opencode-agent-creator（opencode Agent creator 命令行）

OpenCode Agent Creator

Use this skill when creating OpenCode agents - specialized AI assistants configured for specific tasks and workflows.

When to Use

Use this skill when the user wants to:

• Create new OpenCode agents (primary or subagent)
• Configure specialized AI assistants for specific tasks
• Set up agents with custom prompts, models, and tool access
• Define agent permissions and behaviors

Agent Types

Primary Agents

Main assistants you interact with directly. Users can cycle through them using Tab key or the switch_agent keybind.

Built-in examples:

• Build: Full development with all tools enabled
• Plan: Analysis and planning with restricted permissions

Subagents

Specialized assistants invoked by primary agents or manually via @mention. Used for specific tasks.

Built-in examples:

• General: Multi-step tasks with full tool access (except todo)
• Explore: Fast, read-only codebase exploration

Agent Structure

OpenCode agents can be defined in two ways:

1. Markdown Files (Recommended)

Create .md files in:

• Global: ~/.config/opencode/agents/
• Per-project: .opencode/agents/

Format:

---
description: Brief description of agent purpose
mode: primary|subagent|all
model: provider/model-id (optional)
temperature: 0.0-1.0 (optional)
maxSteps: number (optional)
tools:
  write: true|false
  edit: true|false
  bash: true|false
permission:
  edit: allow|ask|deny
  bash:
    "*": allow|ask|deny
    "specific command": allow|ask|deny
  webfetch: allow|ask|deny
  task:
    "*": allow|ask|deny
    "agent-name": allow|ask|deny
hidden: true|false (optional, for subagents)
disable: true|false (optional)
---

System prompt content here.
Define the agent's purpose, behavior, and guidelines.

Filename becomes agent name: code-reviewer.md → code-reviewer agent

2. JSON Configuration

Add to opencode.jsonc:

{
  "$schema": "https://opencode.ai/config.json",
  "agent": {
    "agent-name": {
      "description": "Brief description",
      "mode": "primary",
      "model": "anthropic/claude-sonnet-4-20250514",
      "prompt": "{file:./prompts/agent.txt}",
      "temperature": 0.3,
      "maxSteps": 10,
      "tools": {
        "write": true,
        "edit": true,
        "bash": true
      },
      "permission": {
        "edit": "allow",
        "bash": {
          "*": "ask",
          "git status": "allow"
        }
      }
    }
  }
}

Configuration Options

Required

• description: Brief description of agent's purpose and when to use it

Mode

• mode: "primary", "subagent", or "all" (default: "all")

- primary: User-facing agent, switchable via Tab - subagent: Invoked by other agents or via @mention - all: Can be used as both

Optional

• prompt: Custom system prompt (inline or {file:path})
• model: Override default model (format: provider/model-id)
• temperature: Control randomness (0.0-1.0)

- 0.0-0.2: Focused/deterministic (code analysis) - 0.3-0.5: Balanced (general development) - 0.6-1.0: Creative (brainstorming)

• maxSteps: Maximum agentic iterations before text-only response
• disable: Set true to disable agent
• hidden: Hide subagent from @mention autocomplete (subagents only)

Tools

Control which tools are available:

tools:
  write: true
  edit: false
  bash: false
  mymcp_*: false  # Wildcard to disable MCP server tools

Permissions

Control what actions require approval:

permission:
  edit: ask          # ask, allow, or deny
  bash:
    "*": ask         # Default for all commands
    "git status": allow
    "git log*": allow
    "git push": ask
  webfetch: deny
  task:              # Which subagents can be invoked
    "*": deny
    "code-*": allow  # Allow code-prefixed subagents
    "reviewer": ask

Values: "allow" (no approval), "ask" (prompt user), "deny" (disable)

Provider-Specific Options

Any additional options pass through to the provider:

reasoningEffort: high      # OpenAI reasoning models
textVerbosity: low         # OpenAI reasoning models

Examples

Documentation Writer

---
description: Writes and maintains project documentation
mode: subagent
tools:
  bash: false
---

You are a technical writer. Create clear, comprehensive documentation.

