fix(crewai): add provider metadata, route tools to metadata, drop eager serialization

[starfolkai]

Summary

Align the CrewAI integration with .agents/skills/sdk-integrations/SKILL.md. Audit turned up five issues:

• Missing metadata.provider on crewai.llm spans — spec requires every llm span to carry both metadata.model and metadata.provider. Added _provider_from_source(), which reads the provider string off the emitting CrewAI LLM object ("openai" / "anthropic" / "bedrock" / ...).
• Tools in input instead of metadata.tools — spec says tool definitions live in metadata.tools, not input. Moved both the LLM-call tools and the agent tools onto metadata.
• Token double-counting on crewai.llm — the previous rule only skipped tokens when the LiteLLM integration was patched, but CrewAI 1.x routes gpt-* to a native openai client, claude-* to a native anthropic client, etc. When those integrations are patched (as auto_instrument() does by default), they emit a leaf span with tokens and crewai.llm double-counted at every ancestor in the trace-tree rollup. The SKILL is explicit here: "Do not add 'if OpenAI is patched, skip metrics' checks — define clear ownership instead." CrewAI is always an orchestration layer that delegates to a provider SDK, so the clear rule is: crewai.llm never owns tokens. Same pattern pydantic_ai already uses for its wrapper spans. Dropped _litellm_owns_leaf_span + the token-name/prefix maps + the extra_metrics plumbing on _end_span accordingly.
• Excess serialization — dropped eager _try_to_dict / _normalize_output / _normalize_tools calls across kickoff, task, agent, LLM, and tool output paths. bt_json._to_bt_safe at log time already handles Pydantic v2 (model_dump), Pydantic v1 (dict), and dataclasses; the eager pre-pass was wasted work (same pattern as fix(claude_agent_sdk): add provider metadata, drop deep-copy serialization, allowlist hook fields #585 for claude_agent_sdk and fix(anthropic): drop redundant bt_safe_deep_copy passes in tracing #583 for anthropic).

All other metadata extractors (_causal_metadata, _crew_metadata, _task_metadata, _LLM_CONFIG_FIELDS) were already allowlist-based — no changes needed there.

Test plan

• Extended test_kickoff_llm_event_tree_parents_and_shape to emit with a real LLM source and assert metadata.provider == "openai".
• Added test_llm_tools_route_to_metadata_not_input (positive-in-metadata + negative-not-in-input).
• Replaced the two conditional token tests with a single test_llm_never_emits_token_metrics that asserts no token key (tokens, prompt_tokens, completion_tokens, prompt_cached_tokens, completion_reasoning_tokens) leaks onto crewai.llm regardless of which provider integrations are patched.
• No new cassettes. The test file docstring documents why VCR is impractical for CrewAI (pytest-vcr + httpcore + CrewAI's native OpenAICompletion model_post_init interaction bug). Direct-event tests remain the source of truth here; the LiteLLM mock_response smoke test still exercises the full crew.kickoff() path.
• cd py && BRAINTRUST_TEST_PACKAGE_VERSION=latest pytest src/braintrust/integrations/crewai/test_crewai.py -v — 17 passed
• cd py && nox -s pylint — success
• pre-commit run on the changed files — passed (ruff format, ruff check, codespell, EOF/trailing-whitespace)

🤖 Generated with Claude Code

Created by abhijeet

Slack thread