steveseguin

Refresh the community benchmark snapshot (data/localmaxxing-snapshot.js) from the public Localmaxxing API, understand how gold cases are selected and what the rows mean, triage a failing weekly refresh, and map new models/hardware so their runs become…

Re-fit or re-anchor the planner's decode/prefill physics (FRAMEWORK_PROFILES, LAYER_OVERHEAD_SCALES, device kernelOverheadScale) against the gold-case corpus, triage outliers and roofline violations, and re-pin the exact-value regression tests. Use when…

Add or update a device template in DEVICE_TEMPLATES (engine.js) — GPU, Apple/AMD unified-memory system, CPU, NVMe tier, or accelerator — with official peak specs, the right backend overhead scale, default runtime, picker group, and evidence mapping. Use when…

Add or update an LLM preset in MODEL_PRESETS (engine.js) from its official config.json so the planner's physics (KV cache, FLOPs, fixed per-layer overhead) are right for it; wires the picker, the evidence rules, and the tests. Use whenever a user asks to…

Build, update, validate, and publish neural.download model packages and human-facing deployment guides from repository evidence. Use when adding a model, quantization, GPU topology, OS/container variant, reproducible recipe, featured benchmark,…

Review, reproduce, benchmark, credit, and integrate outside model-optimization work in the B70 optimization lab. Use when a contributor submits or points to a patch, recipe, benchmark, model lead, issue, pull request, fork, or external optimization; when…

Use the installed ninja-p2p CLI when the user wants Codex to coordinate with room peers, keep or wake a sidecar inbox, exchange files, run resumable swarm transfers, or bridge Social Stream Ninja over WebRTC.

Use the installed ninja-p2p CLI when Claude should coordinate with room peers, keep or wake a sidecar inbox, exchange files, run resumable swarm transfers, or bridge Social Stream Ninja over WebRTC.

Send a specific task or selected text file to an already paired VDO.Ninja background agent, retrieve its reply or cancel its requests. Use for work delegated to that paired agent, not ordinary questions or access to the owner's filesystem.

Connect agents or a human browser through VDO.Ninja, exchange messages and selected files, or manage invitations to a separately installed background Codex service. Use for VDO.Ninja peer coordination, not ordinary web browsing or native ChatGPT voice calls.

Inspect, test, and control the Social Stream Ninja standalone Electron app through its opt-in localhost API and MCP adapter, including source events, diagnostics, screenshots, semantic page inspection, and safe human handoff.

Use VDO.Ninja WebRTC data channels as MCP tools for bot-to-bot messaging and reliable file transfer, with clear transport limits, optional TURN, and optional join-token/session-MAC security.

Install, configure, and test the Caption Local transcription and speech-to-English translation service using native Python or Docker on Linux or Windows, including optional NVIDIA acceleration.