Glossary

Concise definitions for terms used throughout this guide. Each links to the chapter where the concept is covered in depth.

agent harness The runtime that drives the model → tool → model loop: reads user input, calls the provider, dispatches tool calls, feeds results back, and repeats until a terminal condition. Caliban is an agent harness. See What Is Caliban?.

auto-memory Per-project notes written by the model itself into a designated memory file. Injected into the system prompt on subsequent sessions. See Auto-Memory.

checkpoint A snapshot of the conversation state (messages + file-tool pre-images) taken before each prompt. Used by /rewind to restore a prior state. See Checkpoints & Rewind.

compaction The process of summarising or truncating conversation history when the context window approaches its limit, allowing the session to continue. See Context & Compaction.

headless / print mode Non-interactive operation via -p / --print. Caliban drives the agent without a TUI and emits text or structured JSON output to stdout. See Print Mode and The stream-json Protocol.

hook An event-driven callback executed by an external command, HTTP endpoint, MCP tool, or in-process handler at defined points in the agent lifecycle (e.g. before_tool, SessionStart). See Hooks.

MCP server A Model Context Protocol server that exposes additional tools to caliban over stdio, HTTP/SSE, or streamable-HTTP transports. Caliban discovers and manages MCP servers via its settings. See MCP Servers.

memory tier One of the three layers of context prepended to the system prompt: global (~/.claude/CLAUDE.md), project (<workspace>/CLAUDE.md), and auto-memory (model-written notes). See Memory Tiers.

message IR The provider-neutral internal representation of conversation messages used by caliban-common. All providers translate to and from this IR so the agent core stays provider-agnostic. See Architecture & ADRs (ADR 0006).

output style A named instruction set (Default, Proactive, Explanatory, Learning, or custom) that shapes how the model formats and explains its responses. See Output Styles.

permission mode A named preset that sets the default disposition for tool-call permission checks. Modes include default, acceptEdits, plan, auto, dontAsk, and bypassPermissions. See Permission Modes.

plugin A self-contained bundle of skills, hooks, agents, MCP server configs, and output styles distributed as a directory with a plugin.json manifest. See Plugins.

provider An adapter that translates caliban's message IR to and from a specific model API (Anthropic, OpenAI, Ollama, Google, Bedrock, Vertex). See Supported Providers.

router The caliban-model-router layer that selects a provider+model for each request based on configured rules, purpose keys, fallback chains, circuit breakers, and capability requirements. See The Model Router.

sandbox An OS-level confinement layer (macOS Seatbelt or Linux bubblewrap) applied to shell and file tools to restrict what they can access on the host. See The OS Sandbox.

session A persisted conversation: a named JSON file on disk containing the full message history for a continuous exchange. See Sessions & Persistence.

skill A markdown file with YAML frontmatter that the model can invoke as a tool. Skills encapsulate reusable workflows without requiring code. See Skills.

sub-agent A nested caliban instance spawned by the parent agent to execute a delegated task, optionally in an isolated git worktree. See Sub-agents.

tool A capability the model can invoke during a turn — built-in tools include Read, Write, Bash, Glob, Grep, Edit, WebSearch, and AgentTool. See Built-in Tools and Tool Execution.