Three faults from one agent-managed onboarding (env-dev-08, 2026-09-09), each of which pointed
the diagnosis away from the actual fault.
1. install.sh granted the agent's read on the ziti identity with a POSIX ACL. ziti-edge-tunnel
rewrites that file on a controller config update and the rewrite drops the ACL: the agent
applied cleanly at 01:21 and was failing every tick by 01:32. Group membership survives the
rewrite (the file stays ziti:ziti 0640), so install.sh and the package postinstall now add
monky-deployd to the `ziti` group, and a default ACL on the identity directory carries the
grant onto a newly created file. The explicit ACLs stay for the boxes that need them.
2. openziti.load() accepts an unreadable or malformed identity: the C SDK logs "configuration is
invalid" and returns a context that only fails at dial, as a bare TypeError, which the
transport reported as a missing intercept or a policy gap. The SDK transport now reads and
parses the identity itself and names the real fault first.
3. The disk pre-flight ran only when the bundle declared disk_need_bytes, so a bundle without one
died mid-pull with containerd's "no space left on device" — which reads as a registry fault.
A bundle that declares no size now has to clear the headroom floor, and the pre-flight measures
containerd's root as well as the docker data-root: docker 29 keeps image layers in the
containerd image store, and on env-dev-08 those sat on different filesystems (93 GiB free where
the agent looked, 2.8 GiB where the pull wrote).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KLB7jieMNRkTsJ2epr4Ds1
Design merged first: monky-design-docs !225 (doc 24 §4a). Pairs with monky-tenancy!40, which
copies the estate-wide read-only Harbor robot into each environment's own prefix and marks the
manifest entry `use: registry-auth`.
- That entry is not an env var (it would otherwise land in .env and therefore in every container's
environment). The agent parses it — JSON, or `username:password` with the new `registry_host` —
and writes `<state_dir>/docker/config.json` 0600 in a directory it owns, with an explicit
DOCKER_CONFIG pointing the docker CLI at it. The unit runs as monky-deployd, so a `docker login`
by a human or by root is invisible to the agent: that is what made env-dev-01 look like it had no
credential at all after the operator had just logged in.
- `compose pull` failures now carry the registry's own message ("no basic auth credentials",
"manifest unknown", DNS) into the journal and the report instead of `rc=1`.
- Tests: both credential shapes, the refusals, 0600/0700 modes, idempotent rewrite, and that the
runner never silently falls back to a human's $HOME.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KLB7jieMNRkTsJ2epr4Ds1
mdella