- Remove patch_tts_tool.py (orphaned)
- Remove run-multi-gateways.sh (replaced by s6 native profile supervision)
- Remove start-hermes.sh (overengineered — the image does this natively)
- compose.yml: command: gateway run, drop HERMES_PROFILES env var
- Dockerfile: revert start-hermes.sh COPY addition
Per-profile gateways are now managed by s6-overlay natively:
profiles with desired_state=running in gateway_state.json are
auto-started by the s6 reconciler on boot.
- New start-hermes.sh: multi-profile launcher that works with s6-overlay's
main-program model (replaces bash->tini->entrypoint.sh chain)
- Dockerfile: COPY start-hermes.sh into /usr/local/bin/ alongside
run-multi-gateways.sh (which is now unused but kept for reference)
- compose.yml: remove entrypoint override, CMD now points at
/usr/local/bin/start-hermes.sh via the image default ENTRYPOINT
Fixes the SIGTERM crash loop caused by S6_CMD_ARG0 being unset when the
deprecated entrypoint.sh shim bypassed s6-overlay's /init.
The custom ENTRYPOINT chained bash -> tini -g -> deprecated entrypoint.sh,
bypassing s6-overlay's /init entirely. This left S6_CMD_ARG0 unset and the
orphan -g flag crashed rc.init with '-g: not found' -> SIGTERM -> restart loop.
Fix:
- Remove the ENTRYPOINT override so the image default is used:
ENTRYPOINT ['/init', '/opt/hermes/docker/main-wrapper.sh']
- Change CMD to point at /opt/data/start-hermes.sh, a new launcher
that starts per-profile gateways in background then the default
gateway in foreground (via s6-overlay's main-program model).
The old /usr/local/bin/run-multi-gateways.sh is no longer called.
- rebased custom commits on upstream v0.16.0
- 125 upstream commits since fork base (e490d91)
- includes dashboard, fleet, dark mode, deps consolidation
- our TS fixes preserved on top
- Install gosu in hermes image so run-multi-gateways.sh can drop
privileges when USER hermes is set at the end of the Dockerfile.
- Add openssh-client + ssh-keyscan to the openconcho-builder stage
so SSH host key verification passes during the first Gitea clone.
b9596 (Dec 2024) has known slot deadlock on first request causing
container freeze. b9890 includes:
- Slot selection consolidation (#24755)
- HIP fast-math support (#23862)
- Hundreds of stability and memory fixes
Compatible with ROCm 6.1 + gfx906 (same build flags).
The Dockerfile was switching to USER root for the final chown but never
switched back to USER hermes. This caused ALL container processes to run
as root (uid 0) instead of the hermes user (uid 10000).
The entrypoint's gosu privilege drop only caught the main exec chain,
leaving backgrounded subprocesses (dashboard PTY sessions, workers with
start_new_session=True) running as root — creating files owned by root
in ExoKortex and breaking Syncthing sync.
Adding USER hermes at the end ensures the container runs unprivileged
and ALL child processes inherit uid 10000 from the start.
- Rename ollama service to ollama-cpu (CPU-only for bge-m3 embeddings)
- Fix llama-cpp-hermes indented under networks instead of as top-level service
- Update hermes OLLAMA_HOST to point to ollama-cpu
Part of PR #54 GPU/ROCm refactor
- Update Dockerfile to clone from code.lazyworkhorse.net/Hermes/honcho.git
(uses build arg HONCHO_REPO, can be overridden at build time)
- Add config.toml volume mount from HoardingCow persistent path
- Use named volume honcho_data instead of host bind mount
- Declare honcho_data as external volume in top-level volumes section
