What problem does it solve? When you bring your own training code, you still need the experiment platform to know a run started, stayed alive, and reached a terminal state — otherwise runs silently time out as lost and results never get recorded. This skill drives the Chronicle run lifecycle (start, heartbeat, succeed/fail) around your own training code and optionally links a Weights & Biases run so metrics can be distilled later. ## Core Features & Use Cases - Run lifecycle management: Marks a committed variation's run as running, then succeeded or failed, with heartbeats for long training jobs. - W&B run linking: When WANDB_API_KEY is set, starts a W&B run first and records its pointer on the Chronicle run so distillation can fetch real metrics. - Pre-flight operational checks: Queries checks distilled from other agents' past execution failures before training starts, and reports new operational failures back to the catalog. - Use Case: You committed variation 2 of an experiment with chronicle-author-variation and want to train it on your own GPU box — this skill wraps your training loop so Chronicle records the run and links the W&B metrics for the later report. ## Quick Start Ask the agent to run variation 2 of the current experiment with my training script and record the run in Chronicle with W&B linked.
