当前维护计划在下方；折叠区仅是历史开发快照，旧版本、旧名称和未同步的当时待办不作为现行操作指令。使用规范见 README、AGENTS.md 与对应 Skill/docs，实际发布以 Git 提交和标签为准。
- 盘点所有受 Git 管理的 Markdown，区分当前指南、作者 Wiki、固定上游材料和历史记录，核查引用与旧入口。
- 对照 CLI、安装脚本、Skill、依赖锁及当前能力，修订现行说明；清理过时表述、补齐有用入口，将旧计划明确划为历史记录。
- 补充可持续执行的文档检查，覆盖根目录导航、失效路径、命令/版本/名称漂移及来源名单。
- 核对知识与 Skill 阅读路线、全量回归和安装同步，完成差异审查。收尾提交 main 并核验远端，实际发布记录以 Git 为准。
范围：本 TrainFlow checkout、任务内审查缓存与已绑定的本机六项 Skill/Python 安装。HCU 大知识库和上游固定文档只读；不执行远端任务，不把历史 release 或上游版本号误改成 TrainFlow 当前版本。
历史开发计划与当时的状态记录（保留追溯）
Scope: reorganize the bound TrainFlow checkout and its installed Skills; HCU-Knowledge is a read-only reference in this task. User authorizes commit, GitHub push and a new tag. No remote training runs or broad source refresh.
- Inventory authored/source pages, generators, evidence, references and HCU overlaps; define project ownership and stable path mappings — complete.
- Migrate official knowledge to projects//knowledgebase/{upstream,local}, preserve stable IDs/evidence; delegate HCU repository knowledge to HCU-Knowledge and retain useful operating experience — complete: 20 projects, 24 official sources; external HCU routing and integration contracts retained.
- Update collection/search/update/compatibility code, metadata, documentation and Skills; prevent excluded HCU knowledge or old paths from reappearing — complete: shared path authority, legacy maintenance mapping, owner guards, navigation and Skill guidance.
- Review migrations and lifecycle behavior; test indexing, PR/doc refresh, old workspace records, package/install and full CPU suite — complete: 1,272 passed, 4 platform skips; 5,185 knowledge pages and all reading/coverage checks pass.
- Synchronize local installation, publish reviewed main and new annotated tag, verify remote refs — installation complete and checked; v0.6.7 ready for publication with remote verification at task close.
Preserve source-reported versus reviewed versus hardware-validated boundaries. Detailed cases are only for substantial optimizations. Keep migration tooling and temporary outputs in ignored .work/wiki-layout; no deletion of original evidence merely to reduce size. Reorganization may retire redundant HCU source copies after verifying their replacement route and references.
- Audit the latest HCU-Knowledge commit, TrainFlow dependency locks, all installed Skills and three-stage usage — complete; read-only HCU audit, only demonstrated drift repaired.
- Read fixed official sources and expand every engine/library/ecosystem to common content dimensions with concrete mechanisms — complete; 22 coverage groups, 27 registered sources including a parent-pinned Loong-Megatron dependency.
- Reconcile inventories, monitored code, official documents, PR bodies/reviews and related workflow reviews — complete: all 27 registered sources checked, all 25 Git PR queues empty with no acquisition errors. GitHub throttling was resumed without dropping cursors; GraphQL discussion acquisition was compared against REST samples and incomplete connections used the existing collector. Pagination accounting was reconciled before publication.
- Cross-review authored claims, run offline/CLI/search/install checks, fix findings and synchronize the existing local installation — complete: 99 pages / 27 workflow targets have current review bindings; 1262 tests passed, 4 platform skips, plus 22 final data/integrity checks. All 5179 Wiki pages validate; representative and default search checks pass. All 11 required Skills and bindings verified; HCU-Knowledge checkout/index reused without source updates.
- Publish reviewed main and annotated v0.6.6, verify remote refs — ready; publication is the remaining operation.
No HCU-Knowledge mutation or new GPU qualification. Source acquisition, authored review and runtime validation remain distinct. Raw source caches and interim scripts stay in ignored .work/wiki-maintenance.
- Rename Skill entry, UI, dispatch and reading/update references; preserve installer migration — complete.
