Coding-agent output is being trimmed and checked before merge

Agent traces are becoming inputs to cleanup rather than disposable logs. TRIM reconstructs repair trajectories and removes edits only when tests still pass, cutting redundant patch content by 17.8%–32.9%. SWE-Pruner Pro reads the coding model’s hidden states to discard irrelevant tool-output lines, reporting token reductions of up to 39.4% while largely preserving benchmark quality. Together, they show two practical ways to reduce residue from long agent runs: minimize the final patch and compress the working context.

Evaluation is becoming more specific about what a successful patch must prove. An analysis of 4,882 agent-generated pull requests found that only 49.6% of code-changing PRs modified tests; existing tests covered 61.5% of changed Java lines and 27.0% of changed Python lines. A separate white-box audit showed that a detailed 24-item checklist passed 10 of 10 runs, versus 5 of 10 for a generic self-check. GitHub’s product response puts coverage metrics and quality gates directly into pull requests, while VNVSpec links machine-readable requirements to executable evidence. The common constraint is explicit evidence: test passage alone can conceal uncovered code or omitted requirements.