Cost: apply LLM cost-mitigation before scaling usage (caching / routing / batch / context)

[atomicdragonranch]

Reference lives in docs/llm-cost-optimization.md (PR #56). Tracking so it surfaces while working on the agent , act on these as usage grows:

• Prompt-cache the stable prefix (system prompt + tool defs). Verify usage.cache_read_input_tokens > 0; audit the prefix for silent invalidators (timestamps/UUIDs/unsorted JSON). ~0.1x cost on cached reads.
• Model routing , move cheap steps (classification/triage/extraction) to Haiku or rules; reserve Opus for hard reasoning. Keep the main loop on one model; delegate cheap sub-steps to Haiku subagents.
• Batch API (50% off) for any non-latency-sensitive work (backfills, bulk/scheduled runs); stacks with caching.
• Context management on long loops , compaction / context-editing + effort/task_budget.
• Measure-first , baseline with count_tokens + usage before optimizing.

Cost impact: $0 to adopt (config/architecture); goal is to cut recurring inference spend (caching + routing alone often >50%).