fix: add workspace path sandboxing and UWF_BUILTIN_ALLOW_SHELL check

- Add resolvePathInWorkspace() to reject paths escaping workspace root
- Apply sandboxing to read_file, write_file, and run_command tools
- Gate run_command behind UWF_BUILTIN_ALLOW_SHELL=1 env var
- Add tests for resolvePathInWorkspace escape detection

.env.example

@@ -1,40 +0,0 @@
-# ──────────────────────────────────────────────
-# Workflow Engine — Environment Variables
-# ──────────────────────────────────────────────
-# Copy this file to .env and fill in the values.
-
-# ── Cursor Agent ──
-
-# CLI command to invoke the Cursor agent (required for develop workflow)
-WORKFLOW_CURSOR_COMMAND=
-
-# Model override for Cursor agent
-WORKFLOW_CURSOR_MODEL=
-
-# Timeout in milliseconds for Cursor agent operations
-WORKFLOW_CURSOR_TIMEOUT=
-
-# ── Hermes Agent (used by develop tester/committer + solve-issue) ──
-
-# CLI command to invoke the Hermes agent (absolute path required)
-WORKFLOW_HERMES_COMMAND=
-
-# Model override for Hermes agent
-WORKFLOW_HERMES_MODEL=
-
-# Timeout in milliseconds for Hermes agent operations
-WORKFLOW_HERMES_TIMEOUT=
-
-# ── Storage ──
-
-# Override the workflow storage root directory
-# Default: ~/.uncaged/workflow
-WORKFLOW_STORAGE_ROOT=
-
-# Gateway secret for the serve command
-WORKFLOW_DASHBOARD_SECRET=
-
-# ── Display ──
-
-# Set to any value to disable colored output
-# NO_COLOR=1

.gitignore

@@ -11,3 +11,5 @@ solve-issue-entry.ts
 packages/workflow-template-develop/develop.esm.js
 .DS_Store
 *.py
+.claude
+tmp

.plan/335-setup-validate.md

@@ -0,0 +1,83 @@
+# Test Spec: uwf setup model connectivity validation (#335)
+
+## Context
+
+File: `packages/cli-workflow/src/commands/setup.ts`
+Test file: `packages/cli-workflow/src/__tests__/setup-validate.test.ts`
+
+After `cmdSetup` writes config, it should send a test chat completion request to verify the configured model is reachable. If validation fails, warn the user (don't abort — config is already saved).
+
+## Implementation Notes
+
+- Add a `validateModel(baseUrl, apiKey, model)` function that sends a minimal chat completion request (`POST /chat/completions` with `messages: [{role:"user",content:"hi"}]`, `max_tokens: 1`)
+- Returns `Result<void, string>` — ok if 2xx response, error with reason string otherwise
+- Use `AbortSignal.timeout(15_000)` for the request
+- Both `cmdSetup` and `cmdSetupInteractive` should call it after saving config
+- `cmdSetup` returns validation result in its return object: `{ ...existing, validation: { ok: true } | { ok: false, error: string } }`
+- `cmdSetupInteractive` prints a warning to console if validation fails, success message if it passes
+- Use the project logger (`createLogger`) — no raw `console.log` except in interactive CLI output (per CLAUDE.md)
+
+## Test Cases (vitest)
+
+### 1. `validateModel` — success path
+- Mock `fetch` to return `{ status: 200, ok: true, json: () => ({}) }`
+- Call `validateModel(baseUrl, apiKey, model)`
+- Assert returns `{ ok: true, value: undefined }`
+- Assert fetch was called with correct URL (`${baseUrl}/chat/completions`), correct headers (`Authorization: Bearer ${apiKey}`), correct body (model, messages, max_tokens: 1)
+
+### 2. `validateModel` — HTTP error (401 unauthorized)
+- Mock `fetch` to return `{ status: 401, ok: false, statusText: "Unauthorized" }`
+- Call `validateModel(baseUrl, apiKey, model)`
+- Assert returns `{ ok: false, error: <string containing "401"> }`
+
+### 3. `validateModel` — HTTP error (404 model not found)
+- Mock `fetch` to return `{ status: 404, ok: false, statusText: "Not Found" }`
+- Assert returns `{ ok: false, error: <string containing "404"> }`
+
+### 4. `validateModel` — network timeout
+- Mock `fetch` to throw `DOMException` with name `AbortError`
+- Assert returns `{ ok: false, error: <string containing "timeout" or "unreachable"> }`
+
+### 5. `validateModel` — network error (DNS failure, connection refused)
+- Mock `fetch` to throw `TypeError("fetch failed")`
+- Assert returns `{ ok: false, error: <string mentioning connectivity> }`
+
+### 6. `cmdSetup` — includes validation result on success
+- Mock global `fetch` for `/chat/completions` to succeed
+- Call `cmdSetup({ provider, baseUrl, apiKey, model, storageRoot })`
+- Assert returned object has `validation: { ok: true, value: undefined }`
+- Assert config files are still written (existing behavior preserved)
+
+### 7. `cmdSetup` — includes validation result on failure (config still saved)
+- Mock global `fetch` for `/chat/completions` to return 401
+- Call `cmdSetup({ ... })`
+- Assert returned object has `validation: { ok: false, error: ... }`
+- Assert `config.yaml` and `.env` are still written (validation failure doesn't prevent saving)
+
+### 8. `cmdSetupInteractive` — prints success message on validation pass
+- Mock `fetch` for both `/models` and `/chat/completions` to succeed
+- Mock stdin to provide valid selections
+- Capture console output
+- Assert output contains a success message like "Model verified" or "✓"
+
+### 9. `cmdSetupInteractive` — prints warning on validation failure
+- Mock `fetch`: `/models` succeeds, `/chat/completions` returns 401
+- Mock stdin for valid selections
+- Capture console output
+- Assert output contains a warning about model not being reachable and suggests trying a different model
+
+### 10. `validateModel` — request body correctness
+- Mock `fetch` to capture the request body
+- Call `validateModel(baseUrl, apiKey, "test-model")`
+- Assert body is `{ model: "test-model", messages: [{role: "user", content: "hi"}], max_tokens: 1 }`
+
+## Export Requirements
+
+- `validateModel` must be exported (for direct unit testing)
+- Signature: `async function validateModel(baseUrl: string, apiKey: string, model: string): Promise<Result<void, string>>`
+- `Result` type: `{ ok: true; value: T } | { ok: false; error: E }` (project convention)
+
+## Files to Create/Modify
+
+- **New**: `packages/cli-workflow/src/__tests__/setup-validate.test.ts` — all test cases above
+- **Modify**: `packages/cli-workflow/src/commands/setup.ts` — add `validateModel`, integrate into `cmdSetup` and `cmdSetupInteractive`

.workflows/solve-issue.yaml

@@ -0,0 +1,196 @@
+name: "solve-issue"
+description: "TDD-driven issue resolution for small, focused changes. Loop protection relies on engine maxRounds."
+roles:
+  planner:
+    description: "Analyzes issue and outputs a TDD test spec"
+    goal: "You are a planning agent. You analyze Gitea issues and produce a TDD test specification that downstream roles will implement and verify."
+    capabilities:
+      - issue-analysis
+      - planning
+    procedure: |
+      On first run (no previous steps):
+      1. Read the issue and all comments from Gitea using `tea issues <number> -r <owner/repo>`
+      2. Read CLAUDE.md (or equivalent project conventions file) to understand coding standards
+      3. Assess whether the issue has enough information to produce a test spec
+      4. If insufficient info: comment on the issue via `echo "..." | tea comment <number> -r <owner/repo>` (skip if you already commented), then output status=insufficient_info and terminate
+      5. If sufficient: produce a detailed TDD test spec in markdown covering all scenarios
+
+      On subsequent runs (bounced back by tester with fix_spec):
+      1. Read the tester's output from the previous step to understand what's wrong with the spec
+      2. Revise the test spec accordingly
+
+      After producing the test spec:
+      1. Store it via `uwf cas put-text "<markdown content>"` and capture the returned hash
+      2. Put the hash in frontmatter.plan (required when status=ready)
+    output: "Output a brief summary of the test spec. Frontmatter must include: status (ready or insufficient_info) and plan (CAS hash of the test spec, required when status=ready)."
+    frontmatter:
+      type: object
+      properties:
+        status:
+          type: string
+          enum: [ready, insufficient_info]
+        plan:
+          type: string
+      required: [status]
+  developer:
+    description: "TDD implementation per test spec"
+    goal: "You are a developer agent. You implement code changes following TDD — write tests first, then implementation."
+    capabilities:
+      - coding
+    procedure: |
+      Before starting any work, ensure a clean worktree:
+      1. `git checkout main && git pull` to get the latest code
+      2. `git checkout -b fix/<issue-number>-<short-description>` to create a fresh branch
+         - If bounced back from reviewer or tester, reuse the existing branch instead
+
+      Then implement TDD:
+      3. Read the test spec from CAS: `uwf cas get <plan hash>` (find the hash from the latest planner step's frontmatter.plan)
+      4. If bounced back from reviewer or tester: read the previous role's output to understand what needs fixing
+      5. Write tests first based on the spec
+      6. Implement the code to make tests pass
+      7. Ensure `bun run build` passes with no errors
+      8. Run `bun test` to verify all tests pass
+    output: "List all files changed and provide a summary. Frontmatter must include: status (done or failed)."
+    frontmatter:
+      type: object
+      properties:
+        status:
+          type: string
+          enum: [done, failed]
+      required: [status]
+  reviewer:
+    description: "Code standards compliance check"
+    goal: "You are a code reviewer. You verify code standards compliance — NOT functionality (that's the tester's job)."
+    capabilities:
+      - code-review
+      - static-analysis
+    procedure: |
+      Before reviewing, verify the git branch:
+      1. Run `git branch --show-current` — confirm the branch name references the issue number being worked on
+      2. If the branch doesn't correspond to the issue, flag it in your output and reject
+
+      Then perform code review:
+      Hard checks (must all pass):
+      3. `bun run build` — no build errors
+      4. `bunx biome check` — no lint violations
+      5. TypeScript strict mode — no type errors
+
+      Soft checks (review against CLAUDE.md conventions):
+      - Functional-first: `function` + `type`, not `class` + `interface`
+      - No optional properties (`?:`) — use `T | null`
+      - Naming conventions (kebab-case files, PascalCase types, camelCase functions)
+      - Module boundary discipline (folder exports via index.ts)
+      - No `console.log` (use structured logger)
+      - No dynamic imports in production code
+
+      Only review standards compliance. Do NOT test functionality.
+      If rejecting, you MUST explain the specific reason in your output.
+    output: "Explain your decision with specific file/line references. Frontmatter must include: approved (true or false)."
+    frontmatter:
+      type: object
+      properties:
+        approved:
+          type: boolean
+      required: [approved]
+  tester:
+    description: "Functional correctness verification"
+    goal: "You are a tester agent. You verify that the implementation correctly satisfies every scenario in the test spec."
+    capabilities:
+      - testing
+    procedure: |
+      1. Run `bun test` for automated test verification
+      2. Read the test spec from CAS: `uwf cas get <plan hash>` (find the hash from the latest planner step's frontmatter.plan)
+      3. Verify each scenario in the spec is covered and passing
+      4. Determine outcome:
+         - passed: all scenarios verified, tests pass
+         - fix_code: tests fail or implementation doesn't match spec → send back to developer
+         - fix_spec: the spec itself is wrong or incomplete → send back to planner
+    output: "Report test results per scenario. Frontmatter must include: status (passed, fix_code, or fix_spec)."
+    frontmatter:
+      type: object
+      properties:
+        status:
+          type: string
+          enum: [passed, fix_code, fix_spec]
+      required: [status]
+  committer:
+    description: "Commits and creates PR"
+    goal: "You are a committer agent. You create a clean commit and push a PR linking the original issue."
+    capabilities: []
+    procedure: |
+      Note: You inherit the developer's worktree and branch. Do NOT create a new branch.
+      1. Stage all changes: `git add -A`
+      2. Commit with a descriptive message referencing the issue: `git commit -m "type: description\n\nFixes #N"`
+      3. Push the branch: `git push -u origin <branch-name>`
+         - If push hook fails: capture the error log in your output, mark hook_failed
+      4. On push success: create a PR via `tea pr create --title "..." --description "..."`
+         - PR description must follow the project template: What / Why / Changes / Ref sections, with `Fixes #N` in Ref
+    output: "Include PR URL on success or error log on failure. Frontmatter must include: success (true or false)."
+    frontmatter:
+      type: object
+      properties:
+        success:
+          type: boolean
+      required: [success]
+conditions:
+  insufficientInfo:
+    description: "Planner determined there's not enough info to proceed"
+    expression: "$last('planner').status = 'insufficient_info'"
+  devFailed:
+    description: "Developer failed to implement"
+    expression: "$last('developer').status = 'failed'"
+  rejected:
+    description: "Reviewer rejected the implementation"
+    expression: "$last('reviewer').approved = false"
+  fixCode:
+    description: "Tester found code issues"
+    expression: "$last('tester').status = 'fix_code'"
+  fixSpec:
+    description: "Tester found spec issues"
+    expression: "$last('tester').status = 'fix_spec'"
+  hookFailed:
+    description: "Push hook failed"
+    expression: "$last('committer').success = false"
+graph:
+  $START:
+    - role: "planner"
+      condition: null
+      prompt: "Analyze the issue and produce an implementation plan."
+  planner:
+    - role: "$END"
+      condition: "insufficientInfo"
+      prompt: "Insufficient information to proceed; end the workflow."
+    - role: "developer"
+      condition: null
+      prompt: "Implement the plan from the planner."
+  developer:
+    - role: "$END"
+      condition: "devFailed"
+      prompt: "Development failed; end the workflow."
+    - role: "reviewer"
+      condition: null
+      prompt: "Send the implementation to the reviewer."
+  reviewer:
+    - role: "developer"
+      condition: "rejected"
+      prompt: "Reviewer rejected the implementation; fix the issues."
+    - role: "tester"
+      condition: null
+      prompt: "Review passed; run tests on the implementation."
+  tester:
+    - role: "developer"
+      condition: "fixCode"
+      prompt: "Tests found code issues; return to developer."
+    - role: "planner"
+      condition: "fixSpec"
+      prompt: "Tests found spec issues; return to planner."
+    - role: "committer"
+      condition: null
+      prompt: "Tests passed; commit and push the changes."
+  committer:
+    - role: "developer"
+      condition: "hookFailed"
+      prompt: "Push hook failed; return to developer to fix."
+    - role: "$END"
+      condition: null
+      prompt: "Commit succeeded; complete the workflow."

