Rendered from docs/process/decisions/0019-ci-runs-actions-from-the-actions-organization-only-and-carries-no-build-state-between-self-hosted-jobs.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

Project process. These notes record how Headwater itself is built and run. You do not need them to use Headwater. Back to the specification

CI runs actions from the actions organization only, and carries no build state between self-hosted jobs

Context

Before this record, the reasons lived in comments of .github/workflows/ci.yml on 2026-09-27. The comments were the paragraph on actions (lines 21 to 24), header rule 5 (lines 52 to 54), and the paragraph on the two runners (lines 68 to 71). Two triggers, one verdict holds the account of what each runner caches.

Decision

Actions from the actions organization only. A step uses an action only from the actions organization, which is GitHub's own. An action from any other source runs code that nobody here reviews, on every push, with a token in the environment. The bar is the organization, and not a count of actions.

No carried build state. No self-hosted job carries engine/target from an earlier job. Two attempts broke actions/checkout and then .githooks/fixtures.sh. sccache makes a cold target directory cheap.

A step reads the runner, not a guess. A step that must act differently on the two runners reads runner.environment, which is self-hosted on the container. Each pip install --user passes --break-system-packages, which the Python of the self-hosted image needs under PEP 668. pip accepts the flag where PEP 668 does not apply.

Reaffirmed on 2026-10-23 with the write cost weighed (#1617). The self-hosted pool ran 1,306 jobs in six days, and one Engine tests debug directory reached 7.5 GB. A job container wrote a mean of 5.07 GB to the NVMe over 737 containers. Containers wrote about half of all NVMe bytes, which is about 0.3 TB a day of 0.6 TB. A persistent per-slot target might halve that. The drive would then last about 3.5 years, not 2.5 to 3. The owner ruled that this gain does not justify the stale-rlib failure of #849 on the runner that gates every merge. The rule stays. A target that lives in memory and ends with its container carries no state, so it does not break this rule. It needs its own measurement first.

The runner itself is not here. The self-hosted containers are provisioned outside this repository. This record states nothing about their lifecycle, their threat model or their incident runbook.

Consequences

An action from another source needs a record that supersedes this one.

A cache that carries engine/target between self-hosted jobs is a defect under this record, even when it makes one run faster.