同语义抽象,即便字面词不重叠。
+  2. **内化分类** —— `node_search` 已内嵌返回 name + description,
+     直接据此 triage。仅对需要看 body 的少数候选用 `read`。
+     对每个候选**内化打 label**(只在思考中分类,不输出):
+     - **same_abstraction** —— body 中的定义 / 原则相同(措辞
+       可不同)→ UPDATE 目标(同思想的略不同表述是 REFINE;真正
+       不同的概念是不同节点)
+     - **related** —— 概念邻近 / 对比 / 上位下位 / 实例互引 →
+       织成 body 内的 synapse wikilink
+     - **unrelated** —— 丢弃
+  3. **决策**(恰好一个 same_abstraction 动作):
+     - 无 same_abstraction 命中 → **CREATE** 在
+       `{digest_dir}/wiki/<slug>.md`。
+     - same_abstraction 命中 → **UPDATE** 最匹配的:
        - **CORROBORATE** —— 原则被新实例再坐实;加
          `derived_from::`,可选强化措辞("跨 N 个来源一致观
          察"、把"似乎"换成"确实");正文实质不变。
@@ -771,6 +846,10 @@ integrate_system_prompt_wiki_zh: |
        - **CORRECT** —— 事实矛盾或夸大;收紧到新旧证据都支持
          的窄形式,或内联标注(`> note: contradicted by
          [[new-material]] — <一句话>`)不仲裁。
+  4. **织突触**(CREATE 与 UPDATE 都要做)—— 把第 2 步所有
+     `related` 候选织入 body 作 `[[Y.md]]`。CREATE:写入时一次性
+     织全。UPDATE:additive `edit`(只增不删,绝不丢已有 wikilink)。
+     宁可多织 —— **这是唯一机会**。
 
   ## 纪律
 
diff --git a/reme4/steps/file_io/daily_list.py b/reme4/steps/file_io/daily_list.py
index 9e9976de..549a6404 100644
--- a/reme4/steps/file_io/daily_list.py
+++ b/reme4/steps/file_io/daily_list.py
@@ -61,3 +61,4 @@ class DailyListStep(BaseStep):
         lines = [self._format_note_line(n) for n in notes]
         self.context.response.answer = "\n".join(lines) if lines else f"No notes found for {day}"
         self.context.response.metadata.update({"date": day, "count": len(notes)})
+        self.logger.info(f"[{self.name}] date={day} notes={len(notes)}")
diff --git a/reme4/steps/file_io/daily_reindex.py b/reme4/steps/file_io/daily_reindex.py
index 8ef38953..e09cf66e 100644
--- a/reme4/steps/file_io/daily_reindex.py
+++ b/reme4/steps/file_io/daily_reindex.py
@@ -44,6 +44,7 @@ class DailyReindexStep(BaseStep):
             self.context.response.success = False
             self.context.response.answer = f"Error: {refreshed['error']}"
             self.context.response.metadata.update(refreshed)
+            self.logger.info(f"[{self.name}] reindex failed error={refreshed['error']!r}")
             return
         notes_count = len(refreshed["notes"])
         self.context.response.success = True
@@ -56,6 +57,10 @@ class DailyReindexStep(BaseStep):
                 "notes_count": notes_count,
             },
         )
+        self.logger.info(
+            f"[{self.name}] date={refreshed['date']} path={refreshed['path']} "
+            f"created={refreshed['created']} notes={notes_count}",
+        )
 
     async def execute(self):
         """Trigger the index rebuild and stamp the response."""
diff --git a/reme4/steps/file_io/delete.py b/reme4/steps/file_io/delete.py
index 7d49a152..abac1fea 100644
--- a/reme4/steps/file_io/delete.py
+++ b/reme4/steps/file_io/delete.py
@@ -47,12 +47,23 @@ class DeleteStep(BaseStep):
         if "error" in payload:
             self.context.response.success = False
             self.context.response.answer = f"Error: {payload['error']}"
+            self.logger.info(f"[{self.name}] delete failed path={path} error={payload['error']!r}")
         elif payload.get("is_dir"):
             self.context.response.success = True
             self.context.response.answer = f"Deleted directory {path} ({len(payload['deleted_files'])} file(s))"
+            self.logger.info(
+                f"[{self.name}] deleted dir path={path} files={len(payload['deleted_files'])} "
+                f"inbound_files={payload['inbound']['files_touched']} "
+                f"inbound_links={payload['inbound']['links_total']}",
+            )
         else:
             self.context.response.success = True
             self.context.response.answer = f"Deleted {path}"
+            self.logger.info(
+                f"[{self.name}] deleted file path={path} "
+                f"inbound_files={payload['inbound']['files_touched']} "
+                f"inbound_links={payload['inbound']['links_total']}",
+            )
         self.context.response.metadata.update(payload)
 
     async def _delete(self, path: str) -> dict:
diff --git a/reme4/steps/file_io/frontmatter_delete.py b/reme4/steps/file_io/frontmatter_delete.py
index 2091d515..d1f8bb3a 100644
--- a/reme4/steps/file_io/frontmatter_delete.py
+++ b/reme4/steps/file_io/frontmatter_delete.py
@@ -59,7 +59,11 @@ class FrontmatterDeleteStep(BaseStep):
         if "error" in payload:
             self.context.response.success = False
             self.context.response.answer = f"Error: {payload['error']}"
+            self.logger.info(f"[{self.name}] delete failed path={path} error={payload['error']!r}")
         else:
             self.context.response.success = True
             self.context.response.answer = f"Deleted {len(payload['deleted'])} key(s) from {path}"
+            self.logger.info(
+                f"[{self.name}] path={path} deleted={payload['deleted']} missing={payload['missing']}",
+            )
         self.context.response.metadata.update(payload)
diff --git a/reme4/steps/file_io/frontmatter_read.py b/reme4/steps/file_io/frontmatter_read.py
index c9c28e82..26dbcc7a 100644
--- a/reme4/steps/file_io/frontmatter_read.py
+++ b/reme4/steps/file_io/frontmatter_read.py
@@ -31,11 +31,13 @@ class FrontmatterReadStep(BaseStep):
             self.context.response.success = False
             self.context.response.answer = f"Error: {path} not found"
             self.context.response.metadata.update({"path": path, "exists": False})
+            self.logger.info(f"[{self.name}] path={path} exists=False")
             return
         if target.suffix != ".md":
             self.context.response.success = False
             self.context.response.answer = "Error: not markdown"
             self.context.response.metadata.update({"path": path, "error": "not markdown"})
+            self.logger.info(f"[{self.name}] path={path} error=not_markdown")
             return
 
         try:
@@ -44,7 +46,9 @@ class FrontmatterReadStep(BaseStep):
             self.context.response.success = False
             self.context.response.answer = f"Error: failed to parse frontmatter in {path}: {exc}"
             self.context.response.metadata.update({"path": path, "exists": True, "error": str(exc)})
+            self.logger.info(f"[{self.name}] path={path} parse_error={exc!r}")
             return
         self.context.response.success = True
         self.context.response.answer = f"Read frontmatter from {path} ({len(meta)} key(s))"
         self.context.response.metadata.update({"path": path, "exists": True, "frontmatter": meta})
+        self.logger.info(f"[{self.name}] path={path} keys={len(meta)}")
diff --git a/reme4/steps/file_io/frontmatter_update.py b/reme4/steps/file_io/frontmatter_update.py
index d1ad5ad4..fce9a1d0 100644
--- a/reme4/steps/file_io/frontmatter_update.py
+++ b/reme4/steps/file_io/frontmatter_update.py
@@ -47,7 +47,9 @@ class FrontmatterUpdateStep(BaseStep):
         if "error" in payload:
             self.context.response.success = False
             self.context.response.answer = f"Error: {payload['error']}"
+            self.logger.info(f"[{self.name}] update failed path={path} error={payload['error']!r}")
         else:
             self.context.response.success = True
             self.context.response.answer = f"Updated {len(metadata)} key(s) on {path}"
+            self.logger.info(f"[{self.name}] path={path} keys={list(metadata.keys())}")
         self.context.response.metadata.update(payload)
diff --git a/reme4/steps/file_io/move.py b/reme4/steps/file_io/move.py
index ffb33c75..de64d050 100644
--- a/reme4/steps/file_io/move.py
+++ b/reme4/steps/file_io/move.py
@@ -52,9 +52,16 @@ class MoveStep(BaseStep):
         if "error" in payload:
             self.context.response.success = False
             self.context.response.answer = f"Error: {payload['error']}"
+            self.logger.info(f"[{self.name}] move failed src={src_path} dst={dst_path} error={payload['error']!r}")
         else:
             self.context.response.success = True
             self.context.response.answer = f"Moved {src_path} → {dst_path}"
+            retarget_info = payload.get("retarget") or {}
+            self.logger.info(
+                f"[{self.name}] moved src={src_path} dst={dst_path} src_removed={payload.get('src_removed')} "
+                f"retarget_files={retarget_info.get('files_touched', 0) if isinstance(retarget_info, dict) else '-'} "
+                f"retarget_links={retarget_info.get('links_changed', 0) if isinstance(retarget_info, dict) else '-'}",
+            )
         self.context.response.metadata.update(payload)
 
     async def _move(self, src_path: str, dst_path: str, overwrite: bool, retarget: bool) -> dict:
diff --git a/reme4/steps/file_io/stat.py b/reme4/steps/file_io/stat.py
index 97a9dcd8..41a82415 100644
--- a/reme4/steps/file_io/stat.py
+++ b/reme4/steps/file_io/stat.py
@@ -42,6 +42,7 @@ class StatStep(BaseStep):
             self.context.response.success = False
             self.context.response.answer = f"stat: {path} not found"
             self.context.response.metadata.update({"path": path, "exists": False})
+            self.logger.info(f"[{self.name}] path={path} exists=False")
             return
 
