feat(engine): add supervisor scene — opt-in LLM-based thread stop (Phase 3)

Supervisor replaces maxRounds as primary stop mechanism. Every N rounds
(configurable via supervisorInterval, default 3), the engine calls a
cheap LLM to evaluate thread progress and decide continue/stop.

- New engine/supervisor.ts: runSupervisor + parseSupervisorDecisionText
- Supervisor is opt-in: no models.supervisor configured = always continue
- WorkflowConfig gains supervisorInterval (default 3, 0 to disable)
- Engine calls supervisor after each supervisorInterval rounds
- 256 tests pass, 14 new tests for supervisor logic

Refs #110

examples/hello-world.ts

@@ -1,4 +1,4 @@
-import { createExtract, createWorkflow, END, type RoleDefinition } from "@uncaged/workflow";
+import { createWorkflow, END, type RoleDefinition } from "@uncaged/workflow";
 import * as z from "zod/v4";
 
 type Roles = {
@@ -32,12 +32,6 @@ const greeter: RoleDefinition<Roles["greeter"]> = {
   extractMode: "single",
 };
 
-const extract = createExtract({
-  baseUrl: "http://127.0.0.1:9",
-  apiKey: "",
-  model: "",
-});
-
 export const run = createWorkflow<Roles>(
   {
     roles: { greeter },
@@ -48,6 +42,4 @@ export const run = createWorkflow<Roles>(
   {
     agent: async (ctx) => `Hello, ${ctx.start.content}`,
   },
-  extract,
-  null,
 );

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

@@ -7,6 +7,7 @@ import { cmdFork, cmdRun } from "../src/commands/thread/index.js";
 import { cmdAdd } from "../src/commands/workflow/index.js";
 import { pathExists } from "../src/fs-utils.js";
 import { addCliArgs } from "./bundle-fixture.js";
+import { ensureTestWorkflowRegistryConfig } from "./workflow-registry-fixture.js";
 
 /** Three-role workflow that respects `input.steps` for fork/resume. */
 const threeRoleBundleSource = `import { putContentMerkleNode } from "@uncaged/workflow";
@@ -77,6 +78,7 @@ describe("cli fork", () => {
     prevEnv = process.env.UNCAGED_WORKFLOW_STORAGE_ROOT;
     storageRoot = await mkdtemp(join(tmpdir(), "uncaged-wf-fork-"));
     process.env.UNCAGED_WORKFLOW_STORAGE_ROOT = storageRoot;
+    await ensureTestWorkflowRegistryConfig(storageRoot);
   });
 
   afterEach(async () => {

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

@@ -19,6 +19,7 @@ import {
 import { cmdAdd } from "../src/commands/workflow/index.js";
 import { pathExists, readTextFileIfExists } from "../src/fs-utils.js";
 import { addCliArgs } from "./bundle-fixture.js";
+import { ensureTestWorkflowRegistryConfig } from "./workflow-registry-fixture.js";
 
 const wfPutImport = `import { putContentMerkleNode } from "@uncaged/workflow";
 `;
@@ -142,6 +143,7 @@ describe("cli thread commands", () => {
     prevEnv = process.env.UNCAGED_WORKFLOW_STORAGE_ROOT;
     storageRoot = await mkdtemp(join(tmpdir(), "uncaged-wf-thread-"));
     process.env.UNCAGED_WORKFLOW_STORAGE_ROOT = storageRoot;
+    await ensureTestWorkflowRegistryConfig(storageRoot);
   });
 
   afterEach(async () => {

packages/cli-workflow/__tests__/workflow-registry-fixture.ts

@@ -0,0 +1,18 @@
+import { writeFile } from "node:fs/promises";
+import { join } from "node:path";
+
+/** Minimal valid global config so {@link executeThread} can resolve the extract scene (CLI integration tests). */
+export const TEST_WORKFLOW_REGISTRY_YAML = `config:
+  maxDepth: 3
+  providers:
+    stub:
+      baseUrl: http://127.0.0.1:9
+      apiKey: test
+  models:
+    default: stub/m
+workflows: {}
+`;
+
+export async function ensureTestWorkflowRegistryConfig(storageRoot: string): Promise<void> {
+  await writeFile(join(storageRoot, "workflow.yaml"), TEST_WORKFLOW_REGISTRY_YAML, "utf8");
+}

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

@@ -13,19 +13,7 @@ import {
   templateTsconfigJson,
 } from "./templates.js";
 import type { CmdInitTemplateSuccess } from "./types.js";
-
-function validateWorkspaceSegment(name: string): Result<void, string> {
-  if (name.length === 0) {
-    return err("workspace name must not be empty");
-  }
-  if (name === "." || name === "..") {
-    return err("invalid workspace name");
-  }
-  if (name.includes("/") || name.includes("\\")) {
-    return err("workspace name must not contain path separators");
-  }
-  return ok(undefined);
-}
+import { validateWorkspaceSegment } from "./validate.js";
 
 function hasTemplatesWorkspaceGlob(workspaces: unknown): boolean {
   return Array.isArray(workspaces) && workspaces.includes("templates/*");

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

@@ -0,0 +1,15 @@
+import { err, ok, type Result } from "@uncaged/workflow";
+
+/** Validates a single path segment for workspace / template names (no separators, not `.` / `..`). */
+export function validateWorkspaceSegment(name: string): Result<void, string> {
+  if (name.length === 0) {
+    return err("workspace name must not be empty");
+  }
+  if (name === "." || name === "..") {
+    return err("invalid workspace name");
+  }
+  if (name.includes("/") || name.includes("\\")) {
+    return err("workspace name must not contain path separators");
+  }
+  return ok(undefined);
+}

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

@@ -5,19 +5,7 @@ import { err, ok, type Result } from "@uncaged/workflow";
 
 import { pathExists } from "../../fs-utils.js";
 import type { CmdInitWorkspaceSuccess } from "./types.js";
-
-function validateWorkspaceSegment(name: string): Result<void, string> {
-  if (name.length === 0) {
-    return err("workspace name must not be empty");
-  }
-  if (name === "." || name === "..") {
-    return err("invalid workspace name");
-  }
-  if (name.includes("/") || name.includes("\\")) {
-    return err("workspace name must not contain path separators");
-  }
-  return ok(undefined);
-}
+import { validateWorkspaceSegment } from "./validate.js";
 
 function rootPackageJson(workspaceName: string): string {
   return `${JSON.stringify(
@@ -119,7 +107,7 @@ Init 生成的骨架：\`templates/\` 下放可复用定义，\`workflows/\` 下
 2. **编写 RoleDefinition**：为每个角色写 Zod \`schema\`，补齐 \`systemPrompt\` / \`extractPrompt\` / \`description\`。
 3. **编写 Moderator**：根据 \`ctx.steps\` 与业务状态返回下一个角色名或 \`END\`。
 4. **组装 WorkflowDefinition**：在模板 \`index\` 中导出 definition（以及必要的角色 / moderator 导出）。
-5. **实例化**：在 workflow 包中使用 \`createWorkflow(def, binding, extract)\`（或项目约定的封装）绑定 **AgentFn** / **ExtractFn**。
+5. **实例化**：在 workflow 包中使用 \`createWorkflow(def, binding)\`（或项目约定的封装）绑定 **AgentFn**；**ExtractFn** 由引擎从 **workflow.yaml** 注入 \`WorkflowFnOptions\`。
 6. **构建**：打包为单个 **.esm.js** bundle，使用 **uncaged-workflow add** 注册。
 
 ## 4. 编码规范

packages/workflow-template-develop/src/index.ts

@@ -1,8 +1,6 @@
 import {
   type AgentBinding,
   createWorkflow,
-  type ExtractFn,
-  type LlmProvider,
   type WorkflowDefinition,
   type WorkflowFn,
 } from "@uncaged/workflow";
@@ -43,10 +41,6 @@ export const developWorkflowDefinition: WorkflowDefinition<DevelopMeta> = {
   moderator: developModerator,
 };
 
-export function createDevelopRun(
-  binding: AgentBinding,
-  extract: ExtractFn,
-  llmProvider: LlmProvider | null,
-): WorkflowFn {
-  return createWorkflow(developWorkflowDefinition, binding, extract, llmProvider);
+export function createDevelopRun(binding: AgentBinding): WorkflowFn {
+  return createWorkflow(developWorkflowDefinition, binding);
 }

packages/workflow-template-solve-issue/__tests__/solve-issue-template.test.ts

@@ -250,17 +250,20 @@ describe("createSolveIssueRun", () => {
     const cas = createCasStore(casDir);
 
     // Override developer so the test does not spin up a child workflow.
-    const run = createSolveIssueRun(
-      {
-        agent: async () => "",
-        overrides: { developer: async () => "stub-root-hash" },
-      },
-      stubExtract,
-      stubLlmProvider,
-    );
+    const run = createSolveIssueRun({
+      agent: async () => "",
+      overrides: { developer: async () => "stub-root-hash" },
+    });
     const gen = run(
       { prompt: "task", steps: [] },
-      { threadId: "01TEST000000000000000000TR", maxRounds: 20, depth: 0, cas },
+      {
+        threadId: "01TEST000000000000000000TR",
+        maxRounds: 20,
+        depth: 0,
+        cas,
+        extract: stubExtract,
+        llmProvider: stubLlmProvider,
+      },
     );
     const first = await gen.next();
     expect(first.done).toBe(false);
@@ -294,33 +297,36 @@ describe("createSolveIssueRun", () => {
     const cas = createCasStore(casDir);
 
     const calls: string[] = [];
-    const run = createSolveIssueRun(
-      {
-        agent: async () => {
-          calls.push("default");
+    const run = createSolveIssueRun({
+      agent: async () => {
+        calls.push("default");
+        return "";
+      },
+      overrides: {
+        preparer: async () => {
+          calls.push("preparer");
           return "";
         },
-        overrides: {
-          preparer: async () => {
-            calls.push("preparer");
-            return "";
-          },
-          developer: async () => {
-            calls.push("developer");
-            return "stub-root-hash";
-          },
-          submitter: async () => {
-            calls.push("submitter");
-            return "";
-          },
+        developer: async () => {
+          calls.push("developer");
+          return "stub-root-hash";
+        },
+        submitter: async () => {
+          calls.push("submitter");
+          return "";
         },
       },
-      stubExtract,
-      stubLlmProvider,
-    );
+    });
     const gen = run(
       { prompt: "task", steps: [] },
-      { threadId: "01TEST000000000000000000TR", maxRounds: 20, depth: 0, cas },
+      {
+        threadId: "01TEST000000000000000000TR",
+        maxRounds: 20,
+        depth: 0,
+        cas,
+        extract: stubExtract,
+        llmProvider: stubLlmProvider,
+      },
     );
     await gen.next();
     expect(calls).toEqual(["preparer"]);
@@ -353,22 +359,25 @@ describe("createSolveIssueRun", () => {
     const cas = createCasStore(casDir);
 