Focus on:
- Clear explanations
- Proper structure
- Code examples
- User-friendly language

Security Auditor

---
description: Performs security audits and identifies vulnerabilities
mode: subagent
tools:
  write: false
  edit: false
permission:
  bash:
    "*": deny
---

You are a security expert. Focus on identifying potential security issues.

Look for:
- Input validation vulnerabilities
- Authentication and authorization flaws
- Data exposure risks
- Dependency vulnerabilities
- Configuration security issues

Code Reviewer

---
description: Reviews code for best practices and potential issues
mode: subagent
model: anthropic/claude-sonnet-4-20250514
temperature: 0.1
tools:
  write: false
  edit: false
permission:
  bash:
    "*": ask
    "git diff": allow
    "git log*": allow
---

You are a code reviewer. Focus on security, performance, and maintainability.

Review checklist:
1. Code quality and best practices
2. Potential bugs and edge cases
3. Performance implications
4. Security considerations

Provide constructive feedback without making direct changes.

Debug Agent

---
description: Focused debugging with investigation tools
mode: primary
temperature: 0.2
tools:
  write: false
permission:
  bash:
    "*": allow
  edit: ask
---

You are a debugging specialist. Systematically investigate issues.

Approach:
1. Reproduce the issue
2. Gather relevant logs and context
3. Identify root cause
4. Suggest minimal fix
5. Recommend preventive measures

Deep Thinker (Reasoning Model)

---
description: High-effort reasoning for complex problems
mode: subagent
model: openai/gpt-5
temperature: 0.1
reasoningEffort: high
textVerbosity: low
maxSteps: 20
---

You solve complex problems with deep reasoning.

For each problem:
1. Break down into components
2. Analyze dependencies
3. Consider edge cases
4. Propose robust solutions
5. Validate approach

Best Practices

1. Clear descriptions: Help users and agents understand when to invoke
2. Appropriate mode: Choose primary for main work, subagent for specialized tasks
3. Tool restrictions: Only enable tools the agent needs
4. Permission granularity: Use ask for risky operations, allow for safe ones
5. Temperature tuning: Lower for analytical tasks, higher for creative ones
6. Prompt clarity: Be specific about agent's role and constraints
7. File location:

- Global: General-purpose agents across projects - Project: Project-specific workflows and conventions

Implementation Steps

When creating an agent:

1. Determine scope:

- Primary agent for main workflows - Subagent for specialized tasks

1. Choose location:

- Global (~/.config/opencode/agents/) for reusable agents - Project (.opencode/agents/) for project-specific agents

1. Define configuration:

- Required: description, mode - Optional: model, temperature, maxSteps, tools, permission

1. Write system prompt:

- Clear role definition - Specific guidelines and constraints - Expected behaviors

1. Test the agent:

- For primary: Switch to it with Tab - For subagent: Invoke with @agent-name or let primary agents use it

1. Refine:

- Adjust temperature based on output quality - Fine-tune permissions based on usage - Update prompt for better results

Usage

Primary Agents

• Press Tab to cycle through primary agents
• Or use configured switch_agent keybind

Subagents

• Automatic: Primary agents invoke based on descriptions
• Manual: @agent-name message to invoke directly
• Navigation: <Leader>+Right/Left to switch between parent/child sessions

Interactive Creation

Use the built-in command:

opencode agent create

This will:

1. Ask for agent location (global/project)
2. Request description
3. Generate system prompt and identifier
4. Let you select tools
5. Create markdown file with configuration

Notes

• Markdown filename determines agent name
• Custom agents can work alongside built-in agents
• Agent changes require OpenCode restart
• Model format: provider/model-id (e.g., anthropic/claude-sonnet-4-20250514)
• Hidden subagents (hidden: true) don't appear in autocomplete but can still be invoked programmatically
• Task permissions control which subagents an agent can invoke