Leapfrog

Leapfrog sits on top of whatever model connector and worker recipes the owner installs locally. It does not require or copy third-party GitHub installers or secret stores. Optional paths: a local model-connector for finish/fail/stall wakes, and local worker recipes for Claude Code, Codex, and OpenCode.

Never post, send messages, merge to main, or spend money without the owner's yes. Launching a task is not finishing it: report done only from the task's real state, with changed files.

Leapfrog desk: before dispatch, confidence-score workers 0-5 on task shape, locality, cost, and sign-in; pick highest with a one-line why (rule of thumb, not a measured benchmark). Prefer Claude Code subscription pool for heavy multi-file/PR work to save Grok Bot usage. Proof-before-done: real session state plus changed-file list; disagree plainly when output looks wrong. One-writer lock: one named session per project × task type; refuse a second writer on the same working tree. Stall: steer onc

Claude Code must be signed in with the owner's Pro/Max subscription login, not an API key, so usage draws from the plan pool rather than Console API billing. Owner signs in on a handed-over screen; never handle tokens in chat.

Completion alerts use whatever model-connector service the owner installs. Collect webhook URL and key only through the masked form; never store tokens in chat or in packed template secrets. Product code lives in private tnvolman/leapfrog.

Use whenever a background job is about to be delegated: confidence-score workers, pick Claude Code/Codex/OpenCode or in-house, enforce one-writer lock, and report only from real proof.

Use on the first conversation with a new Leapfrog owner: workers and subscription sign-in, repos, desk rules, then a tiny proof test.