Agent Audit

Scan your entire OpenClaw setup and get actionable cost/performance recommendations.

What This Skill Does

1. Scans config — reads OpenClaw config to map models to agents/tasks
2. Analyzes cron history — checks every cron job's model, token usage, runtime, success rate
3. Classifies tasks — determines complexity level of each task
4. Calculates costs — per agent, per cron, per task type using provider pricing
5. Recommends changes — with confidence levels and risk warnings
6. Generates report — markdown report with specific savings estimates

Running the Audit

python3 {baseDir}/scripts/audit.py

Options:

python3 {baseDir}/scripts/audit.py --format markdown    # Full report (default)
python3 {baseDir}/scripts/audit.py --format summary     # Quick summary only
python3 {baseDir}/scripts/audit.py --dry-run             # Show what would be analyzed
python3 {baseDir}/scripts/audit.py --output /path/to/report.md  # Save to file

How It Works

Phase 1: Discovery

• Read OpenClaw config (~/.openclaw/openclaw.json or similar)
• List all cron jobs and their configurations
• List all agents and their default models
• Detect provider (Anthropic, OpenAI, Google, xAI) from model names

Phase 2: History Analysis

• Pull cron job run history (last 7 days by default)
• Calculate per-job: avg tokens, avg runtime, success rate, model used
• Pull session history where available
• Calculate total token spend by model tier

Phase 3: Task Classification

Classify each task into complexity tiers:

Tier	Examples	Recommended Models
Simple	Health checks, status reports, reminders, notifications	Cheapest tier (Haiku, GPT-4o-mini, Flash, Grok-mini)
Medium	Content drafts, research, summarization, data analysis	Mid tier (Sonnet, GPT-4o, Pro, Grok)
Complex	Coding, architecture, security review, nuanced writing	Top tier (Opus, GPT-4.5, Ultra, Grok-2)

Classification signals:

• Simple: Short output (<500 tokens), low thinking requirement, repetitive pattern, status/health tasks
• Medium: Medium output, some reasoning needed, creative but templated, research tasks
• Complex: Long output, multi-step reasoning, code generation, security-critical, tasks that previously failed on weaker models

Phase 4: Recommendations

For each task where the model tier doesn't match complexity:

⚠️ RECOMMENDATION: Downgrade "Knox Bot Health Check" from opus to haiku
   Current: anthropic/claude-opus-4 ($15/M input, $75/M output)
   Suggested: anthropic/claude-haiku ($0.25/M input, $1.25/M output)
   Reason: Simple status check averaging 300 output tokens
   Estimated savings: $X.XX/month
   Risk: LOW — task is simple pattern matching
   Confidence: HIGH

Safety Rules — NEVER Recommend Downgrading:

• Coding/development tasks
• Security reviews or audits
• Tasks that have previously failed on weaker models
• Tasks where the user explicitly chose a higher model
• Complex multi-step reasoning tasks
• Anything the user flagged as critical

Phase 5: Report Generation

Output a clean markdown report with:

1. Overview — total agents, crons, monthly spend estimate
2. Per-agent breakdown — model, usage, cost
3. Per-cron breakdown — model, frequency, avg tokens, cost
4. Recommendations — sorted by savings potential
5. Total potential savings — monthly estimate
6. One-liner config changes — exact model strings to swap

Model Pricing Reference

See references/model-pricing.md for current pricing across all providers. Update this file when prices change.

Task Classification Details

See references/task-classification.md for detailed heuristics on how tasks are classified into complexity tiers.

Important Notes

• This skill is read-only — it never changes your config automatically
• All recommendations include risk levels and confidence scores
• When unsure about a task's complexity, it defaults to keeping the current model
• The audit should be re-run periodically (monthly) as usage patterns change
• Token counts are estimates based on cron history — actual costs depend on your provider's billing