- Add pinned optional KernelWiki checkout and opt-in installation; keep the default required set unchanged — complete; kernel-Skill pin aligned with the already reviewed reference workflow.
- Validate routing, optional setup, source exclusion and local upgrade; synchronize installed TrainFlow and publish — complete (1234 passed, 4 platform skips; actual setup and final install checks passed). Publish the reviewed tree to main without a new tag.
No model experiments or HCU-Knowledge content refresh in this task.
The physical qualification round is sufficient for this review. Consolidate reusable decisions and evidence contracts so a fresh Agent can take over a new environment/model through the single coordinator Skill. Do not resume proxy tuning or present bounded drills as full-model convergence. Historical sections below record earlier decisions; current AGENTS.md and published workflow docs take precedence over superseded optional-dependency/naming/release notes.
- Audit requirements against the entry Skills, phase transitions, runtime, installation, Wiki and observation/delivery contracts — complete.
- Repair demonstrated gaps and simplify discovery/handoffs; preserve three independent stage modes, scientific prioritization and evidence limitations — complete.
- Run focused regressions, full CPU suite, knowledge validation, fresh-workspace/install checks and independent behavioral review — complete (1054 passed, 4 platform skips; 490 Wiki pages valid; 68/68 packaged evidence files).
- Update version/changelog and qualification boundaries, synchronize installed Skills — complete. Publish this reviewed tree to main with the explicitly requested v0.4.0-dev.1 development milestone tag; matching remote branch and tag refs are the publication receipt.
Parallel reviews covered workflow/Skill decision quality, runtime/operation state and installation/Wiki; the main Agent owned integration and final acceptance. Independent code review and a fresh-context entry exercise checked the integrated result. No new live GPU experiments were performed. Process notes remain in these three planning files; transferable operating instructions belong in docs/Skills.
Implement the first usable local-centred workflow release. Preserve the 36 requirements and staged roadmap from the approved design. Five Skills: prepare, optimize, operate, wiki search, wiki update. Real HCU/site validation remains explicit and separate from local fixture tests.
- Source/interface research, public boundary and release contracts — complete
- Task/evidence/leases, source snapshots and execution adapters — complete
- Environment/quality gates, profile analysis, experiments and monitoring — complete
- Public upstream Wiki, refresh/impact/search and workflow maintenance — complete
- Five Skills, installation, reproducible demo and tests — complete
- Integrated review, clean installation, public export, GitHub release — complete
- Keep working copies, credentials, site data and original private planning outside the published file set.
- Prefer measured evidence; all-skip, empty trace and mismatched snapshots cannot pass.
- Analysis-only, environment-only and diagnose-only are independent modes.
- Per-iteration correctness and stage-level loss validation are distinct.
- Remote watcher persists when the controller is offline; existing recovery remains single-owner.
- HCU-Knowledge is an optional read-only reference during this development.
- Original planning documents are preserved under ignored .private/planning-original; public documentation will be curated separately.
- Complete: corrected preparation guidance to reuse applicable HCU scripts and the user's deployment, with official model/training semantics as the reference.
- Complete: reviewed related workflow, Megatron topics, profiling and maintenance guidance together; Wiki integrity and all three changed Skills validate successfully.
- Inspect current TE, cuDNN Frontend and TraceLens sources/tutorials and dependency repository layout — complete.
- Add detailed official library Wiki topics, monitored source/tutorial paths, source locks and workflow impact rules — complete.
- Integrate TraceLens execution and reproducible thirdparty acquisition, including optional private HCU-Knowledge and the three Hygon kernel Skills — complete.
- Update installation/workflow guidance, validate fresh-environment behavior and real TraceLens fixture execution, review — complete. Publication is recorded in Git history.
Public repository stores reusable integration and public authored knowledge. HCU-Knowledge contents remain in an ignored checkout and access failure cannot block the public workflow installation.
- cuDNN Frontend shallow clone hit an HTTP/2 early EOF; TraceLens pack transfer stalled. Retry with HTTP/1.1 and sparse acquisition to avoid large non-code payloads.
- TE's historical user-guide URL returns 404; use the current canonical documentation root and pinned in-repository tutorial sources.
