Skip to content

CI Hosted Cutover

Chris Dare edited this page Nov 15, 2026 · 1 revision

Proposed hosted-only CI cutover for CI1.07

Status, 2026-11-08. Hosted routing is on main for every event. At the owner's direction, the four repository runner registrations (ids 27–30) were deleted on 2026-11-08 while all four were idle and no run was queued. The runner API now reports total_count 0. The owner's runners now serve a private mirror instead. That is the separate, owner-approved boundary the last paragraph of this plan requires; see the owner lane. The step-4 negative test (a disposable PR requesting an old self-hosted label set) has not been run. The rest of this file is the plan as written.

This is a staged plan for #1436, not a record of an applied routing change. The public repository currently has four registered Ubuntu runners. Its pull_request and merge_group builds are hosted, while main push and manual builds use those persistent runners. A pull request can change the workflow file used by its own run, including runs-on; the expression in ci.yml cannot keep that PR away from the repository runners. The existing fork approval policy and read-only default token reduce other risks but do not enforce runner placement. The services also share one Unix account, so their separate build directories are not a hostile-code boundary.

The proposed end state is one hosted build job for every ci.yml event, with no self-hosted runner registered to the public repository. The required aggregate check remains named ci, and strict up-to-date branch protection remains enabled. This end state gives up persistent-main and manual-branch builds. It does not by itself prove that the required ci result reflects an unmodified gate inventory; #1434 owns candidate and gate provenance.

Admission before rollout

The owner must explicitly select this routing and its rollback, including the disposition of #1436's self-hosted parity, physical-host pressure and cancellation acceptance items and its #1434/#1435 blockers. The loop may prepare and review this patch, but its authority does not include repository settings or runner registration changes. Do not merge the routing patch as an unattended PR while the four registrations remain available.

Before the maintenance window, obtain four same-commit reviews of the patch. Run the PR workflow on its updated synthetic merge commit; record the PR head, merge commit, workflow blob, required ci result, and build runner name. Confirm that the build, roadmap, and ci jobs still appear once each and that ci remains the strict required context. The PR's hosted result is a preflight of the new YAML, not an enforcement test while runners still exist.

Coordinated rollout

  1. Disable auto-merge for the cutover PR. Pause other merges and manual dispatches, and let current main/manual runner jobs finish. Record their final run IDs. An already-running job must not be killed to establish the boundary.
  2. Stop the four runner services and verify they are offline. Remove all four repository runner registrations using the owner-controlled GitHub runner settings and host removal procedure. Check the repository runner API until total_count is zero. Stopping a service alone leaves its registration and can be reversed accidentally; the zero-registration check is required. During this short interval the old main workflow cannot obtain a build runner, so keep merges and dispatches paused.
  3. Mark the reviewed PR ready, re-check its exact head and required ci, and merge through normal strict protection. Do not bypass a check or use an administrative merge. The resulting main push must run build, roadmap, and ci on hosted runners. Record the main SHA and run ID, confirm the workflow blob is the reviewed one, and compare the tested PR merge tree with the published tree.
  4. Resume merges and manual dispatch only after the main ci succeeds. Confirm again that the repository runner count is zero. For a bounded negative test, submit a disposable PR whose only workflow change requests one of the old self-hosted label sets. Verify it obtains no runner, cancel the queued run, and close the test PR. Do not infer isolation merely from an offline service or a successful benign PR.

If the PR becomes stale or its content changes, get a fresh required run on the updated head. A changed routing diff also needs a fresh four-role review. Keep the exact run, job, runner, tree, and registration evidence with #1436.

Rollback and failure

Keep the public repository runners deregistered. Repair or revert the hosted workflow through a reviewed PR that still uses hosted runners and produces the required ci result. Do not restore repository self-hosted registrations or the old runs-on labels as a rollback: that would reopen the PR-to-runner path. If hosted CI is unavailable, leave merges paused until a hosted repair can pass strict protection. A future self-hosted design requires a separate provider-enforced runner boundary and its own owner-approved rollout.

This cutover does not close #1436 until the owner has accepted the changed routing scope and the exact PR/main and negative-path evidence exists. The three earlier equal-tree, warm-cache hosted/self-hosted observations remain in routing-parity-2026-09-27.md; they do not substitute for the cutover checks.


Exported from docs/ci/hosted-cutover.md at the commit that removed it from the tree.

Home

Architecture

CI and runners

Decision records

Roadmaps

Clone this wiki locally