🤖 Block standalone sleep commands in bash tool (#262)

Blocks `sleep` at the start of bash commands to prevent agents from
wasting time waiting. Agents are instructed via error message to use
polling loops instead (`while ! condition; do sleep 1; done`).

**Pattern:** `/^\s*sleep\s/` - Only blocks sleep at the beginning, not
in loops or after other commands.

**Examples:**
- Blocked: `sleep 5`
- Allowed: `while ! test -f done; do sleep 1; done`
- Allowed: `echo start; sleep 2`

Tests verify blocking behavior and confirm sleep works correctly in
polling loops.

_Generated with `cmux`_

Author: ammar-agent

src/services/tools/bash.test.ts

@@ -268,7 +268,7 @@ describe("bash tool", () => {
     using testEnv = createTestBashTool();
     const tool = testEnv.tool;
     const args: BashToolArgs = {
-      script: "sleep 10",
+      script: "while true; do sleep 0.1; done",
       timeout_secs: 1,
     };
 
@@ -507,15 +507,15 @@ describe("bash tool", () => {
 
     const args: BashToolArgs = {
       // Background process that would block if we waited for it
-      script: "sleep 100 > /dev/null 2>&1 &",
+      script: "while true; do sleep 1; done > /dev/null 2>&1 &",
       timeout_secs: 5,
     };
 
     const result = (await tool.execute!(args, mockToolCallOptions)) as BashToolResult;
     const duration = performance.now() - startTime;
 
     expect(result.success).toBe(true);
-    // Should complete in well under 1 second, not wait for sleep 100
+    // Should complete in well under 1 second, not wait for infinite loop
     expect(duration).toBeLessThan(2000);
   });
 
@@ -527,7 +527,7 @@ describe("bash tool", () => {
     const args: BashToolArgs = {
       // Spawn background process, echo its PID, then exit
       // Should not wait for the background process
-      script: "sleep 100 > /dev/null 2>&1 & echo $!",
+      script: "while true; do sleep 1; done > /dev/null 2>&1 & echo $!",
       timeout_secs: 5,
     };
 
@@ -550,7 +550,7 @@ describe("bash tool", () => {
 
     const args: BashToolArgs = {
       // Background process with output redirected but still blocking
-      script: "sleep 10 & wait",
+      script: "while true; do sleep 0.1; done & wait",
       timeout_secs: 1,
     };
 
@@ -655,6 +655,44 @@ describe("bash tool", () => {
     }
   });
 
+  it("should block sleep command at start of script", async () => {
+    using testEnv = createTestBashTool();
+    const tool = testEnv.tool;
+    const args: BashToolArgs = {
+      script: "sleep 5",
+      timeout_secs: 10,
+    };
+
+    const result = (await tool.execute!(args, mockToolCallOptions)) as BashToolResult;
+
+    expect(result.success).toBe(false);
+    if (!result.success) {
+      expect(result.error).toContain("sleep commands are blocked");
+      expect(result.error).toContain("polling loops");
+      expect(result.error).toContain("while ! condition");
+      expect(result.exitCode).toBe(-1);
+      expect(result.wall_duration_ms).toBe(0);
+    }
+  });
+
+  it("should allow sleep in polling loops", async () => {
+    using testEnv = createTestBashTool();
+    const tool = testEnv.tool;
+    const args: BashToolArgs = {
+      script: "for i in 1 2 3; do echo $i; sleep 0.1; done",
+      timeout_secs: 5,
+    };
+
+    const result = (await tool.execute!(args, mockToolCallOptions)) as BashToolResult;
+
+    expect(result.success).toBe(true);
+    if (result.success) {
+      expect(result.output).toContain("1");
+      expect(result.output).toContain("2");
+      expect(result.output).toContain("3");
+    }
+  });
+
   it("should use default timeout (3s) when timeout_secs is undefined", async () => {
     using testEnv = createTestBashTool();
     const tool = testEnv.tool;

src/services/tools/bash.ts

@@ -52,6 +52,17 @@ export const createBashTool: ToolFactory = (config: ToolConfiguration) => {
         };
       }
 
+      // Block sleep at the beginning of commands - they waste time waiting. Use polling loops instead.
+      if (/^\s*sleep\s/.test(script)) {
+        return {
+          success: false,
+          error:
+            "sleep commands are blocked to minimize waiting time. Instead, use polling loops to check conditions repeatedly (e.g., 'while ! condition; do sleep 1; done' or 'until condition; do sleep 1; done').",
+          exitCode: -1,
+          wall_duration_ms: 0,
+        };
+      }
+
       // Default timeout to 3 seconds for interactivity
       // OpenAI models often don't provide timeout_secs even when marked required,
       // so we make it optional with a sensible default.

