new-experiment

name	new-experiment
description	Scaffold a new model + dataset + config triple inside an existing curryTrain project. Activate when the user asks to "add a new model to curryTrain", "scaffold a new experiment", "generate model.py / dataset.py / config.yaml for X", or describes an idea they want to start training. Delegates the actual generation to the scaffolder agent.

Stage 1 entry point — scaffold a model package

Create a four-file model package (config.py, model.py, checkpoint.py, protocol.py) under curry_train/models/<name>/, plus a starter Hydra config and a minimal dataset adapter. Most of the actual file generation is handed off to the scaffolder agent (see agents/scaffolder.md).

When to invoke

• User says: "add a model called X", "I want to try architecture Y in curryTrain", "scaffold an experiment for Z".
• User has an idea but no code yet — this skill is the entry point for Stage 1 (Skeleton).

Inputs

• <model-name> (required): kebab-case identifier. Becomes the registered name in register_model(...). Reject names with spaces, slashes, or non-ASCII characters.
• --task (optional): one of lm, cls, mt, cv, snn. Drives the dataset adapter template the agent picks.
• --from (optional): a HuggingFace model path the user wants to start from (e.g. --from=Qwen/Qwen2.5-1.5B). The scaffolder uses this to derive shapes and pre-fill config.py.

Procedure

1. Verify the user is inside a curryTrain project. Look for curry_train/__init__.py and curry_train/models.py. If missing, point the user to /curry-train:init.

2. Validate the model name. Lowercase, alphanumeric, hyphens only. If conflict with an existing curry_train/models/<name>/ directory, halt and ask whether to overwrite.

3. Delegate to the scaffolder agent. Spawn the scaffolder subagent with:

   • The model name, task type, and (optional) HF source.
   • A pointer to the four-file template described in skills/stage1-scaffolder/SKILL.md.
   • Instructions to also create a starter Hydra config under configs/<name>.yaml.

4. Run preflight asserts. After scaffolding, immediately invoke the checks listed in skills/stage1-preflight-asserts/SKILL.md (zero_grad / mode / shape / leakage). Surface failures back to the user before they try to train.

5. Suggest the next two stages. Print:

   • Stage 2: try skills/stage2-overfit-single-batch (the gold-standard sanity check).
   • Stage 3 (when ready): pre-validation skills before scaling up.

Boundaries

• This skill does not train. It only scaffolds.
• It does not download datasets. The dataset adapter is a stub that the user wires up.
• It does not propose hyperparameters — that's hpo-proposer agent territory in Stage 4.
• For SNN tasks (--task=snn), the scaffolder uses primitive-lif-neuron and emits a 4D tensor shape contract (B, T, N, D). For non-SNN, shape stays (B, N, D).

Failure modes

• No HF access when --from was passed: the scaffolder must fall back to asking the user to provide a local config / weights path. See primitive-hf-bridge skill for the offline fallback procedure.
• Conflicting registered model name: halt; do not silently rename.
• Unsupported task type: list the supported set and ask which one applies.