         st = target.stat()
@@ -69,3 +70,7 @@ class StatStep(BaseStep):
         self.context.response.success = True
         self.context.response.answer = answer
         self.context.response.metadata.update(payload)
+        self.logger.info(
+            f"[{self.name}] path={path} type={payload['type']} "
+            f"size={payload.get('size', '-')} mime={payload.get('mime', '-')}",
+        )
diff --git a/reme4/steps/index/channel_notify.py b/reme4/steps/index/channel_notify.py
new file mode 100644
index 00000000..099c8d56
--- /dev/null
+++ b/reme4/steps/index/channel_notify.py
@@ -0,0 +1,67 @@
+"""``channel_notify_step`` — push a debounced batch of vault changes as one channel event.
+
+Designed to be slotted into ``watch_changes_step.dispatch_step`` next to
+``update_index_step``: when the watcher emits a batch of changes, this
+step forwards a single human-readable summary to
+``ApplicationContext.metadata["channel_sink"]``. The Claude Code main
+session then sees a ``<channel source="reme" kind="vault_change" ...>``
+tag and reacts per the server's ``instructions``.
+
+One event per batch (not per file) — the watcher already de-bounces and
+de-duplicates, so a batch is a meaningful "things that changed together"
+unit. Putting N events on the wire per batch would multiply session
+turns without adding signal.
+
+No-op (silently) when:
+
+* ``channel_sink`` is absent from the application context metadata
+  (e.g. service wasn't an ``MCPService``), so this step is safe in
+  any pipeline.
+* ``context['changes']`` is missing or empty.
+* The sink itself has no bound session (no client called ``claim_channel``).
+"""
+
+from pathlib import Path
+
+from ..base_step import BaseStep
+from ...components import R
+
+
+@R.register("channel_notify_step")
+class ChannelNotifyStep(BaseStep):
+    """Forward a batch of vault changes to the Claude Code channel."""
+
+    async def execute(self):
+        sink = self.app_context.metadata.get("channel_sink") if self.app_context is not None else None
+        if sink is None:
+            return
+
+        changes = (self.context.get("changes", []) if self.context is not None else []) or []
+        if not changes:
+            return
+
+        # Render paths vault-relative so the agent can pass them directly to
+        # slash commands like /dream <path>. Absolute paths that fall outside
+        # the vault are left as-is rather than erroring — they shouldn't occur,
+        # but a stray entry shouldn't kill the event.
+        vault = self.vault_path
+        lines: list[str] = []
+        for change in changes:
+            try:
+                raw = Path(change["path"])
+            except (KeyError, TypeError):
+                continue
+            try:
+                shown = str(raw.resolve().relative_to(vault))
+            except ValueError:
+                shown = str(raw)
+            lines.append(f"{change.get('change', '?')}: {shown}")
+
+        if not lines:
+            return
+
+        self.logger.info(f"[channel_notify] emit batch count={len(lines)}")
+        await sink.emit(
+            content="Vault 变更:\n" + "\n".join(lines),
+            meta={"kind": "vault_change", "count": str(len(lines))},
+        )
diff --git a/reme4/steps/index/claim_channel.py b/reme4/steps/index/claim_channel.py
new file mode 100644
index 00000000..756b0afa
--- /dev/null
+++ b/reme4/steps/index/claim_channel.py
@@ -0,0 +1,56 @@
+"""``claim_channel_step`` — let an MCP client elect itself as the ``<channel>`` recipient.
+
+The single bind path for every transport (stdio, sse, streamable-http):
+this step uses ``fastmcp.server.dependencies.get_context()`` to grab the
+current request's ``ServerSession`` and ``ChannelSink.bind`` it.
+
+Semantics:
+
+* **Last-claim-wins.** A second client calling ``claim_channel`` silently
+  replaces the previous binding; the prior leader stops receiving events.
+* **Lossy on leader loss.** If the bound session goes away, the next
+  ``send_message`` raises and ``ChannelSink`` swallows it as a warning.
+  Events drop until another client claims.
+* **stdio = trivially the one client.** Under stdio there is exactly one
+  session ever; calling ``claim_channel`` once after init binds it for
+  the rest of the server's life. Until then, channel events drop.
+"""
+
+from ..base_step import BaseStep
+from ...components import R
+
+
+@R.register("claim_channel_step")
+class ClaimChannelStep(BaseStep):
+    """Bind the current MCP session as the ``<channel>`` recipient."""
+
+    async def execute(self):
+        assert self.context is not None
+
+        try:
+            from fastmcp.server.dependencies import get_context
+
+            ctx = get_context()
+            session = ctx.session
+            assert session is not None, "FastMCP context has no ServerSession"
+            assert self.app_context is not None, "claim_channel requires an application context"
+            sink = self.app_context.metadata.get("channel_sink")
+            assert sink is not None, "channel_sink not configured on application context metadata"
+        except Exception as e:
+            self.context.response.answer = {"claimed": False, "reason": f"{type(e).__name__}: {e}"}
+            self.context.response.metadata["claimed"] = False
+            return self.context.response
+
+        sink.bind(session)
+        session_id = ctx.session_id or "<unknown>"
+        self.logger.info(f"[claim_channel] channel bound to session={session_id}")
+        self.context.response.answer = {
+            "claimed": True,
+            "session_id": session_id,
+            "note": (
+                "this session now receives <channel source='reme'> notifications. "
+                "last-claim-wins: another call to claim_channel takes over."
+            ),
+        }
+        self.context.response.metadata["claimed"] = True
+        return self.context.response
diff --git a/reme4/steps/index/node_search.py b/reme4/steps/index/node_search.py
new file mode 100644
index 00000000..cef24c82
--- /dev/null
+++ b/reme4/steps/index/node_search.py
@@ -0,0 +1,158 @@
+"""``node_search_step`` — node-level digest search for dream Phase 2.
+
+Specialized for dream's recall needs; **NOT** a drop-in replacement for the
+general-purpose ``search`` step (which serves external RAG agents).
+
+Five differences vs ``search``:
+
+1. **Node-level results** — same-path chunks aggregated by max score; one
+   row per digest node, not per chunk.
+2. **Frontmatter included** — returns ``name + description`` inline so the
+   caller can triage without a follow-up ``frontmatter_read`` per hit.
+3. **Digest-only filter** — hardcoded to ``<digest_dir>/`` prefix; dream
+   never wants daily / resource hits as recall candidates.
+4. **No expand_links** — dream's synapse recall is looking for nodes that
+   *don't* yet have wikilinks; expansion would surface already-linked
+   neighbors (anti-pattern for synapse construction).
+5. **No body / chunk text in response** — caller follows up with ``read``
+   only on the few candidates that need deep inspection, not all.
+
+A single hybrid (vector + BM25 RRF) recall serves **both** the dedup
+judgment (`same_abstraction` label — is any candidate the same as the
+new unit?) and the synapse judgment (`related` label — which candidates
+should be woven as wikilinks?). They are two LLM-internal labels over
+the **same candidate pool**; there is no need to split into separate
+recall passes (the previous ``mode={hybrid,vector_only}`` toggle was
+spurious — both judgments operate on the same recall output).
+
+Used in dream Phase 2 (dream.yaml ``integrate_system_prompt_*``).
+External agents must keep using ``search`` for chunk-level retrieval +
+link expansion.
+"""
+
+import asyncio
+from collections import defaultdict
+
+from ..base_step import BaseStep
+from ...components import R
+
+_RRF_K = 60
+_MAX_CANDIDATES = 200
+
+
+def _rrf_merge_nodes(
+    vector_paths: list[str],
+    keyword_paths: list[str],
+    vector_weight: float,
+) -> dict[str, float]:
+    """RRF fuse two ranked path lists into node-level scores."""
+    scores: dict[str, float] = defaultdict(float)
+    text_weight = 1.0 - vector_weight
+    for rank, path in enumerate(vector_paths, start=1):
+        scores[path] += vector_weight / (_RRF_K + rank)
+    for rank, path in enumerate(keyword_paths, start=1):
+        scores[path] += text_weight / (_RRF_K + rank)
+    return scores
+
+
+@R.register("node_search_step")
+class NodeSearchStep(BaseStep):
+    """Node-level digest-only hybrid search for dream Phase 2 recall."""
+
+    async def execute(self):
+        assert self.context is not None
+        query: str = (self.context.get("query", "") or "").strip()
+        limit: int = int(self.context.get("limit") or 20)
+        vector_weight: float = float(self.kwargs.get("vector_weight", 0.7))
+        candidate_multiplier: float = float(self.kwargs.get("candidate_multiplier", 5.0))
+
+        if not query:
+            self.context.response.success = False
+            self.context.response.answer = "Error: query cannot be empty"
+            return self.context.response
+        assert limit > 0, f"limit must be positive, got {limit}"
+
+        digest_dir = getattr(self.app_context.app_config, "digest_dir", "digest") or "digest"
+        digest_prefix = digest_dir.rstrip("/") + "/"
+
+        # Over-fetch — digest filter drops a lot of raw hits.
+        candidates = min(_MAX_CANDIDATES, max(1, int(limit * candidate_multiplier)))
+
+        vector_chunks, keyword_chunks = await asyncio.gather(
+            self.file_store.vector_search(query, candidates, {}),
+            self.file_store.keyword_search(query, candidates, {}),
+        )
+
+        def _node_dedup(chunks: list) -> list[str]:
+            """Keep first occurrence per path; respect digest prefix only.
+
+            Self-exclusion (e.g. UPDATE target) is the LLM's job: frontmatter
+            inlining lets the agent recognize self from the candidate list.
+            No need to push it into a mechanical parameter.
+            """
+            seen: set[str] = set()
+            out: list[str] = []
+            for c in chunks:
+                if c.path in seen:
+                    continue
+                if not c.path.startswith(digest_prefix):
+                    continue
+                seen.add(c.path)
+                out.append(c.path)
+            return out
+
+        vector_paths = _node_dedup(vector_chunks)
+        keyword_paths = _node_dedup(keyword_chunks)
+        path_to_score = _rrf_merge_nodes(vector_paths, keyword_paths, vector_weight)
+
+        ranked = sorted(path_to_score.items(), key=lambda kv: -kv[1])[:limit]
+
+        # Attach frontmatter from in-memory FileNode metadata (no extra IO).
+        node_paths = [p for p, _ in ranked]
+        nodes = await self.file_store.get_nodes(node_paths) if node_paths else []
+        path_to_fm: dict[str, dict[str, str]] = {}
+        for n in nodes:
+            fm = n.front_matter
+            path_to_fm[n.path] = {
+                "name": (getattr(fm, "name", "") or "").strip(),
+                "description": (getattr(fm, "description", "") or "").strip(),
+            }
+
+        hits: list[dict] = []
+        for path, score in ranked:
+            fm = path_to_fm.get(path, {})
+            hits.append(
+                {
+                    "path": path,
+                    "score": round(float(score), 4),
+                    "name": fm.get("name", ""),
+                    "description": fm.get("description", ""),
+                },
+            )
+
+        self.logger.info(
+            f"[{self.name}] query={query!r} candidates={candidates} "
+            f"vector_hits={len(vector_chunks)} keyword_hits={len(keyword_chunks)} "
+            f"returned={len(hits)}",
+        )
+
+        lines = [
+            f"=== node_search query={query!r} hits={len(hits)}/{candidates} ===",
+        ]
+        if not hits:
+            lines.append("(no digest hits)")
+        else:
+            for h in hits:
+                lines.append(
+                    f"[{h['score']:.4f}] {h['path']}\n" f"  name: {h['name']}\n" f"  description: {h['description']}",
+                )
+
+        self.context.response.success = True
+        self.context.response.answer = "\n".join(lines)
+        self.context.response.metadata["hits"] = hits
+        self.context.response.metadata["counts"] = {
+            "vector_raw": len(vector_chunks),
+            "keyword_raw": len(keyword_chunks),
+            "returned": len(hits),
+        }
+        return self.context.response
diff --git a/reme4/steps/index/scan_changes.py b/reme4/steps/index/scan_changes.py
index 3c1ee7d9..2a8c8c6e 100644
--- a/reme4/steps/index/scan_changes.py
+++ b/reme4/steps/index/scan_changes.py
@@ -1,61 +1,102 @@
-"""One-shot scan: diff watch_paths vs file_store and write changes into context.
+"""One-shot scan: diff watch_paths vs an indexed-state source, write changes into context.
 
