Replace the single Replan tool with three targeted alternatives:
ReviseStep, RevisePlan, and SkipStep. This gives the orchestration
agent more precise control over plan recovery — revising only the
current step, replacing all remaining steps, or skipping a step
entirely — rather than triggering a full replan each time.
Also give the orchestration agent read access to vault files so it
can resolve execution failures without routing back to the planning
agent, and update the execution agent prompt to stop it from
attempting self-recovery (gap resolution is now the orchestrator's
responsibility).
Clarify execution vs orchestration responsibilities, eliminate conditional step patterns in planning, add explicit plan completion signal, and strengthen guidance on atomic unconditional actions with outcome-based routing.
- Add context_for_next_step parameter to CompleteStep for passing execution history
- Add context parameter to CompleteTask for preserving task completion state
- Update OrchestrationResult to handle context propagation between steps
- Add debug color differentiation for agent types (Main, Execution, Orchestration, Planning)
- Reorganize SearchTypes from Helpers to Types directory
- Add justification requirement for execution deviations
- Support reasonable deviations in orchestration plan validation
- Refactor dependency service with TryResolve utility
- Add whitespace cleanup to Semaphore class
Restructure the AI workflow from a single-agent model to a specialized
multi-agent system with distinct roles:
- Add AgentType enum (Main, Orchestration, Planning, Execution) to define
agent specializations
- Replace AIControllerService and AIFunctionService with modular agent
classes in Services/AIServices/:
- MainAgent: Handles user interaction and delegates to orchestration
- OrchestrationAgent: Coordinates plan execution and step-by-step workflow
- PlanningAgent: Creates and revises execution plans
- ExecutionAgent: Executes individual plan steps
- AIController: Base class providing common agent loop functionality
- Create specialized prompts (OrchestrationPrompt, ExecutionPrompt) for
agent-specific behavior
- Add OrchestrationResult type to communicate workflow control decisions
(continue, abort, replan)
- Introduce agent-specific function scoping:
- ExecuteWorkflow for main agent
- CompleteTask for execution agent
- CompleteStep/Replan/CancelPlan for orchestration agent
- SubmitPlan/AskUserQuestionPlanning for planning agent
- Update BaseAIClass to use agentType property instead of isPlanningAgent
boolean for model selection
- Simplify ExecutionPlan and ExecutionStep types by removing execution
state tracking (moved to agent coordination)
- Remove PlanningEnabledAppendix and ExecutionStatus enum (superseded by
agent architecture)
- Add comprehensive integration tests for agent workflows
This architecture provides better separation of concerns, clearer agent
responsibilities, and more robust plan execution with explicit orchestration
control flow.
