fix: no sleep on exhausted final 429 attempt; correct errors.py doc (v0.1.6)

A 429 retry_after was slept immediately, including on the final attempt
that goes on to exhaust max_retries — a pointless wait right before
giving up. The wait is now carried in a per-call mutable cell and slept
at the start of the next attempt instead, so it's honored before every
attempt that actually runs and never after the last one.

errors.py's docstring claimed the exported error types cover the
missing-[discord]-extra raise path; that raise is a plain RuntimeError,
not one of these types. Reworded to match.

Compressed essay-length docstrings/comments across sender.py; no
behavior change. Verified against the aioproxies twin: aiowebhooks'
proxy-key normalization already routes zero-padded ports through
urlsplit().port (parses to int, no zero-pad on render), so host:080
and host:80 already collapse to one canonical key — no code change
needed there.

Signed-off-by: disqualifier <dev@disqualifier.me>

.gitignore

@@ -1,5 +1,5 @@
 # claude
-CLAUDE.md
+.claude/
 
 # python
 __pycache__/

README.md

@@ -13,15 +13,17 @@ send to the core — inheriting rotation, proxy, retry, and result for free.
 ## Install
 
 ```
-aiowebhooks @ git+ssh://git@git.rethinkstudios.io/rethink-public/aiowebhooks.git@v0.1.2
+aiowebhooks @ git+ssh://git@git.rethinkstudios.io/rethink-public/aiowebhooks.git@v0.1.6
 # discord embeds / identity helpers need the extra:
-aiowebhooks[discord] @ git+ssh://git@git.rethinkstudios.io/rethink-public/aiowebhooks.git@v0.1.2
+aiowebhooks[discord] @ git+ssh://git@git.rethinkstudios.io/rethink-public/aiowebhooks.git@v0.1.6
 ```
 
 The base pulls `aiohttp` and `commons` (for the retry/backoff engine). Only
 `aiowebhooks[discord]` adds `discord.py` (>=2.3, mainline — not discord.py-self), and
 only for `DiscordWebhook`.
 
