Skip to content

AGENTS.md

Latest commit

 

History

History
88 lines (79 loc) · 4.86 KB

AGENTS.md

File metadata and controls

88 lines (79 loc) · 4.86 KB
Raw

AGENTS.md - dot-agents operating guide

This repository stores personal agent configuration. Treat it like dotfiles for agent harnesses: useful to version, but easy to accidentally pollute with local state.

Before Editing

  1. Read README.md and docs/sources.md.
  2. Inspect the relevant agent folder before changing it.
  3. If the change depends on an external agent/config project, use opensrc or the real source repository before copying patterns.
  4. Check git status --short before and after edits.

Rules

  • Do not commit auth material, local histories, sessions, logs, caches, generated images, SQLite state, browser state, or app telemetry.
  • Keep the skills/ directory as a submodule for stanfish06/skillquarium.
  • Put agent-specific configuration under that agent's folder, not at the root, unless the file is meant to guide this repository itself.
  • Put future lifecycle hook experiments under hooks/.
  • Put refresh/install/validation helpers under scripts/.
  • Put spec-driven development notes, templates, and SDD harness experiments under spec/ (see spec/README.md). Do not force third-party SDD products into every agent home by default.
  • Update README.md when the layout changes.
  • After changing the submodule, verify with git submodule status.

Current Sources

  • Global instructions: prompts/AGENTS.md is the only copy. scripts/install.sh symlinks it to ~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, ~/.pi/agent/AGENTS.md, ~/.config/opencode/AGENTS.md, ~/.config/kilo/AGENTS.md, ~/.grok/AGENTS.md, ~/.gemini/GEMINI.md, and ~/.gemini/config/rules/AGENTS.md, and renders it to ~/.cursor/rules/agents.mdc. Do not add a per-harness instructions file.
  • Claude source: ~/.claude/settings*.json and ~/.claude/skills/graphify.
  • Codex source: ~/.codex/config.toml, ~/.codex/rules/default.rules, and ~/.codex/skills/hatch-pet.
  • Pi source: ~/.pi/agent/themes/mypi.json.
  • opencode source: ~/.config/opencode/opencode.jsonc. opencode auto-loads skills from ~/.agents/skills/ and ~/.claude/skills/, so no separate skill wiring is needed.
  • Kilo Code source: ~/.config/kilo/kilo.jsonc, wired for the OpenRouter provider (BYOK). The API key is read from the OPENROUTER_API_KEY env var.
  • Cursor source: harnesses/cursor/cli-config.json (copied to ~/.cursor/cli-config.json). approvalMode is unrestricted (Run Everything). Do not commit the live CLI file; it accumulates auth and caches. Cursor has no ~/.cursor/AGENTS.md; CLI/IDE also read project-root AGENTS.md / CLAUDE.md and <project>/.cursor/rules.
  • Antigravity (agy) source: harnesses/agy/settings.json, harnesses/agy/keybindings.json, and harnesses/agy/skills.json. Global instructions are linked to ~/.gemini/GEMINI.md and ~/.gemini/config/rules/AGENTS.md, specialist personas to ~/.gemini/config/agents/, live prompts to ~/.gemini/config/workflows/, and skills.json discovers ~/.agents/skills/.
  • Skills source: git@github.com:stanfish06/skillquarium.git, checked out at ~/.agents (the live vault every harness's skill symlinks point to) and pinned here as the skills/ submodule. The vault's skillquarium CLI (install, query, grep, build, embed; each takes --help) replaced install-skills.sh, graphify, and obsidian-cli at vault commit 2c2ee30d. scripts/install.sh still calls skills/install-skills.sh, so advancing the submodule pin past that commit requires switching the installer to skills/skillquarium install.
  • APImanac source: apis/ (catalog locally; SKILL.md refreshed from stanfish06/APImanac skill/SKILL.md on each install). Installed as $XDG_CONFIG_HOME/apimanac/config.yaml catalog_root and as skills/apimanac/SKILL.md in each harness home. Not part of the skills/ submodule. The apimanac mcp stdio server is declared in the symlinked Codex/opencode/Kilo configs, added at user scope for Claude (~/.claude.json, written by claude mcp add), and merged into ~/.cursor/mcp.json and ~/.gemini/config/mcp_config.json. Pi has no native MCP support.

Office Files

  • For any task involving Word (.docx), Excel (.xlsx/.csv), or PowerPoint (.pptx) files, check and invoke the OfficeCLI skills before acting. Use officecli-docx, officecli-xlsx, or officecli-pptx for format-specific work and officecli for general or cross-format work.
  • Load the matching specialized skill when applicable: officecli-academic-paper, officecli-data-dashboard, officecli-financial-model, officecli-pitch-deck, officecli-word-form, morph-ppt, or morph-ppt-3d. Scene-layer skills inherit their base OfficeCLI skill, so read both when the specialized skill says to do so.
  • Follow the selected skill's help-first and delivery/visual-validation rules. Use officecli help when command syntax or properties are uncertain instead of guessing.