Fix #5534: propagate Task.input_files when Crew.agents is not set explicitly

[devin-ai-integration]

Summary

Fixes #5534.

When a Task had an agent assigned but the Crew was constructed without an explicit agents=[...] list (e.g. Crew(tasks=[task])), setup_agents iterated an empty list and never set agent.crew = crew. Downstream paths that rely on the crew reference — task.prompt() file injection, CrewAgentExecutor file lookups via get_all_files(crew.id, task.id), delegation tools — then silently no-op'd, so input_files attached at the Task level were ignored by the model (no error raised).

The fix adds a model_validator(mode="after") on Crew that auto-populates self.agents with any agent referenced by a task so the crew behaves identically whether agents=[...] is passed explicitly or inferred from tasks. The manager agent is explicitly excluded to preserve the existing manager_agent_in_agents validation.

Changes:

• <ref_snippet file="/home/ubuntu/repos/crewAI/lib/crewai/src/crewai/crew.py" lines="627-661" /> — new sync_agents_with_tasks validator.
• <ref_snippet file="/home/ubuntu/repos/crewAI/lib/crewai/tests/test_crew_multimodal.py" lines="463-568" /> — regression tests covering the auto-population, no duplication when explicit, agent.crew wiring through prepare_kickoff, and task.prompt() including the input files when only task.agent is set.

Review & Testing Checklist for Human

• Confirm the auto-populated ordering of Crew.agents is acceptable — agents inferred from tasks are appended in task order after any explicitly provided agents.
• Verify the fix with the exact reproduction from the issue: a real multimodal call (Crew(tasks=[task]) + input_files at kickoff time) — the model should now receive the image.
• Confirm there is no downstream consumer that relied on Crew.agents being empty when only task.agent was used (e.g. UI, serialization, or external schemas).

Notes

• No behavior change when agents=[...] is already passed (validator is a no-op: the list length doesn't change).
• The new validator runs before check_manager_llm and check_config, so cache/rpm handlers are wired onto every task agent too.
• Ran ruff check and mypy on the changed files — both clean. 4 new unit tests pass; lib/crewai/tests/test_crew.py (121 tests) and related agent/task suites remain green.

Link to Devin session: https://app.devin.ai/sessions/2258f5fb63894d9db54bbcad01c76281

---

Note

Medium Risk
Changes Crew model validation to mutate agents based on tasks, which can affect agent ordering/identity and any logic that assumed Crew.agents stayed user-provided only. Coverage is added, but this touches core kickoff/agent setup behavior used broadly.

Overview
Fixes a regression where constructing a Crew with only tasks=[...] (and each Task having task.agent) left crew.agents empty, so setup_agents never set agent.crew and task-level input_files/delegation/memory features could silently no-op.

Adds an after model_validator (sync_agents_with_tasks) that auto-populates Crew.agents with any agents referenced by tasks (excluding manager_agent) and adds regression tests ensuring agents are inferred, not duplicated when explicitly provided, and that prepare_kickoff + task.prompt() correctly see stored input_files.

^{Reviewed by Cursor Bugbot for commit 518c9af. Bugbot is set up for automated code reviews on this repo. Configure here.}

[devin-ai-integration]

Prompt hidden (unlisted session)

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring