One-off worker verification passed before this flip (2026-10-26):
production-shaped docker task (apparmor=docker-codex, pids-limit 512,
tenant isolation, egress proxy 8888, --docker-runtime runsc) completed
in 60s with gpt-6.1-sol, worktree quadruple + proof verified, template
and tenant repo copies clean. README now states the production state.
Inner codex sandbox stays danger-full-access by choice (bwrap
path is a separate decision); this hardens everything around it:
- DockerExecutor: --security-opt plumbing (repeatable, dedups
no-new-privileges), --pids-limit (CLI default 512, fork-bomb
guard), per-call extra_env (-e NAME=VALUE, non-secrets only,
e.g. proxy URLs; the model key still goes by -e NAME only).
- Worker: --egress-proxy-port (0 = off). When on, Docker task
containers get HTTP(S)_PROXY pointing at the host filtering
proxy via their network gateway IP (tenant network gateway
or default bridge gateway). Gateway unresolvable => the
attempt fails instead of running unproxied (fail closed,
same rule as tenant network creation failure).
- AppArmor profile deploy/apparmor/docker-codex: docker-default
base + deny network raw/packet, deny ptrace, deny mount/
umount/pivot_root, deny dangerous capabilities, deny writes
to /proc/sys, /sys, /proc/keys. Deliberately NO userns rule
(that would enable the bwrap path, not this one). File rules
mirror docker-default (broad + sensitive denies); a strict
file allowlist would need profiling codex's file access set
and is not claimed here.
- Egress proxy deploy/egress/: Tinyproxy with
FilterDefaultDeny + anchored domain allowlist (relay,
github, pypi, npm). Anchoring matters: an unanchored
'github.com' regex would also match 'evilgithub.com'.
- Seccomp honesty (README): Docker default seccomp already
enforces on all task containers; no new stricter allowlist
is claimed in this change (building one from memory risks
silently breaking the toolchain; it needs syscall profiling
first).
Tests: 3 outer-hardening tests (security-opt dedup, pids
limit, proxy env without leaking the key value, gateway
parse/fail, proxy fail-closed). Full suites pass.
Container-side isolation was project-blind: every task mounted
the shared /app/work (all worktrees + host-state), the shared
codex-home (all thread history), the shared repos dir and the
shared test-repo, and all containers shared the default bridge
network. With --tenant-isolation (Docker only; Local runs on
the host under one user and has no container boundary to
enforce), each project gets its own host dirs under
tenants_root/<project>/{work,codex-home,test-repo,<project>}:
- mounts: container paths stay identical (/app/work,
codex-home, repos/<pid>, test-repo) so host.py config needs
no change; only the host mount sources are per-project.
Static --docker-extra-mounts are replaced per task (not
appended), so a shared mount cannot leak back in.
- codex-home seeds ONLY the template's config.toml (relay
endpoint/provider names, no secret values); thread_history
is never copied (it holds other tasks' prompts/results).
test-repo is a per-project copy of the fixture (worktree
creation needs a writable repo; a shared writable fixture
would let one tenant poison the next task's base code).
- repos: repo_url checkouts go to the tenant dir and task
repo is set to the container path; no-repo_url tasks get
the tenant test-repo copy instead of the shared fixture.
- network: one Docker network per project (codex-<pid>,
created idempotently, already-exists is OK); network
creation failure fails the task instead of silently
falling back to the default network (a silent downgrade
would make isolation fail quietly).
- hardening for all Docker runs: --cap-drop=ALL and
--security-opt=no-new-privileges (non-root user was already
the case); verified live by a good-model task, not assumed.
Honest limit, kept in README: the inner codex sandbox is
still danger-full-access (host kernel blocks user namespaces,
bwrap unusable), so the boundary is container + mounts +
network, not the inner sandbox.
Tests: 7 tenant tests (seeding incl. no-history, per-project
mounts, network sanitization, per-call override, ensure
network ok/fail). Full suites pass.
Layer 1, in-worker: run_all pool thread count is now adaptive
per round (min(capacity, ready_count)) instead of always
capacity threads; ready_count counts only dependency-satisfied
queued tasks via a shared _ready_tasks() helper (pick_next
refactored onto it), so blocked tasks never inflate the pool.
Layer 2, process scaler (cloud/autoscaler.py): a systemd-managed
daemon that starts/stops cloud.worker subprocesses by ready
queue depth. desired = ceil(depth / per_worker_capacity)
clamped to [min_workers, max_workers]; scale-up is immediate,
scale-down waits scale_down_after consecutive low checks
(default 3) to avoid flapping on bursty completions. Scale-down
is LIFO (youngest worker first), SIGTERM for graceful drain
(worker run_loop drains up to drain_timeout), SIGKILL of the
process group only after drain timeout. Depth read failure
keeps the current worker count (never scale on unknown state).
Worker command forces --loop/--worker-id/--worker-capacity
(appended last so argparse last-wins); --db comes from scaler
or worker args; workers inherit scaler env (Postgres etc.).
State file (depth/desired/actual/workers) written atomically
for observability. deploy.sh now restarts codex-autoscaler
instead of the static worker unit when the scaler is active,
so the two never poll the same queue at once.
Tests: 8 autoscaler tests (desired math, ready-only depth,
cycle=0, scale up/down, cooldown streak, forced worker flags).
host