fix(anthropic): surface mid-stream errors on the sync /v1/messages adapter

[revanth-045]

Problem

AnthropicStreamWrapper has two SSE wrappers, and only the async one handles a mid-stream provider failure:

def anthropic_sse_wrapper(self) -> Iterator[bytes]:
    for chunk in self:            # no error handling
        ...

async def async_anthropic_sse_wrapper(self) -> AsyncIterator[bytes]:
    try:
        async for chunk in self:
            ...
    except Exception as e:
        yield _mid_stream_error_sse_event(e)

transformation.py returns the sync one whenever is_async is False, so it is a live path.

The sync case is actually worse than the pre-#43601 async behaviour. __next__ ends with:

except Exception as e:
    verbose_logger.error("Anthropic Adapter - %s\n%s", e, traceback.format_exc())
    raise StopIteration

So an upstream failure (the BedrockError on a ConverseStream that ends without a terminal messageStop, per the existing regression tests) is swallowed and converted into a clean end-of-stream. The client sees a silently truncated answer and no error event — not even the "Connection closed mid-response" that the async fix was written for. __anext__ has no such catch, which is why the async guard works.

Fix

Two lines, mirroring the async path:

• __next__ propagates the exception instead of converting it to StopIteration (the log line stays).
• anthropic_sse_wrapper gets the same try/except as async_anthropic_sse_wrapper, emitting _mid_stream_error_sse_event(e).

On the __next__ change: AnthropicStreamWrapper is constructed in exactly one place (transformation.py:333) and is only ever consumed through one of the two SSE wrappers, so the propagated exception is caught immediately by the guard added here and never reaches a caller that previously saw a clean stop.

Tests

Two cases added to the existing test_streaming_iterator_mid_stream_error.py, as sync mirrors of the async ones already there: a BedrockError after partial content ends the stream with a terminal Anthropic error event carrying the upstream message, and draining the wrapper does not re-raise. Both fail on main.

tests/unit/llms/anthropic/pass_through/adapters 371 passed; ruff check and ruff format --check clean on the source file. (The UP035/UP045 findings in that test file are pre-existing on main and untouched.)

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

revanth-045 seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Fixes error handling in streaming response adapter.

The PR appears safe to merge; no actionable regression was identified.

Summary

The PR makes the synchronous Anthropic Messages SSE adapter report upstream mid-stream failures instead of silently ending a truncated response.

• Re-raises failures from the sync iterator and converts them to terminal Anthropic error events at the SSE boundary.
• Adds sync regression tests for an error after partial content and an error before upstream content.

_{Reviews (1) · Last reviewed commit: "fix(anthropic): surface mid-stream error..."}

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing revanth-045:fix/anthropic-sync-sse-mid-stream-error (3de13bc) with main (7f95b5f)}

[codecov]

Codecov Report

❌ Patch coverage is 90.90909% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...hropic/pass_through/adapters/streaming_iterator.py	90.90%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!