feat(docker): add LaTeX typesetting stack

Add TeXLive packages for PDF generation from Org-mode and LaTeX documents: - texlive-latex-base: Core LaTeX packages - texlive-latex-extra: Additional LaTeX packages - texlive-fonts-recommended: Recommended fonts - texlive-xetex: XeTeX engine for Unicode support - texlive-science: Science/math packages This is PR 3 of 5 for Docker package additions. Depends on PR #10 and #11.
Add AI model optimizer cron job draft and initial state files
2026-04-29 21:00:05 +00:00 · 2026-04-28 17:19:45 +00:00
18 changed files with 739 additions and 672 deletions
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -5,7 +5,6 @@ This document outlines the development conventions for this NixOS-based infrastr
 ## Build & Deployment
 - **Build/Deploy:** Use `nixos-rebuild switch --flake .#<hostname>` to build and deploy the configuration for a specific host.
 - **CRITICAL — Validate before pushing:** Always `nix build --no-link '.#nixosConfigurations.<hostname>.config.system.build.toplevel'` (or `nh os build`) and confirm it succeeds before pushing any changes. Never push untested NixOS configs.
 - **Development Shell:** Activate the development environment with `nix develop`.
 ## Linting & Formatting
--- a/assets/ai-optimizer/CRON_EXECUTION_PROMPT.md
+++ b/assets/ai-optimizer/CRON_EXECUTION_PROMPT.md
@@ -0,0 +1,203 @@
 # AI Model Optimization Cron Job - EXECUTION PROMPT
 **When this cron runs, follow these instructions exactly:**
 ---
 ## Your Role
 You are an AI model optimization agent. Your task is to find the best ollama/llama.cpp configuration for maximum context size and hardware utilization.
 **Hardware:**
 - 2× AMD MI50 GPUs (32GB VRAM each, 64GB total)
 - 128GB system RAM
 - ROCm: HSA_OVERRIDE_GFX_VERSION=9.0.6, HIP_VISIBLE_DEVICES=0,1
 ---
 ## File Locations
 ```
 STATE: /opt/data/infra/assets/ai-optimizer/state.json
 RESULTS: /opt/data/infra/assets/ai-optimizer/results.csv
 INFRA_REPO: /opt/data/infra
 ```
 ---
 ## Model Queues
 ### GPU Track (Coding - prioritize speed + context on GPU)
 1. `devstral-small-2:24b`
 2. `qwen2.5-coder:32b`
 3. `codellama:34b-instruct`
 ### RAM Track (Knowledge - prioritize max context)
 1. `qwen2.5:72b`
 2. `nemotron-3-nano:30b`
 3. `mixtral:8x7b-instruct`
 ---
 ## Context Steps (in order)
 ```
 [32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680]
 ```
 ---
 ## Each Run - Step by Step
 ### 1. Read State
 ```bash
 cd /opt/data/infra
 cat assets/ai-optimizer/state.json
 ```
 ### 2. Determine Next Test
 - Read `track` (gpu or ram)
 - Read `current_model` from queue at `model_index`
 - Read `current_config` for parameters to test
 - Select next context step from `context_steps` based on `phase`
 ### 3. Pull Model (if needed)
 ```bash
 docker exec ollama ollama list | grep -q "<model>" || docker exec ollama ollama pull <model>
 ```
 ### 4. Create Test Modelfile
 ```bash
 docker exec ollama bash -c "cat <<EOF > /root/.ollama/test_${model}.modelfile
 FROM ${model}
 PARAMETER num_ctx ${current_config.num_ctx}
 PARAMETER num_gpu ${current_config.num_gpu}
 PARAMETER flash_attn ${current_config.flash_attn}
 PARAMETER num_predict 4096
 PARAMETER num_keep 1024
 PARAMETER repeat_penalty 1.1
 EOF"
 docker exec ollama ollama create test-model -f /root/.ollama/test_${model}.modelfile
 ```
 ### 5. Run Benchmark
 ```bash
 # Warm up
 docker exec ollama ollama run test-model "Hello" > /dev/null
 # Coding prompt
 START=$(date +%s%N)
 docker exec ollama ollama run test-model "Write a Python async context manager that retries a function with exponential backoff, max 5 retries, and logs each attempt using structlog. Include type hints."
 END=$(date +%s%N)
 # Calculate tokens/sec from output
 ```
 ### 6. Measure VRAM (if possible)
 ```bash
 # Try host first
 rocm-smi --showmeminfo vram 2>/dev/null || \
 # Try via docker
 docker exec --privileged ollama rocm-smi --showmeminfo vram 2>/dev/null || \
 # Fallback
 echo "VRAM measurement unavailable"
 ```
 ### 7. Record Results
 - Parse tokens/sec from ollama output
 - Record VRAM/RAM usage
 - Determine if this is best config so far for this model
 - Update `best_configs` if tokens/sec improved or context increased
 ### 8. Update State
 ```python
 # Logic:
 if test_successful:
    if context_step < max_reached:
        phase = "context_scaling"
        current_config.num_ctx = next_context_step
    else:
        # Move to next model
        model_index += 1
        phase = "context_scaling"
        current_config.num_ctx = context_steps[0]
 else:
    # OOM or error - record last good as best
    best_configs[track][current_model] = last_good_config
    model_index += 1
    phase = "context_scaling"
 ```
 ### 9. Commit to Repo
 ```bash
 cd /opt/data/infra
 git add assets/ai-optimizer/
 git commit -m "ai-optimizer: tested ${model} at ${num_ctx} ctx - ${status}"
 git push origin master
 ```
 ### 10. Matrix Notification (if available)
 ```python
 import os
 if os.getenv("MATRIX_HOME_SERVER") and os.getenv("MATRIX_ACCESS_TOKEN"):
    # Send notification to Matrix room
    # Room ID from env or config
    pass
 # Else: silent
 ```
 ---
 ## Stop Conditions
 1. All models in both queues have `best_configs` recorded
 2. Manual intervention needed (error in state.json `error` field)
 3. No progress for 3 consecutive runs (stuck)
 ---
 ## Error Handling
 If any step fails:
 1. Log error to state.json: `"error": {"message": "...", "timestamp": "..."}`
 2. Do NOT increment model_index (retry next run)
 3. Commit state with error field
 4. Exit gracefully
 ---
 ## Important Notes
 - **No num_parallel**: Do not use this parameter
 - **Two tracks**: Complete GPU track first, then RAM track
 - **Backend**: Start with ollama, llama.cpp testing is optional (requires uncommenting in compose.yml)
 - **Host access**: Some commands need host - use docker exec or SSH if available
 - **Ask before deploy**: If config changes needed in NixOS modules, show diff and wait for user confirmation before `nh os switch`
 ---
 ## Example State Transitions
 **Start:**
 ```json
 {"track": "gpu", "model_index": 0, "current_model": "devstral-small-2:24b", "current_config": {"num_ctx": 32768, ...}}
 ```
 **After successful test at 32k:**
 ```json
 {"track": "gpu", "model_index": 0, "current_model": "devstral-small-2:24b", "current_config": {"num_ctx": 65536, ...}}
 ```
 **After OOM at 131k:**
 ```json
 {
  "track": "gpu",
  "model_index": 1,
  "current_model": "qwen2.5-coder:32b",
  "best_configs": {
    "gpu": {
      "devstral-small-2:24b": {"num_ctx": 98304, "num_gpu": 99, "tokens_per_sec": 11.2}
    }
  }
 }
 ```
--- a/assets/ai-optimizer/CRON_JOB_DRAFT.md
+++ b/assets/ai-optimizer/CRON_JOB_DRAFT.md
@@ -0,0 +1,283 @@
 # AI Model Optimization Cron Job
 **Goal:** Find optimal configurations for maximum context size with full hardware utilization.
 **Hardware:**
 - 2× AMD MI50 GPUs (32GB VRAM each, 64GB total)
 - 128GB system RAM
 - ROCm: HSA_OVERRIDE_GFX_VERSION=9.0.6, HIP_VISIBLE_DEVICES=0,1
 ---
 ## Model Queue
 ### GPU-Optimized (Coding - prioritize speed + context on GPU)
 1. `devstral-small-2:24b` - Best coding model
 2. `qwen2.5-coder:32b` - Strong coder, fits on GPU+offload
 3. `codellama:34b-instruct` - Legacy but solid
 ### RAM-Optimized (Knowledge - prioritize max context, accept slower)
 1. `qwen2.5:72b` - Best knowledge, needs heavy offload
 2. `nemotron-3-nano:30b` - Good general knowledge
 3. `mixtral:8x7b-instruct` - MoE, efficient for knowledge
 ---
 ## Optimization Strategy
 **Two separate tracks:**
 ### Track A: GPU-Focused (Coding)
 ```
 Baseline: num_ctx=32768, num_gpu=99, flash_attn=true
 Steps:
 1. Increase context: 32k → 65k → 98k → 131k → 163k
 2. At each step, verify VRAM usage < 60GB (leave headroom)
 3. If OOM: reduce num_gpu until stable, record best
 4. Measure tokens/sec - if < 5 tok/s, consider context too high
 ```
 ### Track B: RAM-Focused (Knowledge)
 ```
 Baseline: num_ctx=65536, num_gpu=50, flash_attn=true
 Steps:
 1. Increase context: 65k → 131k → 200k → 262k → 327k
 2. Allow heavy RAM offload (system RAM up to 100GB)
 3. If OOM: reduce context or num_gpu
 4. Speed less critical - focus on max stable context
 ```
 ---
 ## Backend-Specific Configs
 ### Ollama (Modelfile parameters)
 ```
 PARAMETER num_ctx <value>
 PARAMETER num_gpu <layers>
 PARAMETER flash_attn true/false
 PARAMETER num_predict 4096
 PARAMETER num_keep 1024
 PARAMETER repeat_penalty 1.1
 ```
 ### Llama.cpp (CLI flags)
 ```
 --ctx-size <value>
 --n-gpu-layers <layers>
 --flash-attn on/off
 --n-predict 4096
 --batch-size 4096
 --ubatch-size 512
 --cache-type-k f16
 --cache-type-v f16
 --split-mode layer
 --no-mmap
 ```
 ---
 ## Host Test Instructions
 **The cron runs inside the hermes container. Some tests require host access:**
 ### 1. VRAM Monitoring (HOST)
 ```bash
 # Run on host to check VRAM usage during/after benchmark
 sudo rocm-smi --showmeminfo vram
 # Or via docker exec if rocm-smi available in container
 docker exec --privileged ollama rocm-smi --showmeminfo vram
 ```
 ### 2. Running Ollama Benchmarks (CONTAINER)
 ```bash
 # Pull model
 docker exec ollama ollama pull <model>
 # Create custom modelfile
 docker exec ollama bash -c 'cat <<EOF > /root/.ollama/test.modelfile
 FROM <model>
 PARAMETER num_ctx 65536
 PARAMETER num_gpu 99
 PARAMETER flash_attn true
 EOF'
 # Create model from modelfile
 docker exec ollama ollama create test-model -f /root/.ollama/test.modelfile
 # Run benchmark (warm model first)
 docker exec ollama ollama run test-model "Write a Python async context manager with exponential backoff"
 # Cleanup
 docker exec ollama ollama rm test-model
 ```
 ### 3. Running Llama.cpp Benchmarks (CONTAINER - needs llama.cpp container)
 ```bash
 # Uncomment llama_cpp_devstral in compose.yml first
 # Then rebuild: sudo nh os switch --flake .#lazyworkhorse
 # Test via HTTP API
 curl http://localhost:8300/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "devstral-2-small-llama_cpp",
    "prompt": "Write a Python function",
    "max_tokens": 100
  }'
 ```
 ### 4. Deploying Changes (HOST via ai-worker)
 ```bash
 # After optimization, commit results
 cd /home/ai-worker/infra
 git add assets/ai-optimizer/
 git commit -m "ai-optimizer: new best config for <model>"
 git push
 # If config changes needed in ollama_init_custom_models.nix:
 # 1. Edit the file
 # 2. nixpkgs-fmt .
 # 3. Show diff to user
 # 4. Wait for confirmation
 # 5. sudo nh os switch --flake .#lazyworkhorse
 ```
 ### 5. Accessing Host from Hermes Container
 ```bash
 # SSH to host as ai-worker (key should be mounted)
 ssh -i /path/to/key ai-worker@host.docker.internal
 # Or via docker socket if mounted
 # (not recommended for security)
 ```
 ---
 ## Benchmark Prompts
 ### Coding (Track A)
 ```
 "Write a Python async context manager that retries a function with exponential backoff, max 5 retries, and logs each attempt using structlog. Include type hints and error handling."
 ```
 ### Knowledge (Track B)
 ```
 "Explain the complete memory hierarchy in modern GPUs, from registers through L1/L2 caches to VRAM, and how data moves between them during matrix multiplication. Include bandwidth considerations for each level."
 ```
 ### Measurement
 - Tokens per second (generation speed)
 - Time to first token (latency)
 - VRAM usage (via rocm-smi)
 - System RAM usage (via free -h)
 - Context success (did it complete without OOM?)
 ---
 ## State File Structure
 `/opt/data/infra/assets/ai-optimizer/state.json`
 ```json
 {
  "track": "gpu",
  "current_model": "devstral-small-2:24b",
  "model_index": 0,
  "phase": "context_scaling",
  "backend": "ollama",
  "current_config": {
    "num_ctx": 65536,
    "num_gpu": 99,
    "flash_attn": true
  },
  "best_configs": {
    "gpu": {
      "devstral-small-2:24b": {
        "backend": "ollama",
        "num_ctx": 131072,
        "num_gpu": 99,
        "flash_attn": true,
        "tokens_per_sec": 12.5,
        "vram_used_gb": 58.2,
        "tested_at": "2026-04-28T17:00:00Z"
      }
    },
    "ram": {}
  },
  "completed_models": [],
  "gpu_queue": ["devstral-small-2:24b", "qwen2.5-coder:32b", "codellama:34b-instruct"],
  "ram_queue": ["qwen2.5:72b", "nemotron-3-nano:30b", "mixtral:8x7b-instruct"]
 }
 ```
 ---
 ## Results CSV
 `/opt/data/infra/assets/ai-optimizer/results.csv`
 ```csv
 timestamp,track,model,backend,phase,num_ctx,num_gpu,flash_attn,tokens_per_sec,vram_gb,ram_gb,status,is_best
 2026-04-28T17:00:00Z,gpu,devstral-small-2:24b,ollama,context_scaling,65536,99,true,15.2,52.1,18.4,success,false
 ```
 ---
 ## Cron Job Flow
 ```
 1. Read state.json
 2. If both queues empty → STOP (all models tested)
 3. Select next model from current track queue
 4. Pull model if needed (docker exec ollama ollama pull)
 5. Create Modelfile / llama.cpp config with current test params
 6. Run benchmark (both prompts)
 7. Measure: tokens/sec, VRAM (rocm-smi), RAM (free -h)
 8. If successful:
   - Increase context (next step)
   - Update current_config in state
 9. If OOM/error:
   - Record last good config as best_configs[track][model]
   - Move to next model in queue
 10. Update state.json
 11. Append to results.csv
 12. Git commit + push to /opt/data/infra
 13. Send Matrix notification if available, else silent
 ```
 ---
 ## Matrix Notification (Optional)
 ```python
 # If matrix credentials available in environment
 if os.getenv("MATRIX_HOME_SERVER") and os.getenv("MATRIX_ACCESS_TOKEN"):
    # Send completion notification
    # Room: !ai-optimizer:lazyworkhorse.net (or similar)
    pass
 # Else: silent, just commit
 ```
 ---
 ## Files to Create
 ```
 /opt/data/infra/assets/ai-optimizer/
 ├── state.json           # Current progress
 ├── results.csv          # All test results
 ├── best_configs.json    # Final best configs (human-readable)
 └── CRON_JOB_DRAFT.md    # This file
 ```
 ---
 ## Notes
 - **No num_parallel**: Removed to avoid limiting other settings
 - **Two tracks**: GPU (coding/speed) vs RAM (knowledge/context)
 - **Both backends**: Test ollama first, then llama.cpp if available
 - **Host tests**: rocm-smi must run on host or privileged container
 - **Deploy**: ai-worker has sudo for nh/nixos-rebuild, must ask user first
--- a/assets/ai-optimizer/results.csv
+++ b/assets/ai-optimizer/results.csv
@@ -0,0 +1 @@
 timestamp,track,model,backend,phase,num_ctx,num_gpu,flash_attn,tokens_per_sec,vram_gb,ram_gb,status,is_best
--- a/assets/ai-optimizer/state.json
+++ b/assets/ai-optimizer/state.json
@@ -0,0 +1,21 @@
 {
  "track": "gpu",
  "current_model": "devstral-small-2:24b",
  "model_index": 0,
  "phase": "context_scaling",
  "backend": "ollama",
  "current_config": {
    "num_ctx": 32768,
    "num_gpu": 99,
    "flash_attn": true
  },
  "best_configs": {
    "gpu": {},
    "ram": {}
  },
  "completed_models": [],
  "gpu_queue": ["devstral-small-2:24b", "qwen2.5-coder:32b", "codellama:34b-instruct"],
  "ram_queue": ["qwen2.5:72b", "nemotron-3-nano:30b", "mixtral:8x7b-instruct"],
  "context_steps": [32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680],
  "last_updated": "2026-04-28T17:00:00Z"
 }
--- a/assets/compose
+++ b/assets/compose
--- a/assets/ollama/Dockerfile
+++ b/assets/ollama/Dockerfile
@@ -1,106 +0,0 @@
 # ollama-gfx906/Dockerfile
 #
 # Custom ollama image with ROCm 6.1 + gfx906 (MI50) support.
 # The official ollama/rocm image ships ROCm 7.2 which dropped gfx906.
 # This uses v0.23.2's native CMake build system with AMDGPU_TARGETS including gfx906.
 #
 # Build: docker build -t ollama/ollama:rocm-gfx906 ai/ollama
 FROM rocm/dev-ubuntu-22.04:6.1.2-complete AS builder
 # Build dependencies (CMake, Ninja, Go)
 ARG CMAKEVERSION=3.31.2
 ARG NINJAVERSION=1.12.1
 ARG GOLANG_VERSION=1.22.0
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    curl git ccache build-essential pkg-config unzip \
    && rm -rf /var/lib/apt/lists/*
 # Install CMake from official binaries
 RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-x86_64.tar.gz \
    | tar xz -C /usr/local --strip-components 1
 # Install Ninja
 RUN curl -fsSL -o /tmp/ninja.zip \
    https://github.com/ninja-build/ninja/releases/download/v${NINJAVERSION}/ninja-linux.zip \
    && unzip /tmp/ninja.zip -d /usr/local/bin && rm /tmp/ninja.zip
 # Install Go
 RUN curl -fsSL https://go.dev/dl/go${GOLANG_VERSION}.linux-amd64.tar.gz \
    | tar xz -C /usr/local
 ENV PATH=/usr/local/go/bin:$PATH
 ARG OLLAMA_VERSION=v0.23.2
 RUN git clone --depth 1 --branch ${OLLAMA_VERSION} https://github.com/ollama/ollama.git /build
 WORKDIR /build
 # ROCm paths
 ENV HIP_PATH=/opt/rocm
 ENV ROCM_PATH=/opt/rocm
 ENV CMAKE_GENERATOR=Ninja
 ENV LDFLAGS=-s
 # Step 1: Build CPU backends with GCC (no ROCm preset)
 # Pre-set CMAKE_HIP_COMPILER="" to prevent check_language(HIP) from
 # finding a HIP compiler (it searches /opt/rocm even without PATH).
 # Remove /opt/rocm from PATH to prevent find_program from finding hipcc.
 RUN mkdir -p build-cpu && \
    PATH=/usr/local/go/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin \
    cmake -B build-cpu -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_HIP_COMPILER="" \
      -DCMAKE_INSTALL_PREFIX=/build/dist && \
    cmake --build build-cpu --target ggml-cpu -- -l $(nproc) && \
    cmake --install build-cpu --component CPU --strip && \
    echo "=== CPU install ===" && \
    (find /build/dist/lib/ollama -type f -o -type l 2>&1 | head -20 || echo "empty")
 # Step 2: Build HIP backend with ROCm preset + gfx906 target only
 # The ROCm 6 preset enables HIP language detection (enable_language(HIP))
 # which ensures GPU kernels are properly compiled for gfx906.
 # OLLAMA_RUNNER_DIR=rocm from the preset, so HIP goes to lib/ollama/rocm/
 # Need CMAKE_PREFIX_PATH so find_package(hip) finds hip-config.cmake
 # at /opt/rocm/lib/cmake/hip/hip-config.cmake.
 RUN mkdir -p build-hip && \
    cmake -B build-hip \
      --preset 'ROCm 6' \
      -DAMDGPU_TARGETS="gfx906:xnack-" \
      -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_PREFIX_PATH="/opt/rocm" && \
    cmake --build build-hip --target ggml-hip -- -l $(nproc) && \
    cmake --install build-hip --component HIP --strip && \
    echo "=== HIP install ===" && \
    find /build/dist/lib/ollama -type f -o -type l | head -20
 # Step 3: Build Go binary (GCC for CGo linking)
 ENV CGO_ENABLED=1
 RUN go build -trimpath -ldflags="-X=github.com/ollama/ollama/version.Version=${OLLAMA_VERSION}" -o /build/dist/ollama .
 # ---------- Runtime image ----------
 FROM ubuntu:24.04
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
    ca-certificates curl libstdc++6 libgomp1 libvulkan1 libopenblas0 \
    && rm -rf /var/lib/apt/lists/*
 # Copy ROCm 6.1 runtime libraries
 # These are needed at runtime by ggml-hip via LD_LIBRARY_PATH
 COPY --from=builder /opt/rocm/lib/ /opt/rocm/lib/
 COPY --from=builder /opt/rocm/share/ /opt/rocm/share/
 # Copy ollama binary + all backends (CPU + HIP)
 # CPU install:  /build/dist/lib/ollama/libggml-*.so
 # HIP install:  /build/dist/lib/ollama/rocm/libggml-hip.so
 COPY --from=builder /build/dist/ollama /usr/bin/ollama
 COPY --from=builder /build/dist/lib/ollama/ /usr/lib/ollama/
 RUN ldconfig
 ENV LD_LIBRARY_PATH=/opt/rocm/lib:/usr/lib/ollama/rocm:/usr/lib/ollama
 ENV HSA_OVERRIDE_GFX_VERSION=9.0.6
 ENV HCC_AMDGPU_TARGET=gfx906
 ENV HSA_ENABLE_SDMA=0
 EXPOSE 11434
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]
--- a/docker/hermes/Dockerfile
+++ b/docker/hermes/Dockerfile
@@ -0,0 +1,67 @@
 FROM ghcr.io/astral-sh/uv:0.11.6-python3.13-trixie@sha256:b3c543b6c4f23a5f2df22866bd7857e5d304b67a564f4feab6ac22044dde719b AS uv_source
 FROM tianon/gosu:1.19-trixie@sha256:3b176695959c71e123eb390d427efc665eeb561b1540e82679c15e992006b8b9 AS gosu_source
 FROM debian:13.4
 # Disable Python stdout buffering to ensure logs are printed immediately
 ENV PYTHONUNBUFFERED=1
 # Store Playwright browsers outside the volume mount so the build-time
 # install survives the /opt/data volume overlay at runtime.
 ENV PLAYWRIGHT_BROWSERS_PATH=/opt/hermes/.playwright
 # Install system dependencies in one layer, clear APT cache
 # tini reaps orphaned zombie processes (MCP stdio subprocesses, git, bun, etc.)
 # that would otherwise accumulate when hermes runs as PID 1. See #15012.
 RUN apt-get update && \
    apt-get install -y --no-install-recommends \
        build-essential nodejs npm python3 ripgrep ffmpeg gcc python3-dev libffi-dev procps git openssh-client docker-cli tini \
        curl poppler-utils imagemagick \
        chromium xvfb fonts-noto-color-emoji fonts-unifont fonts-liberation fonts-ipafont-gothic fonts-wqy-zenhei fonts-tlwg-loma-otf fonts-freefont-ttf \
        libasound2t64 libatk-bridge2.0-0t64 libatk1.0-0t64 libatspi2.0-0t64 libcairo2 libcups2t64 libdbus-1-3 libdrm2 libgbm1 libglib2.0-0t64 libnspr4 libnss3 libpango-1.0-0 libx11-6 libxcb1 libxcomposite1 libxdamage1 libxext6 libxfixes3 libxkbcommon0 libxrandr2 \
        texlive-latex-base texlive-latex-extra texlive-fonts-recommended texlive-xetex texlive-science && \
    rm -rf /var/lib/apt/lists/*
 # Non-root user for runtime; UID can be overridden via HERMES_UID at runtime
 RUN useradd -u 10000 -m -d /opt/data hermes
 COPY --chmod=0755 --from=gosu_source /gosu /usr/local/bin/
 COPY --chmod=0755 --from=uv_source /usr/local/bin/uv /usr/local/bin/uvx /usr/local/bin/
 WORKDIR /opt/hermes
 # ---------- Layer-cached dependency install ----------
 # Copy only package manifests first so npm install + Playwright are cached
 # unless the lockfiles themselves change.
 COPY package.json package-lock.json ./
 COPY web/package.json web/package-lock.json web/
 RUN npm install --prefer-offline --no-audit && \
    npx playwright install --with-deps chromium --only-shell && \
    (cd web && npm install --prefer-offline --no-audit) && \
    npm cache clean --force
 # ---------- Source code ----------
 # .dockerignore excludes node_modules, so the installs above survive.
 COPY --chown=hermes:hermes . .
 # Build web dashboard (Vite outputs to hermes_cli/web_dist/)
 RUN cd web && npm run build
 # ---------- Permissions ----------
 # Make install dir world-readable so any HERMES_UID can read it at runtime.
 # The venv needs to be traversable too.
 USER root
 RUN chmod -R a+rX /opt/hermes
 # Start as root so the entrypoint can usermod/groupmod + gosu.
 # If HERMES_UID is unset, the entrypoint drops to the default hermes user (10000).
 # ---------- Python virtualenv ----------
 RUN uv venv && \
    uv pip install --no-cache-dir -e ".[all]"
 # ---------- Runtime ----------
 ENV HERMES_WEB_DIST=/opt/hermes/hermes_cli/web_dist
 ENV HERMES_HOME=/opt/data
 ENV PATH="/opt/data/.local/bin:${PATH}"
 VOLUME [ "/opt/data" ]
 ENTRYPOINT [ "/usr/bin/tini", "-g", "--", "/opt/hermes/docker/entrypoint.sh" ]
--- a/docker/hermes/entrypoint.sh
+++ b/docker/hermes/entrypoint.sh
@@ -0,0 +1,102 @@
 #!/bin/bash
 # Docker/Podman entrypoint: bootstrap config files into the mounted volume, then run hermes.
 set -e
 HERMES_HOME="${HERMES_HOME:-/opt/data}"
 INSTALL_DIR="/opt/hermes"
 # --- Privilege dropping via gosu ---
 # When started as root (the default for Docker, or fakeroot in rootless Podman),
 # optionally remap the hermes user/group to match host-side ownership, fix volume
 # permissions, then re-exec as hermes.
 if [ "$(id -u)" = "0" ]; then
    if [ -n "$HERMES_UID" ] && [ "$HERMES_UID" != "$(id -u hermes)" ]; then
        echo "Changing hermes UID to $HERMES_UID"
        usermod -u "$HERMES_UID" hermes
    fi
    if [ -n "$HERMES_GID" ] && [ "$HERMES_GID" != "$(id -g hermes)" ]; then
        echo "Changing hermes GID to $HERMES_GID"
        # -o allows non-unique GID (e.g. macOS GID 20 "staff" may already exist
        # as "dialout" in the Debian-based container image)
        groupmod -o -g "$HERMES_GID" hermes 2>/dev/null || true
    fi
    # Fix ownership of the data volume. When HERMES_UID remaps the hermes user,
    # files created by previous runs (under the old UID) become inaccessible.
    # Always chown -R when UID was remapped; otherwise only if top-level is wrong.
    actual_hermes_uid=$(id -u hermes)
    needs_chown=false
    if [ -n "$HERMES_UID" ] && [ "$HERMES_UID" != "10000" ]; then
        needs_chown=true
    elif [ "$(stat -c %u "$HERMES_HOME" 2>/dev/null)" != "$actual_hermes_uid" ]; then
        needs_chown=true
    fi
    if [ "$needs_chown" = true ]; then
        echo "Fixing ownership of $HERMES_HOME to hermes ($actual_hermes_uid)"
        # In rootless Podman the container's "root" is mapped to an unprivileged
        # host UID — chown will fail.  That's fine: the volume is already owned
        # by the mapped user on the host side.
        chown -R hermes:hermes "$HERMES_HOME" 2>/dev/null || \
            echo "Warning: chown failed (rootless container?) — continuing anyway"
    fi
    echo "Dropping root privileges"
    exec gosu hermes "$0" "$@"
 fi
 # --- Running as hermes from here ---
 source "${INSTALL_DIR}/.venv/bin/activate"
 # Create essential directory structure.  Cache and platform directories
 # (cache/images, cache/audio, platforms/whatsapp, etc.) are created on
 # demand by the application — don't pre-create them here so new installs
 # get the consolidated layout from get_hermes_dir().
 # The "home/" subdirectory is a per-profile HOME for subprocesses (git,
 # ssh, gh, npm …).  Without it those tools write to /root which is
 # ephemeral and shared across profiles.  See issue #4426.
 mkdir -p "$HERMES_HOME"/{cron,sessions,logs,hooks,memories,skills,skins,plans,workspace,home}
 # .env
 if [ ! -f "$HERMES_HOME/.env" ]; then
    cp "$INSTALL_DIR/.env.example" "$HERMES_HOME/.env"
 fi
 # config.yaml
 if [ ! -f "$HERMES_HOME/config.yaml" ]; then
    cp "$INSTALL_DIR/cli-config.yaml.example" "$HERMES_HOME/config.yaml"
 fi
 # Ensure the main config file remains accessible to the hermes runtime user
 # even if it was edited on the host after initial ownership setup.
 if [ -f "$HERMES_HOME/config.yaml" ]; then
    chown hermes:hermes "$HERMES_HOME/config.yaml"
    chmod 640 "$HERMES_HOME/config.yaml"
 fi
 # SOUL.md
 if [ ! -f "$HERMES_HOME/SOUL.md" ]; then
    cp "$INSTALL_DIR/docker/SOUL.md" "$HERMES_HOME/SOUL.md"
 fi
 # Sync bundled skills (manifest-based so user edits are preserved)
 if [ -d "$INSTALL_DIR/skills" ]; then
    python3 "$INSTALL_DIR/tools/skills_sync.py"
 fi
 # Final exec: two supported invocation patterns.
 #
 #   docker run <image>                 -> exec `hermes` with no args (legacy default)
 #   docker run <image> chat -q "..."   -> exec `hermes chat -q "..."` (legacy wrap)
 #   docker run <image> sleep infinity  -> exec `sleep infinity` directly
 #   docker run <image> bash            -> exec `bash` directly
 #
 # If the first positional arg resolves to an executable on PATH, we assume the
 # caller wants to run it directly (needed by the launcher which runs long-lived
 # `sleep infinity` sandbox containers — see tools/environments/docker.py).
 # Otherwise we treat the args as a hermes subcommand and wrap with `hermes`,
 # preserving the documented `docker run <image> <subcommand>` behavior.
 if [ $# -gt 0 ] && command -v "$1" >/dev/null 2>&1; then
    exec "$@"
 fi
 exec hermes "$@"
--- a/flake.nix
+++ b/flake.nix
@@ -61,7 +61,6 @@
              ./modules/nixos/services/open_code_server.nix
              ./modules/nixos/services/ollama_init_custom_models.nix
              ./modules/nixos/services/openclaw_node.nix
              ./modules/nixos/security/ai-worker-restricted.nix
              ./users/gortium.nix
              ./users/ai-worker.nix
            ];
--- a/hosts/lazyworkhorse/configuration.nix
+++ b/hosts/lazyworkhorse/configuration.nix
@@ -36,7 +36,7 @@
    "transparent_hugepage=always" # because mucho ram
  ];
  # 2. Load the specific drivers found by sensors-detect
-  boot.kernelModules = [ "nct6775" "lm96163" "iptable_nat" "iptable_filter" ];
+  boot.kernelModules = [ "nct6775" "lm96163" ];
  # 3. Force the nct6775 driver to recognize the chip if it's stubborn
  boot.extraModprobeConfig = ''
    options nct6775 force_id=0xd280
@@ -49,26 +49,6 @@
  networking.networkmanager.enable = true;  # Easiest to use and most distros use this by default.
  networking.hostId = "deadbeef";
  # WireGuard VPN client -- always up, connects to wg-easy server
  # Create age-encrypted secrets before deploying (run on the host):
  #   echo -n "<private_key>" | agenix -e secrets/wireguard_private_key.age
  #   echo -n "<preshared_key>" | agenix -e secrets/wireguard_preshared_key.age
  networking.wireguard.interfaces = {
    wg0 = {
      ips = [ "10.8.0.3/24" ];
      privateKeyFile = config.age.secrets.wireguard_private_key.path;
      peers = [
        {
          publicKey = "rY9zII3AOm8rog2rv02PyA3Bq7zdvTOGkZapfCV1DkE=";
          presharedKeyFile = config.age.secrets.wireguard_preshared_key.path;
          allowedIPs = [ "10.8.0.0/24" ];
          endpoint = "vpn.lazyworkhorse.net:51820";
          persistentKeepalive = 25;
        }
      ];
    };
  };
  # Set your time zone.
  time.timeZone = "America/Montreal";
@@ -178,7 +158,7 @@
    settings = {
      PasswordAuthentication = false;
      KbdInteractiveAuthentication = false;
-      # Additional hardening settings below in SERVER HARDENING section
+      PermitRootLogin = "prohibit-password"; 
    };
    hostKeys = [
      {
@@ -207,7 +187,6 @@
    ai = {
      path = self + "/assets/compose/ai";
      envFile = config.age.secrets.containers_env.path;
      ports = [ 22000 ];  # Syncthing TCP sync
    };
    cloudstorage = {
@@ -242,11 +221,6 @@
      path = self + "/assets/compose/homepage";
    };
    vpn = {
      path = self + "/assets/compose/vpn";
      envFile = config.age.secrets.containers_env.path;
    };
    # tak = {
    #   path = self + "/assets/compose/tak";
    # };
@@ -290,20 +264,6 @@
        mode = "0440";
        path = "/run/secrets/openclaw_gateway_token";
      };
      wireguard_private_key = {
        file = ../../secrets/wireguard_private_key.age;
        owner = "root";
        group = "root";
        mode = "0400";
        path = "/run/secrets/wireguard_private_key";
      };
      wireguard_preshared_key = {
        file = ../../secrets/wireguard_preshared_key.age;
        owner = "root";
        group = "root";
        mode = "0400";
        path = "/run/secrets/wireguard_preshared_key";
      };
    };
  };
@@ -348,203 +308,6 @@
  # Or disable the firewall altogether.
  # networking.firewall.enable = false;
  # =============================================================================
  # SERVER HARDENING - Firewall, Fail2ban, SSH, Kernel
  # =============================================================================
  # Firewall - default deny, explicit allow
  networking.firewall = {
    # Enable firewall with default deny policy (NixOS firewall denies all by default)
    enable = true;
    allowPing = true;
    # Only essential ports exposed to internet
    allowedTCPPorts = [
      2424  # SSH (non-standard port)
      2222  # Gitea (version control)
      80    # HTTP (Traefik redirect)
      443   # HTTPS (Traefik)
      # 8000  # Portainer - REVIEW: internal only?
      # 4242  # Coms - REVIEW: internal only?
      # 5000  # TAK API - REVIEW: internal only?
      # 8087  # TAK Connect - REVIEW: internal only?
      # 8089  # TAK Management - REVIEW: internal only?
    ];
    allowedUDPPorts = [
      51820 # WireGuard VPN
    ];
    # Rate limiting and attack prevention
    extraCommands = ''
      # 1. Wipe the INPUT chain clean at the start of every activation
      iptables -F INPUT
      # Rate limit SSH connections (max 20 new connections per 60 seconds)
      iptables -A INPUT -p tcp --dport 2424 -m state --state NEW -m recent --set
      iptables -A INPUT -p tcp --dport 2424 -m state --state NEW -m recent --update --seconds 60 --hitcount 20 -j DROP
      # Rate limit HTTP/HTTPS (protects Traefik)
      iptables -A INPUT -p tcp --dport 80 -m state --state NEW -m limit --limit 25/minute --limit-burst 100 -j ACCEPT
      iptables -A INPUT -p tcp --dport 443 -m state --state NEW -m limit --limit 25/minute --limit-burst 100 -j ACCEPT
      # Drop invalid packets
      iptables -A INPUT -m state --state INVALID -j DROP
      # Log dropped packets (rate limited)
      iptables -A INPUT -m limit --limit 5/min -j LOG --log-prefix "IPTables-Dropped: " --log-level 4
      # 3. CRITICAL: Re-link the NixOS default firewall chain
      # Without this line, the 'allowedTCPPorts' in your Nix config will be ignored!
      iptables -A INPUT -j nixos-fw
    '';
  };
  # Fail2ban - automatic IP banning
  services.fail2ban = {
    enable = true;
    maxretry = 3;
    bantime = "1h";
    banaction = "iptables-multiport";
    jails = {
      # SSH brute force protection (uses systemd journal backend)
      sshd = {
        enabled = true;
        settings = {
          filter = "sshd";
          port = "2424";
          maxretry = 3;
          bantime = "1h";
        };
      };
      # Recidive - ban repeat offenders for 1 week
      recidive = {
        enabled = true;
        settings = {
          filter = "recidive";
          logpath = "/var/log/fail2ban.log";
          bantime = "1w";
          findtime = "1d";
          maxretry = 3;
        };
      };
      # HTTP authentication failures (Traefik)
      http-auth = {
        enabled = true;
        settings = {
          filter = "traefik-auth";
          port = "80,443";
          logpath = "/var/log/traefik/access.log";
          maxretry = 5;
          bantime = "1h";
        };
      };
      # HTTP scanning/attacks (Traefik)
      http-botsearch = {
        enabled = true;
        settings = {
          filter = "traefik-botsearch";
          port = "80,443";
          logpath = "/var/log/traefik/access.log";
          maxretry = 2;
          bantime = "2h";
        };
      };
    };
  };
  # Custom fail2ban filters for Traefik
  environment.etc."fail2ban/filter.d/traefik-auth.conf".text = ''
    [Definition]
    failregex = ^<HOST> -.*"(GET|POST|HEAD|PUT|DELETE).*" (401|403) \d+.*$
    ignoreregex =
  '';
  environment.etc."fail2ban/filter.d/traefik-botsearch.conf".text = ''
    [Definition]
    failregex = ^<HOST> -.*"(GET|POST|HEAD|PUT|DELETE).*" 404 \d+.*$
                ^<HOST> -.*"(GET|POST|HEAD|PUT|DELETE).*/(\.|wp-|php|admin|login|xmlrpc|\.env|\.git|\.aws|\.azure).*" \d+.*$
    ignoreregex =
  '';
  # SSH hardening
  services.openssh.settings = {
    PermitRootLogin = "no";
    MaxAuthTries = 3;
    MaxSessions = 20;
    LoginGraceTime = 30;
    ClientAliveInterval = 300;
    ClientAliveCountMax = 2;
    PermitEmptyPasswords = "no";
    ChallengeResponseAuthentication = "no";
    UsePAM = true;
    LogLevel = "VERBOSE";
    X11Forwarding = false;
    AllowTcpForwarding = "no";
    AllowAgentForwarding = "no";
    PermitTunnel = "no";
  };
  # Kernel network hardening
  boot.kernel.sysctl = {
    # IP Spoofing protection
    "net.ipv4.conf.all.rp_filter" = 1;
    "net.ipv4.conf.default.rp_filter" = 1;
    # Ignore ICMP broadcasts
    "net.ipv4.icmp_echo_ignore_broadcasts" = 1;
    # Disable source routing
    "net.ipv4.conf.all.accept_source_route" = 0;
    "net.ipv4.conf.default.accept_source_route" = 0;
    "net.ipv6.conf.all.accept_source_route" = 0;
    "net.ipv6.conf.default.accept_source_route" = 0;
    # Disable redirects
    "net.ipv4.conf.all.send_redirects" = 0;
    "net.ipv4.conf.default.send_redirects" = 0;
    # SYN flood protection
    "net.ipv4.tcp_syncookies" = 1;
    "net.ipv4.tcp_max_syn_backlog" = 2048;
    "net.ipv4.tcp_synack_retries" = 2;
    "net.ipv4.tcp_syn_retries" = 5;
    # Log martian packets
    "net.ipv4.conf.all.log_martians" = 1;
    "net.ipv4.conf.default.log_martians" = 1;
    # Ignore redirects
    "net.ipv4.conf.all.accept_redirects" = 0;
    "net.ipv4.conf.default.accept_redirects" = 0;
    "net.ipv4.conf.all.secure_redirects" = 0;
    "net.ipv4.conf.default.secure_redirects" = 0;
    "net.ipv6.conf.all.accept_redirects" = 0;
    "net.ipv6.conf.default.accept_redirects" = 0;
    # Connection tuning
    "net.core.somaxconn" = 4096;
    "net.core.netdev_max_backlog" = 65536;
    "net.ipv4.tcp_max_orphans" = 65536;
    "net.ipv4.tcp_fin_timeout" = 15;
    "net.ipv4.tcp_keepalive_time" = 300;
    "net.ipv4.tcp_keepalive_probes" = 5;
    "net.ipv4.tcp_keepalive_intvl" = 15;
  };
  # Audit logging
  security.auditd.enable = true;
  # Fail2ban log directory
  systemd.tmpfiles.rules = [
    "d /var/log/fail2ban 0755 root root -"
    "d /var/log/traefik 0755 root root -"
  ];
  # Copy the NixOS configuration file and link it from the resulting system
  # (/run/current-system/configuration.nix). This is useful in case you
  # accidentally delete configuration.nix.
--- a/modules/nixos/security/README-ai-worker.md
+++ b/modules/nixos/security/README-ai-worker.md
@@ -1,105 +0,0 @@
 # AI Worker Restricted Access
 This module provides SSH access for the AI worker (hermes-agent) to run ollama benchmarks on the host.
 ## Security Model
 The `ai-worker` user has:
 ### Filesystem Access
 - **Home directory**: `/home/ai-worker` (standard user home)
 - **No bind mounts**: Cannot access `/home/gortium/infra` or other host files
 - **Cannot access**: Any files outside standard system paths
 ### Sudo Access
 - **NONE**: ai-worker has no sudo privileges
 - Cannot run `nh`, `nixos-rebuild`, `nixpkgs-fmt`, or `nix` with elevated permissions
 ### Docker Access
 - Member of `docker` group - can run `docker` and `docker exec` commands
 - Primary use: `docker exec ollama ollama ...` for benchmarking
 - Can run `docker exec --privileged ollama rocm-smi ...` for VRAM monitoring
 ## Workflow: SSH + Docker Benchmarking
 The AI worker connects from the Hermes container to the host via SSH, runs ollama benchmarks, then returns to save results.
 ### Example Workflow
 ```bash
 # From Hermes container, SSH to host
 ssh -i /path/to/ssh/key ai-worker@host.docker.internal
 # On host, run ollama benchmarks via docker
 docker exec ollama ollama pull devstral-small-2:24b
 # Create test modelfile
 docker exec ollama bash -c 'cat <<EOF > /root/.ollama/test.modelfile
 FROM devstral-small-2:24b
 PARAMETER num_ctx 65536
 PARAMETER num_gpu 99
 PARAMETER flash_attn true
 EOF'
 # Create and test model
 docker exec ollama ollama create test-model -f /root/.ollama/test.modelfile
 docker exec ollama ollama run test-model "Write a Python async function"
 # Check VRAM usage
 docker exec --privileged ollama rocm-smi --showmeminfo vram
 # Cleanup
 docker exec ollama ollama rm test-model
 # Exit SSH, return to Hermes container
 exit
 # Save results in Hermes container
 # /opt/data/ai-optimizer/state.json
 # /opt/data/ai-optimizer/results.csv
 ```
 ## SSH Access
 Connect as:
 ```bash
 ssh ai-worker@lazyworkhorse
 ```
 The working directory will be `/home/ai-worker`. No infra repo access.
 ## Verification
 Check ai-worker permissions:
 ```bash
 # On the host, as root or gortium:
 sudo -u ai-worker sudo -l
 # Should show: no sudo access
 # Check docker group membership
 groups ai-worker
 # Should show: ai-worker docker
 ```
 ## Troubleshooting
 If ai-worker cannot run docker commands:
 ```bash
 # Check docker group membership
 groups ai-worker
 # Verify ollama container is running
 docker ps | grep ollama
 # Test docker access
 sudo -u ai-worker docker exec ollama ollama list
 ```
 If SSH connection fails:
 ```bash
 # Check SSH key is authorized
 cat /home/ai-worker/.ssh/authorized_keys
 # Check SSH service
 systemctl status sshd
 ```
--- a/modules/nixos/security/ai-worker-restricted.nix
+++ b/modules/nixos/security/ai-worker-restricted.nix
@@ -1,17 +0,0 @@
 { config, pkgs, lib, ... }:
 with lib;
 {
  options.services.aiWorkerAccess = mkOption {
    type = types.bool;
    default = false;
    description = "Enable AI worker SSH access with docker group membership for ollama benchmarking";
  };
  config = mkIf config.services.aiWorkerAccess {
    # ai-worker is member of docker group - can run docker commands via SSH
    # No bind mounts, no sudo access - docker-only for ollama benchmarking
    users.groups.docker.members = [ "ai-worker" ];
  };
 }
--- a/modules/nixos/services/ollama_init_custom_models.nix
+++ b/modules/nixos/services/ollama_init_custom_models.nix
@@ -1,87 +1,45 @@
 { pkgs, ... }: {
  systemd.services.init-ollama-model = {
    description = "Initialize LLM models with extra context in Ollama Docker";
-    
+    after = [ "docker-ollama.service" ];
    # On s'assure que Docker tourne avant de lancer ce script
    after = [ "docker.service" ];
    wantedBy = [ "multi-user.target" ];
    script = ''
-      # Fonction de création asynchrone pour ne pas bloquer le démarrage
+      # Wait for Ollama
-      (
+      while ! ${pkgs.curl}/bin/curl -s http://localhost:11434/api/tags > /dev/null; do
-        echo "Starting asynchronous Ollama initialization..."
+        sleep 2
      done
-        # Attente d'Ollama (maximum 120 secondes pour éviter une boucle infinie)
+      create_model_if_missing() {
-        TIMEOUT=60
+        local model_name=$1
-        COUNT=0
+        local base_model=$2
-        while ! ${pkgs.curl}/bin/curl -s -f http://127.0.0.1:11434/api/tags > /dev/null; do
+        if ! ${pkgs.docker}/bin/docker exec ollama ollama list | grep -q "$model_name"; then
-          if [ $COUNT -ge $TIMEOUT ]; then
+          echo "$model_name not found, creating from $base_model..."
-            echo "Ollama did not become ready in time. Exiting."
+          ${pkgs.docker}/bin/docker exec ollama sh -c "cat <<EOF > /root/.ollama/$model_name.modelfile
            exit 1
          fi
          echo "Waiting for Ollama API to be reachable..."
          sleep 5
          COUNT=$((COUNT + 5))
        done
        create_model_if_missing() {
          local model_name=$1
          local base_model=$2
          # Vérification robuste via l'API HTTP d'Ollama plutôt que docker exec (évite les conflits de tty)
          if ! ${pkgs.curl}/bin/curl -s http://127.0.0.1:11434/api/tags | ${pkgs.jq}/bin/jq -e ".models[] | select(.name == \"$model_name\")" > /dev/null; then
            echo "$model_name not found, creating from $base_model..."
            # Utilisation d'un fichier temporaire sur l'hôte pour l'injecter proprement dans Docker
            TMP_FILE=$(mktemp)
            cat <<EOF > "$TMP_FILE"
 FROM $base_model
 TEMPLATE """{{- if .System }}
 [SYSTEM_PROMPT]
 {{ .System }}
 [/SYSTEM_PROMPT]
 {{- end }}
 {{- range .Messages }}
 {{- if eq .Role "user" }}
 [INST]
 {{ .Content }}
 [/INST]
 {{- else if eq .Role "assistant" }}
 {{ .Content }}
 {{- end }}
 {{- end }}"""
 PARAMETER num_ctx 131072
 PARAMETER num_predict 4096
 PARAMETER num_keep 1024
 PARAMETER repeat_penalty 1.1
 PARAMETER top_k 40
-PARAMETER stop "[INST]"
+PARAMETER stop \"[INST]\"
-PARAMETER stop "[/INST]"
+PARAMETER stop \"[/INST]\"
-PARAMETER stop "</s>"
+PARAMETER stop \"</s>\"
-EOF
+EOF"
          ${pkgs.docker}/bin/docker exec ollama ollama create "$model_name" -f "/root/.ollama/$model_name.modelfile"
        else
          echo "$model_name already exists, skipping."
        fi
      }
-            # Copie et création dans le conteneur
+      # Create Nemotron
-            ${pkgs.docker}/bin/docker cp "$TMP_FILE" ollama:/tmp/model.modelfile
+      create_model_if_missing "nemotron-3-nano:30b-128k" "nemotron-3-nano:30b"
            ${pkgs.docker}/bin/docker exec ollama ollama create "$model_name" -f /tmp/model.modelfile
            ${pkgs.docker}/bin/docker exec ollama rm /tmp/model.modelfile
            rm -f "$TMP_FILE"
          else
            echo "$model_name already exists, skipping."
          fi
        }
-        # Create Nemotron
+      # Create Devstral
-        create_model_if_missing "nemotron-3-nano:30b-128k" "nemotron-3-nano:30b"
+      create_model_if_missing "devstral-small-2:24b-128k" "devstral-small-2:24b" 
        # Create Devstral
        create_model_if_missing "devstral-small-2:24b-128k" "devstral-small-2:24b" 
      ) &
    '';
    serviceConfig = {
-      Type = "forking"; # Permet à systemd de savoir que le script passe en arrière-plan via '&'
+      Type = "oneshot";
-      User = "root";
+      RemainAfterExit = true;
    };
  };
 }
--- a/secrets/containers.env.age
+++ b/secrets/containers.env.age
@@ -1,36 +1,34 @@
 -----BEGIN AGE ENCRYPTED FILE-----
-YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IHNzaC1lZDI1NTE5IEdoTUQ4QSBWNEpt
+YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IHNzaC1lZDI1NTE5IEdoTUQ4QSBOL29w
-cGFNeVBBaDRqb3pLSEZGQW0wb3VmVnBoZCswUFkzbnBLUnJ0QTNrClRqVkk4RUVO
+eGk1N2xxTHJtaUEvWWZmbkh1bk11Tjk3anNnMDB1cCtPYUMzdTNJCkdhQ08vblNG
-d29KYjd5YUcwankvaTFmVHUxQVpDT2ZYWHRaY3JXTUtQMU0KLT4gKXBtQ3UsXi1n
+UlV1K2xVTGZVTzFWYXAzcjZaMWs0RTFWdStKSmlSTURvK1EKLT4gLC1zKU8zVkgt
-cmVhc2UgNnwxYCBXVyA/KCQmIHt9NAo3OTZVUHR2UXkvaEFwY0ZBdEJsaFpsbHJ0
+Z3JlYXNlIFUiXFcpS302IHByVn5jOy0gRDMKQjV3SHpDWUIybGFyQUg3ZlR0R2hV
-cklKcDVHcEdWMEdPSkpnN0FiRU43RW5hUWFMdjR3WFRRSFBLSGlmClM3cTNJWlNM
+eWM3SFlCVW5mdlpBVUF3a0xpNlZCeGNUd1oxTTlkc1RkTXdZS0lFTmN3Ci0tLSA3
-TExkdHdXUHJISkNIaE1TTUxUc1NUWkV1a09HeFU3bVZwQXMKLS0tIGhOcXFTUElS
+VlBqM1VLWllZc0JnOTMvUFRjMU13OTdzMmhsdGJubkk5eGpERVVLYUk4Cnzh5UbU
-azJJNnEreUhMWTJBaVZGSTJPRUFqQkVYS01KRENUVVpZSDgK8+8onFejroBo7MeO
+FlgqpM8jkJ6XlsaIDCw/G3D6uJ/GRJW4gIekuhAUxpZJrc8eOA8ZuHfGrBbH3acV
-dW+so4lOsq4zJKn3f0cxmCFg1f0X8zt6h4Uc3A5Cvr1uU+6yw1FWmJ7xa3jJz3lO
+tVafX5F0Kr2oOblqZ6gduZOUS52KmWH8stiBJM+e5ZZ7zRQVE4PJUKUPCzi+WdcH
-EEaKQJXYC+xIIKGcA7qILa0SFp4a/4OuYjcg27HrlPhg7u5wDhQrd0LdVEe1Xngp
+zr295T//FOdicrYHdsjfziKEHzBtUCFiATW05+O2zMjYjO6cPzePcCzPWinwiID6
-ZivX7P7HwIna3X8C+TL+K2v/AG2N/z86cdKfRvxyMKNbHhYw+CfHEnWgh8tJ++4h
+V+f6ngfkkQaj3wBGkzaieQJzRcdSwky21aVhGCCX/bvqx61iW2d5QAKxGbtQ2RcG
-G9evNniuNqte6cQaRe7jODfPNW4FuY/Sb7barlJ/M9iAQdYAdyLAzU1LABeHeUfD
+X1okr+xunAM94nzDMv46vyN97KxY7cZd4pAaOxoICc2Tfhtw6F+iS6QkQh1odJzO
-wtHjxy9DUZ55Vg8bB8M2JJU9MkoRT4ewiVd9LeC1GWeVmKsm93wsmrov714i7U2j
+7ZH+sSQCvndG+8z9shXGiHalASF5tdguM+JlEvAGljcaiAUtsQWxr9CoWiEkC6c6
-wHtDkjqEF2MmzuQc18sjNaAHiwz8j6o5xU2L/Q4+Q707yISWG7RGZYh389Cr1rnw
+NCaECSYO8Il+SXBQnSZSGJSNDhuPYCYrsjXGSAONFixuyeslAkq9x2WUaUS4H063
-siUq/Vunqw2wk13+J/4vu9nqt5mMktBaCtp+QiWIurjwB5LUAyChrSm+dg5lb0Mt
+1QvRF7XO2tBPtgCLsSjdiGp0h+ImUaGdu6fDR7zrDsGsaAFCSFeH/rGNNXRQ2vP2
-UhSc0lq1+E3vxAXM2Hmk+vP86VD+6WJvAU82VFApF1s6zG2FU1/AcOVVf54nan/q
+CSfPfDDCqpUSCn0WuA30BtaPLxGmZT6OjFevKzYMNDmdeq9ia/q8K0hmjLUBdN3k
-f+rgSFfASHQCYSblUJHyEtwLNsWEmTGmOEn1buUKD/H0zatPQnc0rYpjlx2V0Sjd
+tdYWbwoaf4gYbUWxSleD768b0Jgxss9Vod+sFQ+NYRksdGIeyND+aQIc312XehfA
-6yB5+wPrZ0AkN1pjcsPKOv8Kaog2DzqIjib+SaSTaRxWHQEb9uzvaReAcYI5HOpE
+qHFBS8nlj7eUF5bdvCYQ64z741mH4cNlGxyjPBH1x8FHnEOocJXYt1l2AZSRJmJA
-gkC040HN33BItATbo4+hz70Im8Ni/VXD+g6yzM6Hj1hJL+PinTKeg5keQRFIZjMx
+c3z0QGXyuCbsrLBXWK1EKa/Juo4PGGsEVoLRhwJAQy9+i1JN0yrfRvSPyzvD4px6
-grzievB2wVBBgLgN3qMdTFmpplaL7iL702JjXZUTTK9Izp+9wiCsV1fTa53FWDht
+wRPzlZ80MQdb2lv84WS/zcOEZmZzlLntszTRRdIfAsuaavP2Rquh4rEXABYeTZwp
-ylFL5SWElqXjK+QBXxAe+Jk6VQov5HI21YDXL67S554ABeRok23wxrQ31TCI4xq9
+5dem79s8bdW2nFsGMNz1OQKQwocyjYu1jJMHu6Gp7Ngdl1xyW7xfg0dezE1c0cIh
-PQV7VtNRjyVud7S29m3OwpWOsgTZhn+JclHj2v4bNJzJkJnZRTmcvGPktzRI5+R4
+xt1aLER9YJp4n5to5cOH16l3mjDHnAvABx38xE9loNL3399J/evw7LxpTYQ4v2Xv
-e5vxVhGnJDzI71txaHl8+xS1lu9VzCQUrxX6TXyTRV4KjIOz0g06JOBgmBRBvJca
+x8xnDHcqJ+deFSwyuUnMS5DkUeYuHmUl0Q2WYcfY+ibCmcgCb2ObTtuN1/ZxNYrL
-7MZbC65xpisl/gyLRbgkVga3t94dPV+dpZsn8eq6427IyRbKslJefatggR9//c6I
+OKrnmfuSvBgyuIOj5e6uWW0+Zs8dHKXu2TgV8WignxOhl5zQgCpCBlqVfO0t+NCu
-5N5fl0fR3gJQMB+HRbipBH2YsdbdWJyb4Nn6STZxIfrqoG/xC6C1raF0xK7hUx6i
+Gi26hU/fhGWQ/1oQa3VkpGsypZbJpgQvfWxfcGHP/MMhnl01zzlP8/aexSY3pAxf
-4DUDSPohM8fOIswQPfE+FH3eygfzu/Ln5+ghsgHTEhgFvmgMvyxaAt6kHIzIUhMX
+fz9v0IVh6xxtu3zbiiVzUsXbfG7t+xY98jMphf4AS2mWva3GWVmhhu0lS3J3P+go
-M3dASr4VPDpIXuXsRWwYLEifhzxsuvwVxfwtsnCaR6XKijsYECWGDdYOWHdleeqx
+YEEP4rOFHeU0Y1/6kLydTXvz4jMH0H92XQIzshd7vzQnEJPUPAzqRmw3LKYGgCI+
-wDPhxEesfFVhKxhrKY9Ir8k9/FFBKQU/3GjW4+SMAg5Al1YEzxshP9vKuVcsei7W
+wZEnxJ6ckqTkGBFnxTpy9LLllwmnz2Ky87nY3XAmqxlhb2Ap1XFAlfgszmGjc+Il
-JDwAwotNXaCm6NBckiyZJE53ou6+gckPY7V9cOfnuH74Z9ywkFzB3HW3ZlonaGyM
+KkIgoWQHTUm6QM9ta++oUTIDneOvxGd0zZsqoEhiC/7E01BNNZ6E58TeJU3fDlA3
-oGmLGcccavFtyhg5s/As4i6X8ARIpDiwe59Pn3GNXMctySqIrrr2ogUoXgrfFCie
+mX6n05XjwPRpgXZfayPoAgBlZc2H4KeiynxwNZ/dWu7qz7L6Ppk6Nvtly8giTbFx
-6GOTdeMW7GeOSdJUxCofghlspS/nq01Og77VI/beWYrIwLubSka6Zaltww9zgObk
