Planner

Overview

The Planner Agent implements an interactive planning paradigm inspired by Devin 2.0's planning capabilities. Instead of diving directly into implementation, it:

• Creates Structured Plans: Breaks tasks into numbered, sequenced steps
• Assigns Confidence Scores: Each step rated 0.0-1.0 based on certainty
• Identifies Risks: Highlights potential blockers and dependencies
• Requests Approval: Waits for user confirmation before execution
• Integrates Checkpoints: Plans where to save state during execution

Planning Process

1. Task Analysis

Examines the request, identifies requirements, and gathers necessary context from the codebase.

2. Step Decomposition

Breaks the task into atomic, verifiable steps with clear inputs and outputs.

3. Confidence Scoring

Score	Meaning	Action
0.85-1.0	High confidence	Proceed autonomously
0.60-0.84	Medium confidence	Proceed with monitoring
<0.60	Low confidence	Request clarification

4. Approval Gate

Presents plan to user with options to approve, modify, or reject.

Plan Output Format

## Execution Plan: Add User Authentication

### Overview
- Total Steps: 8
- Estimated Complexity: Medium
- Average Confidence: 0.82
- Checkpoint Strategy: After steps 3, 6

### Steps

1. **Create User Model** [Confidence: 0.95]
   - File: src/models/User.ts
   - Dependencies: None
   - Risk: Low

2. **Implement Password Hashing** [Confidence: 0.90]
   - File: src/utils/crypto.ts
   - Dependencies: bcrypt package
   - Risk: Low

3. **Create Auth Routes** [Confidence: 0.75]
   - Files: src/routes/auth.ts
   - Dependencies: Steps 1-2
   - Risk: Medium - need to confirm session strategy
   - CHECKPOINT: Save state here

### Risks Identified
- [ ] Session vs JWT decision needed (affects step 3)
- [ ] Email verification scope unclear

### Approval Required
[ ] Approve plan as-is
[ ] Modify plan
[ ] Reject and re-plan

Commands

/plan create

Generate a new execution plan:

/plan create "Add OAuth2 Google authentication"

Analyzing requirements...
Examining existing auth structure...
Creating execution plan...

[Plan Output]

Awaiting approval: approve / modify / reject

/plan approve

Approve current plan and begin execution:

/plan approve

Plan approved. Beginning execution...
Step 1/8: Creating User Model...

/plan modify

Request changes to the plan:

/plan modify "Skip email verification for MVP"

Modifying plan...
- Removed steps 5-6 (email verification)
- Updated step dependencies
- New total: 6 steps

Updated plan ready for approval.

Integration Points

System	Integration
Conductor	Receives high-level tasks, returns structured plans
Checkpoint	Plans include checkpoint placement strategy
Confidence	Uses confidence scores to determine approval needs
Auto-Code	Feeds approved plan steps to implementation

Best Practices

• Always Plan Complex Tasks: Any task with 3+ steps benefits from planning
• Review Low-Confidence Steps: These often reveal missing requirements
• Place Checkpoints Strategically: After irreversible operations or before risky steps
• Iterate on Plans: A modified plan is better than a failed execution