AgentSkillsCN

agent-monitor

在 WezTerm 中对 Pi 代理进行精细化监控，实时展示每个代理的完整任务列表、近期活动记录，以及最后一次工具执行的输出结果。

SKILL.md
--- frontmatter
name: agent-monitor
description: Detailed monitoring of Pi agents in WezTerm. Shows full task, recent activity, and last tool output for each agent.

Agent Monitor

Get detailed status of each Pi agent.

Run Detailed Status Check

bash
echo "🧠 DETAILED AGENT STATUS - $(date '+%H:%M:%S')"
echo ""

PANES=$(bb wezterm list 2>/dev/null)
if [ $? -ne 0 ] || [ -z "$PANES" ]; then
  echo "⚠️ WezTerm not accessible"
  exit 0
fi

echo "$PANES" | jq -r '.data[] | "\(.pane_id)|\(.title)|\(.cwd)"' | while IFS='|' read -r PANE_ID TITLE CWD; do
  CWD_CLEAN=$(echo "$CWD" | sed 's|file://||')
  PROJECT=$(echo "$CWD_CLEAN" | xargs basename 2>/dev/null || echo "unknown")
  
  [[ "$TITLE" != *"π"* ]] && continue
  
  SESSION_PATH=$(echo "$CWD_CLEAN" | sed 's|/|-|g' | sed 's|^-||')
  SESSION=$(ls -t "$HOME/.pi/agent/sessions/--${SESSION_PATH}--"/*.jsonl 2>/dev/null | head -1)
  [ -z "$SESSION" ] && continue
  
  STOP=$(tail -1 "$SESSION" | jq -r '.message.stopReason // "null"')
  ROLE=$(tail -1 "$SESSION" | jq -r '.message.role // "unknown"')
  [ "$STOP" = "stop" ] && [ "$ROLE" = "assistant" ] && STATUS="✅ IDLE" || STATUS="🔄 WORKING"
  
  echo "═══════════════════════════════════════════════════════════════"
  echo "## Pane $PANE_ID - $PROJECT - $STATUS"
  echo ""
  
  echo "**Last user request:**"
  tail -100 "$SESSION" | jq -r 'select(.type=="message" and .message.role=="user") | .message.content | if type=="array" then .[0].text else . end' 2>/dev/null | tail -1
  echo ""
  
  echo "**Recent activity (last 5 messages):**"
  tail -50 "$SESSION" | jq -r 'select(.type=="message") | "\(.message.role): \(.message.content | if type=="array" then ([.[] | if .type=="text" then .text[:150] elif .type=="toolCall" then "[\(.name)]" else .type end] | join(" ")) else .[:150] end)"' 2>/dev/null | tail -5
  echo ""
  
  echo "**Last tool output (truncated):**"
  tail -20 "$SESSION" | jq -r 'select(.type=="message" and .message.role=="toolResult") | .message.content | if type=="array" then .[0].text else . end' 2>/dev/null | tail -1 | head -c 300
  echo ""
  echo ""
done

echo "═══════════════════════════════════════════════════════════════"

Output Explained

For each agent shows:

  • •Pane + Project + Status (idle/working)
  • •Last user request - the full task they were given
  • •Recent activity - last 5 messages (role + truncated content/tool calls)
  • •Last tool output - what the last command returned

After Review

Analyze each agent and decide:

  1. •Is it stuck? Same output for too long → send "status?"
  2. •Is it done? ✅ IDLE → assign new task
  3. •Is it duplicated? Multiple panes same work → stop extras
  4. •Does it need help? Errors in output → intervene

Actions

bash
# Send task
bb wezterm send <pane_id> "your task"

# Check full session
tail -100 <session_file> | jq '.message'

# Interrupt
bb wezterm send <pane_id> "stop, summarize progress"