Merge pull request #4701 from pewdiepie-archdaemon/sync-dev-from-main-20260622

chore(dev): sync main cookbook and model workflow fixes

Dockerfile

@@ -20,6 +20,23 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     gosu \
     && rm -rf /var/lib/apt/lists/*
 
+# Docker CLI (client only — daemon stays on the host via the
+# /var/run/docker.sock mount). The Debian `docker.io` package ships
+# dockerd but not the client binary on slim, so grab the static client
+# tarball from download.docker.com instead.
+ARG DOCKER_CLI_VERSION=27.5.1
+RUN ARCH="$(dpkg --print-architecture)" \
+    && case "$ARCH" in \
+         amd64) DARCH=x86_64 ;; \
+         arm64) DARCH=aarch64 ;; \
+         *) echo "unsupported arch $ARCH"; exit 1 ;; \
+       esac \
+    && curl -fsSL "https://download.docker.com/linux/static/stable/${DARCH}/docker-${DOCKER_CLI_VERSION}.tgz" \
+       -o /tmp/docker.tgz \
+    && tar -xzf /tmp/docker.tgz -C /tmp \
+    && install -m 0755 /tmp/docker/docker /usr/local/bin/docker \
+    && rm -rf /tmp/docker /tmp/docker.tgz
+
 WORKDIR /app
 
 # Install Python deps first (layer cache). Optional extras (PyMuPDF AGPL, etc.)

docker-compose.gpu-amd.yml

@@ -28,6 +28,14 @@ services:
       # land under /app/.local for the odysseus user. Persist them so a
       # container recreate does not silently remove installed serve engines.
       - ${APP_DATA_DIR:-./data}/local:/app/.local:z
+      # Docker socket — lets Cookbook launch commands like
+      # `docker exec ollama-rocm ollama show <tag>` reach the host's
+      # Docker daemon (and sibling containers like ollama-rocm /
+      # ollama-test). The in-container user needs to be in the
+      # socket's owning group — see `group_add` below; the GID
+      # there must match the host's `docker` group (defaults to 963
+      # on Debian, 999 on Ubuntu — override via env if yours differs).
+      - /var/run/docker.sock:/var/run/docker.sock
     extra_hosts:
       # Lets the container reach local services on the Docker host, including
       # Ollama at http://host.docker.internal:11434.
@@ -93,6 +101,7 @@ services:
       - /dev/kfd
       - /dev/dri
     group_add:
+      - "${DOCKER_GID:-963}"
       - video
       - ${RENDER_GID:-render}
 

docker-compose.gpu-nvidia.yml

@@ -27,6 +27,16 @@ services:
       # land under /app/.local for the odysseus user. Persist them so a
       # container recreate does not silently remove installed serve engines.
       - ${APP_DATA_DIR:-./data}/local:/app/.local:z
+      # Docker socket — lets Cookbook launch commands like
+      # `docker exec ollama-rocm ollama show <tag>` reach the host's
+      # Docker daemon (and sibling containers like ollama-rocm /
+      # ollama-test). The in-container user needs to be in the
+      # socket's owning group — see `group_add` below; the GID
+      # there must match the host's `docker` group (defaults to 963
+      # on Debian, 999 on Ubuntu — override via env if yours differs).
+      - /var/run/docker.sock:/var/run/docker.sock
+    group_add:
+      - "${DOCKER_GID:-963}"
     extra_hosts:
       # Lets the container reach local services on the Docker host, including
       # Ollama at http://host.docker.internal:11434.

docker-compose.yml

@@ -16,6 +16,16 @@ services:
       # land under /app/.local for the odysseus user. Persist them so a
       # container recreate does not silently remove installed serve engines.
       - ${APP_DATA_DIR:-./data}/local:/app/.local:z
+      # Docker socket — lets Cookbook launch commands like
+      # `docker exec ollama-rocm ollama show <tag>` reach the host's
+      # Docker daemon (and sibling containers like ollama-rocm /
+      # ollama-test). The in-container user needs to be in the
+      # socket's owning group — see `group_add` below; the GID
+      # there must match the host's `docker` group (defaults to 963
+      # on Debian, 999 on Ubuntu — override via env if yours differs).
+      - /var/run/docker.sock:/var/run/docker.sock
+    group_add:
+      - "${DOCKER_GID:-963}"
     extra_hosts:
       # Lets the container reach local services on the Docker host, including
       # Ollama at http://host.docker.internal:11434.

docker/entrypoint.sh

@@ -26,6 +26,27 @@ if ! getent passwd "$PUID" >/dev/null 2>&1; then
     useradd -u "$PUID" -g "$PGID" -M -s /bin/sh -d /app odysseus
 fi
 
+ODY_USER="$(getent passwd "$PUID" | cut -d: -f1)"
+[ -z "$ODY_USER" ] && ODY_USER=odysseus
+
+# Docker-socket group plumbing. When /var/run/docker.sock is bind-mounted
+# (Cookbook uses docker exec to reach sibling containers), the socket is
+# owned by root:<host docker gid>. Add the app user to that group and later
+# call gosu by username so supplementary groups are retained.
+DOCKER_SOCK="${DOCKER_SOCK:-/var/run/docker.sock}"
+if [ -S "$DOCKER_SOCK" ]; then
+    SOCK_GID="$(stat -c '%g' "$DOCKER_SOCK" 2>/dev/null || echo '')"
+    if [ -n "$SOCK_GID" ] && [ "$SOCK_GID" != "0" ]; then
+        if ! getent group "$SOCK_GID" >/dev/null 2>&1; then
+            groupadd -g "$SOCK_GID" docker_host || true
+        fi
+        SOCK_GROUP="$(getent group "$SOCK_GID" | cut -d: -f1)"
+        if [ -n "$SOCK_GROUP" ]; then
+            usermod -aG "$SOCK_GROUP" "$ODY_USER" 2>/dev/null || true
+        fi
+    fi
+fi
+
 mount_root_for() {
     awk -v target="$1" '$5 == target { print $4; exit }' /proc/self/mountinfo 2>/dev/null || true
 }
@@ -103,6 +124,7 @@ for cu in \
         break
     fi
 done
+
 # Disable the FlashInfer JIT sampler unconditionally — it is sampler-only
 # and has no impact on the attention path, but requires nvcc + matching
 # CUDA headers at startup. Without this, vLLM crashes with "Could not find
@@ -116,9 +138,9 @@ export PATH="/app/.local/bin:$PATH"
 # Run first-time setup as the app user so data/ files get the right ownership.
 # setup.py is idempotent — skips auth.json / .env if they already exist.
 # || true so a setup failure never prevents the container from starting.
-"$GOSU_BIN" "$PUID:$PGID" "$PYTHON_BIN" /app/setup.py || true
+"$GOSU_BIN" "$ODY_USER" "$PYTHON_BIN" /app/setup.py || true
 
 # Drop root and run the actual app. `gosu` is preferred over `su` /
 # `sudo` because it cleans up the process tree (no extra shell layer)
 # so signals (SIGTERM from `docker stop`) reach uvicorn directly.
-exec "$GOSU_BIN" "$PUID:$PGID" "$@"
+exec "$GOSU_BIN" "$ODY_USER" "$@"

routes/chat_helpers.py

@@ -22,6 +22,31 @@ from fastapi import HTTPException
 
 logger = logging.getLogger(__name__)
 
+_CASUAL_OPENING_RE = re.compile(
+    r"^\s*(?:h+i+|hey+|hello+|yo+|sup+|what'?s up|wass?up|hiya|howdy|"
+    r"lol|lmao|haha+|hehe+|thanks?|thank you|ty|idk|dunno|meh|bruh|bro)\b(?P<tail>.*)$",
+    re.IGNORECASE,
+)
+_CASUAL_BLOCKLIST_RE = re.compile(
+    r"\b(?:cookbook|serve|serving|launch|start|vllm|sglang|llama\.?cpp|ollama|"
+    r"download|model|email|document|doc|note|calendar|task|search|web|research|"
+    r"file|folder|repo|git|settings?|endpoint|api|token|mcp)\b",
+    re.IGNORECASE,
+)
+
+
+def _is_casual_low_signal(text: str) -> bool:
+    """Short greetings/slang should not pull memory, skills, RAG, or docs."""
+    s = str(text or "").strip()
+    m = _CASUAL_OPENING_RE.match(s)
+    if not m:
+        return False
+    tail = m.group("tail") or ""
+    if _CASUAL_BLOCKLIST_RE.search(tail):
+        return False
+    tail_words = re.findall(r"[A-Za-z0-9_'-]+", tail)
+    return len(tail_words) <= 2
+
 
 # Strong references to in-flight fire-and-forget tasks scheduled from this
 # module. asyncio only keeps weak references to tasks created via
@@ -588,6 +613,7 @@ async def build_chat_context(
     # bearer-token chat requests use the token owner instead of the "api" sentinel.
     user = effective_user(request)
     uprefs = load_prefs_for_user(user)
+    casual_low_signal = _is_casual_low_signal(message)
 
     # Memory enabled?
     mem_enabled = not incognito and not no_memory and uprefs.get("memory_enabled", True)
@@ -597,6 +623,9 @@ async def build_chat_context(
     if not allow_tool_preprocessing:
         mem_enabled = False
         skills_enabled = False
+    if casual_low_signal:
+        mem_enabled = False
+        skills_enabled = False
     logger.debug(
         "Memory enabled=%s for user=%s (incognito=%s, no_memory=%s, pref=%s)",
         mem_enabled, user, incognito, no_memory, uprefs.get("memory_enabled", "NOT_SET"),
@@ -612,11 +641,11 @@ async def build_chat_context(
 
     # Use RAG?
     use_rag_val = (str(use_rag).lower() != "false") if use_rag is not None else True
-    if incognito or not allow_tool_preprocessing or is_research_spinoff:
+    if incognito or not allow_tool_preprocessing or is_research_spinoff or casual_low_signal:
         use_rag_val = False
 
     # If pre-fetched search context was provided (compare mode), skip live web search
-    skip_web = bool(search_context) or not allow_tool_preprocessing
+    skip_web = bool(search_context) or not allow_tool_preprocessing or casual_low_signal
 
     # Build context preface
     # The stream path uses enhanced_message (with CoT/preprocessing applied),
@@ -635,7 +664,7 @@ async def build_chat_context(
         incognito=incognito,
         use_skills=skills_enabled,
     )
-    if use_rag is not None or is_research_spinoff:
+    if use_rag is not None or is_research_spinoff or casual_low_signal:
         _preface_kwargs["use_rag"] = use_rag_val
     preface, rag_sources, web_sources = chat_processor.build_context_preface(**_preface_kwargs)
 
@@ -643,7 +672,7 @@ async def build_chat_context(
     used_memories = getattr(chat_processor, '_last_used_memories', [])
 
     # Inject pre-fetched search context (compare mode)
-    if search_context and allow_tool_preprocessing:
+    if search_context and allow_tool_preprocessing and not casual_low_signal:
         preface.append(untrusted_context_message("prefetched search context", search_context))
 
     # YouTube transcripts

routes/chat_routes.py

@@ -829,7 +829,11 @@ def setup_chat_routes(
         from src.settings import get_setting
         _global_disabled = get_setting("disabled_tools", [])
         if _global_disabled and isinstance(_global_disabled, list):
-            disabled_tools.update(_global_disabled)
+            explicit_web_allowed = allow_web_search is not None and str(allow_web_search).lower() == "true"
+            if explicit_web_allowed:
+                disabled_tools.update(t for t in _global_disabled if t not in {"web_search", "web_fetch"})
+            else:
+                disabled_tools.update(_global_disabled)
 
         # Light auto-escalation: the user is in chat mode and just expressed a
         # notes/calendar/email intent. Grant the relevant managers but withhold
@@ -1259,6 +1263,10 @@ def setup_chat_routes(
                         _max_rounds = _DEFAULT_ROUNDS
                     _max_rounds = max(1, min(_max_rounds, 200))
 
+                    _forced_tools = None
+                    if allow_web_search is not None and str(allow_web_search).lower() == "true":
+                        _forced_tools = {"web_search", "web_fetch"}
+
                     async for chunk in stream_agent_loop(
                         sess.endpoint_url,
                         sess.model,
@@ -1280,6 +1288,7 @@ def setup_chat_routes(
                         plan_mode=plan_mode,
                         approved_plan=approved_plan or None,
                         workspace=workspace or None,
+                        forced_tools=_forced_tools,
                     ):
                         if chunk.startswith("data: ") and not chunk.startswith("data: [DONE]"):
                             try:

routes/cookbook_helpers.py

@@ -786,25 +786,149 @@ def _append_llama_cpp_linux_accel_build_lines(runner_lines: list[str]) -> None:
     to hard-wire CUDA on Linux. That made ROCm hosts attempt a CUDA configure and
     fail with "CUDA Toolkit not found" instead of building with HIP.
     """
+    # Try a prebuilt binary from llama.cpp's GitHub releases FIRST — no
+    # cmake/build-essential/git/CUDA-headers needed at all. The from-source
+    # build below stays as a fallback (custom flags, esoteric arch, no
+    # internet, etc). 30 seconds vs 5+ minutes of compile, and removes
+    # every OS-package dep from the launch path. Sets _odysseus_have_prebuilt=1
+    # on success; the existing build-tier if/elif chain below is gated on
+    # that variable so we never compile twice or shadow the prebuilt symlink.
+    runner_lines.append('    _odysseus_have_prebuilt=""')
+    runner_lines.append('    _odysseus_arch="$(uname -m)"')
+    runner_lines.append('    _odysseus_prebuilt_url=""')
+    runner_lines.append('    if command -v curl >/dev/null 2>&1 && [ "$_odysseus_arch" = "x86_64" ]; then')
+    runner_lines.append('      _odysseus_pat=""')
+    runner_lines.append('      _odysseus_has_nv_inline() { command -v nvidia-smi >/dev/null 2>&1 && nvidia-smi -L 2>/dev/null | grep -q "GPU "; }')
+    runner_lines.append('      _odysseus_has_vk_inline() { ldconfig -p 2>/dev/null | grep -q "libvulkan\\.so" || command -v vulkaninfo >/dev/null 2>&1 || [ -e /usr/lib/x86_64-linux-gnu/libvulkan.so.1 ]; }')
+    runner_lines.append('      _odysseus_has_vkdev_inline() { ls /dev/dri/renderD* >/dev/null 2>&1 || (lspci 2>/dev/null | grep -Ei \'VGA|3D|Display\' | grep -Eiq \'AMD|ATI|Radeon\'); }')
+    runner_lines.append('      if _odysseus_has_nv_inline; then')
+    runner_lines.append('        _odysseus_pat="ubuntu.*cuda"')
+    runner_lines.append('      elif _odysseus_has_vkdev_inline && _odysseus_has_vk_inline; then')
+    runner_lines.append('        _odysseus_pat="ubuntu.*vulkan"')
+    runner_lines.append('      else')
+    runner_lines.append('        _odysseus_pat="ubuntu-x64\\\\.zip"')
+    runner_lines.append('      fi')
+    runner_lines.append('      _odysseus_prebuilt_url="$(curl -fsSL --max-time 15 https://api.github.com/repos/ggml-org/llama.cpp/releases/latest 2>/dev/null | grep \'"browser_download_url"\' | cut -d\'"\' -f4 | grep -iE "$_odysseus_pat" | grep -iv "arm\\|aarch64" | head -1)"')
+    runner_lines.append('    fi')
+    # Accept any of unzip / bsdtar / python3 -m zipfile as the extractor.
+    # python3 is essentially always present on modern Linux, so this lets
+    # the prebuilt path work on minimal Ubuntu installs that lack `unzip`.
+    runner_lines.append('    if [ -n "$_odysseus_prebuilt_url" ] && (command -v unzip >/dev/null 2>&1 || command -v bsdtar >/dev/null 2>&1 || command -v python3 >/dev/null 2>&1); then')
+    runner_lines.append('      echo "[odysseus] Found prebuilt llama-server: $_odysseus_prebuilt_url"')
+    runner_lines.append('      mkdir -p ~/bin "$HOME/.cache/odysseus/llama-cpp-prebuilt" && cd "$HOME/.cache/odysseus/llama-cpp-prebuilt"')
+    runner_lines.append('      rm -f llama-cpp.zip')
+    runner_lines.append('      if curl -fsSL --max-time 120 "$_odysseus_prebuilt_url" -o llama-cpp.zip && [ -s llama-cpp.zip ]; then')
+    runner_lines.append('        rm -rf build && mkdir -p build')
+    runner_lines.append('        if command -v unzip >/dev/null 2>&1; then unzip -qq -o llama-cpp.zip -d build; elif command -v bsdtar >/dev/null 2>&1; then bsdtar -xf llama-cpp.zip -C build; else python3 -c "import zipfile; zipfile.ZipFile(\\"llama-cpp.zip\\").extractall(\\"build\\")"; fi')
+    runner_lines.append('        _odysseus_extracted="$(find build -type f -name llama-server 2>/dev/null | head -1)"')
+    runner_lines.append('        if [ -n "$_odysseus_extracted" ]; then')
+    runner_lines.append('          chmod +x "$_odysseus_extracted"')
+    runner_lines.append('          ln -sf "$_odysseus_extracted" ~/bin/llama-server')
+    runner_lines.append('          _odysseus_libdir="$(dirname "$_odysseus_extracted")"')
+    runner_lines.append('          mkdir -p ~/.config && echo "export LD_LIBRARY_PATH=\\"$_odysseus_libdir:\\${LD_LIBRARY_PATH:-}\\"" > ~/.config/odysseus-llama-cpp-env')
+    runner_lines.append('          _odysseus_have_prebuilt=1')
+    runner_lines.append('          echo "[odysseus] Prebuilt llama-server installed at $_odysseus_extracted"')
+    runner_lines.append('        fi')
+    runner_lines.append('      fi')
+    runner_lines.append('      [ -z "$_odysseus_have_prebuilt" ] && echo "[odysseus] Prebuilt download/extract failed — falling back to from-source build."')
+    runner_lines.append('    elif [ -z "$_odysseus_prebuilt_url" ]; then')
+    runner_lines.append('      echo "[odysseus] No matching prebuilt llama-server for this host (arch=$_odysseus_arch) — will build from source."')
+    runner_lines.append('    fi')
+    runner_lines.append('  if [ -z "$_odysseus_have_prebuilt" ]; then')
     # Detect pip-installed nvcc (from vLLM/nvidia CUDA wheels) and put it on PATH
-    # so cmake's CUDA configure can find it. We keep this after the ROCm/HIP
-    # check — a machine with both stacks should honor the native HIP toolchain on
-    # AMD hosts instead of accidentally preferring a stray nvcc wheel.
-    runner_lines.append('    for _cudir in ~/.local/lib/python*/site-packages/nvidia/cu13 ~/.local/lib/python*/site-packages/nvidia/cu12 ~/.local/lib/python*/site-packages/nvidia/cuda_nvcc; do')
-    runner_lines.append('      [ -x "$_cudir/bin/nvcc" ] && export CUDA_HOME="$_cudir" && export PATH="$_cudir/bin:$PATH" && break')
-    runner_lines.append('    done')
+    # so cmake's CUDA configure can find it — BUT only when actual NVIDIA
+    # hardware is present. On AMD/Intel hosts the pip nvcc is a misleading
+    # leftover (no libcudart, no GPU it could target) and would otherwise
+    # send the build down the CUDA branch and fail with "CUDA Toolkit not
+    # found" instead of trying Vulkan.
+    runner_lines.append('    _odysseus_has_nvidia_hw() {')
+    runner_lines.append('      command -v nvidia-smi >/dev/null 2>&1 && nvidia-smi -L 2>/dev/null | grep -q "GPU " && return 0')
+    runner_lines.append('      ls /dev/nvidia* >/dev/null 2>&1 && return 0')
+    runner_lines.append('      lspci 2>/dev/null | grep -iE \'VGA|3D|Display\' | grep -iq nvidia && return 0')
+    runner_lines.append('      return 1')
+    runner_lines.append('    }')
+    runner_lines.append('    if _odysseus_has_nvidia_hw; then')
+    runner_lines.append('      for _cudir in ~/.local/lib/python*/site-packages/nvidia/cu13 ~/.local/lib/python*/site-packages/nvidia/cu12 ~/.local/lib/python*/site-packages/nvidia/cuda_nvcc; do')
+    runner_lines.append('        [ -x "$_cudir/bin/nvcc" ] && export CUDA_HOME="$_cudir" && export PATH="$_cudir/bin:$PATH" && break')
+    runner_lines.append('      done')
+    runner_lines.append('    fi')
     # rm -rf build so a prior poisoned CMakeCache.txt (e.g. from a failed CUDA
     # or HIP attempt) doesn't cause the next configure to reuse stale settings.
     runner_lines.append('    mkdir -p ~/bin')
-    runner_lines.append('    cd ~/llama.cpp && rm -rf build')
+    # Try to install cmake / build-essential / git automatically before the
+    # build, but ONLY via passwordless sudo (`sudo -n`) — interactive sudo
+    # would hang a tmux-backgrounded serve task waiting for a password. If
+    # sudo asks for a password the install is skipped silently and the
+    # diagnosis pattern (cookbook_routes.py / cookbook_helpers.py) surfaces
+    # an explicit "install cmake" suggestion in the Cookbook diagnosis
+    # toolbar after the inevitable build failure.
+    runner_lines.append('    _odysseus_apt_bootstrap() {')
+    runner_lines.append('      local _missing=""')
+    runner_lines.append('      command -v cmake >/dev/null 2>&1 || _missing="$_missing cmake"')
+    runner_lines.append('      command -v g++ >/dev/null 2>&1 || command -v gcc >/dev/null 2>&1 || _missing="$_missing build-essential"')
+    runner_lines.append('      command -v git >/dev/null 2>&1 || _missing="$_missing git"')
+    runner_lines.append('      [ -z "$_missing" ] && return 0')
+    runner_lines.append('      if command -v apt-get >/dev/null 2>&1 && sudo -n true 2>/dev/null; then')
+    runner_lines.append('        echo "[odysseus] Auto-installing missing build deps via apt:$_missing"')
+    runner_lines.append('        sudo -n env DEBIAN_FRONTEND=noninteractive apt-get update -qq 2>&1 | tail -3')
+    runner_lines.append('        sudo -n env DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends $_missing 2>&1 | tail -5 || true')
+    runner_lines.append('      elif command -v pacman >/dev/null 2>&1 && sudo -n true 2>/dev/null; then')
+    runner_lines.append('        echo "[odysseus] Auto-installing missing build deps via pacman:$_missing"')
+    runner_lines.append('        local _pacpkgs="$(echo "$_missing" | sed -e \'s/build-essential/base-devel/g\')"')
+    runner_lines.append('        sudo -n pacman -Sy --needed --noconfirm $_pacpkgs 2>&1 | tail -5 || true')
+    runner_lines.append('      elif command -v dnf >/dev/null 2>&1 && sudo -n true 2>/dev/null; then')
+    runner_lines.append('        echo "[odysseus] Auto-installing missing build deps via dnf:$_missing"')
+    runner_lines.append('        local _dnfpkgs="$(echo "$_missing" | sed -e \'s/build-essential/gcc gcc-c++ make/g\')"')
+    runner_lines.append('        sudo -n dnf install -y $_dnfpkgs 2>&1 | tail -5 || true')
+    runner_lines.append('      else')
+    runner_lines.append('        echo "[odysseus] WARNING: missing build deps ($_missing) — passwordless sudo is unavailable, cannot auto-install. Cookbook Diagnosis will explain the fix after the build fails."')
+    runner_lines.append('      fi')
+    runner_lines.append('    }')
+    runner_lines.append('    _odysseus_apt_bootstrap')
+    runner_lines.append('    _odysseus_missing_build_deps=""')
+    runner_lines.append('    command -v cmake >/dev/null 2>&1 || _odysseus_missing_build_deps="$_odysseus_missing_build_deps cmake"')
+    runner_lines.append('    command -v git >/dev/null 2>&1 || _odysseus_missing_build_deps="$_odysseus_missing_build_deps git"')
+    runner_lines.append('    command -v g++ >/dev/null 2>&1 || command -v gcc >/dev/null 2>&1 || _odysseus_missing_build_deps="$_odysseus_missing_build_deps build-essential"')
+    runner_lines.append('    if [ -n "$_odysseus_missing_build_deps" ]; then')
+    runner_lines.append('      echo "ERROR: llama.cpp source build needs missing packages:$_odysseus_missing_build_deps"')
+    runner_lines.append('      if command -v apt-get >/dev/null 2>&1; then')
+    runner_lines.append('        echo "Install on this host: sudo apt-get update && sudo apt-get install -y cmake build-essential git"')
+    runner_lines.append('      elif command -v pacman >/dev/null 2>&1; then')
+    runner_lines.append('        echo "Install on this host: sudo pacman -Sy --needed cmake base-devel git"')
+    runner_lines.append('      elif command -v dnf >/dev/null 2>&1; then')
+    runner_lines.append('        echo "Install on this host: sudo dnf install -y cmake gcc gcc-c++ make git"')
+    runner_lines.append('      fi')
+    runner_lines.append('      echo "Alternative: install a native llama-server on PATH, then relaunch."')
+    runner_lines.append('      ODYSSEUS_PREFLIGHT_EXIT=127')
+    runner_lines.append('    fi')
+    runner_lines.append('    cd ~/llama.cpp')
+    runner_lines.append('    _odysseus_has_vulkan() {')
+    runner_lines.append('      ldconfig -p 2>/dev/null | grep -q \'libvulkan\\.so\' && return 0')
+    runner_lines.append('      [ -e /usr/lib/libvulkan.so.1 ] && return 0')
+    runner_lines.append('      [ -e /usr/lib/x86_64-linux-gnu/libvulkan.so.1 ] && return 0')
+    runner_lines.append('      command -v vulkaninfo >/dev/null 2>&1 && return 0')
+    runner_lines.append('      return 1')
+    runner_lines.append('    }')
+    runner_lines.append('    _odysseus_has_vulkan_device() {')
+    runner_lines.append('      ls /dev/dri/renderD* >/dev/null 2>&1 && return 0')
+    runner_lines.append('      lspci 2>/dev/null | grep -Ei \'VGA|3D|Display\' | grep -Eiq \'AMD|ATI|Radeon\' && return 0')
+    runner_lines.append('      return 1')
+    runner_lines.append('    }')
+    # Backend preference: native ROCm/HIP > native CUDA > Vulkan > CPU.
+    # Vulkan is a portable fallback that works on AMD when ROCm isn't
+    # installed (e.g. Strix Halo) and on any vendor's discrete GPU, but
+    # it's ~30-40% slower than native HIP/CUDA for LLM inference — only
+    # pick it when no native toolchain is present.
     runner_lines.append('    if command -v hipconfig &>/dev/null || [ -d /opt/rocm ] || [ -n "$ROCM_PATH" ] || [ -n "$HIP_PATH" ]; then')
+    runner_lines.append('      rm -rf build')
     runner_lines.append('      if command -v hipconfig &>/dev/null; then')
     runner_lines.append('        export HIPCXX="${HIPCXX:-$(hipconfig -l)/clang}"')
     runner_lines.append('        export HIP_PATH="${HIP_PATH:-$(hipconfig -R)}"')
     runner_lines.append('      fi')
     runner_lines.append('      echo "[odysseus] ROCm/HIP detected — building llama-server with HIP support..."')
     runner_lines.append('      cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_HIP=ON && cmake --build build -j"$NPROC" --target llama-server && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
-    runner_lines.append('    elif command -v nvcc &>/dev/null; then')
+    runner_lines.append('    elif command -v nvcc &>/dev/null && _odysseus_has_nvidia_hw; then')
+    runner_lines.append('      rm -rf build')
     # nvcc alone is not sufficient — pip-installed CUDA wheels or incomplete
     # tooling can expose nvcc without shipping libcudart, causing cmake to fail
     # mid-build with "CUDA runtime library not found". Check cudart explicitly
@@ -828,31 +952,50 @@ def _append_llama_cpp_linux_accel_build_lines(runner_lines: list[str]) -> None:
     runner_lines.append('        echo "[odysseus]   Ensure libcudart is installed (e.g. cuda-runtime package) and visible via ldconfig or CUDA_HOME."')
     runner_lines.append('        cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j"$NPROC" --target llama-server && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
     runner_lines.append('      fi')
+    runner_lines.append('    elif _odysseus_has_vulkan_device && _odysseus_has_vulkan; then')
+    runner_lines.append('      echo "[odysseus] Vulkan-capable GPU detected (no ROCm/CUDA toolchain installed) — building llama-server with Vulkan support..."')
+    runner_lines.append('      rm -rf build-vulkan')
+    runner_lines.append('      cmake -B build-vulkan -DCMAKE_BUILD_TYPE=Release -DGGML_VULKAN=ON && cmake --build build-vulkan -j"$NPROC" --target llama-server && ln -sf ~/llama.cpp/build-vulkan/bin/llama-server ~/bin/llama-server')
     runner_lines.append('    else')
-    runner_lines.append('      echo "[odysseus] WARNING: no HIP/CUDA toolchain found — building llama-server for CPU only."')
+    runner_lines.append('      echo "[odysseus] WARNING: no HIP/CUDA/Vulkan toolchain found — building llama-server for CPU only."')
     runner_lines.append('      echo "[odysseus]   GPU inference will not be available for this llama.cpp build."')
-    runner_lines.append('      echo "[odysseus]   Install ROCm for AMD GPUs or vLLM/CUDA tooling for NVIDIA, then re-launch this serve task."')
+    runner_lines.append('      echo "[odysseus]   Install Vulkan (libvulkan-dev) / ROCm for AMD GPUs or CUDA tooling for NVIDIA, then re-launch this serve task."')
+    runner_lines.append('      rm -rf build')
     runner_lines.append('      cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j"$NPROC" --target llama-server && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
     runner_lines.append('    fi')
+    runner_lines.append('  fi  # end _odysseus_have_prebuilt guard')
 
 
-def _llama_cpp_rebuild_cmd() -> str:
+def _llama_cpp_rebuild_cmd(update_source: bool = False) -> str:
     """Shell command that clears the Cookbook-managed llama.cpp build.
 
-    Removes the cached ``llama-server`` symlink and the ``~/llama.cpp/build``
+    Removes the cached ``llama-server`` symlink and the ``~/llama.cpp/build*``
     directory so the next llama.cpp serve recompiles from source, picking up a
     CUDA or HIP toolchain if one is now available. The serve bootstrap only
     builds when ``llama-server`` is missing from PATH, so without this an
-    existing CPU-only build is reused forever. It deliberately installs and
-    downloads nothing; the rebuild itself happens on the next serve.
+    existing CPU-only build is reused forever. When ``update_source`` is true,
+    the command also fast-forwards the Cookbook-managed ``~/llama.cpp`` checkout
+    if it exists. The rebuild itself happens on the next serve.
     """
+    update_cmd = ''
+    if update_source:
+        update_cmd = (
+            'if [ -d "$HOME/llama.cpp/.git" ]; then '
+            'git -C "$HOME/llama.cpp" pull --ff-only --depth 1 || '
+            'echo "[odysseus] WARNING: llama.cpp source update failed; clearing cached build anyway."; '
+            'elif command -v git >/dev/null 2>&1; then '
+            'git clone --depth 1 https://github.com/ggml-org/llama.cpp "$HOME/llama.cpp" || '
+            'echo "[odysseus] WARNING: llama.cpp clone failed; clearing cached build anyway."; '
+            'fi && '
+        )
     return (
         'mkdir -p "$HOME/bin" && '
+        f'{update_cmd}'
         'rm -f "$HOME/bin/llama-server" && '
-        'rm -rf "$HOME/llama.cpp/build" && '
+        'rm -rf "$HOME/llama.cpp/build" "$HOME/llama.cpp/build-vulkan" && '
         'echo "[odysseus] Cleared the cached llama.cpp build. '
         'Re-launch the serve task to rebuild llama-server from source '
-        '(CUDA or HIP will be used if a toolchain is now available)."'
+        '(Vulkan, HIP, or CUDA will be used if a matching toolchain is now available)."'
     )
 
 
@@ -1115,8 +1258,27 @@ def _diagnose_serve_output(text: str) -> dict | None:
             "SGLang is not installed or not in PATH on this server.",
             [{"label": "install SGLang in Cookbook Dependencies", "op": "dependency", "package": "sglang[all]"}],
         ),
+        # System build deps come BEFORE the generic llama.cpp catch-all so
+        # cmake / build-essential / git missing → a specific OS-package
+        # remediation instead of "install llama-cpp-python[server]" (which
+        # itself fails to compile when cmake is absent).
         (
-            r"llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'|git: command not found|cmake: command not found",
+            r"cmake: command not found|cmake.*not found.*[Cc]ould not",
+            "cmake is required to build llama.cpp from source but isn't installed on this server.",
+            [{"label": "install build deps for llama.cpp (apt: cmake build-essential git / pacman: cmake base-devel git / dnf: cmake gcc-c++ make git / brew: cmake git)", "op": "dependency", "package": "llama-cpp-python[server]"}],
+        ),
+        (
+            r"^(make|g\+\+|gcc): command not found|Could not find C\+\+ compiler",
+            "A C/C++ compiler (build-essential) is required to build llama.cpp from source.",
+            [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
+        ),
+        (
+            r"^git: command not found",
+            "git is required to clone the llama.cpp source tree.",
+            [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
+        ),
+        (
+            r"llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'",
             "llama.cpp / llama-cpp-python dependencies are missing.",
             [{"label": "install llama.cpp dependencies or llama-cpp-python[server]", "op": "dependency", "package": "llama-cpp-python[server]"}],
         ),

routes/cookbook_routes.py

@@ -189,8 +189,27 @@ def setup_cookbook_routes() -> APIRouter:
                 "SGLang is not installed or not in PATH on this server.",
                 [{"label": "install SGLang in Cookbook Dependencies", "op": "dependency", "package": "sglang[all]"}],
             ),
+            # System build deps come BEFORE the generic llama.cpp catch-all
+            # so cmake / build-essential / git missing → a specific OS-package
+            # remediation instead of "install llama-cpp-python[server]" (which
+            # itself fails to compile when cmake is absent).
             (
-                r"llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'|git: command not found|cmake: command not found",
+                r"cmake: command not found|cmake.*not found.*[Cc]ould not",
+                "cmake is required to build llama.cpp from source but isn't installed on this server.",
+                [{"label": "install build deps for llama.cpp (apt: cmake build-essential git / pacman: cmake base-devel git / dnf: cmake gcc-c++ make git / brew: cmake git)", "op": "dependency", "package": "llama-cpp-python[server]"}],
+            ),
+            (
+                r"^(make|g\+\+|gcc): command not found|Could not find C\+\+ compiler",
+                "A C/C++ compiler (build-essential) is required to build llama.cpp from source.",
+                [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
+            ),
+            (
+                r"^git: command not found",
+                "git is required to clone the llama.cpp source tree.",
+                [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
+            ),
+            (
+                r"llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'",
                 "llama.cpp / llama-cpp-python dependencies are missing.",
                 [{"label": "install llama.cpp dependencies or llama-cpp-python[server]", "op": "dependency", "package": "llama-cpp-python[server]"}],
             ),
@@ -254,6 +273,79 @@ def setup_cookbook_routes() -> APIRouter:
     def _load_stored_hf_token() -> str:
         return load_stored_hf_token(state_path=_cookbook_state_path)
 
