Skip to content

[Tracking] Skill library 分类后续:适用契约、来源边界与未尽事项 #264

Closed

Description

@Dong1017

Current state: All three sections are dispositioned. Section 1 was delivered by PR #263. Section 2's content arrived with Batch 3–5 (#283, #285, #282, merged 2026-11-03): applicability and recommendations gaps are closed on main 663f3b32, and the resource-layer decisions are recorded, 7 in sidecars and 8 as real mappings in the SearchSpace Matrix. The closed-allowlist proposal is decided: not adopted. Section 3 is condition-triggered with no pending request.
Decision needed: none
Next action: No assignee and no remaining item in this tracker; it is closed. Metadata delivery is not retrieval or efficacy acceptance, which stay with #255 and #253. Any later source expansion or taxonomy change opens its own proposal under section 3's conditions.

目标与边界

把 skill library 分类相关、已有公开证据支持的未尽事项集中关联,避免开发遗漏。五类 canonical owners(hardware / frameworks / repositories / methods / workflows)及导航已经存在,见 固定版本 README。目录解释本身不等于分类缺陷,本单不预设目录迁移或类别合并。

本单是关联与决策跟踪,实施仍在既有 Issue/PR,ownership 不变。待决定建议不因录入而成为已接受需求。

已交付,不重复立项

未尽事项与归属

1. methods 适用契约:归属 #91(本节已完成)

Batch 2 决策已接受冻结的 17 个 methods IDs,在来源支持下投影 input_contract、produced_artifact、failure_modes、refusal_point。建议产物不能写成已经验证的配置。

#263 已于 2026-10-31 06:42 UTC 合入(merge commit ac4c54bc,head 2e6461a0)。本单建立时记录的「15/17 未合入部分交付」已过期:合入版本交付全部 17 个选定 IDs。以下各项于 2026-11-02 对 main 2baebf6e 核实,核实方式为只读读取仓库文件,未运行生成器或测试。

  • 现有 provenance blocker 已处置:该意见指出 packaged catalog 指向的 source snapshot 058fc76b 无法解析(commit API 返回 422)。该 SHA 至今仍不可解析,但 main 上 packaged README 声明的 source snapshot 已是 8115da15,可解析,且该意见举证的 methods/training-resource-budget/SKILL.md 在此 pin 下可取得。集成/验证意见随该 PR 合入收束,不在此重复同 head review。
  • deferred 的 kernel-writing-guide 与 weight-loading-optimization 已交付:两者 metadata.json 的 applicability 均包含 scope、input_contract、produced_artifact、failure_modes、refusal_point 五项,设备要求保留在各自正文推导的条件中,未由来源项目平台猜测。
  • deferred 的 docs/knowledge-base.md field-ownership 说明已交付:该文档「Field ownership and compatibility」一节与其后的 applicability 键表记录了四个键各自的含义与禁止写法,并明确 failure_modes 不等于 limitations、applicability 键不受机械 allowlist 限制。
  • 已对 [Tracking] SearchSpace distillation coverage and knowledge consolidation #91 冻结清单核对完整范围,未把 partial delivery 标成整个 Batch 2 完成:registered methods 共 20 个,全部具备 applicability;其中 17 个具备完整四键,与冻结选择一致。仅具 scope 的 3 个为 performance-ceiling-analysis、repository-evolution-analysis、video-diffusion-optimization,即 Batch 2 决策明确排除且不触碰的条目。本节完成指 metadata 契约交付,不代表检索改善或 skill 效果。

2. 其余类别与层映射:复用 #91 原计划(本节已完成)

#91 Round 2 已列出后续 hardware、repositories/frameworks 和 layer decision 批次,冻结 ID 选择见 Round 2 Batches 3–5。

三个批次已于 2026-11-03 合入:#283(hardware,ebdbf5af)、#285(repositories/frameworks,5d16e522)、#282(resource layer,1fca842a)。

2026-11-04 对 main 663f3b32 只读核实 66 个冻结 ID 的 metadata.json,与 Round 2 记录的合入前基线对照:

批次 冻结 IDs applicability 缺失 recommendations 缺失
Batch 3 hardware 22 20 → 0 22 → 0
Batch 4 repositories/frameworks 29 4 → 0 29 → 0
Batch 5 resource layer 15 0 8 → 2

全部 66 个 ID 均可在库中定位,无缺失条目。

  • 后续批次的实际 IDs 与已交付范围已核对:缺口按上表清零。四键不完整的条目不计为缺口——Batch 2 的范围 review 明确「do not force all four keys where the body provides no basis」,故正文无依据时保留部分键是既定做法,不是未交付。Batch 5 余下 2 个无 recommendations 的条目为 repository-evolution-analysis 与 video-diffusion-optimization,属 Batch 2 明确排除的三个条目之一,保持原样。
  • 无硬件资源层条目的逐项解释已完成,且区分了 unknown、covered、supplementary:按 [Tracking] SearchSpace distillation coverage and knowledge consolidation #91 的分工约定「真实映射留在 SearchSpace Matrix,有理由的 not-applicable 放 sidecar」,15 个冻结 ID 两侧都有处置。7 个在 sidecar 的 searchspace_not_applicable 字段写明理由并指向具体 SKILL.md 章节,例如 competitor-scan 记录其产出为定性竞品比较而非硬件资源机制;其余 8 个(ascend-video-diffusion-optimization、audit-recurrent-kv-finalization、dwarfstar、husky、lily、repository-evolution-analysis、splash、video-diffusion-optimization)在 docs/searchspace-matrix.md 各有真实映射行,按该约定不应再有 sidecar 的 not-applicable 记录。未按类别虚构覆盖。
  • 闭合 applicability key allowlist 的决定已记录:不采用,保持现状的开放键集。理由有二。其一,与同一批次已接受的「正文无依据不得硬凑四键」规则相冲:字段形状按证据决定,机械闭合清单会把「依证据省略」与「违规」混为一类。其二,库中已有正当的额外键在用,例如 ascend-video-diffusion-optimization 的 hardware 与 prerequisites,闭合清单将使这些条目需要迁移或逐个豁免,而它们并无缺陷。docs/knowledge-base.md 现记录的「No mechanical allowlist restricts applicability keys」即为该决定的落地状态,无需兼容方案。四个键各自的含义与禁止写法仍然适用,本决定只是不增加机械的键名白名单。

