AI Agent Workflow Collection

Use when an agent/runner with its OWN training code needs to make a Chronicle run durable and reproducible — phrases like "snapshot the environment", "upload checkpoints as it trains", "resume training from the last checkpoint", "continue the crashed run". It owns three things: capturing an environment snapshot at run start (library/hardware versions + your custom config state), uploading resumable checkpoints during training, and discovering + downloading the latest checkpoint to resume from. It is training-AGNOSTIC — the training loop is the caller's; this skill never writes it. It is the durable-state companion to chronicle-run-variation (which owns the run lifecycle + W&B linking): run-variation marks the run, this skill makes it survive a crash and continue. Distill afterward with chronicle-write-report.

Use when the user (or an agent flow) wants to take a committed or concluded Chronicle experiment — or a single variation — out of use while preserving the record: phrases like "retract this experiment", "this result turned out to be wrong", "flag this as invalid", "mark these findings as withdrawn", "the bug invalidated those runs". Retraction is a soft flag with a required reason: the row, lineage, and audit trail survive; the experiment's output assets are invalidated; live agent deployments are torn down; the GitHub repo is archived read-only. For deleting an open (uncommitted) draft use chronicle-delete-experiment instead — hard delete is refused once the experiment is committed.