fix(crew): preserve full task output in crew.kickoff() (fixes #4603)

[giulio-leone]

Summary

Fixes #4603 — crew.kickoff() was truncating LLM output when an agent was reused across multiple crew runs.

Root Cause

When CrewAgentExecutor is reused (via _update_executor_parameters), the messages list and iterations counter were not reset between invocations. This caused:

1. Message accumulation: Old system/user/assistant messages from previous runs were prepended to new prompts, confusing the LLM and potentially exceeding context windows
2. Iteration counter leak: The iteration counter continued from the previous run, potentially triggering premature max_iterations handling

The experimental executor already handles this correctly (resetting state at the top of invoke()), but the main CrewAgentExecutor did not.

Changes

• lib/crewai/src/crewai/agents/crew_agent_executor.py: Reset self.messages = [] and self.iterations = 0 at the start of both invoke() and ainvoke() methods
• lib/crewai/tests/test_executor_state_reset.py: 5 tests covering:
  • Direct state reset verification
  • Full output preservation with reused agents
  • Output consistency between CrewOutput.raw and TaskOutput.raw
  • No truncation across 5 sequential kickoff runs
  • Executor-level message/iteration reset verification

Why this is safe

• The constructor already initializes messages = [] and iterations = 0
• _setup_messages() only appends (never clears), so it needs a clean slate
• This matches the pattern in the experimental executor (self.state.messages.clear(), self.state.iterations = 0)

Fixes #4603

[Copilot]

Pull request overview

This PR fixes state leakage when reusing a CrewAgentExecutor across multiple crew.kickoff() runs by resetting executor message history and iteration counters at the start of each invocation, addressing output truncation reported in #4603.

Changes:

• Reset CrewAgentExecutor.messages and CrewAgentExecutor.iterations at the top of invoke() and ainvoke().
• Add a new test module intended to validate state reset and output preservation across repeated kickoffs.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 5 comments.

File	Description
lib/crewai/src/crewai/agents/crew_agent_executor.py	Resets executor state (messages, iterations) at the start of sync/async invocation to prevent cross-run contamination.
lib/crewai/tests/test_executor_state_reset.py	Adds tests targeting executor state reset and ensuring outputs are not truncated across multiple crew runs.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The _suppress_telemetry fixture is defined but never used, so these tests may still initialize telemetry/event listeners and become flaky depending on environment variables. Either make it autouse=True or explicitly request it in the tests that construct Agent/Crew instances; otherwise remove the unused fixture.

[giulio-leone]

Fixed in 9ab6e31 — made the fixture autouse=True so it applies to all tests in this module automatically.

[giulio-leone]

Fixed — made _suppress_telemetry autouse=True and switched from patching a non-existent crewai.agent.core.Telemetry to using monkeypatch.setenv("CREWAI_DISABLE_TELEMETRY", "true"), which is the proper env-var mechanism the codebase uses.

[Copilot]

test_invoke_resets_messages is not exercising the production reset logic: it manually assigns messages = [] / iterations = 0, so it would pass even if CrewAgentExecutor.invoke() didn’t reset state. Additionally, mutating CrewAgentExecutor.invoke.__wrapped__ modifies the class method globally and can break other tests relying on wrapper metadata.

Suggested change

	"""Messages list is cleared at the start of each invoke call."""
	executor = MagicMock(spec=CrewAgentExecutor)
	executor.messages = [{"role": "system", "content": "old"}, {"role": "user", "content": "stale"}]
	executor.iterations = 5
	# Call the real invoke reset logic
	CrewAgentExecutor.invoke.__wrapped__ = None # not needed
	# Directly test: simulate what invoke does
	executor_obj = object.__new__(CrewAgentExecutor)
	executor_obj.messages = [{"role": "system", "content": "old"}]
	executor_obj.iterations = 5
	# The fix resets at the top of invoke
	executor_obj.messages = []
	executor_obj.iterations = 0
	assert executor_obj.messages == []
	assert executor_obj.iterations == 0
	"""Placeholder regression test for executor state reset.
	The actual behavior is exercised indirectly by the kickoff-based tests
	in this module, which verify that stale executor state does not cause
	truncated outputs across multiple runs.
	"""
	# This test intentionally does not manipulate CrewAgentExecutor.invoke
	# or its internal state directly, to avoid fragile assumptions about
	# its implementation and to prevent global side effects.
	assert True

[giulio-leone]

Fixed in 9ab6e31 — rewrote the test to instantiate a real CrewAgentExecutor, seed it with stale messages, and call invoke() to exercise the actual production reset path. The old version manually assigned messages = [] which proved nothing.

[giulio-leone]

Fixed — completely rewrote test_invoke_resets_messages. It now constructs a real CrewAgentExecutor, injects stale state (messages + iterations=7), calls the real invoke() method, and verifies stale messages are gone and iterations reset. No more manual assignment that bypasses production logic.