cost-optimization-background-llm

Cost Optimization — Background LLM Calls

Reduce LLM token spend on background/scheduled work in {{AGENT_NAME}} without touching the user-facing interactive path (CLI / ACP / Discord messages). Terminal-direct calls are off-scope by user preference.

Related skills

• hermes-model-routing — broader framework for model selection across all 4 routing levels (main/delegation/auxiliary/provider_routing). This skill focuses on the cost dimension; the routing skill covers the complete selection framework.

When to use

Trigger words from user: "cost optimization", "background LLM", "cron LLM", "token cost", "cheap model", "스케줄 작업과 백그라운드".

Decision tree

Background LLM call site discovered
    │
    ├── (1) Is the LLM call essential?
    │     │
    │     ├── Yes (report synthesis, MCP query, instruction) → keep LLM
    │     │   but route to cheaper model via config.yaml
    │     │
    │     └── No (simple shell interpretation) → make deterministic
    │
    ├── (2) What cost lever is available?
    │     │
    │     ├── Smart-routing cheap_model != main model
    │     │   → config.yaml smart_model_routing.cheap_model.{provider,model}
    │     │
    │     ├── Auxiliary task model override
    │     │   → config.yaml auxiliary.{task}.{provider,model}
    │     │     tasks: vision, web_extract, compression, session_search,
    │     │            skills_hub, approval, mcp, flush_memories
    │     │
    │     ├── Script-based fast path (cron)
    │     │   → jobs.json add `script:` field + scheduler.py branch
    │     │
    │     └── Task body classification (kanban worker)
    │         → classify first non-empty line; shell-prefix → subprocess
    │
    └── (3) Verify with hard evidence, not just "Done ✅" claims.

Procedure (5 phases)

Phase 1 — Inventory

# 1a) Cron jobs: list all jobs
jq '.jobs[] | {id, name, enabled, schedule, last_status, last_run_at}' \
   ~/.{{AGENT_NAME_LOWER}}/cron/jobs.json

# 1b) Find scheduler entry point
grep -n "run_job\|run_conversation\|AIAgent" \
   ~/.{{AGENT_NAME_LOWER}}/source/{{AGENT_NAME_LOWER}}-agent/cron/scheduler.py

# 1c) Background threads / fire-and-forget
grep -rn "threading.Thread\|daemon=True" \
   ~/.{{AGENT_NAME_LOWER}}/source/{{AGENT_NAME_LOWER}}-agent/ | grep -v test_

# 1d) Kanban worker LLM path
grep -n "AIAgent\|agent.chat\|run_conversation" \
   ~/.{{AGENT_NAME_LOWER}}/scripts/run_kanban_worker.py

# 1e) Auxiliary task consumers
grep -rn "call_llm\|async_call_llm" \
   ~/.{{AGENT_NAME_LOWER}}/source/{{AGENT_NAME_LOWER}}-agent/agent/ | grep -v test_