     let developerInvocations = 0;
-    const run = createSolveIssueRun(
-      {
-        agent: async () => "",
-        overrides: {
-          developer: async () => {
-            developerInvocations += 1;
-            return "stub-root-hash";
-          },
+    const run = createSolveIssueRun({
+      agent: async () => "",
+      overrides: {
+        developer: async () => {
+          developerInvocations += 1;
+          return "stub-root-hash";
         },
       },
-      stubExtract,
-      stubLlmProvider,
-    );
+    });
     const gen = run(
       { prompt: "task", steps: [] },
-      { threadId: "01TEST000000000000000000TR", maxRounds: 20, depth: 0, cas },
+      {
+        threadId: "01TEST000000000000000000TR",
+        maxRounds: 20,
+        depth: 0,
+        cas,
+        extract: stubExtract,
+        llmProvider: stubLlmProvider,
+      },
     );
     // preparer
     await gen.next();

packages/workflow-template-solve-issue/src/index.ts

@@ -1,8 +1,6 @@
 import {
   type AgentBinding,
   createWorkflow,
-  type ExtractFn,
-  type LlmProvider,
   type WorkflowDefinition,
   type WorkflowFn,
   workflowAsAgent,
@@ -46,11 +44,7 @@ export const solveIssueWorkflowDefinition: WorkflowDefinition<SolveIssueMeta> =
  * {@link workflowAsAgent}; if the caller supplies their own `developer` override in
  * `binding.overrides`, it takes precedence so tests and custom hosts can stub it.
  */
-export function createSolveIssueRun(
-  binding: AgentBinding,
-  extract: ExtractFn,
-  llmProvider: LlmProvider | null,
-): WorkflowFn {
+export function createSolveIssueRun(binding: AgentBinding): WorkflowFn {
   const developerOverride = binding.overrides?.developer ?? workflowAsAgent("develop");
   const mergedBinding: AgentBinding = {
     agent: binding.agent,
@@ -59,5 +53,5 @@ export function createSolveIssueRun(
       developer: developerOverride,
     },
   };
-  return createWorkflow(solveIssueWorkflowDefinition, mergedBinding, extract, llmProvider);
+  return createWorkflow(solveIssueWorkflowDefinition, mergedBinding);
 }

packages/workflow/__tests__/engine.test.ts

@@ -1,5 +1,5 @@
 import { afterEach, describe, expect, test } from "bun:test";
-import { mkdir, mkdtemp, readFile, rm } from "node:fs/promises";
+import { mkdir, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import * as z from "zod/v4";
@@ -13,8 +13,7 @@ import {
 } from "../src/cas/merkle.js";
 import { createWorkflow } from "../src/engine/create-workflow.js";
 import { executeThread } from "../src/engine/engine.js";
-import { createExtract } from "../src/extract/extract-fn.js";
-import { END, type LlmProvider } from "../src/types.js";
+import { END } from "../src/types.js";
 import { createLogger } from "../src/util/logger.js";
 
 const plannerMetaSchema = z.object({
@@ -82,11 +81,112 @@ function installMockChatCompletions(sequence: ReadonlyArray<Record<string, unkno
   };
 }
 
-const demoExtract = createExtract({
-  baseUrl: "http://127.0.0.1:9",
-  apiKey: "test",
-  model: "test",
-});
+const EXTRACT_REGISTRY_YAML = `config:
+  maxDepth: 3
+  providers:
+    stub:
+      baseUrl: http://127.0.0.1:9
+      apiKey: test
+  models:
+    default: stub/model
+workflows: {}
+`;
+
+async function writeExtractRegistryConfig(storageRoot: string): Promise<void> {
+  await writeFile(join(storageRoot, "workflow.yaml"), EXTRACT_REGISTRY_YAML, "utf8");
+}
+
+const SUPERVISOR_INTERVAL_REGISTRY_YAML = `config:
+  maxDepth: 3
+  supervisorInterval: 2
+  providers:
+    stub:
+      baseUrl: http://127.0.0.1:9
+      apiKey: test
+  models:
+    extract: stub/model
+    supervisor: stub/supervisor-cheap
+workflows: {}
+`;
+
+const SUPERVISOR_LONG_INTERVAL_REGISTRY_YAML = `config:
+  maxDepth: 3
+  supervisorInterval: 10
+  providers:
+    stub:
+      baseUrl: http://127.0.0.1:9
+      apiKey: test
+  models:
+    extract: stub/model
+    supervisor: stub/supervisor-cheap
+workflows: {}
+`;
+
+async function writeRegistryYaml(storageRoot: string, yaml: string): Promise<void> {
+  await writeFile(join(storageRoot, "workflow.yaml"), yaml, "utf8");
+}
+
+/** Extract rounds use tool_calls; supervisor uses plain `content` (no tools). */
+function installMockExtractThenSupervisor(params: {
+  extractArgs: ReadonlyArray<Record<string, unknown>>;
+  supervisorContent: string;
+  onSupervisorCall?: () => void;
+}): () => void {
+  const origFetch = globalThis.fetch;
+  let extractI = 0;
+  const mockFetch = async (
+    _input: Parameters<typeof fetch>[0],
+    init?: RequestInit,
+  ): Promise<Response> => {
+    const body = init?.body ? (JSON.parse(String(init.body)) as Record<string, unknown>) : {};
+    const tools = body.tools;
+    const hasTools = Array.isArray(tools) && tools.length > 0;
+    if (hasTools) {
+      const args =
+        params.extractArgs[extractI] ?? params.extractArgs[params.extractArgs.length - 1];
+      if (args === undefined) {
+        throw new Error("installMockExtractThenSupervisor: empty extractArgs");
+      }
+      extractI += 1;
+      const firstTool = tools[0] as Record<string, unknown>;
+      const fn = firstTool.function as Record<string, unknown> | undefined;
+      const toolName = typeof fn?.name === "string" ? fn.name : "extract";
+      return new Response(
+        JSON.stringify({
+          choices: [
+            {
+              message: {
+                tool_calls: [
+                  {
+                    type: "function",
+                    function: {
+                      name: toolName,
+                      arguments: JSON.stringify(args),
+                    },
+                  },
+                ],
+              },
+            },
+          ],
+        }),
+        { status: 200, headers: { "Content-Type": "application/json" } },
+      );
+    }
+    params.onSupervisorCall?.();
+    return new Response(
+      JSON.stringify({
+        choices: [{ message: { content: params.supervisorContent } }],
+      }),
+      { status: 200, headers: { "Content-Type": "application/json" } },
+    );
+  };
+  globalThis.fetch = Object.assign(mockFetch, {
+    preconnect: origFetch.preconnect.bind(origFetch),
+  }) as typeof fetch;
+  return () => {
+    globalThis.fetch = origFetch;
+  };
+}
 
 const demoWorkflow = createWorkflow<DemoMeta>(
   {
@@ -125,8 +225,6 @@ const demoWorkflow = createWorkflow<DemoMeta>(
       coder: async () => "code-body",
     },
   },
-  demoExtract,
-  null,
 );
 
 describe("executeThread", () => {
@@ -150,6 +248,7 @@ describe("executeThread", () => {
       const dataPath = join(root, "logs", hash, `${threadId}.data.jsonl`);
       const infoPath = join(root, "logs", hash, `${threadId}.info.jsonl`);
       await mkdir(join(root, "logs", hash), { recursive: true });
+      await writeExtractRegistryConfig(root);
       const cas = createCasStore(join(root, "cas"));
 
       const logger = createLogger({ sink: { kind: "file", path: infoPath } });
@@ -166,6 +265,7 @@ describe("executeThread", () => {
           awaitAfterEachYield: async () => {},
           forkSourceThreadId: null,
           prefilledDiskSteps: null,
+          storageRoot: root,
         },
         { threadId, hash, dataJsonlPath: dataPath, infoJsonlPath: infoPath, cas },
         logger,
@@ -258,6 +358,7 @@ describe("executeThread", () => {
       const dataPath = join(root, "logs", hash, `${threadId}.data.jsonl`);
       const infoPath = join(root, "logs", hash, `${threadId}.info.jsonl`);
       await mkdir(join(root, "logs", hash), { recursive: true });
+      await writeExtractRegistryConfig(root);
       const cas = createCasStore(join(root, "cas"));
       const plannerHash = await cas.put(serializeMerkleNode(createContentMerkleNode("plan-body")));
 
@@ -295,6 +396,7 @@ describe("executeThread", () => {
               timestamp: histTs,
             },
           ],
+          storageRoot: root,
         },
         { threadId, hash, dataJsonlPath: dataPath, infoJsonlPath: infoPath, cas },
         logger,
@@ -354,6 +456,7 @@ describe("executeThread", () => {
           awaitAfterEachYield: async () => {},
           forkSourceThreadId: null,
           prefilledDiskSteps: null,
+          storageRoot: root,
         },
         { threadId, hash, dataJsonlPath: dataPath, infoJsonlPath: infoPath, cas },
         logger,
@@ -391,6 +494,7 @@ describe("executeThread", () => {
       const dataPath = join(root, "logs", hash, `${threadId}.data.jsonl`);
       const infoPath = join(root, "logs", hash, `${threadId}.info.jsonl`);
       await mkdir(join(root, "logs", hash), { recursive: true });
+      await writeExtractRegistryConfig(root);
       const cas = createCasStore(join(root, "cas"));
 
       const logger = createLogger({ sink: { kind: "file", path: infoPath } });
@@ -407,6 +511,7 @@ describe("executeThread", () => {
           awaitAfterEachYield: async () => {},
           forkSourceThreadId: null,
           prefilledDiskSteps: null,
+          storageRoot: root,
         },
         { threadId, hash, dataJsonlPath: dataPath, infoJsonlPath: infoPath, cas },
         logger,
@@ -549,9 +654,6 @@ describe("executeThread", () => {
         { preconnect: origFetch.preconnect.bind(origFetch) },
       ) as typeof fetch;
 
