Skip to content

v1.3.1 — Workspace reliability and safer dependency upgrades

Latest
Latest

Choose a tag to compare

View all tags
@loopx-agent loopx-agent released this 06 Oct 08:10
· 702 commits to main since this release
v1.3.1
7445ae3
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

LoopX v1.3.1

At a Glance

Illustrated release guide: Personal 1.3.1 upgrade and usage guide — verified personal ownership, current-tag screenshots, text/image readback and native PDF inspection; the published package, updater and stable installation are verified. The 1.3.0 guide is its previous-version baseline.

A reliability patch for ordinary Chat, desktop updates and trusted result readback, with a root fix for the dependency API upgrade.

Publication status: All 12 GitHub assets, the PyPI package and signed desktop-stable feed are verified. The CLI stable branch points to the frozen release commit. Actual PyPI and named GitHub wheel upgrades from 1.3.0, plus a fresh public stable installation and version/doctor readback, passed.

Upgrade

loopx update check
loopx update apply
loopx --version
loopx doctor

Highlights

  • Keep read-only Chat within its selected workspace/executor scope; start with loopx chat --project-workspace-grant workspace_read --no-open. #5596.
  • Read published research from Goal Results and handle original owner requests from Needs you. #5575, #5571.
  • Use explicit manual desktop update while App-owned service allocation stays isolated. #5660, #5709.
  • Safely read web links with the adapted linkify-it6.1 dependency and zero dashboard audit findings. #5755.
  • Inspect Portfolio scores with bound settings/cohorts; no automatic job or publishing effect. #5750.

Replan Fixes

  • Receipt-backed progress settles the original host Turn before frontier reentry, while open/waiting work remains open/waiting. #5727.

Contributors

Release Decision

**Who should upgrade:**1.3.0 users affected by Chat scope/replay, desktop update races, result readback or these dependency advisories; users without these paths may wait for their usual update window.

**What this release solves:**A distinct patch identity for merged reliability fixes, faithful original-source results, and the real linkify-it6.x API/type compatibility failure.

Breaking changes: No. No new release-bump state migration; package rollback does not restore stored data. Captured heartbeat transport remains explicit/default-off. Link recognition preserves existing explicit http(s)/inert-HTML policy.

**How to verify:**Expect loopx 1.3.1, an installed-owner-consistent doctor result, then read an existing Goal result or original request in the App.

loopx --version
loopx doctor
loopx chat --project-workspace-grant workspace_read --no-open

**Contributors:**Release maintained by the project maintainer; external contributors @Duang777 and @mikamikasuki, project automation @loopx-agent and dependency automation @dependabot[bot] are credited below from the actual tag range.

State Kernel & Control Plane

  • Accepted semantic progress settles its original host Turn without falsely completing an open/waiting Todo; stopped Goals and conflicting message/wake replays retain their fences. #5727, #5742, #5733, #5730.
  • Canonical Todo callers avoid retired source-writer crossings, peer work facts retain their canonical source, and torn post-writeback journals preserve retries. #5739, #5724, #5710, #5703, #5668, #5705.

Capabilities & Workflows

  • Existing Goal results now expose already published research and faithful review cards; Needs you opens an actionable original user request. App presentation completes this readback slice; it does not automatically create/publish research. #5575, #5571, #5609.
  • Chat respects the selected read-only project executor; registered workspace routing and task/compare recovery preserve original context. #5596, #5732, #5692, #5748.
  • Native child result/parent-review recording accepts compatible original terminal receipts while rejecting conflict or unregistered reporting. #5746.

Quality & Testing

  • Root repair for linkify-it6.1 uses its named class/bundled types and explicit URL authentication policy. Safe complete-address recognition and inert HTML remain tested; dashboard dependency audit is clean. #5755.
  • Browser fixtures retain the authoritative task read and wait for editor hydration; release source records distinguish product and workflow identities. #5737, #5738, #5722, #5740.

