pr-autopilot

A Servers pane for this project's dev servers: what is listening, on which port, since when, with Start, Stop, Restart and Log buttons; names the process holding a port when a command fails with EADDRINUSE; a read-only dev_servers tool for Claude. Scans with netstat (never lsof), one session for all; no model calls.

Notices files that land in Downloads while you work (PDFs, 3MF models, images, notes) and offers them in a band above the prompt: Attach puts @"path" mentions in the prompt box, Dismiss clears them. /downloads lists the newest files and attaches any by number. No model calls.

Routes each turn's effort from your prompt: git housekeeping replies ("merged", "commit and push", "push it") run at low effort, deep asks (review, audit, plan, why does...) at max, everything else (approvals like "yes" or "continue" included) at the session's own. Switches only when the prompt cache makes it worth it, never mid-turn; optional model guard; /route shows and steers it. No model...

Live pane of long-running jobs (training, downloads, extractions): progress, ETA, stalls and finishes, read from their logs with no model calls; jobs_list / jobs_watch / jobs_unwatch tools so Claude follows jobs instead of polling them.

Memory, swap, GPU and the CPU cores macOS is using on the status line, and a machine_status tool for Claude; refuses heavy local jobs (training, inference, rendering, Docker) while memory is critical or the Mac is reserved with /busy.

Watches how your other mods really behave: hook failures and slow hooks (read off the chain's own trace), what each one did (toasts, status lines, commands, processes, model calls, writes), with /mods for a live pane and /mods report for the last day, week or month. Changes nothing they do; no model calls.

Watches your Modal apps from the CLI, in projects that use Modal: running apps and containers on the status line, a toast when one has been up too long, today's spend where the CLI reports it, and a pane to stop an app (after you confirm). One session polls for all the open ones. No model calls.

Makes past sessions recallable: indexes Claude Code and Codex transcripts, subagent runs, memory files, standing orders and second-opinion reviews into a local SQLite index (kept after Claude Code deletes old transcripts), gives Claude search, expand, recap and list tools, and adds /recall (search, last session, timeline, decisions, commands, files, PRs, ask), /remember, a...

Gathers the repo's state at session start (branch, ahead/behind, uncommitted work, recent commits, open PRs with CI, focus-label issues, stale branches) with no model calls, shows it in a band above the prompt and hands Claude the same summary in its first message.

Tells you when a scheduled routine (a scheduled-task run) stops to wait on you: a Mac notification, an optional phone push and a status line once a permission ask has stood 20 s (configurable; auto mode decides most asks by itself sooner) or when it asks you a question, plus a note when it finishes. Optionally lets routines read the web without asking. No model calls.

/second-opinion asks Fable (or the model you set) for a candid senior review of your recent commits, branch, uncommitted diff or a plan file, in the background without switching the session's model. Each run is one paid model call against your usage; the review opens in a pane, is saved under ~/.claude/second-opinions/, and can be handed to Claude with your next prompt.

Holds back a prompt that looks like it carries a secret (API keys, tokens, access-key pairs, private keys, passwords) before it reaches the transcript, puts it back in the box, and offers to save the secret as an env var in ~/.zshrc and send $NAME instead. No model calls.