feat: per-vendor model_name defaults on new chat clients (parity with legacy)

[cosminacho]

Summary

The legacy chat classes had per-vendor model_name defaults so callers could instantiate UiPathChat() or UiPathChatBedrock() with no args. The new classes re-exported from uipath_langchain.chat (backed by uipath_langchain_client) don't — model= is a required pydantic field. This PR restores parity where a legacy default existed.

Per-vendor defaults (verified against the git history of the legacy classes)

Class	Default	Mechanism
UiPathChat, UiPathAzureChatOpenAI, UiPathChatOpenAI	UIPATH_MODEL_NAME → gpt-4.1-mini-2025-04-14	field default_factory
UiPathChatGoogleGenerativeAI, UiPathChatVertex	gemini-2.5-flash	field default_factory
UiPathChatBedrock, UiPathChatAnthropicBedrock	anthropic.claude-haiku-4-5-20251001-v1:0	field default_factory
UiPathChatBedrockConverse	anthropic.claude-haiku-4-5-20251001-v1:0	thin subclass with a model_validator(mode="before")

Classes with NO default (model= stays required)

Class	Reason
UiPathChatAnthropic	No legacy equivalent
UiPathChatAnthropicVertex	No legacy equivalent
UiPathChatFireworks	No legacy equivalent

Only Azure-backed classes ever honored UIPATH_MODEL_NAME; Bedrock and Vertex legacy defaults were hardcoded. Verified in commit 634d9fe feat: add chat models.

Approach

Each existing vendor re-export file is where:

1. The upstream uipath_langchain_client class is first imported
2. The legacy model_name default is attached inline

src/uipath_langchain/chat/
    __init__.py       # lazy __getattr__ delegates to the vendor files
    openai.py         # sets default on UiPathChat, UiPathAzureChatOpenAI, UiPathChatOpenAI
    bedrock.py        # sets default on UiPathChatBedrock + UiPathChatAnthropicBedrock; wraps UiPathChatBedrockConverse with a before-validator subclass
    vertex.py         # sets default on UiPathChatGoogleGenerativeAI (+ UiPathChatVertex alias)
    models.py         # re-exports the OpenAI family from openai.py

For classes where it works, the default is attached via pydantic FieldInfo mutation directly on the upstream class: cls.model_fields["model_name"].default_factory = factory; cls.model_rebuild(force=True). Pydantic V2 gives each class its own FieldInfo instance at class creation, so the mutation is scoped to exactly that class and doesn't leak to siblings or parents in uipath_langchain_client.

UiPathChatBedrockConverse gets a thin subclass with a @model_validator(mode="before") because its upstream ChatBedrockConverse.set_disable_streaming reads model / model_id from the raw input dict before field defaults fire.

Properties preserved

• model=... kwarg still wins over the default.
• For non-wrapped classes, the re-exported class IS the upstream class (identity preserved).
• For UiPathChatBedrockConverse, the re-export is a subclass — issubclass check still holds.
• Tracing span names unchanged.
• Lazy imports preserved: asking for UiPathChat doesn't load Bedrock/Vertex code.

Bookkeeping

• pyproject.toml: 0.10.5 → 0.10.6
• uv.lock refreshed

Test plan

• tests/chat/test_default_model_name.py — 26 tests exercising the public uipath_langchain.chat API only. For each of the 8 defaulted classes: construct cls() with no model kwarg and assert .model_name. For the 3 non-defaulted classes: assert ValidationError is raised. Plus env-var behavior, identity/alias, and subclass-relationship assertions. All pass.
• uv run pytest tests/chat/ — 192 passed
• uv run mypy src/ — clean
• just lint — clean

Generated with Claude Code