[azure-ai-projects] Support an HTTPX async transport for AnyIO/Trio applications

[eavanvalkenburg]

Motivation

Microsoft Agent Framework is assessing support for applications running on either asyncio or Trio through AnyIO. Its Foundry integration uses azure.ai.projects.aio.AIProjectClient (and azure.ai.inference.aio for embeddings). The current azure-ai-projects package explicitly requires aiohttp, and these async clients use azure-core's pipeline, whose default async transport is aiohttp. HTTPX supports both asyncio and Trio; requiring aiohttp keeps this path tied to asyncio even if the calling framework uses AnyIO.

Request

Please make a supported HTTPX-based async transport available end to end for azure-ai-projects (and coordinate with azure-ai-inference and async identity/credential paths where needed), so callers can use Foundry async clients under Trio without requiring aiohttp. Replacing the default aiohttp transport with HTTPX would meet the need; if a change of default is too disruptive, a stable, documented HTTPX transport option that avoids a mandatory aiohttp dependency would also work. Preserve the existing asyncio path.

An application should be able to run the same AIProjectClient operation with anyio.run(main, backend="asyncio") and anyio.run(main, backend="trio"), including authentication, concurrent requests, response streaming, and clean cancellation/close. A transport substitution alone is insufficient if another layer still creates asyncio-only resources.

Related work and references

• Shared azure-core request: Add HttpxTransport support for SDK #41382 (stable HTTPX transport/default selection). This issue is specifically about Foundry SDK usability and end-to-end tests once the core transport is available.
• azure-ai-projects dependency on aiohttp: https://github.com/Azure/azure-sdk-for-python/blob/main/sdk/ai/azure-ai-projects/pyproject.toml
• The generated async client uses azure.core.AsyncPipelineClient: https://github.com/Azure/azure-sdk-for-python/blob/main/sdk/ai/azure-ai-projects/azure/ai/projects/aio/_client.py
• Azure Core Experimental has an AsyncHttpXTransport, but the corehttp HTTPX transport documented separately is not an interchangeable transport for these azure-core clients: https://github.com/Azure/azure-sdk-for-python/tree/main/sdk/core/azure-core-experimental
• HTTPX backend support: https://www.python-httpx.org/async/#supported-async-environments

[github-actions]

🎯 Agentic Issue Triage — Needs Manual Review

🏷️ Label Decision

• Candidate labels considered: Client + AI Projects — the issue explicitly names azure-ai-projects, AIProjectClient.aio, and links to the package's pyproject.toml and generated client source under /sdk/ai/azure-ai-projects/. Mgmt was not considered viable since this is a data-plane client library issue.
• Confidence blockers: None — the package, /sdk/ path, and single service are unambiguous, and the category (Client, data-plane) is clearly implied.
• Outcome: Applied Client + AI Projects — prediction was confident (reference issue Calling MCP server tool not working for Azure AI Agent #42700 also uses this exact Client + AI Projects pairing with customer-reported + issue-addressed).

👥 Owner Routing

• CODEOWNERS scan: Scanning bottom-to-top, the entry at line 51 matches: # ServiceLabel: %AI Projects %Service Attention (ignoring the %Service Attention token, "AI Projects" matches the predicted service label). The block for this entry contains only # ServiceLabel: and # PRLabel: %AI Projects comment lines plus a path-ownership line (/sdk/ai/azure-ai-projects/ @dargilco @glharper @howieleung @kingernupur @nick863 @trangevi @bobogogo1990 @Azure/ai-foundry-devex) — it does not contain explicit # AzureSdkOwners: or # ServiceOwners: comment lines.
• Matched entry: # ServiceLabel: %AI Projects %Service Attention (line 51)
• Owners found: none listed under AzureSdkOwners: or ServiceOwners: comment tags in the matched block
• Outcome: Applied needs-team-triage — matched entry has no AzureSdkOwners/ServiceOwners tags to route to automatically.

Generated by Agentic Triage for #49154 · copilot · sonnet50 · 26.6 AIC · ⌖ 6.76 AIC · ⊞ 13.3K · ◷

[kashifkhan]

Eduard van Valkenburg (@eavanvalkenburg) The supported  transport=  extension point already allows Agent Framework or the calling application to implement an async HTTPX/AnyIO transport and pass it to  AIProjectClient .