packages/agent-hermes/package.json

@@ -19,7 +19,7 @@
   "scripts": {
     "prepublishOnly": "echo 'Use pnpm run release from repo root' && exit 1",
     "test": "vitest run __tests__/",
-    "test:ci": "vitest run __tests__/"
+    "test:ci": "vitest run __tests__/ --exclude __tests__/integration/"
   },
   "dependencies": {
     "@ocas/core": "^0.3.0",

packages/cli/src/__tests__/e2e-mock-agent.test.ts

@@ -241,7 +241,9 @@ describe("E2E mock-agent: full uwf pipeline", () => {
     expect(finalEntry!.head).toBe(step2.head);
   });
 
-  test("2. branching workflow loops developer→reviewer→developer→reviewer→$END", async () => {
+  test("2. branching workflow loops developer→reviewer→developer→reviewer→$END", {
+    timeout: 30_000,
+  }, async () => {
     await writeMockConfig("e2e-loop.mock.yaml");
     const workflowHash = await addWorkflow("e2e-loop.workflow.yaml", "test-loop");
 
@@ -299,7 +301,9 @@ describe("E2E mock-agent: full uwf pipeline", () => {
     expect(finalEntry!.status).toBe("completed");
   });
 
-  test("3. role mismatch in mock data makes the agent exit with an error", async () => {
+  test("3. role mismatch in mock data makes the agent exit with an error", {
+    timeout: 30_000,
+  }, async () => {
     // Reuses the linear workflow but with a mock whose step[1].role is wrong.
     await writeMockConfig("e2e-mismatch.mock.yaml");
     const workflowHash = await addWorkflow("e2e-linear.workflow.yaml", "test-linear");
@@ -325,7 +329,9 @@ describe("E2E mock-agent: full uwf pipeline", () => {
     expect(entry!.head).toBe(step1.head);
   });
 
-  test("4. planner $SUSPEND then resume re-runs planner and reaches $END", async () => {
+  test("4. planner $SUSPEND then resume re-runs planner and reaches $END", {
+    timeout: 30_000,
+  }, async () => {
     await writeMockConfig("e2e-suspend.mock.yaml");
     const workflowHash = await addWorkflow("e2e-suspend.workflow.yaml", "test-suspend");
 
@@ -372,7 +378,9 @@ describe("E2E mock-agent: full uwf pipeline", () => {
     expect(finalEntry!.head).toBe(resumeOut.head);
   });
 
-  test("5. --count 3 runs the whole linear pipeline in one invocation", async () => {
+  test("5. --count 3 runs the whole linear pipeline in one invocation", {
+    timeout: 30_000,
+  }, async () => {
     await writeMockConfig("e2e-count.mock.yaml");
     const workflowHash = await addWorkflow("e2e-count.workflow.yaml", "test-count");
 
@@ -412,7 +420,9 @@ describe("E2E mock-agent: full uwf pipeline", () => {
     expect(finalEntry!.head).toBe(results[2].head);
   });
 
-  test("6. mustache edge prompt renders planner variables into the worker step", async () => {
+  test("6. mustache edge prompt renders planner variables into the worker step", {
+    timeout: 30_000,
+  }, async () => {
     await writeMockConfig("e2e-mustache.mock.yaml");
     const workflowHash = await addWorkflow("e2e-mustache.workflow.yaml", "test-mustache");
 
@@ -441,7 +451,9 @@ describe("E2E mock-agent: full uwf pipeline", () => {
     expect(workerStep.edgePrompt).toBe("Work on branch fix/42-auth in /tmp/my-repo");
   });
 
-  test("7. completed thread can be resumed (衔尾蛇: end → start)", async () => {
+  test("7. completed thread can be resumed (衔尾蛇: end → start)", {
+    timeout: 30_000,
+  }, async () => {
     // Reuse the suspend workflow (planner with ready → $END), but mock data
     // goes straight to ready on first run, then ready again after resume.
     await writeMockConfig("e2e-completed-resume.mock.yaml");

packages/cli/src/__tests__/prompt.test.ts

@@ -88,11 +88,11 @@ describe("prompt commands", () => {
     expect(result).toContain("version");
   });
 
-  test("prompt help subcommand is suppressed", () => {
+  test("prompt help subcommand is suppressed", { timeout: 30_000 }, () => {
-    const output = execFileSync("npx", ["tsx", "src/cli.ts", "prompt", "--help"], {
+    const cliPath = join(__dirname, "..", "..", "dist", "cli.js");
-      cwd: join(__dirname, "..", ".."),
+    const output = execFileSync("node", [cliPath, "prompt", "--help"], {
       encoding: "utf-8",
-      env: { ...process.env, PATH: `/opt/homebrew/bin:${process.env.PATH}` },
+      env: { ...process.env },
     });
     expect(output).not.toMatch(/help\s+\[command\]/i);
     expect(output).toContain("usage");

packages/cli/src/__tests__/store-unified-threads.test.ts

@@ -15,7 +15,7 @@ import {
 async function makeUwfStore(storageRoot: string) {
   const casDir = join(storageRoot, "cas");
   await mkdir(casDir, { recursive: true });
-  process.env.OCAS_DIR = casDir;
+  process.env.OCAS_HOME = casDir;
   return createUwfStore(storageRoot);
 }
 

packages/cli/src/__tests__/thread-resume.test.ts

@@ -491,7 +491,7 @@ describe("uwf thread resume - completed threads", () => {
       cwd: tmpDir,
     });
 
-    process.env.OCAS_DIR = casDir;
+    process.env.OCAS_HOME = casDir;
 
     const workerOutputHash = await store.cas.put(outputSchemaHash, { $status: "_" });
     const reviewerOutputHash = await store.cas.put(outputSchemaHash, { $status: "_" });
@@ -539,9 +539,7 @@ describe("uwf thread resume - completed threads", () => {
     const { createUwfStore, getThread } = await import("../store.js");
     const verifyUwf = await createUwfStore(tmpDir);
     const verifyEntry = getThread(verifyUwf.varStore, THREAD_ID);
-    // biome-ignore lint/suspicious/noConsole: test debugging
     console.log("Seeded entry status:", verifyEntry?.status);
-    // biome-ignore lint/suspicious/noConsole: test debugging
     console.log("Seeded entry:", JSON.stringify(verifyEntry, null, 2));
 
     const promptCapturePath = join(tmpDir, "captured-prompt-completed.txt");
@@ -601,7 +599,6 @@ echo '${adapterJson}'
     );
 
     if (result.status !== 0) {
-      // biome-ignore lint/suspicious/noConsole: test debugging
       console.error("Command failed:", result.stderr);
     }
 
@@ -654,7 +651,7 @@ echo '${adapterJson}'
       cwd: tmpDir,
     });
 
-    process.env.OCAS_DIR = casDir;
+    process.env.OCAS_HOME = casDir;
     await seedThreads(tmpDir, {
       [THREAD_ID]: {
         head: startHash,
@@ -702,7 +699,7 @@ echo '${adapterJson}'
       cwd: tmpDir,
     });
 
-    process.env.OCAS_DIR = casDir;
+    process.env.OCAS_HOME = casDir;
     await seedThreads(tmpDir, { [THREAD_ID]: startHash });
 
     const result = runUwf(["thread", "resume", THREAD_ID], casDir);

packages/cli/src/__tests__/thread-step-count.test.ts

@@ -3,11 +3,11 @@ import { dirname, join } from "node:path";
 import { fileURLToPath } from "node:url";
 import { describe, expect, test } from "vitest";
 
-const CLI_PATH = join(dirname(fileURLToPath(import.meta.url)), "..", "cli.js");
+const CLI_PATH = join(dirname(fileURLToPath(import.meta.url)), "..", "..", "dist", "cli.js");
 
 function runCli(args: string[]): { stdout: string; stderr: string; exitCode: number } {
   try {
-    const stdout = execFileSync("npx", ["tsx", CLI_PATH, ...args], {
+    const stdout = execFileSync("node", [CLI_PATH, ...args], {
       encoding: "utf8",
       env: { ...process.env, UWF_HOME: "/tmp/uwf-test-nonexistent" },
       stdio: ["ignore", "pipe", "pipe"],
@@ -24,19 +24,21 @@ function runCli(args: string[]): { stdout: string; stderr: string; exitCode: num
 }
 
 describe("thread exec --count CLI parsing", () => {
-  test("--help shows -c/--count option", () => {
+  test("--help shows -c/--count option", { timeout: 30_000 }, () => {
     const result = runCli(["thread", "exec", "--help"]);
-    expect(result.stdout).toContain("--count");
+    const combined = result.stdout + result.stderr;
-    expect(result.stdout).toContain("-c");
+    expect(combined).toContain("--count");
+    expect(combined).toContain("-c");
   });
 
-  test("description says 'one or more steps'", () => {
+  test("description says 'one or more steps'", { timeout: 30_000 }, () => {
     const result = runCli(["thread", "exec", "--help"]);
-    expect(result.stdout).toContain("one or more steps");
+    const combined = result.stdout + result.stderr;
+    expect(combined).toContain("one or more steps");
   });
 });
 
-describe("cmdThreadExec count logic", () => {
+describe("cmdThreadExec count logic", { timeout: 30_000 }, () => {
   test("count=0 fails with validation error", () => {
     const result = runCli(["thread", "exec", "FAKE_THREAD_ID", "-c", "0"]);
     expect(result.exitCode).not.toBe(0);