Farnsworth-Loop

Run a Farnsworth Loop tournament in one of two modes. The sigil is @@FL[:N][:M[:Z]] — N (optional) = attempts per round, M = passes (1 single, 2 two), Z = grand loops (Z>=2 = an UNATTENDED chain that, per loop, runs a full tournament, implements the winning proposal into your real repo on a new FL-<loop>-<random7> branch, runs fail-closed verify, and opens one PR — never auto-merged; Z=1 or omitted = today's isolated tournament; Z capped at Z_MAX=5); N may be inferred from a prose model spec like '2 opus, 2 glm 5.2, 1 codex high' (sum of counts = N, the items become the per-attempt Mixed assignment) or the Top Mixed preset ('top mixed' + N spread over opus/glm-5.2/codex-high), and bare @@FL falls back to the interactive model gate. A marker-adjacent task-size word (short/medium/long, e.g. @@FL:5 long) optionally overrides the per-attempt turn/timeout limits, which otherwise scale to the orchestrator's estimate of task size. First ask the user which model quality to use for the attempts (Anthropic Opus, Sonnet

Benchmark generation throughput (cold vs hot tok/s) for every model the farnsworth-loop system can call (Anthropic / GLM / local MLX / codex / MiniMax). Two workload profiles — light (tiny paragraph) and heavy (>5k-token input context + long >5k-token output, representative of coding/agentic work). Thin wrapper over bin/fl-bench.mjs. Use when the user asks to benchmark model speed, measure tokens/second, compare cold vs hot throughput across providers, or run /fl-bench.