+Drop the `@v0.1.6` suffix from the line above to install the latest unpinned.
+
 ## Core sender
 
 ```python
@@ -70,9 +72,17 @@ result.proxy     # canonical proxy string used (host:port:user:pass / host:port)
 
 Status retries run through `commons.aretry` (exponential backoff + cap):
 
-- **429** — honors the `retry_after` from the body first (Discord sends seconds), then
-  the `Retry-After` header, sleeping that exact value; capped by `max_retries`.
+- **429** — always retried, capped by `max_retries`. When a wait is parseable (body
+  `retry_after` first — Discord sends seconds — then the `Retry-After` header) it sleeps
+  that value before retrying; a 429 with no parseable wait (edge/Cloudflare/generic
+  webhook) still retries under aretry's backoff rather than failing one-shot. A
+  non-finite value (`inf`/`nan`) is rejected outright and a finite value is clamped to
+  `MAX_RETRY_AFTER` (300s) — a server-controlled wait can never stall a `send()` past
+  that ceiling, however large or malformed the value it sends.
 - **5xx** — retried with exponential backoff, capped by `max_retries`.
+- **connection/timeout errors** — retried with exponential backoff, capped by
+  `max_retries`, same as a 5xx (with no proxy provider; see below for the
+  proxy-rotation path).
 - **4xx** (other than 429) — fails immediately (no retry), returned as `ok=False`.
 
 Exceeding a cap returns a failed result rather than looping — and the result carries the
@@ -100,7 +110,8 @@ On a timeout/connection error the current proxy is burned and the next is tried,
 `max_proxy_retries`. Hitting the cap, or **any exception from the provider's
 `get()`/`burn()`** (the provider is duck-typed and never imported, so its exception
 types can't be caught by class), returns a failed result — never an infinite loop, never
-an escape. With no provider, a timeout just fails after normal retry.
+an escape. With no provider, a timeout/connection error is retried directly under
+`max_retries` (the same cap and backoff a 5xx gets) instead of burn+rotate.
 
 ## Discord (`aiowebhooks[discord]`)
 
@@ -137,6 +148,46 @@ Without the extra installed, importing `aiowebhooks` still works; constructing o
 
 ## Changelog
 
+### v0.1.6
+
+- **429 `retry_after` no longer sleeps on an exhausted final attempt:** the wait is
+  now carried to the START of the next attempt instead of slept immediately after
+  seeing the 429. Previously the last (exhausted) attempt slept the full
+  `retry_after` before giving up — a pointless wait since no retry followed. The
+  wait is still honored (additive with aretry's backoff) before every attempt that
+  actually runs.
+- Docs: `errors.py` no longer implies the missing-`[discord]`-extra raise is one of
+  the exported error types — it's a plain `RuntimeError`.
+- Docstrings/comments compressed; no behavior change.
+
+### v0.1.5
+
+- **429 `retry_after` bounded:** a non-finite server-controlled wait (`inf`/`nan`, from
+  the body `retry_after` or the `Retry-After` header) is now rejected outright, and a
+  finite wait is clamped to `MAX_RETRY_AFTER` (300s). Previously a bare `float()` parse
+  slept the value verbatim, unbounded and outside `max_retries` accounting — an
+  adversarial or ms-vs-s-misconfigured server could stall a `send()` for hours.
+- **Connection/timeout errors now retry without a proxy provider:** previously
+  `aiohttp.ClientError`/`asyncio.TimeoutError` with no `proxies=` set failed one-shot,
+  contradicting both the README and the single most retry-worthy failure class. Now it
+  retries under `commons.aretry`'s backoff, capped by `max_retries`, same as a 5xx —
+  and still returns `ok=False` (never raises) once retries are exhausted. Proxy-rotation
+  behavior (burn + rotate, capped by `max_proxy_retries`) is unchanged when a provider
+  is set.
+
+### v0.1.4
+
+- **Never-raises net widened:** an unexpected exception that escapes a send attempt (a
+  closed injected session → `RuntimeError`, a malformed proxy URL → `ValueError`) now
+  converts to a falsy `WebhookResult(ok=False, ...)` instead of propagating out of
+  `send()`, restoring the documented contract for those edge triggers.
+
+### v0.1.3
+
+- **429 always retries:** every `429` is now retryable under aretry's backoff + cap, not
+  only those with a parseable `retry_after`. A 429 with no body `retry_after` and no
+  `Retry-After` header (edge/Cloudflare/generic webhook) previously failed one-shot.
+
 ### v0.1.2
 
 - Removed a dead `clock` constructor param (it was stored but never used). Pinned
@@ -155,4 +206,4 @@ Without the extra installed, importing `aiowebhooks` still works; constructing o
 
 ## Versioning
 
-Tagged `vX.Y.Z`. Pin the tag.
+Releases are tagged `vX.Y.Z`. The install line above pins a release; drop the `@vX.Y.Z` suffix to install the latest unpinned. Pin deliberately for reproducible installs.

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "aiowebhooks"
-version = "0.1.2"
+version = "0.1.6"
 description = "async webhook sender (aiohttp) with round-robin urls, retry, and proxy rotation; optional discord.py embeds"
 requires-python = ">=3.10"
 dependencies = [

src/aiowebhooks/__init__.py

@@ -21,4 +21,4 @@ from .sender import Webhook
 
 __all__ = ["Webhook", "WebhookResult", "WebhookError", "NoUrlsError"]
 
-__version__ = "0.1.2"
+__version__ = "0.1.6"

src/aiowebhooks/errors.py

@@ -3,8 +3,9 @@
 these are surfaced for callers that want to branch on a specific failure cause.
 note the core `Webhook.send` does NOT raise on a send failure — it returns a
 `WebhookResult` with `ok=False` and the cause captured in `error`. these types
-exist for the few raise paths (bad construction, missing extra) and as a base for
-any future raising surface.
+cover bad construction (e.g. `NoUrlsError`) and serve as a base for any future
+raising surface; the missing-`[discord]`-extra path raises a plain `RuntimeError`,
+not one of these.
 """
 
 

src/aiowebhooks/sender.py

@@ -1,13 +1,20 @@
 """core async webhook sender (aiohttp only, no discord knowledge).
 
 `Webhook` posts a JSON dict to a url (or round-robins a pool), handling 429/5xx
-retries and optional proxy rotation, and always returns a `WebhookResult` — it
-never raises on a send failure. the discord layer builds payloads and delegates
-the actual POST here so it inherits rotation / proxy / retry / result.
+retries, connection/timeout retries, and optional proxy rotation, and always
+returns a `WebhookResult` — it never raises on a send failure. the discord layer
+builds payloads and delegates the actual POST here so it inherits rotation / proxy
+/ retry / result.
+
+a 429's server-controlled `retry_after` wait is bounded: non-finite values
+(inf/nan) are rejected and finite values are clamped to `MAX_RETRY_AFTER`, so an
+adversarial or misconfigured server can never stall a send() past a bounded
+ceiling.
 """
 
 import asyncio
 import logging
+import math
 from typing import Dict, List, Optional, Union
 from urllib.parse import unquote, urlsplit
 
@@ -19,6 +26,14 @@ from .result import WebhookResult
 
 log = logging.getLogger(__name__)
 
+MAX_RETRY_AFTER = 300.0
+"""ceiling (seconds) honored from a server-controlled 429 retry_after/Retry-After
+
+guards against an arbitrarily large or non-finite wait (inf/nan, or a ms-vs-s unit
+mismatch) stalling a send() past its ClientTimeout. non-finite values are rejected
+outright; finite values are clamped to this ceiling.
+"""
+
 
 class _Retryable(Exception):
     """internal signal: a retryable HTTP status (429/5xx); carries the response
@@ -90,18 +105,26 @@ class Webhook:
 
     @staticmethod
     def _retry_after(status: int, headers, body) -> Optional[float]:
-        """seconds to wait on a 429, from body retry_after then Retry-After header"""
+        """seconds to wait on a 429, from body retry_after then Retry-After header
+
+        non-finite values (inf/nan) are rejected as unparseable; finite values are
+        clamped to MAX_RETRY_AFTER.
+        """
         if status != 429:
             return None
         if isinstance(body, dict) and body.get("retry_after") is not None:
             try:
-                return float(body["retry_after"])
+                value = float(body["retry_after"])
+                if math.isfinite(value):
+                    return max(0.0, min(value, MAX_RETRY_AFTER))
             except (TypeError, ValueError):
                 pass
         header = headers.get("Retry-After") if headers else None
         if header is not None:
             try:
-                return float(header)
+                value = float(header)
+                if math.isfinite(value):
+                    return max(0.0, min(value, MAX_RETRY_AFTER))
             except (TypeError, ValueError):
                 pass
         return None
@@ -110,9 +133,10 @@ class Webhook:
         """post a json payload to the next pool url; always returns a result
 
         handles 429 (wait + retry, capped by max_retries), 5xx (retry, capped), and
-        timeout/connection errors with optional proxy rotation (burn + next, capped
-        by max_proxy_retries). 4xx other than 429 fail immediately. never raises on a
-        send failure.
+        timeout/connection errors: with a proxy provider, burn + rotate to the next
+        proxy (capped by max_proxy_retries) before failing; without one, retried
+        directly under max_retries like a 5xx. 4xx other than 429 fail immediately.
+        never raises on a send failure.
         """
         url = self._next_url()
         session = self._session
@@ -130,41 +154,66 @@ class Webhook:
     ) -> WebhookResult:
         """status-retry (via commons.aretry) wrapping proxy rotation; never raises
 
-        commons.aretry owns the 429/5xx backoff schedule + retry cap (max_retries),
-        retrying on the internal _Retryable signal. on exhaustion it re-raises the
-        last _Retryable, whose carried result is the REAL last response (not a
-        synthetic status-0). proxy rotation on connection errors lives inside the
-        attempt and is capped separately.
+        commons.aretry owns the 429/5xx/no-provider-connection-error backoff schedule
+        + retry cap (max_retries), retrying on the internal _Retryable signal; on
+        exhaustion it re-raises the last _Retryable, whose carried result is the real
+        last response (not a synthetic status-0). proxy rotation on connection errors
+        lives inside the attempt, capped separately (max_proxy_retries); once that cap
+        is hit the failure is terminal, not retried again via aretry.
         """
         counter = [0]
+        pending_wait: List[Optional[float]] = [None]
         try:
             return await aretry(
-                lambda: self._attempt(session, url, payload, counter),
+                lambda: self._attempt(session, url, payload, counter, pending_wait),
                 attempts=self.max_retries + 1,
                 on=(_Retryable,),
             )
         except _Retryable as exhausted:
             return exhausted.result
+        except Exception as error:
+            # never-raises safety net: anything not aiohttp.ClientError/TimeoutError
+            # (closed injected session, malformed proxy url, ...) comes back as a
+            # failed result instead of escaping send()
+            log.warning("webhook send failed unexpectedly on %s: %s", url, error, exc_info=True)
+            return WebhookResult(
+                ok=False, status=None, url=url, attempts=counter[0] or 1,
+                error=f"{type(error).__name__}: {error}",
+            )
 
     async def _attempt(
-        self, session: aiohttp.ClientSession, url: str, payload: Dict, counter: List[int]
+        self, session: aiohttp.ClientSession, url: str, payload: Dict, counter: List[int],
+        pending_wait: List[Optional[float]],
     ) -> WebhookResult:
         """one logical send: proxy rotation + a single POST; may raise _Retryable
 
         raises _Retryable (carrying the real response) on a 429/5xx so the caller's
