chore: type remaining Any fields on BaseAgent and Crew

[greysonlalonde]

Summary

• Replace Any annotations with proper types on BaseAgent (llm, crew, agent_executor, _request_within_rpm_limit) and remove redundant Any from Crew (planning_llm, chat_llm)
• Centralize all model_rebuild() calls in crewai/__init__.py to resolve circular forward refs (Crew, A2A configs)
• Fix pre-existing bugs exposed by stricter types: wrong attr name (stop_words → stop), wrong param type (list[BaseTool] → list[CrewStructuredTool]), missing isinstance narrowing for BaseLLM

Test plan

• python -c "import crewai" passes
• All ruff, ruff-format, and mypy hooks pass
• Core test suites pass (crew, task, agent, flow)

---

Note

Medium Risk
Medium risk because it changes Pydantic model typing/serialization (notably BaseAgent.crew and executor/LLM fields) and centralizes model_rebuild() namespace injection, which could surface new validation/forward-ref issues at import/runtime.

Overview
Tightens typing across core models by replacing several Any fields with concrete types/InstanceOf checks (e.g., BaseAgent.llm, BaseAgent.agent_executor, Crew.planning_llm/chat_llm) and adds custom serialization for BaseAgent.crew to allow either a Crew object or an id string.

Centralizes Pydantic model_rebuild() handling in crewai/__init__.py, including optional A2A config types and module namespace injection to resolve forward refs/circular imports.

Fixes runtime paths revealed by stricter types: guards for crew being a string, enforces BaseLLM instances where required (guardrails, streaming, multimodal), aligns executor stop field differences (stop_words vs stop), and corrects tool/executor parameter typing and OpenAI adapter model casting.

^{Written by Cursor Bugbot for commit d12e8c6. This will update automatically on new commits. Configure here.}

[iris-clawd]

💬 139 — Good direction tightening the Any fields. A few things to address:

🔴 CI: type-checker failing (24 errors across 6 files)

The stricter types (llm: str | BaseLLM | None, crew: Crew | str | None) propagate new mypy errors to files not touched by this PR:

• telemetry.py (lines 326, 430, 843): agent.llm.model — needs isinstance(agent.llm, BaseLLM) guard
• agent/utils.py (lines 140, 247, 358, 384): crew.knowledge, crew._train, crew.aquery_knowledge — needs narrowing for str variant
• task.py (lines 321, 346, 366): passing str | BaseLLM | None to LLMGuardrail which expects BaseLLM
• task.py (lines 649, 650, 764, 765, 816, 819): crew.task_callback, crew.id, llm.supports_multimodal — need narrowing
• crews/utils.py (line 54): llm.stream — needs isinstance guard
• openai_adapter.py (lines 191, 198): incompatible model arg type + executor type mismatch

These are all real — the type narrowing you added in crew.py is the right pattern, it just needs to be applied everywhere agent.llm and agent.crew are accessed as objects.

📝 Minor notes:

1. _validate_crew_ref is a no-op (return value). Either add actual validation or drop it — the BeforeValidator adds overhead for nothing currently.

2. sys.modules[...]\.__dict__.update(_resolve_namespace) — mutating module dicts is fairly aggressive. If another module imports base_agent before crewai.__init__ runs, you get stale references. Worth a comment explaining why this is safe, or consider update_forward_refs patterns instead.

3. calculate_usage_metrics behavior change — the old code had a fallback path (llm._token_process.get_summary()) for non-BaseLLM llms. The new code silently skips them. If that's intentional, worth a note.

Overall the concept is solid — just needs the type narrowing propagated to all downstream access sites to get CI green.

[iris-clawd]

All 24 mypy errors fixed, CI fully green across 3.10–3.13. Nice work addressing the feedback quickly.

Remaining minor nits (non-blocking):

1. _validate_crew_ref is still a passthrough no-op — consider dropping it or adding actual validation (e.g. type check) to justify the BeforeValidator overhead.

2. sys.modules[...]\.__dict__.update() — works but mutates module namespaces globally. A brief comment explaining why this is needed for forward-ref resolution would help future readers.

3. calculate_usage_metrics silently drops non-BaseLLM usage — old code had a _token_process fallback. If intentional (cleaning up dead code path), 👍. If not, worth a follow-up.

4. openai_adapter.py line 198 # type: ignore[assignment] — Runner assigned to agent_executor: InstanceOf[CrewAgentExecutorMixin] | None. The ignore works but the OpenAI adapter's executor pattern fundamentally doesn't fit the new type. Might warrant a dedicated executor type or protocol for adapter-style agents in a future PR.

Overall solid cleanup. LGTM. ✅

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[iris-clawd]

Re-approving after latest commits. CI green, LGTM ✅