agent-framework/dotnet/tests/Microsoft.Agents.AI.Workflows.UnitTests/MagenticE2E_TestPlan.md

Magentic E2E Test Plan

Overview

This document outlines a comprehensive plan for adding end-to-end (E2E) tests for the Magentic orchestrator, similar to the existing HandoffOrchestrationTests.cs and smoke test patterns. These tests will drive a fully-built workflow through MagenticWorkflowBuilder.Build() and verify correct behavior at every logical fork in the orchestrator logic.

Background

Magentic Orchestrator Logic Flow

Based on the analysis of MagenticOrchestrator.cs, the orchestrator follows this decision tree:

TakeTurn (Initial)
  └─> UpdatePlanAndDelegateAsync
       ├─> Create/Update Plan via MagenticManager.UpdatePlanAsync
       ├─> Emit MagenticPlanCreatedEvent or MagenticReplannedEvent
       └─> If requirePlanSignoff: SubmitPlanReviewRequestAsync
           └─> ProcessPlanReviewAsync (on human response)
               ├─> If Approved: DelegateToTeamAsync
               └─> If Revise: Add review to chat, UpdatePlanAndDelegateAsync
       └─> If !requirePlanSignoff: DelegateToTeamAsync

DelegateToTeamAsync
  └─> RunCoordinationRoundAsync
       ├─> CHECK: Hit round limit? → Yield termination message, set IsTerminated
       ├─> CHECK: Hit reset limit? → Yield termination message, set IsTerminated
       ├─> Increment RoundCount
       ├─> UpdateProgressLedgerAsync (with retries)
       │    └─> On exception (non-cancellation): ResetAndReplanAsync
       ├─> Emit MagenticProgressLedgerUpdatedEvent
       ├─> CHECK: IsRequestSatisfied? → PrepareFinalAnswerAsync
       ├─> CHECK: IsInLoop OR !IsProgressBeingMade? → Increment StallCount
       │    └─> Else: Decrement StallCount (min 0)
       ├─> CHECK: IsStalled (StallCount > MaxStallCount)? → ResetAndReplanAsync
       ├─> Validate NextSpeaker → Fallback to first participant if empty
       ├─> CHECK: Invalid NextSpeaker? → Warning + PrepareFinalAnswerAsync
       └─> Send instruction + TurnToken to next agent

PrepareFinalAnswerAsync
  └─> Get final answer from manager, yield output, set IsTerminated

ResetAndReplanAsync
  └─> Reset context, send ResetChatSignal, UpdatePlanAndDelegateAsync

Test Categories

1. Happy Path Tests

Test Name	Description	Expected Behavior
Task_Completes_When_RequestSatisfied	Manager reports task satisfied on first coordination round	Workflow yields final answer and terminates
Task_Delegates_To_Correct_Agent	Manager selects specific agent as next speaker	Selected agent receives TurnToken
Task_Completes_After_Multiple_Rounds	Task requires multiple coordination rounds before completion	Each round delegates to specified agent, final answer produced when satisfied
PlanSignoff_Disabled_Proceeds_Immediately	With requirePlanSignoff=false	Workflow proceeds to team delegation without plan review request

2. Plan Review Tests (Human-in-the-Loop)

Test Name	Description	Expected Behavior
PlanReview_Approved_Proceeds	Human approves initial plan	Workflow continues to DelegateToTeamAsync
PlanReview_Revised_Triggers_Replan	Human requests plan revision	Revision added to chat, plan updated, MagenticReplannedEvent emitted
PlanReview_Multiple_Revisions	Human revises multiple times	Each revision triggers replan until approved
PlanReview_On_Stall_Replan	Stall triggers replan with plan signoff	Plan review request sent with IsStalled=true

3. Limit Enforcement Tests

Test Name	Description	Expected Behavior
MaxRoundLimit_Terminates_Workflow	RoundCount exceeds MaxRoundCount	Workflow terminates with "maximum round count limit" message
MaxResetLimit_Terminates_Workflow	ResetCount exceeds MaxResetCount	Workflow terminates with "maximum reset count limit" message
MaxStallCount_Triggers_Reset	StallCount exceeds MaxStallCount	ResetAndReplanAsync called, ResetChatSignal sent

4. Stall Detection Tests

Test Name	Description	Expected Behavior
Stall_IsInLoop_Increments_StallCount	ProgressLedger reports IsInLoop=true	StallCount incremented
Stall_NoProgress_Increments_StallCount	ProgressLedger reports IsProgressBeingMade=false	StallCount incremented
Progress_Made_Decrements_StallCount	ProgressLedger reports progress being made	StallCount decremented (min 0)
Consecutive_Stalls_Trigger_Reset	Multiple stalls in a row exceed MaxStallCount	Reset and replan triggered

5. Progress Ledger Validation Tests

Test Name	Description	Expected Behavior
ProgressLedger_Retry_On_Parse_Failure	First attempt fails JSON parsing	Retry up to MaxProgressLedgerRetryCount times
ProgressLedger_Max_Retries_Triggers_Reset	All retry attempts fail	ResetAndReplanAsync called
ProgressLedger_Updated_Event_Emitted	Valid progress ledger update	MagenticProgressLedgerUpdatedEvent emitted

