Errors: reported, not raised

An exception in the sandbox's code does not raise in your program. It arrives as an error message and becomes execution.error, which your code must check.

Example

import json

from e2b_code_interpreter.models import Execution, parse_output

execution = Execution()
for message in [
    {"type": "stdout", "text": "starting\n", "timestamp": 1},
    {"type": "error", "name": "ZeroDivisionError", "value": "division by zero",
     "traceback": "Traceback (most recent call last):\n  Cell In[1], line 2\nZeroDivisionError: division by zero"},
]:
    parse_output(execution, json.dumps(message))

Example

print(execution.error.name, "|", execution.error.value)
print(execution.error.traceback.splitlines()[-1])
print(execution.logs.stdout)
print(execution.text)

The cell printed starting, then failed. execution.error has the exception's name, value and the traceback as text; there is no main result. run_code returned normally: the SDK's parser records an error message and moves on.

What does raise

Situation	In your program
The code raises inside the sandbox	No exception; `execution.error` is set
The code runs longer than `timeout`	`TimeoutException`, from the SDK's `httpx.ReadTimeout` handling
The sandbox was killed or expired	An API error from the HTTP response
No key, or an invalid key	`AuthenticationException`

For an agent this split is right: a model should read a KeyError and fix its code, but your program should stop on a sandbox that no longer exists. timeout=0 in run_code removes the time limit, which an agent should never do.

Try it yourself

Print execution.error.to_json().
Find format_execution_timeout_error in the SDK and read its message.
Write an if that returns the error text when there is one and the result otherwise.

Previousrun_code and the Execution object

This is what real progress feels like.