LoopX 2.0.0 — TypeScript control plane · long-horizon research workbench

Two months of control-plane engineering meet a research workbench that connects Goals, delegated work, experiment history, artifacts and trajectory analysis.

2.0 upgrade guide · 86-second workbench film · Illustrated guide: Personal illustrated guide

Release Decision

Who should upgrade: Users continuing research across sessions, inspecting experiment history, or building on the migrated TypeScript owners. Operators with unmigrated Markdown Goals should prepare their stopped-writer/import path before upgrading those writers.

What this release solves: It consolidates migrated control-plane rules under shared TypeScript transaction owners and gives users a continuous path from Goal/work ownership to recorded history, cited trajectory analysis and subsequent decisions.

Breaking changes: Yes. Unpromoted Markdown Goals no longer accept ordinary writes; legacy ownership no longer admits new execution. Keep canonical File/SQLite and supported soft_claim/hard_lease choices. An install is not a state migration. Private owner Codex managers default to host tools and gpt-6.1-sol/xhigh, preserving explicit restricted choices. Checkpoint-enabled controller banks rotate completed points by default; see the activation and retention paths below.

How to verify: Expect loopx 2.0.0, a ready packaged TS runtime and the intended Goal’s own storage/ownership identity. Open its recorded history, select a registered source, and read its artifact/citations before explicitly requesting model analysis.

Contributors: Release maintained by @loopx-agent, with the eight community contributors recognized below.

loopx --version
loopx doctor --deep
loopx configure-goal --goal-id GOAL --format json

State Kernel & Control Plane

• Migrated transactions now have shared TS owners through Inbox/Turn observation, task projection, scheduler arbitration, semantic replan and recovery. Retirement removes ordinary legacy writers while preserving supported historical reads and original-operation recovery. #6396, #6860, #6634, #6613, #6602, #6596.
• Private owner Codex managers now execute ordinary authorized local work with host tools; executor and evidence scope follow the coordination root. Policy changes restore visible conversation in native history. #6859, #6793, #6893.
• Shared hosts receive bounded capability context while retaining complete execution-boundary clauses. #6919. Semantics work makes ownership, values, producers and projections inspectable; real CLI-output budgets protect decision-bearing detail while tracking actual emitted bytes. The engineering history covers the multi-month program and Song’s contribution. The core-only census spans August 15–October 16: 62 days, 1,144 first-parent landings and 1,121 associated merged PRs. Historical core statistics end on October 16; tag-range attribution below is independently derived. #6396, #6725.
• Canonical creation defaults to SQLite/hard leases for unconfigured new Goals. Existing Goals retain explicit choices; import and execution-policy migration have separate admission and recovery. #6396.

Capabilities & Workflows

• Research history exposes Goals, timelines, experiments, evidence/artifacts and trajectory sources. Deep analysis keeps citations, coverage, process dimensions, native outcome and solver cost traceable to their respective sources. Original-runtime analysis and nested/container Goal paths extend the same workbench; installed child Goals retain their original source and shared namespace rule. #6648, #6674, #6700, #6878, #6821, #6811, #6910, #6925.
• Memory settings now start with purpose and Agent scope, reusing independent provider bindings and lifecycle verification. #6711, #6589.
• Local media providers compose titles, diagrams, excerpts, captions, audio and encoded video, including editable Blender scenes and verified passes. This release ships the provider/render/readback layer and its compatibility workbench; integration with ordinary tasks and general Workspace artifact acceptance remains staged. #6857.
• Shared Goal planning context, explicit checkpoint histories and private owner-task notifications improve the available research/coordination paths. Each keeps its own activation and authority boundary. #6654, #6785, #6664, #6533.

Quality & Testing

• History observation overhead and digest boundaries are tightened; bad UTF-8 caches recover from live sources and artifact probe failures remain explicit. #6840, #6845, #6790, #6863.
• Release qualification retains complete packaged scenarios and source checks, with separate whole-job completion headroom. pnpm and portable TS shards keep their frozen dependency and platform requirements. #6881, #6833, #6861, #6852.

Benchmarks & Integrations

• Trajectory and research tools retain the original benchmark/source identity. Checkpoint-enabled controller banks now rotate complete points by default. Full controller capture uses a shared 90-second preparation window capped by the original trial deadline; explicit capture and feedback boundaries remain attached to their owner. #6785, #6618, #6624, #6878.
• Native Codex account handoff is an optional integration with the existing native slot owner; automatic client restart is a separate permission. #6783.
• Source/Windows runtime prerequisites, CLI/App upgrade planning and native Goal-lifetime session recovery are improved. App upgrade opens the selected installed copy even when another copy is already running; proven native routing timeouts recover within the original live Turn. #6851, #6839, #6789, #6747, #6917, #6916.
• Lark recovery retains fairness during discovery backoff. #6914. External research adds bounded Ego/X discovery; financial operation consumers add original-attempt confirmation and retry guards. See each optional surface below. #6832, #6731, #6752.