CLAUDE.md

@@ -2,46 +2,41 @@
 
 ## Project Overview
 
-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`.
+This monorepo implements a stateless workflow engine driven by a single-step CLI (`uwf`). Workflows are **YAML definitions** stored as CAS nodes; threads are immutable chains of CAS-linked step nodes. No daemon — each `uwf thread step` invocation runs one moderator→agent→extract cycle and exits.
 
 ### Key Terms
 
 | Concept | What it is |
 |---------|-----------|
-| **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. 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. |
+| **Workflow** | A YAML definition (`WorkflowPayload`) with roles, conditions, and a routing graph. Stored as a CAS node, identified by its XXH64 hash. |
+| **Thread** | A single execution of a workflow, identified by a ULID. State is an immutable CAS chain; active threads indexed in `threads.yaml`; completed threads in `history.jsonl`. |
+| **Role** | A named actor within a workflow. Each role has a system prompt and a JSON Schema `outputSchema`. |
+| **Moderator** | JSONata-based graph evaluator — determines the next role (or `$END`) with zero LLM cost. |
+| **Agent** | An external CLI command (`uwf-hermes`, etc.) spawned by `uwf thread step`. Produces frontmatter markdown output. |
+| **CAS** | Content-Addressed Storage via `@uncaged/json-cas` — all workflow definitions, thread nodes, and outputs are immutable CAS nodes. |
+| **Registry** | `~/.uncaged/workflow/registry.yaml` — maps workflow names to current CAS hashes. |
 
 ### Monorepo Structure
 
 ```
 workflow/
   packages/
-    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-agent-react/             # @uncaged/workflow-agent-react
-    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-protocol/    # @uncaged/workflow-protocol — shared types (WorkflowPayload, StepNodePayload, WorkflowConfig, etc.)
+    workflow-util/        # @uncaged/workflow-util — Crockford Base32, ULID, logger, frontmatter parsing/validation
+    workflow-moderator/   # @uncaged/workflow-moderator — JSONata graph evaluator
+    workflow-agent-kit/   # @uncaged/workflow-agent-kit — createAgent factory, context builder, extract pipeline
+    workflow-agent-hermes/ # @uncaged/workflow-agent-hermes — uwf-hermes CLI binary (spawns hermes chat)
+    cli-workflow/         # @uncaged/cli-workflow — uwf CLI binary
+  legacy-packages/       # Archived packages (preserved for reference, not active)
+  examples/              # Workflow YAML examples (solve-issue.yaml)
+  docs/                  # Architecture docs
+  biome.json             # root Biome config
+  tsconfig.json          # root TypeScript config
 ```
 
-- Execution stack layers: `workflow-protocol` → (`workflow-runtime`, `workflow-util`, `workflow-reactor`) → (`workflow-cas`, `workflow-register`) → `workflow-execute` → `cli-workflow`
+- Dependency layers: `workflow-protocol` → (`workflow-util`, `workflow-moderator`) → `workflow-agent-kit` → `workflow-agent-hermes` / `cli-workflow`
 - Packages use `workspace:^` protocol (resolves to `^x.y.z` on publish)
+- External CAS: `@uncaged/json-cas` (store API, hashing, schema validation) + `@uncaged/json-cas-fs` (filesystem backend)
 
 ## Language & Paradigm
 
@@ -109,8 +104,6 @@ type WorkflowEntry = {
 - Always named exports, never default exports
 - One module = one responsibility, filename = purpose
 
-Workflow bundles (`.esm.js`) follow the same rule: export `const run` and `const descriptor`, not `export default`.
-
 ### Folder Module Discipline
 
 Every folder under `src/` is a **module boundary**. Four rules:
@@ -136,10 +129,10 @@ export { createCasStore } from "../cas/cas.js";
 
 // ❌ Bad — types defined in index.ts
 // in cas/index.ts:
-export type CasStore = { ... };  // should be in cas/types.ts
+export type CasStore = { ... }; // should be in cas/types.ts
 ```
 
-**Exception**: The package-level `src/index.ts` is the public API surface and re-exports from folder `index.ts` files. Files that remain at `src/` root (e.g. `types.ts`, `workflow-as-agent.ts`) are not inside a folder module and follow normal rules.
+**Exception**: The package-level `src/index.ts` is the public API surface and re-exports from folder `index.ts` files. Files that remain at `src/` root (e.g. `types.ts`) are not inside a folder module and follow normal rules.
 
 ## Naming
 
@@ -160,7 +153,7 @@ Workflow names use **verb-first** kebab-case:
 ### ID Encoding
 
 All IDs use **Crockford Base32**:
-- Bundle hash: XXH64 → 13-char Crockford Base32
+- CAS hash: XXH64 → 13-char Crockford Base32
 - Thread ID: ULID → 26-char Crockford Base32 (10 timestamp + 16 random)
 
 ## Error Handling
@@ -189,7 +182,7 @@ import { createLogger } from "@uncaged/workflow-util";
 const log = createLogger();
 
 // Each call site has a fixed 8-char Crockford Base32 tag
-log("4KNMR2PX", "Loading workflow bundle...");
+log("4KNMR2PX", "Loading workflow...");
 log("7BQST3VW", `Role ${role} started`);
 ```
 
@@ -204,7 +197,7 @@ log("7BQST3VW", `Role ${role} started`);
 
 ### Why fixed tags?
 
-- `grep "4KNMR2PX"` in `.info.jsonl` → instant code location
+- `grep "4KNMR2PX"` in logs → instant code location
 - No need for file/line info in the log — tag is the locator
 - Survives refactoring (tag stays the same when code moves)
 
@@ -221,74 +214,76 @@ console.log(result);
 
 Do NOT use `await import()` in production code. Always use static top-level `import`.
 
-**Exception**: The bundle loader and `extractBundleExports` dynamically import user workflow files at runtime.
-
-```ts
-// Dynamic import required: user bundle path resolved at runtime
-const mod = await import(bundlePath);
-```
-
 Test files (`__tests__/**`) are exempt.
 
 ## Toolchain
 
 | Tool | Purpose |
 |------|---------|
-| **bun** | Package manager + runtime + test runner |
+| **bun** | Package manager + runtime |
 | **TypeScript** | Type checking (strict mode) |
 | **Biome** | Lint + format (replaces ESLint + Prettier) |
+| **vitest** | Test runner (`cli-workflow` uses vitest; other packages use `bun test`) |
 
-### Commands
+### Development Workflow
 
 ```bash
-bun run check       # tsc --build + biome check
-bun run format      # biome format --write
-bun test            # run tests
+# ── Setup ──
+bun install                 # install all workspace dependencies
+
+# ── Daily development ──
+bun run build               # tsc --build (all packages, dependency order)
+bun run check               # tsc --build + biome check + lint-log-tags
+bun run format              # biome format --write
+bun test                    # run tests across all packages
+
+# ── Before committing ──
+bun run check               # must pass — typecheck + lint + log tag validation
+bun test                    # must pass — all package tests
 ```
 
-### Version Management & Publishing
+### Publishing
 
-All public `@uncaged/*` packages are published to **npmjs.org** via `@changesets/cli` with **fixed mode** (all packages share the same version number). `workflow-dashboard` is private and excluded.
+All public `@uncaged/*` packages are published to **npmjs.org** with **fixed mode** (all packages share the same version number).
 
 ```bash
-# 1. After making changes, add a changeset describing the change
+# 1. Add a changeset describing the change
 bun changeset
 
-# 2. Before release, bump all package versions + generate CHANGELOGs
+# 2. Bump all package versions + generate CHANGELOGs
 bun version
 
-# 3. Build, test, and publish to npmjs
+# 3. Build, test, and publish (runs scripts/publish-all.mjs)
 bun release
+
+# Or publish manually with a tag:
+node scripts/publish-all.mjs --tag alpha
+node scripts/publish-all.mjs --dry-run    # preview without publishing
 ```
 
 - `workspace:^` dependencies resolve to `^x.y.z` on publish
+- Publish order defined in `scripts/publish-all.mjs` (dependency order)
 - Changesets config: `.changeset/config.json` (fixed mode, public access)
-- Each package has auto-generated `CHANGELOG.md`
 
-### Consuming @uncaged/* Packages
-
-External workflow repos just `bun install` — packages come from npmjs like any other dependency. No special registry config needed.
-
-### End-to-end: Monorepo → Registry → Workspace → Bundle
+### End-to-end: Author → Register → Run
 
 ```
-workflow/ (monorepo)           — engine, runtime, templates, agents
-  │  bun release               — build + test + changeset publish
+examples/solve-issue.yaml       — write a workflow YAML definition
+  │  uwf workflow put
   ▼
-npmjs.org                      — @uncaged/* scoped packages (public)
-  │  bun install
+~/.uncaged/workflow/cas/        — Workflow stored as CAS node
+~/.uncaged/workflow/registry.yaml — name → hash mapping updated
+  │  uwf thread start <name> -p "..."
   ▼
-my-workflows/ (workspace)     — normal package.json
-  │  bun run build:develop     — bun build → single .esm.js
+~/.uncaged/workflow/threads.yaml — new thread head pointer
+  │  uwf thread step <thread-id>
   ▼
-uncaged-workflow workflow add  — register bundle locally
-uncaged-workflow run           — execute workflow
+moderator → agent → extract      — one step per invocation, repeat until $END
 ```
 
-1. **Monorepo changes** → `bun changeset` (describe change) → `bun version` (bump) → `bun release` (publish)
-2. **Workspace** → `bun install` fetches latest from npmjs
-3. **Build** → produces single-file ESM bundle with `@uncaged/*` as externals
-4. **Register & Run** → `uncaged-workflow workflow add <name> <bundle>` then `uncaged-workflow run <name>`
+1. **Author** — write a workflow YAML file with roles, conditions, and graph
+2. **Register** — `uwf workflow put <file.yaml>` parses YAML, registers output schemas, stores `WorkflowPayload` in CAS
+3. **Run** — `uwf thread start` creates a thread, `uwf thread step` executes one cycle per invocation
 
