Gemini, Google AI Studio, and Vertex AI workflows. This directory includes agent skill packages and prompt templates; each card identifies its type. A prompt template is copied from its source, while a skill package needs the appropriate agent setup. Compatibility may be inferred from the content unless a recorded test is linked.
Validate one plan against the codebase with COLD seats (the owner's standing choice, or the one-shot picker the first time and on request: 1-5 models from families different than the Player that authored the plan (codex/Astra, Claude, Grok CLIs at $0 subscription; OpenRouter doors from a curated roster), so the read is genuinely independent. Every seat runs at high reasoning, dispatched in parallel, and findings are merged across seats with seat attribution, never voted. Read-only but for two residues: the saved seat choice in the repo profile (the first pick, or a change) and the gate: plancheck line, written on request and always for a plan under or headed for the repo's ready-to-build plans folder. Use when the user runs /f-plancheck or /cv, wants a plan checked before building, or asks for cross-validation. This is the cold-seat protocol, not the validator itself, it dispatches plan-validator; reach for that directly only when cross-family independence does not matter. Also answers
Set up Telegram and read or send messages in the owner's personal account through tg. Use when asked to install or connect Telegram, find a chat or person, read a conversation, or send a message.
Gate fine-tuned checkpoints with drift budgets, paired comparison, and forgetting checks before promotion. Use after a training run produces a checkpoint, when deciding whether a tuned model ships, or when a promoted model needs re-gating against updated goldens.
Build the evaluation harness that gates every fine-tuning run — golden sets, per-failure-mode graders, judge calibration, and base-model baselines. Use when starting a fine-tuning effort, when converting traces into an eval set, or when calibrating a judge against human labels.
Train reasoning and verifiable-task behavior with GRPO and reinforcement learning from verifiable rewards (RLVR). Use when task success is algorithmically checkable (math, code, tool calls, structured output), when designing GRPO reward functions, or when a GRPO run diverges or reward-hacks.
Align a fine-tuned model with preference data using DPO, ORPO, KTO, or SimPO. Use when preference pairs or thumbs-up/down feedback exist, when choosing between preference-optimization methods, or when a DPO run needs hyperparameters or debugging.
Convert evaluation traces and production logs into SFT examples and preference pairs. Use when graded traces or failure examples exist and need to become training data, when applying rejection sampling to model outputs, or when building DPO pairs from passing and failing runs.
Fine-tune vision-language models (VLMs) with supervised learning on image+text data. Use when adapting a VLM to a visual domain or task, configuring frozen-vision-tower LoRA, or debugging a VLM fine-tune that trains without learning.