Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.98%. Comparing base (6264834) to head (7005508).
✅ All tests successful. No failed tests found.

@@           Coverage Diff            @@
##           master   #13330    +/-   ##
========================================
  Coverage   98.98%   98.98%            
========================================
  Files         132      132            
  Lines       49073    49174   +101     
  Branches     2553     2557     +4     
========================================
+ Hits        48576    48677   +101     
  Misses        373      373            
  Partials      124      124            

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

Merging this PR will regress 0 benchmarks

⚡ 1 improved benchmark
❌ 1 (👁 1) regressed benchmark
✅ 82 untouched benchmarks
⏩ 83 skipped benchmarks¹

Performance Changes

_{Comparing aiolibsbot:koan/implement-13329 (7005508) with master (6264834)}

The fix looks right. extending it to the case where the request is retried should work. Just didn't plan for that case

@bdraco: The fix looks right. extending it to the case where the request is retried should work. Just didn't ...

Retry path now mirrors 307/308 path: reuses req._body instead of rebuilding from data. IOBasePayload._set_or_restore_start_position() seeks back to recorded start; size stays full-file, so Content-Length matches. AsyncIterablePayload marks itself consumed at first drain, not only at StopAsyncIteration. Unreplayable bodies raise original connection error rather than retry truncated.

Previous review — superseded by a newer review below.

Rebase with requested adjustments

Branch koan/implement-13329 was rebased onto master and review feedback was applied.

Changes applied

• Changes complete. No lint tooling in sandbox; syntax verified.
• Added tests/test_payload.py::test_async_iterable_payload_consumed_on_interrupted_write — drives write_with_length() against a writer that raises after the first chunk and asserts consumed is True, covering the early _consumed flag that suggestion ＃1 showed was untested (reverting payload.py alone left CI green).
• Added MultipartWriter.consumed property (aiohttp/multipart.py) returning True when the writer or any part is consumed, so a multipart/FormData body containing a drained AsyncIterablePayload part now fails fast on both the retry and redirect guards — closes the HIGH silent-truncation bypass (＃2). close() still reads _consumed directly, so close-once semantics are unchanged.
• Added tests/test_multipart.py::test_multipart_writer_consumed_follows_parts covering the new aggregation.
• Added await req._close() before reusing req._body on the retry path (aiohttp/client.py) so the cancelled writer task is awaited and any in-flight executor read cannot race the rewind seek() (＃3).
• Extended CHANGES/13329.bugfix.rst with the multipart replay-guard behaviour.

Not changed (and why)

• MEDIUM "guard relies on a swallowed error" (_set_or_restore_start_position swallowing a failed seek()): disagree on scope. That swallow predates this PR and is on every IOBasePayload read path (size, as_bytes, decode, plus the redirect replay), not just retry. Turning it into a raise is a

Stats

8 files changed, 159 insertions(+), 5 deletions(-)

CI status

CI will be checked asynchronously.

Automated by Kōan

Previous review — superseded by a newer review below.

Rebase with requested adjustments

Branch koan/implement-13329 was rebased onto master and review feedback was applied.

Changes applied

• Trimmed the retry-path comment at aiohttp/client.py:740 from 14 lines to 5, keeping only the two non-obvious facts (rebuilt payload resends just the unread remainder; the cancelled writer must finish before the rewind) — per @bdraco and the follow-up.
• Moved await req._close() above the req._body.consumed guard, so replayability is read after in-flight write state settles — per the HIGH silent-failure finding at client.py:749-756.

Not changed (and why)

• MEDIUM "incomplete synchronization barrier" (track the payload's executor future): disagree on scope. Real gap, but fixing it means new public quiesce/drain API on IOBasePayload plus per-payload future tracking — a separate change, not this bugfix. Current ordering is strictly better than before.
• MEDIUM "advisory flag with no enforcement" in payload.py: disagree. Raising inside AsyncIterablePayload.write_with_length on _consumed changes behaviour for existing non-retry writers and risks breaking multipart re-write paths. Out of scope for this fix.

Stats

8 files changed, 152 insertions(+), 5 deletions(-)

CI status

CI will be checked asynchronously.

Automated by Kōan

Previous review — superseded by a newer review below.

Rebase with requested adjustments

Branch koan/implement-13329 was rebased onto master and review feedback was applied.

Changes applied

• Trimmed retry-path comment in aiohttp/client.py from five lines to three, per @bdraco's "overly verbose prose" request — kept only the two load-bearing facts (rebuilt payload would resend just the unread remainder; rewind happens after the cancelled writer settles) and dropped the prose duplicated in the changelog fragment.

Not changed (and why)

• Three MEDIUM "Silent Failure Analysis" findings (distinguishable ClientPayloadError on unreplayable body; propagating seek failure out of _set_or_restore_start_position; raising in AsyncIterablePayload.write_with_length when consumed) — automated advisory findings, not maintainer change requests, and each widens scope beyond the stated fix. The latter two alter behaviour of the pre-existing redirect-replay path shared by all callers; that belongs in a separate PR.

Stats

8 files changed, 150 insertions(+), 5 deletions(-)

CI status

CI will be checked asynchronously.

Automated by Kōan

Previous review — superseded by a newer review below.

Rebase with requested adjustments

Branch koan/implement-13329 was rebased onto master and review feedback was applied.

Changes applied

• BodyPartReaderPayload.write() now sets _consumed = True before draining the reader (aiohttp/multipart.py:664) — per the HIGH finding, a fully drained proxy body previously reported consumed is False, so the retry guard passed and an empty body was resent.
• Added test_body_part_reader_payload_consumed_after_write and test_multipart_writer_consumed_follows_body_part_reader — the second covers the requested case of a MultipartWriter holding a drained BodyPartReader part rather than only an async generator.
• Changelog fragment notes the body-part-reader behaviour.

Not changed (and why)

• @bdraco "Trim the overly verbose prose" (client.py:754) — already applied in commit 737a606. The retry block at HEAD carries 3 comment lines (client.py:740-742), not the 14 the follow-up bot comment describes; that comment is stale.
• MEDIUM "error state set after the check that reads it" — real, but pre-existing and not introduced here: the 307/308 redirect path (client.py:796) uses the identical consumed-then-write ordering, and IOBasePayload._set_or_restore_start_position() has always swallowed the seek failure. Making it fail-closed means changing IOBasePayload's contract for every write path, which is a separate fix.
• "Unreplayable-body failure distinguishable from a transient disconnect" — deliberate: the request did fail with a connection error, and re-raising the original preserves the cause and traceback. Substituting a synthetic error would hide the real failure;

Stats

8 files changed, 189 insertions(+), 5 deletions(-)

CI status

CI will be checked asynchronously.

Automated by Kōan

Previous review — superseded by a newer review below.

Rebase with requested adjustments

Branch koan/implement-13329 was rebased onto master and review feedback was applied.

Changes applied

• Trimmed CHANGES/13329.bugfix.rst from 11 lines to 4 per @bdraco: kept the fix sentence plus the one behaviour-change line, dropped payload-reuse and multipart mechanism detail (already in the PR body).

Not changed (and why)

• aiohttp/client.py:754 "trim the overly verbose prose" — already trimmed in an earlier pass. Current comment is 3 lines (client.py:740-742) covering exactly the two facts flagged as worth keeping: rebuild resends only the unread remainder, and the cancelled writer must settle before rewind. The follow-up note claiming 14 comment lines was stale.
• Three MEDIUM automated-review findings (_close() inside except block, consumed-flag doubling as cleanup guard, write_with_length not enforcing _consumed) — advisory, not maintainer change requests, and each widens scope beyond the retry fix.

Stats

8 files changed, 183 insertions(+), 5 deletions(-)

CI status

CI will be checked asynchronously.

Automated by Kōan

PR Review — Rewind file bodies when internally retrying a request

Both outstanding maintainer requests are now addressed and the fix holds up under verification — merge-ready with nits.

The core change is the right shape: reusing the failed attempt's payload instead of rebuilding it from data mirrors what the 307/308 path already does at client.py:796-803, so this adds no parallel mechanism. I re-verified the load-bearing details against the branch — IOBasePayload.size (payload.py:534-559) subtracts _start_position rather than reading from the current offset, so Content-Length stays consistent with the rewound body; PayloadRegistry.get returns an existing Payload unchanged (payload.py:113-115), so passing req._body back as data reuses the instance rather than re-wrapping it; _body defaults to the _EMPTY_BODY singleton (client_reqrep.py:1043), never None, so the new .consumed read cannot AttributeError; and no error path closes the body between the failure and the retry (req._close() only awaits the writer task). Ordering is right too — await req._close() precedes the consumed check, so replayability is read after write state settles.

The two _consumed relocations are sound. AsyncIterablePayload setting the flag before draining leaves the cached-chunks branch untouched (it returns at payload.py:1040 before the flag is set), so as_bytes()-then-write still replays. AsyncIterablePayload._autoclose = True, so the multipart close-guard is unaffected there. Tests are well targeted: both functional tests genuinely fail on master, and the new multipart tests cover the drained-BodyPartReader case rather than only the async-generator case.

