By name: find-issues description: Search linq2db/linq2db for issues and PRs matching a topic or the content of an existing ticket. Read-only. Topic mode takes a free-form query and returns ranked candidates. Ticket mode takes an issue/PR number or URL, derives search terms from its title/body/labels, excludes self, and returns candidates with a duplicate verdict. In ticket mode when the verdict is "likely duplicate", suggests /merge-duplicates as follow-up.
find-issues
User-triggered read-only lookup of existing issues and PRs on linq2db/linq2db.
Two modes:
- Topic mode — user provides a free-form description (e.g. "FB5 IF EXISTS support", "DuckDB provider").
- Ticket mode — user provides an issue or PR number/URL; skill derives search context from its title, body, and labels, then excludes the target from results.
Write actions (closing, merging, commenting) are out of scope — see /merge-duplicates for that.
When to run
Only when the user explicitly invokes this skill. Typical prompts:
- "search for issues about X"
- "find duplicates of #N"
- "is this a dupe of anything?"
Steps
1. Classify input
| Input | Mode |
|---|---|
All digits (5483) or #5483 |
ticket (issue number) |
https://github.com/<owner>/<repo>/(issues|pull)/<n> |
ticket (URL — parse owner/repo/n) |
| Anything else | topic |
If a URL's <owner>/<repo> ≠ linq2db/linq2db, flag the mismatch and confirm with user before proceeding.
Ambiguity (input is a single word that could be a topic or a typoed number): ask.
2. Ticket mode — fetch target
gh issue view <n> --repo linq2db/linq2db \
--json number,title,body,labels,state,author,url,closedAt
Note: this endpoint works for PRs too (GitHub models PRs as issues at the API level), but if the URL is /pull/<n>, use gh pr view <n> instead to also capture isDraft, baseRefName, headRefName, and closing-issue references (Fixes #…, Closes #…). Surface any closing references in the output — a PR with Fixes #5412 isn't a dupe of #5412, it's the fix.
Truncate body to ~1000 chars for term extraction.
3. Extract search terms and run searches
Follow the discipline in .claude/docs/issue-search.md:
- Term extraction (step 1 of that doc)
- Confirm terms with user
- Parallel search strategies A–C (topic mode) or A–D (ticket mode with labels)
- Dedup + rank (cap 10)
- Self-exclude (ticket mode)
Use parallel Bash calls for the independent gh invocations.
4. Present results
Compact table per issue-search.md step 5.
Then, depending on mode:
Topic mode — just the table plus:
Want to file a new issue covering this? →
/create-issue
Ticket mode — the table plus a verdict:
| Verdict | When |
|---|---|
| likely duplicate | At least one open issue with ≥2 matching terms in title, same provider/area labels if present, and substantive body |
| related but distinct | Candidates share keywords but title/scope diverges (e.g. "SQL Server IDENTITY bug" vs "SQL Server IDENTITY in TPT mapping") |
| no duplicates found | No candidates survive ranking, or all candidates are clearly unrelated |
If verdict is likely duplicate:
Looks like duplicate of #
. Run /merge-duplicates <canonical> <target>to consolidate and close.
The skill does NOT invoke /merge-duplicates itself — the user does.
5. Interactive follow-ups
After presenting results, accept:
show <n>— fetch full body and comments of candidate<n>:gh issue view <n> --repo linq2db/linq2db --json body,commentsrefine <new terms>— re-run step 3 with adjusted termsdone— exit
6. Do not
- Post, edit, or close any issue or comment. This skill is read-only.
- Search repos other than
linq2db/linq2dbunless the user's ticket URL explicitly names a different one and they've confirmed. - Invent a verdict when candidates are ambiguous — say "uncertain" and list evidence.