codex-session-benchmark-maintainer

What problem does it solve?

It helps you evaluate how efficiently and effectively a specific Codex session performed by benchmarking it against a recent local cohort and turning log evidence into actionable workflow guidance.

Core Features & Use Cases

• Target-vs-cohort proxy benchmarking: Compares a named Codex session against a deduped recent window to estimate efficiency, rework, follow-up, interruptions, and handoff quality without assuming ground-truth success.
• Evidence-first metric extraction: Computes or labels metrics as direct, derived, proxy, or missing using local JSONL/SQLite sources, including validation/commit/PR handoff indicators when present.
• Reusable failure-class outcomes: Classifies the session outcome (e.g., clean handoff, completed with rework, partial/blocked, diagnosis only, ambiguous) and maps underperformance to reusable workflow failure classes, then recommends next skill/workflow changes (often by handing off to skill-optimizer).

Example use cases:

• Assess whether a target Codex session (given by id or prefix) was notably worse or better than the surrounding 30–100 session cohort for the same Rudder development work context.
• Diagnose whether a session’s low effectiveness came from wrong skill routing, weak source-of-truth checks, validation gaps, wrong worktree/runtime, tooling/permission blockers, or unclear user-intent extraction.

Quick Start

Compare the Codex session id 019e3f7c-d54b-71b3-ba08-2ee20ce6be27 against the most recent 100 local Rudder-related Codex sessions and report which proxy metrics and failure classes explain any performance gap.