eval-diagnoselisted

# Eval Diagnose Use this skill for evidence-backed RCA once a packet, URL-derived evidence, or trace/session/log-stream context is available. ## Required Reference Use `skills/eval-engineer/references/rca-recipe.md`, `skills/eval-engineer/references/debug-packets.md`, `skills/eval-engineer/references/evidence-provenance.md`, and `skills/eval-engineer/assets/diagnosis-template.md`. ## Do - Start from fetched evidence, not source-code guesses. - Name the failing metric contract and what it proves. - Label hosted Galileo evidence separately from local deterministic packets before making metric or score claims. - Inspect traces, spans, sessions, tool calls, retrieval context, and scorer status to classify the fix surface. - Classify the fix surface: prompt, tool schema, adapter, retriever, ranker, guardrail, metric, dataset, or SDK wiring. - Write diagnosis and bounded fix plan only when evidence supports it. - Honor read-only requests. If the user says read-only, dry run, no edits, or "do not edit files", do not write `.galileo/` artifacts. Return the RCA inline and include a short "Would write" list for any suggested artifact paths. ## Gotchas - Fetched debug packets are the RCA source of truth when scorer jobs are still settling or runner output disagrees with fetched metrics. - A prompt diff, local score, or code diff is not proof of improvement without before/after Galileo evidence. - Bare correctness or factuality can be a smoke test only. Prefer the