-Designed to be chained before ``update_index_step`` so that the second step
-performs the actual writes and persistence.
+Two symmetric variants — pick the one whose state source matches the loop's
+write target so the index loop and the dream loop never contend on the
+same component:
+
+* :class:`ScanStoreChangesStep` (``scan_store_changes_step``) — diffs
+  against ``file_store``; used by ``update_store_index_loop`` (sole
+  writer of ``file_store``).
+* :class:`ScanCatalogChangesStep` (``scan_catalog_changes_step``) — diffs
+  against ``file_catalog``; used by ``auto_dream_loop`` (sole writer of
+  ``file_catalog``).
+
+Both share the same diff vocabulary (``added`` / ``modified`` / ``deleted``)
+and write into ``context["changes"]`` in the same shape, so downstream
+steps don't care which variant produced the batch.
 """
 
 from pathlib import Path
+from typing import Iterable
 
-from ..base_step import BaseStep
+from ..base_step import BaseStep, Ref
 from ...components import R
+from ...components.file_catalog import BaseFileCatalog
+from ...enumeration import ComponentEnum
+from ...schema import FileNode
 
 
-@R.register("scan_changes_step")
-class ScanChangesStep(BaseStep):
-    """One-shot scan: compute added/modified/deleted vs file_store and write to context."""
+def _collect_existing(
+    raw: list[str] | str,
+    suffixes: list[str],
+    vault_path: Path,
+    recursive: bool,
+) -> dict[str, float]:
+    """Walk watch_paths under ``vault_path`` and return ``{abs_path: st_mtime}``."""
+    paths = [raw] if isinstance(raw, str) else raw
+    watch_paths = [vault_path / x for x in paths if (vault_path / x).exists()]
+
+    existing: dict[str, float] = {}
+    for path in watch_paths:
+        candidates = [path] if path.is_file() else (path.rglob("*") if recursive else path.iterdir())
+        for p in candidates:
+            if not p.is_file():
+                continue
+            if suffixes and not any(str(p).endswith("." + s.strip(".")) for s in suffixes):
+                continue
+            abs_p = p.absolute()
+            existing[str(abs_p)] = abs_p.stat().st_mtime
+    return existing
+
+
+def _diff(
+    existing: dict[str, float],
+    nodes: Iterable[FileNode],
+    vault_path: Path,
+) -> tuple[list[dict], dict[str, int]]:
+    """Compute added/modified/deleted vs ``nodes`` and return (changes, counts)."""
+    indexed: dict[str, float] = {
+        str(Path(n.path) if Path(n.path).is_absolute() else vault_path / n.path): n.st_mtime for n in nodes
+    }
+
+    to_delete = list(indexed.keys() - existing.keys())
+    to_add = list(existing.keys() - indexed.keys())
+    to_modify = [p for p in existing.keys() & indexed.keys() if existing[p] != indexed[p]]
+
+    changes: list[dict] = (
+        [{"change": "added", "path": p} for p in to_add]
+        + [{"change": "modified", "path": p} for p in to_modify]
+        + [{"change": "deleted", "path": p} for p in to_delete]
+    )
+    counts = {"added": len(to_add), "modified": len(to_modify), "deleted": len(to_delete)}
+    return changes, counts
+
+
+class _ScanChangesBase(BaseStep):
+    """Shared scaffolding: collect on-disk state, defer node-loading to the subclass."""
 
     def __init__(self, recursive: bool = True, **kwargs):
         super().__init__(**kwargs)
         self.recursive: bool = recursive
 
+    async def _load_indexed_nodes(self) -> Iterable[FileNode]:
+        raise NotImplementedError
+
     async def execute(self):
         assert self.context is not None
-        if self.file_store is None:
-            raise RuntimeError("file_store is not initialized!")
-
-        raw: list[str] = self.context.get("watch_paths", [])
-        suffixes: list[str] = self.context.get("suffix_filters", ["md"])
         vault_path = self.vault_path
+        raw: list[str] = self.context.get("watch_paths", []) or []
+        suffixes: list[str] = self.context.get("suffix_filters", ["md"]) or ["md"]
 
-        paths = [raw] if isinstance(raw, str) else raw
-        watch_paths = [vault_path / x for x in paths if (vault_path / x).exists()]
-
-        existing: dict[str, float] = {}
-        for path in watch_paths:
-            candidates = [path] if path.is_file() else (path.rglob("*") if self.recursive else path.iterdir())
-            for p in candidates:
-                if not p.is_file():
-                    continue
-                if suffixes and not any(str(p).endswith("." + s.strip(".")) for s in suffixes):
-                    continue
-                abs_p = p.absolute()
-                existing[str(abs_p)] = abs_p.stat().st_mtime
-
-        indexed: dict[str, float] = {
-            str(Path(n.path) if Path(n.path).is_absolute() else vault_path / n.path): n.st_mtime
-            for n in await self.file_store.get_nodes()
-        }
-
-        to_delete = list(indexed.keys() - existing.keys())
-        to_add = list(existing.keys() - indexed.keys())
-        to_modify = [p for p in existing.keys() & indexed.keys() if existing[p] != indexed[p]]
-
-        changes: list[dict] = (
-            [{"change": "added", "path": p} for p in to_add]
-            + [{"change": "modified", "path": p} for p in to_modify]
-            + [{"change": "deleted", "path": p} for p in to_delete]
+        existing = _collect_existing(
+            raw=raw,
+            suffixes=suffixes,
+            vault_path=vault_path,
+            recursive=self.recursive,
         )
-        counts = {"added": len(to_add), "modified": len(to_modify), "deleted": len(to_delete)}
+
+        nodes = await self._load_indexed_nodes()
+        changes, counts = _diff(existing, nodes, vault_path)
 
         self.context["changes"] = changes
         if changes:
@@ -65,3 +106,23 @@ class ScanChangesStep(BaseStep):
 
         self.context.response.metadata["counts"] = counts
         return self.context.response
+
+
+@R.register("scan_store_changes_step")
+class ScanStoreChangesStep(_ScanChangesBase):
+    """Diff vault against ``file_store``; used by the index loop."""
+
+    async def _load_indexed_nodes(self) -> Iterable[FileNode]:
+        if self.file_store is None:
+            raise RuntimeError("file_store is not initialized!")
+        return await self.file_store.get_nodes()
+
+
+@R.register("scan_catalog_changes_step")
+class ScanCatalogChangesStep(_ScanChangesBase):
+    """Diff vault against ``file_catalog``; used by the dream loop."""
+
+    file_catalog: BaseFileCatalog = Ref(BaseFileCatalog, ComponentEnum.FILE_CATALOG)
+
+    async def _load_indexed_nodes(self) -> Iterable[FileNode]:
+        return await self.file_catalog.get_nodes()
diff --git a/reme4/steps/index/traverse.py b/reme4/steps/index/traverse.py
index ac5f1ed4..f55ae597 100644
--- a/reme4/steps/index/traverse.py
+++ b/reme4/steps/index/traverse.py
@@ -104,8 +104,26 @@ class TraverseStep(BaseStep):
         outbound, inbound = await _build_adjacency(self.file_store)
         results = _bfs(seeds, depth, direction, outbound, inbound)
 