-      const llm: LlmProvider = { baseUrl: "http://127.0.0.1:9", apiKey: "test", model: "test" };
-      const extractFn = createExtract(llm);
-
       const dagWorkflow = createWorkflow<DagDemoMeta>(
         {
           roles: {
@@ -568,8 +670,6 @@ describe("executeThread", () => {
           moderator: (ctx) => (ctx.steps.length === 0 ? "walker" : END),
         },
         { agent: async () => dagRootHash },
-        extractFn,
-        llm,
       );
 
       const threadId = "01KQXKW18CT8G75T53R8F4G7YG";
@@ -577,6 +677,7 @@ describe("executeThread", () => {
       const dataPath = join(root, "logs", hash, `${threadId}.data.jsonl`);
       const infoPath = join(root, "logs", hash, `${threadId}.info.jsonl`);
       await mkdir(join(root, "logs", hash), { recursive: true });
+      await writeExtractRegistryConfig(root);
 
       const logger = createLogger({ sink: { kind: "file", path: infoPath } });
       const ac = new AbortController();
@@ -592,6 +693,7 @@ describe("executeThread", () => {
           awaitAfterEachYield: async () => {},
           forkSourceThreadId: null,
           prefilledDiskSteps: null,
+          storageRoot: root,
         },
         { threadId, hash, dataJsonlPath: dataPath, infoJsonlPath: infoPath, cas },
         logger,
@@ -613,4 +715,102 @@ describe("executeThread", () => {
       await rm(root, { recursive: true, force: true });
     }
   });
+
+  test("supervisor stops thread when interval elapses and model returns stop", async () => {
+    restoreFetch = installMockExtractThenSupervisor({
+      extractArgs: [{ plan: "do-it", files: ["a.ts"] }, { diff: "+ok" }],
+      supervisorContent: "stop",
+    });
+
+    const root = await mkdtemp(join(tmpdir(), "wf-engine-sup-stop-"));
+    try {
+      const threadId = "01KQXKW18CT8G75T53R8F4G7YG";
+      const hash = "C9NMV6V2TQT81";
+      const dataPath = join(root, "logs", hash, `${threadId}.data.jsonl`);
+      const infoPath = join(root, "logs", hash, `${threadId}.info.jsonl`);
+      await mkdir(join(root, "logs", hash), { recursive: true });
+      await writeRegistryYaml(root, SUPERVISOR_INTERVAL_REGISTRY_YAML);
+      const cas = createCasStore(join(root, "cas"));
+
+      const logger = createLogger({ sink: { kind: "file", path: infoPath } });
+      const ac = new AbortController();
+
+      const result = await executeThread(
+        demoWorkflow,
+        "demo-flow",
+        { prompt: "supervisor-stop-case", steps: [] },
+        {
+          maxRounds: 20,
+          depth: 0,
+          signal: ac.signal,
+          awaitAfterEachYield: async () => {},
+          forkSourceThreadId: null,
+          prefilledDiskSteps: null,
+          storageRoot: root,
+        },
+        { threadId, hash, dataJsonlPath: dataPath, infoJsonlPath: infoPath, cas },
+        logger,
+      );
+
+      expect(result.returnCode).toBe(0);
+      expect(result.summary).toBe("completed: supervisor stopped thread");
+
+      const dataText = await readFile(dataPath, "utf8");
+      const lines = dataText
+        .trim()
+        .split("\n")
+        .filter((l) => l !== "");
+      expect(lines.length).toBe(3);
+    } finally {
+      await rm(root, { recursive: true, force: true });
+    }
+  });
+
+  test("supervisor is not invoked before supervisorInterval rounds", async () => {
+    let supervisorCalls = 0;
+    restoreFetch = installMockExtractThenSupervisor({
+      extractArgs: [{ plan: "do-it", files: ["a.ts"] }, { diff: "+ok" }],
+      supervisorContent: "stop",
+      onSupervisorCall: () => {
+        supervisorCalls += 1;
+      },
+    });
+
+    const root = await mkdtemp(join(tmpdir(), "wf-engine-sup-skip-"));
+    try {
+      const threadId = "01KQXKW18CT8G75T53R8F4G7YG";
+      const hash = "C9NMV6V2TQT81";
+      const dataPath = join(root, "logs", hash, `${threadId}.data.jsonl`);
+      const infoPath = join(root, "logs", hash, `${threadId}.info.jsonl`);
+      await mkdir(join(root, "logs", hash), { recursive: true });
+      await writeRegistryYaml(root, SUPERVISOR_LONG_INTERVAL_REGISTRY_YAML);
+      const cas = createCasStore(join(root, "cas"));
+
+      const logger = createLogger({ sink: { kind: "file", path: infoPath } });
+      const ac = new AbortController();
+
+      const result = await executeThread(
+        demoWorkflow,
+        "demo-flow",
+        { prompt: "no-supervisor-yet", steps: [] },
+        {
+          maxRounds: 20,
+          depth: 0,
+          signal: ac.signal,
+          awaitAfterEachYield: async () => {},
+          forkSourceThreadId: null,
+          prefilledDiskSteps: null,
+          storageRoot: root,
+        },
+        { threadId, hash, dataJsonlPath: dataPath, infoJsonlPath: infoPath, cas },
+        logger,
+      );
+
+      expect(supervisorCalls).toBe(0);
+      expect(result.returnCode).toBe(0);
+      expect(result.summary).toBe("completed: moderator returned END");
+    } finally {
+      await rm(root, { recursive: true, force: true });
+    }
+  });
 });

packages/workflow/__tests__/extract-provider.test.ts

@@ -1,87 +0,0 @@
-import { describe, expect, test } from "bun:test";
-import { mkdir, mkdtemp, rm, writeFile } from "node:fs/promises";
-import { tmpdir } from "node:os";
-import { join } from "node:path";
-
-import { getExtractProvider } from "../src/extract-provider.js";
-
-describe("getExtractProvider", () => {
-  test("returns provider when config.extract is present", async () => {
-    const root = await mkdtemp(join(tmpdir(), "wf-ext-prov-ok-"));
-    try {
-      await mkdir(root, { recursive: true });
-      await writeFile(
-        join(root, "workflow.yaml"),
-        `config:
-  maxDepth: 3
-  extract:
-    baseUrl: https://dashscope.aliyuncs.com/compatible-mode/v1
-    model: qwen-plus
-    apiKey: literal-key
-workflows: {}
-`,
-        "utf8",
-      );
-      const r = await getExtractProvider(root);
-      expect(r.ok).toBe(true);
-      if (!r.ok) {
-        return;
-      }
-      expect(r.value.baseUrl).toBe("https://dashscope.aliyuncs.com/compatible-mode/v1");
-      expect(r.value.model).toBe("qwen-plus");
-      expect(r.value.apiKey).toBe("literal-key");
-    } finally {
-      await rm(root, { recursive: true, force: true });
-    }
-  });
-
-  test("errs when registry has no config section", async () => {
-    const root = await mkdtemp(join(tmpdir(), "wf-ext-prov-missing-"));
-    try {
-      await mkdir(root, { recursive: true });
-      await writeFile(join(root, "workflow.yaml"), "workflows: {}\n", "utf8");
-      const r = await getExtractProvider(root);
-      expect(r.ok).toBe(false);
-      if (r.ok) {
-        return;
-      }
-      expect(r.error).toContain("no global config");
-    } finally {
-      await rm(root, { recursive: true, force: true });
-    }
-  });
-
-  test("resolves apiKey from env at registry read time", async () => {
-    const root = await mkdtemp(join(tmpdir(), "wf-ext-prov-env-"));
-    const prev = process.env.WF_GET_EXTRACT_PROVIDER_KEY;
-    process.env.WF_GET_EXTRACT_PROVIDER_KEY = "resolved-secret";
-    try {
-      await mkdir(root, { recursive: true });
-      await writeFile(
-        join(root, "workflow.yaml"),
-        `config:
-  maxDepth: 1
-  extract:
-    baseUrl: https://example.com
-    model: m
-    apiKey: env:WF_GET_EXTRACT_PROVIDER_KEY
-workflows: {}
-`,
-        "utf8",
-      );
-      const r = await getExtractProvider(root);
-      expect(r.ok).toBe(true);
-      if (!r.ok) {
-        return;
-      }
-      expect(r.value.apiKey).toBe("resolved-secret");
-    } finally {
-      if (prev === undefined) {
-        delete process.env.WF_GET_EXTRACT_PROVIDER_KEY;
-      } else {
-        process.env.WF_GET_EXTRACT_PROVIDER_KEY = prev;
-      }
-      await rm(root, { recursive: true, force: true });
-    }
-  });
-});

packages/workflow/__tests__/refs-tracking.test.ts

@@ -1,5 +1,5 @@
 import { afterEach, describe, expect, test } from "bun:test";
-import { mkdir, mkdtemp, readFile, rm } from "node:fs/promises";
+import { mkdir, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import * as z from "zod/v4";
@@ -8,7 +8,6 @@ import { createCasStore } from "../src/cas/cas.js";
 import { createWorkflow } from "../src/engine/create-workflow.js";
 import { executeThread } from "../src/engine/engine.js";
 import { buildForkPlan, parseThreadDataJsonl } from "../src/engine/fork-thread.js";
-import { createExtract } from "../src/extract/extract-fn.js";
 import { END } from "../src/types.js";
 import { createLogger } from "../src/util/logger.js";
 
@@ -76,11 +75,16 @@ function installMockChatCompletions(sequence: ReadonlyArray<Record<string, unkno
   };
 }
 
