Judge

Funktionen

• Launches meta-judge to generate evaluation criteria
• Dispatches judge sub-agent for isolated context evaluation
• Supports structured scoring with evidence-based feedback
• Generates multi-dimensional rubrics tailored to artifact type
• Provides report-only evaluation without automatic changes

Anwendungsfälle

• Evaluating code changes for quality and adherence to requirements
• Assessing documentation for clarity and completeness
• Reviewing analysis or reports for accuracy and evidence
• Getting objective feedback on any conversational output

Nicht-Ziele

• Automatically modifying or fixing the evaluated work
• Performing destructive operations
• Handling secrets or sensitive data
• Replacing human review entirely, but augmenting it

Workflow

1. Identify work to evaluate based on conversation history and arguments.
2. Extract evaluation context (original task, output, files, criteria).
3. Present evaluation scope to the user.
4. Dispatch meta-judge agent to generate evaluation specification YAML.
5. Dispatch judge agent with work context and specification YAML.
6. Validate the judge's evaluation report.
7. Present the final evaluation report and verdict to the user.