• Resolved since last round: the news fragment is trimmed to 4 lines (in line with recent bugfix fragments on master, which run 1-4), and the client.py comment is down to 3 lines — both maintainer change-requests are now satisfied.
• await req._close() can re-raise a non-cancelled writer-task exception in place of the retryable error, skipping the retry; reachable in principle via write_eof/expect100, could not reproduce.
• Re-iterable AsyncIterable bodies (not async generators) retried successfully on master and are now refused — the only case where working behaviour is traded for a refusal; the fragment's "partially consumed" wording understates it.
• Marking BodyPartReaderPayload consumed at write start also skips it in MultipartWriter.close()'s per-part close loop; harmless today (inherited no-op close()) but a latent trap on a class explicitly marked _autoclose = False.
• AsyncIterablePayload.write_with_length sets _consumed but never enforces it, so any caller that does not consult .consumed can still emit a short body.
• The comment claims the payload rewinds "once the cancelled writer has settled", but _close() does not wait for the in-flight executor read; narrow race with the retry's seek (unverified).
• multipart.py:915 still says "Can be encoded multiple times" — the exact invariant the new property revokes. (A prior round recorded this fixed; it is still present.)
• The PR description omits the multipart changes entirely, and the trimmed changelog no longer mentions them either.

✅ Resolved since last review (1)

🟢 Suggestions

with contextlib.suppress(Exception):
    await req._close()

                            await req._close()
                            if req._body.consumed:
                                raise

    async def write(self, writer: AbstractStreamWriter) -> None:
        field = self._value
        self._consumed = True
        while chunk := await field.read_chunk(size=DEFAULT_CHUNK_SIZE):

Checklist

• Fix addresses the stated bug and is covered by tests that fail on master
• Retried payload produces a Content-Length consistent with the resent body
• No resource leaks or missed cleanup on the new error paths — suggestion ＃2
• Error handling preserves the original exception on the retry path — suggestion ＃1
• No unintended behavioral regressions for previously working inputs
• New invariant enforced at the point of failure, not only by convention
• No race conditions introduced between the aborted and retried write
• Comments and class-level docs match the new invariants
• PR description and changelog cover all behavioral changes in the diff
• Maintainer change-requests addressed (news fragment and code comment trimmed)
• No hardcoded secrets, injection, or unsafe deserialization
• Tests verify observable behavior rather than inspecting source

Silent Failure Analysis

except (ClientOSError, ServerDisconnectedError):
    ...
    if data is not None:
        await req._close()
        if req._body.consumed:
            raise

await req._close()
if req._body.consumed:
    raise
data = req._body

# ... reuse the payload, which rewinds itself once the
# cancelled writer has settled.
await req._close()

field = self._value
# Reading the part drains the underlying stream irreversibly...
self._consumed = True

if req._body.consumed:
    raise

Automated review by Kōan (Claude) HEAD=ef3a00d 7 min 12s

Rebase with requested adjustments

Branch koan/implement-13329 was rebased onto master and review feedback was applied.

Changes applied

• Both open inline items already applied in HEAD.
• client.py:740-742 — comment now 3 lines, keeps only the two facts @aiolibsbot proposed (trimmed in 737a6066).
• CHANGES/13329.bugfix.rst — now 5 lines: sentence one + one behaviour-change line (trimmed in ef3a00de). Payload-reuse and multipart mechanism prose dropped.
• Reviewer notes citing "14 comment lines" / "11-line fragment" describe pre-ef3a00de state.

Not changed (and why)

• @bdraco "trim verbose prose" on client.py:754 — already done in commit 737a6066; comment is 3 lines / 4 code lines, matching the proposed version.
• @bdraco "trim verbose prose" on CHANGES/13329.bugfix.rst — already done in commit ef3a00de; fragment is sentence one plus one behaviour line, exactly the requested shape.
• Automated HIGH _set_or_restore_start_position() swallows failed seek() — real but pre-existing upstream edge case (tell() succeeds, seek() later fails), and it equally affects the redirect path. Making it raise is a behaviour change beyond the two prose trims requested this round; belongs in its own PR.
• Automated MEDIUM items (executor read race, BodyPartReaderPayload close hook, ClientPayloadError for abandoned retry) — advisory, addressed or deliberately deferred in earlier passes; no human reviewer asked for them.

Stats

8 files changed, 183 insertions(+), 5 deletions(-)

CI status

CI will be checked asynchronously.

Automated by Kōan

Rebase with requested adjustments

Branch koan/implement-13329 was rebased onto master and review feedback was applied.

Changes applied

• Both trims already in HEAD. No changes needed.

Not changed (and why)

• @bdraco "trim overly verbose prose" on aiohttp/client.py:754 — already applied. Comment now 3 lines (client.py:740-742) covering only the two retained facts: rebuilt payload resends unread remainder only, and rewind happens after the cancelled writer settles.
• @bdraco "trim the overly verbose prose" on CHANGES/13329.bugfix.rst — already applied. Fragment now 5 lines: one sentence plus one behaviour-change line. Payload-reuse and multipart mechanism prose dropped, matching repo norm (1-4 line fragments).
• @aiolibsbot follow-up notes citing "14 comment lines" and "11-line fragment" — describe pre-ef3a00de state; both superseded.

Stats

8 files changed, 183 insertions(+), 5 deletions(-)

CI status

CI will be checked asynchronously.

Automated by Kōan

Backport to 3.14: 💔 cherry-picking failed — conflicts found

❌ Failed to cleanly apply b084084 on top of patchback/backports/3.14/b08408400fc1a3566ea284f7b07cd54543dcd370/pr-13330

Backporting merged PR #13330 into master

1. Ensure you have a local repo clone of your fork. Unless you cloned it
   from the upstream, this would be your origin remote.
2. Make sure you have an upstream repo added as a remote too. In these
   instructions you'll refer to it by the name upstream. If you don't
   have it, here's how you can add it:

   $ git remote add upstream https://github.com/aio-libs/aiohttp.git

3. Ensure you have the latest copy of upstream and prepare a branch
   that will hold the backported code:

   $ git fetch upstream
   $ git checkout -b patchback/backports/3.14/b08408400fc1a3566ea284f7b07cd54543dcd370/pr-13330 upstream/3.14

4. Now, cherry-pick PR Rewind file bodies when internally retrying a request #13330 contents into that branch:

   $ git cherry-pick -x b08408400fc1a3566ea284f7b07cd54543dcd370

   If it'll yell at you with something like fatal: Commit b08408400fc1a3566ea284f7b07cd54543dcd370 is a merge but no -m option was given., add -m 1 as follows instead:

   $ git cherry-pick -m1 -x b08408400fc1a3566ea284f7b07cd54543dcd370

5. At this point, you'll probably encounter some merge conflicts. You must
   resolve them in to preserve the patch from PR Rewind file bodies when internally retrying a request #13330 as close to the
   original as possible.
6. Push this branch to your fork on GitHub:

   $ git push origin patchback/backports/3.14/b08408400fc1a3566ea284f7b07cd54543dcd370/pr-13330

7. Create a PR, ensure that the CI is green. If it's not — update it so that
   the tests and any other checks pass. This is it!
   Now relax and wait for the maintainers to process your pull request
   when they have some cycles to do reviews. Don't worry — they'll tell you if
   any improvements are necessary when the time comes!

🤖 @patchback
I'm built with octomachinery and
my source is open — https://github.com/sanitizers/patchback-github-app.

Backport to 3.15: 💔 cherry-picking failed — conflicts found

❌ Failed to cleanly apply b084084 on top of patchback/backports/3.15/b08408400fc1a3566ea284f7b07cd54543dcd370/pr-13330

Backporting merged PR #13330 into master

1. Ensure you have a local repo clone of your fork. Unless you cloned it
   from the upstream, this would be your origin remote.
2. Make sure you have an upstream repo added as a remote too. In these
   instructions you'll refer to it by the name upstream. If you don't
   have it, here's how you can add it:

   $ git remote add upstream https://github.com/aio-libs/aiohttp.git

3. Ensure you have the latest copy of upstream and prepare a branch
   that will hold the backported code:

   $ git fetch upstream
   $ git checkout -b patchback/backports/3.15/b08408400fc1a3566ea284f7b07cd54543dcd370/pr-13330 upstream/3.15

4. Now, cherry-pick PR Rewind file bodies when internally retrying a request #13330 contents into that branch:

   $ git cherry-pick -x b08408400fc1a3566ea284f7b07cd54543dcd370

   If it'll yell at you with something like fatal: Commit b08408400fc1a3566ea284f7b07cd54543dcd370 is a merge but no -m option was given., add -m 1 as follows instead:

   $ git cherry-pick -m1 -x b08408400fc1a3566ea284f7b07cd54543dcd370

5. At this point, you'll probably encounter some merge conflicts. You must
   resolve them in to preserve the patch from PR Rewind file bodies when internally retrying a request #13330 as close to the
   original as possible.
6. Push this branch to your fork on GitHub:

   $ git push origin patchback/backports/3.15/b08408400fc1a3566ea284f7b07cd54543dcd370/pr-13330

7. Create a PR, ensure that the CI is green. If it's not — update it so that
   the tests and any other checks pass. This is it!
   Now relax and wait for the maintainers to process your pull request
   when they have some cycles to do reviews. Don't worry — they'll tell you if
   any improvements are necessary when the time comes!

🤖 @patchback
I'm built with octomachinery and
my source is open — https://github.com/sanitizers/patchback-github-app.

@aiolibsbot create the backports

@aiolibsbot gh_request

Use the backport-failed skill to create the failed backports