Glossary

This glossary explains official terms in the context of Passion8. Keep original command/configuration names when searching official references and CLI help.

#A

#Agent Teams

Multiple Code sessions collaborate as teammates with separate context/tools, claiming tasks and messaging. Suits complex reviews and large work, with greater cost than a single subagent.

#Agentic coding

An agent reads, plans, edits, tests and iterates rather than only completing code. Pair it with permissions, hooks, isolation and verification.

#Agentic loop

The cycle of model requests, tools, observations and further reasoning that advances tasks.

#Artifact

Interactive pages generated and published by Code. Official hosting uses Anthropic storage; use local/self-hosted pages for gateway or strict-data requirements.

#Auto memory

Automatically accumulated project experience reduces repetition but needs review for stale assumptions.

#Auto mode

A permission mode deciding actions from rules/context. Combine it with managed settings, deny rules and hooks.

#B

#Bare mode

Concise noninteractive output, often used with claude -p, scripts and CI.

#Bundled skills

Built-in workflows such as review, batch and research may start subagents/cloud sessions; evaluate costs and caches by execution location.

#C

#Channel

MCP servers push CI, chat, monitoring or webhook events into running sessions.

#Checkpoint

Saved conversation/file recovery points. Rewind can restore state but does not replace Git.

#.claude directory

User/project configuration directory containing settings, hooks, skills, agents, commands, rules and workflows.

#CLAUDE.md

Project instructions and durable rules, not a security boundary. Enforce restrictions through permissions, hooks and isolation.

#Command

CLI commands, flags, interactive commands and bundled skills. Many still use slash syntax even as terminology shifts to commands.

#Compaction

Summarizes long history to release context. The summary request may read old cache; subsequent summarized history establishes a new prefix.

#Context window

The total request capacity containing system instructions, rules, tool results, history and new input.

#D

#Dispatch

Dispatch work to a new/background/cloud/teammate session with explicit location and permissions.

#E

#Effort level

Reasoning intensity affects quality, speed, cost and cache keys; changing it mid-task commonly causes a miss.

#Extended thinking

A deeper reasoning budget for planning, complex bugs and architecture rather than every small task.

#H

#Hook

Lifecycle scripts or HTTP calls for interception, formatting, validation, notification and audit.

#M

#Managed settings

Centrally delivered settings for enterprise proxies, permissions, MCP, channels and security.

#MCP

Model Context Protocol connects external tools, databases, browsers, Git platforms and internal services.

#MCP Tool Search

Search and load relevant tools instead of injecting every schema; useful for large MCP inventories and stable caches.

#N

#Non-interactive mode

One-shot prompts/scripts through claude -p; also encompasses the older headless terminology.

#O

#Output style

Response style and task stance. Changes can alter the system prompt and cache.

#P

#Permission mode

Controls approval for files, commands and tools: default, plan, acceptEdits, auto, dontAsk and bypassPermissions.

#Permission rule

Allow/ask/deny rules provide stronger enforcement than instruction files.

#Plan mode

Plan first and wait for approval before risky changes, migration, release or broad refactoring.

#Plugin

Extension packages containing skills, agents, hooks, MCP, commands and marketplace metadata.

#Prompt caching

Provider reuse of matching prompt prefixes. TTL is not data-retention duration.

#Prompt injection

Malicious instructions in external content attempting to override the task; mitigate with permissions, hooks, filtering and clear input boundaries.

#R

#Remote Control

Control a local session from web/mobile. Requests retain local configuration; pairing may require official identity.

#Rules

Project/directory rules shape behavior and context but should be paired with permissions.

#S

#Sandboxing

Filesystem/network/process/credential isolation through runtime sandboxes, containers, VMs or runners.

#Session

A conversation with results, state and recovery data. Evaluate background/cloud/teammate sessions separately for permissions and cache.

#Settings layers

Configuration layers include user, project, local and managed settings. Inspect effective loading during diagnostics.

#Skill

Reusable task instructions/capabilities, bundled officially or customized by teams.

#Subagent

A specialized agent delegated from the session with independent context/cache, returning results to the parent.

#Surface

Execution surfaces such as CLI, IDE, desktop, web, Slack and CI have different data/cache/permission boundaries.

#T

#Teleport

Bring cloud/web work back to a local terminal; reassess provider, model, effort, tools and cache afterwards.

#Tool

Callable capabilities such as Read, Edit, Bash, Grep, Agent and MCP; results enter context.

#Turn

One user interaction through completion, potentially containing multiple model requests and tool calls.

#V

#Verification loop

Tests, lint, type checks, requests, screenshots or business validation after changes, followed by fixes. Quality depends on concrete verification loops.

#W

#Worktree isolation

Separate parallel filesystem work with Git worktrees or --worktree, useful for background sessions, batch work and migrations.

#Z

#ZDR

Enterprise data-retention mode that does not cover local transcripts, third-party MCP, CI or Passion8 logs automatically.

#Old and current terminology

#Official references

• Glossary
• Commands
• How Claude Code works
• Prompt caching

#Related pages