- The pinned TraceLens CLI imports an unrelated helper that sets csv.field_size_limit(sys.maxsize), overflowing Windows C long. The integration uses its supported report API in an isolated worker, retaining the upstream checkout unmodified.
- Verify account and create/reuse the public yuguo-Jack fork with upstream history — complete.
- Preserve upstream modules and commands; document HCU extension/validation boundaries, and repair only locally reproduced compatibility defects — complete.
- Point TrainFlow to the reviewed fork commit; provide explicit safe migration from the upstream checkout and retain the upstream knowledge source — complete.
- Validate native command entrypoints, report regressions, dependency migration and repository integrity — complete. Publication is recorded in both Git repositories.
User requests a unified autonomous entry Skill, persistent implement/review/correct cycles, file-based human guidance and progress, renaming prepare to adapt and operate to fault-tolerance, accurate development version/README, and no release. Existing three stage workflows plus two Wiki Skills remain callable; one coordinator entry is now explicitly requested. Real HCU validation is deferred until a site is supplied.
- Inspect TrainFlow and pinned Humanize/BBuf/Hyperloom reference implementations — complete.
- Design and implement evidence-bound iteration/review and human board contracts integrated with existing task/execution/monitor state — complete.
- Add unified Skill, rename stage Skills with safe installation migration, and rewrite workflow-first docs — complete.
- Exercise interruption/resume, stale reviews, human guidance and numerical/scale gates; review and validate; no release or new tag — complete.
- Inspect existing assignment, review and execution contracts; identify safe parallel stages — complete.
- Implement dependency-aware assignment lifecycle, bounded dispatch selection and related-agent message receipts; preserve execution uncertainty and quality gates — complete.
- Use the agreed agentic project description, five-minute guidance collection, stage/coordinator Skills and concrete parallelization guidance — complete.
- Test dependency/race/stale-result/message/resource cases, review and install Skills, push development changes without release/tag — complete locally; publication is recorded in Git history and CI.
- Dynamic operator decomposition clarification: use real model/profile operator identity and implementation scope; independent analysis and optimization proceed concurrently without an all-operator analysis barrier — complete.
- Complete HCU-Knowledge online PR/MR fallback and pinned source reading, test/review and publish v0.5.3 first — complete, pushed and tagged.
- Implement official Wiki PR discovery/read/source tracing, persistent public PR pages, full source-tree navigation and resumable source/PR update coverage — complete.
- Expand official engine documentation and source navigation with versioned evidence, deepen Megatron ecosystem; distinguish automatic source maps from authored analysis — complete.
- Add private task experience knowledge: environment/model/version identity, performance/loss evidence, successful and failed optimization milestones, retrieval and cookbook delivery provenance; integrate flow milestones — complete.
- Test offline and real public retrieval; review privacy, pagination, stale evidence and restart behavior; install Skills — complete. Development changes committed; push verification is recorded in Git remote state. No TrainFlow release/tag.
Real training and site-dependent command validation remain pending until hardware is supplied. This does not defer source-based Wiki work. HCU-Knowledge is not updated as a side effect of TrainFlow operation.
- Review execution/monitoring, flow/team state, numerical/performance gates, Wiki and experience persistence against actual source and regression tests — complete.
- Reproduce and repair concrete defects, preserving public/private boundaries and existing workflow scope — complete; independent reproduction and cross-review included.
- Run complete local tests, knowledge validation and relevant CLI/demo smoke checks; synchronize changed Skills if needed — complete: 206 tests, 485 pages/24 sources, all six installed Skills matched.
- Commit and push reviewed development changes; report readiness and site-dependent validation limits. No tag/release — validated for publication; final publication is recorded in Git remote state and CI.
- Map current stages, code gates, optimization loops, Agent coordination, remote execution, monitoring and knowledge lifecycle — complete.
- Add editable diagrams with source links and explicit implementation/site-validation boundaries; connect existing reading entries — complete.
- Render and inspect diagrams, verify links and gate mapping, publish documentation without release/tag — locally validated; publication recorded in Git history and remote state.
- Read the supplied reference implementation and current HCU knowledge — complete.
- Integrate conditional Torch-native and communication guidance into the existing optimization Skill; clarify restoration and DP scale-out — complete.
- Require HCU-Knowledge in full setup, including index and bound Skills; support reuse and explicit permission failures — complete.
- Refresh diagrams, review/test setup and Skills, synchronize local installation, commit/push without release — complete locally; publication recorded in Git history.
- Discover and qualify authorized transport, container, occupancy, tool provenance and computation/communication capabilities.
- Adapt a fixed official model proxy on current HCU software; establish component numerics, real-data baseline and checkpoint restoration.
- Exercise measured training traces, dynamic optimization tasks, private experience, independent review, monitoring and scale-out.
- Generalize and test each reproduced workflow/TraceLens defect, synchronize Skills and review public-safe changes; no release.
Site inventory, model/data definitions, raw measurements and user guidance are retained only in the external private task workspace. Environment and model gates remain incomplete until their actual coverage passes.
- Transport, immutable snapshots, bounded resource admission, training observation, multi-member health, private references and offline plots have reusable implementations and CPU regressions.
- Actual component numerics, bounded training, multiple-rank profiling, an optimization comparison, stage loss observations and collector failure/recovery have been exercised. These are separate evidence cohorts, not a blanket production pass.
- Main HCU delivery integration was qualified separately: the mixed-precision checkpoint path was repaired, and bounded save/load/continued-step checks ran. Distributed and full numerical qualification remain separate; no donor performance result is relabelled as main-branch acceptance.
- Initial public review, Skill synchronization and development commits are complete; actual clean-CI failures were repaired and the four-job follow-up matrix passed. No tag/release; unsupported automatic recovery/notification and full-size model claims stay open.
- Require reference/method/configuration/path checks, HCU knowledge and fixed-source investigation, and authorized bounded trials before accepting a discrepancy investigation as complete — implemented.
- Preserve unresolved evidence-backed gaps through environment-check even when execution succeeds or an older floor is reached; distinguish failure from missing comparison evidence — implemented with regressions.
- Exercise the sequence on a private collective diagnosis and retain per-rank configuration, correctness, physical-port counters and rollback evidence — initial A/B/A executed with per-rank numerical/exit checks; counter/source review and any additional candidate stay in private evidence.
- Review, run the complete test suite and knowledge validator, synchronize managed Skills and publish development changes without release/tag — local review, complete regression and managed Skill synchronization passed; publication is recorded in Git history and remote CI.
User requires engineering detail and real scientific optimization, not ceremony. Prioritize a single coordinator; delegate only an independent substantial task with clear expected benefit or one necessary review.
- Audit actual environment, model/parallelism/profile/operator, recovery/monitor and RLCR evidence; publish a complete private capability/validation matrix with commands, dependencies and gaps. In progress.
- Extend the generated board with evidence-backed technical sections and update rules: environment/results vs expected, model/memory/parallelism, bottleneck conclusions, per-op bounds/headroom, experiments and decisions, scale/recovery/training health, RLCR findings. Pending.
- Execute focused real-model profiling/shape-bound analysis, justified parameter and operator optimization trials on fresh idle resources; quantify numerical/performance outcomes and limitations. Pending.
- Establish actual fault-tool command/dependency ownership; exercise task-local controlled failure, diagnosis, screening where supported, checkpoint relaunch and observation; clearly separate full site capabilities from partial tests. Pending.
- Validate RLCR revise→repair→independent-review on a meaningful candidate, integrate lessons into the existing six Skills, regression/review/sync and push public-safe changes. Pending.
No automatic big knowledge-base update. No new public release. No artificial Agent roster or per-command review. Site traffic/host modification and destructive fault injection remain outside scope; controlled failure of our own test process only. User latest request authorizes onward model/workflow qualification; historical partial communication reference remains an explicit limitation, not a reason to silently stall all independent work.
- 跨任务站点知识与必要原件进入本仓 Wiki；检索、哈希、相对链接、Skill 规则及独立 review。
- donor 原执行分支完成微批 ABBA、系统/算子优化与阶段 loss，维持主仓交付延后。
- Cluster Manager 实际故障触发、任务身份停止、checkpoint 自动恢复和持续监测；CPU控制协议验证不代替此项。
- 将后续可复用调参、loss、故障结果继续沉淀到同仓站点/模型案例。
用户补充：先用 HCU Train Sim 等判断完整模型可行性，卡足优先全参，卡不足才缩 layer；资源富余时独立完整模型实例并行优化，明确分工和及时交流。已核对公开模拟器 b7d8e6f3（CLI/adapter/模板），并修正 proxy/adapt 的旧“优先缩层”措辞；协同协议沿用现有 team/lease/inbox，新增实践、更新阶段 Skill 和可编辑流程图。预测不作为硬件或全模型验证，当前 Kimi K3 模拟覆盖仍需核实，实际实验继续。
按用户要求将HCU Primus Turbo、UCCL、UltraEP、MoonEP纳入按瓶颈选择/复用/修改重编/分层验证/目标PR的链路。已检索HCU-Knowledge保留报告，明确通用与专项分支和接口边界；HCU问题先查询大知识库，同问题已取回证据可复用，不触发大库更新。训练引擎与容错项目不扩名单。
- 共享 Wiki 收紧为官方资料、可跨项目复用环境经验、模型最终优化里程碑；中间候选完整归回任务档案。
- 工作流方法迁入 docs/practices，统一所有引用、维护目标及 Skill 的收录/搜索边界。
- 更新 Skill 更名 hcu-engine-wiki-skill-update；保留旧安装的可恢复迁移，本机六个入口同步。
- 方法只保留通用触发/输入/判断/验证/回退；项目参数与临时绕过留 workspace。
- 完成独立复核修正和提交推送。真实算子及训练质量验证在任务工作区继续；不发布 release。
- 两节点16卡只读属性与管理工具复核，统一硬件基线；产品标签冲突、旧测量执行映射缺口明确保留。
- 非Adam高贡献模型8项CPU检查通过，主控接受分析阶段；每项工作量、条件效率、余量与下一动作已更新看板。完整优化没有据此关闭。
- baseline初版→HIP优化、现有Triton→Triton优化的交接与负责人验收已写通用Skill；本机三项变更Skill已同步。
- 真实CM恢复密封日志经现有解析/绘图得到3张图、9步记录，历史离线视图不进入实时监控。
- 站点硬件知识及精选6原件随公共工程；全新Store可搜读；独立review唯一来源过度断言已修正。72项相关测试和490页/25来源校验通过，无release。
本轮按用户要求将新模型/环境的完整说明整理为工程文档，更新图示并准备发布 0.5.0，不使用 dev 后缀。
- 新增
docs/project-overview.md，覆盖任务接手、环境、准确适配、容量、优化与算子迭代、质量、多 Agent、扩 DP、持续容错、交互、知识与交付。 - README 增加明显阅读入口，更新白底科技风总览图；12 张可编辑 Mermaid 补齐分支、并行分析、算子 Skill 交接和质量门槛。
- 独立文档复核并修正代理扩 DP、守护部署前提、完整 collective 采集口径；保留真实验证范围。
- 源码、包元数据、CLI、能力说明与 changelog 统一为 0.5.0；本机 editable 安装同步。
- 完成本地套件、Wiki、链接、实际 Mermaid 渲染及真实源码包检查。
发布目标：本轮变更提交至 main，推送 origin，并创建 annotated v0.5.0 tag；最终发布回执以 Git 远端 ref 核对为准。本轮不创建 GitHub Release，不执行远端训练或更新大知识库。
- 主页按环境适配、性能优化、扩 DP/持续容错三阶段展开，补齐现场检查、准确模型语义、容量、初始基线及交付。
- 说明 Wiki/Skill 内容维护与本机安装同步的分工；主页移除历史名称迁移叙述。
- 安装器全部内容和绑定验证后清理本次临时备份，失败保留；更新 Skill 明确同步收尾。
- 针对性测试、知识与链接校验，本机同步及再次幂等安装均通过。
本轮提交至 main，不移动已有 v0.5.0 tag；行为变化记入 Unreleased。
- 查询 HCU 大知识库和匹配的官方证据，复核未充分实测的排障与适配/优化衔接。
- 补充按症状分流、证据采集、假设验证、升级和恢复验收的通用诊断指南；明确方法与实测边界。