-const refsDemoExtract = createExtract({
-  baseUrl: "http://127.0.0.1:9",
-  apiKey: "test",
-  model: "test",
-});
+const EXTRACT_REGISTRY_YAML = `config:
+  maxDepth: 3
+  providers:
+    stub:
+      baseUrl: http://127.0.0.1:9
+      apiKey: test
+  models:
+    default: stub/model
+workflows: {}
+`;
 
 const refsDemoWorkflow = createWorkflow<RefsDemoMeta>(
   {
@@ -99,8 +103,6 @@ const refsDemoWorkflow = createWorkflow<RefsDemoMeta>(
   {
     agent: async () => "plan-output",
   },
-  refsDemoExtract,
-  null,
 );
 
 describe("RoleStep refs tracking", () => {
@@ -142,6 +144,7 @@ describe("RoleStep refs tracking", () => {
       const dataPath = join(root, "logs", hash, `${threadId}.data.jsonl`);
       const infoPath = join(root, "logs", hash, `${threadId}.info.jsonl`);
       await mkdir(join(root, "logs", hash), { recursive: true });
+      await writeFile(join(root, "workflow.yaml"), EXTRACT_REGISTRY_YAML, "utf8");
       const cas = createCasStore(join(root, "cas"));
 
       const logger = createLogger({ sink: { kind: "file", path: infoPath } });
@@ -158,6 +161,7 @@ describe("RoleStep refs tracking", () => {
           awaitAfterEachYield: async () => {},
           forkSourceThreadId: null,
           prefilledDiskSteps: null,
+          storageRoot: root,
         },
         { threadId, hash, dataJsonlPath: dataPath, infoJsonlPath: infoPath, cas },
         logger,

packages/workflow/__tests__/registry.test.ts

@@ -105,10 +105,13 @@ describe("workflow registry", () => {
     const yaml = `
 config:
   maxDepth: 3
-  extract:
-    baseUrl: https://example.com/v1
-    model: qwen-plus
-    apiKey: secret-key
+  providers:
+    dashscope:
+      baseUrl: https://example.com/v1
+      apiKey: secret-key
+  models:
+    default: dashscope/qwen-turbo
+    extract: dashscope/qwen-plus
 workflows:
   solve-issue:
     hash: SPVR4BDMSGC1W
@@ -125,9 +128,69 @@ workflows:
       return;
     }
     expect(r.value.config.maxDepth).toBe(3);
-    expect(r.value.config.extract.baseUrl).toBe("https://example.com/v1");
-    expect(r.value.config.extract.model).toBe("qwen-plus");
-    expect(r.value.config.extract.apiKey).toBe("secret-key");
+    expect(r.value.config.providers.dashscope?.baseUrl).toBe("https://example.com/v1");
+    expect(r.value.config.providers.dashscope?.apiKey).toBe("secret-key");
+    expect(r.value.config.models.extract).toBe("dashscope/qwen-plus");
+    expect(r.value.config.models.default).toBe("dashscope/qwen-turbo");
+    expect(r.value.config.supervisorInterval).toBe(3);
+  });
+
+  test("defaults supervisorInterval to 3 when omitted", () => {
+    const yaml = `
+config:
+  maxDepth: 0
+  providers:
+    p:
+      baseUrl: https://example.com
+      apiKey: k
+  models:
+    default: p/m
+workflows: {}
+`;
+    const r = parseWorkflowRegistryYaml(yaml);
+    expect(r.ok).toBe(true);
+    if (!r.ok || r.value.config === null) {
+      return;
+    }
+    expect(r.value.config.supervisorInterval).toBe(3);
+  });
+
+  test("parses explicit supervisorInterval", () => {
+    const yaml = `
+config:
+  maxDepth: 0
+  supervisorInterval: 7
+  providers:
+    p:
+      baseUrl: https://example.com
+      apiKey: k
+  models:
+    default: p/m
+workflows: {}
+`;
+    const r = parseWorkflowRegistryYaml(yaml);
+    expect(r.ok).toBe(true);
+    if (!r.ok || r.value.config === null) {
+      return;
+    }
+    expect(r.value.config.supervisorInterval).toBe(7);
+  });
+
+  test("parse errors when supervisorInterval is negative", () => {
+    const yaml = `
+config:
+  maxDepth: 0
+  supervisorInterval: -1
+  providers:
+    p:
+      baseUrl: https://example.com
+      apiKey: k
+  models:
+    default: p/m
+workflows: {}
+`;
+    const r = parseWorkflowRegistryYaml(yaml);
+    expect(r.ok).toBe(false);
   });
 
   test("parses config apiKey env: prefix from process.env", () => {
@@ -137,10 +200,13 @@ workflows:
       const yaml = `
 config:
   maxDepth: 1
-  extract:
-    baseUrl: https://dashscope.aliyuncs.com/compatible-mode/v1
-    model: qwen-plus
-    apiKey: env:WF_REGISTRY_TEST_API_KEY
+  providers:
+    dashscope:
+      baseUrl: https://dashscope.aliyuncs.com/compatible-mode/v1
+      apiKey: env:WF_REGISTRY_TEST_API_KEY
+  models:
+    default: dashscope/qwen-plus
+    extract: dashscope/qwen-plus
 workflows: {}
 `;
       const r = parseWorkflowRegistryYaml(yaml);
@@ -148,7 +214,7 @@ workflows: {}
       if (!r.ok) {
         return;
       }
-      expect(r.value.config?.extract.apiKey).toBe("from-env");
+      expect(r.value.config?.providers.dashscope?.apiKey).toBe("from-env");
     } finally {
       if (prev === undefined) {
         delete process.env.WF_REGISTRY_TEST_API_KEY;
@@ -165,10 +231,12 @@ workflows: {}
       const yaml = `
 config:
   maxDepth: 1
-  extract:
-    baseUrl: https://example.com
-    model: m
-    apiKey: env:WF_REGISTRY_TEST_API_KEY_UNSET
+  providers:
+    p:
+      baseUrl: https://example.com
+      apiKey: env:WF_REGISTRY_TEST_API_KEY_UNSET
+  models:
+    default: p/m
 workflows: {}
 `;
       const r = parseWorkflowRegistryYaml(yaml);

packages/workflow/__tests__/resolve-model.test.ts

@@ -0,0 +1,104 @@
+import { describe, expect, test } from "bun:test";
+
+import { resolveModel } from "../src/config/resolve-model.js";
+import type { WorkflowConfig } from "../src/registry/index.js";
+
+function sampleConfig(): WorkflowConfig {
+  return {
+    maxDepth: 3,
+    supervisorInterval: 3,
+    providers: {
+      dashscope: {
+        baseUrl: "https://dashscope.aliyuncs.com/compatible-mode/v1",
+        apiKey: "secret",
+      },
+      other: {
+        baseUrl: "https://other.example/v1",
+        apiKey: "k2",
+      },
+    },
+    models: {
+      default: "dashscope/qwen-plus",
+      extract: "other/foo/bar-model",
+    },
+  };
+}
+
+describe("resolveModel", () => {
+  test("uses explicit scene mapping", () => {
+    const config = sampleConfig();
+    const r = resolveModel(config, "extract");
+    expect(r.ok).toBe(true);
+    if (!r.ok) {
+      return;
+    }
+    expect(r.value.baseUrl).toBe("https://other.example/v1");
+    expect(r.value.apiKey).toBe("k2");
+    expect(r.value.model).toBe("foo/bar-model");
+  });
+
+  test("falls back to models.default when scene is missing", () => {
+    const config = sampleConfig();
+    const r = resolveModel(config, "unknown-scene");
+    expect(r.ok).toBe(true);
+    if (!r.ok) {
+      return;
+    }
+    expect(r.value.model).toBe("qwen-plus");
+    expect(r.value.baseUrl).toBe("https://dashscope.aliyuncs.com/compatible-mode/v1");
+  });
+
+  test("errs when scene missing and no default", () => {
+    const config: WorkflowConfig = {
+      maxDepth: 1,
+      supervisorInterval: 3,
+      providers: {
+        p: { baseUrl: "https://x", apiKey: "k" },
+      },
+      models: {
+        extract: "p/m",
+      },
+    };
+    const r = resolveModel(config, "other");
+    expect(r.ok).toBe(false);
+    if (r.ok) {
+      return;
+    }
+    expect(r.error).toContain("no model mapping");
+    expect(r.error).toContain("default");
+  });
+
+  test("errs when provider is unknown", () => {
+    const config: WorkflowConfig = {
+      maxDepth: 1,
+      supervisorInterval: 3,
+      providers: {
+        p: { baseUrl: "https://x", apiKey: "k" },
+      },
+      models: {
+        default: "missing/m",
+      },
+    };
+    const r = resolveModel(config, "any");
+    expect(r.ok).toBe(false);
+    if (r.ok) {
+      return;
+    }
+    expect(r.error).toContain("unknown provider");
+  });
+
+  test("errs on invalid model reference shape", () => {
+    const config: WorkflowConfig = {
+      maxDepth: 1,
+      supervisorInterval: 3,
+      providers: {
+        p: { baseUrl: "https://x", apiKey: "k" },
+      },
+      models: {
+        default: "no-slash-model",
+      },
+    };
+    const r = resolveModel(config, "x");
+    expect(r.ok).toBe(false);
+  });
+});

