Skip to content

The Dialectic Pipeline

The heart of Dial Code's safety system - a multi-stage review process that ensures higher-quality code changes.

Overview

When changes go through Review or Safe mode, they pass through four stages:

graph LR
    A[Your Request] --> B[Planner]
    B --> C[Reviewer]
    C --> D[Resolver]
    D --> E[Learner]
    E --> F[Final Code]

Each stage adds a layer of scrutiny and refinement.

The Four Stages

1. Planner

Creates the implementation plan.

The Planner:

  • Analyzes your request
  • Identifies affected files
  • Creates code patches
  • Documents the approach

Output: A detailed plan with proposed changes.

2. Reviewer

Critically examines the plan.

The Reviewer checks for:

  • Logic errors
  • Security vulnerabilities
  • Edge cases
  • Performance issues
  • Code style violations
  • Missing error handling

Output: List of issues, concerns, and suggestions.

3. Resolver

Reconciles the plan with reviewer feedback.

The Resolver:

  • Addresses each concern
  • Refines the implementation
  • Produces the final code
  • Documents changes made

Output: Final, refined implementation.

4. Learner

Extracts patterns for future reference.

The Learner:

  • Identifies what worked well
  • Notes common issues found
  • Updates context for future tasks
  • Improves future suggestions

Output: Patterns stored for future sessions.

When It Activates

Mode Pipeline Active
Ask No
Quick No
Review Yes (light)
Safe Yes (full)

Review mode: Faster pipeline, focused on major issues. Safe mode: Thorough pipeline, comprehensive review.

Watching the Pipeline

During execution, you'll see progress:

[◎] Review mode active

[Planner] Analyzing request...
[Planner] Creating implementation plan...
[Reviewer] Checking for issues...
[Reviewer] Found 2 suggestions
[Resolver] Addressing feedback...
[Resolver] Finalizing code...
[Learner] Extracting patterns...

Ready for approval.

Benefits

Catches Issues Early

The Reviewer stage catches problems before they reach your codebase:

  • SQL injection vulnerabilities
  • Missing null checks
  • Race conditions
  • Incomplete error handling

Higher Quality Code

Multiple perspectives lead to better implementations:

  • Cleaner code structure
  • Better error messages
  • More robust edge case handling
  • Improved documentation

Continuous Learning

The Learner stage means Dial Code gets better:

  • Remembers your coding patterns
  • Avoids past mistakes
  • Suggests consistent solutions

Configuration

Control pipeline behavior in .dial/settings.json:

{
  "dialectic": {
    "reviewDepth": "standard",
    "learnerEnabled": true
  }
}

Options:

Setting Values Description
reviewDepth light, standard, thorough Review intensity
learnerEnabled true, false Enable pattern learning

Example Flow

Request: "Add rate limiting to the API endpoint"

Planner:

Plan: Add rate limiting middleware
- Create rateLimiter.js
- Import in app.js
- Apply to /api routes

Reviewer:

Issues found:
1. No configuration for rate limits
2. Missing Redis store for distributed systems
3. No bypass for health checks

Resolver:

Updated implementation:
- Added configurable limits
- Added Redis store option
- Excluded /health endpoint

Learner:

Pattern stored: Rate limiting implementations
should include configuration options and
exclude health check endpoints.

Next Steps