| name | agent-satisfaction |
| description | Run the 19-intent satisfaction scorer across all agents, write report to Chartroom, and trigger self-healing actions on threshold breaches |
| version | 1.0.0 |
| author | captain |
| tags | ["satisfaction","scoring","intents","health","self-healing"] |
| intent | Observable [I13], Reliable [I05] |
Agent Satisfaction
Score all 16 agents across 19 intents. Write results to Chartroom as report-agent-satisfaction. Trigger self-healing actions on threshold breaches.
When to use
- Daily check: report is older than 7 days or missing
- After major changes: new agent created, skills added, config changed
- On demand: when Reactor or Robert asks for fleet health
Command
Use MCP tools (works from inside container — no host paths needed):
- Full JSON report: Use
satisfaction_scores MCP tool (no args for all agents)
- Single agent: Use
satisfaction_scores MCP tool with agent: "<agent-id>"
- One-liner for sitrep: Use
satisfaction_summary MCP tool
Fallback (host only):
python3 /root/.openclaw/scripts/agent-satisfaction-score.py --json
Procedure
- Run the scorer: use
satisfaction_scores MCP tool
- Parse the JSON output for fleet average and per-agent scores
- Write summary to Chartroom:
chart update report-agent-satisfaction "<summary>"
- Check thresholds and trigger self-healing:
Self-Healing Thresholds
| Condition | Action |
|---|
| Any agent context > 85% | Reset session via openclaw sessions --agent <id> --reset |
| Any agent context > 70% recurring (3+ snapshots) | Flag for workload evaluation |
| I03 Competent < 5 | Agent needs skills — flag for Dev |
| I02 Understood < 5 | SOUL.md needs rewrite — flag for Reactor |
| I08 Resilient < 5 | Missing fallback model — fix auth-profiles |
| I05 Reliable < 4 | Chronic errors — investigate model/config |
| Fleet avg < 6.0 | System-wide alert to Reactor |
Report Format (Chartroom)
Fleet avg: X.X/10 (16 agents). [date].
Top: [agent]=X.X. Bottom: [agent]=X.X.
Alerts: [threshold breaches].
Actions taken: [self-healing actions].
Trigger
Captain checks report-agent-satisfaction date on first turn each day.
If older than 7 days (or missing), run this skill.
Rules
- No self-scoring (Captain doesn't score itself higher)
- Unknown (?) is honest — never invent scores without data
- Scores are earned from real signals, not defaults
- System-wide metrics are system scores, not agent scores
