Glossary

Agent — The AI coding assistant (Claude Code) doing the actual work in your session. Zwischen monitors the agent's actions.

Block — An enforcement decision by the reviewer that stops the agent from proceeding. Three types: tool block (denies a tool call), write block (rejects a file change), stop block (forces the agent to continue working).

Context cycling — Automatic reset of the reviewer's context window when it approaches capacity. The reviewer writes a handoff summary, and a new session starts with that summary plus recent events.

Daemon — The local HTTP server (localhost:8384) that runs on your machine during a Claude Code session. Manages communication between the hooks, the MCP server, and the Zwischen server.

Determination — The reviewer's decision for a given event: pass, block, or release.

Frank — The reviewer. Named after zwischenzug — the chess "in-between move."

Handoff summary — A compressed account written by the reviewer during context cycling, summarizing what it's seen and any ongoing concerns. Carried into the fresh context.

Hook — A Claude Code hook that fires on specific events (SessionStart, PreToolUse, Stop, etc.). Zwischen installs hooks for all eight Claude Code hook events.

MCP server — The local MCP (Model Context Protocol) server that provides the message_adversary tool to the agent. Runs as a separate process spawned by Claude Code.

Monitoring session — The reviewer's AI conversation with the Zwischen server. Each agent (main or subagent) gets its own independent monitoring session.

Orchestrator — The default name for the main Claude Code agent. Used for profile matching when no custom agent name is set.

Pre-review — Review that happens before a tool executes. The reviewer can block the tool call before it runs.

Profile — A named configuration block in settings.yaml that applies specific reviewer settings to specific agents.

Release — An enforcement decision that lifts a previous block. Issued when the reviewer agrees to let the agent proceed after a dispute.

Reviewer — The second AI model that watches the agent's event stream and issues enforcement decisions. See Frank.

Scrutiny level — A 1–5 scale controlling how aggressively the reviewer intervenes. Higher levels mean more blocks.

Session ID — A unique identifier for a Claude Code session. Groups all agents (main and subagents) in one user task.

Spy transcript — The reviewer's activity log, viewable via /z spy. Shows reviewed events, determinations, and reasoning.

Subagent — An agent spawned by the main Claude Code agent (via the Agent tool). Each subagent gets its own independent reviewer session.

Write review — A specialized form of pre-review for Edit and Write tools. The reviewer sees the proposed file diff and can reject it before the file is modified.

Zwischenzug — A chess term for an "in-between move" — an unexpected intermediate move that disrupts the opponent's expected line. The inspiration for the product name.