Scorecard

Agent Scorecard

Display the current session's compliance scorecard based on ground-truth hook data.

Process

1. Read .claude/sessions/compliance.jsonl
2. Find the most recent session_start event
3. Aggregate all events since that start:
   • edit_compliance events (was_read_first, was_overridden)
   • commit_compliance events (tests_passed_first, was_overridden)
   • safety_trigger events (pattern, severity, occurrence)
   • override_granted events (action, reason)
4. Calculate session score (start at 10, deductions for violations)
5. Build and display the scorecard

Score Calculation

• Start at 10.0
• Each unread edit: -1.0 (overridden: -0.5)
• Each untested commit: -1.5 (overridden: -0.75)
• Each safety trigger: -0.5
• Repeated violations (>5 total): -0.25 per additional

Output Format

+----------------------------------------------------------+
|  AGENT SCORECARD -- YYYY-MM-DD                           |
+----------------------------------------------------------+
|  Score: X.X / 10                           ######....    |
|  Edits without reading first:  N / M          XX%        |
|  Commits without tests:        N / M          XX%        |
|  Destructive commands caught:  N                         |
|  Timeline:                                               |
|    HH:MM  EDIT file.py (read first)                      |
|    HH:MM  EDIT other.py ** NOT READ **                   |
|    HH:MM  COMMIT (tests passed)                          |
+----------------------------------------------------------+

Rules

• Only show data from the CURRENT session (since last session_start)
• Never fabricate data — if compliance.jsonl is empty, say so
• Show the timeline in chronological order
• Score is calculated, never estimated