preserve full ResultMessage field set on response_metadata (both paths)

[jasoncarreira]

Summary

_generate and _astream emit asymmetric generation_info shapes, so calling invoke vs ainvoke/astream against ClaudeCodeChatModel produces AIMessages with different response_metadata keys — even when the underlying SDK ResultMessage is identical.

Field	_generate	_astream
num_turns	✅ preserved	❌ dropped
is_error	✅ preserved	❌ collapsed into binary finish_reason
stop_reason	❌ dropped	❌ dropped
finish_reason (LangChain convention)	❌ missing	✅ emitted

Downstream impact: code that reads ai_message.response_metadata to detect e.g. max_turns truncation or to count SDK-internal turn iterations gets different results depending on which call mode was used. The _astream path loses granularity that ResultMessage originally carries.

Change

Extract a _generation_info_from_result(msg) helper that builds one canonical shape from a ResultMessage, and use it in both _generate and _astream. Net effect:

• Both paths emit the same set of keys.
• stop_reason is preserved when the SDK carries it. Access is getattr-guarded so the package keeps working on the >= 0.1.10 SDK floor (the field was added in a later SDK release).
• LangChain's finish_reason convention is honored uniformly on both paths.
• num_turns / is_error continue to surface as structured fields, not just as derived finish_reason.

Net diff: −19 / +225 across 2 files (most additions are tests + a thorough helper docstring).

Test plan

• Adds 4 new tests in tests/test_claude_chat_model.py:
  • test_astream_preserves_full_result_message_fields — streaming path emits the full set
  • test_astream_finish_reason_error_on_is_error_true — is_error → finish_reason="error" mapping correct
  • test_generate_preserves_full_result_message_fields — non-streaming path emits the matching set
  • test_generation_info_from_result_helper — direct unit test of the helper, covers the omit-when-missing semantics for stop_reason (older SDK) and usage (None)
• All 15 tests pass on SDK 0.1.10 (the version pinned in uv.lock).

Backwards compatibility

The change is purely additive on the streaming path (new keys), and additive on the non-streaming path (one new key: finish_reason). Existing callers that read only the keys that were already emitted continue to work unchanged.

Context

mimir hit this on a real production deployment after migrating to deepagents (which streams via astream). We worked around it with a local monkey-patch in our _langchain_claude_code_patches module — same pattern your existing docstrings describe for the _arun config-kwarg fix. Once this lands, the mimir-side patch can be retired.

Happy to iterate on style / scope if you'd prefer the change shaped differently.