+    def _normalize_minimax_m3_vllm_cmd(cmd: str) -> str:
+        """Patch MiniMax M3 vLLM launches into the known-good local form.
+
+        The browser form can be stale or omit advanced-only fields. MiniMax M3
+        is sensitive to several flags: using the HF repo id with block-size 128
+        fails KV-cache setup, and FlashInfer sampler JIT fails on this host's
+        system nvcc. Normalize server-side before writing the tmux runner.
+        """
+        cmd_lower = (cmd or "").lower()
+        if not cmd or "vllm serve" not in cmd_lower or "minimax" not in cmd_lower or "m3" not in cmd_lower:
+            return cmd
+        try:
+            parts = shlex.split(cmd)
+        except ValueError:
+            return cmd
+        if "serve" not in parts:
+            return cmd
+
+        env_re = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*=")
+        env_parts = [p for p in parts if env_re.match(p)]
+        body = [p for p in parts if not env_re.match(p)]
+        try:
+            serve_i = body.index("serve")
+        except ValueError:
+            return cmd
+        if serve_i + 1 >= len(body):
+            return cmd
+
+        repo_id = "cyankiwi/MiniMax-M3-AWQ-INT4"
+        snapshot = (
+            "/home/pewds/.cache/huggingface/hub/"
+            "models--cyankiwi--MiniMax-M3-AWQ-INT4/"
+            "snapshots/4082acbbec1236d21828d55b6bb0fe02ade4ab5b"
+        )
+        if body[serve_i + 1] == repo_id:
+            body[serve_i + 1] = snapshot
+
+        def add_env(key: str, value: str) -> None:
+            if not any(p.startswith(f"{key}=") for p in env_parts):
+                env_parts.append(f"{key}={value}")
+
+        def has_flag(flag: str) -> bool:
+            return any(p == flag or p.startswith(flag + "=") for p in body)
+
+        def set_flag(flag: str, value: str) -> None:
+            for i, part in enumerate(body):
+                if part == flag:
+                    if i + 1 < len(body):
+                        body[i + 1] = value
+                    else:
+                        body.append(value)
+                    return
+                if part.startswith(flag + "="):
+                    body[i] = f"{flag}={value}"
+                    return
+            body.extend([flag, value])
+
+        def add_bool(flag: str) -> None:
+            if not has_flag(flag):
+                body.append(flag)
+
+        add_env("VLLM_TARGET_DEVICE", "cuda")
+        add_env("VLLM_USE_FLASHINFER_SAMPLER", "0")
+        set_flag("--served-model-name", repo_id)
+        set_flag("--tool-call-parser", "minimax_m3")
+        set_flag("--reasoning-parser", "minimax_m3")
+        set_flag("--attention-backend", "TRITON_ATTN")
+        set_flag("--block-size", "128")
+        add_bool("--language-model-only")
+        add_bool("--disable-custom-all-reduce")
+        add_bool("--enable-expert-parallel")
+        return shlex.join(env_parts + body)
+
     def _cookbook_ssh_dir() -> Path:
         # The Docker image keeps cookbook keys under /app/.ssh; that path only
         # exists inside the container. On Windows (and any non-container host)
@@ -1230,6 +1322,7 @@ def setup_cookbook_routes() -> APIRouter:
         # `TypeError: argument of type 'NoneType'` (a 500 instead of a clean 400).
         req.cmd = _validate_serve_cmd(req.cmd) or ""
         req.cmd = _normalize_llama_cpp_python_cache_types(req.cmd) or ""
+        req.cmd = _normalize_minimax_m3_vllm_cmd(req.cmd)
         req.cmd = _venv_safe_local_pip_install_cmd(
             req.cmd,
             local=not bool(req.remote_host),
@@ -1243,8 +1336,16 @@ def setup_cookbook_routes() -> APIRouter:
             req.cmd = _pip_install_no_cache(req.cmd)
             # Accept common aliases and enforce server extras for llama-cpp so
             # `python -m llama_cpp.server` has all runtime dependencies.
-            req.cmd = re.sub(r"(?<![A-Za-z0-9_.-])llama_cpp(?![A-Za-z0-9_.-])", "llama-cpp-python[server]", req.cmd)
-            req.cmd = re.sub(r"(?<![A-Za-z0-9_.-])llama-cpp-python(?!\[)", "llama-cpp-python[server]", req.cmd)
+            # CRITICAL: the lookbehind / lookahead must also exclude `/` so
+            # the regex DOESN'T mangle a URL path like
+            #   https://abetlen.github.io/llama-cpp-python/whl/cu124
+            # The previous regex turned that URL into
+            #   https://abetlen.github.io/llama-cpp-python[server]/whl/cu124
+            # which pip then couldn't resolve → silent fallback to source
+            # build of the .tar.gz → CPU-only binary (because CMAKE_ARGS
+            # isn't set), defeating the entire purpose of the CUDA index.
+            req.cmd = re.sub(r"(?<![A-Za-z0-9_.\-/])llama_cpp(?![A-Za-z0-9_.\-/])", "llama-cpp-python[server]", req.cmd)
+            req.cmd = re.sub(r"(?<![A-Za-z0-9_.\-/])llama-cpp-python(?![\[/])", "llama-cpp-python[server]", req.cmd)
             if "llama-cpp-python" in req.cmd and "--extra-index-url" not in req.cmd:
                 req.cmd += " --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu"
             # PEP-508-style package spec — letters, digits, `.-_` for the
@@ -1431,6 +1532,69 @@ def setup_cookbook_routes() -> APIRouter:
                 runner_lines.append('  else')
                 _append_llama_cpp_linux_accel_build_lines(runner_lines)
                 runner_lines.append('  fi')
+                # Source the env file the prebuilt-download path writes so
+                # LD_LIBRARY_PATH includes the directory holding libllama.so
+                # and friends. No-op when prebuilt wasn't used.
+                runner_lines.append('  [ -r ~/.config/odysseus-llama-cpp-env ] && . ~/.config/odysseus-llama-cpp-env')
+                # Auto-upgrade pip llama-cpp-python to the CUDA-enabled
+                # wheel when (a) NVIDIA hardware is present and (b) the
+                # currently-installed wheel is CPU-only. Without this the
+                # user gets the Python server happily running at 3 tok/s
+                # because pip's default index ships CPU-only wheels.
+                # Forward-compat: cu124 wheels work on driver/runtime
+                # 12.4+ including the cu13.x line.
+                runner_lines.append('  if command -v nvidia-smi >/dev/null 2>&1 && nvidia-smi -L 2>/dev/null | grep -q "GPU " && python3 -c "import llama_cpp" 2>/dev/null; then')
+                runner_lines.append('    if ! python3 -c "import llama_cpp; import sys; sys.exit(0 if llama_cpp.llama_supports_gpu_offload() else 1)" 2>/dev/null; then')
+                runner_lines.append('      echo "[odysseus] NVIDIA detected but installed llama-cpp-python is CPU-only — reinstalling with CUDA wheel index for GPU offload..."')
+                runner_lines.append('      python3 -m pip install --user --break-system-packages --force-reinstall --no-cache-dir "llama-cpp-python[server]" --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124 2>&1 | tail -8 || echo "[odysseus] WARNING: CUDA wheel reinstall failed — Python server will stay CPU-only (slow). Manual fix: pip install --user --force-reinstall \'llama-cpp-python[server]\' --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124"')
+                runner_lines.append('      if python3 -c "import llama_cpp; import sys; sys.exit(0 if llama_cpp.llama_supports_gpu_offload() else 1)" 2>/dev/null; then')
+                runner_lines.append('        echo "[odysseus] llama-cpp-python now supports GPU offload."')
+                runner_lines.append('      fi')
+                runner_lines.append('    fi')
+                runner_lines.append('  fi')
+                # SHORT-CIRCUIT before the build/pip fallback: if the
+                # native binary is missing but llama_cpp Python is already
+                # installed, drop a wrapper at ~/bin/llama-server that
+                # translates llama-server CLI args to llama_cpp.server's
+                # underscore-style flags. The user's serve command stays
+                # `llama-server ...` and "just works" — no build, no cmake,
+                # no second install. This is the path that unblocks every
+                # remote where pip-installed llama-cpp-python is already
+                # working but Cookbook used to insist on a native binary.
+                runner_lines.append('  if ! command -v llama-server >/dev/null 2>&1 && python3 -c "import llama_cpp" 2>/dev/null; then')
+                runner_lines.append('    mkdir -p ~/bin')
+                runner_lines.append('    cat > ~/bin/llama-server <<\'_ODY_LLAMA_SHIM_EOF\'')
+                runner_lines.append('#!/usr/bin/env bash')
+                runner_lines.append('# Auto-generated by Odysseus Cookbook: a `llama-server` lookalike')
+                runner_lines.append('# that translates the native CLI to `python -m llama_cpp.server`.')
+                runner_lines.append('# Lets cookbook-generated launch commands run unchanged on hosts')
+                runner_lines.append('# where only the pip llama-cpp-python package is installed.')
+                runner_lines.append('ARGS=()')
+                runner_lines.append('while [ $# -gt 0 ]; do')
+                runner_lines.append('  case "$1" in')
+                runner_lines.append('    -ngl|--gpu-layers|--n-gpu-layers) ARGS+=(--n_gpu_layers "$2"); shift 2 ;;')
+                runner_lines.append('    -c|--ctx-size) ARGS+=(--n_ctx "$2"); shift 2 ;;')
+                runner_lines.append('    -b|--batch-size) ARGS+=(--n_batch "$2"); shift 2 ;;')
+                runner_lines.append('    -ub|--ubatch-size) shift 2 ;;  # llama-cpp-python has no separate ubatch')
+                runner_lines.append('    --flash-attn) ARGS+=(--flash_attn true); shift 2 ;;')
+                runner_lines.append('    --cache-type-k) ARGS+=(--type_k "$2"); shift 2 ;;')
+                runner_lines.append('    --cache-type-v) ARGS+=(--type_v "$2"); shift 2 ;;')
+                runner_lines.append('    --n-cpu-moe) ARGS+=(--n_cpu_moe "$2"); shift 2 ;;')
+                runner_lines.append('    --mmproj) ARGS+=(--clip_model_path "$2"); shift 2 ;;')
+                runner_lines.append('    --image-max-tokens) shift 2 ;;  # native-only')
+                runner_lines.append('    --no-mmap) ARGS+=(--no_mmap true); shift ;;')
+                runner_lines.append('    --no-warmup) shift ;;  # native-only')
+                runner_lines.append('    --chat-template) ARGS+=(--chat_format "$2"); shift 2 ;;')
+                runner_lines.append('    --fit|--split-mode|--tensor-split|--main-gpu|--parallel) shift 2 ;;  # native-only')
+                runner_lines.append('    --mlock) ARGS+=(--use_mlock true); shift ;;')
+                runner_lines.append('    *) ARGS+=("$1"); shift ;;')
+                runner_lines.append('  esac')
+                runner_lines.append('done')
+                runner_lines.append('exec python3 -m llama_cpp.server "${ARGS[@]}"')
+                runner_lines.append('_ODY_LLAMA_SHIM_EOF')
+                runner_lines.append('    chmod +x ~/bin/llama-server')
+                runner_lines.append('    echo "[odysseus] Created llama-server shim → python -m llama_cpp.server (no native binary needed)"')
+                runner_lines.append('  fi')
                 runner_lines.append('  # If the native build failed, fall back to the Python bindings.')
                 runner_lines.append('  if ! command -v llama-server &>/dev/null && ! python3 -c "import llama_cpp" 2>/dev/null; then')
                 runner_lines.append('    echo "llama-server build failed — installing Python bindings as fallback..."')
@@ -1494,6 +1658,96 @@ def setup_cookbook_routes() -> APIRouter:
                 runner_lines.append('  echo "ERROR: vLLM is not installed."')
                 runner_lines.append('  ODYSSEUS_PREFLIGHT_EXIT=127')
                 runner_lines.append('fi')
+                runner_lines.append(f"ODYSSEUS_SERVE_CMD='{_bash_squote(req.cmd)}'")
+                runner_lines.append('if [ -z "$ODYSSEUS_PREFLIGHT_EXIT" ]; then')
+                runner_lines.append('  ODYSSEUS_VLLM_HELP_CMD="$(python3 - "$ODYSSEUS_SERVE_CMD" <<\'PY\'')
+                runner_lines.append('import shlex, sys')
+                runner_lines.append('parts = shlex.split(sys.argv[1])')
+                runner_lines.append('try:')
+                runner_lines.append('    serve_i = parts.index("serve")')
+                runner_lines.append('except ValueError:')
+                runner_lines.append('    print("vllm serve --help")')
+                runner_lines.append('else:')
+                runner_lines.append('    print(shlex.join(parts[:serve_i + 1] + ["--help"]))')
+                runner_lines.append('PY')
+                runner_lines.append(')"')
+                runner_lines.append('  ODYSSEUS_VLLM_SUPPORTS_SWAP=0')
+                runner_lines.append('  if eval "$ODYSSEUS_VLLM_HELP_CMD" 2>&1 | grep -q -- "--swap-space"; then ODYSSEUS_VLLM_SUPPORTS_SWAP=1; fi')
+                runner_lines.append('fi')
+                runner_lines.append('if [ -z "$ODYSSEUS_PREFLIGHT_EXIT" ] && [ "${ODYSSEUS_VLLM_SUPPORTS_SWAP:-0}" = "1" ] && ! printf "%s" "$ODYSSEUS_SERVE_CMD" | grep -q -- "--swap-space"; then')
+                runner_lines.append('  echo "[odysseus] Setting vLLM --swap-space 0 so the runtime does not reserve CPU swap per GPU."')
+                runner_lines.append('  ODYSSEUS_SERVE_CMD="${ODYSSEUS_SERVE_CMD} --swap-space 0"')
+                runner_lines.append('fi')
+                runner_lines.append('if [ -z "$ODYSSEUS_PREFLIGHT_EXIT" ] && [ "${ODYSSEUS_VLLM_SUPPORTS_SWAP:-0}" != "1" ]; then')
+                runner_lines.append('  if printf "%s" "$ODYSSEUS_SERVE_CMD" | grep -q -- "--swap-space"; then')
+                runner_lines.append('    echo "[odysseus] vLLM serve does not expose --swap-space; removing the flag and patching the runtime default to 0."')
+                runner_lines.append('    ODYSSEUS_SERVE_CMD="$(python3 - "$ODYSSEUS_SERVE_CMD" <<\'PY\'')
+                runner_lines.append('import shlex, sys')
+                runner_lines.append('parts = shlex.split(sys.argv[1])')
+                runner_lines.append('out = []')
+                runner_lines.append('skip = False')
+                runner_lines.append('for part in parts:')
+                runner_lines.append('    if skip:')
+                runner_lines.append('        skip = False')
+                runner_lines.append('        continue')
+                runner_lines.append('    if part == "--swap-space":')
+                runner_lines.append('        skip = True')
+                runner_lines.append('        continue')
+                runner_lines.append('    if part.startswith("--swap-space="):')
+                runner_lines.append('        continue')
+                runner_lines.append('    out.append(part)')
+                runner_lines.append('print(shlex.join(out))')
+                runner_lines.append('PY')
+                runner_lines.append(')"')
+                runner_lines.append('  fi')
+                runner_lines.append('  ODYSSEUS_SERVE_CMD="$(python3 - "$ODYSSEUS_SERVE_CMD" <<\'PY\'')
+                runner_lines.append('import shlex, sys')
+                runner_lines.append('parts = shlex.split(sys.argv[1])')
+                runner_lines.append('patch = r"""import inspect, sys')
+                runner_lines.append('from vllm.engine.arg_utils import EngineArgs, AsyncEngineArgs')
+                runner_lines.append('def _odysseus_swap0(cls):')
+                runner_lines.append('    params = list(inspect.signature(cls).parameters)')
+                runner_lines.append('    if "swap_space" not in params:')
+                runner_lines.append('        return')
+                runner_lines.append('    idx = params.index("swap_space")')
+                runner_lines.append('    defaults = list(cls.__init__.__defaults__ or ())')
+                runner_lines.append('    if idx < len(defaults):')
+                runner_lines.append('        defaults[idx] = 0')
+                runner_lines.append('        cls.__init__.__defaults__ = tuple(defaults)')
+                runner_lines.append('    fields = getattr(cls, "__dataclass_fields__", {})')
+                runner_lines.append('    if "swap_space" in fields:')
+                runner_lines.append('        fields["swap_space"].default = 0')
+                runner_lines.append('_odysseus_swap0(EngineArgs)')
+                runner_lines.append('_odysseus_swap0(AsyncEngineArgs)')
+                runner_lines.append('try:')
+                runner_lines.append('    from vllm.config import CacheConfig')
+                runner_lines.append('    CacheConfig.swap_space = 0')
+                runner_lines.append('except Exception:')
+                runner_lines.append('    pass')
+                runner_lines.append('_orig_create_engine_config = EngineArgs.create_engine_config')
+                runner_lines.append('def _odysseus_create_engine_config(self, *args, **kwargs):')
+                runner_lines.append('    self.swap_space = 0')
+                runner_lines.append('    return _orig_create_engine_config(self, *args, **kwargs)')
+                runner_lines.append('EngineArgs.create_engine_config = _odysseus_create_engine_config')
+                runner_lines.append('AsyncEngineArgs.create_engine_config = _odysseus_create_engine_config')
+                runner_lines.append('from vllm.entrypoints.cli.main import main')
+                runner_lines.append('sys.exit(main())"""')
+                runner_lines.append('try:')
+                runner_lines.append('    serve_i = parts.index("serve")')
+                runner_lines.append('except ValueError:')
+                runner_lines.append('    print(shlex.join(parts))')
+                runner_lines.append('else:')
+                runner_lines.append('    exe_i = serve_i - 1')
+                runner_lines.append('    exe = parts[exe_i] if exe_i >= 0 else "vllm"')
+                runner_lines.append('    py = "python3"')
+                runner_lines.append('    if exe.endswith("/bin/vllm"):')
+                runner_lines.append('        py = exe[:-len("/bin/vllm")] + "/bin/python"')
+                runner_lines.append('    parts[exe_i:serve_i] = [py, "-c", patch]')
+                runner_lines.append('    print(shlex.join(parts))')
+                runner_lines.append('PY')
+                runner_lines.append(')"')
+                runner_lines.append('  echo "[odysseus] Patched vLLM internal swap_space default to 0 for this runtime."')
+                runner_lines.append('fi')
             elif "sglang.launch_server" in req.cmd:
                 runner_lines.append('export PATH="$HOME/.local/bin:$PATH"')
                 runner_lines.append('if ! command -v sglang &>/dev/null; then')
@@ -1535,7 +1789,10 @@ def setup_cookbook_routes() -> APIRouter:
                     runner_lines,
                     keep_shell_open=not local_windows,
                 )
-                runner_lines.append(req.cmd)
+                if "vllm serve" in req.cmd:
+                    runner_lines.append('eval "$ODYSSEUS_SERVE_CMD"')
+                else:
+                    runner_lines.append(req.cmd)
                 if local_windows:
                     # Detached background process — no interactive shell to keep open.
                     # Print the exit marker the status poller looks for, then stop.
@@ -1839,6 +2096,25 @@ def setup_cookbook_routes() -> APIRouter:
         out, err = await _run_gpu_shell("ls -1 /sys/class/drm 2>/dev/null", host, ssh_port, timeout=4)
         if err is not None or not out:
             return []
+        # Pick the runtime label up-front so each GPU dict gets the
+        # right `backend`. AMD silicon can be driven by ROCm/HIP (native)
+        # OR Vulkan (mesa RADV). Reporting "rocm" on a host where no
+        # ROCm toolchain is installed misleads the frontend env-var
+        # prefix logic — it would emit `HIP_VISIBLE_DEVICES=` for a
+        # Vulkan-only stack, which is a silent no-op at best.
+        rt_out, _ = await _run_gpu_shell(
+            'command -v rocminfo >/dev/null 2>&1 && echo rocm '
+            '|| (command -v hipconfig >/dev/null 2>&1 && echo rocm) '
+            '|| (command -v vulkaninfo >/dev/null 2>&1 && echo vulkan) '
+            '|| echo unknown',
+            host, ssh_port, timeout=4,
+        )
+        _amd_runtime = (rt_out or "").strip().splitlines()[-1:][0].strip() if rt_out else "rocm"
+        if _amd_runtime not in ("rocm", "vulkan"):
+            # Default to rocm so existing ROCm-installed hosts keep
+            # working; "unknown" only happens when neither toolchain is
+            # detected (e.g. minimal sysfs read on a fresh box).
+            _amd_runtime = "rocm"
         gpus = []
         for entry in out.split():
             if not entry.startswith("card") or "-" in entry:
@@ -1882,7 +2158,7 @@ def setup_cookbook_routes() -> APIRouter:
                 "free_mb": free_mb, "total_mb": total_mb, "used_mb": used_mb,
                 "gtt_used_mb": gtt_used_mb,
                 "util_pct": 0, "busy": bool(total_mb and (free_mb / total_mb) < 0.85),
-                "processes": [], "backend": "rocm", "source": "amd-sysfs",
+                "processes": [], "backend": _amd_runtime, "source": "amd-sysfs",
                 "unified_memory": unified,
             })
         if gpus:
@@ -2023,10 +2299,15 @@ def setup_cookbook_routes() -> APIRouter:
 
         amd_gpus = await _probe_amd_sysfs(host, ssh_port)
         if amd_gpus:
+            # The per-GPU dict already carries the runtime label picked by
+            # _probe_amd_sysfs (rocm vs vulkan); mirror that into the
+            # wrapper so the frontend can read `data.backend` directly
+            # without scanning the list.
+            _amd_wrap_backend = str(amd_gpus[0].get("backend") or "rocm")
             return {
                 "ok": True,
                 "gpus": amd_gpus,
-                "backend": "rocm",
+                "backend": _amd_wrap_backend,
                 "source": "amd-sysfs",
                 "fallback_from": "nvidia-smi",
                 "nvidia_error": nvidia_error,
@@ -2166,6 +2447,17 @@ def setup_cookbook_routes() -> APIRouter:
 
             disk_tasks = on_disk.get("tasks") or [] if isinstance(on_disk, dict) else []
             incoming_tasks = data.get("tasks") if isinstance(data.get("tasks"), list) else []
+            incoming_removed = data.get("removedTasks") if isinstance(data.get("removedTasks"), dict) else {}
+            disk_removed = on_disk.get("removedTasks") if isinstance(on_disk, dict) and isinstance(on_disk.get("removedTasks"), dict) else {}
+            removed_tasks = {**disk_removed, **incoming_removed}
+            data["removedTasks"] = removed_tasks
+            removed_ids = set(removed_tasks.keys())
+            if removed_ids:
+                incoming_tasks = [
+                    t for t in incoming_tasks
+                    if not (isinstance(t, dict) and t.get("sessionId") in removed_ids)
+                ]
+                data["tasks"] = incoming_tasks
             # Anti-poisoning guard: a stale browser tab can keep POSTing a
             # download task as status='done' from before the strict-finish
             # fix landed, undoing any server-side correction. For each
@@ -2203,6 +2495,8 @@ def setup_cookbook_routes() -> APIRouter:
                 sid = t.get("sessionId")
                 if not sid or sid in incoming_ids:
                     continue  # client's version wins
+                if sid in removed_ids:
+                    continue  # intentional cross-device clear/remove
                 ts = t.get("ts") or 0
                 if isinstance(ts, (int, float)) and (now_ms - ts) <= RACE_WINDOW_MS:
                     preserved.append(t)
@@ -2309,16 +2603,14 @@ def setup_cookbook_routes() -> APIRouter:
             # Add 30% headroom for KV cache, activations, etc.
             needed_vram = (est_vram * 1.3) if est_vram else None
 
-            if vram_gb > 0 and needed_vram is not None and needed_vram > vram_gb:
-                continue
-            # Unknown-size models (e.g. MiniMax-M2.7, DeepSeek-V4-Flash) have no
-            # "NB" in the repo id, so the regex above can't extract their
-            # param count. Previously we dropped them entirely, which made
-            # brand-new flagship releases silently vanish from this list even
-            # on rigs with hundreds of GB of VRAM. Adapters/LoRAs are already
-            # filtered by _is_excluded(), so what falls through here is
-            # overwhelmingly full models — keep them, just without a size
-            # badge (the frontend handles needed_vram_gb=null gracefully).
+            if vram_gb > 0:
+                if needed_vram is None:
+                    # The "trending models that fit" list must be conservative:
+                    # if we cannot estimate size from the repo id/tags, do not
+                    # present it as runnable on this hardware.
+                    continue
+                if needed_vram > vram_gb:
+                    continue
 
             out.append({
                 "repo_id": repo_id,
@@ -2515,6 +2807,33 @@ def setup_cookbook_routes() -> APIRouter:
             except Exception as e:
                 logger.warning(f"orphan sweep: state write failed: {e}")
 
+    @router.get("/api/cookbook/hf-gguf-files")
+    async def hf_gguf_files(repo_id: str, owner: str = Depends(require_user)):
+        """List GGUF files in a HuggingFace repo for the direct-download picker."""
+        import httpx
+
+        repo_id = _validate_repo_id(repo_id)
+        url = f"https://huggingface.co/api/models/{repo_id}"
+        try:
+            headers = {}
+            token = _load_stored_hf_token()
+            if token:
+                headers["Authorization"] = f"Bearer {token}"
+            async with httpx.AsyncClient(timeout=15, follow_redirects=True) as client:
+                resp = await client.get(url, headers=headers)
+                if resp.status_code != 200:
+                    return {"ok": False, "files": [], "error": f"HF API HTTP {resp.status_code}"}
+                data = resp.json()
+        except Exception:
+            logger.exception("HF GGUF file scan failed for %s", repo)
+            return {"ok": False, "files": [], "error": "HF API request failed"}
+        files = [
+            str(s.get("rfilename") or "")
+            for s in data.get("siblings", [])
+            if str(s.get("rfilename") or "").lower().endswith(".gguf")
+        ]
+        return {"ok": True, "repo_id": repo_id, "files": files}
+
     # In-memory cache for the Ollama library scrape. ollama.com is a public
     # site, but it doesn't expose a stable JSON listing — we fetch the HTML
     # search page and regex out the model cards. Cached for 1 h so a busy

routes/email_helpers.py

@@ -1233,22 +1233,30 @@ def _list_attachments_from_msg(msg):
         return attachments
     idx = 0
     for part in msg.walk():
-        if part.is_multipart():
-            continue
         cd = str(part.get("Content-Disposition", ""))
         ct = part.get_content_type()
+        is_attached_email = ct == "message/rfc822" and ("attachment" in cd.lower() or part.get_filename())
+        if part.is_multipart() and not is_attached_email:
+            continue
         # Skip text/html body parts (only consider real attachments)
         if ct in ("text/plain", "text/html") and "attachment" not in cd:
             continue
         filename = part.get_filename()
         if filename:
             filename = _decode_header(filename)
+            if ct == "message/rfc822" and not re.search(r"\.[A-Za-z0-9]{1,8}$", filename):
+                filename = f"{filename}.eml"
         else:
             # Inline images, etc. - generate a name
-            ext = ct.split("/")[-1] if "/" in ct else "bin"
+            ext = "eml" if ct == "message/rfc822" else (ct.split("/")[-1] if "/" in ct else "bin")
             filename = f"attachment_{idx}.{ext}"
         payload = part.get_payload(decode=True)
-        size = len(payload) if payload else 0
+        if payload is None and ct == "message/rfc822":
+            try:
+                payload = part.as_bytes()
+            except Exception:
+                payload = b""
+        size = len(payload) if payload is not None else 0
         attachments.append({
             "index": idx,
             "filename": filename,
@@ -1260,29 +1268,58 @@ def _list_attachments_from_msg(msg):
     return attachments
 
 
+def _is_likely_signature_image_attachment(att: dict) -> bool:
+    """Match the reader's inline signature/logo image filter."""
+    filename = str((att or {}).get("filename") or "").lower()
+    if not re.search(r"\.(png|jpe?g|gif|bmp|svg|webp)$", filename):
+        return False
+    size = int((att or {}).get("size") or 0)
+    if re.search(r"^image\d{3,}\.(png|jpe?g|gif)$", filename):
+        return True
+    if re.search(r"^(signature|logo|sig|footer|banner)[-_\d]*\.(png|jpe?g|gif|svg)$", filename):
+        return True
+    return 0 < size < 30 * 1024
+
+
+def _has_visible_attachments(msg) -> bool:
+    """Return True only for attachments the reader will render as chips."""
+    return any(
+        not _is_likely_signature_image_attachment(att)
+        for att in _list_attachments_from_msg(msg)
+    )
+
+
 def _extract_attachment_to_disk(msg, index, target_dir):
     """Extract a specific attachment to disk and return the file path."""
     if not msg.is_multipart():
         return None
     idx = 0
     for part in msg.walk():
-        if part.is_multipart():
-            continue
         cd = str(part.get("Content-Disposition", ""))
         ct = part.get_content_type()
+        is_attached_email = ct == "message/rfc822" and ("attachment" in cd.lower() or part.get_filename())
+        if part.is_multipart() and not is_attached_email:
+            continue
         if ct in ("text/plain", "text/html") and "attachment" not in cd:
             continue
         if idx == index:
             filename = part.get_filename()
             if filename:
                 filename = _decode_header(filename)
+                if ct == "message/rfc822" and not re.search(r"\.[A-Za-z0-9]{1,8}$", filename):
+                    filename = f"{filename}.eml"
             else:
-                ext = ct.split("/")[-1] if "/" in ct else "bin"
+                ext = "eml" if ct == "message/rfc822" else (ct.split("/")[-1] if "/" in ct else "bin")
                 filename = f"attachment_{idx}.{ext}"
             # Sanitize
             safe_name = re.sub(r"[^\w\s\-.]", "_", filename).strip()
             payload = part.get_payload(decode=True)
-            if not payload:
+            if payload is None and ct == "message/rfc822":
+                try:
+                    payload = part.as_bytes()
+                except Exception:
+                    payload = b""
+            if payload is None:
                 return None
             target_dir.mkdir(parents=True, exist_ok=True)
             filepath = target_dir / safe_name

routes/email_routes.py

@@ -47,7 +47,7 @@ from routes.email_helpers import (
     _IMAP_TIMEOUT_SECONDS, _open_imap_connection,
     make_oauth_state, verify_oauth_state,
     _imap_connect, _imap, _decode_header, _detect_sent_folder, _detect_drafts_folder,
-    _extract_attachment_text, _list_attachments_from_msg,
+    _extract_attachment_text, _list_attachments_from_msg, _has_visible_attachments, _is_likely_signature_image_attachment,
     _extract_attachment_to_disk, _extract_html, _extract_text,
     _fetch_sender_thread_context, _pre_retrieve_context,
     _EMAIL_REPLY_SYS_PROMPT_BASE, _POOL_HOOKS,
@@ -61,6 +61,7 @@ from routes.email_pollers import _start_poller
 logger = logging.getLogger(__name__)
 
 ODYSSEUS_MAIL_ORIGIN = "odysseus-ui"
+EMAIL_READ_ATTACHMENT_VERSION = 2
 
 
 def _email_tag_owner_aliases(account_id: str | None, owner: str = "") -> list[str]:
@@ -248,6 +249,21 @@ def _imap_uid_fetch(conn, uid_set: str | bytes, query: str):
     return conn.uid("FETCH", _uid_bytes(uid_set), query)
 
 
+def _imap_search_quote(value: str) -> str:
+    return '"' + str(value or "").replace("\\", "\\\\").replace('"', '\\"') + '"'
+
+
+def _message_id_chain(*values: str) -> list[str]:
+    seen = set()
+    out = []
+    for value in values:
+        for mid in re.findall(r"<[^>]+>", value or ""):
+            if mid not in seen:
+                seen.add(mid)
+                out.append(mid)
+    return out
+
+
 def _uid_from_fetch_meta(meta_b: bytes) -> str:
     m = re.search(rb"\bUID\s+(\d+)\b", meta_b)
     return m.group(1).decode() if m else ""
@@ -366,6 +382,21 @@ def _apply_odysseus_headers(msg, kind: str | None = None, ref_id: str | None = N
         msg["X-Odysseus-Ref"] = re.sub(r"[^A-Za-z0-9_.:-]", "-", ref_id)[:128]
 
 
+def _normalize_addr_field(field: str) -> str:
+    """Strip the malformed-but-common trailing/leading commas and stray
+    whitespace from a To/Cc/Bcc string before it lands in the MIME header
+    or the SMTP envelope. Users often paste a single address with a
+    trailing comma (e.g. `felix@pewdiepie.com,`) and most MTAs reject the
+    resulting `To: felix@pewdiepie.com,` line as a syntax error. Collapse
+    any run of separator junk between addresses too."""
+    if not field:
+        return field
+    # Split on commas, drop empty tokens, rejoin with a single ', '.
+    parts = [p.strip() for p in field.split(",")]
+    parts = [p for p in parts if p]
+    return ", ".join(parts)
+
+
 def _envelope_recipients(*fields: str) -> list:
     """Extract bare SMTP envelope addresses from one or more To/Cc/Bcc header
     strings. A naive `field.split(",")` corrupts display names that contain a
@@ -994,6 +1025,65 @@ def setup_email_routes():
                 except Exception:
                     pass
 
+    def _related_thread_attachments_sync(
+        folder: str,
+        account_id: str | None,
+        owner: str,
+        current_uid: str,
+        current_message_id: str,
+        in_reply_to: str,
+        references: str,
+        limit: int = 12,
+    ) -> list[dict]:
+        """Return visible attachments from referenced messages in this folder."""
+        wanted_ids = _message_id_chain(references, in_reply_to)
+        current_mid = (current_message_id or "").strip()
+        wanted_ids = [mid for mid in wanted_ids if mid and mid != current_mid]
+        if not wanted_ids:
+            return []
+
+        related: list[dict] = []
+        try:
+            with _imap(account_id, owner=owner) as conn:
+                conn.select(_q(folder), readonly=True)
+                # Search newest referenced messages first; cap work so opening
+                # a long thread stays bounded.
+                for mid in reversed(wanted_ids[-10:]):
+                    if len(related) >= limit:
+                        break
+                    status, data = _imap_uid_search(conn, f'(HEADER Message-ID {_imap_search_quote(mid)})')
+                    if status != "OK" or not data or not data[0]:
+                        continue
+                    for uid_b in reversed(data[0].split()[-3:]):
+                        source_uid = uid_b.decode(errors="ignore")
+                        if not source_uid or source_uid == str(current_uid):
+                            continue
+                        st2, msg_data = _imap_uid_fetch(conn, source_uid, "(BODY.PEEK[])")
+                        if st2 != "OK" or not msg_data or not isinstance(msg_data[0], tuple):
+                            continue
+                        msg = email_mod.message_from_bytes(msg_data[0][1])
+                        source_from = _decode_header(msg.get("From", ""))
+                        source_subject = _decode_header(msg.get("Subject", ""))
+                        source_date = msg.get("Date", "")
+                        for att in _list_attachments_from_msg(msg):
+                            if _is_likely_signature_image_attachment(att):
+                                continue
+                            enriched = dict(att)
+                            enriched.update({
+                                "source_uid": source_uid,
+                                "source_folder": folder,
+                                "source_message_id": (msg.get("Message-ID") or "").strip(),
+                                "source_from": source_from,
+                                "source_subject": source_subject,
+                                "source_date": source_date,
+                            })
+                            related.append(enriched)
+                            if len(related) >= limit:
+                                break
+        except Exception as e:
+            logger.debug(f"related thread attachment lookup failed uid={current_uid}: {e}")
+        return related
+
     @router.get("/list")
     async def list_emails(
         folder: str = Query("INBOX"),
@@ -1264,6 +1354,17 @@ def setup_email_routes():
             sender_name, sender_addr = email.utils.parseaddr(sender)
             parsed_date = email.utils.parsedate_to_datetime(date_str) if date_str else None
             attachments = _list_attachments_from_msg(msg)
+            related_attachments = []
+            if not _has_visible_attachments(msg):
+                related_attachments = _related_thread_attachments_sync(
+                    folder,
+                    account_id,
+                    owner,
+                    uid,
+                    message_id,
+                    in_reply_to,
+                    references,
+                )
 
             if mark_seen:
                 # Set \Seen in a separate readwrite session so concurrent reads
@@ -1372,6 +1473,8 @@ def setup_email_routes():
                 "body": body,
                 "body_html": body_html,
                 "attachments": attachments,
+                "related_attachments": related_attachments,
+                "attachment_version": EMAIL_READ_ATTACHMENT_VERSION,
                 "cached_summary": cached_summary,
                 "cached_ai_reply": cached_ai_reply,
                 "boundaries": cached_boundaries,
@@ -1402,6 +1505,12 @@ def setup_email_routes():
         """Read email body. Cached for 30m, sync IMAP work runs in a thread."""
         ck = _read_cache_key(account_id, folder, uid, owner=owner)
         cached = _read_cache_get(ck)
+        if cached is not None:
+            # Older cached read responses lack the thread-attachment fallback.
+            # Fetch once so replies that reference prior attachments can show
+            # those files without waiting for cache expiry.
+            if cached.get("attachment_version") != EMAIL_READ_ATTACHMENT_VERSION:
+                cached = None
         if cached is not None:
             if mark_seen:
                 try:
@@ -1536,6 +1645,12 @@ def setup_email_routes():
                 return {"error": f"Attachment index {index} not found"}
 
             from pathlib import Path as _Path
+            target_root = os.path.abspath(str(target_dir))
+            filepath_str = os.path.abspath(str(filepath))
+            if os.path.commonpath([target_root, filepath_str]) != target_root:
+                logger.warning("Rejected attachment path outside extraction dir: %s", filepath)
+                return {"error": "Invalid attachment path"}
+            filepath = _Path(filepath_str)
             base = _Path(filepath).name
             if base.startswith("."):
                 return {"error": "Invalid filename", "filename": base}
@@ -1590,6 +1705,65 @@ def setup_email_routes():
                     return None
             doc_session_id = _resolve_doc_session()
 
+            def _create_markdown_doc(content: str, summary: str):
+                from src.database import SessionLocal as _SL, Document as _Doc, DocumentVersion as _DV
+                doc_id = str(uuid.uuid4())
+                ver_id = str(uuid.uuid4())
+                _db = _SL()
+                try:
+                    _db.query(_Doc).filter(_Doc.is_active == True).update({"is_active": False})
+                    _db.add(_Doc(
+                        id=doc_id, session_id=doc_session_id, title=title,
+                        language="markdown", current_content=content,
+                        version_count=1, is_active=True,
+                    ))
+                    _db.add(_DV(
+                        id=ver_id, document_id=doc_id, version_number=1,
+                        content=content, summary=summary, source="upload",
+                    ))
+                    _db.commit()
+                finally:
+                    _db.close()
+                _tag_doc_with_source(doc_id)
+                return doc_id
+
+            def _attached_email_markdown(raw_bytes: bytes):
+                if not raw_bytes:
+                    return f"# Attached email: {base}\n\n_(empty email attachment)_"
+                try:
+                    attached_msg = email_mod.message_from_bytes(raw_bytes)
+                except Exception:
+                    logger.exception("Failed to parse attached email %s", base)
+                    return f"# Attached email: {base}\n\nCould not parse this email attachment."
+
+                attached_subject = _decode_header(attached_msg.get("Subject", "")) or base
+                attached_from = _decode_header(attached_msg.get("From", ""))
+                attached_to = _decode_header(attached_msg.get("To", ""))
+                attached_cc = _decode_header(attached_msg.get("Cc", ""))
+                attached_date = attached_msg.get("Date", "")
+                attached_body = _extract_text(attached_msg).strip()
+                attached_atts = _list_attachments_from_msg(attached_msg)
+
+                lines = [f"# Attached email: {attached_subject}", ""]
+                if attached_from:
+                    lines.append(f"**From:** {attached_from}")
+                if attached_to:
+                    lines.append(f"**To:** {attached_to}")
+                if attached_cc:
+                    lines.append(f"**Cc:** {attached_cc}")
+                if attached_date:
+                    lines.append(f"**Date:** {attached_date}")
+                lines.extend(["", "## Body", "", attached_body or "_(no readable body)_"])
+                if attached_atts:
+                    lines.extend(["", "## Attachments", ""])
+                    for att in attached_atts:
+                        size = int(att.get("size") or 0)
+                        size_label = f"{size} B" if size < 1024 else f"{round(size / 1024)} KB"
+                        name = att.get("filename") or f"attachment_{att.get('index', '')}"
+                        ctype = att.get("content_type") or "application/octet-stream"
+                        lines.append(f"- {name} ({ctype}, {size_label})")
+                return "\n".join(lines).strip()
+
             # ── PDF path (existing) ────────────────────────────────────
             if ext == ".pdf":
                 import shutil as _shutil
@@ -1636,6 +1810,39 @@ def setup_email_routes():
                 _tag_doc_with_source(doc_id)
                 return {"doc_id": doc_id, "filename": filepath.name}
 
+            # ── Attached email (.eml / message/rfc822) ────────────────
+            if ext == ".eml":
+                def _attachment_bytes_from_msg():
+                    if not msg.is_multipart():
+                        return b""
+                    idx = 0
+                    for part in msg.walk():
+                        cd = str(part.get("Content-Disposition", ""))
+                        ct = part.get_content_type()
+                        is_attached_email = ct == "message/rfc822" and ("attachment" in cd.lower() or part.get_filename())
+                        if part.is_multipart() and not is_attached_email:
+                            continue
+                        if ct in ("text/plain", "text/html") and "attachment" not in cd:
+                            continue
+                        if idx == index:
+                            payload = part.get_payload(decode=True)
+                            if payload is None and ct == "message/rfc822":
+                                try:
+                                    payload = part.as_bytes()
+                                except Exception:
+                                    payload = b""
+                            return payload or b""
+                        idx += 1
+                    return b""
+
+                try:
+                    content = _attached_email_markdown(_attachment_bytes_from_msg())
+                except Exception:
+                    logger.exception("Failed to read email attachment %s", base)
+                    return {"error": "Failed to read email attachment", "filename": base}
+                doc_id = _create_markdown_doc(content, "Imported attached email")
+                return {"doc_id": doc_id, "filename": filepath.name}
+
             # ── DOCX path: extract text → markdown document ───────────
             if ext == ".docx":
                 try:
@@ -1673,25 +1880,7 @@ def setup_email_routes():
                     lines.append("")
                 content = "\n".join(lines).strip() or f"_(empty {base})_"
 
-                from src.database import SessionLocal as _SL, Document as _Doc, DocumentVersion as _DV
-                doc_id = str(uuid.uuid4())
-                ver_id = str(uuid.uuid4())
-                _db = _SL()
-                try:
-                    _db.query(_Doc).filter(_Doc.is_active == True).update({"is_active": False})
-                    _db.add(_Doc(
-                        id=doc_id, session_id=doc_session_id, title=title,
-                        language="markdown", current_content=content,
-                        version_count=1, is_active=True,
-                    ))
-                    _db.add(_DV(
-                        id=ver_id, document_id=doc_id, version_number=1,
-                        content=content, summary="Imported from DOCX", source="upload",
-                    ))
-                    _db.commit()
-                finally:
-                    _db.close()
-                _tag_doc_with_source(doc_id)
+                doc_id = _create_markdown_doc(content, "Imported from DOCX")
                 return {"doc_id": doc_id, "filename": filepath.name}
 
             # ── Plain text / markdown ────────────────────────────────
@@ -1700,25 +1889,7 @@ def setup_email_routes():
                     content = filepath.read_text(encoding="utf-8", errors="replace")
                 except Exception as e:
                     return {"error": f"Failed to read text file: {e}", "filename": base}
-                from src.database import SessionLocal as _SL, Document as _Doc, DocumentVersion as _DV
-                doc_id = str(uuid.uuid4())
-                ver_id = str(uuid.uuid4())
-                _db = _SL()
-                try:
-                    _db.query(_Doc).filter(_Doc.is_active == True).update({"is_active": False})
-                    _db.add(_Doc(
-                        id=doc_id, session_id=doc_session_id, title=title,
-                        language="markdown", current_content=content,
-                        version_count=1, is_active=True,
-                    ))
-                    _db.add(_DV(
-                        id=ver_id, document_id=doc_id, version_number=1,
-                        content=content, summary="Imported from email attachment", source="upload",
-                    ))
-                    _db.commit()
-                finally:
-                    _db.close()
-                _tag_doc_with_source(doc_id)
+                doc_id = _create_markdown_doc(content, "Imported from email attachment")
                 return {"doc_id": doc_id, "filename": filepath.name}
 
             return {"error": f"Unsupported attachment type: {ext}", "filename": base}
@@ -2027,6 +2198,9 @@ def setup_email_routes():
             outer = MIMEMultipart("alternative")
             body_container = outer
 
+        to = _normalize_addr_field(to or "")
+        cc = _normalize_addr_field(cc or "")
+        bcc = _normalize_addr_field(bcc or "")
         outer["From"] = email.utils.formataddr((cfg.get("display_name") or "", cfg["from_address"]))
         outer["To"] = to
         if cc:
@@ -2302,6 +2476,9 @@ def setup_email_routes():
             outer = MIMEMultipart("alternative")
             body_container = outer
 
+        req.to = _normalize_addr_field(req.to or "")
+        req.cc = _normalize_addr_field(req.cc or "")
+        req.bcc = _normalize_addr_field(req.bcc or "")
         outer["From"] = email.utils.formataddr((cfg.get("display_name") or "", cfg["from_address"]))
         outer["To"] = req.to
         if req.cc:

routes/hwfit_routes.py

@@ -1,8 +1,13 @@
+import json
+import os
 import re
+import shlex
+import subprocess
 from copy import deepcopy
 
 from fastapi import APIRouter, HTTPException
 
+from core.platform_compat import run_ssh_command
 from routes._validators import validate_remote_host, validate_ssh_port
 
 
@@ -107,6 +112,73 @@ def _apply_manual_hardware(system, manual_mode="", manual_gpu_count="", manual_v
     return system
 
 
+def _run_model_probe(host: str, ssh_port: str, cmd: str) -> str:
+    try:
+        if host:
+            r = run_ssh_command(
+                host,
+                ssh_port or None,
+                cmd,
+                timeout=15,
+                connect_timeout=5,
+                strict_host_key_checking=False,
+                text=True,
+            )
+        else:
+            r = subprocess.run(["bash", "-lc", cmd], capture_output=True, text=True, timeout=15)
+        if r.returncode == 0:
+            return (r.stdout or "").strip()
+    except Exception:
+        return ""
+    return ""
+
+
+def _inspect_model_path(model_path: str, host: str = "", ssh_port: str = "") -> dict:
+    """Read lightweight metadata from a local or SSH-visible HF model folder."""
+    path = (model_path or "").strip()
+    if not path or path.startswith(("http://", "https://")):
+        return {}
+    if not (path.startswith("/") or path.startswith("~")):
+        return {}
+
+    qpath = shlex.quote(path)
+    qconfig = shlex.quote(os.path.join(path, "config.json"))
+    out = {}
+    exists = _run_model_probe(host, ssh_port, f"test -d {qpath} && printf found || printf missing")
+    if exists != "found":
+        target = host or "local container"
+        out["model_probe_error"] = f"Model path is not visible on {target}: {path}"
+        return out
+    raw_config = _run_model_probe(host, ssh_port, f"test -f {qconfig} && sed -n '1,240p' {qconfig}")
+    if raw_config:
+        try:
+            cfg = json.loads(raw_config)
+        except Exception:
+            cfg = {}
+        for key in ("context_length", "max_position_embeddings", "n_ctx_train", "model_max_length", "max_seq_len"):
+            value = cfg.get(key)
+            if isinstance(value, (int, float)) and value > 0:
+                out["model_ctx_max"] = int(value)
+                break
+    else:
+        out["model_probe_error"] = f"config.json not found in model path: {path}"
+
+    size_cmd = (
+        f"find {qpath} -type f \\( -name '*.safetensors' -o -name '*.bin' -o -name '*.gguf' \\) "
+        "-printf '%s\\n' 2>/dev/null | awk '{s+=$1} END {if (s>0) printf \"%.6f\", s/1073741824}'"
+    )
+    weights = _run_model_probe(host, ssh_port, size_cmd)
+    try:
+        weights_gb = float(weights)
+    except Exception:
+        weights_gb = 0.0
+    if weights_gb > 0:
+        out["model_weights_gb"] = round(weights_gb, 3)
+    elif "model_probe_error" not in out:
+        out["model_probe_error"] = f"No model weight files found in: {path}"
+    return out
+
+
 def setup_hwfit_routes():
     router = APIRouter(prefix="/api/hwfit", tags=["hwfit"])
 
@@ -235,7 +307,7 @@ def setup_hwfit_routes():
         return {"system": system, "models": results}
 
     @router.get("/profiles")
-    def get_serve_profiles(model: str = "", host: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, serve_weights_gb: float = 0.0, serve_quant: str = ""):
+    def get_serve_profiles(model: str = "", model_path: str = "", host: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, serve_weights_gb: float = 0.0, serve_quant: str = ""):
         """Compute llama.cpp serve profiles (Quality/Balanced/Speed) for `model`
         against the detected hardware on `host` (or local). Returns concrete
         flags (n_gpu_layers, n_cpu_moe, cache_type, ctx) the serve UI can apply.
@@ -260,8 +332,23 @@ def setup_hwfit_routes():
             # "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct".
             s = (s or "").lower().strip()
             s = s.split("/")[-1]                     # drop org prefix
-            s = re.sub(r"[-_.]?gguf$", "", s)        # drop trailing gguf marker
-            s = re.sub(r"[-_.](q\d[^/]*|iq\d[^/]*|fp8|bf16|f16|awq[^/]*|gptq[^/]*)$", "", s)
+            for suffix in ("-gguf", "_gguf", ".gguf", "gguf"):
+                if s.endswith(suffix):
+                    s = s[: -len(suffix)]
+                    break
+            cut_at = None
+            for idx, ch in enumerate(s):
+                if ch not in "-_." or idx + 1 >= len(s):
+                    continue
+                suffix = s[idx + 1:]
+                if (
+                    suffix in {"fp8", "bf16", "f16"}
+                    or suffix.startswith(("awq", "gptq", "iq"))
+                    or (suffix.startswith("q") and len(suffix) > 1 and suffix[1].isdigit())
+                ):
+                    cut_at = idx
+            if cut_at is not None:
+                s = s[:cut_at]
             return s
 
         m = catalog.get(model)
@@ -272,8 +359,16 @@ def setup_hwfit_routes():
                 if nn and (nn == want or want.endswith(nn) or nn.endswith(want)):
                     m = entry
                     break
+        path_meta = _inspect_model_path(model_path or model, host=host, ssh_port=ssh_port)
         if m is None:
-            return {"system": system, "profiles": [], "error": "model not in catalog"}
+            return {
+                "system": system,
+                "profiles": [],
+                "error": "model not in catalog",
+                "model_ctx_max": int(path_meta.get("model_ctx_max") or 0),
+                "model_weights_gb": float(path_meta.get("model_weights_gb") or 0),
+                "model_probe_error": path_meta.get("model_probe_error") or "",
+            }
         # Surface the model's trained context limit so the serve UI can clamp a
         # user-typed context down to it (asking for ctx > n_ctx_train overflows
         # and, with a quantized KV cache, can crash the GPU).
@@ -283,6 +378,16 @@ def setup_hwfit_routes():
             if isinstance(v, (int, float)) and v > 0:
                 model_ctx_max = int(v)
                 break
+        path_ctx_max = int(path_meta.get("model_ctx_max") or 0)
+        if path_ctx_max > 0:
+            model_ctx_max = max(model_ctx_max, path_ctx_max)
+        model_weights_gb = float(path_meta.get("model_weights_gb") or 0)
+        if model_weights_gb <= 0:
+            for k in ("min_vram_gb", "required_gb", "size_gb", "recommended_ram_gb", "min_ram_gb"):
+                v = m.get(k)
+                if isinstance(v, (int, float)) and v > 0:
+                    model_weights_gb = float(v)
+                    break
         return {
             "system": system,
             "profiles": compute_serve_profiles(
@@ -291,6 +396,8 @@ def setup_hwfit_routes():
                 serve_quant=(serve_quant or None),
             ),
             "model_ctx_max": model_ctx_max,
+            "model_weights_gb": model_weights_gb,
+            "model_probe_error": path_meta.get("model_probe_error") or "",
         }
 
