dotfiles

Discover other Claude Code sessions running on this machine and message them directly — list who is live and what they are working on, then send a message straight into another session's chat so the two Claudes can talk. Use when context needs to move between sessions instead of the user copy-pasting it by hand, when work in one session depends on what another is doing, or when the user says "ask the other Claude", "tell the session in X", "what is the other Claude doing".

Hand the keyboard to the user so they can edit a file themselves mid-session, in a real nvim popup, then read back what they changed. Use when the user says things like "drop me into <file>", "let me hand-edit X", "open X in nvim for me", "I'll edit this part myself", "pause so I can tweak it", or otherwise wants to make manual edits rather than have you make them. Requires tmux + nvim.

Put a live tqdm progress bar (items/second, ETA) in the status line for any long or repetitive job — batch loops, N-item processing, writing files or JSONL shards, downloads, sweeps, evaluations, migrations — instead of spamming the transcript with "1/200", "2/200". Use whenever a job iterates more than ~20 times or runs longer than ~30 seconds. The `ccbar` CLI can watch a growing file or directory, so the job needs no changes at all.

Quick honest terminal distribution plots (sparkline histogram + box-and-whisker) of numeric data. Use when the user wants a fast in-terminal look at the distribution/spread of some values (e.g. coherence/alignment scores across models) — "histogram of these", "show me the distribution", "plot these numbers", "box plot of", "what do these look like". For throwaway terminal viz, NOT publication plots.

Run the adversarial, read-only codex reviewer on the current diff and fix what it finds. Invoke as /reviewer with optional focus text and scope flags — it kicks off codex immediately (spends codex/OpenAI quota) and injects the report for Claude to act on. codex reads the whole repo to dig but CANNOT edit anything.

Manage RunPod GPU pods and run LLM fine-tuning experiments remotely. Use when user asks to create pods, run training, check training status, manage volumes, or do anything RunPod-related.

Run a shell command in a specified tmux pane and return the pane's output. Use when user asks to run something in a tmux pane, send a command to tmux, or check tmux pane output.

Create W&B reports for experiment reviews. Use when asked to create a report showing recent experiment results. Drafts an outline for approval, then creates the actual report.