CI runners
Route: /ci (sidebar: Checks › CI runners. CI runners is a Lab: switch it on under Settings → Display & sidebar → Labs to keep it in the sidebar and on the Checks grid)
Every fix Mushi Mushi opens is tested by your CI before it can merge. This page shows the machines that run those tests, the jobs waiting for one, and why a fix takes as long as it does to land. This app follows the app you have open. All apps reads every project you can see.
The page reloads every 30 seconds while the tab is visible, and Refresh does the same. Each reload asks GitHub for queued and running jobs. The server does that live read at most once a minute per app.
Spend less on CI
AI agents open a lot of pull requests, and every one runs your tests on GitHub Actions. This page keeps that bill small:
- See what you really pay. Your GitHub Actions bill per repo and per day, with public repos and free minutes shown as $0.
- Find the waste. Tests that run on drafts, schedules that run every hour, docs-only changes that run every test.
- Fix it in one pull request. Mushi Mushi opens a small workflow change for you to review, then shows the minutes saved after it merges.
Mushi Mushi’s own fix pull requests run your tests only after a review says they are worth merging.
How it reads your runs
A collector also runs every 10 minutes. It reads GitHub Actions runs updated in the last 2 hours, on every branch of each app’s repositories, classifies each job, and deletes job rows and samples older than 30 days.
Right now
The Now view opens with plain answers, in this order:
| Part | What it answers |
|---|---|
| The top line | Is CI OK? It names a machine that stopped reporting, jobs waiting a long time for a machine, and a job running slower than usual. For waiting jobs it says why: which machine takes jobs with their labels, and which idle machines are not set up for them. |
| Machines right now | Which machine runs what: one card per real machine, by name, with how busy its CPU and memory are. Jobs on GitHub’s machines show as GitHub’s cloud. Idle machines are listed on one line. |
| Running now | Each running job, the machine it is on, how long it usually takes, and about how long is left. A job type needs 3 runs before it has a usual time. |
| What broke in the last 2 hours | One group per pull request. Each job that broke names the step that failed. A summary check such as gate, which fails whenever another job fails, is listed under the job that broke. A job that failed in a re-run too is listed once. Draft pull requests that skipped CI on purpose are one line below the list, not failures. |
| Where the time goes this week | Per app: how long jobs can wait for a machine, how long a run usually takes, how often checks pass, and the main delay. The line above the table names the worst one. Two repos of one app show their repo name. |
Last 7 days follows: the week’s counts and pass rate. It does not repeat what is waiting or running now; that is the board above.
Test machine or your code
A lost runner, a job still waiting for a runner, and a queue that expired are the test machine. They are never a code failure. The page labels them Test machine stopped. Tests failed means a step failed while a machine was running your code.
Re-run a job the test machine stopped. Open the fix for a job whose tests failed. The line at the top says what needs you: a job has been waiting a long time for a machine, most failures are the test machine, some jobs failed on the code, or every finished job passed.
Queue and wait
Queue is every job, from waiting for a machine to finished.
| Column | Meaning |
|---|---|
| Waiting | Waiting for a free test machine. A long column means too few machines. |
| Running | On a test machine now. |
| Finished | Finished in the last 2 hours. Filter to Tests failed, Test machine stopped, or Passed. |
Above the queue: how many are waiting now, how many are running, the typical wait for a machine, and the typical run time. The wait also names the slow tail, the 1 job in 20 that waits longer. A bar splits the jobs from the last 7 days into passed, tests failed, test machine stopped, and cancelled, skipped or still going.
Timeline is one row per test machine. Gaps are idle time. Striped bars are jobs waiting for a machine.
Machines
Machines shows a card for every machine that runs your checks: what it runs now and how busy it is.
Runner map (advanced) is a diagram of your machines, the runners on them, the repositories they serve, and the jobs waiting or running now. Open it from the Machines view. Click a card for its health and recent jobs.
Runners that live for a single job, named as a label plus a number of seconds, are folded into that label. With Administration: read on the GitHub token, Mushi Mushi reads the repository runner list, including idle and offline runners. Without it, the page says it only sees runners through jobs. Without Actions: read, it cannot see the jobs at all.
What would make fixes land sooner
Speed it up lists changes that would land your fixes sooner, each with the evidence and a copy-ready fix.
Slow jobs is the jobs and steps that take the longest.
Stopped by the test machine lists jobs that never finished testing. Re-run them.
Your real bill
The rest of the cost section estimates money from job times. Your real bill, at the top of it, shows what GitHub actually charged, read from GitHub’s billing:
- this month’s net charge so far, in large type;
- the gross price beside it, because public repos and your plan’s included minutes are free and show up as a discount;
- the net per repository over the last 30 days, and the net per day;
- the change against the 30 days before, shown only when Mushi Mushi holds the whole earlier period;
- how much of your paid GitHub-hosted minutes ran on Mushi Mushi’s own agent branches. This share is estimated from the jobs Mushi Mushi recorded.
Mushi Mushi reads the bill once a day for each GitHub account or organization your repositories belong to. It uses only the GitHub token saved in the project’s settings, never the GitHub App. The token needs billing access:
| Account | Fine-grained token permission |
|---|---|
| Personal account | Plan: read |
| Organization | Administration: read (owners and billing managers) |
If the token can’t read the bill, the card says so and shows how to connect one. Instead of a token, you can upload the bill from a machine where gh is signed in as the paying account:
gh auth refresh -s user # once: a gh login has no billing access by default
mushi runner bill --pushA classic token, which is what gh auth login makes, reads a personal account’s bill only with the user scope. Without it, GitHub answers 404 and mushi runner bill tells you to run the refresh above.
mushi runner bill without --push prints the same bill in the terminal. Your token stays on your machine.
