Ultra Orchestration is an open-source skill repository for disciplined AI-assisted software delivery. It provides a two-plane method for turning ambiguous engineering requests into structured specifications, bounded execution packages, review evidence, QA results, and auditable delivery records.
- OpenSpec specification plane for long-lived program, milestone, change, and archive assets
- Ultra-Orchestrator execution plane for intake, planning, dispatch, execution, review, QA, delivery, risk control, and retro
The repository ships a mainline skill suite and a compatibility preview suite:
skills/is the mainline suite. Start here.skills-vnext/keeps vNext compatibility aliases and preview internals.
Mainline entry points:
ultra-orchestratoris the default public entry point.ultra-vnext-coreremains a compatibility alias.ultra-orchestrator-legacypreserves the older lightweight workflow.
Backstory and whitepaper page:
Ultra Orchestration organizes AI-assisted work through a structured engineering lifecycle:
- clarify what is being built
- turn the request into a reviewable design or spec
- create a decision-complete task graph
- dispatch bounded work packages
- review against spec and engineering quality
- verify behavior through QA
- package the deliverable, evidence, risks, and retro
The mainline ultra-orchestrator now includes the vNext strict control plane:
- design-first clarification and explicit approval gates
- OpenSpec-backed durable specification artifacts
- finite-state-machine loopbacks instead of fragile linear waterfall execution
- artifact-driven handoff instead of full chat-history inheritance
- DAG plus write-lock rules for safe parallelism
- host-driven ledger ownership
- run mode selection:
LIGHT,STANDARD,STRICT,STRICT_OPENSPEC - OpenSpec-first defaults for development work
- machine-checkable JSON artifacts and contract validation
The default experience is:
- start with
ultra-orchestrator - let the entry point choose run mode and route subskills
- prefer
STRICT_OPENSPECfor development work - require ledger-backed planning before controlled dispatch
- proceed through execution, review, QA, risk evidence, and delivery
Under skills/:
ultra-orchestratorultra-orchestrator-legacyopenspec-ultra-bridgeclarify-and-intakedecision-complete-plannerdispatch-and-trackspec-reviewcode-reviewqa-verifyrisk-vettersafety-guarddeliver-and-retroautoplan
Use ultra-orchestrator for new work. Use ultra-orchestrator-legacy only
when you need the original lightweight staged flow.
Under skills-vnext/:
ultra-vnext-coreultra-brainstormingultra-planningultra-execution-controlultra-reviewultra-qaultra-risk-vettingultra-deliveryopenspec-ultra-bridge-v2
Use this track only when you need vNext compatibility names or internal
protocol references. New projects should start from ultra-orchestrator.
- Promoted the vNext strict protocol into
skills/ultra-orchestrator. - Preserved the old stable orchestrator as
skills/ultra-orchestrator-legacy. - Kept
skills-vnext/ultra-vnext-coreas a compatibility alias. - Made development tasks prefer
STRICT_OPENSPECby default. - Added mandatory
control_surface_usedreporting for strict runs.
| Mainline / vNext skill | Legacy relationship | Migration type | Main upgrade |
|---|---|---|---|
ultra-orchestrator |
ultra-orchestrator-legacy |
Mainline replacement | Adds run mode selection, strict OpenSpec preference, ledger-backed execution, JSON artifacts, contract validation, slice DAGs, and control-surface reporting. |
ultra-vnext-core |
ultra-orchestrator |
Compatibility alias | Keeps existing vNext invocations working while following the mainline protocol. |
ultra-brainstorming |
clarify-and-intake + early autoplan |
Enhanced replacement | Adds design-first discovery, hard gates, one-question-at-a-time clarification, approach comparison, and approval before planning or coding. |
ultra-planning |
decision-complete-planner |
Direct vNext replacement | Keeps task graph planning but strengthens DAG dependencies, owned paths, acceptance checks, retry assumptions, and serial/parallel boundaries. |
ultra-execution-control |
dispatch-and-track + part of safety-guard |
Enhanced integration | Combines dispatch, write locks, bounded retries, blockers, freeze/careful/guard behavior, and host-driven ledger updates. |
ultra-review |
spec-review + code-review |
Merged replacement | Keeps spec fidelity and engineering quality as two internal gates, then returns accept, reject-to-execute, or reroute-to-plan. |
ultra-qa |
qa-verify |
Direct vNext replacement | Makes static QA and dynamic QA explicit, with scenario coverage, evidence, pass/fail, and loopback routing. |
ultra-risk-vetting |
risk-vetter + approval-facing part of safety-guard |
Direct vNext replacement | Defines LOW/MEDIUM/HIGH/EXTREME thresholds, approval requirements, guardrail mapping, and vetter report output. |
ultra-delivery |
deliver-and-retro |
Direct vNext replacement | Packages final_deliverable, orchestration_log, vetter_report, QA evidence, residual risks, follow-ups, and retro. |
openspec-ultra-bridge-v2 |
openspec-ultra-bridge |
Enhanced replacement | Keeps OpenSpec as the spec layer and Ultra as the control plane, with bridge scripts that materialize TaskManifest and WorkPackage artifacts. |
- Added
ultra-risk-vetting - Added
ultra-delivery - Updated README with stable-to-vNext mapping
- Added clearer installation, startup, and success-evaluation guidance
- OpenSpec owns long-lived specification assets and status accounting
- Ultra-Orchestrator owns execution control, quality gates, delivery evidence, and risk decisions
Neither replaces the other. They stay synchronized.
