Glossary

Agent — A specialized AI sub-assistant with its own instructions, model override, and optional tool restrictions. Defined in .ai/agents/*.md. See Agents.

Approval Policy — Codex’s coarse-grained access control. One of: suggest (propose only), auto-edit (edit files, ask before commands), or full-auto (no manual approval). dotai maps fine-grained Permissions to the closest Codex policy.

Checkpoint — A snapshot of working tree state saved before making changes, enabling rollback.

Compaction — Summarizing earlier conversation history to free context window space when approaching token limits.

Context Window — The fixed token budget available to the agent for a single session. IgnorePattern and Directive scoping manage what occupies this limited space.

Directive — A persistent, text-based instruction that shapes agent behavior. The unified abstraction for Claude Code’s CLAUDE.md/rules, Cursor’s .mdc rules, Codex’s AGENTS.md, and Copilot’s .github/copilot-instructions.md/.github/instructions/. Defined in .ai/directives/*.md. See Directives.

Enterprise Scope — The highest-precedence scope tier. Organization-wide policy that cannot be overridden.

Frontmatter — YAML metadata at the top of a markdown file, delimited by --- lines. Used to declare properties like scope, alwaysApply, appliesTo.

Glob — A file path pattern using wildcard syntax (e.g., *.tsx, src/**/*.test.ts). Used in directive appliesTo, permission pattern, and ignore patterns.

Hook — An event handler fired at specific agent lifecycle points. Supported by Claude Code (full), GitHub Copilot (most events), and Cursor (partial). Not supported by Codex or OpenCode. Defined in .ai/config.yaml. See Hooks.

IgnorePattern — A gitignore-style pattern excluding files from the agent’s context and indexing. Restricted to project and user scope. See Ignore Patterns.

Local Scope — The lowest-precedence scope tier. Machine-specific overrides not committed to VCS.

Matcher — A filter on a Hook: tool name for tool-use hooks, glob pattern for file-edit hooks.

MCP (Model Context Protocol) — An open protocol for connecting AI agents to external tool/data providers. All six target tools support MCP. See Tool Servers.

Permission — An access control rule: tool name + optional argument pattern + decision (allow/deny/ask). Stored at finest granularity, projected to coarser models during generation. See Permissions.

Project Scope — The second-highest scope tier. Shared config committed to VCS and used by all contributors.

Scope — One of four hierarchical tiers (enterprise > project > user > local) determining where config is defined and its precedence. See Concepts Overview.

Setting — A key-value configuration entry for model selection, feature flags, or tool behavior. Defined in .ai/config.yaml. See Settings.

Skill — A portable package of domain knowledge. A SKILL.md file inside a named directory, optionally with scripts/, references/, assets/. Invoked via /skill-name. Defined in .ai/skills/<name>/SKILL.md. See Skills.

Slash Command — A user action triggered by /<name>. Skills are invoked as slash commands.

ToolServer — An external provider connected via MCP. Three transports: stdio, http, sse. Configured in .ai/config.yaml. See Tool Servers.

Transport — MCP connection mode: stdio (subprocess), http (stateless), or sse (streaming).

User Scope — The third-precedence scope tier. Personal preferences stored outside the project directory.