Documentation & Compatibility

• Start with the 2.0 upgrade guide, full engineering article and 86-second workbench film and transcript. README, hero, current manual and book anchors now present the same release. #6908, #6722, #6668.
• Existing state: stop affected writers and verify exit, settle leases/outbox, inspect the immutable complete import plan, then commit/recover through the original operation. A binary downgrade does not undo a committed writer fence. #6396.
• Private steward scope: existing configured external sender policies cover all active registered local Agents, across current/future Goals. To keep selected enrollment, set local_delivery_scope: selected and exact targets in the existing private policy. Source/sender, read, execution-binding and protected-operation grants remain separate. #6664; canonical contract.

Community Contributors

• @songoow — New maintainer Song led Semantics ownership and vocabulary convergence throughout the reconstruction, with runtime/profile and release qualification work. #6733, #6833, #6881.
• @Duang777 — Retirement of ordinary legacy writers and continued-session ownership. #6396, #6747.
• @hhyykk — TypeScript owners for Inbox/Turn context, scheduler arbitration and semantic replanning. #6860, #6634, #6613.
• @JasonBuildAI — History observation overhead, digest ownership, damaged-projection recovery, workspace source ownership and Lark notice qualification. #6840, #6845, #6802, #6891, #6890, #6893.
• @BigDataDZ — Portable TypeScript shards, frozen pnpm dependencies and Windows/source prerequisites. #6861, #6852, #6851.
• @steven-kid — Explicit artifact-probe failures and recovery from unreadable projection caches. #6863, #6790.
• @LIHUA919 — Scoped monitor gates and scheduler deadline/backoff handling. #6803, #6787.
• @hua03 — Paused Goal visibility and status-widget behavior in the Pi integration. #6109.

Optional Capability Activation & Use

Commands that install an optional source package assume a checkout of v2.0.0; they do not change your other installation owner. Use the Agent/UI to prepare private source/configuration files; uppercase Goal/Agent/file identifiers below are caller-bound templates.

Research workbench and trajectory analysis

Activation: Open a Goal’s Explore history and Deep analysis. An Agent registers the authorized source pack; analysis is per operation. Preview with the command below, then explicitly add --execute when the source and model use are authorized.

Validation: Read the selected source, citations, coverage, native outcome and costs. Use the original analysis status/cancel path for a running operation.

Disable / rollback: Do not execute the preview; cancel a running analysis through its original operation. Reading and refreshing remain available.

Authority boundary: Source access and model execution retain their existing owner and account boundaries; a page read starts no model.

Docs: Versioned guide.

loopx benchmark trajectory-assess --input-json evidence-pack.json --output-dir .local/trajectory-analysis --format json

Agent memory and evidence

Activation: Use Settings → Capabilities → Memory by purpose and Agent. Prepare a private provider config; preview and apply only the intended Agent binding.

Validation: Read experiment-status; require the exact provider canary write/readback and current binding digest.

Disable / rollback: Clear only that Agent binding with the scoped command below; retain other Agent bindings.

Authority boundary: Memory is scoped context, not a task, permission or shared-account instruction. Provider/config remains private.

Docs: Versioned guide.

loopx configure-goal --goal-id GOAL --reward-memory-agent-id AGENT --reward-memory-config .loopx/config/reward-memory/agent.json
loopx reward-memory experiment-status --goal-id GOAL --agent-id AGENT --format json
loopx configure-goal --goal-id GOAL --reward-memory-agent-id AGENT --clear-reward-memory-config --execute

Local video and editable Blender

Activation: From this tag’s source checkout, explicitly install/register loopx-media. Real 3D requires its Blender extra in Python 3.11. An Agent prepares the composition. The tagged compatibility workbench remains at Settings → Capabilities → Content production; ordinary-task creation and general Workspace artifact acceptance are still a staged integration.

Validation: Doctor the pinned executable; read video/captions/source digests and, for Blender, the editable project and RGBA/depth passes.

Disable / rollback: Disable loopx-media to stop new renders. Recover or cancel the original render rather than replacing an ambiguous operation.

Authority boundary: Local rendering creates private files and consumes local compute; account use, cloud generation and publication are separate actions. Blender qualification is macOS/Cycles CPU.