For non-trivial work, use:
ProgramMilestoneChangeSlice
Then run execution in this phase order:
- Intake
- Plan
- Dispatch
- Execute
- Review
- QA
- Deliver
- Retro
Changeis the specification and planning unitSliceis the implementation and verification unit
Do not run milestone-wide engineering without change and slice decomposition.
Canonical slice states:
slice_0_not_openedslice_0_spec_readyslice_1_completedslice_2_in_progressslice_3_qa_pendingslice_4_done
Advance slice status only with verification evidence.
Each active change should include:
proposal.mdtasks.mdultra-bridge.md
Recommended for richer changes:
design.md
After each completed slice:
- update
tasks.md - update
ultra-bridge.md - update any program-facing status rollups if counts changed
- record verification evidence
Recommended:
git clone https://github.com/keithhegit/ultra-orchestration.git
cd ultra-orchestrationYou can also download a GitHub release archive, but the examples below assume a normal cloned repository.
Copy folders from skills/ into your agent's global or
project-level skill directory.
Common targets:
| Agent | Project-level target | User-level target | Notes |
|---|---|---|---|
| Codex | <PROJECT_ROOT>\.agents\skills |
%USERPROFILE%\.codex\skills |
Use project-level install when sharing skills with a repo. |
| Cursor | <PROJECT_ROOT>\.agents\skills or Cursor's configured skill path |
%USERPROFILE%\.cursor\skills |
Cursor setups vary by version and plugin support. |
| Claude Code | <PROJECT_ROOT>\.claude\skills |
%USERPROFILE%\.claude\skills |
Claude Code discovers project and user skills. |
| OpenCode | <PROJECT_ROOT>\.opencode\skills |
platform-specific home config | OpenCode also discovers compatible .claude\skills and .agents\skills folders. |
| OpenClaw, Hermes, other IDE agents | the agent's documented skills directory | the agent's documented user skills directory | If the agent supports SKILL.md folders, copy each skill folder as-is. |
Platform references:
Generic project-level example:
$TARGET = "<PROJECT_ROOT>\.agents\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills\ultra-orchestrator "$TARGET\ultra-orchestrator"
Copy-Item -Recurse -Force .\skills\openspec-ultra-bridge "$TARGET\openspec-ultra-bridge"Claude Code project-level example:
$TARGET = "<PROJECT_ROOT>\.claude\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills\ultra-orchestrator "$TARGET\ultra-orchestrator"
Copy-Item -Recurse -Force .\skills\openspec-ultra-bridge "$TARGET\openspec-ultra-bridge"OpenCode project-level example:
$TARGET = "<PROJECT_ROOT>\.opencode\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills\ultra-orchestrator "$TARGET\ultra-orchestrator"
Copy-Item -Recurse -Force .\skills\openspec-ultra-bridge "$TARGET\openspec-ultra-bridge"Restart or refresh your agent after copying skills.
New users should install and invoke ultra-orchestrator. The compatibility
folders are only for older prompts or migration testing.
