cron-cost-guard（cron 成本守护）

Cron Cost Guard

Prevent silent token budget burns from misconfigured AI agent cron jobs.

Quick Audit

Run this checklist on every cron setup or when investigating a cost spike:

1. Session Isolation (Critical)

Check every cron job for session binding conflicts:

cron list (includeDisabled: true)

Red flags:

• sessionKey: "agent:main:main" with sessionTarget: "isolated" → stale binding, will cause model conflicts
• agentId pointing to a different agent than the session owner → cross-agent model contamination
• consecutiveErrors >= 3 → likely stuck in a retry loop
• lastError containing LiveSessionModelSwitchError → model-switch loop confirmed

Fix: Remove and recreate the job without sessionKey. Set sessionTarget: "isolated".

2. Model Conflicts

In multi-agent setups (e.g., Agent A on Claude, Agent B on GPT), each agent's crons must be scoped to that agent only.

• Set agentId explicitly on every cron job
• Set model explicitly in the payload when available
• Never let Agent B's cron inherit Agent A's session model

3. System Prompt Size

Audit injected workspace files:

wc -c MEMORY.md SOUL.md AGENTS.md TOOLS.md USER.md QUEUE.md

Target: < 20KB total injected. Move large files (playbooks, heartbeat templates, reference docs) to references/ for on-demand reading.

4. Cost Monitoring

sessions_list (limit: 10, messageLimit: 1)

Look for sessions with high estimatedCostUsd but low output tokens — that's a retry loop signature.

Diagnosis: Token Spike

For detailed diagnosis steps and post-incident checklist, read references/diagnosis.md.

Prevention Rules

1. Every cron job: sessionTarget: "isolated", no stale sessionKey
2. Every cron job: explicit timeoutSeconds (never unlimited)
3. Multi-agent: explicit agentId matching the agent that should run it
4. After changing default model: audit all cron jobs for conflicts
5. Weekly: check consecutiveErrors across all jobs — anything ≥ 3 needs investigation