azure-ai-agents==1.1.0 — Async AgentsClient.files.upload_and_poll() fails with DecodeError (sync works)

[AyushTripathi07]

• Package Name: azure-ai-agents
• Package Version: 1.1.0
• Operating System: macOS (ARM64)
• Python Version: 3.12.4

Describe the bug

The async AgentsClient (azure.ai.agents.aio.AgentsClient) fails on files.upload_and_poll() with a DecodeError, while the sync AgentsClient (azure.ai.agents.AgentsClient) succeeds for the exact same operation, endpoint, credentials, and file.

The async pipeline's ContentDecodePolicy.on_response attempts to JSON-deserialize the HTTP response body but receives raw binary/compressed data instead, causing json.JSONDecodeError: Expecting value: line 1 column 1 (char 0).

This appears to be a difference in how the async HTTP transport (aiohttp) handles response decompression compared to the sync transport (requests). The sync transport correctly decompresses the response before JSON parsing, while the async transport does not.

To Reproduce

1. Install azure-ai-agents==1.1.0 and azure-identity
2. Run the following sync script — it succeeds:

from azure.ai.agents import AgentsClient
from azure.ai.agents.models import FilePurpose
from azure.identity import DefaultAzureCredential

client = AgentsClient(
    endpoint="https://<endpoint>.services.ai.azure.com/api/projects/<project>",
    credential=DefaultAzureCredential()
)

# ✅ This succeeds
uploaded_file = client.files.upload_and_poll(
    file_path="any_file.md",
    purpose=FilePurpose.AGENTS
)
print(f"Uploaded: {uploaded_file.id}")

# Cleanup
client.files.delete(file_id=uploaded_file.id)
client.close()

3. Run the following async script — it fails:

import asyncio
from azure.ai.agents.aio import AgentsClient
from azure.ai.agents.models import FilePurpose
from azure.identity import DefaultAzureCredential

async def main():
    client = AgentsClient(
        endpoint="https://<endpoint>.services.ai.azure.com/api/projects/<project>",
        credential=DefaultAzureCredential()
    )

    # ❌ This raises DecodeError
    uploaded_file = await client.files.upload_and_poll(
        file_path="any_file.md",
        purpose=FilePurpose.AGENTS
    )

asyncio.run(main())

Same endpoint, same credentials, same file, same SDK version — only difference is sync vs async client.

Expected behavior

The async files.upload_and_poll() should return a valid FileInfo object with a .id field, identical to the sync version's behavior.

Additional context

Full async traceback:

Traceback (most recent call last):
  File ".../azure/core/pipeline/policies/_universal.py", line 618, in deserialize_from_text
    return json.loads(data_as_str)
json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)

The above exception was the direct cause of the following exception:

  File ".../azure/ai/agents/aio/operations/_patch.py", line 1563, in upload_and_poll
    uploaded_file = await self.upload(file_path=file_path, purpose=purpose, **kwargs)
  File ".../azure/ai/agents/aio/operations/_patch.py", line 1435, in upload
    return await super()._upload_file(body={"file": file_content, "purpose": purpose}, **kwargs)
  File ".../azure/ai/agents/aio/operations/_operations.py", line 2411, in _upload_file
    pipeline_response: PipelineResponse = await self._client._pipeline.run(...)
  ...
  File ".../azure/core/pipeline/policies/_universal.py", line 728, in on_response
    response.context[self.CONTEXT_NAME] = self.deserialize_from_http_generics(...)
  File ".../azure/core/pipeline/policies/_universal.py", line 690, in deserialize_from_http_generics
    return cls.deserialize_from_text(response.text(encoding), mime_type, response=response)
  File ".../azure/core/pipeline/policies/_universal.py", line 620, in deserialize_from_text
    raise DecodeError(...)

azure.core.exceptions.DecodeError: JSON is invalid: Expecting value: line 1 column 1 (char 0)
Content: <binary garbage bytes — appears to be compressed/undecoded response body>

The Content: output in the error shows raw binary data, suggesting the async transport is returning the response body without decompression (gzip/deflate) before ContentDecodePolicy attempts JSON parsing.

Impact: This bug makes the async AgentsClient unusable for any file upload operations, blocking files.upload(), files.upload_and_poll(), and any downstream vector store workflows in async applications (e.g. FastAPI, aiohttp-based services).

[AyushTripathi07]

Root Cause:

I think I have identified the root cause of the issue.

The async transport layer used by AgentsClient in azure.ai.agents.aio receives responses with Content-Encoding: gzip, but the response body is not decompressed before JSON parsing occurs. As a result, the JSON parser encounters the binary gzip header and throws a DecodeError.

Verification

I verified this behavior by forcing the server to return uncompressed responses. After doing so, the error disappeared completely.

This was achieved by injecting a custom SansIOHTTPPolicy that sets the Accept-Encoding header to identity.

Working Workaround

Below is a workaround that keeps everything fully asynchronous without needing to wrap calls in synchronous code.

from azure.core.pipeline.policies import SansIOHTTPPolicy
from azure.ai.agents.aio import AgentsClient
from azure.identity import DefaultAzureCredential


class NoGzipPolicy(SansIOHTTPPolicy):
    def on_request(self, request):
        # Force the server to send uncompressed responses
        request.http_request.headers["Accept-Encoding"] = "identity"


client = AgentsClient(
    endpoint=ENDPOINT,
    credential=DefaultAzureCredential(),
    per_call_policies=[NoGzipPolicy()]
)

Result

With compression disabled, the following operations work correctly:

• files.upload_and_poll
• vector_stores.create_and_poll
• list operations

This confirms that the failure path is specifically tied to gzip-compressed responses not being decompressed in the async pipeline before JSON parsing.

Notes

• This workaround allows continued use of the async client without needing any sync wrappers.
• The issue appears to be related to response decompression handling in the async transport layer.

Hope this helps narrow down the fix.