+        self.logger.info(
+            f"[{self.name}] seeds={seeds!r} depth={depth} direction={direction} "
+            f"nodes={len(outbound) + len(inbound)} edges={len(results)}",
+        )
+
         label = seeds[0] if len(seeds) == 1 else f"{len(seeds)} seeds"
+        if not results:
+            answer = f"No edges found from {label}"
+        else:
+            header = f"Traversed {len(results)} edge(s) from {label}"
+            lines = [header, ""]
+            for r in results:
+                target = r["path"]
+                if r["anchor"]:
+                    target = f"{target}#{r['anchor']}"
+                predicate = r["predicate"] or "-"
+                lines.append(f"[depth={r['depth']}] {r['via']} --{predicate}--> {target}")
+            answer = "\n".join(lines)
+
         self.context.response.success = True
-        self.context.response.answer = f"Traversed {len(results)} edge(s) from {label}"
+        self.context.response.answer = answer
         self.context.response.metadata.update({"edges": results, "count": len(results)})
         return self.context.response
diff --git a/reme4/steps/index/watch_changes.py b/reme4/steps/index/watch_changes.py
index dfb978a2..5f88401c 100644
--- a/reme4/steps/index/watch_changes.py
+++ b/reme4/steps/index/watch_changes.py
@@ -1,4 +1,27 @@
-"""Long-running awatch loop: convert raw changes into update_index calls."""
+"""Long-running awatch loop: convert raw changes into dispatch_step calls.
+
+Two relevant awatch parameters are exposed verbatim:
+
+* ``step`` (default ``50ms``) — awatch yields when the entire watcher
+  has gone this long without new changes (and at least one change is
+  pending). Raise to ``5 minutes``-ish for ``auto_dream_loop`` so
+  half-written sync output isn't dreamed mid-write; keep at default
+  for ``update_store_index_loop`` where every fs change should hit
+  the index promptly.
+
+* ``debounce`` (default ``2000ms``) — per-batch ceiling, regardless
+  of whether activity is still arriving. Set ``debounce > step`` so
+  ``step`` is the operative limit; otherwise the watcher pre-empts
+  long-quiet-window setups under bursty writes.
+
+Both are global to the watcher (not per-path). The two reme watchers
+have disjoint ``watch_paths`` (digest vs daily/resource), so global
+quiet windows are good enough — no per-path bookkeeping needed.
+
+awatch internally deduplicates same-path same-change tuples within
+the yielded batch, so a file ``modified`` ten times during the
+quiet window arrives as one ``(modified, path)`` entry.
+"""
 
 import asyncio
 
@@ -11,13 +34,14 @@ from ...enumeration import ComponentEnum
 
 @R.register("watch_changes_step")
 class WatchChangesStep(BaseStep):
-    """Watch files and forward each batch of raw changes to a downstream step."""
+    """Watch files and forward each yielded batch to a downstream step."""
 
     def __init__(
         self,
         recursive: bool = True,
         force_polling: bool = True,
         debounce: int = 2000,
+        step: int = 50,
         poll_delay_ms: int = 2000,
         dispatch_step: str = "",
         **kwargs,
@@ -26,6 +50,7 @@ class WatchChangesStep(BaseStep):
         self.recursive: bool = recursive
         self.force_polling: bool = force_polling
         self.debounce: int = debounce
+        self.step: int = step
         self.poll_delay_ms: int = poll_delay_ms
         self.dispatch_step: str = dispatch_step
 
@@ -52,13 +77,17 @@ class WatchChangesStep(BaseStep):
             if dispatch_step_cls is None:
                 raise RuntimeError(f"Unregistered step '{self.dispatch_step}'")
 
-        self.logger.info(f"Watching: {[str(p) for p in valid_paths]}")
+        self.logger.info(
+            f"Watching: {[str(p) for p in valid_paths]} step={self.step}ms debounce={self.debounce}ms",
+        )
+
         async for raw_changes in awatch(
             *valid_paths,
             watch_filter=self._filter,
             recursive=self.recursive,
             force_polling=self.force_polling,
             debounce=self.debounce,
+            step=self.step,
             poll_delay_ms=self.poll_delay_ms,
             stop_event=stop_event,
         ):
diff --git a/reme4/steps/transfer/download.py b/reme4/steps/transfer/download.py
index 8dca0732..99db6420 100644
--- a/reme4/steps/transfer/download.py
+++ b/reme4/steps/transfer/download.py
@@ -50,9 +50,14 @@ class DownloadStep(BaseStep):
         if "error" in payload:
             self.context.response.success = False
             self.context.response.answer = f"Error: {payload['error']}"
+            self.logger.info(f"[{self.name}] download failed src={src_path} error={payload['error']!r}")
         else:
             self.context.response.success = True
             self.context.response.answer = f"Downloaded {src_path} → {payload['dst_path']} ({payload['size']} bytes)"
+            self.logger.info(
+                f"[{self.name}] src={src_path} dst={payload['dst_path']} "
+                f"size={payload['size']} mime={payload['mime']}",
+            )
         self.context.response.metadata.update(payload)
 
     async def _download(self, src_path: str, dst_path: str, overwrite: bool) -> dict:
diff --git a/reme4/steps/transfer/ingest.py b/reme4/steps/transfer/ingest.py
index f9f36b52..db638fdf 100644
--- a/reme4/steps/transfer/ingest.py
+++ b/reme4/steps/transfer/ingest.py
@@ -115,6 +115,7 @@ class IngestStep(BaseStep):
         prepared, prep_error = _prepare_inputs(path, channel, description, metadata_raw)
         if prep_error:
             self._fail({"error": prep_error})
+            self.logger.info(f"[{self.name}] ingest failed channel={channel!r} error={prep_error!r}")
             return
 
         try:
@@ -131,14 +132,19 @@ class IngestStep(BaseStep):
             )
         except _DuplicateIngest as e:
             self._fail({"error": str(e)})
+            self.logger.info(f"[{self.name}] ingest duplicate channel={channel!r} error={str(e)!r}")
             return
         except Exception as e:
             self._fail({"error": f"{type(e).__name__}: {e}"})
+            self.logger.info(f"[{self.name}] ingest crashed channel={channel!r} error={type(e).__name__}: {e}")
             return
 
         self.context.response.success = True
         self.context.response.answer = f"Ingested {outcome['name']} to {outcome['path']}"
         self.context.response.metadata.update(outcome)
+        self.logger.info(
+            f"[{self.name}] channel={channel} date={outcome['date']} name={outcome['name']} path={outcome['path']}",
+        )
 
     # ------------------------------------------------------------------
 
diff --git a/reme4/steps/transfer/upload.py b/reme4/steps/transfer/upload.py
index cbcffb32..a60fd367 100644
--- a/reme4/steps/transfer/upload.py
+++ b/reme4/steps/transfer/upload.py
@@ -40,9 +40,15 @@ class UploadStep(BaseStep):
         if "error" in payload:
             self.context.response.success = False
             self.context.response.answer = f"Error: {payload['error']}"
+            self.logger.info(
+                f"[{self.name}] upload failed src={src_path} dst={dst_path} error={payload['error']!r}",
+            )
         else:
             self.context.response.success = True
             self.context.response.answer = f"Uploaded {src_path} → {dst_path} ({payload['size']} bytes)"
+            self.logger.info(
+                f"[{self.name}] src={src_path} dst={dst_path} size={payload['size']} mime={payload['mime']}",
+            )
         self.context.response.metadata.update(payload)
 