Benchmarks & Integrations

  • Captured TurnEnvelope context is an explicit default-off heartbeat experiment; consumers read the same invocation instead of rerunning a guard to reconstruct its receipt. #5665, #5736.
  • Desktop windows isolate their own service allocation; manual update wins its race with automatic checking, and Windows locator diagnostics stay observable. Native SForge planned entry runs inside the timed environment boundary; its real external environment is not qualified by this expedited release. #5660, #5709, #5734, #5701.
  • The Portfolio ten-arm retrospective publishes observed scores bound to runner settings and evaluator cohorts, plus a local export/verification helper. It is not a paired1.3.0→1.3.1 uplift claim. #5750.

Documentation & Compatibility

  • CLI command reference stays available independently of runtime state; late install feed observations and registered workspace recovery have bounded/readable diagnostics. #5706, #5720, #5716.
  • Current version/help/book anchors align on1.3.1; read-only boundaries and local capture instructions are explicit. No new persisted-state migration is introduced by the release bump. #5755, #5736.

Community Contributors

Compatibility & Validation

Expedited qualification. Source 7445ae360aaf330e0067959a5e0084ffd9f76de4, tree 32d08be88fb2b02036735d9f9a4a8a3fc551429f; this is the merged main snapshot with the exact reviewed5755 tree. Version/help/release-doc/book smokes, production frontend compile/build, zero dashboard audit, three safe-rendering regressions,81 focused Python tests plus27 offline adapter passes/62 environment skips, strict quality and risk canary passed. Earlier book-anchor and dependency-only compile failures remain recorded with their repairs. Full Python/public smoke suites, a new paid model matrix, separate Pages, live Lark/Bot, Windows native execution, Apple Developer ID/notarization and unavailable Harbor/SForge/PostgreSQL/NoKV were not newly qualified.1.3.0 results are prior-version context, not a same-source1.3.1 certificate. Hosted package/browser/DSH installation gates passed. All 12 public GitHub assets, four package attestations, four desktop installer attestations, checksums, 1,523 package source files and 3,492 desktop runtime source files match the frozen tag; the signed updater and desktop-stable feed were independently verified. The two queued desktop publisher jobs were cancelled and the maintainer uploaded unchanged original CI bytes and canonical feed/checksums; those overall workflows are recorded as cancelled, with successful producer jobs. Actual public GitHub wheel upgrade from 1.3.0, installed frontend and deep doctor passed. GitHub/PyPI package hashes match; an actual PyPI upgrade from 1.3.0 and a fresh public stable archive installation passed version, frontend/doctor and source-manifest readback. The CLI stable branch was fast-forwarded to the frozen release commit.

Install / Update

Existing installs follow their own owner; check first, then apply. PyPI users can explicitly install the published named package.

loopx update check
loopx update apply
loopx --version
loopx doctor
python -m pip install --upgrade loopx==1.3.1

The immutable named GitHub wheel is also available; use the same Python installation owner:

python -m pip install --upgrade https://github.com/loopx-project/loopx/releases/download/v1.3.1/loopx-1.3.1-py3-none-any.whl
loopx --version
loopx doctor

Rollback:use the previously verified named package/tag or desktop installer under the same install owner. Preserve data and configuration backups; code downgrade is not data restore. New installs follow the versioned installation guide.

Optional Capability Activation & Use

Set example Goal/Agent/Turn variables to your already authorized identities; do not create fake receipts or infer authority from an installed command. Source-only benchmark helper commands run from the corresponding tag checkout.

Explore Harness

Activation: In Goal settings → Capability Center choose evidence or planning. CLI planning opt-in is shown below; the default is off.

Validation: Read turn context and summary for the same Goal/Agent; recorded nodes alone do not prove adoption.

Disable / rollback: Set --explore-mode off --execute; retained evidence stays readable.

Authority boundary: Analysis and planning grant no worker spawning, claims, quota spend, merge or external publication.

