CI

Hosted CI is manual-only. The workflow has no push, pull-request, or schedule triggers: a branch push, a PR (including into production), a tag, or a schedule never starts a hosted runner. The production workflow is also disabled at the repository level (paused manually on October 14, 2026 per the CI cost policy), so all hosted builds are paused until Edi re-enables it and dispatches it himself. See releasing for the release side of the same pause.

Home CI

.github/workflows/home-ci.yml runs the full gate (npm ci, build, check, test) on an isolated self-hosted runner for each push to main or production and for manual runs. It uses no hosted minutes. Because this repository is public, the runner is guarded in three places:

The workflow has a read-only token, uses no secrets, does not keep checkout credentials, and never prints the environment or host details, because its logs are public. scripts/tests/home-ci-policy.test.mjs enforces that and runs as the job's first step. Pull requests from forks get no home CI.

Validation therefore happens locally on every change:

npm run build
npm run check
npm test

npm test runs the suite under a private temporary folder and a private home folder (HOME is <folder>/home and CODEX_HOME is <folder>/home/.codex, overriding an inherited Codex home, so execution hosts keep nothing in ~/.local/state/gw-exec), so the real ones are left as they were. At most eight test files run concurrently because each can start several CLI processes. It also fails a test that hangs (120 s by default for a test with no timeout of its own), lets a test file exit once its tests have ended, and stops every test process when the run is signalled: the test processes exit on the signal, which stops the Copilot CLIs they started (those run in process groups of their own), and the runner removes the stale sockets its run linked to in /tmp/codex-daemon-<uid>, the one place the pinned SDK fixes outside both folders. Every test process, and every process a test starts with its environment, refuses the real per-user roots (gw-exec, gw-native, gw-ssh, the SDK rendezvous and a linked native Codex home) unless they lie in a temporary folder (src/test-root-guard.ts), so a test file run on its own with node --test fails instead of writing there; run single files through node scripts/run-tests.mjs <file>. As a backstop the runner notes the entries of the real ~/.local/state/gw-native and gw-exec before the run and fails it, naming the files, when a new entry's owner.json points its configuration or native home into a temporary folder; it never removes anything there. The tests that need macOS retain their explicit platform guards.

Hosted jobs (opt-in only)

Every job in .github/workflows/ci.yml carries a job-level guard requiring workflow_dispatch on refs/heads/production with the typed boolean run_build input set to true. The input defaults to false, so a plain dispatch runs nothing. Promotion to production happens before a build is requested; historical tags and old workflow definitions must never be dispatched.

workflow_dispatch on refs/heads/production
├── run_build: true                      → verify (Ubuntu 24.04 ARM)
└── run_build: true + run_native: true   → macos-native (macOS 15 ARM)

• verify runs the unchanged source gate on a fresh hosted machine: npm ci, npm run build, npm run check, the executable CI-policy guard, and npm test. Each job has a ten-minute limit, read-only repository permissions, pinned official actions, and an npm download cache.

• macos-native spends paid Apple-silicon minutes only for an intentional native validation: a dispatch from production with both hosted-build and native opt-ins enabled. A skipped job is not native proof. When selected, it runs the complete native files and ownership regressions:

  • mac-service.test.ts: native plist parsing.
  • service-plan.test.ts: managed service inventory.
  • service-tunnel.test.ts: native SSH parsing and supervisor arguments.
  • service-shared-executors.test.ts: shared service migration and lifecycle.
  • service-shared-native.test.ts: native migration and verified artifact guards.
  • native-execution-host.test.ts: shared native ownership and recovery.

  Keep any additional macOS-only test files in that job. When native validation is selected, both jobs must pass; the Linux suite alone does not establish macOS coverage. Fixtures use temporary configuration, mock providers, and owned child processes. CI does not install or activate the gateway, register desktop services, authenticate providers, publish a release, or send paid inference. Signed SDK, installed-service, and rendered desktop checks remain separate release verification described in verification and releasing.

Policy guards

Two executable guards read the actual workflow and fail if anyone widens it:

• node --test scripts/tests/ci-native-policy.test.mjs parses the YAML, evaluates every job's real gate expression against inline event scenarios (automatic events with every opt-in, wrong refs, missing/default opt-ins, the positive production opt-in) and rejects mutations such as an ungated Linux job, a new job, an extra automatic trigger, or a weakened input default. It no longer depends on an external event fixture, so the guard cannot drift from the workflow.
• ruby scripts/test-ci-cost-policy.rb parses the same YAML with the standard library and applies the same manual-only policy to every hosted job, native or not, including mutation coverage for ungated or under-gated jobs and added automatic triggers. See ci-cost-policy.

Run both locally before touching CI; neither requires a hosted run.