Shared multi-agent harness state and coordination for Codex, Claude Code, and Antigravity. Use for repo resume/state, shared memory or active-task handoff, plan claims/recovery, fail-ledger checks, stale `.oms` cleanup, role/executor setup, local provider calls or write delegation, patch admission/landing, artifact provenance, and prior-session handoff.
Discovery-focused chem-bio ML guardrails for small molecules, molecular 3D and physics, proteins and variants, biomolecular interactions, peptides and antibodies, reactions and synthesis, generative design, cellular omics and phenotypic profiling, and biomedical networks. Use before writing, reviewing, evaluating, or training on RDKit/SMILES/SELFIES, QSAR/ADMET, quantum or force data, MMseqs2/MSA/protein language models, PDB/AlphaFold/DMS, docking/DTI/PPI, peptide-MHC or antibody data, reaction/retrosynthesis/yield data, molecular or protein generation, RNA/oligonucleotide/CRISPR, single-cell perturbation/Cell Painting, or biomedical knowledge graphs. Enforce domain splits, leakage checks, label semantics, baselines, metrics, calibration, applicability domain, and provenance.
PyTorch training correctness for optimizers and schedules, distributed/DDP, variable-size masked losses, checkpoint save/load, CUDA runtime settings, and equivariant model checks. Use when writing or reviewing training loops, optimizer/scheduler setup, multi-GPU code, loss reduction, or checkpoints.
Multi-agent code review. Use when the user explicitly requests independent verification, cross-agent/council review, or a release go/no-go or requested ML pre-training gate needs several independent reviewers. High-risk changes alone do not require this skill.
Hypothesis-driven loop for ML and scientific experiments. Use to turn a question into a falsifiable hypothesis, design the smallest disconfirming run, pre-register prediction/baseline/metric, interpret results, coordinate experiment claims, reproduce or compare runs, or trace an output to its run.
Slurm/HPC workflow helper. Use when working on clusters, partitions, nodes, sbatch/srun jobs, GPU/CPU allocations, queues, logs, checkpoints, or any resource-heavy command — e.g. "submit the job", "check the queue", "잡 돌려줘", "큐 확인", "GPU 몇 장", "학습 돌려줘" on a cluster; also digesting a long job log (job-digest, "로그 요약") and reconciling finished Slurm jobs into shared state (run-reconcile, sacct). Reads local cluster reference when generated.
Specification-first workflow for new projects, existing-repo onboarding, unresolved draft specs, and broad or architecture-shaping features. Detect project state before asking questions and resolve only choices that affect implementation. Clear bounded changes do not require an interview.
Delegate a write task to another local agent CLI (Codex, Claude Code, or Antigravity) from inside the current agent session. Use when the user says "have codex do it", "delegate to", "ask another agent to implement", "codex한테 시켜", "위임해줘", "다른 에이전트한테 맡겨", or wants a second agent to execute work in parallel — including working a shared plan task (--plan-task) or driving the worker with a reusable persona (--role). The worker runs in an isolated git worktree and returns a patch that should go through patch-admit/patch-land before touching the main tree.