default_retry_on inverts requests error handling: 4xx are retried, connection errors and timeouts are not

[nazsats]

Checked other resources

• This is a bug, not a usage question.
• I added a clear and descriptive title that summarizes this issue.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangGraph rather than my code.
• The bug is not resolved by updating to the latest stable version of LangGraph.
• This is not related to the langchain-community package.
• I posted a self-contained, minimal, reproducible example.

Related Issues / PRs

• fix(langgraph): make NodeTimeoutError retryable by default #7659 previously modified default_retry_on (made NodeTimeoutError retryable by default).
• fix(langgraph): classify requests exceptions correctly in default_retry_on #8801 — I opened a PR with the fix and regression tests before filing this issue, and it was
  automatically closed for not linking an approved issue. The branch is still up to date with main.
  I can add Fixes #<this issue> to reopen it as soon as a maintainer approves and assigns.

Reproduction Steps / Example Code (Python)

import requests
from requests.models import Response
from typing_extensions import TypedDict

from langgraph.graph import START, StateGraph
from langgraph.types import RetryPolicy


class State(TypedDict):
    n: int


def make_graph(exc_factory):
    attempts = {"count": 0}

    def node(state: State) -> State:
        attempts["count"] += 1
        raise exc_factory()

    g = StateGraph(State)
    g.add_node(
        "call", node, retry_policy=RetryPolicy(initial_interval=0.01, max_attempts=3)
    )
    g.add_edge(START, "call")
    return g.compile(), attempts


def http_404():
    r = Response()
    r.status_code = 404
    return requests.HTTPError("404 Not Found", response=r)


app, attempts = make_graph(http_404)
try:
    app.invoke({"n": 0})
except Exception:
    pass
print(f"requests.HTTPError(404)  -> node ran {attempts['count']} time(s); expected 1")

app, attempts = make_graph(lambda: requests.ConnectionError("connection refused"))
try:
    app.invoke({"n": 0})
except Exception:
    pass
print(f"requests.ConnectionError -> node ran {attempts['count']} time(s); expected 3")

Output:

requests.HTTPError(404)  -> node ran 3 time(s); expected 1
requests.ConnectionError -> node ran 1 time(s); expected 3

Description

default_retry_on — the default value of RetryPolicy.retry_on — misclassifies requests
exceptions, and the two errors are exact inverses of each other: permanent failures are
retried, and transient failures are not.

1. Every requests.HTTPError is retried, including 4xx.

return 500 <= exc.response.status_code < 600 if exc.response else True

requests.Response.__bool__ is an alias for Response.ok, so every error response is falsy:

>>> r = requests.models.Response(); r.status_code = 404
>>> bool(r)
False

The if exc.response guard therefore always takes the else True branch, which makes the
500 <= status < 600 comparison unreachable. A 401, 404 or 422 is retried up to
max_attempts with backoff, re-sending a request that cannot succeed against an API that has
already refused it.

2. requests connection errors and timeouts are never retried.

requests.RequestException subclasses OSError, which is in the non-retryable list, so
requests.ConnectionError, ConnectTimeout and ReadTimeout are treated as permanent — even
though the builtin ConnectionError on the first line of the function retries, and an
unrecognised exception falls through to return True.

>>> isinstance(requests.exceptions.ConnectionError(), ConnectionError)  # builtin
False
>>> isinstance(requests.exceptions.ConnectionError(), OSError)
True

Expected: 4xx is not retried; 5xx, connection errors and timeouts are.
Actual: 4xx is retried; connection errors and timeouts are not.

Note on test coverage. test_should_retry_default_retry_on already asserts the intended
behaviour, but builds the response with Mock(). A bare Mock is truthy, so the test reaches
a code path production never reaches, and passes.

I have a fix with regression tests (using a real requests.models.Response) — 9 new test cases fail
on main and pass with the change, with no regressions. The branch is pushed and was opened as
#8801, which the bot closed for not linking an approved issue. Happy to relink it here once a
maintainer approves and assigns this.

System Info

System Information
------------------
> OS:  Linux
> OS Version:  #30~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Fri Aug  7 13:27:52 UTC 2
> Python Version:  3.12.3 (main, Jun 19 2026, 12:46:00) [GCC 13.3.0]

Package Information
-------------------
> langgraph: 1.2.11 (main @ c0a13bb)
> langchain_core: 1.6.1
> langsmith: 0.12.1
> requests: 2.31.0
> httpx: 0.28.1