界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
设计与创意 · SKILL.md

lingtai-issue-report

lingtai-issue-report

【设计与创意】Protocol router for reporting bugs, stale info, missing capabilities, or design issues in any LingTai skill, capability, preset, or system behavior. Enter through this router, then load the one nested reference you need — evidence-checklist, report-template, or filing-flow. You always assemble a structured report and ask the human for permission befo…

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

Llingtai-ai / lingtaiApache-2.0

技能说明

GitHub

name: lingtai-issue-report description: Protocol router for reporting bugs, stale info, missing capabilities, or design issues in any LingTai skill, capability, preset, or system behavior. Enter through this router, then load the one nested reference you need — evidence-checklist, report-template, or filing-flow. You always assemble a structured report and ask the human for permission before filing. version: 1.4.0 last_changed_at: "2026-09-05T00:00:00Z" maintenance: "If you find stale or incorrect information here, use the lingtai-issue-report skill to assemble evidence and obtain per-issue human consent before filing an issue. Never include secrets, credentials, tokens, or private paths."

Reporting LingTai Issues

This is a reference router. You use the LingTai system as a real user all day — its skills, capabilities, and procedures — so you are uniquely positioned to notice problems humans miss. When you notice something wrong, surface it. This skill is the protocol: enter through this router, pick the one nested reference matching where you are in the report lifecycle, and read that leaf for the full procedure.

The non-negotiables (read before anything else)

Two rules hold across every path and every leaf:

  1. Human consent is required, always. You never open a GitHub issue without an explicit "yes" from the human. The human is the accountable owner of what gets filed under their name. Even if gh is authenticated and you have a shell, per-issue consent is non-negotiable. If they decline, drop it — no nagging, no auto-retry.
  2. Secrets never enter a report. No tokens, keys, or passwords in the body, in chat, in logs, or in files. A human-provided GH_TOKEN stays in the env of the single command that needs it. Redact before you quote.

The nested references elaborate these; they never weaken them.

Nested reference catalog

- name: issue-report-evidence-checklist
  location: reference/evidence-checklist/SKILL.md
  description: When an observation is worth reporting (and when it isn't), what evidence to capture verbatim, and how to keep secrets out of the report.
- name: issue-report-report-template
  location: reference/report-template/SKILL.md
  description: The report skeleton — subject/title, the structured body sections, sending it via mail to your parent and the human, and which repo to target.
- name: issue-report-filing-flow
  location: reference/filing-flow/SKILL.md
  description: The filing decision — human consent boundary, the read-only gh probe, Path A (direct gh filing) and Path B (paste-ready handoff), token hygiene, and proactive surfacing.

Routing table

You will usually move through all three in order, but read one leaf at a time — don't pull the whole protocol into context at once.

Where you areReadWhat it covers
Just noticed somethingreference/evidence-checklist/SKILL.mdWhether it's report-worthy, the evidence to capture before it's gone, scrubbing secrets
Ready to write it upreference/report-template/SKILL.mdThe body/title structure, and mailing it to your parent and the human
Time to filereference/filing-flow/SKILL.mdThe read-only gh probe, the human's permission, then Path A filing or Path B handoff

Found a bug or issue? If you encounter any problems with this skill, load the lingtai-issue-report skill and follow its instructions to report it.

我的收藏