Docs:Versioned documentation.

loopx configure-goal --goal-id "$GOAL_ID" --explore-mode planning --explore-harness-profile adaptive-resilient --execute
loopx explore turn-context --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"
loopx explore summary --goal-id "$GOAL_ID"
loopx configure-goal --goal-id "$GOAL_ID" --explore-mode off --execute

Ordinary workspace Chat

Activation: Run loopx chat; in the App steward conversation choose an already granted workspace in Scope. Host-declared roots now default to workspace_write; use the explicit read-only command below when required.

Validation: Read the selected scope, effective grant and returned result in the same conversation. A changed grant creates a new context; old history remains.

Disable / rollback: Restart the service with --project-workspace-grant workspace_read, or revoke the host workspace grant; new work on the old context is rejected.

Authority boundary: Workspace writes follow AGENTS.md and the actual sandbox. No hidden Goal, portfolio grant, peer delegation or external-send authority is created.

Docs:Versioned documentation.

loopx chat --project-workspace-grant workspace_read --no-open

Governed delegation stop

Activation: Use an already configured exact binding, registered requester and operation; delegation stop --execute explicitly requests stop. CLI/MCP and the App team surface share that owner.

Validation: delegation read distinguishes acknowledgement, proved native process drain and settlement. Missing supervision remains unknown and requires the returned recovery.

Disable / rollback: Remove the exact operator binding to revoke new execution; use stop on an active operation. A stopped/settled receipt is not a resume grant.

Authority boundary: Do not infer Host exit, lease release or non-execution from absent records. Windows/unsupported stopping refuses before launch-side cancellation effects.

Docs:Versioned documentation.

loopx delegation stop --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --execution-config "$DELEGATION_CONFIG" --operation-id "$OPERATION_ID" --execute
loopx delegation read --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --execution-config "$DELEGATION_CONFIG" --operation-id "$OPERATION_ID"

Captured heartbeat decisions

Activation: Opt in per thin generic-CLI heartbeat with --decision-output-root pointing to an existing private directory and an explicit Turn ID. For the shared benchmark heartbeat worker only, kwargs.turn_envelope:true is a separate default-off trial option.

Validation: Read the returned detail_ref.full_decision from the same invocation; verify Goal/Agent/Turn and original source hash. Worker install/execution/wake receipts must echo turn_envelope:true.

Disable / rollback: Omit the capture flag; set turn_envelope:false for the next trial. Do not modify the treatment of a running trial. Remove unneeded private captures by ordinary file management.

Authority boundary: Captures are private observations, not fresh admission, worker creation, scheduler authority, quota or merge permission. Native Goal/full/compact/brief combinations remain excluded. No token-saving or decision-quality claim.

Docs:Versioned documentation.

mkdir -m 700 ./private-decisions
loopx --format json heartbeat-prompt --thin --runtime-profile generic_cli --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --turn-instance-id "$TURN_ID" --decision-output-root ./private-decisions

Native child result readback

Activation: For an already admitted Turn and registered coordinator, use native-child record --execute with the original operation identity and an explicit decision/result/review stage. Compatible terminal guard receipts can now accept the corresponding report without a status-label whitelist.

Validation: Read the same Turn with the command below. Host completion, parent review and adopted evidence remain separate.

Disable / rollback: Stop issuing records for the operation; where a started operation is cancelled, explicitly record the real cancelled result and rejected parent review. Use the previous package/tag for code rollback; preserve historical records.

Authority boundary: Records do not spawn children, prove host provenance, adopt evidence, resume a closed Turn or grant peer/Goal/workspace authority. Conflicting replays still refuse.

Docs:Versioned documentation.

loopx --format json native-child read --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --turn-instance-id "$TURN_ID"

Published research and owner actions

Activation: Open an existing Goal in the App and choose Results to read already published research; open Needs you and its original request to perform the authorized owner action. There is no new automatic execution switch.

Validation: Read the published result/review status and source in the original Goal; owner-action readback must return to the same request. The command below checks the installed service before this UI journey.

Disable / rollback: Close the reader without changing the result; cancel an unapplied request through its existing owner. For code rollback reinstall the previous named release; do not erase stored evidence.

Authority boundary: Reading results does not publish new research, accept a result, modify private sources or grant external-send authority. Owner actions still require their exact original scope and confirmation; unavailable service remains explicit.

Docs:Versioned documentation.

loopx --version
loopx doctor

Desktop update and service ownership

Activation: Install the v1.3.1 desktop artifact for the supported OS, open Settings and explicitly Check for updates / install the offered signed update. Manual update has priority over a simultaneous automatic check.

Validation: Read the installed App version and service diagnostics in Settings. Inspect the public stable feed with the command below; a feed is not proof the App installed it.

Disable / rollback: Close the App-owned service via the App lifecycle; reinstall the previous verified desktop artifact for rollback. Do not stop an unrelated shared CLI service.

Authority boundary: Desktop windows own their private local services and avoid shared CLI ports; opening the App grants no extra Goal/workspace writes or external-send authority. Windows runtime execution and Apple Developer ID/notarization are not newly qualified.

Docs:Versioned documentation.

curl -fsSL https://github.com/loopx-project/loopx/releases/download/desktop-stable/desktop-updater.json

Portfolio report export

Activation: In a source checkout explicitly choose --runs-root, --selection and --output for terminal export. To inspect the shipped ten-arm retrospective, use --verify on its existing report directory.

Validation: Run the verification command below to check settings/data hash bindings; inspect evaluator cohorts and qualified/countable labels separately.

Disable / rollback: Omit the export invocation; remove only an unneeded generated local report. Revert the helper through the previous tag without rewriting existing experiment receipts.

Authority boundary: Export launches no benchmark job, reads no session, qualifies no integrity and uploads nothing. Missing provenance and evaluation-cohort differences remain visible; published scores do not prove paired-version uplift.

Docs:Versioned documentation.

python -m benchmark.edgebench.export_report --verify benchmark/edgebench/studies/portfolio-ten-arm-20261006

SForge native task entry

Activation: Explicitly choose --task-entry loopx-planned with EdgeBench heartbeat-resume or heartbeat-explore for a new trial; default remains seeded-todo. Keep other inputs fixed for a task-entry ablation.

Validation: The offline adapter checks below validate profile/receipt boundaries. In a real authorized SForge environment additionally read the persisted deadline and initial planning receipt/task/Todo lineage before execution; this release has no fresh external-environment qualification.

Disable / rollback: Select --task-entry seeded-todo for the next trial or omit the flag. Do not alter an active trial or rewrite its planning receipt.

Authority boundary: Planning runs inside the existing timed network-isolated worker/API proxy. Failed/missing/stale/blocked plans cannot execute; the option grants no additional model budget, network, scoring, claim, quota or publication authority.

Docs:Versioned documentation.

uv run --extra test python -m pytest -q benchmark/tests/test_sforge_runtime.py benchmark/tests/test_task_entry.py

中文摘要

升级决策

**发布状态：**12个GitHub资产、PyPI命名包和签名desktop-stable feed已核验，CLI stable分支已指向冻结的发布commit。PyPI与命名GitHub wheel的1.3.0→1.3.1实际升级，以及公开stable归档的全新安装、版本与doctor回读均通过。

**谁需要升级：**受 Chat 范围／重放、桌面更新竞态、研究结果读取或依赖安全问题影响的1.3.0用户；未使用这些路径可按原升级窗口更新。

**解决了什么：**将已合并可靠性修复发布为独立补丁版本，恢复原来源结果读取，完成 linkify-it6.x导出／类型变化的根因适配。

**是否有破坏性变更：**无。release bump 不增加持久状态迁移；降级包不恢复数据。heartbeat capture 显式开启且默认关闭，原安全链接策略保持。

**如何验证：**应显示 loopx 1.3.1，doctor 与安装 owner 一致，再在 App 读取既有 Goal 结果／原请求。

loopx update check
loopx update apply
loopx --version
loopx doctor

**贡献者：**项目维护者负责发行；外部贡献者 @Duang777、@mikamikasuki，项目自动化 @loopx-agent 和依赖自动化 @dependabot[bot]，按实际版本范围在下方具体署名。

状态内核与控制面

  • 已接受的语义进展结算原 host Turn，不把仍开放／等待的 Todo 虚报为完成；已停止 Goal 与消息／wake 冲突继续受原 fence 约束。#5727、#5742、#5733、#5730。
  • canonical Todo 入口退出旧 source-writer crossings，peer work 事实保留 canonical 来源，破损 writeback 尾部保留重试。#5739、#5724、#5710、#5703、#5668、#5705。

能力与工作流

  • Goal 结果可读已发布研究及忠实复核卡片，需要你处理的请求可回到原操作；此切片交付 App 阅读／操作入口，不自动创建或发布研究。#5575、#5571、#5609。
  • Chat 遵守所选只读 executor；注册工作区路由与 Task／对比恢复保留原上下文。#5596、#5732、#5692、#5748。
  • native child 原终态回执可接受兼容 result／parent review，冲突和未注册上报仍拒绝。#5746。

质量与测试

  • linkify-it6.1 根因修复使用命名 class／内置类型，显式保留 URL 用户信息策略；完整地址保护、危险协议与惰性 HTML 继续受原断言验证，dashboard 审计清零。#5755。
  • browser 夹具保留权威 Task 来源并等待编辑器初始化；发布来源记录区分产品与 workflow。#5737、#5738、#5722、#5740。

基准与集成

  • captured TurnEnvelope 是默认关闭的 heartbeat 试验，读取同次观察而不是为补回执重跑 guard。#5665、#5736。
  • 桌面窗口隔离自有服务，手动更新优先，Windows locator 诊断保持；SForge planned entry 在计时环境内执行，本次快速发布未资格化外部 SForge 环境。#5660、#5709、#5734、#5701。
  • Portfolio 十臂回顾公开绑定 settings／evaluator cohort 的已观察分数及本地导出核验 helper；不宣称1.3.0→1.3.1配对涨分。#5750。

文档与兼容性

  • CLI 命令参考不依赖运行状态；延迟安装 feed 与注册工作区恢复保留可读诊断。#5706、#5720、#5716。
  • 版本／帮助／当前手册统一1.3.1；发布 bump 未增加持久状态迁移，明确只读和私有 capture 边界。#5755、#5736。

社区贡献者

兼容性与验证

快速资格路径：源码7445ae360aaf330e0067959a5e0084ffd9f76de4与已审查5755的tree一致。版本／帮助／release文档／book、生产前端构建、dashboard零漏洞审计、三组安全渲染回归、81项定向Python及27项离线adapter通过／62环境跳过、strict quality和风险canary通过。初次book漏锚点与依赖-only编译失败保留及修复。未重跑全量Python／public smoke、新付费模型矩阵、独立Pages和原未具备的真实Lark／Bot、Windows原生、Apple签名／公证、Harbor／SForge／PostgreSQL／NoKV环境；1.3.0结果不转写为1.3.1同源码资格。托管包／浏览器／DSH安装门槛已通过；12个公开GitHub资产、4份包attestation、4份桌面installer attestation、校验和、1523个分发源码文件和3492个桌面runtime源码文件符合冻结tag，签名updater与desktop-stable feed已独立核验。两个桌面publisher在排队时取消，由维护者上传未改动的原CI bytes及canonical feed／checksum；整体workflow如实保留cancelled，producer jobs成功。公开GitHub wheel的真实1.3.0→1.3.1升级、前端与deep doctor通过；GitHub／PyPI包哈希一致，PyPI的1.3.0→1.3.1真实升级与公开stable归档全新安装通过版本、前端／doctor和来源manifest回读；CLI stable仅fast-forward到冻结发布commit。

安装与更新

按当前安装owner先check再apply；PyPI用户可显式安装已公开的 loopx==1.3.1。也可在原Python环境使用已核验的GitHub命名wheel：

python -m pip install --upgrade https://github.com/loopx-project/loopx/releases/download/v1.3.1/loopx-1.3.1-py3-none-any.whl
loopx --version
loopx doctor

回退使用旧已核验命名包／tag／桌面制品，保留数据和配置备份；代码降级不等于数据恢复。详见版本化安装文档。

可选能力启用与使用

示例身份变量仅使用已授权的 Goal／Agent／Turn；安装命令不授予权限，不能伪造回执。benchmark helper从对应源码tag运行。

Explore Harness

启用： 在 Goal 设置 → 能力中心选择 evidence 或 planning；下方命令启用 planning，默认关闭。

验证： 读取同一 Goal/Agent 的 turn-context 与 summary；节点存在不等于已采用。

停用 / 回退： 执行 --explore-mode off --execute；保留既有证据。

权限边界： 分析与规划不授予 spawn、claim、扣额、合并或外发权限。

文档：版本化文档.

loopx configure-goal --goal-id "$GOAL_ID" --explore-mode planning --explore-harness-profile adaptive-resilient --execute
loopx explore turn-context --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"
loopx explore summary --goal-id "$GOAL_ID"
loopx configure-goal --goal-id "$GOAL_ID" --explore-mode off --execute

Ordinary workspace Chat

启用： 运行 loopx chat，在管家对话的 Scope 选择已授权 workspace。宿主声明的目录默认 workspace_write；需要只读时用下方命令。

验证： 核对 Scope、生效 grant 与同一对话的结果；grant 改变创建新上下文，旧历史保留。

停用 / 回退： 以 --project-workspace-grant workspace_read 重启服务，或撤销宿主 workspace grant；旧上下文的新工作被拒绝。

权限边界： 写入遵循 AGENTS.md 与实际 sandbox；不创建隐藏 Goal，不借 portfolio、peer 委派或外发权限。

文档：版本化文档.

loopx chat --project-workspace-grant workspace_read --no-open

Governed delegation stop

启用： 使用已配置的精确 binding、注册 requester 与 operation，delegation stop --execute 显式请求停止；CLI/MCP 与 App 团队表面复用 owner。

验证： delegation read 区分确认、已证明的原生进程 drain 与结算；监督缺失保持 unknown，执行返回的恢复路径。

停用 / 回退： 移除精确 operator binding 撤销新执行，对 active operation 请求 stop；停止／结算回执不是 resume grant。

权限边界： 缺少记录不证明 Host 退出、lease 释放或未执行；不支持的停止平台在 launch-side 取消效果之前拒绝。

文档：版本化文档.

loopx delegation stop --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --execution-config "$DELEGATION_CONFIG" --operation-id "$OPERATION_ID" --execute
loopx delegation read --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --execution-config "$DELEGATION_CONFIG" --operation-id "$OPERATION_ID"

Captured heartbeat decisions

启用： 每次 thin generic-cli heartbeat 显式传已有私有目录 --decision-output-root 和明确 Turn ID；benchmark heartbeat worker 另在 kwargs 设 turn_envelope:true，默认关闭。

验证： 读取同次调用返回的 detail_ref.full_decision，核对 Goal／Agent／Turn 和源 hash；worker install／execution／wake 回执应记录 true。

停用 / 回退： 省略 capture 参数；下一试验设 turn_envelope:false，不修改正在运行的试验 treatment；仅按普通文件管理删除不用的私有 capture。

权限边界： 私有观察不提供新准入、worker 创建、调度、quota 或合并权限。原生 Goal 与 full／compact／brief 仍拒绝；不声称降低 token 或改善决策。

文档：版本化文档.

mkdir -m 700 ./private-decisions
loopx --format json heartbeat-prompt --thin --runtime-profile generic_cli --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --turn-instance-id "$TURN_ID" --decision-output-root ./private-decisions

Native child result readback

启用： 仅对已准入 Turn 和注册 coordinator，以原 operation 身份显式执行 native-child record --execute 并选择 decision／result／review；兼容终态 guard 回执现可接受对应上报，不靠状态标签白名单。