Legacy lightweight workflow:
$TARGET = "<PROJECT_ROOT>\.agents\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills\ultra-orchestrator-legacy "$TARGET\ultra-orchestrator-legacy"Older vNext prompt compatibility:
$TARGET = "<PROJECT_ROOT>\.agents\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills-vnext\ultra-vnext-core "$TARGET\ultra-vnext-core"Do not start new projects from the compatibility aliases unless you are testing
backward compatibility. Start new work with $ultra-orchestrator.
ultra-orchestrator is the default public entry point. Users should not need
to name every subskill. The orchestrator chooses run mode, prefers
STRICT_OPENSPEC for development work, and routes to the right sequence:
OpenSpec bootstrap or bridge, planning, risk vetting, execution control,
review, QA, and delivery.
$ultra-orchestrator <task description>
Example:
$ultra-orchestrator Build a workspace settings page for model defaults.
$ultra-orchestrator OpenSpec change <change-id or path>: <task description>
Example:
$ultra-orchestrator OpenSpec change workspace-model-defaults: implement the first slice.
$ultra-orchestrator bugfix: <symptom or failing behavior>
Example:
$ultra-orchestrator bugfix: command execution hangs after approval is granted.
If your agent client exposes slash aliases, use /ultra-orchestrator with the
same forms. $ultra-vnext-core remains available as a compatibility alias.
ultra-orchestrator is the primary router. It should classify run mode first,
then load only the required subskills.
Typical routes:
- development, feature, bugfix, pilot, or multi-file implementation
ultra-orchestrator -> STRICT_OPENSPEC -> openspec-ultra-bridge -> decision-complete-planner -> dispatch-and-track -> review -> qa-verify -> deliver-and-retro
- existing OpenSpec change
ultra-orchestrator -> openspec-ultra-bridge -> decision-complete-planner -> dispatch-and-track -> review -> qa-verify -> deliver-and-retro
- review-only or QA-only
ultra-orchestrator -> LIGHT -> review/QA -> deliver-and-retro
For STRICT and STRICT_OPENSPEC, markdown-only artifacts are not enough. A
run is not execution-ready until it has a ledger, JSON TaskManifest, JSON
WorkPackage artifacts, owned paths, acceptance checks, dependency order,
retry assumptions, and a contract-validation plan.
Initialize a mainline run directory:
python .\skills\ultra-orchestrator\scripts\new_run.py .\runs --run-id run-001 --run-mode STRICT_OPENSPECValidate core JSON artifacts:
python .\skills\ultra-orchestrator\scripts\validate_contracts.py .\runs\run-001\ledger.json --kind executionledgerOptional bridge helper for an OpenSpec change:
python .\skills-vnext\openspec-ultra-bridge-v2\scripts\bridge_change.py .\openspec\changes\<change-id> .\runs\run-001\bridge-outputChinese pilot guide:
A mainline Ultra run is considered healthy when:
- the user only needs to invoke
ultra-orchestrator - development work defaults to
STRICT_OPENSPEC - the run does not silently downgrade to markdown-only orchestration
- OpenSpec change or bootstrap status is explicit before implementation
ledger.json, JSONTaskManifest, and JSONWorkPackageartifacts exist for strict runs- contract validation is run or an explicit blocker explains why it could not run
- slice status, owned paths, review gate, and QA gate are clear
- risk decisions appear in
vetter_report - delivery includes
final_deliverable,orchestration_log,vetter_report, andcontrol_surface_used - skipped control surfaces, residual risks, and follow-ups are separated from completed work
.github/
ISSUE_TEMPLATE/
workflows/
skills/
mainline skills and legacy compatibility
skills-vnext/
vNext compatibility aliases and preview internals
docs/
ultra-vnext-pilot-guide-CN.md
docs.html
README.md
LICENSE
CONTRIBUTING.md
SECURITY.md
CODE_OF_CONDUCT.md
- update or add skills in this repository
- validate the workflow in a real project
- commit and tag a release if needed
- sync released skills into the target global or workspace skill directory
- protocol over improvisation
- evidence over claims
- design before implementation
- safe parallelism over accidental concurrency
- durable artifacts over chat memory
- Ultra remains the control plane
This project is licensed under the Apache License 2.0. See LICENSE.
Ultra Orchestration 是一个面向 AI 辅助软件交付的开源技能仓库。它通过标准化规格、任务拆解、执行控制、审查、QA、风险记录和交付证据,帮助 AI 编程代理更稳定地处理复杂工程任务。
项目采用双平面协作模型:
- OpenSpec 规格平面:负责长期规格资产、里程碑、变更与归档
- Ultra-Orchestrator 编排平面:负责 Intake、Plan、Dispatch、Execute、Review、QA、Deliver、风险控制与 Retro
仓库提供两组技能目录:
skills/是主线技能与 legacy 兼容入口skills-vnext/是 vNext 兼容别名与内部协议参考
主线入口:
ultra-orchestrator是默认公开入口ultra-vnext-core保留为兼容别名ultra-orchestrator-legacy保留旧轻量流程
文档与白皮书:
Ultra Orchestration 通过工程化生命周期管理提升 AI 辅助交付的可控性和可审计性:
- 澄清目标、边界和成功标准
- 把请求变成可审查的设计或规格
- 生成 decision-complete 任务图
- 派发边界明确的工作包
- 做规格一致性和工程质量审查
- 通过 QA 验证行为
- 交付结果、证据、风险和复盘
主线 ultra-orchestrator 已合入 vNext strict 控制面:
- 设计先行的澄清流程与显式审批门
- 基于 OpenSpec 的持久规格资产
- 用有限状态机闭环替代单向执行流程
- 用 artifact handoff 替代全文上下文传递
- 用 DAG + 写锁定义安全并发
- 用 host-driven ledger 管理执行状态
- 支持
LIGHT、STANDARD、STRICT、STRICT_OPENSPEC运行模式 - 开发任务默认优先
STRICT_OPENSPEC - 使用机器可校验 JSON 工件和 contract validation
默认使用方式是:
- 从
ultra-orchestrator启动 - 由主入口判断运行模式并自动路由
- 开发任务优先进入
STRICT_OPENSPEC - 在进入受控 dispatch 前建立 ledger 和 JSON planning artifacts
- 再进入执行、审查、QA、风险记录和交付
位于 skills/:
ultra-orchestratorultra-orchestrator-legacyopenspec-ultra-bridgeclarify-and-intakedecision-complete-plannerdispatch-and-trackspec-reviewcode-reviewqa-verifyrisk-vettersafety-guarddeliver-and-retroautoplan
新工作默认使用 ultra-orchestrator。只有需要旧轻量流程时才使用
ultra-orchestrator-legacy。
位于 skills-vnext/:
ultra-vnext-coreultra-brainstormingultra-planningultra-execution-controlultra-reviewultra-qaultra-risk-vettingultra-deliveryopenspec-ultra-bridge-v2
这组技能保留为兼容入口和内部协议参考。新项目默认从 ultra-orchestrator 启动,不需要手动枚举这些子技能。
| 当前入口或技能 | 旧版或兼容来源 | 迁移关系 | 主要变化 |
|---|---|---|---|
ultra-orchestrator |
ultra-vnext-core + 原 ultra-orchestrator |
新默认主线 | 将 vNext strict protocol 提升为默认入口,自动判断 LIGHT / STANDARD / STRICT / STRICT_OPENSPEC,开发任务优先 strict OpenSpec。 |
ultra-orchestrator-legacy |
原 ultra-orchestrator |
兼容保留 | 保留旧轻量阶段化编排,供老项目或非 strict 场景使用。 |
ultra-vnext-core |
vNext 预览入口 | 兼容别名 | 继续支持已有 vNext 调用,但推荐新项目改用 ultra-orchestrator。 |
ultra-brainstorming |
clarify-and-intake + autoplan 前半段 |
增强版 | 增加 design-first、显式审批门、逐问澄清、方案对比和规划前确认。 |
ultra-planning |
decision-complete-planner |
直接对应 | 强化 DAG、owned paths、acceptance checks、retry 假设和串并行边界。 |
ultra-execution-control |
dispatch-and-track + 部分 safety-guard |
整合版 | 合并 dispatch、写锁、受限重试、blocker、freeze/careful/guard 和 ledger 更新纪律。 |
ultra-review |
spec-review + code-review |
合并版 | 保留规格一致性和工程质量两层审查,并输出 accept / reject / reroute。 |
ultra-qa |
qa-verify |
直接对应 | 明确 static QA / dynamic QA 双模式,补充场景覆盖、证据和失败回流。 |
ultra-risk-vetting |
risk-vetter + safety-guard 的审批部分 |
直接对应 | 定义 LOW/MEDIUM/HIGH/EXTREME、审批门槛、护栏映射和 vetter report。 |
ultra-delivery |
deliver-and-retro |
直接对应 | 统一 final_deliverable、orchestration_log、vetter_report、QA 证据、残余风险和复盘。 |