Docs: Versioned guide.

uv tool install --python 3.11 './packages/loopx-media[blender]'
loopx extension install --manifest packages/loopx-media/extension.toml --execute
loopx extension doctor loopx-media --execute
loopx extension disable loopx-media --execute

Shared Goal planning context

Activation: Enable Show Goal in planning horizon in the existing Goal review/stride editor, or use the explicit Goal switch below.

Validation: Read the Goal configuration and quota planning projection; look for the applicable planning_horizon.core_goal.

Disable / rollback: Explicitly turn the Goal switch off; clearing an override instead restores inheritance.

Authority boundary: This is planning context. Todo selection, quota, claims, leases and execution admission remain with their existing owners.

Docs: Versioned guide.

loopx configure-goal --goal-id GOAL --execution-planning-horizon-goal --execute
loopx quota should-run --goal-id GOAL --agent-id AGENT --format json
loopx configure-goal --goal-id GOAL --no-execution-planning-horizon-goal --execute

Research checkpoints

Activation: CLI/Host stage: prepare a quiescent canonical source, settled Turn, qualified native session and explicit artifact inventory. Capture/admission/adoption preview by default; execute each operation explicitly. New checkpoint-enabled controller banks rotate by default after a completed capture. The shared full-controller capture allowance is now 90 seconds (previously 30), capped by the original remaining phase deadline.

Validation: Verify the original root and digest; inspect the full inventory and destination’s adopted context under a fresh claim/lease.

Disable / rollback: Do not execute capture, or disable checkpoint capture in the owning runner. Use --no-checkpoint-retention when intentionally retaining every bank point; do not reset an uncertain original capture.

Authority boundary: Historical leases, grants, budgets and feedback opportunities remain historical. Ordinary capture does not sweep its parent; the packaged checkpoint editor remains a staged gap.

Docs: Versioned guide.

loopx --format json research-checkpoint verify --root CHECKPOINT_ROOT --expected-sha256 CAPTURE_DIGEST

Native Codex accounts

Activation: Install the tagged source package into the serving LoopX environment, restart the service and open Settings → Model provider → Native Codex accounts. Automatic handoff and automatic client restart require separate explicit switches.

Validation: Read status/allowance and the current selection; check the inspected client/slot owner’s availability and recovery.

Disable / rollback: Turn off automatic handoff and restart permission in the same editor. Recover/rollback a committed selection using its exact revision and selection identity.

Authority boundary: Requires the installed Personal Apps native slot owner and supported macOS route. Installation switches no account; this is not hot credential replacement, turn replay or a new vault.

Docs: Versioned guide.

python3 -m pip install ./packages/loopx-codex-provider-routing
loopx-codex-accounts status

Owner notifications

Activation: Desktop: Settings → Appearance → Desktop notifications, off per browser by default. Private steward: Settings → Lark → Automatically notify new owner tasks, off until explicitly enabled on a verified owner-only binding.

Validation: Read the switch and browser permission. A newly waiting owner decision/task should produce one scoped notification; enabling first establishes a baseline rather than replaying old tasks.

Disable / rollback: Turn off the same switch. Browser permission revocation stops background reads; the steward retains its existing private outbox for original-result recovery.

Authority boundary: Notifications grant neither task execution nor decision authority. Desktop delivery needs the page open; private delivery rechecks the original owner/App/Chat scope.

Docs: Versioned guide.

loopx --version
loopx doctor

External steward context

Activation: An external channel needs a private source policy bound to verified sender IDs. The existing source policy targets active registered Agents across Goals, including future registrations. Set local_delivery_scope: selected and explicit targets to keep selected enrollment.

Validation: Inspect the source policy and the existing inbox readback under the connection’s exact registry/runtime. Receiver adoption and result return are separate observations.

Disable / rollback: Remove the sender/source grant to stop external delivery and future replay; exact Agent/Goal revocations use the existing delivery-target commands.

Authority boundary: A context grant is not evidence read permission, worker launch, a Todo/lease change or a protected-operation grant. Existing execution bindings require their own configured authority.

Docs: Versioned guide.

loopx manager-inbox read --goal-id GOAL --agent-id AGENT --format json

External research discovery

Activation: Opt in per research question. Discover a currently ready provider, bind the decision in a plan and then use the enabled Ego reader or installed Ego browser for actual X results and original-post/media inspection.

Validation: Read back the exact plan, provider receipt and admitted source links; distinguish loading, empty results and login refusal.

Disable / rollback: Do not invoke the selected search; stop the original browser operation or disable its reader through its owner. There is no automatic background X search.

Authority boundary: Keep TaskSpace/account control and source privacy; snippets/search URLs do not prove an actual platform read or publication authority.

