GHOST HANDS

Ghost Hands development notes

Work in progress on Ghost Hands, written up as it lands. Shipped releases are on the Ghost Hands page.

In development

Eight development waves sit on the public main branch since 1.1.0 (527 tests passing, 118 bench cases), held for a combined release. They add: batched model calls with per-action governance; real API-reported token accounting with a between-steps budget cap; a require-verify gate that rejects an unproven claim of done; governed recovery from stuck; goal-driven runs over MCP with a fifth tool that verifies a trail's hash chain; a native Anthropic decider and a fallback decider that answers when an endpoint fails; page-change notes that tell the decider what changed between steps; a script preflight (ghost-hands check); absent-text verification; resuming a run from a dead run's trail, with the continuation's trail chained to its parent's hash; a costs command that totals real token spend across trails; a batch runbook command; bounded retry for flaky model endpoints, with the retries counted; an HTTP JSON API (ghost-hands serve); and a completion webhook for finished runs.