docs/wf-stateless-design.md

@@ -200,7 +200,7 @@ payload:
 
 - `roles` — 内联定义，每个 role 的 `meta` 是独立的 ocas_ref（指向 ocas 内置 JSON Schema 节点）
 - `graph` — `Record<Role | "$START", Record<Status, Target>>`，每个 Target = `{ role, prompt }`
-- Status 来自上一个 role 输出的 `status` 字段，`$START` 用 `_` 作为初始 status
+- Status 来自上一个 role 输出的 `$status` 字段，`$START` 使用 `new`（首次启动）和 `resume`（恢复已完成的 thread）作为 status
 - Prompt 模板使用 Mustache 渲染，变量来自 lastOutput
 - 不含 agent binding — agent 配置在 `~/.uwf/config.yaml` 中管理
 
@@ -208,7 +208,7 @@ Moderator 的求值逻辑：
 
 ```typescript
 evaluate(graph, lastRole, lastOutput) → { role, prompt }
-// 1. status = lastRole === "$START" ? "_" : lastOutput.status
+// 1. status = lastOutput.$status (e.g. "new" for $START first run, "resume" for completed thread resume)
 // 2. target = graph[lastRole][status]
 // 3. prompt = mustache.render(target.prompt, lastOutput)
 ```
@@ -422,8 +422,8 @@ type StepNodePayload = StepRecord & {
 Moderator 使用 `evaluate(graph, lastRole, lastOutput)` 进行同步 status-based routing：
 
 ```typescript
-// graph[lastRole][lastOutput.status] → Target { role, prompt }
+// graph[lastRole][lastOutput.$status] → Target { role, prompt }
-// $START 角色使用 "_" 作为初始 status
+// $START 使用 "new"（首次启动）和 "resume"（恢复已完成 thread）作为 status
 // prompt 通过 Mustache 模板渲染，变量来自 lastOutput
 ```
 

packages/util/src/index.ts

@@ -15,7 +15,7 @@ export {
   validateFrontmatter,
 } from "./frontmatter-markdown/index.js";
 export { createLogger } from "./logger.js";
-export { generateModeratorReference } from "./moderator-reference.js";
+
 export type {
   CreateProcessLoggerOptions,
   ProcessLogFn,
@@ -35,4 +35,3 @@ export { extractUlidTimestamp, generateUlid } from "./ulid.js";
 export { generateUsageReference } from "./usage-reference.js";
 export { VERSION } from "./version.js";
 export { generateWorkflowAuthoringReference } from "./workflow-authoring-reference.js";
-export { generateYamlReference } from "./yaml-reference.js";

packages/util/src/moderator-reference.ts

@@ -1,57 +0,0 @@
-export function generateModeratorReference(): string {
-  return `# Moderator Reference
-
-## Overview
-
-The moderator is the workflow engine's routing component. It evaluates the directed graph defined in the workflow YAML to determine the next role (or \`$END\`) after each step — with zero LLM cost.
-
-## Status-Based Routing
-
-The moderator uses **status-based routing**: it inspects the previous step's extracted output (specifically the \`$status\` field) and looks up the corresponding edge in the graph.
-
-### Graph Structure
-
-The graph is a nested map: \`Record<Role | "$START", Record<Status, Target>>\`. Each role maps its possible \`$status\` values to a target with a \`role\` and \`prompt\`:
-
-\`\`\`yaml
-graph:
-  $START:
-    new: { role: planner, prompt: "Analyze the issue." }
-    resume: { role: planner, prompt: "Review the previous run output and continue." }
-  planner:
-    ready: { role: developer, prompt: "Implement the plan (CAS hash: {{{plan}}})." }
-    insufficient_info: { role: $END, prompt: "Not enough info." }
-  developer:
-    done: { role: reviewer, prompt: "Review branch {{{branch}}} at {{{worktree}}}." }
-    failed: { role: $END, prompt: "Developer failed: {{{reason}}}." }
-  reviewer:
-    approved: { role: tester, prompt: "Run tests on {{{branch}}} at {{{worktree}}}." }
-    rejected: { role: developer, prompt: "Fix issues: {{{comments}}}." }
-\`\`\`
-
-### Routing Algorithm
-
-1. Look up \`graph[lastRole]\` to get the status map for the current role
-2. Look up \`statusMap[lastOutput.$status]\` to get the target
-3. If target role is \`$END\`, mark thread as completed
-4. Otherwise, render the edge prompt (Mustache templates with \`{{{field}}}\` from output) and spawn the next agent
-
-### Edge Prompts and Mustache Templates
-
-Edge prompts use triple-brace Mustache syntax (\`{{{field}}}\`) to interpolate values from the previous step's output into the next agent's task prompt. This passes structured data (branch names, file paths, CAS hashes) between roles without manual wiring.
-
-## Special Nodes
-
-- \`$START\` — entry point; uses status keys \`new\` (first start) and \`resume\` (resuming a completed thread)
-- \`$END\` — terminal node; thread completes when reached and is moved to history
-
-## Integration with Steps
-
-Each \`uwf thread exec\` cycle:
-1. Moderator reads the thread's head step output
-2. Looks up \`graph[lastRole][output.$status]\` to pick the next role
-3. If next is \`$END\`, marks thread as completed
-4. Otherwise, renders the edge prompt and spawns the agent for the selected role
-5. Extract pipeline parses agent output → new step node → append to CAS chain
-`;
-}