 ## Commit Convention
 
@@ -296,5 +291,5 @@ uncaged-workflow run           — execute workflow
 <type>(<scope>): <description>
 
 type: feat | fix | refactor | docs | chore | test
-scope: workflow | cli | rfc-001 | ...
+scope: workflow | cli | moderator | agent-kit | hermes | util | protocol | ...
 ```

README.md

@@ -1,71 +1,93 @@
 # @uncaged/workflow
 
-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).
+A stateless workflow engine driven by a single-step CLI. Workflows are YAML definitions with roles, JSONata routing conditions, and a directed graph. Threads are immutable CAS-linked chains — each `uwf thread step` runs one moderator→agent→extract cycle and exits.
 
-## Core Concepts
+## Package Map
 
-| Concept | Description |
-|---------|-------------|
-| **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. 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. |
+| Package | npm | Role |
+|---------|-----|------|
+| `cli-workflow` | `@uncaged/cli-workflow` | `uwf` CLI binary — thread lifecycle, workflow registry, CAS inspection, setup |
+| `workflow-protocol` | `@uncaged/workflow-protocol` | Shared TypeScript types (`WorkflowPayload`, `StepNodePayload`, `WorkflowConfig`, etc.) |
+| `workflow-moderator` | `@uncaged/workflow-moderator` | JSONata graph evaluator — determines next role or `$END` |
+| `workflow-agent-kit` | `@uncaged/workflow-agent-kit` | `createAgent` factory, context builder, two-layer extract pipeline |
+| `workflow-agent-hermes` | `@uncaged/workflow-agent-hermes` | `uwf-hermes` agent — spawns Hermes chat, captures session |
+| `workflow-util` | `@uncaged/workflow-util` | Crockford Base32, ULID, logger, frontmatter parsing |
 
-## Monorepo Packages
-
-```
-packages/
-  workflow/                      # @uncaged/workflow — core lib (types, engine, hash, ULID, registry)
-  cli-workflow/                  # @uncaged/cli-workflow — CLI (`uncaged-workflow` command)
-  workflow-template-develop/     # @uncaged/workflow-template-develop — develop workflow template (includes roles)
-  workflow-template-solve-issue/ # @uncaged/workflow-template-solve-issue — solve-issue workflow template (includes roles)
-  workflow-agent-hermes/         # @uncaged/workflow-agent-hermes — Hermes agent adapter
-  workflow-agent-cursor/         # @uncaged/workflow-agent-cursor — Cursor agent adapter
-  workflow-agent-llm/            # @uncaged/workflow-agent-llm — LLM agent adapter
-  workflow-util-agent/           # @uncaged/workflow-util-agent — agent utilities (buildAgentPrompt, spawnCli)
-```
-
-Managed with **bun workspace** using the `workspace:*` protocol.
+External: [`@uncaged/json-cas`](https://www.npmjs.com/package/@uncaged/json-cas) (CAS store + JSON Schema validation) + `@uncaged/json-cas-fs` (filesystem backend).
 
 ## Quick Start
 
 ```bash
-# Install dependencies
-bun install
+# 1. Configure provider and model
+uwf setup
 
-# Build all packages
-bun run build
+# 2. Register a workflow from YAML
+uwf workflow put examples/solve-issue.yaml
 
-# Register a workflow bundle
-uncaged-workflow workflow add solve-issue dist/packages/workflow-template-solve-issue/solve-issue.esm.js
+# 3. Start a thread
+uwf thread start solve-issue -p "Fix the login redirect bug"
 
-# Run a workflow
-uncaged-workflow run solve-issue --prompt "Fix bug #42"
+# 4. Execute steps (one at a time, until done)
+uwf thread step <thread-id>
 ```
 
-## CLI Usage
+## CLI Commands
 
-```bash
-uncaged-workflow                   # Print full command usage (exits with status 1)
-uncaged-workflow workflow list     # List registered workflows
-uncaged-workflow run <name>        # Start a workflow thread
-uncaged-workflow thread list       # List all threads
-uncaged-workflow thread show <id>  # Inspect a thread
-uncaged-workflow skill             # Agent-consumable reference docs
-```
+### Thread
 
-Run `uncaged-workflow` with no arguments to print usage, or `uncaged-workflow skill cli` for the full CLI skill reference.
+| Command | Description |
+|---------|-------------|
+| `uwf thread start <workflow> -p <prompt>` | Create a thread (no execution) |
+| `uwf thread step <thread-id> [--agent <cmd>]` | Execute one moderator→agent→extract cycle |
+| `uwf thread show <thread-id>` | Show head pointer and done status |
+| `uwf thread list [--all]` | List threads (`--all` includes archived) |
+| `uwf thread steps <thread-id>` | List all steps chronologically |
+| `uwf thread read <thread-id> [--quota N]` | Render thread as readable markdown |
+| `uwf thread fork <step-hash>` | Fork from a specific step |
+| `uwf thread step-details <step-hash>` | Dump full detail node |
+| `uwf thread kill <thread-id>` | Terminate and archive |
+
+### Workflow
+
+| Command | Description |
+|---------|-------------|
+| `uwf workflow put <file.yaml>` | Register a workflow from YAML |
+| `uwf workflow show <name-or-hash>` | Show workflow definition |
+| `uwf workflow list` | List registered workflows |
+
+### CAS
+
+| Command | Description |
+|---------|-------------|
+| `uwf cas get <hash>` | Read a CAS node |
+| `uwf cas put <type-hash> <data>` | Store a node |
+| `uwf cas has <hash>` | Check existence |
+| `uwf cas refs <hash>` | List direct references |
+| `uwf cas walk <hash>` | Recursive traversal |
+| `uwf cas reindex` | Rebuild type index |
+| `uwf cas schema list` | List schemas |
+| `uwf cas schema get <hash>` | Show a schema |
+
+### Setup
+
+| Command | Description |
+|---------|-------------|
+| `uwf setup` | Interactive provider/model/agent configuration |
+| `uwf setup --provider ... --base-url ... --api-key ... --model ...` | Non-interactive setup |
+
+Config stored in `~/.uncaged/workflow/config.yaml`. API keys in `~/.uncaged/workflow/.env`.
 
 ## Development
 
 ```bash
-bun run check    # Biome lint + format check
-bun run format   # Auto-format with Biome
-bun test         # Run tests
+bun install --no-cache     # Install dependencies
+bun run check              # tsc + biome + lint-log-tags
+bun run format             # Auto-format with Biome
+bun test                   # Run all tests
 ```
 
+Managed with **bun workspace**. See [CLAUDE.md](CLAUDE.md) for coding conventions.
+
 ## Architecture
 
-See [docs/architecture.md](docs/architecture.md) for the full design — three-phase engine loop, bundle contract, storage layout, and design decisions.
+See [docs/architecture.md](docs/architecture.md) for the full design — three-phase engine loop, CAS node types, storage layout, agent CLI protocol, and design decisions.

biome.json

@@ -5,6 +5,8 @@
       "**",
       "!**/dist",
       "!**/node_modules",
+      "!**/legacy-packages",
+      "!scripts",
       "!packages/workflow/workflow",
       "!xiaoju/scripts/bundle.ts"
     ]
@@ -36,7 +38,7 @@
       }
     },
     {
-      "includes": ["**/*.d.ts"],
+      "includes": ["**/*.d.ts", "**/vitest.config.*"],
       "linter": {
         "rules": {
           "style": {
@@ -44,6 +46,16 @@
           }
         }
       }
+    },
+    {
+      "includes": ["**/cli.ts", "**/setup.ts"],
+      "linter": {
+        "rules": {
+          "suspicious": {
+            "noConsole": "off"
+          }
+        }
+      }
     }
   ],
   "linter": {

docs/architecture.md

@@ -1,271 +1,495 @@
-# Uncaged workflow — Architecture
+# Workflow Engine — Architecture
 
-**Last updated:** 2026-05-09
+**Last updated:** 2026-05-19
 
 ---
 
 ## Overview
 
-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.
+A stateless workflow engine driven by a single-step CLI. Workflows are YAML definitions stored as CAS nodes; threads are immutable chains of CAS-linked step nodes. No daemon — each `uwf thread step` invocation runs one moderator→agent→extract cycle and exits.
 
-The implementation lives in **21** Bun workspace packages under `packages/`, using the `workspace:*` protocol.
+The implementation lives in **6** active packages under `packages/`, plus two external CAS packages (`@uncaged/json-cas`, `@uncaged/json-cas-fs`). Legacy packages reside in `legacy-packages/` and are not part of the active stack.
 
 ## Package map
 
-Grouped by responsibility (npm name → folder).
-
 | 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-office` → `workflow-agent-office` | `AdapterFn` via `office-agent` CLI; generates or edits Word documents, stores outputs per threadId. |
-| | `@uncaged/workflow-agent-docx-diff` → `workflow-agent-docx-diff` | `AdapterFn` via `docx-diff` CLI; produces Word-format diff reports for document edit workflows. |
-| | `@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. |
-| | `@uncaged/workflow-template-document` → `workflow-template-document` | Document generation/editing workflow definition (writer + differ roles, moderator table, descriptor). |
-| Dashboard | `@uncaged/workflow-dashboard` → `workflow-dashboard` | Private Vite + React app (`src/main.tsx`); only `react` / `react-dom` dependencies — no workspace packages. |
+|-------|---------|---------------|
+| Contract | `@uncaged/workflow-protocol` → `workflow-protocol` | Shared TypeScript types (`WorkflowPayload`, `StepNodePayload`, `ModeratorContext`, `WorkflowConfig`, etc.). No runtime deps beyond `@uncaged/json-cas-fs`. |
+| Shared infra | `@uncaged/workflow-util` → `workflow-util` | Crockford Base32, ULID generation, `createLogger`, frontmatter parsing/validation. |
+| Moderator | `@uncaged/workflow-moderator` → `workflow-moderator` | JSONata-based graph evaluator: given a `WorkflowPayload` and `ModeratorContext`, returns the next role or `$END`. |
+| Agent framework | `@uncaged/workflow-agent-kit` → `workflow-agent-kit` | `createAgent` entrypoint factory, context builder, frontmatter fast-path extractor, LLM extract fallback, output format instruction builder. |
+| Agent: Hermes | `@uncaged/workflow-agent-hermes` → `workflow-agent-hermes` | `uwf-hermes` CLI binary — spawns `hermes chat`, pipes prompt, captures session detail. |
+| CLI | `@uncaged/cli-workflow` → `cli-workflow` | `uwf` binary — thread lifecycle, workflow registry, CAS inspection, setup. |
 
-## Dependency graph (workspace packages)
+### External dependencies
 
-Bottom-up layering for the execution stack:
+| Package | Role |
+|---------|------|
+| `@uncaged/json-cas` | Content-addressed store API, XXH64 hashing, JSON Schema registration and validation. |
+| `@uncaged/json-cas-fs` | Filesystem backend for `json-cas`. |
+| `jsonata` | JSONata expression evaluator (used by `workflow-moderator`). |
+| `commander` | CLI argument parsing (used by `cli-workflow`). |
+| `dotenv` | Loads `.env` files for API keys. |
+| `yaml` | YAML parse/stringify. |
+
+## Dependency graph
 
 ```mermaid
 flowchart BT
+  subgraph External
+    jcas["@uncaged/json-cas"]
+    jcasfs["@uncaged/json-cas-fs"]
+  end
   subgraph L0["Layer 0 — contract"]
     protocol["@uncaged/workflow-protocol"]
   end
-  subgraph L1["Layer 1 — on protocol"]
-    runtime["@uncaged/workflow-runtime"]
+  subgraph L1["Layer 1 — shared"]
     util["@uncaged/workflow-util"]
-    reactor["@uncaged/workflow-reactor"]
+    moderator["@uncaged/workflow-moderator"]
   end
-  subgraph L2["Layer 2 — protocol + util"]
-    cas["@uncaged/workflow-cas"]
-    register["@uncaged/workflow-register"]
+  subgraph L2["Layer 2 — agent framework"]
+    kit["@uncaged/workflow-agent-kit"]
   end
-  subgraph L3["Layer 3 — engine"]
-    execute["@uncaged/workflow-execute"]
+  subgraph L3["Layer 3 — agent implementations"]
+    hermes["@uncaged/workflow-agent-hermes"]
   end
   subgraph L4["Layer 4 — CLI"]
     cli["@uncaged/cli-workflow"]
   end
-  runtime --> protocol
+  protocol --> jcasfs
   util --> protocol
-  reactor --> protocol
-  cas --> protocol
-  cas --> util
-  register --> protocol
-  register --> util
-  execute --> protocol
-  execute --> runtime
-  execute --> util
-  execute --> cas
-  execute --> reactor
-  execute --> register
+  moderator --> protocol
+  kit --> protocol
+  kit --> util
+  kit --> jcas
+  kit --> jcasfs
+  hermes --> kit
+  hermes --> jcas
   cli --> protocol
   cli --> util
-  cli --> cas
-  cli --> execute
-  cli --> register
-  cli --> runtime
+  cli --> kit
+  cli --> moderator
+  cli --> jcas
+  cli --> jcasfs
 ```
 
-**Adjacent consumers** (not in the main CLI stack):
+## Workflow definition
 
-- `@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`)
+Workflows are **YAML files** (not ESM bundles). `uwf workflow put <file.yaml>` parses the YAML, registers output schemas as JSON Schema CAS nodes, and stores the `WorkflowPayload` as a CAS node.
 
-## Package roles (detail)
+Example (`examples/solve-issue.yaml`):
 
-- **`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`.
+```yaml
+name: "solve-issue"
+description: "End-to-end issue resolution"
+roles:
+  planner:
+    description: "Creates implementation plan"
+    goal: "You are a planning agent. Analyze the issue and create a step-by-step plan."
+    capabilities:
+      - issue-analysis
+      - planning
+    procedure: "Analyze the issue and create a detailed, actionable implementation plan."
+    output: "Output the plan summary and list of concrete steps."
+    meta:
+      type: object
+      properties:
+        plan: { type: string }
+        steps: { type: array, items: { type: string } }
+      required: [plan, steps]
+  developer:
+    description: "Implements code changes"
+    goal: "You are a developer agent. Implement the plan."
+    capabilities:
+      - file-edit
+      - shell
+    procedure: "Implement the plan. Write code, tests, and ensure existing tests pass."
+    output: "List all files changed and provide a summary of the implementation."
+    meta:
+      type: object
+      properties:
+        filesChanged: { type: array, items: { type: string } }
+        summary: { type: string }
+      required: [filesChanged, summary]
+  reviewer:
+    description: "Reviews code changes"
+    goal: "You are a code reviewer. Review the implementation."
+    capabilities:
+      - code-review
+    procedure: "Review the implementation against the plan."
+    output: "Approve or reject with detailed comments."
+    meta:
+      type: object
+      properties:
+        approved: { type: boolean }
+        comments: { type: string }
+      required: [approved, comments]
+conditions:
+  notApproved:
+    description: "Reviewer rejected the implementation"
+    expression: "steps[-1].output.approved = false"
+graph:
+  $START:
+    - role: "planner"
+      condition: null
+  planner:
+    - role: "developer"
+      condition: null
+  developer:
+    - role: "reviewer"
+      condition: null
+  reviewer:
+    - role: "developer"
+      condition: "notApproved"
+    - role: "$END"
+      condition: null
+```
+
+Key properties:
+
+- **`roles`** — inline role definitions; each `meta` is a JSON Schema (stored as its own CAS node on registration)
+- **`conditions`** — named JSONata expressions evaluated against the `ModeratorContext`
+- **`graph`** — `Record<Role | "$START", Transition[]>` — first matching transition wins; `condition: null` = fallback
+- **No agent binding** — agent selection is a deployment concern, configured in `config.yaml`
+- **No Zod** — all schemas are JSON Schema, validated through `@uncaged/json-cas`
 
 ## 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`.
+Each `uwf thread step` runs exactly one cycle: moderator → agent → extract. The CLI orchestrates this in `packages/cli-workflow/src/commands/thread.ts` (`cmdThreadStep`).
 
