perf(huggingface_hub): optimize how dict attributes are set

[starfolkai]

Summary

Audit of py/src/braintrust/integrations/huggingface_hub/ against .agents/skills/sdk-integrations/SKILL.md. Two related fixes:

• Drop eager _try_to_dict in stream tool-call merge. _merge_tool_call_delta was running _try_to_dict(entry) (Pydantic model_dump("python")) on every streaming tool-call delta chunk just to read index / id / type / function fields. SpanImpl.log_internal already runs bt_safe_deep_copy at send time, so the eager conversion was a redundant recursive walk on the hot path. Same pattern as fix(cohere): drop redundant _try_to_dict pass in tool-call iteration #586 (cohere) and fix(google_genai): add provider metadata, route tools to metadata, drop eager serialization #590 (google_genai).
• Remove the _get_field(obj, key) dict-or-attr helper. HuggingFace's BaseInferenceType (parent of every response type and stream chunk on both matrix pins — latest=1.23.0 and floor=0.32.0) inherits from dict:

  >>> from huggingface_hub.inference._generated.types.base import BaseInferenceType
  >>> BaseInferenceType.__mro__
  (BaseInferenceType, dict, object)
  

  So .get() works uniformly on every response, chunk, delta, and nested field — the dict-or-attr dispatch was dead cleverness. The two places that read from an InferenceClient instance now use getattr(instance, key, None) directly.

Net diff: 60 insertions / 89 deletions, one file. Also picked up two small invariants that were previously implicit via _get_field:

• Tool-call-delta loop now checks isinstance(entry, dict) (tighter than is None) and drops the corresponding isinstance(incoming_fn, dict) re-check inside.
• _text_generation_extra_metadata gets an explicit if details is None: return {} guard at the top.

Everything else was already SKILL-aligned and needed no change:

• metadata.provider always set (kwarg > instance > "huggingface" fallback), metadata.model set when known.
• Tool definitions routed to metadata.tools via _CHAT_METADATA_KEYS allowlist, not into input.
• All metadata extraction is allowlist-based — no denylists anywhere in the file.
• context.span_origin.instrumentation.name = "huggingface-hub-auto" via local start_span helper.
• Only spec-listed metric keys.
• One accumulated span per stream, errors propagate, setup + wrap_huggingface_hub idempotent.

Test plan

• CI=1 BRAINTRUST_TEST_PACKAGE_VERSION=latest pytest src/braintrust/integrations/huggingface_hub/test_huggingface_hub.py — 24/24 passed (huggingface-hub 1.23.0)
• CI=1 BRAINTRUST_TEST_PACKAGE_VERSION=0.32.0 pytest src/braintrust/integrations/huggingface_hub/test_huggingface_hub.py — 22 passed, 2 pre-existing skips for text_generation on that pin (_skip_if_text_generation_unavailable)
• No cassettes re-recorded, no new mocks/fakes — used the existing VCR-backed suite.

🤖 Generated with Claude Code

Created by abhijeet

Slack thread