- 阶段名称改为 hcu-train-check-and-adapt / hcu-train-scale-and-stability，更新引用与安装迁移。
- 验证安装迁移、技能结构、文档与 Wiki，独立场景复核后修正。
- 同步本机 Skill，提交推送 main；不移动既有 v0.5.0。
本轮只维护 TrainFlow，不刷新 HCU 大知识库、不操作远端训练或执行故障注入。
- 固定 Hyperloom/GEAK/BBuf 可读版本，深入对照性能方法、代码与当前实现，记录采用/不采用依据。
- 结合 HCU 大库复核诊断、环境适配、性能与长稳薄弱点，完善通用方法和衔接。
- 对影响决策的程序缺口做必要修正及有意义的回归，补齐维护与使用说明。
- 独立审查、项目整体测试与 Wiki/Skill/链接验证，明确未实测边界。
- 版本统一 0.5.1，同步本机 Skill；全部检查完成，按本轮发布操作提交并 push main 和 annotated v0.5.1；不创建 GitHub Release。
仅维护 TrainFlow；不刷新 HCU 大知识库、不启动实际训练或自动升级运行依赖。用户所说 hyperloop 按本会话已明确参考的 Hyperloom 解读。
- 核对并明确 Skill / references / docs / Wiki 的职责与模式阅读路径。
- 分析与优化入口更名，更新引用、维护关联和安装迁移。
- 完善工程维护后的本机包/Skill 同步与只读校验，明确依赖范围。
- 全面回归、安装验收和审查后提交推送 main；本轮不新增 tag。
- 从CLI返回值验证所有独立模式和完整阶段，修正动态Skill名称，补行为回归。
- 明确只交付已有修改的范围，复用阶段交付规范；不增独立Skill/训练状态机模式。
- 审查状态/证据/权限/执行恢复、知识维护/安装与文档的一致性；确认问题后修复。
- 完成场景复核、完整测试、知识与资源检查及本机同步；最终提交推送结果以 Git 回执为准，不新增 tag。
范围：本地TrainFlow代码与方法审查、CPU合成场景。不开实际训练，不变更远程环境，不刷新HCU-Knowledge。
用户要求：补齐并验证 HCU Roofline/效率预测、Graph、EventReplay、Agent 分析及 HCU profiler 输入；参考 Hyperloom 与公开实战，系统复核 TrainFlow 各阶段方法，修复 Wiki/代码/Skill 漂移，最后复核安装与提交推送。
- 固定源版本、独立开发 checkout、功能/依赖/数据/实际环境准入清单。
- Roofline/效率/条件上限：真实硬件参数、算子语义与计数口径，联动热点排序、同 shape replay、优化 Skill。
- Graph/回放/Agent：最小实景验证、失败路径、输入与数值契约、可重跑分析产物。
- HCU profiler 输入：核对 hipprof/XProf/XCompute 实际格式，与 rocprofv3/pftrace 共用解析或增加明确转换，不能只改名称。
- 官方实战与 Hyperloom 方法对照；合理保留，仅对有证据缺口修订；局部 Wiki/来源/维护映射同步。
- 定向与全量回归、实际验证记录、独立复核、本机安装；TraceLens fork 已发布并锁入 TrainFlow，本次提交交付更新，不新增 tag。Git 推送结果以远端回执为准。
执行边界：不更新大 HCU 知识库。复用此前授权集群，先核对连接/镜像/占卡与残留作业，只在空闲隔离资源进行有界测试；本地代码同步到 /cfs/yuguo 下对应目录，保留原始输入和失败证据。不用模拟/CLI help 代替真实功能验收，环境不可达/依赖不支持及时反馈并继续可做的代码与离线验证。
- 清理站点正文中的会话决策、临时设施与任务待办，保留配方、实测身份及原件。
- 审查低相关官方文档，登记有理由的排除规则并整理生成页、ledger 与目录。
- 更新 Skill 和知识规则，防止后续采集恢复已排除页；冲突显式报错，不擅自删除。
- 回归同步/检索、知识与阅读资源校验、本机同步和最终差异复核。
本轮不刷新上游、不修改运行依赖或 HCU 大知识库，不启动远端任务。
- 按当前主线复核首页、工程介绍、架构、快速开始、能力边界及相关 README。
- 用 imagegen 更新白底总览图并保存提示词；同步可编辑细图和模式分流。
- 补齐 TraceLens 证据链、仅分析/仅交付边界、共享 Wiki 筛选和本机同步说明。
- 渲染全部相关 Mermaid，检查链接/章节、知识和 Skill 阅读资源，同步本机安装。
- 最终差异审查后提交推送 main；版本保持 0.5.1，不新增 tag。实际推送以 Git 回执为准。