 ```
 ┌─→ Phase 1: MODERATOR
-│   Context: ModeratorContext { threadId, depth, start, steps }
-│   Action:  moderator(ctx) → role name | END
+│   Input:  WorkflowPayload + ModeratorContext { start, steps[] }
+│   Engine: JSONata conditions evaluated against the graph
+│   Output: next role name | $END
 │
 │   Phase 2: AGENT
-│   Context: AgentContext = ModeratorCtx + { currentRole: { name, systemPrompt } }
-│   Action:  agent(ctx) → raw string
+│   Input:  thread-id + role (via argv)
+│   Engine: agent-kit builds context from CAS chain, prepends
+│           output format instruction to system prompt, spawns agent
+│   Output: raw string (frontmatter markdown)
 │
-│   Phase 3: EXTRACTOR
-│   Context: ExtractContext = AgentCtx + { agentContent }
-│   Action:  runtime.extract(schema, extractPrompt, ctx) → typed meta
+│   Phase 3: EXTRACT
+│   Input:  raw agent output + role's meta schema
+│   Engine: two-layer extract (frontmatter fast path → LLM fallback)
+│   Output: CasRef to structured output node
 │
-│   Merge: RoleStep { role, contentHash, meta, refs, timestamp }
-│   Append to steps
-└─────────────────────────────────────────────────────┘
+│   Persist: StepNode { start, prev, role, output, detail, agent }
+│   Update:  threads.yaml head pointer
+└─────────────────────────────────────────────────────────────────┘
 ```
 
-### Context types (progressive)
+### Context types
 
 Defined in `packages/workflow-protocol/src/types.ts`:
 
 ```typescript
-type ModeratorContext<M> = ThreadContext<M>;
-type AgentContext<M> = ModeratorContext<M> & {
-  currentRole: { name: string; systemPrompt: string };
+type StepContext = {
+  role: string;
+  output: unknown;    // CAS node payload, expanded (not hash)
+  detail: CasRef;
+  agent: string;
+};
+
+type ModeratorContext = {
+  start: StartNodePayload;  // { workflow: CasRef, prompt: string }
+  steps: StepContext[];     // chronological, oldest first
+};
+
+type AgentContext = ModeratorContext & {
+  threadId: ThreadId;
+  role: string;
+  store: Store;
+  workflow: WorkflowPayload;
+  outputFormatInstruction: string;
 };
-type ExtractContext<M> = AgentContext<M> & { agentContent: string };
 ```
 
 ### Key properties
 
-- **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.
+- **Moderator** — pure JSONata evaluation; no LLM call, no I/O beyond CAS reads. Evaluates `workflow.graph[currentRole]` transitions in order, returns first match.
+- **Agent** — receives `AgentContext` with thread history + role system prompt + output format instruction. Raw output is frontmatter markdown.
+- **Extractor** — two-layer: tries frontmatter fast-path first (zero LLM cost), falls back to LLM extract if frontmatter is absent or invalid.
+- **Stateless** — each `uwf thread step` is an atomic, self-contained operation. No in-memory state between steps.
 
-## Agent information sources
+## Agent CLI protocol
 
-An agent has exactly three information sources:
+Each agent is an external command invoked by `uwf thread step`:
 
-1. **Prior knowledge** — LLM training, agent memory, agent skills
-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 obtains it via `ExtractFn` (e.g. Cursor agent).
-
-## Bundle contract
-
-A workflow bundle is a single `.esm.js` file with two named exports (see `WorkflowFn` / `WorkflowDescriptor` in `packages/workflow-protocol/src/types.ts`):
-
-```typescript
-export const descriptor: WorkflowDescriptor;
-export const run: WorkflowFn;
-
-type WorkflowFn = (
-  thread: ThreadContext,
-  runtime: WorkflowRuntime,
-) => AsyncGenerator<RoleOutput, WorkflowCompletion>;
+```bash
+<agent-cmd> <thread-id> <role>
 ```
 
-`RoleOutput` carries `contentHash`, `meta`, and `refs` (agent text lives in CAS, addressed by hash).
+Contract:
+1. `uwf thread step` determines the next role via the moderator
+2. Agent CLI is spawned with `(thread-id, role)` as positional args
+3. `workflow-agent-kit` (`createAgent`) handles the boilerplate:
+   - Parses argv
+   - Loads `.env` from storage root
+   - Builds `AgentContext` by walking the CAS chain from `threads.yaml` head
+   - Resolves the role's `meta` schema and builds `outputFormatInstruction`
+   - Calls the agent's `run` function
+   - Runs two-layer extract on the raw output
+   - Writes `StepNode` to CAS (output + detail + prev link)
+   - Prints the new `StepNode` CAS hash to stdout
+4. `uwf thread step` reads stdout, updates `threads.yaml` head pointer, re-evaluates moderator for `done`
+5. Exit 0 = success, non-zero = failure
 
-### Constraints
+Agent resolution priority: `--agent` CLI override → `config.yaml` per-workflow/role override → `config.yaml` `defaultAgent`.
 
-- Single `.esm.js` file
-- 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
+## Agent output format: frontmatter markdown (RFC #351)
 
-### Why AsyncGenerator?
+Agents produce **frontmatter markdown** — YAML frontmatter for structured meta, followed by a markdown body for content:
 
-- 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
+```markdown
+---
+status: done
+next: reviewer
+confidence: 0.9
+artifacts:
+  - src/auth.ts
+scope: role
+---
+
+## Implementation
+
+Fixed the login redirect by updating the auth middleware...
+```
+
+The `outputFormatInstruction` (built by `buildOutputFormatInstruction` in `workflow-agent-kit`) is prepended to the role's system prompt, so the deliverable format is the first thing the agent sees. It lists the expected frontmatter fields derived from the role's `meta` JSON Schema.
+
+## Two-layer extract
+
+Structured output extraction uses a two-layer strategy (`workflow-agent-kit`):
+
+### Layer 1: frontmatter fast path (`frontmatter.ts`)
+
+1. Parse YAML frontmatter from raw agent output (`parseFrontmatterMarkdown`)
+2. Validate required fields (`validateFrontmatter`)
+3. Build a candidate object from frontmatter fields (`status`, `next`, `confidence`, `artifacts`, `scope`)
+4. `store.put()` the candidate against the role's `meta` schema
+5. Validate with `json-cas` schema validation
+6. If valid → return `outputHash` (zero LLM cost)
+
+### Layer 2: LLM extract fallback (`extract.ts`)
+
+If the fast path returns `null` (no frontmatter, invalid, or doesn't satisfy schema):
+
+1. Resolve extract model alias from config (`modelOverrides.extract` → `models.extract` → `defaultModel`)
+2. Call OpenAI-compatible chat completion with JSON mode
+3. System prompt: "Extract structured data matching this JSON Schema: ..."
+4. User message: the raw agent output
+5. Parse response, `store.put()`, validate
+6. Return `outputHash`
+
+## Prompt injection
+
+`workflow-agent-kit` prepends two pieces of context to the agent's system prompt:
+
+1. **Deliverable format instruction** — generated from the role's `meta` schema, tells the agent exactly what frontmatter fields to produce and the expected format
+2. **Scope constraint** — "Focus exclusively on YOUR role's deliverable. Do not perform actions outside your role's scope."
+
+This ensures agents produce parseable frontmatter output without requiring per-agent format knowledge.
+
+## CAS node types
+
+### Workflow
+
+```yaml
+type: <workflow-schema-hash>
+payload:
+  name: "solve-issue"
+  description: "End-to-end issue resolution"
+  roles:
+    planner:
+      description: "Creates implementation plan"
+      goal: "You are a planning agent..."
+      capabilities: [planning, issue-analysis]
+      procedure: "Analyze the issue and create a plan."
+      output: "Output the plan summary."
+      meta: "5GWKR8TN1V3JA"    # cas_ref → JSON Schema node
+  conditions:
+    notApproved:
+      description: "Reviewer rejected"
+      expression: "steps[-1].output.approved = false"
+  graph:
+    $START:
+      - role: "planner"
+        condition: null
+```
+
+### StartNode
+
+```yaml
+type: <start-node-schema-hash>
+payload:
+  workflow: "4KNM2PXR3B1QW"    # cas_ref → Workflow
+  prompt: "Fix the login bug..."
+```
+
+### StepNode
+
+```yaml
+type: <step-node-schema-hash>
+payload:
+  start: "4TNVW8KR2B3MA"      # cas_ref → StartNode
+  prev: "2MXBG6PN4A8JR"       # cas_ref → previous StepNode (null for first step)
+  role: "developer"
+  output: "9KRVW3TN5F1QA"     # cas_ref → structured output (validated against meta schema)
+  detail: "7BQST3VW9F2MA"     # cas_ref → execution detail (raw turns, session data)
+  agent: "uwf-hermes"         # agent command used (plain string)
+```
+
+### Chain structure
+
+```
+threads.yaml: { "01J7K9...4T": "8FWKR3TN5V1QA" }
+                                    │
+                                    ▼
+                            StepNode (step 3)
+                            ├── start ──→ StartNode
+                            │              ├── workflow → Workflow (CAS)
+                            │              └── prompt: "Fix..."
+                            ├── prev ──→ StepNode (step 2)
+                            │             ├── prev ──→ StepNode (step 1)
+                            │             │             └── prev: null
+                            │             └── ...
+                            ├── role: "reviewer"
+                            ├── output → CAS({ approved: true })
+                            ├── detail → CAS(session turns)
+                            └── agent: "uwf-hermes"
+```
 
 ## Storage layout
 
 ```
 ~/.uncaged/workflow/
-├── cas/                           # Global content-addressed blobs (see getGlobalCasDir)
-├── bundles/
-│   ├── C9NMV6V2TQT81.esm.js       # Crockford Base32 of XXH64
-│   ├── 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.running     # Present while worker executes this thread (optional)
-│       └── 01KQXKW…YG.info.jsonl   # Debug log
-└── workflow.yaml                  # Registry
+├── cas/                          # json-cas filesystem store (all CAS nodes)
+├── config.yaml                   # Provider, model, agent configuration
+├── threads.yaml                  # Active thread head pointers: threadId → CasRef
+├── history.jsonl                 # Archived thread records
+├── registry.yaml                 # Workflow name → CAS hash mapping
+└── .env                          # API keys (loaded by dotenv)
 ```
 
+### Mutable state
+
+Only three files carry mutable state:
+
+| File | Contents |
+|------|----------|
+| `threads.yaml` | `Record<ThreadId, CasRef>` — maps active thread IDs to head node hash |
+| `history.jsonl` | Append-only log of completed threads (`thread`, `workflow`, `head`, `completedAt`) |
+| `registry.yaml` | Workflow name → current CAS hash |
+
+Everything else is immutable CAS content.
+
 ### ID encoding: Crockford Base32
 