+CA+tto7vq+D+CAEJ4bgyq4BCH4GL4APrhPcWp98Mko1WCiRTIKgkZxQCYvlg/LZq
-/FGEMgFkEpq7iyCvYSPA8F46pJKvnMP3S84AWCPmcTcHeg4lwGPvs6btexXBGdoz
+LNhMacP9T1qTvNC+yR1NEMiegE3APzk6CkDpVaO9+5f/sqifNPINCMothenI9ePw
-nkCyq7wdH5Nngm7jUbl88LtaLZPAQkuqXphBVTnrF9Ofbnb4iRZ2Op4xpx9rGyvx
+zjQLI3Mo1m73bkomytUZ7i1VstP5sEZ5LF72Sq7BpR3oQ3Gp0CAN9w==
 mO6UEhL6V1i2YZFNkNMg/W8aoMiUgBdqbkxaxblT9L0aNdlFU9+LbWYolURVEadd
 Qjv0Z1gMA+tsuBbVszwsMfneZ5+B9Q==
 -----END AGE ENCRYPTED FILE-----
--- a/secrets/wireguard_preshared_key.age
+++ b/secrets/wireguard_preshared_key.age
@@ -1,9 +0,0 @@
 -----BEGIN AGE ENCRYPTED FILE-----
 YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IHNzaC1lZDI1NTE5IEdoTUQ4QSA3VG9Z
 MVFPVFc2VVJ3d0h0dmtBUnI3WHl2SzUxTkRZbjFCaGloWmV3dnd3ClcxdnVPeGd6
 SU4zR0Q0K1dtVjRRVHd0VW5XSFI0dVFpTjZnYk1DNjRxTVEKLT4gQzlgRy1ncmVh
 c2UKeUozOWgyUytSTVF0NjY2STBEb2VadwotLS0gblI3bmJCUWxxU3QrYTEyVFBI
 Snc4NC9rTkh0NnZYbUtxUE9hRWRkelpmMAq58fmH6cK13GeD7wGLxKmx10hmJeW4
 b7KqnCD1ZP7uG85s32xzVRwRG8RrG4xZo5nR9Mrtg1CoTSFfUGeFnf5xveN+Ej0X
 wDVB1LwC+Q==
 -----END AGE ENCRYPTED FILE-----
--- a/secrets/wireguard_private_key.age
+++ b/secrets/wireguard_private_key.age
@@ -1,11 +0,0 @@
 -----BEGIN AGE ENCRYPTED FILE-----
 YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IHNzaC1lZDI1NTE5IEdoTUQ4QSA5dzVG
 WUNvT3NlRmcrWS81bzJqSWlTekVYaDFFTE10SkI2dEgzaGpxcUI4Cmk5Y0FGYTRZ
 K0NGYzY3VUp4aS9ZZGRmWTgybDJFUURva2pZNmVOS3QxdEUKLT4gPnVRTCtldGMt
 Z3JlYXNlCk04OTJZeFRNeDI5aGpMVTk1ZTE0Y2FMMnFEMjlJalJpMHRlaTE4ZWIx
 d2lCRGQ5RHVjcktOMGJCb1VERlNWcTYKaSt0L1Z6dVJ0QWIyZkhsYzFEVjZSQWUr
 ZWpwVlo1TmhoUFJZdkEvR0gxNlVhcXF2ZTRnCi0tLSBLcmM2MThNVkdWclpHUXRr
 VTF6QVk2WUZlTXpZMVNLMlpBOFc3M1o5WjZzCs9xbPlIX+u5vRSQ/z9utu+I9S2c
 02DOsIb1kzxzb1OK91b8Kh4JucQSq3qkyEvRucsNn5QW8hIHDnRuND6EbPyN7p4S
 YB/F0dxSqgnq
 -----END AGE ENCRYPTED FILE-----
--- a/users/ai-worker.nix
+++ b/users/ai-worker.nix
@@ -9,85 +9,6 @@
    openssh.authorizedKeys.keys = [
      keys.users.ai-worker.main
    ];
    # No password login - SSH key only
    hashedPassword = "!";
  };
  users.groups.ai-worker = {};
  # Enable restricted AI worker SSH access for ollama benchmarking
  # SECURITY: ai-worker can only:
  #   - SSH into host from Hermes container
  #   - Run docker commands (docker exec ollama ...) via docker group
  #   - Run specific security audit commands
  #   - NO access to infra repo (no bind mount)
  #   - NO sudo access (no nh, nixos-rebuild, nixpkgs-fmt, nix)
  # WORKFLOW: SSH from Hermes container, run docker benchmarks, return and save results to /opt/data/ai-optimizer/
  services.aiWorkerAccess = true;
  # Restricted sudo for ai-worker - security checks only
  security.sudo.extraRules = [
    {
      users = [ "ai-worker" ];
      commands = [
        # Firewall checks
        {
          command = "/run/wrappers/bin/sudo iptables -L -n -v";
          options = [ "NOPASSWD" ];
        }
        {
          command = "/run/wrappers/bin/sudo iptables -S";
          options = [ "NOPASSWD" ];
        }
        # Fail2ban status
        {
          command = "/run/current-system/sw/bin/fail2ban-client status";
          options = [ "NOPASSWD" ];
        }
        {
          command = "/run/current-system/sw/bin/fail2ban-client status *";
          options = [ "NOPASSWD" ];
        }
        {
          command = "/run/current-system/sw/bin/fail2ban-client get * banned";
          options = [ "NOPASSWD" ];
        }
        # Log inspection
        {
          command = "/run/current-system/sw/bin/journalctl -t kernel -n 100";
          options = [ "NOPASSWD" ];
        }
        {
          command = "/run/current-system/sw/bin/journalctl -u fail2ban -n 50";
          options = [ "NOPASSWD" ];
        }
        {
          command = "/run/current-system/sw/bin/journalctl -u firewall -n 50";
          options = [ "NOPASSWD" ];
        }
        # SSH config verification
        {
          command = "/run/current-system/sw/bin/sshd -T";
          options = [ "NOPASSWD" ];
        }
        # Docker service checks
        {
          command = "/run/current-system/sw/bin/docker ps";
          options = [ "NOPASSWD" ];
        }
        {
          command = "/run/current-system/sw/bin/docker inspect *";
          options = [ "NOPASSWD" ];
        }
        # Network diagnostics
        {
          command = "/run/current-system/sw/bin/ss -tlnp";
          options = [ "NOPASSWD" ];
        }
        {
          command = "/run/current-system/sw/bin/cat /proc/net/tcp";
          options = [ "NOPASSWD" ];
        }
      ];
    }
  ];
 }
		`@@ -0,0 +1 @@`
							`timestamp,track,model,backend,phase,num_ctx,num_gpu,flash_attn,tokens_per_sec,vram_gb,ram_gb,status,is_best`