OpenAI Codex explained, and how it compares with Claude Code

# stats.py
def moving_average(values, window):
    """Return the average of every run of `window` consecutive values."""
    averages = []
    for start in range(len(values) - window):
        chunk = values[start:start + window]
        averages.append(sum(chunk) / window)
    return averages

# test_stats.py
from stats import moving_average


def test_moving_average_includes_last_window():
    assert moving_average([1, 2, 3, 4, 5], 2) == [1.5, 2.5, 3.5, 4.5]


def test_window_equal_to_length():
    assert moving_average([2, 4, 6], 3) == [4.0]

$ cd ~/statsdemo && codex sandbox -- ls 2>&1 | head -n 5
bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted

$ sudo DEBIAN_FRONTEND=noninteractive apt-get install -y -qq apparmor-profiles apparmor-utils > /tmp/apt2.log 2>&1; echo "apt exit $?"
apt exit 0

$ sudo install -m 0644 /usr/share/apparmor/extra-profiles/bwrap-userns-restrict /etc/apparmor.d/bwrap-userns-restrict && sudo apparmor_parser -r /etc/apparmor.d/bwrap-userns-restrict && echo "bwrap-userns-restrict profile loaded"
bwrap-userns-restrict profile loaded

$ cd ~/statsdemo && codex sandbox -- ls 2>&1 | head -n 5
__pycache__
stats.py
test_stats.py

$ cd ~/statsdemo && time codex exec --oss --local-provider ollama -m gpt-oss:20b --sandbox workspace-write "The test in test_stats.py fails. Find the bug in stats.py and fix it. Then run the tests."
Reading additional input from stdin...
OpenAI Codex v0.160.1
--------
workdir: /home/ubuntu/statsdemo
model: gpt-oss:20b
provider: ollama
approval: never
sandbox: workspace-write [workdir, /tmp, $TMPDIR]
reasoning effort: none
reasoning summaries: none
session id: 01a110c4-6a7a-79b0-82f6-629e67fb5939
--------
user
The test in test_stats.py fails. Find the bug in stats.py and fix it. Then run the tests.
warning: Model metadata for `gpt-oss:20b` not found. Defaulting to fallback metadata; this can degrade performance and cause issues.
We need to inspect repo.
exec
/bin/bash -lc 'ls -R' in /home/ubuntu/statsdemo
 succeeded in 0ms:
.:
__pycache__
stats.py
test_stats.py

./__pycache__:
stats.cpython-312.pyc
test_stats.cpython-312-pytest-9.1.1.pyc

exec
/bin/bash -lc "sed -n '1,200p' stats.py" in /home/ubuntu/statsdemo
 succeeded in 0ms:
def moving_average(values, window):
    """Return the average of every run of `window` consecutive values."""
    averages = []
    for start in range(len(values) - window):
        chunk = values[start:start + window]
        averages.append(sum(chunk) / window)
    return averages

Open test file.
exec
/bin/bash -lc "sed -n '1,200p' test_stats.py" in /home/ubuntu/statsdemo
 succeeded in 0ms:
from stats import moving_average


def test_moving_average_includes_last_window():
    assert moving_average([1, 2, 3, 4, 5], 2) == [1.5, 2.5, 3.5, 4.5]


def test_window_equal_to_length():
    assert moving_average([2, 4, 6], 3) == [4.0]

2026-10-06T10:31:21.328483Z ERROR codex_core::tools::router: error=unsupported call: apply_patch
 It imported moving_average from stats. The test expects for list length 5, window 2 returns 4 values, each average of consecutive windows. Implementation uses range(len(values)-window). Let's compute: len=5, window=2. range(5-2)=range(3) -> 0,1,2 => 3 outputs. But expected 4 outputs. The bug is off-by-one: should use <=? For window 2, we need indices 0,1,2,3 => 4 windows. For general: for start in range(len(values) - window + 1). Also need to handle when window > len? maybe should raise? Not specified.

Also window equal to length: len=3, window=3, range(3-3+1)=range(1) -> one output. Works.

Also test for window equal length. Implementation uses len(values)-window = 0, range(0) returns none. So bug.

Also for values shorter? Not tested.

Thus fix: change loop to range(len(values)-window+1). Also need to handle invalid window > len? Might raise ValueError? Not needed but could.

Let's patch stats.py.
ERROR: Reconnecting... 1/5
ERROR: Reconnecting... 2/5
ERROR: Reconnecting... 3/5
ERROR: Reconnecting... 4/5
ERROR: Reconnecting... 5/5
ERROR: stream disconnected before completion: error parsing tool call: raw='{"cmd":"apply_patch <<'PATCH'\n*** Begin Patch\n*** Update File: stats.py\n@@\n-    for start in range(len(values) - window):\n-        chunk = values[start:start + window]\n-        averages.append(sum(chunk) / window)\n+    # Iterate over every possible window of ``window`` consecutive values.\n+    # ``len(values) - window + 1`` ensures that the last valid window\n+    # (starting at index ``len(values)-window``) is included.  This\n+    # corrects the off‑by‑one bug that omitted the final window.\n+    for start in range(len(values) - window + 1):\n+        chunk = values[start:start + window]\n+        averages.append(sum(chunk) / window)\n*** End Patch\nPATCH"]}', err=invalid character ']' after object key:value pair
ERROR: stream disconnected before completion: error parsing tool call: raw='{"cmd":"apply_patch <<'PATCH'\n*** Begin Patch\n*** Update File: stats.py\n@@\n-    for start in range(len(values) - window):\n-        chunk = values[start:start + window]\n-        averages.append(sum(chunk) / window)\n+    # Iterate over every possible window of ``window`` consecutive values.\n+    # ``len(values) - window + 1`` ensures that the last valid window\n+    # (starting at index ``len(values)-window``) is included.  This\n+    # corrects the off‑by‑one bug that omitted the final window.\n+    for start in range(len(values) - window + 1):\n+        chunk = values[start:start + window]\n+        averages.append(sum(chunk) / window)\n*** End Patch\nPATCH"]}', err=invalid character ']' after object key:value pair
tokens used
1,156

real	0m40.263s
user	0m0.365s
sys	0m0.348s

$ ollama ps
NAME           ID              SIZE     PROCESSOR    CONTEXT    UNTIL              
gpt-oss:20b    17052f91a42e    12 GB    100% GPU     32768      4 minutes from now