验证： 按下方命令读取同一 Turn；宿主完成、父级复核与证据采用仍分别记录。

停用 / 回退： 停止为该 operation 新增记录；真实取消的已启动操作应明确记录 cancelled result 与 rejected review。代码回退用旧包／tag，保留历史记录。

权限边界： 记录不会 spawn 子执行、证明宿主来源、采用证据、恢复已关闭 Turn 或授予 peer／Goal／workspace 权限；冲突重放仍拒绝。

文档：版本化文档.

loopx --format json native-child read --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --turn-instance-id "$TURN_ID"

Published research and owner actions

启用： 在 App 打开既有 Goal 的结果入口，读取已发布研究；从需要你处理的原请求执行已授权 owner 操作。没有新增自动执行开关。

验证： 在原 Goal 核对结果／复核状态和来源，owner action 应回到同一原请求；下方命令先核对安装服务，再走 UI 旅程。

停用 / 回退： 关闭阅读器不改变结果；尚未应用的请求通过原 owner 取消。代码回退重装旧命名版本，不删除证据。

权限边界： 阅读不等于发布研究、接受结果、修改私有来源或授权外发；owner action 保留精确范围及确认，不可用状态仍明确。

文档：版本化文档.

loopx --version
loopx doctor

Desktop update and service ownership

启用： 安装匹配系统的1.3.1桌面制品，在设置显式检查更新／安装提供的签名更新；手动操作优先于同时发生的自动检查。

验证： 在设置读取实际 App 版本和服务诊断；下方可读取公开 stable feed，feed 存在不证明 App 已安装。

停用 / 回退： 按 App 生命周期关闭 App 自有服务；回退重装旧已核验桌面制品，不停止无关共享 CLI 服务。

权限边界： 各窗口隔离本地服务并避开共享 CLI 端口；打开 App 不增加 Goal／workspace 写入或外发权限。未新增 Windows 原生运行或 Apple Developer ID／公证资格。

文档：版本化文档.

curl -fsSL https://github.com/loopx-project/loopx/releases/download/desktop-stable/desktop-updater.json

Portfolio report export

启用： 在源码 checkout 显式选择 --runs-root、--selection 和 --output 导出终态报告；读取已随包提供的十臂回顾时，对既有报告目录执行 --verify。

验证： 下方命令核对 settings／data hash 绑定；另读 evaluator cohort 与 qualified／countable 标签。

停用 / 回退： 不调用 export 即停用；仅删除不用的本地生成报告。helper 可随旧 tag 回退，不改写原试验回执。

权限边界： 导出不启动 benchmark、不读 session、不资格化完整性、不上传网络；来源缺口与 evaluator 差异保留，分数不证明配对版本涨分。

文档：版本化文档.

python -m benchmark.edgebench.export_report --verify benchmark/edgebench/studies/portfolio-ten-arm-20261006

SForge native task entry

启用： 仅在新试验的 EdgeBench heartbeat-resume／heartbeat-explore 显式选择 --task-entry loopx-planned，默认仍 seeded-todo；消融比较固定其他输入。

验证： 下方离线 adapter 检查验证 profile／回执边界；真实已授权 SForge 环境还需读持久 deadline、初始 planning 回执和 task／Todo lineage。本次未重新资格化外部环境。

停用 / 回退： 下一试验选择 --task-entry seeded-todo 或省略参数；不修改活动试验或改写其 planning 回执。

权限边界： 规划在原计时、网络隔离 worker／API proxy 内运行；失败／缺失／过期／blocked plan 不能执行。不增加模型预算、网络、评分、claim、quota 或发布权限。

文档：Versioned documentation.

uv run --extra test python -m pytest -q benchmark/tests/test_sforge_runtime.py benchmark/tests/test_task_entry.py

图文指南

个人 1.3.1 图文升级指南已核对个人 ownership、当前 tag 截图、正文／图片及原生 PDF 排版；命名包、updater与stable安装均已核验。1.3.0指南是其上一版本基线。

Full comparison

Contributors

dependabot, Duang777, and 2 other contributors
Assets 14
Loading