planner.md

Planner Module

Goal planner for decomposing high-level goals into executable task DAGs.

Design

The planner takes a user goal (e.g. "deploy this project to a server"), breaks it into a directed acyclic graph of Tasks, and executes them in topological order with automatic verification and rollback.

• GoalDecomposer — Breaks goals into subtasks using LLM or rule-based decomposition
• DagScheduler — Topological execution of task dependencies
• TaskExecutor — Executes individual tasks with retry and verification
• VerificationEngine — Checks task success against criteria
• RollbackManager — Snapshot-based state restoration on failure
• ErrorDiagnosisEngine — Root cause analysis and remediation suggestions
• ToolLearningEngine — Learns from successful tool executions for future suggestions
• WorkflowRecorder / WorkflowPlayer — Record and replay task sequences
• PersistentTaskManager — SQLite-backed durable task queue
• TaskScheduler — Scheduled task execution with cron support
• CompositeToolRegistry — Registry for multi-step composite tools

Task Lifecycle

Pending ──▶ Running ──▶ Completed
   │           │
   ▼           ▼
Failed    RolledBack

Key Types

pub struct Task {
    pub id: TaskId,
    pub description: String,
    pub action: DesktopAction,
    pub dependencies: Vec<TaskId>,
    pub verification: Option<VerificationCriteria>,
    pub snapshot_before: bool,
    pub max_retries: u32,
    pub retry_delay: Duration,
    pub status: TaskStatus,
    pub error: Option<String>,
    pub result: Option<String>,
}

pub enum TaskStatus {
    Pending,
    Running,
    Completed,
    Failed,
    RolledBack,
}

pub struct Plan {
    pub goal: String,
    pub tasks: HashMap<TaskId, Task>,
}

pub struct VerificationCriteria {
    pub expected_output: Option<String>,
    pub expected_files: Vec<String>,
    pub timeout: Duration,
}

pub struct ErrorDiagnosisEngine {
    // Root cause analysis and remediation
}

pub struct ToolLearningEngine {
    // Experience-based tool suggestion
}

Data Flow

User Goal
    │
    ▼
GoalDecomposer::decompose()
    │
    ▼
Plan (DAG of Tasks)
    │
    ▼
DagScheduler::execute()
    │
    ├──▶ Topological sort
    ├──▶ Execute ready tasks
    ├──▶ Verify results
    ├──▶ Retry on failure
    └──▶ Rollback if needed

Implemented Features

• Goal decomposition into task DAGs
• Topological task scheduling with dependency resolution
• Task execution with retry and timeout
• Verification engine with multiple criteria types
• Rollback manager with snapshot-based restoration
• Error diagnosis with root cause analysis
• Tool learning from execution history
• Workflow recording and playback
• Persistent task queue with SQLite backend
• Scheduled task execution
• Composite tool registry for reusable multi-step tools
• Task state machine (Pending, Running, Completed, Failed, RolledBack)