packages/util/src/yaml-reference.ts

@@ -1,83 +0,0 @@
-export function generateYamlReference(): string {
-  return `# Workflow YAML Schema Reference
-
-## Top-Level Structure
-
-A workflow YAML file defines the complete workflow specification:
-
-\`\`\`yaml
-name: solve-issue          # verb-first kebab-case identifier
-description: "..."         # human-readable description
-
-roles:                     # named actors in the workflow
-  planner:
-    description: "Analyzes issue and outputs a plan"
-    goal: "You are a planning agent."
-    capabilities:
-      - issue-analysis
-      - planning
-    procedure: |
-      1. Read the issue
-      2. Produce a test spec
-    output: "Output the plan summary. Set $status to ready or insufficient_info."
-    frontmatter:           # JSON Schema for structured output (drives routing)
-      oneOf:
-        - properties:
-            $status: { const: ready }
-            plan: { type: string }
-          required: [$status, plan]
-        - properties:
-            $status: { const: insufficient_info }
-          required: [$status]
-
-graph:                     # status-based routing (nested map)
-  $START:
-    new: { role: planner, prompt: "Analyze the issue." }
-    resume: { role: planner, prompt: "Review the previous run output and continue." }
-  planner:
-    ready: { role: developer, prompt: "Implement plan {{{plan}}}." }
-    insufficient_info: { role: $END, prompt: "Not enough info." }
-\`\`\`
-
-## roles
-
-Each role defines an actor in the workflow:
-
-| Field | Type | Description |
-|-------|------|-------------|
-| \`description\` | string | Short description of the role's purpose |
-| \`goal\` | string | System-level goal statement for the agent |
-| \`capabilities\` | string[] | Tags describing what the role can do |
-| \`procedure\` | string | Step-by-step instructions for the agent |
-| \`output\` | string | Description of expected output format |
-| \`frontmatter\` | JSON Schema | Defines the structured output the agent must produce |
-
-### frontmatter
-
-The \`frontmatter\` field is a standard JSON Schema object. The extract pipeline validates agent output against it. Key conventions:
-- \`$status\` field drives routing decisions in the graph
-- Use \`const\` or \`enum\` to constrain status values
-- Use \`oneOf\` to define multiple valid output shapes (one per status)
-- All \`required\` fields must appear in the agent's frontmatter output
-
-## graph
-
-The graph is a nested map defining status-based routing:
-
-\`\`\`
-Record<Role | "$START", Record<Status, { role: string, prompt: string }>>
-\`\`\`
-
-| Level | Key | Value |
-|-------|-----|-------|
-| Outer | Role name or \`$START\` | Status map for that role |
-| Inner | \`$status\` value | Target: \`{ role, prompt }\` |
-
-### Special Nodes
-- \`$START\` — entry point; uses status keys \`new\` (first start) and \`resume\` (resuming a completed thread)
-- \`$END\` — terminal node; thread completes when reached
-
-### Edge Prompts
-Prompts use triple-brace Mustache templates (\`{{{field}}}\`) to interpolate values from the previous step's output. Example: \`"Implement plan {{{plan}}} in repo {{{repoPath}}}."\`
-`;
-}