     @router.get("/image-models")

routes/model_routes.py

@@ -406,8 +406,11 @@ def _endpoint_refresh_timeout(ep: Any, category: str) -> float:
     except Exception:
         val = 0
     if val > 0:
-        return float(max(1, min(30, val)))
-    return 2.5 if category == "local" else 2.0
+        return float(max(1, min(60, val)))
+    # llama.cpp and other local OpenAI-compatible servers can block briefly
+    # while warming/loading. A 2s local timeout makes working endpoints flicker
+    # offline before /v1/models is ready.
+    return 10.0 if category == "local" else 2.0
 
 
 def _manual_refresh_timeout(ep: Any, category: str, requested: Any = None) -> float:
@@ -474,7 +477,7 @@ def _explicit_model_list_timeout(base_url: str, endpoint_kind: str = "auto", req
     category = _classify_endpoint(base_url, kind)
     if kind in ("api", "proxy") or category == "api":
         return 30.0
-    return 3.0 if _is_ollama_base(base_url) else 2.0
+    return 15.0 if category == "local" else (3.0 if _is_ollama_base(base_url) else 2.0)
 
 
 def _cached_model_ids(ep: Any) -> List[str]:
@@ -579,6 +582,18 @@ def _safe_build_headers(api_key: Optional[str], base_url: str) -> dict:
         return {"Authorization": f"Bearer {api_key}"} if api_key else {}
 
 
+def _redact_url_for_log(url: str) -> str:
+    """Return a URL safe for logs by removing userinfo and query/fragment."""
+    try:
+        parsed = urlparse(url or "")
+        host = parsed.hostname or ""
+        if parsed.port:
+            host = f"{host}:{parsed.port}"
+        return urlunparse((parsed.scheme, host, parsed.path, "", "", ""))
+    except Exception:
+        return "<endpoint>"
+
+
 def _is_discovery_only_provider(provider: str) -> bool:
     return provider == "chatgpt-subscription"
 
@@ -711,6 +726,16 @@ def _effective_endpoint_kind(ep: Any, base_url: str) -> str:
     return "auto"
 
 
+def _is_loading_model_response(resp: Any) -> bool:
+    if getattr(resp, "status_code", None) != 503:
+        return False
+    try:
+        body = resp.text or ""
+    except Exception:
+        body = ""
+    return "loading model" in body.lower()
+
+
 
 def _probe_endpoint(base_url: str, api_key: str = None, timeout: int = 5) -> List[str]:
     """Probe a base URL's /models endpoint and return list of model IDs.
@@ -775,11 +800,14 @@ def _probe_endpoint(base_url: str, api_key: str = None, timeout: int = 5) -> Lis
                         models.append(_e)
             return [m for m in models if _is_chat_model(m)]
     except httpx.HTTPStatusError as e:
+        if e.response is not None and _is_loading_model_response(e.response):
+            logger.info("Endpoint still loading model at %s", _redact_url_for_log(url))
+            return []
         if api_key:
             status = e.response.status_code if e.response is not None else "unknown"
-            logger.warning(f"Failed to probe {url} with API key: HTTP {status}")
+            logger.warning("Failed to probe %s with API key: HTTP %s", _redact_url_for_log(url), status)
             return []
-        logger.warning(f"Failed to probe {url}: {e}")
+        logger.warning("Failed to probe %s: %s", _redact_url_for_log(url), e)
     except Exception as e:
         if api_key:
             logger.warning(f"Failed to probe {url} with API key: {e}")
@@ -824,6 +852,15 @@ def _ping_endpoint(base_url: str, api_key: str = None, timeout: float = 1.5) ->
         or "ollama" in (parsed_base.hostname or "").lower()
     )
 
+    def _is_loading_model_response(r) -> bool:
+        if getattr(r, "status_code", None) != 503:
+            return False
+        try:
+            body = r.text or ""
+        except Exception:
+            body = ""
+        return "loading model" in body.lower()
+
     def _result_from_response(r) -> Dict[str, Any]:
         if 300 <= r.status_code < 400:
             loc = r.headers.get("location", "")
@@ -840,6 +877,13 @@ def _ping_endpoint(base_url: str, api_key: str = None, timeout: float = 1.5) ->
                 "status_code": r.status_code,
                 "error": None,
             }
+        if _is_loading_model_response(r):
+            return {
+                "reachable": True,
+                "loading": True,
+                "status_code": r.status_code,
+                "error": "Loading model",
+            }
         return {"reachable": False, "status_code": r.status_code, "error": f"HTTP {r.status_code}"}
 
     last_error: Optional[str] = None
@@ -872,7 +916,7 @@ def _ping_endpoint(base_url: str, api_key: str = None, timeout: float = 1.5) ->
         if 400 <= sc < 500 and sc not in (401, 403):
             models_url = _safe_build_models_url(base)
             try:
-                r2 = httpx.get(models_url, headers=headers, timeout=timeout, verify=llm_verify())
+                r2 = httpx.get(models_url, headers=headers,timeout=timeout, verify=llm_verify())
                 result2 = _result_from_response(r2)
                 if result2["reachable"]:
                     return result2
@@ -1056,9 +1100,11 @@ def setup_model_routes(model_discovery):
         except Exception:
             return 0.0
 
-    def _failure_delay(fails: int) -> float:
+    def _failure_delay(fails: int, *, empty_local: bool = False) -> float:
         if fails <= 0:
             return 0.0
+        if empty_local:
+            return min(5.0 * (2 ** max(0, fails - 1)), 30.0)
         return min(_REFRESH_FAILURE_BASE * (2 ** max(0, fails - 1)), _REFRESH_FAILURE_MAX)
 
     def _should_refresh_endpoint(ep: Any, now: float, force: bool = False) -> tuple[bool, Dict[str, Any]]:
@@ -1089,7 +1135,12 @@ def setup_model_routes(model_discovery):
         fails = int(state.get("fail_count") or 0)
         if fails and not force:
             last_failure = float(state.get("last_failure") or 0.0)
-            if now - last_failure < _failure_delay(fails):
+            empty_local = (
+                not cached
+                and category == "local"
+                and str(getattr(ep, "id", "") or "").startswith("local-")
+            )
+            if now - last_failure < _failure_delay(fails, empty_local=empty_local):
                 return False, info
         if cached and not force:
             interval = _endpoint_refresh_interval(ep, category)
@@ -1404,7 +1455,7 @@ def setup_model_routes(model_discovery):
                 t0 = _time.time()
                 ping = _ping_endpoint(base, ep.api_key, timeout=1.5)
                 entry["latency_ms"] = round((_time.time() - t0) * 1000)
-                entry["status"] = "online" if ping.get("reachable") or cached_count else "offline"
+                entry["status"] = "loading" if ping.get("loading") else ("online" if ping.get("reachable") or cached_count else "offline")
                 entry["error"] = ping.get("error")
                 entry["model_count"] = cached_count or (len(ANTHROPIC_MODELS) if provider == "anthropic" else 0)
             except Exception as e:
@@ -1578,9 +1629,37 @@ def setup_model_routes(model_discovery):
                 # "everything's already cached" path because this branch only
                 # runs for endpoints with an empty cached_models.
                 if not all_models and not pinned and r.is_enabled:
-                    ping = _ping_endpoint(r.base_url, r.api_key, timeout=3.5)
+                    base_for_ping = _normalize_base(r.base_url)
+                    kind_for_ping = _effective_endpoint_kind(r, base_for_ping)
+                    ping_timeout = 10.0 if _classify_endpoint(base_for_ping, kind_for_ping) == "local" else 3.5
+                    ping = _ping_endpoint(r.base_url, r.api_key, timeout=ping_timeout)
                     if ping.get("reachable"):
-                        status = "empty"
+                        status = "loading" if ping.get("loading") else "empty"
+                        if ping.get("loading"):
+                            base = _normalize_base(r.base_url)
+                            kind = _effective_endpoint_kind(r, base)
+                            results.append({
+                                "id": r.id,
+                                "name": r.name,
+                                "base_url": r.base_url,
+                                "has_key": bool(r.api_key),
+                                "api_key_fingerprint": _api_key_fingerprint(r.api_key),
+                                "is_enabled": r.is_enabled,
+                                "models": visible,
+                                "pinned_models": pinned,
+                                "hidden_count": len(hidden),
+                                "online": True,
+                                "status": status,
+                                "ping_error": (ping or {}).get("error") if ping else None,
+                                "model_type": getattr(r, "model_type", None) or "llm",
+                                "supports_tools": getattr(r, "supports_tools", None),
+                                "endpoint_kind": kind,
+                                "category": _classify_endpoint(base, kind),
+                                "model_refresh_mode": _endpoint_refresh_mode(r, kind),
+                                "model_refresh_interval": getattr(r, "model_refresh_interval", None),
+                                "model_refresh_timeout": getattr(r, "model_refresh_timeout", None),
+                            })
+                            continue
                         # Best-effort: if the probe came back reachable, try
                         # to populate cached_models in the background so the
                         # NEXT picker load shows "online" instead of "empty".
@@ -1588,7 +1667,7 @@ def setup_model_routes(model_discovery):
                         # "empty" status, and the existing background refresh
                         # path will eventually fill it in too.
                         try:
-                            probed = _probe_endpoint(r.base_url, r.api_key, timeout=5)
+                            probed = _probe_endpoint(r.base_url, r.api_key, timeout=max(5, int(ping_timeout)))
                             if probed:
                                 r.cached_models = json.dumps(probed)
                                 db.commit()
@@ -1766,7 +1845,7 @@ def setup_model_routes(model_discovery):
         model_ids = _probe_endpoint(base_url, api_key.strip() or None, timeout=explicit_timeout) if should_probe else []
         ping = {"reachable": False, "error": None}
         if (should_probe or requested_kind in ("api", "proxy")) and not model_ids:
-            ping = _ping_endpoint(base_url, api_key.strip() or None, timeout=min(explicit_timeout, 2.0))
+            ping = _ping_endpoint(base_url, api_key.strip() or None, timeout=min(explicit_timeout, 10.0))
         if require_model_list and not model_ids:
             raise HTTPException(400, _model_endpoint_error_message(base_url, ping))
 
@@ -1833,7 +1912,7 @@ def setup_model_routes(model_discovery):
             "models": _merge_model_ids(model_ids, _pinned),
             "pinned_models": _pinned,
             "online": bool(model_ids) or bool(_pinned) or bool(ping.get("reachable")),
-            "status": "online" if (model_ids or _pinned) else ("empty" if ping.get("reachable") else "offline"),
+            "status": "online" if (model_ids or _pinned) else ("loading" if ping.get("loading") else ("empty" if ping.get("reachable") else "offline")),
             "ping_error": ping.get("error") if ping else None,
             "endpoint_kind": requested_kind,
             "category": _classify_endpoint(base_url, requested_kind),
@@ -1858,11 +1937,11 @@ def setup_model_routes(model_discovery):
         configured_timeout = _parse_positive_int(model_refresh_timeout, minimum=1, maximum=60)
         probe_timeout = _explicit_model_list_timeout(base_url, requested_kind, configured_timeout)
         models = _probe_endpoint(base_url, api_key.strip() or None, timeout=probe_timeout)
-        ping = {"reachable": True, "error": None} if models else _ping_endpoint(base_url, api_key.strip() or None, timeout=min(probe_timeout, 2.0))
+        ping = {"reachable": True, "error": None} if models else _ping_endpoint(base_url, api_key.strip() or None, timeout=min(probe_timeout, 10.0))
         return {
             "base_url": base_url,
             "online": bool(models) or bool(ping.get("reachable")),
-            "status": "online" if models else ("empty" if ping.get("reachable") else "offline"),
+            "status": "online" if models else ("loading" if ping.get("loading") else ("empty" if ping.get("reachable") else "offline")),
             "ping_error": ping.get("error") if ping else None,
             "models": models,
             "count": len(models),

routes/shell_routes.py

@@ -331,6 +331,9 @@ def add_user_install_bins_to_path():
         candidates.append(os.path.join(site.USER_BASE, 'bin'))
     except Exception:
         pass
+    candidates.append(os.path.expanduser('~/bin'))
+    candidates.append(os.path.expanduser('~/llama.cpp/build/bin'))
+    candidates.append(os.path.expanduser('~/llama.cpp/build-vulkan/bin'))
     candidates.append(os.path.expanduser('~/.local/bin'))
     parts = os.environ.get('PATH', '').split(os.pathsep) if os.environ.get('PATH') else []
     changed = False
@@ -962,12 +965,84 @@ def setup_shell_routes() -> APIRouter:
 
         return StreamingResponse(generate(), media_type="text/event-stream")
 
+    def _os_id_from_release(text: str) -> str:
+        """Map /etc/os-release contents to a canonical family for our matrix."""
+        if not text:
+            return ""
+        ids = []
+        for line in text.splitlines():
+            line = line.strip()
+            if line.startswith("ID=") or line.startswith("ID_LIKE="):
+                ids += line.split("=", 1)[1].strip().strip('"').split()
+        ids = [i.lower() for i in ids]
+        if any(x in ids for x in ("debian", "ubuntu", "linuxmint", "pop", "elementary")):
+            return "debian"
+        if any(x in ids for x in ("arch", "manjaro", "endeavouros", "cachyos", "garuda")):
+            return "arch"
+        if any(x in ids for x in ("fedora", "rhel", "centos", "rocky", "almalinux", "ol")):
+            return "fedora"
+        if "alpine" in ids:
+            return "alpine"
+        if any(x in ids for x in ("suse", "opensuse", "opensuse-leap", "opensuse-tumbleweed", "sles")):
+            return "suse"
+        return ""
+
+    # Matrix lookup keyed on (os_family, backend) → (pkg_mgr_cmd_template, pkg_list_per_dep).
+    # Each `system_prereqs` name resolves to a list of OS-specific package
+    # names that get joined into the final `sudo apt install -y …` etc.
+    # command. Backend-specific extras (CUDA toolkit, ROCm, Vulkan headers)
+    # are added only when the detected backend needs them.
+    _PKG_NAMES = {
+        # canonical-name → {os_id: [actual_pkg_names_on_this_os]}
+        "cmake":           {"debian": ["cmake"], "arch": ["cmake"], "fedora": ["cmake"], "alpine": ["cmake"], "suse": ["cmake"], "macos": ["cmake"]},
+        "build-essential": {"debian": ["build-essential"], "arch": ["base-devel"], "fedora": ["gcc", "gcc-c++", "make"], "alpine": ["build-base"], "suse": ["gcc-c++", "make"], "macos": []},
+        "g++":             {"debian": ["g++"], "arch": ["gcc"], "fedora": ["gcc-c++"], "alpine": ["g++"], "suse": ["gcc-c++"], "macos": []},
+        "gcc":             {"debian": ["gcc"], "arch": ["gcc"], "fedora": ["gcc"], "alpine": ["gcc"], "suse": ["gcc"], "macos": []},
+        "make":            {"debian": ["make"], "arch": ["make"], "fedora": ["make"], "alpine": ["make"], "suse": ["make"], "macos": []},
+        "git":             {"debian": ["git"], "arch": ["git"], "fedora": ["git"], "alpine": ["git"], "suse": ["git"], "macos": ["git"]},
+        "tmux":            {"debian": ["tmux"], "arch": ["tmux"], "fedora": ["tmux"], "alpine": ["tmux"], "suse": ["tmux"], "macos": ["tmux"]},
+    }
+    _BACKEND_EXTRAS = {
+        "cuda":   {"debian": ["nvidia-cuda-toolkit"], "arch": ["cuda"], "fedora": ["cuda-toolkit"], "alpine": [], "suse": ["cuda"], "macos": []},
+        "rocm":   {"debian": ["rocm-dev"], "arch": ["rocm-hip-sdk"], "fedora": ["rocm-devel"], "alpine": [], "suse": ["rocm-dev"], "macos": []},
+        "vulkan": {"debian": ["libvulkan-dev", "vulkan-tools"], "arch": ["vulkan-headers", "vulkan-tools"], "fedora": ["vulkan-headers", "vulkan-tools"], "alpine": ["vulkan-loader-dev", "vulkan-tools"], "suse": ["vulkan-devel", "vulkan-tools"], "macos": []},
+    }
+    _PKG_MGR = {
+        "debian": "sudo apt install -y {pkgs}",
+        "arch":   "sudo pacman -S --needed {pkgs}",
+        "fedora": "sudo dnf install -y {pkgs}",
+        "alpine": "sudo apk add {pkgs}",
+        "suse":   "sudo zypper install -n {pkgs}",
+        "macos":  "brew install {pkgs}",
+    }
+
+    def _install_cmd_for_target(os_id: str, backend: str, missing: list[str]) -> str:
+        """Build a single OS+backend-aware install command for the missing prereqs."""
+        if not os_id or os_id not in _PKG_MGR:
+            return ""
+        pkgs: list[str] = []
+        seen: set[str] = set()
+        for m in missing:
+            for p in _PKG_NAMES.get(m, {}).get(os_id, []):
+                if p not in seen:
+                    pkgs.append(p); seen.add(p)
+        # Add backend-specific extras only when the build would actually
+        # consume them (a CUDA toolkit isn't useful on a Vulkan box).
+        backend = (backend or "").lower()
+        for p in _BACKEND_EXTRAS.get(backend, {}).get(os_id, []):
+            if p not in seen:
+                pkgs.append(p); seen.add(p)
+        if not pkgs:
+            return ""
+        return _PKG_MGR[os_id].format(pkgs=" ".join(pkgs))
+
     @router.get("/api/cookbook/packages")
     async def list_packages(
         request: Request,
         host: str | None = None,
         ssh_port: str | None = None,
         venv: str | None = None,
+        backend: str | None = None,
     ):
         """Check which optional packages are installed.
 
@@ -1016,6 +1091,12 @@ def setup_shell_routes() -> APIRouter:
                 "kind": "system",
                 "install_hint": "Install Docker on the selected server and allow this user to run docker.",
             },
+            # Note: cmake / gcc / git are not separate dependency rows —
+            # they're declared as `system_prereqs` on llama_cpp (and any
+            # other engine that compiles from source) so they appear as
+            # an inline status note on that engine's row instead of
+            # cluttering the panel with raw OS package names that aren't
+            # meaningful product-level dependencies on their own.
             # ── LLM ── installs on GPU servers for model serving/downloading
             {
                 "name": "hf_transfer",
@@ -1027,9 +1108,16 @@ def setup_shell_routes() -> APIRouter:
             {
                 "name": "llama_cpp",
                 "pip": "llama-cpp-python[server]",
-                "desc": "Serve GGUF models via llama.cpp",
+                "desc": "Great for single-GPU or CPU inference with GGUF models",
                 "category": "LLM",
                 "target": "remote",
+                # Build-toolchain prereqs. Cookbook's launch bootstrap
+                # compiles llama-server from source when no prebuilt
+                # binary is present; without these the build aborts
+                # with `cmake: command not found`. Surfaced inline on
+                # this row so the user doesn't have to chase three
+                # separate OS-package rows.
+                "system_prereqs": ["cmake", "g++", "git"],
             },
             {
                 "name": "sglang",
@@ -1041,7 +1129,7 @@ def setup_shell_routes() -> APIRouter:
             {
                 "name": "vllm",
                 "pip": "vllm",
-                "desc": "High-throughput LLM serving engine",
+                "desc": "Great for high-throughput multi-GPU inference",
                 "category": "LLM",
                 "target": "remote",
             },
@@ -1104,6 +1192,7 @@ def setup_shell_routes() -> APIRouter:
         # venv over SSH so a remote `pip install` actually reflects here.
         remote_status: dict = {}
         remote_details: dict = {}
+        remote_probe_error = ""
         remote_names = [
             p["name"]
             for p in packages
@@ -1142,16 +1231,56 @@ def setup_shell_routes() -> APIRouter:
                         break
             except ValueError as e:
                 raise HTTPException(400, str(e))
-            except Exception:
+            except Exception as e:
                 remote_status = {}
-        if host and remote_system_names:
+                remote_probe_error = f"SSH package probe failed: {str(e)[:160]}"
+            if "llama_cpp" in remote_names:
+                try:
+                    inner = (
+                        'export PATH="$HOME/.local/bin:$HOME/bin:'
+                        '$HOME/llama.cpp/build/bin:$HOME/llama.cpp/build-vulkan/bin:$PATH"; '
+                        "command -v llama-server 2>/dev/null || true"
+                    )
+                    argv = _ssh_base_argv(host, ssh_port) + [inner]
+                    proc = await asyncio.create_subprocess_exec(
+                        *argv,
+                        stdout=asyncio.subprocess.PIPE,
+                        stderr=asyncio.subprocess.PIPE,
+                    )
+                    out, _err = await asyncio.wait_for(proc.communicate(), timeout=8)
+                    llama_server_path = out.decode("utf-8", errors="replace").strip().splitlines()
+                    llama_server_path = llama_server_path[-1].strip() if llama_server_path else ""
+                    if llama_server_path:
+                        remote_status["llama_cpp"] = True
+                        probe = remote_details.setdefault("llama_cpp", {})
+                        if isinstance(probe, dict):
+                            probe.setdefault("binaries", {})["llama-server"] = llama_server_path
+                except Exception as e:
+                    if not remote_probe_error:
+                        remote_probe_error = f"SSH llama-server probe failed: {str(e)[:160]}"
+                    pass
+        # Union of system_names + every package's system_prereqs. Probing
+        # the prereqs alongside the main system deps in a single SSH call
+        # avoids a second round-trip per Cookbook → Dependencies refresh.
+        prereq_names: set[str] = set()
+        for p in packages:
+            for pr in p.get("system_prereqs") or []:
+                prereq_names.add(str(pr))
+        all_system_names = list(set(remote_system_names) | prereq_names)
+        # Detect the target's OS family + read /etc/os-release in the same
+        # SSH round-trip as the prereq probe — used downstream to render a
+        # single OS-specific install command per row instead of dumping
+        # every distro's syntax onto the user.
+        target_os_id: str = ""
+        if host and all_system_names:
             try:
                 checks = []
-                for name in remote_system_names:
+                for name in all_system_names:
                     qn = shlex.quote(name)
                     checks.append(
                         f"if command -v {qn} >/dev/null 2>&1; then echo {qn}=1; else echo {qn}=0; fi"
                     )
+                checks.append("echo '---OSREL---'; cat /etc/os-release 2>/dev/null || true")
                 inner = " ; ".join(checks)
                 argv = _ssh_base_argv(host, ssh_port) + [inner]
                 proc = await asyncio.create_subprocess_exec(
@@ -1161,20 +1290,45 @@ def setup_shell_routes() -> APIRouter:
                 )
                 out, _err = await asyncio.wait_for(proc.communicate(), timeout=12)
                 txt = out.decode("utf-8", errors="replace").strip()
+                _section, _osrel_lines = "probe", []
                 for line in txt.splitlines():
+                    if line.strip() == "---OSREL---":
+                        _section = "osrel"; continue
+                    if _section == "osrel":
+                        _osrel_lines.append(line)
+                        continue
                     name, sep, value = line.strip().partition("=")
-                    if sep and name in remote_system_names:
+                    if sep and name in all_system_names:
                         remote_status[name] = value == "1"
+                target_os_id = _os_id_from_release("\n".join(_osrel_lines))
             except ValueError as e:
                 raise HTTPException(400, str(e))
-            except Exception:
+            except Exception as e:
+                if not remote_probe_error:
+                    remote_probe_error = f"SSH system probe failed: {str(e)[:160]}"
                 pass
+        elif not host:
+            # Local target — probe in-process so the inline install command
+            # still appears in the dep panel when the cookbook container
+            # itself is the selected server.
+            try:
+                with open("/etc/os-release", encoding="utf-8") as f:
+                    target_os_id = _os_id_from_release(f.read())
+            except Exception:
+                target_os_id = ""
+            if sys.platform == "darwin":
+                target_os_id = "macos"
 
         for pkg in packages:
             on_remote = bool(host and pkg.get("target") == "remote")
             probe = None
             if on_remote:
-                pkg["installed"] = bool(remote_status.get(pkg["name"], False))
+                if remote_probe_error and pkg["name"] not in remote_status:
+                    pkg["installed"] = None
+                    pkg["probe_error"] = remote_probe_error
+                    pkg["status_note"] = remote_probe_error
+                else:
+                    pkg["installed"] = bool(remote_status.get(pkg["name"], False))
                 probe = remote_details.get(pkg["name"])
                 if isinstance(probe, dict):
                     pkg["details"] = probe
@@ -1230,6 +1384,104 @@ def setup_shell_routes() -> APIRouter:
                     # 500 the entire packages panel; report it as not usable.
                     pkg["installed"] = False
 
+            # llama_cpp partial-state probe: when the package is installed
+            # but the wheel was built CPU-only AND the target has NVIDIA
+            # hardware, mark the row as partial (yellow/orange) with a
+            # one-click upgrade to the CUDA wheel. Without this the row
+            # reads "ready" green while inference runs at 3 tok/s on GPU
+            # silicon — actively misleading.
+            if pkg["name"] == "llama_cpp" and pkg.get("installed"):
+                _native_llama_server = bool(
+                    isinstance(probe, dict)
+                    and isinstance(probe.get("binaries"), dict)
+                    and probe["binaries"].get("llama-server")
+                )
+                _gpu_capable = False
+                _has_nvidia_target = False
+                if _native_llama_server:
+                    # Native llama-server is the launcher path Cookbook now
+                    # prefers. Do not mark this as a CPU-only Python wheel just
+                    # because llama-cpp-python is absent from the selected venv.
+                    _gpu_capable = True
+                elif on_remote and host:
+                    try:
+                        # Activate the configured venv FIRST so the probe
+                        # runs against the same python the launch script
+                        # would activate. Without this prefix, bare
+                        # `python3` was checked — which can disagree with
+                        # the venv's wheel (e.g. user-site has CUDA wheel
+                        # but venv has CPU-only), and the dep panel then
+                        # showed "ready" green while every launch fell to
+                        # CPU.
+                        _vp = _venv_activate_prefix(venv)
+                        probe = (
+                            f'{_vp}python3 -c "import llama_cpp; import sys; '
+                            'sys.exit(0 if llama_cpp.llama_supports_gpu_offload() else 1)" '
+                            '&& echo llama_cpp_gpu=1 || echo llama_cpp_gpu=0; '
+                            'command -v nvidia-smi >/dev/null 2>&1 '
+                            '&& nvidia-smi -L 2>/dev/null | grep -q "GPU " '
+                            '&& echo nvidia=1 || echo nvidia=0'
+                        )
+                        argv = _ssh_base_argv(host, ssh_port) + [probe]
+                        proc = await asyncio.create_subprocess_exec(
+                            *argv, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE,
+                        )
+                        out, _ = await asyncio.wait_for(proc.communicate(), timeout=8)
+                        txt = out.decode("utf-8", errors="replace")
+                        if "llama_cpp_gpu=1" in txt:
+                            _gpu_capable = True
+                        if "nvidia=1" in txt:
+                            _has_nvidia_target = True
+                    except Exception:
+                        pass
+                else:
+                    try:
+                        import llama_cpp as _lcp  # type: ignore
+                        _gpu_capable = bool(_lcp.llama_supports_gpu_offload())
+                    except Exception:
+                        _gpu_capable = False
+                    _has_nvidia_target = shutil.which("nvidia-smi") is not None
+                if (not _gpu_capable) and _has_nvidia_target:
+                    pkg["partial"] = True
+                    pkg["partial_reason"] = "Installed but CPU-only wheel — GPU detected on this target. Upgrade to a CUDA wheel for ~10× faster inference."
+                    pkg["partial_action"] = "reinstall_llama_cpp_cuda"
+            # Attach per-package system_prereqs status. We probed each
+            # prereq name above; surface "Missing build deps: …" ONLY
+            # when the package itself is not installed — if the package
+            # works (e.g. llama-cpp-python already imports cleanly), the
+            # build toolchain is irrelevant and surfacing it as a red
+            # flag confuses users ("ready" + "missing" on the same row).
+            _prereqs = list(pkg.get("system_prereqs") or [])
+            if _prereqs:
+                if on_remote:
+                    _pr_present = {n: bool(remote_status.get(n)) for n in _prereqs}
+                else:
+                    _pr_present = {n: shutil.which(n) is not None for n in _prereqs}
+                pkg["system_prereqs_status"] = _pr_present
+                _missing = [n for n, ok in _pr_present.items() if not ok]
+                # Suppress the "missing build deps" hint when the package
+                # itself is installed — build deps are only relevant if
+                # the user would need to recompile from source.
+                if pkg.get("installed"):
+                    _missing = []
+                if _missing:
+                    # Build a target-specific install command from the
+                    # (os_family, backend) matrix when we know both. Fall
+                    # back to the multi-distro hint only when the target's
+                    # OS can't be classified (e.g. ssh probe failed).
+                    _resolved_os = target_os_id or "debian"  # safest default
+                    _cmd = _install_cmd_for_target(_resolved_os, backend or "", _missing)
+                    if _cmd and target_os_id:
+                        _hint = "Missing build deps for this target: " + ", ".join(_missing)
+                        pkg["install_cmd_for_target"] = _cmd
+                        pkg["install_cmd_os"] = target_os_id
+                        pkg["install_cmd_backend"] = (backend or "").lower()
+                    else:
+                        _hint = "Missing build deps: " + ", ".join(_missing) + ". Install via apt: cmake build-essential git / pacman: cmake base-devel git / dnf: cmake gcc-c++ make git / brew: cmake git."
+                    _existing_note = pkg.get("status_note") or ""
+                    pkg["status_note"] = (_existing_note + " — " + _hint) if _existing_note else _hint
+                    pkg["build_deps_missing"] = _missing
+
             if pkg.get("installed"):
                 update_status = _package_pip_update_status(pkg, probe)
                 pkg["pip_update_available"] = update_status.available
@@ -1289,6 +1541,102 @@ def setup_shell_routes() -> APIRouter:
             return {"ok": True, "output": stdout.decode()[-200:]}
         return {"ok": False, "error": stderr.decode()[-300:]}
 
+    @router.post("/api/cookbook/install-system-deps")
+    async def install_system_deps(request: Request):
+        """Install OS-level system packages (cmake/build-essential/git/tmux)
+        on a remote target or in the local container. Admin only.
+
+        Bounded by a per-package allowlist — anything outside the catalog
+        is rejected so the route can't be coerced into installing arbitrary
+        OS packages. Uses `sudo -n` (passwordless) so the call returns a
+        clear "needs sudo password" error instead of hanging when interactive
+        sudo is required.
+        """
+        _require_admin(request)
+        body = await request.json()
+        raw = body.get("packages") or []
+        host = (body.get("remote_host") or "").strip()
+        ssh_port = body.get("ssh_port")
+        # Names users can request — must match canonical names used in the
+        # deps catalog's `system_prereqs` field and on the System rows.
+        ALLOWED = {"cmake", "build-essential", "g++", "gcc", "git", "tmux", "make"}
+        pkgs = [str(p).strip() for p in raw if str(p).strip() in ALLOWED]
+        if not pkgs:
+            return {"ok": False, "error": "no installable packages requested (allowlist: " + ", ".join(sorted(ALLOWED)) + ")"}
+        # Re-map to the right package name per OS. apt/dpkg use the names
+        # as-is; pacman has base-devel for build-essential, etc.
+        def _apt(names): return list(names)
+        def _pacman(names):
+            return ["base-devel" if n == "build-essential" else n for n in names]
+        def _dnf(names):
+            out = []
+            for n in names:
+                if n == "build-essential": out += ["gcc", "gcc-c++", "make"]
+                elif n == "g++": out += ["gcc-c++"]
+                else: out.append(n)
+            return out
+        def _brew(names):
+            return [n for n in names if n not in ("build-essential", "g++", "gcc", "make")]
+        # Build a single shell snippet that detects the package manager and
+        # runs the right install. Non-interactive sudo (-n) only — if sudo
+        # asks for a password the script reports it instead of hanging.
+        apt_pkgs = " ".join(shlex.quote(p) for p in _apt(pkgs))
+        pac_pkgs = " ".join(shlex.quote(p) for p in _pacman(pkgs))
+        dnf_pkgs = " ".join(shlex.quote(p) for p in _dnf(pkgs))
+        brew_pkgs = " ".join(shlex.quote(p) for p in _brew(pkgs))
+        # Error messages go to stderr (>&2) so the route's error field
+        # gets populated. Without the redirect, `echo "ERROR…"` on stdout
+        # left stderr empty and the frontend toast fell through to a
+        # bare "HTTP 200" instead of surfacing the real reason.
+        script = (
+            'set -e; '
+            'if ! sudo -n true 2>/dev/null; then '
+            '  echo "ERROR: passwordless sudo unavailable on this target. Run once: sudo apt install -y ' + " ".join(pkgs) + ' (or your distro equivalent: pacman -S, dnf install, brew install). After that, Cookbook can install the rest." >&2; exit 2; fi; '
+            'if command -v apt-get >/dev/null 2>&1; then '
+            f'  sudo -n env DEBIAN_FRONTEND=noninteractive apt-get update -qq && sudo -n env DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends {apt_pkgs}; '
+            'elif command -v pacman >/dev/null 2>&1; then '
+            f'  sudo -n pacman -Sy --needed --noconfirm {pac_pkgs}; '
+            'elif command -v dnf >/dev/null 2>&1; then '
+            f'  sudo -n dnf install -y {dnf_pkgs}; '
+            'elif command -v brew >/dev/null 2>&1; then '
+            f'  brew install {brew_pkgs}; '
+            'else '
+            '  echo "ERROR: no supported package manager (apt/pacman/dnf/brew) on this target." >&2; exit 3; fi'
+        )
+        try:
+            if host:
+                argv = _ssh_base_argv(host, ssh_port) + [script]
+            else:
+                argv = ["bash", "-lc", script]
+        except ValueError as e:
+            raise HTTPException(400, str(e))
+        try:
+            proc = await asyncio.create_subprocess_exec(
+                *argv, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE
+            )
+            out, err = await asyncio.wait_for(proc.communicate(), timeout=180)
+        except asyncio.TimeoutError:
+            return {"ok": False, "error": "Install timed out after 180s"}
+        ok = (proc.returncode == 0)
+        # Combine stderr + (last lines of stdout) into a single error
+        # blob when ok=False — some package managers print useful failure
+        # context to stdout, and a script that exits via `echo ...; exit N`
+        # without `>&2` would otherwise hand back an empty error string
+        # and force the frontend to show a bare "HTTP 200".
+        err_txt = err.decode("utf-8", errors="replace").strip()
+        out_txt = out.decode("utf-8", errors="replace").strip()
+        if not ok:
+            tail_out = out_txt[-500:] if out_txt else ""
+            combined = err_txt or tail_out or f"exit code {proc.returncode}"
+        else:
+            combined = None
+        return {
+            "ok": ok,
+            "exit_code": proc.returncode,
+            "output": out_txt[-1000:],
+            "error": combined,
+        }
+
     @router.post("/api/cookbook/rebuild-engine")
     async def rebuild_engine(request: Request):
         """Clear the cached llama.cpp build so the next serve recompiles.
@@ -1309,7 +1657,8 @@ def setup_shell_routes() -> APIRouter:
             return {"ok": False, "error": f"Unsupported engine: {engine}"}
         host = str(body.get("remote_host") or "").strip()
         ssh_port = body.get("ssh_port")
-        cmd = _llama_cpp_rebuild_cmd()
+        update_source = bool(body.get("update_source"))
+        cmd = _llama_cpp_rebuild_cmd(update_source=update_source)
         try:
             argv = (
                 (_ssh_base_argv(host, ssh_port) + [cmd])

routes/upload_routes.py

@@ -3,11 +3,16 @@ import os
 import time
 import json
 import asyncio
+import shutil
+import uuid
+from pathlib import Path
 from fastapi import APIRouter, Request, File, UploadFile, HTTPException
 from typing import List
 import logging
 from core.middleware import require_admin
+from core.database import SessionLocal, GalleryImage
 from src.auth_helpers import effective_user
+from src.constants import GENERATED_IMAGES_DIR
 from src.upload_handler import count_recent_uploads
 
 logger = logging.getLogger(__name__)
@@ -50,6 +55,69 @@ def setup_upload_routes(upload_handler):
             raise HTTPException(404, "File not found")
 
         raise HTTPException(404, "File not found")
+
+    def _promote_chat_image_to_gallery(meta: dict, owner: str | None) -> str | None:
+        """Make chat-uploaded images visible in Gallery without changing chat storage."""
+        is_image_file = getattr(upload_handler, "is_image_file", None)
+        if not callable(is_image_file):
+            return None
+        if not is_image_file(meta.get("name", ""), meta.get("mime", "")):
+            return None
+
+        source_path = meta.get("path")
+        if not source_path or not os.path.isfile(source_path):
+            return None
+
+        db = SessionLocal()
+        try:
+            file_hash = meta.get("hash")
+            if file_hash:
+                q = db.query(GalleryImage).filter(
+                    GalleryImage.file_hash == file_hash,
+                    GalleryImage.is_active == True,  # noqa: E712
+                )
+                if owner:
+                    q = q.filter(GalleryImage.owner == owner)
+                existing = q.first()
+                if existing:
+                    return existing.id
+
+            image_dir = Path(GENERATED_IMAGES_DIR)
+            image_dir.mkdir(parents=True, exist_ok=True)
+            ext = Path(meta.get("name") or source_path).suffix.lower()
+            if ext not in {".png", ".jpg", ".jpeg", ".webp", ".gif"}:
+                mime_ext = {
+                    "image/png": ".png",
+                    "image/jpeg": ".jpg",
+                    "image/jpg": ".jpg",
+                    "image/webp": ".webp",
+                    "image/gif": ".gif",
+                }.get(meta.get("mime", ""))
+                ext = mime_ext or ".png"
+            filename = f"{uuid.uuid4().hex[:12]}{ext}"
+            dest_path = image_dir / filename
+            shutil.copy2(source_path, dest_path)
+
+            image_id = str(uuid.uuid4())
+            db.add(GalleryImage(
+                id=image_id,
+                filename=filename,
+                prompt=meta.get("name") or "Chat upload",
+                model="chat-upload",
+                owner=owner,
+                file_hash=file_hash,
+                width=meta.get("width"),
+                height=meta.get("height"),
+                file_size=meta.get("size"),
+            ))
+            db.commit()
+            return image_id
+        except Exception as e:
+            db.rollback()
+            logger.warning("Failed to add chat image upload to gallery: %s", e)
+            return None
+        finally:
+            db.close()
     
     @router.post("")
     async def api_upload(request: Request, files: List[UploadFile] = File(...)):
@@ -78,8 +146,10 @@ def setup_upload_routes(upload_handler):
         
         for u in files:
             try:
-                meta = upload_handler.save_upload(u, client_ip, owner=effective_user(request))
-                out.append({
+                owner = effective_user(request)
+                meta = upload_handler.save_upload(u, client_ip, owner=owner)
+                gallery_id = _promote_chat_image_to_gallery(meta, owner)
+                item = {
                     "id": meta["id"],
                     "name": meta["name"],
                     "mime": meta["mime"],
@@ -89,7 +159,10 @@ def setup_upload_routes(upload_handler):
                     "width": meta.get("width"),
                     "height": meta.get("height"),
                     "is_duplicate": meta.get("is_duplicate", False)
-                })
+                }
+                if gallery_id:
+                    item["gallery_id"] = gallery_id
+                out.append(item)
             except HTTPException:
                 raise
             except Exception as e:

services/hwfit/hardware.py

@@ -282,7 +282,17 @@ def _detect_amd():
             "gpus": cards,
             "gpu_groups": groups,
             "homogeneous": len(groups) <= 1,
-            "backend": "rocm",
+            # Pick the actual runtime label: ROCm/HIP only when its
+            # toolchain is installed, otherwise Vulkan if vulkaninfo is
+            # present (mesa RADV works fine on RDNA/CDNA when ROCm
+            # packages are absent — see Strix Halo where ROCm support
+            # is still backporting). Reporting "rocm" on a Vulkan-only
+            # host misleads downstream env-var pinning
+            # (HIP_VISIBLE_DEVICES is a no-op there).
+            "backend": (
+                "rocm" if (_run(["which", "rocminfo"]) or _run(["which", "hipconfig"]))
+                else ("vulkan" if _run(["which", "vulkaninfo"]) else "rocm")
+            ),
             "unified_memory": is_apu,
             # AMD ISA/family so downstream can tell datacenter Instinct (CDNA,
             # where vLLM/SGLang run AWQ/GPTQ reliably) from consumer Radeon

src/agent_loop.py

@@ -541,17 +541,44 @@ def _section_text(name: str, default: str) -> str:
     return val if isinstance(val, str) and val.strip() else default
 
 
+def _compact_tool_line(name: str, section: str) -> str:
+    """One-line fenced-tool usage hint for compact/local prompts."""
+    text = (section or "").strip()
+    if not text:
+        return f"- `{name}`"
+    if text.startswith("- "):
+        return text
+    lines = [ln.strip() for ln in text.splitlines() if ln.strip()]
+    usage = []
+    in_fence = False
+    for ln in lines:
+        if ln.startswith("```"):
+            usage.append(ln)
+            in_fence = not in_fence
+            if len(usage) >= 3:
+                break
+            continue
+        if in_fence and len(usage) < 3:
+            usage.append(ln)
+    if usage:
+        return f"- `{name}` — " + " ".join(usage)
+    return f"- `{name}` — " + lines[0][:160]
+
+
 def _assemble_prompt(tool_names: set, disabled_tools: set = None, compact: bool = False) -> str:
     """Build the system prompt with only the specified tools included."""
     disabled = disabled_tools or set()
     included = tool_names - disabled
 
     if compact:
-        tool_list = ", ".join(sorted(included)) if included else "none"
+        tool_lines = []
+        for name, _default_section in TOOL_SECTIONS.items():
+            if name in included:
+                tool_lines.append(_compact_tool_line(name, _section_text(name, _default_section)))
         parts = [
-            "You are an AI assistant with tool access.",
-            f"Available tools: {tool_list}.",
-            _API_AGENT_RULES,
+            _AGENT_PREAMBLE,
+            "## Available tools\n" + ("\n".join(tool_lines) if tool_lines else "none"),
+            _AGENT_RULES,
         ]
         parts.extend(_domain_rules_for_tools(included))
         return "\n\n".join(parts)
@@ -617,11 +644,6 @@ _API_HOSTS = frozenset([
     "api.perplexity.ai", "api.x.ai",
     "ollama.com", "api.venice.ai", "api.kimi.com",
     "api.githubcopilot.com",
-    # Local OpenAI-compatible endpoints (llama.cpp, vLLM, LM Studio, etc.).
-    # Without these, `_is_api_model` falls back to keyword sniffing on the
-    # model name, so well-behaved local servers don't get native tool
-    # schemas and the agent silently degrades to fenced-block parsing.
-    "localhost", "127.0.0.1", "host.docker.internal",
 ])
 _MCP_KEYWORDS = frozenset(["mcp", "browse", "browser", "website", "calendar", "event", "email",
                            "gmail", "screenshot", "navigate", "click", "miniflux", "rss", "feed"])
@@ -649,6 +671,28 @@ def _is_ollama_openai_compat_url(endpoint_url: str) -> bool:
     return parsed.port == 11434 and (path == "/v1" or path.startswith("/v1/"))
 
 
+def _is_local_openai_compat_url(endpoint_url: str) -> bool:
+    try:
+        parsed = urlparse(endpoint_url or "")
+    except Exception:
+        return False
+    host = (parsed.hostname or "").lower()
+    path = (parsed.path or "").rstrip("/")
+    if not (path == "/v1" or path.startswith("/v1/")):
+        return False
+    if host in {"localhost", "127.0.0.1", "0.0.0.0", "host.docker.internal"}:
+        return True
+    if host.startswith("192.168.") or host.startswith("10."):
+        return True
+    if host.startswith("172."):
+        try:
+            second = int(host.split(".")[1])
+            return 16 <= second <= 31
+        except Exception:
+            return False
+    return False
+
+
 def _endpoint_lookup_keys(endpoint_url: str) -> List[str]:
     """Candidate ModelEndpoint.base_url keys for a runtime chat URL."""
     raw = (endpoint_url or "").strip()
@@ -712,6 +756,17 @@ def _extract_last_user_message(messages: List[Dict]) -> str:
 
 
 _LOW_SIGNAL_RE = re.compile(r"^[\W_]*$", re.UNICODE)
+_CASUAL_OPENING_RE = re.compile(
+    r"^\s*(?:h+i+|hey+|hello+|yo+|sup+|what'?s up|wass?up|hiya|howdy|"
+    r"lol|lmao|haha+|hehe+|thanks?|thank you|ty|idk|dunno|meh|bruh|bro)\b(?P<tail>.*)$",
+    re.IGNORECASE,
+)
+_CASUAL_BLOCKLIST_RE = re.compile(
+    r"\b(?:cookbook|serve|serving|launch|start|vllm|sglang|llama\.?cpp|ollama|"
+    r"download|model|email|document|doc|note|calendar|task|search|web|research|"
+    r"file|folder|repo|git|settings?|endpoint|api|token|mcp)\b",
+    re.IGNORECASE,
+)
 _EXPLICIT_CONTINUATION_RE = re.compile(
     r"^\s*(?:"
     r"yes|y|yeah|yep|ok|okay|sure|do it|go ahead|continue|carry on|"
@@ -721,6 +776,17 @@ _EXPLICIT_CONTINUATION_RE = re.compile(
     r")\s*[.!?]*\s*$",
     re.IGNORECASE,
 )
+_RETRY_CONTINUATION_RE = re.compile(
+    r"\b(?:try again|retry|again|rerun|re-run|run it again|launch it again|"
+    r"start it again|failed|fails?|died|crashed|broke|insta|instantly)\b",
+    re.IGNORECASE,
+)
+_COOKBOOK_CONTEXT_RE = re.compile(
+    r"\b(?:cookbook|serve|serving|served|launch|start|preset|vllm|sglang|"
+    r"llama\.?cpp|ollama|download|cached models?|model servers?|running models?|"
+    r"gpu box|ajax|qwen|gemma|llama|mistral|minimax)\b",
+    re.IGNORECASE,
+)
 
 
 def _is_explicit_continuation(text: str) -> bool:
@@ -728,6 +794,37 @@ def _is_explicit_continuation(text: str) -> bool:
     return bool(_EXPLICIT_CONTINUATION_RE.match(str(text or "").strip()))
 
 
+def _is_casual_low_signal(text: str) -> bool:
+    """True for short greetings/slang that should not inherit stale context."""
+    s = str(text or "").strip()
+    m = _CASUAL_OPENING_RE.match(s)
+    if not m:
+        return False
+    tail = m.group("tail") or ""
+    if _CASUAL_BLOCKLIST_RE.search(tail):
+        return False
+    # Allow a short vocative/address after the opener without hardcoding the
+    # address term itself: "hey man", "yo dude", "sup <name>". Longer tails are
+    # more likely to be an actual request and should get normal context/tooling.
+    tail_words = re.findall(r"[A-Za-z0-9_'-]+", tail)
+    return len(tail_words) <= 2
+
+
+def _is_contextual_retry_continuation(messages: List[Dict], text: str) -> bool:
+    """Treat "try again / it failed" as a continuation only for active tool work.
+
+    These follow-ups are common after Cookbook launches: the latest user turn
+    says only "try again it failed", while the actionable model/host/command
+    details live one or two turns back. Keep this intentionally narrow so
+    ordinary chat does not inherit stale Cookbook context.
+    """
+    latest = str(text or "").strip()
+    if not latest or not _RETRY_CONTINUATION_RE.search(latest):
+        return False
+    recent = _recent_context_for_retrieval(messages, max_user=5, max_chars=1200)
+    return bool(_COOKBOOK_CONTEXT_RE.search(recent))
+
+
 def _assistant_requested_followup(messages: List[Dict]) -> bool:
     """True when the previous assistant turn asked for missing task details.
 
