Skip to content

[Roadmap]: LoopX overall delivery program / LoopX 整体路线与交付总跟踪 #4574

Description

@huangruiteng

方向讨论 / Portfolio discussion: #4575

总目标 / Objective

通过本地前端或 Lark 提出、修订和验收复杂目标,由持久管家协调多个长程 LoopX Agent,在本地 managed 与云端 runtime 上完成可验证工作。先证明小团队协作、handoff 和恢复,再分级资格化百 Agent 并发。

Let an owner express, revise and accept complex goals through the local frontend or Lark, while a persistent steward coordinates long-running LoopX Agents across local managed and cloud runtimes. Prove small-team collaboration, handoff and recovery before hundred-Agent concurrency.

Canonical plan: 中文总纲 · English roadmap · merged #4570.

本 Issue 跟踪跨领域依赖与组合验收;领域 Issue/实际 Todo 跟踪可领取工作,避免重复建账。总纲覆盖原有 30 个主 RFC、13 条工作流及 RFC 外的重要领域,仍保留各领域 Draft/Accepted、权限、迁移和 provider 晋升门槛。

This umbrella tracks dependencies and portfolio acceptance. Domain issues/canonical Todos own claimable work. The roadmap covers all 30 pre-existing primary RFCs, 13 streams and important non-RFC areas without changing domain acceptance, permission or promotion gates.

组合里程碑 / Portfolio gates

  • G0 — 可信基线 / Trustworthy baseline: 修复 R1 确认承诺丢失、stale basis、空成功与部分提交恢复;明确单 Agent 发布包旅程及成本未知项。 Repair R1 commitment/freshness/empty-success/recovery gaps and qualify the basic release-artifact journey.
  • G1 — 小团队交付 / Small-team delivery: 管家 + 2–3 个长程 worker,至少两轮真实依赖产物和 peer handoff,含故障、用户补充与独立验收。 Two real cycles with dependencies, peer handoff, recovery, owner correction and independent acceptance.
  • G2 — 可恢复协作工作台 / Recoverable workspace: R3 通用 peer 请求/复核/自动回报,R4 对齐与受治理修订,R5 本地 durable authority;frontend/Lark 同源回读。 General collaboration, alignment and qualified local durability with shared readback.
  • G3 — 本地/云端协作 / Local-cloud work: R6 至少两真实 host,认证作用域、撤销、共享预算、网络恢复与旧 worker fence。 Two real hosts with authenticated scope, shared budget and recovery.
  • G4 — 百 Agent 资格 / Hundred-Agent qualification: R7 按 10→30→100+ 活跃 cohort 验收预先冻结的成本、吞吐、延迟和恢复指标;注册数不等于并发执行。 Qualify active cohorts against frozen SLOs; registrations are not concurrency evidence.
  • G5 — 可重复采用 / Repeatable adoption: 独立安装、升级、导出/回滚,工程及知识/研究两类案例,维护与支持成本可追溯;可从 G0 并行推进。 Independent adoption and repeatable cases can progress alongside G0, without waiting for G4.

首批执行顺序 / First execution sequence

  1. R1 / P0: 在最新 main 重现总纲 F1–F4,合并已存在修复,逐字段核对确认→持久工作→回读;再补同一路径的源基线、幂等和持久恢复。 Reproduce on current main, reconcile existing fixes, preserve confirmed semantics and qualify retry/recovery through the owning path.
  2. R2 + minimum R3 / P0: 使用真实合格 binding 验 2–3 worker。必须通过 Agent 间请求/产物交接推进依赖,不能以创建多条 Todo 或发送消息作为成功。 Require real inter-Agent dependency handoff, not just multiple Todos or sent messages.
  3. R3–R5 / P1: 完成 peer 协作/结果返回、共享修订及相关 TS 整笔事务/本地存储资格;不等待全量 TS 重写,也不复制 work/lease/effect authority。 Complete bounded transactions and qualifications without a second authority or full-rewrite prerequisite.
  4. R6–R7 / P2: 本地云端真实协作后扩规模;只读研究、材料、诊断、采用和发布质量按依赖同期推进。 Expand after real cross-host evidence; parallel streams retain their own gates.

Goal lifetime integration / Goal 生命周期联动

#5389 已合并 quota_settlement 实例归属隔离及双语规划更新;生命周期与恢复继续由 #5206 负责。沿用 R1–R3/G1 的创建 Goal → 创建/复用 Agent → 两轮依赖产物与 peer handoff → 中断纠偏 → 独立验收和原入口回报旅程。P1 接共享预算与 R5/M4 孤儿恢复,P2 接 R6/G3 跨主机;没有新增并行里程碑。

规范路线与验收顺序 · 具体 golden-query 场景。先验已支持 profile 的普通小团队;选定 source profile 的 owner/compatibility/drain 门槛通过后,再验隔离的同名重建和迟到结果。#5389 的组件验证不关闭 G1、M3/M4/M5,不激活 source-session profile,也不代替 packaged App、CLI、Lark 的组合回读。

Existing owners and trackers / 复用现有跟踪

Area Existing tracker
Manager and semantic peer handoff #4339; continuation #4085
TS kernel and shared authority #3225, #3245; local durability #4224
Shared intent and direction #3836, #2831
Goal lifetime, orphan recovery and creation/team integration #5206; quota foundation #5389
Frontend/Lark and review UX #3244, #3786; computer use #4114
Memory and cost #3214, #3824, #3163
Daemon and hooks #3930, #3479
Research and benchmark evidence #3243, #3246, #4391

未列出的工作流仍在总纲 S1–S13 中有 owner/下一步;本表不声称每项 tracker 都已完成或立即可领取。其余已有 Issue/PR 先查重再拆分。

Other streams retain owners/next steps in S1–S13. Links do not establish completion or immediate claimability; reconcile existing work before opening child tasks.

Completion contract / 完成合同

每个后继 Issue/PR 写明 S/G/R 与领域验收编号、真实 caller/唯一 owner、当前 baseline、独立语义预期、失败/重试/越权反例、frontend/Lark/CLI 回读、验证范围、回滚及一个有界后继。记录已完成/失败/未测;字段、消息、receipt 或测试数量都不能单独证明完整能力。

Each successor names S/G/R and domain acceptance IDs, real caller/single owner, baseline, independent expected semantics, negative/recovery cases, entrypoint readback, evidence limits, rollback and a bounded successor. Mark passed/failed/untested honestly. Update the canonical roadmap when dependencies or portfolio gates change; do not copy long execution logs here.

Activity

  1. pinned this issue on Sep 16, 2026
  2. huangruiteng commented on Sep 27, 2026

    @huangruiteng
    CollaboratorAuthor

    RFC delivery task map / RFC 交付任务对应表(2026-09-28)

    One RFC → one delivery issue/retained issue record. 39 accepted RFCs mapped: 17 new task issues and 22 reused records. This follows the contributor-task format used by the semantic-convergence program, without creating another roadmap or splitting every milestone into a new issue.

    Audit basis: public main at ce3862e33, open issues/PRs and current implementation ownership. RFC headings and old issue descriptions can lag shipped code; new tasks explicitly separate implemented foundations, remaining acceptance and promotion. Source inspection here is not a fresh runtime qualification.

    当前按“每份 RFC 一个任务”组织;已存在的任务直接复用,已交付部分不重新招领。新任务含当前实现、剩余工作、现有 PR 衔接与验收条件。没有新增子任务树;本表也不改变 RFC 的权限、预算、兼容和验收边界。