tests/ipcMain/executeBash.test.ts

@@ -143,7 +143,7 @@ describeIntegration("IpcMain executeBash integration tests", () => {
         const timeoutResult = await env.mockIpcRenderer.invoke(
           IPC_CHANNELS.WORKSPACE_EXECUTE_BASH,
           workspaceId,
-          "sleep 10",
+          "while true; do sleep 0.1; done",
           { timeout_secs: 1 }
         );
 

tests/ipcMain/renameWorkspace.test.ts

@@ -247,7 +247,7 @@ describeIntegration("IpcMain rename workspace integration tests", () => {
         void sendMessageWithModel(
           env.mockIpcRenderer,
           workspaceId,
-          "Run this bash command: sleep 30 && echo done"
+          "Run this bash command: for i in {1..60}; do sleep 0.5; done && echo done"
         );
 
         // Wait for stream to start

tests/ipcMain/resumeStream.test.ts

@@ -35,7 +35,7 @@ describeIntegration("IpcMain resumeStream integration tests", () => {
         void sendMessageWithModel(
           env.mockIpcRenderer,
           workspaceId,
-          `Run this bash command: sleep 5 && echo '${expectedWord}'`,
+          `Run this bash command: for i in 1 2 3; do sleep 0.5; done && echo '${expectedWord}'`,
           "anthropic",
           "claude-sonnet-4-5"
         );

tests/ipcMain/sendMessage.test.ts

@@ -88,7 +88,7 @@ describeIntegration("IpcMain sendMessage integration tests", () => {
         const { env, workspaceId, cleanup } = await setupWorkspace(provider);
         try {
           // Start a long-running stream with a bash command that takes time
-          const longMessage = "Run this bash command: sleep 60 && echo done";
+          const longMessage = "Run this bash command: while true; do sleep 1; done";
           void sendMessageWithModel(env.mockIpcRenderer, workspaceId, longMessage, provider, model);
 
           // Wait for stream to start
@@ -263,11 +263,11 @@ describeIntegration("IpcMain sendMessage integration tests", () => {
 
         const { env, workspaceId, cleanup } = await setupWorkspace(provider);
         try {
-          // Start a stream with tool call that takes 10 seconds
+          // Start a stream with tool call that takes a long time
           void sendMessageWithModel(
             env.mockIpcRenderer,
             workspaceId,
-            "Run this bash command: sleep 10",
+            "Run this bash command: while true; do sleep 0.1; done",
             provider,
             model
           );
@@ -279,7 +279,7 @@ describeIntegration("IpcMain sendMessage integration tests", () => {
 
           await collector1.waitForEvent("tool-call-start", 10000);
 
-          // At this point, bash sleep is running (will take 10 seconds if abort doesn't work)
+          // At this point, bash loop is running (will run forever if abort doesn't work)
           // Get message ID for verification
           collector1.collect();
           const messageId =
@@ -344,8 +344,8 @@ describeIntegration("IpcMain sendMessage integration tests", () => {
             expect(partialMessages.length).toBe(0);
           }
 
-          // Note: If test completes quickly (~5s), abort signal worked and killed sleep
-          // If test takes ~10s, abort signal didn't work and sleep ran to completion
+          // Note: If test completes quickly (~5s), abort signal worked and killed the loop
+          // If test takes much longer, abort signal didn't work
         } finally {
           await cleanup();
         }
@@ -448,7 +448,7 @@ describeIntegration("IpcMain sendMessage integration tests", () => {
           const result1 = await sendMessageWithModel(
             env.mockIpcRenderer,
             workspaceId,
-            "Run this bash command: sleep 10 && echo done",
+            "Run this bash command: for i in {1..20}; do sleep 0.5; done && echo done",
             provider,
             model
           );
@@ -467,7 +467,7 @@ describeIntegration("IpcMain sendMessage integration tests", () => {
           const result2 = await sendMessageWithModel(
             env.mockIpcRenderer,
             workspaceId,
-            "Run this bash command: sleep 5 && echo second",
+            "Run this bash command: for i in {1..10}; do sleep 0.5; done && echo second",
             provider,
             model,
             { editMessageId: (firstUserMessage as { id: string }).id }

tests/ipcMain/truncate.test.ts

@@ -261,7 +261,7 @@ describeIntegration("IpcMain truncate integration tests", () => {
         void sendMessageWithModel(
           env.mockIpcRenderer,
           workspaceId,
-          "Run this bash command: sleep 30 && echo done"
+          "Run this bash command: for i in {1..60}; do sleep 0.5; done && echo done"
         );
 
         // Wait for stream to start