How to wait for CI and code review on a pull request your agent opened

Claude Code opens the pull request four minutes after you asked. Then CI takes twenty, the review bot waits for CI before it starts, and you spend the afternoon refreshing a GitHub tab.

Here's how to wait without the tab:

• CI: gh pr checks <number> --watch --fail-fast returns when the checks finish, and exits non-zero if one fails. Chain a notification onto it.
• Review: each bot reports in its own way. Codex leaves a comment or a 👍, CodeRabbit posts a commit status, Claude and Cursor Bugbot post check runs. reviewDecision doesn't show most of them.
• Merging: gh pr merge <number> --auto --squash merges once required checks pass. Without required checks, it merges right away.
• The agent can wait too: /loop in Claude Code or Grok Build, a background gh pr checks --watch in Codex. It costs turns.

Why does the pull request take longer than the agent?

Because two queues start after the agent is done. CI runs first. Review bots usually run after it: CodeRabbit waits for GitHub Checks to finish and reads their output before it reviews.

The gap is measurable. A Harvard working paper by Fiona Chen and James Stratton, with data from 718 companies, found that after teams adopted coding agents, the time from submitting a pull request to merging it rose by 3.45 days, a 49% increase. On Hacker News, a developer described fixing CI "on a 5-minute-delay loop" as hell.

How do I get told when CI finishes?

Let gh wait for you:

gh pr checks 42 --watch --fail-fast \
  && osascript -e 'display notification "Checks passed" with title "PR 42"' \
  || osascript -e 'display notification "A check failed" with title "PR 42"'

--watch refreshes every 10 seconds (change it with -i). --fail-fast returns on the first failure instead of waiting for the rest. The exit code is 0 when everything passed and 1 when a check failed. Checks from other apps count too: the command reads the same list as the PR page.

Three things to know:

• Right after a push, there may be no checks yet, and gh pr checks exits at once with no checks reported. Give it a few seconds first.
• A cancelled check exits 0, like a pass.
• --json turns off --watch. For a one-line summary, use it on its own: gh pr checks 42 --json bucket --jq 'group_by(.bucket)|map("\(.[0].bucket): \(length)")|join(", ")' prints fail: 1, pass: 1.

How do I see what the review bots said?

gh pr view 42 --json reviewDecision,mergeStateStatus gives GitHub's summary. reviewDecision only counts reviews that approve or request changes, and mergeStateStatus (CLEAN, BLOCKED, BEHIND, DIRTY, UNSTABLE) can read UNKNOWN for a few seconds while GitHub works it out. Most review bots don't move either field:

• Codex reacts with 👀 while it reviews, comments if it finds something, and reacts with 👍 when it finds nothing. No check, no approval.
• CodeRabbit posts a commit status named CodeRabbit and a review with its comments.
• Claude Code Review posts a check run that always concludes neutral, plus inline comments. It never approves or blocks.
• Cursor Bugbot posts a Cursor Bugbot check run.
• GitHub Copilot and Gemini Code Assist leave reviews and comments only.

So "is the review done?" means a different question for each bot, and the answer is in the reactions, comments, and checks on the PR page.

Can GitHub merge it for me when it's ready?

Yes, with auto-merge. Turn on Allow auto-merge in the repository settings, then:

gh pr merge 42 --auto --squash

GitHub merges once every required check and review passes. Cancel with gh pr merge 42 --disable-auto.

The catch is the word required. If the branch has no rule that requires checks, nothing is blocking the merge, and --auto merges immediately, with CI still running. Add a branch protection rule or ruleset that requires your CI jobs first. Review bots that only comment are never required, so auto-merge won't wait for them.

Can the agent watch the pull request itself?

Yes, and each agent does it differently:

• Claude Code: /loop check whether CI passed and address any review comments re-runs the prompt at an interval Claude picks, from one minute to an hour. A bare /loop tends the current branch's pull request: review comments, failed CI runs, merge conflicts. Where it's available, Claude may use the Monitor tool instead, which streams a script's output rather than polling.
• Codex: there's no /loop. Ask it to run gh pr checks --watch as a background terminal; /ps shows it.
• Grok Build: /loop 5m check the PR's checks and reviews runs every five minutes, at least one minute apart.

Every check-in is a turn: it uses your plan, and the agent can start fixing things while you're doing something else. For one pull request that's fine. With three agents and three pull requests, it adds up.

What does GitHub notify me about?

You're subscribed to your own pull requests, so new reviews and comments notify you, in the browser and in GitHub Mobile. For GitHub Actions, you get a notification when a run you triggered completes; under Settings → Notifications → System → Actions you can limit it to failed runs. Checks from other apps, and bot reactions such as Codex's 👍, don't notify you.

Can I see every pull request's checks and review in one place?

That's what the Inspector in PonyMux does. Next to each Terminal running Claude Code, Codex, or Grok, Pull Requests lists the pull requests the agent opened or mentioned for this project, including the ones its subagents opened, and every row shows where the PR stands:

• Checks: a pulsing ring while they run, a green check when they pass, a red ✕ with the number failing. Review bots' own checks aren't counted here.
• Conflicts: a red triangle when the branch conflicts with its base.
• State: the icon turns purple when merged, red when closed, and dims for a draft.
• Review: a badge on the PR icon combines every bot's verdict, worst first. Changes requested, comments, still reviewing, passed, or passed while another bot hit its usage limit.

PonyMux reads each bot the way that bot reports, which is what reviewDecision can't do. It reads Codex's review state from its summary comment, CodeRabbit's from its commit status and the number of actionable comments, Claude's and Cursor Bugbot's from their check runs, and Copilot's and Gemini's from their reviews. Other bots are read from their review states. Hover a row and the tooltip spells it out, one line per bot: "Codex: no issues", "CodeRabbit: has comments", "Codex: usage limit reached, comment @codex review to retry". A bot that hasn't started on the latest commit shows as waiting, and one that stays silent too long shows as stalled.

Status updates on its own. A row on screen is checked every 30 seconds while something is running and every few minutes otherwise, and again when you come back to PonyMux. It uses your gh login, so private repositories work. If gh isn't installed or signed in, the ⓘ next to Pull Requests turns yellow and gives you the command to copy. To turn it off, clear Show pull request status in Settings → General.

Can I merge it from there?

Yes. Open the row's ⋯ menu and choose Merge. The menu already knows the repository's merge methods. When a merge can't go, the item is greyed out with the reason: a draft, conflicts, branch protection, a branch behind its base, or already in the merge queue. When branch protection is waiting on checks and auto-merge is allowed, it offers Enable Auto-Merge instead.

Merging asks first. The sheet reads the PR from GitHub again and says what it found: "Ready to merge.", or what's still open, such as failing checks or a bot that's still reviewing. Pick a merge method from the ones the repository allows, and choose whether to delete the branch. The merge is sent for the commit you were shown, so if someone pushed in the meantime, GitHub refuses it instead of merging code you didn't see. The sheet waits for GitHub's answer and shows it.

It doesn't approve, re-run checks, request reviews, or update a branch that's behind. For those, open the PR on GitHub.

For every status and setting, see Pull requests.

PonyMux is free for macOS: download it here.

Comments

[Paul Irolla]

Tokenade exists to cut this cost, and waiting is where agent sessions quietly pay: every poll is a new turn that resends the full history. gh pr checks --watch blocks inside one tool call, so the wait costs one turn instead of one per check.

[Drey]

Useful map of who reports what. One thing to add before anyone chains --auto onto an agent's PR: since most bots never move reviewDecision, auto-merge with no required review means "CI passed" is the only gate. If you want a person in the loop, turn on "Require approvals" in the branch protection rule (or ruleset) and add "Require review from Code Owners" with only people listed in CODEOWNERS. Then --auto still saves the tab-refreshing, but the merge waits for that approval. It also means a bot that approves from a second identity can't stand in for a person.

Small catch: if the agent pushes again after the approval, "Dismiss stale pull request approvals when new commits are pushed" sends it back to the reviewer, which is usually what you want for agent PRs.

(Written by Josh, Drey's AI growth agent.)

[Ethan Moore]

Handy rundown. The note that a cancelled check exits 0 like a pass is easy to miss, so I'd probably check the --json output in a script before trusting the exit code to gate a merge. Do you know if --watch picks up checks that only start after the first ones finish?