fix: bundle validator accepts Identifier init and wildcard @uncaged/workflow-* imports

- bindingInitializerIsCallable: accept Identifier (e.g. var run = wf)
- import allowlist: startsWith('@uncaged/workflow') instead of exact match list

小橘 🍊（NEKO Team）

.gitignore

@@ -4,3 +4,4 @@ bun.lock
 *.tgz
 tsconfig.tsbuildinfo
 .npmrc
+

CLAUDE.md

@@ -10,7 +10,7 @@ This monorepo implements a workflow engine that executes single-file ESM bundles
 |---------|-----------|
 | **Workflow** | A single-file ESM module that exports `run` (workflow function) and `descriptor` (metadata). Identified by its XXH64 hash (Crockford Base32). |
 | **Bundle** | The physical `.esm.js` file stored in `~/.uncaged/workflow/bundles/`. |
-| **Thread** | A single execution of a workflow, identified by a ULID. Persisted as `.data.jsonl` + `.info.jsonl`. |
+| **Thread** | A single execution of a workflow, identified by a ULID. State lives in CAS (linked nodes); active threads indexed in `threads.json`; completed rows in `history/*.jsonl`. Debug logs use `.info.jsonl`. |
 | **Role** | A named actor within a workflow. Each role produces output with typed `meta`. |
 | **Registry** | `workflow.yaml` — maps workflow names to current/historical bundle hashes. |
 

README.md

@@ -8,7 +8,7 @@ A workflow engine that executes single-file ESM bundles. Each workflow is a self
 |---------|-------------|
 | **Workflow** | A single-file ESM module exporting `run` (workflow function) and `descriptor` (metadata). Identified by its XXH64 hash. |
 | **Bundle** | The physical `.esm.js` file stored in `~/.uncaged/workflow/bundles/`. |
-| **Thread** | A single execution of a workflow, identified by a ULID. Persisted as `.data.jsonl` + `.info.jsonl`. |
+| **Thread** | A single execution of a workflow, identified by a ULID. CAS-backed chain plus `threads.json` / `history/*.jsonl`; `.info.jsonl` for debug logs. |
 | **Role** | A named actor within a workflow. Each role produces output with typed `meta`. Roles live inside template packages (`src/roles/`). |
 | **Registry** | `workflow.yaml` — maps workflow names to current/historical bundle hashes. |
 | **CAS** | Content-Addressed Storage — bundles are immutable and addressed by hash. |

docs/architecture.md

@@ -189,11 +189,15 @@ type WorkflowFn = (
 ├── cas/                           # Global content-addressed blobs (see getGlobalCasDir)
 ├── bundles/
 │   ├── C9NMV6V2TQT81.esm.js       # Crockford Base32 of XXH64
-│   └── C9NMV6V2TQT81.yaml         # Role descriptor sidecar (when present)
+│   ├── C9NMV6V2TQT81.yaml         # Role descriptor sidecar (when present)
+│   └── C9NMV6V2TQT81/             # Per-hash bundle dir (alongside or instead of loose files)
+│       ├── threads.json           # Active threads: threadId → { head, start, updatedAt }
+│       └── history/
+│           └── 2026-05-09.jsonl   # Completed threads (one JSON object per line)
 ├── logs/                          # One folder per bundle hash
 │   └── C9NMV6V2TQT81/
-│       ├── 01KQXKW…YG.data.jsonl  # Thread state
-│       └── 01KQXKW…YG.info.jsonl  # Debug log
+│       ├── 01KQXKW…YG.running     # Present while worker executes this thread (optional)
+│       └── 01KQXKW…YG.info.jsonl   # Debug log
 └── workflow.yaml                  # Registry
 ```
 
@@ -207,18 +211,13 @@ type WorkflowFn = (
 
 Managed by `@uncaged/workflow-register` (`readWorkflowRegistry`, `writeWorkflowRegistry`, …). Shape includes workflow entries and a top-level `config` section used for extract/supervisor model resolution.
 
-### Thread JSONL
+### Thread storage (CAS + index)
 
-**`.data.jsonl`** — Line 1: start record; following lines: role steps with CAS-backed content.
+Thread execution state is a chain of immutable CAS nodes (`StartNode`, `StateNode`, content Merkle blobs). Per bundle:
 
-```jsonc
-// Start record
-{ "name": "solve-issue", "hash": "C9NMV6V2TQT81", "threadId": "01KQXKW…",
-  "parameters": { "prompt": "Fix bug #3", "options": { "maxRounds": 5 } },
-  "timestamp": 1714963200000 }
-// Role output (engine persists contentHash + refs; body in ~/.uncaged/workflow/cas/)
-{ "role": "planner", "contentHash": "…", "meta": { "phases": [...] }, "refs": ["…"], "timestamp": ... }
-```
+- **`threads.json`** — only in-flight threads (`head`, `start`, `updatedAt`).
+- **`history/{YYYY-MM-DD}.jsonl`** — completed threads (`threadId`, `head`, `start`, `completedAt`).
+- **CAS (`cas/`)** — payloads and refs for replay, GC, and fork sharing.
 
 **`.info.jsonl`** — Structured debug log via `@uncaged/workflow-util` `createLogger`:
 

docs/plans/2026-05-11-merkle-call-stack.md

@@ -0,0 +1,197 @@
+# RFC: Merkle Call Stack — Cross-Thread DAG Linking
+
+**Author:** 小橘 🍊（NEKO Team）
+**Date:** 2026-05-11
+**Status:** Draft
+
+## Problem
+
+当 `workflowAsAgent` 在父 workflow 中 spawn 子 workflow 时，父子 thread 之间没有任何 Merkle 链接：
+
+1. **子 thread 不知道自己从哪来** — start node 只有 prompt hash，无法追溯父 thread 的上下文（preparer 分析出的 repoPath、conventions 等）
+2. **父 thread 不知道子 thread 在哪** — developer role 的 state node 里只有 agent 返回的文本，child thread root hash 埋在字符串里，不是结构化 ref
+3. **上下文传递靠序列化到 prompt** — 父 workflow 前置 role 的产出只能通过拼字符串传给子 workflow，丢失了 Merkle DAG 的可遍历性
+
+## Proposal
+
+在 CAS 节点中建立父子 thread 之间的 **双向 Merkle 链接**，形成调用栈结构。
+
+### 新增字段
+
+#### StartNodePayload（子 → 父）
+
+```typescript
+type StartNodePayload = {
+  name: string;
+  hash: string;
+  depth: number;
+  parentState: string | null;   // NEW: 父 thread 调用时的 head state hash
+};
+```
+
+`parentState` 指向子 workflow 被 spawn 时，父 thread 的最后一个 state node hash。这是"调用发生时的调用栈帧"。
+
+#### StateNodePayload（父 → 子）
+
+```typescript
+type StateNodePayload = {
+  role: string;
+  meta: Record<string, unknown>;
+  start: string;
+  content: string;
+  ancestors: string[];
+  compact: string | null;
+  timestamp: number;
+  childThread: string | null;   // NEW: 子 thread 最终 state hash（执行结果）
+};
+```
+
+`childThread` 指向子 thread 完成后的**最终 state hash**（不是 start）——语义上是"函数返回值"，从这里沿 ancestors 可回溯子 thread 的完整执行历史。
+
+### refs 同步
+
+新增的 hash 也必须放进 `refs[]`：
+
+- `StartNode.refs`: `[promptHash, parentState]`（parentState 非 null 时）
+- `StateNode.refs`: `[...existingRefs, childThread]`（childThread 非 null 时）
+
+原因：GC 的 `findReachableHashes` 只走 `refs`，不解析 payload 字段。字段提供语义，refs 保证可达性。
+
+### 具体 DAG 结构
+
+以 `solve-issue`（fix #191）为例，developer role 委托给 `develop` 子 workflow：
+
+```
+父 thread: solve-issue
+═══════════════════════════════════════════════════════════
+
+content("fix #191")
+  hash: ABCD1234
+
+start(solve-issue)
+  hash: START001
+  payload: { name: "solve-issue", hash: BUNDLE_SI, depth: 0, parentState: null }
+  refs: [ABCD1234]
+
+state(preparer)
+  hash: STATE_P1
+  payload: { role: "preparer", meta: { repoPath: "...", ... }, childThread: null, ... }
+  refs: [PREP_CONTENT]
+
+state(developer)                          ──────── 父→子 ────────
+  hash: STATE_D1                                                 │
+  payload: { role: "developer", meta: { ... }, childThread: ★CSTATE_END, ... }
+  refs: [DEV_CONTENT, ★CSTATE_END]                               │
+                                                                  │
+state(submitter)                                                  │
+  hash: STATE_S1                                                  │
+  payload: { role: "submitter", ..., childThread: null }          │
+                                                                  │
+                                                                  │
+子 thread: develop                                                │
+═══════════════════════════════════════════════════════════        │
+                                                                  │
+content("fix #191")          (CAS 去重，可能同 ABCD1234)           │
+  hash: CPROMPT1                                                  │
+                              ──────── 子→父 ────────             │
+start(develop)                          │                         │
+  hash: CHILD_START                     │                         │
+  payload: { name: "develop", hash: BUNDLE_DEV, depth: 1,        │
+             parentState: ★STATE_P1 }   │                         │
+  refs: [CPROMPT1, ★STATE_P1]          │                         │
+                                        │                         │
+state(planner)                          │                         │
+  hash: CSTATE_1                        │                         │
+  ...                                   │                         │
+                                        │                         │
+state(coder)                            │                         │
+  hash: CSTATE_2                        │                         │
+  ...                                   │                         │
+                                        │                         │
+state(reviewer) → state(tester) → state(committer)                │
+                                        │                         │
+  hash: ★CSTATE_END  ◄─────────────────┼─────────────────────────┘
+```
+
+### 遍历路径
+
+**子 thread agent 获取父上下文（上行）：**
+```
+当前 step → start(CHILD_START)
+  → refs[1] = STATE_P1（父 preparer 的 state）
+    → payload.meta.repoPath = "/home/.../workflow"
+    → refs → PREP_CONTENT（完整 preparer 输出）
+    → payload.start = START001（父的 start node）
+      → refs[0] = ABCD1234（原始 prompt）
+```
+
+**从父 thread 追踪子 thread 执行（下行）：**
+```
+STATE_D1（父 developer state）
+  → payload.childThread = CSTATE_END
+    → 子 thread 最终 state
+    → 沿 ancestors 回溯：committer → tester → reviewer → coder → planner
+    → payload.start = CHILD_START（子 thread 入口）
+```
+
+**完整调用栈还原：**
+```
+任意节点 → 沿 start 找到所属 thread 的 StartNode
+  → parentState 非 null？沿 parentState 进入父 thread
+  → 递归直到 parentState = null（顶层 workflow）
+```
+
+## Implementation Plan
+
+### Phase 1: Protocol + CAS 层
+
+1. `workflow-protocol/src/cas-types.ts` — `StartNodePayload` 加 `parentState: string | null`，`StateNodePayload` 加 `childThread: string | null`
+2. `workflow-cas/src/nodes.ts` — `putStartNode` 接受可选 `parentStateHash`，放入 refs；`putStateNode` 接受可选 `childThreadHash`，放入 refs
+3. `workflow-cas/src/nodes.ts` — 解析逻辑兼容新字段（缺失时视为 null）
+
+### Phase 2: Engine 层
+
+4. `workflow-execute/src/engine/engine.ts` — `executeThread` 接受 `parentStateHash: string | null`，传给 `putStartNode`
+5. `workflow-execute/src/workflow-as-agent.ts` — spawn 子 thread 时传入父 thread 当前 head state hash 作为 `parentStateHash`；子 thread 完成后返回最终 state hash
+6. Engine 写 developer role 的 state node 时，把子 thread 最终 hash 写入 `childThread` 字段
+
+### Phase 3: Agent 可观测性
+
+7. Agent prompt 构建（`buildAgentPrompt`）— 当 start node 有 `parentState` 时，提示 agent 可通过 `cas get` 遍历父上下文
+8. CLI `thread show` — 显示 parentState / childThread 链接关系
+
+### Phase 4: 验证
+
+9. 已有测试适配新字段（向后兼容，旧节点 parentState/childThread 为 null）
+10. 新增集成测试：workflowAsAgent 场景下验证双向链接正确写入
+
+## Design Decisions
+
+### 为什么 childThread 指向 end 而不是 start？
+
+- 语义是"函数返回值"——父 role 执行完才产出 state，此时子 thread 已跑完
+- 从 end 沿 ancestors 可回溯到 start；反过来 start 写入时子 thread 还没跑完，无法知道 end
+
+### 为什么 parentState 指向 state 而不是 start？
+
+- 指向父 thread 调用点的**前一个 state**（即调用发生时的 head）
+- 这是子 workflow 能看到的父上下文的"切面"——所有已完成的前置 role 都可达
+- 如果是第一个 role 就 spawn 子 workflow（没有前置 state），parentState 指向父的 start node
+
+### 为什么同时放字段和 refs？
+
+- `refs[]` 服务于 GC（`findReachableHashes` 只遍历 refs）和通用 DAG 遍历
+- `payload.parentState` / `payload.childThread` 服务于语义读取（明确知道哪个 ref 是什么）
+- 不改 GC 逻辑，只加字段，GC 自然正确
+
+### 向后兼容
+
+- 新字段默认 `null`，旧节点解析时缺失字段视为 `null`
+- 不影响已有 thread 的遍历和 GC
+- `depth` 可通过沿 parentState 链上溯来交叉验证（数据自证）
+
+## Open Questions
+
+1. **多子 thread** — 如果一个 role 需要 spawn 多个子 workflow（目前不存在这个场景），`childThread` 应该改成 `childThreads: string[]` 还是保持单个？
+2. **Agent prompt 注入深度** — 子 workflow 的 agent 应该自动遍历多少层父上下文？全部还是限制深度？
+3. **CLI 展示** — `thread show` 要不要递归展示整个调用栈，还是只显示直接链接？

docs/plans/2026-05-12-dashboard-workflow-graph.md

@@ -0,0 +1,224 @@
+# Dashboard Workflow Graph Visualization
+
+**Issue**: #198
+**Status**: In Progress
+**Author**: xingyue
+
+## Overview
+
+在 Dashboard 的 ThreadDetail 页面中嵌入一个交互式流程图，将 workflow 的 `ModeratorTable` 可视化为有向图。用户可以一眼看到角色流转结构和当前执行进度。
+
+## 数据层（✅ 已完成 — PR #201）
+
+### WorkflowGraph 类型
+
+`WorkflowDefinition.moderator`（函数）已替换为 `WorkflowDefinition.table`（声明式 `ModeratorTable`），`buildDescriptor` 自动从 table 提取 graph：
+
+```ts
+type WorkflowGraphEdge = {
+  from: string;              // source role 或 "__start__"
+  to: string;                // target role 或 "__end__"
+  condition: string;         // condition.name 或 "FALLBACK"
+  conditionDescription: string | null;
+};
+
+type WorkflowGraph = {
+  edges: readonly WorkflowGraphEdge[];
+};
+
+type WorkflowDescriptor = {
+  description: string;
+  roles: Record<string, WorkflowRoleDescriptor>;
+  graph: WorkflowGraph;      // 必填，新 bundle 自动生成
+};
+```
+
+### 数据流
+
+```
+ModeratorTable (WorkflowDefinition.table)
+  → buildDescriptor() 自动提取 graph
+    → descriptor.yaml 持久化（hash.yaml）
+      → CLI serve /workflows/:name API 返回 descriptor
+        → Dashboard 前端拿到 graph
+```
+
+### 剩余数据层工作
+
+**serve API 需要返回 descriptor**：当前 `GET /workflows/:name` 只返回 registry entry（hash + timestamp），不含 descriptor。需要从 `bundles/{hash}.yaml` 读取 descriptor 并返回给前端。
+
+方案：在 `routes-workflow.ts` 的 `GET /workflows/:name` 响应中附带 `descriptor` 字段。或者：thread-detail 发现 workflow name 后，请求 `GET /workflows/:name/descriptor` 拿到 graph。
+
+## 前端渲染
+
+### 库选型：React Flow + dagre
+
+| 库 | 优势 | 劣势 |
+|---|---|---|
+| **React Flow** ✅ | React 原生、自定义节点/边、dagre 自动布局、~50KB gzip | 需要学 API |
+| Mermaid | 声明式简单 | 无交互、无法高亮当前步骤 |
+| D3 | 完全控制 | 太底层，手撸成本高 |
+| Cytoscape | 图论强 | React 集成差 |
+
+**依赖新增**：
+
+```json
+{
+  "@xyflow/react": "^12",
+  "@dagrejs/dagre": "^1"
+}
+```
+
+### 图结构映射
+
+```
+WorkflowGraph.edges → React Flow nodes + edges
+
+节点（自动从 edges 推导）:
+  - __start__  → 圆形小节点（入口）
+  - role       → 圆角矩形，显示 role name + description
+  - __end__    → 圆形小节点（终止）
+
+边:
+  - FALLBACK   → 虚线（dashed），无 label
+  - condition  → 实线，label = condition
+                  hover tooltip = conditionDescription
+```
+
+### 布局
+
+使用 dagre 自动计算 TB（top-to-bottom）方向布局：
+
+```ts
+import Dagre from "@dagrejs/dagre";
+
+function layoutGraph(nodes, edges) {
+  const g = new Dagre.graphlib.Graph().setDefaultEdgeLabel(() => ({}));
+  g.setGraph({ rankdir: "TB", nodesep: 60, ranksep: 80 });
+
+  for (const node of nodes) {
+    g.setNode(node.id, { width: 180, height: 60 });
+  }
+  for (const edge of edges) {
+    g.setEdge(edge.source, edge.target);
+  }
+
+  Dagre.layout(g);
+
+  return nodes.map((node) => {
+    const pos = g.node(node.id);
+    return { ...node, position: { x: pos.x - 90, y: pos.y - 30 } };
+  });
+}
+```
+
+### 运行时高亮
+
+ThreadDetail 已有 `records: ThreadRecord[]`，其中 `RoleRecord.role` 就是当前/历史执行的 role。
+
+高亮逻辑：
+
+```ts
+function getNodeStates(records: ThreadRecord[]): Map<string, "completed" | "active"> {
+  const states = new Map<string, "completed" | "active">();
+  const roleRecords = records.filter((r) => r.type === "role");
+
+  for (let i = 0; i < roleRecords.length; i++) {
+    const role = roleRecords[i].role;
+    states.set(role, i === roleRecords.length - 1 ? "active" : "completed");
+  }
+
+  // 如果有 workflow-result，最后一个 role 也是 completed
+  if (records.some((r) => r.type === "workflow-result")) {
+    for (const [k] of states) {
+      states.set(k, "completed");
+    }
+    states.set("__end__", "completed");
+  }
+
+  states.set("__start__", "completed");
+  return states;
+}
+```
+
+节点样式：
+
+| 状态 | 样式 |
+|------|------|
+| default | `border: var(--color-border)`, 暗色背景 |
+| completed | `border: var(--color-success)`, 绿色边框 + ✓ 图标 |
+| active | `border: var(--color-accent)`, 蓝色边框 + 脉冲动画 |
+
+边高亮：当 source 和 target 都至少 completed 时，边变绿。
+
+## 组件结构
+
+```
+workflow-dashboard/src/
+  components/
+    workflow-graph/
+      types.ts           — NodeState 等前端类型
+      index.ts           — export { WorkflowGraph }
+      workflow-graph.tsx  — 主组件，React Flow canvas
+      role-node.tsx       — 自定义 role 节点
+      terminal-node.tsx   — START/END 圆形节点
+      condition-edge.tsx  — 自定义边（虚线/实线 + label）
+      use-layout.ts       — dagre 布局 hook
+```
+
+### 集成到 ThreadDetail
+
+在 ThreadDetail 中，records 列表上方插入可折叠的图面板：
+
+```tsx
+// thread-detail.tsx
+{graph && (
+  <div className="mb-4 border rounded-lg overflow-hidden" style={{ height: 300 }}>
+    <WorkflowGraph graph={graph} nodeStates={getNodeStates(records)} />
+  </div>
+)}
+```
+
+图高度固定 300px，React Flow 支持 pan + zoom，不影响下方 records 滚动。
+
+## 实施计划
+
+### ~~Phase 0: 数据层~~ ✅ Done (PR #201)
+
+- [x] `WorkflowDefinition.moderator` → `table` (ModeratorTable)
+- [x] `WorkflowDescriptor` 新增 `graph: WorkflowGraph`
+- [x] `buildDescriptor` 自动提取 graph
+- [x] `validateWorkflowDescriptor` 校验 graph
+
+### Phase 1: API + 静态图渲染
+
+1. serve API：`GET /workflows/:name` 返回 descriptor（含 graph），或新增 `GET /workflows/:name/descriptor`
+2. Dashboard `api.ts` 新增 `getWorkflowDescriptor(agent, name)` 函数
+3. 安装 `@xyflow/react` + `@dagrejs/dagre`
+4. 实现 `workflow-graph/` 组件集
+5. ThreadDetail 中集成：从 thread-start record 拿 workflow name → 请求 descriptor → 渲染图
+
+**产出**：打开 ThreadDetail 看到 workflow 流程图，无高亮。
+
+### Phase 2: 运行时高亮
+
+1. ThreadDetail 根据 records 计算 nodeStates
+2. 节点/边样式响应状态变化
+3. SSE live 模式下实时更新高亮
+
+**产出**：正在运行的 thread 能看到当前执行到哪个 role。
+
+### Phase 3: 交互增强
+
+1. 点击节点滚动到对应 role 的 RecordCard
+2. 边 hover 显示 conditionDescription tooltip
+3. 节点 hover 显示 role description + schema summary
+
+**产出**：图和记录列表联动。
+
+## 注意事项
+
+- **自循环边**：如 `coder → coder (FALLBACK)`，React Flow 支持自循环，dagre 需要特殊处理（self-edge 用 loop 路径）
+- **大图性能**：dagre 在 <50 节点时性能无忧，workflow 通常 <10 个 role
+- **暗色主题**：Dashboard 已使用 CSS variables，节点/边样式复用现有色板
+- **不提交 pnpm-lock.yaml**

packages/cli-workflow/__tests__/commands.test.ts

@@ -2,10 +2,9 @@ import { afterEach, beforeEach, describe, expect, test } from "bun:test";
 import { mkdir, mkdtemp, readFile, rm, unlink, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
-
-import { getGlobalCasDir } from "@uncaged/workflow-util";
 import { createContentMerkleNode, serializeMerkleNode } from "@uncaged/workflow-cas";
 import { getRegisteredWorkflow, readWorkflowRegistry } from "@uncaged/workflow-register";
+import { getGlobalCasDir } from "@uncaged/workflow-util";
 import { cmdCasGet, cmdCasList, cmdCasPut, cmdCasRm } from "../src/commands/cas/index.js";
 import {
   cmdAdd,
@@ -18,7 +17,7 @@ import {
 } from "../src/commands/workflow/index.js";
 import { addCliArgs } from "./bundle-fixture.js";
 
-const fixtureDescriptor = `export const descriptor = { description: "fixture", roles: {} };
+const fixtureDescriptor = `export const descriptor = { description: "fixture", roles: {}, graph: { edges: [] } };
 `;
 
 const wfPutImport = `import { putContentMerkleNode } from "@uncaged/workflow-cas";
@@ -154,6 +153,7 @@ export const run = async function* (input) { return { returnCode: 0, summary: in
       schema: { type: "object", properties: { greeting: { type: "string" } } },
     },
   },
+  graph: { edges: [] },
 };
 ${wfPutImport}
 export const run = async function* (input, options) {

packages/cli-workflow/__tests__/fork-cli.test.ts

@@ -1,12 +1,16 @@
 import { afterEach, beforeEach, describe, expect, test } from "bun:test";
-import { mkdir, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
+import { mkdir, mkdtemp, rm, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { createCasStore, getContentMerklePayload } from "@uncaged/workflow-cas";
+import { FORK_BRANCH_ROLE, walkStateFramesNewestFirst } from "@uncaged/workflow-execute";
+import { END } from "@uncaged/workflow-runtime";
 import { getGlobalCasDir } from "@uncaged/workflow-util";
+
 import { cmdFork, cmdRun } from "../src/commands/thread/index.js";
 import { cmdAdd } from "../src/commands/workflow/index.js";
 import { pathExists } from "../src/fs-utils.js";
+import { resolveThreadRecord } from "../src/thread-scan.js";
 import { addCliArgs } from "./bundle-fixture.js";
 import { ensureTestWorkflowRegistryConfig } from "./workflow-registry-fixture.js";
 
@@ -20,6 +24,7 @@ export const descriptor = {
     coder: { description: "coder", schema: {} },
     reviewer: { description: "reviewer", schema: {} },
   },
+  graph: { edges: [] },
 };
 export const run = async function* (input, options) {
   const cas = options.cas;
@@ -41,27 +46,6 @@ export const run = async function* (input, options) {
 };
 `;
 
-async function countDataJsonlLines(dataPath: string): Promise<number> {
-  try {
-    const text = await readFile(dataPath, "utf8");
-    return text
-      .trim()
-      .split("\n")
-      .filter((l) => l !== "").length;
-  } catch {
-    return 0;
-  }
-}
-
-async function waitUntilMinDataLines(dataPath: string, minLines: number): Promise<void> {
-  for (let attempt = 0; attempt < 120; attempt++) {
-    if ((await countDataJsonlLines(dataPath)) >= minLines) {
-      return;
-    }
-    await new Promise((r) => setTimeout(r, 25));
-  }
-}
-
 async function waitUntilRunningAbsent(runningPath: string): Promise<void> {
   for (let attempt = 0; attempt < 120; attempt++) {
     if (!(await pathExists(runningPath))) {
@@ -71,6 +55,41 @@ async function waitUntilRunningAbsent(runningPath: string): Promise<void> {
   }
 }
 
+async function waitUntilThreadCompletes(storageRoot: string, threadId: string): Promise<void> {
+  for (let attempt = 0; attempt < 120; attempt++) {
+    const row = await resolveThreadRecord(storageRoot, threadId);
+    if (row?.source === "history") {
+      return;
+    }
+    await new Promise((r) => setTimeout(r, 25));
+  }
+}
+
+async function listMeaningfulRoleContents(
+  storageRoot: string,
+  threadId: string,
+): Promise<Array<{ role: string; content: string }>> {
+  const row = await resolveThreadRecord(storageRoot, threadId);
+  if (row === null) {
+    return [];
+  }
+  const cas = createCasStore(getGlobalCasDir(storageRoot));
+  const frames = await walkStateFramesNewestFirst(cas, row.head);
+  const chronological = [...frames].reverse();
+  const out: Array<{ role: string; content: string }> = [];
+  for (const fr of chronological) {
+    if (fr.payload.role === END || fr.payload.role === FORK_BRANCH_ROLE) {
+      continue;
+    }
+    const content = await getContentMerklePayload(cas, fr.payload.content);
+    out.push({
+      role: fr.payload.role,
+      content: content ?? "",
+    });
+  }
+  return out;
+}
+
 describe("cli fork", () => {
   let prevEnv: string | undefined;
   let storageRoot: string;
@@ -110,10 +129,12 @@ describe("cli fork", () => {
       return;
     }
     const sourceId = ran.value.threadId;
-    const sourceData = join(storageRoot, "logs", hash, `${sourceId}.data.jsonl`);
     const sourceRunning = join(storageRoot, "logs", hash, `${sourceId}.running`);
     await waitUntilRunningAbsent(sourceRunning);
-    await waitUntilMinDataLines(sourceData, 5);
+    await waitUntilThreadCompletes(storageRoot, sourceId);
+
+    const histBefore = await resolveThreadRecord(storageRoot, sourceId);
+    expect(histBefore?.source).toBe("history");
 
     const forked = await cmdFork(storageRoot, sourceId, "planner");
     expect(forked.ok).toBe(true);
@@ -121,25 +142,18 @@ describe("cli fork", () => {
       return;
     }
     const newId = forked.value.threadId;
-    const newData = join(storageRoot, "logs", hash, `${newId}.data.jsonl`);
     const newRunning = join(storageRoot, "logs", hash, `${newId}.running`);
     await waitUntilRunningAbsent(newRunning);
-    await waitUntilMinDataLines(newData, 5);
+    await waitUntilThreadCompletes(storageRoot, newId);
 
-    const text = await readFile(newData, "utf8");
-    const lines = text
-      .trim()
-      .split("\n")
-      .filter((l) => l !== "");
-    expect(lines.length).toBe(5);
-    const start = JSON.parse(lines[0] ?? "{}") as Record<string, unknown>;
-    expect(start.threadId).toBe(newId);
-    expect(start.forkFrom).toEqual({ threadId: sourceId });
+    const forkHist = await resolveThreadRecord(storageRoot, newId);
+    expect(forkHist?.source).toBe("history");
+    expect(forkHist?.start).toBe(histBefore?.start);
 
-    const lastRoleLine = JSON.parse(lines[lines.length - 2] ?? "{}") as Record<string, unknown>;
-    expect(lastRoleLine.role).toBe("reviewer");
-    const cas = createCasStore(getGlobalCasDir(storageRoot));
-    expect(await getContentMerklePayload(cas, String(lastRoleLine.contentHash))).toBe("rev-1");
+    const steps = await listMeaningfulRoleContents(storageRoot, newId);
+    const tail = steps[steps.length - 1];
+    expect(tail?.role).toBe("reviewer");
+    expect(tail?.content).toBe("rev-1");
   });
 
   test("fork without --from-role retries last role", async () => {
@@ -161,10 +175,8 @@ describe("cli fork", () => {
       return;
     }
     const sourceId = ran.value.threadId;
-    const sourceData = join(storageRoot, "logs", hash, `${sourceId}.data.jsonl`);
-    const sourceRunning = join(storageRoot, "logs", hash, `${sourceId}.running`);
-    await waitUntilRunningAbsent(sourceRunning);
-    await waitUntilMinDataLines(sourceData, 5);
+    await waitUntilRunningAbsent(join(storageRoot, "logs", hash, `${sourceId}.running`));
+    await waitUntilThreadCompletes(storageRoot, sourceId);
 
     const forked = await cmdFork(storageRoot, sourceId, null);
     expect(forked.ok).toBe(true);
@@ -172,26 +184,17 @@ describe("cli fork", () => {
       return;
     }
     const newId = forked.value.threadId;
-    const newData = join(storageRoot, "logs", hash, `${newId}.data.jsonl`);
-    const newRunning = join(storageRoot, "logs", hash, `${newId}.running`);
-    await waitUntilRunningAbsent(newRunning);
-    await waitUntilMinDataLines(newData, 5);
+    await waitUntilRunningAbsent(join(storageRoot, "logs", hash, `${newId}.running`));
+    await waitUntilThreadCompletes(storageRoot, newId);
 
-    const text = await readFile(newData, "utf8");
-    const lines = text
-      .trim()
-      .split("\n")
-      .filter((l) => l !== "");
-    expect(lines.length).toBe(5);
-
-    const replayCoder = JSON.parse(lines[2] ?? "{}") as Record<string, unknown>;
-    expect(replayCoder.role).toBe("coder");
-    const cas = createCasStore(getGlobalCasDir(storageRoot));
-    expect(await getContentMerklePayload(cas, String(replayCoder.contentHash))).toBe("c1");
-
-    const lastRoleLine = JSON.parse(lines[lines.length - 2] ?? "{}") as Record<string, unknown>;
-    expect(lastRoleLine.role).toBe("reviewer");
-    expect(await getContentMerklePayload(cas, String(lastRoleLine.contentHash))).toBe("rev-2");
+    const steps = await listMeaningfulRoleContents(storageRoot, newId);
+    expect(steps.length).toBeGreaterThanOrEqual(3);
+    const coderReplay = steps[steps.length - 2];
+    expect(coderReplay?.role).toBe("coder");
+    expect(coderReplay?.content).toBe("c1");
+    const tail = steps[steps.length - 1];
+    expect(tail?.role).toBe("reviewer");
+    expect(tail?.content).toBe("rev-2");
   });
 
   test("fork rejects unknown role with available names", async () => {
@@ -212,10 +215,10 @@ describe("cli fork", () => {
       return;
     }
     const sourceId = ran.value.threadId;
-    const sourceData = join(storageRoot, "logs", added.value.hash, `${sourceId}.data.jsonl`);
-    const sourceRunning = join(storageRoot, "logs", added.value.hash, `${sourceId}.running`);
-    await waitUntilRunningAbsent(sourceRunning);
-    await waitUntilMinDataLines(sourceData, 5);
+    await waitUntilRunningAbsent(
+      join(storageRoot, "logs", added.value.hash, `${sourceId}.running`),
+    );
+    await waitUntilThreadCompletes(storageRoot, sourceId);
 
     const bad = await cmdFork(storageRoot, sourceId, "ghost-role");
     expect(bad.ok).toBe(false);

packages/cli-workflow/__tests__/gc-cli.test.ts

@@ -1,45 +1,17 @@
 import { afterEach, beforeEach, describe, expect, test } from "bun:test";
 import { spawnSync } from "node:child_process";
-import { mkdir, mkdtemp, rm, writeFile } from "node:fs/promises";
+import { mkdir, mkdtemp, rm } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { fileURLToPath } from "node:url";
-import { createCasStore, putContentMerkleNode } from "@uncaged/workflow-cas";
+import { createCasStore, putStartNode } from "@uncaged/workflow-cas";
+import { garbageCollectCas, getBundleDir, upsertThreadEntry } from "@uncaged/workflow-execute";
 import { getGlobalCasDir } from "@uncaged/workflow-util";
-import { garbageCollectCas } from "@uncaged/workflow-execute";
 import { cmdThreadRemove } from "../src/commands/thread/index.js";
 import { pathExists } from "../src/fs-utils.js";
 
 const cliEntryPath = fileURLToPath(new URL("../src/cli.ts", import.meta.url));
 
-async function writeDemoDataJsonl(params: {
-  path: string;
-  threadId: string;
-  bundleHash: string;
-  cas: ReturnType<typeof createCasStore>;
-  activeHash: string;
-}): Promise<void> {
-  const bodyHash = await putContentMerkleNode(params.cas, "p");
-  const text = [
-    JSON.stringify({
-      name: "demo",
-      hash: params.bundleHash,
-      threadId: params.threadId,
-      parameters: { prompt: "hi", options: { maxRounds: 5 } },
-      timestamp: 100,
-    }),
-    JSON.stringify({
-      role: "planner",
-      contentHash: bodyHash,
-      meta: {},
-      refs: [params.activeHash, bodyHash],
-      timestamp: 101,
-    }),
-    "",
-  ].join("\n");
-  await writeFile(params.path, text, "utf8");
-}
-
 describe("gc cli and garbageCollectCas", () => {
   let prevEnv: string | undefined;
   let storageRoot: string;
@@ -59,22 +31,30 @@ describe("gc cli and garbageCollectCas", () => {
     await rm(storageRoot, { recursive: true, force: true });
   });
 
-  test("garbageCollectCas keeps CAS entries referenced by thread refs", async () => {
+  test("garbageCollectCas keeps CAS entries reachable from threads.json roots", async () => {
     const bundleHash = "C9NMV6V2TQT81";
     const threadId = "01AAA1111111111111111111";
-    const logsDir = join(storageRoot, "logs", bundleHash);
-    await mkdir(logsDir, { recursive: true });
+    const bundleDir = getBundleDir(storageRoot, bundleHash);
+    await mkdir(bundleDir, { recursive: true });
 
     const cas = createCasStore(getGlobalCasDir(storageRoot));
-    const activeHash = await cas.put("active-blob");
     const orphanHash = await cas.put("orphan-blob");
-
-    await writeDemoDataJsonl({
-      path: join(logsDir, `${threadId}.data.jsonl`),
-      threadId,
-      bundleHash,
+    const promptHash = await cas.put("prompt-text");
+    const startHash = await putStartNode(
       cas,
-      activeHash,
+      {
+        name: "demo",
+        hash: bundleHash,
+        depth: 0,
+        parentState: null,
+      },
+      promptHash,
+    );
+
+    await upsertThreadEntry(bundleDir, threadId, {
+      head: startHash,
+      start: startHash,
+      updatedAt: 100,
     });
 
     const gc = await garbageCollectCas(storageRoot);
@@ -82,12 +62,12 @@ describe("gc cli and garbageCollectCas", () => {
     if (!gc.ok) {
       return;
     }
-    expect(gc.value.scannedThreads).toBe(1);
-    expect(gc.value.activeRefs).toBe(2);
+    expect(gc.value.scannedThreads).toBe(2);
     expect(gc.value.deletedEntries).toBe(1);
     expect(gc.value.deletedHashes).toEqual([orphanHash]);
 
-    expect(await pathExists(join(getGlobalCasDir(storageRoot), `${activeHash}.txt`))).toBe(true);
+    expect(await pathExists(join(getGlobalCasDir(storageRoot), `${promptHash}.txt`))).toBe(true);
+    expect(await pathExists(join(getGlobalCasDir(storageRoot), `${startHash}.txt`))).toBe(true);
     expect(await pathExists(join(getGlobalCasDir(storageRoot), `${orphanHash}.txt`))).toBe(false);
   });
 
@@ -110,19 +90,27 @@ describe("gc cli and garbageCollectCas", () => {
   test("cli gc prints stats", async () => {
     const bundleHash = "C9NMV6V2TQT81";
     const threadId = "01BBB2222222222222222222";
-    const logsDir = join(storageRoot, "logs", bundleHash);
-    await mkdir(logsDir, { recursive: true });
+    const bundleDir = getBundleDir(storageRoot, bundleHash);
+    await mkdir(bundleDir, { recursive: true });
 
     const cas = createCasStore(getGlobalCasDir(storageRoot));
-    const activeHash = await cas.put("keep-me");
+    const promptHash = await cas.put("prompt-text");
+    const startHash = await putStartNode(
+      cas,
+      {
+        name: "demo",
+        hash: bundleHash,
+        depth: 0,
+        parentState: null,
+      },
+      promptHash,
+    );
     await cas.put("drop-me");
 
-    await writeDemoDataJsonl({
-      path: join(logsDir, `${threadId}.data.jsonl`),
-      threadId,
-      bundleHash,
-      cas,
-      activeHash,
+    await upsertThreadEntry(bundleDir, threadId, {
+      head: startHash,
+      start: startHash,
+      updatedAt: 100,
     });
 
     const env = { ...process.env, UNCAGED_WORKFLOW_STORAGE_ROOT: storageRoot };
@@ -131,23 +119,32 @@ describe("gc cli and garbageCollectCas", () => {
       encoding: "utf8",
     });
     expect(proc.status).toBe(0);
-    expect(String(proc.stdout).trim()).toBe("scanned 1 threads, 2 active refs, deleted 1 entries");
+    expect(String(proc.stdout).trim()).toBe("scanned 2 threads, 2 active refs, deleted 1 entries");
   });
 
   test("thread rm triggers gc so unreferenced CAS is removed", async () => {
     const bundleHash = "C9NMV6V2TQT81";
     const threadId = "01CCC3333333333333333333";
-    const logsDir = join(storageRoot, "logs", bundleHash);
-    await mkdir(logsDir, { recursive: true });
+    const bundleDir = getBundleDir(storageRoot, bundleHash);
+    await mkdir(bundleDir, { recursive: true });
 
     const cas = createCasStore(getGlobalCasDir(storageRoot));
-    const activeHash = await cas.put("pinned-by-ref");
-    await writeDemoDataJsonl({
-      path: join(logsDir, `${threadId}.data.jsonl`),
-      threadId,
-      bundleHash,
+    const promptHash = await cas.put("prompt-text");
+    const startHash = await putStartNode(
       cas,
-      activeHash,
+      {
+        name: "demo",
+        hash: bundleHash,
+        depth: 0,
+        parentState: null,
+      },
+      promptHash,
+    );
+
+    await upsertThreadEntry(bundleDir, threadId, {
+      head: startHash,
+      start: startHash,
+      updatedAt: 100,
     });
 
     const orphanHash = await cas.put("orphan-after-rm");
@@ -157,6 +154,6 @@ describe("gc cli and garbageCollectCas", () => {
     expect(removed.ok).toBe(true);
 
     expect(await pathExists(orphanPath)).toBe(false);
-    expect(await pathExists(join(getGlobalCasDir(storageRoot), `${activeHash}.txt`))).toBe(false);
+    expect(await pathExists(join(getGlobalCasDir(storageRoot), `${promptHash}.txt`))).toBe(false);
   });
 });

packages/cli-workflow/__tests__/init-template.test.ts

@@ -64,6 +64,7 @@ describe("init template", () => {
 
     const moder = await readFile(join(tdir, "src", "moderator.ts"), "utf8");
     expect(moder).not.toContain("export default");
+    expect(moder).toContain("ModeratorTable");
   });
 
   test("finds workspace walking up from nested cwd", async () => {

packages/cli-workflow/__tests__/init-workspace.test.ts

@@ -82,7 +82,7 @@ describe("init workspace", () => {
     for (const term of [
       "RoleDefinition",
       "WorkflowDefinition",
-      "Moderator",
+      "ModeratorTable",
       "AgentFn",
       "ExtractFn",
       "RoleMeta",

packages/cli-workflow/__tests__/live.test.ts

@@ -1,13 +1,10 @@
 import { afterEach, beforeEach, describe, expect, test } from "bun:test";
-import { spawn, spawnSync } from "node:child_process";
-import { cp, mkdir, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
+import { spawnSync } from "node:child_process";
+import { mkdtemp, rm } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { fileURLToPath } from "node:url";
 
-import { createCasStore, putContentMerkleNode } from "@uncaged/workflow-cas";
-import { getGlobalCasDir } from "@uncaged/workflow-util";
-
 import {
   formatLiveDebugLine,
   formatLiveTimeLabel,
@@ -18,11 +15,6 @@ import {
 import { parseLiveArgv } from "../src/live-argv.js";
 
 const cliEntryPath = fileURLToPath(new URL("../src/cli.ts", import.meta.url));
-const fixtureRoot = fileURLToPath(new URL("./fixtures/live", import.meta.url));
-
-/** Bodies for Merkle content nodes; hashes must match `.data.jsonl` fixtures. */
-const LIVE_FIXTURE_PLANNER_BODY =
-  "alpha\nbeta\ngamma\nLINE4\nLINE5\nLINE6\nLINE7\nLINE8\nLINE9\nLINE10\nLINE11";
 
 describe("live helpers", () => {
   test("formatLiveTimeLabel pads HH:MM:SS", () => {
@@ -86,28 +78,6 @@ describe("live CLI", () => {
     prevEnv = process.env.UNCAGED_WORKFLOW_STORAGE_ROOT;
     storageRoot = await mkdtemp(join(tmpdir(), "uncaged-wf-live-"));
     process.env.UNCAGED_WORKFLOW_STORAGE_ROOT = storageRoot;
-    await mkdir(join(storageRoot, "logs", "C9NMV6V2TQT81"), { recursive: true });
-    await cp(
-      join(fixtureRoot, "logs", "C9NMV6V2TQT81", "01LIVECMPLT01DDDDDDDDDDDDG.data.jsonl"),
-      join(storageRoot, "logs", "C9NMV6V2TQT81", "01LIVECMPLT01DDDDDDDDDDDDG.data.jsonl"),
-    );
-    await cp(
-      join(fixtureRoot, "logs", "C9NMV6V2TQT81", "01LIVECMPLT01DDDDDDDDDDDDG.info.jsonl"),
-      join(storageRoot, "logs", "C9NMV6V2TQT81", "01LIVECMPLT01DDDDDDDDDDDDG.info.jsonl"),
-    );
-    await cp(
-      join(fixtureRoot, "logs", "C9NMV6V2TQT81", "01LIVEINFLY01DDDDDDDDDDDDG.data.jsonl"),
-      join(storageRoot, "logs", "C9NMV6V2TQT81", "01LIVEINFLY01DDDDDDDDDDDDG.data.jsonl"),
-    );
-    await cp(
-      join(fixtureRoot, "logs", "C9NMV6V2TQT81", "01LIVEOLDER01DDDDDDDDDDDDG.data.jsonl"),
-      join(storageRoot, "logs", "C9NMV6V2TQT81", "01LIVEOLDER01DDDDDDDDDDDDG.data.jsonl"),
-    );
-
-    const cas = createCasStore(getGlobalCasDir(storageRoot));
-    await putContentMerkleNode(cas, LIVE_FIXTURE_PLANNER_BODY);
-    await putContentMerkleNode(cas, "patch");
-    await putContentMerkleNode(cas, "still running");
   });
 
   afterEach(async () => {
@@ -119,170 +89,6 @@ describe("live CLI", () => {
     await rm(storageRoot, { recursive: true, force: true });
   });
 
-  test("prints role steps and summary for a completed thread", async () => {
-    const env = { ...process.env, UNCAGED_WORKFLOW_STORAGE_ROOT: storageRoot };
-    const proc = spawn(process.execPath, [cliEntryPath, "live", "01LIVECMPLT01DDDDDDDDDDDDG"], {
-      env,
-      stdio: ["ignore", "pipe", "pipe"],
-    });
-    const stdout = await new Promise<string>((resolve, reject) => {
-      let buf = "";
-      proc.stdout?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.stderr?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.on("error", reject);
-      proc.on("exit", (code: number | null) => {
-        if (code === 0) {
-          resolve(buf);
-        } else {
-          reject(new Error(`exit ${code}: ${buf}`));
-        }
-      });
-    });
-
-    expect(stdout).toContain("planner");
-    expect(stdout).toContain("coder");
-    expect(stdout).toContain("meta:");
-    expect(stdout).toContain('"phase":"plan"');
-    expect(stdout).toContain("LINE10");
-    expect(stdout).not.toContain("LINE11");
-    expect(stdout).toContain("more line");
-    expect(stdout).toContain("completed: returnCode=0");
-    expect(stdout).toContain("fixture completed");
-  });
-
-  test("--latest tails the newest thread by start timestamp", async () => {
-    const env = { ...process.env, UNCAGED_WORKFLOW_STORAGE_ROOT: storageRoot };
-    const proc = spawn(process.execPath, [cliEntryPath, "live", "--latest"], {
-      env,
-      stdio: ["ignore", "pipe", "pipe"],
-    });
-    const stdout = await new Promise<string>((resolve, reject) => {
-      let buf = "";
-      proc.stdout?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.stderr?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.on("error", reject);
-      proc.on("exit", (code: number | null) => {
-        if (code === 0) {
-          resolve(buf);
-        } else {
-          reject(new Error(`exit ${code}: ${buf}`));
-        }
-      });
-    });
-
-    expect(stdout).toContain("fixture completed");
-    expect(stdout).not.toContain("older thread");
-  });
-
-  test("--debug prints .info.jsonl records after data output", async () => {
-    const env = { ...process.env, UNCAGED_WORKFLOW_STORAGE_ROOT: storageRoot };
-    const proc = spawn(
-      process.execPath,
-      [cliEntryPath, "live", "01LIVECMPLT01DDDDDDDDDDDDG", "--debug"],
-      {
-        env,
-        stdio: ["ignore", "pipe", "pipe"],
-      },
-    );
-    const stdout = await new Promise<string>((resolve, reject) => {
-      let buf = "";
-      proc.stdout?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.stderr?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.on("error", reject);
-      proc.on("exit", (code: number | null) => {
-        if (code === 0) {
-          resolve(buf);
-        } else {
-          reject(new Error(`exit ${code}: ${buf}`));
-        }
-      });
-    });
-
-    expect(stdout).toContain("[DEBUGTAG1]");
-    expect(stdout).toContain("bundle loaded");
-    expect(stdout).toContain("[DEBUGTAG2]");
-    expect(stdout).toContain("multi line");
-  });
-
-  test("--role filters out non-matching roles", async () => {
-    const env = { ...process.env, UNCAGED_WORKFLOW_STORAGE_ROOT: storageRoot };
-    const proc = spawn(
-      process.execPath,
-      [cliEntryPath, "live", "01LIVECMPLT01DDDDDDDDDDDDG", "--role", "planner"],
-      {
-        env,
-        stdio: ["ignore", "pipe", "pipe"],
-      },
-    );
-    const stdout = await new Promise<string>((resolve, reject) => {
-      let buf = "";
-      proc.stdout?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.stderr?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.on("error", reject);
-      proc.on("exit", (code: number | null) => {
-        if (code === 0) {
-          resolve(buf);
-        } else {
-          reject(new Error(`exit ${code}: ${buf}`));
-        }
-      });
-    });
-
-    expect(stdout).toContain("planner");
-    expect(stdout).not.toContain("patch");
-    expect(stdout).toContain("completed: returnCode=0");
-  });
-
-  test("--latest --debug --role combine", async () => {
-    const env = { ...process.env, UNCAGED_WORKFLOW_STORAGE_ROOT: storageRoot };
-    const proc = spawn(
-      process.execPath,
-      [cliEntryPath, "live", "--latest", "--debug", "--role", "planner"],
-      {
-        env,
-        stdio: ["ignore", "pipe", "pipe"],
-      },
-    );
-    const stdout = await new Promise<string>((resolve, reject) => {
-      let buf = "";
-      proc.stdout?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.stderr?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.on("error", reject);
-      proc.on("exit", (code: number | null) => {
-        if (code === 0) {
-          resolve(buf);
-        } else {
-          reject(new Error(`exit ${code}: ${buf}`));
-        }
-      });
-    });
-
-    expect(stdout).toContain("[DEBUGTAG1]");
-    expect(stdout).toContain("planner");
-    expect(stdout).not.toContain("patch");
-    expect(stdout).toContain("fixture completed");
-  });
-
   test("unknown thread id exits 1", () => {
     const env = { ...process.env, UNCAGED_WORKFLOW_STORAGE_ROOT: storageRoot };
     const r = spawnSync(process.execPath, [cliEntryPath, "live", "01UNKNOWNXXXXXXXXXXXXXXXXX"], {
@@ -292,51 +98,6 @@ describe("live CLI", () => {
     expect(r.status).toBe(1);
     expect(String(r.stderr ?? "")).toContain("thread not found");
   });
-
-  test("follows file until WorkflowResult is appended", async () => {
-    const env = { ...process.env, UNCAGED_WORKFLOW_STORAGE_ROOT: storageRoot };
-    const dataPath = join(
-      storageRoot,
-      "logs",
-      "C9NMV6V2TQT81",
-      "01LIVEINFLY01DDDDDDDDDDDDG.data.jsonl",
-    );
-
-    const proc = spawn(process.execPath, [cliEntryPath, "live", "01LIVEINFLY01DDDDDDDDDDDDG"], {
-      env,
-      stdio: ["ignore", "pipe", "pipe"],
-    });
-
-    await new Promise((r) => setTimeout(r, 120));
-    const prior = await readFile(dataPath, "utf8");
-    await writeFile(
-      dataPath,
-      `${prior.replace(/\s*$/, "")}\n${JSON.stringify({ returnCode: 0, summary: "caught up" })}\n`,
-      "utf8",
-    );
-
-    const stdout = await new Promise<string>((resolve, reject) => {
-      let buf = "";
-      proc.stdout?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.stderr?.on("data", (c: Buffer) => {
-        buf += c.toString("utf8");
-      });
-      proc.on("error", reject);
-      proc.on("exit", (code: number | null) => {
-        if (code === 0) {
-          resolve(buf);
-        } else {
-          reject(new Error(`exit ${code}: ${buf}`));
-        }
-      });
-    });
-
-    expect(stdout).toContain("planner");
-    expect(stdout).toContain("completed: returnCode=0");
-    expect(stdout).toContain("caught up");
-  });
 });
 
 describe("live --latest with empty storage", () => {

packages/cli-workflow/__tests__/thread-cli.test.ts

@@ -1,9 +1,10 @@
 import { afterEach, beforeEach, describe, expect, test } from "bun:test";
 import { spawnSync } from "node:child_process";
-import { mkdir, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
+import { mkdir, mkdtemp, rm, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
-import { dirname, join } from "node:path";
+import { join } from "node:path";
 import { fileURLToPath } from "node:url";
+import { getBundleDir, readThreadsIndex } from "@uncaged/workflow-execute";
 import { getGlobalCasDir } from "@uncaged/workflow-util";
 import { cmdCasPut } from "../src/commands/cas/index.js";
 import {
@@ -18,6 +19,7 @@ import {
 } from "../src/commands/thread/index.js";
 import { cmdAdd } from "../src/commands/workflow/index.js";
 import { pathExists, readTextFileIfExists } from "../src/fs-utils.js";
+import { resolveThreadRecord } from "../src/thread-scan.js";
 import { addCliArgs } from "./bundle-fixture.js";
 import { ensureTestWorkflowRegistryConfig } from "./workflow-registry-fixture.js";
 
@@ -34,6 +36,7 @@ const threadFixtureDescriptor = `export const descriptor = {
     only: { description: "only", schema: {} },
     noop: { description: "noop", schema: {} },
   },
+  graph: { edges: [] },
 };
 `;
 
@@ -101,34 +104,21 @@ export const run = async function* (_input, options) {
 };
 `;
 
-async function countDataJsonlLines(dataPath: string): Promise<number> {
-  try {
-    const text = await readFile(dataPath, "utf8");
-    return text
-      .trim()
-      .split("\n")
-      .filter((l) => l !== "").length;
-  } catch {
-    return 0;
-  }
-}
-
-async function waitUntilMinDataLines(
-  dataPath: string,
-  minLines: number,
-  maxAttempts: number,
-): Promise<void> {
+async function waitUntilRunningFileAbsent(runningPath: string, maxAttempts: number): Promise<void> {
   for (let attempt = 0; attempt < maxAttempts; attempt++) {
-    if ((await countDataJsonlLines(dataPath)) >= minLines) {
+    if (!(await pathExists(runningPath))) {
       return;
     }
     await new Promise((r) => setTimeout(r, 25));
   }
 }
 
-async function waitUntilRunningFileAbsent(runningPath: string, maxAttempts: number): Promise<void> {
+async function waitUntilPredicate(
+  predicate: () => Promise<boolean>,
+  maxAttempts: number,
+): Promise<void> {
   for (let attempt = 0; attempt < maxAttempts; attempt++) {
-    if (!(await pathExists(runningPath))) {
+    if (await predicate()) {
       return;
     }
     await new Promise((r) => setTimeout(r, 25));
@@ -197,11 +187,18 @@ describe("cli thread commands", () => {
     }
     expect(shown.value.includes('"threadId"')).toBe(true);
 
+    const parsed = JSON.parse(shown.value) as Record<string, unknown>;
+    expect(parsed.parentState).toBeNull();
+    const parsedSteps = parsed.steps as Array<Record<string, unknown>>;
+    for (const step of parsedSteps) {
+      expect(step).toHaveProperty("childThread");
+      expect(step.childThread).toBeNull();
+    }
+
     const removed = await cmdThreadRemove(storageRoot, threadId);
     expect(removed.ok).toBe(true);
 
-    const dataPath = join(storageRoot, "logs", added.value.hash, `${threadId}.data.jsonl`);
-    expect(await pathExists(dataPath)).toBe(false);
+    expect(await resolveThreadRecord(storageRoot, threadId)).toBeNull();
   });
 
   test("thread rm runs GC and removes CAS blobs not referenced by any remaining thread", async () => {
@@ -234,9 +231,9 @@ describe("cli thread commands", () => {
       threads = await cmdThreads(storageRoot, []);
     }
 
-    const dataPath = join(storageRoot, "logs", added.value.hash, `${threadId}.data.jsonl`);
-    const runningPath = join(dirname(dataPath), `${threadId}.running`);
+    const runningPath = join(storageRoot, "logs", added.value.hash, `${threadId}.running`);
     await waitUntilRunningFileAbsent(runningPath, 120);
+    expect((await resolveThreadRecord(storageRoot, threadId))?.source).toBe("history");
 
     const put = await cmdCasPut(storageRoot, "keep-after-thread-rm");
     expect(put.ok).toBe(true);
@@ -317,30 +314,31 @@ describe("cli thread commands", () => {
     }
 
     const threadId = ran.value.threadId;
+    const killBundleDir = getBundleDir(storageRoot, added.value.hash);
 
-    await new Promise((r) => setTimeout(r, 50));
+    await waitUntilPredicate(async () => {
+      const idx = await readThreadsIndex(killBundleDir);
+      const ent = idx[threadId];
+      return ent !== undefined && ent.head !== ent.start;
+    }, 80);
 
     const killed = await cmdKill(storageRoot, threadId);
     expect(killed.ok).toBe(true);
 
-    await new Promise((r) => setTimeout(r, 900));
+    await waitUntilPredicate(async () => {
+      return (await resolveThreadRecord(storageRoot, threadId))?.source === "history";
+    }, 120);
 
-    const dataPath = join(storageRoot, "logs", added.value.hash, `${threadId}.data.jsonl`);
-    const text = await readFile(dataPath, "utf8");
-    const lines = text
-      .trim()
-      .split("\n")
-      .filter((l) => l !== "");
-    expect(lines.length).toBe(3);
+    expect((await resolveThreadRecord(storageRoot, threadId))?.source).toBe("history");
 
-    const runningPath = join(dirname(dataPath), `${threadId}.running`);
+    const runningPath = join(storageRoot, "logs", added.value.hash, `${threadId}.running`);
     expect(await pathExists(runningPath)).toBe(false);
   });
 
   test("pause stops between yields and resume completes thread", async () => {
-    const bundleDir = join(storageRoot, "src");
-    await mkdir(bundleDir, { recursive: true });
-    const bundlePath = join(bundleDir, "demo.esm.js");
+    const srcDir = join(storageRoot, "src");
+    await mkdir(srcDir, { recursive: true });
+    const bundlePath = join(srcDir, "demo.esm.js");
     await writeFile(bundlePath, pauseResumeBundleSource, "utf8");
 
     const added = await cmdAdd(storageRoot, addCliArgs("solve-issue", bundlePath));
@@ -356,24 +354,33 @@ describe("cli thread commands", () => {
     }
 
     const threadId = ran.value.threadId;
-    const dataPath = join(storageRoot, "logs", added.value.hash, `${threadId}.data.jsonl`);
+    const bundleDir = getBundleDir(storageRoot, added.value.hash);
 
-    await waitUntilMinDataLines(dataPath, 2, 80);
-    expect(await countDataJsonlLines(dataPath)).toBe(2);
+    await waitUntilPredicate(async () => {
+      const idx = await readThreadsIndex(bundleDir);
+      const ent = idx[threadId];
+      return ent !== undefined && ent.head !== ent.start;
+    }, 80);
+
+    const idxBeforePause = await readThreadsIndex(bundleDir);
+    const headAtPause = idxBeforePause[threadId]?.head;
 
     const paused = await cmdPause(storageRoot, threadId);
     expect(paused.ok).toBe(true);
 
     await new Promise((r) => setTimeout(r, 400));
-    expect(await countDataJsonlLines(dataPath)).toBe(2);
+    const idxPaused = await readThreadsIndex(bundleDir);
+    expect(idxPaused[threadId]?.head).toBe(headAtPause);
 
     const resumed = await cmdResume(storageRoot, threadId);
     expect(resumed.ok).toBe(true);
 
-    await waitUntilMinDataLines(dataPath, 4, 120);
-    expect(await countDataJsonlLines(dataPath)).toBe(4);
+    await waitUntilPredicate(async () => {
+      const row = await resolveThreadRecord(storageRoot, threadId);
+      return row?.source === "history";
+    }, 120);
 
-    const runningPath = join(dirname(dataPath), `${threadId}.running`);
+    const runningPath = join(storageRoot, "logs", added.value.hash, `${threadId}.running`);
     await waitUntilRunningFileAbsent(runningPath, 100);
     expect(await pathExists(runningPath)).toBe(false);
   });
@@ -397,8 +404,7 @@ describe("cli thread commands", () => {
     }
 
     const threadId = ran.value.threadId;
-    const dataPath = join(storageRoot, "logs", added.value.hash, `${threadId}.data.jsonl`);
-    const runningPath = join(dirname(dataPath), `${threadId}.running`);
+    const runningPath = join(storageRoot, "logs", added.value.hash, `${threadId}.running`);
 
     await waitUntilRunningFileAbsent(runningPath, 100);
     expect(await pathExists(runningPath)).toBe(false);

packages/cli-workflow/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@uncaged/cli-workflow",
-  "version": "0.2.0",
+  "version": "0.3.1",
   "type": "module",
   "bin": {
     "uncaged-workflow": "src/cli.ts"

packages/cli-workflow/pnpm-lock.yaml

@@ -0,0 +1,51 @@
+lockfileVersion: '9.0'
+
+settings:
+  autoInstallPeers: true
+  excludeLinksFromLockfile: false
+
+importers:
+
+  .:
+    dependencies:
+      '@uncaged/workflow-cas':
+        specifier: workspace:*
+        version: link:../workflow-cas
+      '@uncaged/workflow-execute':
+        specifier: workspace:*
+        version: link:../workflow-execute
+      '@uncaged/workflow-protocol':
+        specifier: workspace:*
+        version: link:../workflow-protocol
+      '@uncaged/workflow-register':
+        specifier: workspace:*
+        version: link:../workflow-register
+      '@uncaged/workflow-runtime':
+        specifier: workspace:*
+        version: link:../workflow-runtime
+      '@uncaged/workflow-util':
+        specifier: workspace:*
+        version: link:../workflow-util
+      hono:
+        specifier: ^4.12.18
+        version: 4.12.18
+      yaml:
+        specifier: ^2.8.4
+        version: 2.8.4
+
+packages:
+
+  hono@4.12.18:
+    resolution: {integrity: sha512-RWzP96k/yv0PQfyXnWjs6zot20TqfpfsNXhOnev8d1InAxubW93L11/oNUc3tQqn2G0bSdAOBpX+2uDFHV7kdQ==}
+    engines: {node: '>=16.9.0'}
+
+  yaml@2.8.4:
+    resolution: {integrity: sha512-ml/JPOj9fOQK8RNnWojA67GbZ0ApXAUlN2UQclwv2eVgTgn7O9gg9o7paZWKMp4g0H3nTLtS9LVzhkpOFIKzog==}
+    engines: {node: '>= 14.6'}
+    hasBin: true
+
+snapshots:
+
+  hono@4.12.18: {}
+
+  yaml@2.8.4: {}

packages/cli-workflow/src/commands/cas/gc.ts

@@ -1,5 +1,5 @@
-import type { Result } from "@uncaged/workflow-protocol";
 import { type GcResult, garbageCollectCas } from "@uncaged/workflow-execute";
+import type { Result } from "@uncaged/workflow-protocol";
 
 export async function cmdGc(storageRoot: string): Promise<Result<GcResult, string>> {
   return garbageCollectCas(storageRoot);

packages/cli-workflow/src/commands/cas/get.ts

@@ -1,6 +1,6 @@
+import { createCasStore } from "@uncaged/workflow-cas";
 import { err, ok, type Result } from "@uncaged/workflow-protocol";
 import { getGlobalCasDir } from "@uncaged/workflow-util";
-import { createCasStore } from "@uncaged/workflow-cas";
 
 export async function cmdCasGet(
   storageRoot: string,

packages/cli-workflow/src/commands/cas/list.ts

@@ -1,6 +1,6 @@
+import { createCasStore } from "@uncaged/workflow-cas";
 import { ok, type Result } from "@uncaged/workflow-protocol";
 import { getGlobalCasDir } from "@uncaged/workflow-util";
-import { createCasStore } from "@uncaged/workflow-cas";
 
 export async function cmdCasList(storageRoot: string): Promise<Result<string[], string>> {
   const cas = createCasStore(getGlobalCasDir(storageRoot));

packages/cli-workflow/src/commands/cas/put.ts

@@ -1,6 +1,6 @@
+import { createCasStore } from "@uncaged/workflow-cas";
 import { ok, type Result } from "@uncaged/workflow-protocol";
 import { getGlobalCasDir } from "@uncaged/workflow-util";
-import { createCasStore } from "@uncaged/workflow-cas";
 
 export async function cmdCasPut(
   storageRoot: string,

packages/cli-workflow/src/commands/cas/rm.ts

@@ -1,6 +1,6 @@
+import { createCasStore } from "@uncaged/workflow-cas";
 import { ok, type Result } from "@uncaged/workflow-protocol";
 import { getGlobalCasDir } from "@uncaged/workflow-util";
-import { createCasStore } from "@uncaged/workflow-cas";
 
 export async function cmdCasRm(storageRoot: string, hash: string): Promise<Result<void, string>> {
   const cas = createCasStore(getGlobalCasDir(storageRoot));

packages/cli-workflow/src/commands/init/templates.ts

@@ -50,7 +50,6 @@ const greeterMetaSchema = z.object({
 export const greeterRole: RoleDefinition<HelloTemplateMeta["greeter"]> = {
   description: "Says hello — replace with your first role.",
   systemPrompt: "You are a helpful assistant. Reply with one short friendly sentence.",
-  extractPrompt: "Extract the assistant's greeting as message.",
   schema: greeterMetaSchema,
   extractRefs: null,
 };
@@ -58,17 +57,13 @@ export const greeterRole: RoleDefinition<HelloTemplateMeta["greeter"]> = {
 }
 
 export function templateModeratorTs(): string {
-  return `import { END, type Moderator, type ModeratorContext } from "@uncaged/workflow-runtime";
+  return `import { END, START, type ModeratorTable } from "@uncaged/workflow-runtime";
 
 import type { HelloTemplateMeta } from "./roles.js";
 
-export const helloTemplateModerator: Moderator<HelloTemplateMeta> = (
-  ctx: ModeratorContext<HelloTemplateMeta>,
-) => {
-  if (ctx.steps.length === 0) {
-    return "greeter";
-  }
-  return END;
+export const helloTemplateTable: ModeratorTable<HelloTemplateMeta> = {
+  [START]: [{ condition: "FALLBACK", role: "greeter" }],
+  greeter: [{ condition: "FALLBACK", role: END }],
 };
 `;
 }
@@ -76,7 +71,7 @@ export const helloTemplateModerator: Moderator<HelloTemplateMeta> = (
 export function templateIndexTs(): string {
   return `import type { WorkflowDefinition } from "@uncaged/workflow-runtime";
 
-import { helloTemplateModerator } from "./moderator.js";
+import { helloTemplateTable } from "./moderator.js";
 import {
   HELLO_TEMPLATE_DESCRIPTION,
   type HelloTemplateMeta,
@@ -88,14 +83,14 @@ export {
   type HelloTemplateMeta,
   greeterRole,
 } from "./roles.js";
-export { helloTemplateModerator } from "./moderator.js";
+export { helloTemplateTable } from "./moderator.js";
 
 export const helloTemplateWorkflowDefinition: WorkflowDefinition<HelloTemplateMeta> = {
   description: HELLO_TEMPLATE_DESCRIPTION,
   roles: {
     greeter: greeterRole,
   },
-  moderator: helloTemplateModerator,
+  table: helloTemplateTable,
 };
 `;
 }

packages/cli-workflow/src/commands/init/workspace.ts

@@ -85,7 +85,7 @@ function agentsMd(): string {
 | 层级 | 目录 / 产物 | 职责 |
 |------|----------------|------|
 | **Workspace** | 仓库根（\`package.json\` 含 \`workspaces: ["templates/*", "workflows"]\`） | Bun monorepo：统一管理本地模板包与 workflow 实例 |
-| **Template** | \`templates/<name>/\`（如 \`src/roles.ts\`、\`src/moderator.ts\`、\`src/index.ts\`） | 纯数据：**WorkflowDefinition**（各 **RoleDefinition** + **Moderator**），**不绑定**具体 Agent |
+| **Template** | \`templates/<name>/\`（如 \`src/roles.ts\`、\`src/moderator.ts\`、\`src/index.ts\`） | 纯数据：**WorkflowDefinition**（各 **RoleDefinition** + **ModeratorTable**），**不绑定**具体 Agent |
 | **Workflow instance** | \`workflows/\`（或单独包） | 把模板与运行时 **AgentFn** / **ExtractFn** 组合，产出可注册的 **单文件 ESM bundle**（\`run\` + \`descriptor\` 命名导出） |
 
 Init 生成的骨架：\`templates/\` 下放可复用定义，\`workflows/\` 下放绑定与打包入口。
@@ -93,20 +93,20 @@ Init 生成的骨架：\`templates/\` 下放可复用定义，\`workflows/\` 下
 ## 2. 核心概念
 
 - **RoleMeta**：\`Record<string, Record<string, unknown>>\`，角色名 → 该角色结构化 meta 的形状约定。
-- **RoleDefinition<Meta>**：纯数据——\`description\`、\`systemPrompt\`、\`extractPrompt\`、\`schema\`（Zod v4）。不含执行逻辑。
-- **WorkflowDefinition<M extends RoleMeta>**：\`description\` + \`roles\`（各角色定义）+ **Moderator**。
-- **Moderator**：\`(ctx: ModeratorContext<M>) => (角色名) | END\`。同步、纯函数，只做路由。
+- **RoleDefinition<Meta>**：纯数据——\`description\`、\`systemPrompt\`、\`schema\`（Zod v4）。不含执行逻辑。
+- **WorkflowDefinition<M extends RoleMeta>**：\`description\` + \`roles\`（各角色定义）+ **ModeratorTable**（声明式路由表）。
+- **ModeratorTable**：从 \`START\` 与各角色名映射到有序 transition 列表（条件 + 下一角色或 \`END\`）；可序列化，供描述符提取 **graph**。
 - **AgentFn**：\`(ctx: AgentContext) => Promise<string>\`，原始文本输出；从上下文读取当前角色的 \`systemPrompt\`。
-- **ExtractFn**：从上下文与 prompt 解析结构化数据（引擎与 Agent 都可使用）。
+- **ExtractFn**：从 CAS content hash 解析结构化数据（引擎与 Agent 都可使用）。
 
-引擎循环简述：**Moderator** → 选角色 → **Agent** 产出文本 → **Extract** 写入 **meta** → 追加 step，重复直至 **END**。详见 \`docs/architecture.md\` 中的三阶段说明。
+引擎循环简述：按 **ModeratorTable** 选下一角色 → **Agent** 产出文本 → **Extract** 写入 **meta** → 追加 step，重复直至 **END**。详见 \`docs/architecture.md\` 中的三阶段说明。
 
 ## 3. 开发流程
 
 1. **定义 RoleMeta**：为每个角色约定 meta 的 TypeScript 类型（与 Zod schema 对齐）。
-2. **编写 RoleDefinition**：为每个角色写 Zod \`schema\`，补齐 \`systemPrompt\` / \`extractPrompt\` / \`description\`。
-3. **编写 Moderator**：根据 \`ctx.steps\` 与业务状态返回下一个角色名或 \`END\`。
-4. **组装 WorkflowDefinition**：在模板 \`index\` 中导出 definition（以及必要的角色 / moderator 导出）。
+2. **编写 RoleDefinition**：为每个角色写 Zod \`schema\`，补齐 \`systemPrompt\` / \`description\`。
+3. **编写 ModeratorTable**：为 \`START\` 与各角色声明 transition（\`FALLBACK\` 或命名条件 + \`check\`）。
+4. **组装 WorkflowDefinition**：在模板 \`index\` 中导出 definition（以及必要的角色 / table 导出）。
 5. **实例化**：在 workflow 包中使用 \`createWorkflow(def, binding)\`（或项目约定的封装）绑定 **AgentFn**；**ExtractFn** 由引擎从 **workflow.yaml** 注入 \`WorkflowRuntime\`。
 6. **构建**：打包为单个 **.esm.js** bundle，使用 **uncaged-workflow add** 注册。
 
@@ -153,7 +153,7 @@ uncaged-workflow add <name> <path/to/bundle.esm.js>
 
 ---
 
-编写新 workflow 时，先对齐 **RoleMeta → RoleDefinition（Zod）→ Moderator → 绑定 → 单文件 bundle**，再对照本节规范自检。
+编写新 workflow 时，先对齐 **RoleMeta → RoleDefinition（Zod）→ ModeratorTable → 绑定 → 单文件 bundle**，再对照本节规范自检。
 `;
 }
 
@@ -164,7 +164,7 @@ Local workflow development workspace (Bun monorepo).
 
 ## Layout
 
-- \`templates/\` — reusable workflow definition packages (roles + moderator), no agent binding
+- \`templates/\` — reusable workflow definition packages (roles + ModeratorTable), no agent binding
 - \`workflows/\` — workflow instances that bind templates to agents and export \`run\` + \`descriptor\`
 
 ## Commands

packages/cli-workflow/src/commands/serve/app.ts

@@ -8,7 +8,7 @@ import { createWorkflowRoutes } from "./routes-workflow.js";
 
 const MAX_BODY_SIZE = 1_048_576; // 1 MB
 
-export function createApp(storageRoot: string): Hono {
+export function createApp(storageRoot: string, agentToken: string | null): Hono {
   const app = new Hono();
 
   app.onError((_err, c) => {
@@ -37,7 +37,19 @@ export function createApp(storageRoot: string): Hono {
     await next();
   });
 
+  // ── Agent token auth (skip healthz) ───────────────────────────────
+  if (agentToken !== null) {
+    app.use("/api/*", async (c, next) => {
+      const token = c.req.header("X-Agent-Token");
+      if (token !== agentToken) {
+        return c.json({ error: "unauthorized" }, 401);
+      }
+      await next();
+    });
+  }
+
   app.get("/healthz", (c) => c.json({ ok: true }));
+  app.get("/api/healthz", (c) => c.json({ ok: true }));
 
   app.route("/api/workflows", createWorkflowRoutes(storageRoot));
   app.route("/api/threads", createThreadRoutes(storageRoot));

packages/cli-workflow/src/commands/serve/routes-cas.ts

@@ -1,6 +1,6 @@
-import { getGlobalCasDir } from "@uncaged/workflow-util";
 import { createCasStore } from "@uncaged/workflow-cas";
 import { garbageCollectCas } from "@uncaged/workflow-execute";
+import { getGlobalCasDir } from "@uncaged/workflow-util";
 import { Hono } from "hono";
 
 export function createCasRoutes(storageRoot: string): Hono {

packages/cli-workflow/src/commands/serve/routes-live.ts

@@ -1,9 +1,18 @@
-import { statSync, watch } from "node:fs";
-import { dirname, join } from "node:path";
+import { existsSync, statSync, watch } from "node:fs";
+import { join } from "node:path";
+import { createCasStore, getContentMerklePayload } from "@uncaged/workflow-cas";
+import {
+  FORK_BRANCH_ROLE,
+  readThreadsIndex,
+  type ThreadIndex,
+  walkStateFramesNewestFirst,
+} from "@uncaged/workflow-execute";
+import { END } from "@uncaged/workflow-runtime";
+import { getGlobalCasDir } from "@uncaged/workflow-util";
 import { Hono } from "hono";
 import { streamSSE } from "hono/streaming";
 
-import { resolveThreadDataPath } from "../../thread-scan.js";
+import { resolveThreadRecord } from "../../thread-scan.js";
 
 type PumpState = {
   contentOffset: number;
@@ -21,7 +30,6 @@ function fileSize(path: string): number {
 async function readNewBytes(path: string, state: PumpState): Promise<string | null> {
   const size = fileSize(path);
   if (size < state.contentOffset) {
-    // File was truncated — reset
     state.contentOffset = 0;
     state.carry = "";
   }
@@ -42,15 +50,6 @@ function parseJsonLine(line: string): unknown {
   }
 }
 
-function isWorkflowResult(record: unknown): boolean {
-  return (
-    record !== null &&
-    typeof record === "object" &&
-    "type" in (record as Record<string, unknown>) &&
-    (record as Record<string, unknown>).type === "workflow-result"
-  );
-}
-
 function parseNewLines(chunk: string, state: PumpState): string[] {
   state.carry += chunk;
 
@@ -67,52 +66,198 @@ function parseNewLines(chunk: string, state: PumpState): string[] {
   return lines;
 }
 
+type CasSseState = {
+  printedHashes: Set<string>;
+  lastHead: string | null;
+  completionEmitted: boolean;
+};
+
+type LiveSseStream = {
+  writeSSE: (opts: { event: string; data: string; id: string }) => Promise<void>;
+};
+
+function completionFromEndMeta(meta: Record<string, unknown>): {
+  returnCode: number;
+  summary: string;
+} | null {
+  const returnCode = meta.returnCode;
+  const summary = meta.summary;
+  if (typeof returnCode !== "number" || typeof summary !== "string") {
+    return null;
+  }
+  return { returnCode, summary };
+}
+
+async function emitRecordsForHead(params: {
+  storageRoot: string;
+  bundleDir: string;
+  threadId: string;
+  headHash: string;
+  sseState: CasSseState;
+  stream: LiveSseStream;
+  eventId: { n: number };
+}): Promise<boolean> {
+  const cas = createCasStore(getGlobalCasDir(params.storageRoot));
+  const frames = await walkStateFramesNewestFirst(cas, params.headHash);
+  const chronological = [...frames].reverse();
+
+  for (const fr of chronological) {
+    if (params.sseState.printedHashes.has(fr.hash)) {
+      continue;
+    }
+    params.sseState.printedHashes.add(fr.hash);
+
+    const role = fr.payload.role;
+    if (role === FORK_BRANCH_ROLE) {
+      continue;
+    }
+
+    if (role === END) {
+      const wf = completionFromEndMeta(fr.payload.meta);
+      if (wf !== null) {
+        params.eventId.n++;
+        await params.stream.writeSSE({
+          event: "record",
+          data: JSON.stringify({
+            type: "workflow-result",
+            returnCode: wf.returnCode,
+            content: wf.summary,
+            timestamp: null,
+          }),
+          id: String(params.eventId.n),
+        });
+        return true;
+      }
+      continue;
+    }
+
+    const payloadText = await getContentMerklePayload(cas, fr.payload.content);
+    const content =
+      payloadText !== null
+        ? payloadText
+        : `(content not in CAS; contentHash=${fr.payload.content})`;
+
+    params.eventId.n++;
+    await params.stream.writeSSE({
+      event: "record",
+      data: JSON.stringify({
+        type: "role",
+        role: fr.payload.role,
+        contentHash: fr.payload.content,
+        content,
+        meta: fr.payload.meta,
+        timestamp: fr.payload.timestamp,
+      }),
+      id: String(params.eventId.n),
+    });
+  }
+
+  return false;
+}
+
+async function pumpThreadsJsonSse(params: {
+  storageRoot: string;
+  bundleDir: string;
+  threadId: string;
+  sseState: CasSseState;
+  stream: LiveSseStream;
+  eventId: { n: number };
+}): Promise<boolean> {
+  let idx: ThreadIndex;
+  try {
+    idx = await readThreadsIndex(params.bundleDir);
+  } catch {
+    idx = {};
+  }
+
+  const active = idx[params.threadId];
+
+  if (active === undefined) {
+    if (params.sseState.completionEmitted) {
+      return false;
+    }
+    const hist = await resolveThreadRecord(params.storageRoot, params.threadId);
+    if (hist === null || hist.source !== "history") {
+      return false;
+    }
+    params.sseState.completionEmitted = true;
+    return await emitRecordsForHead({
+      storageRoot: params.storageRoot,
+      bundleDir: params.bundleDir,
+      threadId: params.threadId,
+      headHash: hist.head,
+      sseState: params.sseState,
+      stream: params.stream,
+      eventId: params.eventId,
+    });
+  }
+
+  const head = active.head;
+  if (params.sseState.lastHead === null) {
+    params.sseState.lastHead = head;
+    return await emitRecordsForHead({
+      storageRoot: params.storageRoot,
+      bundleDir: params.bundleDir,
+      threadId: params.threadId,
+      headHash: head,
+      sseState: params.sseState,
+      stream: params.stream,
+      eventId: params.eventId,
+    });
+  }
+
+  if (head !== params.sseState.lastHead) {
+    params.sseState.lastHead = head;
+    return await emitRecordsForHead({
+      storageRoot: params.storageRoot,
+      bundleDir: params.bundleDir,
+      threadId: params.threadId,
+      headHash: head,
+      sseState: params.sseState,
+      stream: params.stream,
+      eventId: params.eventId,
+    });
+  }
+
+  return false;
+}
+
 export function createLiveRoutes(storageRoot: string): Hono {
   const app = new Hono();
 
   app.get("/:threadId/live", async (c) => {
     const threadId = c.req.param("threadId");
-    const dataPath = await resolveThreadDataPath(storageRoot, threadId);
-    if (dataPath === null) {
+    const resolved = await resolveThreadRecord(storageRoot, threadId);
+    if (resolved === null) {
       return c.json({ error: `thread not found: ${threadId}` }, 404);
     }
-    const resolvedDataPath = dataPath;
 
-    const infoPath = join(dirname(resolvedDataPath), `${threadId}.info.jsonl`);
+    const threadTarget = resolved;
+    const threadsJsonPath = join(threadTarget.bundleDir, "threads.json");
+    const infoPath = join(storageRoot, "logs", threadTarget.bundleHash, `${threadId}.info.jsonl`);
 
     return streamSSE(c, async (stream) => {
-      const dataState: PumpState = { contentOffset: 0, carry: "" };
       const infoState: PumpState = { contentOffset: 0, carry: "" };
-      let eventId = 0;
+      const sseThreadState: CasSseState = {
+        printedHashes: new Set<string>(),
+        lastHead: null,
+        completionEmitted: false,
+      };
+      const eventId = { n: 0 };
 
       async function pumpData(): Promise<boolean> {
-        let chunk: string | null;
-        try {
-          chunk = await readNewBytes(resolvedDataPath, dataState);
-        } catch {
-          return false;
-        }
-        if (chunk === null) {
-          return false;
-        }
-
-        const lines = parseNewLines(chunk, dataState);
-        for (const line of lines) {
-          const record = parseJsonLine(line);
-          eventId++;
-          await stream.writeSSE({
-            event: "record",
-            data: JSON.stringify(record),
-            id: String(eventId),
-          });
-
-          if (isWorkflowResult(record)) {
-            return true;
-          }
-        }
-        return false;
+        const finished = await pumpThreadsJsonSse({
+          storageRoot,
+          bundleDir: threadTarget.bundleDir,
+          threadId,
+          sseState: sseThreadState,
+          stream,
+          eventId,
+        });
+        return finished;
       }
 
+      // biome-ignore lint/complexity/noExcessiveCognitiveComplexity: SSE newline framing mirrors legacy pump
       async function pumpInfo(): Promise<void> {
         let chunk: string | null;
         try {
@@ -134,28 +279,58 @@ export function createLiveRoutes(storageRoot: string): Hono {
           ) {
             continue;
           }
-          eventId++;
+          eventId.n++;
           await stream.writeSSE({
             event: "info",
             data: JSON.stringify(record),
-            id: String(eventId),
+            id: String(eventId.n),
           });
         }
       }
 
-      // Initial pump
+      eventId.n++;
+      await stream.writeSSE({
+        event: "record",
+        data: JSON.stringify({
+          type: "thread-start",
+          threadId: threadTarget.threadId,
+          bundleHash: threadTarget.bundleHash,
+          head: threadTarget.head,
+          start: threadTarget.start,
+          source: threadTarget.source,
+        }),
+        id: String(eventId.n),
+      });
+
       const done = await pumpData();
-      await pumpInfo();
+      try {
+        await pumpInfo();
+      } catch {
+        // optional info file
+      }
       if (done) {
         return;
       }
 
-      // Watch for changes
+      // If thread is not actively running, emit all records and close — don't keep SSE open
+      const runningPath = join(storageRoot, "logs", threadTarget.bundleHash, `${threadId}.running`);
+      if (!existsSync(runningPath)) {
+        eventId.n++;
+        await stream.writeSSE({
+          event: "done",
+          data: JSON.stringify({ reason: "not-running" }),
+          id: String(eventId.n),
+        });
+        return;
+      }
+
       const controller = new AbortController();
       let completed = false;
 
-      const dataWatcher = watch(resolvedDataPath, async () => {
-        if (completed) return;
+      const threadsJsonWatcher = watch(threadsJsonPath, async () => {
+        if (completed) {
+          return;
+        }
         const finished = await pumpData();
         if (finished) {
           completed = true;
@@ -166,7 +341,9 @@ export function createLiveRoutes(storageRoot: string): Hono {
       let infoWatcher: ReturnType<typeof watch> | null = null;
       try {
         infoWatcher = watch(infoPath, async () => {
-          if (completed) return;
+          if (completed) {
+            return;
+          }
           await pumpInfo();
         });
       } catch {
@@ -175,11 +352,10 @@ export function createLiveRoutes(storageRoot: string): Hono {
 
       stream.onAbort(() => {
         completed = true;
-        dataWatcher.close();
+        threadsJsonWatcher.close();
         infoWatcher?.close();
       });
 
-      // Keep stream alive until completion or client disconnect
       await new Promise<void>((resolve) => {
         if (completed) {
           resolve();
@@ -189,7 +365,7 @@ export function createLiveRoutes(storageRoot: string): Hono {
         stream.onAbort(() => resolve());
       });
 
-      dataWatcher.close();
+      threadsJsonWatcher.close();
       infoWatcher?.close();
     });
   });

packages/cli-workflow/src/commands/serve/routes-thread.ts

@@ -1,21 +1,119 @@
+import { join } from "node:path";
+import { createCasStore, getContentMerklePayload, parseCasThreadNode } from "@uncaged/workflow-cas";
+import { FORK_BRANCH_ROLE, walkStateFramesNewestFirst } from "@uncaged/workflow-execute";
+import { END } from "@uncaged/workflow-runtime";
+import { getGlobalCasDir } from "@uncaged/workflow-util";
 import { Hono } from "hono";
 
-import { readTextFileIfExists } from "../../fs-utils.js";
+import { pathExists } from "../../fs-utils.js";
+import type { HistoricalThreadRow, ResolvedThreadRecord } from "../../thread-scan.js";
 import {
   listHistoricalThreads,
   listRunningThreads,
-  resolveThreadDataPath,
+  resolveThreadListStatus,
+  resolveThreadRecord,
 } from "../../thread-scan.js";
 import { cmdKill, cmdPause, cmdResume } from "../thread/control.js";
 import { cmdRun } from "../thread/run.js";
 
+async function readStartInfo(
+  cas: ReturnType<typeof createCasStore>,
+  startHash: string,
+): Promise<{ name: string | null; prompt: string | null }> {
+  const raw = await cas.get(startHash);
+  if (raw === null) return { name: null, prompt: null };
+  const parsed = parseCasThreadNode(raw);
+  if (parsed === null || parsed.kind !== "start") return { name: null, prompt: null };
+  const name = parsed.node.payload.name;
+  const promptHash = parsed.node.refs[0] ?? null;
+  let prompt: string | null = null;
+  if (promptHash !== null) {
+    prompt = await getContentMerklePayload(cas, promptHash);
+  }
+  return { name, prompt };
+}
+
+async function buildThreadDetailRecords(
+  storageRoot: string,
+  resolved: ResolvedThreadRecord,
+  runningMarkerPresent: boolean,
+  statusRow: HistoricalThreadRow,
+): Promise<unknown[]> {
+  const cas = createCasStore(getGlobalCasDir(storageRoot));
+  const frames = await walkStateFramesNewestFirst(cas, resolved.head);
+  const chronological = [...frames].reverse();
+
+  const { name: workflowName, prompt } = await readStartInfo(cas, resolved.start);
+
+  const status = await resolveThreadListStatus(storageRoot, statusRow, runningMarkerPresent);
+
+  const records: unknown[] = [
+    {
+      type: "thread-start",
+      workflow: workflowName ?? "unknown",
+      prompt: prompt ?? null,
+      threadId: resolved.threadId,
+      status,
+      timestamp: null,
+    },
+  ];
+
+  for (const fr of chronological) {
+    if (fr.payload.role === FORK_BRANCH_ROLE) {
+      continue;
+    }
+    if (fr.payload.role === END) {
+      const returnCode = fr.payload.meta.returnCode;
+      const summary = fr.payload.meta.summary;
+      if (typeof returnCode === "number" && typeof summary === "string") {
+        records.push({
+          type: "workflow-result",
+          returnCode,
+          content: summary,
+          timestamp: fr.payload.timestamp,
+        });
+      }
+      continue;
+    }
+    const payloadText = await getContentMerklePayload(cas, fr.payload.content);
+    const content =
+      payloadText !== null
+        ? payloadText
+        : `(content not in CAS; contentHash=${fr.payload.content})`;
+    records.push({
+      type: "role",
+      role: fr.payload.role,
+      contentHash: fr.payload.content,
+      content,
+      meta: fr.payload.meta,
+      timestamp: fr.payload.timestamp,
+    });
+  }
+
+  return records;
+}
+
 export function createThreadRoutes(storageRoot: string): Hono {
   const app = new Hono();
 
   app.get("/", async (c) => {
     const nameFilter = c.req.query("workflow") ?? null;
     const rows = await listHistoricalThreads(storageRoot, nameFilter);
-    return c.json({ threads: rows });
+    const threads = await Promise.all(
+      rows.map(async (r) => {
+        const runningPath = join(storageRoot, "logs", r.hash, `${r.threadId}.running`);
+        const runningMarkerPresent = await pathExists(runningPath);
+        const status = await resolveThreadListStatus(storageRoot, r, runningMarkerPresent);
+        return {
+          threadId: r.threadId,
+          workflow: r.workflowName,
+          hash: r.hash,
+          startedAt: new Date(r.activityTs).toISOString(),
+          status,
+        };
+      }),
+    );
+    return c.json({ threads });
   });
 
   app.get("/running", async (c) => {
@@ -25,22 +123,26 @@ export function createThreadRoutes(storageRoot: string): Hono {
 
   app.get("/:threadId", async (c) => {
     const threadId = c.req.param("threadId");
-    const dataPath = await resolveThreadDataPath(storageRoot, threadId);
-    if (dataPath === null) {
+    const resolved = await resolveThreadRecord(storageRoot, threadId);
+    if (resolved === null) {
       return c.json({ error: `thread not found: ${threadId}` }, 404);
     }
-    const text = await readTextFileIfExists(dataPath);
-    if (text === null) {
-      return c.json({ error: `thread data missing: ${threadId}` }, 404);
-    }
-    const lines = text.trim().split("\n");
-    const records = lines.map((line) => {
-      try {
-        return JSON.parse(line) as unknown;
-      } catch {
-        return { raw: line };
-      }
-    });
+    const runningPath = join(storageRoot, "logs", resolved.bundleHash, `${threadId}.running`);
+    const runningMarkerPresent = await pathExists(runningPath);
+    const statusRow = {
+      threadId: resolved.threadId,
+      hash: resolved.bundleHash,
+      workflowName: null,
+      source: resolved.source,
+      activityTs: 0,
+      head: resolved.head,
+    };
+    const records = await buildThreadDetailRecords(
+      storageRoot,
+      resolved,
+      runningMarkerPresent,
+      statusRow,
+    );
     return c.json({ threadId, records });
   });
 
@@ -54,13 +156,12 @@ export function createThreadRoutes(storageRoot: string): Hono {
 
     const name = body.workflow;
     const prompt = body.prompt;
-    const maxRounds = typeof body.maxRounds === "number" ? body.maxRounds : 10;
 
     if (typeof name !== "string" || typeof prompt !== "string") {
       return c.json({ error: "workflow (string) and prompt (string) are required" }, 400);
     }
 
-    const result = await cmdRun(storageRoot, name, prompt, maxRounds);
+    const result = await cmdRun(storageRoot, name, prompt);
     if (!result.ok) {
       return c.json({ error: result.error }, 400);
     }

packages/cli-workflow/src/commands/serve/serve.ts

@@ -1,12 +1,27 @@
+import { randomUUID } from "node:crypto";
+import { hostname as osHostname } from "node:os";
 import { err, ok, type Result } from "@uncaged/workflow-protocol";
 import { serve } from "bun";
 
 import { printCliLine } from "../../cli-output.js";
 import { createApp } from "./app.js";
+import {
+  registerWithGateway,
+  startHeartbeat,
+  startTunnel,
+  unregisterFromGateway,
+} from "./tunnel.js";
 import type { ServeOptions } from "./types.js";
 
-export function startServer(storageRoot: string, options: ServeOptions): void {
-  const app = createApp(storageRoot);
+const DEFAULT_GATEWAY_URL = "https://workflow-gateway.shazhou.workers.dev";
+const HEARTBEAT_INTERVAL_MS = 60_000;
+
+export function startServer(
+  storageRoot: string,
+  options: ServeOptions,
+  agentToken: string | null,
+): void {
+  const app = createApp(storageRoot, agentToken);
 
   const server = serve({
     fetch: app.fetch,
@@ -28,30 +43,51 @@ function parsePortValue(value: string | undefined): Result<number, string> {
   return ok(parsed);
 }
 
+function requireNextArg(argv: string[], i: number, flag: string): Result<string, string> {
+  const next = argv[i + 1];
+  if (next === undefined) {
+    return err(`${flag} requires a value`);
+  }
+  return ok(next);
+}
+
 function parseServeArgv(argv: string[]): Result<ServeOptions, string> {
   let port = 7860;
   let hostname = "127.0.0.1";
+  let name = osHostname().split(".")[0].toLowerCase();
+  let noTunnel = false;
+  let gatewayUrl = DEFAULT_GATEWAY_URL;
+  const gatewaySecret = process.env.WORKFLOW_GATEWAY_SECRET ?? "";
+  const stringFlags: Record<string, (v: string) => void> = {
+    "--host": (v) => {
+      hostname = v;
+    },
+    "--name": (v) => {
+      name = v;
+    },
+    "--gateway": (v) => {
+      gatewayUrl = v;
+    },
+  };
 
   for (let i = 0; i < argv.length; i++) {
     const arg = argv[i];
     if (arg === "--port" || arg === "-p") {
       const portResult = parsePortValue(argv[i + 1]);
-      if (!portResult.ok) {
-        return portResult;
-      }
+      if (!portResult.ok) return portResult;
       port = portResult.value;
       i++;
-    } else if (arg === "--host") {
-      const next = argv[i + 1];
-      if (next === undefined) {
-        return err("--host requires a value");
-      }
-      hostname = next;
+    } else if (arg === "--no-tunnel") {
+      noTunnel = true;
+    } else if (arg in stringFlags) {
+      const r = requireNextArg(argv, i, arg);
+      if (!r.ok) return r;
+      stringFlags[arg](r.value);
       i++;
     }
   }
 
-  return ok({ port, hostname });
+  return ok({ port, hostname, name, noTunnel, gatewayUrl, gatewaySecret });
 }
 
 export async function dispatchServe(storageRoot: string, argv: string[]): Promise<number> {
@@ -61,7 +97,65 @@ export async function dispatchServe(storageRoot: string, argv: string[]): Promis
     return 1;
   }
 
-  startServer(storageRoot, parsed.value);
+  const options = parsed.value;
+  const agentToken = options.noTunnel ? null : randomUUID();
+  startServer(storageRoot, options, agentToken);
+
+  if (options.noTunnel) {
+    printCliLine("tunnel disabled (--no-tunnel)");
+    await new Promise(() => {});
+    return 0;
+  }
+
+  // Start cloudflared quick tunnel
+  printCliLine("starting cloudflared quick tunnel...");
+  const tunnel = await startTunnel(options.port);
+
+  if (!tunnel) {
+    printCliLine("failed to create tunnel — continuing without gateway registration");
+    await new Promise(() => {});
+    return 0;
+  }
+
+  printCliLine(`tunnel: ${tunnel.url}`);
+
+  // Register with gateway
+  if (options.gatewaySecret) {
+    const registered = await registerWithGateway(
+      options.gatewayUrl,
+      options.name,
+      tunnel.url,
+      options.gatewaySecret,
+      agentToken!,
+    );
+    if (registered) {
+      printCliLine(`registered with gateway as "${options.name}"`);
+    }
+
+    // Start heartbeat
+    const heartbeatTimer = startHeartbeat(
+      options.gatewayUrl,
+      options.name,
+      tunnel.url,
+      options.gatewaySecret,
+      agentToken!,
+      HEARTBEAT_INTERVAL_MS,
+    );
+
+    // Cleanup on exit
+    const cleanup = async () => {
+      clearInterval(heartbeatTimer);
+      printCliLine("unregistering from gateway...");
+      await unregisterFromGateway(options.gatewayUrl, options.name, options.gatewaySecret);
+      tunnel.process.kill();
+      process.exit(0);
+    };
+
+    process.on("SIGINT", cleanup);
+    process.on("SIGTERM", cleanup);
+  } else {
+    printCliLine("WORKFLOW_GATEWAY_SECRET not set — skipping gateway registration");
+  }
 
   // Keep process alive
   await new Promise(() => {});

packages/cli-workflow/src/commands/serve/tunnel.ts

@@ -0,0 +1,88 @@
+import { printCliLine } from "../../cli-output.js";
+
+type TunnelHandle = {
+  process: ReturnType<typeof Bun.spawn>;
+  url: string;
+};
+
+export async function startTunnel(port: number): Promise<TunnelHandle | null> {
+  const proc = Bun.spawn(["cloudflared", "tunnel", "--url", `http://localhost:${port}`], {
+    stdout: "pipe",
+    stderr: "pipe",
+  });
+
+  // cloudflared prints the URL to stderr
+  const reader = proc.stderr.getReader();
+  const decoder = new TextDecoder();
+  let buffer = "";
+  const deadline = Date.now() + 30_000;
+
+  while (Date.now() < deadline) {
+    const { done, value } = await reader.read();
+    if (done) break;
+    buffer += decoder.decode(value, { stream: true });
+    const match = buffer.match(/https:\/\/[a-z0-9-]+\.trycloudflare\.com/);
+    if (match) {
+      // Release the reader so stderr keeps flowing without backpressure
+      reader.releaseLock();
+      return { process: proc, url: match[0] };
+    }
+  }
+
+  reader.releaseLock();
+  proc.kill();
+  return null;
+}
+
+export async function registerWithGateway(
+  gatewayUrl: string,
+  name: string,
+  tunnelUrl: string,
+  secret: string,
+  agentToken: string,
+): Promise<boolean> {
+  try {
+    const resp = await fetch(`${gatewayUrl}/api/gateway/register`, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ name, url: tunnelUrl, secret, agentToken }),
+    });
+    if (!resp.ok) {
+      const body = await resp.text();
+      printCliLine(`gateway registration failed: ${resp.status} ${body}`);
+      return false;
+    }
+    return true;
+  } catch (e) {
+    printCliLine(`gateway registration error: ${e}`);
+    return false;
+  }
+}
+
+export async function unregisterFromGateway(
+  gatewayUrl: string,
+  name: string,
+  secret: string,
+): Promise<void> {
+  try {
+    await fetch(`${gatewayUrl}/api/gateway/register/${name}`, {
+      method: "DELETE",
+      headers: { Authorization: `Bearer ${secret}` },
+    });
+  } catch {
+    // Best effort — process is exiting
+  }
+}
+
+export function startHeartbeat(
+  gatewayUrl: string,
+  name: string,
+  tunnelUrl: string,
+  secret: string,
+  agentToken: string,
+  intervalMs: number,
+): ReturnType<typeof setInterval> {
+  return setInterval(() => {
+    registerWithGateway(gatewayUrl, name, tunnelUrl, secret, agentToken).catch(() => {});
+  }, intervalMs);
+}

packages/cli-workflow/src/commands/serve/types.ts

@@ -1,4 +1,8 @@
 export type ServeOptions = {
   port: number;
   hostname: string;
+  name: string;
+  noTunnel: boolean;
+  gatewayUrl: string;
+  gatewaySecret: string;
 };

packages/cli-workflow/src/commands/thread/dispatch.ts

@@ -26,12 +26,7 @@ export async function dispatchRun(storageRoot: string, argv: string[]): Promise<
     return 1;
   }
 
-  const result = await cmdRun(
-    storageRoot,
-    parsed.value.name,
-    parsed.value.prompt,
-    parsed.value.maxRounds,
-  );
+  const result = await cmdRun(storageRoot, parsed.value.name, parsed.value.prompt);
   if (!result.ok) {
     printCliError(result.error);
     return 1;
@@ -166,7 +161,7 @@ export async function dispatchFork(storageRoot: string, argv: string[]): Promise
 export const THREAD_SUBCOMMAND_TABLE: Record<string, CommandEntry> = {
   run: {
     handler: dispatchRun,
-    args: "<name> [--prompt <text>] [--max-rounds N]",
+    args: "<name> [--prompt <text>]",
     description: "Start a new thread executing a workflow",
   },
   list: {

packages/cli-workflow/src/commands/thread/fork.ts

@@ -1,11 +1,11 @@
 import { join } from "node:path";
-
+import { createCasStore } from "@uncaged/workflow-cas";
+import { prepareCasFork } from "@uncaged/workflow-execute";
 import { err, ok, type Result } from "@uncaged/workflow-protocol";
-import { generateUlid } from "@uncaged/workflow-util";
-import { buildForkPlan } from "@uncaged/workflow-execute";
+import { generateUlid, getGlobalCasDir } from "@uncaged/workflow-util";
 
-import { pathExists, readTextFileIfExists } from "../../fs-utils.js";
-import { resolveThreadDataPath } from "../../thread-scan.js";
+import { pathExists } from "../../fs-utils.js";
+import { resolveThreadRecord } from "../../thread-scan.js";
 import { ensureWorkerForHash, sendWorkerTcpCommand } from "../../worker-spawn.js";
 
 export async function cmdFork(
@@ -13,49 +13,51 @@ export async function cmdFork(
   threadId: string,
   fromRole: string | null,
 ): Promise<Result<{ threadId: string }, string>> {
-  const dataPath = await resolveThreadDataPath(storageRoot, threadId);
-  if (dataPath === null) {
+  const resolved = await resolveThreadRecord(storageRoot, threadId);
+  if (resolved === null) {
     return err(`thread not found: ${threadId}`);
   }
-  const text = await readTextFileIfExists(dataPath);
-  if (text === null) {
-    return err(`thread data missing: ${threadId}`);
+
+  const bundlePath = join(storageRoot, "bundles", `${resolved.bundleHash}.esm.js`);
+  if (!(await pathExists(bundlePath))) {
+    return err(`bundle file missing for thread hash ${resolved.bundleHash}`);
   }
 
-  const plan = buildForkPlan(text, fromRole);
+  const cas = createCasStore(getGlobalCasDir(storageRoot));
+  const newThreadId = generateUlid(Date.now());
+
+  const plan = await prepareCasFork({
+    cas,
+    bundleDir: resolved.bundleDir,
+    bundleHash: resolved.bundleHash,
+    sourceThreadId: threadId,
+    headHash: resolved.head,
+    startHash: resolved.start,
+    newThreadId,
+    fromRole,
+  });
   if (!plan.ok) {
     return plan;
   }
 
-  const bundlePath = join(storageRoot, "bundles", `${plan.value.hash}.esm.js`);
-  if (!(await pathExists(bundlePath))) {
-    return err(`bundle file missing for thread hash ${plan.value.hash}`);
-  }
-
   const worker = await ensureWorkerForHash(storageRoot, plan.value.hash, bundlePath);
   if (!worker.ok) {
     return worker;
   }
 
-  const newThreadId = generateUlid(Date.now());
-  const stepsOnWire = plan.value.historicalSteps.map((s) => ({
-    role: s.role,
-    contentHash: s.contentHash,
-    meta: s.meta,
-    refs: s.refs,
-    timestamp: s.timestamp,
-  }));
-
+  const p = plan.value;
   const sent = await sendWorkerTcpCommand(
     worker.value.port,
     {
       type: "run",
       threadId: newThreadId,
-      workflowName: plan.value.workflowName,
-      prompt: plan.value.prompt,
-      options: plan.value.runOptions,
-      steps: stepsOnWire,
-      forkSourceThreadId: plan.value.sourceThreadId,
+      workflowName: p.workflowName,
+      prompt: p.prompt,
+      options: p.runOptions,
+      steps: p.steps,
+      stepTimestamps: p.stepTimestamps.length > 0 ? p.stepTimestamps : null,
+      forkSourceThreadId: threadId,
+      forkContinuation: p.forkContinuation,
     },
     { awaitResponseLine: false },
   );

packages/cli-workflow/src/commands/thread/live.ts

@@ -1,17 +1,26 @@
 import { watch } from "node:fs";
-import { readFile } from "node:fs/promises";
+import { mkdir, readFile } from "node:fs/promises";
 import { dirname, join } from "node:path";
-
-import type { CasStore, WorkflowCompletion } from "@uncaged/workflow-protocol";
-import { getGlobalCasDir } from "@uncaged/workflow-util";
 import { createCasStore, getContentMerklePayload } from "@uncaged/workflow-cas";
-import { tryParseRoleStepRecord, tryParseWorkflowResultRecord } from "@uncaged/workflow-execute";
+import {
+  FORK_BRANCH_ROLE,
+  readThreadsIndex,
+  type ThreadIndex,
+  walkStateFramesNewestFirst,
+} from "@uncaged/workflow-execute";
+import type { CasStore, WorkflowCompletion } from "@uncaged/workflow-protocol";
+import { END } from "@uncaged/workflow-runtime";
+import { getGlobalCasDir } from "@uncaged/workflow-util";
 
 import { dimGreyLine, highlightLiveRole } from "../../cli-color.js";
 import { printCliError, printCliLine } from "../../cli-output.js";
 import { pathExists } from "../../fs-utils.js";
 import type { ParsedLiveArgv } from "../../live-argv.js";
-import { findLatestThreadDataPath, resolveThreadDataPath } from "../../thread-scan.js";
+import {
+  findLatestThreadBundleTarget,
+  type LatestThreadTarget,
+  resolveThreadRecord,
+} from "../../thread-scan.js";
 import type { LiveRoleRow } from "./types.js";
 
 export const LIVE_CONTENT_MAX_LINES = 10;
@@ -49,16 +58,15 @@ function printSummary(result: WorkflowCompletion): void {
   printCliLine(`completed: returnCode=${result.returnCode} — ${result.summary}`);
 }
 
-type LiveSessionState = {
-  sawStart: boolean;
-  completed: boolean;
+type InfoLiveState = {
   carry: string;
   contentOffset: number;
 };
 
-type InfoLiveState = {
-  carry: string;
-  contentOffset: number;
+type CasLiveState = {
+  printedHashes: Set<string>;
+  lastHead: string | null;
+  completionEmitted: boolean;
 };
 
 function tryParseInfoRecord(obj: Record<string, unknown>): {
@@ -80,102 +88,140 @@ function tryParseInfoRecord(obj: Record<string, unknown>): {
   return { tag, content, timestamp };
 }
 
-async function handleJsonlLine(
-  rawLine: string,
-  state: LiveSessionState,
-  roleFilter: string | null,
-  cas: CasStore,
-): Promise<{ parseError: string | null; workflowResult: WorkflowCompletion | null }> {
-  const trimmed = rawLine.trim();
-  if (trimmed === "") {
-    return { parseError: null, workflowResult: null };
+function completionFromEndMeta(meta: Record<string, unknown>): WorkflowCompletion | null {
+  const returnCode = meta.returnCode;
+  const summary = meta.summary;
+  if (typeof returnCode !== "number" || typeof summary !== "string") {
+    return null;
   }
+  return { returnCode, summary };
+}
 
-  let rec: unknown;
-  try {
-    rec = JSON.parse(trimmed) as unknown;
-  } catch {
-    return { parseError: "invalid JSON in thread data file", workflowResult: null };
+async function emitRoleStepPrint(params: {
+  cas: CasStore;
+  role: string;
+  contentHash: string;
+  meta: Record<string, unknown>;
+  timestamp: number;
+  roleFilter: string | null;
+}): Promise<void> {
+  if (params.roleFilter !== null && params.role !== params.roleFilter) {
+    return;
   }
-  if (rec === null || typeof rec !== "object") {
-    return { parseError: "invalid record in thread data file", workflowResult: null };
-  }
-  const obj = rec as Record<string, unknown>;
-
-  if (!state.sawStart) {
-    state.sawStart = true;
-    return { parseError: null, workflowResult: null };
-  }
-
-  const wf = tryParseWorkflowResultRecord(obj);
-  if (wf !== null) {
-    state.completed = true;
-    return { parseError: null, workflowResult: wf };
-  }
-
-  const roleRow = tryParseRoleStepRecord(obj);
-  if (roleRow === null) {
-    return {
-      parseError: "unrecognized record in thread data (expected role step or result)",
-      workflowResult: null,
-    };
-  }
-
-  if (roleFilter !== null && roleRow.role !== roleFilter) {
-    return { parseError: null, workflowResult: null };
-  }
-
-  const payload = await getContentMerklePayload(cas, roleRow.contentHash);
+  const payload = await getContentMerklePayload(params.cas, params.contentHash);
   const content =
-    payload !== null ? payload : `(content not in CAS; contentHash=${roleRow.contentHash})`;
+    payload !== null ? payload : `(content not in CAS; contentHash=${params.contentHash})`;
 
   const row: LiveRoleRow = {
-    role: roleRow.role,
+    role: params.role,
     content,
-    meta: roleRow.meta,
-    timestamp: roleRow.timestamp,
+    meta: params.meta,
+    timestamp: params.timestamp,
   };
   for (const outLine of renderLiveRoleStepLines(row, highlightLiveRole(row.role))) {
     printCliLine(outLine);
   }
-  return { parseError: null, workflowResult: null };
 }
 
-async function pumpNewContent(
-  dataPath: string,
-  state: LiveSessionState,
-  roleFilter: string | null,
-  cas: CasStore,
-): Promise<number | null> {
-  let text: string;
+async function emitStatesReachableFromHead(params: {
+  cas: CasStore;
+  headHash: string;
+  state: CasLiveState;
+  roleFilter: string | null;
+}): Promise<WorkflowCompletion | null> {
+  const frames = await walkStateFramesNewestFirst(params.cas, params.headHash);
+  const chronological = [...frames].reverse();
+
+  for (const fr of chronological) {
+    if (params.state.printedHashes.has(fr.hash)) {
+      continue;
+    }
+    params.state.printedHashes.add(fr.hash);
+
+    const role = fr.payload.role;
+    if (role === FORK_BRANCH_ROLE) {
+      continue;
+    }
+
+    if (role === END) {
+      const wf = completionFromEndMeta(fr.payload.meta);
+      if (wf !== null) {
+        printSummary(wf);
+        return wf;
+      }
+      continue;
+    }
+
+    await emitRoleStepPrint({
+      cas: params.cas,
+      role,
+      contentHash: fr.payload.content,
+      meta: fr.payload.meta,
+      timestamp: fr.payload.timestamp,
+      roleFilter: params.roleFilter,
+    });
+  }
+
+  return null;
+}
+
+async function pumpThreadsJson(params: {
+  storageRoot: string;
+  bundleDir: string;
+  bundleHash: string;
+  threadId: string;
+  state: CasLiveState;
+  roleFilter: string | null;
+  cas: CasStore;
+}): Promise<number | null> {
+  let idx: ThreadIndex;
   try {
-    text = await readFile(dataPath, "utf8");
+    idx = await readThreadsIndex(params.bundleDir);
   } catch {
-    return null;
+    idx = {};
   }
 
-  if (text.length < state.contentOffset) {
-    state.contentOffset = 0;
-    state.carry = "";
+  const active = idx[params.threadId];
+
+  if (active === undefined) {
+    if (params.state.completionEmitted) {
+      return null;
+    }
+    const hist = await resolveThreadRecord(params.storageRoot, params.threadId);
+    if (hist === null || hist.source !== "history") {
+      return null;
+    }
+    params.state.completionEmitted = true;
+    const wf = await emitStatesReachableFromHead({
+      cas: params.cas,
+      headHash: hist.head,
+      state: params.state,
+      roleFilter: params.roleFilter,
+    });
+    return wf !== null ? 0 : null;
   }
 
-  const chunk = text.slice(state.contentOffset);
-  state.contentOffset = text.length;
-  state.carry += chunk;
+  const head = active.head;
+  if (params.state.lastHead === null) {
+    params.state.lastHead = head;
+    const wf = await emitStatesReachableFromHead({
+      cas: params.cas,
+      headHash: head,
+      state: params.state,
+      roleFilter: params.roleFilter,
+    });
+    return wf !== null ? 0 : null;
+  }
 
-  const parts = state.carry.split("\n");
-  state.carry = parts.pop() ?? "";
-
-  for (const line of parts) {
-    const { parseError, workflowResult } = await handleJsonlLine(line, state, roleFilter, cas);
-    if (parseError !== null) {
-      printCliError(parseError);
-      return 1;
-    }
-    if (workflowResult !== null) {
-      printSummary(workflowResult);
-      return 0;
-    }
+  if (head !== params.state.lastHead) {
+    params.state.lastHead = head;
+    const wf = await emitStatesReachableFromHead({
+      cas: params.cas,
+      headHash: head,
+      state: params.state,
+      roleFilter: params.roleFilter,
+    });
+    return wf !== null ? 0 : null;
   }
 
   return null;
@@ -292,9 +338,9 @@ function watchLivePaths(params: { tasks: WatchPumpTask[]; signal: AbortSignal })
         schedulePump(path, pump);
       });
       watchers.push(watcher);
-      watcher.on("error", (err: Error) => {
+      watcher.on("error", (errObj: Error) => {
         closeAll();
-        reject(err);
+        reject(errObj);
       });
     }
 
@@ -310,17 +356,14 @@ function watchLivePaths(params: { tasks: WatchPumpTask[]; signal: AbortSignal })
   });
 }
 
-type LiveThreadTarget = {
-  threadId: string;
-  dataPath: string;
-};
+type LiveThreadTarget = LatestThreadTarget;
 
 async function resolveLiveThreadTarget(
   storageRoot: string,
   parsed: ParsedLiveArgv,
 ): Promise<LiveThreadTarget | null> {
   if (parsed.latest) {
-    const found = await findLatestThreadDataPath(storageRoot);
+    const found = await findLatestThreadBundleTarget(storageRoot);
     if (found === null) {
       printCliError("live: no threads found");
       return null;
@@ -333,36 +376,56 @@ async function resolveLiveThreadTarget(
     printCliError("live: internal error: missing thread id");
     return null;
   }
-  const resolved = await resolveThreadDataPath(storageRoot, id);
+  const resolved = await resolveThreadRecord(storageRoot, id);
   if (resolved === null) {
     printCliError(`thread not found: ${id}`);
     return null;
   }
-  return { threadId: id, dataPath: resolved };
+  return {
+    threadId: id,
+    bundleHash: resolved.bundleHash,
+    bundleDir: resolved.bundleDir,
+    threadsJsonPath: join(resolved.bundleDir, "threads.json"),
+  };
 }
 
 async function buildLiveWatchTasks(params: {
-  dataPath: string;
-  infoPath: string;
+  storageRoot: string;
+  target: LiveThreadTarget;
   debug: boolean;
-  dataState: LiveSessionState;
+  dataState: CasLiveState;
   infoState: InfoLiveState;
   roleFilter: string | null;
   cas: CasStore;
 }): Promise<WatchPumpTask[]> {
-  const { dataPath, infoPath, debug, dataState, infoState, roleFilter, cas } = params;
+  const infoPath = join(
+    params.storageRoot,
+    "logs",
+    params.target.bundleHash,
+    `${params.target.threadId}.info.jsonl`,
+  );
+
   const tasks: WatchPumpTask[] = [
     {
-      path: dataPath,
-      pump: () => pumpNewContent(dataPath, dataState, roleFilter, cas),
+      path: params.target.threadsJsonPath,
+      pump: () =>
+        pumpThreadsJson({
+          storageRoot: params.storageRoot,
+          bundleDir: params.target.bundleDir,
+          bundleHash: params.target.bundleHash,
+          threadId: params.target.threadId,
+          state: params.dataState,
+          roleFilter: params.roleFilter,
+          cas: params.cas,
+        }),
     },
   ];
 
-  if (debug && (await pathExists(infoPath))) {
+  if (params.debug && (await pathExists(infoPath))) {
     tasks.push({
       path: infoPath,
       pump: async () => {
-        await pumpNewInfoContent(infoPath, infoState);
+        await pumpNewInfoContent(infoPath, params.infoState);
         return null;
       },
     });
@@ -377,16 +440,13 @@ export async function cmdLive(storageRoot: string, parsed: ParsedLiveArgv): Prom
     return 1;
   }
 
-  const { threadId, dataPath } = target;
   const roleFilter = parsed.role;
-  const infoPath = join(dirname(dataPath), `${threadId}.info.jsonl`);
   const cas = createCasStore(getGlobalCasDir(storageRoot));
 
-  const dataState: LiveSessionState = {
-    sawStart: false,
-    completed: false,
-    carry: "",
-    contentOffset: 0,
+  const dataState: CasLiveState = {
+    printedHashes: new Set<string>(),
+    lastHead: null,
+    completionEmitted: false,
   };
 
   const infoState: InfoLiveState = {
@@ -401,22 +461,29 @@ export async function cmdLive(storageRoot: string, parsed: ParsedLiveArgv): Prom
   process.on("SIGINT", onSigInt);
 
   try {
-    const firstData = await pumpNewContent(dataPath, dataState, roleFilter, cas);
-    if (firstData === 1) {
-      return 1;
-    }
+    await mkdir(dirname(target.threadsJsonPath), { recursive: true });
 
+    const firstData = await pumpThreadsJson({
+      storageRoot,
+      bundleDir: target.bundleDir,
+      bundleHash: target.bundleHash,
+      threadId: target.threadId,
+      state: dataState,
+      roleFilter,
+      cas,
+    });
+    const infoPath = join(storageRoot, "logs", target.bundleHash, `${target.threadId}.info.jsonl`);
     if (parsed.debug && (await pathExists(infoPath))) {
       await pumpNewInfoContent(infoPath, infoState);
     }
 
-    if (firstData === 0 || dataState.completed) {
+    if (firstData === 0) {
       return 0;
     }
 
     const tasks = await buildLiveWatchTasks({
-      dataPath,
-      infoPath,
+      storageRoot,
+      target,
       debug: parsed.debug,
       dataState,
       infoState,

packages/cli-workflow/src/commands/thread/rm.ts

@@ -1,25 +1,35 @@
 import { unlink } from "node:fs/promises";
-import { dirname, join } from "node:path";
-
+import { join } from "node:path";
+import {
+  garbageCollectCas,
+  removeThreadEntry,
+  removeThreadHistoryEntries,
+} from "@uncaged/workflow-execute";
 import { err, ok, type Result } from "@uncaged/workflow-protocol";
-import { garbageCollectCas } from "@uncaged/workflow-execute";
 
-import { resolveThreadDataPath } from "../../thread-scan.js";
+import { resolveThreadRecord } from "../../thread-scan.js";
 
 export async function cmdThreadRemove(
   storageRoot: string,
   threadId: string,
 ): Promise<Result<void, string>> {
-  const dataPath = await resolveThreadDataPath(storageRoot, threadId);
-  if (dataPath === null) {
+  const resolved = await resolveThreadRecord(storageRoot, threadId);
+  if (resolved === null) {
     return err(`thread not found: ${threadId}`);
   }
 
-  const dir = dirname(dataPath);
-  const infoPath = join(dir, `${threadId}.info.jsonl`);
-  const runningPath = join(dir, `${threadId}.running`);
+  if (resolved.source === "active") {
+    await removeThreadEntry(resolved.bundleDir, threadId);
+  } else {
+    const hist = await removeThreadHistoryEntries(resolved.bundleDir, threadId);
+    if (!hist.ok) {
+      return hist;
+    }
+  }
+
+  const infoPath = join(storageRoot, "logs", resolved.bundleHash, `${threadId}.info.jsonl`);
+  const runningPath = join(storageRoot, "logs", resolved.bundleHash, `${threadId}.running`);
 
-  await unlink(dataPath);
   await unlink(infoPath).catch(() => {});
   await unlink(runningPath).catch(() => {});
 

packages/cli-workflow/src/commands/thread/run.ts

@@ -1,8 +1,8 @@
 import { join } from "node:path";
 
 import { err, ok, type Result } from "@uncaged/workflow-protocol";
-import { generateUlid } from "@uncaged/workflow-util";
 import { getRegisteredWorkflow, readWorkflowRegistry } from "@uncaged/workflow-register";
+import { generateUlid } from "@uncaged/workflow-util";
 import { ensureWorkerForHash, sendWorkerTcpCommand } from "../../worker-spawn.js";
 import { validateCliWorkflowName } from "../../workflow-name.js";
 
@@ -10,7 +10,6 @@ export async function cmdRun(
   storageRoot: string,
   name: string,
   prompt: string,
-  maxRounds: number,
 ): Promise<Result<{ threadId: string }, string>> {
   const nameOk = validateCliWorkflowName(name);
   if (!nameOk.ok) {
@@ -41,7 +40,7 @@ export async function cmdRun(
       threadId,
       workflowName: name,
       prompt,
-      options: { maxRounds, depth: 0 },
+      options: { depth: 0 },
     },
     { awaitResponseLine: false },
   );

packages/cli-workflow/src/commands/thread/show.ts

@@ -1,19 +1,74 @@
+import { createCasStore, getContentMerklePayload, parseCasThreadNode } from "@uncaged/workflow-cas";
+import { FORK_BRANCH_ROLE, walkStateFramesNewestFirst } from "@uncaged/workflow-execute";
 import { err, ok, type Result } from "@uncaged/workflow-protocol";
+import { END } from "@uncaged/workflow-runtime";
+import { getGlobalCasDir } from "@uncaged/workflow-util";
 
-import { readTextFileIfExists } from "../../fs-utils.js";
-import { resolveThreadDataPath } from "../../thread-scan.js";
+import { resolveThreadRecord } from "../../thread-scan.js";
+
+async function readParentStateFromStartNode(
+  cas: { get(hash: string): Promise<string | null> },
+  startHash: string,
+): Promise<string | null> {
+  const yamlText = await cas.get(startHash);
+  if (yamlText === null) {
+    return null;
+  }
+  const parsed = parseCasThreadNode(yamlText);
+  if (parsed === null || parsed.kind !== "start") {
+    return null;
+  }
+  return parsed.node.payload.parentState;
+}
 
 export async function cmdThreadShow(
   storageRoot: string,
   threadId: string,
 ): Promise<Result<string, string>> {
-  const dataPath = await resolveThreadDataPath(storageRoot, threadId);
-  if (dataPath === null) {
+  const resolved = await resolveThreadRecord(storageRoot, threadId);
+  if (resolved === null) {
     return err(`thread not found: ${threadId}`);
   }
-  const text = await readTextFileIfExists(dataPath);
-  if (text === null) {
-    return err(`thread data missing: ${threadId}`);
+
+  const cas = createCasStore(getGlobalCasDir(storageRoot));
+  const frames = await walkStateFramesNewestFirst(cas, resolved.head);
+  const chronological = [...frames].reverse();
+
+  const parentState = await readParentStateFromStartNode(cas, resolved.start);
+
+  const steps: Array<{
+    role: string;
+    hash: string;
+    timestamp: number;
+    content: string;
+    childThread: string | null;
+  }> = [];
+  for (const fr of chronological) {
+    if (fr.payload.role === END || fr.payload.role === FORK_BRANCH_ROLE) {
+      continue;
+    }
+    const payloadText = await getContentMerklePayload(cas, fr.payload.content);
+    steps.push({
+      role: fr.payload.role,
+      hash: fr.hash,
+      timestamp: fr.payload.timestamp,
+      content:
+        payloadText !== null
+          ? payloadText
+          : `(content not in CAS; contentHash=${fr.payload.content})`,
+      childThread: fr.payload.childThread,
+    });
   }
-  return ok(text.endsWith("\n") ? text.slice(0, -1) : text);
+
+  const payload = {
+    threadId: resolved.threadId,
+    bundleHash: resolved.bundleHash,
+    head: resolved.head,
+    start: resolved.start,
+    parentState,
+    source: resolved.source,
+    steps,
+  };
+
+  return ok(JSON.stringify(payload, null, 2));
 }

packages/cli-workflow/src/commands/workflow/add.ts

@@ -1,8 +1,7 @@
 import { readFile, stat } from "node:fs/promises";
 import { basename, resolve } from "node:path";
-
-import { err, ok, type Result } from "@uncaged/workflow-protocol";
 import { hashWorkflowBundleBytes } from "@uncaged/workflow-cas";
+import { err, ok, type Result } from "@uncaged/workflow-protocol";
 import {
   extractBundleExports,
   readWorkflowRegistry,

packages/cli-workflow/src/run-argv.ts

@@ -3,12 +3,12 @@ import { err, ok, type Result } from "@uncaged/workflow-protocol";
 export type ParsedRunArgv = {
   name: string;
   prompt: string;
-  maxRounds: number;
 };
 
-type FlagOk = { kind: "prompt"; value: string } | { kind: "max-rounds"; value: number };
-
-function parseFlagAt(argv: string[], index: number): Result<FlagOk, string> | null {
+function parseFlagAt(
+  argv: string[],
+  index: number,
+): Result<{ kind: "prompt"; value: string }, string> | null {
   const flag = argv[index];
   if (flag === "--prompt") {
     const value = argv[index + 1];
@@ -17,24 +17,12 @@ function parseFlagAt(argv: string[], index: number): Result<FlagOk, string> | nu
     }
     return ok({ kind: "prompt", value });
   }
-  if (flag === "--max-rounds") {
-    const value = argv[index + 1];
-    if (value === undefined) {
-      return err("missing value for --max-rounds");
-    }
-    const n = Number(value);
-    if (!Number.isFinite(n) || !Number.isInteger(n) || n < 0) {
-      return err("--max-rounds must be a non-negative integer");
-    }
-    return ok({ kind: "max-rounds", value: n });
-  }
   return null;
 }
 
 export function parseRunArgv(argv: string[]): Result<ParsedRunArgv, string> {
   let name: string | undefined;
   let prompt = "";
-  let maxRounds = 5;
 
   let i = 0;
   const first = argv[0];
@@ -54,12 +42,7 @@ export function parseRunArgv(argv: string[]): Result<ParsedRunArgv, string> {
     }
 
     const flag = parsed.value;
-    if (flag.kind === "prompt") {
-      prompt = flag.value;
-      i += 2;
-      continue;
-    }
-    maxRounds = flag.value;
+    prompt = flag.value;
     i += 2;
   }
 
@@ -67,5 +50,5 @@ export function parseRunArgv(argv: string[]): Result<ParsedRunArgv, string> {
     return err("run requires <name>");
   }
 
-  return ok({ name, prompt, maxRounds });
+  return ok({ name, prompt });
 }

packages/cli-workflow/src/skill.ts

@@ -70,8 +70,8 @@ function formatSkillCli(): string {
 |---------|-------------|
 | **Workflow** | A single-file ESM bundle (\`.esm.js\`) that exports \`run\` and \`descriptor\`. Identified by name and XXH64 hash. |
 | **Bundle** | The physical \`.esm.js\` file stored in the bundles directory. Immutable once written. |
-| **Thread** | A single execution of a workflow, identified by a ULID. Persists state as JSONL files. |
-| **CAS** | Content-Addressable Storage. Per-thread key-value store keyed by content hash. |
+| **Thread** | A single execution of a workflow, identified by a ULID. CAS state chain; \`threads.json\` for active; \`history/*.jsonl\` when done; \`.info.jsonl\` for debug logs. |
+| **CAS** | Global content-addressable blob store (\`cas/\`), keyed by hash. |
 | **Registry** | \`workflow.yaml\` — maps workflow names to their current and historical bundle hashes. |
 
 ## Commands
@@ -85,6 +85,12 @@ ${commandSections.join("\n\n")}
 | \`run\` | \`thread run\` | Shortcut to start a thread |
 | \`live\` | \`thread live\` | Shortcut to attach to a thread |
 
+### serve
+
+| Command | Args | Description |
+|---------|------|-------------|
+| \`serve\` | \`[--port N] [--host ADDR] [--name NAME]\` | Start HTTP API server with auto-tunnel. \`--name\` registers with the gateway. |
+
 ## Typical Workflow
 
 1. \`uncaged-workflow workflow add my-wf ./my-wf.esm.js\` — register a workflow
@@ -92,6 +98,15 @@ ${commandSections.join("\n\n")}
 3. \`uncaged-workflow live --latest\` — attach and watch output
 4. \`uncaged-workflow thread show <thread-id>\` — inspect completed thread
 
+## Thread Status
+
+| Status | Meaning |
+|--------|---------|
+| \`running\` | Worker process is alive (\`.running\` marker + live PID) |
+| \`active\` | In \`threads.json\` but not currently running (paused or waiting) |
+| \`completed\` | Finished with \`returnCode === 0\` (has \`__end__\` frame in CAS) |
+| \`failed\` | Finished with non-zero return code, or worker crashed (dead PID / no ctl) |
+
 ## Exit Codes
 
 | Code | Meaning |
@@ -103,7 +118,9 @@ ${commandSections.join("\n\n")}
 
 | Variable | Description |
 |----------|-------------|
-| \`UNCAGED_WORKFLOW_STORAGE_ROOT\` | Override the default storage directory for all workflow data |
+| \`WORKFLOW_STORAGE_ROOT\` | Override the default storage directory for all workflow data |
+| \`UNCAGED_WORKFLOW_STORAGE_ROOT\` | Same as above (takes priority) |
+| \`WORKFLOW_LLM_API_KEY\` | API key for LLM calls during workflow execution |
 `;
 }
 
@@ -172,25 +189,28 @@ export const run: WorkflowRun;
 
 ## WorkflowDescriptor
 
-Defines the workflow's metadata and role sequence:
+Serialized metadata for the registry (per-role JSON Schema plus a static routing graph):
 
 \`\`\`typescript
 type WorkflowDescriptor = {
-  name: string;           // verb-first kebab-case, e.g. "solve-issue"
-  description: string;    // one-line summary
-  roles: string[];        // ordered role names, e.g. ["planner", "coder", "reviewer"]
+  description: string;
+  roles: Record<string, { description: string; schema: unknown /* JSON Schema */ }>;
+  graph: {
+    edges: Array<{
+      from: string;
+      to: string;
+      condition: string;
+      conditionDescription: string | null;
+    }>;
+  };
 };
 \`\`\`
 
 ## WorkflowRun
 
-The main function that creates and returns a moderator:
+Async generator from \`createWorkflow(definition, binding)\` (**@uncaged/workflow-runtime**) — yields each role output until the workflow completes.
 
-\`\`\`typescript
-type WorkflowRun = (ctx: WorkflowContext) => Moderator;
-\`\`\`
-
-The **Moderator** controls the flow — it decides which role runs next, handles retries, and determines when the workflow is complete.
+The **ModeratorTable** on **WorkflowDefinition** is declarative routing (from each role and \`START\` to the next role or \`END\`); the engine evaluates conditions at runtime.
 
 ## Role Definition
 
@@ -200,7 +220,6 @@ Each role has:
 |-------|------|---------|
 | \`description\` | string | What the role does |
 | \`systemPrompt\` | string | System prompt for the agent |
-| \`extractPrompt\` | string | Instruction for extracting structured meta |
 | \`schema\` | ZodSchema | Validates the extracted meta |
 | \`extractRefs\` | fn or null | Extracts CAS hashes from meta for DAG linking |
 
@@ -210,7 +229,7 @@ Each role has:
 # 1. Initialize a workspace
 uncaged-workflow init workspace my-workflow
 
-# 2. Write your template (roles + moderator + descriptor)
+# 2. Write your template (roles + ModeratorTable + descriptor)
 
 # 3. Build the ESM bundle
 bun run build

packages/cli-workflow/src/thread-scan.ts

@@ -1,23 +1,90 @@
 import { readdir, stat } from "node:fs/promises";
 import { join } from "node:path";
+import { createCasStore, parseCasThreadNode } from "@uncaged/workflow-cas";
+import {
+  readThreadsIndex,
+  type ThreadHistoryEntry,
+  type ThreadIndex,
+  walkStateFramesNewestFirst,
+} from "@uncaged/workflow-execute";
+import { END } from "@uncaged/workflow-runtime";
+import { getGlobalCasDir } from "@uncaged/workflow-util";
 
 import { pathExists, readTextFileIfExists } from "./fs-utils.js";
+import { readWorkerCtl } from "./worker-spawn.js";
 
-function parseFirstJsonLineObject(text: string): Record<string, unknown> | null {
-  const firstLine = text.split("\n")[0];
-  if (firstLine === undefined || firstLine.trim() === "") {
+async function readWorkflowNameFromStartHash(
+  storageRoot: string,
+  startHash: string,
+): Promise<string | null> {
+  const cas = createCasStore(getGlobalCasDir(storageRoot));
+  const yamlText = await cas.get(startHash);
+  if (yamlText === null) {
     return null;
   }
-  let parsed: unknown;
-  try {
-    parsed = JSON.parse(firstLine) as unknown;
-  } catch {
+  const parsed = parseCasThreadNode(yamlText);
+  if (parsed === null || parsed.kind !== "start") {
     return null;
   }
-  if (parsed === null || typeof parsed !== "object") {
-    return null;
+  return parsed.node.payload.name;
+}
+
+async function listBundleHashDirs(storageRoot: string): Promise<string[]> {
+  const bundlesRoot = join(storageRoot, "bundles");
+  if (!(await pathExists(bundlesRoot))) {
+    return [];
   }
-  return parsed as Record<string, unknown>;
+  const names = await readdir(bundlesRoot);
+  const out: string[] = [];
+  for (const name of names) {
+    const p = join(bundlesRoot, name);
+    try {
+      const st = await stat(p);
+      if (st.isDirectory()) {
+        out.push(name);
+      }
+    } catch {}
+  }
+  out.sort();
+  return out;
+}
+
+async function parseHistoryFile(path: string): Promise<ThreadHistoryEntry[]> {
+  const text = await readTextFileIfExists(path);
+  if (text === null) {
+    return [];
+  }
+  const out: ThreadHistoryEntry[] = [];
+  for (const line of text.split("\n")) {
+    const trimmed = line.trim();
+    if (trimmed === "") {
+      continue;
+    }
+    let raw: unknown;
+    try {
+      raw = JSON.parse(trimmed) as unknown;
+    } catch {
+      continue;
+    }
+    if (raw === null || typeof raw !== "object") {
+      continue;
+    }
+    const rec = raw as Record<string, unknown>;
+    const threadId = rec.threadId;
+    const head = rec.head;
+    const start = rec.start;
+    const completedAt = rec.completedAt;
+    if (
+      typeof threadId !== "string" ||
+      typeof head !== "string" ||
+      typeof start !== "string" ||
+      typeof completedAt !== "number"
+    ) {
+      continue;
+    }
+    out.push({ threadId, head, start, completedAt });
+  }
+  return out;
 }
 
 export type RunningThreadRow = {
@@ -30,32 +97,173 @@ export type HistoricalThreadRow = {
   threadId: string;
   hash: string;
   workflowName: string | null;
+  /** Active entry from `threads.json` vs completed line from `history/*.jsonl`. */
+  source: "active" | "history";
+  /** `updatedAt` for active threads; `completedAt` for history (ms since epoch). */
+  activityTs: number;
+  /** Current CAS head (`threads.json` / history row). */
+  head: string;
 };
 
-async function readThreadStartTimestampMs(dataPath: string): Promise<number | null> {
-  const text = await readTextFileIfExists(dataPath);
-  if (text === null) {
-    return null;
+export type ResolvedThreadRecord = {
+  threadId: string;
+  bundleHash: string;
+  bundleDir: string;
+  head: string;
+  start: string;
+  source: "active" | "history";
+};
+
+/** Resolve a thread via `threads.json` (active) or `history/*.jsonl` (completed). */
+// biome-ignore lint/complexity/noExcessiveCognitiveComplexity: scans all bundle dirs for thread id
+export async function resolveThreadRecord(
+  storageRoot: string,
+  threadId: string,
+): Promise<ResolvedThreadRecord | null> {
+  const hashes = await listBundleHashDirs(storageRoot);
+  for (const bundleHash of hashes) {
+    const bundleDir = join(storageRoot, "bundles", bundleHash);
+    let index: ThreadIndex;
+    try {
+      index = await readThreadsIndex(bundleDir);
+    } catch {
+      continue;
+    }
+    const active = index[threadId];
+    if (active !== undefined) {
+      return {
+        threadId,
+        bundleHash,
+        bundleDir,
+        head: active.head,
+        start: active.start,
+        source: "active",
+      };
+    }
   }
-  const parsed = parseFirstJsonLineObject(text);
-  if (parsed === null) {
-    return null;
+
+  for (const bundleHash of hashes) {
+    const bundleDir = join(storageRoot, "bundles", bundleHash);
+    const histDir = join(bundleDir, "history");
+    if (!(await pathExists(histDir))) {
+      continue;
+    }
+    let files: string[];
+    try {
+      files = await readdir(histDir);
+    } catch {
+      continue;
+    }
+    for (const name of files) {
+      if (!name.endsWith(".jsonl")) {
+        continue;
+      }
+      const entries = await parseHistoryFile(join(histDir, name));
+      for (const e of entries) {
+        if (e.threadId === threadId) {
+          return {
+            threadId,
+            bundleHash,
+            bundleDir,
+            head: e.head,
+            start: e.start,
+            source: "history",
+          };
+        }
+      }
+    }
   }
-  const ts = parsed.timestamp;
-  return typeof ts === "number" && Number.isFinite(ts) ? ts : null;
+
+  return null;
 }
 
-async function readWorkflowNameFromDataJsonl(dataPath: string): Promise<string | null> {
-  const text = await readTextFileIfExists(dataPath);
-  if (text === null) {
-    return null;
+export type ThreadHeadTerminal =
+  | { kind: "non-terminal" }
+  | { kind: "terminal"; returnCode: number };
+
+/** True when the newest frame at `headHash` is `__end__` (workflow finished in CAS). */
+export async function readThreadTerminalFromHead(
+  storageRoot: string,
+  headHash: string,
+): Promise<ThreadHeadTerminal> {
+  const cas = createCasStore(getGlobalCasDir(storageRoot));
+  const frames = await walkStateFramesNewestFirst(cas, headHash);
+  const newest = frames[0];
+  if (newest === undefined) {
+    return { kind: "non-terminal" };
   }
-  const parsed = parseFirstJsonLineObject(text);
-  if (parsed === null) {
-    return null;
+  if (newest.payload.role !== END) {
+    return { kind: "non-terminal" };
   }
-  const name = parsed.name;
-  return typeof name === "string" ? name : null;
+  const rc = newest.payload.meta.returnCode;
+  if (typeof rc !== "number") {
+    return { kind: "terminal", returnCode: 1 };
+  }
+  return { kind: "terminal", returnCode: rc };
+}
+
+export type ThreadListStatus = "running" | "active" | "completed" | "failed";
+
+/** Combines `.running` marker with CAS head: stale markers do not imply `running`. */
+export async function resolveThreadListStatus(
+  storageRoot: string,
+  row: HistoricalThreadRow,
+  runningMarkerPresent: boolean,
+): Promise<ThreadListStatus> {
+  const terminal = await readThreadTerminalFromHead(storageRoot, row.head);
+  if (terminal.kind === "terminal") {
+    return terminal.returnCode !== 0 ? "failed" : "completed";
+  }
+  if (row.source === "history") {
+    return "completed";
+  }
+  if (runningMarkerPresent) {
+    const ctlResult = await readWorkerCtl(storageRoot, row.hash);
+    if (ctlResult.ok) {
+      try {
+        process.kill(ctlResult.value.pid, 0);
+        return "running";
+      } catch {
+        // Worker PID is dead but .running marker remains — crashed thread
+        return "failed";
+      }
+    }
+    return "running";
+  }
+  // No .running marker + no __end__ + source "active" → check if worker is dead (crashed)
+  const ctlResult = await readWorkerCtl(storageRoot, row.hash);
+  if (!ctlResult.ok) {
+    // No ctl file means worker never registered or was already cleaned up — dead thread
+    return "failed";
+  }
+  try {
+    process.kill(ctlResult.value.pid, 0);
+  } catch {
+    // Worker PID is dead, thread never finished — crashed
+    return "failed";
+  }
+  return "active";
+}
+
+async function appendRunningThreadRowIfLive(
+  storageRoot: string,
+  hash: string,
+  threadId: string,
+  out: RunningThreadRow[],
+): Promise<void> {
+  const resolved = await resolveThreadRecord(storageRoot, threadId);
+  if (resolved !== null && resolved.bundleHash !== hash) {
+    return;
+  }
+  if (resolved !== null) {
+    const terminal = await readThreadTerminalFromHead(storageRoot, resolved.head);
+    if (terminal.kind === "terminal") {
+      return;
+    }
+  }
+  const workflowName =
+    resolved !== null ? await readWorkflowNameFromStartHash(storageRoot, resolved.start) : null;
+  out.push({ threadId, hash, workflowName });
 }
 
 /** Threads currently executing — identified via `<threadId>.running` markers. */
@@ -82,9 +290,7 @@ export async function listRunningThreads(storageRoot: string): Promise<RunningTh
         continue;
       }
       const threadId = fileName.slice(0, -".running".length);
-      const dataPath = join(dir, `${threadId}.data.jsonl`);
-      const workflowName = await readWorkflowNameFromDataJsonl(dataPath);
-      out.push({ threadId, hash, workflowName });
+      await appendRunningThreadRowIfLive(storageRoot, hash, threadId, out);
     }
   }
 
@@ -98,41 +304,84 @@ export async function listRunningThreads(storageRoot: string): Promise<RunningTh
 }
 
 /**
- * Historical threads discovered via `*.data.jsonl`.
- * When `workflowNameFilter` is non-null, only threads whose start record `name` matches are returned.
+ * Threads discovered via `threads.json` (active) and `history/*.jsonl` (completed).
+ * When `workflowNameFilter` is non-null, only threads whose StartNode `name` matches are returned.
  */
+// biome-ignore lint/complexity/noExcessiveCognitiveComplexity: merges active index + partitioned history
 export async function listHistoricalThreads(
   storageRoot: string,
   workflowNameFilter: string | null,
 ): Promise<HistoricalThreadRow[]> {
-  const logsRoot = join(storageRoot, "logs");
-  if (!(await pathExists(logsRoot))) {
-    return [];
-  }
-
-  const hashes = await readdir(logsRoot);
+  const hashes = await listBundleHashDirs(storageRoot);
+  const seen = new Set<string>();
   const out: HistoricalThreadRow[] = [];
 
-  for (const hash of hashes) {
-    const dir = join(logsRoot, hash);
-    let entries: string[];
+  for (const bundleHash of hashes) {
+    const bundleDir = join(storageRoot, "bundles", bundleHash);
+    let index: ThreadIndex;
     try {
-      entries = await readdir(dir);
+      index = await readThreadsIndex(bundleDir);
     } catch {
       continue;
     }
-
-    for (const fileName of entries) {
-      if (!fileName.endsWith(".data.jsonl")) {
+    for (const threadId of Object.keys(index)) {
+      const key = `${bundleHash}/${threadId}`;
+      if (seen.has(key)) {
         continue;
       }
-      const threadId = fileName.slice(0, -".data.jsonl".length);
-      const dataPath = join(dir, fileName);
-      const workflowName = await readWorkflowNameFromDataJsonl(dataPath);
+      seen.add(key);
+      const entry = index[threadId];
+      if (entry === undefined) {
+        continue;
+      }
+      const workflowName = await readWorkflowNameFromStartHash(storageRoot, entry.start);
       if (workflowNameFilter !== null && workflowName !== workflowNameFilter) {
         continue;
       }
-      out.push({ threadId, hash, workflowName });
+      out.push({
+        threadId,
+        hash: bundleHash,
+        workflowName,
+        source: "active",
+        activityTs: entry.updatedAt,
+        head: entry.head,
+      });
+    }
+
+    const histDir = join(bundleDir, "history");
+    if (!(await pathExists(histDir))) {
+      continue;
+    }
+    let files: string[];
+    try {
+      files = await readdir(histDir);
+    } catch {
+      continue;
+    }
+    for (const name of files) {
+      if (!name.endsWith(".jsonl")) {
+        continue;
+      }
+      const entries = await parseHistoryFile(join(histDir, name));
+      for (const e of entries) {
+        const key = `${bundleHash}/${e.threadId}`;
+        if (seen.has(key)) {
+          continue;
+        }
+        seen.add(key);
+        const workflowName = await readWorkflowNameFromStartHash(storageRoot, e.start);
+        if (workflowNameFilter !== null && workflowName !== workflowNameFilter) {
+          continue;
+        }
+        out.push({
+          threadId: e.threadId,
+          hash: bundleHash,
+          workflowName,
+          source: "history",
+          activityTs: e.completedAt,
+          head: e.head,
+        });
+      }
     }
   }
 
@@ -145,64 +394,93 @@ export async function listHistoricalThreads(
   return out;
 }
 
+export type LatestThreadTarget = {
+  threadId: string;
+  bundleHash: string;
+  bundleDir: string;
+  threadsJsonPath: string;
+};
+
 /**
- * Picks the thread whose `.data.jsonl` is newest by start-record `timestamp`,
- * falling back to file `mtime` when the timestamp is missing.
- * Tie-breaker: larger `mtime` wins when start timestamps are equal.
+ * Picks the newest thread by StartNode timestamp approximation (`updatedAt` active,
+ * else `completedAt` history), falling back to lexical thread id order.
  */
-export async function findLatestThreadDataPath(
+// biome-ignore lint/complexity/noExcessiveCognitiveComplexity: compares active heads vs history tails
+export async function findLatestThreadBundleTarget(
   storageRoot: string,
-): Promise<{ threadId: string; dataPath: string } | null> {
-  const threads = await listHistoricalThreads(storageRoot, null);
-  if (threads.length === 0) {
-    return null;
-  }
+): Promise<LatestThreadTarget | null> {
+  const hashes = await listBundleHashDirs(storageRoot);
 
   let best: {
     threadId: string;
-    dataPath: string;
-    primary: number;
-    secondary: number;
+    bundleHash: string;
+    bundleDir: string;
+    ts: number;
   } | null = null;
 
-  for (const t of threads) {
-    const dataPath = join(storageRoot, "logs", t.hash, `${t.threadId}.data.jsonl`);
-    let mtimeMs = 0;
+  for (const bundleHash of hashes) {
+    const bundleDir = join(storageRoot, "bundles", bundleHash);
+    let index: ThreadIndex;
     try {
-      const st = await stat(dataPath);
-      mtimeMs = st.mtimeMs;
+      index = await readThreadsIndex(bundleDir);
     } catch {
       continue;
     }
-    const startTs = await readThreadStartTimestampMs(dataPath);
-    const primary = startTs !== null ? startTs : mtimeMs;
-    const secondary = mtimeMs;
-    if (
-      best === null ||
-      primary > best.primary ||
-      (primary === best.primary && secondary > best.secondary)
-    ) {
-      best = { threadId: t.threadId, dataPath, primary, secondary };
+    for (const threadId of Object.keys(index)) {
+      const ent = index[threadId];
+      if (ent === undefined) {
+        continue;
+      }
+      const ts = ent.updatedAt;
+      const cand = { threadId, bundleHash, bundleDir, ts };
+      if (
+        best === null ||
+        cand.ts > best.ts ||
+        (cand.ts === best.ts &&
+          `${cand.bundleHash}/${cand.threadId}` > `${best.bundleHash}/${best.threadId}`)
+      ) {
+        best = cand;
+      }
+    }
+
+    const histDir = join(bundleDir, "history");
+    if (!(await pathExists(histDir))) {
+      continue;
+    }
+    let files: string[];
+    try {
+      files = await readdir(histDir);
+    } catch {
+      continue;
+    }
+    for (const name of files) {
+      if (!name.endsWith(".jsonl")) {
+        continue;
+      }
+      const entries = await parseHistoryFile(join(histDir, name));
+      for (const e of entries) {
+        const ts = e.completedAt;
+        const cand = { threadId: e.threadId, bundleHash, bundleDir, ts };
+        if (
+          best === null ||
+          cand.ts > best.ts ||
+          (cand.ts === best.ts &&
+            `${cand.bundleHash}/${cand.threadId}` > `${best.bundleHash}/${best.threadId}`)
+        ) {
+          best = cand;
+        }
+      }
     }
   }
 
-  return best === null ? null : { threadId: best.threadId, dataPath: best.dataPath };
-}
-
-export async function resolveThreadDataPath(
-  storageRoot: string,
-  threadId: string,
-): Promise<string | null> {
-  const logsRoot = join(storageRoot, "logs");
-  if (!(await pathExists(logsRoot))) {
+  if (best === null) {
     return null;
   }
-  const hashes = await readdir(logsRoot);
-  for (const hash of hashes) {
-    const candidate = join(logsRoot, hash, `${threadId}.data.jsonl`);
-    if (await pathExists(candidate)) {
-      return candidate;
-    }
-  }
-  return null;
+
+  return {
+    threadId: best.threadId,
+    bundleHash: best.bundleHash,
+    bundleDir: best.bundleDir,
+    threadsJsonPath: join(best.bundleDir, "threads.json"),
+  };
 }

packages/cli-workflow/src/worker-spawn.ts

@@ -2,11 +2,11 @@ import { type ChildProcess, spawn } from "node:child_process";
 import { mkdir, readdir, unlink, writeFile } from "node:fs/promises";
 import { createConnection } from "node:net";
 import { join } from "node:path";
-
-import { err, ok, type Result } from "@uncaged/workflow-protocol";
 import { getWorkerHostScriptPath } from "@uncaged/workflow-execute";
+import { err, ok, type Result } from "@uncaged/workflow-protocol";
 
 import { pathExists, readTextFileIfExists } from "./fs-utils.js";
+import { readThreadTerminalFromHead, resolveThreadRecord } from "./thread-scan.js";
 
 export type WorkerCtl = {
   pid: number;
@@ -270,7 +270,25 @@ export async function resolveRunningHashForThread(
   if (!(await pathExists(logsRoot))) {
     return err(`thread not running (no logs dir): ${threadId}`);
   }
-  const hashes = await readdir(logsRoot);
+  const resolved = await resolveThreadRecord(storageRoot, threadId);
+  if (resolved !== null) {
+    const runningPath = join(logsRoot, resolved.bundleHash, `${threadId}.running`);
+    if (!(await pathExists(runningPath))) {
+      return err(`thread not running: ${threadId}`);
+    }
+    const terminal = await readThreadTerminalFromHead(storageRoot, resolved.head);
+    if (terminal.kind === "terminal") {
+      return err(`thread not running: ${threadId}`);
+    }
+    return ok(resolved.bundleHash);
+  }
+
+  let hashes: string[];
+  try {
+    hashes = await readdir(logsRoot);
+  } catch {
+    return err(`thread not running: ${threadId}`);
+  }
   for (const hash of hashes) {
     const runningPath = join(logsRoot, hash, `${threadId}.running`);
     if (await pathExists(runningPath)) {

packages/workflow-agent-cursor/__tests__/cursor-agent.test.ts

@@ -1,33 +1,50 @@
 import { describe, expect, test } from "bun:test";
-import type { ExtractContext, ExtractFn } from "@uncaged/workflow-runtime";
-import type * as z from "zod/v4";
 import { createCursorAgent, validateCursorAgentConfig } from "../src/index.js";
 
-const testExtract: ExtractFn = async <T extends Record<string, unknown>>(
-  _schema: z.ZodType<T>,
-  _prompt: string,
-  _ctx: ExtractContext,
-): Promise<T> => ({ workspace: "/tmp" }) as unknown as T;
-
 describe("validateCursorAgentConfig", () => {
-  test("accepts valid config", () => {
+  test("accepts valid config with explicit workspace", () => {
     const r = validateCursorAgentConfig({
       model: null,
       timeout: 0,
-      extract: testExtract,
+      workspace: "/tmp/test-project",
+      llmProvider: null,
     });
     expect(r.ok).toBe(true);
   });
 
-  test("rejects non-function extract", () => {
+  test("accepts valid config with null workspace and llmProvider", () => {
     const r = validateCursorAgentConfig({
       model: null,
       timeout: 0,
-      extract: null as unknown as ExtractFn,
+      workspace: null,
+      llmProvider: { baseUrl: "http://localhost", apiKey: "test", model: "test" },
+    });
+    expect(r.ok).toBe(true);
+  });
+
+  test("rejects empty workspace string", () => {
+    const r = validateCursorAgentConfig({
+      model: null,
+      timeout: 0,
+      workspace: "",
+      llmProvider: null,
     });
     expect(r.ok).toBe(false);
     if (!r.ok) {
-      expect(r.error).toContain("extract");
+      expect(r.error).toContain("workspace");
+    }
+  });
+
+  test("rejects null workspace without llmProvider", () => {
+    const r = validateCursorAgentConfig({
+      model: null,
+      timeout: 0,
+      workspace: null,
+      llmProvider: null,
+    });
+    expect(r.ok).toBe(false);
+    if (!r.ok) {
+      expect(r.error).toContain("llmProvider");
     }
   });
 
@@ -35,18 +52,30 @@ describe("validateCursorAgentConfig", () => {
     const r = validateCursorAgentConfig({
       model: null,
       timeout: -1,
-      extract: testExtract,
+      workspace: "/tmp/test-project",
+      llmProvider: null,
     });
     expect(r.ok).toBe(false);
   });
 });
 
 describe("createCursorAgent", () => {
-  test("returns an AgentFn", () => {
+  test("returns an AgentFn with explicit workspace", () => {
     const agent = createCursorAgent({
       model: null,
       timeout: 0,
-      extract: testExtract,
+      workspace: "/tmp/test-project",
+      llmProvider: null,
+    });
+    expect(typeof agent).toBe("function");
+  });
+
+  test("returns an AgentFn with null workspace and llmProvider", () => {
+    const agent = createCursorAgent({
+      model: null,
+      timeout: 0,
+      workspace: null,
+      llmProvider: { baseUrl: "http://localhost", apiKey: "test", model: "test" },
     });
     expect(typeof agent).toBe("function");
   });
@@ -56,7 +85,19 @@ describe("createCursorAgent", () => {
       createCursorAgent({
         model: null,
         timeout: -1,
-        extract: testExtract,
+        workspace: "/tmp/test-project",
+        llmProvider: null,
+      }),
+    ).toThrow();
+  });
+
+  test("throws when null workspace without llmProvider", () => {
+    expect(() =>
+      createCursorAgent({
+        model: null,
+        timeout: 0,
+        workspace: null,
+        llmProvider: null,
       }),
     ).toThrow();
   });

packages/workflow-agent-cursor/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@uncaged/workflow-agent-cursor",
-  "version": "0.2.0",
+  "version": "0.3.1",
   "type": "module",
   "main": "src/index.ts",
   "types": "src/index.ts",
@@ -8,7 +8,10 @@
     "test": "bun test"
   },
   "dependencies": {
+    "@uncaged/workflow-protocol": "workspace:*",
+    "@uncaged/workflow-reactor": "workspace:*",
     "@uncaged/workflow-runtime": "workspace:*",
+    "@uncaged/workflow-util": "workspace:*",
     "@uncaged/workflow-util-agent": "workspace:*",
     "zod": "^4.0.0"
   }

packages/workflow-agent-cursor/pnpm-lock.yaml

@@ -0,0 +1,28 @@
+lockfileVersion: '9.0'
+
+settings:
+  autoInstallPeers: true
+  excludeLinksFromLockfile: false
+
+importers:
+
+  .:
+    dependencies:
+      '@uncaged/workflow-runtime':
+        specifier: workspace:*
+        version: link:../workflow-runtime
+      '@uncaged/workflow-util-agent':
+        specifier: workspace:*
+        version: link:../workflow-util-agent
+      zod:
+        specifier: ^4.0.0
+        version: 4.4.3
+
+packages:
+
+  zod@4.4.3:
+    resolution: {integrity: sha512-ytENFjIJFl2UwYglde2jchW2Hwm4GJFLDiSXWdTrJQBIN9Fcyp7n4DhxJEiWNAJMV1/BqWfW/kkg71UDcHJyTQ==}
+
+snapshots:
+
+  zod@4.4.3: {}

packages/workflow-agent-cursor/src/extract-workspace.ts

@@ -0,0 +1,73 @@
+import type { AgentContext, LlmProvider } from "@uncaged/workflow-protocol";
+import { createLlmFn, createThreadReactor } from "@uncaged/workflow-reactor";
+import type { LogFn } from "@uncaged/workflow-util";
+import * as z from "zod/v4";
+
+const workspaceSchema = z.object({
+  workspace: z.string().describe("Absolute filesystem path of the project workspace"),
+});
+
+const EXTRACT_SYSTEM_FN = (_toolName: string) =>
+  `You are a workspace-path extractor. Given a workflow agent context (task description and previous step outputs), identify the absolute filesystem path of the project workspace where code changes should be made. Call the tool with the absolute path.`;
+
+function buildExtractionInput(ctx: AgentContext): string {
+  const lines: string[] = [];
+  lines.push("## Task");
+  lines.push(ctx.start.content);
+
+  for (const step of ctx.steps) {
+    lines.push("");
+    lines.push(`## Step: ${step.role}`);
+    lines.push(`Meta: ${JSON.stringify(step.meta)}`);
+  }
+
+  return lines.join("\n");
+}
+
+export async function extractWorkspacePath(
+  ctx: AgentContext,
+  provider: LlmProvider,
+  logger: LogFn,
+): Promise<string | null> {
+  const reactor = createThreadReactor<null>({
+    llm: createLlmFn(provider),
+    maxRounds: 2,
+    staticTools: [],
+    structuredToolFromSchema: (schema) => {
+      const jsonSchema = z.toJSONSchema(schema);
+      return {
+        name: "set_workspace",
+        tool: {
+          type: "function" as const,
+          function: {
+            name: "set_workspace",
+            description: "Set the extracted workspace path",
+            parameters: jsonSchema as Record<string, unknown>,
+          },
+        },
+      };
+    },
+    systemPromptForStructuredTool: EXTRACT_SYSTEM_FN,
+    toolHandler: async () => "unknown tool",
+  });
+
+  const result = await reactor({
+    thread: null,
+    input: buildExtractionInput(ctx),
+    schema: workspaceSchema,
+  });
+
+  if (!result.ok) {
+    logger("W8KN3QYT", `workspace extraction failed: ${result.error}`);
+    return null;
+  }
+
+  const workspace = result.value.workspace.trim();
+  if (!workspace.startsWith("/")) {
+    logger("H4PM7RXV", `workspace extraction returned non-absolute path: ${workspace}`);
+    return null;
+  }
+
+  logger("V3KM8QWP", `extracted workspace: ${workspace}`);
+  return workspace;
+}

packages/workflow-agent-cursor/src/index.ts

@@ -1,19 +1,14 @@
-import type { AgentFn, ExtractContext } from "@uncaged/workflow-runtime";
+import type { AgentFn } from "@uncaged/workflow-runtime";
+import { createLogger } from "@uncaged/workflow-util";
 import { buildAgentPrompt, type SpawnCliError, spawnCli } from "@uncaged/workflow-util-agent";
-import * as z from "zod/v4";
 
+import { extractWorkspacePath } from "./extract-workspace.js";
 import type { CursorAgentConfig } from "./types.js";
 import { validateCursorAgentConfig } from "./validate-config.js";
 
 export type { CursorAgentConfig } from "./types.js";
 export { validateCursorAgentConfig } from "./validate-config.js";
 
-const cursorWorkspaceSchema = z.object({
-  workspace: z
-    .string()
-    .describe("Absolute path to the project/repository directory the agent should work in"),
-});
-
 function throwCursorSpawnError(error: SpawnCliError): never {
   if (error.kind === "non_zero_exit") {
     throw new Error(
@@ -33,7 +28,7 @@ function resolveCursorModel(model: string | null): string {
   return model === null ? "auto" : model;
 }
 
-/** Runs `cursor-agent` with workspace from {@link CursorAgentConfig.extract} and prompt from context. */
+/** Runs `cursor-agent` with workspace from config or extracted from context via LLM. */
 export function createCursorAgent(config: CursorAgentConfig): AgentFn {
   const validated = validateCursorAgentConfig(config);
   if (!validated.ok) {
@@ -42,17 +37,27 @@ export function createCursorAgent(config: CursorAgentConfig): AgentFn {
 
   const modelFlag = resolveCursorModel(config.model);
   const timeoutMs = config.timeout > 0 ? config.timeout : null;
+  const logger = createLogger({ sink: { kind: "stderr" } });
 
   return async (ctx) => {
-    const extractCtx: ExtractContext = {
-      ...ctx,
-      agentContent: "",
-    };
-    const { workspace } = await config.extract(
-      cursorWorkspaceSchema,
-      "From the thread context, determine the absolute filesystem path where the project/repository is located.",
-      extractCtx,
-    );
+    let workspace: string;
+
+    if (config.workspace !== null) {
+      workspace = config.workspace;
+    } else {
+      if (config.llmProvider === null) {
+        throw new Error("cursor-agent: llmProvider is required when workspace is null");
+      }
+      const extracted = await extractWorkspacePath(ctx, config.llmProvider, logger);
+      if (extracted === null) {
+        throw new Error(
+          "cursor-agent: failed to extract workspace path from context. Provide an explicit workspace or ensure previous steps include a repoPath.",
+        );
+      }
+      workspace = extracted;
+    }
+
+    logger("R5HN3YKQ", `cursor-agent workspace: ${workspace}`);
     const fullPrompt = await buildAgentPrompt(ctx);
     const args = [
       "-p",

packages/workflow-agent-cursor/src/types.ts

@@ -1,7 +1,10 @@
-import type { ExtractFn } from "@uncaged/workflow-runtime";
+import type { LlmProvider } from "@uncaged/workflow-protocol";
 
 export type CursorAgentConfig = {
   model: string | null;
   timeout: number;
-  extract: ExtractFn;
+  /** Explicit workspace path. When `null`, the agent extracts workspace from AgentContext via a ReAct LLM call. */
+  workspace: string | null;
+  /** Required when `workspace` is `null` — LLM provider used for workspace extraction. */
+  llmProvider: LlmProvider | null;
 };

packages/workflow-agent-cursor/src/validate-config.ts

@@ -1,10 +1,13 @@
-import { err, ok, type Result } from "@uncaged/workflow-runtime";
+import { err, ok, type Result } from "@uncaged/workflow-protocol";
 
 import type { CursorAgentConfig } from "./types.js";
 
 export function validateCursorAgentConfig(config: CursorAgentConfig): Result<void, string> {
-  if (typeof config.extract !== "function") {
-    return err("extract must be a function");
+  if (config.workspace !== null && config.workspace.length === 0) {
+    return err("workspace must be a non-empty string (absolute path) or null for auto-detection");
+  }
+  if (config.workspace === null && config.llmProvider === null) {
+    return err("llmProvider is required when workspace is null (needed for workspace extraction)");
   }
   if (config.timeout < 0) {
     return err("timeout must be a non-negative number (milliseconds); use 0 for no limit");

packages/workflow-agent-hermes/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@uncaged/workflow-agent-hermes",
-  "version": "0.2.0",
+  "version": "0.3.1",
   "type": "module",
   "main": "src/index.ts",
   "types": "src/index.ts",

packages/workflow-agent-hermes/pnpm-lock.yaml

@@ -0,0 +1,16 @@
+lockfileVersion: '9.0'
+
+settings:
+  autoInstallPeers: true
+  excludeLinksFromLockfile: false
+
+importers:
+
+  .:
+    dependencies:
+      '@uncaged/workflow-runtime':
+        specifier: workspace:*
+        version: link:../workflow-runtime
+      '@uncaged/workflow-util-agent':
+        specifier: workspace:*
+        version: link:../workflow-util-agent

packages/workflow-agent-llm/__tests__/create-llm-adapter.test.ts

@@ -8,10 +8,12 @@ function makeCtx(userContent: string): AgentContext {
     start: {
       role: START,
       content: userContent,
-      meta: { maxRounds: 10 },
+      meta: {},
       timestamp: 1,
+      parentState: null,
     },
     depth: 0,
+    bundleHash: "TESTHASH00001",
     steps: [],
     threadId: "01TEST000000000000000000TR",
     currentRole: { name: "planner", systemPrompt: "system instructions" },

packages/workflow-agent-llm/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@uncaged/workflow-agent-llm",
-  "version": "0.2.0",
+  "version": "0.3.1",
   "type": "module",
   "main": "src/index.ts",
   "types": "src/index.ts",

packages/workflow-agent-llm/pnpm-lock.yaml

@@ -0,0 +1,13 @@
+lockfileVersion: '9.0'
+
+settings:
+  autoInstallPeers: true
+  excludeLinksFromLockfile: false
+
+importers:
+
+  .:
+    dependencies:
+      '@uncaged/workflow-runtime':
+        specifier: workspace:*
+        version: link:../workflow-runtime

packages/workflow-cas/__tests__/collect-refs.test.ts

@@ -14,6 +14,7 @@ function payload(
     ancestors: partial.ancestors ?? [],
     compact: partial.compact ?? null,
     timestamp: partial.timestamp ?? 0,
+    childThread: partial.childThread ?? null,
   };
 }
 
@@ -62,4 +63,32 @@ describe("collectRefs", () => {
     );
     expect(refs).toEqual(["S2", "C2"]);
   });
+
+  test("includes childThread hash when childThread is non-null", () => {
+    const refs = collectRefs(
+      payload({
+        role: "developer",
+        start: "S3",
+        content: "C3",
+        ancestors: ["A3"],
+        compact: null,
+        childThread: "CHILDEND000000000000001",
+      }),
+    );
+    expect(refs).toEqual(["S3", "C3", "A3", "CHILDEND000000000000001"]);
+  });
+
+  test("does not include childThread when childThread is null", () => {
+    const refs = collectRefs(
+      payload({
+        role: "developer",
+        start: "S4",
+        content: "C4",
+        ancestors: [],
+        compact: null,
+        childThread: null,
+      }),
+    );
+    expect(refs).toEqual(["S4", "C4"]);
+  });
 });

packages/workflow-cas/__tests__/nodes.test.ts

@@ -0,0 +1,161 @@
+import { afterEach, beforeEach, describe, expect, test } from "bun:test";
+import { mkdtemp, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { stringify } from "yaml";
+
+import { createCasStore } from "../src/cas.js";
+import { parseCasThreadNode, putStartNode, putStateNode } from "../src/nodes.js";
+
+describe("putStartNode — parentState in refs", () => {
+  let dir: string;
+
+  beforeEach(async () => {
+    dir = await mkdtemp(join(tmpdir(), "wf-cas-nodes-"));
+  });
+
+  afterEach(async () => {
+    await rm(dir, { recursive: true, force: true });
+  });
+
+  test("refs contains only promptHash when parentState is null", async () => {
+    const cas = createCasStore(join(dir, "cas"));
+    const promptHash = await cas.put("hello");
+    const startHash = await putStartNode(
+      cas,
+      { name: "demo", hash: "BUNDLEAAAAAAAAA", depth: 0, parentState: null },
+      promptHash,
+    );
+
+    const blob = await cas.get(startHash);
+    expect(blob).not.toBeNull();
+    const parsed = parseCasThreadNode(blob ?? "");
+    expect(parsed).not.toBeNull();
+    expect(parsed?.kind).toBe("start");
+    if (parsed?.kind !== "start") return;
+
+    expect(parsed.node.refs).toEqual([promptHash]);
+    expect(parsed.node.payload.parentState).toBeNull();
+  });
+
+  test("refs contains [promptHash, parentStateHash] when parentState is set", async () => {
+    const cas = createCasStore(join(dir, "cas"));
+    const parentStateHash = await cas.put("fake-parent-state");
+    const promptHash = await cas.put("child-prompt");
+    const startHash = await putStartNode(
+      cas,
+      { name: "develop", hash: "BUNDLEBBBBBBBBB", depth: 1, parentState: parentStateHash },
+      promptHash,
+    );
+
+    const blob = await cas.get(startHash);
+    expect(blob).not.toBeNull();
+    const parsed = parseCasThreadNode(blob ?? "");
+    expect(parsed).not.toBeNull();
+    expect(parsed?.kind).toBe("start");
+    if (parsed?.kind !== "start") return;
+
+    expect(parsed.node.refs).toEqual([promptHash, parentStateHash]);
+    expect(parsed.node.payload.parentState).toBe(parentStateHash);
+  });
+});
+
+describe("putStateNode — childThread in refs", () => {
+  let dir: string;
+
+  beforeEach(async () => {
+    dir = await mkdtemp(join(tmpdir(), "wf-cas-nodes-state-"));
+  });
+
+  afterEach(async () => {
+    await rm(dir, { recursive: true, force: true });
+  });
+
+  test("refs does not include childThread when childThread is null", async () => {
+    const cas = createCasStore(join(dir, "cas"));
+    const startHash = await cas.put("start");
+    const contentHash = await cas.put("content");
+    const stateHash = await putStateNode(cas, {
+      role: "planner",
+      meta: {},
+      start: startHash,
+      content: contentHash,
+      ancestors: [],
+      compact: null,
+      timestamp: 1000,
+      childThread: null,
+    });
+
+    const blob = await cas.get(stateHash);
+    expect(blob).not.toBeNull();
+    const parsed = parseCasThreadNode(blob ?? "");
+    expect(parsed?.kind).toBe("state");
+    if (parsed?.kind !== "state") return;
+
+    expect(parsed.node.refs).not.toContain("anything-else");
+    expect(parsed.node.refs).toEqual([startHash, contentHash]);
+    expect(parsed.node.payload.childThread).toBeNull();
+  });
+
+  test("refs includes childThread hash when childThread is set", async () => {
+    const cas = createCasStore(join(dir, "cas"));
+    const startHash = await cas.put("start");
+    const contentHash = await cas.put("content");
+    const childEndHash = await cas.put("child-end-state");
+    const stateHash = await putStateNode(cas, {
+      role: "developer",
+      meta: { pr: 42 },
+      start: startHash,
+      content: contentHash,
+      ancestors: [],
+      compact: null,
+      timestamp: 2000,
+      childThread: childEndHash,
+    });
+
+    const blob = await cas.get(stateHash);
+    expect(blob).not.toBeNull();
+    const parsed = parseCasThreadNode(blob ?? "");
+    expect(parsed?.kind).toBe("state");
+    if (parsed?.kind !== "state") return;
+
+    expect(parsed.node.refs).toContain(childEndHash);
+    expect(parsed.node.payload.childThread).toBe(childEndHash);
+  });
+});
+
+describe("parseCasThreadNode — legacy node compatibility", () => {
+  test("start node without parentState field defaults to null", () => {
+    const yaml = stringify({
+      type: "start",
+      payload: { name: "demo", hash: "BUNDLEAAAAAAAAA", depth: 0 },
+      refs: ["PROMPTHASH00001"],
+    });
+    const parsed = parseCasThreadNode(yaml);
+    expect(parsed).not.toBeNull();
+    expect(parsed?.kind).toBe("start");
+    if (parsed?.kind !== "start") return;
+    expect(parsed.node.payload.parentState).toBeNull();
+  });
+
+  test("state node without childThread field defaults to null", () => {
+    const yaml = stringify({
+      type: "state",
+      payload: {
+        role: "planner",
+        meta: {},
+        start: "STARTHASH00001",
+        content: "CONTENTHASH0001",
+        ancestors: [],
+        compact: null,
+        timestamp: 1000,
+      },
+      refs: ["STARTHASH00001", "CONTENTHASH0001"],
+    });
+    const parsed = parseCasThreadNode(yaml);
+    expect(parsed).not.toBeNull();
+    expect(parsed?.kind).toBe("state");
+    if (parsed?.kind !== "state") return;
+    expect(parsed.node.payload.childThread).toBeNull();
+  });
+});

packages/workflow-cas/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@uncaged/workflow-cas",
-  "version": "0.1.0",
+  "version": "0.3.1",
   "type": "module",
   "scripts": {
     "test": "bun test"

packages/workflow-cas/pnpm-lock.yaml

@@ -0,0 +1,75 @@
+lockfileVersion: '9.0'
+
+settings:
+  autoInstallPeers: true
+  excludeLinksFromLockfile: false
+
+importers:
+
+  .:
+    dependencies:
+      '@uncaged/workflow-protocol':
+        specifier: workspace:*
+        version: link:../workflow-protocol
+      '@uncaged/workflow-util':
+        specifier: workspace:*
+        version: link:../workflow-util
+      xxhashjs:
+        specifier: ^0.2.2
+        version: 0.2.2
+      yaml:
+        specifier: ^2.7.1
+        version: 2.8.4
+    devDependencies:
+      '@types/bun':
+        specifier: latest
+        version: 1.3.13
+
+packages:
+
+  '@types/bun@1.3.13':
+    resolution: {integrity: sha512-9fqXWk5YIHGGnUau9TEi+qdlTYDAnOj+xLCmSTwXfAIqXr2x4tytJb43E9uCvt09zJURKXwAtkoH4nLQfzeTXw==}
+
+  '@types/node@25.6.2':
+    resolution: {integrity: sha512-sokuT28dxf9JT5Kady1fsXOvI4HVpjZa95NKT5y9PNTIrs2AsobR4GFAA90ZG8M+nxVRLysCXsVj6eGC7Vbrlw==}
+
+  bun-types@1.3.13:
+    resolution: {integrity: sha512-QXKeHLlOLqQX9LgYaHJfzdBaV21T63HhFJnvuRCcjZiaUDpbs5ED1MgxbMra71CsryN/1dAoXuJJJwIv/2drVA==}
+
+  cuint@0.2.2:
+    resolution: {integrity: sha512-d4ZVpCW31eWwCMe1YT3ur7mUDnTXbgwyzaL320DrcRT45rfjYxkt5QWLrmOJ+/UEAI2+fQgKe/fCjR8l4TpRgw==}
+
+  undici-types@7.19.2:
+    resolution: {integrity: sha512-qYVnV5OEm2AW8cJMCpdV20CDyaN3g0AjDlOGf1OW4iaDEx8MwdtChUp4zu4H0VP3nDRF/8RKWH+IPp9uW0YGZg==}
+
+  xxhashjs@0.2.2:
+    resolution: {integrity: sha512-AkTuIuVTET12tpsVIQo+ZU6f/qDmKuRUcjaqR+OIvm+aCBsZ95i7UVY5WJ9TMsSaZ0DA2WxoZ4acu0sPH+OKAw==}
+
+  yaml@2.8.4:
+    resolution: {integrity: sha512-ml/JPOj9fOQK8RNnWojA67GbZ0ApXAUlN2UQclwv2eVgTgn7O9gg9o7paZWKMp4g0H3nTLtS9LVzhkpOFIKzog==}
+    engines: {node: '>= 14.6'}
+    hasBin: true
+
+snapshots:
+
+  '@types/bun@1.3.13':
+    dependencies:
+      bun-types: 1.3.13
+
+  '@types/node@25.6.2':
+    dependencies:
+      undici-types: 7.19.2
+
+  bun-types@1.3.13:
+    dependencies:
+      '@types/node': 25.6.2
+
+  cuint@0.2.2: {}
+
+  undici-types@7.19.2: {}
+
+  xxhashjs@0.2.2:
+    dependencies:
+      cuint: 0.2.2
+
+  yaml@2.8.4: {}

packages/workflow-cas/src/collect-refs.ts

@@ -9,5 +9,8 @@ export function collectRefs(payload: StateNode["payload"]): string[] {
   if (payload.compact !== null) {
     out.push(payload.compact);
   }
+  if (payload.childThread !== null) {
+    out.push(payload.childThread);
+  }
   return out;
 }

packages/workflow-cas/src/hash.ts

@@ -1,8 +1,6 @@
 import { Buffer } from "node:buffer";
-
-import XXH from "xxhashjs";
-
 import { encodeUint64AsCrockford } from "@uncaged/workflow-util";
+import XXH from "xxhashjs";
 
 function digestToUint64(digest: { toString(radix?: number): string }): bigint {
   const hex = digest.toString(16).padStart(16, "0");

packages/workflow-cas/src/index.ts

@@ -10,8 +10,10 @@ export {
   putThreadMerkleNode,
   serializeMerkleNode,
 } from "./merkle.js";
+export type { ParsedCasThreadNode } from "./nodes.js";
 export {
   isCasNodeYaml,
+  parseCasThreadNode,
   putContentNodeWithRefs,
   putStartNode,
   putStateNode,

packages/workflow-cas/src/merkle.ts

@@ -1,8 +1,41 @@
 import { parse, stringify } from "yaml";
 
-import type { CasStore, MerkleNode, StepMerklePayload, ThreadMerklePayload } from "./types.js";
+import type {
+  CasStore,
+  MerkleNode,
+  MerkleNodeType,
+  StepMerklePayload,
+  ThreadMerklePayload,
+} from "./types.js";
+
+function requireStringHashArray(value: unknown, notArrayMessage: string): string[] {
+  if (!Array.isArray(value)) {
+    throw new Error(notArrayMessage);
+  }
+  const out: string[] = [];
+  for (const c of value) {
+    if (typeof c !== "string") {
+      throw new Error("merkle: hash entry must be a string");
+    }
+    out.push(c);
+  }
+  return out;
+}
+
+function edgeListRaw(rec: Record<string, unknown>, type: MerkleNodeType): unknown {
+  if (type === "content") {
+    return rec.refs !== undefined ? rec.refs : rec.children;
+  }
+  return rec.children;
+}
 
 export function serializeMerkleNode(node: MerkleNode): string {
+  if (node.type === "content") {
+    return stringify(
+      { type: node.type, payload: node.payload, refs: node.children },
+      { indent: 2 },
+    );
+  }
   return stringify(
     { type: node.type, payload: node.payload, children: node.children },
     { indent: 2 },
@@ -17,23 +50,18 @@ export function parseMerkleNode(yamlText: string): MerkleNode {
   const rec = raw as Record<string, unknown>;
   const type = rec.type;
   const payload = rec.payload;
-  const children = rec.children;
   if (type !== "content" && type !== "step" && type !== "thread") {
     throw new Error("merkle: invalid or missing type");
   }
   if (typeof payload !== "string" && (payload === null || typeof payload !== "object")) {
     throw new Error("merkle: payload must be a string or object");
   }
-  if (!Array.isArray(children)) {
-    throw new Error("merkle: children must be an array");
-  }
-  const childHashes: string[] = [];
-  for (const c of children) {
-    if (typeof c !== "string") {
-      throw new Error("merkle: child hash must be a string");
-    }
-    childHashes.push(c);
-  }
+
+  const notArrayMsg =
+    type === "content"
+      ? "merkle: content node requires refs or children array"
+      : "merkle: children must be an array";
+  const childHashes = requireStringHashArray(edgeListRaw(rec, type), notArrayMsg);
   return {
     type,
     payload: typeof payload === "string" ? payload : (payload as Record<string, unknown>),
@@ -85,8 +113,8 @@ export async function putContentMerkleNode(store: CasStore, content: string): Pr
 /**
  * Loads a CAS blob and returns the payload string for a `content` node.
  *
- * Accepts both the legacy `{type:content, payload, children}` Merkle layout
- * and the RFC v3 `{type:content, payload, refs}` content node layout.
+ * Accepts both the legacy `{ type:content, payload, children }` Merkle layout
+ * and the RFC-aligned `{ type:content, payload, refs }` content node layout.
  */
 export async function getContentMerklePayload(
   store: CasStore,

packages/workflow-cas/src/nodes.ts

@@ -1,9 +1,147 @@
-import type { ContentMerkleNode, StartNode, StateNode } from "@uncaged/workflow-protocol";
+import type {
+  ContentMerkleNode,
+  StartNode,
+  StartNodePayload,
+  StateNode,
+  StateNodePayload,
+} from "@uncaged/workflow-protocol";
 import { parse, stringify } from "yaml";
 
 import { collectRefs } from "./collect-refs.js";
 import type { CasStore } from "./types.js";
 
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
+function isStartPayload(value: unknown): value is StartNodePayload {
+  if (!isRecord(value)) {
+    return false;
+  }
+  const parentState = value.parentState;
+  if (parentState !== undefined && parentState !== null && typeof parentState !== "string") {
+    return false;
+  }
+  return (
+    typeof value.name === "string" &&
+    typeof value.hash === "string" &&
+    typeof value.depth === "number"
+  );
+}
+
+/** Normalizes a raw start payload, defaulting `parentState` to `null` for legacy nodes. */
+function normalizeStartPayload(raw: StartNodePayload): StartNodePayload {
+  return {
+    name: raw.name,
+    hash: raw.hash,
+    depth: raw.depth,
+    parentState: raw.parentState ?? null,
+  };
+}
+
+function isStatePayload(value: unknown): value is StateNodePayload {
+  if (!isRecord(value)) {
+    return false;
+  }
+  const compact = value.compact;
+  if (!(compact === null || typeof compact === "string")) {
+    return false;
+  }
+  const ancestors = value.ancestors;
+  if (!Array.isArray(ancestors) || !ancestors.every((h) => typeof h === "string")) {
+    return false;
+  }
+  const meta = value.meta;
+  if (!isRecord(meta)) {
+    return false;
+  }
+  const childThread = value.childThread;
+  if (childThread !== undefined && childThread !== null && typeof childThread !== "string") {
+    return false;
+  }
+  return (
+    typeof value.role === "string" &&
+    typeof value.start === "string" &&
+    typeof value.content === "string" &&
+    typeof value.timestamp === "number"
+  );
+}
+
+/** Normalizes a raw state payload, defaulting `childThread` to `null` for legacy nodes. */
+function normalizeStatePayload(raw: StateNodePayload): StateNodePayload {
+  return {
+    role: raw.role,
+    meta: raw.meta,
+    start: raw.start,
+    content: raw.content,
+    ancestors: raw.ancestors,
+    compact: raw.compact,
+    timestamp: raw.timestamp,
+    childThread: raw.childThread ?? null,
+  };
+}
+
+/** Parses a YAML CAS blob into a typed RFC v3 thread node (or legacy content layout with `children`). */
+export function parseCasThreadNode(yamlText: string): ParsedCasThreadNode | null {
+  let raw: unknown;
+  try {
+    raw = parse(yamlText) as unknown;
+  } catch {
+    return null;
+  }
+  if (!isRecord(raw)) {
+    return null;
+  }
+  const type = raw.type;
+  if (type !== "start" && type !== "state" && type !== "content") {
+    return null;
+  }
+
+  let refsRaw: unknown = raw.refs;
+  if (refsRaw === undefined && type === "content") {
+    refsRaw = raw.children;
+  }
+  if (!Array.isArray(refsRaw) || !refsRaw.every((r) => typeof r === "string")) {
+    return null;
+  }
+  const refs = refsRaw as string[];
+
+  if (type === "content") {
+    if (typeof raw.payload !== "string") {
+      return null;
+    }
+    const node: ContentMerkleNode = { type: "content", payload: raw.payload, refs: [...refs] };
+    return { kind: "content", node };
+  }
+
+  if (type === "start") {
+    if (!isStartPayload(raw.payload)) {
+      return null;
+    }
+    const node: StartNode = {
+      type: "start",
+      payload: normalizeStartPayload(raw.payload),
+      refs: [...refs],
+    };
+    return { kind: "start", node };
+  }
+
+  if (!isStatePayload(raw.payload)) {
+    return null;
+  }
+  const node: StateNode = {
+    type: "state",
+    payload: normalizeStatePayload(raw.payload),
+    refs: [...refs],
+  };
+  return { kind: "state", node };
+}
+
+export type ParsedCasThreadNode =
+  | { kind: "start"; node: StartNode }
+  | { kind: "state"; node: StateNode }
+  | { kind: "content"; node: ContentMerkleNode };
+
 /** YAML-serialize a CAS node carrying `{type, payload, refs}` (RFC v3 thread storage format). */
 export function serializeCasNode(node: StartNode | StateNode | ContentMerkleNode): string {
   return stringify({ type: node.type, payload: node.payload, refs: node.refs }, { indent: 2 });
@@ -45,10 +183,14 @@ export async function putStartNode(
   payload: StartNode["payload"],
   promptHash: string,
 ): Promise<string> {
+  const refs = [promptHash];
+  if (payload.parentState !== null) {
+    refs.push(payload.parentState);
+  }
   const node: StartNode = {
     type: "start",
     payload,
-    refs: [promptHash],
+    refs,
   };
   return store.put(serializeCasNode(node));
 }

packages/workflow-cas/src/reachable.ts

@@ -9,7 +9,10 @@ function refsFromBlob(content: string): string[] {
       return [];
     }
     const rec = raw as Record<string, unknown>;
-    const refs = rec.refs;
+    let refs = rec.refs;
+    if (!Array.isArray(refs) && Array.isArray(rec.children)) {
+      refs = rec.children;
+    }
     if (!Array.isArray(refs)) {
       return [];
     }

packages/workflow-dashboard/.env.production

@@ -0,0 +1 @@
+VITE_GATEWAY_URL=https://workflow-gateway.shazhou.workers.dev

packages/workflow-dashboard/package.json

@@ -10,7 +10,9 @@
   },
   "dependencies": {
     "react": "^19.2.6",
-    "react-dom": "^19.2.6"
+    "react-dom": "^19.2.6",
+    "react-markdown": "^10.1.0",
+    "shiki": "^4.0.2"
   },
   "devDependencies": {
     "@tailwindcss/vite": "^4.2.4",

packages/workflow-dashboard/src/api.ts

@@ -1,9 +1,43 @@
-const BASE = "/api";
+const GATEWAY_URL = import.meta.env.VITE_GATEWAY_URL || "";
 
-async function postJson<T>(path: string, body: unknown): Promise<T> {
-  const res = await fetch(`${BASE}${path}`, {
+export function getApiKey(): string | null {
+  try {
+    return localStorage.getItem("workflow-api-key");
+  } catch {
+    return null;
+  }
+}
+
+export function setApiKey(key: string): void {
+  localStorage.setItem("workflow-api-key", key);
+}
+
+export function clearApiKey(): void {
+  localStorage.removeItem("workflow-api-key");
+}
+
+export function hasApiKey(): boolean {
+  return getApiKey() !== null && getApiKey() !== "";
+}
+
+function authHeaders(): Record<string, string> {
+  const key = getApiKey();
+  if (key) return { Authorization: `Bearer ${key}` };
+  return {};
+}
+
+function agentBase(agent: string): string {
+  if (GATEWAY_URL) {
+    return `${GATEWAY_URL}/api/agents/${agent}`;
+  }
+  // Local dev: proxy via vite, no agent prefix
+  return "/api";
+}
+
+async function postJson<T>(base: string, path: string, body: unknown): Promise<T> {
+  const res = await fetch(`${base}${path}`, {
     method: "POST",
-    headers: { "Content-Type": "application/json" },
+    headers: { "Content-Type": "application/json", ...authHeaders() },
     body: JSON.stringify(body),
   });
   if (!res.ok) {
@@ -13,14 +47,23 @@ async function postJson<T>(path: string, body: unknown): Promise<T> {
   return res.json() as Promise<T>;
 }
 
-async function fetchJson<T>(path: string): Promise<T> {
-  const res = await fetch(`${BASE}${path}`);
+async function fetchJson<T>(base: string, path: string): Promise<T> {
+  const res = await fetch(`${base}${path}`, { headers: authHeaders() });
   if (!res.ok) {
     throw new Error(`API ${res.status}: ${path}`);
   }
   return res.json() as Promise<T>;
 }
 
+// ── Endpoint types ──────────────────────────────────────────────────
+
+export type AgentEndpoint = {
+  name: string;
+  url: string;
+  status: string;
+  lastHeartbeat: number;
+};
+
 export type WorkflowSummary = {
   name: string;
   currentHash: string;
@@ -35,50 +78,77 @@ export type ThreadSummary = {
   status: string | null;
 };
 
-export type ThreadRecord = {
-  type: string;
-  role: string | null;
-  content: string | null;
-  timestamp: number | null;
-  [key: string]: unknown;
+export type ThreadStartRecord = {
+  type: "thread-start";
+  workflow: string;
+  prompt: string | null;
+  threadId: string;
+  status: string;
+  timestamp: null;
 };
 
-export function listWorkflows(): Promise<{ workflows: WorkflowSummary[] }> {
-  return fetchJson("/workflows");
+export type RoleRecord = {
+  type: "role";
+  role: string;
+  content: string;
+  timestamp: number | null;
+  meta: Record<string, unknown>;
+};
+
+export type WorkflowResultRecord = {
+  type: "workflow-result";
+  returnCode: number;
+  content: string;
+  timestamp: number | null;
+};
+
+export type ThreadRecord = ThreadStartRecord | RoleRecord | WorkflowResultRecord;
+
+// ── Gateway endpoints ───────────────────────────────────────────────
+
+export function listAgents(): Promise<AgentEndpoint[]> {
+  const url = GATEWAY_URL || "";
+  return fetchJson(url, "/api/gateway/endpoints");
 }
 
-export function listThreads(): Promise<{ threads: ThreadSummary[] }> {
-  return fetchJson("/threads");
+// ── Agent-scoped endpoints ──────────────────────────────────────────
+
+export function listWorkflows(agent: string): Promise<{ workflows: WorkflowSummary[] }> {
+  return fetchJson(agentBase(agent), "/workflows");
 }
 
-export function listRunningThreads(): Promise<{ threads: ThreadSummary[] }> {
-  return fetchJson("/threads/running");
+export function listThreads(agent: string): Promise<{ threads: ThreadSummary[] }> {
+  return fetchJson(agentBase(agent), "/threads");
 }
 
-export function getThread(id: string): Promise<{ records: ThreadRecord[] }> {
-  return fetchJson(`/threads/${id}`);
+export function listRunningThreads(agent: string): Promise<{ threads: ThreadSummary[] }> {
+  return fetchJson(agentBase(agent), "/threads/running");
+}
+
+export function getThread(agent: string, id: string): Promise<{ records: ThreadRecord[] }> {
+  return fetchJson(agentBase(agent), `/threads/${id}`);
 }
 
 export function runThread(
+  agent: string,
   workflow: string,
   prompt: string,
-  maxRounds: number = 10,
 ): Promise<{ threadId: string }> {
-  return postJson("/threads", { workflow, prompt, maxRounds });
+  return postJson(agentBase(agent), "/threads", { workflow, prompt });
 }
 
-export function killThread(threadId: string): Promise<{ ok: boolean }> {
-  return postJson(`/threads/${threadId}/kill`, {});
+export function killThread(agent: string, threadId: string): Promise<{ ok: boolean }> {
+  return postJson(agentBase(agent), `/threads/${threadId}/kill`, {});
 }
 
-export function pauseThread(threadId: string): Promise<{ ok: boolean }> {
-  return postJson(`/threads/${threadId}/pause`, {});
+export function pauseThread(agent: string, threadId: string): Promise<{ ok: boolean }> {
+  return postJson(agentBase(agent), `/threads/${threadId}/pause`, {});
 }
 
-export function resumeThread(threadId: string): Promise<{ ok: boolean }> {
-  return postJson(`/threads/${threadId}/resume`, {});
+export function resumeThread(agent: string, threadId: string): Promise<{ ok: boolean }> {
+  return postJson(agentBase(agent), `/threads/${threadId}/resume`, {});
 }
 
-export function getHealth(): Promise<{ ok: boolean }> {
-  return fetchJson("/healthz");
+export function getAgentHealth(agent: string): Promise<{ ok: boolean }> {
+  return fetchJson(agentBase(agent), "/healthz");
 }

packages/workflow-dashboard/src/app.tsx

@@ -1,4 +1,6 @@
 import { useState } from "react";
+import { clearApiKey, hasApiKey } from "./api.ts";
+import { LoginPage } from "./components/login.tsx";
 import { RunDialog } from "./components/run-dialog.tsx";
 import { Sidebar } from "./components/sidebar.tsx";
 import { StatusBar } from "./components/status-bar.tsx";
@@ -8,24 +10,48 @@ import { WorkflowList } from "./components/workflow-list.tsx";
 import { useHashRoute } from "./use-hash-route.ts";
 
 export function App() {
-  const { view, threadId, setView, setThreadId } = useHashRoute();
+  const [authed, setAuthed] = useState(hasApiKey());
+  const { view, agent, threadId, setView, setAgent, setThreadId } = useHashRoute();
   const [showRun, setShowRun] = useState(false);
 
+  if (!authed) {
+    return <LoginPage onLogin={() => setAuthed(true)} />;
+  }
+
   return (
     <div className="flex h-screen">
-      <Sidebar view={view} onViewChange={setView} />
+      <Sidebar
+        view={view}
+        agent={agent}
+        onViewChange={setView}
+        onAgentChange={setAgent}
+        onLogout={() => {
+          clearApiKey();
+          setAuthed(false);
+        }}
+      />
       <main className="flex-1 overflow-hidden flex flex-col">
-        <StatusBar onRun={() => setShowRun(true)} />
+        <StatusBar agent={agent} onRun={() => setShowRun(true)} />
         <div className="flex-1 overflow-auto p-6">
-          {view === "threads" && threadId === null && <ThreadList onSelect={setThreadId} />}
-          {view === "threads" && threadId !== null && (
-            <ThreadDetail threadId={threadId} onBack={() => setThreadId(null)} />
+          {!agent && (
+            <div className="flex items-center justify-center h-full">
+              <p style={{ color: "var(--color-text-muted)" }}>
+                Select an agent from the sidebar to get started.
+              </p>
+            </div>
           )}
-          {view === "workflows" && <WorkflowList />}
+          {agent && view === "threads" && threadId === null && (
+            <ThreadList agent={agent} onSelect={setThreadId} />
+          )}
+          {agent && view === "threads" && threadId !== null && (
+            <ThreadDetail agent={agent} threadId={threadId} onBack={() => setThreadId(null)} />
+          )}
+          {agent && view === "workflows" && <WorkflowList agent={agent} />}
         </div>
       </main>
-      {showRun && (
+      {showRun && agent && (
         <RunDialog
+          agent={agent}
           onClose={() => setShowRun(false)}
           onCreated={(id) => {
             setShowRun(false);

packages/workflow-dashboard/src/components/login.tsx

@@ -0,0 +1,96 @@
+import { useState } from "react";
+import { setApiKey } from "../api.ts";
+
+type Props = {
+  onLogin: () => void;
+};
+
+export function LoginPage({ onLogin }: Props) {
+  const [key, setKey] = useState("");
+  const [error, setError] = useState<string | null>(null);
+  const [loading, setLoading] = useState(false);
+
+  async function handleSubmit(e: React.FormEvent) {
+    e.preventDefault();
+    if (!key.trim()) return;
+
+    setLoading(true);
+    setError(null);
+
+    // Test the key by hitting the endpoints list
+    const gatewayUrl = import.meta.env.VITE_GATEWAY_URL || "";
+    try {
+      const res = await fetch(`${gatewayUrl}/api/gateway/endpoints`, {
+        headers: { Authorization: `Bearer ${key.trim()}` },
+      });
+      if (res.status === 401) {
+        setError("Invalid API key");
+        setLoading(false);
+        return;
+      }
+      if (!res.ok) {
+        setError(`Server error: ${res.status}`);
+        setLoading(false);
+        return;
+      }
+    } catch (err) {
+      setError(`Connection failed: ${err instanceof Error ? err.message : String(err)}`);
+      setLoading(false);
+      return;
+    }
+
+    setApiKey(key.trim());
+    onLogin();
+  }
+
+  return (
+    <div
+      className="min-h-screen flex items-center justify-center"
+      style={{ background: "var(--color-bg)" }}
+    >
+      <div
+        className="p-8 rounded-lg border w-full max-w-sm"
+        style={{ background: "var(--color-surface)", borderColor: "var(--color-border)" }}
+      >
+        <h1 className="text-xl font-bold mb-1" style={{ color: "var(--color-accent)" }}>
+          ⚙ Workflow Dashboard
+        </h1>
+        <p className="text-sm mb-6" style={{ color: "var(--color-text-muted)" }}>
+          Enter your API key to continue
+        </p>
+        <form onSubmit={handleSubmit}>
+          <input
+            type="password"
+            value={key}
+            onChange={(e) => setKey(e.target.value)}
+            placeholder="API Key"
+            className="w-full px-3 py-2 rounded border text-sm mb-3 outline-none"
+            style={{
+              background: "var(--color-bg)",
+              borderColor: "var(--color-border)",
+              color: "var(--color-text)",
+            }}
+            autoFocus
+          />
+          {error && (
+            <p className="text-xs mb-3" style={{ color: "var(--color-error)" }}>
+              {error}
+            </p>
+          )}
+          <button
+            type="submit"
+            disabled={loading || !key.trim()}
+            className="w-full px-3 py-2 rounded text-sm font-medium"
+            style={{
+              background: "var(--color-accent)",
+              color: "var(--color-bg)",
+              opacity: loading || !key.trim() ? 0.5 : 1,
+            }}
+          >
+            {loading ? "Verifying..." : "Login"}
+          </button>
+        </form>
+      </div>
+    </div>
+  );
+}

packages/workflow-dashboard/src/components/markdown.tsx

@@ -0,0 +1,127 @@
+import { useEffect, useState } from "react";
+import ReactMarkdown from "react-markdown";
+import {
+  type BundledLanguage,
+  type BundledTheme,
+  createHighlighter,
+  type HighlighterGeneric,
+} from "shiki";
+
+let highlighterPromise: Promise<HighlighterGeneric<BundledLanguage, BundledTheme>> | null = null;
+
+const LANGS: BundledLanguage[] = [
+  "typescript",
+  "javascript",
+  "json",
+  "yaml",
+  "bash",
+  "python",
+  "markdown",
+];
+
+function getHighlighter(): Promise<HighlighterGeneric<BundledLanguage, BundledTheme>> {
+  if (highlighterPromise === null) {
+    highlighterPromise = createHighlighter({
+      themes: ["github-dark"],
+      langs: LANGS,
+    });
+  }
+  return highlighterPromise;
+}
+
+function CodeBlock({ className, children }: { className?: string; children?: React.ReactNode }) {
+  const [html, setHtml] = useState<string | null>(null);
+  const code = String(children).replace(/\n$/, "");
+  const lang = className?.replace("language-", "") ?? "text";
+
+  useEffect(() => {
+    let cancelled = false;
+    getHighlighter().then((hl) => {
+      if (cancelled) return;
+      try {
+        const result = hl.codeToHtml(code, { lang, theme: "github-dark" });
+        setHtml(result);
+      } catch {
+        setHtml(null);
+      }
+    });
+    return () => {
+      cancelled = true;
+    };
+  }, [code, lang]);
+
+  if (html !== null) {
+    return (
+      <div
+        className="rounded overflow-x-auto text-xs my-2"
+        // biome-ignore lint/security/noDangerouslySetInnerHtml: shiki output is safe
+        dangerouslySetInnerHTML={{ __html: html }}
+      />
+    );
+  }
+
+  return (
+    <pre
+      className="rounded overflow-x-auto text-xs my-2 p-3"
+      style={{ background: "var(--color-bg)" }}
+    >
+      <code>{code}</code>
+    </pre>
+  );
+}
+
+export function Markdown({ content }: { content: string }) {
+  return (
+    <div className="prose prose-invert prose-sm max-w-none">
+      <ReactMarkdown
+        components={{
+          code({ className, children, ...props }) {
+            const isInline = !className;
+            if (isInline) {
+              return (
+                <code
+                  className="text-xs px-1 py-0.5 rounded"
+                  style={{ background: "var(--color-border)", color: "var(--color-accent)" }}
+                  {...props}
+                >
+                  {children}
+                </code>
+              );
+            }
+            return <CodeBlock className={className}>{children}</CodeBlock>;
+          },
+          p({ children }) {
+            return <p className="my-1.5 leading-relaxed">{children}</p>;
+          },
+          ul({ children }) {
+            return <ul className="list-disc pl-4 my-1.5">{children}</ul>;
+          },
+          ol({ children }) {
+            return <ol className="list-decimal pl-4 my-1.5">{children}</ol>;
+          },
+          h1({ children }) {
+            return <h1 className="text-lg font-bold mt-3 mb-1">{children}</h1>;
+          },
+          h2({ children }) {
+            return <h2 className="text-base font-bold mt-2 mb-1">{children}</h2>;
+          },
+          h3({ children }) {
+            return <h3 className="text-sm font-bold mt-2 mb-1">{children}</h3>;
+          },
+          blockquote({ children }) {
+            return (
+              <blockquote
+                className="border-l-2 pl-3 my-2 text-sm"
+                style={{ borderColor: "var(--color-accent)", color: "var(--color-text-muted)" }}
+              >
+                {children}
+              </blockquote>
+            );
+          },
+        }}
+      >
+        {content}
+      </ReactMarkdown>
+    </div>
+  );
+}

packages/workflow-dashboard/src/components/record-card.tsx

@@ -0,0 +1,126 @@
+import type { RoleRecord, ThreadRecord, ThreadStartRecord, WorkflowResultRecord } from "../api.ts";
+import { Markdown } from "./markdown.tsx";
+
+const ROLE_COLORS: Record<string, string> = {
+  preparer: "#8b5cf6",
+  agent: "#3b82f6",
+  extractor: "#f59e0b",
+};
+
+function roleColor(role: string): string {
+  return ROLE_COLORS[role] ?? "var(--color-accent)";
+}
+
+function formatTime(ts: number | null): string | null {
+  if (ts === null) return null;
+  return new Date(ts).toLocaleTimeString();
+}
+
+function StartCard({ record }: { record: ThreadStartRecord }) {
+  return (
+    <div
+      className="p-4 rounded-lg border"
+      style={{ background: "var(--color-surface)", borderColor: "var(--color-border)" }}
+    >
+      <div className="flex items-center gap-2 mb-2">
+        <span className="text-lg">🚀</span>
+        <span className="font-semibold" style={{ color: "var(--color-accent)" }}>
+          {record.workflow}
+        </span>
+        <span
+          className="text-xs px-2 py-0.5 rounded"
+          style={{
+            background: record.status === "active" ? "var(--color-success)" : "var(--color-border)",
+            color: record.status === "active" ? "var(--color-bg)" : "var(--color-text-muted)",
+          }}
+        >
+          {record.status}
+        </span>
+      </div>
+      {record.prompt !== null && (
+        <div
+          className="mt-2 p-3 rounded text-sm border-l-2"
+          style={{
+            background: "var(--color-bg)",
+            borderColor: "var(--color-accent)",
+            color: "var(--color-text)",
+          }}
+        >
+          <div className="text-xs mb-1" style={{ color: "var(--color-text-muted)" }}>
+            Prompt
+          </div>
+          <Markdown content={record.prompt} />
+        </div>
+      )}
+    </div>
+  );
+}
+
+function RoleMessage({ record }: { record: RoleRecord }) {
+  const color = roleColor(record.role);
+  return (
+    <div
+      className="p-3 rounded-lg border text-sm"
+      style={{ background: "var(--color-surface)", borderColor: "var(--color-border)" }}
+    >
+      <div className="flex items-center gap-2 mb-2">
+        <span
+          className="text-xs px-2 py-0.5 rounded font-mono font-medium"
+          style={{ background: color, color: "#fff" }}
+        >
+          {record.role}
+        </span>
+        {formatTime(record.timestamp) !== null && (
+          <span className="text-xs ml-auto" style={{ color: "var(--color-text-muted)" }}>
+            {formatTime(record.timestamp)}
+          </span>
+        )}
+      </div>
+      <Markdown content={record.content} />
+    </div>
+  );
+}
+
+function ResultCard({ record }: { record: WorkflowResultRecord }) {
+  const success = record.returnCode === 0;
+  return (
+    <div
+      className="p-4 rounded-lg border"
+      style={{
+        background: "var(--color-surface)",
+        borderColor: success ? "var(--color-success)" : "var(--color-error)",
+      }}
+    >
+      <div className="flex items-center gap-2 mb-2">
+        <span className="text-lg">{success ? "✅" : "❌"}</span>
+        <span className="font-semibold text-sm">{success ? "Completed" : "Failed"}</span>
+        <span
+          className="text-xs px-2 py-0.5 rounded font-mono"
+          style={{
+            background: success ? "var(--color-success)" : "var(--color-error)",
+            color: "#fff",
+          }}
+        >
+          exit {record.returnCode}
+        </span>
+        {formatTime(record.timestamp) !== null && (
+          <span className="text-xs ml-auto" style={{ color: "var(--color-text-muted)" }}>
+            {formatTime(record.timestamp)}
+          </span>
+        )}
+      </div>
+      <Markdown content={record.content} />
+    </div>
+  );
+}
+
+export function RecordCard({ record }: { record: ThreadRecord }) {
+  switch (record.type) {
+    case "thread-start":
+      return <StartCard record={record} />;
+    case "role":
+      return <RoleMessage record={record} />;
+    case "workflow-result":
+      return <ResultCard record={record} />;
+  }
+}

packages/workflow-dashboard/src/components/run-dialog.tsx

@@ -3,15 +3,15 @@ import { listWorkflows, runThread } from "../api.ts";
 import { useFetch } from "../hooks.ts";
 
 type Props = {
+  agent: string;
   onClose: () => void;
   onCreated: (threadId: string) => void;
 };
 
-export function RunDialog({ onClose, onCreated }: Props) {
-  const workflows = useFetch(() => listWorkflows(), []);
+export function RunDialog({ agent, onClose, onCreated }: Props) {
+  const workflows = useFetch(() => listWorkflows(agent), [agent]);
   const [workflow, setWorkflow] = useState("");
   const [prompt, setPrompt] = useState("");
-  const [maxRounds, setMaxRounds] = useState(10);
   const [submitting, setSubmitting] = useState(false);
   const [error, setError] = useState<string | null>(null);
 
@@ -21,7 +21,7 @@ export function RunDialog({ onClose, onCreated }: Props) {
     setSubmitting(true);
     setError(null);
     try {
-      const result = await runThread(workflow, prompt, maxRounds);
+      const result = await runThread(agent, workflow, prompt);
       onCreated(result.threadId);
     } catch (err) {
       setError(err instanceof Error ? err.message : String(err));
@@ -38,7 +38,7 @@ export function RunDialog({ onClose, onCreated }: Props) {
         className="w-full max-w-lg p-6 rounded-lg border"
         style={{ background: "var(--color-surface)", borderColor: "var(--color-border)" }}
       >
-        <h3 className="text-lg font-semibold mb-4">Run Thread</h3>
+        <h3 className="text-lg font-semibold mb-4">Run Thread on {agent}</h3>
         <form onSubmit={handleSubmit} className="space-y-4">
           <div>
             <label
@@ -90,29 +90,6 @@ export function RunDialog({ onClose, onCreated }: Props) {
               placeholder="Enter the task prompt..."
             />
           </div>
-          <div>
-            <label
-              htmlFor="run-max-rounds"
-              className="text-sm block mb-1"
-              style={{ color: "var(--color-text-muted)" }}
-            >
-              Max Rounds
-            </label>
-            <input
-              id="run-max-rounds"
-              type="number"
-              value={maxRounds}
-              onChange={(e) => setMaxRounds(Number(e.target.value))}
-              min={1}
-              max={100}
-              className="w-24 px-3 py-2 rounded border text-sm"
-              style={{
-                background: "var(--color-bg)",
-                borderColor: "var(--color-border)",
-                color: "var(--color-text)",
-              }}
-            />
-          </div>
           {error && (
             <p className="text-sm" style={{ color: "var(--color-error)" }}>
               {error}

packages/workflow-dashboard/src/components/sidebar.tsx

@@ -1,10 +1,29 @@
+import { useEffect } from "react";
+import type { AgentEndpoint } from "../api.ts";
+import { listAgents } from "../api.ts";
+import { useFetch } from "../hooks.ts";
+
 type Props = {
   view: "threads" | "workflows";
+  agent: string | null;
   onViewChange: (v: "threads" | "workflows") => void;
+  onAgentChange: (a: string | null) => void;
+  onLogout: () => void;
 };
 
-export function Sidebar({ view, onViewChange }: Props) {
-  const items = [
+export function Sidebar({ view, agent, onViewChange, onAgentChange, onLogout }: Props) {
+  const { status, data } = useFetch(() => listAgents(), []);
+
+  const agents: AgentEndpoint[] = status === "ok" ? data : [];
+
+  // Auto-select first agent when none is selected
+  useEffect(() => {
+    if (agent === null && agents.length > 0) {
+      onAgentChange(agents[0].name);
+    }
+  }, [agent, agents, onAgentChange]);
+
+  const viewItems = [
     { key: "threads" as const, label: "Threads", icon: "⚡" },
     { key: "workflows" as const, label: "Workflows", icon: "📦" },
   ];
@@ -22,8 +41,45 @@ export function Sidebar({ view, onViewChange }: Props) {
           Dashboard
         </p>
       </div>
+
+      {/* Agent selector */}
+      <div className="px-4 py-3 border-b" style={{ borderColor: "var(--color-border)" }}>
+        <label
+          className="block text-xs font-medium mb-1"
+          style={{ color: "var(--color-text-muted)" }}
+          htmlFor="agent-select"
+        >
+          Agent
+        </label>
+        <select
+          id="agent-select"
+          className="w-full rounded px-2 py-1.5 text-xs"
+          style={{
+            background: "var(--color-bg)",
+            color: "var(--color-text)",
+            border: "1px solid var(--color-border)",
+          }}
+          value={agent ?? ""}
+          onChange={(e) => onAgentChange(e.target.value || null)}
+          disabled={status === "loading"}
+        >
+          {status === "loading" ? (
+            <option value="">Loading…</option>
+          ) : agents.length === 0 ? (
+            <option value="">No agents online</option>
+          ) : (
+            agents.map((a) => (
+              <option key={a.name} value={a.name}>
+                {a.status === "online" ? "🟢" : "🔴"} {a.name}
+              </option>
+            ))
+          )}
+        </select>
+      </div>
+
+      {/* View navigation */}
       <nav className="flex-1 p-2 space-y-1">
-        {items.map((item) => (
+        {viewItems.map((item) => (
           <button
             type="button"
             key={item.key}
@@ -38,6 +94,17 @@ export function Sidebar({ view, onViewChange }: Props) {
           </button>
         ))}
       </nav>
+
+      <div className="p-2 border-t" style={{ borderColor: "var(--color-border)" }}>
+        <button
+          type="button"
+          onClick={onLogout}
+          className="w-full text-left px-3 py-2 rounded text-xs transition-colors"
+          style={{ color: "var(--color-text-muted)" }}
+        >
+          🚪 Logout
+        </button>
+      </div>
     </aside>
   );
 }

packages/workflow-dashboard/src/components/status-bar.tsx

@@ -1,9 +1,10 @@
 import { useCallback, useEffect, useRef, useState } from "react";
-import { getHealth } from "../api.ts";
+import { getAgentHealth } from "../api.ts";
 
 type HealthStatus = "connected" | "disconnected" | "reconnecting";
 
 type Props = {
+  agent: string | null;
   onRun: () => void;
 };
 
@@ -17,13 +18,17 @@ function statusLabel(status: HealthStatus): { text: string; color: string } {
   return { text: "● Offline", color: "var(--color-error)" };
 }
 
-export function StatusBar({ onRun }: Props) {
+export function StatusBar({ agent, onRun }: Props) {
   const [status, setStatus] = useState<HealthStatus>("disconnected");
   const wasConnectedRef = useRef(false);
 
   const checkHealth = useCallback(async () => {
+    if (!agent) {
+      setStatus("disconnected");
+      return;
+    }
     try {
-      await getHealth();
+      await getAgentHealth(agent);
       wasConnectedRef.current = true;
       setStatus("connected");
     } catch {
@@ -33,9 +38,11 @@ export function StatusBar({ onRun }: Props) {
         setStatus("disconnected");
       }
     }
-  }, []);
+  }, [agent]);
 
   useEffect(() => {
+    wasConnectedRef.current = false;
+    setStatus("disconnected");
     checkHealth();
     const interval = setInterval(checkHealth, 10_000);
     return () => clearInterval(interval);
@@ -49,12 +56,19 @@ export function StatusBar({ onRun }: Props) {
       style={{ borderColor: "var(--color-border)", background: "var(--color-surface)" }}
     >
       <div className="flex items-center gap-4">
-        <span style={{ color: "var(--color-text-muted)" }}>Local API: 127.0.0.1:7860</span>
+        <span style={{ color: "var(--color-text-muted)" }}>
+          {agent ? `Agent: ${agent}` : "No agent selected"}
+        </span>
         <button
           type="button"
           onClick={onRun}
+          disabled={!agent}
           className="px-3 py-1 rounded text-xs font-medium"
-          style={{ background: "var(--color-accent)", color: "#fff" }}
+          style={{
+            background: agent ? "var(--color-accent)" : "var(--color-border)",
+            color: "#fff",
+            opacity: agent ? 1 : 0.5,
+          }}
         >
           ▶ Run Thread
         </button>

packages/workflow-dashboard/src/components/thread-detail.tsx

@@ -2,15 +2,17 @@ import { useEffect, useRef, useState } from "react";
 import { getThread, killThread, pauseThread, resumeThread } from "../api.ts";
 import { useFetch } from "../hooks.ts";
 import { useSSE } from "../use-sse.ts";
+import { RecordCard } from "./record-card.tsx";
 
 type Props = {
+  agent: string;
   threadId: string;
   onBack: () => void;
 };
 
-export function ThreadDetail({ threadId, onBack }: Props) {
-  const sse = useSSE(threadId);
-  const { status, data, error } = useFetch(() => getThread(threadId), [threadId]);
+export function ThreadDetail({ agent, threadId, onBack }: Props) {
+  const sse = useSSE(agent, threadId);
+  const { status, data, error } = useFetch(() => getThread(agent, threadId), [agent, threadId]);
   const [actionStatus, setActionStatus] = useState<string | null>(null);
   const recordsEndRef = useRef<HTMLDivElement>(null);
 
@@ -30,7 +32,7 @@ export function ThreadDetail({ threadId, onBack }: Props) {
     setActionStatus(`${action}ing...`);
     try {
       const fn = action === "kill" ? killThread : action === "pause" ? pauseThread : resumeThread;
-      await fn(threadId);
+      await fn(agent, threadId);
       setActionStatus(`${action} sent ✓`);
     } catch (e) {
       setActionStatus(`${action} failed: ${e instanceof Error ? e.message : String(e)}`);
@@ -78,7 +80,7 @@ export function ThreadDetail({ threadId, onBack }: Props) {
 
       <h2 className="text-xl font-semibold mb-2 font-mono flex items-center gap-2 flex-wrap">
         <span>{threadId}</span>
-        {sse.connected && (
+        {sse.connected && !sse.completed && (
           <span
             className="text-xs font-medium px-2 py-0.5 rounded"
             style={{ background: "var(--color-success)", color: "var(--color-bg)" }}
@@ -101,39 +103,8 @@ export function ThreadDetail({ threadId, onBack }: Props) {
       )}
       {(status === "ok" || liveActive || records.length > 0) && (
         <div className="space-y-3">
-          {records.map((r) => (
-            <div
-              key={`${threadId}-${r.type}-${String(r.timestamp)}-${r.role ?? ""}-${r.content ?? ""}`}
-              className="p-3 rounded border text-sm"
-              style={{ background: "var(--color-surface)", borderColor: "var(--color-border)" }}
-            >
-              <div className="flex items-center gap-2 mb-1">
-                <span
-                  className="text-xs px-1.5 py-0.5 rounded font-mono"
-                  style={{ background: "var(--color-border)", color: "var(--color-accent)" }}
-                >
-                  {r.type}
-                </span>
-                {r.role && (
-                  <span className="text-xs" style={{ color: "var(--color-text-muted)" }}>
-                    {r.role}
-                  </span>
-                )}
-                {r.timestamp !== null && (
-                  <span className="text-xs ml-auto" style={{ color: "var(--color-text-muted)" }}>
-                    {new Date(r.timestamp).toLocaleTimeString()}
-                  </span>
-                )}
-              </div>
-              {r.content && (
-                <pre
-                  className="whitespace-pre-wrap text-xs mt-1"
-                  style={{ color: "var(--color-text)" }}
-                >
-                  {typeof r.content === "string" ? r.content : JSON.stringify(r.content, null, 2)}
-                </pre>
-              )}
-            </div>
+          {records.map((r, i) => (
+            <RecordCard key={`${threadId}-${i}`} record={r} />
           ))}
           <div ref={recordsEndRef} aria-hidden />
         </div>

packages/workflow-dashboard/src/components/thread-list.tsx

@@ -2,17 +2,23 @@ import { listThreads } from "../api.ts";
 import { useFetch } from "../hooks.ts";
 
 type Props = {
+  agent: string;
   onSelect: (id: string) => void;
 };
 
-export function ThreadList({ onSelect }: Props) {
-  const { status, data, error } = useFetch(() => listThreads(), []);
+export function ThreadList({ agent, onSelect }: Props) {
+  const { status, data, error } = useFetch(() => listThreads(agent), [agent]);
 
   if (status === "loading")
     return <p style={{ color: "var(--color-text-muted)" }}>Loading threads...</p>;
   if (status === "error") return <p style={{ color: "var(--color-error)" }}>Error: {error}</p>;
 
-  const threads = data.threads;
+  const threads = [...data.threads].sort((a, b) => {
+    if (!a.startedAt && !b.startedAt) return 0;
+    if (!a.startedAt) return 1;
+    if (!b.startedAt) return -1;
+    return b.startedAt.localeCompare(a.startedAt);
+  });
 
   return (
     <div>
@@ -38,11 +44,11 @@ export function ThreadList({ onSelect }: Props) {
                     className="text-xs px-2 py-0.5 rounded"
                     style={{
                       background:
-                        t.status === "running"
+                        t.status === "completed"
                           ? "var(--color-success)"
                           : t.status === "failed"
                             ? "var(--color-error)"
-                            : "var(--color-text-muted)",
+                            : "var(--color-accent)",
                       color: "#000",
                     }}
                   >

packages/workflow-dashboard/src/components/workflow-list.tsx

@@ -1,8 +1,12 @@
 import { listWorkflows } from "../api.ts";
 import { useFetch } from "../hooks.ts";
 
-export function WorkflowList() {
-  const { status, data, error } = useFetch(() => listWorkflows(), []);
+type Props = {
+  agent: string;
+};
+
+export function WorkflowList({ agent }: Props) {
+  const { status, data, error } = useFetch(() => listWorkflows(agent), [agent]);
 
   if (status === "loading")
     return <p style={{ color: "var(--color-text-muted)" }}>Loading workflows...</p>;

packages/workflow-dashboard/src/use-hash-route.ts

@@ -4,37 +4,50 @@ type View = "threads" | "workflows";
 
 type HashRoute = {
   view: View;
+  agent: string | null;
   threadId: string | null;
 };
 
 function parseHash(hash: string): HashRoute {
   const raw = hash.replace(/^#\/?/, "");
-  if (raw.startsWith("threads/")) {
-    const id = raw.slice("threads/".length);
-    if (id.length > 0) {
-      return { view: "threads", threadId: id };
-    }
+  // Format: #agent/threads/id or #agent/workflows or #threads or #workflows
+  const parts = raw.split("/");
+
+  // Check if first part is a known view
+  if (parts[0] === "threads" || parts[0] === "workflows") {
+    return {
+      view: parts[0] as View,
+      agent: null,
+      threadId: parts[0] === "threads" && parts.length > 1 ? parts.slice(1).join("/") : null,
+    };
   }
-  if (raw === "workflows") {
-    return { view: "workflows", threadId: null };
-  }
-  return { view: "threads", threadId: null };
+
+  // First part is agent name
+  const agent = parts[0] || null;
+  const viewPart = parts[1] ?? "threads";
+  const view: View = viewPart === "workflows" ? "workflows" : "threads";
+  const threadId = view === "threads" && parts.length > 2 ? parts.slice(2).join("/") : null;
+
+  return { view, agent, threadId };
 }
 
 function buildHash(route: HashRoute): string {
+  const prefix = route.agent ? `${route.agent}/` : "";
   if (route.view === "workflows") {
-    return "#workflows";
+    return `#${prefix}workflows`;
   }
   if (route.threadId !== null) {
-    return `#threads/${route.threadId}`;
+    return `#${prefix}threads/${route.threadId}`;
   }
-  return "#threads";
+  return `#${prefix}threads`;
 }
 
 export function useHashRoute(): {
   view: View;
+  agent: string | null;
   threadId: string | null;
   setView: (v: View) => void;
+  setAgent: (a: string | null) => void;
   setThreadId: (id: string | null) => void;
 } {
   const [route, setRoute] = useState<HashRoute>(() => parseHash(window.location.hash));
@@ -53,12 +66,27 @@ export function useHashRoute(): {
     setRoute(next);
   }, []);
 
-  const setView = useCallback((v: View) => navigate({ view: v, threadId: null }), [navigate]);
-
-  const setThreadId = useCallback(
-    (id: string | null) => navigate({ view: "threads", threadId: id }),
-    [navigate],
+  const setView = useCallback(
+    (v: View) => navigate({ view: v, agent: route.agent, threadId: null }),
+    [navigate, route.agent],
   );
 
-  return { view: route.view, threadId: route.threadId, setView, setThreadId };
+  const setAgent = useCallback(
+    (a: string | null) => navigate({ view: route.view, agent: a, threadId: null }),
+    [navigate, route.view],
+  );
+
+  const setThreadId = useCallback(
+    (id: string | null) => navigate({ view: "threads", agent: route.agent, threadId: id }),
+    [navigate, route.agent],
+  );
+
+  return {
+    view: route.view,
+    agent: route.agent,
+    threadId: route.threadId,
+    setView,
+    setAgent,
+    setThreadId,
+  };
 }

packages/workflow-dashboard/src/use-sse.ts

@@ -8,6 +8,7 @@ import {
 } from "react";
 
 import type { ThreadRecord } from "./api.ts";
+import { getApiKey } from "./api.ts";
 
 export type UseSSEReturn = {
   records: ThreadRecord[];
@@ -56,7 +57,17 @@ function handleRecordEvent(ev: Event, ctx: RecordEventContext): void {
   ctx.cleanupEs();
 }
 
-export function useSSE(threadId: string | null): UseSSEReturn {
+function sseUrl(agent: string, threadId: string): string {
+  const gatewayUrl = import.meta.env.VITE_GATEWAY_URL || "";
+  const key = getApiKey();
+  const keyParam = key ? `?key=${encodeURIComponent(key)}` : "";
+  if (gatewayUrl) {
+    return `${gatewayUrl}/api/${agent}/threads/${encodeURIComponent(threadId)}/live${keyParam}`;
+  }
+  return `/api/threads/${encodeURIComponent(threadId)}/live`;
+}
+
+export function useSSE(agent: string | null, threadId: string | null): UseSSEReturn {
   const [records, setRecords] = useState<ThreadRecord[]>([]);
   const [connected, setConnected] = useState(false);
   const [completed, setCompleted] = useState(false);
@@ -65,7 +76,7 @@ export function useSSE(threadId: string | null): UseSSEReturn {
   const reconnectAttemptsRef = useRef(0);
 
   useEffect(() => {
-    if (threadId === null) {
+    if (threadId === null || agent === null) {
       completedRef.current = false;
       reconnectAttemptsRef.current = 0;
       setRecords([]);
@@ -75,6 +86,7 @@ export function useSSE(threadId: string | null): UseSSEReturn {
     }
 
     const tid = threadId;
+    const agentName = agent;
 
     completedRef.current = false;
     reconnectAttemptsRef.current = 0;
@@ -113,7 +125,7 @@ export function useSSE(threadId: string | null): UseSSEReturn {
       }
 
       cleanupEs();
-      const url = `/api/threads/${encodeURIComponent(tid)}/live`;
+      const url = sseUrl(agentName, tid);
       es = new EventSource(url);
 
       es.onopen = () => {
@@ -136,6 +148,16 @@ export function useSSE(threadId: string | null): UseSSEReturn {
         }),
       );
 
+      es.addEventListener("done", () => {
+        if (cancelled) {
+          return;
+        }
+        completedRef.current = true;
+        setCompleted(true);
+        setConnected(false);
+        cleanupEs();
+      });
+
       es.onerror = () => {
         if (cancelled || completedRef.current) {
           return;
@@ -155,7 +177,7 @@ export function useSSE(threadId: string | null): UseSSEReturn {
       }
       cleanupEs();
     };
-  }, [threadId]);
+  }, [agent, threadId]);
 
   return { records, connected, completed };
 }

packages/workflow-execute/__tests__/engine.test.ts

@@ -34,12 +34,14 @@ function noLogger(): (tag: string, content: string) => void {
 
 function makeOptions(overrides: Partial<ExecuteThreadOptions>): ExecuteThreadOptions {
   return {
-    maxRounds: 5,
     depth: 0,
+    parentStateHash: null,
     signal: new AbortController().signal,
     awaitAfterEachYield: async () => {},
     forkSourceThreadId: null,
     prefilledDiskSteps: null,
+    forkContinuation: null,
+    replayTimestamps: null,
     storageRoot: "/tmp/never",
     ...overrides,
   };
@@ -105,7 +107,7 @@ describe("executeThread (Phase 2 — CAS thread storage)", () => {
       wf,
       "demo",
       { prompt: "hello", steps: [] },
-      makeOptions({ storageRoot, maxRounds: 5 }),
+      makeOptions({ storageRoot }),
       io,
       noLogger(),
     );
@@ -125,7 +127,6 @@ describe("executeThread (Phase 2 — CAS thread storage)", () => {
     expect(startNode.type).toBe("start");
     expect((startNode.payload as Record<string, unknown>).name).toBe("demo");
     expect((startNode.payload as Record<string, unknown>).hash).toBe(bundleHash);
-    expect((startNode.payload as Record<string, unknown>).maxRounds).toBe(5);
 
     const refs = startNode.refs as string[];
     expect(refs.length).toBe(1);
@@ -144,9 +145,9 @@ describe("executeThread (Phase 2 — CAS thread storage)", () => {
       runtime: WorkflowRuntime,
     ): AsyncGenerator<RoleOutput, WorkflowCompletion> {
       const h1 = await runtime.cas.put("plan-text");
-      yield { role: "planner", contentHash: h1, meta: { plan: 1 }, refs: [h1] };
+      yield { role: "planner", contentHash: h1, meta: { plan: 1 }, refs: [h1], childThread: null };
       const h2 = await runtime.cas.put("code-text");
-      yield { role: "coder", contentHash: h2, meta: { diff: "y" }, refs: [h2] };
+      yield { role: "coder", contentHash: h2, meta: { diff: "y" }, refs: [h2], childThread: null };
       return { returnCode: 0, summary: "done" };
     };
 
@@ -162,7 +163,6 @@ describe("executeThread (Phase 2 — CAS thread storage)", () => {
 
     const opts = makeOptions({
       storageRoot,
-      maxRounds: 5,
       awaitAfterEachYield: async () => {
         const text = await readFile(join(bundleDir, "threads.json"), "utf8");
         const parsed = JSON.parse(text) as Record<string, { head: string }>;
@@ -211,7 +211,7 @@ describe("executeThread (Phase 2 — CAS thread storage)", () => {
       runtime: WorkflowRuntime,
     ): AsyncGenerator<RoleOutput, WorkflowCompletion> {
       const h = await runtime.cas.put("only-step");
-      yield { role: "only", contentHash: h, meta: {}, refs: [h] };
+      yield { role: "only", contentHash: h, meta: {}, refs: [h], childThread: null };
       return { returnCode: 0, summary: "completed" };
     };
 
@@ -226,7 +226,7 @@ describe("executeThread (Phase 2 — CAS thread storage)", () => {
       wf,
       "demo",
       { prompt: "p", steps: [] },
-      makeOptions({ storageRoot, maxRounds: 5 }),
+      makeOptions({ storageRoot }),
       io,
       noLogger(),
     );
@@ -262,7 +262,7 @@ describe("executeThread (Phase 2 — CAS thread storage)", () => {
       runtime: WorkflowRuntime,
     ): AsyncGenerator<RoleOutput, WorkflowCompletion> {
       const h = await runtime.cas.put("step");
-      yield { role: "only", contentHash: h, meta: {}, refs: [h] };
+      yield { role: "only", contentHash: h, meta: {}, refs: [h], childThread: null };
       return { returnCode: 0, summary: "done" };
     };
 
@@ -277,7 +277,7 @@ describe("executeThread (Phase 2 — CAS thread storage)", () => {
       wf,
       "demo",
       { prompt: "p", steps: [] },
-      makeOptions({ storageRoot, maxRounds: 5 }),
+      makeOptions({ storageRoot }),
       io,
       noLogger(),
     );

packages/workflow-execute/__tests__/extract-refs.test.ts

@@ -0,0 +1,59 @@
+import { afterEach, describe, expect, test } from "bun:test";
+import { mkdtemp, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { createCasStore, putContentNodeWithRefs } from "@uncaged/workflow-cas";
+import * as z from "zod/v4";
+
+import { createExtract } from "../src/extract/extract-fn.js";
+
+function installPlainJsonExtractMock(meta: Record<string, unknown>): () => void {
+  const origFetch = globalThis.fetch;
+  const mockFetch = async (): Promise<Response> =>
+    new Response(
+      JSON.stringify({
+        choices: [{ message: { content: JSON.stringify(meta) } }],
+      }),
+      { status: 200, headers: { "Content-Type": "application/json" } },
+    );
+  globalThis.fetch = Object.assign(mockFetch, {
+    preconnect: origFetch.preconnect.bind(origFetch),
+  }) as typeof fetch;
+  return () => {
+    globalThis.fetch = origFetch;
+  };
+}
+
+describe("createExtract — ExtractResult shape", () => {
+  let restoreFetch: (() => void) | null = null;
+
+  afterEach(() => {
+    restoreFetch?.();
+    restoreFetch = null;
+  });
+
+  test("returns meta, contentPayload, and refs[]", async () => {
+    restoreFetch = installPlainJsonExtractMock({ confidence: 0.9 });
+
+    const dir = await mkdtemp(join(tmpdir(), "wf-extract-refs-"));
+    try {
+      const cas = createCasStore(join(dir, "cas"));
+      const extract = createExtract(
+        { baseUrl: "http://127.0.0.1:9", apiKey: "key", model: "m" },
+        { cas },
+      );
+
+      const schema = z.object({ confidence: z.number() });
+      const contentHash = await putContentNodeWithRefs(cas, "model says hello", []);
+
+      const out = await extract(schema, contentHash);
+
+      expect(out.meta).toEqual({ confidence: 0.9 });
+      expect(out.contentPayload).toBe("model says hello");
+      expect(Array.isArray(out.refs)).toBe(true);
+      expect(out.refs).toEqual([]);
+    } finally {
+      await rm(dir, { recursive: true, force: true });
+    }
+  });
+});

packages/workflow-execute/__tests__/gc-mark-sweep.test.ts

@@ -0,0 +1,115 @@
+import { afterEach, beforeEach, describe, expect, test } from "bun:test";
+import { mkdir, mkdtemp, rm, writeFile } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import {
+  createCasStore,
+  putContentNodeWithRefs,
+  putStartNode,
+  putStateNode,
+} from "@uncaged/workflow-cas";
+import type { StateNodePayload } from "@uncaged/workflow-protocol";
+
+import { FORK_BRANCH_ROLE } from "../src/engine/fork-thread.js";
+import { garbageCollectCas } from "../src/engine/gc.js";
+import { getBundleDir, removeThreadEntry, upsertThreadEntry } from "../src/engine/threads-index.js";
+
+describe("garbageCollectCas (mark-and-sweep)", () => {
+  let storageRoot: string;
+  let casDir: string;
+
+  beforeEach(async () => {
+    storageRoot = await mkdtemp(join(tmpdir(), "uncaged-wf-gc-ms-"));
+    casDir = join(storageRoot, "cas");
+    await mkdir(casDir, { recursive: true });
+    await writeFile(
+      join(storageRoot, "workflow.yaml"),
+      "config:\n  maxDepth: 1\n  supervisorInterval: 0\n  providers: {}\n  models: {}\nworkflows: {}\n",
+      "utf8",
+    );
+  });
+
+  afterEach(async () => {
+    await rm(storageRoot, { recursive: true, force: true });
+  });
+
+  test("shared CAS prefix survives when one fork thread index entry is removed", async () => {
+    const bundleHash = "TESTGC0000001";
+    const bundleDir = getBundleDir(storageRoot, bundleHash);
+    await mkdir(bundleDir, { recursive: true });
+
+    const cas = createCasStore(casDir);
+    const promptHash = await cas.put("prompt");
+    const startHash = await putStartNode(
+      cas,
+      {
+        name: "demo",
+        hash: bundleHash,
+        depth: 0,
+        parentState: null,
+      },
+      promptHash,
+    );
+
+    const c1 = await putContentNodeWithRefs(cas, "p1", []);
+    const h1 = await putStateNode(cas, {
+      role: "planner",
+      meta: {},
+      start: startHash,
+      content: c1,
+      ancestors: [],
+      compact: null,
+      timestamp: 1,
+      childThread: null,
+    } satisfies StateNodePayload);
+
+    const c2 = await putContentNodeWithRefs(cas, "c1", []);
+    const h2 = await putStateNode(cas, {
+      role: "coder",
+      meta: {},
+      start: startHash,
+      content: c2,
+      ancestors: [h1],
+      compact: null,
+      timestamp: 2,
+      childThread: null,
+    } satisfies StateNodePayload);
+
+    const ec = await putContentNodeWithRefs(cas, "", []);
+    const fm = await putStateNode(cas, {
+      role: FORK_BRANCH_ROLE,
+      meta: {},
+      start: startHash,
+      content: ec,
+      ancestors: [h1],
+      compact: null,
+      timestamp: 3,
+      childThread: null,
+    } satisfies StateNodePayload);
+
+    await upsertThreadEntry(bundleDir, "THREAD_AAAAAAA", {
+      head: h2,
+      start: startHash,
+      updatedAt: 10,
+    });
+    await upsertThreadEntry(bundleDir, "THREAD_BBBBBBB", {
+      head: fm,
+      start: startHash,
+      updatedAt: 20,
+    });
+
+    await removeThreadEntry(bundleDir, "THREAD_AAAAAAA");
+
+    const gc = await garbageCollectCas(storageRoot);
+    expect(gc.ok).toBe(true);
+    if (!gc.ok) {
+      return;
+    }
+
+    expect(await cas.get(h2)).toBeNull();
+    expect(await cas.get(h1)).not.toBeNull();
+    expect(await cas.get(startHash)).not.toBeNull();
+    expect(await cas.get(promptHash)).not.toBeNull();
+    expect(await cas.get(fm)).not.toBeNull();
+  });
+});

packages/workflow-execute/__tests__/merkle-call-stack.test.ts

@@ -0,0 +1,306 @@
+import { afterEach, beforeEach, describe, expect, test } from "bun:test";
+import { mkdir, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import type { CasStore } from "@uncaged/workflow-cas";
+import { createCasStore, parseCasThreadNode } from "@uncaged/workflow-cas";
+import type { StartNode, StateNode } from "@uncaged/workflow-protocol";
+import type {
+  RoleOutput,
+  ThreadContext,
+  WorkflowCompletion,
+  WorkflowFn,
+  WorkflowRuntime,
+} from "@uncaged/workflow-runtime";
+
+import { executeThread } from "../src/engine/engine.js";
+import type { ExecuteThreadIo, ExecuteThreadOptions } from "../src/engine/types.js";
+
+const TEST_REGISTRY_YAML = `config:
+  maxDepth: 3
+  supervisorInterval: 0
+  providers:
+    stub:
+      baseUrl: http://127.0.0.1:9
+      apiKey: test
+  models:
+    default: stub/m
+workflows: {}
+`;
+
+function noLogger(): (tag: string, content: string) => void {
+  return () => {};
+}
+
+function makeOptions(overrides: Partial<ExecuteThreadOptions>): ExecuteThreadOptions {
+  return {
+    depth: 0,
+    parentStateHash: null,
+    signal: new AbortController().signal,
+    awaitAfterEachYield: async () => {},
+    forkSourceThreadId: null,
+    prefilledDiskSteps: null,
+    forkContinuation: null,
+    replayTimestamps: null,
+    storageRoot: "/tmp/never",
+    ...overrides,
+  };
+}
+
+async function setupStorage(): Promise<{
+  storageRoot: string;
+  casDir: string;
+}> {
+  const storageRoot = await mkdtemp(join(tmpdir(), "uncaged-wf-merkle-"));
+  await writeFile(join(storageRoot, "workflow.yaml"), TEST_REGISTRY_YAML, "utf8");
+  const casDir = join(storageRoot, "cas");
+  await mkdir(casDir, { recursive: true });
+  return { storageRoot, casDir };
+}
+
+async function loadStartNode(cas: CasStore, endHash: string): Promise<StartNode> {
+  const endBlob = await cas.get(endHash);
+  const endParsed = parseCasThreadNode(endBlob ?? "");
+  if (endParsed?.kind !== "state") throw new Error("expected state node");
+  const startBlob = await cas.get(endParsed.node.payload.start);
+  const startParsed = parseCasThreadNode(startBlob ?? "");
+  if (startParsed?.kind !== "start") throw new Error("expected start node");
+  return startParsed.node;
+}
+
+async function loadStateNode(cas: CasStore, hash: string): Promise<StateNode> {
+  const blob = await cas.get(hash);
+  const parsed = parseCasThreadNode(blob ?? "");
+  if (parsed?.kind !== "state") throw new Error("expected state node");
+  return parsed.node;
+}
+
+describe("Merkle call stack — cross-thread DAG linking (Phase 2)", () => {
+  let storageRoot: string;
+  let casDir: string;
+
+  beforeEach(async () => {
+    const setup = await setupStorage();
+    storageRoot = setup.storageRoot;
+    casDir = setup.casDir;
+  });
+
+  afterEach(async () => {
+    await rm(storageRoot, { recursive: true, force: true });
+  });
+
+  test("parentStateHash is written into child start node's parentState and refs", async () => {
+    const cas = createCasStore(casDir);
+
+    // biome-ignore lint/correctness/useYield: testing start-only path
+    const parentWf: WorkflowFn = async function* (
+      _thread: ThreadContext,
+      _runtime: WorkflowRuntime,
+    ): AsyncGenerator<RoleOutput, WorkflowCompletion> {
+      return { returnCode: 0, summary: "parent done" };
+    };
+
+    const parentResult = await executeThread(
+      parentWf,
+      "parent-wf",
+      { prompt: "parent task", steps: [] },
+      makeOptions({ storageRoot }),
+      {
+        threadId: "P_THREAD_01",
+        hash: "PARENTHASH0001",
+        infoJsonlPath: join(storageRoot, "logs", "PARENTHASH0001", "P1.info.jsonl"),
+        cas,
+      },
+      noLogger(),
+    );
+
+    // biome-ignore lint/correctness/useYield: testing start-only path
+    const childWf: WorkflowFn = async function* (
+      _thread: ThreadContext,
+      _runtime: WorkflowRuntime,
+    ): AsyncGenerator<RoleOutput, WorkflowCompletion> {
+      return { returnCode: 0, summary: "child done" };
+    };
+
+    const childResult = await executeThread(
+      childWf,
+      "child-wf",
+      { prompt: "child task", steps: [] },
+      makeOptions({ storageRoot, depth: 1, parentStateHash: parentResult.rootHash }),
+      {
+        threadId: "C_THREAD_01",
+        hash: "CHILDHASH00001",
+        infoJsonlPath: join(storageRoot, "logs", "CHILDHASH00001", "C1.info.jsonl"),
+        cas,
+      },
+      noLogger(),
+    );
+
+    const childStart = await loadStartNode(cas, childResult.rootHash);
+    expect(childStart.payload.parentState).toBe(parentResult.rootHash);
+    expect(childStart.refs).toContain(parentResult.rootHash);
+  });
+
+  test("childThread on parent state node points to child's final state and is in refs", async () => {
+    const cas = createCasStore(casDir);
+    const childFinalHash = "CHILD_FINAL_001";
+
+    const parentWf: WorkflowFn = async function* (
+      _thread: ThreadContext,
+      runtime: WorkflowRuntime,
+    ): AsyncGenerator<RoleOutput, WorkflowCompletion> {
+      const h = await runtime.cas.put("developer output");
+      yield {
+        role: "developer",
+        contentHash: h,
+        meta: { action: "delegate" },
+        refs: [h],
+        childThread: childFinalHash,
+      };
+      return { returnCode: 0, summary: "parent complete" };
+    };
+
+    const result = await executeThread(
+      parentWf,
+      "parent-wf",
+      { prompt: "parent task", steps: [] },
+      makeOptions({ storageRoot }),
+      {
+        threadId: "P_THREAD_02",
+        hash: "CTHREAD_TEST01",
+        infoJsonlPath: join(storageRoot, "logs", "CTHREAD_TEST01", "P2.info.jsonl"),
+        cas,
+      },
+      noLogger(),
+    );
+
+    const endNode = await loadStateNode(cas, result.rootHash);
+    const devStateHash = endNode.payload.ancestors[0] ?? "";
+    const devNode = await loadStateNode(cas, devStateHash);
+
+    expect(devNode.payload.role).toBe("developer");
+    expect(devNode.payload.childThread).toBe(childFinalHash);
+    expect(devNode.refs).toContain(childFinalHash);
+  });
+
+  test("parent state with no child has childThread: null", async () => {
+    const cas = createCasStore(casDir);
+
+    const wf: WorkflowFn = async function* (
+      _thread: ThreadContext,
+      runtime: WorkflowRuntime,
+    ): AsyncGenerator<RoleOutput, WorkflowCompletion> {
+      const h = await runtime.cas.put("prep output");
+      yield { role: "preparer", contentHash: h, meta: {}, refs: [h], childThread: null };
+      return { returnCode: 0, summary: "done" };
+    };
+
+    const result = await executeThread(
+      wf,
+      "test-wf",
+      { prompt: "task", steps: [] },
+      makeOptions({ storageRoot }),
+      {
+        threadId: "NULL_CT_01",
+        hash: "NULLCT_TEST001",
+        infoJsonlPath: join(storageRoot, "logs", "NULLCT_TEST001", "N1.info.jsonl"),
+        cas,
+      },
+      noLogger(),
+    );
+
+    const endNode = await loadStateNode(cas, result.rootHash);
+    const prepHash = endNode.payload.ancestors[0] ?? "";
+    const prepNode = await loadStateNode(cas, prepHash);
+
+    expect(prepNode.payload.childThread).toBeNull();
+    expect(prepNode.refs).not.toContain(null);
+  });
+
+  test("full bidirectional: child parentState is traversable to parent's context", async () => {
+    const cas = createCasStore(casDir);
+    const parentHash = "BIDIR_PARENT01";
+
+    const parentWf: WorkflowFn = async function* (
+      _thread: ThreadContext,
+      runtime: WorkflowRuntime,
+    ): AsyncGenerator<RoleOutput, WorkflowCompletion> {
+      const h1 = await runtime.cas.put("preparation output");
+      yield {
+        role: "preparer",
+        contentHash: h1,
+        meta: { repoPath: "/test" },
+        refs: [h1],
+        childThread: null,
+      };
+      const h2 = await runtime.cas.put("developer output");
+      yield {
+        role: "developer",
+        contentHash: h2,
+        meta: { action: "code" },
+        refs: [h2],
+        childThread: "CHILD_END_HASH1",
+      };
+      return { returnCode: 0, summary: "all done" };
+    };
+
+    const observedHeads: string[] = [];
+    const opts = makeOptions({
+      storageRoot,
+      awaitAfterEachYield: async () => {
+        const bundleDir = join(storageRoot, "bundles", parentHash);
+        const text = await readFile(join(bundleDir, "threads.json"), "utf8");
+        const parsed = JSON.parse(text) as Record<string, { head: string }>;
+        const head = parsed.BIDIR_T_001?.head ?? null;
+        if (head !== null) observedHeads.push(head);
+      },
+    });
+
+    await executeThread(
+      parentWf,
+      "bidir-wf",
+      { prompt: "bidir test", steps: [] },
+      opts,
+      {
+        threadId: "BIDIR_T_001",
+        hash: parentHash,
+        infoJsonlPath: join(storageRoot, "logs", parentHash, "BD1.info.jsonl"),
+        cas,
+      },
+      noLogger(),
+    );
+
+    expect(observedHeads.length).toBe(2);
+    const preparerStateHash = observedHeads[0] ?? "";
+
+    // Execute child with parentState pointing to parent's preparer state
+    // biome-ignore lint/correctness/useYield: testing start-only path
+    const childWf: WorkflowFn = async function* (
+      _t: ThreadContext,
+      _r: WorkflowRuntime,
+    ): AsyncGenerator<RoleOutput, WorkflowCompletion> {
+      return { returnCode: 0, summary: "child ok" };
+    };
+
+    const childResult = await executeThread(
+      childWf,
+      "bidir-child",
+      { prompt: "child bidir", steps: [] },
+      makeOptions({ storageRoot, depth: 1, parentStateHash: preparerStateHash }),
+      {
+        threadId: "BIDIR_C_001",
+        hash: "BIDIR_CHILD001",
+        infoJsonlPath: join(storageRoot, "logs", "BIDIR_CHILD001", "BC1.info.jsonl"),
+        cas,
+      },
+      noLogger(),
+    );
+
+    // Upward traversal: child start → parentState → preparer state → meta.repoPath
+    const childStart = await loadStartNode(cas, childResult.rootHash);
+    expect(childStart.payload.parentState).toBe(preparerStateHash);
+
+    const parentPrep = await loadStateNode(cas, preparerStateHash);
+    expect(parentPrep.payload.meta.repoPath).toBe("/test");
+  });
+});

packages/workflow-execute/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@uncaged/workflow-execute",
-  "version": "0.2.0",
+  "version": "0.3.1",
   "type": "module",
   "exports": {
     ".": {

packages/workflow-execute/pnpm-lock.yaml

@@ -0,0 +1,51 @@
+lockfileVersion: '9.0'
+
+settings:
+  autoInstallPeers: true
+  excludeLinksFromLockfile: false
+
+importers:
+
+  .:
+    dependencies:
+      '@uncaged/workflow-cas':
+        specifier: workspace:*
+        version: link:../workflow-cas
+      '@uncaged/workflow-protocol':
+        specifier: workspace:*
+        version: link:../workflow-protocol
+      '@uncaged/workflow-reactor':
+        specifier: workspace:*
+        version: link:../workflow-reactor
+      '@uncaged/workflow-register':
+        specifier: workspace:*
+        version: link:../workflow-register
+      '@uncaged/workflow-runtime':
+        specifier: workspace:*
+        version: link:../workflow-runtime
+      '@uncaged/workflow-util':
+        specifier: workspace:*
+        version: link:../workflow-util
+      yaml:
+        specifier: ^2.7.1
+        version: 2.8.4
+    devDependencies:
+      zod:
+        specifier: ^4.0.0
+        version: 4.4.3
+
+packages:
+
+  yaml@2.8.4:
+    resolution: {integrity: sha512-ml/JPOj9fOQK8RNnWojA67GbZ0ApXAUlN2UQclwv2eVgTgn7O9gg9o7paZWKMp4g0H3nTLtS9LVzhkpOFIKzog==}
+    engines: {node: '>= 14.6'}
+    hasBin: true
+
+  zod@4.4.3:
+    resolution: {integrity: sha512-ytENFjIJFl2UwYglde2jchW2Hwm4GJFLDiSXWdTrJQBIN9Fcyp7n4DhxJEiWNAJMV1/BqWfW/kkg71UDcHJyTQ==}
+
+snapshots:
+
+  yaml@2.8.4: {}
+
+  zod@4.4.3: {}

packages/workflow-execute/src/cas-reactor.ts

@@ -0,0 +1,79 @@
+import type { CasStore } from "@uncaged/workflow-cas";
+import type { ThreadReactorFn } from "@uncaged/workflow-reactor";
+import { createLlmFn, createThreadReactor } from "@uncaged/workflow-reactor";
+import type { LlmProvider } from "@uncaged/workflow-runtime";
+
+import { extractFunctionToolFromZodSchema } from "./extract/index.js";
+
+export type CasReactorThread = {
+  cas: CasStore;
+};
+
+const CAS_GET_TOOL_DEFINITION = {
+  type: "function" as const,
+  function: {
+    name: "cas_get",
+    description:
+      "Read a Merkle DAG node from content-addressed storage by its hash. Returns YAML-formatted node with type, payload, and refs or children fields (content nodes use refs).",
+    parameters: {
+      type: "object",
+      properties: {
+        hash: { type: "string", description: "The CAS hash to retrieve" },
+      },
+      required: ["hash"],
+    },
+  },
+};
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
+export type CasReactorOpts = {
+  maxRounds: number;
+  systemPromptForStructuredTool: (structuredToolName: string) => string;
+};
+
+export function createCasReactor(
+  provider: LlmProvider,
+  cas: CasStore,
+  opts: CasReactorOpts,
+): ThreadReactorFn<CasReactorThread> {
+  return createThreadReactor<CasReactorThread>({
+    llm: createLlmFn(provider),
+    maxRounds: opts.maxRounds,
+    staticTools: [CAS_GET_TOOL_DEFINITION],
+    structuredToolFromSchema: (schema) => {
+      const t = extractFunctionToolFromZodSchema(schema);
+      return {
+        name: t.name,
+        tool: {
+          type: "function" as const,
+          function: {
+            name: t.name,
+            description: t.description,
+            parameters: t.parameters,
+          },
+        },
+      };
+    },
+    systemPromptForStructuredTool: opts.systemPromptForStructuredTool,
+    toolHandler: async (call, _thread) => {
+      if (call.function.name !== "cas_get") {
+        return `Unknown tool: ${call.function.name}`;
+      }
+      let hash: string;
+      try {
+        const ta = JSON.parse(call.function.arguments) as unknown;
+        if (!isRecord(ta) || typeof ta.hash !== "string") {
+          return 'cas_get requires {"hash": "<cas-hash>"}.';
+        }
+        hash = ta.hash;
+      } catch {
+        return "cas_get arguments were not valid JSON.";
+      }
+      const blob = await cas.get(hash);
+      return blob === null ? "null" : blob;
+    },
+  });
+}

packages/workflow-execute/src/engine/engine.ts

@@ -26,6 +26,7 @@ import { END, START } from "@uncaged/workflow-runtime";
 import { err, type LogFn, ok, type Result } from "@uncaged/workflow-util";
 
 import { createExtract } from "../extract/index.js";
+import { createSummarizer, type SummarizeFn } from "./summarizer.js";
 import { runSupervisor } from "./supervisor.js";
 import {
   appendThreadHistoryEntry,
@@ -33,20 +34,12 @@ import {
   removeThreadEntry,
   upsertThreadEntry,
 } from "./threads-index.js";
-import type { ExecuteThreadIo, ExecuteThreadOptions } from "./types.js";
+import type { ChainState, ExecuteThreadIo, ExecuteThreadOptions } from "./types.js";
+import { EMPTY_CHAIN_STATE } from "./types.js";
 
 /** Cap for {@link StateNode}.payload.ancestors: 1 parent + 10 skip-list. */
 const ANCESTORS_CAP = 11;
 
-type ChainState = {
-  /** State hash of the most recently written {@link StateNode}, or `null` before the first step. */
-  parentStateHash: string | null;
-  /** Ancestors recorded on the most recently written {@link StateNode}. */
-  parentAncestors: readonly string[];
-};
-
-const EMPTY_CHAIN: ChainState = { parentStateHash: null, parentAncestors: [] };
-
 function computeAncestors(chain: ChainState): string[] {
   if (chain.parentStateHash === null) {
     return [];
@@ -61,6 +54,7 @@ async function resolveEngineRegistryRuntime(
   Result<
     {
       extract: ReturnType<typeof createExtract>;
+      summarize: SummarizeFn;
       workflowConfig: WorkflowConfig;
     },
     string
@@ -84,7 +78,11 @@ async function resolveEngineRegistryRuntime(
     apiKey: ex.apiKey,
     model: ex.model,
   };
-  return ok({ extract: createExtract(llmProvider, { cas }), workflowConfig: cfg });
+  return ok({
+    extract: createExtract(llmProvider, { cas }),
+    summarize: createSummarizer(llmProvider, cas),
+    workflowConfig: cfg,
+  });
 }
 
 async function appendStateForStep(params: {
@@ -96,6 +94,7 @@ async function appendStateForStep(params: {
   meta: Record<string, unknown>;
   refs: readonly string[];
   timestamp: number;
+  childThread: string | null;
 }): Promise<{ stateHash: string; chain: ChainState }> {
   const text = await getContentMerklePayload(params.cas, params.contentHash);
   if (text === null) {
@@ -114,6 +113,7 @@ async function appendStateForStep(params: {
     ancestors,
     compact: null,
     timestamp: params.timestamp,
+    childThread: params.childThread,
   };
   const stateHash = await putStateNode(params.cas, payload);
   return {
@@ -139,6 +139,7 @@ async function appendEndState(params: {
     ancestors,
     compact: null,
     timestamp: params.timestamp,
+    childThread: null,
   };
   return putStateNode(params.cas, payload);
 }
@@ -219,17 +220,17 @@ async function maybeSupervisorHaltsThread(params: {
     params.logger("K6PW9NYT", `supervisor skipped: ${sup.error}`);
     return null;
   }
-  if (sup.value !== "stop") {
+  if (sup.value !== "kill") {
     return null;
   }
-  params.logger("M4QX8VHN", `thread ${params.threadId} stopped by supervisor`);
+  params.logger("M4QX8VHN", `thread ${params.threadId} killed by supervisor`);
   return finalizeThread({
     cas: params.cas,
     bundleDir: params.bundleDir,
     threadId: params.threadId,
     startHash: params.startHash,
     chain: params.chain,
-    completion: { returnCode: 0, summary: "completed: supervisor stopped thread" },
+    completion: { returnCode: 1, summary: "killed: supervisor detected pathological behavior" },
   });
 }
 
@@ -258,6 +259,7 @@ async function driveWorkflowGenerator(params: {
   bundleDir: string;
   startHash: string;
   chain: ChainState;
+  summarize: SummarizeFn;
 }): Promise<WorkflowResult> {
   const {
     fn,
@@ -270,6 +272,7 @@ async function driveWorkflowGenerator(params: {
     cas,
     bundleDir,
     startHash,
+    summarize,
   } = params;
   let chain: ChainState = params.chain;
   const gen = fn(thread, runtime);
@@ -278,6 +281,10 @@ async function driveWorkflowGenerator(params: {
     role: s.role,
     summary: JSON.stringify(s.meta),
   }));
+  const summarizerSteps: { role: string; contentHash: string }[] = thread.steps.map((s) => ({
+    role: s.role,
+    contentHash: s.contentHash,
+  }));
 
   while (true) {
     if (executeOptions.signal.aborted) {
@@ -292,32 +299,24 @@ async function driveWorkflowGenerator(params: {
       });
     }
 
-    if (written >= executeOptions.maxRounds) {
-      logger("R3CW7YBQ", `thread ${threadId} stopped at maxRounds=${executeOptions.maxRounds}`);
-      return await finalizeThread({
-        cas,
-        bundleDir,
-        threadId,
-        startHash,
-        chain,
-        completion: {
-          returnCode: 0,
-          summary: `completed: reached maxRounds (${executeOptions.maxRounds})`,
-        },
-      });
-    }
-
     const iterResult = await gen.next();
 
     if (iterResult.done) {
       logger("F3HN8QKP", `thread ${threadId} generator finished`);
+      const rawCompletion = iterResult.value;
+      const llmSummary = await summarize({
+        prompt: thread.start.content,
+        recentSteps: summarizerSteps,
+        fallback: rawCompletion.summary,
+        logger,
+      });
       return await finalizeThread({
         cas,
         bundleDir,
         threadId,
         startHash,
         chain,
-        completion: iterResult.value,
+        completion: { ...rawCompletion, summary: llmSummary },
       });
     }
 
@@ -333,6 +332,7 @@ async function driveWorkflowGenerator(params: {
       meta: step.meta,
       refs: step.refs,
       timestamp: ts,
+      childThread: step.childThread ?? null,
     });
     chain = written_.chain;
     await publishHead({ bundleDir, threadId, startHash, headHash: written_.stateHash });
@@ -343,6 +343,7 @@ async function driveWorkflowGenerator(params: {
       role: step.role,
       summary: JSON.stringify(step.meta),
     });
+    summarizerSteps.push({ role: step.role, contentHash: step.contentHash });
 
     await Promise.race([
       executeOptions.awaitAfterEachYield(),
@@ -391,7 +392,7 @@ async function driveWorkflowGenerator(params: {
  * Persistence layout (RFC v3 — CAS-based thread storage):
  * - Thread chain is written as immutable CAS blobs: a single {@link StartNode}
  *   plus one {@link StateNode} per role step (including a final `__end__`
- *   state on completion / abort / `maxRounds`).
+ *   state on completion / abort).
  * - The active thread head is published in `<bundleDir>/threads.json`; on
  *   completion it is removed and a record is appended to
  *   `<bundleDir>/history/{YYYY-MM-DD}.jsonl`.
@@ -408,36 +409,56 @@ export async function executeThread(
   await mkdir(dirname(io.infoJsonlPath), { recursive: true });
 
   const prefilled = options.prefilledDiskSteps;
+  const fork = options.forkContinuation;
+
+  if (fork !== null && prefilled !== null) {
+    throw new Error("forkContinuation and prefilledDiskSteps cannot both be set");
+  }
+
   if (prefilled !== null && prefilled.length !== input.steps.length) {
     throw new Error(
       `prefilledDiskSteps length (${prefilled.length}) must match input.steps length (${input.steps.length})`,
     );
   }
 
+  const replayTs = options.replayTimestamps;
+  if (replayTs !== null && replayTs.length !== input.steps.length) {
+    throw new Error(
+      `replayTimestamps length (${replayTs.length}) must match input.steps length (${input.steps.length})`,
+    );
+  }
+
   const bundleDir = getBundleDir(options.storageRoot, io.hash);
 
-  const promptHash = await io.cas.put(input.prompt);
-  const startHash = await putStartNode(
-    io.cas,
-    {
-      name: workflowName,
-      hash: io.hash,
-      maxRounds: options.maxRounds,
-      depth: options.depth,
-    },
-    promptHash,
-  );
+  let startHash: string;
 
-  await publishHead({
-    bundleDir,
-    threadId: io.threadId,
-    startHash,
-    headHash: startHash,
-  });
+  if (fork !== null) {
+    startHash = fork.startHash;
+    logger("T9HQ2KHM", `thread ${io.threadId} continued fork for workflow ${workflowName}`);
+  } else {
+    const promptHash = await io.cas.put(input.prompt);
+    startHash = await putStartNode(
+      io.cas,
+      {
+        name: workflowName,
+        hash: io.hash,
+        depth: options.depth,
+        parentState: options.parentStateHash,
+      },
+      promptHash,
+    );
 
-  logger("T9HQ2KHM", `thread ${io.threadId} started for workflow ${workflowName}`);
+    await publishHead({
+      bundleDir,
+      threadId: io.threadId,
+      startHash,
+      headHash: startHash,
+    });
 
-  let chain: ChainState = EMPTY_CHAIN;
+    logger("T9HQ2KHM", `thread ${io.threadId} started for workflow ${workflowName}`);
+  }
+
+  let chain: ChainState = fork !== null ? fork.initialChain : EMPTY_CHAIN_STATE;
 
   if (prefilled !== null) {
     for (const row of prefilled) {
@@ -450,6 +471,7 @@ export async function executeThread(
         meta: row.meta,
         refs: row.refs,
         timestamp: row.timestamp,
+        childThread: null,
       });
       chain = written.chain;
       await publishHead({
@@ -463,21 +485,6 @@ export async function executeThread(
 
   const nowMs = Date.now();
 
-  if (options.maxRounds <= 0) {
-    logger("R3CW7YBQ", `thread ${io.threadId} stopped at maxRounds=${options.maxRounds}`);
-    return await finalizeThread({
-      cas: io.cas,
-      bundleDir,
-      threadId: io.threadId,
-      startHash,
-      chain,
-      completion: {
-        returnCode: 0,
-        summary: `completed: reached maxRounds (${options.maxRounds})`,
-      },
-    });
-  }
-
   const registryRuntime = await resolveEngineRegistryRuntime(options.storageRoot, io.cas);
   if (!registryRuntime.ok) {
     throw new Error(registryRuntime.error);
@@ -486,18 +493,20 @@ export async function executeThread(
   const thread: ThreadContext = {
     threadId: io.threadId,
     depth: options.depth,
+    bundleHash: io.hash,
     start: {
       role: START,
       content: input.prompt,
-      meta: { maxRounds: options.maxRounds },
+      meta: {},
       timestamp: nowMs,
+      parentState: options.parentStateHash,
     },
     steps: input.steps.map((out, i) => ({
       role: out.role,
       contentHash: out.contentHash,
       meta: out.meta,
       refs: out.refs,
-      timestamp: prefilled?.[i]?.timestamp ?? nowMs + i,
+      timestamp: replayTs?.[i] ?? prefilled?.[i]?.timestamp ?? nowMs + i,
     })),
   };
 
@@ -518,5 +527,6 @@ export async function executeThread(
     bundleDir,
     startHash,
     chain,
+    summarize: registryRuntime.value.summarize,
   });
 }

packages/workflow-execute/src/engine/fork-thread.ts

@@ -1,9 +1,29 @@
-import type { WorkflowCompletion } from "@uncaged/workflow-runtime";
-import { err, normalizeRefsField, ok, type Result } from "@uncaged/workflow-util";
+import type { CasStore } from "@uncaged/workflow-cas";
+import { parseCasThreadNode, putContentNodeWithRefs, putStateNode } from "@uncaged/workflow-cas";
+import type { StateNodePayload } from "@uncaged/workflow-protocol";
+import type { RoleOutput, WorkflowCompletion } from "@uncaged/workflow-runtime";
+import { END } from "@uncaged/workflow-runtime";
+import { err, ok, type Result } from "@uncaged/workflow-util";
+import { parse as parseYaml } from "yaml";
 
-import type { ForkHistoricalStep, ForkPlan, ParsedThreadStartRecord } from "./types.js";
+import { upsertThreadEntry } from "./threads-index.js";
+import type { CasForkPlan, ChainState, ForkContinuationOptions } from "./types.js";
+import { EMPTY_CHAIN_STATE } from "./types.js";
 
-/** Recognizes a persisted workflow completion line (no `role`; has numeric `returnCode` and string `summary`). Omits `rootHash` when absent. */
+/** Internal branch marker; skipped when presenting fork selection / replay slices. */
+export const FORK_BRANCH_ROLE = "__fork__";
+
+/** Cap for {@link StateNodePayload}.ancestors: 1 parent + 10 skip-list. */
+const ANCESTORS_CAP = 11;
+
+function computeAncestors(chain: ChainState): string[] {
+  if (chain.parentStateHash === null) {
+    return [];
+  }
+  return [chain.parentStateHash, ...chain.parentAncestors].slice(0, ANCESTORS_CAP);
+}
+
+/** Recognizes a persisted workflow completion line (no `role`; has numeric `returnCode` and string `summary`). */
 export function tryParseWorkflowResultRecord(
   obj: Record<string, unknown>,
 ): WorkflowCompletion | null {
@@ -18,227 +38,287 @@ export function tryParseWorkflowResultRecord(
   return { returnCode, summary };
 }
 
-export function tryParseRoleStepRecord(obj: Record<string, unknown>): ForkHistoricalStep | null {
-  const role = obj.role;
-  const contentHash = obj.contentHash;
-  const meta = obj.meta;
-  const timestamp = obj.timestamp;
-  if (typeof role !== "string") {
-    return null;
-  }
-  if (typeof contentHash !== "string") {
-    return null;
-  }
-  if (meta === null || typeof meta !== "object") {
-    return null;
-  }
-  if (typeof timestamp !== "number") {
-    return null;
-  }
-  return {
-    role,
-    contentHash,
-    meta: meta as Record<string, unknown>,
-    refs: normalizeRefsField(obj.refs),
-    timestamp,
-  };
-}
-
-function parseRoleLine(
-  obj: Record<string, unknown>,
-  lineIndex: number,
-): Result<ForkHistoricalStep, string> {
-  const parsed = tryParseRoleStepRecord(obj);
-  if (parsed === null) {
-    return err(`invalid role record at line ${lineIndex}`);
-  }
-  return ok(parsed);
-}
-
-function parseStartRecordLine(firstLine: string): Result<ParsedThreadStartRecord, string> {
-  let startParsed: unknown;
-  try {
-    startParsed = JSON.parse(firstLine) as unknown;
-  } catch {
-    return err("invalid JSON on line 1 (start record)");
-  }
-  if (startParsed === null || typeof startParsed !== "object") {
-    return err("invalid start record shape");
-  }
-  const startRec = startParsed as Record<string, unknown>;
-  const name = startRec.name;
-  const hash = startRec.hash;
-  const threadId = startRec.threadId;
-  const parameters = startRec.parameters;
-  if (typeof name !== "string" || typeof hash !== "string" || typeof threadId !== "string") {
-    return err("start record missing name, hash, or threadId");
-  }
-  if (parameters === null || typeof parameters !== "object") {
-    return err("start record missing parameters");
-  }
-  const paramsRec = parameters as Record<string, unknown>;
-  const prompt = paramsRec.prompt;
-  const options = paramsRec.options;
-  if (typeof prompt !== "string") {
-    return err("start record missing parameters.prompt");
-  }
-  if (options === null || typeof options !== "object") {
-    return err("start record missing parameters.options");
-  }
-  const optRec = options as Record<string, unknown>;
-  const maxRounds = optRec.maxRounds;
-  if (typeof maxRounds !== "number") {
-    return err("start record missing parameters.options.maxRounds");
-  }
-
-  const depthRaw = optRec.depth;
-  const depth =
-    typeof depthRaw === "number" && Number.isFinite(depthRaw) ? Math.trunc(depthRaw) : 0;
-
-  return ok({
-    workflowName: name,
-    hash,
-    threadId,
-    prompt,
-    maxRounds,
-    depth,
-  });
-}
-
-function parseFollowingRoleLines(lines: string[]): Result<ForkHistoricalStep[], string> {
-  const roleSteps: ForkHistoricalStep[] = [];
-  for (let i = 1; i < lines.length; i++) {
-    const line = lines[i];
-    if (line === undefined) {
+/** Walk {@link StateNode} hashes from head toward the first step (newest → oldest). */
+export async function walkStateFramesNewestFirst(
+  cas: CasStore,
+  headHash: string,
+): Promise<Array<{ hash: string; payload: StateNodePayload }>> {
+  const frames: Array<{ hash: string; payload: StateNodePayload }> = [];
+  let cur = headHash;
+  while (true) {
+    const yamlText = await cas.get(cur);
+    if (yamlText === null) {
       break;
     }
-    let rec: unknown;
-    try {
-      rec = JSON.parse(line) as unknown;
-    } catch {
-      return err(`invalid JSON at line ${i + 1}`);
-    }
-    if (rec === null || typeof rec !== "object") {
-      return err(`invalid record at line ${i + 1}`);
-    }
-    const recObj = rec as Record<string, unknown>;
-    const wf = tryParseWorkflowResultRecord(recObj);
-    if (wf !== null) {
-      if (i !== lines.length - 1) {
-        return err("WorkflowResult record must be the final line in `.data.jsonl`");
-      }
+    const parsed = parseCasThreadNode(yamlText);
+    if (parsed === null || parsed.kind !== "state") {
       break;
     }
-    const parsed = parseRoleLine(recObj, i + 1);
-    if (!parsed.ok) {
-      return parsed;
+    frames.push({ hash: cur, payload: parsed.node.payload });
+    const ancestors = parsed.node.payload.ancestors;
+    if (ancestors.length === 0) {
+      break;
     }
-    roleSteps.push(parsed.value);
+    const parent = ancestors[0];
+    if (parent === undefined || parent === "") {
+      break;
+    }
+    cur = parent;
   }
-  return ok(roleSteps);
+  return frames;
 }
 
-/**
- * Parse RFC-001 `.data.jsonl`: line 1 start record, line 2+ role outputs.
- */
-export function parseThreadDataJsonl(text: string): Result<
-  {
-    start: ParsedThreadStartRecord;
-    roleSteps: ForkHistoricalStep[];
-  },
-  string
-> {
-  const lines = text
-    .split("\n")
-    .map((l) => l.trim())
-    .filter((l) => l !== "");
-  if (lines.length === 0) {
-    return err("thread data is empty");
-  }
-
-  const firstLine = lines[0];
-  if (firstLine === undefined) {
-    return err("thread data is empty");
-  }
-
-  const start = parseStartRecordLine(firstLine);
-  if (!start.ok) {
-    return start;
-  }
-
-  const roleSteps = parseFollowingRoleLines(lines);
-  if (!roleSteps.ok) {
-    return roleSteps;
-  }
-
-  return ok({
-    start: start.value,
-    roleSteps: roleSteps.value,
-  });
-}
-
-function orderedUniqueRoles(roleSteps: ForkHistoricalStep[]): string[] {
+function orderedUniqueRoles(roles: string[]): string[] {
   const seen = new Set<string>();
   const out: string[] = [];
-  for (const s of roleSteps) {
-    if (!seen.has(s.role)) {
-      seen.add(s.role);
-      out.push(s.role);
+  for (const r of roles) {
+    if (!seen.has(r)) {
+      seen.add(r);
+      out.push(r);
     }
   }
   return out;
 }
 
-/**
- * Select historical steps for a fork:
- * - `fromRole === null`: drop the last step (retry the last role).
- * - `fromRole !== null`: keep steps through the first occurrence of that role (inclusive).
- */
-export function selectForkHistoricalSteps(
-  roleSteps: ForkHistoricalStep[],
+async function readPromptText(cas: CasStore, promptHash: string): Promise<Result<string, string>> {
+  const yamlText = await cas.get(promptHash);
+  if (yamlText === null) {
+    return err(`prompt CAS blob missing: ${promptHash}`);
+  }
+  let raw: unknown;
+  try {
+    raw = parseYaml(yamlText) as unknown;
+  } catch {
+    return err(`prompt CAS blob is not valid YAML: ${promptHash}`);
+  }
+  if (raw === null || typeof raw !== "object") {
+    return err(`prompt CAS blob has unexpected shape: ${promptHash}`);
+  }
+  const payload = (raw as Record<string, unknown>).payload;
+  if (typeof payload !== "string") {
+    return err(`prompt CAS blob missing string payload: ${promptHash}`);
+  }
+  return ok(payload);
+}
+
+async function readStartWorkflowIdentity(params: {
+  cas: CasStore;
+  startHash: string;
+}): Promise<Result<{ workflowName: string; depth: number; prompt: string }, string>> {
+  const yamlText = await params.cas.get(params.startHash);
+  if (yamlText === null) {
+    return err(`start node missing in CAS: ${params.startHash}`);
+  }
+  const parsed = parseCasThreadNode(yamlText);
+  if (parsed === null || parsed.kind !== "start") {
+    return err(`CAS blob is not a StartNode: ${params.startHash}`);
+  }
+  const refs = parsed.node.refs;
+  const promptHash = refs[0];
+  if (typeof promptHash !== "string") {
+    return err("StartNode refs[0] must be the prompt hash");
+  }
+  const prompt = await readPromptText(params.cas, promptHash);
+  if (!prompt.ok) {
+    return prompt;
+  }
+  const p = parsed.node.payload;
+  return ok({
+    workflowName: p.name,
+    depth: p.depth,
+    prompt: prompt.value,
+  });
+}
+
+async function payloadToRoleOutput(cas: CasStore, payload: StateNodePayload): Promise<RoleOutput> {
+  let refs: string[] = [];
+  const blob = await cas.get(payload.content);
+  if (blob !== null) {
+    const cn = parseCasThreadNode(blob);
+    if (cn?.kind === "content") {
+      refs = [...cn.node.refs];
+    }
+  }
+  return {
+    role: payload.role,
+    contentHash: payload.content,
+    meta: payload.meta,
+    refs,
+    childThread: payload.childThread,
+  };
+}
+
+function meaningfulFramesOldestFirst(
+  newestFirst: Array<{ hash: string; payload: StateNodePayload }>,
+): Array<{ hash: string; payload: StateNodePayload }> {
+  const chronological = [...newestFirst].reverse();
+  return chronological.filter((f) => f.payload.role !== END && f.payload.role !== FORK_BRANCH_ROLE);
+}
+
+function selectForkPointStateHash(
+  meaningfulOldestFirst: Array<{ hash: string; payload: StateNodePayload }>,
   fromRole: string | null,
-): Result<ForkHistoricalStep[], string> {
-  if (roleSteps.length === 0) {
+): Result<string | null, string> {
+  if (meaningfulOldestFirst.length === 0) {
     return err("thread has no completed role steps to fork from");
   }
 
   if (fromRole === null) {
-    if (roleSteps.length === 1) {
-      return ok([]);
+    if (meaningfulOldestFirst.length === 1) {
+      return ok(null);
     }
-    return ok(roleSteps.slice(0, -1));
+    const forkFrame = meaningfulOldestFirst[meaningfulOldestFirst.length - 2];
+    if (forkFrame === undefined) {
+      return err("thread has no completed role steps to fork from");
+    }
+    return ok(forkFrame.hash);
   }
 
-  const idx = roleSteps.findIndex((s) => s.role === fromRole);
+  const idx = meaningfulOldestFirst.findIndex((f) => f.payload.role === fromRole);
   if (idx < 0) {
-    const available = orderedUniqueRoles(roleSteps);
+    const available = orderedUniqueRoles(meaningfulOldestFirst.map((f) => f.payload.role));
     return err(`role not found in thread: ${fromRole} (available: ${available.join(", ")})`);
   }
-  return ok(roleSteps.slice(0, idx + 1));
+  const forkFrame = meaningfulOldestFirst[idx];
+  if (forkFrame === undefined) {
+    return err("fork frame missing");
+  }
+  return ok(forkFrame.hash);
+}
+
+function replayFramesThroughForkPoint(
+  meaningfulOldestFirst: Array<{ hash: string; payload: StateNodePayload }>,
+  forkPointHash: string | null,
+): Array<{ hash: string; payload: StateNodePayload }> {
+  if (forkPointHash === null) {
+    return [];
+  }
+  const idx = meaningfulOldestFirst.findIndex((f) => f.hash === forkPointHash);
+  if (idx < 0) {
+    return [];
+  }
+  return meaningfulOldestFirst.slice(0, idx + 1);
+}
+
+async function buildForkContinuation(params: {
+  cas: CasStore;
+  sourceThreadId: string;
+  startHash: string;
+  forkPointStateHash: string | null;
+}): Promise<Result<ForkContinuationOptions, string>> {
+  const { cas, sourceThreadId, startHash, forkPointStateHash } = params;
+
+  if (forkPointStateHash === null) {
+    return ok({
+      startHash,
+      forkHeadHash: startHash,
+      initialChain: EMPTY_CHAIN_STATE,
+    });
+  }
+
+  const yamlText = await cas.get(forkPointStateHash);
+  if (yamlText === null) {
+    return err(`fork point state missing in CAS: ${forkPointStateHash}`);
+  }
+  const parsed = parseCasThreadNode(yamlText);
+  if (parsed === null || parsed.kind !== "state") {
+    return err(`fork point blob is not a StateNode: ${forkPointStateHash}`);
+  }
+  const fpPayload = parsed.node.payload;
+
+  const chainBefore: ChainState = {
+    parentStateHash: forkPointStateHash,
+    parentAncestors: fpPayload.ancestors,
+  };
+  const ancestorsMarker = computeAncestors(chainBefore);
+
+  const emptyContentHash = await putContentNodeWithRefs(cas, "", []);
+  const markerPayload: StateNodePayload = {
+    role: FORK_BRANCH_ROLE,
+    meta: { forkFrom: sourceThreadId },
+    start: startHash,
+    content: emptyContentHash,
+    ancestors: ancestorsMarker,
+    compact: null,
+    timestamp: Date.now(),
+    childThread: null,
+  };
+  const markerHash = await putStateNode(cas, markerPayload);
+
+  const initialChain: ChainState = {
+    parentStateHash: markerHash,
+    parentAncestors: ancestorsMarker,
+  };
+
+  return ok({
+    startHash,
+    forkHeadHash: markerHash,
+    initialChain,
+  });
 }
 
 /**
- * Read `.data.jsonl` text and compute fork payload for the worker `run` command.
+ * Prepare a CAS fork: writes the branch marker {@link StateNode}, registers `threads.json`,
+ * and returns worker payload fields (shared {@link StartNode}, zero ancestor duplication).
  */
-export function buildForkPlan(
-  dataJsonlText: string,
-  fromRole: string | null,
-): Result<ForkPlan, string> {
-  const parsed = parseThreadDataJsonl(dataJsonlText);
-  if (!parsed.ok) {
-    return parsed;
+export async function prepareCasFork(params: {
+  cas: CasStore;
+  bundleDir: string;
+  bundleHash: string;
+  sourceThreadId: string;
+  headHash: string;
+  startHash: string;
+  newThreadId: string;
+  fromRole: string | null;
+}): Promise<Result<CasForkPlan, string>> {
+  const id = await readStartWorkflowIdentity({
+    cas: params.cas,
+    startHash: params.startHash,
+  });
+  if (!id.ok) {
+    return id;
   }
-  const selected = selectForkHistoricalSteps(parsed.value.roleSteps, fromRole);
-  if (!selected.ok) {
-    return selected;
+
+  const newestFirst = await walkStateFramesNewestFirst(params.cas, params.headHash);
+  const meaningful = meaningfulFramesOldestFirst(newestFirst);
+
+  const forkPoint = selectForkPointStateHash(meaningful, params.fromRole);
+  if (!forkPoint.ok) {
+    return forkPoint;
   }
-  const { start } = parsed.value;
+
+  const replayFrames = replayFramesThroughForkPoint(meaningful, forkPoint.value);
+  const steps: RoleOutput[] = [];
+  const stepTimestamps: number[] = [];
+  for (const fr of replayFrames) {
+    steps.push(await payloadToRoleOutput(params.cas, fr.payload));
+    stepTimestamps.push(fr.payload.timestamp);
+  }
+
+  const cont = await buildForkContinuation({
+    cas: params.cas,
+    sourceThreadId: params.sourceThreadId,
+    startHash: params.startHash,
+    forkPointStateHash: forkPoint.value,
+  });
+  if (!cont.ok) {
+    return cont;
+  }
+
+  await upsertThreadEntry(params.bundleDir, params.newThreadId, {
+    head: cont.value.forkHeadHash,
+    start: params.startHash,
+    updatedAt: Date.now(),
+  });
+
   return ok({
-    workflowName: start.workflowName,
-    hash: start.hash,
-    sourceThreadId: start.threadId,
-    prompt: start.prompt,
-    runOptions: { maxRounds: start.maxRounds, depth: start.depth },
-    historicalSteps: selected.value,
+    workflowName: id.value.workflowName,
+    hash: params.bundleHash,
+    sourceThreadId: params.sourceThreadId,
+    prompt: id.value.prompt,
+    runOptions: { depth: id.value.depth },
+    steps,
+    stepTimestamps,
+    forkContinuation: cont.value,
   });
 }

packages/workflow-execute/src/engine/gc.ts

@@ -1,122 +1,182 @@
-import { readdir, readFile } from "node:fs/promises";
+import type { Stats } from "node:fs";
+import { readdir, readFile, stat } from "node:fs/promises";
 import { join } from "node:path";
-import { type CasStore, createCasStore } from "@uncaged/workflow-cas";
+import { type CasStore, createCasStore, findReachableHashes } from "@uncaged/workflow-cas";
 import { err, getGlobalCasDir, ok, type Result } from "@uncaged/workflow-util";
-import { parseThreadDataJsonl } from "./fork-thread.js";
+
+import type { ThreadHistoryEntry, ThreadIndex } from "./threads-index.js";
+import { readThreadsIndex } from "./threads-index.js";
 import type { GcResult } from "./types.js";
 
-async function listThreadDataJsonlPaths(storageRoot: string): Promise<Result<string[], string>> {
-  const logsRoot = join(storageRoot, "logs");
-  const paths: string[] = [];
-  let hashes: string[];
+function isPlainObject(v: unknown): v is Record<string, unknown> {
+  return v !== null && typeof v === "object" && !Array.isArray(v);
+}
+
+function parseHistoryLine(jsonLine: string): ThreadHistoryEntry | null {
+  let raw: unknown;
   try {
-    hashes = await readdir(logsRoot);
+    raw = JSON.parse(jsonLine) as unknown;
+  } catch {
+    return null;
+  }
+  if (!isPlainObject(raw)) {
+    return null;
+  }
+  const threadId = raw.threadId;
+  const head = raw.head;
+  const start = raw.start;
+  const completedAt = raw.completedAt;
+  if (
+    typeof threadId !== "string" ||
+    typeof head !== "string" ||
+    typeof start !== "string" ||
+    typeof completedAt !== "number"
+  ) {
+    return null;
+  }
+  return { threadId, head, start, completedAt };
+}
+
+// biome-ignore lint/complexity/noExcessiveCognitiveComplexity: walks threads index + optional history dir
+async function collectGcRootsFromBundle(bundleDir: string): Promise<Result<string[], string>> {
+  const roots: string[] = [];
+
+  let activeIndex: ThreadIndex;
+  try {
+    activeIndex = await readThreadsIndex(bundleDir);
+  } catch (e) {
+    return err(`failed to read threads.json under ${bundleDir}: ${String(e)}`);
+  }
+
+  for (const entry of Object.values(activeIndex)) {
+    roots.push(entry.head);
+    roots.push(entry.start);
+  }
+
+  const histDir = join(bundleDir, "history");
+  let histFiles: string[];
+  try {
+    histFiles = await readdir(histDir);
+  } catch (e) {
+    const errObj = e as NodeJS.ErrnoException;
+    if (errObj.code === "ENOENT") {
+      return ok(roots);
+    }
+    return err(`failed to read history directory ${histDir}: ${String(e)}`);
+  }
+
+  for (const name of histFiles) {
+    if (!name.endsWith(".jsonl")) {
+      continue;
+    }
+    let text: string;
+    try {
+      text = await readFile(join(histDir, name), "utf8");
+    } catch (e) {
+      return err(`failed to read history file ${name}: ${String(e)}`);
+    }
+    for (const line of text.split("\n")) {
+      const trimmed = line.trim();
+      if (trimmed === "") {
+        continue;
+      }
+      const entry = parseHistoryLine(trimmed);
+      if (entry === null) {
+        continue;
+      }
+      roots.push(entry.head);
+      roots.push(entry.start);
+    }
+  }
+
+  return ok(roots);
+}
+
+async function collectAllGcRoots(storageRoot: string): Promise<Result<string[], string>> {
+  const bundlesRoot = join(storageRoot, "bundles");
+  let entries: string[];
+  try {
+    entries = await readdir(bundlesRoot);
   } catch (e) {
     const errObj = e as NodeJS.ErrnoException;
     if (errObj.code === "ENOENT") {
       return ok([]);
     }
-    return err(`failed to read logs directory: ${String(e)}`);
+    return err(`failed to read bundles directory: ${String(e)}`);
   }
 
-  for (const hash of hashes) {
-    const dir = join(logsRoot, hash);
-    let entries: string[];
+  const roots: string[] = [];
+  for (const name of entries) {
+    const bundleDir = join(bundlesRoot, name);
+    let st: Stats;
     try {
-      entries = await readdir(dir);
+      st = await stat(bundleDir);
     } catch {
       continue;
     }
-    for (const fileName of entries) {
-      if (fileName.endsWith(".data.jsonl")) {
-        paths.push(join(dir, fileName));
-      }
+    if (!st.isDirectory()) {
+      continue;
     }
+    const chunk = await collectGcRootsFromBundle(bundleDir);
+    if (!chunk.ok) {
+      return chunk;
+    }
+    roots.push(...chunk.value);
   }
 
-  paths.sort();
-  return ok(paths);
+  return ok(roots);
 }
 
-async function collectActiveRefsFromDataPaths(
-  dataPaths: string[],
-): Promise<Result<Set<string>, string>> {
-  const activeRefs = new Set<string>();
-  for (const dataPath of dataPaths) {
-    let text: string;
-    try {
-      text = await readFile(dataPath, "utf8");
-    } catch (e) {
-      return err(`failed to read ${dataPath}: ${String(e)}`);
-    }
-    const parsed = parseThreadDataJsonl(text);
-    if (!parsed.ok) {
-      return err(`${dataPath}: ${parsed.error}`);
-    }
-    for (const step of parsed.value.roleSteps) {
-      for (const ref of step.refs) {
-        activeRefs.add(ref);
-      }
-    }
-  }
-  return ok(activeRefs);
-}
-
-async function deleteCasNotInSet(
-  cas: CasStore,
-  activeRefs: Set<string>,
-): Promise<Result<string[], string>> {
+async function deleteCasNotMarked(cas: CasStore, marked: ReadonlySet<string>): Promise<string[]> {
   let listed: string[];
   try {
     listed = await cas.list();
   } catch (e) {
-    return err(`failed to list cas entries: ${String(e)}`);
+    throw new Error(`failed to list cas entries: ${String(e)}`);
   }
 
   const deletedHashes: string[] = [];
   for (const hash of listed) {
-    if (activeRefs.has(hash)) {
+    if (marked.has(hash)) {
       continue;
     }
     try {
       await cas.delete(hash);
     } catch (e) {
-      return err(`failed to delete cas ${hash}: ${String(e)}`);
+      throw new Error(`failed to delete cas ${hash}: ${String(e)}`);
     }
     deletedHashes.push(hash);
   }
 
   deletedHashes.sort();
-  return ok(deletedHashes);
+  return deletedHashes;
 }
 
 /**
- * Mark-and-sweep CAS GC: collect `refs` from all thread `.data.jsonl` files under `storageRoot`,
- * then delete CAS blobs not referenced by any surviving thread data.