     async def _upload(self, src_path: str, dst_path: str, overwrite: bool) -> dict:
diff --git a/reme4/utils/service_utils.py b/reme4/utils/service_utils.py
index 0bc29c78..35a077cd 100644
--- a/reme4/utils/service_utils.py
+++ b/reme4/utils/service_utils.py
@@ -73,6 +73,7 @@ def precheck_start(svc_config: dict | None) -> bool:
     """Pre-flight check for `start`: False if reme is up, exits 1 on port conflict."""
     host = (svc_config or {}).get("host") or REME_DEFAULT_HOST
     port = (svc_config or {}).get("port") or REME_DEFAULT_PORT
+    port = int(port)
     status = asyncio.run(find_reme(host, port))
     if status == "reme":
         print(f"reme already running at {host}:{port}")
diff --git a/reme4/utils/transcript.py b/reme4/utils/transcript.py
new file mode 100644
index 00000000..cbc96ebf
--- /dev/null
+++ b/reme4/utils/transcript.py
@@ -0,0 +1,195 @@
+"""Parse Claude Code transcript JSONL into a plain message slice.
+
+Both routes (``reme-driver`` external, ``reme-service`` internal) accept
+``transcript_path`` as the canonical ``sync`` input — Claude Code hooks
+hand us the path, not the message list. Centralising the parse here
+keeps the two routes producing identical slices for the same transcript.
+
+The JSONL format is one record per line. The records we care about:
+
+    {"type": "user",      "message": {"role": "user",      "content": str | list[block]}, ...}
+    {"type": "assistant", "message": {"role": "assistant", "content": list[block]}, ...}
+
+Everything else (``ai-title`` / ``mode`` / ``permission-mode`` /
+``file-history-snapshot`` / ``attachment`` / ``last-prompt`` /
+``queue-operation`` / ``system``) is metadata or harness chatter and is
+ignored.
+
+Content blocks we recognise (shape from the Anthropic message format):
+
+* ``{"type": "text", "text": str}``                    — appended verbatim
+* ``{"type": "tool_use", "name": str, "input": ...}``  — rendered as ``[tool <name>(<json excerpt>)]``
+* ``{"type": "tool_result", "content": ...}``          — rendered as ``[tool_result <excerpt>]``
+* ``{"type": "thinking", "thinking": str}``            — dropped (private reasoning)
+
+User content frequently contains Claude-Code-injected boilerplate that
+isn't part of the real conversation:
+
+* ``<local-command-caveat>...``  — `bash` command warnings prepended to first user turn
+* ``<local-command-stdout>...``  — output of slash commands
+* ``<command-name>...``          — slash command label
+* ``<command-message>...``       — slash command description
+* ``<system-reminder>...``       — periodic harness reminders
+
+These are filtered out (whole-message drop if the text is only injected
+markers, partial strip otherwise) so the synchronizer sees the actual
+user/assistant dialogue.
+"""
+
+from __future__ import annotations
+
+import json
+import re
+from pathlib import Path
+from typing import Any
+
+# Whole-message-drop when the user content is only one of these wrappers.
+_INJECTED_TAGS = (
+    "<local-command-caveat>",
+    "<local-command-stdout>",
+    "<local-command-stderr>",
+    "<command-name>",
+    "<command-message>",
+    "<command-args>",
+    "<system-reminder>",
+    "<bash-input>",
+    "<bash-stdout>",
+    "<bash-stderr>",
+)
+
+# Heuristic: if the message starts with one of these AND is short / mostly markup,
+# drop it. We keep the regex permissive — false negatives (a real message that
+# looks like markup) are recoverable downstream; false positives (dropping real
+# user text) are silent and worse.
+_DROP_IF_STARTS_WITH = tuple(_INJECTED_TAGS)
+
+
+def load_messages_from_transcript(
+    transcript_path: str | Path,
+    *,
+    include_thinking: bool = False,
+    tool_input_excerpt: int = 200,
+) -> list[dict[str, str]]:
+    """Read a Claude Code transcript JSONL → list of role/content dicts.
+
+    Returns ``[{role, name, content}, ...]`` in source order, where
+    ``role`` is ``"user"`` or ``"assistant"`` and ``name`` mirrors role
+    (so the dicts are directly consumable by AgentScope's ``Msg``, which
+    requires a ``name`` field). Empty list when the file is missing,
+    empty, or contains no user/assistant turns.
+
+    Parameters
+    ----------
+    transcript_path:
+        Absolute or relative path to the transcript JSONL file.
+    include_thinking:
+        When True, include ``thinking`` blocks (assistant private reasoning).
+        Default False — the synchronizer wants observable dialogue.
+    tool_input_excerpt:
+        Max chars of a ``tool_use`` input JSON to render inline. Default 200.
+    """
+    path = Path(transcript_path)
+    if not path.is_file():
+        return []
+
+    messages: list[dict[str, str]] = []
+    for line in path.read_text(encoding="utf-8").splitlines():
+        line = line.strip()
+        if not line:
+            continue
+        try:
+            record = json.loads(line)
+        except json.JSONDecodeError:
+            continue
+
+        record_type = record.get("type")
+        if record_type not in ("user", "assistant"):
+            continue
+
+        message = record.get("message") or {}
+        role = message.get("role")
+        if role not in ("user", "assistant"):
+            continue
+
+        text = _render_content(
+            message.get("content", ""),
+            include_thinking=include_thinking,
+            tool_input_excerpt=tool_input_excerpt,
+        )
+        if not text:
+            continue
+        if _is_injected_only(text):
+            continue
+
+        messages.append({"role": role, "name": role, "content": text})
+
+    return messages
+
+
+def _render_content(
+    content: Any,
+    *,
+    include_thinking: bool,
+    tool_input_excerpt: int,
+) -> str:
+    if isinstance(content, str):
+        return content.strip()
+
+    if not isinstance(content, list):
+        return ""
+
+    parts: list[str] = []
+    for block in content:
+        if not isinstance(block, dict):
+            continue
+        btype = block.get("type")
+
+        if btype == "text":
+            t = (block.get("text") or "").strip()
+            if t:
+                parts.append(t)
+
+        elif btype == "thinking":
+            if include_thinking:
+                t = (block.get("thinking") or "").strip()
+                if t:
+                    parts.append(f"[thinking]\n{t}")
+
+        elif btype == "tool_use":
+            name = block.get("name", "?")
+            try:
+                inp = json.dumps(block.get("input"), ensure_ascii=False)[:tool_input_excerpt]
+            except (TypeError, ValueError):
+                inp = str(block.get("input"))[:tool_input_excerpt]
+            parts.append(f"[tool {name}({inp})]")
+
+        elif btype == "tool_result":
+            inner = block.get("content")
+            if isinstance(inner, list):
+                excerpt = _render_content(
+                    inner,
+                    include_thinking=False,
+                    tool_input_excerpt=tool_input_excerpt,
+                )
+            else:
+                excerpt = str(inner or "")
+            excerpt = excerpt.strip()
+            if len(excerpt) > tool_input_excerpt:
+                excerpt = excerpt[:tool_input_excerpt] + "..."
+            parts.append(f"[tool_result {excerpt}]")
+
+    return "\n".join(p for p in parts if p).strip()
+
+
+def _is_injected_only(text: str) -> bool:
+    """True if the text is composed entirely of Claude-Code-injected markers
+    (no genuine user/assistant dialogue around them).
+    """
+    stripped = text.strip()
+    if not stripped.startswith(_DROP_IF_STARTS_WITH):
+        return False
+    # If it starts with an injected tag, peek whether anything substantive
+    # follows the closing tag. Cheap heuristic: strip all wrapped <tag>...</tag>
+    # blocks and see what's left.
+    remaining = re.sub(r"<([a-z-]+)>.*?</\1>", "", stripped, flags=re.DOTALL)
+    return len(remaining.strip()) < 16  # arbitrary "essentially empty" threshold
diff --git a/tests4/integration/test_dreamer_cli.sh b/tests4/integration/test_dreamer_cli.sh
deleted file mode 100755
index 9072dea3..00000000
--- a/tests4/integration/test_dreamer_cli.sh
+++ /dev/null
@@ -1,68 +0,0 @@
-#!/usr/bin/env bash
-# dreamer CLI integration test (option B).
-#
-# Seeds a rich workspace via _dreamer_fixture.py, starts `reme start`
-# bound to that vault, reindexes so Phase 2 recall can hit the pre-
-# seeded digest nodes, then calls `reme dream`.
-#
-# Usage (from anywhere):
-#   VAULT_PATH=/tmp/reme-dreamer-test bash tests4/integration/test_dreamer_cli.sh
-#   VAULT_PATH=/tmp/reme-dreamer-test bash tests4/integration/test_dreamer_cli.sh daily/2026-05-28/auth-refactor/notes.md
-#
-# Defaults:
-#   VAULT_PATH unset → /tmp/reme-dreamer-test
-#   Workspace seeded on first run (idempotent).
-#
-# Required env (from .env or shell):
-#   LLM_API_KEY, LLM_BASE_URL, LLM_MODEL_NAME
-set -euo pipefail
-
-VAULT="${VAULT_PATH:-/tmp/reme-dreamer-test}"
-INTEGRATION_DIR="$(cd "$(dirname "$0")" && pwd)"
-REPO="$(cd "$INTEGRATION_DIR/../.." && pwd)"
-LOG="/tmp/test_dreamer_cli_server.log"
-
-# Resolve input from arg, else default to fixture's input path.
-DEFAULT_INPUT="$(python -c "import sys; sys.path.insert(0, '$INTEGRATION_DIR'); from _dreamer_fixture import INPUT_PATH; print(INPUT_PATH)")"
-INPUT="${1:-$DEFAULT_INPUT}"
-
-mkdir -p "$VAULT"
-echo "--- seeding fixture under $VAULT"
-python "$INTEGRATION_DIR/_dreamer_fixture.py" "$VAULT"
-
-cd "$REPO"
-
-echo ""
-echo "--- starting reme server (log: $LOG)"
-reme start "vault_dir=$VAULT" >"$LOG" 2>&1 &
-SERVER_PID=$!
-trap 'echo "--- stopping reme server (pid $SERVER_PID)"; kill "$SERVER_PID" 2>/dev/null || true; wait "$SERVER_PID" 2>/dev/null || true' EXIT
-
-echo "--- waiting for server"
-for _ in $(seq 1 30); do
-  if curl -s -o /dev/null http://localhost:8000/docs 2>/dev/null; then
-    echo "--- server up"
-    break
-  fi
-  sleep 0.5
-done
-
-echo ""
-echo "--- reindexing vault so Phase 2 recall has something to hit"
-reme reindex
-
-echo ""
-echo "=== reme dream path=$INPUT ==="
-reme dream "path=$INPUT"
-echo ""
-
-echo "=== digest/ tree after dream ==="
-if [ -d "$VAULT/digest" ]; then
-  find "$VAULT/digest" -name "*.md" | sort | while read -r f; do
-    echo ""
-    echo "--- ${f#$VAULT/} ---"
-    cat "$f"
-  done
-else
-  echo "  (no digest/ created)"
-fi
diff --git a/tests4/integration/test_dreamer_inproc.py b/tests4/integration/test_dreamer_inproc.py
index 03cac9a1..a5102efa 100644
--- a/tests4/integration/test_dreamer_inproc.py
+++ b/tests4/integration/test_dreamer_inproc.py
@@ -11,12 +11,12 @@ wiki}; Phase 2 dispatches to the bucket-specific integrate prompt
 and writes via the canonical `write` / `edit` tools.
 
