Update the default HTTP client to httpx-aiohttp (httpx with aiohttp transport) insted of httpx.AsyncClient for better concurrency

[samuel-cahyawijaya]

Summary

This change switches the default async HTTP stack for cohere.AsyncClient (and anything built on AsyncBaseCohere) from a plain httpx.AsyncClient to httpx_aiohttp.HttpxAiohttpClient, which subclasses httpx.AsyncClient but uses aiohttp under the hood for the actual I/O.

That matches the kind of problem discussed for other Python SDKs that rely on httpx for highly concurrent async workloads: the default httpx async path can lag behind aiohttp when many requests run in parallel (see openai-python#1596 and the discussion there, including this comment). We keep httpx’s request/response surface (types, retries, cookies, auth patterns, etc.) while improving scalability for parallel async calls.

Backward compatibility

• Fully backward compatible for callers who still want the previous behavior: pass your own httpx.AsyncClient via httpx_client=.... It is used verbatim (not wrapped or replaced).
  • HttpxAiohttpClient is a subclass of httpx.AsyncClient, so existing type hints and code that expect an httpx.AsyncClient remain valid for the default client.
• Sync cohere.Client continues to use httpx.Client unchanged.

Dependencies

• Add httpx-aiohttp.
• Bump httpx to >=0.27.0 to satisfy httpx-aiohttp.

Tests

Test: tests/test_async_client.py — asserts a custom httpx.AsyncClient is the same object wired into the wrapper for both AsyncClient and AsyncClientV2.
Demo: scripts/demo_async_client_httpx.py — runs chat with the default client vs an explicit httpx.AsyncClient.

---

Note

Medium Risk
Changes the default async HTTP transport from httpx.AsyncClient to httpx_aiohttp.HttpxAiohttpClient, which can alter runtime networking behavior (connection handling, timeouts, redirects) under load. Risk is mitigated by preserving the ability to pass a custom httpx.AsyncClient verbatim and adding coverage for that path.

Overview
Switches AsyncClient/AsyncBaseCohere to use httpx_aiohttp.HttpxAiohttpClient (aiohttp-backed) as the default async HTTP client to improve concurrency, while still accepting a caller-provided httpx.AsyncClient unchanged.

Updates dependencies to support this (httpx>=0.27.0, adds httpx-aiohttp) and regenerates poetry.lock accordingly. Adds a small async demo script and new tests asserting custom httpx.AsyncClient instances are wired through for both AsyncClient and AsyncClientV2.

^{Written by Cursor Bugbot for commit df83506. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

[cursor]

Top-level import forces aiohttp for sync-only users

Medium Severity

The from httpx_aiohttp import HttpxAiohttpClient import is at the module top level, unconditionally executed when base_client.py is loaded. Since client.py imports BaseCohere and AsyncBaseCohere from this module, and cohere.__init__ imports from client.py, any import cohere — even for purely sync usage via cohere.Client — triggers loading of httpx_aiohttp and transitively all of aiohttp (including its C extensions). HttpxAiohttpClient is only used inside AsyncBaseCohere.__init__, so the import could be deferred there or placed behind typing.TYPE_CHECKING with a lazy import at the call site.