huangruiteng · 2026-08-04 · notable
LoopX 0.4.1 — durable state kernel for long-running coding agents
LoopX is an MIT-licensed Python control plane that gives Codex, Claude Code, and other agents durable goals, executable todos, and evidence logs across long sessions. v0.4.1 adds persistent todo routes and goal-continuation contracts.
LoopX is a small durable-state kernel that stops long agent runs from forgetting their own plan.
What is it?
LoopX is an MIT-licensed Python control plane that sits alongside coding agents like Codex and Claude Code. It stores goals, executable todos, evidence logs, and handoff contracts in a durable store, so an agent's plan survives crashes, quota resets, or a new turn starting cold.
How does it work?
The kernel treats each agent turn as a bounded step against a Goal host — a persistent objective with selected-todo context. v0.4.1 adds Todo routes that survive across turns, so an agent resumes work without re-inferring context, plus a quota-aware auto-wake that reactivates a goal at real state transitions instead of polling.
Why does it matter?
Long agent runs fail today because the model's scratchpad is the only memory — restart the loop and the plan disappears. LoopX moves that plan out of chat memory into a small kernel, so a paused or crashed run can resume against the same objective without re-reasoning from scratch.
Who is it for?
Engineers running Codex or Claude Code sessions for hours or days
Try it
git clone https://github.com/huangruiteng/loopx