packages/workflow/__tests__/supervisor.test.ts

@@ -0,0 +1,136 @@
+import { afterEach, describe, expect, test } from "bun:test";
+
+import { parseSupervisorDecisionText, runSupervisor } from "../src/engine/supervisor.js";
+import type { WorkflowConfig } from "../src/registry/index.js";
+import type { LogFn } from "../src/util/index.js";
+
+const noopLogger: LogFn = () => {};
+
+function supervisorOnlyConfig(): WorkflowConfig {
+  return {
+    maxDepth: 3,
+    supervisorInterval: 3,
+    providers: {
+      stub: { baseUrl: "http://127.0.0.1:9/v1", apiKey: "k" },
+    },
+    models: {
+      extract: "stub/extract-model",
+      supervisor: "stub/supervisor-model",
+    },
+  };
+}
+
+describe("parseSupervisorDecisionText", () => {
+  test("reads continue and stop case-insensitively", () => {
+    expect(parseSupervisorDecisionText("continue")).toBe("continue");
+    expect(parseSupervisorDecisionText("CONTINUE")).toBe("continue");
+    expect(parseSupervisorDecisionText("stop")).toBe("stop");
+    expect(parseSupervisorDecisionText("STOP.")).toBe("stop");
+  });
+
+  test("finds token inside a sentence", () => {
+    expect(parseSupervisorDecisionText("Answer: continue")).toBe("continue");
+    expect(parseSupervisorDecisionText("I recommend stop now")).toBe("stop");
+  });
+
+  test("when both appear, earlier token wins", () => {
+    expect(parseSupervisorDecisionText("continue then stop")).toBe("continue");
+    expect(parseSupervisorDecisionText("stop then continue")).toBe("stop");
+  });
+
+  test("defaults to continue when unclear", () => {
+    expect(parseSupervisorDecisionText("maybe later")).toBe("continue");
+  });
+});
+
+describe("runSupervisor", () => {
+  let restoreFetch: (() => void) | null = null;
+
+  afterEach(() => {
+    restoreFetch?.();
+    restoreFetch = null;
+  });
+
+  test("returns continue when supervisor model cannot be resolved (no fetch)", async () => {
+    const origFetch = globalThis.fetch;
+    restoreFetch = () => {
+      globalThis.fetch = origFetch;
+    };
+    globalThis.fetch = Object.assign(
+      async () => {
+        throw new Error("fetch should not run when supervisor is not configured");
+      },
+      { preconnect: origFetch.preconnect.bind(origFetch) },
+    ) as typeof fetch;
+
+    const config: WorkflowConfig = {
+      maxDepth: 1,
+      supervisorInterval: 3,
+      providers: {
+        stub: { baseUrl: "http://127.0.0.1:9/v1", apiKey: "k" },
+      },
+      models: {
+        extract: "stub/m",
+      },
+    };
+
+    const r = await runSupervisor({
+      config,
+      prompt: "task",
+      recentSteps: [{ role: "planner", summary: "{}" }],
+      logger: noopLogger,
+    });
+    expect(r.ok).toBe(true);
+    if (!r.ok) {
+      return;
+    }
+    expect(r.value).toBe("continue");
+  });
+
+  test("returns stop from chat/completions assistant content", async () => {
+    const origFetch = globalThis.fetch;
+    restoreFetch = () => {
+      globalThis.fetch = origFetch;
+    };
+    globalThis.fetch = Object.assign(
+      async () =>
+        new Response(
+          JSON.stringify({
+            choices: [{ message: { content: "stop" } }],
+          }),
+          { status: 200, headers: { "Content-Type": "application/json" } },
+        ),
+      { preconnect: origFetch.preconnect.bind(origFetch) },
+    ) as typeof fetch;
+
+    const r = await runSupervisor({
+      config: supervisorOnlyConfig(),
+      prompt: "do X",
+      recentSteps: [{ role: "a", summary: "{}" }],
+      logger: noopLogger,
+    });
+    expect(r.ok).toBe(true);
+    if (!r.ok) {
+      return;
+    }
+    expect(r.value).toBe("stop");
+  });
+
+  test("returns err on invalid JSON body", async () => {
+    const origFetch = globalThis.fetch;
+    restoreFetch = () => {
+      globalThis.fetch = origFetch;
+    };
+    globalThis.fetch = Object.assign(async () => new Response("not-json", { status: 200 }), {
+      preconnect: origFetch.preconnect.bind(origFetch),
+    }) as typeof fetch;
+
+    const r = await runSupervisor({
+      config: supervisorOnlyConfig(),
+      prompt: "p",
+      recentSteps: [],
+      logger: noopLogger,
+    });
+    expect(r.ok).toBe(false);
+  });
+});

packages/workflow/__tests__/worker.test.ts

@@ -9,6 +9,17 @@ import { createCasStore } from "../src/cas/cas.js";
 import { createContentMerkleNode, serializeMerkleNode } from "../src/cas/merkle.js";
 import { getWorkerHostScriptPath } from "../src/engine/worker-entry-path.js";
 
+const WORKER_REGISTRY_YAML = `config:
+  maxDepth: 3
+  providers:
+    stub:
+      baseUrl: http://127.0.0.1:9
+      apiKey: test
+  models:
+    default: stub/model
+workflows: {}
+`;
+
 const bundleSource = `import { putContentMerkleNode } from "@uncaged/workflow";
 
 export const descriptor = {
@@ -89,6 +100,7 @@ describe("worker process", () => {
     try {
       const hash = "C9NMV6V2TQT81";
       await mkdir(join(root, "bundles"), { recursive: true });
+      await writeFile(join(root, "workflow.yaml"), WORKER_REGISTRY_YAML, "utf8");
       const bundlePath = join(root, "bundles", `${hash}.esm.js`);
       await writeFile(bundlePath, bundleSource, "utf8");
 
@@ -136,6 +148,7 @@ describe("worker process", () => {
     try {
       const hash = "C9NMV6V2TQT81";
       await mkdir(join(root, "bundles"), { recursive: true });
+      await writeFile(join(root, "workflow.yaml"), WORKER_REGISTRY_YAML, "utf8");
       const bundlePath = join(root, "bundles", `${hash}.esm.js`);
       await writeFile(bundlePath, bundleSource, "utf8");
 

packages/workflow/__tests__/workflow-as-agent-integration.test.ts

@@ -9,7 +9,6 @@ import { hashWorkflowBundleBytes } from "../src/cas/hash.js";
 import { getContentMerklePayload, parseMerkleNode } from "../src/cas/merkle.js";
 import { createWorkflow } from "../src/engine/create-workflow.js";
 import { executeThread } from "../src/engine/engine.js";
-import { createExtract } from "../src/extract/extract-fn.js";
 import {
   readWorkflowRegistry,
   registerWorkflowVersion,
@@ -76,11 +75,16 @@ function installMockChatCompletions(sequence: ReadonlyArray<Record<string, unkno
   };
 }
 
-const parentExtract = createExtract({
-  baseUrl: "http://127.0.0.1:9",
-  apiKey: "test",
-  model: "test",
-});
+const PARENT_REGISTRY_WITH_CONFIG = `config:
+  maxDepth: 3
+  providers:
+    stub:
+      baseUrl: http://127.0.0.1:9
+      apiKey: test
+  models:
+    default: stub/m
+workflows: {}
+`;
 
 const childBundleSource = `import { putContentMerkleNode } from "@uncaged/workflow";
 
@@ -131,6 +135,8 @@ describe("workflowAsAgent integration", () => {
 
     const root = await mkdtemp(join(tmpdir(), "wf-waa-int-"));
     try {
+      await mkdir(root, { recursive: true });
+      await writeFile(join(root, "workflow.yaml"), PARENT_REGISTRY_WITH_CONFIG, "utf8");
       const { hash: childHash } = await installChildWorkflow(root);
 
       const parentWorkflow = createWorkflow<ParentMeta>(
@@ -148,8 +154,6 @@ describe("workflowAsAgent integration", () => {
           moderator: (ctx) => (ctx.steps.length === 0 ? "caller" : END),
         },
         { agent: workflowAsAgent("child-wf", { storageRoot: root }) },
-        parentExtract,
-        null,
       );
 
       const threadId = "01KQXKW18CT8G75T53R8F4G7YG";
@@ -173,6 +177,7 @@ describe("workflowAsAgent integration", () => {
           awaitAfterEachYield: async () => {},
           forkSourceThreadId: null,
           prefilledDiskSteps: null,
+          storageRoot: root,
         },
         { threadId, hash: parentHash, dataJsonlPath: dataPath, infoJsonlPath: infoPath, cas },
         logger,

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

@@ -93,6 +93,21 @@ describe("workflowAsAgent", () => {
   test("runs registered workflow and returns child thread root CAS hash", async () => {
     const root = await mkdtemp(join(tmpdir(), "wf-waa-ok-"));
     try {
+      await mkdir(root, { recursive: true });
+      await writeFile(
+        join(root, "workflow.yaml"),
+        `config:
+  maxDepth: 3
+  providers:
+    stub:
+      baseUrl: http://127.0.0.1:9
+      apiKey: test
+  models:
+    default: stub/m
+workflows: {}
+`,
+        "utf8",
+      );
       await installChildWorkflow(root);
       const agent = workflowAsAgent("child-wf", { storageRoot: root });
       const out = await agent(
@@ -140,10 +155,16 @@ describe("workflowAsAgent", () => {
         ...reg.value,
         config: {
           maxDepth: 2,
-          extract: {
-            baseUrl: "http://127.0.0.1:9",
-            model: "m",
-            apiKey: "k",
+          supervisorInterval: 3,
+          providers: {
+            local: {
+              baseUrl: "http://127.0.0.1:9",
+              apiKey: "k",
+            },
+          },
+          models: {
+            default: "local/m",
+            extract: "local/m",
           },
         },
       };

packages/workflow/src/config/index.ts

@@ -0,0 +1,3 @@
+export { resolveModel } from "./resolve-model.js";
+export { splitProviderModelRef } from "./split-provider-model-ref.js";
+export type { ProviderConfig, ResolvedModel } from "./types.js";

packages/workflow/src/config/resolve-model.ts

@@ -0,0 +1,30 @@
+import type { WorkflowConfig } from "../registry/index.js";
+import { err, ok, type Result } from "../util/index.js";
+import { splitProviderModelRef } from "./split-provider-model-ref.js";
+import type { ResolvedModel } from "./types.js";
+
+/** Resolves scene → provider endpoint + model using {@link WorkflowConfig.providers} and {@link WorkflowConfig.models}. */
+export function resolveModel(config: WorkflowConfig, scene: string): Result<ResolvedModel, string> {
+  const models = config.models;
+  let ref = models[scene] ?? null;
+  if (ref === null) {
+    ref = models.default ?? null;
+  }
+  if (ref === null) {
+    return err(`no model mapping for scene "${scene}" and no models.default fallback`);
+  }
+  const split = splitProviderModelRef(ref);
+  if (!split.ok) {
+    return split;
+  }
+  const { providerName, modelName } = split.value;
+  const provider = config.providers[providerName] ?? null;
+  if (provider === null) {
+    return err(`unknown provider "${providerName}" referenced by scene "${scene}"`);
+  }
+  return ok({
+    baseUrl: provider.baseUrl,
+    apiKey: provider.apiKey,
+    model: modelName,
+  });
+}

packages/workflow/src/config/split-provider-model-ref.ts

@@ -0,0 +1,17 @@
+import { err, ok, type Result } from "../util/index.js";
+
+/** Parses `providerName/modelName` references used in {@link WorkflowConfig.models}. */
+export function splitProviderModelRef(
+  ref: string,
+): Result<{ providerName: string; modelName: string }, string> {
+  const idx = ref.indexOf("/");
+  if (idx <= 0 || idx === ref.length - 1) {
+    return err(`invalid model reference "${ref}": expected providerName/modelName`);
+  }
+  const providerName = ref.slice(0, idx);
+  const modelName = ref.slice(idx + 1);
+  if (providerName === "" || modelName === "") {
+    return err(`invalid model reference "${ref}": expected providerName/modelName`);
+  }
+  return ok({ providerName, modelName });
+}

packages/workflow/src/config/types.ts

@@ -0,0 +1,10 @@
+export type ProviderConfig = {
+  baseUrl: string;
+  apiKey: string;
+};
+
+export type ResolvedModel = {
+  baseUrl: string;
+  apiKey: string;
+  model: string;
+};

