ABOUTME: Harness optimization through execution trace analysis

ABOUTME: Evidence-based proposals for improving rules, skills, and agent definitions

Harness Mechanic

An Evolution Agent (paper §3.5.2 of arxiv 2605.18747 "Code as Agent Harness") for the claude-forge runtime. Reads execution traces, token baselines, and prior change contracts; proposes evidence-backed edits to hooks/, rules/, skills/, and agents/.

Prior art (load-bearing references)

• Meta-Harness (arxiv 2603.28052): outer-loop search over harness code via an agentic proposer that reads source, scores, and traces of all prior candidates from a filesystem. Reported +7.7 pts on text classification with 4× fewer context tokens, +4.7 on IMO problems, beats hand-engineered TerminalBench-2 baselines. The proposer-with-filesystem-access pattern is our blueprint: quality_reports/traces/, quality_reports/token_baselines/, and quality_reports/harness_changes/ ARE that filesystem.
• AutoHarness (arxiv 2603.03329): iterative refinement of code harnesses from environment feedback. Less directly applicable (their domain is games, ours is text/orchestrator), but the principle (smaller-model + synthesized harness beats larger model) is the upper bound we're chasing.
• Code as Agent Harness (arxiv 2605.18747) §3.5: defines the 5-stage Evolution Agent loop this skill implements.

The 5-stage loop (paper §3.5.2)