feat: Phase 1 — CAS thread storage types + helpers

- Add StartNode, StateNode, ContentMerkleNode types to workflow-protocol
- Add collectRefs() to workflow-cas — extracts CAS hashes from StateNode payload
- Add findReachableHashes() to workflow-cas — recursive mark traversal via refs[]
- Tests: 7 pass (collect-refs + reachable)

Refs #155, closes #156

小橘 <xiaoju@shazhou.work>

CLAUDE.md

@@ -2,7 +2,7 @@
 
 ## Project Overview
 
-**@uncaged/workflow** is a workflow engine that executes single-file ESM bundles. Each workflow is a self-contained `.esm.js` file with an XXH64 hash as its version identifier.
+This monorepo implements a workflow engine that executes single-file ESM bundles. Each workflow is a self-contained `.esm.js` file with an XXH64 hash as its version identifier. Shared types live in `@uncaged/workflow-protocol`; bundle authors typically depend on `@uncaged/workflow-runtime`.
 
 ### Key Terms
 
@@ -19,14 +19,27 @@
 ```
 workflow/
   packages/
-    workflow/       # @uncaged/workflow — core lib (types, hash, ULID, JSONL, registry)
-    cli-workflow/   # @uncaged/cli-workflow — CLI (uncaged-workflow command)
+    workflow-protocol/              # @uncaged/workflow-protocol — shared types + Result
+    workflow-runtime/               # @uncaged/workflow-runtime — createWorkflow, type re-exports
+    workflow-util/                  # @uncaged/workflow-util — Base32, ULID, logger, storage paths, refs helpers
+    workflow-reactor/               # @uncaged/workflow-reactor — LLM fn + thread reactor (tool calls)
+    workflow-cas/                   # @uncaged/workflow-cas — CAS store, hash, Merkle
+    workflow-register/              # @uncaged/workflow-register — bundle validation, registry YAML, model resolution
+    workflow-execute/               # @uncaged/workflow-execute — engine, extract, fork, GC, workflowAsAgent
+    cli-workflow/                   # @uncaged/cli-workflow — uncaged-workflow CLI
+    workflow-agent-cursor/          # @uncaged/workflow-agent-cursor
+    workflow-agent-hermes/          # @uncaged/workflow-agent-hermes
+    workflow-agent-llm/             # @uncaged/workflow-agent-llm
+    workflow-util-agent/            # @uncaged/workflow-util-agent — buildAgentPrompt, spawnCli
+    workflow-template-develop/      # @uncaged/workflow-template-develop
+    workflow-template-solve-issue/  # @uncaged/workflow-template-solve-issue
+    workflow-dashboard/             # @uncaged/workflow-dashboard — React dashboard (private app)
   docs/             # RFCs, conventions
   biome.json        # root Biome config
   tsconfig.json     # root TypeScript config
 ```
 
-- `workflow` is the core; `cli-workflow` depends on it
+- Execution stack layers: `workflow-protocol` → (`workflow-runtime`, `workflow-util`, `workflow-reactor`) → (`workflow-cas`, `workflow-register`) → `workflow-execute` → `cli-workflow`
 - Packages use `workspace:*` protocol
 
 ## Language & Paradigm
@@ -167,10 +180,10 @@ type Result<T, E = Error> = { ok: true; value: T } | { ok: false; error: E };
 
 Never use `console.log/warn/error` directly — Biome's `noConsole` rule enforces this.
 
-All logging goes through the structured logger from `@uncaged/workflow`:
+All logging goes through the structured logger from `@uncaged/workflow-util`:
 
 ```typescript
-import { createLogger } from "@uncaged/workflow";
+import { createLogger } from "@uncaged/workflow-util";
 
 const log = createLogger();
 

docs/architecture.md

@@ -1,6 +1,6 @@
-# @uncaged/workflow — Architecture
+# Uncaged workflow — Architecture
 
-**Last updated:** 2026-05-06 by 小橘 🍊（NEKO Team）
+**Last updated:** 2026-05-09
 
 ---
 
@@ -8,72 +8,106 @@
 
 A workflow engine that executes single-file ESM bundles. Each workflow is a self-contained `.esm.js` file identified by its XXH64 hash (Crockford Base32). No daemon — processes start on demand and exit when done.
 
-## Package Structure
+The implementation lives in **15** Bun workspace packages under `packages/`, using the `workspace:*` protocol.
 
-| Package | npm Name | Purpose |
-|---------|----------|---------|
-| `workflow` | `@uncaged/workflow` | Core: types, engine, ExtractFn, hash/ULID/registry |
-| `cli-workflow` | `@uncaged/cli-workflow` | CLI: `uncaged-workflow` command |
-| `workflow-agent-cursor` | `@uncaged/workflow-agent-cursor` | Cursor CLI agent (extracts workspace from ctx) |
-| `workflow-agent-hermes` | `@uncaged/workflow-agent-hermes` | Hermes CLI agent |
-| `workflow-agent-llm` | `@uncaged/workflow-agent-llm` | OpenAI-compatible LLM agent |
-| `workflow-template-develop` | `@uncaged/workflow-template-develop` | Develop workflow template (roles in `src/roles/`) |
-| `workflow-template-solve-issue` | `@uncaged/workflow-template-solve-issue` | Solve-issue workflow template (roles in `src/roles/`) |
-| `workflow-util-agent` | `@uncaged/workflow-util-agent` | `buildAgentPrompt` + `spawnCli` utilities |
+## Package map
 
-Monorepo with **bun workspace**, `workspace:*` protocol.
+Grouped by responsibility (npm name → folder).
 
-## Core Types
+| Layer | Package | One-line role |
+|-------|---------|----------------|
+| Contract | `@uncaged/workflow-protocol` → `workflow-protocol` | Shared TypeScript types and `Result` helpers; peer `zod` only — no other workspace deps. |
+| Author API | `@uncaged/workflow-runtime` → `workflow-runtime` | `createWorkflow` and re-exports of protocol workflow types for bundle authors. |
+| Shared infra | `@uncaged/workflow-util` → `workflow-util` | Base32/ULID, logger, storage root paths, global CAS dir, ref-field helpers. |
+| LLM plumbing | `@uncaged/workflow-reactor` → `workflow-reactor` | `createLlmFn`, `createThreadReactor`, and related tool-call types for threaded LLM invocation. |
+| CAS | `@uncaged/workflow-cas` → `workflow-cas` | `CasStore` implementation, XXH64 hashing, Merkle helpers over CAS payloads. |
+| Registry / bundles | `@uncaged/workflow-register` → `workflow-register` | Bundle validation & dynamic export extraction, `workflow.yaml` registry I/O, provider/model resolution. |
+| Engine | `@uncaged/workflow-execute` → `workflow-execute` | Thread execution, worker entry path, fork/GC, extract pipeline, `workflowAsAgent`. |
+| CLI | `@uncaged/cli-workflow` → `cli-workflow` | `uncaged-workflow` binary (depends on engine, registry, CAS, protocol, util, runtime). |
+| Agent adapters | `@uncaged/workflow-agent-cursor` → `workflow-agent-cursor` | `AgentFn` via `cursor-agent` CLI + workspace extraction. |
+| | `@uncaged/workflow-agent-hermes` → `workflow-agent-hermes` | `AgentFn` via `hermes chat` CLI. |
+| | `@uncaged/workflow-agent-llm` → `workflow-agent-llm` | `AgentFn` via OpenAI-compatible HTTP (`LlmProvider` from runtime). |
+| Agent shared | `@uncaged/workflow-util-agent` → `workflow-util-agent` | `buildAgentPrompt`, `spawnCli` for CLI-backed agents. |
+| Templates | `@uncaged/workflow-template-develop` → `workflow-template-develop` | Develop workflow definition, roles, descriptor builder. |
+| | `@uncaged/workflow-template-solve-issue` → `workflow-template-solve-issue` | Solve-issue workflow definition, roles, descriptor builder. |
+| Dashboard | `@uncaged/workflow-dashboard` → `workflow-dashboard` | Private Vite + React app (`src/main.tsx`); only `react` / `react-dom` dependencies — no workspace packages. |
 
-```typescript
-// --- Sentinel values ---
-const START = "__start__";
-const END = "__end__";
+## Dependency graph (workspace packages)
 
-// --- RoleMeta: maps role names → their meta types ---
-type RoleMeta = Record<string, Record<string, unknown>>;
+Bottom-up layering for the execution stack:
 
-// --- Role Definition: pure data, no execution logic ---
-type RoleDefinition<Meta> = {
-  description: string;      // human-readable
-  systemPrompt: string;     // given to agent
-  extractPrompt: string;    // given to extractor
-  schema: z.ZodType<Meta>;  // meta shape (Zod v4)
-};
-
-// --- Workflow Definition: pure data, no agent binding ---
-type WorkflowDefinition<M extends RoleMeta> = {
-  description: string;
-  roles: { [K in keyof M & string]: RoleDefinition<M[K]> };
-  moderator: Moderator<M>;
-};
-
-// --- Agent: raw string output, reads role info from context ---
-type AgentFn = (ctx: AgentContext) => Promise<string>;
-
-// --- Agent Binding: runtime assignment ---
-type AgentBinding = {
-  agent: AgentFn;
-  overrides?: Partial<Record<string, AgentFn>>;
-};
-
-// --- Extract: structured data from context ---
-type ExtractFn = <T>(schema: z.ZodType<T>, prompt: string, ctx: ExtractContext) => Promise<T>;
-
-// --- Moderator: pure routing function ---
-type Moderator<M extends RoleMeta> = (ctx: ModeratorContext<M>) => (keyof M & string) | typeof END;
-
-// --- Composition ---
-// createWorkflow(def, binding, extract) => WorkflowFn
+```mermaid
+flowchart BT
+  subgraph L0["Layer 0 — contract"]
+    protocol["@uncaged/workflow-protocol"]
+  end
+  subgraph L1["Layer 1 — on protocol"]
+    runtime["@uncaged/workflow-runtime"]
+    util["@uncaged/workflow-util"]
+    reactor["@uncaged/workflow-reactor"]
+  end
+  subgraph L2["Layer 2 — protocol + util"]
+    cas["@uncaged/workflow-cas"]
+    register["@uncaged/workflow-register"]
+  end
+  subgraph L3["Layer 3 — engine"]
+    execute["@uncaged/workflow-execute"]
+  end
+  subgraph L4["Layer 4 — CLI"]
+    cli["@uncaged/cli-workflow"]
+  end
+  runtime --> protocol
+  util --> protocol
+  reactor --> protocol
+  cas --> protocol
+  cas --> util
+  register --> protocol
+  register --> util
+  execute --> protocol
+  execute --> runtime
+  execute --> util
+  execute --> cas
+  execute --> reactor
+  execute --> register
+  cli --> protocol
+  cli --> util
+  cli --> cas
+  cli --> execute
+  cli --> register
+  cli --> runtime
 ```
 