openspec-ultra-bridge-v2 |
openspec-ultra-bridge |
增强版 | 保持 OpenSpec 为规格层、Ultra 为控制平面,并用脚本生成 TaskManifest 与 WorkPackage 工件。 |
- 新增
ultra-risk-vetting - 新增
ultra-delivery - README 增加稳定版到 vNext 对应表
- README 增加安装、启动和成功评估指南
- OpenSpec 负责长期规格资产与状态核算
- Ultra-Orchestrator 负责执行控制、质量门禁、交付证据与风险决策
两者同步协作,互不替代。
对非平凡任务,先按以下结构组织:
ProgramMilestoneChangeSlice
然后按以下阶段执行:
- Intake
- Plan
- Dispatch
- Execute
- Review
- QA
- Deliver
- Retro
Change是规格与计划单位Slice是实现与验证单位
不要直接以整段 Milestone 作为粗粒度开发单位。
使用以下标准状态:
slice_0_not_openedslice_0_spec_readyslice_1_completedslice_2_in_progressslice_3_qa_pendingslice_4_done
只有存在验证证据时,才推进 slice 状态。
每个活跃 change 建议至少包含:
proposal.mdtasks.mdultra-bridge.md
更复杂的变更建议补上:
design.md
每完成一个 slice,至少同步四件事:
- 更新
tasks.md - 更新
ultra-bridge.md - 如果项目级统计变化,更新汇总状态文档
- 记录验证证据
推荐使用 git clone:
git clone https://github.com/keithhegit/ultra-orchestration.git
cd ultra-orchestration也可以下载 GitHub release 或源码压缩包,但下面的命令默认你已经进入克隆后的仓库目录。
把 skills/ 里的技能目录复制到目标 agent 的全局或项目级技能目录。
常见目标目录:
| Agent | 项目级目录 | 用户级目录 | 说明 |
|---|---|---|---|
| Codex | <PROJECT_ROOT>\.agents\skills |
%USERPROFILE%\.codex\skills |
如果希望技能随项目共享,优先使用项目级目录。 |
| Cursor | <PROJECT_ROOT>\.agents\skills 或 Cursor 当前配置的 skills 目录 |
%USERPROFILE%\.cursor\skills |
Cursor 的插件和技能支持会随版本与配置变化。 |
| Claude Code | <PROJECT_ROOT>\.claude\skills |
%USERPROFILE%\.claude\skills |
Claude Code 支持项目级和用户级 skills。 |
| OpenCode | <PROJECT_ROOT>\.opencode\skills |
按 OpenCode 当前平台配置 | OpenCode 也会发现兼容的 .claude\skills 和 .agents\skills。 |
| OpenClaw、Hermes、其他 IDE Agent | 对应工具文档中的 skills 目录 | 对应工具文档中的用户级 skills 目录 | 如果工具支持 SKILL.md 文件夹结构,保持目录原样复制即可。 |
平台参考:
通用项目级示例:
$TARGET = "<PROJECT_ROOT>\.agents\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills\ultra-orchestrator "$TARGET\ultra-orchestrator"
Copy-Item -Recurse -Force .\skills\openspec-ultra-bridge "$TARGET\openspec-ultra-bridge"Claude Code 项目级示例:
$TARGET = "<PROJECT_ROOT>\.claude\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills\ultra-orchestrator "$TARGET\ultra-orchestrator"
Copy-Item -Recurse -Force .\skills\openspec-ultra-bridge "$TARGET\openspec-ultra-bridge"OpenCode 项目级示例:
$TARGET = "<PROJECT_ROOT>\.opencode\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills\ultra-orchestrator "$TARGET\ultra-orchestrator"
Copy-Item -Recurse -Force .\skills\openspec-ultra-bridge "$TARGET\openspec-ultra-bridge"复制完成后,请重启或刷新对应 agent 的上下文。
新用户只需要安装并调用 ultra-orchestrator。下面两个入口只用于老提示词兼容或迁移测试。
旧轻量流程:
$TARGET = "<PROJECT_ROOT>\.agents\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills\ultra-orchestrator-legacy "$TARGET\ultra-orchestrator-legacy"旧 vNext 提示词兼容:
$TARGET = "<PROJECT_ROOT>\.agents\skills"
New-Item -ItemType Directory -Force $TARGET