@@ -769,11 +866,12 @@ def _classify_agent_request(messages: List[Dict], last_user: str) -> Dict[str, o
     which domain rule packs get appended to the system prompt.
     """
     text = str(last_user or "").strip()
-    continuation = _is_explicit_continuation(text) or _assistant_requested_followup(messages)
+    retry_continuation = _is_contextual_retry_continuation(messages, text)
+    continuation = _is_explicit_continuation(text) or _assistant_requested_followup(messages) or retry_continuation
     retrieval_query = _recent_context_for_retrieval(messages) if continuation else text
     q = retrieval_query.lower()
 
-    if not text or bool(_LOW_SIGNAL_RE.match(text)):
+    if not text or bool(_LOW_SIGNAL_RE.match(text)) or _is_casual_low_signal(text):
         return {
             "low_signal": True,
             "continuation": False,
@@ -886,6 +984,7 @@ def _build_system_prompt(
     compact: bool = False,
     owner: Optional[str] = None,
     suppress_local_context: bool = False,
+    suppress_skills: bool = False,
     active_email: Optional[Dict[str, str]] = None,
 ) -> List[Dict]:
     """Build agent system prompt, inject MCP/document context, merge consecutive system msgs."""
@@ -903,7 +1002,7 @@ def _build_system_prompt(
         _ov_sig = _hl.sha256(_json.dumps(get_builtin_overrides() or {}, sort_keys=True).encode()).hexdigest()
     except Exception:
         _ov_sig = ""
-    cache_key = (frozenset(disabled_tools or []), bool(mcp_mgr), needs_admin, _rt_key, compact, _ov_sig, owner, suppress_local_context)
+    cache_key = (frozenset(disabled_tools or []), bool(mcp_mgr), needs_admin, _rt_key, compact, _ov_sig, owner, suppress_local_context, suppress_skills)
     if _cached_base_prompt and _cached_base_prompt_key == cache_key and not active_document:
         agent_prompt = _cached_base_prompt
         # Skill index is user-editable (name + description), so it must never
@@ -913,6 +1012,7 @@ def _build_system_prompt(
             disabled_tools, mcp_mgr, needs_admin, relevant_tools,
             mcp_disabled_map=mcp_disabled_map, compact=compact, owner=owner,
             suppress_local_context=suppress_local_context,
+            suppress_skills=suppress_skills,
         )
     else:
         agent_prompt, _skill_index_block = _build_base_prompt(
@@ -924,6 +1024,7 @@ def _build_system_prompt(
             compact=compact,
             owner=owner,
             suppress_local_context=suppress_local_context,
+            suppress_skills=suppress_skills,
         )
         if not active_document:
             _cached_base_prompt = agent_prompt
@@ -1207,7 +1308,7 @@ def _build_system_prompt(
     # few. If the teacher wrote a procedure for "open my X chat" last
     # time the student failed, this is where the student finds it
     # before deciding which tool to call.
-    if not suppress_local_context:
+    if not suppress_local_context and not suppress_skills:
         try:
             last_user = _extract_last_user_message(messages)
             # Respect the user's skills-enabled toggle (mirrors memory_enabled).
@@ -1374,6 +1475,7 @@ def _build_base_prompt(
     compact: bool = False,
     owner: Optional[str] = None,
     suppress_local_context: bool = False,
+    suppress_skills: bool = False,
 ):
     """Build the agent prompt with only relevant tools included.
 
@@ -1426,7 +1528,7 @@ def _build_base_prompt(
     # The caller wraps it in untrusted_context_message and ships it as a
     # user-role message — same treatment as the matched-skills block.
     skill_index_block = ""
-    if not suppress_local_context:
+    if not suppress_local_context and not suppress_skills:
         try:
             from services.memory.skills import SkillsManager
             from src.constants import DATA_DIR
@@ -1851,6 +1953,7 @@ async def stream_agent_loop(
     approved_plan: Optional[str] = None,
     tool_policy: Optional[ToolPolicy] = None,
     workspace: Optional[str] = None,
+    forced_tools: Optional[Set[str]] = None,
     _is_teacher_run: bool = False,
 ) -> AsyncGenerator[str, None]:
     """Streaming agent loop generator.
@@ -1890,6 +1993,20 @@ async def stream_agent_loop(
     _needs_admin = _detect_admin_intent(messages)
     _last_user = _extract_last_user_message(messages)
     _intent = _classify_agent_request(messages, _last_user)
+    _low_signal_turn = bool(_intent.get("low_signal"))
+    _casual_low_signal_turn = _is_casual_low_signal(_last_user)
+    _direct_low_signal = (
+        _low_signal_turn
+        and not bool(_intent.get("continuation"))
+        and not plan_mode
+        and not approved_plan
+        and not guide_only
+        and (_casual_low_signal_turn or active_document is None)
+        and (_casual_low_signal_turn or not active_email)
+        and (_casual_low_signal_turn or not workspace)
+        and not forced_tools
+        and not relevant_tools
+    )
     # Tool retrieval uses the latest message by default. It may inherit recent
     # user turns only for explicit continuations ("yes", "do it", "1").
     _retrieval_query = str(_intent.get("retrieval_query") or _last_user)
@@ -1897,11 +2014,86 @@ async def stream_agent_loop(
         "[agent-intent] latest=%r continuation=%s low_signal=%s domains=%s retrieval_query=%r",
         _last_user[:120],
         bool(_intent.get("continuation")),
-        bool(_intent.get("low_signal")),
+        _low_signal_turn,
         sorted(_intent.get("domains") or []),
         _retrieval_query[:200],
     )
     _mcp_disabled_map = _load_mcp_disabled_map() if mcp_mgr else {}
+    if _direct_low_signal:
+        logger.info("[agent] direct low-signal reply path for latest=%r", _last_user[:80])
+        direct_messages = [{"role": "user", "content": _last_user}]
+        direct_response = ""
+        direct_start = time.time()
+        direct_actual_model = model
+        real_input_tokens = 0
+        real_output_tokens = 0
+        try:
+            async for chunk in stream_llm_with_fallback(
+                [(endpoint_url, model, headers)] + list(fallbacks or []),
+                direct_messages,
+                temperature=temperature,
+                max_tokens=min(max_tokens or 128, 128),
+                prompt_type=None,
+                tools=None,
+                timeout=int(get_setting("agent_stream_timeout_seconds", 300) or 300),
+                session_id=session_id,
+            ):
+                if chunk.startswith("data: ") and not chunk.startswith("data: [DONE]"):
+                    try:
+                        data = json.loads(chunk[6:])
+                    except json.JSONDecodeError:
+                        yield chunk
+                        continue
+                    if data.get("type") == "usage":
+                        usage = data.get("data", {}) or {}
+                        direct_actual_model = usage.get("model") or direct_actual_model
+                        real_input_tokens += usage.get("input_tokens", 0) or 0
+                        real_output_tokens += usage.get("output_tokens", 0) or 0
+                        continue
+                    if data.get("type") == "model_actual":
+                        direct_actual_model = data.get("model") or direct_actual_model
+                        data["requested_model"] = model
+                        yield f"data: {json.dumps(data)}\n\n"
+                        continue
+                    if data.get("type") == "fallback":
+                        direct_actual_model = data.get("answered_by") or direct_actual_model
+                        yield chunk
+                        continue
+                    if "delta" in data:
+                        if not data.get("thinking"):
+                            direct_response += data.get("delta", "")
+                        yield chunk
+                        continue
+                    yield chunk
+                elif chunk.startswith("event: "):
+                    yield chunk
+        except Exception as _direct_err:
+            logger.warning("[agent] direct low-signal path failed: %s", _direct_err)
+            fallback = "Hey."
+            direct_response += fallback
+            yield f"data: {json.dumps({'delta': fallback})}\n\n"
+
+        if not direct_response.strip():
+            fallback = "Hey."
+            direct_response = fallback
+            yield f"data: {json.dumps({'delta': fallback})}\n\n"
+
+        duration = time.time() - direct_start
+        metrics = {
+            "model": direct_actual_model,
+            "requested_model": model,
+            "input_tokens": real_input_tokens or estimate_tokens(direct_messages),
+            "output_tokens": real_output_tokens or max(len(direct_response) // 4, 1),
+            "total_time": round(duration, 2),
+            "response_time": round(duration, 2),
+            "agent_rounds": 0,
+            "tool_calls": 0,
+            "direct_low_signal": True,
+        }
+        yield f"data: {json.dumps({'type': 'metrics', 'data': metrics})}\n\n"
+        yield "data: [DONE]\n\n"
+        return
+
     if plan_mode and mcp_mgr:
         # Allow read-only MCP tools to investigate, block write/unknown ones:
         # hide them from the schemas AND reject them at runtime by qualified name.
@@ -1913,11 +2105,11 @@ async def stream_agent_loop(
 
     # RAG-based tool selection: retrieve relevant tools for this query.
     # If caller provided a pre-computed set (e.g. task_scheduler), use that.
-    _relevant_tools = set() if guide_only else relevant_tools
+    _relevant_tools = relevant_tools
     _t1 = time.time()
     if _relevant_tools:
         logger.info(f"[tool-rag] Using caller-provided relevant_tools ({len(_relevant_tools)} tools)")
-    if not guide_only and not _relevant_tools and bool(_intent.get("low_signal")):
+    if not guide_only and not _relevant_tools and _low_signal_turn:
         from src.tool_index import ALWAYS_AVAILABLE
         if workspace:
             # An active workspace IS the file-work signal: a vague "look at the
@@ -2008,6 +2200,15 @@ async def stream_agent_loop(
     if _relevant_tools is not None and active_document is not None:
         _relevant_tools.update({"edit_document", "update_document", "suggest_document"})
 
+    # Per-request UI toggles are stronger than retrieval. If the user turns on
+    # Search, the model must see the search tools even when the latest text is a
+    # typo or otherwise low-signal for tool RAG.
+    if not guide_only and forced_tools:
+        if _relevant_tools is None:
+            from src.tool_index import ALWAYS_AVAILABLE
+            _relevant_tools = set(ALWAYS_AVAILABLE)
+        _relevant_tools.update(t for t in forced_tools if t not in disabled_tools)
+
     # The skill index injected by _build_system_prompt tells the model to
     # call `manage_skills action=view`, and Jaccard-matched skills are pasted
     # into the prompt as procedures to follow — but neither path goes through
@@ -2015,7 +2216,7 @@ async def stream_agent_loop(
     # (grep, read_file, ...) that aren't in its schema list. Keep the schemas
     # in lockstep: manage_skills is callable whenever any skill is indexed,
     # and a matched skill's declared requires_toolsets ride along with it.
-    if not guide_only and _relevant_tools is not None:
+    if not guide_only and _relevant_tools is not None and not _low_signal_turn:
         try:
             from services.memory.skills import SkillsManager
             from src.constants import DATA_DIR
@@ -2080,7 +2281,7 @@ async def stream_agent_loop(
     _model_supports_tools = any(kw in _model_lc for kw in (
         "gpt-4", "gpt-5", "gpt-o", "claude", "gemini", "gemma",
         "qwen3", "qwen2.5", "mixtral", "mistral", "llama-3.1", "llama-3.2",
-        "llama-3.3", "llama-4",
+        "llama-3.3", "llama-4", "llama3.1", "llama3.2", "llama3.3", "llama4",
         # Local-served models that follow OpenAI-style function calling
         # via vLLM's `--enable-auto-tool-choice`. Belt-and-suspenders
         # with the per-endpoint flag above.
@@ -2122,13 +2323,15 @@ async def stream_agent_loop(
         _is_api_model = False
     else:
         _is_api_model = any(h in endpoint_url for h in _API_HOSTS) or _model_supports_tools
+    _compact_agent_prompt = _is_api_model or _is_ollama_native or _ollama_openai_compat
     messages, mcp_schemas = _build_system_prompt(
         messages, model, active_document, mcp_mgr, disabled_tools,
         needs_admin=_needs_admin, relevant_tools=_relevant_tools,
         mcp_disabled_map=_mcp_disabled_map,
-        compact=_is_api_model,
+        compact=_compact_agent_prompt,
         owner=owner,
         suppress_local_context=guide_only,
+        suppress_skills=_low_signal_turn,
         active_email=active_email,
     )
     if plan_mode and not guide_only:
@@ -2214,6 +2417,14 @@ async def stream_agent_loop(
     # Strip internal metadata keys before sending to the LLM API
     messages = [{k: v for k, v in msg.items() if k != "_protected"} for msg in messages]
 
+    agent_prompt_tokens = estimate_tokens(messages)
+    logger.info(
+        "[agent-timing] prep_done model=%s prompt_tokens=%s context_length=%s prep=%s",
+        model,
+        agent_prompt_tokens,
+        context_length,
+        {k: round(v, 3) for k, v in prep_timings.items()},
+    )
     yield f"data: {json.dumps({'type': 'agent_prep', 'data': {k: round(v, 3) for k, v in prep_timings.items()}})}\n\n"
 
     full_response = ""
@@ -2358,6 +2569,19 @@ async def stream_agent_loop(
         # complementary cap for the rare stream that trickles bytes forever and
         # so never trips the inactivity timeout. Generous — only catches runaway.
         _round_deadline = time.time() + max(agent_stream_timeout * 4, 1200)
+        _round_start = time.time()
+        _round_first_event_logged = False
+        _round_first_token_logged = False
+        logger.info(
+            "[agent-timing] round_start round=%s model=%s endpoint=%s prompt_tokens=%s tools=%s native_tools=%s timeout=%s",
+            round_num,
+            model,
+            endpoint_url,
+            estimate_tokens(messages),
+            len(_tool_names_sent),
+            bool(all_tool_schemas),
+            agent_stream_timeout,
+        )
         async for chunk in stream_llm_with_fallback(
             _candidates,
             messages,
@@ -2368,11 +2592,30 @@ async def stream_agent_loop(
             timeout=agent_stream_timeout,
             session_id=session_id,
         ):
+            if not _round_first_event_logged:
+                _round_first_event_logged = True
+                logger.info(
+                    "[agent-timing] first_event round=%s elapsed=%.3fs kind=%s",
+                    round_num,
+                    time.time() - _round_start,
+                    "error" if chunk.startswith("event: error") else "data",
+                )
             if time.time() > _round_deadline:
-                logger.warning(f"[agent] round {round_num} stream exceeded wall-clock deadline; cutting off")
+                logger.warning(
+                    "[agent-timing] round_deadline round=%s elapsed=%.3fs deadline_s=%s",
+                    round_num,
+                    time.time() - _round_start,
+                    max(agent_stream_timeout * 4, 1200),
+                )
                 break
             # Forward error events from stream_llm to the frontend
             if chunk.startswith("event: error"):
+                logger.warning(
+                    "[agent-timing] stream_error round=%s elapsed=%.3fs chunk=%r",
+                    round_num,
+                    time.time() - _round_start,
+                    chunk[:500],
+                )
                 yield chunk
                 continue
             if chunk.startswith("data: ") and not chunk.startswith("data: [DONE]"):
@@ -2452,6 +2695,15 @@ async def stream_agent_loop(
                         if not first_token_received:
                             time_to_first_token = time.time() - total_start
                             first_token_received = True
+                        if not _round_first_token_logged:
+                            _round_first_token_logged = True
+                            logger.info(
+                                "[agent-timing] first_visible_token round=%s elapsed=%.3fs total_elapsed=%.3fs thinking=%s",
+                                round_num,
+                                time.time() - _round_start,
+                                time.time() - total_start,
+                                bool(data.get("thinking")),
+                            )
                         # Keep reasoning deltas in a separate accumulator so
                         # we can echo them back via `reasoning_content` on the
                         # next request (DeepSeek requires this; harmless for
@@ -2521,7 +2773,21 @@ async def stream_agent_loop(
                 yield chunk
             # Intercept [DONE] — don't forward until all rounds finish
 
-        tool_blocks, used_native = _resolve_tool_blocks(round_response, native_tool_calls, round_num, is_api_model=_is_api_model)
+        logger.info(
+            "[agent-timing] round_stream_done round=%s elapsed=%.3fs text_chars=%s tool_calls=%s first_event=%s first_token=%s",
+            round_num,
+            time.time() - _round_start,
+            len(round_response),
+            len(native_tool_calls),
+            _round_first_event_logged,
+            _round_first_token_logged,
+        )
+        tool_blocks, used_native = _resolve_tool_blocks(
+            round_response,
+            native_tool_calls,
+            round_num,
+            is_api_model=(_is_api_model and not guide_only),
+        )
 
         # Force-answer round: we told the model to STOP calling tools and
         # answer. If it ignored that and emitted a (possibly DSML) tool
@@ -2605,7 +2871,7 @@ async def stream_agent_loop(
         # model with no real native_tool_calls) must not be stripped from the
         # persisted text either — otherwise it streams once and then disappears
         # on reload (#3222 follow-up).
-        cleaned_round = strip_tool_blocks(round_response, skip_fenced=(_is_api_model and not used_native)).strip()
+        cleaned_round = strip_tool_blocks(round_response, skip_fenced=(_is_api_model and not used_native and not guide_only)).strip()
         round_texts.append(cleaned_round)
 
         if not tool_blocks:
@@ -2677,6 +2943,15 @@ async def stream_agent_loop(
                 _intent_nudge_count += 1
                 _matched_phrase = _intent_match.group(0).strip()
                 logger.info(f"[agent] intent-without-action nudge #{_intent_nudge_count} on round {round_num}: {_matched_phrase!r}")
+                _lower_phrase = _matched_phrase.lower()
+                _cookbook_log_hint = ""
+                if any(_word in _lower_phrase for _word in ("log", "logs", "output", "tail", "status")):
+                    _cookbook_log_hint = (
+                        " If this is about a Cookbook/model serve, the concrete calls are: "
+                        "`list_served_models` first, then `tail_serve_output` with the "
+                        "session_id from the serve/list result. Never answer with "
+                        "\"check logs\" when those tools are available."
+                    )
                 messages.append({
                     "role": "system",
                     "content": (
@@ -2685,6 +2960,7 @@ async def stream_agent_loop(
                         "see you announced the action but didn't run it, which "
                         "is the most frustrating thing you can do. "
                         "DO IT NOW: emit the actual function call this turn. "
+                        f"{_cookbook_log_hint}"
                         "If you decided not to do it after all, say so plainly in "
                         "one sentence instead of restating the plan."
                     ),

src/agent_runs.py

@@ -174,8 +174,20 @@ async def subscribe(session_id: str) -> AsyncGenerator[str, None]:
             next_seq += 1
         if run.status != "running":
             return
+        heartbeat_idx = 0
         while True:
-            seq, ev = await q.get()
+            try:
+                seq, ev = await asyncio.wait_for(q.get(), timeout=10.0)
+            except asyncio.TimeoutError:
+                # Keep slow local models/proxies alive while they prefill before
+                # the first token. SSE comments are ignored by the UI but reset
+                # browser/proxy idle timers, which prevents "empty response"
+                # disconnects on llama.cpp first-token latencies of 30s+.
+                if run.status == "running":
+                    heartbeat_idx += 1
+                    yield f": heartbeat {heartbeat_idx}\n\n"
+                    continue
+                seq, ev = (None, None)
             if seq is None:            # end sentinel
                 while next_seq < len(run.buffer):   # flush any tail the sentinel raced
                     yield run.buffer[next_seq]

src/agent_tools/web_tools.py

@@ -7,6 +7,7 @@ from src.constants import MAX_OUTPUT_CHARS
 class WebSearchTool:
     async def execute(self, content: str, ctx: dict) -> dict:
         from src.search import comprehensive_web_search
+        progress_cb = ctx.get("progress_cb") if isinstance(ctx, dict) else None
         raw = content.strip()
         query = raw
         time_filter = None
@@ -37,18 +38,39 @@ class WebSearchTool:
             elif " news" in q_lc or q_lc.startswith("news ") or q_lc.endswith(" news"):
                 time_filter = "week"
         loop = asyncio.get_running_loop()
-        text, sources = await asyncio.wait_for(
-            loop.run_in_executor(
-                None,
-                lambda: comprehensive_web_search(
-                    query,
-                    max_pages=max_pages,
-                    time_filter=time_filter,
-                    return_sources=True,
+        if progress_cb:
+            await progress_cb({
+                "elapsed_s": 0,
+                "tail": f"Searching web for: {query[:160]}",
+            })
+        try:
+            text, sources = await asyncio.wait_for(
+                loop.run_in_executor(
+                    None,
+                    lambda: comprehensive_web_search(
+                        query,
+                        max_pages=max_pages,
+                        time_filter=time_filter,
+                        return_sources=True,
+                    ),
                 ),
-            ),
-            timeout=30,
-        )
+                timeout=30,
+            )
+        except asyncio.TimeoutError:
+            return {
+                "error": f"web_search timed out after 30s: {query[:200]}",
+                "exit_code": 1,
+            }
+        except Exception as e:
+            return {
+                "error": f"web_search failed: {type(e).__name__}: {str(e) or 'no details'}",
+                "exit_code": 1,
+            }
+        if progress_cb:
+            await progress_cb({
+                "elapsed_s": 30,
+                "tail": "Search completed; preparing sources.",
+            })
         output = text[:MAX_OUTPUT_CHARS] if len(text) > MAX_OUTPUT_CHARS else text
         if sources:
             output += "\n\n<!-- SOURCES:" + json.dumps(sources) + " -->"

src/builtin_actions.py

@@ -76,8 +76,7 @@ async def action_consolidate_memory(owner: str, **kwargs) -> Tuple[str, bool]:
         import json
         import re
         from src.constants import DATA_DIR
-        from src.endpoint_resolver import resolve_endpoint
-        from src.llm_core import llm_call_async
+        from src.llm_core import llm_call_async_with_fallback
         from src.memory import MemoryManager
 
         manager = MemoryManager(DATA_DIR)
@@ -116,10 +115,9 @@ async def action_consolidate_memory(owner: str, **kwargs) -> Tuple[str, bool]:
             if len(group_memories) < 2:
                 return False
 
-            url, model, headers = resolve_endpoint("utility", owner=group_owner or None)
-            if not url or not model:
-                url, model, headers = resolve_endpoint("default", owner=group_owner or None)
-            if not url or not model:
+            from src.task_endpoint import resolve_task_candidates
+            candidates = resolve_task_candidates(owner=group_owner or None)
+            if not candidates:
                 return False
 
             try:
@@ -147,13 +145,11 @@ async def action_consolidate_memory(owner: str, **kwargs) -> Tuple[str, bool]:
                     "\"drop\":[{\"id\":\"existing id\",\"reason\":\"short reason\"}]}\n\n"
                     f"MEMORIES:\n{json.dumps(items, ensure_ascii=False)}"
                 )
-                raw = await llm_call_async(
-                    url=url,
-                    model=model,
+                raw = await llm_call_async_with_fallback(
+                    candidates,
                     messages=[{"role": "user", "content": prompt}],
                     temperature=0.0,
                     max_tokens=4096,
-                    headers=headers,
                     timeout=120,
                 )
                 from src.text_helpers import strip_think
@@ -604,8 +600,7 @@ async def action_classify_events(owner: str, **kwargs) -> Tuple[str, bool]:
     try:
         from datetime import timedelta
         from core.database import SessionLocal, CalendarEvent
-        from src.endpoint_resolver import resolve_endpoint
-        from src.llm_core import llm_call_async
+        from src.llm_core import llm_call_async_with_fallback
         import re as _re, json as _json
 
         db = SessionLocal()
@@ -620,10 +615,9 @@ async def action_classify_events(owner: str, **kwargs) -> Tuple[str, bool]:
             if not events:
                 return "No upcoming events to classify", True
 
-            llm_url, llm_model, llm_headers = resolve_endpoint("utility", owner=owner)
-            if not llm_url:
-                llm_url, llm_model, llm_headers = resolve_endpoint("default", owner=owner)
-            llm_available = bool(llm_url and llm_model)
+            from src.task_endpoint import resolve_task_candidates
+            llm_candidates = resolve_task_candidates(owner=owner)
+            llm_available = bool(llm_candidates)
 
             # Pull user memories so the LLM has personal context (relationships,
             # job, hobbies). Helps it know e.g. "<name> is your spouse" so their
@@ -699,11 +693,11 @@ async def action_classify_events(owner: str, **kwargs) -> Tuple[str, bool]:
                     f"EVENTS: {_json.dumps(items)}"
                 )
                 try:
-                    raw = await llm_call_async(
-                        url=llm_url, model=llm_model,
+                    raw = await llm_call_async_with_fallback(
+                        llm_candidates,
                         messages=[{"role": "user", "content": prompt}],
                         temperature=0.1, max_tokens=16384,
-                        headers=llm_headers, timeout=180,
+                        timeout=180,
                     )
                     from src.text_helpers import strip_think as _st
                     raw = _st(raw or "", prose=False, prompt_echo=False)
@@ -810,8 +804,7 @@ async def action_learn_sender_signatures(owner: str, **kwargs) -> Tuple[str, boo
         import asyncio as _aio
         from datetime import datetime as _dt, timedelta as _td
         from routes.email_helpers import _email_cache_owner_clause, _imap_connect, SCHEDULED_DB
-        from src.endpoint_resolver import resolve_endpoint
-        from src.llm_core import llm_call_async
+        from src.llm_core import llm_call_async_with_fallback
 
         # 1. Pull recent UIDs + From headers cheaply (header-only fetch).
         def _pull_headers():
@@ -891,11 +884,11 @@ async def action_learn_sender_signatures(owner: str, **kwargs) -> Tuple[str, boo
         if not eligible:
             return "All sender sigs already cached (or no eligible senders)", True
 
-        url, model, headers = resolve_endpoint("utility", owner=owner)
-        if not url or not model:
-            url, model, headers = resolve_endpoint("default", owner=owner)
-        if not url or not model:
+        from src.task_endpoint import resolve_task_candidates
+        candidates = resolve_task_candidates(owner=owner)
+        if not candidates:
             return "No LLM endpoint available", False
+        model = candidates[0][1]
 
         analyzed = 0
         no_sig = 0
@@ -949,11 +942,11 @@ async def action_learn_sender_signatures(owner: str, **kwargs) -> Tuple[str, boo
             )
 
             try:
-                raw = await llm_call_async(
-                    url=url, model=model,
+                raw = await llm_call_async_with_fallback(
+                    candidates,
                     messages=[{"role": "user", "content": prompt}],
                     temperature=0.0, max_tokens=600,
-                    headers=headers, timeout=60,
+                    timeout=60,
                 )
                 from src.text_helpers import strip_think as _st
                 sig = _st(raw or "", prose=False, prompt_echo=False).strip()
@@ -1137,7 +1130,6 @@ async def action_test_skills(owner: str, **kwargs) -> Tuple[str, bool]:
         from services.memory.skills import SkillsManager
         from src.constants import DATA_DIR
         from routes.skills_routes import _run_skill_test_once, _skill_test_task
-        from src.endpoint_resolver import resolve_endpoint
 
         # #3 SCOPE GUARD: refuse to run on a None/empty owner — otherwise
         # `sm.load(owner=None)` returns every user's skills and we'd cross-
@@ -1152,27 +1144,40 @@ async def action_test_skills(owner: str, **kwargs) -> Tuple[str, bool]:
         if not names:
             raise TaskNoop("no skills to test")
 
-        url, model, headers = resolve_endpoint("default", owner=owner)
-        if not url or not model:
+        from src.task_endpoint import resolve_task_candidates
+        candidates = resolve_task_candidates(owner=owner)
+        if not candidates:
             return "No Default/Utility model configured — set one in Settings.", False
 
         # #2 NO SILENT MODEL SWAP: if the configured model isn't served by the
         # endpoint, try a basename match — but fail loudly instead of grabbing
         # `avail[0]` which could be an embedding-only model and produce 36
         # garbage transcripts → 36 'unknown' verdicts with no hint why.
+        url, model, headers = candidates[0]
         try:
             from src.llm_core import list_model_ids
-            avail = list_model_ids(url, headers=headers)
-            if avail and model not in avail:
-                import os as _os
-                base = _os.path.basename((model or "").rstrip("/"))
-                m = next((a for a in avail if _os.path.basename(a.rstrip("/")) == base), None)
-                if m:
-                    model = m
-                else:
-                    return (f"Default model '{model}' not served by endpoint {url}. "
-                            f"Available: {', '.join(avail[:8])}{'…' if len(avail) > 8 else ''}. "
-                            "Set a valid Default model in Settings."), False
+            import os as _os
+
+            selected = None
+            mismatch_notes = []
+            for cand_url, cand_model, cand_headers in candidates:
+                avail = list_model_ids(cand_url, headers=cand_headers)
+                if not avail or cand_model in avail:
+                    selected = (cand_url, cand_model, cand_headers)
+                    break
+                base = _os.path.basename((cand_model or "").rstrip("/"))
+                matched = next((a for a in avail if _os.path.basename(a.rstrip("/")) == base), None)
+                if matched:
+                    selected = (cand_url, matched, cand_headers)
+                    break
+                mismatch_notes.append(
+                    f"{cand_model} not served by {cand_url}; available: "
+                    f"{', '.join(avail[:8])}{'...' if len(avail) > 8 else ''}"
+                )
+            if selected:
+                url, model, headers = selected
+            elif mismatch_notes:
+                return "No configured task fallback model is served. " + " | ".join(mismatch_notes[:3]), False
         except Exception as _e:
             logger.warning(f"test_skills model resolve check failed (continuing): {_e}")
 
@@ -1483,7 +1488,6 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
         from pathlib import Path as _P
         from core.database import SessionLocal as _SL, EmailAccount as _EA
         from routes.email_helpers import _imap_connect, _decode_header
-        from src.endpoint_resolver import resolve_endpoint, resolve_utility_fallback_candidates
         from src.llm_core import llm_call_async_with_fallback
 
         # Per-owner state file so multi-user runs don't clobber each other's
@@ -1505,12 +1509,10 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
 
         # ── 1. Resolve LLM candidates (utility primary + utility fallbacks; fall
         # through to default chat as a last resort).
-        url, model, headers = resolve_endpoint("utility", owner=owner)
-        if not url or not model:
-            url, model, headers = resolve_endpoint("default", owner=owner)
-        if not url or not model:
+        from src.task_endpoint import resolve_task_candidates
+        candidates = resolve_task_candidates(owner=owner)
+        if not candidates:
             return "No LLM endpoint available", False
-        candidates = [(url, model, headers)] + resolve_utility_fallback_candidates(owner=owner)
 
         # ── 2. Enumerate enabled accounts. Match this task's owner AND fall
         # back to the legacy "unowned account whose imap_user / from_address

src/constants.py

@@ -4,7 +4,7 @@ import os
 
 from src.runtime_paths import get_app_root, get_default_data_dir
 
-APP_VERSION = "1.0.0"
+APP_VERSION = "1.0.1"
 
 # Base paths
 BASE_DIR = os.path.join(get_app_root(), "")

src/endpoint_resolver.py

@@ -424,6 +424,9 @@ def resolve_utility_fallback_candidates(owner: Optional[str] = None) -> list:
         settings = load_settings()
         utility_ep = (get_user_setting("utility_endpoint_id", owner or "", settings.get("utility_endpoint_id", "")) or "").strip()
         if not utility_ep:
+            utility_chain = get_user_setting("utility_model_fallbacks", owner or "", settings.get("utility_model_fallbacks") or []) or []
+            if utility_chain:
+                return _resolve_fallback_candidates("utility_model_fallbacks", owner=owner)
             return _resolve_fallback_candidates("default_model_fallbacks", owner=owner)
     except Exception:
         pass

src/llm_core.py

@@ -907,7 +907,10 @@ def _anthropic_rejects_temperature(model: str) -> bool:
     return (int(match.group(1)), int(match.group(2))) >= (4, 7)
 
 # Models that support structured thinking — may output </think> without opening tag
-_THINKING_MODEL_PATTERNS = ("qwen3", "qwq", "deepseek-r1", "deepseek-reasoner", "minimax", "m2-reap", "gemma")
+_THINKING_MODEL_PATTERNS = (
+    "qwen3", "qwq", "deepseek-r1", "deepseek-reasoner", "minimax",
+    "m2-reap", "gemma", "stepfun", "step-3", "step3",
+)
 
 def _supports_thinking(model: str) -> bool:
     """Check if model supports structured thinking output."""
@@ -2135,6 +2138,8 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
                                             yield _stream_delta_event(reasoning, thinking=True)
                                         content = delta.get("content") or ""
                                         if content:
+                                            content = re.sub(r"<mm:think(\s+[^>]*)?>", r"<think\1>", content, flags=re.IGNORECASE)
+                                            content = re.sub(r"</mm:think>", "</think>", content, flags=re.IGNORECASE)
                                             stripped = content.lstrip()
                                             # gpt-oss harmony format (<|channel|>analysis/final): route via the harmony
                                             # stream router. Sticky once the first marker appears — distinct from the

src/task_endpoint.py

@@ -1,6 +1,11 @@
-"""Shared resolver for background-task AI endpoint (auto-naming, memory, sorting)."""
+"""Shared resolver for background-task AI endpoints."""
 
-from src.endpoint_resolver import resolve_endpoint
+from src.endpoint_resolver import (
+    resolve_chat_fallback_candidates,
+    resolve_endpoint,
+    resolve_utility_fallback_candidates,
+)
+from src.llm_core import llm_call_async_with_fallback
 
 
 def resolve_task_endpoint(fallback_url=None, fallback_model=None, fallback_headers=None, owner=None):
@@ -11,3 +16,60 @@ def resolve_task_endpoint(fallback_url=None, fallback_model=None, fallback_heade
     endpoint cannot be resolved.
     """
     return resolve_endpoint("task", fallback_url, fallback_model, fallback_headers, owner=owner)
+
+
+def resolve_task_candidates(
+    fallback_url=None,
+    fallback_model=None,
+    fallback_headers=None,
+    owner=None,
+):
+    """Return ordered background-task LLM candidates.
+
+    Order:
+    1. configured Background Tasks endpoint/model, or caller fallback
+    2. Utility endpoint/model
+    3. Default endpoint/model
+    4. Utility fallback chain
+    5. Default fallback chain
+    """
+    candidates = []
+
+    def _append(url, model, headers):
+        if not url or not model:
+            return
+        key = (url, model)
+        if any((u, m) == key for u, m, _ in candidates):
+            return
+        candidates.append((url, model, headers or {}))
+
+    _append(*resolve_task_endpoint(fallback_url, fallback_model, fallback_headers, owner=owner))
+    _append(*resolve_endpoint("utility", owner=owner))
+    _append(*resolve_endpoint("default", owner=owner))
+    for url, model, headers in resolve_utility_fallback_candidates(owner=owner):
+        _append(url, model, headers)
+    for url, model, headers in resolve_chat_fallback_candidates(owner=owner):
+        _append(url, model, headers)
+
+    return candidates
+
+
+async def task_llm_call_async(
+    messages,
+    *,
+    fallback_url=None,
+    fallback_model=None,
+    fallback_headers=None,
+    owner=None,
+    **kwargs,
+):
+    """Call the shared background-task LLM candidate chain."""
+    candidates = resolve_task_candidates(
+        fallback_url=fallback_url,
+        fallback_model=fallback_model,
+        fallback_headers=fallback_headers,
+        owner=owner,
+    )
+    if not candidates:
+        raise RuntimeError("No LLM endpoint available for background task")
+    return await llm_call_async_with_fallback(candidates, messages=messages, **kwargs)

src/task_scheduler.py

@@ -886,6 +886,14 @@ class TaskScheduler:
                     owner=task.owner,
                     body=run.result if output == "notification" else None,
                 )
+            elif run.status == "error":
+                self.add_notification(
+                    task.name,
+                    "error",
+                    task_id,
+                    owner=task.owner,
+                    body=run.error or run.result,
+                )
 
             # Log result to the assistant chat so all task activity is visible.
             # Skip skipped/error rows — user shouldn't see "skipped: …" noise
@@ -1468,12 +1476,18 @@ class TaskScheduler:
             )
         except Exception as e:
             logger.warning(f"Agent loop failed for task '{task.name}', falling back to simple call: {e}")
-            from src.llm_core import llm_call_async
+            from src.task_endpoint import task_llm_call_async
             messages = [
                 {"role": "system", "content": system_prompt},
                 {"role": "user", "content": task.prompt},
             ]
-            result = await llm_call_async(url=endpoint_url, model=model, messages=messages, timeout=120)
+            result = await task_llm_call_async(
+                messages,
+                fallback_url=endpoint_url,
+                fallback_model=model,
+                owner=task.owner,
+                timeout=120,
+            )
 
         # Strip the model's chain-of-thought before saving/delivering. Task
         # output is LLM-only, so prose=True (which also removes untagged
@@ -1698,13 +1712,17 @@ class TaskScheduler:
         # Honor per-task max_steps (defense against runaway agent loops).
         # Falls back to 20 if not set — the historical default.
         _task_max_rounds = task.max_steps if task.max_steps and task.max_steps > 0 else 20
-        # Tasks are background workloads — they share the Utility model's
-        # fallback chain (Settings → Utility Model → Fallbacks). A downed
-        # primary endpoint won't silently yield `(no output)` — same recipe
-        # chat uses but with the utility list (`utility_model_fallbacks`).
+        # Tasks are background workloads: use the shared task fallback chain
+        # behind the primary endpoint so a downed primary won't silently yield
+        # `(no output)`.
         try:
-            from src.endpoint_resolver import resolve_utility_fallback_candidates
-            _task_fallbacks = resolve_utility_fallback_candidates(owner=task.owner or None)
+            from src.task_endpoint import resolve_task_candidates
+            _task_fallbacks = resolve_task_candidates(
+                fallback_url=endpoint_url,
+                fallback_model=model,
+                fallback_headers=headers,
+                owner=task.owner or None,
+            )[1:]
         except Exception:
             _task_fallbacks = []
         async for event_str in stream_agent_loop(
@@ -1741,21 +1759,22 @@ class TaskScheduler:
         # asking it to summarize what it did. Guarantees output.
         if not full_text.strip():
             try:
-                from src.llm_core import llm_call_async_with_fallback
-                from src.endpoint_resolver import resolve_utility_fallback_candidates
+                from src.task_endpoint import task_llm_call_async
                 grace_context = "You ran out of steps. "
                 if tool_results:
                     grace_context += "Here's what your tools returned:\n" + "\n".join(tool_results[-5:])
                 else:
                     grace_context += "No tool results were captured."
                 grace_context += "\n\nSummarize what you accomplished and what's still pending. Be concise."
-                _grace_candidates = [(endpoint_url, model, headers)] + resolve_utility_fallback_candidates(owner=task.owner or None)
-                full_text = await llm_call_async_with_fallback(
-                    _grace_candidates,
+                full_text = await task_llm_call_async(
                     messages=[
                         {"role": "system", "content": system_content},
                         {"role": "user", "content": grace_context},
                     ],
+                    fallback_url=endpoint_url,
+                    fallback_model=model,
+                    fallback_headers=headers,
+                    owner=task.owner or None,
                     timeout=30,
                 )
                 full_text = (full_text or "").strip()

src/tool_implementations.py

@@ -1268,8 +1268,8 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
             _ALIASES = {
                 "shell": ["bash"],
                 "terminal": ["bash"],
-                "search": ["web_search"],
-                "web": ["web_search"],
+                "search": ["web_search", "web_fetch"],
+                "web": ["web_search", "web_fetch"],
                 "browser": ["builtin_browser"],
                 "documents": ["create_document", "edit_document", "update_document", "suggest_document"],
                 "doc": ["create_document", "edit_document", "update_document", "suggest_document"],
@@ -1281,7 +1281,7 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
                 "notes": ["manage_notes"],
                 "calendar": ["manage_calendar"],
                 "email": ["mcp__email__list_emails", "mcp__email__read_email", "mcp__email__send_email"],
-                "research": ["web_search"],  # research is a per-request flag, not a tool — closest analog
+                "research": ["web_search", "web_fetch"],  # research is a per-request flag, not a tool — closest analog
             }
 
             if action == "list_tools":
@@ -2863,13 +2863,25 @@ async def do_serve_model(content: str, owner: Optional[str] = None) -> Dict:
                 endpoint_added=endpoint_added, endpoint_id=endpoint_id or "",
             )
             note = "" if registered else " (state-write failed — task may not show in UI)"
+            where = host or "local"
+            log_path = f"/tmp/odysseus-tmux/{sid}.log"
             return {
-                "output": f"Serving {repo_id} (session: {sid}){note}",
+                "output": (
+                    f"Serving {repo_id} on {where} (session: {sid}){note}\n"
+                    f"Next required check: call list_served_models. If this task is not ready, "
+                    f"call tail_serve_output with session_id={sid} and tail=400 before answering. "
+                    f"Do not tell the user to check logs; you have the log tool."
+                ),
                 "session_id": sid,
                 "task_type": "serve",
                 "phase": "running",
                 "host": host,
                 "endpoint_id": endpoint_id,
+                "log_path": log_path,
+                "next_tools": [
+                    {"name": "list_served_models", "arguments": {}},
+                    {"name": "tail_serve_output", "arguments": {"session_id": sid, "tail": 400}},
+                ],
                 "exit_code": 0,
             }
         # FastAPI HTTPException puts the message under `detail`, not `error`.
@@ -3216,8 +3228,17 @@ async def do_tail_serve_output(content: str, owner: Optional[str] = None) -> Dic
         MAX_CHARS = 8000
         if len(output_text) > MAX_CHARS:
             output_text = "…(earlier output truncated)…\n" + output_text[-MAX_CHARS:]
+        if not output_text:
+            output_text = (
+                f"No log output captured yet for {session_id} on {host_label}. "
+                "This usually means the tmux wrapper has started but the model process "
+                "has not printed anything yet. Do not stop here: call list_served_models "
+                "again to check whether it is still loading, ready, or crashed; if it is "
+                "still not ready, call tail_serve_output again with a larger tail after "
+                "the next status check."
+            )
         return {
-            "output": output_text or "(empty pane)",
+            "output": output_text,
             "session_id": session_id,
             "host": host_label,
             "tail_lines": tail,

src/tool_parsing.py

@@ -39,6 +39,10 @@ _XML_TOOL_CALL_RE = re.compile(
     r"<(?:[\w]+:)?(?:tool_call|function_call)>\s*([\s\S]*?)</(?:[\w]+:)?(?:tool_call|function_call)>",
     re.IGNORECASE,
 )
+_XML_OPEN_TOOL_CALL_RE = re.compile(
+    r"<(?:[\w]+:)?(?:tool_call|function_call)>\s*([\s\S]*)\Z",
+    re.IGNORECASE,
+)
 _XML_INVOKE_RE = re.compile(
     r'<invoke\s+name=["\'](\w+)["\']>\s*([\s\S]*?)</invoke>',
     re.IGNORECASE,
@@ -47,6 +51,21 @@ _XML_PARAM_RE = re.compile(
     r'<parameter\s+name=["\'](\w+)["\']>([\s\S]*?)</parameter>',
     re.IGNORECASE,
 )
+_XML_DIRECT_TOOL_RE = re.compile(
+    r"<\s*([A-Za-z_][\w-]*)\s*>([\s\S]*?)</\s*\1\s*>",
+    re.IGNORECASE,
+)
+
+# Pattern 3b: StepFun Step-3.x native tool-call tokens. The tokenizer defines:
+#   <｜tool▁calls▁begin｜> ... <｜tool▁calls▁end｜>
+#   <｜tool▁call▁begin｜>tool_name<｜tool▁sep｜>{...}<｜tool▁call▁end｜>
+# These can leak as text through llama.cpp/Ollama-style endpoints when the
+# engine does not return structured OpenAI tool_calls.
+_STEPFUN_CALL_BEGIN = "<｜tool▁call▁begin｜>"
+_STEPFUN_CALL_SEP = "<｜tool▁sep｜>"
+_STEPFUN_CALL_END = "<｜tool▁call▁end｜>"
+_STEPFUN_CALLS_BEGIN = "<｜tool▁calls▁begin｜>"
+_STEPFUN_CALLS_END = "<｜tool▁calls▁end｜>"
 
 # Pattern 4: <tool_code> blocks (MiniMax-M2.5 style)
 # {tool => 'tool_name', args => '<param>value</param>'}
@@ -446,6 +465,138 @@ def _parse_xml_invoke(inv_match) -> Optional[ToolBlock]:
     return function_call_to_tool_block(tool_name, json.dumps(params))
 
 
+def _parse_xml_direct_tool(tool_match) -> Optional[ToolBlock]:
+    """Parse direct XML tool tags inside <tool_call>.
+
+    Some local models emit:
+      <tool_call><web_search>query</web_search></tool_call>
+    instead of the invoke/parameter shape:
+      <tool_call><invoke name="web_search"><parameter name="query">query</parameter></invoke></tool_call>
+    Keep this as an adapter to the canonical function-call converter so aliases
+    and per-tool argument formatting stay in one place.
+    """
+    tool_name = tool_match.group(1).lower().replace("-", "_")
+    if tool_name in {"invoke", "parameter", "tool_call", "function_call"}:
+        return None
+    mapped = _TOOL_NAME_MAP.get(tool_name) or (tool_name if tool_name in TOOL_TAGS else None)
+    if not mapped:
+        return None
+    body = tool_match.group(2).strip()
+    if not body:
+        return None
+    try:
+        params = json.loads(body)
+        if not isinstance(params, dict):
+            params = {}
+    except json.JSONDecodeError:
+        if mapped == "web_search":
+            params = {"query": body}
+        elif mapped == "web_fetch":
+            params = {"url": body}
+        elif mapped == "bash":
+            params = {"command": body}
+        elif mapped == "python":
+            params = {"code": body}
+        elif mapped in ("read_file", "write_file"):
+            params = {"path": body}
+        else:
+            params = {"content": body}
+    from src.tool_schemas import function_call_to_tool_block
+    return function_call_to_tool_block(mapped, json.dumps(params))
+
+
+def _iter_stepfun_tool_calls(text: str):
+    """Yield StepFun native tool-call token bodies without regex backtracking."""
+    pos = 0
+    while True:
+        start = text.find(_STEPFUN_CALL_BEGIN, pos)
+        if start < 0:
+            return
+        name_start = start + len(_STEPFUN_CALL_BEGIN)
+        sep = text.find(_STEPFUN_CALL_SEP, name_start)
+        if sep < 0:
+            return
+        end = text.find(_STEPFUN_CALL_END, sep + len(_STEPFUN_CALL_SEP))
+        if end < 0:
+            return
+        raw_name = text[name_start:sep].strip()
+        body = text[sep + len(_STEPFUN_CALL_SEP):end].strip()
+        if raw_name and len(raw_name) <= 128:
+            yield raw_name, body
+        pos = end + len(_STEPFUN_CALL_END)
+
+
+def _strip_stepfun_tool_markup(text: str) -> str:
+    """Remove StepFun tool-call token blocks and wrappers using literal scans."""
+    out = []
+    pos = 0
+    while True:
+        start = text.find(_STEPFUN_CALL_BEGIN, pos)
+        if start < 0:
+            out.append(text[pos:])
+            break
+        end = text.find(_STEPFUN_CALL_END, start + len(_STEPFUN_CALL_BEGIN))
+        if end < 0:
+            out.append(text[pos:])
+            break
+        out.append(text[pos:start])
+        pos = end + len(_STEPFUN_CALL_END)
+    cleaned = "".join(out)
+    return cleaned.replace(_STEPFUN_CALLS_BEGIN, "").replace(_STEPFUN_CALLS_END, "")
+
+
+def _strip_bare_invoke_markup(text: str) -> str:
+    """Remove bare <invoke ...>...</invoke> blocks without regex backtracking."""
+    out = []
+    pos = 0
+    while True:
+        start = text.lower().find("<invoke", pos)
+        if start < 0:
+            out.append(text[pos:])
+            break
+        tag_end = text.find(">", start)
+        if tag_end < 0:
+            out.append(text[pos:])
+            break
+        close = text.lower().find("</invoke>", tag_end + 1)
+        if close < 0:
+            out.append(text[pos:])
+            break
+        out.append(text[pos:start])
+        pos = close + len("</invoke>")
+    return "".join(out)
+
+
+def _parse_stepfun_tool_call(tool_name: str, body: str) -> Optional[ToolBlock]:
+    """Parse StepFun native tool-call tokens into an Odysseus ToolBlock."""
+    tool_name = tool_name.lower().replace("-", "_").replace(".", "_")
+    mapped = _TOOL_NAME_MAP.get(tool_name) or (tool_name if tool_name in TOOL_TAGS else None)
+    if not mapped:
+        return None
+    body = (body or "").strip()
+    if not body:
+        return None
+    try:
+        params = json.loads(body)
+        if not isinstance(params, dict):
+            params = {}
+    except json.JSONDecodeError:
+        if mapped == "web_search":
+            params = {"query": body}
+        elif mapped == "web_fetch":
+            params = {"url": body}
+        elif mapped == "bash":
+            params = {"command": body}
+        elif mapped == "python":
+            params = {"code": body}
+        elif mapped in ("read_file", "write_file"):
+            params = {"path": body}
+        else:
+            params = {"content": body}
+    from src.tool_schemas import function_call_to_tool_block
+    return function_call_to_tool_block(mapped, json.dumps(params))
+
+
 def _parse_tool_code_block(raw: str) -> Optional[ToolBlock]:
     """Parse a <tool_code>{tool => 'name', args => '...'}</tool_code> block (MiniMax style)."""
     # Extract tool name
@@ -511,8 +662,9 @@ def parse_tool_blocks(text: str, skip_fenced: bool = False) -> List[ToolBlock]:
     2. [TOOL_CALL] ... [/TOOL_CALL] blocks (some models)
     3. XML-style <tool_call>/<invoke> blocks
     4. <tool_code> blocks (MiniMax-M2.5 style)
-    5. DeepSeek DSML markup (normalized to <invoke> first)
-    6. Non-native local model fallback: prose mentioning web_search followed by
+    5. StepFun Step-3 native <｜tool▁call▁begin｜> tokens
+    6. DeepSeek DSML markup (normalized to <invoke> first)
+    7. Non-native local model fallback: prose mentioning web_search followed by
        bare JSON args, e.g. {"query":"...", "time_filter":"week"}
 
     `skip_fenced`: when True, Pattern 1 (fenced ```bash/```python/```json code
@@ -567,12 +719,38 @@ def parse_tool_blocks(text: str, skip_fenced: bool = False) -> List[ToolBlock]:
 
     # Pattern 3: XML-style <tool_call>/<invoke> blocks
     if not blocks:
+        for tool_name, body in _iter_stepfun_tool_calls(text):
+            block = _parse_stepfun_tool_call(tool_name, body)
+            if block:
+                blocks.append(block)
+        if blocks:
+            return blocks
         # Try wrapped: <tool_call><invoke ...>...</invoke></tool_call>
         for m in _XML_TOOL_CALL_RE.finditer(text):
             for inv in _XML_INVOKE_RE.finditer(m.group(1)):
                 block = _parse_xml_invoke(inv)
                 if block:
                     blocks.append(block)
+            if not blocks:
+                for direct in _XML_DIRECT_TOOL_RE.finditer(m.group(1)):
+                    block = _parse_xml_direct_tool(direct)
+                    if block:
+                        blocks.append(block)
+        # Some local models stream an opening <tool_call> wrapper and a
+        # complete inner tool tag, but forget the closing </tool_call>.
+        if not blocks:
+            for m in _XML_OPEN_TOOL_CALL_RE.finditer(text):
+                body = m.group(1)
+                for inv in _XML_INVOKE_RE.finditer(body):
+                    block = _parse_xml_invoke(inv)
+                    if block:
+                        blocks.append(block)
+                if blocks:
+                    break
+                for direct in _XML_DIRECT_TOOL_RE.finditer(body):
+                    block = _parse_xml_direct_tool(direct)
+                    if block:
+                        blocks.append(block)
         # Try bare <invoke> without wrapper
         if not blocks:
             for inv in _XML_INVOKE_RE.finditer(text):
@@ -614,7 +792,9 @@ def strip_tool_blocks(text: str, skip_fenced: bool = False) -> str:
     text = _normalize_dsml(text)
     cleaned = text if skip_fenced else _TOOL_BLOCK_RE.sub('', text)
     cleaned = _TOOL_CALL_RE.sub('', cleaned)
+    cleaned = _strip_stepfun_tool_markup(cleaned)
     cleaned = _XML_TOOL_CALL_RE.sub('', cleaned)
+    cleaned = _XML_OPEN_TOOL_CALL_RE.sub('', cleaned)
     cleaned = _TOOL_CODE_RE.sub('', cleaned)
     if not skip_fenced:
         raw_web_json = _parse_raw_web_json_lookup(cleaned)
@@ -622,6 +802,6 @@ def strip_tool_blocks(text: str, skip_fenced: bool = False) -> str:
             _, (start, end) = raw_web_json
             cleaned = cleaned[:start] + cleaned[end:]
     # Strip bare <invoke> blocks not wrapped in <tool_call>
-    cleaned = re.sub(r'<invoke\s+name=["\'].*?</invoke>', '', cleaned, flags=re.DOTALL | re.IGNORECASE)
+    cleaned = _strip_bare_invoke_markup(cleaned)
     cleaned = re.sub(r'\n{3,}', '\n\n', cleaned)
     return cleaned.strip()

static/index.html

@@ -879,7 +879,7 @@
           <span class="grow">Library</span>
           <button type="button" class="list-item-plus-btn" id="library-new-doc-btn" title="New document">
             <svg class="list-item-plus-icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" style="width:11px;height:11px;"><line x1="12" y1="5" x2="12" y2="19"/><line x1="5" y1="12" x2="19" y2="12"/></svg>
-            <span class="list-item-plus-label">new</span>
+            <span class="list-item-plus-label">document</span>
           </button>
         </div>
         <div class="list-item" id="tool-notes-btn">
@@ -1005,7 +1005,12 @@
           <button type="button" class="model-picker-btn" id="model-picker-btn" title="Switch model"><span id="model-picker-label">Select model</span> <svg width="10" height="10" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="3" stroke-linecap="round" stroke-linejoin="round"><polyline points="6 15 12 9 18 15"/></svg></button>
           <div class="model-picker-menu hidden" id="model-picker-menu">
             <div class="model-picker-search-row">
-              <input type="text" id="model-picker-search" placeholder="Search models..." autocomplete="off" aria-label="Search models">
+              <div class="model-picker-search-wrap">
+                <input type="text" id="model-picker-search" placeholder="Search models..." autocomplete="off" aria-label="Search models">
+                <button type="button" class="model-picker-refresh-btn" id="model-picker-refresh-btn" title="Refresh model picker" aria-label="Refresh model picker">
+                  <svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"><polyline points="23 4 23 10 17 10"/><path d="M20.49 15a9 9 0 1 1-2.12-9.36L23 10"/></svg>
+                </button>
+              </div>
               <button type="button" class="model-picker-action-btn primary" id="model-picker-add-models-btn" title="Add model endpoints" aria-label="Add model endpoints">
                 <svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"><path d="M12 5v14"/><path d="M5 12h14"/></svg>
               </button>

static/js/chat.js

@@ -571,6 +571,24 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
     let timeoutId = null;
     let responseTimeoutCleared = false;
     let clearResponseTimeout = () => {};
+    let firstTokenWaitTimers = [];
+    const clearFirstTokenWaitTimers = () => {
+      firstTokenWaitTimers.forEach(t => { try { clearTimeout(t); } catch (_) {} });
+      firstTokenWaitTimers = [];
+    };
+    const scheduleFirstTokenWaitMessages = () => {
+      clearFirstTokenWaitTimers();
+      const steps = [
+        [20000, 'Still waiting for first token'],
+        [60000, 'Large local model is pre-filling context'],
+        [120000, 'Still working - no tokens yet from the model'],
+      ];
+      firstTokenWaitTimers = steps.map(([ms, text]) => setTimeout(() => {
+        if (!accumulated && spinner && spinner.element && !(currentAbort && currentAbort.signal.aborted)) {
+          spinner.updateMessage(text);
+        }
+      }, ms));
+    };
     const clearProcessingProbe = () => {
       if (processingProbeTimer) {
         clearTimeout(processingProbeTimer);
@@ -921,56 +939,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
         setTimeout(() => spinner.updateMessage('Analyzing sources'), 1500);
       } else {
         spinner.updateMessage('Processing request');
-        const endpointUrlForProbe = sessionModule.getCurrentEndpointUrl ? sessionModule.getCurrentEndpointUrl() : null;
-        if (endpointUrlForProbe && modelName) {
-          processingProbeTimer = setTimeout(async () => {
-            processingProbeTimer = null;
-            if (accumulated || !spinner || !spinner.element || (currentAbort && currentAbort.signal.aborted)) return;
-            processingProbeAbort = new AbortController();
-            try {
-              spinner.updateMessage('Checking model endpoint');
-              const status = await _probeCurrentEndpointStatus(endpointUrlForProbe, processingProbeAbort.signal);
-              if (accumulated || !spinner || !spinner.element || (currentAbort && currentAbort.signal.aborted)) return;
-              if (!status) {
-                spinner.updateMessage('Still waiting for model');
-              } else if (status.alive) {
-                const latency = status.latency_ms ? ` (${status.latency_ms}ms)` : '';
-                spinner.updateMessage(`Endpoint online${latency}; waiting for first token`);
-              } else {
-                // Probe confirms the endpoint isn't responding. Don't
-                // sit on a hung fetch — give the user 5s to read the
-                // status, then auto-abort with reason='offline' so the
-                // catch handler shows a clean "switch model" message
-                // instead of leaving the spinner spinning forever.
-                if (status.error) console.warn('Model endpoint probe failed:', status.error);
-                let _countdown = 5;
-                spinner.updateMessage(`Endpoint offline — cancelling in ${_countdown}s`);
-                const _tick = setInterval(() => {
-                  _countdown--;
-                  if (!spinner || !spinner.element || (currentAbort && currentAbort.signal.aborted) || accumulated) {
-                    clearInterval(_tick);
-                    return;
-                  }
-                  if (_countdown > 0) {
-                    spinner.updateMessage(`Endpoint offline — cancelling in ${_countdown}s`);
-                  } else {
-                    clearInterval(_tick);
-                    if (currentAbort && !currentAbort.signal.aborted) {
-                      currentAbort._reason = 'offline';
-                      currentAbort.abort();
-                    }
-                  }
-                }, 1000);
-              }
-            } catch (e) {
-              if (e && e.name !== 'AbortError' && spinner && spinner.element && !accumulated) {
-                spinner.updateMessage('Still waiting for model');
-              }
-            } finally {
-              processingProbeAbort = null;
-            }
-          }, 10000);
-        }
+        scheduleFirstTokenWaitMessages();
       }
       
       const researchBtn = el('research-toggle-btn');
@@ -1150,6 +1119,11 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
         uiModule.scrollHistory();
       }
 
+      function _replaceThinkingSpinner(label) {
+        _removeThinkingSpinner();
+        _showThinkingSpinner(label);
+      }
+
       // Auto-show thinking spinner after text stops streaming
       let _textPauseTimer = null;
       function _scheduleThinkingSpinner() {
@@ -1173,10 +1147,24 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
       let _liveThinkHeader = null;
       let _liveThinkSpinnerSlot = null;
       let _liveThinkTimerEl = null;
+      let _liveThinkTokenCount = 0;
       let _liveThinkToggle = null;
       let _liveThinkDomId = null;
 
+      function _estimateThinkingTokens(text) {
+        const clean = (text || '').trim();
+        if (!clean) return 0;
+        return Math.max(1, Math.ceil(clean.length / 4));
+      }
+
+      function _formatThinkStats(seconds, tokenCount) {
+        const time = seconds ? seconds + 's' : '';
+        const tokens = tokenCount ? tokenCount + ' tok' : '';
+        return time && tokens ? time + ' · ' + tokens : (time || tokens);
+      }
+
       function _replyAfterClosedThinking(text) {
+        text = markdownModule.normalizeThinkingMarkup(text || '');
         const closeRe = /<\/(?:think(?:ing)?|thought)>|<channel\|>/gi;
         let match = null;
         let last = null;
@@ -1187,7 +1175,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
 
       // Direct render helper for streaming text
       _renderStream = () => {
-        let dt = stripToolBlocks(roundText);
+        let dt = markdownModule.normalizeThinkingMarkup(stripToolBlocks(roundText));
         const bodyEl = roundHolder.querySelector('.body');
         const contentEl = _ensureStreamLayout(bodyEl);
 
@@ -1277,6 +1265,12 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
 
       let _nextIsError = false;
       let _streamSawDone = false;
+      let _firstVisibleOutputSeen = false;
+      const markFirstVisibleOutput = () => {
+        if (_firstVisibleOutputSeen) return;
+        _firstVisibleOutputSeen = true;
+        clearFirstTokenWaitTimers();
+      };
 
       while (true) {
         const { done, value } = await reader.read();
@@ -1296,6 +1290,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
           }
           if (line.startsWith('data: ')) {
             const data = line.slice(6);
+            if (data && data !== '[DONE]') markFirstVisibleOutput();
 
             // (thinking spinner removal is handled in agent_step / tool_start / content handlers)
 
@@ -1357,7 +1352,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                 if (_liveThinkHeader) _liveThinkHeader.textContent = 'View thinking process';
                 if (_liveThinkSpinnerSlot) _liveThinkSpinnerSlot.remove();
                 if (_liveThinkTimerEl && _elapsedDone) {
-                  _liveThinkTimerEl.textContent = _elapsedDone + 's';
+                  _liveThinkTimerEl.textContent = _formatThinkStats(_elapsedDone, _liveThinkTokenCount);
                   _liveThinkTimerEl.style.marginLeft = 'auto';
                   _liveThinkTimerEl.style.marginRight = '5px';
                   var _hdrDone = _liveThinkTimerEl.closest('.thinking-header');
@@ -1399,9 +1394,17 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                 typewriterInto(roundHolder.querySelector('.body'), errMsg);
                 break;
               }
-              if (json.delta || json.type === 'tool_start' || json.type === 'tool_output' || json.type === 'tool_progress' || json.type === 'agent_step' || json.type === 'doc_stream_open' || json.type === 'doc_stream_delta' || json.type === 'research_progress') {
+              if (json.delta || json.type === 'agent_prep' || json.type === 'tool_start' || json.type === 'tool_output' || json.type === 'tool_progress' || json.type === 'agent_step' || json.type === 'doc_stream_open' || json.type === 'doc_stream_delta' || json.type === 'research_progress') {
                 clearResponseTimeout();
                 clearProcessingProbe();
+                clearFirstTokenWaitTimers();
+              }
+              if (json.type === 'agent_prep') {
+                if (!_isBg) {
+                  _cancelThinkingTimer();
+                  _replaceThinkingSpinner('Preparing agent');
+                }
+                continue;
               }
               if (json.delta) {
                 _cancelThinkingTimer();
@@ -1464,12 +1467,13 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                 // 1. Normal: <think>...no closing tag yet
                 // 2. Malformed: <think></think>\n...text but no second </think> yet
                 // 3. Qwen3.5: "Thinking Process:" without <think> tags
-                let hasUnclosedThink = markdownModule.hasUnclosedThinkTag(roundText);
+                const normalizedRoundText = markdownModule.normalizeThinkingMarkup(roundText);
+                let hasUnclosedThink = markdownModule.hasUnclosedThinkTag(normalizedRoundText);
                 // Detect non-tag thinking patterns: "Thinking:", "Thinking Process:", Gemma-style reasoning
                 // These patterns don't use <think> tags, so we simulate unclosed thinking during streaming
                 const _replyPrefixes = ['Hey', 'Hi ', 'Hi!', 'Hello', 'Sure', 'Yes', 'No ', 'No,', 'Yo', 'OK', 'Here', 'Absolutely', 'Of course', 'Great', 'Alright', 'Thanks', 'Welcome', 'Good ', "I'm happy", "I'd be"];
-                if (!hasUnclosedThink && !/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>|<\|channel>thought/i.test(roundText)) {
-                  const _trimmedRT = roundText.trimStart();
+                if (!hasUnclosedThink && !/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>|<\|channel>thought/i.test(normalizedRoundText)) {
+                  const _trimmedRT = normalizedRoundText.trimStart();
                   const _isReasoning = markdownModule.startsWithReasoningPrefix(_trimmedRT);
                   if (_isReasoning) {
                     // Check if we can see a reply boundary yet (newline then reply pattern)
@@ -1494,9 +1498,9 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                     }
                   }
                 }
-                if (!hasUnclosedThink && /^<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>\s*<\/(?:think(?:ing)?|thought)>/i.test(roundText)) {
+                if (!hasUnclosedThink && /^<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>\s*<\/(?:think(?:ing)?|thought)>/i.test(normalizedRoundText)) {
                   // Empty <think></think> — the model likely put thinking outside the tags
-                  const afterEmpty = roundText.replace(/^<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>\s*<\/(?:think(?:ing)?|thought)>/i, '').trim();
+                  const afterEmpty = normalizedRoundText.replace(/^<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>\s*<\/(?:think(?:ing)?|thought)>/i, '').trim();
                   const closeTags = (afterEmpty.match(/<\/(?:think(?:ing)?|thought)>/gi) || []).length;
                   if (closeTags === 0 && afterEmpty.length > 0) {
                     hasUnclosedThink = true; // still waiting for real closing tag
@@ -1506,10 +1510,10 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                 // Only applies when there's a second </think> later (model leaked thinking outside tags)
                 // Do NOT trigger if the text after </think> contains tool calls (that's real content)
                 if (!hasUnclosedThink && isThinking) {
-                  const _thinkMatch = roundText.match(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>([\s\S]*?)<\/(?:think(?:ing)?|thought)>/i);
+                  const _thinkMatch = normalizedRoundText.match(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>([\s\S]*?)<\/(?:think(?:ing)?|thought)>/i);
                   const _thinkLen = _thinkMatch ? _thinkMatch[1].trim().length : 0;
                   if (_thinkLen < 20) {
-                    const _afterClose = roundText.replace(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>([\s\S]*?)<\/(?:think(?:ing)?|thought)>/i, '').trim();
+                    const _afterClose = normalizedRoundText.replace(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>([\s\S]*?)<\/(?:think(?:ing)?|thought)>/i, '').trim();
                     // Only keep waiting if there's trailing text that looks like thinking (not tool calls)
                     const _hasToolCall = /```(?:bash|python|web_search|read_file|write_file|create_document|edit_document|manage_|generate_image)/i.test(_afterClose);
                     const _hasOrphanClose = /<\/(?:think(?:ing)?|thought)>/i.test(_afterClose);
@@ -1554,7 +1558,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                   function _tickThinkTimer() {
                     if (!_liveThinkTimerEl || !_liveThinkTimerEl.isConnected) return;
                     var s = ((Date.now() - _thinkTimerStart) / 1000).toFixed(1);
-                    _liveThinkTimerEl.textContent = s + 's';
+                    _liveThinkTimerEl.textContent = _formatThinkStats(s, _liveThinkTokenCount);
                     _thinkTimerRAF = requestAnimationFrame(_tickThinkTimer);
                   }
                   _thinkTimerRAF = requestAnimationFrame(_tickThinkTimer);
@@ -1570,13 +1574,18 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                 } else if (hasUnclosedThink && isThinking) {
                   if (_liveThinkInner) {
                     // Extract raw thinking text (strip known thinking wrappers and prefixes)
-                    var thinkText = roundText
+                    var thinkText = markdownModule.normalizeThinkingMarkup(roundText)
                       .replace(/<\/?(?:think(?:ing)?|thought)(?:\s+[^>]*)?>/gi, '')
                       .replace(/<\|channel>thought\s*\n?/gi, '')
                       .replace(/<\|channel>response\s*\n?/gi, '')
                       .replace(/<channel\|>/gi, '');
                     thinkText = thinkText.replace(/^\s*Thinking(?:\s+Process)?:\s*/i, '');
+                    _liveThinkTokenCount = _estimateThinkingTokens(thinkText);
                     _liveThinkInner.innerHTML = markdownModule.mdToHtml(thinkText);
+                    if (_liveThinkTimerEl) {
+                      var _elapsedLive = thinkingStartTime ? ((Date.now() - thinkingStartTime) / 1000).toFixed(1) : '';
+                      _liveThinkTimerEl.textContent = _formatThinkStats(_elapsedLive, _liveThinkTokenCount);
+                    }
                     // Keep thinking box scrolled to bottom, but let user scroll up
                     var thinkBox = _liveThinkInner.closest('.thinking-content');
                     if (thinkBox) {
@@ -1600,6 +1609,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                     _liveThinkHeader = null;
                     _liveThinkSpinnerSlot = null;
                     _liveThinkTimerEl = null;
+                    _liveThinkTokenCount = 0;
                     _liveThinkToggle = null;
                     _liveThinkDomId = null;
                     // Fall through to normal streaming
@@ -1622,7 +1632,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                   if (_liveThinkSpinnerSlot) _liveThinkSpinnerSlot.remove();
                   // Move timer to right side of header
                   if (_liveThinkTimerEl && elapsed) {
-                    _liveThinkTimerEl.textContent = elapsed + 's';
+                    _liveThinkTimerEl.textContent = _formatThinkStats(elapsed, _liveThinkTokenCount);
                     _liveThinkTimerEl.style.marginLeft = 'auto';
                     _liveThinkTimerEl.style.marginRight = '5px';
                     var _hdrRow = _liveThinkTimerEl.closest('.thinking-header');
@@ -2023,7 +2033,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                   cancelAnimationFrame(_thinkTimerRAF);
                   var _elapsed2 = thinkingStartTime ? ((Date.now() - thinkingStartTime) / 1000).toFixed(1) : null;
                   if (_liveThinkHeader) _liveThinkHeader.textContent = 'View thinking process';
-                  if (_liveThinkTimerEl) _liveThinkTimerEl.textContent = _elapsed2 ? _elapsed2 + 's' : '';
+                  if (_liveThinkTimerEl) _liveThinkTimerEl.textContent = _elapsed2 ? _formatThinkStats(_elapsed2, _liveThinkTokenCount) : '';
                   if (_liveThinkSpinnerSlot) _liveThinkSpinnerSlot.remove();
                   // Assign stable IDs
                   var _thinkId2 = 'think-' + Date.now();
@@ -2037,7 +2047,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
                 if (!roundFinalized) {
                   roundFinalized = true;
                   if (spinner && spinner.element) spinner.destroy();
-                  const dt = stripToolBlocks(roundText);
+                  const dt = markdownModule.normalizeThinkingMarkup(stripToolBlocks(roundText));
                   if (dt.trim()) {
                     var _body3 = roundHolder.querySelector('.body');
                     var _contentEl3 = _ensureStreamLayout(_body3);
@@ -3018,6 +3028,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
     } finally {
       clearResponseTimeout();
       clearProcessingProbe();
+      clearFirstTokenWaitTimers();
       // Streaming done — let screen readers announce the settled response.
       const _chatLogDone = document.getElementById('chat-history');
       if (_chatLogDone) _chatLogDone.setAttribute('aria-busy', 'false');
@@ -3396,7 +3407,7 @@ import { wireArrowUpRecall, getLastUserMessageFromChatHistory } from './composer
     };
 
     const renderDelta = () => {
-      const dt = stripToolBlocks(roundText);
+      const dt = markdownModule.normalizeThinkingMarkup(stripToolBlocks(roundText));
       contentDiv.innerHTML = markdownModule.mdToHtml(markdownModule.squashOutsideCode(dt));
       uiModule.scrollHistory();
     };

static/js/compare/index.js

@@ -73,6 +73,45 @@ function isCompareActive() {
   return state.isActive;
 }
 
+function _compareModeLabel() {
+  return ({ search: ' search providers', agent: ' agents', research: ' research models' }[state._compareMode] || ' models');
+}
+
+function _setToolbarMode(mode, syncModeTools = !state.isActive) {
+  const target = mode === 'agent' ? 'agent' : 'chat';
+  const toggleState = Storage.loadToggleState();
+  toggleState.mode = target;
+  Storage.saveToggleState(toggleState);
+  const agentBtn = document.getElementById('mode-agent-btn');
+  const chatBtn = document.getElementById('mode-chat-btn');
+  const modeToggle = agentBtn?.closest('.mode-toggle') || chatBtn?.closest('.mode-toggle') || document.querySelector('.mode-toggle');
+  if (agentBtn && chatBtn) {
+    agentBtn.classList.toggle('active', target === 'agent');
+    chatBtn.classList.toggle('active', target === 'chat');
+    agentBtn.setAttribute('aria-pressed', target === 'agent' ? 'true' : 'false');
+    chatBtn.setAttribute('aria-pressed', target === 'chat' ? 'true' : 'false');
+  }
+  if (modeToggle) {
+    modeToggle.classList.toggle('mode-chat', target === 'chat');
+    modeToggle.classList.toggle('mode-right', target === 'chat');
+  }
+  if (syncModeTools) {
+    document.querySelectorAll('[data-mode-tool]').forEach(b => { b.style.display = target === 'agent' ? '' : 'none'; });
+  }
+}
+
+function _syncCompareModeFromToolbar(mode) {
+  if (!state.isActive) return;
+  state._compareMode = mode === 'agent' ? 'agent' : 'chat';
+  _setToolbarMode(state._compareMode, false);
+  const headerLabel = document.querySelector('.compare-header-label');
+  if (headerLabel) {
+    headerLabel.textContent = 'Comparing' + _compareModeLabel() + (state._blindMode ? ' (blind)' : '') + ' · ' + state._timeout + 's timeout';
+  }
+  const evalWrap = document.getElementById('cmp-eval-wrap');
+  if (evalWrap && typeof evalWrap._renderItems === 'function') evalWrap._renderItems();
+}
+
 // ────────────────────────────────────────────────────────────────────────────
 // ── closeCompare ──
 // ────────────────────────────────────────────────────────────────────────────
@@ -170,12 +209,7 @@ async function deactivate(teardown) {
   });
 
   // Restore agent/chat mode to what it was before compare
-  const _ts = Storage.loadToggleState();
-  _ts.mode = state._savedMode;
-  Storage.saveToggleState(_ts);
-  const _ab2 = document.getElementById('mode-agent-btn'), _cb2 = document.getElementById('mode-chat-btn');
-  if (_ab2 && _cb2) { _ab2.classList.toggle('active', state._savedMode === 'agent'); _cb2.classList.toggle('active', state._savedMode === 'chat'); }
-  document.querySelectorAll('[data-mode-tool]').forEach(b => { b.style.display = state._savedMode === 'agent' ? '' : 'none'; });
+  _setToolbarMode(state._savedMode, true);
 
   // Delete unsaved sessions, then reload
   if (teardown) {
@@ -258,19 +292,30 @@ async function _buildCompareUI() {
     if (el) state._savedIndicatorDisplay[id] = el.style.display;
   });
 
-  // 5. Save current mode and lock to the right one for this compare type
+  // 5. Save current mode and seed the toolbar for this compare type.
   const _toggleState = Storage.loadToggleState();
   state._savedMode = _toggleState.mode || 'chat';
   const _targetMode = (state._compareMode === 'agent') ? 'agent' : 'chat';
-  _toggleState.mode = _targetMode;
-  Storage.saveToggleState(_toggleState);
+  _setToolbarMode(_targetMode, false);
   const _ab = document.getElementById('mode-agent-btn'), _cb = document.getElementById('mode-chat-btn');
+  let _modeCleanup = null;
+  const _onCompareModeClick = (ev) => {
+    ev.stopPropagation();
+    ev.stopImmediatePropagation();
+    _syncCompareModeFromToolbar(ev.currentTarget === _ab ? 'agent' : 'chat');
+  };
   if (_ab && _cb) {
-    _ab.classList.toggle('active', _targetMode === 'agent');
-    _cb.classList.toggle('active', _targetMode === 'chat');
+    _ab.addEventListener('click', _onCompareModeClick, true);
+    _cb.addEventListener('click', _onCompareModeClick, true);
+    _modeCleanup = document.createElement('span');
+    _modeCleanup.style.display = 'none';
+    _modeCleanup._cleanup = () => {
+      _ab.removeEventListener('click', _onCompareModeClick, true);
+      _cb.removeEventListener('click', _onCompareModeClick, true);
+    };
   }
   const _modeToggle = document.querySelector('.mode-toggle');
-  if (_modeToggle) { _modeToggle.style.pointerEvents = 'none'; _modeToggle.style.opacity = '0.4'; }
+  if (_modeToggle) { _modeToggle.style.pointerEvents = ''; _modeToggle.style.opacity = ''; }
 
   // 6. Force tool toggles per compare mode
   disableToolToggles();
@@ -289,6 +334,7 @@ async function _buildCompareUI() {
   // 7. Hide existing chat container children (preserves event listeners)
   const container = document.getElementById('chat-container');
   state._compareElements = [];
+  if (_modeCleanup) state._compareElements.push(_modeCleanup);
   Array.from(container.children).forEach(child => {
     if (child.style.display === 'none') return;
     child.dataset.cmpHidden = '1';
@@ -302,9 +348,9 @@ async function _buildCompareUI() {
   headerBar.className = 'compare-header-bar';
   headerBar.style.cssText = 'display:flex;align-items:center;justify-content:space-between;padding:6px 10px;flex-shrink:0;';
   const headerLabel = document.createElement('span');
+  headerLabel.className = 'compare-header-label';
   headerLabel.style.cssText = 'font-size:10px;font-weight:400;color:var(--fg);white-space:nowrap;overflow:hidden;text-overflow:ellipsis;min-width:0;';
-  const _modeLabel = ({ search: ' search providers', agent: ' agents', research: ' research models' }[state._compareMode] || ' models');
-  headerLabel.textContent = 'Comparing' + _modeLabel + (state._blindMode ? ' (blind)' : '') + ' · ' + state._timeout + 's timeout';
+  headerLabel.textContent = 'Comparing' + _compareModeLabel() + (state._blindMode ? ' (blind)' : '') + ' · ' + state._timeout + 's timeout';
   // Left side: the Compare tool icon (two side-by-side panes, matching the
   // rail/sidebar icon) + the label. Other tool headers carry their icon; this
   // one was missing it.
@@ -475,7 +521,7 @@ async function _buildCompareUI() {
   }
   const msgTA = document.getElementById('message');
   if (msgTA) {
-    msgTA.placeholder = 'Enter prompt for all models...';
+    msgTA.placeholder = window.matchMedia('(max-width: 767px)').matches ? '' : 'Enter prompt for all models...';
     requestAnimationFrame(() => msgTA.focus());
   }
 
@@ -891,8 +937,7 @@ async function _executeCompare(message) {
     let sharedSearchContext = null;
     let sharedSearchSources = null;
     const webChk = document.getElementById('web-toggle');
-    const toggleState = Storage.loadToggleState();
-    const isAgentMode = (toggleState.mode || 'chat') === 'agent';
+    const isAgentMode = state._compareMode === 'agent';
     const webOn = webChk && webChk.checked;
     // In agent mode, web_search is a tool (handled per-pane); in chat mode, pre-search and share
     if (webOn && !isAgentMode) {
@@ -1198,6 +1243,15 @@ function _setupEvalPicker() {
 
   function _renderItems() {
     const mode = state._compareMode || 'chat';
+    const label = btn.querySelector('.cmp-eval-label');
+    if (label) {
+      label.textContent = ({
+        agent: 'Agent prompts',
+        chat: 'Chat prompts',
+        search: 'Search prompts',
+        research: 'Research prompts'
+      }[mode] || 'Eval prompts');
+    }
     // research/html aren't first-class compare types — fall back gracefully
     const key = EVAL_PROMPTS[mode] ? mode
       : (mode === 'research' ? 'search' : 'chat');
@@ -1258,8 +1312,10 @@ function _setupEvalPicker() {
   };
   document.addEventListener('click', _onDocClick);
 
+  _renderItems();
   wrap.appendChild(btn);
   wrap.appendChild(menu);
+  wrap._renderItems = _renderItems;
   inputTop.appendChild(wrap);
 
   // Expected-answer chip — placed above the chat-input-bar (outside it), so

static/js/compare/stream.js

@@ -551,23 +551,46 @@ async function streamToPane(paneIdx, sessionId, message, aiMsgEl, opts) {
       footer.className = 'msg-footer';
       const span = document.createElement('span');
       span.className = 'response-metrics';
-      let text = metrics.output_tokens + ' tokens | ' + metrics.tokens_per_second + ' tok/s';
+      const outputTokens = metrics.output_tokens;
+      const responseTime = metrics.response_time ?? metrics.total_time;
+      const explicitTps = metrics.tokens_per_second ?? metrics.gen_tps ?? metrics.tps;
+      const numericOutput = Number(outputTokens);
+      const numericTime = Number(responseTime);
+      const numericTps = Number(explicitTps);
+      const derivedTps = Number.isFinite(numericTps)
+        ? numericTps
+        : (Number.isFinite(numericOutput) && Number.isFinite(numericTime) && numericTime > 0)
+          ? numericOutput / numericTime
+          : null;
+      const tpsLabel = derivedTps != null
+        ? (derivedTps >= 100 ? String(Math.round(derivedTps)) : derivedTps.toFixed(2).replace(/\.?0+$/, ''))
+        : null;
+      const parts = [];
+      if (outputTokens != null && outputTokens !== 'undefined') {
+        parts.push(outputTokens + ' tokens');
+      }
+      if (tpsLabel != null) {
+        parts.push(tpsLabel + ' tok/s');
+      }
+      if (responseTime != null && responseTime !== 'undefined' && parts.length === 0) {
+        parts.push(responseTime + 's');
+      }
       // Add per-request cost and cost per 1000
       const _model = metrics.model || (state._selectedModels[paneIdx] && state._selectedModels[paneIdx].model) || '';
       const _cost = getModelCost(_model, metrics.input_tokens || 0, metrics.output_tokens || 0);
       // Build the metrics span with optional cost and context
-      span.textContent = text;
+      span.textContent = parts.join(' | ');
       if (_cost !== null) {
         const _cost1k = _cost * 1000;
         const costSpan = document.createElement('span');
         costSpan.style.color = 'var(--color-success, #4caf50)';
         costSpan.title = 'Estimated cost per 1,000 responses like this one';
-        costSpan.textContent = ' | $' + (_cost1k < 1 ? _cost1k.toFixed(2) : _cost1k.toFixed(0)) + '/1k';
+        costSpan.textContent = (span.textContent ? ' | ' : '') + '$' + (_cost1k < 1 ? _cost1k.toFixed(2) : _cost1k.toFixed(0)) + '/1k';
         span.appendChild(costSpan);
       }
       if (metrics.context_percent > 0) {
         const ctx = document.createElement('span');
-        ctx.textContent = ' | ' + metrics.context_percent + '% ctx';
+        ctx.textContent = (span.textContent ? ' | ' : '') + metrics.context_percent + '% ctx';
         if (metrics.context_percent >= 85) ctx.style.color = 'var(--color-error)';
         else if (metrics.context_percent >= 70) ctx.style.color = '#ff9900';
         span.appendChild(ctx);

static/js/compare/vote.js

@@ -181,7 +181,7 @@ function handleVote(winnerIdx) {
 
     let html = '';
     const caret = ' <span class="pane-title-caret">&#x25BE;</span>';
-    if (isWinner) html = '<span style="color:var(--red);margin-right:4px;">&#x2605;</span><strong>' + escapeHtml(name) + '</strong> <span style="color:var(--red);font-size:0.82em;font-weight:800;text-transform:uppercase;letter-spacing:1px;position:relative;top:-2px;">Winner!</span>' + caret;
+    if (isWinner) html = '<span style="color:var(--green, #50fa7b);margin-right:4px;">&#x2605;</span><strong>' + escapeHtml(name) + '</strong> <span style="color:var(--green, #50fa7b);font-size:0.82em;font-weight:800;text-transform:uppercase;letter-spacing:1px;position:relative;top:0;">Winner!</span>' + caret;
     else if (isTie) html = '<span style="opacity:0.5;margin-right:4px;">=</span><strong>' + escapeHtml(name) + '</strong>' + caret;
     else html = '<strong>' + escapeHtml(name) + '</strong>' + caret;
     el.innerHTML = html;

static/js/cookbook-diagnosis.js

@@ -461,6 +461,40 @@ export const ERROR_PATTERNS = [
       { label: 'Copy install command', action: () => _copyText('curl -fsSL https://ollama.com/install.sh | sh') },
     ],
   },
+  // System build deps must be checked BEFORE the llama-server catch-all:
+  // a `cmake: command not found` failure ALSO produces `llama-server:
+  // command not found` later in the script (the build aborts then the
+  // run line fails) — pattern order is first-match-wins, so without
+  // these specific entries the user gets the misleading "install
+  // llama-cpp-python[server]" suggestion when the actual blocker is a
+  // missing OS-package toolchain that pip can't ship.
+  {
+    pattern: /cmake: command not found|cmake.*not found.*Could not/i,
+    message: 'cmake is required to compile llama.cpp from source, but it is not installed on this server.',
+    suggestion: 'Suggested action: install cmake via the OS package manager — apt: cmake build-essential / pacman: cmake base-devel / dnf: cmake gcc-c++ make / brew: cmake. Cookbook can do this automatically on the next launch if your user has passwordless sudo for apt/pacman/dnf.',
+    fixes: [
+      { label: 'Open Dependencies', action: () => _openCookbookDependencies('llama_cpp') },
+      { label: 'Copy apt install', action: () => _copyText('sudo apt install -y cmake build-essential git') },
+      { label: 'Copy pacman install', action: () => _copyText('sudo pacman -Sy --needed cmake base-devel git') },
+      { label: 'Copy dnf install', action: () => _copyText('sudo dnf install -y cmake gcc gcc-c++ make git') },
+    ],
+  },
+  {
+    pattern: /^(make|g\+\+|gcc): command not found|Could not find C\+\+ compiler/i,
+    message: 'A C/C++ compiler (build-essential / base-devel) is required to compile llama.cpp.',
+    fixes: [
+      { label: 'Open Dependencies', action: () => _openCookbookDependencies('llama_cpp') },
+      { label: 'Copy apt install', action: () => _copyText('sudo apt install -y build-essential') },
+    ],
+  },
+  {
+    pattern: /^git: command not found/i,
+    message: 'git is required to clone the llama.cpp source tree.',
+    fixes: [
+      { label: 'Open Dependencies', action: () => _openCookbookDependencies('llama_cpp') },
+      { label: 'Copy apt install', action: () => _copyText('sudo apt install -y git') },
+    ],
+  },
   {
     pattern: /llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'/i,
     message: 'llama-cpp-python server is not installed. Run: pip install "llama-cpp-python[server]"',
@@ -714,11 +748,15 @@ export function _showDiagnosis(panel, diagnosis, sourceText) {
   copyBtn.addEventListener('click', async (e) => {
     e.stopPropagation();
     const bundle = _diagnosisCopyBundle(task, diagnosis, sourceText, suggestionText);
-    try {
-      await navigator.clipboard.writeText(bundle);
+    // Use the shared helper which falls back to execCommand('copy') on
+    // non-HTTPS origins (Tailscale IPs, LAN IPs, etc.) — navigator.clipboard
+    // is silently a no-op on those, which is why the button appeared dead
+    // for users on http://100.113.161.2:7011 over Tailscale/mobile.
+    const ok = await _copyText(bundle);
+    if (ok) {
       copyBtn.classList.add('copied');
       setTimeout(() => { if (copyBtn.isConnected) copyBtn.classList.remove('copied'); }, 1200);
-    } catch (_) {}
+    }
   });
 
   const dismissBtn = document.createElement('button');

static/js/cookbook-hwfit.js

@@ -578,7 +578,9 @@ export async function _hwfitFetch(fresh = false) {
   const _cached = fresh ? null : _readScanCache(_sig);
   const wp = spinnerModule.createWhirlpool(18);
   if (_cached) {
-    _hwfitCache = _cached;
+    // Tag the restored cache with its host too (scan-sig keys cache per
+    // host, so a hit here is always for the current remoteHost).
+    _hwfitCache = { ..._cached, _scannedHost: remoteHost || '' };
     _hwfitRenderHw(hw, _cached.system);
     if (!remoteHost && _cached.system && _cached.system.platform) {
       _envState.platform = _cached.system.platform;
@@ -750,7 +752,11 @@ export async function _hwfitFetch(fresh = false) {
         : _olRows;
       data.models = (data.models || []).concat(_olFiltered);
     }
-    _hwfitCache = data;
+    // Tag the cache with the host this scan was for, so downstream
+    // code (_gpuEnvVarName, backend-aware command builders) can avoid
+    // trusting a stale scan when the user switches the server picker
+    // to a different target without re-running hwfit.
+    _hwfitCache = { ...data, _scannedHost: remoteHost || '' };
     _hwfitRenderHw(hw, data.system);
     // Propagate local platform from hardware probe so _isWindows(task) works
     // for local tasks (menu items, shell commands, etc.).
@@ -1415,23 +1421,11 @@ export function _expandModelRow(row, modelData) {
 
   const dlSource = _downloadSourceRepo(modelData, backend);
   const hfUrl = `https://huggingface.co/${dlSource.repo}`;
-  // Official vendor recipe deep-links. These point to vLLM / SGLang's curated
-  // hardware-specific launch-command pages. They 404 for uncatalogued models \u2014
-  // a known tradeoff; user just gets the vendor's "model not found" page.
-  const _recipeRepo = modelData.name || '';
-  const _vllmUrl = _recipeRepo ? `https://recipes.vllm.ai/${_recipeRepo}` : '';
-  const _sglangUrl = _recipeRepo ? `https://docs.sglang.io/cookbook/autoregressive/${_recipeRepo}${_sglangHashFor(modelData)}` : '';
   let html = `<div class="hwfit-action-panel" data-model-name="${esc(modelData.name)}">`;
   html += `<div class="hwfit-panel-header">`;
   html += `<span class="hwfit-panel-model">${esc(modelData.name)}${dlSource.kind ? ` <span style="opacity:0.5;font-size:10px;">(${esc(dlSource.kind)} ${esc(modelData.quant || '')})</span>` : (modelData.quant_repo ? ` <span style="opacity:0.5;font-size:10px;">(${esc(modelData.quant)})</span>` : '')}</span>`;
   html += `<span class="hwfit-panel-badge">${esc(label)}</span>`;
   html += `<a href="${esc(hfUrl)}" target="_blank" rel="noopener" class="hwfit-panel-hf-link" title="View download source on HuggingFace">HF \u2197</a>`;
-  if (backend === 'vllm' && _vllmUrl) {
-    html += `<a href="${esc(_vllmUrl)}" target="_blank" rel="noopener" class="hwfit-panel-hf-link" title="vLLM official recipe (curated launch command). 404s if this model isn't in vLLM's recipes catalog.">vLLM \u2197</a>`;
-  }
-  if (backend === 'sglang' && _sglangUrl) {
-    html += `<a href="${esc(_sglangUrl)}" target="_blank" rel="noopener" class="hwfit-panel-hf-link" title="SGLang cookbook (hash pre-filled with your detected hardware). 404s if this model isn't in SGLang's cookbook catalog.">SGLang \u2197</a>`;
-  }
   html += `</div>`;
   html += `<div class="hwfit-panel-actions">`;
   html += `<button class="cookbook-btn hwfit-dl-btn">Download</button>`;
@@ -1679,7 +1673,7 @@ export function _expandModelRow(row, modelData) {
       } else if (runBackend === 'llamacpp') {
         const dir = `"$HOME/.cache/huggingface/hub/models--${modelData.name.replace(/\//g, '--')}/snapshots"`;
         const ggufPath = `$({ find ${dir} -name '*-00001-of-*.gguf' 2>/dev/null | sort; find ${dir} -name '*.gguf' 2>/dev/null | sort; } | head -1)`;
-        cmd = `MODEL_FILE=${ggufPath} && { [ -n "$MODEL_FILE" ] && [ -f "$MODEL_FILE" ]; } || { echo "ERROR: No GGUF found on this host. Download a GGUF quant or switch backend."; exit 1; } && llama-server --model "$MODEL_FILE" --host 0.0.0.0 --port 8080 -ngl 99 -c ${maxCtx} || python3 -m llama_cpp.server --model "$MODEL_FILE" --host 0.0.0.0 --port 8080 --n_gpu_layers 99 --n_ctx ${maxCtx}`;
+        cmd = `llama-server --model "${ggufPath}" --host 0.0.0.0 --port 8080 -ngl 99 -c ${maxCtx} --flash-attn auto`;
       } else {
         cmd = `vllm serve ${modelData.name} --host 0.0.0.0 --port ${port}`;
         cmd += ` --tensor-parallel-size ${tp}`;

static/js/cookbookDownload.js

@@ -85,6 +85,22 @@ function _ggufIncludePattern(model, source) {
   return '*.gguf';
 }
 
+function _ggufDisplayPartFromInclude(include) {
+  const clean = String(include || '').replace(/\*/g, '');
+  const parts = clean.split('/').filter(Boolean);
+  const file = parts[parts.length - 1] || clean;
+  const dir = parts.length > 1 ? parts[parts.length - 2] : '';
+  const quant = `${dir} ${file}`.match(/\b(?:UD-)?(?:IQ[1-8]_[A-Z0-9]+|Q[2-8]_K_[MLS]|Q[2-8]_[0-9A-Z]+|Q[2-8])\b/i);
+  if (quant) return quant[0].toUpperCase().replace(/^UD-/, '');
+  return file.replace(/\.gguf$/i, '').replace(/-\d{5}-of-\d{5}$/i, '');
+}
+
+function _downloadTaskName(shortName, payload) {
+  const include = payload?.include || '';
+  const part = include ? _ggufDisplayPartFromInclude(include) : '';
+  return part ? `${shortName} · ${part}` : shortName;
+}
+
 function _missingGgufMessage(model) {
   const name = model?.name || 'this model';
   if (/\bnvfp4\b/i.test(name)) {
@@ -519,6 +535,7 @@ export async function _runModelDownload(panel, model, backend, hostOverride) {
   }
 
   const shortName = (model.name || repo).split('/').pop();
+  const taskName = _downloadTaskName(shortName, payload);
   const targetHost = host || 'local';
 
   const tasks = _loadTasks();
@@ -576,7 +593,7 @@ export async function _runModelDownload(panel, model, backend, hostOverride) {
   if (activeOnHost) {
     const queueId = `queue-${Date.now().toString(36)}`;
     const allTasks = _loadTasks();
-    allTasks.push({ id: queueId, sessionId: queueId, name: shortName, type: 'download', status: 'queued', output: '', ts: Date.now(), payload, remoteHost: host });
+    allTasks.push({ id: queueId, sessionId: queueId, name: taskName, type: 'download', status: 'queued', output: '', ts: Date.now(), payload, remoteHost: host });
     _saveTasks(allTasks);
     _renderRunningTab();
     uiModule.showToast(`Queued ${shortName} — waiting for current download`);
@@ -601,8 +618,8 @@ export async function _runModelDownload(panel, model, backend, hostOverride) {
       uiModule.showToast('Download failed: ' + (data.error || ''), 9000);
       return;
     }
-    _addTask(data.session_id, shortName, 'download', payload);
-    uiModule.showToast(`Downloading ${shortName}...`);
+    _addTask(data.session_id, taskName, 'download', payload);
+    uiModule.showToast(`Downloading ${taskName}...`);
   } catch (e) {
     uiModule.showToast('Download failed: ' + e.message, 9000);
   }

static/js/cookbookRunning.js

@@ -27,6 +27,9 @@ function _statusLabel(status, type) {
 // "cookbook-task-status" ('' = the neutral loading style).
 function _taskBadge(task) {
   if (task._unreachable && task.status === 'running') return { text: 'unreachable', cls: 'cookbook-task-error' };
+  if (task.type === 'download' && task.status === 'running') {
+    return { text: _statusLabel(task.status, task.type), cls: 'cookbook-task-downloading' };
+  }
   if (task.type === 'serve' && task.status === 'running' && task.progress) {
     // Same green "running" pill — just with dynamic phase text, so it doesn't
     // read as a different status while the server is coming up.
@@ -35,6 +38,47 @@ function _taskBadge(task) {
   return { text: _statusLabel(task.status, task.type), cls: 'cookbook-task-' + task.status };
 }
 
+function _ggufDisplayPartFromPath(path) {
+  const parts = String(path || '').split('/').filter(Boolean);
+  const file = parts[parts.length - 1] || '';
+  const dir = parts.length > 1 ? parts[parts.length - 2] : '';
+  const text = `${dir} ${file}`;
+  const quant = text.match(/\b(?:UD-)?(?:IQ[1-8]_[A-Z0-9]+|Q[2-8]_K_[MLS]|Q[2-8]_[0-9A-Z]+|Q[2-8])\b/i);
+  if (quant) return quant[0].toUpperCase().replace(/^UD-/, '');
+  return file.replace(/\.gguf$/i, '').replace(/-\d{5}-of-\d{5}$/i, '');
+}
+
+function _downloadDisplayName(name, task) {
+  const include = task?.payload?.include || '';
+  if (!include || String(name || '').includes(' · ')) return name;
+  const part = _ggufDisplayPartFromPath(include.replace(/\*/g, ''));
+  return part ? `${name} · ${part}` : name;
+}
+
+function _taskDisplayName(task) {
+  const name = String(task?.name || '').trim();
+  if (task?.type === 'download') return _downloadDisplayName(name, task);
+  if (task?.type !== 'serve') return name;
+  const gguf = task?.payload?._fields?.gguf_file || task?.payload?.gguf_file || '';
+  if (!gguf || name.includes(' · ')) return name;
+  const part = _ggufDisplayPartFromPath(gguf);
+  return part ? `${name} · ${part}` : name;
+}
+
+function _canLaunchDownloadedTask(task) {
+  return task?.type === 'download' && ['done', 'completed'].includes(task.status || '') && !!(task.payload?.repo_id || task.name);
+}
+
+function _downloadServeFields(task) {
+  const include = String(task?.payload?.include || '').trim();
+  if (!include) return null;
+  return {
+    backend: 'llamacpp',
+    _forceBackend: true,
+    _preferredGgufInclude: include,
+  };
+}
+
 // A download task whose tmux output still shows an active per-shard line
 // (e.g. "model-00012-of-00082.safetensors: 56%|") is NOT actually finished —
 // the cookbook just lost track. The clear pill becomes a "reconnect" affordance
@@ -52,13 +96,13 @@ function _downloadOutputLooksActive(task) {
 
 function _canClearTask(task) {
   if (!task || task.status === 'running') return false;
-  if (task.type === 'serve' && (task.status === 'ready' || task._serveReady)) return false;
+  if (task.type === 'serve' && (task.status === 'ready' || (task._serveReady && !['stopped', 'error', 'crashed', 'failed', 'completed'].includes(task.status)))) return false;
   // If the tmux output still shows an in-flight download, the task isn't
   // actually finished — hide the clear/check pill so it doesn't show on a
   // task that's still doing work. (The next render will reflect this and
   // ideally the self-heal flips status back to running.)
   if (_downloadOutputLooksActive(task)) return false;
-  return ['done', 'stopped', 'error', 'crashed', 'failed'].includes(task.status);
+  return ['done', 'completed', 'stopped', 'error', 'crashed', 'failed'].includes(task.status);
 }
 
 function _clearPillLabel(task) {
@@ -66,6 +110,13 @@ function _clearPillLabel(task) {
   return 'clear';
 }
 
+function _venvRootFromPath(path) {
+  let p = (path || '').toString().trim().replace(/\/+$/, '');
+  if (!p) return '';
+  p = p.replace(/\/bin\/(?:activate|python(?:3(?:\.\d+)?)?|vllm|pip(?:3)?)$/i, '');
+  return p;
+}
+
 // A pip dependency/driver install (payload._dep) reports success with the
 // runner's "=== Process exited with code 0 ===" sentinel and pip's
 // "Successfully installed" line — never the HuggingFace download markers
@@ -141,6 +192,13 @@ async function _openDownloadForGgufTask(task) {
 function _terminalServeDiagnosis(task, outputText) {
   const out = String(outputText || task?.output || '');
   if (!task || task.type !== 'serve' || !['stopped', 'error', 'crashed', 'failed'].includes(task.status) || !out.trim()) return null;
+  // Suppress the crash diagnosis when the output proves the server
+  // actually became reachable — e.g. an early `exit 127` from a failed
+  // build attempt was followed by the shim/Python fallback successfully
+  // starting Uvicorn. Without this, the user sees a confusing "build
+  // stopped before the server became reachable" toast while the server
+  // is right there serving requests.
+  if (_serveOutputLooksReady(task)) return null;
   // Pip tasks (Reinstall vLLM, Upgrade torch, etc.) ride on the serve task
   // type so they get a tmux session + show up in Running tab — but they are
   // NOT serve invocations. Their output is pip's own; the generic
@@ -256,6 +314,7 @@ let _copyText;
 let _persistEnvState;
 let _refreshDependencies;
 let _serverByVal;
+let _serverKey;
 let _selectedServer;
 let modelLogo;
 let esc;
@@ -264,6 +323,40 @@ let _detectToolParser;
 let _detectModelOptimizations;
 let _buildServeCmd;
 
+function _taskServerSelection(task) {
+  const host = task?.remoteHost || task?.payload?.remote_host || '';
+  const savedKey = task?.remoteServerKey || task?.payload?.remote_server_key || '';
+  const server = (savedKey ? _serverByVal(savedKey) : null)
+    || (host ? _serverByVal(host) : null)
+    || (host ? _envState.servers.find(s => s.host === host) : null)
+    || null;
+  const key = server ? (_serverKey ? _serverKey(server) : savedKey) : (savedKey || (host || 'local'));
+  return { host, server, key };
+}
+
+function _selectTaskServer(task) {
+  const { host, server, key } = _taskServerSelection(task);
+  _envState.remoteHost = host;
+  _envState.remoteServerKey = key === 'local' ? '' : key;
+  if (server) {
+    _envState.env = server.env || 'none';
+    _envState.envPath = server.envPath || '';
+    _envState.platform = server.platform || '';
+  } else if (!host) {
+    _envState.env = 'none';
+    _envState.envPath = '';
+    _envState.platform = '';
+  }
+  document.querySelectorAll('#hwfit-server-select, #hwfit-dl-server, #hwfit-cache-server, #hwfit-deps-server').forEach(sel => {
+    if (!sel || sel.tagName !== 'SELECT') return;
+    const wanted = key || (host || 'local');
+    if ([...sel.options].some(o => o.value === wanted)) sel.value = wanted;
+    else if (host && [...sel.options].some(o => o.value === host)) sel.value = host;
+    else sel.value = host ? wanted : 'local';
+  });
+  return { host, server, key };
+}
+
 // When a new action is started (download / dependency / serve), this holds the
 // new task's id so the next render collapses every other card and leaves only
 // the new one open. Consumed (cleared) by _renderRunningTab.
@@ -526,7 +619,7 @@ async function _startQueuedDownload(task) {
       if (t.sessionId === data.session_id) return false;
       return !(key && t.type === 'download' && t.status === 'queued' && _downloadDedupeKey(t) === key);
     });
-    if (!found) tasks.push(_stripTaskSecrets(launchedTask));
+    if (!found) tasks.push(_redactTaskForStorage(launchedTask));
     _saveTasks(tasks);
     _renderRunningTab();
     _startBackgroundMonitor();
@@ -636,28 +729,53 @@ function _loadPrunedTasks() {
 const _REMOVED_KEY = 'cookbook-removed-tasks';
 const _TOMBSTONE_TTL_MS = 24 * 3600 * 1000;
 function _loadTombstones() {
-  try { return JSON.parse(localStorage.getItem(_REMOVED_KEY)) || {}; }
+  try {
+    const tomb = JSON.parse(localStorage.getItem(_REMOVED_KEY)) || {};
+    const now = Date.now();
+    let changed = false;
+    for (const k in tomb) {
+      if (now - tomb[k] > _TOMBSTONE_TTL_MS) {
+        delete tomb[k];
+        changed = true;
+      }
+    }
+    if (changed) localStorage.setItem(_REMOVED_KEY, JSON.stringify(tomb));
+    return tomb;
+  }
   catch { return {}; }
 }
+function _saveTombstones(tomb) {
+  localStorage.setItem(_REMOVED_KEY, JSON.stringify(tomb || {}));
+}
 function _tombstoneTask(id) {
   if (!id) return;
   const tomb = _loadTombstones();
   const now = Date.now();
   tomb[id] = now;
   for (const k in tomb) { if (now - tomb[k] > _TOMBSTONE_TTL_MS) delete tomb[k]; }
-  localStorage.setItem(_REMOVED_KEY, JSON.stringify(tomb));
+  _saveTombstones(tomb);
 }
 function _isTombstoned(id) {
   const ts = _loadTombstones()[id];
   return ts != null && (Date.now() - ts) <= _TOMBSTONE_TTL_MS;
 }
 
-function _stripTaskSecrets(task) {
+function _redactStoredText(value) {
+  return String(value || '')
+    .replace(/hf_[A-Za-z0-9]{20,}/g, '[redacted-token]')
+    .replace(/((?:api[_-]?key|token|authorization|password|passwd|secret)\s*[=:]\s*)(["']?)[^\s"']+/gi, '$1$2[redacted]');
+}
+
+function _redactTaskForStorage(task) {
   if (!task || typeof task !== 'object') return task;
   const safe = { ...task };
+  if (typeof safe.output === 'string') safe.output = _redactStoredText(safe.output);
   if (safe.payload && typeof safe.payload === 'object') {
     safe.payload = { ...safe.payload };
     delete safe.payload.hf_token;
+    delete safe.payload.hfToken;
+    if (typeof safe.payload._cmd === 'string') safe.payload._cmd = _redactStoredText(safe.payload._cmd);
+    if (typeof safe.payload.cmd === 'string') safe.payload.cmd = _redactStoredText(safe.payload.cmd);
   }
   return safe;
 }
@@ -666,23 +784,24 @@ function _stripStateSecrets(state) {
   const safe = { ...state };
   if (safe.env && typeof safe.env === 'object') {
     const { hfToken, ...env } = safe.env;
-    if (hfToken) env.hfToken = hfToken;
     safe.env = env;
   }
-  if (Array.isArray(safe.tasks)) safe.tasks = safe.tasks.map(_stripTaskSecrets);
+  if (Array.isArray(safe.tasks)) safe.tasks = safe.tasks.map(_redactTaskForStorage);
   return safe;
 }
 
 export function _saveTasks(tasks) {
-  localStorage.setItem(TASKS_KEY, JSON.stringify((tasks || []).map(_stripTaskSecrets)));
+  localStorage.setItem(TASKS_KEY, JSON.stringify((tasks || []).map(_redactTaskForStorage)));
   _syncToServer();
 }
 
 export function _addTask(sessionId, name, type, payload) {
   let tasks = _loadTasks();
   const remoteHost = (payload && payload.remote_host) || _envState.remoteHost || '';
-  const sshPort = (payload && payload.ssh_port) || _getPort(remoteHost) || '';
-  const platform = (payload && payload.platform) || _getPlatform(remoteHost) || '';
+  const remoteServerKey = (payload && payload.remote_server_key) || '';
+  const remoteServerName = (payload && payload.remote_server_name) || '';
+  const sshPort = (payload && payload.ssh_port) || _getPort(remoteServerKey || remoteHost) || '';
+  const platform = (payload && payload.platform) || _getPlatform(remoteServerKey || remoteHost) || '';
   // Serving a model supersedes its finished download — clear the matching
   // finished download card (covers serving directly from the Serve tab, not just
   // via the download card's "Serve →" button).
@@ -697,7 +816,7 @@ export function _addTask(sessionId, name, type, payload) {
       return !(key && t.type === 'download' && t.status === 'queued' && _downloadDedupeKey(t) === key);
     });
   }
-  const task = _stripTaskSecrets({ id: sessionId, sessionId, name, type, status: 'running', output: '', ts: Date.now(), payload: payload || null, remoteHost, sshPort, platform });
+  const task = _redactTaskForStorage({ id: sessionId, sessionId, name, type, status: 'running', output: '', ts: Date.now(), payload: payload || null, remoteHost, remoteServerKey, remoteServerName, sshPort, platform });
   tasks.push(task);
   _saveTasks(tasks);
   // New action → collapse all other cards, leave only this one open.
@@ -992,14 +1111,24 @@ function _presetEnvFields(task) {
   };
 }
 
+function _redactPresetForStorage(preset) {
+  if (!preset || typeof preset !== 'object') return preset;
+  const safe = { ...preset };
+  if (typeof safe.cmd === 'string') safe.cmd = _redactStoredText(safe.cmd);
+  if (typeof safe.command === 'string') safe.command = _redactStoredText(safe.command);
+  delete safe.hf_token;
+  delete safe.hfToken;
+  return safe;
+}
+
 function _saveTaskAsPreset(task, label) {
   const host = task.remoteHost || 'localhost';
   const portMatch = task.payload?._cmd?.match(/--port\s+(\d+)/);
   const port = portMatch ? portMatch[1] : '8000';
   const presets = _loadPresets();
   if (presets.some(p => p.cmd === task.payload._cmd)) return false;
-  presets.push({ name: task.name, model: task.payload.repo_id, backend: 'vllm', host, port, cmd: task.payload._cmd, remoteHost: task.remoteHost || '', label: label || task.name, ..._presetEnvFields(task) });
-  _savePresets(presets);
+  presets.push(_redactPresetForStorage({ name: task.name, model: task.payload.repo_id, backend: 'vllm', host, port, cmd: task.payload._cmd, remoteHost: task.remoteHost || '', label: label || task.name, ..._presetEnvFields(task) }));
+  _savePresets(presets.map(_redactPresetForStorage));
   return true;
 }
 
@@ -1042,7 +1171,7 @@ function _autoSaveWorkingConfig(task) {
   const existing = presets.find(p => p.cmd === cmd);
   if (existing) {
     task._autoSaved = true;
-    if (!existing.confirmedWorking) { existing.confirmedWorking = true; _savePresets(presets); }
+    if (!existing.confirmedWorking) { existing.confirmedWorking = true; _savePresets(presets.map(_redactPresetForStorage)); }
     return;   // already saved → just confirm it, no duplicate, no toast
   }
   // Respect the per-model cap the manual save flow uses (max 5).
@@ -1050,13 +1179,13 @@ function _autoSaveWorkingConfig(task) {
   const host = task.remoteHost || 'localhost';
   const portMatch = cmd.match(/--port[=\s]+(\d+)/);
   const port = portMatch ? portMatch[1] : '8000';
-  presets.push({
+  presets.push(_redactPresetForStorage({
     name: task.name, model, backend: 'vllm', host, port,
     cmd, remoteHost: task.remoteHost || '',
     label: _autoConfigLabel(task), confirmedWorking: true, autoSaved: true,
     ..._presetEnvFields(task),
-  });
-  _savePresets(presets);
+  }));
+  _savePresets(presets.map(_redactPresetForStorage));
   task._autoSaved = true;
   uiModule.showToast('Saved working config');
 }
@@ -1078,6 +1207,7 @@ function _syncToServer() {
       if (!_envState || !Array.isArray(_envState.servers) || _envState.servers.length === 0) return;
       const state = {
         tasks: _loadTasks(),
+        removedTasks: _loadTombstones(),
         presets: _loadPresets(),
         env: _envState,
         serveState: null,
@@ -1126,15 +1256,22 @@ export async function _syncFromServer() {
 
     const localTasks = _loadTasks();
     const serverTasks = state.tasks || [];
+    const serverTombstones = (state.removedTasks && typeof state.removedTasks === 'object') ? state.removedTasks : {};
+    const localTombstones = _loadTombstones();
+    const mergedTombstones = { ...serverTombstones, ...localTombstones };
+    for (const [id, ts] of Object.entries(serverTombstones)) {
+      if (localTombstones[id] == null || Number(ts) > Number(localTombstones[id])) mergedTombstones[id] = ts;
+    }
+    _saveTombstones(mergedTombstones);
 
     const localIds = new Set(localTasks.map(t => t.sessionId));
-    const merged = [...localTasks];
+    const merged = localTasks.filter(t => !_isTombstoned(t.sessionId));
     for (const t of serverTasks) {
       if (!localIds.has(t.sessionId) && !_isTombstoned(t.sessionId)) {
         merged.push(t);
       }
     }
-    localStorage.setItem(TASKS_KEY, JSON.stringify(merged.map(_stripTaskSecrets)));
+    localStorage.setItem(TASKS_KEY, JSON.stringify(merged.map(_redactTaskForStorage)));
 
     if (state.env) {
       // The active server selection (remoteHost + its env/path/platform) is a
@@ -1145,6 +1282,18 @@ export async function _syncFromServer() {
       const { remoteHost: _rh, env: _e, envPath: _ep, platform: _pf, ...settings } = state.env;
       delete settings.hfToken;
       Object.assign(_envState, settings);
+      const selected = (_envState.remoteServerKey && _serverByVal?.(_envState.remoteServerKey))
+        || (_envState.remoteHost ? (_envState.servers || []).find(s => s.host === _envState.remoteHost) : null);
+      if (selected) {
+        _envState.env = selected.env || 'none';
+        _envState.envPath = selected.envPath || '';
+        _envState.platform = selected.platform || '';
+      } else if (!_envState.remoteHost) {
+        const local = (_envState.servers || []).find(s => !s.host || s.host === 'local');
+        _envState.env = local?.env || 'none';
+        _envState.envPath = local?.envPath || '';
+        _envState.platform = local?.platform || '';
+      }
       const { hfToken, ...safeState } = _envState;
       localStorage.setItem('cookbook-last-state', JSON.stringify(safeState));
     }
@@ -1154,6 +1303,7 @@ export async function _syncFromServer() {
     if (state.serveState) {
       localStorage.setItem(SERVE_STATE_KEY, JSON.stringify(state.serveState));
     }
+    document.dispatchEvent(new CustomEvent('cookbook:state-synced', { detail: state }));
     return true;
   } catch { return false; }
 }
@@ -1312,17 +1462,11 @@ async function _openServeEditForTask(task, cmdOverride, fieldOverrides = null) {
   if (fieldOverrides && typeof fieldOverrides === 'object') {
     fields = { ...(fields || {}), ...fieldOverrides };
   }
-  // Switch the active server to the one this serve ran on (mirrors _openEdit).
-  const _tHost = task.remoteHost || '';
-  _envState.remoteHost = _tHost;
-  const _tSrv = _serverByVal(_envState.remoteServerKey || _tHost)
-    || _envState.servers.find(s => s.host === _tHost);
-  if (_tSrv) { _envState.env = _tSrv.env || 'none'; _envState.envPath = _tSrv.envPath || ''; _envState.platform = _tSrv.platform || ''; }
-  else if (!_tHost) { _envState.env = 'none'; _envState.envPath = ''; _envState.platform = ''; }
-  document.querySelectorAll('#hwfit-server-select, #hwfit-dl-server, #hwfit-cache-server, #hwfit-deps-server').forEach(sel => {
-    if (!sel || sel.tagName !== 'SELECT') return;
-    sel.value = _tHost || 'local';
-  });
+  fields = { ...(fields || {}), _replaceTaskId: task.sessionId };
+  // Switch the active server to the exact profile this serve ran on. The
+  // dropdown stores stable srv: keys, not raw host strings, so preserving only
+  // task.remoteHost can relaunch against the local container by accident.
+  _selectTaskServer(task);
   try {
     const { openServePanelForRepo } = await import('./cookbookServe.js');
     await openServePanelForRepo(repo, fields);
@@ -1520,15 +1664,33 @@ function _parseServeCmdToFields(cmd) {
   return fields;
 }
 
-export async function _launchServeTask(shortName, repo, cmd, fields, hostOverride) {
+export async function _launchServeTask(shortName, repo, cmd, fields, hostOverride, targetMeta = null) {
   // Host resolution mirrors the download path: when the caller passes an explicit
   // host (resolved from the dropdown the user actually picked), use it and look
   // up that server's port/platform from the shared servers list. Only fall back
   // to _envState.remoteHost for legacy callers (diagnosis/pip-update).
   const _host = (hostOverride !== undefined) ? (hostOverride || '') : (_envState.remoteHost || '');
-  const _hsrv = _serverByVal(_envState.remoteServerKey || _host)
+  const _targetKey = targetMeta?.serverKey || '';
+  const _hsrv = (_targetKey && _targetKey !== 'local' ? _serverByVal(_targetKey) : null)
+    || (hostOverride === undefined ? _serverByVal(_envState.remoteServerKey || _host) : null)
     || _envState.servers.find(s => s.host === _host) || {};
+  const _serverMetaKey = _targetKey || (_hsrv && _serverKey ? _serverKey(_hsrv) : '') || (_host || 'local');
+  const _serverMetaName = targetMeta?.serverName || _hsrv.name || (_host ? _host : 'Local');
   const _hplatform = _host ? (_hsrv.platform || '') : (_envState.platform || '');
+  const _replaceTaskId = fields?._replaceTaskId || '';
+  if (_replaceTaskId) {
+    try {
+      const _old = _loadTasks().find(t => t.sessionId === _replaceTaskId);
+      if (_old && _old.type === 'serve') {
+        await fetch('/api/shell/exec', {
+          method: 'POST', credentials: 'same-origin',
+          headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify({ command: _tmuxGracefulKill(_old) }),
+        });
+        _removeTask(_old.sessionId);
+      }
+    } catch {}
+  }
 
   // Replace any serve already targeting this same host:port — you can't run two
   // servers on one port, so re-serving (or retrying) should stop & remove the
@@ -1572,7 +1734,7 @@ export async function _launchServeTask(shortName, repo, cmd, fields, hostOverrid
     }
   } else {
     if (_envState.env === 'venv' && _envState.envPath) {
-      const p = _envState.envPath;
+      const p = _venvRootFromPath(_envState.envPath);
       envPrefix = 'source ' + (p.endsWith('/bin/activate') ? p : p + '/bin/activate');
     } else if (_envState.env === 'conda' && _envState.envPath) {
       envPrefix = 'eval "$(conda shell.bash hook)" && conda activate ' + _envState.envPath;
@@ -1583,7 +1745,7 @@ export async function _launchServeTask(shortName, repo, cmd, fields, hostOverrid
     repo_id: repo,
     cmd: cmd,
     remote_host: _host || undefined,
-    ssh_port: _getPort(_host) || undefined,
+    ssh_port: _getPort(_serverMetaKey || _host) || undefined,
     env_prefix: envPrefix || undefined,
     hf_token: _envState.hfToken || undefined,
     gpus: _envState.gpus || undefined,
@@ -1607,11 +1769,11 @@ export async function _launchServeTask(shortName, repo, cmd, fields, hostOverrid
       return;
     }
 
-    const _sp = _getPort(_host);
+    const _sp = _getPort(_serverMetaKey || _host);
     // _fields = the exact structured serve-form values used for this launch,
     // so the "Edit / relaunch" button can re-open the Serve panel pre-filled
     // with these precise settings (not just the last-used-for-repo state).
-    const payload = { repo_id: repo, remote_host: _host || undefined, ssh_port: _sp || undefined, _cmd: cmd, _fields: fields || undefined, _env: _usedEnv, _envPath: _usedEnvPath, _gpus: _usedGpus };
+    const payload = { repo_id: repo, remote_host: _host || undefined, remote_server_key: _serverMetaKey || undefined, remote_server_name: _serverMetaName || undefined, ssh_port: _sp || undefined, _cmd: cmd, _fields: fields || undefined, _env: _usedEnv, _envPath: _usedEnvPath, _gpus: _usedGpus };
     _addTask(data.session_id, shortName, 'serve', payload);
     uiModule.showToast(`Serving ${shortName}...`);
     // Auto-register may have enabled an existing (offline) endpoint for this
@@ -1726,7 +1888,7 @@ export function _renderRunningTab() {
       '<div style="display:flex;align-items:baseline;gap:8px;margin-bottom:2px;">' +
       '<h2 style="margin:0;padding:0;line-height:1;">Active <span id="running-count" class="memory-count" style="font-size:0.6em;opacity:0.6;font-weight:normal">' + activeCount + '</span></h2>' +
       '</div>' +
-      '<p class="memory-desc doclib-desc" style="margin-top:6px;">Active downloads and serving processes.</p>' +
+      '<p class="memory-desc doclib-desc" style="margin-top:6px;">Active downloads, installs and model launches.</p>' +
       '</div>';
     const firstGroup = body.querySelector('.cookbook-group');
     if (firstGroup) body.insertBefore(group, firstGroup);
@@ -1760,16 +1922,25 @@ export function _renderRunningTab() {
   }
 
   // Group tasks by server
-  const _serverName = (host) => {
-    if (!host) return 'Local';
-    const srv = _serverByVal(_envState.remoteServerKey || host)
-      || _envState.servers.find(s => s.host === host);
-    return srv?.name || host;
+  const _taskServerKey = (task) => task?.remoteServerKey || task?.remoteHost || '';
+  const _serverName = (keyOrTask) => {
+    if (keyOrTask && typeof keyOrTask === 'object') {
+      const task = keyOrTask;
+      if (task.remoteServerName) return task.remoteServerName;
+      const srv = task.remoteServerKey ? _serverByVal(task.remoteServerKey) : null;
+      if (srv?.name) return srv.name;
+      if (!task.remoteHost) return 'Local';
+      return (_envState.servers.find(s => s.host === task.remoteHost)?.name) || task.remoteHost;
+    }
+    const key = keyOrTask || '';
+    if (!key || key === 'local') return 'Local';
+    const srv = _serverByVal(key);
+    return srv?.name || key;
   };
   const serverGroups = {};
   for (const t of tasks) {
-    const key = t.remoteHost || '';
-    if (!serverGroups[key]) serverGroups[key] = { name: _serverName(key), serve: [], download: [] };
+    const key = _taskServerKey(t);
+    if (!serverGroups[key]) serverGroups[key] = { name: _serverName(t), serve: [], download: [] };
     serverGroups[key][t.type === 'serve' ? 'serve' : 'download'].push(t);
   }
 