-## Three-Phase Engine Loop
+**Adjacent consumers** (not in the main CLI stack):
 
-Each role execution has three distinct phases with progressive context:
+- `@uncaged/workflow-util-agent` → `@uncaged/workflow-runtime`
+- `@uncaged/workflow-agent-llm` → `@uncaged/workflow-runtime`
+- `@uncaged/workflow-agent-cursor` → `@uncaged/workflow-runtime`, `@uncaged/workflow-util-agent`, `zod`
+- `@uncaged/workflow-agent-hermes` → `@uncaged/workflow-runtime`, `@uncaged/workflow-util-agent`
+- `@uncaged/workflow-template-develop` → `@uncaged/workflow-register`, `@uncaged/workflow-runtime`, `zod`
+- `@uncaged/workflow-template-solve-issue` → `@uncaged/workflow-register`, `@uncaged/workflow-runtime`, `zod` (dev-only workspace deps: `@uncaged/workflow-cas`, `@uncaged/workflow-execute` for tests/tooling per `package.json`)
+
+## Package roles (detail)
+
+- **`workflow-protocol`** — Pure types (`WorkflowFn`, contexts, `CasStore` interface, descriptor shapes), `START` / `END`, `ok` / `err`. Depends only on peer `zod` for schema-related types in signatures.
+- **`workflow-runtime`** — Workflow author surface: `createWorkflow` from `src/create-workflow.js`, re-exports protocol types/constants used when authoring bundles.
+- **`workflow-util`** — Cross-cutting utilities: Crockford Base32, ULID, `createLogger`, `getDefaultWorkflowStorageRoot`, `getGlobalCasDir`, ref normalization; re-exports `ok`/`err` from protocol.
+- **`workflow-cas`** — Filesystem CAS (`createCasStore`), `hashString` / `hashWorkflowBundleBytes`, Merkle node serialization and helpers (`merkle.js`).
+- **`workflow-register`** — Bundle pipeline (`validateWorkflowBundle`, `extractBundleExports`, descriptor builders), registry YAML read/write, `resolveModel` / `splitProviderModelRef`.
+- **`workflow-execute`** — `executeThread`, supervisor/worker wiring (`engine/`), fork/GC/pause gate, `createExtract` + LLM extract helpers (`extract/`), `workflowAsAgent`. Imports `@uncaged/workflow-reactor` for LLM-backed extract/supervisor paths (`extract-fn.ts`, `supervisor.ts`).
+- **`workflow-reactor`** — `createLlmFn`, `createThreadReactor`, and thread tool-invocation types — consumed by `workflow-execute`.
+- **`cli-workflow`** — CLI commands and HTTP/dashboard-related wiring (`hono`, `yaml`); composes register + execute + CAS + util.
+- **`workflow-agent-*`** — Replaceable `AgentFn` implementations (Cursor / Hermes CLIs, or HTTP LLM).
+- **`workflow-util-agent`** — Shared prompt assembly and subprocess spawning for CLI agents.
+- **`workflow-template-*`** — Concrete `WorkflowDefinition` graphs + Zod role schemas + descriptor builders for publishing bundles.
+- **`workflow-dashboard`** — Standalone React UI; no published library entry matching `src/index.ts`.
+
+## Three-phase engine loop
+
+Each role round is implemented in `packages/workflow-runtime/src/create-workflow.ts` (`advanceOneRound`): moderator → agent → extractor, with progressive context types from `@uncaged/workflow-protocol`.
 
 ```
 ┌─→ Phase 1: MODERATOR
-│   Context: ModeratorContext { threadId, start, steps }
+│   Context: ModeratorContext { threadId, depth, start, steps }
 │   Action:  moderator(ctx) → role name | END
 │
 │   Phase 2: AGENT
@@ -82,90 +116,80 @@ Each role execution has three distinct phases with progressive context:
 │
 │   Phase 3: EXTRACTOR
 │   Context: ExtractContext = AgentCtx + { agentContent }
-│   Action:  extract(schema, extractPrompt, ctx) → typed meta
+│   Action:  runtime.extract(schema, extractPrompt, ctx) → typed meta
 │
-│   Merge: RoleStep { role, content, meta, timestamp }
+│   Merge: RoleStep { role, contentHash, meta, refs, timestamp }
 │   Append to steps
 └─────────────────────────────────────────────────────┘
 ```
 
-### Context Types (progressive)
+### Context types (progressive)
+
+Defined in `packages/workflow-protocol/src/types.ts`:
 
 ```typescript
-// Phase 1: Moderator sees accumulated state only
-type ModeratorContext<M> = {
-  threadId: string;
-  start: StartStep;
-  steps: RoleStep<M>[];
-};
-
-// Phase 2: Agent knows its identity
+type ModeratorContext<M> = ThreadContext<M>;
 type AgentContext<M> = ModeratorContext<M> & {
   currentRole: { name: string; systemPrompt: string };
 };
-
-// Phase 3: Extractor has agent output
-type ExtractContext<M> = AgentContext<M> & {
-  agentContent: string;
-};
-
-// ThreadContext is an alias for AgentContext (backward compat)
-type ThreadContext<M> = AgentContext<M>;
+type ExtractContext<M> = AgentContext<M> & { agentContent: string };
 ```
 
-### Key Properties
+### Key properties
 
-- **Moderator is synchronous and pure** — no I/O, no state mutation
-- **Agent gets context, not instructions** — reads `ctx.currentRole.systemPrompt`
-- **Extractor is a general tool** — not limited to post-agent extraction; agents can use it too (e.g. Cursor agent extracts workspace path before execution)
-- **extractPrompt is a call parameter**, not context state — different callers use different prompts
+- **Moderator is synchronous and pure** — no I/O, no state mutation inside `createWorkflow`’s moderator call path.
+- **Agent receives `AgentContext`** — reads `ctx.currentRole.systemPrompt`; raw output becomes `agentContent` for extract.
+- **Extractor is `WorkflowRuntime.extract`** — supplied by the engine from registry-resolved LLM config (`workflow-execute`); stores agent body in CAS and yields `contentHash` + `refs` on each step (`create-workflow.ts`).
+- **`extractPrompt` is a call parameter** on `RoleDefinition`, not implicit context state.
 
-## Agent Information Sources
+## Agent information sources
 
 An agent has exactly three information sources:
 
 1. **Prior knowledge** — LLM training, agent memory, agent skills
-2. **Thread context** — `AgentContext` (start, steps, currentRole)
+2. **Thread context** — `AgentContext` (`start`, `steps`, `currentRole`)
 3. **Derived information** — from 1 & 2 (e.g. tool calls, shell commands)
 
-No hidden environment parameters. If an agent needs something (like a workspace path), it extracts it from context using `ExtractFn`.
+No hidden environment parameters. If an agent needs something (like a workspace path), it obtains it via `ExtractFn` (e.g. Cursor agent).
 
-## Bundle Contract
+## Bundle contract
 
-A workflow bundle is a single `.esm.js` file with two named exports:
+A workflow bundle is a single `.esm.js` file with two named exports (see `WorkflowFn` / `WorkflowDescriptor` in `packages/workflow-protocol/src/types.ts`):
 
 ```typescript
-// Named exports (no default export)
 export const descriptor: WorkflowDescriptor;
 export const run: WorkflowFn;
 
 type WorkflowFn = (
-  input: { prompt: string; steps: RoleOutput[] },
-  options: { threadId: string; maxRounds: number },
-) => AsyncGenerator<RoleOutput, WorkflowResult>;
+  thread: ThreadContext,
+  runtime: WorkflowRuntime,
+) => AsyncGenerator<RoleOutput, WorkflowCompletion>;
 ```
 
+`RoleOutput` carries `contentHash`, `meta`, and `refs` (agent text lives in CAS, addressed by hash).
+
 ### Constraints
 
 - Single `.esm.js` file
-- No dynamic `import()`
-- All static imports must be Node built-in modules only
-- XXH64 hash (Crockford Base32) = globally unique version ID
+- No dynamic `import()` in bundles (loader exempt in engine)
+- Portable bundle static imports are constrained by validation in `@uncaged/workflow-register` (`validateWorkflowBundle`)
+- XXH64 hash (Crockford Base32) = version ID
 
 ### Why AsyncGenerator?
 
-- Each `yield` → engine writes to `.data.jsonl`, checks abort/pause
-- `return` → engine marks thread complete
-- Fork = pass historical steps as `input.steps` to a new generator
-- Zero injection — bundle doesn't import from the engine
+- Each `yield` lets `workflow-execute` persist state, CAS rows, and enforce pause/abort
+- `return` supplies `WorkflowCompletion`
+- Fork replays historical steps into a new thread context
+- Bundle does not import the engine — only protocol/runtime types at build time
 
-## Storage Layout
+## Storage layout
 
 ```
 ~/.uncaged/workflow/
+├── cas/                           # Global content-addressed blobs (see getGlobalCasDir)
 ├── bundles/
-│   ├── C9NMV6V2TQT81.esm.js     # Crockford Base32 of XXH64
-│   └── C9NMV6V2TQT81.yaml       # Role descriptor
+│   ├── C9NMV6V2TQT81.esm.js       # Crockford Base32 of XXH64
+│   └── C9NMV6V2TQT81.yaml         # Role descriptor sidecar (when present)
 ├── logs/                          # One folder per bundle hash
 │   └── C9NMV6V2TQT81/
 │       ├── 01KQXKW…YG.data.jsonl  # Thread state
@@ -173,7 +197,7 @@ type WorkflowFn = (
 └── workflow.yaml                  # Registry
 ```
 
-### ID Encoding: Crockford Base32
+### ID encoding: Crockford Base32
 
 - Case-insensitive, filesystem-safe, no ambiguous chars (0/O, 1/I/L)
 - Bundle hash: XXH64 → 13-char
@@ -181,45 +205,36 @@ type WorkflowFn = (
 
 ### Registry (`workflow.yaml`)
 
-```yaml
-workflows:
-  solve-issue:
-    hash: "C9NMV6V2TQT81"
-    timestamp: 1714963200000
-    history:
-      - hash: "A7BKR3M1NPQ40"
-        timestamp: 1714876800000
-```
+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
 
-**`.data.jsonl`** — Line 1: start record, Line 2+: role outputs
+**`.data.jsonl`** — Line 1: start record; following lines: role steps with CAS-backed content.
 
 ```jsonc
 // Start record
 { "name": "solve-issue", "hash": "C9NMV6V2TQT81", "threadId": "01KQXKW…",
   "parameters": { "prompt": "Fix bug #3", "options": { "maxRounds": 5 } },
   "timestamp": 1714963200000 }
-// Role output
-{ "role": "planner", "content": "...", "meta": { "phases": [...] }, "timestamp": ... }
+// Role output (engine persists contentHash + refs; body in ~/.uncaged/workflow/cas/)
+{ "role": "planner", "contentHash": "…", "meta": { "phases": [...] }, "refs": ["…"], "timestamp": ... }
 ```
 
-**`.info.jsonl`** — Structured debug log
+**`.info.jsonl`** — Structured debug log via `@uncaged/workflow-util` `createLogger`:
 
 ```jsonc
 { "tag": "4KNMR2PX", "content": "Loading bundle...", "timestamp": ... }
 ```
 
-Tags are 8-char Crockford Base32 (40-bit random), one per call site. `grep "4KNMR2PX"` → instant code location.
+Tags are 8-char Crockford Base32 (40-bit random), one per call site. `grep "4KNMR2PX"` → code location.
 
-## Execution Model
+## Execution model
 
-- **No daemon.** `uncaged-workflow run <name>` starts a worker process
-- Same bundle's threads share one process (memory efficiency)
-- Process exits when all threads complete
-- Thread termination via IPC within the process
+- **No daemon.** `uncaged-workflow run <name>` starts a worker process (`workflow-execute` worker entry via `getWorkerHostScriptPath`)
+- Threads share bundle-scoped workers as implemented in CLI/engine
+- Pause/resume/abort via engine IPC and pause gate (`createThreadPauseGate`)
 
-## CLI Commands
+## CLI commands
 
 | Priority | Command | Description |
 |----------|---------|-------------|
@@ -239,18 +254,16 @@ Tags are 8-char Crockford Base32 (40-bit random), one per call site. `grep "4KNM
 | P2 | `resume <thread-id>` | Resume a paused thread |
 | P3 | `fork <thread-id> [--from-role <role>]` | Fork from historical state |
 
-All commands implemented and tested. ✅
-
-## Design Decisions
+## Design decisions
 
 | Decision | Rationale |
 |----------|-----------|
 | **Role = pure data** | Decouples definition from execution; same role with different agents |
-| **Agent bound at runtime** | WorkflowDefinition is reusable; agent choice is deployment concern |
-| **Three-phase context** | Each phase sees only what it needs; clean separation |
-| **ExtractFn as general tool** | Agents use it for pre-execution extraction; engine uses it for meta |
-| **Single-file ESM** | Hash = version, no dependency hell, self-contained |
-| **No daemon** | OS handles process lifecycle; unnecessary complexity |
+| **Agent bound at runtime** | `WorkflowDefinition` is reusable; agent choice is deployment concern |
+| **Three-phase context** | Each phase sees only what it needs; types live in `workflow-protocol` |
+| **`WorkflowRuntime.extract` + CAS `contentHash`** | Large agent bodies deduplicated globally; Merkle roots summarize threads |
+| **`workflow-reactor` split** | LLM tool-calling loop isolated from filesystem/registry concerns |
+| **Single-file ESM** | Hash = version, self-contained bundle |
+| **No daemon** | OS handles process lifecycle |
 | **Crockford Base32** | Filesystem-safe, readable, compact |
-| **No concurrency in registry** | Different workflows have different constraints; belongs at workflow/role level |
-| **No dryRun** | Tests use mock agents + mock fetch; simpler architecture |
+| **15-package split** | Clear boundaries: protocol ↔ runtime author API ↔ util/CAS/register ↔ execute ↔ CLI ↔ agents/templates/UI |

docs/plans/2025-05-07-workflow-as-agent.md

@@ -1,5 +1,7 @@
 # Workflow-as-Agent Implementation Plan
 
+> ⚠️ This plan references the pre-split package structure. File paths have changed.
+
 > **For Hermes:** Use subagent-driven-development skill to implement this plan task-by-task.
 
 **Goal:** Enable workflows to invoke other workflows as agents, backed by global CAS and refs tracking.

docs/plans/2026-05-09-cas-thread-storage.md

@@ -0,0 +1,262 @@
+# RFC: CAS-Based Thread Storage
+
+> Status: Draft
+> Author: 小橘 🍊（NEKO Team）
+> Date: 2026-05-09
+
+## Summary
+
+Replace `.data.jsonl` with a fully CAS-based thread state chain. Threads become linked lists of immutable CAS nodes, indexed by a per-bundle `threads.json`.
+
+## Motivation
+
+`.data.jsonl` is a flat append-only file with three different row formats (start, role step, end). This makes forking expensive (copy file), deduplication impossible (forked threads repeat shared history), and GC complex (must parse every row to find CAS refs).
+
+Threads are inherently immutable append-only sequences — a natural fit for CAS hash chains, similar to git's commit DAG.
+
+## Design
+
+### Node Types
+
+Two CAS node types, using the existing `{ type, payload, refs }` CAS blob structure:
+
+#### StartNode
+
+Contains workflow-level parameters. **No threadId** (because the same StartNode can be shared across forks). Prompt is stored as a CAS blob and referenced via `refs[0]`.
+
+```
+CAS blob:
+{
+  type: "start",
+  payload: {
+    name: "solve-issue",
+    hash: "BUNDLE_HASH",
+    maxRounds: 10,
+    depth: 0
+  },
+  refs: [
+    <prompt_hash>    // refs[0]: initial task prompt (CAS blob)
+  ]
+}
+```
+
+- No `role`, `content`, `meta` — this is not a step, it's workflow metadata
+- Prompt is **not** inline — it lives in CAS and is referenced by hash
+
+#### StateNode
+
+One per role step (including `__end__`).
+
+```
+CAS blob:
+{
+  type: "state",
+  payload: {
+    role: "coder",
+    meta: { ... },
+    start: "<start_hash>",
+    content: "<content_merkle_hash>",
+    ancestors: ["<parent_hash>", "<grandparent_hash>", ...],
+    compact: null,
+    timestamp: 1234567890
+  },
+  refs: [<start_hash>, <content_hash>, <parent_hash>, ...]
+}
+```
+
+**Payload is the source of truth.** Application code reads named fields from payload. `refs[]` is a **GC index** — automatically derived from payload by collecting all CAS hashes. GC only scans `refs[]` without understanding payload structure.
+
+**Payload fields:**
+
+| Field | Type | Meaning |
+|-------|------|---------|
+| `role` | `string` | Role name, or `"__end__"` for completion |
+| `meta` | `object` | Structured metadata extracted from agent output |
+| `start` | `string` | StartNode hash |
+| `content` | `string` | Content Merkle node hash (carries role artifact refs) |
+| `ancestors` | `string[]` | `[parent, grandparent, ...]` — up to 11 entries (1 parent + 10 skip-list). Empty for first step after start. `ancestors[0]` is the direct parent. |
+| `compact` | `string \| null` | CAS hash of a compacted summary of all nodes before this one. When present, LLM context assembly can use this instead of walking the full chain. |
+| `timestamp` | `number` | Unix timestamp in ms |
+
+### Content Merkle Node
+
+The content at `refs[2]` of each StateNode is itself a CAS Merkle node. This is where **role artifact references** live:
+
+```
+CAS blob:
+{
+  type: "content",
+  payload: "<role output text>",
+  refs: [
+    <artifact_hash_1>,   // e.g. a commit, a file, a sub-result
+    <artifact_hash_2>,
+    ...
+  ]
+}
+```
+
+The Extractor is responsible for producing both `meta` and `refs` from raw agent output:
+
+```
+Agent raw output
+    ↓
+Extractor → { meta, contentPayload, refs[] }
+    ↓
+CAS put content Merkle: { type: "content", payload: contentPayload, refs }
+    ↓ contentHash
+StateNode: { ..., refs: [start, parent, contentHash, ...ancestors] }
+```
+
+This keeps StateNode refs fixed and simple. All role-specific artifact references are encapsulated in the content Merkle node. GC follows: `thread head → StateNode.refs → content Merkle.refs → artifacts`, full chain recursive.
+
+### End Node
+
+An end is just a StateNode with `role: "__end__"`:
+
+```
+{
+  type: "state",
+  payload: {
+    role: "__end__",
+    meta: { returnCode: 0, summary: "completed successfully" },
+    start: "<start_hash>",
+    content: "<content_hash>",
+    ancestors: ["<parent_hash>", ...],
+    compact: null,
+    timestamp: 1234567891
+  },
+  refs: [<start_hash>, <content_hash>, <parent_hash>, ...]
+}
+```
+
+### Thread Index: `threads.json`
+
+Per-bundle directory, one `threads.json` file. **Only active (in-progress) threads** live here:
+
+```
+~/.uncaged/workflow/bundles/<hash>/threads.json
+```
+
+```json
+{
+  "01JTHREAD1AAAAAAAAAAAAAAA": {
+    "head": "<latest_state_node_hash>",
+    "start": "<start_node_hash>",
+    "updatedAt": 1234567891
+  }
+}
+```
+
+When a thread completes (`__end__`), it is **removed from `threads.json`** and appended to a date-partitioned history file:
+
+```
+~/.uncaged/workflow/bundles/<hash>/history/{YYYY-MM-DD}.jsonl
+```
+
+Each line:
+
+```json
+{"threadId":"01JTHREAD1AAAAAAAAAAAAAAA","head":"<end_node_hash>","start":"<start_node_hash>","completedAt":1234567891}
+```
+
+Benefits:
+- `threads.json` stays small — only in-flight threads
+- Dashboard watches `threads.json` for real-time updates; completed threads don't trigger watches
+- History is queryable by date but not actively monitored
+- GC roots = all heads from `threads.json` + all heads from `history/*.jsonl`
+
+### Ancestor Skip-List
+
+Each StateNode carries up to 11 entries in `payload.ancestors` (1 parent + 10 skip-list, newest first):
+
+```
+Node 15: ancestors = [node14, node13, node12, node11, node10, node9, node8, node7, node6, node5, node4]
+                      ^parent  ^--- skip-list (10 most recent) ---^
+```
+
+This enables:
+- **Paginated fetch**: jump to any recent ancestor without walking the full chain
+- **Partial replay**: fetch last N steps without loading the entire history
+- The list is capped at 10 to keep node size bounded
+
+### Fork
+
+Forking a thread at step N:
+
+1. Create new threadId
+2. Create a new StateNode whose `parent` (refs[1]) points to the fork point's StateNode
+3. Register the new threadId in `threads.json` with its own head
+4. **Zero data duplication** — the forked thread shares all ancestor nodes via CAS
+
+### Compact
+
+When a StateNode has `payload.compact` set:
+
+```json
+{
+  "type": "state",
+  "payload": {
+    "role": "coder",
+    "meta": { ... },
+    "compact": "<cas_hash_of_summary>",
+    "timestamp": 1234
+  },
+  "refs": [...]
+}
+```
+
+This means: "everything before this node has been summarized into the blob at `compact`". When building LLM context:
+
+1. Walk back from head
+2. If a node has `compact`, stop walking — use the compact summary + all nodes after it
+3. If no compact found, use full chain
+
+This enables long-running threads without unbounded context growth.
+
+### GC
+
+Simple mark-and-sweep:
+
+1. **Roots**: all `head` and `start` hashes from `threads.json` + all `history/*.jsonl` files
+2. **Mark**: from each root, recursively mark all reachable hashes via `refs[]` (including content Merkle → artifact refs)
+3. **Sweep**: delete unmarked CAS blobs
+
+No per-row format parsing needed. GC only needs to understand `refs[]`.
+
+### refs[] Derivation
+
+`refs[]` is auto-derived from payload at write time via a `collectRefs(payload)` function that extracts all CAS hash strings from named fields (`start`, `content`, `ancestors`, `compact`). Application code never reads `refs[]` — it reads named payload fields. This makes `refs[]` a pure GC optimization with zero semantic coupling.
+
+### Extract Phase
+
+The Extractor is expanded from the current design. Currently it only extracts `meta` from agent output. In the new design it extracts:
+
+| Output | Purpose |
+|--------|---------|
+| `meta` | Structured metadata (same as before) |
+| `contentPayload` | The text payload for the content Merkle node |
+| `refs[]` | CAS hashes of artifacts produced by this role step |
+
+The `refs[]` become the content Merkle node's refs, enabling GC to trace all role-produced artifacts.
+
+## What Stays Unchanged
+
+- `.info.jsonl` — debug logging stays as-is (high-frequency append, not suitable for CAS)
+- CAS blob storage format (`~/.uncaged/workflow/cas/`)
+- Bundle registry (`workflow.yaml`)
+
+## Migration
+
+Breaking change. Old `.data.jsonl` files become incompatible. No backward compat fallback (per project convention).
+
+## Changes by Package
+
+| Package | Changes |
+|---------|---------|
+| `workflow-protocol` | Replace `StartStep`, `RoleStep` types with `StartNode`, `StateNode`. Add `ContentMerkleNode` type. Expand `ExtractResult` to include `refs[]`. |
+| `workflow-cas` | Add `findReachableHashes(roots)` for GC mark phase |
+| `workflow-execute` | Rewrite engine to write CAS nodes + update `threads.json` instead of appending JSONL. Move completed threads to `history/`. Simplify `gc.ts`. Simplify `fork-thread.ts`. Expand extract phase to produce refs. |
+| `workflow-runtime` | `ThreadContext` built by walking chain from head. `start.prompt` resolved from CAS via StartNode.refs[0]. |
+| `cli-workflow` | `thread list/show/rm` read from `threads.json` + `history/`. SSE watches `threads.json`. |
+| `workflow-dashboard` | Watch `threads.json` instead of `.data.jsonl` |
+| Templates & Agents | Update extract definitions to produce `refs[]`. Update `ctx.start.content` → CAS resolved. |

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

@@ -9,9 +9,6 @@ import { createThreadDispatcher, dispatchLive, dispatchRun } from "./commands/th
 import { createWorkflowDispatcher } from "./commands/workflow/index.js";
 import { formatSkillIndex, formatSkillTopic, getSkillTopics } from "./skill.js";
 
-export type { CommandEntry, CommandGroup, DispatchFn } from "./cli-command-types.js";
-export { getCommandRegistry } from "./cli-registry.js";
-
 function dispatchGroup(
   tableName: string,
   table: Record<string, CommandEntry>,

packages/workflow-agent-cursor/README.md

@@ -7,10 +7,10 @@ The agent builds a full prompt (system + task + step history via `@uncaged/workf
 ## Install
 
 ```bash
-bun add @uncaged/workflow-agent-cursor @uncaged/workflow @uncaged/workflow-util-agent zod
+bun add @uncaged/workflow-agent-cursor @uncaged/workflow-runtime @uncaged/workflow-util-agent zod
 ```
 
-In this monorepo: `"@uncaged/workflow-agent-cursor": "workspace:*"` plus `workspace:*` for `@uncaged/workflow` and `@uncaged/workflow-util-agent`.
+In this monorepo: `"@uncaged/workflow-agent-cursor": "workspace:*"` plus `workspace:*` for `@uncaged/workflow-runtime` and `@uncaged/workflow-util-agent`, and `zod` ^4.
 
 ## Usage
 
@@ -28,9 +28,8 @@ const agent = createCursorAgent({
 
 | Export | Description |
 |--------|-------------|
-| `createCursorAgent(config)` | Returns `AgentFn` that runs `cursor-agent` with `buildAgentPrompt(ctx)` |
+| `createCursorAgent(config)` | Returns `AgentFn` that runs `cursor-agent` with `buildAgentPrompt(ctx)` from `@uncaged/workflow-util-agent` |
 | `CursorAgentConfig` | `model`, `timeout`, `extract` (must supply workspace path) |
 | `validateCursorAgentConfig` | Config validation result |
-| `buildAgentPrompt` | Re-exported from `@uncaged/workflow-util-agent` |
 
 Requires `cursor-agent` on `PATH` at runtime.

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

@@ -5,7 +5,6 @@ import * as z from "zod/v4";
 import type { CursorAgentConfig } from "./types.js";
 import { validateCursorAgentConfig } from "./validate-config.js";
 
-export { buildAgentPrompt } from "@uncaged/workflow-util-agent";
 export type { CursorAgentConfig } from "./types.js";
 export { validateCursorAgentConfig } from "./validate-config.js";
 

packages/workflow-agent-hermes/README.md

@@ -7,10 +7,10 @@ The agent composes the same thread-aware prompt as other CLI-backed agents via `
 ## Install
 
 ```bash
-bun add @uncaged/workflow-agent-hermes @uncaged/workflow @uncaged/workflow-util-agent
+bun add @uncaged/workflow-agent-hermes @uncaged/workflow-runtime @uncaged/workflow-util-agent
 ```
 
-In this monorepo: use `workspace:*` for all three `@uncaged/*` packages.
+In this monorepo: use `workspace:*` for `@uncaged/workflow-agent-hermes`, `@uncaged/workflow-runtime`, and `@uncaged/workflow-util-agent`.
 
 ## Usage
 
@@ -30,6 +30,5 @@ const agent = createHermesAgent({
 | `createHermesAgent(config)` | Returns `AgentFn` wrapping `hermes chat -q ...` |
 | `HermesAgentConfig` | `model`, `timeout` |
 | `validateHermesAgentConfig` | Config validation result |
-| `buildAgentPrompt` | Re-exported from `@uncaged/workflow-util-agent` |
 
 Requires `hermes` on `PATH` at runtime.

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

@@ -6,7 +6,6 @@ import { validateHermesAgentConfig } from "./validate-config.js";
 
 const HERMES_DEFAULT_MAX_TURNS = 90;
 
-export { buildAgentPrompt } from "@uncaged/workflow-util-agent";
 export type { HermesAgentConfig } from "./types.js";
 export { validateHermesAgentConfig } from "./validate-config.js";
 

packages/workflow-agent-llm/README.md

@@ -1,16 +1,16 @@
 # @uncaged/workflow-agent-llm
 
-`AgentFn` adapter that calls an OpenAI-compatible `POST /chat/completions` endpoint using `@uncaged/workflow`’s `LlmProvider` (base URL, API key, model).
+`AgentFn` adapter that calls an OpenAI-compatible `POST /chat/completions` endpoint using `LlmProvider` from `@uncaged/workflow-runtime`.
 
 Single-turn: system text is the current role’s `systemPrompt`, user text is the thread’s initial prompt (`ctx.start.content`). Errors from HTTP, JSON, or empty choices are thrown as `Error` with a JSON payload string.
 
