supervisor-monitor

name	supervisor-monitor
description	Live tree view of supervisor via cron — agent collects stats every 30s
user_invocable	true

/supervisor-monitor — Live Monitoring

Sets up a recurring cron that dispatches the supervisor-monitor agent every minute to collect stats as JSON, then renders the tree.

Implementation

When the user invokes /supervisor-monitor:

1. Create a cron job:

   CronCreate(cron="*/1 * * * *", prompt="/supervisor-monitor-tick", recurring=true)
   

2. Run the first tick immediately (see tick logic below).

3. Report the cron job ID so the user can cancel with CronDelete.

What each tick does

1. Dispatch the supervisor-monitor agent:

   Agent(subagent_type="supervisor-monitor", prompt="Collect stats and return JSON.")
   

   The agent returns structured JSON (see .claude/agents/supervisor-monitor.md for schema).

2. Parse the JSON and render the tree in this EXACT format:

Hub: ~\/.claude-<profile>
└── Supervisor: ~\/.claude-<profile>  (started 17:00, running 1h 15m)
    │
    ├── Researcher (main): ~\/.claude-<profile>  (started 17:00, running 1h 15m)
    │   PID: 95600  running
    │   │
    │   └── Target: P@5=0.65  R@5=0.575
    │       Runs (last 3 of 5: 2 keep, 3 discard):
    │         #3  112ba8b  0.65  baseline
    │         #4  1cc81ef  0.35  discard  chunk overlap
    │         #5  eca93b6  0.60  discard  swap RRF
    │
    ├── Researcher Variant rv-001 (precision-safe): ~\/.claude-<profile>  (started 17:05, running 1h 10m)
    │   PID: 96001  running
    │   │
    │   ├── Target Variant rv-001-tv-1: P@5=0.70  R@5=0.62
    │   └── Target Variant rv-001-tv-2: P@5=0.68  R@5=0.64
    │   Runs (last 3 of 8: 3 keep, 5 discard):
    │     #6  ccc9012  0.68  keep     BM25 stop words
    │     #7  ddd3456  0.62  discard  heading-aware
    │     #8  eee7890  0.64  keep     overlap 300
    │
    └── Researcher Variant rv-002 (evaluator-direct): ~\/.claude-<profile>  (started 17:10, stopped after 25m)
        PID: 96002  stopped
        │
        └── Target Variant rv-002-tv-1: P@5=0.55  R@5=0.50
        Runs (2: 0 keep, 2 discard):
          #1  fff1234  0.65  baseline
          #2  ggg5678  0.55  discard  alternate eval

Tree rendering rules

Given the JSON from the agent:

• Hub line: Hub: {hub_profile}
• Supervisor line: └── Supervisor: {supervisor_profile} (started HH:MM, running Xh Ym) — compute duration from now - supervisor_started_at
• Each researcher in the researchers array becomes a child node:
  • First entry is always Researcher (main)
  • Others are Researcher Variant {id} ({name})
  • Use ├── for non-last, └── for last
  • Show profile + timing: (started HH:MM, running Xh Ym) if running, (started HH:MM, stopped after Xh Ym) if stopped
  • Show PID: {pid} {running/stopped}
  • Target: MRR={mrr:.4f} R@5={recall_at_5:.4f} or no report yet if target is null
  • Runs: Runs (last 3 of {total}: {kept} keep, {discarded} discard): or no runs yet if runs is null
  • Each run: #{n} {commit} {metric} {status} {description}
  • Target variants: children of the researcher, show metrics for each

Duration formatting

Compute duration = now - started_at from the JSON:

• Under 1 hour: Xm (e.g. 45m)
• 1+ hours: Xh Ym (e.g. 1h 15m)
• If started_at is null: show (no start time)

For a one-shot snapshot

Use /supervisor-list instead, or dispatch the agent directly:

Agent(subagent_type="supervisor-monitor", prompt="Collect stats and return JSON.")

Then render the tree from the returned JSON.