Create Pr

Create PR

Push the current branch to the user's fork of yugabyte/yugabyte-db and create a cross-repo GitHub pull request against the upstream using the gh CLI.

Contributors do not have push access to yugabyte/yugabyte-db. The branch must be pushed to a personal fork, and the PR is opened cross-repo from that fork into the upstream repo.

Confidentiality — read before you write anything public

The upstream repo, every GitHub issue, every PR title/description/comment, and every commit message are PUBLIC on the internet. Anything you (the agent) write into them is permanent and indexed by search engines. A leak cannot be unsent.

Never put any of the following into the GitHub issue (title/body/labels), the PR title, PR description, PR comments, commit messages, code comments, test code, test fixtures, golden output, branch names, or filenames:

• Customer-identifying data — company names, account IDs, universe/cluster names or UUIDs, support case numbers, environment names, region/zone names tied to a customer deployment. Substitute with customer-1, acme-corp, or a generic description. This includes the branch name — when you push to your fork and open the PR, the branch becomes the PR's public head ref (shown in the URL and on the PR page), so name it after the change (fix-replication-lag), never the customer (fix-acme-replication-lag); rename an existing customer-named branch with git branch -m first.
• PII — real names, real email addresses, real phone numbers, real postal addresses, real IP addresses (public or private). In tests use only the documentation ranges: IPv4 192.0.2.0/24 / 198.51.100.0/24 / 203.0.113.0/24 (RFC 5737), IPv6 2001:db8::/32 (RFC 3849), hostnames example.com / example.org / example.net (RFC 2606), names Alice/Bob/Carol, phone numbers 555-0100–555-0199.
• Unanonymized customer schema or queries — table names, column names, SQL text, query plans, or sample rows pulled from a real customer report. Reconstruct a synthetic minimal reproducer using generic identifiers (t1, users, id, value) that demonstrates the same defect.
• Secrets / credentials — API keys, tokens, passwords, TLS certificates, private keys, license keys, kubeconfigs, production S3 buckets, internal Slack/JIRA/Linear URLs, internal hostnames, vault paths. Use placeholders like AKIAIOSFODNN7EXAMPLE in mocks.
• YugabyteDB-internal information not yet public — unreleased roadmap, internal SLAs, embargoed security findings, internal infra hostnames.

This applies to the test code and test data you write too — .cc, .py, .java, .sql, golden .out files, YAML fixtures, mock responses, anything. If a customer's reproducer uses acme_orders with a customer_email column populated with real addresses, you must rewrite it as t1/email with user@example.com before the test goes anywhere near a PR.

When the source is a customer report: read the original from the internal source (JIRA / support ticket / Slack) — never paste or link it from a public artifact. Reproduce locally with synthetic inputs. Land only the synthetic reproducer. Reference the issue by internal ticket ID (PLAT-20518) only; don't quote customer-facing text.

If you're unsure whether a string is sensitive, don't write it down — ask the user. A clarifying question costs nothing; a leaked customer name in a public PR cannot be retracted.

This rule applies at every step below — when you draft the commit message in Step 1, when you draft the issue body in Step 3, when you write the title in Step 4, and when you compose the description / test plan / upgrade notes in Step 5. See src/AGENTS.md § Confidentiality for the canonical list.

Prerequisites

This skill requires the gh CLI installed and authenticated:

1. Install gh: see https://cli.github.com/.
2. Authenticate: gh auth login (or gh auth status to verify).
3. Fork: the authenticated user must have a fork of yugabyte/yugabyte-db. If they don't, gh repo fork --remote=false yugabyte/yugabyte-db creates one without mutating local remotes.

Workflow

Step 1: Commit pending changes

create-pr.sh (Step 5) refuses to run against a dirty working copy and only pushes committed changes — so commit first.

Run git status. If there are uncommitted changes:

1. Stage the relevant files (avoid git add -A/. — pick files explicitly so secrets or stray artifacts don't get included).
2. Draft a commit message based on the changes and confirm it with the user before committing.
3. Create the commit. Use git commit -F <message-file> rather than -m "..." or a quoted heredoc — apostrophes, backticks, or $ in the message will silently break a <<'EOF' ... EOF heredoc with a confusing unexpected EOF shell error. Write the message to a temp file (e.g., /tmp/claude/commit-msg-<issue>.txt) and pass it via -F.

If the branch already has at least one commit ahead of the base and the working copy is clean, skip to Step 2.

Step 1.5: PG code style pass (only when src/postgres files changed)

If the branch changes anything under src/postgres, load the pg-code-style-guide skill and check the diff against it (yb prefixes, YB marker blocks, upstream lines untouched, regress test rules). Fix violations now -- this is the last point where cleanup is free; after publishing they come back as review comments.

Commit the resulting fixes -- amend Step 1's commit or add a new one -- before Step 5; the script refuses to run against a dirty tree.

Step 2: Base branch

The PR targets master. Do not prompt the user for a base branch — create-pr.sh always rebases and pushes against master. Backports are not opened with this skill — use /backport-commit instead.

Step 3: Gather PR metadata

Infer as much as you can from context first, then batch-confirm with the user in a single prompt if all three fields below are inferable with high confidence. Only fall back to per-field prompts when something is missing or ambiguous.

1. Issue tracker reference — before prompting, look for candidates in this order:

   1. The current branch name (git branch --show-current). Branches frequently embed the issue number directly: e.g. Fix31329, fix-31151, pr/31151, or PLAT-20518-thing → #31329, #31151, PLAT-20518. Match #?\d{4,} for GitHub issues and [A-Z]+-\d+ for JIRA keys.
   2. The conversation history for any issue reference the user has mentioned (GitHub issue numbers like #31151, JIRA keys like PLAT-20518, or URLs pointing at either).
   3. Recent commit messages on the branch (git log <upstream>/master..HEAD).

   If you find a candidate, surface it to the user and ask them to confirm it's the one to use before asking them to supply a fresh one. Only prompt for a new reference if no candidate is found or the user rejects the candidate. Acceptable forms are:

   • A GitHub issue number (e.g., #31151)

   • A JIRA ticket (e.g., PLAT-20518)

   • Offer to auto-create a GitHub issue or JIRA ticket if the user doesn't have one yet. If they accept, invoke the /create-issue skill to file it (GitHub issue for core DB code, JIRA ticket for managed/ platform work), then capture the issue number / JIRA key it reports back. Since the issue tracks this PR's work, tell /create-issue to assign it to the invoker (@me) rather than leaving it unassigned.

   • Don't reuse an issue across multiple in-flight master PRs — create a fresh issue (via /create-issue) when starting unrelated work, even if a related closed/merged PR used a similar issue.

2. Component — the component tag that will appear in the title (e.g., DocDB, YSQL, YBA, CDC, xCluster). Infer from the files changed (git diff --name-only <upstream>/master..HEAD):

   • src/yb/ → DocDB
   • src/postgres/ → YSQL
   • src/yb/yql/cql/ → YCQL
   • managed/ → YBA
   • yugabyted-ui/ → yugabyted
   • docs/ → Docs
   • .claude/, AGENTS.md, CLAUDE.md → ClaudeCode
   • build-support/, cmake_modules/, .github/ → Build
   • mixed/cross-cutting → ask the user

3. Title — a short description of the change. If the branch has a single commit, use its subject (minus the [<issue>] <Component>: prefix if present). Confirm with the user.

Step 4: Build the title

Construct the title in the format:

[<#GH issue> or <JIRA>] <Component>: <Title>

Examples:

• [#31151] DocDB: Fix SamplingProfilerTest.EstimatedBytesAndCount in release
• [PLAT-20518] YBA: Fix full move edit universe to populate userIntentOverrides during configure call

Pass only the <Component>: <Title> part to create-pr.sh -t — the script prepends [<issue>] from -i automatically. The script validates both inputs and rejects invalid ones with exit 1:

• -i: must be a GH issue (#NNNN or bare NNNN) or a JIRA key (PROJECT-NNN, where PROJECT is uppercase letters only — no digits). A comma-separated list is also accepted for a change that closes several issues (e.g. 31151,#31152) — the script joins the tokens with , and renders them as [#a,#b] <Component>: <Title>. The list must be all GH or all JIRA; mixing them, or putting a space after the comma, is rejected by .github/workflows/pr-title.yml (which gates every PR that touches code — it carries paths-ignore for README.md and docs/**). Anything else is rejected by the script.
• -t: must match ^[A-Za-z][A-Za-z0-9]+:[[:space:]].+ — i.e. a Component prefix (letter-led, alphanumeric, ≥2 chars), then : , then a non-empty description. Known components: DocDB, YSQL, YCQL, YBA, CDC, xCluster, yugabyted, Docs, ClaudeCode, Build.

