metrics

Load the metrics-timeseries.csv history file and compare it against recent point-in-time snapshots to detect deteriorating trends like rising PR latency or growing open issues. Then investigate the single most significant trend with hypothesis testing.

Develop competing hypotheses such as reviewer overload, author inaction, or CI delays, then gather evidence using the gh CLI and GraphQL queries to support or refute each one. Select the cause most strongly supported by the collected data.

Only when the orchestrator's System Directive enables PR creation for the run. In that case it activates the prs skill to generate a PR description and stage changes; otherwise it only reports findings without staging files.

Prefer deterministic TypeScript or Git logic whenever it achieves equivalent quality. Use the Gemini CLI only for semantic classification tasks like sentiment analysis or triage labeling, never for execution or decision-making.

Analysis depends on the quality and coverage of the recorded time-series data and cannot modify the output format of existing metrics scripts. Cost-saving recommendations are the lowest priority and only apply after health and latency concerns are addressed.