Skip to content

LoopX 2.0: TypeScript control plane, long-horizon research workbench

中文指南 · Release and verified assets

LoopX keeps goals, ownership, decisions and evidence durable while your Agent host supplies the model and tools. 2.0 connects a two-month control-plane reconstruction with a workbench for inspecting research across sessions.

Choose your upgrade path

2.0 is a breaking release for unpromoted Markdown Goals and retired legacy execution policy. Reading retained history and recovering a supported original operation remain separate from permission to start a new write. An installation never silently converts a Goal, changes its provider or releases its leases.

Your current state Before continuing work in 2.0
Canonical File/SQLite with soft_claim or hard_lease Preserve the selected provider, update through its installation owner and verify package/runtime identity.
Unpromoted Markdown Goal Stop its writers, verify exit, settle leases and pending capture/outbox through their owners, then review and import the original source.
Canonical store using legacy ownership Review a claim-preserving ownership-policy migration to soft_claim or hard_lease; this is separate from storage conversion.
Existing target-only settings or explicitly disabled canonical creation Inspect and explicitly select compatible creation settings before creating a writable Goal; disabled does not mean enabled.

If you cannot perform the reviewed migration yet, retain the verified 1.4.2 installation for those Goals. Do not run two writers against the same state.

Install and verify

Require Python 3.11+ and Node.js 22.22.3+; Node 24 LTS is recommended. Keep one pip, pipx, uv-tool or archive installation owner. In an appropriate Python environment, install this exact release:

node --version
python3 -m pip install --upgrade 'loopx==2.0.0'
loopx workflow-skills --install
loopx --version
loopx doctor --deep

Expect loopx 2.0.0 and a ready packaged TypeScript runtime. Restart the Agent host after first installing workflow skills. For existing installations, inspect loopx update check and loopx update plan before loopx update apply. After updating, read back enabled extensions with loopx extension doctor --all-enabled --execute.

The browser/PWA entry is loopx dashboard. Native desktop downloads remain preview shells and depend on the complete platform build assets; consult the desktop guide. A CLI upgrade and an App upgrade have independent identity readbacks.

Import an old Markdown Goal

Open Goal settings → Task ownership → Data storage. Stop the affected writers and verify process exit, settle retained leases, then use Back up and preview import. Review the complete active/archive inventory and confirm Import reviewed Markdown source. A checkbox cannot bypass an unsettled lease or changed source. The same transaction is available to an Agent:

loopx --format json coordination-shadow prepare-import --goal-id GOAL --operation-id IMPORT \
  --provider sqlite --target-handoff-mode hard_lease
# Review the returned immutable plan and its exact plan_sha256 before applying.
loopx --format json coordination-shadow apply-import --goal-id GOAL --operation-id IMPORT \
  --plan-sha256 SAVED-SHA256 --writers-stopped --execute

Prepare is not a writer stop or an import. It creates a private scoped authority archive and plan. That archive is not a complete disaster-recovery backup. After a lost response, use Read original preview and current storage, or coordination-shadow recover-import with the same operation and digest. Do not replace an ambiguous operation with a fresh import. Replacing the binary does not remove a committed writer fence or roll state back. Follow the full migration and recovery contract for source verification, File/SQLite alternatives and supported historical input.

What changed in the control plane

The migration began with the August 15 RFC, then moved from leaf helpers to complete transactions. Effect execution, settlement, canonical authority and migrated ownership rules share TypeScript owners. Python retains real provider and Host IO, compatibility codecs and specialized computation. This is not a claim that every Python module has been rewritten.

Three engineering details matter when using or extending it:

  • Check before effects. Preflight validates the exact operation before a provider acts; final reduction records its result. Fewer RPCs are not a reason to move rejection after an irreversible operation.
  • Recover the original result. An operation receipt can recover committed work without renewing old execution authority or overwriting later writes. An uncertain external effect stays uncertain without independent readback.
  • Preserve decision semantics. The Semantics work traces definition owners, values, producers and projections. CLI output budgets compare real stdout and growth while preserving obligations, stop conditions and detail routes.

Read the full engineering history and core-only statistics, including new maintainer Song's Semantics work and Duang777's contributor-led retirement. The historical statistics end on October 16; they are not release-tag totals, unique developer counts, productivity measurements or benchmark gains.