Docs: Versioned guide.

loopx external-evidence discover --connector-registry --format json

Finance confirmation and retry guards

Activation: From the tagged checkout, explicitly install/register the optional simulator. Prepare/deliver an immutable operation card through Goal Channel; the installed executor still needs its separately authenticated confirmation and consumed claim.

Validation: Read the human-confirmed executor capability. Summary/risk/reconciliation commands inspect caller-retained facts; transient retry policy stays bound to the original attempt and budget.

Disable / rollback: Disable the extension to stop new simulated execution; reconcile an ambiguous original attempt instead of resubmitting.

Authority boundary: The bundled M1 provider is a simulator. It contains no venue client, signer or real-order permission; installing it grants none of those.

Docs: Versioned guide.

python3 -m pip install ./packages/loopx-finance-execution
loopx extension install --manifest packages/loopx-finance-execution/extension.toml --execute
loopx capability show human-confirmed-operation-executor --format json
loopx extension disable loopx-finance-execution --execute

Private owner host tools

Activation: A private owner Codex manager defaults to trusted_owner with gpt-6.1-sol/xhigh. Existing explicit restricted choices and machine/service model overrides retain precedence. Inspect the effective machine policy and current Session in the capability workbench.

Validation: Run machine-config inspect and read the current Session profile/sandbox. A changed manager policy rotates its upstream thread while restoring visible conversation in native history.

Disable / rollback: Apply an explicit restricted profile through the existing preview/apply transaction below. Removing the override restores the trusted_owner installation default.

Authority boundary: Only local private owner and freshly verified native private-steward audiences qualify. Other external audiences remain restricted. Current requests, protected operations, provider permissions and durable-state owners retain their boundaries; non-Codex endpoints refuse this profile.

Docs: Versioned guide.

cat > manager-runtime-restricted.json <<'JSON'
{"schema_version":"manager_runtime_profile_v0","runtime_profile":"restricted"}
JSON
loopx machine-config inspect --format json
loopx machine-config preview --namespace manager_runtime --config-json manager-runtime-restricted.json
loopx machine-config apply --namespace manager_runtime --config-json manager-runtime-restricted.json --expected-plan-revision PLAN_REVISION --execute

Install / Update

Use one installation owner. New package installs can pin the release; existing installations inspect their plan before applying. Native App and CLI versions need separate readbacks.

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

macOS/Windows desktop downloads are preview shells. Inspect the actual platform assets and desktop guide; macOS’s supported update route and Windows’s manual route retain their current boundaries.

Validation

The final source is clean commit 774d0c5f7, tree cd47a8b3a7d8af9707092e1a2441cc5cd3348493, version/tag 2.0.0 / v2.0.0. All eight exact-source qualification lanes are ready.

• Final-source checks: 1,430 tests across 36 complete affected modules passed; Ruff, configured mypy and TypeScript checks passed. All 514 public smokes passed with no failures, timeouts or skips. Risk validation passed its 6 direct and 19 selected checks; the raw benchmark manual hold is retained alongside its separate contract audit.
• Real paths: isolated PostgreSQL store/service/archive suites passed 366/11/4 cases. A real PyPI 1.4.2 → candidate-wheel upgrade preserved the existing Goal, Todo and registry; scoped settings preview/apply/remove, managed host-skill lifecycle and the packaged research-history and Goal-hierarchy interfaces passed. History used its real read-only HTTP owner; analysis and hierarchy presentation fixtures were explicitly synthetic. Both frontend asset generations and all bundled manifest hashes were verified. The changed-path public-boundary scan covered 1,649 files without violations.
• Full-suite history: the preceding immutable b93f65be0 sweep recorded 25,079 passed, 3 failed and 650 skipped, plus 2,242 passing subtests. The observed-session fixture was repaired in #6920. The other two complete modules passed unchanged at the final source with their original startup deadlines. This is scoped final-source qualification; it does not relabel the earlier full sweep as passing.
• Actual-default model behavior: one complete DeepSeek Flash/high run (deepseek-flash) passed the unchanged 22 scenarios × 2 repeats and 6 contrasts, using 87 HTTP calls over the documented streaming interface, with zero failed or skipped scenarios. Separate transport diagnostics, with two repeats per selected case, are excluded from release qualification. Earlier Flash/high failures and aborted attempts, plus the preceding Pro/high failure at source 5760c67ef after 8 HTTP calls (conservative cost upper USD 0.22926816), and the interrupted non-streaming run at the final source after 33 HTTP calls (14 scenarios not run; conservative cost upper USD 1.93756596), remain recorded. A later 6-HTTP streaming attempt also failed (one Goal-field mismatch and incomplete transport; 20 scenarios not run, conservative cost upper USD 1.07107572). The complete direct-stream Pro/high comparison recorded 89 HTTP calls, 2 failures and 0 unrun scenarios (descriptor presence and surfacing a non-blocking notice remain unqualified); it was not combined with this profile. Partial results were not combined. The authenticated native Codex live regression was explicitly skipped because an isolated API profile was unavailable; local native-CLI tests used a local mock Responses server.
• Limits: no matched stable-versus-candidate benchmark or long-horizon outcome uplift is claimed, so that expensive outcome gate was not required. Fresh-guest and Linux writer-barrier qualification remain open. Hosted package checksums, provenance and platform preview assets are produced by this release’s workflows and must be inspected before installation.

