Correction Mode

Producer-reviewer workflow for quality assurance

Correction mode implements a producer-reviewer pattern where one model generates output and another reviews it, optionally followed by fixes based on feedback.

Usage

TUI (Interactive)

/correct <producer> <reviewer> "<prompt>"
/settings        # toggle fix mode

CLI

puzldai correct "<prompt>" --producer <agent> --reviewer <agent> [--fix]

Examples

# TUI: Claude produces, Gemini reviews
/correct claude gemini "Write a REST API endpoint"

# CLI: With automatic fix
puzldai correct "Implement binary search" --producer claude --reviewer gemini --fix

How It Works

1. Production

   The producer agent generates the initial output for the task.

2. Review

   The reviewer agent critiques the output, identifying issues, suggesting improvements, and rating quality (1-10).

3. Fix (Optional)

   If --fix is enabled, the producer receives the review feedback and creates an improved version.

Configuration

TUI Settings

Configure fix mode via /settings → Collaboration tab:

Setting	Description
Fix	Enable automatic fix after review (default: off)

CLI Flags

Flag	Description
--producer	Agent that generates the initial output (required)
--reviewer	Agent that reviews and critiques (required)
--fix or -f	Enable automatic fix based on review

Use Cases

Code Review

# TUI
/correct claude codex "Implement a cache class"

# CLI
puzldai correct "Implement a cache class" --producer claude --reviewer codex --fix

The reviewer checks for:

• Correctness and edge cases
• Performance issues
• Code style and best practices
• Security concerns

Writing Review

# TUI
/correct gemini claude "Write documentation for the API"

# CLI
puzldai correct "Write documentation for the API" --producer gemini --reviewer claude

The reviewer evaluates:

• Clarity and completeness
• Technical accuracy
• Structure and organization

Fact Checking

# TUI
/correct ollama claude "Explain how SSL/TLS works"

# CLI
puzldai correct "Explain how SSL/TLS works" --producer ollama --reviewer claude

The reviewer verifies:

• Technical accuracy
• Completeness of explanation
• Common misconceptions

Example Output

─── Correction ───

Producer: claude
━━━━━━━━━━━━━━━━
```python
def binary_search(arr, target):
    left, right = 0, len(arr)
    while left < right:
        mid = (left + right) // 2
        if arr[mid] == target:
            return mid
        elif arr[mid] < target:
            left = mid + 1
        else:
            right = mid
    return -1

Reviewer: gemini ━━━━━━━━━━━━━━━━ Rating: 7/10

Issues:

1. Integer overflow possible with (left + right) // 2 for large arrays
2. Missing input validation
3. No docstring

Suggestions:

• Use left + (right - left) // 2 to prevent overflow
• Add type hints and docstring

Fixed Output (claude): ━━━━━━━━━━━━━━━━━━━━━━

def binary_search(arr: list, target: int) -> int:
    """Binary search for target in sorted array."""
    if not arr:
        return -1
    left, right = 0, len(arr)
    while left < right:
        mid = left + (right - left) // 2
        ...

Keyboard Shortcuts

Key	Action
Escape	Hide view
Ctrl+E	Return to view
Ctrl+C	Cancel

Tips

Use different model families - Cross-model review catches more issues than same-model review.

Enable —fix for iterative improvement - The producer learns from feedback and creates better output.