 ## Install
 
 ```bash
-bun add @uncaged/workflow-agent-llm @uncaged/workflow
+bun add @uncaged/workflow-agent-llm @uncaged/workflow-runtime zod
 ```
 
-In this monorepo: `"@uncaged/workflow-agent-llm": "workspace:*"`, `"@uncaged/workflow": "workspace:*"`.
+In this monorepo: `"@uncaged/workflow-agent-llm": "workspace:*"`, `"@uncaged/workflow-runtime": "workspace:*"` (and satisfy `zod` ^4 as required by `@uncaged/workflow-runtime`).
 
 ## Usage
 

packages/workflow-cas/README.md

@@ -0,0 +1,31 @@
+# @uncaged/workflow-cas
+
+Content-addressable storage implementation, bundle hashing, and Merkle helpers.
+
+## What This Package Does
+
+It implements `CasStore` from `@uncaged/workflow-protocol`, hashes workflow bundle bytes and strings with XXH64, and builds serializable Merkle nodes for thread/step/content payloads used when persisting execution artifacts.
+
+## Key Exports
+
+From `src/index.ts`:
+
+- **CAS:** `createCasStore`
+- **Hash:** `hashString`, `hashWorkflowBundleBytes`
+- **Merkle:** `createContentMerkleNode`, `getContentMerklePayload`, `parseMerkleNode`, `putContentMerkleNode`, `putStepMerkleNode`, `putThreadMerkleNode`, `serializeMerkleNode`
+- **Types:** `CasStore`, `MerkleNode`, `MerkleNodeType`, `StepMerklePayload`, `ThreadMerklePayload`
+
+## Dependencies
+
+- **Workspace:** `@uncaged/workflow-protocol` (`CasStore` contract), `@uncaged/workflow-util`
+- **npm:** `xxhashjs`, `yaml`
+
+## Usage
+
+```typescript
+import { createCasStore, hashWorkflowBundleBytes } from "@uncaged/workflow-cas";
+import { getGlobalCasDir } from "@uncaged/workflow-util";
+
+const store = createCasStore(getGlobalCasDir());
+const hash = await hashWorkflowBundleBytes(esmJsBytes);
+```

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

@@ -0,0 +1,65 @@
+import { describe, expect, test } from "bun:test";
+import type { StateNode } from "@uncaged/workflow-protocol";
+
+import { collectRefs } from "../src/collect-refs.js";
+
+function payload(
+  partial: Partial<StateNode["payload"]> & Pick<StateNode["payload"], "role">,
+): StateNode["payload"] {
+  return {
+    role: partial.role,
+    meta: partial.meta ?? {},
+    start: partial.start ?? "STARTHASH000000000000001",
+    content: partial.content ?? "CONTENTHASH00000000000001",
+    ancestors: partial.ancestors ?? [],
+    compact: partial.compact ?? null,
+    timestamp: partial.timestamp ?? 0,
+  };
+}
+
+describe("collectRefs", () => {
+  test("collects start, content, ancestors, and compact hashes in order", () => {
+    const refs = collectRefs(
+      payload({
+        role: "coder",
+        start: "01START00000000000000001",
+        content: "01CONTENT0000000000000001",
+        ancestors: ["01PARENT0000000000000001", "01GRAND000000000000000001"],
+        compact: "01COMPACT0000000000000001",
+      }),
+    );
+    expect(refs).toEqual([
+      "01START00000000000000001",
+      "01CONTENT0000000000000001",
+      "01PARENT0000000000000001",
+      "01GRAND000000000000000001",
+      "01COMPACT0000000000000001",
+    ]);
+  });
+
+  test("does not collect compact when compact is null", () => {
+    const refs = collectRefs(
+      payload({
+        role: "coder",
+        start: "S1",
+        content: "C1",
+        ancestors: ["A1"],
+        compact: null,
+      }),
+    );
+    expect(refs).toEqual(["S1", "C1", "A1"]);
+  });
+
+  test("returns only start and content when ancestors is empty", () => {
+    const refs = collectRefs(
+      payload({
+        role: "coder",
+        start: "S2",
+        content: "C2",
+        ancestors: [],
+        compact: null,
+      }),
+    );
+    expect(refs).toEqual(["S2", "C2"]);
+  });
+});

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

@@ -0,0 +1,69 @@
+import { describe, expect, test } from "bun:test";
+import type { CasStore } from "@uncaged/workflow-protocol";
+import { stringify } from "yaml";
+
+import { findReachableHashes } from "../src/reachable.js";
+
+function yamlBlob(refs: readonly string[]): string {
+  return stringify({ type: "node", payload: {}, refs: [...refs] }, { indent: 2 });
+}
+
+function memoryCas(entries: Record<string, string>): CasStore {
+  const map = { ...entries };
+  return {
+    async put(): Promise<string> {
+      throw new Error("memoryCas.put not used in tests");
+    },
+    async get(hash: string): Promise<string | null> {
+      return map[hash] ?? null;
+    },
+    async delete(): Promise<void> {},
+    async list(): Promise<string[]> {
+      return Object.keys(map);
+    },
+  };
+}
+
+describe("findReachableHashes", () => {
+  test("walks refs recursively from a single root", async () => {
+    const cas = memoryCas({
+      R1: yamlBlob(["R2"]),
+      R2: yamlBlob(["R3"]),
+      R3: yamlBlob([]),
+    });
+    const reachable = await findReachableHashes(["R1"], cas);
+    expect([...reachable].sort()).toEqual(["R1", "R2", "R3"]);
+  });
+
+  test("union of reachability from multiple roots", async () => {
+    const cas = memoryCas({
+      A: yamlBlob(["X"]),
+      B: yamlBlob(["Y"]),
+      X: yamlBlob([]),
+      Y: yamlBlob(["Z"]),
+      Z: yamlBlob([]),
+    });
+    const reachable = await findReachableHashes(["A", "B"], cas);
+    expect([...reachable].sort()).toEqual(["A", "B", "X", "Y", "Z"]);
+  });
+
+  test("handles cycles via visited set", async () => {
+    const cas = memoryCas({
+      C1: yamlBlob(["C2"]),
+      C2: yamlBlob(["C1"]),
+    });
+    const reachable = await findReachableHashes(["C1"], cas);
+    expect(reachable.size).toBe(2);
+    expect(reachable.has("C1")).toBe(true);
+    expect(reachable.has("C2")).toBe(true);
+  });
+
+  test("does not throw when a ref points to a missing blob", async () => {
+    const cas = memoryCas({
+      H1: yamlBlob(["MISSINGHASH0000000000001"]),
+    });
+    const reachable = await findReachableHashes(["H1"], cas);
+    expect(reachable.has("H1")).toBe(true);
+    expect(reachable.has("MISSINGHASH0000000000001")).toBe(false);
+  });
+});

packages/workflow-cas/package.json

@@ -2,6 +2,9 @@
   "name": "@uncaged/workflow-cas",
   "version": "0.1.0",
   "type": "module",
+  "scripts": {
+    "test": "bun test"
+  },
   "exports": {
     ".": {
       "types": "./dist/index.d.ts",