中文摘要

升级指南： LoopX 2.0 · 图文使用指南： Personal illustrated guide

LoopX 2.0 把两个月的 TypeScript 控制面工程，与长程研究工作台连接起来：从 Goal、委派与任务归属，追到实验历史、产物、轨迹分析和下一步依据。

升级决策

谁需要升级： 跨会话持续研究、查看实验历史，或基于迁移后的 TS owner 开发的用户。旧 Markdown Goal 的 writer 要先准备迁移路径。

解决了什么： 将已迁移的控制面规则收敛到共享 TS 事务 owner，并打通 Goal / 任务、研究历史、带引用的轨迹分析和后续决策入口。

是否有破坏性变更： 有。 未迁移 Markdown Goal 不再接受普通写入，legacy policy 不再准入新执行。canonical File/SQLite 保留支持的 soft_claim/hard_lease 选择；安装不会替代迁移。私人 owner 的 Codex 管家默认提供本机工具与 gpt-6.1-sol/xhigh，保留显式 restricted。checkpoint bank 的保留策略见下文。

如何验证： 版本为 loopx 2.0.0，打包 TS runtime 就绪，Goal 使用自己预期的存储与执行策略；从已记录历史选择轨迹，读取来源与引用后再显式运行分析。

贡献者： 发布维护者 @loopx-agent，社区贡献者与具体工作见下方八位贡献者。

loopx --version
loopx doctor --deep
loopx configure-goal --goal-id GOAL --format json

状态内核与控制面

共享 TS owner 贯穿 Inbox/Turn 观测、任务投影、调度仲裁、语义重规划与恢复；共享宿主接收有界能力上下文，并保留完整执行边界条款；普通旧 writer 退役，历史读取与原操作恢复分别保留。Python 继续承担真实 Host/provider IO 与专业计算。Semantics 与 CLI output 工程保护 owner、值域、生产者、投影和决策细节。#6396, #6860, #6634, #6613, #6725, #6919。

能力与工作流

研究工作台连接 Goal、时间线、实验、产物与轨迹来源，安装后的子 Goal 保留原来源与共享 namespace 规则；深度分析展示引用、覆盖、过程维度、原生结果与 solver 成本。记忆按用途与 Agent 组织。媒体路线交付本地排版、字幕、片段、可编辑 Blender 工程及渲染/读回层，保留兼容工作台；普通任务内制作与 Workspace 通用验收仍待集成。共享 Goal 规划上下文、显式 checkpoint 和本人任务提醒各有原有入口。#6648, #6674, #6700, #6878, #6711, #6589, #6857, #6654, #6785, #6664, #6533, #6910, #6925。

私人 owner 的 Codex 管家可直接完成已有授权的本机工作，执行器与证据 scope 沿 coordination root 解析；策略轮换沿原生历史恢复可见对话。#6859, #6793, #6893。

质量与测试

降低历史观测开销，收敛摘要 owner，保留产物探测失败，并从不可解码的投影缓存恢复真实读取。发布 job 为完整验证留出总时长，保留场景、来源与资格化检查；pnpm 和 TS 分片保留锁定依赖及平台前提。#6840, #6845, #6863, #6790, #6881, #6833, #6861, #6852。

基准与集成

checkpoint bank 在已启用捕获的 controller 中默认轮转完整点；完整 controller 捕获共用 90 秒准备窗口，仍受原 trial 截止时间约束。原始来源和反馈边界由原 owner 管理。原生 Codex 账号接力与客户端重启分别授权，CLI/App 升级与会话恢复保留 Goal lifetime。App 更新会打开选定安装副本，已有其他副本运行时也保留此选择；已证明的原生路由超时沿原 live Turn 恢复，Lark discovery backoff 保留恢复公平性。Ego/X 发现、财务确认与重试是各自可选路径。#6785, #6618, #6624, #6878, #6783, #6851, #6839, #6789, #6747, #6832, #6731, #6752, #6917, #6916, #6914。

