Highlights
Full-breadth model routing with a frontier escalation ladder. The cost-aware selector now ships a default capability ladder grounded in the live OpenRouter catalog (harvested-free / cheap → mid → frontier) and routes each task by quality-per-dollar across the full breadth of models. The strongest tier is reserved for the final attempt only — a hard task starts mid-ceiling and escalates to a frontier model as the safety net, not on the first try. The gate quality floor is unchanged, so routing only ever moves cost/latency, never shipped quality. Verified live: a task cheaper tiers stalled on was resolved by the frontier tier on its final attempt, for pennies.
Reproduction-first with pre-patch validation. The spec-as-tests workflow runs its generated reproduction test on the clean tree first and keeps it only if it actually fails there — a test that reproduces nothing is a false gate and is discarded. Engaged automatically on both SWE-bench execution paths (host and containerized), where the judged tests are hidden.
Two-timescale evolution — self-authored tool library (consolidation layer). A new ToolLibrary over the existing MAP-Elites + held-out promotion machinery: a runtime-invented tool is admitted to a persistent, best-per-capability library only if it passes the same held-out generalization gate that guards scaffold self-edits, and promoted tools seed future runs. Capability compounds across runs instead of being reinvented each task. Design: docs/two-timescale-evolution.md.
Test suite: 1,897 passing. Full changelog in CHANGELOG.md.