本节的交付均为 metadata 与来源契约。检索效率、准确度与 skill 效果分别由 #255 与 #253 验收,不因本节完成而成立。

3. 来源深度与分类边界:有具体需求再决定扩展(按条件触发,当前无待办)

#260 的有界完成不代表所有底层来源均已覆盖;其保留的 native NCCL、更广 CUTLASS/DeepGEMM/FlashInfer dispatch,以及有目标条件的 UCX/ROCm 扩展,仍须按具体需求决定。

截至 2026-11-04 无已提出的来源扩展或目录调整请求,故本节不计入当前缺口。以下两项在出现具体提案时适用,提案应各自立项,不由本单持有:

  • 对拟启动的来源扩展,先核对 readsets、既有 owner 和未读边界,明确接受或暂缓;不把 backlog 自动变成全平台实施计划。
  • 若提出目录调整或类别合并,先提供具体冲突实例、影响路径及处置理由。当前不预设已接受全库迁移。

相关但独立的验收

本单不复制以上任务的完整验收清单,不重开 #260,也不启动实验。

完成口径

每项关联工作有可追溯处置:已有任务链接、明确接受/暂缓/拒绝理由,或基于具体新证据的独立提案。仅完成本单的信息整理,不代表底层实现或 efficacy 已完成。

本单的完成口径已满足:第 1、2 节各项均有可追溯处置,第 3 节两项为条件触发且当前无提案。关闭本单不改变 #91、#255、#253 的范围或验收状态。

Activity

  1. added
    area:knowledgeKnowledge line: packs, skill library, knowledge admission
    kind:trackingTracking umbrella that follows child work
    on Oct 30, 2026
  2. Dong1017 commented on Nov 4, 2026

    @Dong1017
    CollaboratorAuthor

    Closing: every associated item has a traceable disposition

    This tracker associates skill-library classification work; it never held implementation. Closing it records that each listed item now points somewhere, not that the underlying retrieval or efficacy questions are settled.

    Section 1 — methods applicability contract. #263 merged on 2026-10-31 (ac4c54bc) and delivered all 17 selected IDs, which supersedes this body's original "15/17, unmerged" record. Verified read-only on main: the provenance blocker's unresolvable 058fc76b pin is replaced by 8115da15, which resolves and serves the file that review cited; both deferred skills carry all five applicability keys; docs/knowledge-base.md documents field ownership and the key contract.

    Section 2 — remaining categories and layer mapping. Batches 3–5 merged on 2026-11-03: #283 ebdbf5af, #285 5d16e522, #282 1fca842a. Re-auditing all 66 frozen IDs' metadata.json against main 663f3b32 on 2026-11-04, compared with the pre-merge baseline recorded in Round 2:

    Batch Frozen IDs applicability absent recommendations absent
    3 hardware 22 20 → 0 22 → 0
    4 repositories/frameworks 29 4 → 0 29 → 0
    5 resource layer 15 0 8 → 2

    Two points on what is not counted as a gap. Entries with fewer than four applicability keys are not gaps: the Batch 2 scope review requires not forcing all four keys where the body gives no basis, so omission by evidence is the accepted practice. The two Batch 5 entries still without recommendations, repository-evolution-analysis and video-diffusion-optimization, are among the three IDs Batch 2 explicitly excluded.

    The resource-layer decisions are recorded on both sides of #91's split — real mappings in the matrix, reasoned not-applicable in the sidecar. Seven IDs state a reason in searchspace_not_applicable pointing at specific SKILL.md sections; the other eight each have a real mapping row in docs/searchspace-matrix.md, where a sidecar not-applicable record would be wrong.

    The closed applicability-key allowlist: not adopted. Two reasons. It conflicts with the accepted rule above — if field shape follows the evidence, a mechanical allowlist cannot tell "omitted for lack of basis" from "non-compliant". And legitimate extra keys are already in use, such as hardware and prerequisites on ascend-video-diffusion-optimization; a closed list would force those entries to migrate or be exempted one by one, for no defect. docs/knowledge-base.md's existing "No mechanical allowlist restricts applicability keys" is this decision's landed state, so no compatibility work follows. The four keys' documented meanings and prohibitions still apply; only a mechanical name whitelist is declined.

    Section 3 — source depth and classification boundaries. No source expansion or taxonomy change has been proposed. Both items are condition-triggered and stay unchecked by design; a future proposal opens its own Issue with the conflict instances and affected paths that section asks for, rather than waiting here.

    What this closure does not claim. Metadata completeness and a clear taxonomy are not retrieval improvement and not skill efficacy — those remain with #255 and #253, both open and unchanged by this closure. #91 stays open as the implementation record. No experiment, measurement or device run was performed for this disposition: the audit read repository files and the GitHub API only.

    Acting under the Issue-line delegation recorded for this tracker. Reopen if an associated item turns out to lack the disposition claimed here.

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

    area:knowledgeKnowledge line: packs, skill library, knowledge admissionkind:trackingTracking umbrella that follows child work

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions