kaggle

# Kaggle Build a public-readable Kaggle notebook with an evidence-backed problem profile, visual EDA, stage-level sanity checks, reproducible training, inference, and submission validation. Write notebook scripts only; use `develop` for packages or production modules and `research` for literature surveys. ## Input Schema ```json { "competition": "required output slug", "context": "competition URL, pasted description, local dataset metadata, or existing notebook path", "problem_type": "optional classification|regression|segmentation|detection|tabular|time-series|point-cloud|mixed", "mode": "full|eda-only|inference-only", "offline_setup": "optional boolean", "resume": "optional existing Jupytext .py path", "keep": "optional user-specified content that must survive regeneration", "done_when": "the grounded notebook is written, structurally verified, and recorded in a validated result artifact" } ``` Default `mode` to `full`. Resolve mode behavior and output paths only through `references/composition.md`. ## Workflow ### 01: Create the run and normalize input Create `.reports/codex/kaggle/<timestamp>/` and keep the active plan current. Record normalized inputs in `profile.md`. - Require a filesystem-safe lowercase slug containing only letters, digits, and hyphens. - Reject conflicting or unsupported mode inputs. - Require `resume` to exist, be readable, and use Jupytext cell markers. - Treat unknown options as blocking until the user confirms whether to ig...