Cost advice that reads your workflows
Each item of Cost advice gives its evidence and a saving per month. The saving is the billed minutes Mushi Mushi recorded over the window, at GitHub’s list price. It is zero for public repos and self-hosted runners, which GitHub doesn’t charge for. An item with an Open fix PR button can open a pull request that makes the change.
| Advice | When it appears | One-click fix |
|---|---|---|
| Scheduled workflow runs too often | A schedule runs more often than weekly, and its runs billed minutes | Yes: weekly |
| Documentation-only pull requests run CI | A pull request workflow has no paths or paths-ignore filter, and pull requests that changed only docs ran it | Yes: ignore docs paths |
| Checks run again after the merge | Jobs that ran on the pull request ran again on the push to main | No: use a merge queue or release trains |
| The same check runs twice | Two jobs on the same event run the same command | No |
| Every package builds on each pull request | A turbo or nx monorepo runs all tasks on pull requests, without --affected or --filter | No. The minutes at stake are shown, but not priced. |
| Pull requests nobody merged | Minutes billed to pull requests closed without merging | No |
| macOS on pull requests | GitHub-hosted macOS jobs on pull requests, at about ten times the Linux price | No |
| Old runs keep going | A push or pull request workflow doesn’t cancel the run it replaces | Yes, when it has no concurrency group yet |
| Draft pull requests run the full CI | A pull request workflow doesn’t skip drafts | Yes: skip drafts |
Advice that needs data Mushi Mushi doesn’t have yet stays hidden. It never guesses.
Fix it in one PR
Some advice is one edit to one workflow file. Those cards have an Open fix PR button for each workflow they found:
| Fix | What changes in the workflow |
|---|---|
| Cancel superseded runs | A concurrency group. A new push to a pull request cancels the run it replaces. Pushes, manual and scheduled runs always finish. Mushi Mushi refuses on a workflow that deploys or releases. |
| Skip drafts | ready_for_review is added to the pull_request types, and every job skips draft pull requests. A job whose check your branch requires keeps running and fails at a first step named “Draft PRs run no CI”: a skipped required check reads as passing, and a draft must never merge untested. |
| Run the schedule weekly | A schedule that runs more than once a week runs once a week, at the same hour. Mushi Mushi refuses on a workflow that deploys or releases, because some daily schedules roll out releases on purpose. |
| Skip docs-only changes | paths-ignore for Markdown files, docs/, the license and issue templates on pull_request and push. Mushi Mushi refuses when the workflow produces a required check, or a job looks like one: a skipped workflow never reports, so the pull request could not merge. It also refuses when a trigger already lists paths. |
Mushi Mushi reads which checks your default branch requires (branch protection and rulesets). When it cannot, it refuses the draft and docs fixes rather than guess.
Mushi Mushi asks you to confirm first. It then edits only that file, keeps your comments, and opens a draft pull request titled ci: …. Nothing changes until you review it, mark it ready and merge it. The card shows the job’s progress and the link. If the workflow already has the fix, it says so and opens nothing.
After the merge, the card shows the saving Mushi Mushi measured. That is the workflow’s billed minutes per day after the merge, against the 14 days before. Like every number on this page, it is estimated from job times.
Writing a workflow file needs the GitHub App’s Workflows: Read and write permission, or a token with the workflow scope. Without it, the card says so. You can still copy the fix and commit it yourself.
From the terminal, run mushi runner fix --repo <owner/repo> --workflow .github/workflows/ci.yml --kind skip_drafts, and follow it with mushi runner fix-status. In your editor, use the MCP tool open_ci_cost_fix.
Apply to every app
Every pull request runs CI once, and each run costs minutes. So when a repo has several fixes, it’s cheaper to open them as one pull request. Apply to every app, next to the advice list, does that for all your apps at once:
- Preview. Mushi Mushi lists every fix it can open in the apps you may change, grouped by repo, with the estimated saving per fix and per repo. A fix whose pull request is still open, or is being opened, can’t be picked again. Mushi Mushi can’t price every fix: those say “not priced”, never $0. A repo where Mushi Mushi already knows its GitHub access can’t edit workflow files starts unticked, with the reason. Mushi Mushi knows this from the setup check or the last attempt.
- Pick and confirm. Tick the fixes you want, then confirm once. Mushi Mushi opens at most 10 repos at a time and 20 fixes per repo. Each repo counts once against its app’s 10 CI fix pull requests an hour.
- Results. Each repo gets one draft pull request with a
ci:title. The body lists every change with what it changed and why, plus the workflow’s minutes over the last 14 days. Each fix keeps the refusals above. A fix Mushi Mushi refused is listed in the pull request and in the result, never dropped. When Mushi Mushi refuses every fix in a repo, it opens no pull request there and says why.
Before opening anything, Mushi Mushi rebuilds the advice. It opens only fixes it still advises, so a stale preview can’t open something the advice no longer says.
After the merge, Mushi Mushi measures each changed workflow once, even when two fixes changed the same file. The saving isn’t counted twice.
From the terminal, run mushi runner fix --all to preview and mushi runner fix --all --apply to open. Follow each repo with mushi runner fix-status --bundle <id>. In your editor, use the MCP tools preview_ci_cost_fix_rollout, apply_ci_cost_fix_rollout and get_ci_cost_fix_bundle.
When nothing has reported in
The page stays empty until GitHub reports a job for this app’s repositories, or a machine that runs your tests reports in. On that machine, run:
mushi runner key
mushi runner hooks install --dir <runner dir>Host health arrives through mushi runner heartbeat and the job hooks, as POST /v1/ingest/runner with a runner:write key only. The setup steps are in CI runners.