How yubuhui/opencode-self-evolving works
This repository is a skills-centered configuration kit that equips OpenCode (and compatible agents) with self-improvement guidance and engineering standards. Its documented task workflow runs from plan review through phased implementation and test verification, then feeds reusable learning back into the skill index. Separate skills define quality, rule-writing, and PowerShell encoding practices. Workflow links reflect the README; unsampled skill internals are represented by their exact paths, without inferring implementation details.
Main entry points: plan-review/SKILL.md Plan review · phased-refactor-and-report/SKILL.md Phased refactor · test-verification/SKILL.md Test verification · auto-skill-creator/SKILL.md Auto skill creator
Components 13
Task workflow
- Plan review
skills/plan-review/SKILL.md - Phased refactor
skills/phased-refactor-and-report/SKILL.md - Test verification
skills/test-verification/SKILL.md - Completion report
skills/phased-refactor-and-report/SKILL.md
Skill evolution
- Auto skill creator
skills/auto-skill-creator/SKILL.md - Skill index
skills/INDEX.md
Engineering standards
- Quality baseline
skills/quality-baseline/SKILL.md - Safe rule writing
skills/white-bear-safe-rules/SKILL.md - PowerShell UTF-8
skills/powershell-utf8-safety/SKILL.md
Agent configuration
- Always-on guidance
AGENTS.md - Safety permissions
opencode.json.example
Other components
- User task
- OpenCode
Connections 10
- User task → Plan review: enters review
- Plan review → Plan review: revise up to twice
- Plan review → Phased refactor: approves plan
- Phased refactor → Test verification: requests verification
- Test verification → Completion report: gates completion
- Completion report → Auto skill creator: captures learning
- Auto skill creator → Skill index: registers skill
- OpenCode → Always-on guidance: loads guidance
- OpenCode → Skill index: loads skills
- OpenCode → Safety permissions: uses permissions