- Add OpenViking service (knowledge graph) using official GHCR image
- Add Honcho stack (user modeling): API + PostgreSQL pgvector + Redis
- Add Holographic config to Hermes (local SQLite, no server needed)
- Hermes: install httpx for OpenViking client
- Hermes: auto-generate config.yaml + honcho.json on first boot
- All data 100% local, zero cloud dependencies
Without this, is empty and entrypoint.sh runs bare 'hermes'
which defaults to interactive chat mode. With a non-TTY stdin
this exits immediately with prompt_toolkit's 'Input is not a
terminal' warning, causing a container restart loop.
The profile gateways (run-multi-gateways.sh) were unaffected
because the script passes 'gateway run' explicitly.
- Add libolm-dev system dep (required by mautrix[encryption])
- Add mautrix[encryption] + openai pip packages to build
- These were previously installed inline at container startup and
persisted via the fragile venv volume mount (now removed)
The volume mount /mnt/HoardingCow_docker_data/Hermes/venv overrides the
container's built-in .venv with whatever is on the host. On a fresh start
or after a clean build, an empty/missing venv directory causes entrypoint.sh
line 62 (source .venv/bin/activate) to fail with set -e.
The Docker image already builds a complete venv — persisting it on the host
is unnecessary and fragile.
- Add run-multi-gateways.sh to /opt/data/hermes-tools/ that reads
HERMES_PROFILES env var and spawns one gateway per profile
- Update entrypoint to call the script before the main entrypoint
- Set HERMES_PROFILES=ashley,claire,finn,matt,paul (was default)
Closes PR #47 (feat/multi-profile-gateways). Builds on 548e15d's cleaner
env-var-driven approach — compose.yml stays declarative, logic in script.
The chown -R hermes:hermes was running as non-root user 'hermes'
since USER hermes was set earlier. The new upstream base image
(v0.12.0+) has tools/ owned by root, so the chown fails.
Previous base image happened to have tools/ owned by hermes,
making the chown a silent no-op.
- Replace rsync with cp -a (rsync unavailable in latest upstream base image)
- Remove npm run build step (fork's package.json has no build script)
- Remove himalaya-ro.sh from build context (deployed via install.sh)
- Add hermes to ai_net network for Traefik access
- Add Traefik labels routing hermes.lazyworkhorse.net to dashboard port 9119
- Switch Dockerfile to clone from gortium/hermes-agent (Gitea fork)
- Add SSH agent forwarding for private repo clone at build time
- Set CHROME_EXECUTABLE for Playwright Chromium
- Remove patch_tts_tool.py (Piper patch now in fork source)
- Enable Gitea Actions in versioncontrol compose
- Use --preset 'ROCm 6' for HIP build step (enables enable_language(HIP))
- Remove /opt/rocm from PATH for CPU build to prevent check_language(HIP)
- Add CMAKE_PREFIX_PATH=/opt/rocm so find_package(hip) finds hip-config.cmake
- cmake --install --component HIP now works correctly with OLLAMA_RUNNER_DIR=rocm
gfx940/gfx1010/gfx1030/gfx1100 cause C++ narrowing errors in ollama's
mma.cuh with hipcc. Since we only have MI50 (gfx906) cards, compile
for gfx906 only. Reduces build time and avoids upstream code bugs.
CPU backends compiled with GCC (fixes AVX intrinsic errors from hipcc).
HIP backend compiled with hipcc (legacy mode skips enable_language(HIP)).
Go binary built with GCC for CGo linking.
This avoids both CMAKE_HIP_COMPILER rejection and CXX=hipcc CPU failures.
The old Dockerfile used the deprecated llama.cpp/ subdirectory approach
which doesn't exist in ollama v0.23.2. Now using the official CMake
presets (ROCm 6 preset) with AMDGPU_TARGETS including gfx906:xnack-.
- Add ollama/Dockerfile that builds ollama from source with AMDGPU_TARGETS=gfx906
- Uses ROCm 6.1 (rocm/dev-ubuntu-22.04:6.1.2-complete) for MI50 support
- Builds llama.cpp runner with HIPBLAS for gfx906 architecture
- Updates compose.yml to build from this Dockerfile instead of pulling ollama/ollama:latest
