Fork and pr

Fork and PR

Walk a contributor through the standard "fork + branch + commit + push + PR" flow against a GitHub repo they don't have write access to. The user knows what they want to change; this skill codifies the boring gh/git syntax around it so they don't have to remember it.

When this applies

Trigger when the user wants to contribute upward to a repo they don't own. Common phrasings: "I want to make a PR to X", "contribute a fix to X", "fork and PR", "open a PR against the upstream X repo", or right after a git push fails with 403 on a clone of someone else's repo.

Skip when:

• The user owns the repo (just git push + gh pr create – no fork needed).
• The user is hitting a SAML SSO 403 on gh itself (Resource protected by organization SAML enforcement). That's a one-time per-org token authorization at https://github.com/settings/tokens, not a per-PR step. Tell them to grant SSO on their gh CLI token and retry; don't try to drive that through this skill.
• The contribution is a stacked-diff or multi-PR series (this skill assumes one branch → one PR).

Inputs from the user

Usually one or more of:

• Upstream repo (owner/name or full URL, e.g. SalesforceCommerceCloud/plugin_sitemap).
• What they want to change (a one-line intent, used to suggest a branch name).
• Where they're starting from – fresh, or already inside a clone.

If the upstream repo isn't named, ask. If the intent isn't named, ask before step 3 (branch naming) – not before steps 1–2.

Flow

The skill runs as five steps with a deliberate pause between steps 3 and 4. Steps 1–3 set up the workspace; the user makes their edits and commits; steps 4–5 publish.

Step 1: State check

Figure out which of four starting states the user is in. Run, in the directory the user is operating from:

git rev-parse --show-toplevel 2>/dev/null
git remote -v 2>/dev/null
gh repo view <upstream> --json viewerPermission,parent 2>&1

Interpret:

• Not in a git repo (git rev-parse fails) → state A: no clone yet.
• In a git repo, origin points at upstream → state B: cloned upstream directly. Will need to swap remotes during fork.
• In a git repo, origin points at user's fork, upstream points at the target → state D: fully set up. Skip to step 3.
• gh repo view shows viewerPermission: ADMIN/WRITE → user owns it or already has push access. Skill doesn't apply – tell them so and stop.
• gh repo view errors with Resource protected by organization SAML enforcement → SSO grant missing. Stop and tell them to authorize at https://github.com/settings/tokens; don't try to fork through it.

Step 2: Fork (if needed)

For state A or B, run:

gh repo fork <upstream> --clone --remote

--clone clones the fork into the current directory if state A; if state B (already in a clone of upstream), gh swaps origin to point at the new fork and renames the existing remote to upstream automatically. --remote ensures the rename happens.

For state C (fork on GitHub but not cloned locally), gh repo clone <user>/<repo> then git remote add upstream <upstream>.

For state D, no-op.

After this step, git remote -v should show origin = user's fork, upstream = the target repo. Confirm before proceeding.

Step 3: Branch

Suggest a branch name based on the user's stated intent. Conventions vary by upstream repo, but reasonable defaults:

• fix/<short-slug> for bug fixes
• feat/<short-slug> for features
• chore/<short-slug> for maintenance/dependency/cleanup
• docs/<short-slug> for docs-only changes

If the upstream repo has a CONTRIBUTING.md with a different convention, follow that instead. Run:

git checkout -b <branch-name>

Step 4: PAUSE — user edits and commits

This is the handoff. Tell the user explicitly:

"Branch <name> is ready. Make your edits, commit them (git add <files> then git commit), and tell me when you're done – I'll push and open the PR."

Then stop. Do not poll git status or proactively check on their progress; wait for them to come back with "done", "committed", "ready", or similar.

While paused, if the user asks code questions or wants help with the change itself, that's a different task – help them, but don't push or open the PR until they say they're done.

Step 5: Push and open PR

When the user confirms they've committed:

git push -u origin <branch-name>
gh pr create --repo <upstream> --web

--web opens the PR creation page in the browser with the branch and base pre-filled. This is preferable to authoring title/body on the CLI for first-time contributions to a repo, because the upstream may have a PR template (.github/PULL_REQUEST_TEMPLATE.md) that only renders in the web UI.

If the user explicitly wants to skip the browser, drop --web and gh will prompt for title and body inline – or pass --title / --body directly if the user already knows what they want.

Disambiguation

• User says "make a PR" but is in a repo they own: skip the fork; just push and gh pr create. Confirm by checking gh repo view --json viewerPermission.
• User has multiple GitHub accounts: gh auth status will show which one is active. The fork lands under the active account. If they want it under a different account, gh auth switch first.
• Upstream uses a non-main default branch (master, develop): gh pr create auto-detects it. Don't hardcode main anywhere.
• User already has an old fork that's behind upstream: gh repo fork is idempotent (it'll reuse the existing fork) but won't sync it. If their fork is stale and they want a clean branch off current upstream, git fetch upstream && git checkout -b <branch> upstream/<default-branch> before step 4.

What this skill doesn't do

• Authenticate gh (gh auth login – one-time setup, not per-PR).
• Grant SAML SSO on gh tokens (per-org, browser-driven, one-time).
• Resolve merge conflicts on rebase/merge.
• Sync a stale fork with upstream beyond a single fresh branch.
• Author the actual commit message or PR body content – that's the user's call.

Style

• Cite commands the user can run, not internal paths.
• One step per assistant turn during the active phases (1, 2, 3, 5). The pause at step 4 is the only multi-turn boundary.
• After step 5, return the PR URL gh printed and stop. Don't ask follow-ups.