 Usage (from anywhere):
-    VAULT_PATH=/tmp/reme-dreamer-test python tests4/integration/test_dreamer_inproc.py
-    VAULT_PATH=/tmp/reme-dreamer-test python tests4/integration/test_dreamer_inproc.py \\
+    VAULT_PATH=tests4/integration/vault python tests4/integration/test_dreamer_inproc.py
+    VAULT_PATH=tests4/integration/vault python tests4/integration/test_dreamer_inproc.py \\
         daily/2026-05-28/auth-refactor/notes.md
 
 Defaults:
-    VAULT_PATH unset → /tmp/reme-dreamer-test
+    VAULT_PATH unset → tests4/integration/vault
     Each run wipes `daily/`, `digest/`, and `reme_metadata/` under the
     vault before reseeding, so the dreamer always starts from the same
     fixture state. See _dreamer_fixture.py for what gets created and
@@ -27,10 +27,13 @@ Required env (from .env or shell):
 """
 
 import asyncio
+import json
 import os
 import sys
 from pathlib import Path
 
+from agentscope.agent import Agent
+
 # Make `reme4` importable regardless of the caller's cwd; and make the
 # fixture module importable as a top-level name.
 REPO_ROOT = Path(__file__).resolve().parents[2]
@@ -41,7 +44,56 @@ sys.path.insert(0, str(INTEGRATION_DIR))
 # pylint: disable=wrong-import-position
 from _dreamer_fixture import clean_vault, seed_vault, INPUT_PATH  # noqa: E402
 
-VAULT = os.environ.get("VAULT_PATH", "/tmp/reme-dreamer-test")
+VAULT = os.environ.get("VAULT_PATH", "tests4/integration/vault")
+
+
+class _AgentMemoryRecorder:
+    """Monkey-patches Agent.__init__ to capture every agent created inside
+    the ``with`` block, then dumps each agent's context history to a jsonl
+    file under ``<vault>/agent_logs/`` on dump().
+
+    Used to inspect the actual ReAct trace of Phase 1 extract + Phase 2
+    integrate (per sub-unit) — what tools were called in what order, what
+    candidates were recalled, what the LLM decided.
+    """
+
+    def __init__(self, vault: Path, prefix: str = "dream"):
+        self.dump_dir = vault / "agent_logs"
+        self.prefix = prefix
+        self.agents: list[Agent] = []
+        self._orig_init = None
+        self.dumped_paths: list[Path] = []
+
+    def __enter__(self):
+        self._orig_init = Agent.__init__
+        agents = self.agents
+        orig = self._orig_init
+
+        def _capturing_init(agent_self, *args, **kwargs):
+            orig(agent_self, *args, **kwargs)
+            agents.append(agent_self)
+
+        Agent.__init__ = _capturing_init
+        return self
+
+    def __exit__(self, *exc):
+        Agent.__init__ = self._orig_init
+
+    async def dump(self) -> list[Path]:
+        """Dump all captured agents' context to <vault>/agent_logs/."""
+        self.dump_dir.mkdir(parents=True, exist_ok=True)
+        for stale in self.dump_dir.glob(f"{self.prefix}_*.jsonl"):
+            stale.unlink()
+
+        for idx, agent in enumerate(self.agents, 1):
+            messages = agent.state.context
+            name = getattr(agent, "name", "agent") or "agent"
+            out_path = self.dump_dir / f"{self.prefix}_{idx:02d}_{name}.jsonl"
+            with out_path.open("w", encoding="utf-8") as f:
+                for msg in messages:
+                    f.write(json.dumps(msg.model_dump(), ensure_ascii=False, default=str) + "\n")
+            self.dumped_paths.append(out_path)
+        return self.dumped_paths
 
 
 async def main() -> None:
@@ -78,7 +130,12 @@ async def main() -> None:
         await app.run_job("reindex")
 
         print(f"\n--- running dream path={rel_input}")
-        resp = await app.run_job("dream", path=rel_input)
+        with _AgentMemoryRecorder(vault, prefix="dream") as recorder:
+            resp = await app.run_job("dream", path=rel_input)
+        dumped = await recorder.dump()
+        print(f"\n--- dumped {len(dumped)} agent memory file(s) to {recorder.dump_dir}")
+        for p in dumped:
+            print(f"  {p.relative_to(vault)}")
 
