Add Foundry Local provider

[justinchuby]

Adds support for Microsoft Foundry Local, an on-device runtime that exposes an OpenAI-compatible API.

FoundryLocalProvider (named after the product) subclasses OpenaiProvider (same pattern as ollama/lmstudio) and is selected with the foundry_local: model prefix. A Foundry_localProvider alias is kept so the provider factory (which derives the class name from "foundry_local".capitalize()) resolves the key. The provider operates in one of two modes:

• Managed (default): uses the Foundry Local Python SDK to start the service on demand, download/load the model, and resolve a model alias to its concrete id. Bootstrap is deferred to the first request, since the endpoint uses a dynamic port and the alias is only known at call time. Both SDK generations are supported via runtime detection: the current foundry-local-sdk 1.x (imported as foundry_local_sdk, using the Configuration/catalog/start_web_service singleton API) and the legacy 0.x package (imported as foundry_local, using FoundryLocalManager(alias)).
• Explicit endpoint: when api_url/base_url or FOUNDRY_LOCAL_API_URL is set, points the OpenAI SDK directly at a running endpoint (no SDK required); the host is normalized to /v1 idempotently. A friendly model alias is resolved to the concrete served id by querying the endpoint's /v1/models (exact id → parent alias → alias-boundary prefix), falling back to the original string when the lookup fails or there is no confident match — so the same foundry_local:<alias> works in both modes.

Validation

Tested end-to-end on Windows against a real Foundry Local install (0.8.119) with foundry-local-sdk 1.2.3:

• Managed mode (foundry_local:qwen2.5-0.5b): auto-started the service, downloaded + loaded the model, resolved the alias to qwen2.5-0.5b-instruct-generic-gpu:4, returned a valid completion (finish_reason=stop).
• Explicit-endpoint mode: friendly alias qwen2.5-0.5b resolved against the running endpoint to qwen2.5-0.5b-instruct-generic-cpu:4 and returned a completion.
• Tool calling: returned a structured tool_calls response (finish_reason=tool_calls). Note: tool calling through ai.Client currently requires the mcp extra (pip install 'aisuite[mcp]') due to an unrelated upstream bug where is_mcp_config is gated behind the MCP import.
• Unit tests: 14 passed.

Changes

• aisuite/providers/foundry_local_provider.py — new provider (FoundryLocalProvider, with a Foundry_localProvider alias for factory compatibility); explicit-endpoint alias resolution.
• pyproject.toml — foundry-local extra installs openai and foundry-local-sdk (^1.2.3, declared with a python = ">=3.11" marker since the SDK requires Python ≥3.11 while aisuite supports ^3.10; it's also added to the all extra). The SDK pulls in native ONNX Runtime/CUDA wheels, so users on Python 3.10 (or who need a hardware-specific variant) can still install it directly with pip install foundry-local-sdk.
• tests/providers/test_foundry_local_provider.py — explicit-endpoint mode (incl. alias resolution by prefix/parent, concrete-id passthrough, ambiguous-alias error, unreachable-endpoint fallback), both managed SDK backends (1.x and 0.x), alias→id resolution, multi-alias loading, missing-SDK error.
• guides/foundry_local.md (+ index link in guides/README.md) — usage for both modes and the aisuite[mcp] note for tool calling.

Usage

import aisuite as ai

client = ai.Client()
response = client.chat.completions.create(
    model="foundry_local:phi-3.5-mini",   # alias resolved + service started automatically
    messages=[{"role": "user", "content": "What is the golden ratio?"}],
)
print(response.choices[0].message.content)

Explicit endpoint (SDK not required; alias resolved against the running endpoint):

client = ai.Client(provider_configs={"foundry_local": {"api_url": "http://localhost:5273"}})
client.chat.completions.create(
    model="foundry_local:phi-3.5-mini",
    messages=[{"role": "user", "content": "Hi"}],
)