exp-down
01Command Claude Code
Power off (or release) the AutoDL instance for the current experiment to stop billing.
Python package + Claude Code MCP server for running ML experiments across a local dev machine and remote AutoDL GPU containers — git worktrees, mutagen sync, AutoDL instance lifecycle, manifest bookkeeping, and a paired docs repo.
This repository also configures its own agents. See what projtool tells them →
Command Claude Code
Power off (or release) the AutoDL instance for the current experiment to stop billing.
Command Claude Code
Start a training run on the current experiment branch. Provisions an instance if needed, syncs code, and launches in tmux.
Command Claude Code
Show the status of all in-flight runs and instances for this project.
Command Claude Code
Adopt a pre-registry docs layout (plans/, exp/) into the research-thread registry — scan for drift, dry-run the migration proposal, confirm the mapping WITH THE USER, then apply. Never guess history.
Command Claude Code
Read the experiment record (journal, findings, reports, run table) and propose the next 1-3 experiments with rationale and cost — the user decides, then execution goes through startruns/starttraining.
Command Claude Code
Start a new analysis report in the independent docs repo, prefilled with frontmatter from the cited runs (single branch or cross-branch summary).
Command Claude Code
Distill this work session's experiment outcomes into the experiment's findings.md (what was tried, what failed and why, what's now established) and commit it to the docs repo.
Command Claude Code
Close a research thread — conclude it with an evidence-citing conclusion.md, or abandon it with a one-line reason. Terminal is terminal; to revive, open a successor thread with a supersedes edge.
Command Claude Code
Open a new research thread (a research line) in the docs registry — pick a slug, create the record + idea.md/plan.md skeletons, interview the user to fill them, and point forward to the first experiment.
Command Claude Code
Babysit every active run to completion — staged polling, auto-finalize, progress report; no raw SSH loops.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: