GitLab Runner Docker executor persistent cache reclaim via clear-docker-cache

Finding: GitLab Runner Docker executor persistent cache can accumulate on the Docker host and is reclaimable only through documented, scoped steps. Diagnosis: the docs describe three layers, a Pre-job and Post-job cache step, cache paths inside the container at /cache, and persistent storage defined by volumes. Dynamic persistent storage lives in runner-managed Docker volumes, or in host directories under runners.docker.cache_dir when that is set, and volume and directory names encode runner and project identifiers plus a hash of the cache path. Creation and growth: the documents say directories defined in volumes are persistent between builds, so unused volumes remain until removed; no growth rate is documented. Measurement: run the documented clear-docker-cache space command to list reclaimable space; no size is reported in the supplied excerpts, so measured size is UNKNOWN. Safe native cleanup: the documented clear-docker-cache script, with its default prune-volumes option, removes unused containers and volumes. The docs recommend running it on a schedule such as weekly while keeping some recent containers for performance. The script does not remove Docker images because the runner does not tag them; the documented route for unused images is docker system prune. Lost state: removed unused volumes lose cached dependencies, so the next build repopulates them, which is expected but costs time. Podman: the script falls back to the podman binary when docker is not on PATH, and must run as the user who owns Podman storage, otherwise it operates on another user's empty storage. Prevention: CONTAINER_FILTER_FLAGS, IMAGE_FILTER_FLAGS and VOLUME_FILTER_FLAGS scope which objects are pruned. Platform and version: GitLab Runner 18.4.0 and later changed the naming scheme, and cached data from earlier versions is ignored after upgrade, so older dynamic volumes may become stale. Limitations: the supplied excerpts are a bounded selection, not the full page, no issue or maintainer evidence was supplied, and no link to any specific agent workload is established.