What problem does it solve? Optimizing ICFPC Littleman .man programs blindly is slow and error-prone: without knowing which blank cells the runner actually traversed or where I/O wait stalls occur, rewrites are guesswork. This Skill profiles representative public cases with the official WASM per-tick trace, separates traversed from unobserved blank cells, and pinpoints execution and I/O-wait hotspots so local structural rewrites can be benchmarked and judged before adoption. ## Core Features & Use Cases - Trace-guided profiling: Runs profile-trace.mjs against benchmark cases from trace-guided-v1.json to produce agent-readable JSON maps with room/cell execution coverage and s/S/r/R/U receive/send wait statistics. - Local rewrite exploration: Uses optimize-trace.mjs to search non-rigid neighborhoods around a critical pipe, emitting a candidate only when official analysis, structural validation, full public judging, and exact score improvement all pass. - Diagnostic judging and telemetry: Validates candidates with judge.mjs in diagnostic-only mode and records every run, artifact, and metric through skill-telemetry.mjs. - Use Case: Given a slow History Lesson submission, profile its representative public case, identify the longest runner wait hotspot by room/cell coordinates, generate a local structural rewrite candidate, and confirm an exact score improvement via diagnostic judging before handing off for submission. ## Quick Start Ask the agent to profile a chosen .man solution's representative public case with the official trace, identify the I/O-wait hotspots, and benchmark a local structural rewrite candidate.
