githubrepos

Picking GitHub repositories and cloning one

internal/githubrepos lists the GitHub repositories the signed-in account can reach, so the person can pick some, and clones a picked repository into a folder they can change. It is the only place that lists the account's repositories; the GitHub calls are github.Client.ViewerRepositories and gitlocal.Clone, both with the account's GitHub token (config.ResolveGitHubToken: explicit token, else the saved device sign-in).

Listing

ViewerRepositories reads one viewer.topRepositories page ordered by latest push, plus the accessible totalCount from viewer.repositories in the same request. The first page has eight nodes, matching the visible rows. Once it returns, one background request asks for one hundred nodes after its cursor. Later pages of one hundred load only when asked for (more=1, below). Names already returned are skipped. Archives are excluded after each response. The top list is cached for five minutes; when it reloads, the earlier extra rows stay listed after the new first page until the new extra page replaces them, so the list never shrinks to eight rows and grows back.

Search calls GitHub's search(type: REPOSITORY, first: 100) directly. Queries search repository names and include forks but exclude archives. GitHub search covers all of GitHub, so a search is limited to repositories the person has a permission on, read in the same request as the top list: every repository of the viewer's login and of each organization it belongs to (viewer.organizations(first: 100)), as user: qualifiers, and each repository it collaborates on elsewhere (viewer.repositories(affiliations: [COLLABORATOR], first: 100), most recently pushed first), as exact repo: qualifiers. GitHub combines the qualifiers with OR; at most 300 are sent. The owner of a collaboration, or of a top repository the viewer only contributed to, is never searched as a whole: the viewer has no permission on that owner's other repositories. Top repositories still match from the cache. With nothing in scope a search returns nothing rather than searching all of GitHub. Owners a query names (owner/name, or user:, org: or repo: qualifiers) narrow the scope to theirs and their names become search terms, so acme/wid searches wid among acme's repositories in scope, and a stranger's owner/name finds nothing. No accessible-repository enumeration runs beyond those first hundred collaborations. A search made before the top list exists loads its first page first, since the scope comes from it.

GET /api/github/repositories?query=&refresh=1&more=1 answers

{"fetched_at": "2026-07-01T10:00:00Z", "total_count": 147, "searching": true,
 "has_more": true,
 "repositories": [{"repository": "owner/name", "private": false, "top": true,
   "fork": false, "pushed_at": "2026-07-01T09:00:00Z", "description": "",
   "picked": true, "clone": "/work/code/name"}]}

A non-empty query returns matching cached top rows immediately while its GitHub search runs in the background. The response also includes top_repositories, the full cached top list, so the picker can filter it on each keystroke before any network request or debounce. On completion the search results merge with matching top rows without duplicates. An empty search preserves the order returned by topRepositories. Search sorting is exact repository name or owner/name match first, then membership in the top list, then latest push time; names break remaining ties. Search results are cached for five minutes by normalized query (case-insensitive, trimmed and collapsed whitespace), with at most 64 cached queries. Concurrent requests for one query share a request.

more=1 starts loading the next page of the current listing in the background, unless one is loading or there is none: one hundred more top repositories after the last cursor for an empty query, or the search's next page of one hundred (with the owners it started with) for a query. New rows are appended without duplicates. has_more says another page exists. total_count is the accessible, non-archived total for an empty query; for a query it counts every match: the rows returned plus the search results not loaded yet (GitHub returns at most 1,000 per search). A failed next page keeps the loaded rows and reports search_error; refresh=1 starts the listing again from its first page.

searching: true means a search or the extra top page is still pending. Cached rows remain pickable. A failed request with cached matches returns those rows and a sanitized search_error; with no matches it uses the HTTP errors below. refresh=1 retries the current query or refreshes the top list when the query is empty. The native relay's named retry-list action uses this option, and its more-list action sends more=1. Account changes cancel background requests and discard both caches; late responses cannot refill them. Shutdown cancels background requests and waits for them. picked comes from the registry; clone is an available registered path or "". earlier_names are the registry's recorded earlier repository names (see internal/repos).

A failure answers a JSON {"code","message"} (never a token or an upstream body) with the codes the app already explains for the GitHub connection:

A request GitHub refused is classified by the sign-in's runtime health (config.GitHubAuth.Health), which the GraphQL client feeds from every response.

Cloning

A clone state is {"repository","dir","state","error"?,"code"?}, state being running, done or failed; error is sanitized text. code: "url_rewritten" means the person's Git configuration changes the requested clone URL, including to another repository on GitHub or to SSH. State is in memory.

Default folder. The most common parent directory of the registered clones that lie below the home directory (ties: flat layout first, then alphabetical), joined with the repository name. If clones follow <root>/<owner>/<name> (the parent folder is named after the registered owner), that layout is kept for the new owner. With no such clone, ~/code/<owner>/<name>. While that path exists, -2, -3, ... is appended to the name.

Starting. The body is strict JSON (an empty body means the default folder). dir must be absolute, inside the home directory but not the home directory itself, and either missing or an empty directory. A symlink directly under home (such as ~/code on another disk) is supported; deeper symlinks cannot lead outside home or that code folder. A symlinked home works by either path. The repository is picked if it is not (asking for a clone chooses it, and it stays picked if the clone fails). The clone runs in the background under the daemon's context, at most one per repository and with no overlapping target folders (including symlink aliases and case variants), with gitlocal.Clone's 30 minute timeout. On success the clone is registered with Registry.AddFor(dir, repository, nil) and the app is told the registry changed. New clones are refused during shutdown. A failed or cancelled clone removes its own target and any empty parent folders it created; a target that already existed is left for git to clean up.

gitlocal.Clone builds its URL only from a validated owner/name (gitlocal.GitHubCloneURL), passes the token in git's private environment only, and disables hooks, templates, fsmonitor, submodules and every protocol but HTTPS. Certificate verification is forced on with a token. Git URL rewrites that change the requested URL are refused before creating folders. Configured checkout filters such as git-lfs still run and may contact their own storage hosts. See SECURITY.md.