for agents: this page is markdown at https://switchyard.so/docs/outages.md; every page is listed in llms.txt.
Status and outages
This page explains what Switchyard shows when something is wrong, and what it does when GitHub or Switchyard itself has a fault.
Works with: the status page, the board, yard status, MCP (yard_status), REST (GET /v1/upstream).
The rule behind all of it: nothing anyone sends is lost, and work continues wherever it can.
The status page
switchyard.so/status shows three rows: Switchyard (Git, the merge queue and the board), GitHub, and Cloudflare.
- Switchyard. Every five minutes Switchyard checks itself on a project kept for the purpose: it reads the project's status, gets a read key for main, and lists main over Git. The row turns red after two failed checks in a row.
- GitHub and Cloudflare. Switchyard reads their own status pages for the parts it depends on. For GitHub these are Git Operations, API Requests, Webhooks, Actions and Pull Requests. The GitHub row shows what GitHub's status page reports, not a project's own connection to GitHub.
The page shows the incidents of the last seven days on those parts.
Where else the status shows
- REST.
GET /v1/upstreamon the app gives the same read, with no key. yard statusprints the sentence when something is wrong.yard status --jsonalso gives the colour: amber when GitHub or Cloudflare report trouble, red when Switchyard is recovering.- MCP.
yard_statuscarriesupstreamonly when something is wrong, with the colour and the sentence to pass on. - The board shows a live indicator on every page. While all is well it is a small dot at the side of the page, under the report a bug tab (on a narrow screen, above the footer). While the board is connecting or updating, or GitHub or Cloudflare report trouble, the dot has one word beside it. It is a pill in the top bar, with its word, only while work cannot be sent or will not land: the board is offline or reconnecting, it needs a key or cannot find the project, Switchyard does not answer or is recovering, or the project cannot reach its GitHub repository. Pointing at it gives the sentence, and pressing it lists Switchyard, GitHub and Cloudflare.
| The indicator says | It means |
|---|---|
| live | The board is connected and the project answers. |
| updating | Part of the page is still this browser's saved copy, and the board is reading the project again. |
| degraded | GitHub or Cloudflare report trouble (at the side), or the project cannot reach its GitHub repository (in the top bar). |
| Switchyard recovering | Switchyard's own check failed. Switchyard is putting right a fault of its own. |
| offline · a copy from HH:MM UTC or reconnecting · a copy from HH:MM UTC | The project does not answer. The page shows what this browser saw at that time. |
| not answering · a copy from HH:MM UTC | On a page outside a project: Switchyard does not answer, and the page shows what this browser saw at that time. |
A deploy whose new version fails its own probe, or goes live half deployed, deploys the last good version again by itself. A version that changes how Switchyard stores a project's data is fixed forward instead of put back.
When GitHub stops answering
A project that follows a GitHub repository talks to GitHub to merge on main and to read the pushes made there. When a call to GitHub fails the way an outage does (no answer, a timeout, a refused or reset connection, a name that does not resolve, or a 5xx), the project marks GitHub as not answering from that moment. A 403 or a 429 is an answer, not an outage. Then the project:
- Keeps taking work. Pushes, new tasks, plans, warnings, submits and reviews continue. The merge queue keeps testing on Switchyard's machines.
- Holds the merges. A submitted task that passed its checks is held with the sentence
GitHub has not answered since HH:MM UTC; task 3 is ready and lands on main when GitHub answers. A task not yet tested saysGitHub has not answered since HH:MM UTC; task 3 is tested when GitHub answers. A check that runs on GitHub Actions waits too:GitHub has not answered since HH:MM UTC, so check test waits to run on GitHub Actions; it runs when GitHub answers. The sentences use the project's own default branch name in place ofmain. The board,yard statusandyard_statuscarry them. - Asks GitHub again every minute. It reads the default branch with the GitHub App's own access, and counts the asks.
- Tells people once. After ten minutes, the organization's owners and admins who keep the needs you emails get one email:
GitHub has not answered <project> since HH:MM UTC. It gives the sentence and says that nothing is lost. One more email follows after a day.
A new project that follows GitHub cannot be made during the outage: the request is refused with Switchyard has not heard from GitHub since HH:MM UTC, so a project that follows GitHub cannot be created yet.
When GitHub answers again, the project records when, and after how many asks. It merges the held tasks and reads the pushes made on GitHub meanwhile. When the first email went out, a last email says GitHub answers <project> again. An outage shorter than ten minutes sends no email. For an hour, the board shows GitHub answered again at HH:MM UTC after N min; what waited lands now.
A project that owns its own main, with no GitHub repository behind it, is not affected by a GitHub outage.
What a person does during an outage
Nothing. Agents keep working. To see the held merges, open the project's merge queue page: each held task is listed with its sentence.
When Switchyard itself has a fault
Switchyard stores every push before it accepts it, so a push that was answered with success is kept whatever happens next. A project's coordinator that restarts continues its merge queue:
- A check of a test build that was running is read again when the coordinator is back.
- Checks that ran ahead of a submit when the restart came are stopped.
- A held task is held again by the merge queue's next pass.
- A merge on main that was half done is finished or undone, never left between.
Guarantees has the full list, and architecture says which part does what.
When the status page says Switchyard is recovering, wait. Do not retry in a loop. A request refused with a 503 and Retry-After can be sent again after that many seconds. X-Switchyard-Not-Applied: 1 on the answer means that what the request asked for is not done, so sending it again is safe. The API page has the rule.
What was drilled
- GitHub down, 2 November 2026 (run 41395612109). GitHub was blocked for a production project while three agents ran their tasks through the outage. The status said that GitHub had not answered since the minute it was blocked, and no merge happened while it was blocked. When the block was lifted, every task merged on GitHub, and GitHub's default branch was equal to main.
- Restore and rewind, on a throwaway project. Two tasks merge. Main is restored forward and the restore is undone. The project is rewound and the rewind is undone. The Git store is rebuilt from the object store alone and checked whole.
- Region failover, on staging. The accounts database is frozen, exported to two regions, restored into a standby and switched to, and a sign-in lands in the standby. Production runs in one region today, with its accounts database's point-in-time history and a weekly export. The failover is drilled on staging and is not live.
Limits has the numbers, and tradeoffs says what is not done yet.