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

@@ -1,12 +1,10 @@
-import type { CasStore } from "../cas/index.js";
 import { putContentMerkleNode } from "../cas/index.js";
-import { buildExtractUserContent, type ExtractFn, reactExtract } from "../extract/index.js";
+import { buildExtractUserContent, reactExtract } from "../extract/index.js";
 import {
   type AgentBinding,
   type AgentContext,
   END,
   type ExtractContext,
-  type LlmProvider,
   type ModeratorContext,
   type RoleDefinition,
   type RoleMeta,
@@ -41,14 +39,12 @@ function resolveExtractedRefs(
 async function resolveRoleMeta<M extends RoleMeta>(
   roleDef: RoleDefinition<Record<string, unknown>>,
   extractCtx: ExtractContext<M>,
-  extract: ExtractFn,
-  llmProvider: LlmProvider | null,
-  cas: CasStore,
+  options: WorkflowFnOptions,
 ): Promise<Record<string, unknown>> {
   if (roleDef.extractMode === "react") {
-    if (llmProvider === null) {
+    if (options.llmProvider === null) {
       throw new Error(
-        'createWorkflow: llmProvider is required when a role uses extractMode "react"',
+        'createWorkflow: WorkflowFnOptions.llmProvider is required when a role uses extractMode "react"',
       );
     }
     const text = await buildExtractUserContent(
@@ -58,15 +54,15 @@ async function resolveRoleMeta<M extends RoleMeta>(
     const reactResult = await reactExtract({
       text,
       schema: roleDef.schema,
-      provider: llmProvider,
-      cas,
+      provider: options.llmProvider,
+      cas: options.cas,
     });
     if (!reactResult.ok) {
       throw new Error(`react extract failed: ${reactResult.error}`);
     }
     return reactResult.value as Record<string, unknown>;
   }
-  return (await extract(
+  return (await options.extract(
     roleDef.schema,
     roleDef.extractPrompt,
     extractCtx as unknown as ExtractContext,
@@ -74,15 +70,13 @@ async function resolveRoleMeta<M extends RoleMeta>(
 }
 
 /**
- * Binds pure role definitions + moderator to runtime agents and structured extraction.
- * Assign with `export const run = createWorkflow(def, binding, extract, llmProvider)`.
- * Pass the same {@link LlmProvider} as {@link createExtract} when any role uses `extractMode: "react"`.
+ * Binds pure role definitions + moderator to runtime agents.
+ * Assign with `export const run = createWorkflow(def, binding)`.
+ * The engine supplies {@link WorkflowFnOptions.extract} and {@link WorkflowFnOptions.llmProvider} from workflow.yaml.
  */
 export function createWorkflow<M extends RoleMeta>(
   def: Pick<WorkflowDefinition<M>, "roles" | "moderator">,
   binding: AgentBinding,
-  extract: ExtractFn,
-  llmProvider: LlmProvider | null,
 ): WorkflowFn {
   return async function* workflowLoop(
     input: ThreadInput,
@@ -149,9 +143,7 @@ export function createWorkflow<M extends RoleMeta>(
       const meta = await resolveRoleMeta(
         roleDef as unknown as RoleDefinition<Record<string, unknown>>,
         extractCtx,
-        extract,
-        llmProvider,
-        options.cas,
+        options,
       );
 
       const contentHash = await putContentMerkleNode(options.cas, raw);

packages/workflow/src/engine/engine.ts

@@ -7,17 +7,53 @@ import {
   putStepMerkleNode,
   putThreadMerkleNode,
 } from "../cas/index.js";
+import { resolveModel } from "../config/index.js";
+import { createExtract } from "../extract/index.js";
+import { readWorkflowRegistry, type WorkflowConfig } from "../registry/index.js";
 import type {
+  LlmProvider,
   ThreadInput,
   WorkflowCompletion,
   WorkflowFn,
   WorkflowFnOptions,
   WorkflowResult,
 } from "../types.js";
-import { type LogFn, normalizeRefsField } from "../util/index.js";
+import { err, type LogFn, normalizeRefsField, ok, type Result } from "../util/index.js";
 
+import { runSupervisor } from "./supervisor.js";
 import type { ExecuteThreadIo, ExecuteThreadOptions } from "./types.js";
 
+async function resolveEngineRegistryRuntime(storageRoot: string): Promise<
+  Result<
+    {
+      extract: ReturnType<typeof createExtract>;
+      llmProvider: LlmProvider;
+      workflowConfig: WorkflowConfig;
+    },
+    string
+  >
+> {
+  const reg = await readWorkflowRegistry(storageRoot);
+  if (!reg.ok) {
+    return err(reg.error.message);
+  }
+  const cfg = reg.value.config;
+  if (cfg === null) {
+    return err("workflow registry has no global config section");
+  }
+  const resolved = resolveModel(cfg, "extract");
+  if (!resolved.ok) {
+    return resolved;
+  }
+  const ex = resolved.value;
+  const llmProvider: LlmProvider = {
+    baseUrl: ex.baseUrl,
+    apiKey: ex.apiKey,
+    model: ex.model,
+  };
+  return ok({ extract: createExtract(llmProvider), llmProvider, workflowConfig: cfg });
+}
+
 async function appendDataLine(path: string, record: unknown): Promise<void> {
   const line = `${JSON.stringify(record)}\n`;
   await appendFile(path, line, "utf8");
@@ -49,9 +85,66 @@ async function finalizeThreadResult(params: {
   };
 }
 
+async function finalizeAbortedThread(params: {
+  cas: CasStore;
+  workflowName: string;
+  threadId: string;
+  stepMerkleHashes: string[];
+  logger: LogFn;
+  abortLogTag: string;
+}): Promise<WorkflowResult> {
+  params.logger(params.abortLogTag, `thread ${params.threadId} aborted`);
+  return finalizeThreadResult({
+    cas: params.cas,
+    workflowName: params.workflowName,
+    threadId: params.threadId,
+    stepMerkleHashes: params.stepMerkleHashes,
+    completion: { returnCode: 130, summary: "thread aborted" },
+  });
+}
+
+async function maybeSupervisorHaltsThread(params: {
+  workflowConfig: WorkflowConfig;
+  input: ThreadInput;
+  written: number;
+  recentSupervisorSteps: readonly { role: string; summary: string }[];
+  logger: LogFn;
+  threadId: string;
+  cas: CasStore;
+  workflowName: string;
+  stepMerkleHashes: string[];
+}): Promise<WorkflowResult | null> {
+  const interval = params.workflowConfig.supervisorInterval;
+  if (interval <= 0 || params.written % interval !== 0) {
+    return null;
+  }
+  const sup = await runSupervisor({
+    config: params.workflowConfig,
+    prompt: params.input.prompt,
+    recentSteps: params.recentSupervisorSteps,
+    logger: params.logger,
+  });
+  if (!sup.ok) {
+    params.logger("K6PW9NYT", `supervisor skipped: ${sup.error}`);
+    return null;
+  }
+  if (sup.value !== "stop") {
+    return null;
+  }
+  params.logger("M4QX8VHN", `thread ${params.threadId} stopped by supervisor`);
+  return finalizeThreadResult({
+    cas: params.cas,
+    workflowName: params.workflowName,
+    threadId: params.threadId,
+    stepMerkleHashes: params.stepMerkleHashes,
+    completion: { returnCode: 0, summary: "completed: supervisor stopped thread" },
+  });
+}
+
 async function driveWorkflowGenerator(params: {
   fn: WorkflowFn;
   workflowName: string;
+  workflowConfig: WorkflowConfig;
   input: ThreadInput;
   bundleOptions: WorkflowFnOptions;
   executeOptions: ExecuteThreadOptions;
@@ -64,6 +157,7 @@ async function driveWorkflowGenerator(params: {
   const {
     fn,
     workflowName,
+    workflowConfig,
     input,
     bundleOptions,
     executeOptions,
@@ -75,16 +169,20 @@ async function driveWorkflowGenerator(params: {
   } = params;
   const gen = fn(input, bundleOptions);
   let written = 0;
+  const recentSupervisorSteps: { role: string; summary: string }[] = input.steps.map((s) => ({
+    role: s.role,
+    summary: JSON.stringify(s.meta),
+  }));
 
   while (true) {
     if (executeOptions.signal.aborted) {
-      logger("V8JX4NP2", `thread ${threadId} aborted`);
-      return await finalizeThreadResult({
+      return await finalizeAbortedThread({
         cas,
         workflowName,
         threadId,
         stepMerkleHashes,
-        completion: { returnCode: 130, summary: "thread aborted" },
+        logger,
+        abortLogTag: "V8JX4NP2",
       });
     }
 
@@ -142,6 +240,11 @@ async function driveWorkflowGenerator(params: {
 
     logger("N7BW4YHQ", `thread ${threadId} wrote role ${step.role}`);
 
+    recentSupervisorSteps.push({
+      role: step.role,
+      summary: JSON.stringify(step.meta),
+    });
+
     await Promise.race([
       executeOptions.awaitAfterEachYield(),
       new Promise<void>((resolve) => {
@@ -154,15 +257,30 @@ async function driveWorkflowGenerator(params: {
     ]);
 
     if (executeOptions.signal.aborted) {
-      logger("V8JX4NP4", `thread ${threadId} aborted`);
-      return await finalizeThreadResult({
+      return await finalizeAbortedThread({
         cas,
         workflowName,
         threadId,
         stepMerkleHashes,
-        completion: { returnCode: 130, summary: "thread aborted" },
+        logger,
+        abortLogTag: "V8JX4NP4",
       });
     }
+
+    const supervised = await maybeSupervisorHaltsThread({
+      workflowConfig,
+      input,
+      written,
+      recentSupervisorSteps,
+      logger,
+      threadId,
+      cas,
+      workflowName,
+      stepMerkleHashes,
+    });
+    if (supervised !== null) {
+      return supervised;
+    }
   }
 }
 
