What problem does it solve? After an e2e genealogy benchmark run finishes, the raw run log is hard to read: it mixes harness facts, tool calls, narration, and judge output. This Skill reads those artifacts and explains in plain language what the agent recovered, what it missed, why it stopped, and the most likely cause — without leaking the judge's grade so the run can still be graded blind afterward. ## Core Features & Use Cases - Recall analysis: Compares the agent's final GedcomX tree against the fixture's expected-findings.json to state which required findings were recovered, missed, or recorded off-shape. - Stop-reason diagnosis: Translates stop_reason values (completed, natural_end, inactivity, timeout, tool_cap, cost_cap, error, mcp_unavailable) into actionable explanations with pointers to the exact tool call and narration turn. - Cause attribution: Distinguishes agent reasoning regressions, /research routing regressions, sub-skill regressions, FamilySearch data drift, and single-run jitter, while staying blind to judge_output, verdict, and outcome. - Use Case: A maintainer asks "why did this fixture fail?" and receives a grounded summary — e.g., the agent looped on place_search until the tool cap — plus a recommendation to diff against the last passing run and grade the run blind with /grade-e2e-run. ## Quick Start Interpret the latest e2e run log for the fixture in eval/runlogs/e2e and tell me what the agent recovered, why it stopped, and the most likely cause.
