v5.3.0 — Rate-Limit Detection, has_output Fix, Absolute Paths

What's New

Rate-limit detection

run_eval.py detects when >30% of runs hit 429/rate-limit errors. Reports rate_limited: true + count in JSON output + warning to stderr. Score 0.0 from quota exhaustion is now distinguishable from actual agent failures.

has_output excluded from combined score

setup.py defaults evaluator_weights to {"has_output": 0}. has_output is still tracked but excluded from the combined score. Previously, any print("Starting...") gave 1.0, inflating scores artificially.

Absolute paths in skill cp commands

Skill instructions now use $(pwd) for absolute paths. No more 2>/dev/null hiding errors. Removed redundant manual cp from eval step (run_eval.py auto-copies since v4.6.0).

Correct --config usage documented

Skill now passes --config "$(pwd)/.evolver.json" (project main) not --config "$WT/.evolver.json" (worktree). The auto-copy in run_eval.py handles the rest.

🤖 Claude + Codex/GPT-5.4