tune-agent-task-costlisted

# Tuning Agent Task Cost Myco's phased agent tasks are the largest LLM-spend surface in the system. They run unattended on a schedule, often dozens of times per day, on the user's own credits. A task that is correct but cost-careless will burn through a monthly budget in days. This skill captures the patterns that have emerged across `skill-evolve` and `vault-evolve` for keeping these tasks cheap *and* predictable, without sacrificing the work they're meant to do. Cost-tuning is a discipline, not a single trick. The same task can cost $0.40 or $3.00 depending on whether these patterns are applied. Use this skill any time you author a new phased task or revisit an existing one. ## Prerequisites - Read `myco:author-and-debug-agent-pipeline-tasks` first for the mechanics (YAML anatomy, scheduling, budgets, preConditions). This skill is the cost-discipline layer on top of those mechanics. - Familiarity with `agent_runs` table fields (`actual_cost_usd`, `checkpoints`, `error`) — the audit trail is where you measure whether your tuning worked. - Understand the difference between `reasoningLevel` (a tier: `low | default | high`) and a concrete model name (`haiku`, `sonnet`, `opus`). Tasks should specify tiers; the provider's `reasoning_map` chooses the model. --- ## Pattern 1: Origin Discipline on Batch Reads **Problem.** `prompt_batches` has an `origin` column: `human`, `system`, `agent_dispatch`. The `system` origin captures environment-context wrappers, tool announcements,