FAQ¶
Does Zwischen store my code?
No. Tool output content is streamed to the reviewer for live evaluation but is not retained. Your code, file contents, and command output are ephemeral to the review session. The review trail (viewable via /z spy) contains the reviewer's determinations and session metadata, not your code.
Can I use Zwischen with models other than Claude Code?
Zwischen currently works with Claude Code only. The hook system and MCP integration are specific to Claude Code's architecture.
Does the reviewer slow down my session?
Pre-reviewed tools (those in pre_review and write_review) have a small latency cost because the reviewer evaluates them before execution. Post-execution review happens asynchronously and does not block the agent.
You can minimize latency impact by keeping the pre_review list short — only include tools with irreversible side effects that genuinely need pre-execution review.
Can I use the same model for the agent and the reviewer?
Yes, but it's not recommended. Using a different model for the reviewer provides genuine adversarial value — different models have different blind spots and biases. The whole point is to introduce a second perspective.
What happens if the Zwischen server is unreachable?
If the client can't reach the Zwischen server at session start, the session runs without monitoring. If the connection drops mid-session, the daemon handles reconnection. The agent is never blocked by server unavailability — monitoring fails open.
Can I run Zwischen without an internet connection?
No. The reviewer session runs on the Zwischen server, which manages the reviewer's AI conversation. The server and the reviewer's decision-making process are fully hosted.
Is the reviewer's API key sent to Zwischen?
No. Your model provider API key is held in process memory on your machine and passed directly to the model provider. It is never sent to the Zwischen server.
Can the agent bypass Zwischen?
No. Zwischen installs as Claude Code hooks, which are enforced by the Claude Code runtime. The agent cannot disable or skip hooks. The only way to bypass Zwischen is to disable it via /z disable or by removing the hooks from your Claude Code settings.
How many subagents can Zwischen monitor?
Each subagent gets its own independent reviewer session. There's no hard limit on the number of concurrent subagent sessions.
Can I use different models for different agents?
Yes. Use profiles to assign different models to different agents. Each profile can specify its own model and api_key.
What's the difference between /z ask and /z direct?
/z ask sends a question to the reviewer without changing its behavior. The reviewer answers and continues monitoring as before.
/z direct sends a high-priority directive that overrides the reviewer's behavior for the rest of the session. Use /z direct when you need the reviewer to change what it's doing, not just answer a question.