Ordewell uses a read-only coding agent as planner, then spawns per-task agent sessions gated by a marker-based verdict engine

GitHub·low signalXBlueskyLinkedInCopy link

Ordewell turns one natural-language goal into an ordered plan of tasks, each assigned its own runner, model, thinking-effort level and execution mode, then executes them respecting a dependency graph across VS Code extension, terminal UI and CLI frontends over a shared core engine. The planner is itself a coding agent (Claude Code, Codex or OpenCode) running read-only over the workspace, interleaving clarifying questions with research in one persistent conversation, which avoids needing separate API credentials. The verification design is the interesting bit: the VerdictEngine only completes a task once its marker appears, so an agent session that exits without one fails visibly instead of silently passing. Early project at 104 stars and 91 commits, surfaced on Hacker News 2026-09-15.