    RFC Task / retained record Current routing
    RFC: Agent IM, LoopX, And OpenViking Collaboration v0 #5198 New · bounded contribution available
    RFC: Agent Loop Effect Interpreter #5199 New · coordinate with active implementation/PRs
    RFC: Agent Session Execution Modes (v0) #5200 New · coordinate with active implementation/PRs
    App conversations and reusable asynchronous work delivery #5201 New · coordinate with active implementation/PRs
    RFC: Automatic Execution Admission (v0) #5202 New · coordinate with active implementation/PRs
    RFC: Benchmark Study Upload and Dashboard Projection v0 #5203 New · bounded contribution available
    RFC: Capable Agent Manager and Semantic Work Handoff (v0) #4339 Active implementation; reuse current handoff/return work
    Explicit Todo continuation: Stage A #4085 Stage A shipped; preserve compatibility and integrate through manager handoff
    RFC: LoopX Desktop Execution Frontends v0 #5204 New · coordinate with active implementation/PRs
    RFC: External Evidence Research Capability v0 #5205 New · bounded contribution available
    RFC: Frontier Science Research Program v0 #4391 Research portfolio; each experiment retains its own selection/budget gates
    RFC: Goal Artifact Lifecycle Projection (milestone / guard / next-transition) v0 #4128 Read-model implementation exists; reconcile remaining acceptance before new code
    RFC: Goal Channel Collaboration v0 #2817 Closed v0 design/delivery record; Lark vertical exists, #3964 tracks provisioning follow-on
    RFC: Goal Direction Baseline (v0) #2831 Existing direction tracker; pure builder/consumer retain the RFC decision gates
    Goal Instance Identity and Orphan Recovery (v0) #5206 New · coordinate with active implementation/PRs
    RFC: Goal-scoped Capability Portfolio and Connector Lifecycle (v0) #5207 New · coordinate with active implementation/PRs
    RFC: Per-Goal Usage, Token, and Cost Surfacing v0 #3163 Capture/aggregation foundations exist; provider coverage/readback work active
    DSH / Pi: L1 Observation and Managed Runtime Selection #5208 New · coordinate with active implementation/PRs
    RFC: Hierarchical Agent Stride Control v0 #3203 Observation foundation shipped; adaptive selection remains research qualification
    RFC: Human Attention Wishlist v0 #3179 Existing implementation task; preserve non-blocking attention semantics
    RFC: Human-confirmed domain operations (v0) #5209 New · coordinate with active implementation/PRs
    RFC: Intelligent Review and Dynamic Presentation Surfaces v0 #3786 Action/delivery-review verticals exist; shared disclosure and cross-channel work remain
    RFC: Live Team Workspace v0 #5210 New · coordinate with active implementation/PRs
    RFC: Long-Horizon Harness Benchmark and Research Program v0 #3243 Existing evidence program; no duplicate live benchmark launches
    RFC: Long-Running Agent Reliability Diagnostics and Governed Delivery v0 #5211 New · bounded contribution available
    LoopX Overall Roadmap v0: Product, Collaboration, Technology and Delivery #4574 Portfolio coordination; this index is a projection, not another roadmap
    Manager runtime profile v0 #5212 New · coordinate with active implementation/PRs
    RFC: Monorepo Distribution Split (v0) #5072 M0 module-count gate shipped in #5191; M1+ and distribution decisions remain
    RFC: Obelisk Session Evidence Provider v0 #3792 Optional package exists; qualify current adapter and scoped integration, not installation by default
    RFC: Agent Judgment and Optional Independent Assessment — Jev as a Candidate (v0) #5213 New · decision brief first; experiments/adoption not yet authorized
    RFC: Post-Outcome Memory Utility Attribution v0 #3214 Stage 1 exists; reuse #3824 / PR #3829 and current attribution work
    RFC: Provider-Neutral Post-Writeback Capability Hooks v0 #3479 Existing hook/report continuation task; preserve post-commit isolation
    RFC: Provider-Neutral Turn-Start Inbox Hook v0 #3844 Implemented opt-in; closed #3844 records read-obligation repair, no new backlog invented
    RFC: Research Exploration Control Plane v0 #5214 New · bounded contribution available
    RFC: Semantic Vocabulary Convergence and Commit-Time Drift Checks (v0) #4447 Existing convergence tracker; #4743 separately owns coin-time decisions
    RFC: Shared Goal Alignment and Governed Amendment Protocol (v0) #3836 Existing alignment/amendment tracker; retained proposals are not canonical commits
    RFC: LoopX Shared Control-Plane Authority and Pluggable State Providers (v0) #3245 Native TS authority substantially delivered; recovery/cutover/qualification remain active
    RFC: Single-Owner Local Daemon (v0) #3930 Existing daemon proposal; identity/readiness precedes supervised migration
    RFC: TypeScript Control-Plane Migration Direction v0 #3225 Whole-transaction replacement and bounded retirement active; no duplicate Python owner

    Use the RFC for design and acceptance, the linked task for milestone/PR evidence, and current execution ownership for assignment. Update the linked record when a boundary changes; do not infer completion from PR counts, stale “not implemented” text, or a design merge. A demonstrated complete scope can close with evidence instead of manufacturing follow-up work.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    direction/architecture-evolutionArchitecture evolution and research-incubator work.direction/operator-surface-imOperator surfaces, frontend control plane, and bounded IM integration.direction/shared-coordinationShared-goal authority and cross-host coordination contracts.enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions