create-opencode-agents

Create OpenCode Agents

Overview

OpenCode agents are specialized AI assistants configured for specific tasks with custom prompts, models, tools, permissions, and message contracts. Creating effective SGAI agents requires bounded context packets, explicit work splitting, and strict return envelopes.

When to Use

Use this skill when:

• You need a new agent for a specialized task like security auditing or code review
• Built-in agents (Build, Plan, General) don't meet your requirements
• You need to restrict tools for safety or focus
• You want custom prompts or models for specific workflows
• You're configuring agents for team use or automation

Don't use for:

• Simple tasks that built-in agents can handle
• When you just need to modify existing agent configs

Core Pattern

1. Define Purpose: Clearly state what the agent does and when to use it
2. Research Options: Review available configuration options from docs
3. Choose Configuration: Select appropriate model, tools, permissions, prompt
4. Define Protocol: Add packet intake, splitting rules, and return envelopes
5. Create Config: Write the agent configuration file
6. Test and Refine: Verify the agent works as intended

Implementation

Step 1: Define Agent Purpose

Start with a clear description of what the agent does:

description: Performs security audits and identifies vulnerabilities

Include when to use it in the description for better discoverability.

Step 2: Research Configuration Options

Key options from OpenCode docs:

• mode: "primary" (main agent) or "subagent" (specialized helper)
• model: Choose based on task (faster for planning, more capable for complex work)
• temperature: 0.0-0.2 for focused tasks, 0.6-1.0 for creative
• disable: Set true only when an agent configuration must be temporarily unavailable
• hidden: Hide internal subagents from user autocomplete
• permissions: Control tool usage (allow, ask, deny, or path/command maps)
• prompt: Custom system prompt file

Step 3: Configure for Task

For security auditing agents:

• Use subagent mode
• Restrict destructive tools (write: false, edit: false)
• Allow read-only tools (grep, read, list)
• Set permissions to deny risky operations
• Use focused prompt emphasizing security analysis

For SGAI wrapper agents:

• Use mode: primary
• Deny direct edits in the prompt; do not add parent permission.edit: deny if child developer subagents must edit
• Allow only the developer/reviewer subagents that wrapper may invoke through permission.task
• Require every delegated task to use a bounded TASK_PACKET
• Require a WORK_SPLIT map before broad fanout
• Reconcile developer/reviewer envelopes before reporting completion to the coordinator

For developer subagents:

• Use mode: subagent and usually hidden: true
• Allow edit tools only when the role actually changes files
• Read packet paths first; return NEEDS_CONTEXT instead of rediscovering broadly
• Return the strict STATUS envelope

For reviewer/verifier subagents:

• Use mode: subagent, usually hidden: true, and read-only permissions
• Review only the provided scope unless discovery is explicitly requested
• Return the strict REVIEW_STATUS envelope

For forbidden flows:

• Delete the agent file
• Remove references to it from prompts, skills, templates, examples, and GOAL composers
• Do not leave empty replacement agent files

Step 4: Create Configuration File

Create markdown file in .sgai/agent/ (project) or ~/.config/opencode/agent/ (global). For skeleton agents, edit cmd/sgai/skel/.sgai/agent/; for project overlays, edit sgai/agent/ with the complete replacement file.

---
description: Performs security audits and identifies vulnerabilities
mode: subagent
temperature: 0.1
permissions:
  edit: deny
  bash: deny
  webfetch: allow
---

You are a security expert specializing in code analysis. Focus on identifying potential security issues including:
- Input validation vulnerabilities
- Authentication and authorization flaws
- Data exposure risks
- Dependency vulnerabilities
- Configuration security issues

Provide detailed analysis with specific code locations and severity levels. Do not make changes to code.

Step 5: Test Configuration

• Invoke the agent with @mention
• Test with sample code or scenarios
• Verify tool restrictions work
• Refine prompt and config as needed

Common Mistakes

• Over-restricting: Denying too many tools makes agent useless
• Under-restricting: Allowing dangerous tools in security contexts
• Vague descriptions: Agents won't know when to invoke subagents
• No custom prompt: Agent behaves like general assistant
• Wrong mode: Primary agents for everything clutters interface
• No packet protocol: Subagents waste context rediscovering the repository
• No split map: Coordinators parallelize unsafe work or serialize independent work without reason
• No return envelope: Callers cannot reconcile completion, blockers, and verification evidence
• Legacy routing: Agents call workflow-state tools or write project-management files despite denied permissions

Required Message Contracts

Every delegating agent must know how to split, delegate, and reconcile work. Coordinator and wrapper-agent prompts must load and follow subagent-delegation-protocol skill for Task Packets, Work Splits, parallelization, and reconciliation.

Developer/execution agents return:

Load and follow subagent-return-protocol skill, then return its Execution Envelope exactly.

Reviewer/verifier agents return:

Load and follow subagent-return-protocol skill, then return its Review Envelope exactly.

Splitting and Reconciliation Rules

Use subagent-delegation-protocol skill for splitting and reconciliation rules. Do not duplicate the packet, split-map, or reconciliation schema in each agent prompt.

Quick Reference

Real-World Impact

Properly configured agents:

• Reduce security risks through restricted tools
• Improve efficiency with task-specific prompts
• Enable better collaboration with specialized roles
• Prevent accidental changes in analysis contexts