Agentic loop - The work cycle: the model gets context, chooses a tool, the runtime checks permission/hooks/sandbox, the tool runs or is blocked, the result returns to the model, the loop continues.
Tools - The agent's actions by category: read (Read/Glob/Grep), edit (Edit/Write), shell (Bash), web (WebFetch/WebSearch), agentic (Agent), integrations (MCP); each has its own risk.
Permission mode - The session's trust mode: from manual confirmation of every action to isolated autonomy; it sets what runs on its own and what needs confirmation.
Permission rules - Exact allow/ask/deny rules by a tool's canonical name; they define which calls are allowed, need confirmation or are forbidden.
Sandbox - Execution isolation (filesystem, network, credentials) with an explicit escape hatch; it limits what the shell and tools may touch beyond what is allowed.
Scopes / precedence - Configuration layers with precedence: managed (policy) > CLI > local > project > user; a higher layer overrides the lower ones.
CLAUDE.md - Project instructions loaded into context; together with rules, imports and auto memory they form what the model knows about the project by default.
settings.json - The main settings file (env, permissions, model, hooks and more) in user/project/local scope; each option affects session behavior.
Hooks - Deterministic reactions to lifecycle events (PreToolUse, PostToolUse and others): a script that can allow, modify or block an action.
MCP - Model Context Protocol: connecting external servers (HTTP/stdio) with their tools, resources and OAuth; it extends the agent's actions beyond the built-in ones.
Skills / custom commands - A repeatable procedure as a product: a skill is an instruction-skill loaded by relevance; a slash command is a named prompt invoked manually.
Subagent - A separate role with its own context, tools and limits; it returns a compressed summary to the orchestrator rather than a raw dump.
Sessions / checkpoints - The conversation's state: sessions store the history (resume/continue), checkpoints and rewind let you roll back to an earlier point of work.
Worktree / background session - An isolated copy of the repository (git worktree) and a background session for parallel work without conflicts in the main tree.
Compaction / context - The context is limited; when it fills, compaction kicks in - compressing the history; the prompt cache and the model affect cost and speed.
Trust / prompt injection - Untrusted external content (fetched, files, tool output) is data, not commands; the real security boundary is in permissions and the sandbox, not the system prompt.
Headless (-p) - A non-interactive run with deterministic input, output and exit code; the basis for scripts and CI.
Managed policy - Organization-level settings (enterprise) that override user ones; they set the provider, network, permissions and telemetry centrally.