Skip to content

LoopX 1.1.0 — Governed collaboration, new organization

Latest
Latest

Choose a tag to compare

View all tags
@huangruiteng huangruiteng released this 19 Sep 17:06
· 426 commits to main since this release
v1.1.0
607c11d
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

LoopX v1.1.0 — governed agent teams, durable handoffs, and a new project home

At a Glance

LoopX now lives at loopx-project/loopx, with its website, documentation, installer and update feeds under the project organization. The repository history, issues, pull requests and stars are preserved. The package is still named loopx.

This release also brings the work since 1.0.5 together: inspectable team plans, owner-configured acceptance, durable local delegation, canonical coordination, and a clearer Personal Workspace. It does not claim that a preview is execution, a completed task is an accepted Goal, or a registered Agent is authorized to launch work.

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

Existing managed installs can instead use loopx update check, loopx update plan, then loopx update apply; the update path preserves the pip, pipx or archive installation owner.

Release Decision

Who should upgrade: Users of Personal Workspace, steward team planning, local delegation, canonical Todo/Monitor coordination, and archive or desktop update paths.

What this release solves: Team proposals now have an explicit confirmation-to-canonical-work path; delegated work has durable operation identity, recovery and checked returns; acceptance and replan obligations retain their existing authority owners. Installation and public entry points now use the transferred repository.

Breaking changes: Yes, for callers of retired CLI entry points and old Pages URLs; no automatic state migration or provider promotion is introduced. Acceptance and execution bindings remain opt-in. The retired todo suggest and todo capture-followups commands are removed; scripts must use current documented Todo creation/completion and follow-up planning flows rather than those retired entry points (#4725, #4723). GitHub repository links redirect, but the old GitHub Pages hostname does not: update website, blog, documentation and installer bookmarks to https://loopx-project.github.io/loopx/. Older executables that print the old installer URL should use the pip command above or the new archive installer.

How to verify: Confirm loopx --version reports 1.1.0, run loopx doctor --deep, then inspect the directory and acceptance state for an existing registered Goal. Inspection does not enable optional contracts.

Contributors: Community work from @songoow, @Duang777, @DJC1412 and the other contributors listed below is credited with concrete changes and source PRs. Release qualification distinguishes the prior full-suite baseline from candidate-specific checks; see Validation.

loopx --format json agent-directory --goal-id "$GOAL_ID"
loopx --format json goal-acceptance inspect --goal-id "$GOAL_ID"

Agent Teams & Personal Workspace

  • Inspect, confirm, then materialize. Steward team plans expose lanes, staffing gaps and bounded commitments. Owner confirmation carries the reviewed plan into canonical Todos and assignments; an unconfirmed preview cannot write or launch work (#4519, #4535, #4547, #4633).
  • A conversation is not a scheduler. Goal Chat exposes inline LoopX mode and coordinated team inspection; normal conversation persistence and autonomous execution remain separate. Workspace first screens focus on the current Goal and inspectable work (#4700, #4736).
  • Delegate from an existing Agent. Governed CLI/MCP delegation uses exact operator bindings, stable operation IDs, detached execution and bounded read/wait/resume. Recovering operations after context loss does not dispatch replacements; accepted returns revalidate the bound Todo and artifacts (#4688, #4706, #4731).
  • Make available routes legible. Goal-owned configuration projects requester-scoped execution readiness without granting new authority or adding background operation polling (#4741).
  • Align explicit host capacity. An opt-in configuration action can raise the active Codex host's child-thread limit to the Goal maximum, preserving higher values and requiring a new Session after a write (#4733).

State, Acceptance & Recovery

  • Canonical lease acquisition, no-change retries, atomic handover, source-witness-bound writes and reviewed Todo/Monitor edits share the authoritative path (#4666, #4669, #4670, #4686, #4689, #4693, #4699).
  • Owner-reviewed Goal acceptance contracts bind existing tasks to bounded artifact checks. Configure, verify, inspect and disable are explicit; disabled or absent contracts retain baseline behavior (#4683).
  • Required-vision replans share typed outcome policy between projection and write-time admission. Agents can inspect, author, correct, write back and settle the original obligation. The associated PR reports two successful live-model trials; that bounded result is not a new release-wide live qualification (#4730).
  • Checkpoint acknowledgements survive history trimming; periodic review uses lane-complete history and deferred selection recovery remains executable (#4554, #4654, #4655).
  • Work counts expose incomplete coverage rather than inventing hidden advancement work. Monitor no-change polls avoid spend, periodic reports cover Goal activity without cross-lane duplicates, and repair dispatch uses typed action classification (#4713, #4739, #4729, #4734, #4740).

Performance, Integrations & Public Evidence

  • Status and runtime hot paths reuse run-history, event, source-fingerprint, registry and readiness reads (#4632, #4685, #4690, #4694, #4695).
  • Runtime and extension diagnostics identify the selected interpreter and manifest source. DeepSeek Harness uses its typed 0.1.5 host/client boundary; NoKV seed-routing and SDK mismatch handling improve without promoting it to default authority (#4482, #4509, #4649, #4724).
  • The public site adds application scenarios, bilingual Agent-facing Kanban articles, and the LHTB research brief with explicit baselines, accounting corrections and counterexamples. These are bounded studies, not evidence that this release universally outperforms other harnesses (#4660, #4673, #4703, #4709, #4714).

Community Contributors

Thank you to the external contributors whose merged work is included in this release:

  • @BigDataDZ: typed unstaffable-lane documentation (#4641).
  • @Chumaniac, first merged contributions in this range: installation versus update authority (#4659, #4657).
  • @DJC1412: Goal-wide periodic reporting, cross-lane deduplication and shared-authority RFC ledger (#4729, #4734, #4728).
  • @Duang777: bounded Chat event retention, migration collision safety and repeated-read reduction (#4463, #4584, #4695).
  • @Jim-jimu: non-default Git transport ports and contributor-scoped DCO checks (#4634, #4697).
  • @YZJF: refresh/promotion modularization, serialized history repair, typed decisions and blocked-priority notices (#4438, #4513, #4640, #4692).
  • @gwh6669999, first merged contribution in this range: archived DeepSWE v1 placement and verification (#4502).
  • @shangzh0, first merged contributions in this range: LHTB runner, five-arm report, Pages brief and token accounting (#4504, #4698, #4703, #4714).
  • @songoow: acceptance contracts, typed Turn/lifecycle semantics, runtime boundaries, semantic inventories and repair classification (#4360, #4499, #4683, #4740).
  • @superwesleyhys-ux, first merged contributions in this range: recorded-rejection fixture and authority documentation (#4549, #4550).
  • @wchwawa: NoKV seed-routing and SDK compatibility handling (#4724).
  • @xiaods: worker-lifecycle state projection (#4678).
  • @yilin-succeed: autonomous replan settlement and visible blocked-priority work (#4536, #4543).

Optional Capability Activation & Use

Use existing registered Goal and Agent IDs in the following commands. No example authorizes production actions or credential sharing.

Steward Team Plan Confirmation

Activation: Run loopx dashboard --no-open, open LoopX Manager, request a team plan, inspect every lane, and explicitly confirm the preview card.

Validation: loopx --format json todo list --goal-id "$GOAL_ID"; check that the returned plan and assignments reference the same Goal and lanes.

Disable / rollback: Reject an unconfirmed preview. After confirmation, loopx goal-lifecycle --goal-id "$GOAL_ID" --operation stop --execute stops new scheduling; reconcile created Todos separately. Downgrading does not erase existing work.

Authority boundary: A preview grants no write, launch, quota, lease or publication authority. Confirmation materializes only the inspected plan through existing owners; unstaffable lanes remain gaps.

Docs: https://github.com/loopx-project/loopx/blob/v1.1.0/docs/product/use-cases/steward/README.md

loopx dashboard --no-open
loopx --format json todo list --goal-id "$GOAL_ID"

Goal Acceptance Contracts

Activation: Inspect provider revision; prepare the owner-reviewed acceptance.json described in the documentation. Preview loopx goal-acceptance configure --goal-id "$GOAL_ID" --document acceptance.json --expected-provider-revision "$PROVIDER_REVISION", then repeat with --execute.

Validation: loopx goal-acceptance verify --goal-id "$GOAL_ID" --execute, then loopx --format json goal-acceptance inspect --goal-id "$GOAL_ID".

Disable / rollback: Inspect the latest provider revision, then loopx goal-acceptance disable --goal-id "$GOAL_ID" --expected-provider-revision "$CURRENT_PROVIDER_REVISION" --execute. Historical receipts remain intact.

Authority boundary: Existing canonical authority is a prerequisite. Acceptance checks grant no task claim, provider promotion, publication or Goal completion permission. Local filesystem/process permissions remain the trust boundary.

Docs: https://github.com/loopx-project/loopx/blob/v1.1.0/docs/reference/goal-acceptance-observations.md

loopx --format json goal-acceptance inspect --goal-id "$GOAL_ID"

Governed Local Delegation

Activation: Prepare an operator-owned binding file outside all delegated member workspaces, using the versioned documentation. Supply it explicitly through --execution-config to loopx delegation or python -m loopx.collaboration_mcp. Without a configured grant, there is no execution authority.

Validation: Use the following read-only commands with the same registry, runtime and requester identity used by the caller:

loopx --registry "$REGISTRY" --runtime-root "$RUNTIME_ROOT" --format json delegation list --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --execution-config "$DELEGATION_CONFIG"
loopx --registry "$REGISTRY" --runtime-root "$RUNTIME_ROOT" --format json delegation operations --limit 10 --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --execution-config "$DELEGATION_CONFIG"

Disable / rollback: Revoke the relevant caller grant and stop/reconcile workers already running. Restart MCP without --execution-config to remove execution tools. If route projection was enabled using configure-goal --subagent-execution-config, clear it with loopx configure-goal --goal-id "$GOAL_ID" --clear-subagent-execution-config --execute; clearing a pointer alone does not revoke the underlying grant.

Authority boundary: Registration, route readiness and coordinator role do not grant execution. Bindings cannot bypass task, claim, lease, quota, acceptance, workspace isolation or credential custody. Operation inventory is requester-scoped and does not mean Goal completion.

Docs: https://github.com/loopx-project/loopx/blob/v1.1.0/docs/reference/local-delegation.md

loopx --registry "$REGISTRY" --runtime-root "$RUNTIME_ROOT" --format json delegation list --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --execution-config "$DELEGATION_CONFIG"

Goal Chat LoopX Mode

Activation: Run loopx dashboard, open a local managed Codex Goal Chat and choose Enable LoopX. In Settings, select a registered sender, existing project-local delegation configuration and a positive total coordinator token allowance; save and enable explicitly.

Validation: Inspect the native state in the compact bar and open Team execution for read-only operation inventory and member prerequisite checks. loopx --format json agent-directory --goal-id "$GOAL_ID" checks the registered identities; registration alone does not prove execution.

Disable / rollback: Pause coordinator stops only the coordinator; Exit mode restores ordinary conversation. Neither cancels children or completes the canonical Goal. Pause/close Chat before downgrading, and reconcile admitted member work through its original execution controls.

Authority boundary: This requires Codex's experimental native Goal APIs and a local managed conversation. The coordinator retains its read-only sandbox; member permissions come only from operator bindings. Initial tool enrollment may replace an idle executor, never an unfinished native Goal. Host completion is not canonical acceptance, and the driver remains subject to Chat timeout rather than becoming an unattended daemon.

Docs: https://github.com/loopx-project/loopx/blob/v1.1.0/docs/reference/goal-chat-continuation.md

loopx dashboard --no-open
loopx --format json agent-directory --goal-id "$GOAL_ID"

Codex Child Capacity Alignment

Activation: Preview loopx configure-goal --goal-id "$GOAL_ID" --multi-subagent-feature enabled --max-children 2 --align-codex-subagent-capacity, then repeat with --execute only after inspecting the resolved host and proposed change. The optional Dashboard control requires loopx dashboard --enable-goal-subagent-configuration.

Validation: Repeat the same command without --execute and inspect capacity/readback. A changed host configuration requires a new Codex Session; existing Sessions are not hot-reloaded.

Disable / rollback: Disable Goal child orchestration with loopx configure-goal --goal-id "$GOAL_ID" --multi-subagent-feature disabled --execute. This does not lower the host setting. To restore that setting, close affected Sessions and restore the owner-only backup identified by the local receipt after checking for later edits; restart without the Dashboard flag to hide its optional mutation surface.

Authority boundary: Host config and Goal admission remain separate. Alignment only raises a known explicit limit, preserves unrelated settings and does not spawn children, grant credentials or bypass work-scope, quota or acceptance restrictions. Unknown implicit defaults remain unknown.

Docs: https://github.com/loopx-project/loopx/blob/v1.1.0/docs/integrations/codex-subagent-orchestration.md

loopx configure-goal --goal-id "$GOAL_ID" --multi-subagent-feature enabled --max-children 2 --align-codex-subagent-capacity

Validation

Release preparation uses an explicitly authorized baseline-plus-incremental qualification strategy. The full baseline at a064de4521d849d0c9e01d0349efe73c90a433c1 passed 10,242 tests with 53 skips and the Full Public Smokes sweep. Those results are not relabelled as final-commit results. Later implementation changes have their own linked PR evidence; this release candidate separately checks migrated entry points, package/update behavior, corrected smoke contracts, frontend builds and PR CI.

Release source: 607c11d75e9d608e44d9caf6b675b32c0abb8267 (#4751); its Git tree is identical to the validated candidate a615b266679ebb9455f9f43356164bc837ae3222. Candidate checks passed: 19 selected risk checks plus 5 direct checks; 57 focused Python tests plus 3 subtests; baseline/head CLI and pure-model regression comparisons; archive install/update; frontend builds; promotion canary with its dashboard subcheck explicitly skipped and without qualification writeback. Separate PR dashboard acceptance, macOS/Windows desktop builds and package builds passed.

The maintainer explicitly merged without waiting for the remaining PR CI: at merge, 15 checks passed, 8 were still running and 8 were conditionally skipped. This is not an all-green CI claim. An earlier installer attempt timed out at 120 seconds; another was invalidated by an in-flight source change. The fixed-head retry with a 360-second per-check allowance passed without weakening product checks. The all-lanes exact-release-commit qualification is not claimed. No new matched benchmark outcome improvement is claimed, so an expensive comparative outcome campaign is not part of this release.

中文摘要

LoopX 已迁至 loopx-project/loopx。历史、Issue、PR 和 Stars 保留，PyPI 包名仍为 loopx。1.1.0 同时整合 1.0.5 之后的团队协作、验收、委托恢复与工作区改进。

升级决策

**谁需要升级：**使用 Personal Workspace、管家团队计划、本地委托、canonical Todo/Monitor，以及归档安装和桌面更新路径的用户。

**解决了什么：**把团队预览、显式确认与 canonical 工作连接起来；让委托拥有可恢复的 operation 和可核验返回；收敛验收与 replan 语义，并修复仓库迁移后的安装与公共入口。

**是否有破坏性变更：**有，影响已退役 CLI 调用和旧 Pages 地址；不自动迁移状态或提升 provider；验收与执行授权继续显式启用。已退役的 todo suggest、todo capture-followups 被移除（#4725、#4723）。仓库旧链接可跳转，但旧 GitHub Pages 域名不会自动跳转；网站、博客、文档和安装脚本请改用 https://loopx-project.github.io/loopx/。旧版命令若仍输出旧安装地址，使用上方 pip 升级命令或新域名的安装脚本。

**如何验证：**确认 loopx --version 为 1.1.0，运行 loopx doctor --deep；对已有 Goal 查询 agent-directory 与 goal-acceptance inspect，只读检查不会启用可选能力。

**贡献者：**本版本包含 @songoow、@Duang777、@DJC1412 等下列外部开发者的实质贡献；对应 PR 与具体改动见社区贡献者章节。

Agent 团队与个人工作区

  • 管家先展示可检查的 lane、人员缺口与承诺，经 owner 确认才形成 canonical Todos；预览不是写入或执行（#4519、#4535、#4547、#4633）。
  • Goal Chat 提供内联 LoopX 模式与团队工作检查；对话持续存在不等于已启用自主调度，首页聚焦当前 Goal（#4700、#4736）。
  • 已有 Agent 可经 CLI/MCP 显式委托、读取、等待、恢复工作；丢失上下文后找回原 operation，而不是重复启动。执行路由只读投影复用既有授权，不新增调度器（#4688、#4706、#4731、#4741）。
  • 显式 opt-in 可把当前 Codex 宿主的子线程上限提升至 Goal 所需容量，保留更高设置；写入后需新建 Session（#4733）。

状态、验收与恢复

  • 租约、幂等重试、原子交接、source witness 与经过审阅的 Todo/Monitor 编辑收敛至 canonical owner（#4666、#4669、#4670、#4686、#4689、#4693、#4699）。
  • owner 可为已有任务配置有界 artifact 验收；缺失或关闭时保持原行为（#4683）。
  • required-vision replan 的投影与写入共用 typed outcome policy；对应 PR 报告两次真实模型完整链路通过，不等同于本版本全量 live qualification。检查点保留、lane 历史与延迟恢复也得到修复（#4730、#4554、#4654、#4655）。
  • 不完整投影不再猜测隐藏工作类型；无变化 monitor poll 不消耗 quota，周期报告覆盖 Goal 并跨 lane 去重，repair 使用 typed 分类（#4713、#4739、#4729、#4734、#4740）。

性能、集成与公开证据

  • 状态与运行时热路径复用历史、事件、指纹与 readiness 读取（#4632、#4685、#4690、#4694、#4695）。
  • 解释器与扩展诊断更明确；DSH 使用 typed 0.1.5 边界，NoKV 改善 seed routing 和 SDK 兼容处理，不默认提升为 authority（#4482、#4509、#4649、#4724）。
  • 网站补充应用场景、双语 Kanban 文章与 LHTB 研究简报，保留基线、核算修正和反例；不宣称普遍优于其他 harness（#4660、#4673、#4703、#4709、#4714）。

社区贡献者

可选能力启用与使用

Steward Team Plan Confirmation

**启用：**运行 loopx dashboard --no-open，打开 LoopX Manager，请求团队计划，逐 lane 检查并显式确认卡片。

验证：loopx --format json todo list --goal-id "$GOAL_ID"，核对 plan、assignment 和 Todo 属于同一 Goal/lane。

**停用 / 回退：**确认前拒绝预览；确认后用 loopx goal-lifecycle --goal-id "$GOAL_ID" --operation stop --execute 停止新调度，并经正常 owner 处理已创建 Todo。降级不会抹去已有工作。

**权限边界：**预览不授予写入、启动、quota、lease 或发布权限；确认仅物化已检查的计划，不可配备的 lane 仍显示缺口。

**文档：**https://github.com/loopx-project/loopx/blob/v1.1.0/docs/product/use-cases/steward/README.md

loopx dashboard --no-open
loopx --format json todo list --goal-id "$GOAL_ID"

Goal Acceptance Contracts

**启用：**读取 provider revision，按文档准备 owner 审阅的 acceptance.json；先预览 loopx goal-acceptance configure --goal-id "$GOAL_ID" --document acceptance.json --expected-provider-revision "$PROVIDER_REVISION"，再加 --execute 执行。

**验证：**运行 loopx goal-acceptance verify --goal-id "$GOAL_ID" --execute，再运行 loopx --format json goal-acceptance inspect --goal-id "$GOAL_ID"。

**停用 / 回退：**读取最新 provider revision，再运行 loopx goal-acceptance disable --goal-id "$GOAL_ID" --expected-provider-revision "$CURRENT_PROVIDER_REVISION" --execute；不改写历史回执。

**权限边界：**前提是已有 canonical authority；验收不授予任务 claim、provider promotion、发布或 Goal 完成权限。本地文件与进程权限仍是信任边界。

**文档：**https://github.com/loopx-project/loopx/blob/v1.1.0/docs/reference/goal-acceptance-observations.md

loopx --format json goal-acceptance inspect --goal-id "$GOAL_ID"

Governed Local Delegation

**启用：**按文档在所有被委托成员工作区之外准备 operator binding 文件，通过 loopx delegation 或 python -m loopx.collaboration_mcp 的 --execution-config 显式提供；无授权配置不获得执行权限。

**验证：**使用英文对应章节的 delegation list、delegation operations --limit 10 命令，保持 registry、runtime、requester 与原调用者一致；均为只读。

**停用 / 回退：**撤销相应 requester grant，停止或处理仍在运行的 worker；不带 --execution-config 重启 MCP 可移除执行工具。已配置的路由指针可用 loopx configure-goal --goal-id "$GOAL_ID" --clear-subagent-execution-config --execute 清除，但只清指针不等于撤销实际授权。

**权限边界：**注册、路由 ready 或协调者角色都不授予执行权；不绕过任务、claim、lease、quota、验收、工作区隔离或凭据托管。operation 清单限于 requester，不代表 Goal 完成。

**文档：**https://github.com/loopx-project/loopx/blob/v1.1.0/docs/reference/local-delegation.md

loopx --registry "$REGISTRY" --runtime-root "$RUNTIME_ROOT" --format json delegation list --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --execution-config "$DELEGATION_CONFIG"

Goal Chat LoopX Mode

**启用：**运行 loopx dashboard，打开本机 managed Codex Goal 对话，选择 Enable LoopX；设置已注册身份、项目内已有执行配置和协调员总 token 额度，保存后显式开启。

**验证：**查看顶部原生状态，打开 Team execution 只读查询 operation 与成员前置条件；loopx --format json agent-directory --goal-id "$GOAL_ID" 可核对身份，但注册不代表执行成功。

停用 / 回退： Pause coordinator 仅停止协调员，Exit mode 恢复普通对话，都不取消子任务或完成 canonical Goal。降级前暂停/关闭 Chat，并通过原执行控制处理已派发成员。

**权限边界：**限本机 managed Codex 对话及实验性原生 Goal API；协调员保持只读沙箱，成员权限来自 operator binding。首次增配工具只可替换闲置执行器，不能替换未结束 Goal。原生完成不等于 canonical 验收；仍受 Chat 超时约束，不是无人值守 daemon。

**文档：**https://github.com/loopx-project/loopx/blob/v1.1.0/docs/reference/goal-chat-continuation.md

loopx dashboard --no-open
loopx --format json agent-directory --goal-id "$GOAL_ID"

Codex Child Capacity Alignment

**启用：**先预览 loopx configure-goal --goal-id "$GOAL_ID" --multi-subagent-feature enabled --max-children 2 --align-codex-subagent-capacity，检查实际宿主与变更后才加 --execute。可选 Dashboard 控件需 loopx dashboard --enable-goal-subagent-configuration。

**验证：**不带 --execute 重复命令，检查容量和读回结果；宿主配置变更后需新建 Codex Session，已有会话不热加载。

停用 / 回退：loopx configure-goal --goal-id "$GOAL_ID" --multi-subagent-feature disabled --execute 关闭 Goal 子任务编排，但不会降低宿主上限。恢复宿主配置前关闭受影响会话，核对后续编辑，再恢复本地回执标识的 owner-only 备份；不带 Dashboard 启动开关可隐藏其可选修改入口。

**权限边界：**宿主配置和 Goal admission 分别拥有权威；对齐仅提升已知显式上限、保留无关设置，不启动子任务、不授予凭据，也不绕过 scope、quota 或验收。未知隐式默认值保持 unknown。

**文档：**https://github.com/loopx-project/loopx/blob/v1.1.0/docs/integrations/codex-subagent-orchestration.md

loopx configure-goal --goal-id "$GOAL_ID" --multi-subagent-feature enabled --max-children 2 --align-codex-subagent-capacity

发布验证

本次经明确授权采用“已有全量基线 + 增量验证”。a064de4521d849d0c9e01d0349efe73c90a433c1 上的 10,242 passed / 53 skipped 与 Full Public Smokes 作为历史基线，不改写成最终提交的结果。之后的功能变化保留各自 PR 证据，本次另验迁移入口、安装更新、修复后的 smoke、前端构建和 PR CI。

发布源码为 607c11d75e9d608e44d9caf6b675b32c0abb8267（#4751），Git tree 与实测候选 a615b266679ebb9455f9f43356164bc837ae3222 完全一致。候选通过 19 项风险检查、5 项直接检查、57 项聚焦 Python 测试及 3 项 subtests、CLI/纯模型基线对照、归档安装更新与前端构建；promotion canary 通过，但显式跳过其 dashboard 子检查，未写资格回执。独立 PR dashboard acceptance、macOS/Windows 桌面构建和发行包构建均通过。

维护者明确选择不等待剩余 PR CI 而合并：当时 15 项通过、8 项仍运行、8 项按条件跳过；不宣称 CI 全绿。此前 installer 的 120 秒超时和一次执行中提交变化导致的来源校验拒绝均保留；固定 head 后以每项 360 秒等待重跑通过，未放宽产品校验。不宣称通过全套 exact-release-commit qualification，也不宣称新的统一 benchmark 收益，因此未重新开展昂贵的对照实验。

Publication Readback / 发布读回

发行包、PyPI、macOS/Windows 安装包、签名更新源和网站部署均完成；stable 已快进到发布提交。从公网 PyPI 新建隔离环境安装成功，控制面探针通过。这些分发验证不改变上文对历史基线、合并时未完成 PR CI 和资格范围的披露。

Contributors

songoow, DJC1412, and Duang777
Assets 13
Loading