Cost Control: Model Routing and Effort Tiers

PROJECT CONTEXT:
My agent runs every subtask - trivial or gnarly - on the most expensive model at full effort, and the bill shows it. Harness evolution analyses show mature systems route by task class and tune effort tiers, without letting quality quietly rot.

GOAL:
Implement model routing and effort tiers in the harness: task classification, per-class model/effort assignment, budget tracking, and quality-guarded rollout.

REQUIREMENTS:
- A task/subtask classification the router can act on (main reasoning, summarization, classification, repair, sub-agent research)
- Per-class routing config: model + effort/reasoning tier + token limits
- Escalation paths: failed cheap attempts retry on stronger settings
- Per-session and per-task cost tracking visible in logs and summaries
- A quality-guard eval comparing routed vs premium-everything on a fixed task set

CONSTRAINTS:
- Routing is config, not scattered conditionals - one table rules all call sites
- Quality regressions beyond an agreed threshold roll the class back to the stronger setting

DEFINITION OF DONE:
- Auxiliary calls (summaries, classifications) run on the economy tier
- A cheap-tier failure escalates automatically and logs the escalation
- Session summaries report cost by class, model, and tier
- The quality-guard eval shows acceptable deltas for every routed class

COMMON FAILURES TO AVOID:
- Premium models summarizing tool output at 20x the needed cost
- Cheap models silently degrading main-task quality with no eval to notice
- Routing logic duplicated across call sites, drifting independently
- No escalation, so cheap-tier failures just become task failures