-        aretry applies backoff; honors an explicit 429 retry_after by sleeping it
-        before signalling. returns a final WebhookResult on success or a terminal
-        (non-retryable) failure — never lets a provider/connection error escape.
+        aretry applies backoff. an explicit 429 retry_after is carried over and slept
+        at the START of the next attempt, never after the one that raises, so an
+        exhausted final attempt never sleeps a wait it won't use. a connection/timeout
+        error also raises _Retryable when no proxy provider is set (retries under
+        aretry like a 5xx); with a provider it instead drives burn+rotate up to
+        max_proxy_retries and only becomes terminal once that cap is hit. always
+        returns a final WebhookResult or raises _Retryable — never lets a provider
+        error escape.
 
-        `counter` is a per-call mutable cell ([0]) owned by the calling `_send_loop`,
-        so the attempt count is local to one `send()` and concurrent sends on the
-        same instance don't corrupt each other's tally.
+        `counter`/`pending_wait` are per-call mutable cells ([0] / [None]) owned by
+        `_send_loop` and threaded through because aretry calls `_attempt` fresh on
+        every top-level retry, so a plain local wouldn't survive across calls.
+        `counter` tallies attempts; `pending_wait` carries a 429 retry_after to the
+        next attempt's sleep. keeps both local to one `send()`, safe for concurrent
+        sends on the same instance.
         """
         timeout = aiohttp.ClientTimeout(total=self.timeout)
         last_proxy: Optional[str] = None
         proxy_tries = 0
 
         while True:
+            if pending_wait[0] is not None:
+                wait = pending_wait[0]
+                pending_wait[0] = None
+                log.warning("webhook 429 on %s; honoring retry_after %.3fs", url, wait)
+                await asyncio.sleep(wait)
             counter[0] += 1
             attempts = counter[0]
             proxy_url = None
@@ -172,8 +221,7 @@ class Webhook:
                 try:
                     proxy_dict = self._proxies.get()
                 except Exception:
-                    # duck-typed provider; any error from get() means no proxy is
-                    # available — fail cleanly rather than escaping send().
+                    # duck-typed provider: any get() error means no proxy available
                     log.warning("webhook: proxy get() failed; no proxy available",
                                 exc_info=True)
                     return WebhookResult(
@@ -201,17 +249,27 @@ class Webhook:
                         error=f"http {status}", response=body, proxy=last_proxy,
                     )
 
+                    if status == 429:
+                        # every 429 is retryable; an explicit retry_after is carried
+                        # to the NEXT attempt's sleep (never slept here), so an
+                        # exhausted final attempt never sleeps a wait it won't use.
+                        # no parseable wait still retries under aretry's backoff+cap;
+                        # an honored wait is additive with that backoff (over-waits,
+                        # never under-waits). _retry_after rejects non-finite values
+                        # and clamps finite ones to MAX_RETRY_AFTER.
                         wait = self._retry_after(status, resp.headers, body)
                         if wait is not None:
-                        log.warning("webhook 429 on %s; honoring retry_after %.3fs", url, wait)
-                        await asyncio.sleep(wait)
+                            pending_wait[0] = wait
+                        else:
+                            log.warning("webhook 429 on %s; no retry_after, backing off", url)
                         raise _Retryable(result)
                     if status >= 500:
                         raise _Retryable(result)
                     return result
 
             except (aiohttp.ClientError, asyncio.TimeoutError) as error:
-                if self._proxies is not None and proxy_tries < self.max_proxy_retries:
+                if self._proxies is not None:
+                    if proxy_tries < self.max_proxy_retries:
                         if self._burn(last_proxy):
                             proxy_tries += 1
                             continue
@@ -223,15 +281,21 @@ class Webhook:
                         ok=False, status=None, url=url, attempts=attempts,
                         error=f"{type(error).__name__}: {error}", proxy=last_proxy,
                     )
+                # no proxy provider: retry a connection/timeout error like a 5xx,
+                # under aretry's backoff + max_retries instead of failing one-shot;
+                # _send_loop's `except _Retryable` returns the carried result once
+                # exhausted, so never-raises still holds.
+                raise _Retryable(WebhookResult(
+                    ok=False, status=None, url=url, attempts=attempts,
+                    error=f"{type(error).__name__}: {error}", proxy=last_proxy,
+                ))
 
     def _burn(self, proxy: Optional[str]) -> bool:
         """burn the current proxy; return False if it can't be rotated
 
-        the provider is duck-typed and never imported, so we cannot catch its
-        exception types by class. ANY exception from burn (a ProxiesExhaustedError
-        on a dead pool, a ValueError when the proxy isn't in the pool, etc.) means
-        we can't rotate — return False so the caller ends the loop with a failed
-        result rather than letting it escape send() (which must never raise).
+        the provider is duck-typed and never imported, so any exception from burn
+        (dead pool, unknown proxy, ...) is caught broadly and means "can't rotate" —
+        return False rather than let it escape send().
         """
         try:
             self._proxies.burn(proxy)