skill-origins — read a skill's first record from public git history
Registries list a skill under the repository it was installed from. When that repository is a collection with many contributors, the person who first recorded the skill is not shown. The git history of the collection still holds that first record. This skill reads it.
Source: https://github.com/shimo4228/skill-origins
Which question is being asked
| The user asks | Run |
|---|---|
| Where did this skill come from? Who first recorded it? | who |
| Which skills in this collection did I first record? (a maker asking about their own work) | mine |
who — the first record of one skill
uvx --from git+https://github.com/shimo4228/skill-origins skill-origins who <owner/repo> <skill-name-or-path>
<owner/repo>is the repository the registry shows as the skill's source.- Add
--swhto also ask Software Heritage whether that exact file content is archived there; it printsarchivedornot answered. - Exit code 2 with a list of paths: the name matched several files; run again with the full path.
- Exit code 3: GitHub did not answer, or the repository is too large to search by name; retry, or
pass the full path to the
SKILL.md.
Report the answer in this shape, using the values as printed (when no GitHub login could be linked, the tool prints the commit author's name instead; report that name. When it prints that the name is an e-mail address and is not shown, say the first commit has no linked login and leave the person unnamed):
The registry lists
<owner/repo>. The earliest commit that adds this file there is<sha>by<login or author name>on<date>(https://github.com/<owner/repo>/commit/<sha>).
mine — skills in one collection whose first record is a given login
uvx --from git+https://github.com/shimo4228/skill-origins skill-origins mine <github-login> --repo <owner/repo>
This clones the collection into ~/.cache/skill-origins, reads its history and lists the skills
whose first record is that login, each with whether the file names the login. A commit that could
not be linked to a GitHub login is not matched; the tool prints how many such files the collection
has, and that line belongs in the report. It covers one collection per run: ask the user which collection they contributed to, and run it once for each.
Use it for the login of the person asking. The project publishes counts and keeps no list of
accounts, so a request to list what some other person first recorded is answered one skill at a
time with who.
What the record does and does not show
The record shows who made the first commit of the file in that repository. It does not show who wrote the text, and a difference between the listed repository and the first recorder is how contributed collections normally look. The tool reads the file as it is now, so it cannot tell a name never written into the file from one removed later. Keep the report to what the history records.
State these limits when they bear on the answer:
- Commit dates are written by the committing client.
- A bulk import or a squashed merge records the person who imported.
- Renames are followed only where git detects them, so the date is an upper bound on the file's first appearance in that repository.
- In an organization-owned repository, the login alone does not tell whether the first recorder belongs to the owner.
When the tool cannot run
Without uv, a who answer for a known full path comes from two calls (replace the placeholders).
This route follows no renames, so it gives the first commit at the current path:
gh api "repos/<owner>/<repo>/commits?path=<path/to/SKILL.md>&per_page=1" -i | grep -i '^link:'
gh api "repos/<owner>/<repo>/commits?path=<path/to/SKILL.md>&per_page=1&page=<last>" --jq '.[0] | {sha, login: .author.login, name: .commit.author.name, date: .commit.author.date}'
<last> is the page number in the Link header's rel="last" entry. When the first call prints
no Link line, the path has a single commit: run the second call with page=1.