[BUG] Empty structured task output files fall back to raw text

Description

When a task configured with output_json successfully converts its result to an empty object, TaskOutput.json_dict is {} but the output file contains the raw model text instead of {}.

Both sync and native-async execution select the file content with a truthiness check. That treats a valid empty dictionary the same as None (no structured result) and falls back to raw.

This report was prepared with AI assistance. I cannot apply repository labels as an external contributor; please add the required llm-generated label.

Steps to reproduce

1. Create a task with output_json and output_file.
2. Let structured conversion produce {} while the raw response is non-JSON text.
3. Execute the task.
4. Observe that task.output.json_dict == {}, but the file contains the raw response.

Minimal equivalent of the affected selection:

json_dict = {}
raw = "raw fallback"
content = json_dict if json_dict else raw
assert content == "raw fallback"

Expected behavior

A present structured result should be authoritative even when empty. The output file should contain {}. Only json_dict is None should fall back to Pydantic or raw output.

Screenshots/Code snippets

The same condition exists in Task._execute_core() and Task._aexecute_core():

content = (
    task_output.json_dict
    if task_output.json_dict
    else task_output.raw
)

Operating System

macOS Sonoma

Python Version

3.13 (the repository currently supports it; the issue template does not list it)

crewAI Version

Current main (3831e8b6c86f)

crewAI Tools Version

Not applicable

Virtual Environment

Venv (uv)

Evidence

A deterministic sync/async regression test produces TaskOutput(json_dict={}). On current main, both paths write raw fallback; changing the presence check to is not None makes both write valid {} JSON.

Possible Solution

Use task_output.json_dict is not None in both output-file selection paths and add sync/async behavior tests that read the actual written file.

Additional context

No open issue or pull request matching task_output.json_dict, _save_file, or empty structured task output was found before filing.

Start at Task._execute_core() and Task._aexecute_core(), where output-file content is selected, then review the deterministic sync/async regression coverage described in the issue. Done means both execution paths write {} when json_dict is an empty dictionary and still fall back only when no structured result is present.