fix: enhance shell() to know when it is interactive (#66)

Co-authored-by: Bradley Axen <baxen@squareup.com>

src/goose/toolkit/developer.py

@@ -1,10 +1,15 @@
-from pathlib import Path
-from subprocess import CompletedProcess, run
-from typing import List, Dict
 import os
-from goose.utils.check_shell_command import is_dangerous_command
+import re
+import subprocess
+import time
+from pathlib import Path
+from typing import Dict, List
 
 from exchange import Message
+from goose.toolkit.base import Toolkit, tool
+from goose.toolkit.utils import get_language, render_template
+from goose.utils.ask import ask_an_ai
+from goose.utils.check_shell_command import is_dangerous_command
 from rich import box
 from rich.markdown import Markdown
 from rich.panel import Panel
@@ -12,9 +17,6 @@ from rich.prompt import Confirm
 from rich.table import Table
 from rich.text import Text
 
-from goose.toolkit.base import Toolkit, tool
-from goose.toolkit.utils import get_language, render_template
-
 
 def keep_unsafe_command_prompt(command: str) -> bool:
     command_text = Text(command, style="bold red")
@@ -136,7 +138,7 @@ class Developer(Toolkit):
     @tool
     def shell(self, command: str) -> str:
         """
-        Execute a command on the shell (in OSX)
+        Execute a command on the shell
 
         This will return the output and error concatenated into a single string, as
         you would see from running on the command line. There will also be an indication
@@ -146,11 +148,7 @@ class Developer(Toolkit):
             command (str): The shell command to run. It can support multiline statements
                 if you need to run more than one at a time
         """
-        self.notifier.status("planning to run shell command")
         # Log the command being executed in a visually structured format (Markdown).
-        # The `.log` method is used here to log the command execution in the application's UX
-        # this method is dynamically attached to functions in the Goose framework to handle user-visible
-        # logging and integrates with the overall UI logging system
         self.notifier.log(Panel.fit(Markdown(f"```bash\n{command}\n```"), title="shell"))
 
         if is_dangerous_command(command):
@@ -159,16 +157,87 @@ class Developer(Toolkit):
             if not keep_unsafe_command_prompt(command):
                 raise RuntimeError(
                     f"The command {command} was rejected as dangerous by the user."
-                    + " Do not proceed further, instead ask for instructions."
+                    " Do not proceed further, instead ask for instructions."
                 )
             self.notifier.start()
         self.notifier.status("running shell command")
-        result: CompletedProcess = run(command, shell=True, text=True, capture_output=True, check=False)
+
-        if result.returncode == 0:
+        # Define patterns that might indicate the process is waiting for input
-            output = "Command succeeded"
+        interaction_patterns = [
+            r"Do you want to",  # Common prompt phrase
+            r"Enter password",  # Password prompt
+            r"Are you sure",  # Confirmation prompt
+            r"\(y/N\)",  # Yes/No prompt
+            r"Press any key to continue",  # Awaiting keypress
+            r"Waiting for input",  # General waiting message
+            r"\?\s",  # Prompts starting with '? '
+        ]
+        compiled_patterns = [re.compile(pattern, re.IGNORECASE) for pattern in interaction_patterns]
+
+        proc = subprocess.Popen(
+            command,
+            shell=True,
+            stdin=subprocess.DEVNULL,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+            text=True,
+        )
+        # this enables us to read lines without blocking
+        os.set_blocking(proc.stdout.fileno(), False)
+
+        # Accumulate the output logs while checking if it might be blocked
+        output_lines = []
+        last_output_time = time.time()
+        cutoff = 10
+        while proc.poll() is None:
+            self.notifier.status("running shell command")
+            line = proc.stdout.readline()
+            if line:
+                output_lines.append(line)
+                last_output_time = time.time()
+
+            # If we see a clear pattern match, we plan to abort
+            exit_criteria = any(pattern.search(line) for pattern in compiled_patterns)
+
+            # and if we haven't seen a new line in 10+s, check with AI to see if it may be stuck
+            if not exit_criteria and time.time() - last_output_time > cutoff:
+                self.notifier.status("checking on shell status")
+                response = ask_an_ai(
+                    input="\n".join([command] + output_lines),
+                    prompt=(
+                        "You will evaluate the output of shell commands to see if they may be stuck."
+                        " Look for commands that appear to be awaiting user input, or otherwise running indefinitely (such as a web service)."  # noqa
+                        " A command that will take a while, such as downloading resources is okay."  # noqa
+                        " return [Yes] if stuck, [No] otherwise."
+                    ),
+                    exchange=self.exchange_view.processor,
+                    with_tools=False,
+                )
+                exit_criteria = "[yes]" in response.content[0].text.lower()
+                # We add exponential backoff for how often we check for the command being stuck
+                cutoff *= 10
+
+            if exit_criteria:
+                proc.terminate()
+                raise ValueError(
+                    f"The command `{command}` looks like it will run indefinitely or is otherwise stuck."
+                    f"You may be able to specify inputs if it applies to this command."
+                    f"Otherwise to enable continued iteration, you'll need to ask the user to run this command in another terminal."  # noqa
+                )
+
+        # read any remaining lines
+        while line := proc.stdout.readline():
+            output_lines.append(line)
+        output = "".join(output_lines)
+
+        # Determine the result based on the return code
+        if proc.returncode == 0:
+            result = "Command succeeded"
         else:
-            output = f"Command failed with returncode {result.returncode}"
+            result = f"Command failed with returncode {proc.returncode}"
-        return "\n".join([output, result.stdout, result.stderr])
+
+        # Return the combined result and outputs if we made it this far
+        return "\n".join([result, output])
 
     @tool
     def write_file(self, path: str, content: str) -> str:

src/goose/utils/ask.py

@@ -1,7 +1,13 @@
 from exchange import Exchange, Message, CheckpointData
 
 
-def ask_an_ai(input: str, exchange: Exchange, prompt: str = "", no_history: bool = True) -> Message:
+def ask_an_ai(
+    input: str,
+    exchange: Exchange,
+    prompt: str = "",
+    no_history: bool = True,
+    with_tools: bool = True,
+) -> Message:
     """Sends a separate message to an LLM using a separate Exchange than the one underlying the Goose session.
 
     Can be used to summarize a file, or submit any other request that you'd like to an AI. The Exchange can have a
@@ -36,6 +42,9 @@ def ask_an_ai(input: str, exchange: Exchange, prompt: str = "", no_history: bool
     if no_history:
         exchange = clear_exchange(exchange)
 
+    if not with_tools:
+        exchange = exchange.replace(tools=())
+
     if prompt:
         exchange = replace_prompt(exchange, prompt)