Harness

Automated fix loops. Each harness runs a make target, spawns sub-agents to fix failures, commits on success, reverts on exhaustion.

Pattern: .claude/harness/<name>/run + fix.prompt.md

Escalation: sonnet:think → opus:think → opus:ultrathink

History: Each harness maintains history.md for cross-attempt learning. Truncated per-file, accumulates across escalation. Agents append summaries after each attempt so higher-level models can avoid repeating failed approaches.

CLI: .claude/scripts/check-<name> symlinks to harness run scripts

Commands: /check-<name> runs the corresponding harness with --no-spinner

Quality: quality/run orchestrates all harnesses in sequence until stable