AgentSystem Stability Ops

Run fast reliability checks before changing cron/model/memory behavior.

Quick Start

1. Run reliability audit:

python3 scripts/reliability_audit.py --hours 24

2. Run scheduler contract audit:

bash ~/.agent-evolution/scripts/agent-system-cron-audit.sh --json

3. Regenerate health report:

bash ~/scripts/cron-health-monitor.sh

Workflow

1. Check scheduler contract first. If isolated + payload.kind != agentTurn exists, fix or remove those jobs before any other tuning.

2. Check agent consistency. If runAgent != job.agentId, treat as high severity and verify session routing + explicit agentId for every enabled job.

3. Check failure quality. Separate transport/provider failures from logic failures. Timeout-heavy clusters usually require timeout/model/prompt changes, not blind retries.

4. Enforce output hygiene. Any summary containing raw tool logs (Exec, stack trace, raw JSON) is a regression.

5. Re-run audits after changes. Do not declare stable until the last 6-12h window has zero contract violations and zero agent mismatches.

Resources

• scripts/reliability_audit.py: Computes reliability metrics from jobs.json + run logs.
• references/runbook.md: Safe remediation order and command cookbook.