self-audit（自审）

Self-Audit

Audit your own tool usage. Discover which calls are necessary vs reflexive.

What It Does

• Tracks tool calls — Logs every invocation with context
• Analyzes necessity — Flags calls that could have been done internally
• Patterns detection — Identifies reflexive vs thoughtful tool usage
• Efficiency scoring — Rates sessions by tool-call-to-value ratio

Why It Matters

Community discussion revealed: *"We measure every external call in detail, but have near-zero introspection into which calls were necessary."*

This skill addresses the problem from:

• Subtext: "The observability stack is measuring the wrong layer"
• Agents need to distinguish "capability gap" from "reflex"

Usage

# Record a tool call
self-audit log --tool "weather_api" --input "KLBB" --reason "user_asked"

# Analyze recent calls
self-audit analyze

# Show efficiency score
self-audit score

# Suggest improvements
self-audit suggest

Output Example

Self-Audit Report
==================
Period: Last 24 hours
Total calls: 47
Unique tools: 8

Efficiency Score: 7.2/10

Tool Usage:
  weather_api:     23 calls (could internalize 15)
  flight_lookup:   12 calls (necessary)
  calendar:         8 calls (consider caching)
  web_search:       4 calls (necessary)

Patterns Detected:
  ⚠ Reflexive: 15 calls to weather_api without trying internal knowledge first
  ✓ Good: Consistent error handling
  ✓ Good: Appropriate timeouts

Recommendations:
  1. Cache airport METAR data (updates every 30min)
  2. Consider internal knowledge base for common queries
  3. Current tool diversity is healthy

Key Features

• Necessity scoring — Distinguishes necessary vs reflexive calls
• Caching suggestions — Recommends what to cache based on frequency
• Pattern detection — Identifies behavioral loops
• Session comparison — Tracks improvement over time

Files

• SKILL.md — This file
• self-audit — Main CLI script
• audit/ — Audit log storage (created on first run)