Publishing deliverables

Publish Deliverable to Feishu

Turn a project deliverable into a technical-report-styled Feishu doc under the team wiki node. Style is constrained by the ml-paper-writing skill (structure, argument, citation discipline) plus the local style spec references/technical-report-style.md (borrowed from Vale's "style-as-rules" idea).

When to use

• A stage (research / data / training / algo / deploy) produced a deliverable (方案 / 调研结果 / 实验报告 / 设计) to file to the team wiki.
• Trigger phrases: "发飞书", "publish to Feishu", "归档到 wiki".

Inputs

• deliverable type (proposal | research | experiment | design), title, and content (path or text).
• deliverable domain (declared by the owning task; one of the instance paths keys).
• target Feishu wiki node: domain-routed — project_profile.yaml -> deliverables.paths[domain] (instance value). If the domain has no mapped path, fall back to deliverables.feishu_wiki_root (= deliverables.routing.fallback) and flag it (per task_spec §2 "Domain-routed publishing"). Path partition ≠ framework partition. Domain names are whatever the instance paths defines.
• reviewer notification config: project_profile.yaml -> deliverables.notify (instance value) — send_as (user|bot), group_chat_id, roles (role → reviewer open_ids), domain_reviewers (domain → roles to @). Everything is read from instance; hardcode no id / chat / name in this skill (decoupling red line, skill_spec §9).

Procedure

1. Apply style (generation): follow the ml-paper-writing discipline to shape the deliverable into a technical report (clear structure, evidence-backed claims, verified citations); then check it against references/technical-report-style.md.
2. Build doc content (XML): title + a summary callout; use tables for comparisons/specs and <whiteboard type="mermaid"> for flows/architectures (per lark-doc rules). Cite sources; tag every data claim as 实测 / 文献 / 推测.
3. Publish via lark-doc: first read ../lark-shared/SKILL.md + the lark-doc references, then resolve the target node = paths[domain] (root fallback + flag if unmapped), then lark-cli docs +create --api-version v2 --parent-token <paths[domain]> --content "<xml>"; fill sections via block_replace if a skeleton was created first.
4. Record: write the doc URL into the owning task's flow-ledger entry and the deliverable's domain folder.
5. Notify reviewers (lark-im): after a successful publish, read deliverables.notify from the instance.
   • Resolve recipients: roles = notify.domain_reviewers[domain]; for each role gather its open_ids from notify.roles[role]. Drop roles whose list is empty (unassigned → skip, leave blank — never error). A role whose value is the literal @all → use the group "@所有人" mention.
   • Compose a message in the instance's outward-deliverable language as a post: title + 简介 + 路径（域名）
     • 链接 + an 评审： line. ⚠️ @ mentions only work in post format, NOT --markdown (markdown wraps everything in one md tag and <at…> stays literal text). Build a post payload {"<lang>":{"title":…,"content":[[…lines…]]}} where the 评审 line uses {"tag":"at","user_id":"ou_xxx"} per resolved reviewer (use "user_id":"all" for @all), interleaved with {"tag":"text","text":" "} spacers; links use {"tag":"a","text":url,"href":url}.
   • Send to notify.group_chat_id using identity notify.send_as: lark-cli im +messages-send --as <send_as> --chat-id <group_chat_id> --msg-type post --content '<post-json>' --idempotency-key <deliverable-id>. Note: --content does not accept @file — inline the JSON (e.g. shell --content "$(cat tmp.json)").
   • Identity preconditions (surface, don't fail silently): send_as: user needs scope im:message.send_as_user (if missing, emit the auth link and pause the notify, keep the publish); send_as: bot needs the app bot to be a group member (else "Bot can NOT be out of the chat").
   • Record the notification (group + @'d reviewers) in the task flow-ledger.

Rules

• Language: deliverable body language follows the project language policy (instance-owned). See references/technical-report-style.md R13, which reads the policy from the instance; if the policy assigns Chinese to outward deliverables, the polishing-chinese-prose skill is the Chinese-prose authority (its 翻译腔 band R14–R20 applies to the body). The skill hardcodes no language — it reads the instance switch.
• Style is constrained, not free-form: every published deliverable follows the style spec.
• Diagrams via native Feishu whiteboard (mermaid); structured data via tables — not walls of text.
• The wiki path is domain-routed and instance-owned (project_profile.yaml -> deliverables.paths[domain]); never hardcode a node token elsewhere. Unmapped domain → root fallback + flag.
• Publishing is outward-facing — confirm with the user before creating/overwriting when unsure.
• Notify after publish: every successful publish notifies the deliverable domain's reviewers in the team group (summary + link + @). Recipients/group/identity are instance-owned (deliverables.notify); unassigned role → skip. The notification is outward-facing too — same confirm-when-unsure rule.
• Decoupling (skill_spec §9): this body is generic; all project values (wiki paths, notify config, language policy/activation) are read from the business-repo instance at runtime. Zero hardcode.

References

• references/technical-report-style.md — generic report style rules: R1–R13 content (Vale concept) + R21–R28 presentation/scannability (Microsoft Style Guide "Scannable content" CC-BY-4.0; R26 Google dev-docs).
• Chinese-prose style is not a reference of this skill — it lives in the standalone polishing-chinese-prose skill (cited by name as the authority): 翻译腔 band R14–R20 (instance-activated) + 文牍腔 band R33–R39 (agent-layer always-on). Do not relative-path into another skill; refer to it by name.
• references/readme-style.md — entry-document style module R29+ (repo README / onboarding; standard-readme section-checklist, MIT). Physically separate from R1–R28: README rules and report-body rules do not cross-apply.