majiayu000 Config Tuning
Analyze and improve OpenCode configuration based on observed patterns.
Announce: "I'm using config-tuning to analyze this pattern and propose config improvements."
When to Use
- User describes a pattern they noticed (good or bad)
- User asks to improve agent behavior
- User wants to add a new workflow or skill
- After noticing repeated friction in workflows
Process
Phase 1: Understand the Pattern
Ask clarifying questions if needed:
- What behavior did you observe?
- When does this happen?
- What would you prefer instead?
Phase 2: Analyze Current Config
Read relevant files:
opencode.jsonc # Main config
AGENTS.md # Root rules
.opencode/agent/*.md # Agent definitions
.opencode/skills/*/ # Existing skills
.opencode/command/*.md # Existing commands
Phase 3: Research
If the pattern relates to:
- Agent behavior → Research OpenCode agent docs
- Prompting patterns → Research Claude best practices
- Tool usage → Research OpenCode tools docs
- Workflows → Check existing skills for similar patterns
Use exa_web_search or webfetch for OpenCode docs if needed.
Phase 4: Propose Changes
Present a structured proposal:
## Observation
[What pattern was noticed]
## Analysis
[Why this happens with current config]
## Proposed Changes
### Option A: [Name]
- File: [path]
- Change: [specific edit]
- Tradeoff: [pros/cons]
### Option B: [Name] (if applicable)
- File: [path]
- Change: [specific edit]
- Tradeoff: [pros/cons]
## Recommendation
[Which option and why]
Phase 5: Apply (with approval)
ALWAYS ASK before modifying config files.
After approval:
- Make the changes
- Explain what was changed
- Suggest how to test the improvement
Config Locations
| Config Type | Location | Purpose |
|---|---|---|
| Main | opencode.jsonc |
Models, tools, permissions |
| Rules | AGENTS.md, */AGENTS.md |
Instructions for all agents |
| Agents | .opencode/agent/*.md |
Specialized agent definitions |
| Skills | .opencode/skills/*/SKILL.md |
Workflow guides |
| Commands | .opencode/command/*.md |
Quick shortcuts |
Common Patterns
Agent not using skills enough
→ Add <skill_awareness> block to AGENTS.md
→ Improve skill descriptions to be more triggering
Agent doing X when it should do Y
→ Add explicit rule to AGENTS.md or agent definition → Consider permission restrictions
Repeated workflow friction
→ Create a new skill to guide the workflow → Or create a command for quick access
Context getting too large
→ Use progressive disclosure (read docs on-demand) → Consider DCP plugin for context pruning
Red Flags - STOP
If you catch yourself:
- Modifying config without user approval
- Making changes that affect security (permissions)
- Removing existing functionality
STOP. Always get explicit approval for config changes.