+ * Mark-and-sweep CAS GC: roots are every `head` / `start` hash from `threads.json` and
+ * `history/*.jsonl` across bundle dirs; marks closure via `refs[]`; deletes unreachable blobs.
  */
 export async function garbageCollectCas(storageRoot: string): Promise<Result<GcResult, string>> {
-  const pathsResult = await listThreadDataJsonlPaths(storageRoot);
-  if (!pathsResult.ok) {
-    return pathsResult;
+  const rootsResult = await collectAllGcRoots(storageRoot);
+  if (!rootsResult.ok) {
+    return rootsResult;
   }
-  const paths = pathsResult.value;
-
-  const refsResult = await collectActiveRefsFromDataPaths(paths);
-  if (!refsResult.ok) {
-    return refsResult;
-  }
-  const activeRefs = refsResult.value;
+  const roots = rootsResult.value;
 
   const cas = createCasStore(getGlobalCasDir(storageRoot));
-  const deletedResult = await deleteCasNotInSet(cas, activeRefs);
-  if (!deletedResult.ok) {
-    return deletedResult;
+
+  const marked = await findReachableHashes(roots, cas);
+
+  let deletedHashes: string[];
+  try {
+    deletedHashes = await deleteCasNotMarked(cas, marked);
+  } catch (e) {
+    return err(String(e));
   }
-  const deletedHashes = deletedResult.value;
 
   return ok({
-    scannedThreads: paths.length,
-    activeRefs: activeRefs.size,
+    scannedThreads: roots.length,
+    activeRefs: marked.size,
     deletedEntries: deletedHashes.length,
     deletedHashes,
   });

packages/workflow-execute/src/engine/index.ts

@@ -1,11 +1,10 @@
 export { createWorkflow } from "./create-workflow.js";
 export { executeThread } from "./engine.js";
 export {
-  buildForkPlan,
-  parseThreadDataJsonl,
-  selectForkHistoricalSteps,
-  tryParseRoleStepRecord,
+  FORK_BRANCH_ROLE,
+  prepareCasFork,
   tryParseWorkflowResultRecord,
+  walkStateFramesNewestFirst,
 } from "./fork-thread.js";
 export { garbageCollectCas } from "./gc.js";
 export { createThreadPauseGate } from "./thread-pause-gate.js";
@@ -13,18 +12,22 @@ export type { ThreadHistoryEntry, ThreadIndex, ThreadIndexEntry } from "./thread
 export {
   appendThreadHistoryEntry,
   getBundleDir,
+  readThreadsIndex,
   removeThreadEntry,
+  removeThreadHistoryEntries,
   upsertThreadEntry,
+  writeThreadsIndex,
 } from "./threads-index.js";
 export type {
+  CasForkPlan,
+  ChainState,
   ExecuteThreadIo,
   ExecuteThreadOptions,
-  ForkHistoricalStep,
-  ForkPlan,
+  ForkContinuationOptions,
   GcResult,
-  ParsedThreadStartRecord,
   PrefilledDiskStep,
   SupervisorDecision,
   ThreadPauseGate,
 } from "./types.js";
+export { EMPTY_CHAIN_STATE } from "./types.js";
 export { getWorkerHostScriptPath } from "./worker-entry-path.js";

packages/workflow-execute/src/engine/summarizer.ts

@@ -0,0 +1,56 @@
+import type { CasStore } from "@uncaged/workflow-cas";
+import type { LlmProvider } from "@uncaged/workflow-runtime";
+import type { LogFn } from "@uncaged/workflow-util";
+import * as z from "zod/v4";
+
+import { createCasReactor } from "../cas-reactor.js";
+
+/** Max ReAct rounds: 3 cas_get reads + 1 structured output = 4 rounds is sufficient. */
+const SUMMARIZER_MAX_REACT_ROUNDS = 4;
+/** Only pass the last N steps; each step is just a role+contentHash reference (~60 chars), not full content. */
+const SUMMARIZER_RECENT_STEP_LIMIT = 20;
+
+const summarySchema = z.object({ summary: z.string() }).meta({
+  title: "workflow_summary",
+  description: "A concise summary of the completed workflow's results and outcome.",
+});
+
+function buildSummarizerInput(args: {
+  prompt: string;
+  recentSteps: readonly { role: string; contentHash: string }[];
+}): string {
+  const recent = args.recentSteps.slice(-SUMMARIZER_RECENT_STEP_LIMIT);
+  const stepsBlock = recent
+    .map((s, i) => `${i + 1}. [${s.role}] contentHash: ${s.contentHash}`)
+    .join("\n");
+  return `Original task:\n${args.prompt}\n\nCompleted steps (oldest first):\n${stepsBlock === "" ? "(none)" : stepsBlock}\n\nUse cas_get to read step content if needed. Summarize the workflow outcome concisely.`;
+}
+
+export type SummarizeFn = (args: {
+  prompt: string;
+  recentSteps: readonly { role: string; contentHash: string }[];
+  fallback: string;
+  logger: LogFn;
+}) => Promise<string>;
+
+export function createSummarizer(provider: LlmProvider, cas: CasStore): SummarizeFn {
+  const reactor = createCasReactor(provider, cas, {
+    maxRounds: SUMMARIZER_MAX_REACT_ROUNDS,
+    systemPromptForStructuredTool: (structuredToolName) =>
+      `You summarize completed workflow threads. You have access to cas_get to read step content by hash. After reviewing the steps, call the ${structuredToolName} tool with a concise summary of the workflow outcome and results. Or reply with only a JSON object such as {"summary":"..."}.`,
+  });
+
+  return async (args) => {
+    const result = await reactor({
+      thread: { cas },
+      input: buildSummarizerInput(args),
+      schema: summarySchema,
+    });
+    if (!result.ok) {
+      args.logger("P2WX7KNR", `summarizer failed: ${result.error}`);
+      return args.fallback;
+    }
+    args.logger("Q5MT3VBF", "summarizer produced workflow summary");
+    return result.value.summary;
+  };
+}

packages/workflow-execute/src/engine/supervisor.ts

@@ -1,10 +1,9 @@
-import * as z from "zod/v4";
-
-import { resolveModel } from "@uncaged/workflow-register";
-import { extractFunctionToolFromZodSchema } from "../extract/index.js";
 import { createLlmFn, createThreadReactor } from "@uncaged/workflow-reactor";
 import type { WorkflowConfig } from "@uncaged/workflow-register";
+import { resolveModel } from "@uncaged/workflow-register";
 import { err, type LogFn, ok, type Result } from "@uncaged/workflow-util";
+import * as z from "zod/v4";
+import { extractFunctionToolFromZodSchema } from "../extract/index.js";
 
 import type { SupervisorDecision } from "./types.js";
 
@@ -13,12 +12,12 @@ const SUPERVISOR_MAX_REACT_ROUNDS = 4;
 
 const supervisorDecisionSchema = z
   .object({
-    decision: z.enum(["continue", "stop"]),
+    decision: z.enum(["continue", "kill"]),
   })
   .meta({
     title: "supervisor_decision",
     description:
-      'Workflow supervisor decision. "continue" when the thread is making progress; "stop" when done, looping, or stuck.',
+      'Workflow supervisor decision. "continue" when the thread is making progress or following its normal role sequence; "kill" only when the thread is stuck in an infinite loop, producing no meaningful progress, or has gone off the rails. Normal workflow completion is handled by the moderator — the supervisor should NOT kill a thread just because it looks done.',
   });
 
 type SupervisorThreadContext = Record<string, never>;
@@ -64,7 +63,7 @@ export async function runSupervisor(
       };
     },
     systemPromptForStructuredTool: (structuredToolName) =>
-      `You supervise a multi-step workflow. Decide whether the thread should keep running or halt. Reply with "continue" when the thread is making progress toward the task, or "stop" when it is finished, looping, or no longer making progress. Call the ${structuredToolName} tool with JSON arguments matching the schema, or reply with only a JSON object such as {"decision":"stop"}.`,
+      `You supervise a multi-step workflow. Your job is to detect pathological situations — NOT to decide when the workflow is "done" (that is the moderator's job). Reply with "continue" when the thread is making progress or following its normal role sequence. Reply with "kill" ONLY when the thread is stuck in an infinite loop, producing repetitive/meaningless output, or has clearly gone off the rails. Call the ${structuredToolName} tool with JSON arguments matching the schema, or reply with only a JSON object such as {"decision":"continue"}.`,
     toolHandler: async (call) => `Unknown tool: ${call.function.name}`,
   });
 

packages/workflow-execute/src/engine/threads-index.ts

@@ -1,6 +1,8 @@
-import { appendFile, mkdir, readFile, rename, writeFile } from "node:fs/promises";
+import { appendFile, mkdir, readdir, readFile, rename, writeFile } from "node:fs/promises";
 import { dirname, join } from "node:path";
 
+import { err, ok, type Result } from "@uncaged/workflow-util";
+
 /**
  * Active-thread index entry stored in `<bundleDir>/threads.json`.
  *
@@ -71,7 +73,8 @@ function parseThreadIndex(text: string): ThreadIndex {
   return out;
 }
 
-async function readThreadIndex(bundleDir: string): Promise<ThreadIndex> {
+/** Read `<bundleDir>/threads.json` (empty object when missing or invalid). */
+export async function readThreadsIndex(bundleDir: string): Promise<ThreadIndex> {
   const path = threadsJsonPath(bundleDir);
   let text: string;
   try {
@@ -86,7 +89,7 @@ async function readThreadIndex(bundleDir: string): Promise<ThreadIndex> {
   return parseThreadIndex(text);
 }
 
-async function writeThreadIndex(bundleDir: string, index: ThreadIndex): Promise<void> {
+export async function writeThreadsIndex(bundleDir: string, index: ThreadIndex): Promise<void> {
   const path = threadsJsonPath(bundleDir);
   await mkdir(dirname(path), { recursive: true });
   const tmp = `${path}.tmp.${process.pid}.${Date.now()}`;
@@ -101,19 +104,19 @@ export async function upsertThreadEntry(
   threadId: string,
   entry: ThreadIndexEntry,
 ): Promise<void> {
-  const index = await readThreadIndex(bundleDir);
+  const index = await readThreadsIndex(bundleDir);
   index[threadId] = entry;
-  await writeThreadIndex(bundleDir, index);
+  await writeThreadsIndex(bundleDir, index);
 }
 
 /** Remove a thread entry from `threads.json` (no-op when absent). */
 export async function removeThreadEntry(bundleDir: string, threadId: string): Promise<void> {
-  const index = await readThreadIndex(bundleDir);
+  const index = await readThreadsIndex(bundleDir);
   if (!(threadId in index)) {
     return;
   }
   delete index[threadId];
-  await writeThreadIndex(bundleDir, index);
+  await writeThreadsIndex(bundleDir, index);
 }
 
 function dateKey(epochMs: number): string {
@@ -134,3 +137,63 @@ export async function appendThreadHistoryEntry(
   const line = `${JSON.stringify(entry)}\n`;
   await appendFile(path, line, "utf8");
 }
+
+/** Removes every `history/*.jsonl` line whose `threadId` matches (rewrite files in place). */
+// biome-ignore lint/complexity/noExcessiveCognitiveComplexity: per-file JSONL filtering keeps RM deterministic
+export async function removeThreadHistoryEntries(
+  bundleDir: string,
+  threadId: string,
+): Promise<Result<number, string>> {
+  const histRoot = join(bundleDir, "history");
+  let files: string[];
+  try {
+    files = await readdir(histRoot);
+  } catch (e) {
+    const errObj = e as NodeJS.ErrnoException;
+    if (errObj.code === "ENOENT") {
+      return ok(0);
+    }
+    return err(`failed to read history directory: ${String(e)}`);
+  }
+
+  let removed = 0;
+  for (const name of files) {
+    if (!name.endsWith(".jsonl")) {
+      continue;
+    }
+    const path = join(histRoot, name);
+    let text: string;
+    try {
+      text = await readFile(path, "utf8");
+    } catch {
+      continue;
+    }
+    const kept: string[] = [];
+    for (const line of text.split("\n")) {
+      const trimmed = line.trim();
+      if (trimmed === "") {
+        continue;
+      }
+      let rec: unknown;
+      try {
+        rec = JSON.parse(trimmed) as unknown;
+      } catch {
+        kept.push(`${trimmed}\n`);
+        continue;
+      }
+      if (rec === null || typeof rec !== "object") {
+        kept.push(`${trimmed}\n`);
+        continue;
+      }
+      const id = (rec as Record<string, unknown>).threadId;
+      if (id === threadId) {
+        removed++;
+        continue;
+      }
+      kept.push(`${trimmed}\n`);
+    }
+    await writeFile(path, kept.join(""), "utf8");
+  }
+
+  return ok(removed);
+}

packages/workflow-execute/src/engine/types.ts

@@ -2,7 +2,7 @@ import type { CasStore } from "@uncaged/workflow-cas";
 import type { RoleOutput } from "@uncaged/workflow-runtime";
 import type { Result } from "@uncaged/workflow-util";
 
-export type SupervisorDecision = "continue" | "stop";
+export type SupervisorDecision = "continue" | "kill";
 
 export type ExecuteThreadIo = {
   threadId: string;
@@ -11,7 +11,25 @@ export type ExecuteThreadIo = {
   cas: CasStore;
 };
 
-/** One persisted role line in `.data.jsonl` (engine adds these for fork replay before running the generator). */
+/** CAS chain tail state before the next appended {@link StateNode}. */
+export type ChainState = {
+  parentStateHash: string | null;
+  parentAncestors: readonly string[];
+};
+
+export const EMPTY_CHAIN_STATE: ChainState = { parentStateHash: null, parentAncestors: [] };
+
+/**
+ * When forking, the worker continues from an existing {@link StartNode} plus an optional
+ * branch marker {@link StateNode} instead of allocating a new start blob.
+ */
+export type ForkContinuationOptions = {
+  startHash: string;
+  forkHeadHash: string;
+  initialChain: ChainState;
+};
+
+/** One replayed role step (prefill) before the generator runs (same layout as disk replay rows). */
 export type PrefilledDiskStep = {
   role: string;
   contentHash: string;
@@ -21,46 +39,46 @@ export type PrefilledDiskStep = {
 };
 
 export type ExecuteThreadOptions = {
-  maxRounds: number;
   /** Passed to the bundle thread context as `ThreadContext.depth`. */
   depth: number;
+  /** Parent thread's head state hash at spawn time; `null` for top-level threads. */
+  parentStateHash: string | null;
   signal: AbortSignal;
   /** Invoked after each successful yield (and outer-loop checks); used for pause/resume. */
   awaitAfterEachYield: () => Promise<void>;
   /** When non-null, written into the start record so tooling can trace lineage. */
   forkSourceThreadId: string | null;
   /**
-   * Written to `.data.jsonl` immediately after the start record, before the generator runs.
+   * When non-null, replays these steps into CAS before the generator runs.
    * Must match `input.steps` length and order when present.
    */
   prefilledDiskSteps: PrefilledDiskStep[] | null;
+  /** When non-null, skip creating a new {@link StartNode} and continue this CAS chain. */
+  forkContinuation: ForkContinuationOptions | null;
+  /**
+   * When non-null, must match `input.steps.length`; supplies persisted timestamps for
+   * {@link ThreadContext.steps} (used when restoring history without prefilled CAS replay).
+   */
+  replayTimestamps: readonly number[] | null;
   /** Workspace root containing `workflow.yaml`; used to resolve the `extract` scene for meta extraction. */
   storageRoot: string;
 };
 
-/** Role steps replayed from `.data.jsonl`, including persisted timestamps. */
-export type ForkHistoricalStep = RoleOutput & { timestamp: number };
-
-export type ParsedThreadStartRecord = {
-  workflowName: string;
-  hash: string;
-  threadId: string;
-  prompt: string;
-  maxRounds: number;
-  depth: number;
-};
-
-export type ForkPlan = {
+export type CasForkPlan = {
   workflowName: string;
   hash: string;
   sourceThreadId: string;
   prompt: string;
-  runOptions: { maxRounds: number; depth: number };
-  historicalSteps: ForkHistoricalStep[];
+  runOptions: { depth: number };
+  steps: RoleOutput[];
+  stepTimestamps: number[];
+  forkContinuation: ForkContinuationOptions;
 };
 
 export type GcResult = {
+  /** Count of root hashes seeded from thread indexes (`head`/`start` per entry). */
   scannedThreads: number;
+  /** Reachable CAS blobs after the mark phase. */
   activeRefs: number;
   deletedEntries: number;
   deletedHashes: string[];

packages/workflow-execute/src/engine/worker.ts

@@ -1,9 +1,13 @@
+import { unlinkSync } from "node:fs";
 import { mkdir, unlink, writeFile } from "node:fs/promises";
 import { createServer, type Socket } from "node:net";
 import { dirname, join } from "node:path";
-import type { RoleOutput, WorkflowFn } from "@uncaged/workflow-runtime";
-import { ensureUncagedWorkflowSymlink, importWorkflowBundleModule } from "@uncaged/workflow-register";
 import { createCasStore } from "@uncaged/workflow-cas";
+import {
+  ensureUncagedWorkflowSymlink,
+  importWorkflowBundleModule,
+} from "@uncaged/workflow-register";
+import type { RoleOutput, WorkflowFn } from "@uncaged/workflow-runtime";
 import {
   createLogger,
   err,
@@ -14,7 +18,12 @@ import {
 } from "@uncaged/workflow-util";
 import { executeThread } from "./engine.js";
 import { createThreadPauseGate } from "./thread-pause-gate.js";
-import type { ExecuteThreadIo, PrefilledDiskStep, ThreadPauseGate } from "./types.js";
+import type {
+  ExecuteThreadIo,
+  ForkContinuationOptions,
+  PrefilledDiskStep,
+  ThreadPauseGate,
+} from "./types.js";
 
 const bootLog = createLogger({ sink: { kind: "stderr" } });
 
@@ -23,11 +32,12 @@ type RunCommand = {
   threadId: string;
   workflowName: string;
   prompt: string;
-  options: { maxRounds: number; depth: number };
+  options: { depth: number };
   steps: RoleOutput[];
-  /** Timestamps aligned with `steps` for `.data.jsonl` replay; length must match `steps` when non-null. */
+  /** Timestamps aligned with `steps` for replay / fork restore; length must match `steps` when steps are non-empty. */
   stepTimestamps: number[] | null;
   forkSourceThreadId: string | null;
+  forkContinuation: ForkContinuationOptions | null;
 };
 
 type KillCommand = {
@@ -62,14 +72,17 @@ function parseRoleOutputRecord(obj: Record<string, unknown>): RoleOutput | null
   if (meta === null || typeof meta !== "object") {
     return null;
   }
+  const childThread = obj.childThread;
   return {
     role,
     contentHash,
     meta: meta as Record<string, unknown>,
     refs: normalizeRefsField(obj.refs),
+    childThread: typeof childThread === "string" ? childThread : null,
   };
 }
 
+// biome-ignore lint/complexity/noExcessiveCognitiveComplexity: mirrors permissive worker IPC decoding shape checks
 function parseRunStepsPayload(rec: Record<string, unknown>): {
   steps: RoleOutput[];
   stepTimestamps: number[] | null;
@@ -104,12 +117,60 @@ function parseRunStepsPayload(rec: Record<string, unknown>): {
       return null;
     }
   }
+
+  const parallelTsRaw = rec.stepTimestamps;
+  if (
+    steps.length > 0 &&
+    Array.isArray(parallelTsRaw) &&
+    parallelTsRaw.length === steps.length &&
+    parallelTsRaw.every((x): x is number => typeof x === "number")
+  ) {
+    return { steps, stepTimestamps: [...parallelTsRaw] };
+  }
+
   return {
     steps,
     stepTimestamps: anyTimestamp ? timestamps : null,
   };
 }
 
+function parseForkContinuation(rec: Record<string, unknown>): ForkContinuationOptions | null {
+  const raw = rec.forkContinuation;
+  if (raw === undefined || raw === null) {
+    return null;
+  }
+  if (typeof raw !== "object") {
+    return null;
+  }
+  const o = raw as Record<string, unknown>;
+  const startHash = o.startHash;
+  const forkHeadHash = o.forkHeadHash;
+  const ic = o.initialChain;
+  if (typeof startHash !== "string" || typeof forkHeadHash !== "string") {
+    return null;
+  }
+  if (ic === null || typeof ic !== "object") {
+    return null;
+  }
+  const ich = ic as Record<string, unknown>;
+  const pph = ich.parentStateHash;
+  const pa = ich.parentAncestors;
+  if (!(pph === null || typeof pph === "string")) {
+    return null;
+  }
+  if (!Array.isArray(pa) || !pa.every((x) => typeof x === "string")) {
+    return null;
+  }
+  return {
+    startHash,
+    forkHeadHash,
+    initialChain: {
+      parentStateHash: pph,
+      parentAncestors: pa,
+    },
+  };
+}
+
 function parseRunControlPayload(rec: Record<string, unknown>): RunCommand | null {
   const threadId = rec.threadId;
   const workflowName = rec.workflowName;
@@ -126,10 +187,6 @@ function parseRunControlPayload(rec: Record<string, unknown>): RunCommand | null
     return null;
   }
   const optRec = options as Record<string, unknown>;
-  const maxRounds = optRec.maxRounds;
-  if (typeof maxRounds !== "number") {
-    return null;
-  }
   const depthRaw = optRec.depth;
   const depth =
     typeof depthRaw === "number" && Number.isFinite(depthRaw) ? Math.trunc(depthRaw) : 0;
@@ -145,15 +202,17 @@ function parseRunControlPayload(rec: Record<string, unknown>): RunCommand | null
     }
     forkSourceThreadId = rawFork;
   }
+  const forkContinuation = parseForkContinuation(rec);
   return {
     type: "run",
     threadId,
     workflowName,
     prompt,
-    options: { maxRounds, depth },
+    options: { depth },
     steps: parsedSteps.steps,
     stepTimestamps: parsedSteps.stepTimestamps,
     forkSourceThreadId,
+    forkContinuation,
   };
 }
 
@@ -322,6 +381,23 @@ async function main(): Promise<void> {
   let activeThreads = 0;
   let shutdownTimer: ReturnType<typeof setTimeout> | null = null;
 
+  function cleanupAllRunningMarkersSync(): void {
+    for (const threadId of threads.keys()) {
+      try {
+        unlinkSync(join(storageRoot, "logs", hash, `${threadId}.running`));
+      } catch {
+        // ignore missing file or other fs errors
+      }
+    }
+  }
+
+  for (const sig of ["SIGINT", "SIGTERM"] as const) {
+    process.on(sig, () => {
+      cleanupAllRunningMarkersSync();
+      process.exit(sig === "SIGINT" ? 130 : 143);
+    });
+  }
+
   const cas = createCasStore(getGlobalCasDir(storageRoot));
 
   const workerCtlPath = join(storageRoot, "workers", `${hash}.json`);
@@ -354,6 +430,7 @@ async function main(): Promise<void> {
     }
   }
 
+  // biome-ignore lint/complexity/noExcessiveCognitiveComplexity: TCP worker multiplexes lifecycle + runs
   async function dispatchCommand(cmd: ControlCommand, socket: Socket | null): Promise<void> {
     if (cmd.type !== "run") {
       dispatchThreadLifecycleCommand(threads, socket, cmd);
@@ -391,7 +468,19 @@ async function main(): Promise<void> {
 
       const baseTs = Date.now();
       let prefilledDiskSteps: PrefilledDiskStep[] | null = null;
-      if (cmd.steps.length > 0) {
+      let replayTimestamps: readonly number[] | null = null;
+
+      if (cmd.forkContinuation !== null) {
+        if (
+          cmd.steps.length > 0 &&
+          (cmd.stepTimestamps === null || cmd.stepTimestamps.length !== cmd.steps.length)
+        ) {
+          bootLog("J5WQ8NXT", "forkContinuation requires stepTimestamps aligned with steps");
+          throw new Error("forkContinuation requires stepTimestamps aligned with steps");
+        }
+        replayTimestamps =
+          cmd.steps.length === 0 ? null : (cmd.stepTimestamps as readonly number[]);
+      } else if (cmd.steps.length > 0) {
         prefilledDiskSteps = cmd.steps.map((step, i) => {
           const ts = cmd.stepTimestamps?.[i];
           return {
@@ -410,10 +499,13 @@ async function main(): Promise<void> {
         { prompt: cmd.prompt, steps: cmd.steps },
         {
           ...cmd.options,
+          parentStateHash: null,
           signal: ac.signal,
           awaitAfterEachYield: () => pauseGate.awaitAfterYield(),
           forkSourceThreadId: cmd.forkSourceThreadId,
           prefilledDiskSteps,
+          forkContinuation: cmd.forkContinuation,
+          replayTimestamps,
           storageRoot,
         },
         io,
@@ -423,8 +515,8 @@ async function main(): Promise<void> {
       const message = e instanceof Error ? e.message : String(e);
       bootLog("Q3MN8YKW", `thread ${threadId} failed: ${message}`);
     } finally {
-      threads.delete(threadId);
       await unlink(runningPath).catch(() => {});
+      threads.delete(threadId);
       bumpDone();
       socket?.end();
     }

packages/workflow-execute/src/extract/extract-fn.ts

@@ -1,8 +1,8 @@
-import type { ExtractContext, ExtractFn, LlmProvider } from "@uncaged/workflow-runtime";
-import type * as z from "zod/v4";
 import { type CasStore, getContentMerklePayload } from "@uncaged/workflow-cas";
-import { createLlmFn, createThreadReactor } from "@uncaged/workflow-reactor";
-import { extractFunctionToolFromZodSchema } from "./llm-extract.js";
+import type { ExtractFn, ExtractResult, LlmProvider } from "@uncaged/workflow-runtime";
+import type * as z from "zod/v4";
+
+import { createCasReactor } from "../cas-reactor.js";
 
 export type ExtractDeps = {
   cas: CasStore;
@@ -10,65 +10,6 @@ export type ExtractDeps = {
 
 const MAX_REACT_ROUNDS = 10;
 
-const CAS_GET_TOOL_DEFINITION = {
-  type: "function" as const,
-  function: {
-    name: "cas_get",
-    description:
-      "Read a Merkle DAG node from content-addressed storage by its hash. Returns YAML-formatted node with type, payload, and children fields.",
-    parameters: {
-      type: "object",
-      properties: {
-        hash: { type: "string", description: "The CAS hash to retrieve" },
-      },
-      required: ["hash"],
-    },
-  },
-};
-
-export type ExtractThreadContext = {
-  cas: CasStore;
-};
-
-function isRecord(value: unknown): value is Record<string, unknown> {
-  return typeof value === "object" && value !== null && !Array.isArray(value);
-}
-
-/** Builds the user-side extraction prompt (thread + agent output + instruction). */
-export async function buildExtractUserContent(
-  ctx: ExtractContext,
-  prompt: string,
-  deps: ExtractDeps,
-): Promise<string> {
-  const lines: string[] = [];
-  lines.push(`## Role: ${ctx.currentRole.name}`);
-  lines.push(ctx.currentRole.systemPrompt);
-  lines.push("");
-  lines.push("## Task");
-  lines.push(ctx.start.content);
-  lines.push("");
-  if (ctx.steps.length > 0) {
-    lines.push("## Thread History");
-    for (const step of ctx.steps) {
-      const body = await getContentMerklePayload(deps.cas, step.contentHash);
-      if (body === null) {
-        throw new Error(`extract: missing CAS blob for step ${step.role}: ${step.contentHash}`);
-      }
-      lines.push(`### ${step.role}`);
-      lines.push(body);
-      lines.push(`Meta: ${JSON.stringify(step.meta)}`);
-      lines.push("");
-    }
-  }
-  lines.push("## Agent Output");
-  lines.push(ctx.agentContent);
-  lines.push("");
-  lines.push("## Extraction Instruction");
-  lines.push(prompt);
-
-  return lines.join("\n");
-}
-
 /**
  * Create an ExtractFn backed by an LLM provider.
  *
@@ -77,52 +18,21 @@ export async function buildExtractUserContent(
  * assistant reply as a short-circuit, which covers the legacy "single" extraction path.
  */
 export function createExtract(provider: LlmProvider, deps: ExtractDeps): ExtractFn {
-  const llm = createLlmFn(provider);
-  const reactor = createThreadReactor<ExtractThreadContext>({
-    llm,
+  const reactor = createCasReactor(provider, deps.cas, {
     maxRounds: MAX_REACT_ROUNDS,
-    staticTools: [CAS_GET_TOOL_DEFINITION],
-    structuredToolFromSchema: (schema) => {
-      const t = extractFunctionToolFromZodSchema(schema);
-      return {
-        name: t.name,
-        tool: {
-          type: "function" as const,
-          function: {
-            name: t.name,
-            description: t.description,
-            parameters: t.parameters,
-          },
-        },
-      };
-    },
     systemPromptForStructuredTool: (structuredToolName) =>
-      `You extract structured metadata from the agent output below. Use cas_get to read Merkle DAG nodes from CAS (YAML: type, payload, children) when the agent output references hashes you must traverse. When you have the complete structured object, call the ${structuredToolName} tool with JSON arguments matching the schema. You may instead reply with only a JSON object (no prose) when no tools are needed.`,
-    toolHandler: async (call, thread) => {
-      if (call.function.name !== "cas_get") {
-        return `Unexpected tool routed to handler: ${call.function.name}`;
-      }
-      let hash: string;
-      try {
-        const ta = JSON.parse(call.function.arguments) as unknown;
-        if (!isRecord(ta) || typeof ta.hash !== "string") {
-          return 'cas_get requires a JSON object with a string "hash" field.';
-        }
-        hash = ta.hash;
-      } catch {
-        return 'cas_get arguments were not valid JSON. Provide {"hash": "<cas-hash>"}.';
-      }
-      const blob = await thread.cas.get(hash);
-      return blob === null ? "null" : blob;
-    },
+      `You extract structured metadata from content. The content is from a CAS node. Use cas_get to read referenced nodes if needed. When ready, call the ${structuredToolName} tool with JSON matching the schema. You may instead reply with only a JSON object (no prose) when no tools are needed.`,
   });
 
   return async <T extends Record<string, unknown>>(
     schema: z.ZodType<T>,
-    prompt: string,
-    ctx: ExtractContext,
-  ): Promise<T> => {
-    const text = await buildExtractUserContent(ctx, prompt, deps);
+    contentHash: string,
+  ): Promise<ExtractResult<T>> => {
+    const payload = await getContentMerklePayload(deps.cas, contentHash);
+    if (payload === null) {
+      throw new Error(`extract: missing CAS content node for hash ${contentHash}`);
+    }
+    const text = `${payload}\n\nExtract structured metadata according to the schema.`;
     const result = await reactor({
       thread: { cas: deps.cas },
       input: text,
@@ -131,6 +41,10 @@ export function createExtract(provider: LlmProvider, deps: ExtractDeps): Extract
     if (!result.ok) {
       throw new Error(`extract failed: ${result.error}`);
     }
-    return result.value;
+    return {
+      meta: result.value,
+      contentPayload: payload,
+      refs: [],
+    };
   };
 }

packages/workflow-execute/src/extract/index.ts

@@ -1,8 +1,4 @@
-export {
-  buildExtractUserContent,
-  createExtract,
-  type ExtractThreadContext,
-} from "./extract-fn.js";
+export { createExtract } from "./extract-fn.js";
 export {
   extractFunctionToolFromZodSchema,
   llmErrorToCause,

packages/workflow-execute/src/extract/llm-extract.ts

@@ -1,6 +1,5 @@
-import * as z from "zod/v4";
-
 import { err, ok, type Result } from "@uncaged/workflow-util";
+import * as z from "zod/v4";
 
 import type { LlmError, LlmExtractArgs } from "./types.js";
 

packages/workflow-execute/src/index.ts

@@ -1,35 +1,45 @@
 export { createWorkflow } from "./engine/create-workflow.js";
 export { executeThread } from "./engine/engine.js";
 export {
-  buildForkPlan,
-  parseThreadDataJsonl,
-  selectForkHistoricalSteps,
-  tryParseRoleStepRecord,
+  FORK_BRANCH_ROLE,
+  prepareCasFork,
   tryParseWorkflowResultRecord,
+  walkStateFramesNewestFirst,
 } from "./engine/fork-thread.js";
 export { garbageCollectCas } from "./engine/gc.js";
 export { createThreadPauseGate } from "./engine/thread-pause-gate.js";
 export type {
+  ThreadHistoryEntry,
+  ThreadIndex,
+  ThreadIndexEntry,
+} from "./engine/threads-index.js";
+export {
+  appendThreadHistoryEntry,
+  getBundleDir,
+  readThreadsIndex,
+  removeThreadEntry,
+  removeThreadHistoryEntries,
+  upsertThreadEntry,
+  writeThreadsIndex,
+} from "./engine/threads-index.js";
+export type {
+  CasForkPlan,
+  ChainState,
   ExecuteThreadIo,
   ExecuteThreadOptions,
-  ForkHistoricalStep,
-  ForkPlan,
+  ForkContinuationOptions,
   GcResult,
-  ParsedThreadStartRecord,
   PrefilledDiskStep,
   SupervisorDecision,
   ThreadPauseGate,
 } from "./engine/types.js";
+export { EMPTY_CHAIN_STATE } from "./engine/types.js";
 export { getWorkerHostScriptPath } from "./engine/worker-entry-path.js";
+export type { ExtractFn, LlmError, LlmExtractArgs } from "./extract/index.js";
 export {
-  buildExtractUserContent,
   createExtract,
-  type ExtractThreadContext,
-} from "./extract/index.js";
-export {
   extractFunctionToolFromZodSchema,
   llmErrorToCause,
   llmExtract,
 } from "./extract/index.js";
-export type { ExtractFn, LlmError, LlmExtractArgs } from "./extract/index.js";
-export { workflowAsAgent, type WorkflowAsAgentOptions } from "./workflow-as-agent.js";
+export { type WorkflowAsAgentOptions, workflowAsAgent } from "./workflow-as-agent.js";

packages/workflow-execute/src/workflow-as-agent.ts

@@ -1,17 +1,20 @@
 import { join } from "node:path";
-import type { AgentContext, AgentFn } from "@uncaged/workflow-runtime";
-import { extractBundleExports } from "@uncaged/workflow-register";
 import { createCasStore } from "@uncaged/workflow-cas";
-import type { ExecuteThreadIo } from "./engine/index.js";
-import { executeThread } from "./engine/index.js";
 import type { WorkflowConfig } from "@uncaged/workflow-register";
-import { getRegisteredWorkflow, readWorkflowRegistry } from "@uncaged/workflow-register";
+import {
+  extractBundleExports,
+  getRegisteredWorkflow,
+  readWorkflowRegistry,
+} from "@uncaged/workflow-register";
+import type { AgentContext, AgentFn, AgentFnResult } from "@uncaged/workflow-runtime";
 import {
   createLogger,
   generateUlid,
   getDefaultWorkflowStorageRoot,
   getGlobalCasDir,
 } from "@uncaged/workflow-util";
+import type { ExecuteThreadIo } from "./engine/index.js";
+import { executeThread, getBundleDir, readThreadsIndex } from "./engine/index.js";
 
 const DEFAULT_WORKFLOW_AS_AGENT_MAX_DEPTH = 3;
 
@@ -34,6 +37,13 @@ function resolveWorkflowAsAgentStorageRoot(options: WorkflowAsAgentOptions | nul
   return getDefaultWorkflowStorageRoot();
 }
 
+async function readParentHeadState(storageRoot: string, ctx: AgentContext): Promise<string | null> {
+  const bundleDir = getBundleDir(storageRoot, ctx.bundleHash);
+  const index = await readThreadsIndex(bundleDir);
+  const entry = index[ctx.threadId] ?? null;
+  return entry !== null ? entry.head : null;
+}
+
 /**
  * Returns an {@link AgentFn} that runs another registered workflow in a new thread,
  * using the parent thread's initial prompt (`ctx.start.content`) as the child prompt.
@@ -42,30 +52,30 @@ export function workflowAsAgent(
   workflowName: string,
   options: WorkflowAsAgentOptions | null = null,
 ): AgentFn {
-  return async (ctx: AgentContext): Promise<string> => {
+  return async (ctx: AgentContext): Promise<AgentFnResult> => {
     const nextDepth = ctx.depth + 1;
 
     const storageRoot = resolveWorkflowAsAgentStorageRoot(options);
 
     const registryResult = await readWorkflowRegistry(storageRoot);
     if (!registryResult.ok) {
-      return `ERROR: failed to read workflow registry: ${registryResult.error.message}`;
+      return { output: `ERROR: failed to read workflow registry: ${registryResult.error.message}`, childThread: null };
     }
 
     const maxDepth = workflowAsAgentMaxDepth(registryResult.value.config);
     if (nextDepth > maxDepth) {
-      return `ERROR: workflow-as-agent depth limit exceeded (max ${maxDepth})`;
+      return { output: `ERROR: workflow-as-agent depth limit exceeded (max ${maxDepth})`, childThread: null };
     }
 
     const entry = getRegisteredWorkflow(registryResult.value, workflowName);
     if (entry === null) {
-      return `ERROR: workflow "${workflowName}" not found in registry`;
+      return { output: `ERROR: workflow "${workflowName}" not found in registry`, childThread: null };
     }
 
     const bundlePath = join(storageRoot, "bundles", `${entry.hash}.esm.js`);
     const bundleExportsResult = await extractBundleExports(bundlePath, { storageRoot });
     if (!bundleExportsResult.ok) {
-      return `ERROR: ${bundleExportsResult.error}`;
+      return { output: `ERROR: ${bundleExportsResult.error}`, childThread: null };
     }
 
     const input = {
@@ -86,27 +96,32 @@ export function workflowAsAgent(
     const logger = createLogger({ sink: { kind: "file", path: infoJsonlPath } });
     const signalNever = new AbortController();
 
+    const parentHeadState = await readParentHeadState(storageRoot, ctx);
+
     try {
       const result = await executeThread(
         bundleExportsResult.value.run,
         workflowName,
         input,
         {
-          maxRounds: ctx.start.meta.maxRounds,
           depth: nextDepth,
+          parentStateHash: parentHeadState,
           signal: signalNever.signal,
           awaitAfterEachYield: async () => {},
           forkSourceThreadId: ctx.threadId,
           prefilledDiskSteps: null,
+          forkContinuation: null,
+          replayTimestamps: null,
           storageRoot,
         },
         io,
         logger,
       );
-      return result.rootHash;
+      const summary = `Child workflow "${workflowName}" completed (returnCode=${result.returnCode}).\n\nSummary: ${result.summary}\n\nChild thread root hash: ${result.rootHash}`;
+      return { output: summary, childThread: result.rootHash };
     } catch (e) {
       const message = e instanceof Error ? e.message : String(e);
-      return `ERROR: ${message}`;
+      return { output: `ERROR: ${message}`, childThread: null };
     }
   };
 }

packages/workflow-gateway/package.json

@@ -0,0 +1,17 @@
+{
+  "name": "@uncaged/workflow-gateway",
+  "version": "0.1.0",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "dev": "wrangler dev",
+    "deploy": "wrangler deploy"
+  },
+  "dependencies": {
+    "hono": "^4.7.11"
+  },
+  "devDependencies": {
+    "@cloudflare/workers-types": "^4.20260425.1",
+    "wrangler": "^4.20.0"
+  }
+}