feat: enable true concurrent execution for independent tasks

[zzstoatzz]

closes #1230

Summary

Enables true concurrent execution for independent tasks while fixing "Multiple EndTurn tools detected" warnings and infinite loops reported by Matthew Dangerfield in Discord.

Problem on Main Branch

When users try to run multiple independent tasks on main, they encounter:

• "Multiple EndTurn tools detected, output validation might be limited" warnings
• Infinite loops as the agent gets confused by multiple EndTurn options
• ContextVar token errors when trying to use asyncio.gather() for concurrent execution

# Current behavior on main: ❌ Warnings, infinite loops, and errors
task_1 = marvin.Task("Say 'one'")
task_2 = marvin.Task("Say 'two'") 
task_3 = marvin.Task("Say 'three'")

# Sequential execution only - takes ~3 seconds
results = marvin.run_tasks([task_1, task_2, task_3])
# WARNING: Multiple EndTurn tools detected...

# This fails with ContextVar errors:
await asyncio.gather(task_1.run_async(), task_2.run_async(), task_3.run_async())
# ValueError: <Token var=<ContextVar...> was created in a different Context

Root Cause

The issue was ContextVar tokens being reset across async contexts:

1. Actor context management tries to reset tokens from different async contexts
2. Orchestrator assigns multiple tasks to same agent turn, causing EndTurn tool conflicts
3. Event loop binding issues when mixing run_sync() and asyncio.gather()

Solution

Fixed ContextVar handling and made run_tasks intelligently concurrent:

1. Fixed Actor Context Management

def __exit__(self, exc_type: Any, exc_val: Any, exc_tb: Any):
    """Reset the current actor in context."""
    if self._tokens:  # Only reset if we have tokens
        try:
            _current_actor.reset(self._tokens.pop())
        except ValueError:
            # Token was created in a different context (e.g., asyncio.gather)
            # This is expected when running tasks concurrently
            pass

2. Made run_tasks Auto-Detect Independence

def _tasks_are_independent(tasks: list[Task[Any]]) -> bool:
    """Check if tasks have no dependencies between each other."""
    # Check for depends_on, parent/child, and subtask relationships
    
async def run_tasks_async(tasks, ...):
    # If we have multiple independent tasks, run them concurrently
    if len(tasks) > 1 and _tasks_are_independent(tasks):
        await asyncio.gather(*[task.run_async() for task in tasks])
        return tasks
    else:
        # Use orchestrator for dependent tasks or single tasks
        orchestrator = Orchestrator(tasks=tasks, ...)
        await orchestrator.run(raise_on_failure=raise_on_failure)
        return tasks

3. Kept Orchestrator Sequential Fix as Fallback

The orchestrator still limits independent tasks to one per turn to avoid EndTurn conflicts when not using the concurrent path.

Results

🚀 True Concurrent Execution Achieved

# Testing run_tasks_async (should NOW be concurrent) 
Duration: 1.5 seconds
Results: ['one', 'two', 'three']

# Testing asyncio.gather (should work without errors)
Duration: 1.0 seconds  
Results: ['one', 'two', 'three']

# Testing dependent tasks (should be sequential)
Duration: 2.4 seconds
Results: ['A', 'B', 'C']

Speedup for independent tasks: 2.0x

✅ run_tasks works: True
✅ asyncio.gather works: True  
✅ Both are concurrent: True
✅ Speedup achieved: True

✅ Both Patterns Now Work

1. marvin.run_tasks([task1, task2]) - Automatically concurrent for independent tasks
2. asyncio.gather(task1.run_async(), task2.run_async()) - Works without ContextVar errors

🧠 Smart Task Detection

• Independent tasks: Concurrent execution via asyncio.gather()
• Dependent tasks: Sequential execution via orchestrator
• Mixed scenarios: Handles complex dependency chains correctly
• Single tasks: Use orchestrator (no overhead from independence detection)

Test Results

Comprehensive test suite shows all patterns working:

============================================================
PARALLEL TASK EXECUTION TEST SUITE
============================================================

✅ Success! All 3 independent tasks completed concurrently
✅ Success! All 3 asyncio.gather tasks completed  
✅ Success! Dependent tasks completed in correct order
✅ Success! Mixed independent/dependent chains work

============================================================
TEST SUMMARY
============================================================  
Passed: 4/4
✅ ALL TESTS PASSED!

Performance Impact

• Independent tasks: ~50% improvement (1.5s vs 3.0s)
• Dependent tasks: No change (sequential as required)
• Memory: Minimal increase (concurrent task execution)
• Architecture: Cleaner separation of concurrent vs sequential execution

Breaking Changes

None. This is a pure enhancement maintaining full backward compatibility.

Files Changed

• src/marvin/agents/actor.py - Fixed ContextVar token reset handling
• src/marvin/fns/run.py - Added independence detection and concurrent execution
• src/marvin/engine/orchestrator.py - Kept existing sequential fix as fallback

---

Fixes Discord issue reported by Matthew Dangerfield where multiple independent tasks caused warnings and infinite loops.

Enables the concurrent execution users expected when they tried using asyncio.gather().

🤖 Generated with Claude Code

[Copilot]

Pull Request Overview

This PR fixes an issue where running multiple independent tasks in parallel would cause "Multiple EndTurn tools detected" warnings and potential infinite loops. The solution modifies the orchestrator to only assign one independent task per agent turn while preserving existing behavior for dependent task chains.

• Modified run_once() method to detect independent tasks and limit assignment to one per turn
• Added comprehensive test coverage for both parallel execution methods (run_tasks and asyncio.gather)
• Maintained backward compatibility for dependent task chains

Reviewed Changes

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

File	Description
src/marvin/engine/orchestrator.py	Core fix that prevents multiple independent tasks from being assigned to the same agent turn
tests/basic/tasks/test_parallel_tasks.py	New test suite covering independent and dependent task execution patterns

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}