Pr Local Preview Maintainer

PR Local Preview Maintainer

Run pull requests in a disposable local worktree and give the user a stable preview link. The point is to let the user review the PR while protecting their current working tree and making UI changes visible.

Default Workflow

1. Preserve the current workspace

Start by checking the repository state:

bashCopy code
git status --short --branch
gh pr view <number> --json number,title,state,headRefName,headRepositoryOwner,url,isCrossRepository

If the current worktree has local changes, do not run gh pr checkout there. Create or reuse a sibling worktree instead. This avoids overwriting unrelated work and keeps the PR preview easy to remove later.

Use a predictable path:

textCopy code
<repo-parent>/<repo-name>-pr-<number>

For this repository, PR 8 should become:

textCopy code
/Users/zeeland/projects/rudder-oss-pr-8

If the directory already exists, inspect it before reusing it:

bashCopy code
git -C <worktree-path> status --short --branch
git -C <worktree-path> branch --show-current

If it is dirty, ask only if the dirty changes conflict with the user's goal. Otherwise, pick a suffixed path such as <repo-name>-pr-<number>-preview.

2. Check out the PR in the worktree

Create the worktree from the remote base, then check out the PR branch inside it:

bashCopy code
git worktree add --detach <worktree-path> origin/main
gh pr checkout <number> --branch zeelandc/pr-<number>

If main is not the base branch, use the PR's actual base ref. Prefer a local branch name that makes the preview purpose clear and follows the user's branch prefix convention when the repository defines one.

After checkout, verify:

bashCopy code
git status --short --branch

3. Install dependencies only when needed

Check whether dependencies are present:

bashCopy code
test -d node_modules && echo node_modules-present || echo node_modules-missing

If missing, install them from the worktree:

bashCopy code
pnpm install

Report install warnings only when they affect the preview. Routine peer or bin link warnings can be mentioned briefly.

4. Choose isolated runtime settings

Before starting, inspect occupied ports and existing local runtimes:

bashCopy code
lsof -nP -iTCP:<candidate-port> -sTCP:LISTEN

For Rudder, prefer:

bashCopy code
RUDDER_INSTANCE_ID=pr-<number>
PORT=<free-api-port>
RUDDER_EMBEDDED_POSTGRES_PORT=<free-pg-port>

Use 3100 only if it is free and the user clearly wants the default instance. If another Rudder dev runtime is already running, leave it alone and use a separate port and instance ID.

Example:

bashCopy code
env RUDDER_INSTANCE_ID=pr-8 PORT=3118 RUDDER_EMBEDDED_POSTGRES_PORT=54358 pnpm dev

5. Start the preview so it survives hand-off

Use tmux when available. It keeps the preview alive after the assistant turn ends and gives the user a simple way to inspect logs.

bashCopy code
tmux new-session -d -s rudder-pr-<number> -c <worktree-path> \
  'env RUDDER_INSTANCE_ID=pr-<number> PORT=<port> RUDDER_EMBEDDED_POSTGRES_PORT=<pg-port> pnpm dev'

If tmux is not available, use another managed background process and capture logs to /tmp/<repo-name>-pr-<number>.log. Do not leave an opaque process running without telling the user how to stop it.

6. Wait for readiness

Poll the health endpoint or the app's equivalent readiness check:

bashCopy code
curl -fsS http://127.0.0.1:<port>/api/health

If the app has no health endpoint, use the first meaningful route and check for a successful response. For frontend-only apps, load the dev server URL and confirm it serves HTML.

When startup fails, inspect the session logs before returning:

bashCopy code
tmux capture-pane -pt rudder-pr-<number> -S -160

Fix straightforward startup issues yourself when they are local environment problems, such as missing dependencies or port conflicts. Do not edit PR source code unless the user asked for a fix.

7. Decide whether UI screenshots are required

Screenshots are required when any of these are true:

• the PR title, files, or user request mentions UI, layout, pages, components, CSS, design, screenshots, visual review, browser behavior, or interaction
• changed files live under ui/, desktop/, frontend route files, component directories, stylesheets, or public assets
• the user says they want to "see" the effect

Use Browser Use for local browser inspection when available. Capture the important screen or flow after the preview is running. Store temporary screenshots outside the repository, for example:

textCopy code
/tmp/rudder-pr-<number>-<view>.png

For non-UI backend/API PRs, a health check plus a task-specific API smoke check is usually enough. If the user still needs to review manually, provide the URL.

8. Hand-off format

Keep the final response short and concrete:

textCopy code
PR <number> is running in <worktree-path>.

Open: http://127.0.0.1:<port>
Logs: tmux attach -t rudder-pr-<number>
Stop: tmux kill-session -t rudder-pr-<number>

Health check passed: <brief status>
Screenshots: <paths or embedded images when UI changed>

If screenshots were required but could not be captured, say exactly why and what verification did run.

Cleanup Requests

When the user asks to close, stop, or clean up the preview, stop only the PR preview session and verify its port is free:

bashCopy code
tmux kill-session -t rudder-pr-<number>
lsof -nP -iTCP:<port> -sTCP:LISTEN

Do not remove the worktree unless the user explicitly asks for cleanup. If they do ask to remove it, verify the worktree has no uncommitted changes first.

Safety Notes

• Never use destructive Git commands in the user's main worktree to make a PR checkout succeed.
• Do not stop unrelated Rudder dev sessions just because the default port is in use; choose an isolated port instead.
• Keep PR preview data isolated with a dedicated RUDDER_INSTANCE_ID.
• If source edits become necessary, pause and state that the task has shifted from previewing a PR to fixing it.