@@ -250,16 +368,24 @@ export async function executeThread(
     });
   }
 
+  const registryRuntime = await resolveEngineRegistryRuntime(options.storageRoot);
+  if (!registryRuntime.ok) {
+    throw new Error(registryRuntime.error);
+  }
+
   const bundleOptions: WorkflowFnOptions = {
     threadId: io.threadId,
     maxRounds: options.maxRounds,
     depth: options.depth,
     cas: io.cas,
+    extract: registryRuntime.value.extract,
+    llmProvider: registryRuntime.value.llmProvider,
   };
 
   return await driveWorkflowGenerator({
     fn,
     workflowName,
+    workflowConfig: registryRuntime.value.workflowConfig,
     input,
     bundleOptions,
     executeOptions: options,

packages/workflow/src/engine/index.ts

@@ -17,6 +17,7 @@ export type {
   GcResult,
   ParsedThreadStartRecord,
   PrefilledDiskStep,
+  SupervisorDecision,
   ThreadPauseGate,
 } from "./types.js";
 export { getWorkerHostScriptPath } from "./worker-entry-path.js";

packages/workflow/src/engine/supervisor.ts

@@ -0,0 +1,140 @@
+import { resolveModel } from "../config/index.js";
+import type { WorkflowConfig } from "../registry/index.js";
+import { err, type LogFn, ok, type Result } from "../util/index.js";
+
+import type { SupervisorDecision } from "./types.js";
+
+const SUPERVISOR_RECENT_STEP_LIMIT = 12;
+
+function chatCompletionsUrl(baseUrl: string): string {
+  const trimmed = baseUrl.replace(/\/+$/, "");
+  return `${trimmed}/chat/completions`;
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
+function readAssistantContent(parsed: unknown): string | null {
+  if (!isRecord(parsed)) {
+    return null;
+  }
+  const choices = parsed.choices;
+  if (!Array.isArray(choices) || choices.length === 0) {
+    return null;
+  }
+  const first = choices[0];
+  if (!isRecord(first)) {
+    return null;
+  }
+  const messageObj = first.message;
+  if (!isRecord(messageObj)) {
+    return null;
+  }
+  const content = messageObj.content;
+  if (typeof content !== "string") {
+    return null;
+  }
+  return content;
+}
+
+/** Lenient: accepts STOP/stop/stop. as prose; prefers {@link SupervisorDecision.stop} when both tokens appear. */
+export function parseSupervisorDecisionText(text: string): SupervisorDecision {
+  const lower = text.toLowerCase();
+  const stopWord = /\bstop\b/.test(lower);
+  const continueWord = /\bcontinue\b/.test(lower);
+  if (stopWord && continueWord) {
+    const si = lower.search(/\bstop\b/);
+    const ci = lower.search(/\bcontinue\b/);
+    return si <= ci ? "stop" : "continue";
+  }
+  if (stopWord) {
+    return "stop";
+  }
+  if (continueWord) {
+    return "continue";
+  }
+  if (lower.includes("stop")) {
+    return "stop";
+  }
+  if (lower.includes("continue")) {
+    return "continue";
+  }
+  return "continue";
+}
+
+type RunSupervisorArgs = {
+  config: WorkflowConfig;
+  prompt: string;
+  recentSteps: readonly { role: string; summary: string }[];
+  logger: LogFn;
+};
+
+/** Calls the `supervisor` scene LLM; opt-out when {@link resolveModel} fails (returns ok(`continue`)). */
+export async function runSupervisor(
+  args: RunSupervisorArgs,
+): Promise<Result<SupervisorDecision, string>> {
+  const resolved = resolveModel(args.config, "supervisor");
+  if (!resolved.ok) {
+    return ok("continue");
+  }
+  const provider = resolved.value;
+  const recent = args.recentSteps.slice(-SUPERVISOR_RECENT_STEP_LIMIT);
+  const stepsBlock = recent.map((s, index) => `${index + 1}. [${s.role}] ${s.summary}`).join("\n");
+
+  const body = {
+    model: provider.model,
+    messages: [
+      {
+        role: "system" as const,
+        content:
+          'You supervise a multi-step workflow. Decide if the thread should keep running or halt.\n\nReply with exactly one token: either "continue" (progress toward the goal, not obviously stuck) or "stop" (done, looping, or no progress). Do not add explanation.',
+      },
+      {
+        role: "user" as const,
+        content: `Original task:\n${args.prompt}\n\nRecent steps (oldest first):\n${stepsBlock === "" ? "(none)" : stepsBlock}`,
+      },
+    ],
+  };
+
+  let response: Response;
+  try {
+    response = await fetch(chatCompletionsUrl(provider.baseUrl), {
+      method: "POST",
+      headers: {
+        Authorization: `Bearer ${provider.apiKey}`,
+        "Content-Type": "application/json",
+      },
+      body: JSON.stringify(body),
+    });
+  } catch (cause) {
+    const message = cause instanceof Error ? cause.message : String(cause);
+    args.logger("R9CW4PLM", `supervisor request failed: ${message}`);
+    return err(`supervisor network error: ${message}`);
+  }
+
+  const responseText = await response.text();
+  if (!response.ok) {
+    args.logger("T3HN8VKQ", `supervisor HTTP ${response.status}: ${responseText.slice(0, 200)}`);
+    return err(`supervisor HTTP ${response.status}: ${responseText.slice(0, 500)}`);
+  }
+
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(responseText) as unknown;
+  } catch (cause) {
+    const message = cause instanceof Error ? cause.message : String(cause);
+    args.logger("W7BQ2NXM", `supervisor response is not JSON: ${message}`);
+    return err(`supervisor invalid JSON: ${message}`);
+  }
+
+  const content = readAssistantContent(parsed);
+  if (content === null || content.trim() === "") {
+    args.logger("Y4JX9PKW", "supervisor returned empty assistant content");
+    return err("supervisor empty assistant content");
+  }
+
+  const decision = parseSupervisorDecisionText(content);
+  args.logger("Z8KM5QWT", `supervisor says ${decision}`);
+  return ok(decision);
+}

packages/workflow/src/engine/types.ts

@@ -2,6 +2,8 @@ import type { CasStore } from "../cas/index.js";
 import type { RoleOutput } from "../types.js";
 import type { Result } from "../util/index.js";
 
+export type SupervisorDecision = "continue" | "stop";
+
 export type ExecuteThreadIo = {
   threadId: string;
   hash: string;
@@ -33,6 +35,8 @@ export type ExecuteThreadOptions = {
    * Must match `input.steps` length and order when present.
    */
   prefilledDiskSteps: PrefilledDiskStep[] | null;
+  /** Workspace root containing `workflow.yaml`; used to resolve the `extract` scene for meta extraction. */
+  storageRoot: string;
 };
 
 /** Role steps replayed from `.data.jsonl`, including persisted timestamps. */

packages/workflow/src/engine/worker.ts

@@ -417,6 +417,7 @@ async function main(): Promise<void> {
           awaitAfterEachYield: () => pauseGate.awaitAfterYield(),
           forkSourceThreadId: cmd.forkSourceThreadId,
           prefilledDiskSteps,
+          storageRoot,
         },
         io,
         logger,

packages/workflow/src/extract-provider.ts

@@ -1,34 +0,0 @@
-import type { WorkflowConfig } from "./registry/index.js";
-import { readWorkflowRegistry } from "./registry/index.js";
-import type { LlmProvider } from "./types.js";
-import { err, getDefaultWorkflowStorageRoot, ok, type Result } from "./util/index.js";
-
-const DEFAULT_WORKFLOW_AS_AGENT_MAX_DEPTH = 3;
-
-export function getWorkflowAsAgentMaxDepth(config: WorkflowConfig | null): number {
-  if (config === null) {
-    return DEFAULT_WORKFLOW_AS_AGENT_MAX_DEPTH;
-  }
-  return config.maxDepth;
-}
-
-/** Loads `config.extract` from workflow.yaml (apiKey already resolved at registry parse time). */
-export async function getExtractProvider(
-  storageRoot: string | undefined,
-): Promise<Result<LlmProvider, string>> {
-  const root = storageRoot ?? getDefaultWorkflowStorageRoot();
-  const regResult = await readWorkflowRegistry(root);
-  if (!regResult.ok) {
-    return err(regResult.error.message);
-  }
-  const cfg = regResult.value.config;
-  if (cfg === null) {
-    return err("workflow registry has no global config section");
-  }
-  const ex = cfg.extract;
-  return ok({
-    baseUrl: ex.baseUrl,
-    apiKey: ex.apiKey,
-    model: ex.model,
-  });
-}

packages/workflow/src/index.ts

@@ -28,6 +28,11 @@ export {
   serializeMerkleNode,
   type ThreadMerklePayload,
 } from "./cas/index.js";
+export {
+  type ProviderConfig,
+  type ResolvedModel,
+  resolveModel,
+} from "./config/index.js";
 export {
   buildForkPlan,
   createThreadPauseGate,
@@ -43,6 +48,7 @@ export {
   type ParsedThreadStartRecord,
   type PrefilledDiskStep,
   parseThreadDataJsonl,
+  type SupervisorDecision,
   selectForkHistoricalSteps,
   type ThreadPauseGate,
   tryParseRoleStepRecord,
@@ -58,9 +64,7 @@ export {
   type ReactExtractArgs,
   reactExtract,
 } from "./extract/index.js";
-export { getExtractProvider } from "./extract-provider.js";
 export {
-  type ExtractProviderConfig,
   getRegisteredWorkflow,
   listRegisteredWorkflowNames,
   parseWorkflowRegistryYaml,

packages/workflow/src/registry/index.ts

@@ -11,7 +11,6 @@ export {
   writeWorkflowRegistry,
 } from "./registry.js";
 export type {
-  ExtractProviderConfig,
   WorkflowConfig,
   WorkflowHistoryEntry,
   WorkflowRegistryEntry,

packages/workflow/src/registry/registry-normalize.ts

@@ -1,49 +1,107 @@
-import { err, ok, type Result } from "../util/index.js";
+import { type ProviderConfig, splitProviderModelRef } from "../config/index.js";
+import { createLogger, err, ok, type Result } from "../util/index.js";
 import type {
-  ExtractProviderConfig,
   WorkflowConfig,
   WorkflowHistoryEntry,
   WorkflowRegistryEntry,
   WorkflowRegistryFile,
 } from "./types.js";
 