文档与兼容性

2.0 双语升级指南、重构工程长文与影片及双语稿连接同一条主线。旧 Goal 先停 writer、确认退出、结清租约/outbox，再审阅完整计划；响应不确定时恢复原操作。外部管家既有 sender 配置覆盖活跃已登记及未来 Agent；保留名单边界时显式用 selected scope。#6908, #6722, #6668, #6396；管家合同。

社区贡献者

• @songoow — 新 maintainer Song 主导重构中的 Semantics owner 与词表收敛，并推进 runtime/profile 与发布资格化. #6733, #6833, #6881.
• @Duang777 — 普通旧 writer 退役与持续会话的 Goal 归属. #6396, #6747.
• @hhyykk — Inbox/Turn context、调度仲裁与语义重规划的 TS owner. #6860, #6634, #6613.
• @JasonBuildAI — 历史观测开销、摘要 owner、损坏投影恢复、工作台来源归属与 Lark 通知资格检查. #6840, #6845, #6802, #6891, #6890, #6893.
• @BigDataDZ — 可移植 TS 分片、锁定 pnpm 依赖与 Windows/source 前置条件. #6861, #6852, #6851.
• @steven-kid — 明确产物探测失败，并从不可解码的投影缓存恢复. #6863, #6790.
• @LIHUA919 — 有范围的 monitor 门禁与调度 deadline/backoff 处理. #6803, #6787.
• @hua03 — Pi 集成中暂停 Goal 的可见性与状态组件. #6109.

可选能力启用与使用

以下安装命令的源码目录绑定本次 v2.0.0 tag；私有配置与来源由 Agent/UI 按现有 owner 准备，保留各自安装渠道。

Research workbench and trajectory analysis

启用： 打开 Goal 的 Explore 历史与 深度分析。Agent 登记已有授权的来源包；每次分析单独执行。下方命令先预览，核对来源和模型使用后再显式加 --execute。

验证： 查看当前来源、引用、覆盖、原生结果与成本；运行中的分析沿原操作读取状态或取消。

停用 / 回退： 不执行预览即可保持只读；已运行的分析沿原操作取消，仍可读取已记录结果。

权限边界： 来源读取与模型调用分别遵循现有 owner、账号和权限边界；打开页面不会启动模型。

文档： 版本化指南。

loopx benchmark trajectory-assess --input-json evidence-pack.json --output-dir .local/trajectory-analysis --format json

Agent memory and evidence

启用： 在设置中按用途与 Agent 找到记忆配置。准备私有 provider 配置，预览后只应用所需 Agent 的绑定。

验证： 读取 experiment-status，核对当前绑定摘要与 provider canary 写入、原文读回。

停用 / 回退： 用下方带 Agent scope 的清除命令停用该绑定，保留其他 Agent 的独立连接。

权限边界： 记忆是有范围的上下文，不授予任务或执行权限；provider 配置与内容保持私有。

文档： 版本化指南。

loopx configure-goal --goal-id GOAL --reward-memory-agent-id AGENT --reward-memory-config .loopx/config/reward-memory/agent.json
loopx reward-memory experiment-status --goal-id GOAL --agent-id AGENT --format json
loopx configure-goal --goal-id GOAL --reward-memory-agent-id AGENT --clear-reward-memory-config --execute

Local video and editable Blender

启用： 从本 tag 源码显式安装并登记 loopx-media；真实 3D 需要 Python 3.11 的 Blender extra。由 Agent 准备 composition。本 tag 的兼容工作台仍在设置 → 能力 → 内容制作；普通任务内制作与 Workspace 通用产物验收的集成仍待交付。

验证： Doctor 核对已登记的可执行环境，读取影片、字幕与来源摘要；Blender 另读回可编辑工程及 RGBA/depth passes。

停用 / 回退： 停用 loopx-media 后停止新渲染；取消和恢复继续使用原操作身份。

权限边界： 本地渲染写私有文件、消耗本机算力；账号操作、云生成与发布另有边界。Blender 当前资格范围为 macOS/Cycles CPU。

文档： 版本化指南。

uv tool install --python 3.11 './packages/loopx-media[blender]'
loopx extension install --manifest packages/loopx-media/extension.toml --execute
loopx extension doctor loopx-media --execute
loopx extension disable loopx-media --execute

Shared Goal planning context

启用： 在 Goal review/stride 编辑器启用「Show Goal in planning horizon」，或使用下方显式开关。

验证： 读取配置与 quota 规划投影，检查适用场景中的 planning_horizon.core_goal。

停用 / 回退： 显式关闭该 Goal 开关；清除 override 则恢复继承值。

权限边界： 这里只增加规划上下文，任务选择、quota、claim、lease 和执行准入仍由原 owner 决定。

文档： 版本化指南。

loopx configure-goal --goal-id GOAL --execution-planning-horizon-goal --execute
loopx quota should-run --goal-id GOAL --agent-id AGENT --format json
loopx configure-goal --goal-id GOAL --no-execution-planning-horizon-goal --execute

Research checkpoints

启用： CLI/Host 阶段：先准备静止的 canonical 来源、已结算 Turn、合格原生会话与显式产物清单。capture/admit/adopt 默认预览，各步显式执行。新启用 checkpoint 的 controller bank 在捕获完成后默认轮转。完整 controller 捕获共用 90 秒窗口（原为 30 秒），仍受原 phase 剩余截止时间约束。

验证： 核验原 root 与 digest，读取完整清单，并在新 claim/lease 下检查接收端采纳的上下文。

停用 / 回退： 不执行捕获，或在原 runner 停用捕获；需要保留所有点时用 --no-checkpoint-retention。响应不确定时仍恢复原捕获。

权限边界： 历史 lease、grant、预算和反馈机会保留为历史。普通 capture 不扫描清理父目录；打包 checkpoint 编辑器仍属后续产品阶段。

文档： 版本化指南。

loopx --format json research-checkpoint verify --root CHECKPOINT_ROOT --expected-sha256 CAPTURE_DIGEST

Native Codex accounts

启用： 将本 tag 的源码 package 装进提供服务的 LoopX 环境，重启服务，打开设置 → Model provider → Native Codex accounts。自动接力与自动退出/重启客户端是两个独立开关。

验证： 读回 status、allowance、当前 selection，以及已识别客户端和 slot owner 的可用性与恢复状态。

停用 / 回退： 在同一编辑器关闭自动接力和重启许可；已提交 selection 沿精确 revision/selection 恢复或回退。

权限边界： 依赖已安装的 Personal Apps 原生 slot owner 与支持的 macOS 路径；安装不会切账号，也不热替换登录或重放 Turn。

文档： 版本化指南。

python3 -m pip install ./packages/loopx-codex-provider-routing
loopx-codex-accounts status

Owner notifications

启用： 桌面：设置 → 外观 → 桌面通知，按浏览器默认关闭。私聊管家：设置 → Lark → Automatically notify new owner tasks，在已验证的 owner-only 绑定中显式启用。

验证： 读回开关和浏览器权限；新出现的待本人决定/处理事项应产生一次有范围的提醒。首次开启先建立基线，不补发历史任务。

停用 / 回退： 关闭原开关；浏览器撤销通知权限会停止后台读取。管家保留原私有 outbox，用于已有结果的恢复。

权限边界： 通知不授予任务执行或决策权限。桌面提醒要求页面保持打开，私聊发送会重验原 owner/App/Chat。

文档： 版本化指南。

loopx --version
loopx doctor

External steward context

启用： 外部渠道需要绑定已验证 sender ID 的私有 source policy。既有 sender 配置覆盖跨 Goal 的活跃已登记 Agent，包括未来登记；保留选定名单时显式设置 local_delivery_scope: selected 与 targets。

验证： 在连接自己的 registry/runtime 下检查私有 policy 与原 inbox 读回；接收、采纳和结果回传分别观察。

停用 / 回退： 移除 sender/source grant 即停用外部交付与未来重放；精确 Agent/Goal 撤销沿现有 delivery-target 命令处理。

权限边界： 上下文交付许可不等于证据读取、worker 启动、任务/租约变更或受保护操作许可；execution binding 另有既有配置准入。

文档： 版本化指南。

loopx manager-inbox read --goal-id GOAL --agent-id AGENT --format json

External research discovery

启用： 按研究问题选择来源：先发现当前 ready 的 provider、在 plan 绑定决策，再使用已有 Ego reader 或已安装 Ego 浏览器读取真实 X 结果、原帖与媒体。

验证： 读取精确 plan、provider receipt 与采纳的原始来源，区分加载、空结果和登录拒绝。

停用 / 回退： 不调用搜索即可停用该次发现；原浏览器操作沿其 owner 停止，reader 按自身生命周期停用，没有自动后台搜索。

权限边界： 保留 TaskSpace、账号控制权和来源隐私；搜索摘要或 URL 不是平台实读或发布授权。

文档： 版本化指南。

loopx external-evidence discover --connector-registry --format json

Finance confirmation and retry guards

启用： 从本 tag 源码显式安装、登记可选模拟器；经 Goal Channel 准备并交付不可变操作卡，再走原有认证确认与 claim 消费。

验证： 读回 human-confirmed executor 能力；summary/risk/reconcile 检查调用方保留的事实，短暂拒绝重试仍绑定原 attempt 与预算。