If the user supplies a title that already includes [<issue>] or doesn't match the format, fix it before calling the script — don't rely on the script's auto-strip (it only handles the leading [*] case) and don't surprise the user with an exit 1 after they confirmed the metadata.

Step 4b: Cut the prose the branch added

Re-read the text the branch adds — comments, architecture/ docs, agent docs — and apply AGENTS.md § Prose discipline. It is a gate here because it is easy to hold at the start of a task and gone by the end of one, and this is the last point where cutting is free.

git diff <upstream>/master...HEAD -- '*.md'   # doc prose; also skim added comments in the code diff

Land any resulting edits as a new commit before Step 5 — the script refuses to run against a dirty tree.

Step 5: Run create-pr.sh to rebase, lint, push, and open the PR

Confidentiality — final scrub before publishing. The repo and every PR are public. Before invoking the script, re-read the description, test plan, upgrade-rollback notes, the commit messages on the branch, the branch name itself (it becomes the public PR head ref), and the test code being added, and confirm none of the following appear: customer names or identifiers (universe UUIDs, account IDs, support cases, environment names, region/zone names); PII (real names / emails / phone numbers / postal addresses / IP addresses — use RFC 5737/3849 documentation ranges in tests); unanonymized customer schemas (table / column / query text / query plans / sample rows from a real customer — reconstruct a synthetic reproducer); credentials, tokens, certificates, private keys, or license keys; internal-only hostnames, URLs, Grafana/Slack/Linear links, or vault paths; unreleased internal information (roadmap, SLAs, embargoed security findings, internal infra hostnames). See the top of this skill and src/AGENTS.md § Confidentiality for the full rule. If unsure whether a string is sensitive, don't write it down — ask the user.

Once you have the issue (Step 3.1), title (Step 4), and reviewers (Step 3 if user-supplied), write the description and test plan to separate temp files and hand everything to the script.

Say why, always — a reviewer who has to reverse-engineer the motivation is the expensive case — then what changed, and whatever the reader must act on (new gflags, upgrade/rollback consequences, migration steps). Not a narration of the diff. Keep the test plan to what was actually run. See AGENTS.md § Prose discipline.

.agents/scripts/create-pr.sh \
  -i <issue> \
  -t "<Component>: <Title>" \
  -d /tmp/claude/pr-desc-<issue>.md \
  -T /tmp/claude/pr-testplan-<issue>.md \
  [-U /tmp/claude/pr-upgrade-<issue>.md] \
  [-D] \
  -r <reviewers>

The script rebases on <upstream>/master, runs lint.sh --rev <upstream>/master and refuses to push if it isn't clean, pushes to your fork, assembles the PR body as ## Summary (from -d) followed by ## Test plan (from -T), runs gh pr create, and adds reviewers via the REST requested_reviewers endpoint (which correctly routes user logins to reviewers[] and team slugs to team_reviewers[]). It auto-detects the upstream and fork remotes.

Inputs:

• -i: bare GH number (31151), #-prefixed (#31151), or a JIRA key (PLAT-20518). Pass a comma-separated list to track multiple issues in one PR (e.g. 31151, #31152); the script normalizes whitespace, prepends # to bare digits, and joins with , so the title renders as [#31151,#31152] <Component>: <Title>. The list must be all GH or all JIRA — see Step 4.
• -t: title body without the [<issue>] prefix but with the Component: prefix (e.g. DocDB: Fix flake).
• -d (required): path to a markdown file with the PR description (the "what / why" prose derived from branch commits). The script makes this the ## Summary section.
• -U (optional, sometimes required): path to a markdown file with upgrade/rollback notes. The script makes this the ## Upgrade/Rollback safety section, inserted between Summary and Test plan. Pass this whenever the branch makes an upgrade-relevant decision. The script enforces it mechanically when any .proto file changes (exit 1 if -U is missing and a *.proto diff exists) — wire-format changes have to spell out forward and backward behavior on a mixed-version cluster and what rollback looks like. Also include for (script can't detect, but the src/AGENTS.md rule expects it): gflag default flips that change observable behavior, catalog schema bumps, on-disk-format changes, RPC-versioning tweaks, migration scripts. When unsure, pass -U with a brief note rather than skipping.
• -T (required): path to a markdown file with the test plan (typically a checkbox list). Always supply this — the script errors out if -T is missing or the file is empty. If the change is trivial enough that you think no test plan applies, write an explicit one-line checkbox saying so (e.g., No runtime behavior change; visually inspected diff) and confirm with the user before proceeding.
• -r: comma-separated handles and/or team slugs (alice,bob,yugabyte/db-approvers). Optional.
• -D (optional): open the PR as a GitHub draft (gh pr create --draft). Use when the user wants to read the rendered PR before notifications fire and before reviewers are auto-pinged. After the user reviews, convert with gh pr ready <num>. Use whenever the user asks for "draft PR(s)" or says they want to inspect the PR on GitHub before sending it out. When -D is set, the script still requests reviewers via the REST endpoint as usual — GitHub silences review-requested notifications until the PR is marked ready, so the two settings combine cleanly.

Exit codes:

• 0 — PR created. Last stdout line is the PR URL.
• 2 — rebase conflict; resolve, git rebase --continue, then re-run.
• 3 — lint failed; fix as a NEW commit (do not amend a pushed commit, per src/AGENTS.md), then re-run.
• 1 — pre-flight failure (dirty tree, missing remote, etc.).

Confirm the title and body with the user before invoking the script.

Step 5b (optional): Auto-recover from trivial rebase conflicts and lint errors

If Step 5 exits 2 (rebase conflict) or 3 (lint), inspect the failure and try once to fix automatically before going back to the user:

• Rebase conflicts — only auto-resolve trivial conflicts (whitespace-only differences, adjacent-but-non-overlapping hunks, conflicts where both sides are byte-identical after whitespace normalization). For each such file, accept the resolution that preserves the branch's intent, git add it, and git rebase --continue. Anything involving renamed identifiers, signature changes, or refactors must be escalated to the user — do not guess.
• Lint errors — if the linter reports auto-fixable issues (e.g. trailing whitespace, missing newlines), apply the fix as a new commit (Fix lint), not an amend. If the errors require judgment (logic changes, unused-variable removal that might be load-bearing), escalate.

After the auto-fix, re-run Step 5 once. If it fails again, stop and surface the conflict/lint output to the user.

Step 6: Report back to the user

Output:

• The PR URL (printed by gh pr create)
• The constructed title
• The base branch the PR targets

Then clean up any temp files created during this run (e.g., /tmp/claude/commit-msg-<issue>.txt, /tmp/claude/pr-body-<issue>.md).

Notes

• Never push to yugabyte/yugabyte-db, or use gh pr create. Always use the create-pr.sh script.
• The title format is strict: [<issue>] <Component>: <Title>. Don't deviate.
• Never force-push without explicit user permission; when authorized, prefer --force-with-lease.
• CI runs automatically on GitHub PRs, so there is no trigger jenkins step (unlike the Phorge create-review skill).
• gh pr create --repo yugabyte/yugabyte-db opens the PR in the upstream repo even when the branch lives on a fork — the head: field is inferred from the tracking branch.
• gh pr edit is broken on this repo — it errors with GraphQL: Projects (classic) is being deprecated... (repository.pullRequest.projectCards). This affects --body-file, --add-reviewer, --add-label, and other post-creation edit flags. For any post-creation update to PR body / reviewers / labels, use the REST API directly:
  • Body update: jq -Rs '{body: .}' < new-body.md | gh api -X PATCH /repos/yugabyte/yugabyte-db/pulls/<num> --input -
  • Add reviewers: gh api -X POST /repos/yugabyte/yugabyte-db/pulls/<num>/requested_reviewers -f 'reviewers[]=user1' -f 'reviewers[]=user2' (and -f 'team_reviewers[]=team-slug' for org/team reviewers — strip the org/ prefix; team_reviewers takes the slug only).