[BUG] kickoff_for_each clears latest replay records before returning

Description

The documented synchronous, non-streaming Crew.kickoff_for_each replay contract is not preserved at current main / CrewAI 1.15.6 (b3aaaab023a53a08db4d36c9430b3463f1efcc7d).

The replay guide says kickoff_for_each supports replay from the most recent crew run:

• docs/edge/.../replay-tasks-from-latest-crew-kickoff.mdx
• docs/v1.15.6/.../replay-tasks-from-latest-crew-kickoff.mdx

Each copied kickoff already resets the shared latest-output store before executing and then persists its records. Because synchronous kickoff_for_each runs the copies sequentially, the final copy leaves the valid latest run. The parent method then performs another self._task_output_handler.reset() immediately before returning, deleting those records.

A deterministic regression on unmodified source receives both expected batch outputs but observes [] in the replay store after return. An ordinary-kickoff replay control passes.

Steps to Reproduce

1. Pin the repository and create its Python 3.12 environment:

   git clone https://github.com/crewAIInc/crewAI.git
   git -C crewAI checkout --detach b3aaaab023a53a08db4d36c9430b3463f1efcc7d
   cd crewAI
   uv sync --python 3.12 --frozen --group dev
   

2. Copy test_kickoff_for_each_replay.py from the public evidence repository into lib/crewai/tests/.

3. Run the regression on unmodified production code:

   ./.venv/Scripts/python.exe -m pytest -q -n 0 --tb=short --disable-warnings \
     lib/crewai/tests/test_kickoff_for_each_replay.py::test_kickoff_for_each_persists_only_the_most_recent_run_for_replay
   

4. Run the ordinary-kickoff control:

   ./.venv/Scripts/python.exe -m pytest -q -n 0 --tb=short --disable-warnings \
     lib/crewai/tests/test_kickoff_for_each_replay.py::test_regular_kickoff_persists_and_replays_latest_run
   

Observed on unmodified source:

• regression: 1 failed; len(stored) is 0, expected 1;
• ordinary-kickoff control: 1 passed;
• combined module: 1 failed, 1 passed.

Expected behavior

After one synchronous kickoff_for_each call over inputs first and latest:

• both batch outputs should return;
• the replay store should contain exactly the final run with {"topic": "latest"};
• Crew.replay(final_task_id) without an input override should execute again using the final stored input.

Screenshots/Code snippets

Failing assertion on the unmodified pinned source:

assert len(stored) == 1
E       assert 0 == 1
E        +  where 0 = len([])

The root-cause-aligned candidate removes only the post-loop reset:

         if not self.stream:
             self.usage_metrics = total_usage_metrics
-        self._task_output_handler.reset()
         return results

With that deletion, the strengthened test verifies:

• final stored inputs;
• a third Task.execute_sync call during replay;
• the final interpolated description and returned output;
• a truthy persisted replay marker.

Operating System

Windows 11

Python Version

3.12.10

crewAI Version

1.15.6

crewAI Tools Version

1.15.6 (installed by the locked workspace; not involved in this test)

Virtual Environment

Venv (uv sync --python 3.12 --frozen --group dev)

Evidence

Public evidence repository:

https://github.com/HarperZ9/crewai-kickoff-replay-receipt

Read-back verification at commit cfe410c8cdce4e999bfeb9938193a45bfc9f22c8:

• public bundle verifier: MATCH, 362 checks;
• canonical result SHA-256: 2c9a17d1c6f049cecb8ef10917d4369ff68e626011f8fe222d329407d21e63aa;
• Project Telos action receipt SHA-256: e5b000658a704f1e099296a3005ddc13a8f3e6f620554c3c23365a7e38062aec;
• candidate regression/control/combined: 1/1, 1/1, 2/2;
• selected existing kickoff_for_each tests: 8/8;
• selected existing replay tests: 9/9;
• Ruff, formatting, git diff --check, and reverse patch checks: pass.

The test substitutes Task.execute_sync with deterministic TaskOutput generation, so it exercises Crew copying, interpolation, SQLite persistence, stored-input recovery, and replay without a model call. Repository network blocking remains active on the exercised Python paths, with exact loopback hosts allowed for Windows socket.socketpair.

Possible Solution

Remove the terminal task-output reset from synchronous kickoff_for_each and add the focused regression/control coverage in the evidence repository.

I would be glad to submit the one-line production patch and tests if maintainers want this behavior.

Additional context

• Scope is synchronous, non-streaming, one-task execution. Async, streaming, multi-task, and real-model behavior were not validated.
• The selected nearby suites passed; the full repository suite was not run.
• Removing the reset means kickoff_for_each([]) preserves the previous replay snapshot instead of clearing it. That empty-input contract should be an explicit maintainer decision.
• RED chronology is recorded but not cryptographically time-sealed beside an unmodified-source hash.
• This issue was prepared with AI assistance under human authorization. Per .github/CONTRIBUTING.md, it must carry the llm-generated label; I am requesting/applying that label explicitly.

Read the synchronous kickoff_for_each implementation in lib/crewai/src/crewai/crew.py, especially the post-loop task-output reset near lines 1106-1110. Copy test_kickoff_for_each_replay.py into lib/crewai/tests/ and run the two focused pytest commands from the issue. Done means both batch outputs return and the final replay record remains available for Crew.replay without an input override.