Skip to content

Live Monitor

The live monitor opens a separate terminal window that streams the reviewer's verdicts in real time, side by side with your session. It turns Frank's otherwise-invisible reasoning into something you can watch as work happens.

/z monitor

This spawns a new terminal window running the live feed and returns immediately – your session is not blocked. Close the window (or Ctrl-C inside it) when you're done.

Both hosts

The monitor works in Claude Code and in Codex. Codex invokes it as $z monitor. The window, the stream, and the reconnect behavior are identical – but a Codex session produces fewer events to show. See What a Codex session streams.

Live only

The monitor shows verdicts from the moment it connects onward. It does not replay history – if you open it mid-session, you'll see everything from that point forward, but nothing that happened before. To review earlier activity, use the /z spy command, which prints the reviewer's transcript for a turn.

What the stream shows

Each line is one verdict from the reviewer, tagged with who it's from:

  • Work verdicts – the reviewer's pass/block decision on a tool call, a proposed write, or a stop attempt, with its one–two sentence reasoning.
  • Side conversations – the agent disputing a block via message_adversary, and the reviewer's replies to the agent and to your /z ask questions.

Most of the stream is pass – the reviewer silently approving the agent's work. The point of the monitor is to make the moments that aren't pass easy to spot.

Reading the stream

The monitor uses structure and color to separate signal from the firehose:

  • Blocks render in red. Any line where the reviewer blocks – a tool block, a write block, or a stop block – is highlighted so it stands out of the wall of passes.
  • Side conversations render as indented threads. The agent↔reviewer dispute is indented behind a pipe gutter and labelled AGENT and FRANK, so a back-and-forth reads as a conversation instead of scattered lines.
  • A status bar pinned to the bottom row shows the reviewer's model.

Note

Color and the indented gutter render on an interactive terminal. Exactly how a given color paints depends on your terminal emulator and theme. Piped or redirected output falls back to plain text.

What a Codex session streams

The monitor itself is identical in Codex, but it can only show what the host reports, and Codex reports less. Codex fires five hook events to Claude Code's nine. Two absences are visible in the feed:

  • No failed-tool-call event. When a tool errors, the reviewer sees that the call happened but not the text of the error, so you will not see a verdict reacting to a failure message.
  • No in-progress narration. Claude Code streams the agent's interstitial commentary as it works; Codex does not. The feed shows what the agent did, not its running account of why.

Blocks, disputes, planning holds, and stop reviews all appear normally. See Codex for the full list of host differences.

Read-only

The monitor is a passive observer. It never affects enforcement – it's a parallel view of verdicts the reviewer is already producing. Opening or closing it changes nothing about how the agent is monitored.

Opening from a terminal

/z monitor runs the underlying CLI command for you. You can also attach a monitor to a session directly from a terminal:

zwischen monitor --session-id <session_id>

Use /z session-id to find the current session ID.

Window spawning

The monitor opens a new window using whatever terminal your platform provides – Windows Terminal (mirroring your current profile) or a console fallback on Windows, Terminal.app or iTerm on macOS, a common emulator on Linux, and Windows Terminal back into the distro under WSL. If it can't spawn a window automatically, it prints the exact command to run so you can open one yourself.

The feed reconnects on its own if the connection drops (for example, during a server deploy) – because it's live-only, there's nothing to replay, so it simply resumes.