停用 / 回退： 停用扩展即停止新模拟执行；状态不确定时核对原 attempt，而不是重新提交。

权限边界： 随包 M1 provider 是模拟器，没有 venue client、签名器或真实下单许可；安装不会创建这些能力。

文档： 版本化指南。

python3 -m pip install ./packages/loopx-finance-execution
loopx extension install --manifest packages/loopx-finance-execution/extension.toml --execute
loopx capability show human-confirmed-operation-executor --format json
loopx extension disable loopx-finance-execution --execute

Private owner host tools

启用： 私人 owner 的 Codex 管家默认采用 trusted_owner 与 gpt-6.1-sol/xhigh；现有显式 restricted 和机器／服务模型覆盖继续优先。在能力工作台核对机器策略与当前 Session。

验证： 运行 machine-config inspect，读回当前 Session 的 profile/sandbox。管家策略变化会轮换上游 thread，并将可见对话恢复到原生历史。

停用 / 回退： 用下方 preview/apply 原事务显式应用 restricted；删除 override 会恢复 trusted_owner 安装默认值。

权限边界： 仅限本地私人 owner 与新鲜核验的原生私聊管家受众，其他外部受众保持 restricted。当前请求、受保护操作、provider 权限与 durable state 各由原 owner 管理；非 Codex endpoint 会明确拒绝该 profile。

文档： 版本化指南。

cat > manager-runtime-restricted.json <<'JSON'
{"schema_version":"manager_runtime_profile_v0","runtime_profile":"restricted"}
JSON
loopx machine-config inspect --format json
loopx machine-config preview --namespace manager_runtime --config-json manager-runtime-restricted.json
loopx machine-config apply --namespace manager_runtime --config-json manager-runtime-restricted.json --expected-plan-revision PLAN_REVISION --execute

安装、更新与验证

使用同一安装 owner。Python 3.11+、Node 22.22.3+，推荐 Node 24 LTS。首次安装 skills 后重启宿主。CLI 与 App 分别核对身份。

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

最终源码固定为干净的 774d0c5f7，tree cd47a8b3a7d8af9707092e1a2441cc5cd3348493，版本/tag 为 2.0.0 / v2.0.0；8 项精确源码资格均已就绪。

• 最终源码的 36 个完整受影响模块共 1,430 项测试通过，Ruff、既有 mypy 范围与 TypeScript 检查通过；514 项公共 smoke 全部通过，无失败、超时或跳过。风险验证的 6 项直接检查与 19 项选定检查通过；原始 benchmark 人工 hold 与独立合同审计分别保留。
• 隔离真实 PostgreSQL 的 store/service/archive 检查通过。真实 PyPI 1.4.2 → 候选 wheel 升级保留原 Goal、Todo、registry；设置预览/应用/清除、宿主 skills 生命周期、打包研究历史与 Goal 层级界面通过；历史使用真实只读 HTTP owner，分析与层级为明确标注的合成界面 fixture。前端两代资源及 manifest 摘要已核验；公开边界扫描 1,649 个变更文件，无违规。
• 前一个不可变源码 b93f65be0 的全量结果为 25,079 通过、3 失败、650 跳过，另有 2,242 个子测试通过。#6920 修复已观测会话 fixture；另两个完整模块在最终源码以原有启动时限原样通过。上述记录不改写为「最终全量全绿」。
• 修复后的单次完整 DeepSeek Flash/high（deepseek-flash）验证通过原有 22 场景 × 2 次和 6 组对照，经官方流式接口共 87 次 HTTP 调用，失败和跳过均为 0。另行进行的传输诊断按每个选定场景双次运行，不计入发布资格。此前 Flash/high 失败与中止，以及旧候选 5760c67ef 的 Pro/high 8 次 HTTP 调用后失败（保守费用上界 0.22926816 美元），以及最终源码上因非流式响应中断而失败的 33 次 HTTP 调用（14 个场景未运行，保守费用上界 1.93756596 美元）全部保留。随后 6 次 HTTP 的流式整组也失败（一次 Goal 字段不一致、传输不完整；20 个场景未运行，保守费用上界 1.07107572 美元）。完整的直接连接 Pro/high 对照记录 89 次 HTTP、2 项失败、0 个未运行场景（描述符存在性、非阻塞通知呈现仍未合格），未与本组拼接。 没有拼接部分通过结果。独立 API profile 不可用，真实登录的原生 Codex 回归明确跳过；本地原生 CLI 测试使用本地 mock Responses 服务。
• 本次不宣称匹配条件下的 benchmark 或长程成果增益，因此未要求昂贵的成果对照。fresh guest 与 Linux writer barrier 仍待资格化。安装前核对本次发布工作流产出的校验和、来源证明与各平台预览资产。

完整 tag 差异