[Bug]: retry_request misses ConnectTimeout/WriteTimeout and crashes on non-numeric Retry-After

[tejajakarulloh]

Description

retry_request in src/agentos/channels/_util.py:297 catches (httpx.ConnectError, httpx.ReadTimeout), but httpx.ConnectTimeout (DNS/TLS handshake timeout) is not a subclass of httpx.ConnectError — it inherits from httpx.TimeoutException via a completely separate hierarchy:

• ConnectError → NetworkError → TransportError
• ConnectTimeout → TimeoutException → TransportError

When a channel API call (Slack, Discord, Telegram, webhook) times out during DNS resolution or TLS negotiation, the exception passes through retry_request without any retry/backoff — the caller gets an unhandled crash instead of a resilient retry.

Similarly, WriteTimeout and PoolTimeout (both TimeoutException subclasses) are also not caught.

Additionally, the Retry-After header is parsed via a bare float() call (_util.py:287). Per RFC 7231 §7.1.3, Retry-After may contain an HTTP-date string (e.g. Wed, 21 Oct 2026 07:28:00 GMT) instead of delay-seconds. A non-numeric value causes a ValueError crash in the retry loop.

Steps to Reproduce

import httpx
# ConnectTimeout is NOT caught by (ConnectError, ReadTimeout)
try:
    raise httpx.ConnectTimeout("dns timed out")
except (httpx.ConnectError, httpx.ReadTimeout):
    print("caught")  # never reached
except httpx.TimeoutException:
    print("missed by retry_request")  # this fires

Expected Behavior

All transient timeout types should be retried with exponential backoff, and non-numeric Retry-After values should fall back to the standard backoff delay instead of crashing.

Environment

• Python 3.12
• httpx (current pinned version)

[andreapn]

Accepted, p2. Both defects confirmed at src/agentos/channels/_util.py.

Timeout hierarchy. The handler is except (httpx.ConnectError, httpx.ReadTimeout). The analysis of the class tree is correct — ConnectTimeout, WriteTimeout and PoolTimeout descend from TimeoutException, not from ConnectError, so none of them reach the backoff path.

Retry-After. float(resp.headers.get("Retry-After", base_delay * (2**attempt))) is a bare parse of a caller-controlled header. RFC 7231 §7.1.3 permits an HTTP-date, and a server sending one turns a rate-limit into a ValueError inside the retry loop.

Catching httpx.TimeoutException alongside ConnectError covers all four timeout types in one clause, and Retry-After should fall back to the computed backoff when it does not parse as a number (a sane upper clamp would also be welcome — a server can otherwise ask us to sleep for hours).

Note #599 reports a third defect in this same function (the 429 branch not returning the response when retries are exhausted). Please coordinate so all three land coherently.

Two PRs target this (#645, #643) — we will pick one.