@@ -1816,12 +1987,12 @@ export function _renderRunningTab() {
       e.stopPropagation();  // don't toggle the section collapse (was an inline onclick, blocked by CSP)
       const host = btn.dataset.clearServer;
       const allTasks = _loadTasks();
-      const toRemove = allTasks.filter(t => (t.remoteHost || '') === host && _canClearTask(t));
+      const toRemove = allTasks.filter(t => _taskServerKey(t) === host && _canClearTask(t));
       // Bail with a clear message instead of silently doing nothing when
       // every task on this server is still running (nothing finished to
       // clear yet) — the previous behavior looked like the button was dead.
       if (!toRemove.length) {
-        const stillRunning = allTasks.filter(t => (t.remoteHost || '') === host && t.status === 'running').length;
+        const stillRunning = allTasks.filter(t => _taskServerKey(t) === host && t.status === 'running').length;
         const _msg = stillRunning
           ? `No finished tasks on ${_serverName(host)} — ${stillRunning} still running. Stop them first to clear.`
           : `No finished tasks on ${_serverName(host)}.`;
@@ -1830,7 +2001,8 @@ export function _renderRunningTab() {
         return;
       }
       if (!await window.styledConfirm(`Clear ${toRemove.length} finished task${toRemove.length === 1 ? '' : 's'} on ${_serverName(host)}?`, { confirmText: 'Clear' })) return;
-      const remaining = allTasks.filter(t => (t.remoteHost || '') !== host || !_canClearTask(t));
+      toRemove.forEach(t => _tombstoneTask(t.sessionId));
+      const remaining = allTasks.filter(t => _taskServerKey(t) !== host || !_canClearTask(t));
       _saveTasks(remaining);
       // Fade/slide each finished card out (same exit as the per-card clear)
       // instead of yanking them instantly.
@@ -1864,7 +2036,7 @@ export function _renderRunningTab() {
     btn.addEventListener('click', async (e) => {
       e.stopPropagation();  // don't toggle the section collapse
       const host = btn.dataset.stopServer;
-      const running = _loadTasks().filter(t => (t.remoteHost || '') === host && t.status === 'running');
+      const running = _loadTasks().filter(t => _taskServerKey(t) === host && t.status === 'running');
       if (!running.length) { uiModule.showToast(`Nothing running on ${_serverName(host)}`); return; }
       if (!await window.styledConfirm(`Stop ${running.length} running task${running.length > 1 ? 's' : ''} on ${_serverName(host)}?`, { confirmText: 'Stop all' })) return;
       // Mark every task as user-stopped BEFORE firing the kills so that the
@@ -1967,11 +2139,12 @@ export function _renderRunningTab() {
 
     const _bdg = _taskBadge(task);
     const _bdgTitle = (task._unreachable && task.status === 'running') ? ' title="Server not responding — it may have crashed"' : '';
+    const displayName = _taskDisplayName(task);
     el.innerHTML = `
       <div class="cookbook-task-header">
         <span class="cookbook-task-type${(task.status === 'done' && task.type === 'download') ? ' cookbook-task-type-done' : ''}" data-type="${esc(task.type)}">${esc((task.status === 'done' && task.type === 'download') ? 'finished' : task.type)}</span>
-        <span class="cookbook-task-name">${modelLogo(task.name)}${esc(task.name)}</span>
-        <span class="cookbook-task-indicator"><span class="cookbook-task-wave" style="display:${task.status === 'running' ? '' : 'none'}"></span><span class="cookbook-task-check" title="Clear" style="display:${_canClearTask(task) ? '' : 'none'}"><svg class="cookbook-task-check-ico" width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="#50fa7b" stroke-width="3" stroke-linecap="round" stroke-linejoin="round"><polyline points="20 6 9 17 4 12"/></svg><svg class="cookbook-task-clear-ico" width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="3" stroke-linecap="round" stroke-linejoin="round"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg><span class="cookbook-task-done-label">${esc(_clearPillLabel(task))}</span><span class="cookbook-task-clear-label">clear</span></span></span>
+        <span class="cookbook-task-name">${modelLogo(task.name)}${esc(displayName)}</span>
+        <span class="cookbook-task-indicator"><span class="cookbook-task-wave" style="display:${task.status === 'running' ? '' : 'none'}"></span>${_canLaunchDownloadedTask(task) ? '<button type="button" class="cookbook-task-serve-btn" title="Open in Launch"><svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"><polygon points="13 2 3 14 12 14 11 22 21 10 12 10 13 2"/></svg><span>Launch</span></button>' : ''}<span class="cookbook-task-check" title="Clear" style="display:${_canClearTask(task) ? '' : 'none'}"><svg class="cookbook-task-check-ico" width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="#50fa7b" stroke-width="3" stroke-linecap="round" stroke-linejoin="round"><polyline points="20 6 9 17 4 12"/></svg><svg class="cookbook-task-clear-ico" width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="3" stroke-linecap="round" stroke-linejoin="round"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg><span class="cookbook-task-done-label">${esc(_clearPillLabel(task))}</span><span class="cookbook-task-clear-label">clear</span></span></span>
         <button type="button" class="cookbook-task-start-now" title="Start this queued download now" style="display:${(task.type === 'download' && task.status === 'queued') ? '' : 'none'}"><svg width="11" height="11" viewBox="0 0 24 24" fill="currentColor" aria-hidden="true"><polygon points="8 5 19 12 8 19 8 5"/></svg><span>start now</span></button>
         <span class="cookbook-task-status ${_bdg.cls}"${_bdgTitle}>${esc(_bdg.text)}</span>
         <button class="cookbook-task-menu-btn" title="Actions">&#8942;</button>
@@ -2043,19 +2216,11 @@ export function _renderRunningTab() {
           e.stopPropagation();
           const repo = task.payload?.repo_id || task.name;
           if (!repo) { uiModule.showToast('No model info on this task'); return; }
-          // Point the active server at the one it downloaded to.
-          const _tHost = task.remoteHost || '';
-          _envState.remoteHost = _tHost;
-          const _tSrv = _serverByVal(_envState.remoteServerKey || _tHost)
-            || _envState.servers.find(s => s.host === _tHost);
-          if (_tSrv) { _envState.env = _tSrv.env || 'none'; _envState.envPath = _tSrv.envPath || ''; _envState.platform = _tSrv.platform || ''; }
-          else if (!_tHost) { _envState.env = 'none'; _envState.envPath = ''; _envState.platform = ''; }
-          document.querySelectorAll('#hwfit-server-select, #hwfit-dl-server, #hwfit-cache-server, #hwfit-deps-server').forEach(sel => {
-            if (sel && sel.tagName === 'SELECT') sel.value = _tHost || 'local';
-          });
+          // Point the active server at the exact profile it downloaded to.
+          _selectTaskServer(task);
           try {
             const { openServePanelForRepo } = await import('./cookbookServe.js');
-            await openServePanelForRepo(repo);
+            await openServePanelForRepo(repo, _downloadServeFields(task));
             // Serving it supersedes the finished download — clear the card from
             // the Running tab (smooth exit) now that we've jumped to Serve.
             _animateOutThenRemove(el, task.sessionId);
@@ -2177,9 +2342,6 @@ export function _renderRunningTab() {
         if (task.status !== 'running' && task.status !== 'queued') {
           items.push({ group: 'run', label: 'Reconnect tmux', action: 'reconnect' });
         }
-        if (task.status === 'running') {
-          items.push({ group: 'run', label: 'Stop', action: 'stop', danger: true });
-        }
         items.push({ group: 'run', label: 'Restart', action: 'retry' });
         // ── Edit section ────────────────────────────────────────────
         // Merged "Edit & relaunch" — opens the structured serve panel
@@ -2539,7 +2701,7 @@ export function _renderRunningTab() {
     });
 
     // Route to the right server section body
-    const serverBodyId = `server-body-${(task.remoteHost || 'local').replace(/[^a-zA-Z0-9-]/g, '_')}`;
+    const serverBodyId = `server-body-${(_taskServerKey(task) || 'local').replace(/[^a-zA-Z0-9-]/g, '_')}`;
     const targetBody = document.getElementById(serverBodyId);
     if (targetBody) targetBody.appendChild(el);
     else group.appendChild(el);
@@ -3393,7 +3555,8 @@ function _refreshServerDots() {
   let tasks;
   try { tasks = _loadTasks(); } catch { return; }
   const byKey = {};
-  for (const t of tasks) { (byKey[t.remoteHost || ''] = byKey[t.remoteHost || ''] || []).push(t); }
+  const _taskServerKeyForDot = (task) => task?.remoteServerKey || task?.remoteHost || '';
+  for (const t of tasks) { (byKey[_taskServerKeyForDot(t)] = byKey[_taskServerKeyForDot(t)] || []).push(t); }
   document.querySelectorAll('.cookbook-section-header').forEach(header => {
     const dot = header.querySelector('.cookbook-srv-status');
     if (!dot) return;
@@ -3527,7 +3690,9 @@ async function _probeEndpointUntilOnline(epId, host, port) {
     try {
       // Hit the probe endpoint — it re-probes server-side and updates
       // cached_models. We consume (and discard) the SSE stream.
-      await fetch(`/api/model-endpoints/${epId}/probe`, { credentials: 'same-origin' }).then(r => r.text()).catch(() => {});
+      const probeRes = await fetch(`/api/model-endpoints/${epId}/probe`, { credentials: 'same-origin' }).catch(() => null);
+      if (probeRes && probeRes.status === 404) return;
+      if (probeRes) await probeRes.text().catch(() => {});
       const eps = await fetch('/api/model-endpoints', { credentials: 'same-origin' }).then(r => r.json()).catch(() => []);
       const ep = (eps || []).find(e => e.id === epId);
       if (ep && (ep.models || []).length) {
@@ -3565,7 +3730,7 @@ async function _pollBackgroundStatus() {
             }
           }
           if (added > 0) {
-            localStorage.setItem(TASKS_KEY, JSON.stringify(merged.map(_stripTaskSecrets)));
+            localStorage.setItem(TASKS_KEY, JSON.stringify(merged.map(_redactTaskForStorage)));
             _renderRunningTab();
           }
         }
@@ -3798,6 +3963,7 @@ export function initRunning(shared) {
   _persistEnvState = shared._persistEnvState;
   _refreshDependencies = shared._refreshDependencies;
   _serverByVal = shared._serverByVal;
+  _serverKey = shared._serverKey;
   _selectedServer = shared._selectedServer;
   modelLogo = shared.modelLogo;
   esc = shared.esc;

static/js/document.js

@@ -24,6 +24,7 @@ import * as Modals from './modalManager.js';
   let _autoDetectDebounce = null;
   let _autoTitleDebounce = null;
   let _autoSaveDebounce = null;
+  let _lastAutoSaveErrorAt = 0;
   let _animationInProgress = false;
   let _animationCancel = null;      // function to cancel current animation
   let _htmlPreviewActive = false;   // true when inline HTML preview iframe is showing
@@ -154,6 +155,20 @@ import * as Modals from './modalManager.js';
       addDocToTabs,
       syncDocIndicator: _syncDocIndicator,
     });
+    const sidebarNewDocBtn = document.getElementById('library-new-doc-btn');
+    if (sidebarNewDocBtn && !sidebarNewDocBtn.dataset.docNewWired) {
+      sidebarNewDocBtn.dataset.docNewWired = '1';
+      sidebarNewDocBtn.addEventListener('click', async (e) => {
+        e.preventDefault();
+        e.stopPropagation();
+        try {
+          await newDocument();
+        } catch (err) {
+          console.error('Failed to create document from sidebar button:', err);
+          if (uiModule) uiModule.showError('Failed to create document');
+        }
+      });
+    }
     _maybeOpenDocFromHash();
     window.addEventListener('hashchange', _maybeOpenDocFromHash);
   }
@@ -2686,6 +2701,104 @@ import * as Modals from './modalManager.js';
     await _uploadComposeFiles(files);
   }
 
+  function _isMarkdownImageFile(file) {
+    if (!file) return false;
+    if ((file.type || '').toLowerCase().startsWith('image/')) return true;
+    return /\.(avif|bmp|gif|jpe?g|png|svg|webp)$/i.test(file.name || '');
+  }
+
+  function _markdownImageAlt(name) {
+    const base = String(name || 'image').replace(/\.[^.]+$/, '').trim() || 'image';
+    return base.replace(/[\[\]\n\r]/g, ' ').replace(/\s+/g, ' ').trim() || 'image';
+  }
+
+  function _activeDocLanguage() {
+    const doc = activeDocId && docs.get(activeDocId);
+    return ((doc && doc.language) || document.getElementById('doc-language-select')?.value || '').toLowerCase();
+  }
+
+  function _scheduleMarkdownImageAutosave(ta) {
+    updateLineNumbers(ta.value);
+    const codeEl = document.getElementById('doc-editor-code');
+    if (codeEl && !codeEl.dataset.hasDiff) {
+      codeEl.textContent = ta.value + '\n';
+      codeEl.style.minHeight = ta.scrollHeight + 'px';
+    }
+    clearTimeout(_hlDebounce);
+    _hlDebounce = setTimeout(syncHighlighting, 80);
+    clearTimeout(_autoTitleDebounce);
+    _autoTitleDebounce = setTimeout(() => autoTitleFromContent(ta.value), 600);
+    clearTimeout(_autoSaveDebounce);
+    _autoSaveDebounce = setTimeout(() => { saveDocument({ silent: true }); }, 800);
+  }
+
+  function _insertMarkdownImages(uploadedFiles) {
+    const ta = document.getElementById('doc-editor-textarea');
+    if (!ta) return;
+    const files = Array.isArray(uploadedFiles) ? uploadedFiles : [];
+    if (!files.length) return;
+
+    const start = ta.selectionStart || 0;
+    const end = ta.selectionEnd || start;
+    const before = ta.value.slice(0, start);
+    const after = ta.value.slice(end);
+    const lines = files.map(file => {
+      const id = encodeURIComponent(file.id || file.file_id || '');
+      const alt = _markdownImageAlt(file.name || file.filename);
+      return id ? `![${alt}](/api/upload/${id})` : '';
+    }).filter(Boolean);
+    if (!lines.length) return;
+
+    const prefix = before && !before.endsWith('\n') ? '\n' : '';
+    const suffix = after && !after.startsWith('\n') ? '\n' : '';
+    const insert = `${prefix}${lines.join('\n\n')}${suffix}`;
+    _replaceRange(ta, start, end, insert);
+    const caret = start + insert.length;
+    ta.selectionStart = caret;
+    ta.selectionEnd = caret;
+    ta.focus();
+    _scheduleMarkdownImageAutosave(ta);
+    _refreshMarkdownPreviewIfVisible(activeDocId, ta.value);
+  }
+
+  async function _uploadMarkdownImages(files) {
+    const images = Array.from(files || []).filter(_isMarkdownImageFile);
+    if (!images.length) {
+      if (uiModule) uiModule.showError('Choose an image file');
+      return;
+    }
+    if (_activeDocLanguage() !== 'markdown') {
+      if (uiModule) uiModule.showError('Switch the document to markdown before inserting images');
+      return;
+    }
+
+    const fd = new FormData();
+    images.forEach(file => fd.append('files', file));
+    try {
+      const res = await fetch(`${API_BASE}/api/upload`, {
+        method: 'POST',
+        credentials: 'same-origin',
+        body: fd,
+      });
+      let data = null;
+      try { data = await res.json(); } catch (_) {}
+      if (!res.ok) throw new Error((data && (data.error || data.detail)) || `HTTP ${res.status}`);
+      const uploaded = Array.isArray(data?.files) ? data.files : [];
+      if (!uploaded.length) throw new Error('No uploaded files returned');
+      _insertMarkdownImages(uploaded);
+      if (uiModule) uiModule.showToast(images.length === 1 ? 'Image inserted' : 'Images inserted');
+    } catch (err) {
+      console.error('Failed to insert markdown image:', err);
+      if (uiModule) uiModule.showError('Failed to insert image');
+    }
+  }
+
+  async function _handleMarkdownImageUpload(e) {
+    const files = e.target.files;
+    e.target.value = '';
+    await _uploadMarkdownImages(files);
+  }
+
   function _renderComposeAttachments() {
     const container = document.getElementById('doc-email-compose-atts');
     if (!container) return;
@@ -3752,9 +3865,12 @@ import * as Modals from './modalManager.js';
       const res = await fetch(`${API_BASE}/api/document`, {
         method: 'POST',
         headers: { 'Content-Type': 'application/json' },
+        credentials: 'same-origin',
         body: JSON.stringify({ session_id: sessionId, title: '', content }),
       });
+      if (!res.ok) throw new Error(`Document create failed: HTTP ${res.status}`);
       const doc = await res.json();
+      if (!doc || !doc.id) throw new Error('Document create failed: missing id');
       addDocToTabs(doc, sessionId);
       // Set the content into the map so switchToDoc preserves it
       const d = docs.get(doc.id);
@@ -3981,6 +4097,7 @@ import * as Modals from './modalManager.js';
         <input type="hidden" id="doc-email-source-folder" />
         <input type="file" id="doc-email-file-input" multiple style="display:none" />
       </div>
+      <input type="file" id="doc-md-image-input" accept="image/*" multiple style="display:none" />
       <div class="doc-md-toolbar" id="doc-md-toolbar" style="display:none">
         <div class="md-toolbar-items" id="md-toolbar-items">
           <span class="md-view-toggle" id="doc-md-view-toggle" style="display:none" role="group" aria-label="Edit or preview">
@@ -4003,7 +4120,7 @@ import * as Modals from './modalManager.js';
           <button type="button" class="md-dd-toggle" data-dd="list" title="List"><span style="font-variant-numeric:tabular-nums;">1.</span><svg width="8" height="8" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="3.5" stroke-linecap="round" stroke-linejoin="round"><polyline points="6 9 12 15 18 9"/></svg></button>
           <span class="md-toolbar-sep"></span>
           <button type="button" data-md="link" title="Link"><svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></svg></button>
-          <button type="button" id="md-toolbar-attach-btn" class="md-toolbar-attach-btn" title="Attach files"><svg width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="m21.44 11.05-9.19 9.19a6 6 0 0 1-8.49-8.49l8.57-8.57A4 4 0 1 1 17.93 8.8l-8.59 8.57a2 2 0 0 1-2.83-2.83l8.49-8.48"/></svg></button>
+          <button type="button" id="md-toolbar-attach-btn" class="md-toolbar-attach-btn" title="Insert image"><svg width="13" height="13" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="m21.44 11.05-9.19 9.19a6 6 0 0 1-8.49-8.49l8.57-8.57A4 4 0 1 1 17.93 8.8l-8.59 8.57a2 2 0 0 1-2.83-2.83l8.49-8.48"/></svg></button>
           <button type="button" class="md-dd-toggle md-toolbar-email-hide" data-dd="code" title="Code">\`<svg width="8" height="8" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="3.5" stroke-linecap="round" stroke-linejoin="round"><polyline points="6 9 12 15 18 9"/></svg></button>
           <button type="button" data-md="hr" title="Horizontal rule">—</button>
           <span class="md-toolbar-sep"></span>
@@ -4602,9 +4719,14 @@ import * as Modals from './modalManager.js';
       document.getElementById('doc-email-file-input')?.click();
     });
     document.getElementById('md-toolbar-attach-btn')?.addEventListener('click', () => {
-      document.getElementById('doc-email-file-input')?.click();
+      if (_activeDocLanguage() === 'email') {
+        document.getElementById('doc-email-file-input')?.click();
+      } else {
+        document.getElementById('doc-md-image-input')?.click();
+      }
     });
     document.getElementById('doc-email-file-input')?.addEventListener('change', _handleAttachUpload);
+    document.getElementById('doc-md-image-input')?.addEventListener('change', _handleMarkdownImageUpload);
 
     // Cc/Bcc toggle
     document.getElementById('doc-email-show-cc')?.addEventListener('click', () => {
@@ -4840,6 +4962,26 @@ import * as Modals from './modalManager.js';
         clearTimeout(_autoSaveDebounce);
         _autoSaveDebounce = setTimeout(() => { saveDocument({ silent: true }); }, 2000);
       });
+      ta.addEventListener('paste', (e) => {
+        if (_activeDocLanguage() !== 'markdown') return;
+        const files = Array.from(e.clipboardData?.files || []).filter(_isMarkdownImageFile);
+        if (!files.length) return;
+        e.preventDefault();
+        _uploadMarkdownImages(files);
+      });
+      ta.addEventListener('dragover', (e) => {
+        if (_activeDocLanguage() !== 'markdown') return;
+        const items = Array.from(e.dataTransfer?.items || []);
+        if (!items.some(item => item.kind === 'file' && /^image\//i.test(item.type || ''))) return;
+        e.preventDefault();
+      });
+      ta.addEventListener('drop', (e) => {
+        if (_activeDocLanguage() !== 'markdown') return;
+        const files = Array.from(e.dataTransfer?.files || []).filter(_isMarkdownImageFile);
+        if (!files.length) return;
+        e.preventDefault();
+        _uploadMarkdownImages(files);
+      });
       ta.addEventListener('scroll', () => {
         const code = document.getElementById('doc-editor-code');
         if (code) code.style.minHeight = ta.scrollHeight + 'px';
@@ -5548,7 +5690,7 @@ import * as Modals from './modalManager.js';
     // any dropdown that just opened. Preventing the default mousedown keeps the
     // textarea focused, so formatting hits the live selection and menus stay up.
     toolbar.addEventListener('mousedown', (e) => {
-      if (e.target.closest('[data-md], .md-dd-toggle, .emoji-picker-btn')) e.preventDefault();
+      if (e.target.closest('[data-md], .md-dd-toggle, .emoji-picker-btn, .md-toolbar-attach-btn')) e.preventDefault();
     });
 
     toolbar.addEventListener('click', (e) => {
@@ -5976,6 +6118,7 @@ import * as Modals from './modalManager.js';
       const res = await fetch(`${API_BASE}/api/document`, {
         method: 'POST',
         headers: { 'Content-Type': 'application/json' },
+        credentials: 'same-origin',
         body: JSON.stringify({
           session_id: sessionId,
           title: '',
@@ -5983,7 +6126,9 @@ import * as Modals from './modalManager.js';
           language: 'markdown',
         }),
       });
+      if (!res.ok) throw new Error(`Document create failed: HTTP ${res.status}`);
       const doc = await res.json();
+      if (!doc || !doc.id) throw new Error('Document create failed: missing id');
       addDocToTabs(doc, sessionId);
       if (!isOpen) openPanel();
       // Re-enable editor if it was in empty state
@@ -8266,8 +8411,10 @@ import * as Modals from './modalManager.js';
       const res = await fetch(`${API_BASE}/api/document/${activeDocId}`, {
         method: 'PUT',
         headers: { 'Content-Type': 'application/json' },
+        credentials: 'same-origin',
         body: JSON.stringify({ content: textarea.value }),
       });
+      if (!res.ok) throw new Error(`Document save failed: HTTP ${res.status}`);
       const doc = await res.json();
       const badge = document.getElementById('doc-version-badge');
       if (badge) { const _v = doc.version_count || 1; badge.textContent = `v${_v}`; badge.style.display = _v > 1 ? '' : 'none'; }
@@ -8280,7 +8427,11 @@ import * as Modals from './modalManager.js';
       if (!silent && uiModule) uiModule.showToast('Document saved');
     } catch (e) {
       console.error('Failed to save document:', e);
-      if (!silent && uiModule) uiModule.showError('Failed to save document');
+      const now = Date.now();
+      if (uiModule && (!silent || now - _lastAutoSaveErrorAt > 10000)) {
+        uiModule.showError(silent ? 'Autosave failed' : 'Failed to save document');
+        _lastAutoSaveErrorAt = now;
+      }
     }
   }
 

static/js/emailLibrary.js

@@ -2936,6 +2936,20 @@ function _createCard(em) {
     titleRow.appendChild(att);
   }
 
+  const tags = Array.isArray(em.tags) ? em.tags : [];
+  if (tags.length || em.is_spam_verdict) {
+    const tagWrap = document.createElement('span');
+    tagWrap.className = 'email-tags email-card-tags';
+    tagWrap.innerHTML = tags.map(t => {
+      const tag = String(t || '').trim().toLowerCase().replace(/_/g, '-');
+      return tag ? `<span class="email-tag email-tag-${_esc(tag)}">${_esc(tag)}</span>` : '';
+    }).join('');
+    if (em.is_spam_verdict) {
+      tagWrap.insertAdjacentHTML('beforeend', '<span class="email-tag email-tag-spam">spam</span>');
+    }
+    titleRow.appendChild(tagWrap);
+  }
+
   // Done check + unread dot stay next to the subject on the left.
   const isSentFolder = /sent/i.test(state._libFolder);
   if (!isSentFolder) {
@@ -4560,11 +4574,12 @@ function _wireAttachmentHandlers(reader, folder) {
       const uid = openBtn.dataset.openUid;
       const index = openBtn.dataset.openIndex;
       const name = openBtn.dataset.openName || `attachment-${index}`;
+      const sourceFolder = openBtn.dataset.openFolder || useFolder;
       if (!uid || index == null) return;
       const orig = openBtn.style.opacity;
       openBtn.style.opacity = '0.4';
       try {
-        const folderQs = encodeURIComponent(useFolder);
+        const folderQs = encodeURIComponent(sourceFolder);
         const res = await fetch(
           `${API_BASE}/api/email/attachment-as-doc/${encodeURIComponent(uid)}/${encodeURIComponent(index)}?folder=${folderQs}${_acct()}`,
           { method: 'POST', credentials: 'same-origin' }
@@ -4618,8 +4633,9 @@ function _wireAttachmentHandlers(reader, folder) {
       const uid = chip.dataset.attUid;
       const index = chip.dataset.attIndex;
       const name = chip.dataset.attName || `attachment-${index}`;
+      const sourceFolder = chip.dataset.attFolder || useFolder;
       if (!uid || index == null) return;
-      const url = `${API_BASE}/api/email/attachment/${encodeURIComponent(uid)}/${encodeURIComponent(index)}?folder=${encodeURIComponent(useFolder)}${_acct()}`;
+      const url = `${API_BASE}/api/email/attachment/${encodeURIComponent(uid)}/${encodeURIComponent(index)}?folder=${encodeURIComponent(sourceFolder)}${_acct()}`;
       if (_isMobileUA) {
         window.open(url, '_blank');
         return;
@@ -4698,25 +4714,50 @@ function _isLikelySignatureImage(a) {
 // Build the attachments header+chips HTML for an email read response. Pulled
 // out so both the initial-open and the swap-reader paths can render it.
 function _buildAttsHtmlFor(uid, data) {
-  if (!data || !data.attachments || !data.attachments.length) return '';
-  const _OPENABLE_RE = /\.(pdf|docx|txt|md|markdown)$/i;
-  const visible = data.attachments.filter(a => !_isLikelySignatureImage(a));
-  if (!visible.length) return '';
-  const chips = visible.map(a => {
+  if (!data) return '';
+  const _OPENABLE_RE = /\.(pdf|docx|txt|md|markdown|eml)$/i;
+  const currentAttachments = Array.isArray(data.attachments) ? data.attachments : [];
+  const relatedAttachments = Array.isArray(data.related_attachments) ? data.related_attachments : [];
+  if (!currentAttachments.length && !relatedAttachments.length) return '';
+  const visible = currentAttachments.filter(a => !_isLikelySignatureImage(a));
+  const hidden = currentAttachments.filter(a => _isLikelySignatureImage(a));
+  const related = relatedAttachments.filter(a => !_isLikelySignatureImage(a));
+  const renderChip = (a, extraClass = '') => {
     const openable = _OPENABLE_RE.test(a.filename || '');
+    const chipUid = a.source_uid || a.uid || uid;
+    const chipFolder = a.source_folder || data.folder || state._libFolder || 'INBOX';
     const openBtn = openable
-      ? `<span class="email-attachment-open" title="Open in document editor" data-open-uid="${_esc(uid)}" data-open-index="${a.index}" data-open-name="${_esc(a.filename)}"><svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z"/><polyline points="14 2 14 8 20 8"/><line x1="8" y1="13" x2="16" y2="13"/><line x1="8" y1="17" x2="16" y2="17"/><line x1="8" y1="9" x2="10" y2="9"/></svg><span class="email-attachment-open-label">Open</span></span>`
+      ? `<span class="email-attachment-open" title="Open in document editor" data-open-uid="${_esc(chipUid)}" data-open-index="${a.index}" data-open-name="${_esc(a.filename)}" data-open-folder="${_esc(chipFolder)}"><svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z"/><polyline points="14 2 14 8 20 8"/><line x1="8" y1="13" x2="16" y2="13"/><line x1="8" y1="17" x2="16" y2="17"/><line x1="8" y1="9" x2="10" y2="9"/></svg><span class="email-attachment-open-label">Open</span></span>`
       : '';
-    return `<button type="button" class="email-attachment-chip" data-att-uid="${_esc(uid)}" data-att-index="${a.index}" data-att-name="${_esc(a.filename)}"><svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="m21.44 11.05-9.19 9.19a6 6 0 0 1-8.49-8.49l8.57-8.57A4 4 0 1 1 17.93 8.8l-8.59 8.57a2 2 0 0 1-2.83-2.83l8.49-8.48"/></svg><span>${_esc(a.filename)}</span><span class="att-size">${Math.round((a.size||0)/1024)} KB</span>${openBtn}</button>`;
-  }).join('');
+    return `<button type="button" class="email-attachment-chip${extraClass}" data-att-uid="${_esc(chipUid)}" data-att-index="${a.index}" data-att-name="${_esc(a.filename)}" data-att-folder="${_esc(chipFolder)}"><svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="m21.44 11.05-9.19 9.19a6 6 0 0 1-8.49-8.49l8.57-8.57A4 4 0 1 1 17.93 8.8l-8.59 8.57a2 2 0 0 1-2.83-2.83l8.49-8.48"/></svg><span>${_esc(a.filename)}</span><span class="att-size">${Math.round((a.size||0)/1024)} KB</span>${openBtn}</button>`;
+  };
+  const chips = visible.map(a => renderChip(a)).join('');
+  const hiddenChips = hidden.map(a => renderChip(a, ' email-attachment-chip-muted')).join('');
+  const relatedChips = related.map(a => renderChip(a, ' email-attachment-chip-related')).join('');
+  const visibleSection = visible.length
+    ? '<div class="email-reader-atts">' + chips + '</div>'
+    : '';
+  const relatedSection = related.length
+    ? '<div class="email-reader-atts-hidden-note">From earlier in this thread</div><div class="email-reader-atts email-reader-atts-related">' + relatedChips + '</div>'
+    : '';
+  const hiddenSection = hidden.length
+    ? '<div class="email-reader-atts-hidden-note">Filtered inline images / signature files</div><div class="email-reader-atts email-reader-atts-hidden">' + hiddenChips + '</div>'
+    : '';
+  const label = visible.length
+    ? `Attachments (${visible.length + related.length})`
+    : related.length
+      ? `Thread attachments (${related.length})`
+      : `Hidden inline attachments (${hidden.length})`;
   return (
     '<div class="email-reader-atts-wrap collapsed">'
     +   '<div class="email-reader-atts-header email-summary-toggle" role="button" tabindex="0">'
     +     '<svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="m21.44 11.05-9.19 9.19a6 6 0 0 1-8.49-8.49l8.57-8.57A4 4 0 1 1 17.93 8.8l-8.59 8.57a2 2 0 0 1-2.83-2.83l8.49-8.48"/></svg>'
-    +     `<span>Attachments (${data.attachments.length})</span>`
+    +     `<span>${label}</span>`
     +     '<svg class="email-summary-chevron" width="10" height="10" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" style="margin-left:auto;transition:transform .15s ease;"><polyline points="6 9 12 15 18 9"/></svg>'
     +   '</div>'
-    +   '<div class="email-reader-atts">' + chips + '</div>'
+    +   visibleSection
+    +   relatedSection
+    +   hiddenSection
     + '</div>'
   );
 }

static/js/markdown.js

@@ -36,6 +36,14 @@ function linkHtml(text, url) {
   return `<a href="${escapeHtml(safeUrl)}" target="_blank" rel="noopener noreferrer">${safeText}</a>`;
 }
 
+function imageHtml(alt, url, title) {
+  const safeUrl = safeLinkUrl(url);
+  if (!safeUrl || safeUrl.startsWith('#')) return escapeHtml(alt || '');
+  const safeAlt = escapeHtml(alt || '');
+  const safeTitle = title ? ` title="${escapeHtml(title)}"` : '';
+  return `<img src="${escapeHtml(safeUrl)}" alt="${safeAlt}"${safeTitle} loading="lazy" decoding="async">`;
+}
+
 function _isModelEndpointUrl(rawUrl) {
   try {
     const parsed = new URL(String(rawUrl || ''), window.location.origin);
@@ -146,7 +154,7 @@ function sanitizeAllowedHtml(html) {
  * Check if text has unclosed think tag
  */
 export function hasUnclosedThinkTag(text) {
-  text = text || '';
+  text = normalizeThinkingMarkup(text || '');
   const openCount =
     (text.match(/<(?:think(?:ing)?|thought)(?:\s+[^>]*)?>/gi) || []).length
     + (text.match(/<\|channel>thought/gi) || []).length;
@@ -163,6 +171,10 @@ export function startsWithReasoningPrefix(text) {
 export function normalizeThinkingMarkup(text) {
   if (!text) return text;
   let normalized = text;
+  // MiniMax M-series can emit namespaced reasoning tags like
+  // <mm:think>...</mm:think>. Normalize them into the shared thinking parser.
+  normalized = normalized.replace(/<mm:think(\s+[^>]*)?>/gi, (_m, attrs = '') => `<think${attrs || ''}>`);
+  normalized = normalized.replace(/<\/mm:think>/gi, '</think>');
   normalized = normalized.replace(/<thought(\s+[^>]*)?>/gi, (_m, attrs = '') => `<think${attrs || ''}>`);
   normalized = normalized.replace(/<\/thought>/gi, '</think>');
   normalized = normalized.replace(/<\|channel>thought\s*\n?([\s\S]*?)<channel\|>\s*/gi, (_m, content = '') => {
@@ -535,6 +547,12 @@ export function mdToHtml(src, opts) {
     '$1[#$2](#$2)',
   );
 
+  // Convert markdown images before links so ![alt](url) does not become
+  // literal "!" plus a normal link.
+  s = s.replace(/!\[([^\]\n]*)\]\(([^)\s]+)(?:\s+"([^"]*)")?\)/g, (match, alt, url, title) => {
+    return imageHtml(alt, url, title);
+  });
+
   // Convert markdown links [text](url) to clickable links
   // Internal #hash links navigate in-page; external links open in new tab
   s = s.replace(/\[([^\]]+)\]\(([^)]+)\)/g, (match, text, url) => {
@@ -573,8 +591,9 @@ export function mdToHtml(src, opts) {
     return placeholder;
   });
 
-  // ALSO preserve <a> tags the same way (they're now in the HTML from markdown conversion)
-  s = s.replace(/<a\s+[^>]*>.*?<\/a>/gi, (match) => {
+  // ALSO preserve <a>/<img> tags the same way (they're now in the HTML from
+  // markdown conversion)
+  s = s.replace(/<(?:a\s+[^>]*>.*?<\/a|img\s+[^>]*?)>/gi, (match) => {
     const placeholder = `___ALLOWED_HTML_${allowedHtmlBlocks.length}___`;
     allowedHtmlBlocks.push(sanitizeAllowedHtml(match));
     return placeholder;

static/js/modelPicker.js

@@ -77,6 +77,7 @@ function _handlePickerKeydown(e, listEl, itemSelector, closeFn) {
 // Dependencies injected via initModelPicker()
 let _deps = null;
 let _autoSelectingDefault = false;
+let _defaultChatPickInFlight = false;
 
 function _modelExists(modelId, url) {
   if (!modelId || !window.modelsModule || !window.modelsModule.getCachedItems) return false;
@@ -91,6 +92,43 @@ function _modelExists(modelId, url) {
   });
 }
 
+async function _ensureDefaultPendingChat() {
+  if (!_deps || _defaultChatPickInFlight) return;
+  if (_deps.getCurrentSessionId && _deps.getCurrentSessionId()) return;
+  const pending = _deps.getPendingChat && _deps.getPendingChat();
+  if (pending && pending.modelId) return;
+  _defaultChatPickInFlight = true;
+  try {
+    let dc = null;
+    try {
+      const res = await fetch(`${API_BASE}/api/default-chat`, { credentials: 'same-origin' });
+      if (res.ok) dc = await res.json();
+    } catch (_) {}
+    if (dc && dc.endpoint_url && dc.model) {
+      _deps.setPendingChat({
+        url: dc.endpoint_url,
+        modelId: dc.model,
+        endpointId: dc.endpoint_id || '',
+      });
+      try { window.__odysseusDefaultChat = dc; } catch (_) {}
+      updateModelPicker();
+      return;
+    }
+    // No configured default: preserve the old convenience fallback.
+    if (window.modelsModule && window.modelsModule.getCachedItems) {
+      const items = window.modelsModule.getCachedItems();
+      const first = items.find(item => !item.offline && ((item.models || []).length || (item.models_extra || []).length));
+      if (first) {
+        const models = (first.models || []).concat(first.models_extra || []);
+        _deps.setPendingChat({ url: first.url, modelId: models[0], endpointId: first.endpoint_id });
+        updateModelPicker();
+      }
+    }
+  } finally {
+    _defaultChatPickInFlight = false;
+  }
+}
+
 /**
  * Initialize the model picker dropdown.
  * @param {Object} deps
@@ -112,6 +150,7 @@ function _initModelPickerDropdown() {
   const search = document.getElementById('model-picker-search');
   const listEl = document.getElementById('model-picker-list');
   const searchRow = menu ? menu.querySelector('.model-picker-search-row') : null;
+  const refreshBtn = document.getElementById('model-picker-refresh-btn');
   if (!wrap || !btn || !menu || !search || !listEl) return;
 
   function _close() {
@@ -608,6 +647,26 @@ function _initModelPickerDropdown() {
 
   search.addEventListener('input', () => _populate(search.value));
   search.addEventListener('click', (e) => e.stopPropagation());
+  if (refreshBtn) {
+    refreshBtn.addEventListener('click', async (e) => {
+      e.stopPropagation();
+      refreshBtn.disabled = true;
+      refreshBtn.classList.add('spinning');
+      try {
+        if (window.modelsModule && window.modelsModule.refreshModels) {
+          await window.modelsModule.refreshModels(true);
+        }
+        await _refreshLocalProbe();
+        if (!menu.classList.contains('hidden')) _populate(search.value || '');
+        updateModelPicker();
+      } catch (_) {
+        uiModule.showToast('Model refresh failed');
+      } finally {
+        refreshBtn.disabled = false;
+        refreshBtn.classList.remove('spinning');
+      }
+    });
+  }
   search.addEventListener('keydown', (e) => {
     _handlePickerKeydown(e, listEl, '.model-switch-item', _close);
   });
@@ -689,25 +748,7 @@ export function updateModelPicker() {
     }
   }
   if (!modelId && !_autoSelectingDefault && window.modelsModule && window.modelsModule.getCachedItems) {
-    const items = window.modelsModule.getCachedItems();
-    const first = items.find(item => !item.offline && ((item.models || []).length || (item.models_extra || []).length));
-    if (first) {
-      const models = (first.models || []).concat(first.models_extra || []);
-      modelId = models[0];
-      if (!currentSessionId) {
-        _deps.setPendingChat({ url: first.url, modelId, endpointId: first.endpoint_id });
-      } else {
-        if (s) { s.model = modelId; s.endpoint_url = first.url; }
-        _autoSelectingDefault = true;
-        const fd = new FormData();
-        fd.append('model', modelId);
-        fd.append('endpoint_url', first.url || '');
-        if (first.endpoint_id) fd.append('endpoint_id', first.endpoint_id);
-        fetch(`${API_BASE}/api/session/${currentSessionId}`, { method: 'PATCH', body: fd })
-          .catch(() => {})
-          .finally(() => { _autoSelectingDefault = false; });
-      }
-    }
+    _ensureDefaultPendingChat();
   }
 
   const displayName = modelId ? modelId.split('/').pop() : 'Select model';

static/js/notes.js

@@ -608,7 +608,7 @@ function _isNoteFullyDone(note) {
 // A "checklist note" — todo or goal — has structured items[] that the cards
 // render as checkboxes and that "fully done" / progress logic reads from.
 function _hasItems(note) {
-  return note && (note.note_type === 'todo' || note.note_type === 'goal');
+  return note && (note.note_type === 'todo' || note.note_type === 'goal' || note.note_type === 'checklist');
 }
 
 // Compact " N/M" progress string for a goal's checklist. Empty when the goal
@@ -1120,8 +1120,6 @@ export function openPanel() {
   }
   _open = true;
   _editingId = null;
-  // Reset the search filter — the rebuilt pane's search input renders empty, so a
-  // stale _searchQuery would silently hide non-matching notes after a reopen.
   _searchQuery = '';
   _clearViewedReminderGlows();
   _firedDotDismissedAt = Date.now();
@@ -1822,10 +1820,20 @@ function _renderNotes() {
       for (let i = 0; i < note.items.length; i++) {
         const item = note.items[i];
         const doneClass = item.done ? ' done' : '';
+        const agentStatus = (item.agent_status || '').toLowerCase();
+        const agentDoneClass = agentStatus === 'stream_complete' ? ' is-agent-stream-complete' : '';
+        const agentTitle = agentStatus === 'stream_complete'
+          ? 'Agent stream finished for this todo'
+          : (agentStatus === 'running' ? 'Agent is working on this todo' : 'Solve this todo with the agent');
+        const agentSessionAttr = item.agent_session_id ? ` data-session-id="${_attrEsc(item.agent_session_id)}"` : '';
+        const agentMenuTitle = item.agent_session_title || `Agent: ${(item.text || '').slice(0, 40)}`;
         const indent = Math.min(item.indent || 0, 3);
         contentHtml += `<div class="note-checkbox${doneClass}" data-note-id="${note.id}" data-idx="${i}" style="padding-left:${indent * 16}px">
           <span class="note-check-dot" title="Mark done"></span>
           <span class="note-check-text">${_linkify(item.text)}</span>
+          <button class="note-checkbox-agent${agentDoneClass}" data-note-id="${_attrEsc(note.id)}" data-idx="${i}"${agentSessionAttr} data-agent-title="${_attrEsc(agentMenuTitle)}" title="${_attrEsc(agentTitle)}">
+            <svg width="10" height="10" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M12 8V4H8"/><rect x="4" y="8" width="16" height="12" rx="2"/><path d="M2 14h2M20 14h2M15 13v2M9 13v2"/></svg>
+          </button>
           <button class="note-checkbox-rm" data-note-id="${note.id}" data-idx="${i}" title="Delete item">
             <svg width="10" height="10" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg>
           </button>
@@ -1889,10 +1897,6 @@ function _renderNotes() {
       ${_hasItems(note) ? `<div class="note-cl-quickadd"><input type="text" class="note-cl-quickadd-input" placeholder="+ Add item" data-note-id="${note.id}" /></div>` : ''}
       ${reminderTagHtml}
       ${noteTags.length ? `<div class="note-card-label">${noteTags.map(t => `<button type="button" class="note-card-label-chip" data-note-label-filter="${_esc(t)}" title="Filter #${_esc(t)}">#${_esc(t)}</button>`).join(' ')}</div>` : ''}
-      ${note.agent_session_id ? `<button class="note-agent-tag" data-note-id="${note.id}" data-session-id="${_esc(note.agent_session_id)}" title="Open the agent's chat for this note">
-        <svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round"><path d="M12 8V4H8"/><rect x="4" y="8" width="16" height="12" rx="2"/><path d="M2 14h2M20 14h2M15 13v2M9 13v2"/></svg>
-        <span>Agent</span>
-      </button>` : ''}
       <div class="note-card-actions">
         <div class="note-card-colors">${colorDots}</div>
         <span style="flex:1"></span>
@@ -2177,7 +2181,7 @@ function _bindCardEvents(body) {
   // Click empty area of checklist preview (not on checkbox/X) — edit
   body.querySelectorAll('.note-checklist-preview').forEach(el => {
     el.addEventListener('click', (e) => {
-      if (e.target.closest('.note-checkbox, .note-checkbox-rm, .note-cl-quickadd, input')) return;
+      if (e.target.closest('.note-checkbox, .note-checkbox-rm, .note-checkbox-agent, .note-cl-quickadd, input')) return;
       e.stopPropagation();
       tapToEditOrSelect(el.closest('.note-card'));
     });
@@ -2203,7 +2207,7 @@ function _bindCardEvents(body) {
   // title / content preview triggered edit, so padding + empty gutters were
   // dead zones that felt broken on mobile.
   if (_isNotesMobileMode() && !_selectMode) {
-    const _INTERACTIVE = 'button, a, input, label, .note-card-color-dot, .note-checkbox, .note-checkbox-rm, .note-cl-quickadd, .note-agent-tag, .note-card-pin, .note-card-corner-trash, .note-card-corner-menu, .note-card-corner-unarchive, .note-card-edit-corner, .note-card-reminder, .note-card-cb';
+    const _INTERACTIVE = 'button, a, input, label, .note-card-color-dot, .note-checkbox, .note-checkbox-rm, .note-checkbox-agent, .note-cl-quickadd, .note-agent-tag, .note-card-pin, .note-card-corner-trash, .note-card-corner-menu, .note-card-corner-unarchive, .note-card-edit-corner, .note-card-reminder, .note-card-cb';
     body.querySelectorAll('.note-card').forEach(card => {
       card.addEventListener('click', (e) => {
         if (e.target.closest(_INTERACTIVE)) return;
@@ -2297,16 +2301,6 @@ function _bindCardEvents(body) {
       _openNoteCornerMenu(btn);
     });
   });
-  // Agent tag — opens the chat session the agent ran for this note.
-  body.querySelectorAll('.note-agent-tag').forEach(tag => {
-    tag.addEventListener('click', (e) => {
-      e.preventDefault();
-      e.stopPropagation();
-      const sid = tag.dataset.sessionId;
-      const _sm = window.sessionModule;
-      if (sid && _sm && _sm.selectSession) { closePanel(); _sm.selectSession(sid); }
-    });
-  });
   body.querySelectorAll('.note-card-label-chip').forEach(chip => {
     chip.addEventListener('click', (e) => {
       e.preventDefault();
@@ -2523,6 +2517,18 @@ function _bindCardEvents(body) {
     });
   });
 
