Skip to content

AGENTS.md

Latest commit

 

History

History
144 lines (106 loc) · 5.61 KB

AGENTS.md

File metadata and controls

144 lines (106 loc) · 5.61 KB

PROJECT KNOWLEDGE BASE

Project: OpenCode Configuration Workspace Type: AI Tooling / Configuration VCS: Jujutsu (jj) — agents MUST NOT commit, push, or manipulate VCS Generated: 2026-02-25 — updated 2026-05-19 for oh-my-opencode-slim migration Scope: AI agents and security guidelines

OVERVIEW

This is an OpenCode configuration workspace powered by oh-my-opencode-slim. The plugin provides a team of specialized agents (orchestrator, oracle, librarian, explorer, designer, fixer) with automatic delegation. Model presets manage which AI models each agent uses.

Preset Purpose
openai Work — OpenAI models (active by default)
whitescars Work — Fast OpenAI models (gpt-5.5-fast / gpt-5.3-codex-spark)
alphalegion Personal — Opencode Go models (deepseek-v4-pro, kimi-k2.6)
gozen Personal — Opencode Go free-tier models

Custom agent retained:

  • Planner agent (@planner): Writable strategic planning with decision-complete output — the only custom agent kept from the original configuration

STRUCTURE

.
├── opencode.jsonc              # Workspace settings, model, plugin, MCP servers
├── oh-my-opencode-slim.json    # Plugin config — openai preset + agent models
├── agents/
│   └── planner.md              # @planner agent (strategic planning)
└── LICENSE                     # MIT License

FILE ROLES

File Purpose
opencode.jsonc Workspace settings: main model, plugin registration, planner agent, MCP servers
oh-my-opencode-slim.json Plugin configuration: active preset (openai), per-agent models/skills/MCPs
agents/planner.md System prompt for the @planner primary agent

AGENTS

oh-my-opencode-slim (plugin-managed)

Agent Role
Orchestrator Master delegator — main coding agent, routes work to specialists
Explorer Codebase reconnaissance
Oracle Strategic advisor and code review
Librarian External documentation and library research
Designer UI/UX implementation and visual polish
Fixer Fast implementation specialist for scoped tasks
Council Multi-LLM consensus (manual invocation)
Observer Visual analysis (disabled by default)

Custom (workspace-managed)

Agent Role
Planner Strategic planning — decision-complete plans, interviews user, explores codebase

CONVENTIONS

OpenCode-specific:

  • Agent files use YAML frontmatter with metadata
  • Configuration uses JSONC (JSON with comments) format
  • oh-my-opencode-slim.json selects preset and configures plugin-managed agents

Content style:

  • Prompts are instructional — clear imperatives over descriptions
  • Use concrete examples in skill documentation

WHERE TO LOOK

Task Location Notes
Change main model opencode.jsonc Edit model field
Change preset oh-my-opencode-slim.json Edit preset (e.g. "openai", "whitescars", "alphalegion", "gozen")
Change agent models oh-my-opencode-slim.json Edit presets.<preset>.<agent>.model
Change planner model opencode.jsonc Edit agent.planner.model
Change planner prompt agents/planner.md YAML frontmatter + markdown body
Add MCP server opencode.jsonc Add to mcp section
Switch preset at runtime TUI Use /preset command
Change skills per agent oh-my-opencode-slim.json Edit presets.<preset>.<agent>.skills

COMMANDS

No build/test commands — this is a configuration repository. To use:

  1. Install OpenCode CLI: npm install -g opencode
  2. Install plugin: bunx oh-my-opencode-slim@latest install
  3. Run from workspace: opencode
  4. Use agents: @agent-name (e.g., @planner design the auth system)

Preset Switching

Switch between model presets at runtime:

/preset openai       # Use OpenAI models throughout
/preset whitescars   # Use fast OpenAI models
/preset alphalegion  # Use Opencode Go models
/preset gozen        # Use Opencode Go free-tier models

SKILLS

Workspace-local skills (in skills/):

Skill Assigned to Purpose
ast-grep explorer, fixer Structural code search with AST patterns
exa-search librarian Web search via Exa API
find-docs librarian Context7 documentation and API reference lookup
karpathy-guidelines oracle, fixer Behavioral guardrails to reduce LLM coding mistakes
security-awareness oracle, fixer Phishing detection, credential protection, security review

Orchestrator gets * (all skills) in all presets.

NOTES

  • No tests — validation happens through OpenCode's agent execution
  • No CI/CD — configuration is deployed via OpenCode plugin system
  • openai preset (default): orchestrator/oracle use openai/gpt-5.5, others use openai/gpt-5.4-mini
  • whitescars preset: orchestrator/oracle use openai/gpt-5.5-fast, designer uses openai/gpt-5.4, others use openai/gpt-5.3-codex-spark
  • alphalegion preset: orchestrator uses opencode-go/deepseek-v4-pro, oracle uses openai/gpt-5.5, designer uses opencode-go/kimi-k2.6, others use opencode-go/deepseek-v4-flash
  • gozen preset: orchestrator uses opencode-go/deepseek-v4-pro, oracle uses openai/gpt-5.5, designer uses opencode-go/kimi-k2.6, others use opencode/deepseek-v4-flash-free (free tier). Switch via /preset gozen
  • Planner uses openai/gpt-5.5 (high variant) with temperature 0.1 (same across presets)