recsys-debugging
Diagnose recommendation quality, traffic, feature, model, retrieval, ranking, latency, timeout, overload, degradation, or logging regressions using an evidence-first decision tree.
SKILL.md.agents/skills/recsys-debugging/SKILL.md
GitHubRecommendation debugging
- State symptom, first bad time, affected traffic, metric definition, and recent changes.
- Run
python3 tools/recsys_agent.py debug "<symptom>". - Follow the emitted order: experiment/config, supply funnel, features, release versions, calibration/logging, then serving health.
- At each branch collect the named evidence before moving on. Compare against a known-good cohort or version.
- Reproduce the fault where safe; fix the root cause, then rerun the same scenario.
- Output timeline, evidence, root cause, corrective change, verification, and prevention. Never suppress an alert as the fix.