-function resolveRegistryApiKey(raw: string): Result<string, Error> {
+const registryNormalizeLog = createLogger({ sink: { kind: "stderr" } });
+
+function resolveRegistryApiKey(raw: string, ctx: string): Result<string, Error> {
   if (raw.startsWith("env:")) {
     const name = raw.slice("env:".length);
     if (name === "") {
-      return err(new Error('config.extract.apiKey "env:" reference must name a variable'));
+      return err(new Error(`${ctx}: "env:" apiKey reference must name a variable`));
     }
     const value = process.env[name];
     if (value === undefined) {
-      return err(new Error(`config.extract.apiKey: environment variable "${name}" is not set`));
+      return err(new Error(`${ctx}: environment variable "${name}" is not set`));
     }
     return ok(value);
   }
   return ok(raw);
 }
 
-function normalizeExtractProviderConfig(raw: unknown): Result<ExtractProviderConfig, Error> {
-  if (raw === null || typeof raw !== "object") {
-    return err(new Error('registry config must contain an "extract" mapping'));
+function normalizeProviderEntry(name: string, entryRaw: unknown): Result<ProviderConfig, Error> {
+  if (name === "") {
+    return err(new Error("config.providers must not contain an empty provider name"));
   }
-  const e = raw as Record<string, unknown>;
+  if (entryRaw === null || typeof entryRaw !== "object" || Array.isArray(entryRaw)) {
+    return err(new Error(`config.providers.${name} must be a mapping`));
+  }
+  const e = entryRaw as Record<string, unknown>;
   const baseUrl = e.baseUrl;
-  const model = e.model;
   const apiKeyRaw = e.apiKey;
   if (typeof baseUrl !== "string" || baseUrl === "") {
-    return err(new Error("config.extract.baseUrl must be a non-empty string"));
-  }
-  if (typeof model !== "string" || model === "") {
-    return err(new Error("config.extract.model must be a non-empty string"));
+    return err(new Error(`config.providers.${name}.baseUrl must be a non-empty string`));
   }
   if (typeof apiKeyRaw !== "string" || apiKeyRaw === "") {
-    return err(new Error("config.extract.apiKey must be a non-empty string"));
+    return err(new Error(`config.providers.${name}.apiKey must be a non-empty string`));
   }
-  const apiKeyResult = resolveRegistryApiKey(apiKeyRaw);
+  const apiKeyCtx = `config.providers.${name}.apiKey`;
+  const apiKeyResult = resolveRegistryApiKey(apiKeyRaw, apiKeyCtx);
   if (!apiKeyResult.ok) {
     return apiKeyResult;
   }
-  return ok({ baseUrl, model, apiKey: apiKeyResult.value });
+  return ok({ baseUrl, apiKey: apiKeyResult.value });
+}
+
+function normalizeProviders(raw: unknown): Result<Record<string, ProviderConfig>, Error> {
+  if (raw === null || typeof raw !== "object" || Array.isArray(raw)) {
+    return err(new Error('registry config must contain a "providers" mapping'));
+  }
+  const root = raw as Record<string, unknown>;
+  const providers: Record<string, ProviderConfig> = {};
+  for (const [name, entryRaw] of Object.entries(root)) {
+    const next = normalizeProviderEntry(name, entryRaw);
+    if (!next.ok) {
+      return next;
+    }
+    providers[name] = next.value;
+  }
+  return ok(providers);
+}
+
+function normalizeModels(
+  raw: unknown,
+  providers: Record<string, ProviderConfig>,
+): Result<Record<string, string>, Error> {
+  if (raw === null || typeof raw !== "object" || Array.isArray(raw)) {
+    return err(new Error('registry config must contain a "models" mapping'));
+  }
+  const root = raw as Record<string, unknown>;
+  const models: Record<string, string> = {};
+  const providerKeys = new Set(Object.keys(providers));
+  for (const [scene, refRaw] of Object.entries(root)) {
+    if (scene === "") {
+      return err(new Error("config.models must not contain an empty scene name"));
+    }
+    if (typeof refRaw !== "string" || refRaw === "") {
+      return err(new Error(`config.models.${scene} must be a non-empty string (provider/model)`));
+    }
+    const ctx = `config.models.${scene}`;
+    const parsed = splitProviderModelRef(refRaw);
+    if (!parsed.ok) {
+      return err(new Error(`${ctx}: ${parsed.error}`));
+    }
+    if (!providerKeys.has(parsed.value.providerName)) {
+      return err(
+        new Error(
+          `${ctx}: unknown provider "${parsed.value.providerName}" (not listed under config.providers)`,
+        ),
+      );
+    }
+    models[scene] = refRaw;
+  }
+  if (!Object.hasOwn(models, "default")) {
+    registryNormalizeLog(
+      "Z2KP9NWQ",
+      'registry config: models mapping has no "default" key; scenes without explicit model mappings may fail at resolveModel',
+    );
+  }
+  return ok(models);
 }
 
 function normalizeWorkflowConfig(raw: unknown): Result<WorkflowConfig, Error> {
@@ -52,15 +110,37 @@ function normalizeWorkflowConfig(raw: unknown): Result<WorkflowConfig, Error> {
   }
   const c = raw as Record<string, unknown>;
   const maxDepth = c.maxDepth;
-  const extractRaw = c.extract;
+  const supervisorIntervalRaw = c.supervisorInterval;
+  const providersRaw = c.providers;
+  const modelsRaw = c.models;
   if (typeof maxDepth !== "number" || !Number.isInteger(maxDepth) || maxDepth < 0) {
     return err(new Error("config.maxDepth must be a non-negative integer"));
   }
-  const extractResult = normalizeExtractProviderConfig(extractRaw);
-  if (!extractResult.ok) {
-    return extractResult;
+  let supervisorInterval = 3;
+  if (supervisorIntervalRaw !== undefined) {
+    if (
+      typeof supervisorIntervalRaw !== "number" ||
+      !Number.isInteger(supervisorIntervalRaw) ||
+      supervisorIntervalRaw < 0
+    ) {
+      return err(new Error("config.supervisorInterval must be a non-negative integer"));
+    }
+    supervisorInterval = supervisorIntervalRaw;
   }
-  return ok({ maxDepth, extract: extractResult.value });
+  const providersResult = normalizeProviders(providersRaw);
+  if (!providersResult.ok) {
+    return providersResult;
+  }
+  const modelsResult = normalizeModels(modelsRaw, providersResult.value);
+  if (!modelsResult.ok) {
+    return modelsResult;
+  }
+  return ok({
+    maxDepth,
+    supervisorInterval,
+    providers: providersResult.value,
+    models: modelsResult.value,
+  });
 }
 
 export function normalizeWorkflowHistoryEntry(

packages/workflow/src/registry/types.ts

@@ -1,3 +1,5 @@
+import type { ProviderConfig } from "../config/index.js";
+
 export type WorkflowHistoryEntry = {
   hash: string;
   timestamp: number;
@@ -9,16 +11,12 @@ export type WorkflowRegistryEntry = {
   history: WorkflowHistoryEntry[];
 };
 
-/** LLM provider settings under `config.extract` in workflow.yaml (apiKey resolved after parse). */
-export type ExtractProviderConfig = {
-  baseUrl: string;
-  model: string;
-  apiKey: string;
-};
-
 export type WorkflowConfig = {
   maxDepth: number;
-  extract: ExtractProviderConfig;
+  /** Run supervisor LLM every N completed role rounds (0 = disabled). Default from YAML: 3. */
+  supervisorInterval: number;
+  providers: Record<string, ProviderConfig>;
+  models: Record<string, string>;
 };
 
 export type WorkflowRegistryFile = {

packages/workflow/src/types.ts

@@ -1,6 +1,7 @@
 import type * as z from "zod/v4";
 
 import type { CasStore } from "./cas/index.js";
+import type { ExtractFn } from "./extract/types.js";
 
 /** Sentinel values for automaton control flow. */
 export const START = "__start__" as const;
@@ -54,6 +55,10 @@ export type WorkflowFnOptions = {
   depth: number;
   /** Global CAS store for Merkle content blobs (role step bodies). */
   cas: CasStore;
+  /** Structured meta extraction; resolved from workflow.yaml `extract` scene by the engine. */
+  extract: ExtractFn;
+  /** Provider for `extractMode: "react"` roles; same backing config as `extract`. */
+  llmProvider: LlmProvider | null;
 };
 
 /** Bundle contract — named export `run` is a function returning an AsyncGenerator. */

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

@@ -4,7 +4,7 @@ import { extractBundleExports } from "./bundle/index.js";
 import { createCasStore } from "./cas/index.js";
 import type { ExecuteThreadIo } from "./engine/index.js";
 import { executeThread } from "./engine/index.js";
-import { getWorkflowAsAgentMaxDepth } from "./extract-provider.js";
+import type { WorkflowConfig } from "./registry/index.js";
 import { getRegisteredWorkflow, readWorkflowRegistry } from "./registry/index.js";
 import type { AgentContext, AgentFn, ThreadInput } from "./types.js";
 import {
@@ -14,6 +14,15 @@ import {
   getGlobalCasDir,
 } from "./util/index.js";
 
+const DEFAULT_WORKFLOW_AS_AGENT_MAX_DEPTH = 3;
+
+function workflowAsAgentMaxDepth(config: WorkflowConfig | null): number {
+  if (config === null) {
+    return DEFAULT_WORKFLOW_AS_AGENT_MAX_DEPTH;
+  }
+  return config.maxDepth;
+}
+
 export type WorkflowAsAgentOptions = {
   /** When `null`, uses `getDefaultWorkflowStorageRoot()`. */
   storageRoot: string | null;
@@ -44,7 +53,7 @@ export function workflowAsAgent(
       return `ERROR: failed to read workflow registry: ${registryResult.error.message}`;
     }
 
-    const maxDepth = getWorkflowAsAgentMaxDepth(registryResult.value.config);
+    const maxDepth = workflowAsAgentMaxDepth(registryResult.value.config);
     if (nextDepth > maxDepth) {
       return `ERROR: workflow-as-agent depth limit exceeded (max ${maxDepth})`;
     }
@@ -92,6 +101,7 @@ export function workflowAsAgent(
           awaitAfterEachYield: async () => {},
           forkSourceThreadId: ctx.threadId,
           prefilledDiskSteps: null,
+          storageRoot,
         },
         io,
         logger,