Follow a research question through the workbench

Open a Goal, inspect its Agents and tasks, then follow recorded Explore history to an experiment, artifact and evidence. Deep analysis reads a registered trajectory with citations, six separate process judgments, coverage and missing data. Native outcomes and measured solver costs have separate source bindings. An Agent can prepare the source pack; the owner need not enter private paths.

Reading pages, refreshing, switching sources and opening citations do not run analysis. To run it, select an authorized source, review the requested model operation and explicitly execute it. Missing or changed sources remain visible; a failed source switch never shows another source's result under the new name.

The film combines labelled recorded excerpts, synthetic examples and illustrations.

中文指南

LoopX 2.0 的两条主线是 TypeScript 控制面重构与长程研究工作台： 让目标、任务归属、决策和证据接续起来，让研究者能追到产物、轨迹与下一步依据。 模型和工具仍由你选择的 Agent 宿主提供。

先判断是否需要迁移

这是有兼容性变化的版本。 未迁移的 Markdown Goal 不再接受普通写入； legacy 也不再授权新的执行。历史可读、原操作可恢复、当前执行获准，是三个独立条件。 安装新版不会自动迁移 Goal、切换 provider 或释放租约。

  • 已用 canonical File/SQLite 且为 soft_claim / hard_lease：保留配置，按原安装渠道更新并核对 runtime。
  • 未迁移 Markdown Goal：先停止相关 writer / Host、确认进程退出、结清租约与 capture/outbox，再预览、备份并导入完整原来源。
  • canonical store 仍用 legacy：单独审阅保留 claim 的执行策略迁移；不必因此再次转换存储。
  • 仍用旧目标配置或显式关闭 canonical creation：先明确兼容的新建设置，不将关闭解释为开启。

暂时不能迁移的 Goal，可保留已验证的 1.4.2 环境；不要让两个 writer 同时操作同一状态。

安装、升级与读回

需要 Python 3.11+、Node.js 22.22.3+，推荐 Node 24 LTS。在合适的 Python 环境中执行：

node --version
python3 -m pip install --upgrade 'loopx==2.0.0'
loopx workflow-skills --install
loopx --version
loopx doctor --deep

预期版本为 loopx 2.0.0，打包 TypeScript runtime 就绪。首次安装 skills 后重启宿主。 既有安装先执行 loopx update check、loopx update plan，再按计划执行 loopx update apply； 之后运行 loopx extension doctor --all-enabled --execute 核对已启用扩展。 浏览器 / PWA 入口为 loopx dashboard；桌面包仍属预览，CLI 与 App 要分别核对版本。

旧 Goal 在 Goal settings → Task ownership → Data storage 中操作： 停止 writer 并确认退出、结清租约，再选择 Back up and preview import， 检查完整 active/archive 清单后确认 Import reviewed Markdown source。 Agent 可复用上方同一组 coordination-shadow prepare-import / apply-import 命令。 预览不会自动停进程或导入；范围内 authority archive 不等于完整灾备。 响应丢失时读取原预览及当前存储，沿原 operation / digest 恢复。 不能靠重装旧二进制清除已提交的 writer fence，也不能覆盖后来的写入。

看懂重构，再使用研究工作台

迁移从 8 月 15 日 RFC 开始，经历完整事务、共享权威、真实调用方、存储成本、 Semantics 与旧 writer 退役。已迁移的领域规则统一由 TS owner 解释； Python 保留真实宿主和 provider IO、兼容 codec 与专业计算。 前置准入发生在副作用之前；原回执恢复不续租、不覆盖新写入；CLI 压缩保护决策语义。

中文重构长文与核心统计 保留新 maintainer Song 的 Semantics 贡献，以及 Duang777 主导的收尾。 文中的历史统计截止 10 月 16 日，不能当作本发布 tag 的统计或模型生产率。

打开 Goal 追踪任务、Agent 和 Explore 历史，再通过 深度分析 / Deep analysis 查看 登记轨迹、引用、六项过程判断、覆盖与缺失。原生结果、solver 成本与分析器判断各有来源。 Agent 准备来源，页面不要求用户填写私有路径。打开、刷新、切换来源和读引用都不会启动模型； 运行需要显式选择已有授权来源和模型操作。失败、来源变化与未知不伪装成通过。

影片与双语文字稿 · 轨迹分析合同

影片中的实录、合成示例和示意动画各自标注。