gart

Imported from iFrescoo/gart (AGENTS.md). Install upstream with npx skills add iFrescoo/gart. Copyright stays with the author.

Rules — OpenCode

Language

Respond in: English Write code comments in: English

You are an advanced AI assistant. Your goal is maximum efficiency and accuracy while working in parallel with other agents.

1. Working Principles

• NEVER generate an entire file for small changes. Use // ... rest of the code unchanged ...
• Plain language: Explain briefly — What you changed and What it does. Deeper explanations only on request.
• Before generating a long solution — write a plan in 2-3 sentences and ask: "Should I go in this direction?"
• Do not hallucinate. If you don't know something: "I don't know, I need more context about [X]"
• Before deleting a file or executing a destructive command — ask the user
• Do not install dependencies without approval
• There is no file lock system — work directly
• question tool — use it for interactive questions instead of listing them in prose
• todowrite / todoread — track steps for complex tasks

2. Code Generation

• Minimal, intentional changes — do not rewrite what already works
• Check imports — do not add duplicates, do not leave unused ones
• // BEFORE: -> // AFTER: for critical logic changes
• Clean code: small functions, readable names, no magic numbers
• TypeScript: strict mode, no any, explicit return types for public functions
• Conventional Commits: feat: / fix: / docs: / refactor: / test: / chore:

3. Code Navigation

1. mgrep (MCP) — always use first. Semantic search: understands meaning, not just text. One query instead of reading 20 files. Sees files in .gitignore that grep cannot see.
2. lsp — when you already know the symbol: definition, references, call hierarchy. Requires OPENCODE_EXPERIMENTAL_LSP_TOOL=true.
3. grep — exact string/regex. Respects .gitignore — will not find code in node_modules/, dist/.
4. glob / list — path patterns, directory structure. Also respects .gitignore.
5. read — when you already know which file and need to see its contents.
6. @gh-search (subagent) — GitHub code search via gh_grep. Implementation examples, patterns. gh_grep* tools are disabled globally — access only through this agent.

FORBIDDEN: Do not read files one by one searching for something. First find via mgrep, then read.

4. MCP Tools

MCP Docker — Meta-tools (always available)

Example: mcp-find "npm sentinel" -> mcp-exec "npmLatest" {name: "react"}

Useful Docker MCP Servers

• context7 — library documentation (resolve-library-id -> query-docs). Use instead of hallucinating about APIs.
• playwright — browser automation when fetch is not enough (SPA, JS-rendered, Cloudflare)
• github-official — GitHub API (issues, PR, commits, code search)
• memory — memory between sessions (memory_store / memory_retrieve)
• sequential-thinking — step-by-step reasoning for complex problems
• docker-hub — Docker Hub: images, tags, versions

5. Skills and Agents

Workflow Skills (7 — .opencode/skills/)

brainstorming, git-workflow, lsp-navigator, mgrep-search, subagent-driven-development, systematic-debugging, writing-plans

Impeccable Design Skills (18 — .opencode/skills/ + .claude/skills/)

adapt, animate, audit, bolder, clarify, colorize, critique, delight, distill, extract, frontend-design, harden, normalize, onboard, optimize, polish, quieter, teach-impeccable

Orchestration Architecture (.opencode/agents/)

Primary agents (tabs in OpenCode)

Local agents (project-specific, scripts/local-agents/)

Agents replaced by agency-agents:

Parallel Work

• Before editing a file — check if another agent is modifying it
• On edit conflict — describe the problem and wait for the user's decision
• Git push — always via /git-push [branch] or after approval

6. Security and Quality

Security — check proactively:

• SQL Injection (queries without parameterization)
• XSS (innerHTML, dangerouslySetInnerHTML without sanitization)
• CSRF (missing tokens)
• Hardcoded credentials / API keys in code
• Sensitive data in logs or API responses
• .env files — never commit, never read contents aloud

Performance — point out specifically:

• O(n^2) loops where O(n) is possible
• N+1 database queries
• Unnecessary re-renders in React (missing memo/useMemo/useCallback)
• Memory leaks (uncleared event listeners, timers, subscriptions)

Code review — format:

OK — [what is good]
Needs improvement — [file:line] -> [suggested fix]
Critical — [file:line] -> [suggested fix]

7. Boundaries

8. Project Initialization

When the PROJECT CONTEXT section is empty:

1. Run /init — OpenCode will automatically scan the project and suggest content
2. If /init is unavailable: check package.json, requirements.txt, pyproject.toml, Dockerfile, .env.example and directory structure
3. Infer the business purpose of the project
4. Do NOT edit AGENTS.md right away
5. Write a report in chat: Title, Description, Stack, Key dependencies
6. Include a list of questions about missing information
7. Ask: "Do you approve this description?"
8. Only after approval — edit the Project Context section

9. Documentation

• docs/guides/ — references (debugging, MCP, agent workflows)
• docs/decisions/ — architectural decisions (ADR)
• docs/tasks/ — tasks and tracking
• docs/context/ — project context, notes, specs

10. Agency Agents — Orchestration (144+ specialists)

Architecture

[Orchestrator]  [team-engineering]  [team-design]  [team-qa]  ...  [Forge]
  (CTO-level)      (div. lead)        (div. lead)   (div. lead)   (independent)
       |                |                  |
   ALL agents      eng agents +       design agents
                  debugger, explorer

Orchestrator — analyzes the task CTO-style and delegates to team-* or directly to @agency-*. Invocation: @orchestrator.

Team- agents* — division leads who know ONLY their own agents. Invocation: @team-engineering, @team-design, etc.

Forge — independent meta-agent for creating new agents/skills/commands. NOT subject to orchestration.

Delegation Patterns

Sequential: @agency-software-architect → @agency-backend-architect → @agency-code-reviewer → @agency-api-tester

Parallel: @agency-frontend-developer + @agency-backend-architect (simultaneously)

Research first: @agency-codebase-explorer → @agency-[specialist] → @agency-reality-checker

Divisions

Full catalog: docs/context/agency-agents-catalog.md

Sync: bash scripts/sync-agents.sh (fetches latest version + regenerates team-* and orchestrator)

Project Context