fix(instructor): add model/provider metadata, drop __future__ import

[starfolkai]

Summary

Audit of py/src/braintrust/integrations/instructor/ against .agents/skills/sdk-integrations/SKILL.md turned up two deviations. Everything else was already spec-compliant (task-typed parent, no token metrics on parent to avoid double-counting the OpenAI/Anthropic llm child, allowlist-only metadata, no eager model_dump()/bt_safe_deep_copy, no try/except around span.log, instrumentation.name = "instructor-auto", CompositeFunctionWrapperPatcher over the stable instructor.core.client surface, VCR-backed tests, subprocess auto_instrument test).

• Missing metadata.model / metadata.provider on the parent task span. SKILL's "Framework llm-like spans that delegate to a separately instrumented client" paragraph says the framework span stays task but keeps metadata.model / metadata.provider for attribution. Instructor is exactly that shape — it calls into a wrapped OpenAI/Anthropic client which emits the llm leaf. Added:
  • _model_name(kwargs, instance) — reads kwargs["model"] (Instructor passes it straight through) with instance.default_model fallback.
  • _provider_name(instance) — reads instance.provider.value from Instructor's built-in Provider enum. Value is already the short string we want ("openai", "anthropic", "cohere", "bedrock", "mistral", …). Falls back to .name.lower() for older builds.
  • _build_metadata now emits model + provider alongside the existing 5 keys (response_model, mode, max_retries, retry_count, validation_errors). Still a strict allowlist — no raw kwargs / no denylist.
• from __future__ import annotations at tracing.py:24 violates CLAUDE.md rule 8 ("Do not add … unless absolutely required"). Only integration doing this. Removed; requires-python = ">=3.10" means str | None and dict[str, Any] already work at runtime unchanged.

Explicitly not changed (verified against SKILL and left alone):

• Parent span type stays task, not llm — provider integrations own the llm leaf and its token accounting.
• input still allowlist {response_model, mode, messages}; no dump of raw kwargs.
• Pydantic outputs still passed through raw — Braintrust serializes at send time; no model_dump() in tracing code.
• Existing tuple/list unwrap in _extract_output is structural (unwraps create_with_completion's (model, raw_completion) tuple), not serialization.

Test plan

• cd py && nox -s "test_instructor(latest)" — 8/8 passed (instructor 1.15.4, openai 2.45.0)
• cd py && nox -s "test_instructor(1.11.0)" — 8/8 passed
• Extended TestInstructorOpenAISpans.test_instructor_openai_single_success and test_instructor_openai_retries_then_succeeds to assert meta["model"] == "gpt-4o-mini" and meta["provider"] == "openai".
• Existing VCR cassettes reused unchanged — attribution reads from client instance + kwargs, not the HTTP response.
• Subprocess auto_instrument test still passes.

🤖 Generated with Claude Code

Created by abhijeet

Slack thread