hermes-s6-container-supervision skill

Hermes s6-overlay Container Supervision

When to use this skill

Load this skill when you're working on:

• Adding or removing a static service in the Hermes Docker image (something that should be supervised at every container start, like the dashboard)
• Diagnosing why a per-profile gateway isn't starting, restarting, or surviving docker restart
• Understanding why the container's CMD is /opt/hermes/docker/main-wrapper.sh and how leading-dash args reach the user's program
• Modifying cont-init.d boot scripts (UID remap, volume seeding, profile reconciliation)
• Changing the rendered run-script for per-profile gateways (Phase 4)

If you're just running the Hermes Agent and want to use Docker, see website/docs/user-guide/docker.md instead.

Architecture at a glance

/init                                  ← PID 1 (s6-overlay v3.2.3.0)
├── cont-init.d                        ← oneshot setup, runs as root
│   ├── 01-hermes-setup                ← docker/stage2-hook.sh
│   │   ├── UID/GID remap
│   │   ├── chown /opt/data
│   │   ├── chown /opt/data/profiles (every boot)
│   │   ├── seed .env / config.yaml / SOUL.md
│   │   └── skills_sync.py
│   └── 02-reconcile-profiles          ← hermes_cli.container_boot
│       ├── mkdir+chown /run/hermes-service (hermes-owned gateway scandir; /run/service stays root)
│       └── walk $HERMES_HOME/profiles/<name>/gateway_state.json
│           → recreate /run/hermes-service/gateway-<name>/
│           → auto-start only those with prior_state == "running"
│
├── s6-rc.d (static services, in /etc/s6-overlay/s6-rc.d/)
│   ├── main-hermes/run                ← exec sleep infinity (no-op slot)
│   ├── dashboard/run                  ← if HERMES_DASHBOARD=1, runs `hermes dashboard`
│   └── profile-gateways/run           ← `s6-setuidgid hermes s6-svscan /run/hermes-service`
│
├── /run/hermes-service (hermes s6-svscan from s6-rc `profile-gateways` watches; tmpfs)
│   ├── gateway-coder/                 ← runtime-registered

Key files

Why Architecture B (CMD as main program, not s6-supervised)

The original plan (v1–v3) called for main hermes to run as a supervised s6-rc service. Two real s6-overlay v3 mechanics blocked that:

1. cont-init.d scripts receive no CMD args — so the stage2 hook can't parse docker run chat -q "hi" to set HERMES_ARGS for a service run script to consume.
2. /run/s6/basedir/bin/halt does NOT propagate the exit code written to /run/s6-linux-init-container-results/exitcode. Containers always exit 143 (SIGTERM) regardless. Confirmed by skarnet (s6 author) in issue #477: "if you want a container shutdown, you need to either have your CMD exit, or, if you have no CMD, write the container exit code you want then call halt".

So we use the s6-overlay-native CMD pattern via the dispatcher: ENTRYPOINT ["/opt/hermes/docker/entrypoint-dispatch.sh"], which under PID 1 exec's /init /opt/hermes/docker/main-wrapper.sh "$@". The wrapper is prepended to user args automatically — so docker run --version becomes /init main-wrapper.sh --version, and --version doesn't get intercepted by /init's POSIX shell. The wrapper drops to hermes via s6-setuidgid, then exec's the chosen program. The program's exit code becomes the container exit code, exactly matching the pre-s6 tini contract. When the entrypoint is NOT PID 1 (Fly Machines, docker run --init), the dispatcher skips /init entirely (it would abort with can only run as pid 1), restores the s6 helper PATH, runs stage2-hook.sh, and exec's main-wrapper.sh directly — no supervised services on that path (#38349).

Trade-off: main hermes is unsupervised under s6. That exactly matches its behavior under tini (the pre-s6 image). Dashboard supervision is the only new guarantee — and per-profile gateways under /run/hermes-service/ get full supervision.

Quick recipes

Verify s6 is PID 1 in a running container

docker exec <c> sh -c 'cat /proc/1/comm; readlink /proc/1/exe'
# Expect: s6-svscan or init / /package/admin/s6/.../s6-svscan

Inspect a profile gateway service

# /command/ isn't on docker-exec PATH — use absolute path
docker exec <c> /command/s6-svstat /run/hermes-service/gateway-<name>
# "up (pid …) … seconds"            → running
# "down (exitcode N) … seconds, normally up, want up, …" → s6 wants it up but the process keeps exiting (crash loop)
# "down … normally up, ready …"     → user stopped it

Bring a service up/down manually

docker exec <c> /command/s6-svc -u /run/hermes-service/gateway-<name>   # up
docker exec <c> /command/s6-svc -d /run/hermes-service/gateway-<name>   # down
docker exec <c> /command/s6-svc -t /run/hermes-service/gateway-<name>   # SIGTERM (restart)

Watch the cont-init reconciler log

docker exec <c> tail -n 50 /opt/data/logs/container-boot.log
# 2026-05-21T06:18:05+0000 profile=coder prior_state=running action=started
# 2026-05-21T06:18:05+0000 profile=writer prior_state=stopped action=registered

Add a new static service

1. Create docker/s6-rc.d//type with longrun\n and docker/s6-rc.d//run (use #!/command/with-contenv sh + # shellcheck shell=sh).
2. Drop to hermes via s6-setuidgid hermes at the top of run (unless you specifically need root).
3. Create empty docker/s6-rc.d//dependencies.d/base so it waits for the base bundle.
4. Create empty docker/s6-rc.d/user/contents.d/ so it joins the user bundle.
5. The COPY docker/s6-rc.d/ in the Dockerfile picks it up automatically — no other changes.

Change the per-profile gateway run command

Edit S6ServiceManager.renderrunscript in hermescli/servicemanager.py. The function is also called by hermescli/containerboot.py::registerservice during boot reconciliation, so it's the single source of truth. Update the corresponding assertion in tests/hermescli/testservicemanager.py::tests6registercreatesservicedirandtriggersscan.

Run the docker test harness

docker build -t hermes-agent-harness:latest .
HERMES_TEST_IMAGE=hermes-agent-harness:latest scripts/run_tests.sh tests/docker/ -v
# Expect 19 passed, 0 xfailed against the s6 image

The harness lives in tests/docker/ and skips when Docker isn't available. The per-test timeout is bumped to 180s (see tests/docker/conftest.py).

Common pitfalls

"command not found" via docker exec

/command/ (where s6-overlay puts its binaries) is on PATH only for processes spawned by the supervision tree — services, cont-init.d, main-wrapper.sh. docker exec s6-svstat … will fail with "command not found"; always use the absolute path /command/s6-svstat. The hermes binary works because the Dockerfile adds /opt/hermes/.venv/bin to the runtime ENV PATH.

Profile directory ownership

The cont-init reconciler runs as hermes (s6-setuidgid hermes in 02-reconcile-profiles). If a profile dir ends up root-owned (e.g. because docker exec hermes profile create … ran as root by default), the reconciler can't read SOUL.md and fails with PermissionError. Mitigation: stage2-hook.sh chowns $HERMESHOME/profiles to hermes on every** boot, idempotently. Don't remove that block.

Files written by docker exec are root-owned

docker exec defaults to root. Either pass --user hermes or rely on the stage2 chown sweep next reboot. Don't write files under $HERMES_HOME/profiles// as root manually — the next reconcile pass will sweep them but in-flight operations may hit perm errors.

Root code must not work inside /run/hermes-service

The gateway scandir is hermes-owned, so hermes can rename anything in it and leave a symlink behind. Root writes, chmods and chowns there follow that symlink. Root registration therefore builds each slot in /run/.hermes-service.staging (root, 0700, same tmpfs) and publishes it with one rename; S6ServiceManager runs s6-svc, s6-svwait and s6-svscanctl through s6-setuidgid hermes when called as root. Keep new root-side code on those paths.

Service slot exists but s6-svstat says "s6-supervise not running"

The service directory is on tmpfs and was wiped on container restart. Either the cont-init reconciler hasn't run yet (give it a moment after docker restart) or it failed. Check docker logs | grep '02-reconcile'.

Gateway starts then immediately exits (down (exitcode 1) in svstat)

Most likely the profile has no model or auth configured. The service slot is correct — the gateway itself is unconfigured. Run hermes -p setup first. The s6 supervisor will keep restarting it; that's the desired behavior (when you fix the config, the next attempt succeeds and stays up).

Reconciler skipped a profile

The reconciler keys on the presence of SOUL.md as the "real profile" marker. hermes profile create always seeds it. If a profile dir is missing SOUL.md (stray directory, partial restore, backup-in-progress), the reconciler skips it intentionally. Add a SOUL.md (even empty) to opt back in.

"Help, the container exits 143!"

Check whether something is invoking s6-svscanctl -t or /run/s6/basedir/bin/halt — both cause /init to begin stage 3 shutdown but return 143 (SIGTERM) rather than the desired exit code. This was the Phase 2 architecture pivot from A to B. For container shutdown with a real exit code, you must let the CMD (main-wrapper.sh) exit normally; do not try to control exit from a finish script.

Related skills

• hermes-agent-dev: General hermes-agent codebase navigation
• hermes-tool-quirks: Specific Hermes-tool workarounds (sed/grep/etc.) — load when debugging the s6 stack's interaction with hermes built-in tools.