         print("\n=== Response.success ===")
         print(resp.success)
diff --git a/tests4/unit/test_auto_dream.py b/tests4/unit/test_auto_dream.py
new file mode 100644
index 00000000..80e8398f
--- /dev/null
+++ b/tests4/unit/test_auto_dream.py
@@ -0,0 +1,360 @@
+"""Tests for AutoDreamStep — daily-tick + file_catalog dedup.
+
+AutoDreamStep walks ``daily/<today>.md`` + ``daily/<today>/**`` and
+diffs the result against ``file_catalog`` (same shape as
+``scan_catalog_changes_step``):
+
+* on-disk path missing from catalog          → dream
+* on-disk mtime != catalog mtime             → dream (modified)
+* on-disk mtime == catalog mtime             → skip (unchanged)
+* catalog entry under today's prefix missing → drop from catalog (deleted)
+
+Successful dreams (and Phase 1 vacuous skips) upsert the current
+``st_mtime`` so the next tick re-dreams only what actually changed.
+Failures leave the catalog untouched.
+
+We mock ``dream_one`` (needs an LLM) and inject a fake ``file_catalog``
+recording every get / upsert / delete / dump.
+"""
+
+# pylint: disable=protected-access
+
+import asyncio
+import os
+import tempfile
+import warnings
+from pathlib import Path
+from unittest.mock import AsyncMock, MagicMock, patch
+
+from reme4.components.file_catalog import BaseFileCatalog
+from reme4.components.runtime_context import RuntimeContext
+from reme4.schema import FileNode
+from reme4.steps import AutoDreamStep
+from reme4.steps.evolve.dream import DreamResult
+
+warnings.filterwarnings("ignore", category=DeprecationWarning, module="jieba")
+warnings.filterwarnings("ignore", category=DeprecationWarning, module="pkg_resources")
+
+
+class temp_chdir:
+    """Context manager that temporarily ``chdir``s into a directory and restores cwd on exit."""
+
+    def __init__(self, path):
+        self.path = path
+        self.old = None
+
+    def __enter__(self):
+        self.old = os.getcwd()
+        os.chdir(self.path)
+        return self
+
+    def __exit__(self, *exc):
+        os.chdir(self.old)
+
+
+def _touch(path: Path, content: str = "x") -> Path:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(content, encoding="utf-8")
+    return path
+
+
+def _make_step(vault: Path, today: str, existing_nodes: list[FileNode] | None = None) -> AutoDreamStep:
+    """AutoDreamStep with vault forced, daily_dir='daily', and a fake catalog."""
+    fake_catalog = MagicMock(spec=BaseFileCatalog)
+    fake_catalog.get_nodes = AsyncMock(return_value=list(existing_nodes or []))
+    fake_catalog.upsert = AsyncMock()
+    fake_catalog.delete = AsyncMock()
+    fake_catalog.dump = AsyncMock()
+
+    class _Fixed(AutoDreamStep):
+        @property
+        def vault_path(self):
+            return vault
+
+        def _vault_dir(self):
+            return vault
+
+        def _now(self):
+            import datetime
+
+            return datetime.datetime.fromisoformat(f"{today}T00:00:00")
+
+    step = _Fixed(file_catalog=fake_catalog, persist=True)
+    cfg = MagicMock()
+    cfg.daily_dir = "daily"
+    cfg.resource_dir = ""
+    step.app_context = MagicMock()
+    step.app_context.app_config = cfg
+    return step
+
+
+def test_scans_date_md_and_date_folder():
+    """Both ``daily/<today>.md`` and files under ``daily/<today>/`` are picked up;
+    date.md is dreamed first so the day-index leads."""
+
+    async def run():
+        with tempfile.TemporaryDirectory() as tmpdir, temp_chdir(tmpdir):
+            vault = Path(tmpdir).resolve()
+            today = "2026-06-04"
+            _touch(vault / "daily" / f"{today}.md")
+            _touch(vault / "daily" / today / "session-a.md")
+            _touch(vault / "daily" / today / "session-b.md")
+            step = _make_step(vault, today)
+            ctx = RuntimeContext()
+
+            seen: list[str] = []
+
+            async def _fake_dream(rel, _hint):
+                seen.append(rel)
+                return DreamResult(used_llm=True, path=rel, summary="ok")
+
+            with patch.object(step, "dream_one", side_effect=_fake_dream):
+                resp = await step(ctx)
+
+            assert resp.success
+            assert resp.metadata["files_scanned"] == 3
+            assert resp.metadata["files_dreamed"] == 3
+            assert seen[0] == f"daily/{today}.md"
+            assert seen[1:] == [f"daily/{today}/session-a.md", f"daily/{today}/session-b.md"]
+        print("✓ test_scans_date_md_and_date_folder passed")
+
+    asyncio.run(run())
+
+
+def test_resource_dir_is_not_scanned():
+    """resource/<today>/ files are NOT picked up by AutoDreamStep anymore."""
+
+    async def run():
+        with tempfile.TemporaryDirectory() as tmpdir, temp_chdir(tmpdir):
+            vault = Path(tmpdir).resolve()
+            today = "2026-06-04"
+            _touch(vault / "daily" / f"{today}.md")
+            _touch(vault / "resource" / today / "spec.pdf")
+            step = _make_step(vault, today)
+            step.app_context.app_config.resource_dir = "resource"
+            ctx = RuntimeContext()
+
+            async def _fake_dream(rel, _hint):
+                return DreamResult(used_llm=True, path=rel, summary="ok")
+
+            with patch.object(step, "dream_one", side_effect=_fake_dream) as dream_mock:
+                await step(ctx)
+
+            paths = [c.args[0] for c in dream_mock.call_args_list]
+            assert paths == [f"daily/{today}.md"]
+        print("✓ test_resource_dir_is_not_scanned passed")
+
+    asyncio.run(run())
+
+
+def test_unchanged_files_skipped_via_catalog_mtime():
+    """A file whose catalog mtime matches on-disk mtime is NOT dreamed."""
+
+    async def run():
+        with tempfile.TemporaryDirectory() as tmpdir, temp_chdir(tmpdir):
+            vault = Path(tmpdir).resolve()
+            today = "2026-06-04"
+            note = _touch(vault / "daily" / today / "note.md")
+            mtime = note.stat().st_mtime
+            existing = [FileNode(path=f"daily/{today}/note.md", st_mtime=mtime)]
+            step = _make_step(vault, today, existing_nodes=existing)
+            ctx = RuntimeContext()
+
+            with patch.object(step, "dream_one") as dream_mock:
+                resp = await step(ctx)
+                dream_mock.assert_not_called()
+
+            assert resp.success
+            assert resp.metadata["files_unchanged"] == 1
+            assert resp.metadata["files_dreamed"] == 0
+            step.file_catalog.upsert.assert_not_awaited()
+            step.file_catalog.dump.assert_not_awaited()
+        print("✓ test_unchanged_files_skipped_via_catalog_mtime passed")
+
+    asyncio.run(run())
+
+
+def test_changed_file_dreamed_and_catalog_updated():
+    """Stale catalog mtime → file is dreamed + catalog upserts new mtime."""
+
+    async def run():
+        with tempfile.TemporaryDirectory() as tmpdir, temp_chdir(tmpdir):
+            vault = Path(tmpdir).resolve()
+            today = "2026-06-04"
+            note = _touch(vault / "daily" / today / "note.md")
+            mtime = note.stat().st_mtime
+            stale = mtime - 999.0
+            existing = [FileNode(path=f"daily/{today}/note.md", st_mtime=stale)]
+            step = _make_step(vault, today, existing_nodes=existing)
+            ctx = RuntimeContext()
+
+            async def _fake_dream(rel, _hint):
+                return DreamResult(used_llm=True, path=rel, summary="ok")
+
+            with patch.object(step, "dream_one", side_effect=_fake_dream):
+                resp = await step(ctx)
+
+            assert resp.success
+            assert resp.metadata["files_dreamed"] == 1
+            assert resp.metadata["files_unchanged"] == 0
+            step.file_catalog.upsert.assert_awaited_once()
+            (nodes,), _ = step.file_catalog.upsert.call_args
+            assert len(nodes) == 1
+            assert nodes[0].path == f"daily/{today}/note.md"
+            assert nodes[0].st_mtime == mtime  # post-dream mtime, not stale
+            step.file_catalog.dump.assert_awaited_once()
+        print("✓ test_changed_file_dreamed_and_catalog_updated passed")
+
+    asyncio.run(run())
+
+
+def test_deleted_file_dropped_from_catalog():
+    """Catalog entry under today's prefix with no on-disk file → catalog.delete."""
+
+    async def run():
+        with tempfile.TemporaryDirectory() as tmpdir, temp_chdir(tmpdir):
+            vault = Path(tmpdir).resolve()
+            today = "2026-06-04"
+            # No on-disk files for today; catalog has a stale entry for today.
+            existing = [FileNode(path=f"daily/{today}/gone.md", st_mtime=123.0)]
+            step = _make_step(vault, today, existing_nodes=existing)
+            ctx = RuntimeContext()
+
+            with patch.object(step, "dream_one") as dream_mock:
+                resp = await step(ctx)
+                dream_mock.assert_not_called()
+
+            assert resp.success
+            assert resp.metadata["files_deleted"] == 1
+            step.file_catalog.delete.assert_awaited_once_with([f"daily/{today}/gone.md"])
+            step.file_catalog.upsert.assert_not_awaited()
+            step.file_catalog.dump.assert_awaited_once()
+        print("✓ test_deleted_file_dropped_from_catalog passed")
+
+    asyncio.run(run())
+
+
+def test_other_days_catalog_entries_untouched():
+    """Catalog entries OUTSIDE today's prefix must not be dropped, even if
+    they don't appear in today's on-disk scan."""
+
+    async def run():
+        with tempfile.TemporaryDirectory() as tmpdir, temp_chdir(tmpdir):
+            vault = Path(tmpdir).resolve()
+            today = "2026-06-04"
+            yesterday = "2026-06-03"
+            # Today: nothing on disk. Catalog has yesterday's entry.
+            existing = [FileNode(path=f"daily/{yesterday}/note.md", st_mtime=99.0)]
+            step = _make_step(vault, today, existing_nodes=existing)
+            ctx = RuntimeContext()
+
+            resp = await step(ctx)
+
+            assert resp.success
+            assert resp.metadata["files_scanned"] == 0
+            assert resp.metadata["files_deleted"] == 0
+            step.file_catalog.delete.assert_not_awaited()
+            step.file_catalog.upsert.assert_not_awaited()
+        print("✓ test_other_days_catalog_entries_untouched passed")
+
+    asyncio.run(run())
+
+
+def test_failure_does_not_upsert():
+    """A dream error must leave the catalog untouched so the next tick retries."""
+
+    async def run():
+        with tempfile.TemporaryDirectory() as tmpdir, temp_chdir(tmpdir):
+            vault = Path(tmpdir).resolve()
+            today = "2026-06-04"
+            _touch(vault / "daily" / today / "note.md")
+            step = _make_step(vault, today)
+            ctx = RuntimeContext()
+
+            async def _fake_dream(rel, _hint):
+                return DreamResult(used_llm=False, path=rel, error="boom")
+
+            with patch.object(step, "dream_one", side_effect=_fake_dream):
+                resp = await step(ctx)
+
+            assert not resp.success
+            assert resp.metadata["files_failed"] == 1
+            step.file_catalog.upsert.assert_not_awaited()
+            step.file_catalog.dump.assert_not_awaited()
+        print("✓ test_failure_does_not_upsert passed")
+
+    asyncio.run(run())
+
+
+def test_phase1_empty_still_upserts():
+    """Phase 1 skipped (no abstractions) is still success — still catalogued so
+    the next tick doesn't redo Phase 1 unnecessarily."""
+
+    async def run():
+        with tempfile.TemporaryDirectory() as tmpdir, temp_chdir(tmpdir):
+            vault = Path(tmpdir).resolve()
+            today = "2026-06-04"
+            note = _touch(vault / "daily" / today / "note.md")
+            mtime = note.stat().st_mtime
+            step = _make_step(vault, today)
+            ctx = RuntimeContext()
+
+            async def _fake_dream(rel, _hint):
+                return DreamResult(used_llm=True, path=rel, skipped=True, summary="empty")
+
+            with patch.object(step, "dream_one", side_effect=_fake_dream):
+                resp = await step(ctx)
+
+            assert resp.success
+            assert resp.metadata["files_skipped"] == 1
+            step.file_catalog.upsert.assert_awaited_once()
+            (nodes,), _ = step.file_catalog.upsert.call_args
+            assert nodes[0].st_mtime == mtime
+        print("✓ test_phase1_empty_still_upserts passed")
+
+    asyncio.run(run())
+
+
+def test_partial_failure_does_not_block_other_files():
+    """One file's failure must not stop other files from being dreamed + catalogued."""
+
+    async def run():
+        with tempfile.TemporaryDirectory() as tmpdir, temp_chdir(tmpdir):
+            vault = Path(tmpdir).resolve()
+            today = "2026-06-04"
+            _touch(vault / "daily" / today / "a.md")
+            _touch(vault / "daily" / today / "b.md")
+            step = _make_step(vault, today)
+            ctx = RuntimeContext()
+
+            async def _fake_dream(rel, _hint):
+                if rel.endswith("a.md"):
+                    return DreamResult(used_llm=False, path=rel, error="boom")
+                return DreamResult(used_llm=True, path=rel, summary="ok")
+
+            with patch.object(step, "dream_one", side_effect=_fake_dream):
+                resp = await step(ctx)
+
+            assert not resp.success
+            assert resp.metadata["files_dreamed"] == 1
+            assert resp.metadata["files_failed"] == 1
+            step.file_catalog.upsert.assert_awaited_once()
+            (nodes,), _ = step.file_catalog.upsert.call_args
+            assert [n.path for n in nodes] == [f"daily/{today}/b.md"]
+        print("✓ test_partial_failure_does_not_block_other_files passed")
+
+    asyncio.run(run())
+
+
+if __name__ == "__main__":
+    print("\n=== AutoDreamStep Tests ===")
+    test_scans_date_md_and_date_folder()
+    test_resource_dir_is_not_scanned()
+    test_unchanged_files_skipped_via_catalog_mtime()
+    test_changed_file_dreamed_and_catalog_updated()
+    test_deleted_file_dropped_from_catalog()
+    test_other_days_catalog_entries_untouched()
+    test_failure_does_not_upsert()
+    test_phase1_empty_still_upserts()
+    test_partial_failure_does_not_block_other_files()
+    print("\n所有测试通过!")
diff --git a/tests4/unit/test_background_steps.py b/tests4/unit/test_background_steps.py
index b3cd50bd..afc83dff 100644
--- a/tests4/unit/test_background_steps.py
+++ b/tests4/unit/test_background_steps.py
@@ -1,11 +1,15 @@
-"""Tests for background steps: ScanChangesStep + WatchChangesStep.
+"""Tests for background steps: ScanStoreChangesStep + WatchChangesStep.
 
 Both steps are subclasses of BaseStep. To exercise them without spinning up the
 full ApplicationContext, we pass real (started) file_store/file_parser via the
 step's kwargs (so the BaseStep _resolve() machinery returns them).
 