 - Case-insensitive, filesystem-safe, no ambiguous chars (0/O, 1/I/L)
-- Bundle hash: XXH64 → 13-char
-- Thread ID: ULID → 26-char (10 timestamp + 16 random)
+- CAS hash: XXH64 → 13-char Crockford Base32
+- Thread ID: ULID → 26-char Crockford Base32 (10 timestamp + 16 random)
 
-### Registry (`workflow.yaml`)
+### Config (`config.yaml`)
 
-Managed by `@uncaged/workflow-register` (`readWorkflowRegistry`, `writeWorkflowRegistry`, …). Shape includes workflow entries and a top-level `config` section used for extract/supervisor model resolution.
+```yaml
+providers:
+  openrouter:
+    baseUrl: "https://openrouter.ai/api/v1"
+    apiKeyEnv: "OPENROUTER_API_KEY"
 
-### Thread storage (CAS + index)
+models:
+  sonnet:
+    provider: "openrouter"
+    name: "anthropic/claude-sonnet-4"
+  gpt4o-mini:
+    provider: "openai"
+    name: "gpt-4o-mini"
 
-Thread execution state is a chain of immutable CAS nodes (`StartNode`, `StateNode`, content Merkle blobs). Per bundle:
+agents:
+  hermes:
+    command: "uwf-hermes"
+    args: []
+  cursor:
+    command: "uwf-cursor"
+    args: []
 
-- **`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.
+defaultAgent: "hermes"
+agentOverrides:
+  solve-issue:
+    developer: "cursor"
 