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

@@ -0,0 +1,13 @@
+import type { StateNode } from "@uncaged/workflow-protocol";
+
+/** Collects CAS hashes from {@link StateNode} payload fields for GC `refs[]` derivation. */
+export function collectRefs(payload: StateNode["payload"]): string[] {
+  const out: string[] = [payload.start, payload.content];
+  for (const h of payload.ancestors) {
+    out.push(h);
+  }
+  if (payload.compact !== null) {
+    out.push(payload.compact);
+  }
+  return out;
+}

packages/workflow-cas/src/index.ts

@@ -1,4 +1,5 @@
 export { createCasStore } from "./cas.js";
+export { collectRefs } from "./collect-refs.js";
 export { hashString, hashWorkflowBundleBytes } from "./hash.js";
 export {
   createContentMerkleNode,
@@ -9,6 +10,7 @@ export {
   putThreadMerkleNode,
   serializeMerkleNode,
 } from "./merkle.js";
+export { findReachableHashes } from "./reachable.js";
 export type {
   CasStore,
   MerkleNode,

packages/workflow-cas/src/reachable.ts

@@ -0,0 +1,55 @@
+import { parse } from "yaml";
+
+import type { CasStore } from "./types.js";
+
+function refsFromBlob(content: string): string[] {
+  try {
+    const raw = parse(content) as unknown;
+    if (raw === null || typeof raw !== "object") {
+      return [];
+    }
+    const rec = raw as Record<string, unknown>;
+    const refs = rec.refs;
+    if (!Array.isArray(refs)) {
+      return [];
+    }
+    const out: string[] = [];
+    for (const r of refs) {
+      if (typeof r === "string") {
+        out.push(r);
+      }
+    }
+    return out;
+  } catch {
+    return [];
+  }
+}
+
+/** Recursively collects all CAS hashes reachable from `roots` via each blob's `refs[]`. */
+export async function findReachableHashes(
+  roots: readonly string[],
+  cas: CasStore,
+): Promise<ReadonlySet<string>> {
+  const visited = new Set<string>();
+  const stack = [...roots];
+  while (stack.length > 0) {
+    const hash = stack.pop();
+    if (hash === undefined) {
+      break;
+    }
+    if (visited.has(hash)) {
+      continue;
+    }
+    const blob = await cas.get(hash);
+    if (blob === null) {
+      continue;
+    }
+    visited.add(hash);
+    for (const ref of refsFromBlob(blob)) {
+      if (!visited.has(ref)) {
+        stack.push(ref);
+      }
+    }
+  }
+  return visited;
+}

packages/workflow-cas/tsconfig.json

@@ -5,8 +5,5 @@
     "outDir": "dist"
   },
   "include": ["src"],
-  "references": [
-    { "path": "../workflow-protocol" },
-    { "path": "../workflow-util" }
-  ]
+  "references": [{ "path": "../workflow-protocol" }, { "path": "../workflow-util" }]
 }

packages/workflow-execute/README.md

@@ -0,0 +1,33 @@
+# @uncaged/workflow-execute
+
+Thread engine: execution, fork/GC, extract pipeline, supervisor/worker wiring, and workflow-as-agent.
+
+## What This Package Does
+
+It runs `WorkflowFn` generators against disk-backed threads, integrates CAS and registry-backed extract (`createExtract`), coordinates LLM tool usage via `@uncaged/workflow-reactor`, handles fork plans and garbage collection, and exposes `workflowAsAgent` for nesting workflows.
+
+## Key Exports
+
+From `src/index.ts`:
+
+- **Engine:** `createWorkflow` (engine-local re-export), `executeThread`, `getWorkerHostScriptPath`
+- **Fork / parse:** `buildForkPlan`, `parseThreadDataJsonl`, `selectForkHistoricalSteps`, `tryParseRoleStepRecord`, `tryParseWorkflowResultRecord`
+- **GC / pause:** `garbageCollectCas`, `createThreadPauseGate`
+- **Engine types:** `ExecuteThreadIo`, `ExecuteThreadOptions`, `ForkHistoricalStep`, `ForkPlan`, `GcResult`, `ParsedThreadStartRecord`, `PrefilledDiskStep`, `SupervisorDecision`, `ThreadPauseGate`
+- **Extract:** `buildExtractUserContent`, `createExtract`, `extractFunctionToolFromZodSchema`, `llmErrorToCause`, `llmExtract`, types `ExtractFn`, `ExtractThreadContext`, `LlmError`, `LlmExtractArgs`
+- **Agent composition:** `workflowAsAgent`, `WorkflowAsAgentOptions`
+
+## Dependencies
+
+- **Workspace:** `@uncaged/workflow-protocol`, `@uncaged/workflow-runtime`, `@uncaged/workflow-util`, `@uncaged/workflow-cas`, `@uncaged/workflow-reactor`, `@uncaged/workflow-register`
+- **npm:** `yaml`
+- **Peer:** `zod` ^4
+
+`@uncaged/workflow-reactor` is used for LLM-backed extract and supervisor flows (`extract-fn.ts`, `supervisor.ts`).
+
+## Usage
+
+```typescript
+import { executeThread } from "@uncaged/workflow-execute";
+// Typical callers are CLI/tests that supply ExecuteThreadIo (paths, CAS, abort, logger, …).
+```

packages/workflow-protocol/README.md

@@ -0,0 +1,29 @@
+# @uncaged/workflow-protocol
+
+Shared workflow types, sentinel constants, and `Result` helpers.
+
+## What This Package Does
+
+It defines the cross-package contract for bundles and the engine: thread/step shapes, `WorkflowFn`, agent/extract contexts, descriptor types, and `CasStore` as an interface. Implementations (CAS store, CLI, extract) depend on these types so bundles stay decoupled from Node APIs.
+
+## Key Exports
+
+From `src/index.ts`:
+
+- **Types:** `Result`, `CasStore`, `WorkflowRoleSchema`, `WorkflowRoleDescriptor`, `WorkflowDescriptor`, `RoleMeta`, `RoleOutput`, `StartStep`, `RoleStep`, `ThreadContext`, `ModeratorContext`, `AgentContext`, `ExtractContext`, `WorkflowCompletion`, `WorkflowResult`, `LlmProvider`, `ProviderConfig`, `ResolvedModel`, `WorkflowConfig`, `ExtractFn`, `AgentFn`, `AgentBinding`, `WorkflowRuntime`, `WorkflowFn`, `RoleDefinition`, `Moderator`, `WorkflowDefinition`, `AdvanceOutcome`
+- **Constants:** `START`, `END`
+- **Functions:** `ok`, `err`
+
+## Dependencies
+
+- **Peer:** `zod` ^4 — used in type positions for schemas (`ExtractFn`, `RoleDefinition`, etc.)
+
+No workspace packages; this is the bottom layer.
+
+## Usage
+
+```typescript
+import { END, START, type WorkflowFn, type ThreadContext } from "@uncaged/workflow-protocol";
+```
+
+Concrete `WorkflowFn` implementations are built with `@uncaged/workflow-runtime` (`createWorkflow`).

packages/workflow-protocol/src/cas-types.ts

@@ -0,0 +1,36 @@
+// ── CAS thread chain nodes (RFC: CAS-based thread storage) ──────────
+
+export type StartNodePayload = {
+  name: string;
+  hash: string;
+  maxRounds: number;
+  depth: number;
+};
+
+export type StartNode = {
+  type: "start";
+  payload: StartNodePayload;
+  refs: string[];
+};
+
+export type StateNodePayload = {
+  role: string;
+  meta: Record<string, unknown>;
+  start: string;
+  content: string;
+  ancestors: string[];
+  compact: string | null;
+  timestamp: number;
+};
+
+export type StateNode = {
+  type: "state";
+  payload: StateNodePayload;
+  refs: string[];
+};
+
+export type ContentMerkleNode = {
+  type: "content";
+  payload: string;
+  refs: string[];
+};

packages/workflow-protocol/src/index.ts

@@ -1,40 +1,46 @@
 // ── Types ──────────────────────────────────────────────────────────
 
 export type {
-	Result,
-	CasStore,
-	WorkflowRoleSchema,
-	WorkflowRoleDescriptor,
-	WorkflowDescriptor,
-	RoleMeta,
-	RoleOutput,
-	StartStep,
-	RoleStep,
-	ThreadContext,
-	ModeratorContext,
-	AgentContext,
-	ExtractContext,
-	WorkflowCompletion,
-	WorkflowResult,
-	LlmProvider,
-	ProviderConfig,
-	ResolvedModel,
-	WorkflowConfig,
-	ExtractFn,
-	AgentFn,
-	AgentBinding,
-	WorkflowRuntime,
-	WorkflowFn,
-	RoleDefinition,
-	Moderator,
-	WorkflowDefinition,
-	AdvanceOutcome,
+  ContentMerkleNode,
+  StartNode,
+  StateNode,
+} from "./cas-types.js";
+
+export type {
+  AdvanceOutcome,
+  AgentBinding,
+  AgentContext,
+  AgentFn,
+  CasStore,
+  ExtractContext,
+  ExtractFn,
+  LlmProvider,
+  Moderator,
+  ModeratorContext,
+  ProviderConfig,
+  ResolvedModel,
+  Result,
+  RoleDefinition,
+  RoleMeta,
+  RoleOutput,
+  RoleStep,
+  StartStep,
+  ThreadContext,
+  WorkflowCompletion,
+  WorkflowConfig,
+  WorkflowDefinition,
+  WorkflowDescriptor,
+  WorkflowFn,
+  WorkflowResult,
+  WorkflowRoleDescriptor,
+  WorkflowRoleSchema,
+  WorkflowRuntime,
 } from "./types.js";
 
