mirror of
https://github.com/ruvnet/ruflo.git
synced 2026-09-14 14:01:28 +08:00
0fd61e3e5b
Wires `npm agent-booster` (Tier 1, WASM, $0/edit, sub-millisecond) into the cost-tracker plugin and proves the speedup against frontier LLMs running on the deployed GCP endpoints — Gemini 2.0 Flash, Claude Sonnet 4.6, and Claude Opus 4.7. What lands: - ADR-0002 (Accepted): 6 decisions covering booster routing, optimizer bridge, model-outcome feedback loop, loop-worker consumption, tier-aware reporting, and the version+keywords bump. - 4 new skills: cost-booster-route (router-side classifier), cost-booster-edit (direct apply via agent-booster.apply()), cost-benchmark (corpus harness), cost-compact-context (token-optimizer bridge with graceful fallback). - Updates to 2 existing skills: cost-optimize closes the hooks_model-outcome loop; cost-report now reads measured booster data from runs/latest.json and aggregates by tier. - New `cost benchmark [--llm] [--anthropic]` subcommand. - 16-case verification corpus (12 Tier 1 + 4 adversarial), bench.mjs harness with split metrics (winRate / escalationRate / overallCorrect) and gcloud-secret-aware LLM/Anthropic baselines. - Smoke contract grew 10 → 27 checks, ~85 ms wall-time. Verified live (16 cases, 4 endpoints, 2026-05-05): Endpoint Tier 1 Adversarial Avg latency Cost/edit Booster (WASM) 12/12 escalates 4/4 0.50 ms $0 Gemini 2.0 Flash 12/12 3/4 762.13 ms $0.000027 Claude Sonnet 4.6 12/12 3/4 1158.06 ms $0.000982 Claude Opus 4.7 12/12 3/4 1517.94 ms $0.006049 Measured speedups (booster vs LLM): 1524× / 2316× / 3036× — exceed upstream's 350× claim against every endpoint. All three LLMs (including Opus 4.7) failed on the same adversarial case; booster correctly refused all four with min confidence 0.000. Proof: https://github.com/ruvnet/ruflo/issues/1743 https://gist.github.com/ruvnet/15fae7a5495026f025fb3baf721c20ea Co-Authored-By: RuFlo <ruv@ruv.net>
19 KiB
Executable File
19 KiB
Executable File