Credit Optimization — Recurring AI Model Usage

Decision

5 config changes to reduce recurring Anthropic credit burn by ~71%:

  1. OpenClaw heartbeat interval: 60m → 480m (8 hours) (Prometheus covers infrastructure gaps)
  2. Weekly trending intelligence model: Opus 4.6 → Sonnet 4.6 (research task, Sonnet sufficient with high thinking)
  3. Hermes Consul reasoning: xhigh → high (structured 6-phase framework doesn’t need max thinking)
  4. Hermes Consul subagents: Haiku 4.5 → Sonnet 4.6 (quality upgrade, no Haiku policy)
  5. Hermes Consul scan schedule: daily → Mon/Wed/Fri (enterprise evolution is gradual) Bonus: CXO optimizer timeout 600→900s, content scoring timeout confirmed at 600s.

Rationale

Claude Code Max subscription extra usage credits triggered by recurring automated processes. Audit identified 2 systems burning credits 24/7: OpenClaw NOVA (heartbeat + 10 cron jobs) and Hermes Consul (daily scan). Optimization matched model tier and thinking level to actual cognitive demand per task. Opus preserved for strategic tasks (night-shift, strategic brief, CXO optimizer, Hermes main brain). Sonnet for research and operational tasks. Per-job thinking levels applied to all 10 cron jobs (low/medium/high) via OpenClaw CLI —thinking flag.

Alternatives Rejected

  • Keep everything as-is (rejected: unnecessary credit burn with no quality benefit)
  • Downgrade night-shift to Sonnet (rejected: strategic task selection is Opus’s differentiator)
  • Reduce Hermes Consul to weekly (rejected: too sparse for evolution detection)
  • Use Haiku for Hermes subagents (rejected: user’s explicit no-Haiku policy)
  • Eliminate heartbeat entirely (rejected: OpenClaw-specific checks have no Prometheus equivalent)

Outcome

Pending