-ScanChangesStep writes its result into ``context["changes"]`` for a downstream
+ScanStoreChangesStep writes its result into ``context["changes"]`` for a downstream
 ``update_index_step`` to consume; tests assert against that key directly.
+
+The catalog-side sibling (ScanCatalogChangesStep) shares the same diff helper
+and is exercised through the dream-loop integration tests; covering it here
+would duplicate the file_store-diff assertions without adding signal.
 """
 
 # pylint: disable=protected-access
@@ -21,7 +25,7 @@ from watchfiles import Change
 from reme4.components.file_parser import ChunkedFileParser
 from reme4.components.file_store import LocalFileStore
 from reme4.components.runtime_context import RuntimeContext
-from reme4.steps import ScanChangesStep, WatchChangesStep
+from reme4.steps import ScanStoreChangesStep, WatchChangesStep
 
 warnings.filterwarnings("ignore", category=DeprecationWarning, module="jieba")
 warnings.filterwarnings("ignore", category=DeprecationWarning, module="pkg_resources")
@@ -51,7 +55,7 @@ def write_file(path: Path, content: str = "x") -> Path:
 
 
 # ---------------------------------------------------------------------------
-# ScanChangesStep
+# ScanStoreChangesStep
 # ---------------------------------------------------------------------------
 
 
@@ -59,12 +63,12 @@ async def _make_scan_step(
     watch_paths: list[str] | str = "vault",
     suffix_filters: list[str] | None = None,
     recursive: bool = True,
-) -> tuple[ScanChangesStep, RuntimeContext, LocalFileStore, ChunkedFileParser]:
+) -> tuple[ScanStoreChangesStep, RuntimeContext, LocalFileStore, ChunkedFileParser]:
     fs = LocalFileStore(name="test_store", embedding_store="")
     parser = ChunkedFileParser()
     await fs.start()
     await parser.start()
-    step = ScanChangesStep(
+    step = ScanStoreChangesStep(
         recursive=recursive,
         file_store=fs,
         file_parser=parser,
@@ -246,7 +250,7 @@ def test_watch_changes_filter_only_passes_md():
 
 if __name__ == "__main__":
     print("\n=== Background Steps Tests ===")
-    # ScanChangesStep
+    # ScanStoreChangesStep
     test_scan_changes_initial_all_added()
     test_scan_changes_no_changes_emits_empty_list()
     test_scan_changes_detects_modify_and_delete()
diff --git a/tests4/unit/test_channel_notify.py b/tests4/unit/test_channel_notify.py
new file mode 100644
index 00000000..c8a8e689
--- /dev/null
+++ b/tests4/unit/test_channel_notify.py
@@ -0,0 +1,104 @@
+"""Tests for ``ChannelNotifyStep`` — vault-watcher batch → channel event."""
+
+import asyncio
+from pathlib import Path
+
+from reme4.components.application_context import ApplicationContext
+from reme4.components.service.mcp_service import ChannelSink
+from reme4.components.runtime_context import RuntimeContext
+from reme4.steps.index.channel_notify import ChannelNotifyStep
+
+
+class _StubSession:
+    """Capture ``send_message`` payloads instead of writing them to a transport."""
+
+    def __init__(self) -> None:
+        self.sent: list = []
+
+    async def send_message(self, message) -> None:
+        """Record the outbound ``SessionMessage`` for later assertions."""
+        self.sent.append(message)
+
+
+def _run(coro):
+    """Drive a coroutine on a fresh event loop (tests don't share one)."""
+    return asyncio.new_event_loop().run_until_complete(coro)
+
+
+def _ctx(changes: list[dict]) -> RuntimeContext:
+    """Build a ``RuntimeContext`` pre-populated with the step's ``changes`` input."""
+    ctx = RuntimeContext()
+    ctx["changes"] = changes
+    return ctx
+
+
+def _app_ctx_with_sink(vault: Path, stub: _StubSession | None) -> tuple[ApplicationContext, ChannelSink | None]:
+    """Build an ``ApplicationContext`` rooted at ``vault``; attach a sink bound to ``stub`` if given."""
+    app_ctx = ApplicationContext(vault_dir=str(vault), app_name="reme-test")
+    if stub is None:
+        return app_ctx, None
+    sink = ChannelSink()
+    sink.bind(stub)
+    app_ctx.metadata["channel_sink"] = sink
+    return app_ctx, sink
+
+
+def test_emits_one_event_per_batch_with_relative_paths(tmp_path):
+    """A batch of changes → exactly one notification with relative paths and a count meta."""
+    vault = tmp_path
+    (vault / "resource" / "2026-06-03").mkdir(parents=True)
+    f1 = vault / "resource" / "2026-06-03" / "a.md"
+    f1.write_text("x")
+
+    stub = _StubSession()
+    app_ctx, _ = _app_ctx_with_sink(vault, stub)
+
+    step = ChannelNotifyStep(app_context=app_ctx)
+    _run(
+        step(
+            context=_ctx(
+                [
+                    {"change": "added", "path": str(f1)},
+                    {"change": "modified", "path": str(f1)},
+                ],
+            ),
+        ),
+    )
+
+    assert len(stub.sent) == 1
+    params = stub.sent[0].message.root.params
+    assert params["meta"] == {"kind": "vault_change", "count": "2"}
+    assert "added: resource/2026-06-03/a.md" in params["content"]
+    assert "modified: resource/2026-06-03/a.md" in params["content"]
+
+
+def test_noop_when_no_changes(tmp_path):
+    """An empty changes list must not produce any notification."""
+    stub = _StubSession()
+    app_ctx, _ = _app_ctx_with_sink(tmp_path, stub)
+    step = ChannelNotifyStep(app_context=app_ctx)
+    _run(step(context=_ctx([])))
+    assert not stub.sent
+
+
+def test_noop_when_sink_not_bound(tmp_path):
+    """Step must run cleanly when no ``ChannelSink`` is configured."""
+    app_ctx, _ = _app_ctx_with_sink(tmp_path, None)
+    step = ChannelNotifyStep(app_context=app_ctx)
+    # Should run without raising even though channel_sink is absent from metadata
+    _run(step(context=_ctx([{"change": "added", "path": "/tmp/x.md"}])))
+
+
+def test_path_outside_vault_passes_through_as_is(tmp_path):
+    """Paths not under the vault are emitted verbatim instead of crashing."""
+    stub = _StubSession()
+    app_ctx, _ = _app_ctx_with_sink(tmp_path, stub)
+    step = ChannelNotifyStep(app_context=app_ctx)
+    _run(
+        step(
+            context=_ctx([{"change": "added", "path": "/elsewhere/wild.md"}]),
+        ),
+    )
+    # Stray absolute path → emitted verbatim, no crash, still one event
+    assert len(stub.sent) == 1
+    assert "added: /elsewhere/wild.md" in stub.sent[0].message.root.params["content"]
diff --git a/tests4/unit/test_channel_sink.py b/tests4/unit/test_channel_sink.py
new file mode 100644
index 00000000..8b8c89e2
--- /dev/null
+++ b/tests4/unit/test_channel_sink.py
@@ -0,0 +1,100 @@
+"""Tests for ``ChannelSink`` — outbound ``notifications/claude/channel`` plumbing.
+
+Strategy: stub a session-like object with an async ``send_message`` capture
+list and exercise three paths:
+
+* not bound → emit is a no-op (no exception, no captured message)
+* bound + valid meta → captured JSON-RPC notification carries our method
+  + content + the meta we passed
+* bound + meta with non-identifier keys → the bad keys are dropped, the
+  rest passes through verbatim
+"""
+
+import asyncio
+
+from reme4.components.service.mcp_service import ChannelSink
+
+
+class _StubSession:
+    """Capture ``send_message`` payloads instead of writing them to a transport."""
+
+    def __init__(self) -> None:
+        self.sent: list = []
+
+    async def send_message(self, message) -> None:
+        """Record the outbound ``SessionMessage`` for later assertions."""
+        self.sent.append(message)
+
+
+def _run(coro):
+    """Drive a coroutine on a fresh event loop (tests don't share one)."""
+    return asyncio.new_event_loop().run_until_complete(coro)
+
+
+def test_emit_without_bind_is_noop():
+    """Emitting before any session is bound must silently no-op."""
+    sink = ChannelSink()
+    _run(sink.emit("hello", {"k": "v"}))  # must not raise
+
+
+def test_emit_after_bind_sends_channel_notification():
+    """A bound session receives a JSON-RPC notification carrying content + meta verbatim."""
+    sink = ChannelSink()
+    stub = _StubSession()
+    sink.bind(stub)
+
+    _run(sink.emit("ingest done", {"path": "resource/2026-06-03/x.md", "kind": "ingest"}))
+
+    assert len(stub.sent) == 1
+    payload = stub.sent[0].message.root
+    assert payload.method == "notifications/claude/channel"
+    assert payload.params["content"] == "ingest done"
+    assert payload.params["meta"] == {"path": "resource/2026-06-03/x.md", "kind": "ingest"}
+
+
+def test_emit_filters_non_identifier_meta_keys():
+    """Meta keys that aren't pure ``[A-Za-z0-9_]`` identifiers are dropped before send."""
+    sink = ChannelSink()
+    stub = _StubSession()
+    sink.bind(stub)
+
+    _run(
+        sink.emit(
+            "x",
+            {
+                "good_key": "ok",
+                "bad-key": "dropped",  # hyphen
+                "also.bad": "dropped",  # dot
+                "Number9": "kept",
+            },
+        ),
+    )
+
+    meta = stub.sent[0].message.root.params["meta"]
+    assert meta == {"good_key": "ok", "Number9": "kept"}
+
+
+def test_unbind_returns_to_noop():
+    """After ``unbind``, subsequent emits stop reaching the previously bound session."""
+    sink = ChannelSink()
+    stub = _StubSession()
+    sink.bind(stub)
+    sink.unbind()
+
+    _run(sink.emit("x", {}))
+    assert not stub.sent
+
+
+def test_emit_swallows_send_failures():
+    """A failing send_message must not bubble out (notification is best-effort)."""
+
+    class _BoomSession:
+        """Session whose ``send_message`` always raises, to exercise the failure path."""
+
+        async def send_message(self, message):
+            """Raise to simulate a broken transport."""
+            raise RuntimeError("transport broke")
+
+    sink = ChannelSink()
+    sink.bind(_BoomSession())
+    _run(sink.emit("x", {}))  # must not raise