sentence-to-issueSAFE

Sentence to Issue — 一句话提 GitHub issue. Evidence, privacy scan, gh.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: sentence-to-issue
description: >-
  Use when the user says "file an issue", "report this bug to the repo",
  "one sentence issue", or the Chinese equivalents "一句话提 GitHub issue" /
  "帮我把这个 bug 传上去" - turn one sentence of description into a filed GitHub
  issue on any public repository. Gathers evidence, privacy-checks the draft,
  writes the report, submits with the gh CLI, then reads it back to verify.
  Works for any software or game, not only AI tools.
license: MIT
---

# Sentence to Issue · 一句话提 GitHub issue

Turn one sentence about something broken into a filed GitHub issue — and hand the user the link.

The user is not a developer by default. Explain in plain language, keep it short, and never
assume they know GitHub vocabulary.

---

## Language rules (hard)

1. **Reply in the language the user writes in.** Chinese in → Chinese out. English in → English out.
2. **The issue body follows the target repository's language** — check its README / recent issues.
   English open-source project → English body. Default to English when unsure.
3. **This file and `references/` are for you, not for the user.** Never paste them into the chat.
   Lines marked `ZH:` / `EN:` are user-facing scripts — pick the one matching the user's language.

---

## The flow (12 steps)

### 1. Parse the one sentence
Extract three things: **what broke**, **which repository**, **what evidence the user already has**.
Ask only for what is missing. Never guess the repository.

### 2. Resolve the target repository — three sources
1. `git remote get-url origin` in the current working directory
2. The user's own words: an `owner/repo`, or a full `https://github.com/...` link
3. **Explicit repository URLs in the project's own config** — `package.json`'s `repository`
   field, `pyproject.toml`'s `repository` / `homepage`. These name a repository outright.
   A `LICENSE` copyright line names at most an **owner**, never a repository: it may
   *confirm* a repo you already locked onto, it must never *choose* one
   (`Never guess the repository`, Step 1).

There is **no keyword search** — if none of the three yields a repository, ask the user for it.
Anything resolved from source 3 is confirmed with the user before it is used to file.

### 3. Capability probe
Inventory what you can actually reach: terminal, file read, browser, screenshots, installed skills.
That decides how deep you can capture (see `references/evidence.md`). Chat window only → layer L0.
**Never invent evidence you did not read.**

### 4. Freeze the scene — say this first, in the user's language
> **ZH:** **正在搜集相关信息 —— 期间别刷新、别重启、别更新、别关窗口。** 这期间如果你有与 bug 相关的细节要补充,直接发给我就行。
> **EN:** **Gathering evidence — please don't refresh, restart, update, or close anything.** If you have more detail about the bug to add meanwhile, just send it over.

That line replaces any "when did it happen / what were you doing" interrogation.
This applies to **any software or game**, not only AI tools — as long as the user's
environment is reachable and they grant the permission.

### 5. Capture evidence (L0 → L3)
Follow `references/evidence.md`. Read-only commands only: no writes, no restarts, no config edits.
Anything you could not reach gets marked `not captured` in the report — never filled in from memory.

### 6. Deduplicate
```bash
gh issue list --repo <owner/repo> --state all --search "<keywords>"
```
A hit → ask the user whether to add a comment to the existing issue instead of filing a new one.

### 7. Check the target repository's issue templates
```bash
gh api repos/<owner/repo>/contents/.github/ISSUE_TEMPLATE --jq '.[].name'
```
- **Template fits** → fetch it and fold its headings into your draft:
  ```bash
  gh api repos/<owner/repo>/contents/.github/ISSUE_TEMPLATE/<file> --jq .content | base64 -d
  ```
  Fill those headings from your evidence, then file with `--body-file` **only**.
  ⚠️ `gh issue create` **rejects** `--template` combined with `--body`/`--body-file`
  (`--template is not supported when using --body or --body-file`) — merge first, then file.
  Filing through `gh` also skips web-only required fields (forms that force a log upload).
- **Template listing quirks** (all seen in the wild):
  - `config.yml` appears in the listing but is **not a template** (`blank_issues_enabled` +
    `contact_links`) — skip it.
  - `.yml` issue **forms** are not markdown: field labels live in `body[].attributes.label`.
    There is no heading to fold — read them as a field list instead.
  - `.md` templates carry YAML frontmatter (`--- name/about/title/labels/assignees ---`).
    Strip it first, or you will send `labels: ''` and friends verbatim.
- **Template demands uploading logs/debug output** → **ask the user first**:

> **ZH:** 这个仓库的模板要求附日志。**要我帮你把日志附上吗?**
> · 日志文件没法用命令行直接传;我可以把相关片段**贴进正文**(你在网页上传也行)——
>   发出去后凭链接的人都能看到,贴之前我会先扫一遍敏感信息。
> · 不附也行,我照常提 issue,正文里标 `logs not attached`,维护者需要时会再找你。
> 你决定。
>
> **EN:** This repo's template asks for logs. **Want me to include them?**
> · Log files cannot be attached from the CLI — I can paste the relevant excerpt into the
>   body (or you can upload it on the web). It becomes public to anyone holding the link,
>   and I'll scan it for secrets first.
> · Or skip it: I'll file with `logs not attached` and maintainers can ask later.
> Your call.

`--attach` takes **images and video only** — a log file cannot ride along that way.
Scan any excerpt for secrets **before** it goes in. If the user declines, proceed without them.

### 8. Privacy check — reminder, not a veto
Run `scripts/privacy_scan.py <draft-file>`. Two levels (see `references/privacy.md`):

| Level | Examples | What you do |
|---|---|---|
| **1 — identity / location / history** | username, local paths, chat logs, ids | Warn **once**, continue on yes |
| **2 — secrets** | password, API key, token, private key | **Second confirmation**: state the consequence, then wait for an explicit reply |

Content the user pasted themselves gets the same warnings.
**You never upload

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (1)

| **2 — secrets** | things that unlock access | passwords, API keys, tokens, session cookies, private keys | **Second confirmation**: state the consequence first, then wait for an explicit reply. |

Gates applied: no_behavioural_pass.