Janissary (0.12.0)

Appearance

Monitoring with personas ​

monitor starts a persona-driven AI session that watches tab activity and surfaces suggestions, either inline in the tab you started it from or in a dedicated reporting tab:

monitor security bilal
monitor activity group:2

Monitors are tool-less by default and receive no filesystem or terminal access. See Persona web tools below for the one narrow exception.

Inline vs. reporting-tab mode ​

monitor <persona> [target...] behaves differently depending on whether you give it targets:

  • No targets (inline mode): the monitor watches the tab you started it from and posts suggestions right into that tab's own transcript, prefixed with the persona name and a sparkle.
  • One or more targets (reporting-tab mode): the monitor watches the named tabs (or groups) instead, and its suggestions go to a dedicated reporting tab, colored after the tab(s) it watches.

A target is a tab label or group:<n>. A tab may also be named by its display alias (see rename), matched case-insensitively; unmonitor's target argument works the same way. Once the monitor's session connects, the owner tab's transcript shows a line naming the monitor, its model, and a one-sentence summary of the persona's role.

The reporting tab ​

A reporting-mode monitor's tab carries a metadata line above its suggestion feed: the persona name, the tab(s)/group(s) it watches, and the total size sent to and received from its session so far (shown in bytes/kilobytes/megabytes). Two buttons sit at the right of that line:

The reporting tab gutter has an up/down-arrow button at the right. Drag it vertically to resize the reporting area: drag up for more reporting space, or down for more main-tab space. Neither the reporting area nor the main tab area can shrink below 15% of the window height.

  • A reset button discards the monitor's accumulated context and reloads just its persona priming, the same recovery the monitor performs automatically after a session error.
  • A context snapshot button opens the monitor's current accumulated context (persona priming, batched updates, questions, and replies) as a point-in-time snapshot in a text tab.

If two owner tabs share one reporting tab (the same persona started from two different tabs), resetting it resets every monitor feeding that tab.

What a monitor sees ​

A monitor receives each target's full existing history the moment it starts, not just activity from then on:

  • An agent tab feeds its complete transcript, in order.
  • A harness tab feeds its normalized session transcript first, then its latest on-screen text. The transcript includes subagent activity and arrives in history order; the screen follows so the monitor can see current interactive state such as a permission prompt, spinner, or TUI panel. The transcript is sent incrementally after the monitor's previous flush. An SSH tab has no session transcript, so it feeds only its latest on-screen text, refreshed when the screen changes.
  • An editor tab feeds its live (possibly unsaved) buffer content: the first feed is the full content, and every feed after that is a diff against what that monitor last saw.
  • A page tab feeds only the text currently visible in its viewport, the same way: full content first, diffs after.

Group membership is re-checked continuously, so a tab that joins a monitored group later is picked up without restarting the monitor.

Content from monitored tabs is treated as data, not as instructions to the monitoring persona. This keeps text from a file, web page, harness, or transcript from changing what the persona was asked to do.

The flush cycle ​

New activity from every target is buffered and sent to the monitor's session as a single batch prompt every 30 seconds. Nothing is sent when the buffer is empty or a previous prompt is still in flight.

Suggestions and summaries ​

A monitor's reply is parsed for two possible marker lines: [SUMMARY]: recaps activity with no action attached, and [SUGGESTION]: offers something actionable, optionally paired with a [COMMAND]: line. Marker text can continue across multiple lines. If a reply includes both markers, the suggestion is delivered and the summary is ignored. A reply with neither marker (like a bare OK) delivers nothing.

In a reporting tab, a suggestion that carries a command shows it as a clickable line: clicking runs it in the tab the suggestion is about, queuing behind anything already queued there. Each suggestion also carries thumbs-up / thumbs-down buttons; either one removes the suggestion from the feed.

Asking a monitor directly ​

monitor ask <persona> <question> sends a question straight to a running monitor's session, skipping the batch buffer. The reply lands in the owner tab's transcript.

Only one direct question or scheduled flush can be in flight for a monitor at a time. If it is busy, wait and try again.

Persona web tools ​

A monitor has no tools by default. A persona file can opt into web_search and web_fetch with a second directive line:

[//]: # claude:Sonnet:high
[//]: # tools: web_search, web_fetch

These are the only two tools a persona can ever request this way. No filesystem or terminal access is ever granted to a monitor.

Stopping a monitor ​

unmonitor security          stop the "security" monitor started from this tab
unmonitor --all             stop every monitor started from this tab
monitors                    list active monitors with their targets and suggestion counts

A monitor's session also ends on its own when its owner tab closes or every one of its tab targets has been removed. A reporting tab stays open as long as at least one monitor still feeds it, and closes once the last one stops.

Closing a reporting tab stops the monitors feeding it. If two owners share a reporting tab, closing one owner stops only that owner's monitor; the reporting tab remains while another monitor still feeds it.