6. Next Speaker Validation Tests

Test Name	Description	Expected Behavior
NextSpeaker_Empty_Falls_Back_To_First	ProgressLedger returns empty next_speaker	First team member selected, warning emitted
NextSpeaker_Invalid_Triggers_FinalAnswer	next_speaker doesn't match any team member	Warning emitted, final answer prepared
NextSpeaker_Valid_Delegates_Correctly	Valid team member specified	TurnToken sent to correct executor

7. Event Emission Tests

Test Name	Description	Expected Behavior
Initial_Plan_Emits_PlanCreatedEvent	First plan creation	MagenticPlanCreatedEvent emitted with full task ledger
Replan_Emits_ReplannedEvent	Plan update after initial	MagenticReplannedEvent emitted
Warning_Events_On_Errors	Various warning conditions	WorkflowWarningEvent emitted with appropriate message

8. Checkpoint/Resume Tests

Note: Direct checkpoint-state inspection tests (Checkpoint_Saves_TaskContext, Checkpoint_Preserves_ProgressLedger) are skipped — the serialized checkpoint format is an internal implementation detail. Checkpoint resume is instead verified behaviorally through plan-review workflows that pause and resume across checkpoint boundaries.

Test Name	Description	Expected Behavior
Checkpoint_Saves_TaskContext	Workflow checkpointed mid-execution	Skipped — internal format
Checkpoint_Resume_Continues_Correctly	Resume from checkpoint	TaskContext restored, execution continues from saved state (behaviorally covered by plan-review tests)
Checkpoint_Preserves_ProgressLedger	Resume preserves ledger state	Skipped — internal format

9. Edge Cases

Test Name	Description	Expected Behavior
Empty_Team_Handling	No participants added	Appropriate error or fallback behavior
Single_Agent_Team	Only one participant	Workflow functions correctly with single agent
Instruction_Message_Sent_When_Present	ProgressLedger has instruction_or_question	Instruction added to chat history and sent
Terminated_Context_Rejects_New_Messages	After termination, new messages arrive	InvalidOperationException thrown

Implementation Approach

Test Infrastructure

Following the patterns established in HandoffOrchestrationTests.cs:

1. TestReplayAgent for Manager: Configure to return specific plan/progress ledger responses
2. TestEchoAgent for Participants: Simple agents that echo or return controlled responses
3. MockChatClient: For fine-grained control of LLM responses
4. WorkflowRunResult: Capture updates, outputs, checkpoints, and pending requests
5. RunWorkflowAsync helper: Execute workflow and collect results

Key Test Helpers Needed

// Helper to create manager agent with specific responses
internal static TestReplayAgent CreateManagerAgent(
    List<List<ChatMessage>> planResponses,
    List<List<ChatMessage>> progressLedgerResponses,
    List<List<ChatMessage>> finalAnswerResponses);

// Helper to create progress ledger JSON responses
internal static ChatMessage CreateProgressLedgerResponse(
    bool isRequestSatisfied,
    bool isInLoop,
    bool isProgressBeingMade,
    string nextSpeaker,
    string instructionOrQuestion);

// Helper to build and run Magentic workflow
internal static Task<WorkflowRunResult> RunMagenticWorkflowAsync(
    AIAgent manager,
    List<AIAgent> team,
    List<ChatMessage> input,
    bool requirePlanSignoff = false,
    TaskLimits? limits = null,
    ExecutionEnvironment environment = ExecutionEnvironment.InProcess_Lockstep);

Test File Structure

Create new test file: MagenticOrchestrationTests.cs

dotnet/tests/Microsoft.Agents.AI.Workflows.UnitTests/
├── MagenticOrchestrationTests.cs  (NEW - E2E tests)
├── MagenticOrchestratorTests.cs   (existing - protocol tests)
├── MagenticManagerTests.cs        (existing - manager unit tests)
├── MagenticProgressLedgerTests.cs (existing - ledger unit tests)
└── TestProgressLedgerState.cs     (existing - reuse for E2E)

Test Execution Order

1. Phase 1: Basic Flow Tests

   • Happy path completion
   • Single agent delegation
   • Multiple rounds

2. Phase 2: Plan Review Tests

   • Approval flow
   • Revision flow

3. Phase 3: Limit and Stall Tests

   • Round limits
   • Reset limits
   • Stall detection and recovery

4. Phase 4: Error Handling Tests

   • Invalid next speaker
   • Progress ledger failures

5. Phase 5: Checkpoint Tests

   • Save and restore

Success Criteria

• All logical forks in MagenticOrchestrator are covered by at least one test
• Tests use the same patterns as HandoffOrchestrationTests
• Tests run against fully-built workflows (not isolated components)
• Each test verifies specific event emissions and state changes
• Tests cover both requirePlanSignoff=true and false paths
• Checkpoint/resume functionality is verified

Estimated Implementation

Component	Estimated Tests	Complexity
Happy Path	4	Low
Plan Review	4	Medium
Limits	3	Low
Stall Detection	4	Medium
Progress Ledger	3	Medium
Next Speaker	3	Low
Events	3	Low
Checkpoints	3	High
Edge Cases	4	Medium
Total	~31 tests

Dependencies

• Existing test helpers: TestReplayAgent, TestEchoAgent, TestProgressLedgerState
• Existing execution infrastructure: RunWorkflowAsync, CheckpointManager
• Extensions may be needed for Magentic-specific assertions