-**`.info.jsonl`** — Structured debug log via `@uncaged/workflow-util` `createLogger`:
-
-```jsonc
-{ "tag": "4KNMR2PX", "content": "Loading bundle...", "timestamp": ... }
+defaultModel: "sonnet"
+modelOverrides:
+  extract: "gpt4o-mini"
 ```
 
-Tags are 8-char Crockford Base32 (40-bit random), one per call site. `grep "4KNMR2PX"` → code location.
-
-## Execution model
-
-- **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
 
-| Priority | Command | Description |
-|----------|---------|-------------|
-| P1 | `add <name> <file.esm.js>` | Register a bundle |
-| P1 | `list` | List registered workflows |
-| P1 | `show <name>` | Show workflow details |
-| P1 | `remove <name>` | Remove a workflow |
-| P1 | `run <name> [--prompt] [--max-rounds]` | Start a thread |
-| P1 | `threads [name]` | List threads |
-| P1 | `thread <id>` | Show thread state |
-| P1 | `thread rm <id>` | Delete a thread |
-| P1 | `ps` | List running threads |
-| P1 | `kill <thread-id>` | Terminate a running thread |
-| P2 | `history <name>` | Show version history |
-| P2 | `rollback <name> [hash]` | Switch to a previous version |
-| P2 | `pause <thread-id>` | Pause a running thread |
-| P2 | `resume <thread-id>` | Resume a paused thread |
-| P3 | `fork <thread-id> [--from-role <role>]` | Fork from historical state |
+Binary: `uwf`
+
+### Thread commands
+
+| Command | Description |
+|---------|-------------|
+| `uwf thread start <workflow> -p <prompt>` | Create a thread (StartNode → CAS, head → threads.yaml). No execution. |
+| `uwf thread step <thread-id> [--agent <cmd>]` | Execute one moderator→agent→extract cycle. |
+| `uwf thread show <thread-id>` | Show thread head pointer and done status. |
+| `uwf thread list [--all]` | List active threads (`--all` includes archived). |
+| `uwf thread steps <thread-id>` | List all steps in chronological order. |
+| `uwf thread read <thread-id> [--quota <chars>] [--before <hash>]` | Render thread as human-readable markdown. |
+| `uwf thread fork <step-hash>` | Fork a thread from a specific CAS node. |
+| `uwf thread step-details <step-hash>` | Dump full detail node as YAML. |
+| `uwf thread kill <thread-id>` | Terminate and archive a thread. |
+
+### Workflow commands
+
+| Command | Description |
+|---------|-------------|
+| `uwf workflow put <file.yaml>` | Register a workflow from YAML definition. |
+| `uwf workflow show <id>` | Show workflow by name or CAS hash. |
+| `uwf workflow list` | List registered workflows. |
+
+### CAS commands
+
+| Command | Description |
+|---------|-------------|
+| `uwf cas get <hash>` | Read a CAS node. |
+| `uwf cas put <type-hash> <data>` | Store a node, print its hash. |
+| `uwf cas has <hash>` | Check if a hash exists. |
+| `uwf cas refs <hash>` | List direct CAS references. |
+| `uwf cas walk <hash>` | Recursive traversal from a node. |
+| `uwf cas reindex` | Rebuild type index from all nodes. |
+| `uwf cas schema list` | List registered schemas. |
+| `uwf cas schema get <hash>` | Show a schema by type hash. |
+
+### Setup
+
+| Command | Description |
+|---------|-------------|
+| `uwf setup [--provider --base-url --api-key --model --agent]` | Configure provider/model/agent (interactive if no flags). |
+
+## Toolchain
+
+| Tool | Purpose |
+|------|---------|
+| **bun** | Package manager + runtime |
+| **TypeScript** | Type checking (strict mode) |
+| **Biome** | Lint + format |
+| **vitest** | Test runner |
 
 ## 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; 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 |
-| **21-package split** | Clear boundaries: protocol ↔ runtime author API ↔ util/CAS/register ↔ execute ↔ CLI ↔ agents/templates/UI |
+| **YAML workflow definitions** | Human-readable, versionable, no build step required. JSON Schema inline in YAML, registered as CAS nodes on `workflow put`. |
+| **Stateless single-step CLI** | Each `uwf thread step` is atomic — no in-memory state, no daemon, no long-running process. OS handles lifecycle. |
+| **CAS-backed thread state** | Immutable linked nodes enable fork, replay, and GC without copying data. Content-addressed deduplication across threads. |
+| **JSONata moderator** | Declarative condition expressions evaluated against thread history. No LLM cost for routing decisions. |
+| **Frontmatter markdown output** | Agents produce structured meta (YAML frontmatter) alongside free-form content (markdown body). Enables zero-cost extraction when frontmatter is well-formed. |
+| **Two-layer extract** | Fast path avoids LLM calls when agents follow the format; LLM fallback handles messy output gracefully. |
+| **Prompt injection for format** | Output format instruction prepended to system prompt ensures agents produce parseable output without per-agent configuration. |
+| **JSON Schema (not Zod)** | Schemas are CAS-native data — storable, hashable, validatable through `json-cas`. No code generation, no runtime library dependency. |
+| **Agent as external command** | Agents are independent CLI binaries (`uwf-hermes`, `uwf-cursor`). Swappable per workflow/role via config. No tight coupling to the engine. |
+| **No daemon** | Process starts, does one step, exits. Simpler failure model, no connection management. |
+| **Crockford Base32** | Filesystem-safe, case-insensitive, readable, compact. |

docs/builtin-agent-research.md

@@ -0,0 +1,779 @@
+# Built-in Role Agent 调研
+
+## 目标
+
+实现一个内置的 role agent（暂称 `uwf-builtin`），不依赖 hermes/openclaw 等外部 agent 进程。
+直接使用 workflow config 中配置的 model，自己实现 agent run loop 和关键 toolkit。
+
+---
+
+## 关键问题
+
+### Q1: Agent 接口协议
+
+现有 agent 是怎么被 CLI 调用的？输入（argv、环境变量）和输出（stdout、CAS）格式是什么？
+
+**调研要点：**
+- `cli-workflow` 里 `spawnAgent` 的完整实现
+- AgentConfig 类型定义
+- agent 进程的 exit code 约定
+- 环境变量传递（UWF_STORAGE_ROOT 等）
+
+**答案：**
+
+#### 调用链
+
+`uwf thread step` → `cmdThreadStepOnce` → moderator 求值下一 role → `resolveAgentConfig` → `spawnAgent`。
+
+#### AgentConfig 类型
+
+```146:149:packages/workflow-protocol/src/types.ts
+export type AgentConfig = {
+  command: string;
+  args: string[];
+};
+```
+
+在 `config.yaml` 的 `agents` 段注册，例如 `hermes: { command: "uwf-hermes", args: [] }`。
+
+#### spawnAgent 行为
+
+```627:653:packages/cli-workflow/src/commands/thread.ts
+function spawnAgent(agent: AgentConfig, threadId: ThreadId, role: string): CasRef {
+  const argv = [...agent.args, threadId, role];
+  let stdout: string;
+  try {
+    stdout = execFileSync(agent.command, argv, {
+      encoding: "utf8",
+      env: process.env,
+      stdio: ["ignore", "pipe", "pipe"],
+    });
+  } catch (e) {
+  // ... stderr 拼进 fail 消息
+  }
+
+  const line = stdout.trim().split("\n").pop()?.trim() ?? "";
+  if (!isCasRef(line)) {
+    fail(`agent stdout is not a valid CAS hash: ${line || "(empty)"}`);
+  }
+  return line;
+}
+```
+
+| 项目 | 约定 |
+|------|------|
+| **argv** | `[...agent.args, <thread-id>, <role>]`，即 `process.argv[2]`=threadId，`process.argv[3]`=role（与 `createAgent` 的 `parseArgv` 一致） |
+| **stdin** | 忽略 |
+| **stdout** | 纯文本，**最后一行**必须是新 `StepNode` 的 CAS hash（13 字符 Crockford Base32） |
+| **stderr** | 失败时 CLI 会附带 stderr；成功时无约定 |
+| **exit code** | `0` = 成功；非 0 时 `execFileSync` 抛错，step 失败 |
+| **环境变量** | 继承父进程 `process.env`（含 storage root、API key 等） |
+| **链头更新** | **不由 agent 负责**；agent 只写 CAS StepNode，CLI 在拿到 stdout hash 后更新 `threads.yaml` |
+
+Agent 解析优先级（`resolveAgentConfig`）：
+
+1. CLI `--agent` override（整段 command + args 字符串）
+2. `config.agentOverrides[workflow.name][role]`
+3. `config.defaultAgent`
+
+#### 环境变量：Storage Root
+
+文档中写的 `UWF_STORAGE_ROOT` **在当前代码中不存在**。实际优先级（`workflow-agent-kit` / `cli-workflow` 一致）：
+
+```33:43:packages/workflow-agent-kit/src/storage.ts
+export function resolveStorageRoot(): string {
+  const internal = process.env.UNCAGED_WORKFLOW_STORAGE_ROOT;
+  if (internal !== undefined && internal !== "") {
+    return internal;
+  }
+  const userOverride = process.env.WORKFLOW_STORAGE_ROOT;
+  if (userOverride !== undefined && userOverride !== "") {
+    return userOverride;
+  }
+  return getDefaultStorageRoot();
+}
+```
+
+Agent 子进程通过继承的 `process.env` 与父 CLI 共享同一 storage root；`createAgent` 内还会 `loadDotenv({ path: getEnvPath(storageRoot) })` 加载 `~/.uncaged/workflow/.env`。
+
+#### Agent 侧职责（设计文档 + 实现）
+
+- 读 `threads.yaml` 链头，构建 context，执行 role
+- 将 `StepNode` 写入 CAS（`output` / `detail` / `agent` / `prev` / `start`）
+- stdout 打印 step hash
+- **不**更新 `threads.yaml`
+
+---
+
+### Q2: createAgent 工厂
+
+workflow-agent-kit 的 `createAgent` 做了什么？它的完整生命周期是什么？
+
+**调研要点：**
+- `AgentOptions` 类型的 `run` 和 `continue` 回调签名
+- `AgentRunResult` 的完整定义
+- retry 逻辑（frontmatter 校验失败后的重试机制）
+- `persistStep` 写入 CAS 的 StepNode 结构
+
+**答案：**
+
+#### 类型定义
+
+```4:35:packages/workflow-agent-kit/src/types.ts
+export type AgentContext = ModeratorContext & {
+  threadId: ThreadId;
+  role: string;
+  store: Store;
+  workflow: WorkflowPayload;
+  outputFormatInstruction: string;
+};
+
+export type AgentRunResult = {
+  output: string;
+  detailHash: CasRef;
+  sessionId: string;
+};
+
+export type AgentContinueFn = (
+  sessionId: string,
+  message: string,
+  store: AgentContext["store"],
+) => Promise<AgentRunResult>;
+
+export type AgentRunFn = (ctx: AgentContext) => Promise<AgentRunResult>;
+
+export type AgentOptions = {
+  name: string;
+  run: AgentRunFn;
+  continue: AgentContinueFn;
+};
+```
+
+- **`run(ctx)`**：首次执行，返回原始 agent 文本 `output`、审计用 `detailHash`、用于续聊的 `sessionId`。
+- **`continue(sessionId, message, store)`**：在同一 session 上追加用户消息（用于 frontmatter 纠错），再次返回 `AgentRunResult`。
+
+`createAgent(options)` 返回 `() => Promise<void>`，作为 agent CLI 的 `main`（见 `uwf-hermes` 的 `cli.ts`）。
+
+#### 生命周期（按执行顺序）
+
+```101:152:packages/workflow-agent-kit/src/run.ts
+export function createAgent(options: AgentOptions): () => Promise<void> {
+  return async function main(): Promise<void> {
+    const { threadId, role } = parseArgv(process.argv);
+    const storageRoot = resolveStorageRoot();
+    loadDotenv({ path: getEnvPath(storageRoot) });
+
+    const ctx = await buildContextWithMeta(threadId, role);
+    // 1. 校验 role 存在
+    // 2. 从 CAS 取 frontmatter JSON Schema → buildOutputFormatInstruction → ctx.outputFormatInstruction
+
+    let agentResult = await options.run(ctx);
+
+    let outputHash = await tryExtractOutput(agentResult.output, roleDef.frontmatter, ctx);
+
+    for (let retry = 0; retry < MAX_FRONTMATTER_RETRIES && outputHash === null; retry++) {
+      const correctionMessage = "Your previous response did not contain valid YAML frontmatter...";
+      agentResult = await options.continue(agentResult.sessionId, correctionMessage, ctx.meta.store);
+      outputHash = await tryExtractOutput(agentResult.output, roleDef.frontmatter, ctx);
+    }
+
+    if (outputHash === null) { fail(...); }
+
+    const stepHash = await persistStep({ ctx, outputHash, detailHash: agentResult.detailHash, agentName });
+    process.stdout.write(`${stepHash}\n`);
+  };
+}
+```
+
+| 阶段 | 行为 |
+|------|------|
+| 解析 argv | `argv[2]=threadId`, `argv[3]=role`，缺失则 `stderr` + `exit(1)` |
+| Context | `buildContextWithMeta` + 可选 `outputFormatInstruction` |
+| Run | `options.run(ctx)` |
+| Extract | **仅** `tryFrontmatterFastPath`（见 Q4）；**不**调用 `extract()` LLM fallback |
+| Retry | 最多 `MAX_FRONTMATTER_RETRIES = 2` 次 `continue` + 再试 fast-path |
+| Persist | `persistStep` → `writeStepNode` |
+| 输出 | stdout 一行 step CAS hash |
+
+#### StepNode 写入结构
+
+```44:68:packages/workflow-agent-kit/src/run.ts
+async function writeStepNode(options: {
+  store: AgentStore["store"];
+  schemas: AgentStore["schemas"];
+  startHash: CasRef;
+  prevHash: CasRef | null;
+  role: string;
+  outputHash: CasRef;
+  detailHash: CasRef;
+  agentName: string;
+}): Promise<CasRef> {
+  const payload: StepNodePayload = {
+    start: options.startHash,
+    prev: options.prevHash,
+    role: options.role,
+    output: options.outputHash,
+    detail: options.detailHash,
+    agent: options.agentName,
+  };
+  // store.put(stepNode schema) + validate
+}
+```
+
+`agentName` 经 `agentLabel(name)` 规范化：已有 `uwf-` 前缀则原样，否则加 `uwf-`（如 `hermes` → `uwf-hermes`）。
+
+`prevHash`：若链头仍是 `StartNode` 则为 `null`，否则为当前 head step hash。
+
+---
+
+### Q3: Context Builder
+
+`buildContextWithMeta` 构建了什么上下文给 agent？
+
+**调研要点：**
+- `AgentContext` 完整类型定义（所有字段）
+- context 构建过程（CAS chain walk）
+- `outputFormatInstruction` 怎么生成的
+- role definition 怎么获取（从 workflow YAML）
+
+**答案：**
+
+#### AgentContext 字段
+
+继承 `ModeratorContext`：
+
+```60:68:packages/workflow-protocol/src/types.ts
+export type ModeratorContext = {
+  start: StartNodePayload;
+  steps: StepContext[];
+};
+```
+
+```48:51:packages/workflow-protocol/src/types.ts
+export type StartNodePayload = {
+  workflow: CasRef;
+  prompt: string;
+};
+```
+
+```61:63:packages/workflow-protocol/src/types.ts
+export type StepContext = Omit<StepRecord, "output"> & {
+  output: unknown;
+};
+```
+
+`AgentContext` 额外字段：
+
+| 字段 | 类型 | 含义 |
+|------|------|------|
+| `threadId` | `ThreadId` | 当前线程 |
+| `role` | `string` | 本步要执行的角色名 |
+| `store` | `Store` | CAS store（读写节点） |
+| `workflow` | `WorkflowPayload` | 已从 CAS 加载的 workflow 定义 |
+| `outputFormatInstruction` | `string` | 由 `createAgent` 根据 role 的 frontmatter schema 生成；`buildContext*` 初始为 `""` |
+
+`buildContextWithMeta` 还返回 `meta`：
+
+```148:154:packages/workflow-agent-kit/src/context.ts
+export type BuildContextMeta = {
+  storageRoot: string;
+  store: Store;
+  schemas: AgentStore["schemas"];
+  headHash: CasRef;
+  chain: ChainState;
+};
+```
+
+#### CAS chain walk
+
+1. 从 `threads.yaml[threadId]` 取 `headHash`
+2. `walkChain`：若 head 是 `StartNode`，`stepsNewestFirst=[]`；否则沿 `prev` 收集所有 `StepNode`， newest-first
+3. `buildHistory`：反转为时间序，`expandOutput` 把每步 `output` CasRef 展开为 JSON payload（供 prompt / JSONata 使用）
+4. `loadWorkflow`：从 `start.workflow` CasRef 加载 `WorkflowPayload`
+
+#### Role definition 来源
+
+- 作者写在 workflow YAML 的 `roles.<name>`（`goal`, `capabilities`, `procedure`, `output`, `frontmatter` 等）
+- `uwf workflow put` 时 `frontmatter` 内联 JSON Schema 经 `putSchema` 存入 CAS，workflow 里存的是 **CasRef**
+- Agent 运行时：`ctx.workflow.roles[ctx.role]` → `RoleDefinition`
+
+#### outputFormatInstruction
+
+在 `createAgent` 中，若 `getSchema(store, roleDef.frontmatter)` 非空，则：
+
+```typescript
+ctx.outputFormatInstruction = buildOutputFormatInstruction(frontmatterSchema);
+```
+
+`buildOutputFormatInstruction` 根据 JSON Schema 的 `properties` 生成「必须以 `---` YAML frontmatter 开头」的说明和示例字段列表（见 `build-output-format-instruction.ts`）。
+
+各 agent 实现（Hermes / Claude Code）在组装 prompt 时把该块放在最前，再接 `buildRolePrompt(roleDef)`。
+
+---
+
+### Q4: Extract Pipeline
+
+agent 输出怎么被处理成结构化数据？
+
+**调研要点：**
+- frontmatter fast-path 的完整逻辑
+- LLM extract fallback 的实现（`extract.ts`）
+- frontmatter schema 从哪里来（role 定义里的 `frontmatter` 字段）
+- 校验失败时的 correction prompt 是什么
+
+**答案：**
+
+#### Schema 来源
+
+Workflow YAML 中每个 role 的 `frontmatter:` 段是 JSON Schema 对象；注册时：
+
+```66:76:packages/cli-workflow/src/commands/workflow.ts
+async function resolveFrontmatterRef(..., frontmatter: unknown): Promise<CasRef> {
+  // 校验为 JSON Schema → putSchema → 返回 CasRef
+}
+```
+
+运行时 `roleDef.frontmatter` 即该 schema 的 CAS hash；structured `output` 节点用**同一 schema** 写入 CAS。
+
+#### Frontmatter fast-path（createAgent 实际使用的路径）
+
+```148:195:packages/workflow-agent-kit/src/frontmatter.ts
+export async function tryFrontmatterFastPath(
+  raw: string,
+  outputSchema: CasRef,
+  store: Store,
+): Promise<FrontmatterFastPathResult | null>
+```
+
+流程：
+
+1. `parseFrontmatterMarkdown(raw)` → 标准 agent 字段（`status`, `next`, `confidence`, `artifacts`, `scope`）+ body
+2. `validateFrontmatter` 失败 → `null`
+3. `getSchema(store, outputSchema)` + `extractSchemaFields` 得到 role 需要的属性名
+4. `buildCandidate`：从标准 frontmatter + YAML 原始字段拼出符合 schema 的对象
+5. `store.put(outputSchema, candidate)` + `validate` → 成功则 `{ body, outputHash }`
+
+**永不抛错**，失败返回 `null`。
+
+#### LLM extract fallback（已实现但未接入 createAgent）
+
+```135:181:packages/workflow-agent-kit/src/extract.ts
+export async function extract(
+  rawOutput: string,
+  outputSchema: CasRef,
+  config: WorkflowConfig,
+): Promise<ExtractResult>
+```
+
+- 模型：`resolveExtractModelAlias(config)` → `modelOverrides.extract` → `models.extract` → `models.default` → `defaultModel`
+- HTTP：`POST {baseUrl}/chat/completions`，`response_format: { type: "json_object" }`
+- System：要求按 JSON Schema 从 agent 输出提取单个 JSON 对象
+- 校验通过后 `store.put(outputSchema, structured)`
+
+**重要：`createAgent` 当前未调用 `extract()`**。fast-path 失败且 2 次 `continue` 仍失败则直接 `fail()`。builtin agent 若希望无 frontmatter 也能跑，需在 kit 或 builtin 层显式接入 `extract()`。
+
+#### Correction prompt（retry）
+
+```125:128:packages/workflow-agent-kit/src/run.ts
+const correctionMessage =
+  "Your previous response did not contain valid YAML frontmatter matching the role schema.\n" +
+  "You MUST begin your response with a YAML frontmatter block (--- delimited).\n" +
+  "Please output ONLY the corrected frontmatter block followed by your work.";
+```
+
+通过 `options.continue(sessionId, correctionMessage, store)` 发给外部 agent；builtin 需在自有 message 历史里 append 同等语义的 user 消息。
+
+---
+
+### Q5: Model 配置与 LLM 调用
+
+workflow 怎么配置和使用 model？
+
+**调研要点：**
+- `WorkflowConfig` 中 providers/models/defaultModel/modelOverrides 的完整定义
+- `resolveModel` 函数的实现
+- `chatCompletionText` 的实现（OpenAI 兼容 HTTP 客户端）
+- 有没有 streaming 支持？tool calling 支持？
+
+**答案：**
+
+#### WorkflowConfig
+
+```136:160:packages/workflow-protocol/src/types.ts
+export type ProviderConfig = {
+  baseUrl: string;
+  apiKeyEnv: string;
+};
+
+export type ModelConfig = {
+  provider: ProviderAlias;
+  name: string;
+};
+
+export type WorkflowConfig = {
+  providers: Record<ProviderAlias, ProviderConfig>;
+  models: Record<ModelAlias, ModelConfig>;
+  agents: Record<AgentAlias, AgentConfig>;
+  defaultAgent: AgentAlias;
+  agentOverrides: Record<WorkflowName, Record<RoleName, AgentAlias>> | null;
+  defaultModel: ModelAlias;
+  modelOverrides: Record<Scenario, ModelAlias> | null;
+};
+```
+
+示例见 `docs/architecture.md`（`providers` / `models` / `defaultModel` / `modelOverrides.extract`）。
+
+#### resolveModel
+
+```32:50:packages/workflow-agent-kit/src/extract.ts
+export function resolveModel(config: WorkflowConfig, alias: ModelAlias): ResolvedLlmProvider {
+  const modelEntry = config.models[alias];
+  const providerEntry = config.providers[modelEntry.provider];
+  const apiKey = process.env[providerEntry.apiKeyEnv];
+  return { baseUrl: providerEntry.baseUrl, apiKey, model: modelEntry.name };
+}
+```
+
+`ResolvedLlmProvider = { baseUrl, apiKey, model }`。
+
+Extract 专用别名解析：
+
+```18:30:packages/workflow-agent-kit/src/extract.ts
+export function resolveExtractModelAlias(config: WorkflowConfig): ModelAlias {
+  return config.modelOverrides?.extract ?? (config.models.extract ? "extract" : config.models.default ? "default" : config.defaultModel);
+}
+```
+
+**尚无** `modelOverrides` 按 role/workflow 解析 agent 主模型的函数；builtin 首版可用 `config.defaultModel`，扩展时可加 `modelOverrides.agent` 或与 `agentOverrides` 对称的表。
+
+#### chatCompletionText
+
+```87:124:packages/workflow-agent-kit/src/extract.ts
+async function chatCompletionText(
+  provider: ResolvedLlmProvider,
+  messages: Array<{ role: "system" | "user"; content: string }>,
+): Promise<string>
+```
+
+| 能力 | 现状 |
+|------|------|
+| 协议 | OpenAI 兼容 `POST /chat/completions` |
+| Streaming | **无**（一次性 `response.text()`） |
+| Tool calling | **无**（无 `tools` / `tool_calls` 字段） |
+| 多模态 | **无**（仅 text `content`） |
+| Extract 专用 | `response_format: { type: "json_object" }` |
+
+builtin agent 的 run loop 需要**新写**带 `tools` 的 completion 客户端（可放在 `workflow-agent-builtin` 或扩展 `workflow-agent-kit` 的 `llm/` 模块），不能复用当前 `chatCompletionText` 而不改。
+
+---
+
+### Q6: Hermes Agent 参考实现
+
+`uwf-hermes` 是怎么实现 `run` 和 `continue` 的？
+
+**调研要点：**
+- prompt 怎么组装的（outputFormatInstruction + rolePrompt + task + history）
+- hermes CLI 的调用参数
+- session management（resume）
+- 输出怎么捕获
+
+**答案：**
+
+#### Prompt 组装
+
+```40:53:packages/workflow-agent-hermes/src/hermes.ts
+export function buildHermesPrompt(ctx: AgentContext): string {
+  const roleDef = ctx.workflow.roles[ctx.role];
+  const rolePrompt = roleDef !== undefined ? buildRolePrompt(roleDef) : "";
+  const parts: string[] = [];
+  if (ctx.outputFormatInstruction !== "") {
+    parts.push(ctx.outputFormatInstruction, "");
+  }
+  parts.push(rolePrompt, "", "## Task", ctx.start.prompt);
+  const historyBlock = buildHistorySummary(ctx.steps);
+  if (historyBlock !== "") {
+    parts.push("", historyBlock);
+  }
+  return parts.join("\n");
+}
+```
+
+`buildRolePrompt` 生成 `## Goal` / `## Capabilities` / `## Prepare`（含 `generateCliReference()`）/ `## Procedure` / `## Output`。
+
+`buildHistorySummary`：每步 `role`、`JSON.stringify(step.output)`、`agent`。
+
+Hermes 把**整段 prompt 作为单条 user 消息**传给 `hermes chat -q`（无独立 system channel）。
+
+#### Hermes CLI 参数
+
+首次：
+
+```88:97:packages/workflow-agent-hermes/src/hermes.ts
+spawnHermes(["chat", "-q", prompt, "--yolo", "--max-turns", "90", "--quiet"]);
+```
+
+续聊：
+
+```100:114:packages/workflow-agent-hermes/src/hermes.ts
+spawnHermes(["chat", "--resume", sessionId, "-q", message, "--yolo", "--max-turns", "90", "--quiet"]);
+```
+
+#### Session
+
+- stdout/stderr 中解析 `session_id: <id>`（`parseSessionIdFromStdout`）
+- 会话文件：`~/.hermes/sessions/session_<id>.json`
+- `loadHermesSession` → `storeHermesSessionDetail`：每 assistant/tool 消息写成 CAS turn 节点，汇总为 `detail`；**output 文本** = 最后一条非空 `assistant` 的 `content`
+
+#### 与 createAgent 的衔接
+
+```157:164:packages/workflow-agent-hermes/src/hermes.ts
+export function createHermesAgent(): () => Promise<void> {
+  return createAgent({ name: "hermes", run: runHermes, continue: continueHermes });
+}
+```
+
+`uwf-hermes` 入口：`createHermesAgent()` 即 main。
+
+Claude Code 包（`workflow-agent-claude-code`）结构相同：`buildClaudeCodePrompt` 同构，`claude -p` + `--resume` + JSON stdout 解析。
+
+---
+
+### Q7: Toolkit 需求分析
+
+要实现一个自给自足的 agent，最少需要哪些 tool？
+
+**调研要点：**
+- 现有 workflow example（solve-issue.yaml）里 role 都做什么任务
+- hermes agent 在 workflow 场景下常用哪些 tool
+- 哪些 tool 是 agent loop 必须的（如 file read/write、shell exec、web fetch）
+
+**答案：**
+
+#### solve-issue.yaml 角色能力
+
+| Role | capabilities | 隐含需求 |
+|------|----------------|----------|
+| planner | issue-analysis, planning | 读上下文/仓库、总结，通常不需写代码 |
+| developer | file-edit, shell, testing | **读文件、写文件、执行命令** |
+| reviewer | code-review, static-analysis | 读 diff/文件、静态分析（可读+可选 shell） |
+
+#### Hermes 侧
+
+Hermes 自带完整 agent runtime（`--yolo`、max-turns），tool 集由 Hermes 项目定义，workflow 不配置。从 session JSON 可见 `tool_calls` 被记入 detail，常见包括文件与 shell 类工具。
+
+#### Builtin 最小 toolkit 建议
+
+| 优先级 | Tool | 用途 |
+|--------|------|------|
+| P0 | `read_file` | 读仓库/配置/issue 上下文 |
+| P0 | `write_file` / `edit_file` | developer 改代码 |
+| P0 | `run_command` | 测试、构建、git（需 cwd + timeout + 输出截断） |
+| P1 | `list_dir` / `glob` | 导航代码库 |
+| P1 | `grep` | 搜索符号/引用 |
+| P2 | `fetch_url` | 查文档（planner 偶尔需要） |
+
+**不需要**在 builtin 里实现 moderator / workflow 路由工具——仍由 `uwf thread step` + JSONata 负责。
+
+#### Agent loop 必须能力
+
+1. 多轮 LLM 调用 + **OpenAI-style tool_calls** 解析与执行
+2. 将 tool 结果 append 回 messages
+3. 终止条件：模型不再请求 tool，或达到 `maxTurns`
+4. 最终响应须含合法 YAML frontmatter（满足 Q4），供 `createAgent` fast-path
+
+---
+
+## 方案草案
+
+（调研完成后基于以上答案撰写）
+
+### 架构设计
+
+```mermaid
+flowchart TB
+  subgraph cli ["cli-workflow"]
+    Step["uwf thread step"]
+    Spawn["spawnAgent(uwf-builtin, threadId, role)"]
+    Step --> Spawn
+  end
+
+  subgraph builtin_pkg ["@uncaged/workflow-agent-builtin"]
+    Main["createBuiltinAgent() = createAgent({...})"]
+    Prompt["buildBuiltinPrompt(ctx)"]
+    Loop["runBuiltinLoop(provider, messages, tools)"]
+    Tools["Toolkit: read/write/exec/..."]
+    Detail["storeBuiltinDetail(turns)"]
+    Main --> Prompt
+    Main --> Loop
+    Loop --> Tools
+    Loop --> Detail
+  end
+
+  subgraph kit ["workflow-agent-kit"]
+    Ctx["buildContextWithMeta"]
+    FM["tryFrontmatterFastPath"]
+    Persist["persistStep"]
+    Ctx --> Main
+    Main --> FM
+    FM --> Persist
+  end
+
+  subgraph cas ["CAS / config"]
+    Config["config.yaml models/providers"]
+    CAS["cas/ + threads.yaml"]
+  end
+
+  Spawn --> Main
+  Config --> Loop
+  CAS --> Ctx
+  Persist --> CAS
+  Spawn -->|"stdout: step hash"| Step
+```
+
+**新包**：`packages/workflow-agent-builtin`，bin `uwf-builtin`，仅依赖 `workflow-agent-kit`、`workflow-protocol`、`workflow-util`（可选 `@uncaged/json-cas` 写 detail schema）。
+
+**分层**：
+
+| 层 | 职责 |
+|----|------|
+| `createAgent`（kit） | argv、context、frontmatter extract、StepNode、stdout 协议 — **不变** |
+| `builtin/agent.ts` | `run` / `continue` 实现 |
+| `builtin/llm.ts` | OpenAI 兼容 chat + tools（可后续抽到 kit） |
+| `builtin/tools/*.ts` | 各 tool 的 JSON Schema + handler |
+| `builtin/prompt.ts` | 复用 Hermes 的 prompt 拼接逻辑（或抽到 kit 的 `buildAgentPrompt`） |
+| `builtin/detail.ts` | 类似 Hermes：每轮 assistant/tool 写入 CAS detail |
+
+**配置集成**：
+
+```yaml
+agents:
+  builtin:
+    command: "uwf-builtin"
+    args: []
+defaultAgent: "builtin"   # 或 agentOverrides 按 role 指定
+```
+
+模型：首版 `resolveModel(config, config.defaultModel)`；后续可增加 `modelOverrides.agent` 或 per-role 映射。
+
+---
+
+### Agent Run Loop
+
+伪代码（单次 `run(ctx)`）：
+
+```
+1. provider ← resolveModel(loadWorkflowConfig(), defaultModel)
+2. system ← buildBuiltinPrompt(ctx)   // outputFormatInstruction + buildRolePrompt + Task + History
+3. messages ← [{ role: "system", content: system }]
+4. sessionId ← newULID()              // 内存或临时目录，供 continue 使用
+5. turns ← []
+
+6. for turn in 1..MAX_TURNS:
+     response ← chatCompletionWithTools(provider, messages, TOOL_DEFINITIONS)
+     record assistant message + tool_calls in turns
+
+     if response has no tool_calls:
+       finalText ← response.content
+       break
+
+     for each tool_call:
+       result ← executeTool(tool_call, { cwd: process.cwd() })
+       messages.push tool result
+       record in turns
+
+7. if no finalText with valid frontmatter after loop:
+     optionally one-shot "finalize" message without tools
+
+8. detailHash ← storeBuiltinDetail(store, sessionId, turns, metadata)
+9. return { output: finalText, detailHash, sessionId }
+```
+
+**`continue(sessionId, message, store)`**：
+
+- 从内存/磁盘恢复 `messages` + `turns`
+- `messages.push({ role: "user", content: message })`（correction 或续聊）
+- 从步骤 6 继续，步数上限可单独设小一点（如 3）
+- 返回新的 `AgentRunResult`
+
+**与 frontmatter 的配合**：
+
+- system prompt 已含 `outputFormatInstruction`；最后一轮可强制 user：`Now output your final answer with YAML frontmatter only if you have not yet.`
+- 仍依赖 `createAgent` 的 fast-path + 最多 2 次 continue
+
+**安全**：
+
+- `run_command`：白名单或需 `UWF_BUILTIN_ALLOW_SHELL=1`，默认工作区限定在 `process.cwd()` 或 `start` 中将来扩展的 `workspace` 字段
+- 路径：禁止 `..` 逃逸出 workspace root
+
+---
+
+### Toolkit 设计
+
+统一注册表：
+
+```typescript
+type BuiltinTool = {
+  name: string;
+  description: string;
+  parameters: JSONSchema; // object type
+  execute: (args: unknown, ctx: ToolContext) => Promise<string>;
+};
+
+type ToolContext = {
+  cwd: string;
+  storageRoot: string;
+};
+```
+
+| Tool name | OpenAI function | 行为摘要 |
+|-----------|-----------------|----------|
+| `read_file` | `read_file` | `{ path }` → UTF-8 文本，大小上限 |
+| `write_file` | `write_file` | `{ path, content }` → 写盘，返回确认 |
+| `edit_file` | 可选 | search/replace 块，减少 token |
+| `run_command` | `run_command` | `{ command, cwd? }` → stdout/stderr 截断 |
+| `list_dir` | `list_dir` | `{ path }` → 条目列表 |
+| `grep` | `grep` | `{ pattern, path? }` → 匹配行 |
+
+**LLM 请求形状**（扩展 extract 客户端）：
+
+```json
+{
+  "model": "...",
+  "messages": [...],
+  "tools": [{ "type": "function", "function": { "name", "description", "parameters" } }],
+  "tool_choice": "auto"
+}
+```
+
+解析 `choices[0].message.tool_calls`，执行后以 `{ role: "tool", tool_call_id, content }` 回传。
+
+**不提供** streaming 首版；detail CAS 记录每轮 tool 名/参数/结果摘要供 `uwf thread step-details` 调试。
+
+---
+
+### 与现有架构的集成
+
+| 集成点 | 方式 |
+|--------|------|
+| CLI 协议 | 实现标准 agent CLI：`uwf-builtin <thread-id> <role>`，stdout 一行 step hash，exit 0/1 |
+| 工厂 | `export function createBuiltinAgent()` → `createAgent({ name: "builtin", run, continue })` |
+| Context / Prompt | 复用 `buildContextWithMeta`、`buildRolePrompt`、`buildOutputFormatInstruction`；prompt 布局对齐 `buildHermesPrompt` |
+| 结构化输出 | 优先 YAML frontmatter fast-path；可选后续在 `createAgent` 增加 `extract()` fallback 开关 |
+| 配置 | `config.yaml` 增加 `agents.builtin`；`uwf setup` 可选默认 agent |
+| 存储 | `resolveStorageRoot()` + `loadWorkflowConfig` + `getEnvPath`；与 Hermes 相同，**不**改 `threads.yaml` 写入方 |
+| 测试 | 单元测试：tool handlers、prompt 组装、mock LLM tool loop；集成测试：临时 storage root + fake provider |
+| 发布 | 新包 `@uncaged/workflow-agent-builtin`，bin `uwf-builtin`，加入 `scripts/publish-all.mjs` |
+
+**明确不做**：
+
+- 不替代 moderator / 不在 agent 内调用 `uwf thread step`
+- 不依赖 Hermes/OpenClaw/Claude Code 二进制
+- 首版不实现 streaming、不实现 MCP
+
+**建议实现顺序**：
+
+1. `llm.ts`：tool calling HTTP 客户端 + 单测
+2. P0 tools + `runBuiltinLoop`
+3. `createBuiltinAgent` + detail CAS
+4. `config` / docs / `examples` 可选 `agentOverrides` 演示
+5. （可选）`createAgent` 接入 `extract()` fallback

docs/wf-stateless-design.md

@@ -112,8 +112,8 @@ uwf-hermes <thread-id> <role>
 
 **约定：**
 - `uwf step` 负责 moderator 决策，将 role 传给 agent CLI
-- agent-kit 根据 thread + role 从 CAS 读 systemPrompt / outputSchema
-- agent-kit 组装完整 prompt（role systemPrompt + thread context + user prompt from StartNode）
+- agent-kit 根据 thread + role 从 CAS 读 goal / capabilities / procedure / output / meta
+- agent-kit 组装完整 prompt（role goal/capabilities/procedure/output + thread context + user prompt from StartNode）
 - agent 执行实际逻辑，agent-kit 负责 extract
 - agent 将 StepNode 写入 CAS（含 output、detail、agent、prev），但**不挪链头指针**
 - stdout 输出新 StepNode 的 CAS hash（纯文本，一行）
@@ -143,7 +143,7 @@ uwf-hermes <thread-id> <role>
 
 #### `Workflow`
 
-Roles 和 moderator 内联在 Workflow 中，只有 outputSchema 独立为 CAS 节点（方便 json-cas 校验）。
+Roles 和 moderator 内联在 Workflow 中，只有 meta 独立为 CAS 节点（方便 json-cas 校验）。
 
 ```yaml
 type: <workflow-schema-hash>
@@ -153,16 +153,25 @@ payload:
   roles:
     planner:
       description: "Creates implementation plan"
-      systemPrompt: "You are a planning agent..."
-      outputSchema: "5GWKR8TN1V3JA"    # cas_ref → JSON Schema 节点（json-cas 内置）
+      goal: "You are a planning agent..."
+      capabilities: [planning, issue-analysis]
+      procedure: "Analyze the issue and create a plan."
+      output: "Output the plan summary."
+      meta: "5GWKR8TN1V3JA"    # cas_ref → JSON Schema 节点（json-cas 内置）
     developer:
       description: "Implements code changes"
-      systemPrompt: "You are a developer agent..."
-      outputSchema: "8CNWT4KR6D1HV"    # cas_ref → JSON Schema 节点
+      goal: "You are a developer agent..."
+      capabilities: [file-edit, shell]
+      procedure: "Implement the plan."
+      output: "List all files changed."
+      meta: "8CNWT4KR6D1HV"    # cas_ref → JSON Schema 节点
     reviewer:
       description: "Reviews code changes"
-      systemPrompt: "You are a code reviewer..."
-      outputSchema: "1VPBG9SM5E7WK"    # cas_ref → JSON Schema 节点
+      goal: "You are a code reviewer..."
+      capabilities: [code-review]
+      procedure: "Review the implementation."
+      output: "Approve or reject with comments."
+      meta: "1VPBG9SM5E7WK"    # cas_ref → JSON Schema 节点
   conditions:
     needsClarification:
       description: "Planner requests clarification from user"
@@ -189,7 +198,7 @@ payload:
         condition: null
 ```
 
-- `roles` — 内联定义，每个 role 的 `outputSchema` 是独立的 cas_ref（指向 json-cas 内置 JSON Schema 节点）
+- `roles` — 内联定义，每个 role 的 `meta` 是独立的 cas_ref（指向 json-cas 内置 JSON Schema 节点）
 - `conditions` — `Record<Name, JSONata>`，命名条件，方便画图描述
 - `graph` — `Record<Role | "$START", Transition[]>`，每个 Transition = `{ role, condition }`
 - `condition` 引用 conditions 中的 key，`null` = fallback
@@ -234,14 +243,14 @@ payload:
   start: "4TNVW8KR2B3MA"          # cas_ref → StartNode（每个 step 都引用）
   prev: "2MXBG6PN4A8JR"           # cas_ref → 前一个 StepNode，第一步为 null
   role: "developer"
-  output: "9KRVW3TN5F1QA"         # cas_ref → 结构化输出节点（符合 role 的 outputSchema）
+  output: "9KRVW3TN5F1QA"         # cas_ref → 结构化输出节点（符合 role 的 meta schema）
   detail: "7BQST3VW9F2MA"         # cas_ref → 执行详情（content node / 子 workflow terminal StepNode / ...）
   agent: "uwf-cursor"              # 实际使用的 agent 命令（纯字符串）
 ```
 