本轮只更新介绍性文档与图片，不变更执行逻辑或上游来源，不新增硬件验证声明。
- 清除流程图及其说明中的工程版本号，核对全部介绍/导航与当前实现。
- 核对统一入口、独立模式、阶段门槛、并行交接、执行不确定性、优化证据、监控恢复及知识维护。
- 对可复现问题修复并补必要回归；保留合理流程及真实硬件验证边界。
- 完成全量测试、合成演示、知识/阅读资源/链接/图渲染及发布元数据检查。
- 完成本机安装同步与发布检查；提交推送 main 和 annotated v0.6.0 的最终结果以远端 Git 回执为准。
一名独立审查成员只读检查 flow/core/team 与交接文档；主控检查其余模块、介绍和发布。开发过程仍记在本组三文件，不更新上游或 HCU 大知识库，不启动真实训练。
- 逐段检查环境/适配、分析优化、规模长稳、Wiki 与交付的重复门槛和实验顺序。
- 修正可复现的程序性重复审核与迟报缺证据；明确初筛、候选和阶段验收的粒度。
- 将证据复用、按影响重验、渐进分析和增量记录贯彻现有 Skill/指南，保留必要质量和资源边界。
- 有针对性的回归、整体检查、本机同步与最终差异审查完成；提交推送按 Git 回执确认，本轮不新增 tag，不引入 Humanize。
只维护 TrainFlow，不刷新大知识库或启动远端实验。程序问题用本地场景验证，避免为审查增加重复硬件实验。
- 核对现有五分钟维护、缓存 pin/计划/隔离与阶段阅读链。
- 扩展显式登记的临时产物退役、容量检查与执行回执；保留现有缓存兼容性。
- 同步主控/三个阶段职责、远端与共享存储部署、checkpoint/证据保护和 README。
- 用隔离临时目录验证删除/保护/中断/大文件路径，完成全量及最终定向回归和安装检查。
- 升级 0.6.1，完成提交/推送/tag 发布准备；main 与 annotated v0.6.1 的最终结果以远端 Git 回执为准。
本轮不清理真实任务目录、不启动远端训练、不刷新大知识库。复用原维护入口；按用户补充改为阶段收尾/默认6小时低频扫描，不随五分钟交互轮询。实际自动执行由部署 policy 和既有调用者负责。
- 核对菜单元数据、实际 Skill ID、安装迁移与 Wiki/经验现有能力。
- 英文显示名对齐 ID，两个 Wiki 入口改为 hcu-trainflow-wiki-*，更新所有执行引用和迁移。
- 明确任务初期官方资料新鲜度检查、适时定向更新与最终经验筛选；补调用提醒，保留非阻塞边界。
- 补 README 多场景详细 prompt，复核知识导航、功能表述和流程图文字。
- 检查迁移/路由/检索/全量回归、同步本机安装。
- 发布收尾：提交并推送 main，远端 CI 通过后创建并推送 v0.6.2，核验远端 SHA。
不刷新上游/大知识库、不改活动训练依赖、不启动远端实验。描述性中文继续便于阅读，菜单名称使用英文标识。
- 核对 capabilities 历史、登记来源和实际搜索能力；暂不开展 Wiki 深度扩展。
- 补齐两 Wiki 入口与维护时机；明确官方 release/roadmap、SFT/RL 模型教程和底层源码的按需网页查证。
- 文档阅读资源、知识验证、链接检查通过；按现有安装同步协议收尾并提交 main，不移动既有 v0.6.2 标签。
- 核对当前三阶段、质量门槛与 HCU Probe 固定源码/公开数值指南。
- 增加共用精度诊断方法：NV 对照、HCU 自查、Probe 采集、首差定位和分布式/RL/恢复。
- 衔接主控/三阶段/更新 Skill、维护关联、阅读入口和描述性流程图。
- 详细审查质量代码、安装与阶段衔接，运行相关及全量回归、文档/图表检查。
- 同步本机安装，准备 0.6.3 元数据与发布范围；提交、推送和 annotated 标签按远端回执确认。
不启动现场训练、不刷新 HCU 大知识库；按需工具的登记和方法验证不冒充硬件实测。
- 总览概括算子/通信库复用及补实现原则，详细操作规则继续由库联动指南承载。
- 扩容、恢复与长稳观测明确精度/loss；交付明确训练引擎及相关库/工具的 PR 归属和跨仓依赖。
- 使用内置 imagegen 重绘白底、细线和低饱和总览图，同步 README、工程介绍、工作流与可编辑流程图。
- 知识、阅读入口、文档链接与 Mermaid 渲染检查通过；本机安装同步及 Git 发布结果见收尾记录。本轮不新增版本标签。
- 固定官方主线，核对权限、完整源码目录、官方文档、发布及关键 PR/review。
- 按统一标准整理 TorchTitan 定位/结构/安装/运行测试/调用链/并行优化/精度与恢复/模型及 RL 专题和精选案例。
- 注册源码、文档及 PR 更新，补导航、搜索、维护关联；逐页复核并验证增量更新可续接。
- 完成针对性检查及本机安装同步，更新版本为 0.6.4；收尾发布 main 与 annotated v0.6.4，结果以远端 Git 回执为准。
本轮只维护局部官方知识及必要工作流入口，不更新 HCU 大库、不安装上游训练依赖或启动 GPU 实验。公共仓库事实、源码审查和 HCU 实测分开表达。
- 审查本地、远端、容器的映射、命令执行、源码传输、缓存与清理边界。
- 建立共用写入约定，贯穿六个 Skill、独立模式、成员交接与 Wiki/安装维护；必要例外按已有授权处理。
- 修复源码传输目标链接可能导致的意外写入；在隔离目录验证 symlink/junction、父路径及正常传输。
- 详细复核文档与执行限制，完成相关及全量回归、本机安装同步。
- 更新版本并完成发布准备；提交推送 main 与 annotated v0.6.5 的最终结果以远端 Git 回执确认。
只维护工作流工程，不操作实际远端训练目录、不清理真实数据、不更新 HCU 大知识库。默认工作空间边界是 Agent 执行约定；不得将路径预检描述成任意子进程的文件系统沙箱。
- 审计全部来源、作者页深度、更新覆盖和 HCU 底层库归属。
- 按相同工程内容维度深化各官方仓，补调用链、配置、机制、诊断和 PR 证据。
- 全部登记源增量更新，逐源检查采集/复核/版本锁与缺口，保持 HCU 实测边界。
- 统一 README/Skill 参考路由与底层修改归属，完善长期维护检查。
- 全量测试、链接/知识/检索 review、本机同步、提交推送和 annotated v0.6.6。
授权写入：D:/codex/projects/HCU-TrainFlow；私有维护缓存 .work/wiki-maintenance；系统 Python editable 安装及 C:/Users/Administrator/.codex/skills 下既有六项 TrainFlow Skill。HCU-Knowledge 仅查询，不修改；不启动远端训练。
- Sync clean development fork to published hcu and merge reviewed latest upstream.
- Inventory every native module/CLI, inspect HCU assumptions and repair applicable gaps.
- Verify remote admission and mappings; run native regressions plus real HCU report/diff/index/source/graph/replay/collective/profiler paths where available.
- Publish precise HCU README, reproducible validation and limitations; commit/push TraceLens as yuguo-Jack.
- Inspect Hyperloom/GEAK TraceLens usage, update TrainFlow integration, workflow and affected official Wiki/routes.
- Review, relevant/full tests, dependency and installed Skill sync; prepare TrainFlow v0.6.8 publication.
Scope: TrainFlow checkout and .private/development/TraceLens (separate development repository); task evidence .work/tracelens-parity-20261013; remote /cfs/yuguo/tracelens-parity-20261013 after mount verification. User authorizes code, remote experiments, commits/pushes. Use idle devices only; do not update HCU-Knowledge; no historical author rewrite; approximate workload-specific roofline suffices, deep modeling only for unexplained limits.