Self Improve

Self-Improve

Weekly meta-analysis of swarm performance. Identifies systematic improvements to skill prompts, scheduling, and coordination patterns.

Steps

1. Collect data:

• Read all files in memory/skill-health/
• Read memory/repair-log.json for recent repairs
• Read memory/token-usage.csv for cost analysis
• Read memory/cron-state.json for timing patterns
• Read memory/tx-log.json for trade outcomes

1. Analyze patterns:

• Which skills consistently score high? Why?
• Which skills are expensive but low-value?
• Are there coordination bottlenecks (analyst waiting too long for scout)?
• Is the guardian over/under-triggering halts?
• Are there time-of-day patterns in success rates?

1. Generate recommendations:

• Schedule optimizations (move skills to their optimal time)
• Prompt improvements (based on what works in high-scoring skills)
• Cost optimizations (downgrade model for simple skills)
• Coordination improvements (adjust trigger conditions)

1. Apply safe changes (optional, if confidence > 0.9):

• Adjust var: values in stoa.yml for optimization
• Add clarifying instructions to skill prompts
• Do NOT change schedules, models, or core behavior without logging

1. Write report: memory/improvement-report.json:

``json { "timestamp": "ISO", "period": "last 7 days", "metrics": { "avg_score": 4.1, "total_cost_usd": 2.34, "total_runs": 156, "failure_rate_pct": 5.2 }, "insights": ["..."], "applied_changes": ["..."], "proposed_changes": ["..."] } ``

Constraints

• This runs weekly. Take time to be thorough.
• Never reduce safety margins (stop-loss, max_drawdown, position limits)
• Proposed changes that affect trading behavior MUST be logged, not applied
• Maximum 3 auto-applied changes per run

Commit Message

guardian: self-improve weekly analysis

Exit Codes

• 0: Report generated
• 1: Insufficient data (< 20 runs in the period)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: stoaaadev
• Source: stoaaadev/stoa
• License: MIT
• Homepage: https://x.com/stoaframework

Install and usage instructions live in the source repository linked above.