by Emery Berger, Stephen Freund, Kyla Levin, Nicolas van Kempen (ordered alphabetically)

ChatDBG is an AI-based debugging assistant for C/C /Python/Rust code that integrates large language models into a standard debugger (pdb, lldb, gdb, and windbg) to help debug your code. With ChatDBG, you can engage in a dialog with your debugger, asking open-ended questions about your program, like why is x null?. ChatDBG will take the wheel and steer the debugger to answer your queries. ChatDBG can provide error diagnoses and suggest fixes.

As far as we are aware, ChatDBG is the first debugger to automatically perform root cause analysis and to provide suggested fixes.

Watch ChatDBG in action!

LLDB on test-overflow.cpp	GDB on test-overflow.cpp	Pdb on bootstrap.py

For technical details and a complete evaluation, see our arXiv paper, ChatDBG: An AI-Powered Debugging Assistant (PDF).

export OPENAI_API_KEY=<your-api-key>

Install ChatDBG using pip (you need to do this whether you are debugging Python, C, or C code):

python3 -m pip install chatdbg

If you are using ChatDBG to debug Python programs, you are done. If you want to use ChatDBG to debug native code with gdb or lldb, follow the installation instructions below.

python3 -m pip install ChatDBG
python3 -c 'import chatdbg; print(f"command script import {chatdbg.__path__[0]}/chatdbg_lldb.py")' >> ~/.lldbinit

sudo apt install -y lsb-release wget software-properties-common gnupg
curl -sSf https://apt.llvm.org/llvm.sh | sudo bash -s -- 18 all
# LLDB now available as `lldb-18`.

xcrun python3 -m pip install ChatDBG
xcrun python3 -c 'import chatdbg; print(f"command script import {chatdbg.__path__[0]}/chatdbg_lldb.py")' >> ~/.lldbinit

python3 -m pip install ChatDBG
python3 -c 'import chatdbg; print(f"source {chatdbg.__path__[0]}/chatdbg_gdb.py")' >> ~/.gdbinit

To use ChatDBG to debug Python programs, simply run your Python script as follows:

chatdbg -c continue yourscript.py

ChatDBG is an extension of the standard Python debugger pdb. Like pdb, when your script encounters an uncaught exception, ChatDBG will enter post mortem debugging mode.

Unlike other debuggers, you can then use the why command to ask ChatDBG why your program failed and get a suggested fix. After the LLM responds, you may issue additional debugging commands or continue the conversation by entering any other text.

To use ChatDBG as the default debugger for IPython or inside Jupyter Notebooks, create a IPython profile and then add the necessary exensions on startup. (Modify these lines as necessary if you already have a customized profile file.)

ipython profile create
echo "c.InteractiveShellApp.extensions = ['chatdbg.chatdbg_pdb', 'ipyflow']" >> ~/.ipython/profile_default/ipython_config.py

On the command line, you can then run:

ipython --pdb yourscript.py

Inside Jupyter, run your notebook with the ipyflow kernel and include this line magic at the top of the file.

%pdb

To use ChatDBG with lldb or gdb, just run native code (compiled with -g for debugging symbols) with your choice of debugger; when it crashes, ask why. This also works for post mortem debugging (when you load a core with the -c option).

The native debuggers work slightly differently than Pdb. After the debugger responds to your question, you will enter into ChatDBG's command loop, as indicated by the (ChatDBG chatting) prompt. You may continue issuing debugging commands and you may send additional messages to the LLM by starting those messages with "chat". When you are done, type quit to return to the debugger's main command loop.

[dependencies]
chatdbg = "0.6.2"

[profile.dev]
panic = "abort"

[profile.release]
panic = "abort"

#[chatdbg::main]
fn main() {

(ChatDBG lldb) run
Process 85494 launched: '/Users/emery/git/ChatDBG/test/a.out' (arm64)
TEST 1
TEST -422761288
TEST 0
TEST 0
TEST 0
TEST 0
TEST 0
TEST 0
Process 85494 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x100056200)
    frame #0: 0x0000000100002f64 a.out`foo(n=8, b=1) at test.cpp:7:22
   4     int x[] = { 1, 2, 3, 4, 5 };
   5     
   6     void foo(int n, float b) {
-> 7       cout << "TEST " << x[n * 10000] << endl;
   8     }
   9     
   10    int main()
Target 0: (a.out) stopped.

(ChatDBG lldb) why
The root cause of this error is accessing an index of the array `x`
that is out of bounds. In `foo()`, the index is calculated as `n *
10000`, which can be much larger than the size of the array `x` (which
is only 5 elements). In the given trace, the program is trying to
access the memory address `0x100056200`, which is outside of the range
of allocated memory for the array `x`.

To fix this error, we need to ensure that the index is within the
bounds of the array. One way to do this is to check the value of `n`
before calculating the index, and ensure that it is less than the size
of the array divided by the size of the element. For example, we can
modify `foo()` as follows:

    ```
    void foo(int n, float b) {
      if (n < 0 || n >= sizeof(x)/sizeof(x[0])) {
        cout << "ERROR: Invalid index" << endl;
        return;
      }
      cout << "TEST " << x[n] << endl;
    }
    ```

This code checks if `n` is within the valid range, and prints an error
message if it is not. If `n` is within the range, the function prints
the value of the element at index `n` of `x`. With this modification,
the program will avoid accessing memory outside the bounds of the
array, and will print the expected output for valid indices.

Traceback (most recent call last):
  File "yourscript.py", line 9, in <module>
    print(tryme(100))
  File "yourscript.py", line 4, in tryme
    if x / i > 2:
ZeroDivisionError: division by zero
Uncaught exception. Entering post mortem debugging
Running 'cont' or 'step' will restart the program
> yourscript.py(4)tryme()
-> if x / i > 2:

(ChatDBG Pdb) why
The root cause of the error is that the code is attempting to
divide by zero in the line "if x / i > 2". As i ranges from 0 to 99,
it will eventually reach the value of 0, causing a ZeroDivisionError.

A possible fix for this would be to add a check for i being equal to
zero before performing the division. This could be done by adding an
additional conditional statement, such as "if i == 0: continue", to
skip over the iteration when i is zero. The updated code would look
like this:

def tryme(x):
    count = 0
    for i in range(100):
        if i == 0:
            continue
        if x / i > 2:
            count  = 1
    return count

if __name__ == '__main__':
    print(tryme(100))