 - `start` — 每个 StepNode 都直接引用 StartNode，方便随机访问
 - `prev` — 前一个 StepNode 的 cas_ref，第一步为 `null`（不指向 StartNode）
-- `output` — cas_ref，指向符合 role outputSchema 的 CAS 节点，可用 json-cas 校验
+- `output` — cas_ref，指向符合 role meta schema 的 CAS 节点，可用 json-cas 校验
 - `detail` — cas_ref，指向执行详情。可以是原始 agent 输出（content node），也可以是子 workflow thread 的 terminal StepNode（workflowAsAgent 场景）
 - `agent` — 纯字符串，不是 CAS 节点
 
@@ -340,12 +349,12 @@ OPENROUTER_API_KEY=sk-or-...
 
 ```
 packages/
-├── cli-uwf/              # @uncaged/cli-uwf — uwf CLI（thread/workflow 命令）
-├── uwf-moderator/        # @uncaged/uwf-moderator — JSONata moderator 引擎
-├── uwf-agent-kit/        # @uncaged/uwf-agent-kit — Agent CLI 框架（含 extractor）
-├── uwf-agent-hermes/     # @uncaged/uwf-agent-hermes — uwf-hermes CLI
-├── uwf-agent-cursor/     # @uncaged/uwf-agent-cursor — uwf-cursor CLI
-└── uwf-protocol/         # @uncaged/uwf-protocol — 共享类型定义
+├── cli-workflow/              # @uncaged/cli-workflow — uwf CLI（thread/workflow 命令）
+├── workflow-moderator/        # @uncaged/workflow-moderator — JSONata moderator 引擎
+├── workflow-agent-kit/        # @uncaged/workflow-agent-kit — Agent CLI 框架（含 extractor）
+├── workflow-agent-hermes/     # @uncaged/workflow-agent-hermes — uwf-hermes CLI
+├── workflow-agent-cursor/ # @uncaged/workflow-agent-cursor — uwf-cursor CLI
+└── workflow-protocol/         # @uncaged/workflow-protocol — 共享类型定义
 ```
 
 **外部依赖：**
@@ -372,7 +381,7 @@ type ThreadId = string;
 /** 一个 step 的核心数据，被 StepNode payload 和 JSONata 上下文共享 */
 type StepRecord = {
   role: string;
-  output: CasRef;                    // cas_ref → 结构化输出节点（符合 role outputSchema）
+  output: CasRef;                    // cas_ref → 结构化输出节点（符合 role meta schema）
   detail: CasRef;                    // cas_ref → 执行详情（content node / 子 workflow terminal StepNode）
   agent: string;                     // 实际使用的 agent 命令（纯字符串）
 };
@@ -383,8 +392,11 @@ type StepRecord = {
 ```typescript
 type RoleDefinition = {
   description: string;
-  systemPrompt: string;
-  outputSchema: CasRef;              // cas_ref → json-cas 内置 JSON Schema 节点
+  goal: string;
+  capabilities: string[];
+  procedure: string;
+  output: string;
+  meta: CasRef;                      // cas_ref → json-cas 内置 JSON Schema 节点
 };
 
 type Transition = {

examples/analyze-topic.yaml

@@ -0,0 +1,43 @@
+name: "analyze-topic"
+description: "Single-role topic analysis using four-phase role description"
+roles:
+  analyst:
+    description: "Analyzes a given topic and produces a structured summary"
+    goal: |
+      You are a research analyst with expertise in breaking down complex topics
+      into clear, structured summaries. You think critically and cite key points.
+    capabilities:
+      - research
+      - critical-thinking
+      - structured-writing
+    procedure: |
+      Analyze the topic by:
+      1. Identifying the main thesis or question
+      2. Listing 3-5 key points with brief explanations
+      3. Noting any counterarguments or caveats
+      Keep your analysis concise (under 500 words).
+    output: |
+      Provide your analysis as markdown under the frontmatter.
+      The frontmatter must include your structured findings.
+    frontmatter:
+      type: object
+      properties:
+        thesis:
+          type: string
+        keyPoints:
+          type: array
+          items:
+            type: string
+        caveats:
+          type: string
+      required: [thesis, keyPoints]
+conditions: {}
+graph:
+  $START:
+    - role: "analyst"
+      condition: null
+      prompt: "Analyze the topic in the task and produce a structured summary with key points."
+  analyst:
+    - role: "$END"
+      condition: null
+      prompt: "Analysis complete. Finish the workflow."

examples/solve-issue.yaml

@@ -0,0 +1,80 @@
+name: "solve-issue"
+description: "End-to-end issue resolution"
+roles:
+  planner:
+    description: "Creates implementation plan"
+    goal: "You are a planning agent. You analyze issues and create step-by-step plans."
+    capabilities:
+      - issue-analysis
+      - planning
+    procedure: "Analyze the issue and create a detailed, actionable implementation plan."
+    output: "Output the plan summary and list of concrete steps."
+    frontmatter:
+      type: object
+      properties:
+        plan:
+          type: string
+        steps:
+          type: array
+          items:
+            type: string
+      required: [plan, steps]
+  developer:
+    description: "Implements code changes"
+    goal: "You are a developer agent. You implement code changes according to plans."
+    capabilities:
+      - file-edit
+      - shell
+      - testing
+    procedure: "Implement the plan. Write code, tests, and ensure existing tests pass."
+    output: "List all files changed and provide a summary of the implementation."
+    frontmatter:
+      type: object
+      properties:
+        filesChanged:
+          type: array
+          items:
+            type: string
+        summary:
+          type: string
+      required: [filesChanged, summary]
+  reviewer:
+    description: "Reviews code changes"
+    goal: "You are a code reviewer. You review implementations for correctness and quality."
+    capabilities:
+      - code-review
+      - static-analysis
+    procedure: "Review the implementation against the plan. Check for bugs, edge cases, and style."
+    output: "Approve or reject with detailed comments explaining your decision."
+    frontmatter:
+      type: object
+      properties:
+        approved:
+          type: boolean
+        comments:
+          type: string
+      required: [approved, comments]
+conditions:
+  notApproved:
+    description: "Reviewer rejected the implementation"
+    expression: "$last('reviewer').approved = false"
+graph:
+  $START:
+    - role: "planner"
+      condition: null
+      prompt: "Analyze the issue described in the task and produce a detailed implementation plan."
+  planner:
+    - role: "developer"
+      condition: null
+      prompt: "Implement the plan from the planner. Write code, tests, and ensure existing tests pass."
+  developer:
+    - role: "reviewer"
+      condition: null
+      prompt: "Review the developer's implementation against the plan for correctness and quality."
+  reviewer:
+    - role: "developer"
+      condition: "notApproved"
+      prompt: "The reviewer rejected your implementation. Read their feedback and fix the issues."
+    - role: "$END"
+      condition: null
+      prompt: "The review passed. Complete the workflow."

legacy-packages/cli-workflow/package.json

@@ -0,0 +1,30 @@
+{
+  "name": "@uncaged/cli-workflow",
+  "version": "0.5.0-alpha.4",
+  "files": [
+    "src",
+    "dist",
+    "package.json"
+  ],
+  "type": "module",
+  "bin": {
+    "uncaged-workflow": "src/cli.ts"
+  },
+  "dependencies": {
+    "@uncaged/workflow-gateway": "workspace:^",
+    "@uncaged/workflow-protocol": "workspace:^",
+    "@uncaged/workflow-util": "workspace:^",
+    "@uncaged/workflow-cas": "workspace:^",
+    "@uncaged/workflow-execute": "workspace:^",
+    "@uncaged/workflow-register": "workspace:^",
+    "@uncaged/workflow-runtime": "workspace:^",
+    "hono": "^4.12.18",
+    "yaml": "^2.8.4"
+  },
+  "scripts": {
+    "test": "bun test"
+  },
+  "publishConfig": {
+    "access": "public"
+  }
+}

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

@@ -0,0 +1,9 @@
+#!/usr/bin/env bun
+
+import { runCli } from "./cli-dispatch.js";
+import { resolveWorkflowStorageRoot } from "./storage-env.js";
+
+const argv = process.argv.slice(2);
+const storageRoot = resolveWorkflowStorageRoot();
+const code = await runCli(storageRoot, argv);
+process.exit(code);