 // ── Constants ──────────────────────────────────────────────────────
 
-export { START, END } from "./types.js";
+export { END, START } from "./types.js";
 
 // ── Constructor functions ──────────────────────────────────────────
 
-export { ok, err } from "./result.js";
+export { err, ok } from "./result.js";

packages/workflow-reactor/README.md

@@ -0,0 +1,26 @@
+# @uncaged/workflow-reactor
+
+LLM calling abstraction and thread “reactor” for structured tool invocation.
+
+## What This Package Does
+
+It exposes `createLlmFn` (chat completion wrapper) and `createThreadReactor` (multi-turn tool loop configuration) plus supporting message/tool types. `@uncaged/workflow-execute` consumes this for extractor and supervisor paths that talk to OpenAI-style APIs with tools.
+
+## Key Exports
+
+From `src/index.ts`:
+
+- **Functions:** `createLlmFn`, `createThreadReactor`
+- **Types:** `ChatMessage`, `LlmFn`, `StructuredToolSpec`, `ThreadReactorConfig`, `ThreadReactorFn`, `ThreadReactorInvokeArgs`, `ToolCall`, `ToolDefinition`
+
+## Dependencies
+
+- **Workspace:** `@uncaged/workflow-protocol`
+- **Peer:** `zod` ^4
+
+## Usage
+
+```typescript
+import { createLlmFn, createThreadReactor } from "@uncaged/workflow-reactor";
+// Usually composed inside @uncaged/workflow-execute rather than directly by applications.
+```

packages/workflow-register/README.md

@@ -0,0 +1,38 @@
+# @uncaged/workflow-register
+
+Bundle validation, dynamic export extraction, registry YAML, and model/provider resolution.
+
+## What This Package Does
+
+It validates workflow `.esm.js` bundles, extracts `descriptor` / `run` exports at runtime, reads and writes `workflow.yaml`, and resolves which LLM endpoint/model to use from registry config (`resolveModel`, `splitProviderModelRef`).
+
+## Key Exports
+
+From `src/index.ts`:
+
+- **Bundle:** `buildDescriptor`, `importWorkflowBundleModule`, `validateWorkflowBundle`, `ensureUncagedWorkflowSymlink`, `extractBundleExports`, `stringifyWorkflowDescriptor`, `validateWorkflowDescriptor`
+- **Bundle types:** `ExtractBundleExportsOptions`, `ExtractedBundleExports`, `WorkflowBundleValidationInput`, `WorkflowDescriptor`, `WorkflowRoleDescriptor`, `WorkflowRoleSchema`
+- **Registry:** `getRegisteredWorkflow`, `listRegisteredWorkflowNames`, `parseWorkflowRegistryYaml`, `readWorkflowRegistry`, `registerWorkflowVersion`, `rollbackWorkflowToHistoryHash`, `stringifyWorkflowRegistryYaml`, `unregisterWorkflow`, `workflowRegistryPath`, `writeWorkflowRegistry`
+- **Registry types:** `WorkflowConfig`, `WorkflowHistoryEntry`, `WorkflowRegistryEntry`, `WorkflowRegistryFile`
+- **Config:** `resolveModel`, `splitProviderModelRef`, types `ProviderConfig`, `ResolvedModel`
+
+## Dependencies
+
+- **Workspace:** `@uncaged/workflow-protocol`, `@uncaged/workflow-util`
+- **Peer:** `acorn`, `yaml`, `zod` ^4 — parsing/validation at runtime for consumers
+
+## Usage
+
+```typescript
+import { readFile } from "node:fs/promises";
+import { readWorkflowRegistry, validateWorkflowBundle } from "@uncaged/workflow-register";
+import { getDefaultWorkflowStorageRoot } from "@uncaged/workflow-util";
+
+const reg = await readWorkflowRegistry(getDefaultWorkflowStorageRoot());
+if (!reg.ok) throw new Error(reg.error.message);
+
+const path = "./my.esm.js";
+const source = await readFile(path, "utf8");
+const v = validateWorkflowBundle({ filePath: path, source });
+if (!v.ok) throw new Error(v.error);
+```

packages/workflow-runtime/README.md

@@ -0,0 +1,28 @@
+# @uncaged/workflow-runtime
+
+Workflow author API: `createWorkflow` plus re-exports of protocol workflow types.
+
+## What This Package Does
+
+Bundle code imports `createWorkflow` to turn a `WorkflowDefinition` plus `AgentBinding` into a `WorkflowFn` generator. It re-exports the protocol types and constants most authors need so workflows rarely import `@uncaged/workflow-protocol` directly.
+
+## Key Exports
+
+From `src/index.ts`:
+
+- **Functions:** `createWorkflow`, `ok`, `err`
+- **Types:** `AgentBinding`, `AgentContext`, `AgentFn`, `CasStore`, `ExtractContext`, `ExtractFn`, `LlmProvider`, `Moderator`, `ModeratorContext`, `Result`, `RoleDefinition`, `RoleMeta`, `RoleOutput`, `RoleStep`, `StartStep`, `ThreadContext`, `WorkflowCompletion`, `WorkflowDefinition`, `WorkflowDescriptor`, `WorkflowFn`, `WorkflowResult`, `WorkflowRoleDescriptor`, `WorkflowRoleSchema`, `WorkflowRuntime`
+- **Constants:** `END`, `START`
+
+## Dependencies
+
+- **Workspace:** `@uncaged/workflow-protocol` — contract types and helpers
+- **Peer:** `zod` ^4 — matches schema usage on role definitions
+
+## Usage
+
+```typescript
+import { createWorkflow, type WorkflowDefinition, type AgentBinding } from "@uncaged/workflow-runtime";
+
+export const run = createWorkflow(myDefinition, myBinding);
+```

packages/workflow-util/README.md

@@ -0,0 +1,32 @@
+# @uncaged/workflow-util
+
+Shared utilities: encoding, IDs, logging, storage paths, and ref-field normalization.
+
+## What This Package Does
+
+It provides filesystem-safe Base32 and ULID generation, the structured logger used across packages, helpers for the default workflow data directory and global CAS path, and utilities to merge/normalize `refs` on steps. It re-exports `ok`/`err` from protocol for convenience.
+
+## Key Exports
+
+From `src/index.ts`:
+
+- **Base32:** `CROCKFORD_BASE32_ALPHABET`, `decodeCrockfordBase32Bits`, `decodeCrockfordToUint64`, `encodeCrockfordBase32Bits`, `encodeUint64AsCrockford`
+- **Logger:** `createLogger`
+- **Refs:** `mergeRefsWithContentHash`, `normalizeRefsField`
+- **Result:** `ok`, `err` (from `@uncaged/workflow-protocol`)
+- **Paths:** `getDefaultWorkflowStorageRoot`, `getGlobalCasDir`
+- **ULID:** `generateUlid`
+- **Types:** `CreateLoggerOptions`, `LogFn`, `LoggerSink`, `Result`
+
+## Dependencies
+
+- **Workspace:** `@uncaged/workflow-protocol` — `Result` and shared types used by helpers
+
+## Usage
+
+```typescript
+import { createLogger, getDefaultWorkflowStorageRoot, generateUlid } from "@uncaged/workflow-util";
+
+const log = createLogger();
+log("4KNMR2PX", "example");
+```