Анализ траекторий агентов через Inspect Scout

testing

check-trajectories-workflow is a Claude Code skill that automates the analysis of agent trajectories from evaluation log files using Inspect Scout. It runs five built-in scanners — outcome_summary, external_failure, formatting_failure, reward_hacking_success, and ethical_refusal — and supports adding custom scanners via InspectEvalScanner objects. The 11-step workflow covers everything from selecting the log file and creating an agent_artefacts/trajectory_analysis/<eval_name> folder to running run_all_scanners.py, extract_results.py, and analyze_validity.py, then producing a final ANALYSIS.md report. It is designed for AI evaluation and quality assurance teams who need to detect infrastructure failures, formatting errors, and reward hacking faster than manual log review allows.