/AGENTS

AGENTS.md — commands/

Local rules for this directory. Read the repo-root AGENTS.md first.

What lives here

One Markdown file per /gerdsenai: slash command. A command file is a prompt with front matter, not code. The body tells Claude Code what to do; heavy lifting is delegated to a script in ../scripts/ (invoked via ${CLAUDE_PLUGIN_ROOT}/scripts/…) or to an autonomous agent in ../agents/.

Rules for editing / adding a command

• Front matter is a contract. Every command declares description and allowed-tools. Keep allowed-tools least-privilege — list only the tools the command actually needs. Read-only/tour commands (help.md) get Read, Bash only and must have zero side effects.
• Delegate, don't inline. Put real logic in a script (one-JSON-line contract, see ../scripts/AGENTS.md) or an agent. The command orchestrates; it shouldn't reimplement.
• Heavy commands fan out. red-team, research-report, sprint-execute, agentic-ui describe a fan-out/verify/consolidate phase graph that runs through the Workflow tool in ultracode mode and sequentially otherwise. Always honor orchestration_mode / orchestration_max_agents from settings.
• Clarify-intent (opt-out). Commands should clarify the user's goal up front (1–2 questions) and auto-skip when intent is explicit/unambiguous. The canonical opt-out flag is --yes (alias -y); where --force already exists (init/agentic-ui/dashboard) it keeps its overwrite meaning AND implies the opt-out. Security carve-out: --yes alone never auto-confirms an install or an overwrite of existing files — those still require --force or a direct confirmation.
• Settings come from resolve_settings_path() (per-repo .claude/gerdsenai.local.md overrides global). Never hard-code a settings path.

When you change this directory, also update

• The command count and the commands/ tree in ../CLAUDE.md.
• The command table in ../README.md.
• ../.claude-plugin/marketplace.json only if the command set is user-visible there.