Copy-Item -Recurse -Force .\skills-vnext\ultra-vnext-core "$TARGET\ultra-vnext-core"除非你正在测试兼容性,否则新项目不要从兼容别名启动。新工作默认使用 $ultra-orchestrator。
ultra-orchestrator 是默认公开入口。用户不需要手动列出所有子技能;主入口会先判断运行模式,开发任务默认优先 STRICT_OPENSPEC,再自动路由到 OpenSpec bootstrap 或 bridge、planning、risk vetting、execution control、review、QA 和 delivery。
$ultra-orchestrator <任务描述>
示例:
$ultra-orchestrator 构建一个工作区模型默认值设置页。
$ultra-orchestrator OpenSpec change <change-id 或路径>: <任务描述>
示例:
$ultra-orchestrator OpenSpec change workspace-model-defaults: 实现第一个 slice。
$ultra-orchestrator bugfix: <异常现象或失败行为>
示例:
$ultra-orchestrator bugfix: 审批通过后命令执行卡住。
如果你的 agent 客户端提供 slash alias,也可以用 /ultra-orchestrator 加同样的参数启动。ultra-vnext-core 仍保留为兼容别名。
ultra-orchestrator 是主路由器。它应先判断运行模式,再按需加载最少必要的子技能。
常见路由:
- 开发任务、feature、bugfix、pilot 或多文件实现
ultra-orchestrator -> STRICT_OPENSPEC -> openspec-ultra-bridge -> decision-complete-planner -> dispatch-and-track -> review -> qa-verify -> deliver-and-retro
- 已有 OpenSpec change
ultra-orchestrator -> openspec-ultra-bridge -> decision-complete-planner -> dispatch-and-track -> review -> qa-verify -> deliver-and-retro
- review-only 或 QA-only
ultra-orchestrator -> LIGHT -> review/QA -> deliver-and-retro
对于 STRICT 和 STRICT_OPENSPEC,Markdown-only 工件不够。只有当运行具备 ledger、JSON TaskManifest、JSON WorkPackage、owned paths、acceptance checks、依赖顺序、retry 假设和 contract validation 计划后,才算 execution-ready。
初始化主线 run 目录:
python .\skills\ultra-orchestrator\scripts\new_run.py .\runs --run-id run-001 --run-mode STRICT_OPENSPEC校验核心 JSON 工件:
python .\skills\ultra-orchestrator\scripts\validate_contracts.py .\runs\run-001\ledger.json --kind executionledger可选:桥接 OpenSpec change:
python .\skills-vnext\openspec-ultra-bridge-v2\scripts\bridge_change.py .\openspec\changes\<change-id> .\runs\run-001\bridge-output中文 pilot 指南:
一次主线 Ultra 运行可以认为健康,当它满足:
- 用户只需要调用
ultra-orchestrator - 开发任务默认进入
STRICT_OPENSPEC - 运行不会静默降级成 Markdown-only 编排
- 实现前存在明确 OpenSpec change 或 bootstrap 状态
- strict run 具备
ledger.json、JSONTaskManifest和 JSONWorkPackage - contract validation 已运行,或有明确 blocker 说明为什么无法运行
- slice status、owned paths、review gate 和 QA gate 明确
vetter_report里有风险决策- delivery 包含
final_deliverable、orchestration_log、vetter_report和control_surface_used - skipped control surfaces、残余风险和 follow-up 没有混进“已完成”结论
.github/
ISSUE_TEMPLATE/
workflows/
skills/
mainline skills and legacy compatibility
skills-vnext/
vNext compatibility aliases and preview internals
docs/
ultra-vnext-pilot-guide-CN.md
docs.html
README.md
LICENSE
CONTRIBUTING.md
SECURITY.md
CODE_OF_CONDUCT.md
- 在本仓库中更新技能
- 在真实项目中验证
- 按需要提交并打 tag
- 将发布版本同步到目标全局或工作空间技能目录
- 协议优先于即兴发挥
- 证据优先于口头结论
- 设计先于实现
- 安全并发优先于偶然并发
- 持久工件优先于聊天记忆
- Ultra 始终是控制平面
本项目使用 Apache License 2.0。 详见 LICENSE。