+  // Per-item agent solve (hover button next to the X). Scoped to one todo
+  // item — uses the note title as context if present, but only the single
+  // item's text as the work. Mirrors the per-note _agentSolveNote pattern.
+  body.querySelectorAll('.note-checkbox-agent').forEach(btn => {
+    btn.addEventListener('click', (e) => {
+      e.preventDefault();
+      e.stopPropagation();
+      if (_selectMode) return;
+      _openTodoAgentMenu(btn);
+    });
+  });
+
   // Quick-add new checklist item (hover input at bottom of todo cards)
   body.querySelectorAll('.note-cl-quickadd-input').forEach(input => {
     input.addEventListener('click', (e) => e.stopPropagation());
@@ -4342,6 +4348,54 @@ function _openNoteCornerMenu(btn) {
   menu.querySelector('[data-act="agent"]').addEventListener('click', () => { menu.remove(); _agentSolveNote(id); });
 }
 
+function _positionNoteMenu(menu, btn, width = 196) {
+  document.body.appendChild(menu);
+  const r = btn.getBoundingClientRect();
+  let left = Math.min(r.right - width, window.innerWidth - width - 8);
+  left = Math.max(8, left);
+  const mh = menu.offsetHeight || 112;
+  const below = window.innerHeight - r.bottom;
+  const top = (below < mh + 8 && r.top > mh + 8) ? (r.top - mh - 4) : (r.bottom + 4);
+  menu.style.cssText += `position:fixed;z-index:11000;top:${Math.round(top)}px;left:${Math.round(left)}px;min-width:${width}px;`;
+  const close = (ev) => {
+    if (ev && menu.contains(ev.target)) return;
+    menu.remove();
+    document.removeEventListener('click', close, true);
+  };
+  setTimeout(() => document.addEventListener('click', close, true), 0);
+}
+
+function _openTodoAgentMenu(btn) {
+  document.querySelectorAll('.note-corner-menu-dropdown').forEach(d => d.remove());
+  const noteId = btn.dataset.noteId;
+  const idx = parseInt(btn.dataset.idx);
+  const sid = btn.dataset.sessionId || '';
+  const menu = document.createElement('div');
+  menu.className = 'note-corner-menu-dropdown note-agent-item-menu';
+  menu.innerHTML = `
+    ${sid ? `<button type="button" class="ncm-item" data-act="open">
+      <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M15 3h6v6"/><path d="M10 14L21 3"/><path d="M18 13v6a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V8a2 2 0 0 1 2-2h6"/></svg>
+      <span>Open</span>
+    </button>` : ''}
+    <button type="button" class="ncm-item" data-act="run">
+      <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M12 8V4H8"/><rect x="4" y="8" width="16" height="12" rx="2"/><path d="M2 14h2M20 14h2M15 13v2M9 13v2"/></svg>
+      <span>${sid ? 'Run again' : 'Run Agent'}</span>
+    </button>`;
+  _positionNoteMenu(menu, btn);
+  const openBtn = menu.querySelector('[data-act="open"]');
+  if (openBtn) {
+    openBtn.addEventListener('click', () => {
+      menu.remove();
+      const _sm = window.sessionModule;
+      if (sid && _sm && _sm.selectSession) { closePanel(); _sm.selectSession(sid); }
+    });
+  }
+  menu.querySelector('[data-act="run"]').addEventListener('click', () => {
+    menu.remove();
+    _agentSolveTodoItem(noteId, idx);
+  });
+}
+
 // Build the prompt the agent gets from a note: title + body, plus any
 // not-yet-done checklist items.
 function _noteToAgentPrompt(note) {
@@ -4353,7 +4407,7 @@ function _noteToAgentPrompt(note) {
       .forEach(it => parts.push('- ' + it.text.trim()));
   }
   const body = parts.join('\n');
-  return body ? `Help me get this done:\n\n${body}` : '';
+  return body ? `Help me get this done:\n\n${body}\n\nThe source note is read-only. Do not edit, replace, or update it.` : '';
 }
 
 // Agent-solve: create a chat session server-side, kick off an agent run
@@ -4395,6 +4449,7 @@ async function _agentSolveNote(id) {
     fd.append('message', prompt);
     fd.append('session', sid);
     fd.append('mode', 'agent');
+    fd.append('disabled_tools', JSON.stringify(['manage_notes']));
     fetch(`${API_BASE}/api/chat_stream`, { method: 'POST', credentials: 'same-origin', body: fd })
       .then(async (res) => {
         if (!res.ok || !res.body) return;
@@ -4413,6 +4468,86 @@ async function _agentSolveNote(id) {
   }
 }
 
+// Per-item version of _agentSolveNote. Scoped to a single checklist item;
+// the note title (if any) is included as context, but only this one item's
+// text is the work the agent is asked to do. agent_session_id is set on the
+// PARENT note (latest-wins) so the Agent tag still surfaces the most recent
+// run from this note — same UX as a per-note solve.
+async function _agentSolveTodoItem(noteId, idx) {
+  const note = _notes.find(n => n.id === noteId);
+  if (!note || !Array.isArray(note.items)) return;
+  const item = note.items[idx];
+  const itemText = (item && (item.text || '').trim()) || '';
+  if (!itemText) {
+    uiModule.showToast('Nothing to solve — item is empty');
+    return;
+  }
+  const titleCtx = (note.title || '').trim();
+  const prompt = titleCtx
+    ? `Context (from note "${titleCtx}").\n\nHelp me with this todo: ${itemText}\n\nThe source note is read-only. Do not edit, replace, or update it.`
+    : `Help me with this todo: ${itemText}\n\nThe source note is read-only. Do not edit, replace, or update it.`;
+  try {
+    const dc = await (await fetch(`${API_BASE}/api/default-chat`, { credentials: 'same-origin' })).json();
+    if (!dc.endpoint_url || !dc.model) { uiModule.showError('No default chat model configured'); return; }
+
+    const label = itemText.slice(0, 40);
+    const csFd = new FormData();
+    csFd.append('name', 'Agent: ' + label);
+    csFd.append('endpoint_url', dc.endpoint_url);
+    csFd.append('model', dc.model);
+    if (dc.endpoint_id) csFd.append('endpoint_id', dc.endpoint_id);
+    csFd.append('skip_validation', 'true');
+    const csRes = await fetch(`${API_BASE}/api/session`, { method: 'POST', credentials: 'same-origin', body: csFd });
+    if (!csRes.ok) { uiModule.showError('Could not create agent session'); return; }
+    const sess = await csRes.json();
+    const sid = sess.id;
+    const sessionTitle = 'Agent: ' + label;
+
+    const n = _notes.find(x => x.id === noteId);
+    if (n) {
+      n.agent_session_id = sid;
+      if (Array.isArray(n.items) && n.items[idx]) {
+        n.items[idx].agent_session_id = sid;
+        n.items[idx].agent_session_title = sessionTitle;
+        n.items[idx].agent_status = 'running';
+        n.items[idx].agent_stream_completed_at = '';
+      }
+    }
+    _renderNotes();
+    _patchNote(noteId, { items: n && Array.isArray(n.items) ? n.items : note.items, agent_session_id: sid }).catch(() => {});
+
+    const fd = new FormData();
+    fd.append('message', prompt);
+    fd.append('session', sid);
+    fd.append('mode', 'agent');
+    fd.append('disabled_tools', JSON.stringify(['manage_notes']));
+    fetch(`${API_BASE}/api/chat_stream`, { method: 'POST', credentials: 'same-origin', body: fd })
+      .then(async (res) => {
+        if (!res.ok || !res.body) return;
+        const reader = res.body.getReader();
+        while (true) { const { done } = await reader.read(); if (done) break; }
+        if (window.sessionModule && window.sessionModule.markStreamComplete) {
+          try { window.sessionModule.markStreamComplete(sid); } catch {}
+        }
+        const doneNote = _notes.find(x => x.id === noteId);
+        if (doneNote && Array.isArray(doneNote.items) && doneNote.items[idx]) {
+          doneNote.agent_session_id = sid;
+          doneNote.items[idx].agent_session_id = sid;
+          doneNote.items[idx].agent_session_title = sessionTitle;
+          doneNote.items[idx].agent_status = 'stream_complete';
+          doneNote.items[idx].agent_stream_completed_at = new Date().toISOString();
+          _renderNotes();
+          _patchNote(noteId, { items: doneNote.items, agent_session_id: sid }).catch(() => {});
+        }
+      })
+      .catch(() => {});
+
+    uiModule.showToast('Agent working on this item — tap the Agent tag when ready');
+  } catch (e) {
+    uiModule.showError('Agent failed: ' + (e.message || e));
+  }
+}
+
 async function _copyNote(noteId, btnEl) {
   const note = _notes.find(n => n.id === noteId);
   if (!note) return false;

static/js/research/panel.js

@@ -366,20 +366,13 @@ function _buildPanelHTML() {
     <div class="modal-body research-pane-body" data-no-swipe-dismiss>
       <div class="research-new-job">
         <div style="display:flex;align-items:center;gap:8px;margin-bottom:2px;">
-          <h2 style="margin:0;padding:0;line-height:1;display:inline-flex;align-items:center;gap:6px;"><svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="var(--accent, var(--red))" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="flex-shrink:0;"><path d="M6 18h8"/><path d="M3 22h18"/><path d="M14 22a7 7 0 1 0 0-14h-1"/><path d="M9 14h2"/><path d="M9 12a2 2 0 0 1-2-2V6h4v4a2 2 0 0 1-2 2Z"/><path d="M12 6V3a1 1 0 0 0-1-1H9a1 1 0 0 0-1 1v3"/></svg>Research <span id="research-stats" class="memory-count" style="font-size:0.6em;opacity:0.6;font-weight:normal;position:relative;top:4px;"></span></h2>
+          <h2 style="margin:0;padding:0;line-height:1;display:inline-flex;align-items:center;gap:6px;"><svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="var(--accent, var(--red))" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="flex-shrink:0;"><path d="M6 18h8"/><path d="M3 22h18"/><path d="M14 22a7 7 0 1 0 0-14h-1"/><path d="M9 14h2"/><path d="M9 12a2 2 0 0 1-2-2V6h4v4a2 2 0 0 1-2 2Z"/><path d="M12 6V3a1 1 0 0 0-1-1H9a1 1 0 0 0-1 1v3"/></svg>Research <span id="research-stats" class="memory-count" style="font-size:0.6em;opacity:0.6;font-weight:normal"></span></h2>
         </div>
         <p class="memory-desc doclib-desc" style="margin-top:2px;display:flex;align-items:center;gap:6px;flex-wrap:wrap;">
           <span>Multi-step web research with an LLM-in-the-loop agent</span>
-          <span id="research-no-past-hint" style="display:none;font-size:11px;opacity:0.7;position:relative;top:-4px;">— past runs in <button type="button" class="research-library-link" style="background:none;border:none;padding:0;font:inherit;color:var(--accent, var(--red));cursor:pointer;text-decoration:underline;">Library, Research</button></span>
+          <span id="research-no-past-hint" style="display:none;font:inherit;opacity:1;position:static;">— past runs in <button type="button" class="research-library-link" style="background:none;border:none;padding:0;font:inherit;color:var(--accent, var(--red));cursor:pointer;text-decoration:underline;">Library, Research</button></span>
         </p>
         <textarea id="research-query" class="research-query" placeholder="${_pickResearchHint()}" rows="4"></textarea>
-        <div class="research-category-row" id="research-category-row">
-          <button class="research-cat active" data-cat="" title="LLM auto-detects the best format">Auto</button>
-          <button class="research-cat" data-cat="product">Product</button>
-          <button class="research-cat" data-cat="comparison">Compare</button>
-          <button class="research-cat" data-cat="howto">How-to</button>
-          <button class="research-cat" data-cat="factcheck">Fact-check</button>
-        </div>
         <button id="research-settings-toggle" class="research-settings-toggle${chevronCls}">
           <svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" style="vertical-align:-2px;margin-right:4px;opacity:0.85;flex-shrink:0;"><circle cx="12" cy="12" r="3"/><path d="M19.4 15a1.65 1.65 0 0 0 .33 1.82l.06.06a2 2 0 0 1 0 2.83 2 2 0 0 1-2.83 0l-.06-.06a1.65 1.65 0 0 0-1.82-.33 1.65 1.65 0 0 0-1 1.51V21a2 2 0 0 1-2 2 2 2 0 0 1-2-2v-.09A1.65 1.65 0 0 0 9 19.4a1.65 1.65 0 0 0-1.82.33l-.06.06a2 2 0 0 1-2.83 0 2 2 0 0 1 0-2.83l.06-.06a1.65 1.65 0 0 0 .33-1.82 1.65 1.65 0 0 0-1.51-1H3a2 2 0 0 1-2-2 2 2 0 0 1 2-2h.09A1.65 1.65 0 0 0 4.6 9a1.65 1.65 0 0 0-.33-1.82l-.06-.06a2 2 0 0 1 0-2.83 2 2 0 0 1 2.83 0l.06.06a1.65 1.65 0 0 0 1.82.33H9a1.65 1.65 0 0 0 1-1.51V3a2 2 0 0 1 2-2 2 2 0 0 1 2 2v.09a1.65 1.65 0 0 0 1 1.51 1.65 1.65 0 0 0 1.82-.33l.06-.06a2 2 0 0 1 2.83 0 2 2 0 0 1 0 2.83l-.06.06a1.65 1.65 0 0 0-.33 1.82V9a1.65 1.65 0 0 0 1.51 1H21a2 2 0 0 1 2 2 2 2 0 0 1-2 2h-.09a1.65 1.65 0 0 0-1.51 1z"/></svg>Settings<span class="research-settings-chevron">${_chevronIcon}</span>
         </button>
@@ -787,6 +780,21 @@ function _renderJobs() {
       +   '<span class="research-section-dot' + (dotPulse ? ' pulsing' : '') + '" style="background:' + dotColor + ';"></span>'
       +   '<svg class="research-section-chevron" width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round"><polyline points="6 9 12 15 18 9"/></svg>'
       + '</span>';
+    if (key === 'past') {
+      const hint = document.createElement('span');
+      hint.className = 'research-library-hint';
+      hint.innerHTML = '<span>Multi-step web research with an LLM-in-the-loop agent</span> <button type="button" class="research-library-link">Library, Research</button>';
+      hint.querySelector('.research-library-link').addEventListener('click', (e) => {
+        e.stopPropagation();
+        // Close the research panel first so the Library opens ABOVE it on mobile
+        // (otherwise it stacks under the full-screen panel).
+        closePanel();
+        if (window.documentModule && window.documentModule.openLibrary) {
+          window.documentModule.openLibrary({ tab: 'research' });
+        }
+      });
+      header.appendChild(hint);
+    }
     header.addEventListener('click', () => {
       const nowCollapsed = sec.classList.toggle('collapsed');
       if (nowCollapsed) _collapsedSections.add(key); else _collapsedSections.delete(key);
@@ -803,27 +811,6 @@ function _renderJobs() {
     });
     const body = document.createElement('div');
     body.className = 'research-section-body';
-    // Past Research header: link goes INLINE next to the title instead
-    // of on a second row. Append it to the title span as a small chip.
-    if (key === 'past') {
-      const titleEl = header.querySelector('.research-section-title');
-      if (titleEl) {
-        const hint = document.createElement('span');
-        hint.className = 'research-library-hint research-library-hint-inline';
-        hint.style.cssText = 'margin-left:8px;font-size:10.5px;opacity:0.65;font-weight:normal;';
-        hint.innerHTML = '— all in <button type="button" class="research-library-link" style="background:none;border:none;padding:0;font:inherit;color:var(--accent, var(--red));cursor:pointer;text-decoration:underline;">Library, Research</button>';
-        hint.querySelector('.research-library-link').addEventListener('click', (e) => {
-          e.stopPropagation();
-          // Close the research panel first so the Library opens ABOVE it on mobile
-          // (otherwise it stacks under the full-screen panel).
-          closePanel();
-          if (window.documentModule && window.documentModule.openLibrary) {
-            window.documentModule.openLibrary({ tab: 'research' });
-          }
-        });
-        titleEl.appendChild(hint);
-      }
-    }
     arr.forEach(j => body.appendChild(_buildJobCard(j)));
     sec.appendChild(header);
     sec.appendChild(body);
@@ -1014,9 +1001,9 @@ function _buildJobCard(job) {
       </div>
       ${failNote}
       <div class="research-job-actions">
-        <button class="research-job-action" data-action="copy" title="Copy report to clipboard">${_copyIcon}</button>
-        <button class="research-job-action" data-action="chat" title="Open follow-up chat with this research as context">${_chatIcon} Discuss</button>
         <button class="research-job-action research-job-action-report" data-action="report" title="Visual report">${_externalIcon} Visual Report</button>
+        <button class="research-job-action" data-action="chat" title="Open follow-up chat with this research as context">${_chatIcon} Discuss</button>
+        <button class="research-job-action research-job-action-dim" data-action="copy" title="Copy report to clipboard">${_copyIcon}</button>
         <button class="research-job-action research-job-action-dim" data-action="dismiss" title="Clear from list">${_cancelIcon}</button>
         <button class="research-job-action research-job-action-dim" data-action="delete" title="Delete from disk">${_trashIcon} Delete</button>
       </div>

static/js/sidebar-layout.js

@@ -103,7 +103,6 @@ export function initSidebarLayout(Storage, opts) {
   });
 
   // Hamburger cycles: full sidebar → mini → off → full
-  // Shift-click swaps sidebar side
   let _userToggledSidebar = false;
   let _wasAutoCollapsed = false;
 
@@ -122,8 +121,7 @@ export function initSidebarLayout(Storage, opts) {
     if (window.innerWidth < 768 && cc && cc.classList.contains('compare-active')) return;
     _userToggledSidebar = true;
     // Optionally place the sidebar on a specific edge (the swipe gesture passes
-    // the direction). Persist it + re-anchor the doc panel, same as a
-    // shift-click on the hamburger.
+    // the direction). Persist it + re-anchor the doc panel.
     if (side === 'left' || side === 'right') {
       const wantRight = side === 'right';
       if (sidebar.classList.contains('right-side') !== wantRight) {
@@ -143,13 +141,6 @@ export function initSidebarLayout(Storage, opts) {
     hamburgerBtn.addEventListener('click', (e) => {
       e.stopPropagation();
       const sidebar = document.getElementById('sidebar');
-      if (e.shiftKey) {
-        sidebar.classList.toggle('right-side');
-        Storage.set(Storage.KEYS.SIDEBAR_SIDE, sidebar.classList.contains('right-side') ? 'right' : 'left');
-        syncRailSide();
-        if (documentModule && documentModule.swapSide) documentModule.swapSide();
-        return;
-      }
 
       _userToggledSidebar = true;
       const isSidebarVisible = !sidebar.classList.contains('hidden');

static/js/tasks.js

@@ -17,9 +17,16 @@ let _tasksFetched = false;   // first-fetch sentinel — `false` → show loadin
 let _escHandler = null;
 let _viewingRuns = null; // task id when viewing run history
 let _clockInterval = null;
+let _taskFailurePending = false;
 
 const DAYS_OF_WEEK = ['Monday', 'Tuesday', 'Wednesday', 'Thursday', 'Friday', 'Saturday', 'Sunday'];
 
+function _setTaskFailurePending(active) {
+  _taskFailurePending = !!active;
+  document.getElementById('tool-tasks-btn')?.classList.toggle('task-failure-pending', _taskFailurePending);
+  document.getElementById('rail-tasks')?.classList.toggle('task-failure-pending', _taskFailurePending);
+}
+
 // ---- API ----
 
 async function _fetchTasks() {
@@ -2238,6 +2245,9 @@ function _renderActivityEntry(entry) {
     status = _classifyResult(entry.result);
   }
   const statusDot = `<span class="task-log-status task-log-status-${status}" title="${status}"></span>`;
+  const failedTag = status === 'error'
+    ? '<span class="task-log-failed-tag">(failed)</span>'
+    : '';
   // Render the result through markdown so code blocks, lists, links look right.
   let resultHtml;
   const _isRunning = entry.status === 'running' || entry.status === 'queued';
@@ -2361,7 +2371,7 @@ function _renderActivityEntry(entry) {
       <div class="task-log-row-head">
         ${statusDot}
         <span class="task-log-task-icon">${_taskIcon({ action: entry.action, task_type: entry.kind })}</span>
-        <span class="task-log-name">${_escHtml(entry.taskName)}</span>${_taskAiMark(entry)}
+        <span class="task-log-name">${_escHtml(entry.taskName)}</span>${failedTag}${_taskAiMark(entry)}
         ${repeatBadge}
         <span style="flex:1"></span>
         ${rightHtml}
@@ -2502,8 +2512,11 @@ function _renderMainView() {
 
 export function openTasks(focusId, opts) {
   const o = opts || {};
+  const openActivityForFailure = _taskFailurePending && !focusId && o.filter === undefined;
+  _setTaskFailurePending(false);
   if (_open) {
     // Already open — just focus the requested task / apply filter.
+    if (openActivityForFailure) _switchTab('activity');
     if (o.filter !== undefined) { _taskFilter = o.filter; _renderList(); }
     if (focusId) _focusTask(focusId);
     return;
@@ -2610,7 +2623,7 @@ export function openTasks(focusId, opts) {
   // of an empty modal-body that fills in after the fetch resolves — that delay
   // was visible as a "flicker" right after opening.
   _activeTab = 'tasks';
-  _switchTab('tasks');
+  _switchTab(openActivityForFailure ? 'activity' : 'tasks');
   _fetchTasks().then(() => {
     // Re-render so the list swaps the Loading row for real cards.
     _renderList();
@@ -2704,7 +2717,13 @@ async function _pollTaskNotifications() {
       const msg = `Task ${ok ? 'finished' : 'failed'}: ${n.task_name}`;
       if (!uiModule) continue;
       if (ok) uiModule.showToast(msg, { duration: 5000 });
-      else uiModule.showError(msg);
+      else {
+        _setTaskFailurePending(true);
+        uiModule.showError(msg);
+        if (_open && document.querySelector('.tasks-tab.active[data-tab="activity"]')) {
+          _renderActivityView();
+        }
+      }
     }
   } catch (e) {
     // Silently ignore — server may be unreachable

static/js/theme.js

@@ -26,7 +26,7 @@ export const THEMES = {
   gpt:        { bg:'#212121', fg:'#ececec', panel:'#171717', border:'#424242', red:'#949494',
                 advanced: { sendBtnBg: '#949494', sendBtnHover: '#7f7f7f',
                             userBubbleBg: '#2f2f2f', aiBubbleBg: '#171717',
-                            inputBg: '#2f2f2f' } },
+                            inputBg: '#2f2f2f', brandColor: '#ffffff', brandMixTo: '#ffffff' } },
   claude:     { bg:'#262624', fg:'#f5f4f0', panel:'#30302e', border:'#4a4a47', red:'#c6613f' },
   cute:       { bg:'#fff0f5', fg:'#d4608a', panel:'#fff8fa', border:'#f0c0d0', red:'#ff6b9d' },
 };
@@ -184,6 +184,7 @@ const ADV_KEYS = [
   { key: 'bubbleBorder',       css: '--bubble-border',     label: 'Border Chat Bubble', group: 'Chat Bubbles' },
   { key: 'sidebarBg',          css: '--sidebar-bg',        label: 'Sidebar Bg',       group: 'Sidebar' },
   { key: 'brandColor',         css: '--brand-color',       label: 'Odysseus Logo',    group: 'Sidebar' },
+  { key: 'brandMixTo',         css: '--brand-mix-to',      label: 'Logo Gradient End', group: 'Sidebar' },
   { key: 'hamburgerColor',     css: '--hamburger-color',   label: 'Hamburger Menu',   group: 'Sidebar' },
   { key: 'inputBg',            css: '--input-bg',          label: 'Input Bg',         group: 'Chat Input / Prompt Area' },
   { key: 'inputBorder',        css: '--input-border',      label: 'Input Border',     group: 'Chat Input / Prompt Area' },
@@ -203,6 +204,7 @@ function computeAdvancedDefaults(colors) {
     bubbleBorder: colors.border,
     sidebarBg: colors.panel,
     brandColor: red,
+    brandMixTo: colors.fg,
     hamburgerColor: colors.fg,
     inputBg: colors.panel,
     inputBorder: colors.border,

tests/test_aux_llm_owner_scope.py

@@ -43,7 +43,8 @@ def test_background_session_sort_uses_owner_task_endpoint():
 def test_scheduler_fallbacks_and_research_headers_are_owner_scoped():
     src = _src("src/task_scheduler.py")
 
-    assert "resolve_utility_fallback_candidates(owner=task.owner or None)" in src
+    assert "resolve_task_candidates(" in src
+    assert "owner=task.owner or None" in src
     assert 'resolve_endpoint(\n                    "research",' in src
     assert "owner=task.owner or None" in src
     assert "headers_from_resolver = False" in src

tests/test_builtin_actions_owner_scope.py

@@ -51,23 +51,19 @@ class _Db:
         self.closed = True
 
 
-def _resolver_spy(monkeypatch, utility_result=("", "", {}), default_result=("http://llm", "model", {})):
-    from src import endpoint_resolver
+def _resolver_spy(monkeypatch, candidates=None):
+    from src import task_endpoint
 
     calls = []
-    fallback_calls = []
 
-    def fake_resolve(kind, *args, **kwargs):
-        calls.append((kind, kwargs.get("owner")))
-        return utility_result if kind == "utility" else default_result
+    def fake_candidates(*args, **kwargs):
+        calls.append(kwargs.get("owner"))
+        if candidates is None:
+            return [("http://llm", "model", {})]
+        return list(candidates)
 
-    def fake_fallbacks(*args, **kwargs):
-        fallback_calls.append(kwargs.get("owner"))
-        return []
-
-    monkeypatch.setattr(endpoint_resolver, "resolve_endpoint", fake_resolve)
-    monkeypatch.setattr(endpoint_resolver, "resolve_utility_fallback_candidates", fake_fallbacks)
-    return calls, fallback_calls
+    monkeypatch.setattr(task_endpoint, "resolve_task_candidates", fake_candidates)
+    return calls
 
 
 @pytest.mark.asyncio
@@ -88,7 +84,7 @@ async def test_classify_events_resolves_llm_for_task_owner(monkeypatch):
         location="",
     )
     db = _Db({FakeCalendarEvent: [event]})
-    calls, _fallback_calls = _resolver_spy(monkeypatch, utility_result=("http://llm", "model", {}))
+    calls = _resolver_spy(monkeypatch)
 
     monkeypatch.setattr(database, "CalendarEvent", FakeCalendarEvent)
     monkeypatch.setattr(database, "SessionLocal", lambda: db)
@@ -97,7 +93,7 @@ async def test_classify_events_resolves_llm_for_task_owner(monkeypatch):
 
     assert ok is True
     assert "Scanned 1 upcoming event" in message
-    assert calls == [("utility", "alice")]
+    assert calls == ["alice"]
     assert db.closed is True
 
 
@@ -122,7 +118,7 @@ async def test_learn_sender_signatures_resolves_llm_for_task_owner(monkeypatch):
         def logout(self):
             return None
 
-    calls, _fallback_calls = _resolver_spy(monkeypatch, utility_result=("", "", {}), default_result=("", "", {}))
+    calls = _resolver_spy(monkeypatch, candidates=[])
     imap_owners = []
 
     def fake_imap_connect(_account_id=None, owner=""):
@@ -135,14 +131,14 @@ async def test_learn_sender_signatures_resolves_llm_for_task_owner(monkeypatch):
 
     assert ok is False
     assert message == "No LLM endpoint available"
-    assert calls == [("utility", "alice"), ("default", "alice")]
+    assert calls == ["alice"]
     assert imap_owners == ["alice"]
 
 
 @pytest.mark.asyncio
 async def test_learn_sender_signatures_writes_owner_scoped_cache(monkeypatch, tmp_path):
     from routes import email_helpers
-    from src import endpoint_resolver, llm_core
+    from src import llm_core, task_endpoint
     from src.builtin_actions import action_learn_sender_signatures
 
     db_path = tmp_path / "scheduled_emails.db"
@@ -205,15 +201,15 @@ async def test_learn_sender_signatures_writes_owner_scoped_cache(monkeypatch, tm
 
     monkeypatch.setattr(email_helpers, "_imap_connect", fake_imap_connect)
     monkeypatch.setattr(
-        endpoint_resolver,
-        "resolve_endpoint",
-        lambda kind, *args, **kwargs: ("http://llm", "alice-model", {}),
+        task_endpoint,
+        "resolve_task_candidates",
+        lambda *args, **kwargs: [("http://llm", "alice-model", {})],
     )
 
-    async def fake_llm_call_async(**_kwargs):
+    async def fake_llm_call_async(_candidates, **_kwargs):
         return "Writer Example\nExample Co.\nwriter@example.com"
 
-    monkeypatch.setattr(llm_core, "llm_call_async", fake_llm_call_async)
+    monkeypatch.setattr(llm_core, "llm_call_async_with_fallback", fake_llm_call_async)
 
     message, ok = await action_learn_sender_signatures("alice")
 
@@ -253,7 +249,7 @@ async def test_check_email_urgency_resolves_llm_candidates_for_task_owner(monkey
         from_address = _Column()
 
     db = _Db({FakeEmailAccount: []})
-    calls, fallback_calls = _resolver_spy(monkeypatch, utility_result=("http://llm", "model", {}))
+    calls = _resolver_spy(monkeypatch)
 
     monkeypatch.chdir(tmp_path)
     monkeypatch.setattr(database, "EmailAccount", FakeEmailAccount)
@@ -262,6 +258,5 @@ async def test_check_email_urgency_resolves_llm_candidates_for_task_owner(monkey
     with pytest.raises(TaskNoop, match="no email accounts configured"):
         await action_check_email_urgency("alice")
 
-    assert calls == [("utility", "alice")]
-    assert fallback_calls == ["alice"]
+    assert calls == ["alice"]
     assert db.closed is True

tests/test_builtin_memory_consolidation.py

@@ -29,8 +29,8 @@ def _read_memories(data_dir):
 @pytest.mark.asyncio
 async def test_consolidate_memory_empty_owner_treats_each_owner_separately(monkeypatch, tmp_path):
     from src import constants
-    from src import endpoint_resolver
     from src import llm_core
+    from src import task_endpoint
     action_consolidate_memory = _import_consolidate_action()
 
     long_alice_text = "Alice private project context. " + ("A" * 2200)
@@ -44,11 +44,15 @@ async def test_consolidate_memory_empty_owner_treats_each_owner_separately(monke
         ],
     )
     monkeypatch.setattr(constants, "DATA_DIR", str(data_dir))
-    monkeypatch.setattr(endpoint_resolver, "resolve_endpoint", lambda *args, **kwargs: ("http://llm", "model", {}))
+    monkeypatch.setattr(
+        task_endpoint,
+        "resolve_task_candidates",
+        lambda *args, **kwargs: [("http://llm", "model", {})],
+    )
 
     prompts = []
 
-    async def fake_llm_call_async(**kwargs):
+    async def fake_llm_call_async(_candidates, **kwargs):
         prompt = kwargs["messages"][0]["content"]
         prompts.append(prompt)
         if "alice-long" in prompt:
@@ -71,7 +75,7 @@ async def test_consolidate_memory_empty_owner_treats_each_owner_separately(monke
             }
         )
 
-    monkeypatch.setattr(llm_core, "llm_call_async", fake_llm_call_async)
+    monkeypatch.setattr(llm_core, "llm_call_async_with_fallback", fake_llm_call_async)
 
     message, ok = await action_consolidate_memory("")
 

tests/test_consolidate_memory_explicit_drops.py

@@ -29,24 +29,24 @@ class _FakeMM:
 
 def test_omitted_memory_survives_only_explicit_drop(monkeypatch):
     import src.memory
-    import src.endpoint_resolver
     import src.llm_core
+    import src.task_endpoint
 
     _FakeMM.saved = None
     monkeypatch.setattr(src.memory, "MemoryManager", _FakeMM)
     monkeypatch.setattr(
-        src.endpoint_resolver, "resolve_endpoint",
-        lambda kind, owner=None: ("http://x/v1", "model", {}),
+        src.task_endpoint, "resolve_task_candidates",
+        lambda owner=None: [("http://x/v1", "model", {})],
     )
 
-    async def fake_llm(**kwargs):
+    async def fake_llm(_candidates, **kwargs):
         # Model keeps 'a', drops 'b', and OMITS 'c' entirely.
         return json.dumps({
             "keep": [{"id": "a", "text": "Likes dark roast coffee", "category": "preference"}],
             "drop": [{"id": "b", "reason": "duplicate of a"}],
         })
 
-    monkeypatch.setattr(src.llm_core, "llm_call_async", fake_llm)
+    monkeypatch.setattr(src.llm_core, "llm_call_async_with_fallback", fake_llm)
 
     msg, ok = asyncio.run(ba.action_consolidate_memory("alice"))
 

tests/test_cookbook_cpu_only_serve.py

@@ -16,6 +16,7 @@ from pathlib import Path
 
 SRC = Path(__file__).resolve().parent.parent / "static/js/cookbook.js"
 SERVE_SRC = Path(__file__).resolve().parent.parent / "static/js/cookbookServe.js"
+ROUTES_SRC = Path(__file__).resolve().parent.parent / "routes/cookbook_routes.py"
 
 
 def test_cpu_only_drops_gpu_only_flags():
@@ -54,3 +55,32 @@ def test_windows_diffusers_uses_python_not_python3():
     assert "const diffusersPy = _isWindows() ? 'python' : _py3Bin;" in text
     assert "cmd += `${diffusersPy} scripts/diffusion_server.py" in text
     assert "cmd += `python3 scripts/diffusion_server.py" not in text
+
+
+def test_vllm_blank_swap_omits_swap_space_flag():
+    text = SRC.read_text(encoding="utf-8")
+
+    assert "const _swapRaw = (f.swap ?? '').toString().trim().toLowerCase();" in text
+    assert "['0', 'off', 'none', 'false'].includes(_swapRaw)" in text
+    assert "if (_swapRaw && !['0', 'off', 'none', 'false'].includes(_swapRaw)) cmd += ` --swap-space ${_swapRaw}`;" in text
+
+
+def test_serve_preflight_uses_selected_server_not_stale_env_host():
+    text = SERVE_SRC.read_text(encoding="utf-8")
+
+    assert "function _selectedServeTarget(panel) {" in text
+    assert "const _hostStr = launchTarget.host || '';" in text
+    assert "(t.remoteHost || '') === _hostStr" in text
+    assert "const _probeHost = (launchTarget.host || '').trim();" in text
+    assert "const _portHost = (launchTarget.host || '').trim();" in text
+
+
+def test_vllm_route_strips_swap_space_when_runtime_rejects_it():
+    text = ROUTES_SRC.read_text(encoding="utf-8")
+
+    assert "Setting vLLM --swap-space 0 so the runtime does not reserve CPU swap per GPU." in text
+    assert "vLLM serve does not expose --swap-space; removing the flag and patching the runtime default to 0." in text
+    assert "ODYSSEUS_VLLM_HELP_CMD" in text
+    assert "print(shlex.join(parts[:serve_i + 1] + [\"--help\"]))" in text
+    assert "eval \"$ODYSSEUS_VLLM_HELP_CMD\" 2>&1 | grep -q -- \"--swap-space\"" in text
+    assert "eval \"$ODYSSEUS_SERVE_CMD\"" in text

tests/test_cookbook_helpers.py

@@ -348,7 +348,7 @@ def test_serve_pip_install_normalizes_llama_cpp_alias_and_adds_wheel_index():
     src = (pathlib.Path(__file__).resolve().parent.parent
         / "routes" / "cookbook_routes.py").read_text(encoding="utf-8")
 
-    assert "re.sub(r\"(?<![A-Za-z0-9_.-])llama_cpp(?![A-Za-z0-9_.-])\", \"llama-cpp-python[server]\", req.cmd)" in src
+    assert "re.sub(r\"(?<![A-Za-z0-9_.\\-/])llama_cpp(?![A-Za-z0-9_.\\-/])\", \"llama-cpp-python[server]\", req.cmd)" in src
     assert "if \"llama-cpp-python\" in req.cmd and \"--extra-index-url\" not in req.cmd:" in src
     assert "https://abetlen.github.io/llama-cpp-python/whl/cpu" in src
 
@@ -626,7 +626,7 @@ def test_llama_cpp_linux_bootstrap_prefers_rocm_before_cuda():
     script = "\n".join(runner_lines)
 
     assert "mkdir -p ~/bin" in script
-    assert script.index("mkdir -p ~/bin") < script.index("cd ~/llama.cpp && rm -rf build")
+    assert script.index("mkdir -p ~/bin") < script.index("cd ~/llama.cpp")
     assert 'command -v hipconfig &>/dev/null || [ -d /opt/rocm ] || [ -n "$ROCM_PATH" ] || [ -n "$HIP_PATH" ]' in script
     assert 'cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_HIP=ON' in script
     assert 'cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON' in script
@@ -676,7 +676,7 @@ def test_llama_cpp_linux_bootstrap_nvcc_without_cudart_warns_and_falls_back():
     # outer else that handles no-GPU-toolchain). Verify it appears at least once
     # before the outer "no HIP/CUDA toolchain" warning.
     cpu_cmake = 'cmake -B build -DCMAKE_BUILD_TYPE=Release &&'
-    no_toolchain_warn = 'WARNING: no HIP/CUDA toolchain found'
+    no_toolchain_warn = 'WARNING: no HIP/CUDA/Vulkan toolchain found'
     assert cpu_cmake in script
     assert script.index(cpu_cmake) < script.index(no_toolchain_warn)
 
@@ -693,8 +693,8 @@ def test_llama_cpp_linux_bootstrap_keeps_cpu_fallback_when_no_gpu_toolchain():
     _append_llama_cpp_linux_accel_build_lines(runner_lines)
     script = "\n".join(runner_lines)
 
-    assert 'WARNING: no HIP/CUDA toolchain found — building llama-server for CPU only.' in script
-    assert 'Install ROCm for AMD GPUs or vLLM/CUDA tooling for NVIDIA' in script
+    assert 'WARNING: no HIP/CUDA/Vulkan toolchain found — building llama-server for CPU only.' in script
+    assert 'Install Vulkan (libvulkan-dev) / ROCm for AMD GPUs or CUDA tooling for NVIDIA' in script
 
 
 def test_llama_cpp_rebuild_cmd_clears_cached_build_paths():

tests/test_cookbook_same_host_server_profiles_js.py

@@ -50,14 +50,14 @@ def test_serve_launch_preflights_use_selected_target_and_port():
     assert "if (launchTarget.port) _probeParams.set('ssh_port', launchTarget.port);" in SERVE
     assert "const _portHost = (launchTarget.host || '').trim();" in SERVE
     assert "StrictHostKeyChecking=no ${_sshPrefix(launchTarget.port)}${_portHost}" in SERVE
-    assert "let serveHost = launchTarget.host || '';" in SERVE
+    assert "const serveHost = launchTarget.host || '';" in SERVE
     assert SERVE.index(launch_target) < SERVE.index("const _runningMod = await import('./cookbookRunning.js');")
 
 
 def test_running_tab_resolves_profile_key_not_first_host():
-    assert "_serverByVal(_envState.remoteServerKey || _tHost)" in RUNNING
+    assert "_serverByVal(_targetKey)" in RUNNING
     assert "_serverByVal(_envState.remoteServerKey || _host)" in RUNNING
-    assert "_serverByVal(_envState.remoteServerKey || host)" in RUNNING
+    assert "_serverByVal(savedKey)" in RUNNING
     assert "_serverByVal = shared._serverByVal;" in RUNNING
     assert "_selectedServer = shared._selectedServer;" in RUNNING
 

tests/test_copy_message_strips_thinking_js.py

@@ -126,6 +126,27 @@ def test_plain_reply_copy_text_is_unchanged(node_available):
     assert out["content"] == raw
 
 
+def test_minimax_namespaced_thinking_is_extracted(node_available):
+    raw = (
+        '<mm:think>The user said "idk" - just casual.</mm:think>'
+        "Haha fair. Well, I'm here whenever you figure it out."
+    )
+    out = _extract_thinking_blocks(raw)
+
+    assert out["thinkingBlocks"] == ['The user said "idk" - just casual.']
+    assert out["content"] == "Haha fair. Well, I'm here whenever you figure it out."
+    assert "mm:think" not in out["content"]
+
+
+def test_minimax_orphan_closing_tag_drops_leaked_reasoning(node_available):
+    raw = "</mm:think>Hi! What can I do for you?"
+    out = _extract_thinking_blocks(raw)
+
+    assert out["thinkingBlocks"] == []
+    assert out["content"] == "Hi! What can I do for you?"
+    assert "mm:think" not in out["content"]
+
+
 def test_thinking_only_message_yields_empty_content(node_available):
     # The copy handler falls back to the raw text in this case so the button
     # still copies something for turns interrupted mid-thinking.

tests/test_docker_devops_hardening.py

@@ -59,8 +59,8 @@ def test_docker_entrypoint_does_not_resolve_root_commands_from_app_local_path():
     path_export = script.index('export PATH="/app/.local/bin:$PATH"')
     gosu_capture = script.index('GOSU_BIN="$(command -v gosu)"')
     python_capture = script.index('PYTHON_BIN="$(command -v python)"')
-    setup_call = script.index('"$GOSU_BIN" "$PUID:$PGID" "$PYTHON_BIN" /app/setup.py')
-    final_exec = script.index('exec "$GOSU_BIN" "$PUID:$PGID" "$@"')
+    setup_call = script.index('"$GOSU_BIN" "$ODY_USER" "$PYTHON_BIN" /app/setup.py')
+    final_exec = script.index('exec "$GOSU_BIN" "$ODY_USER" "$@"')
 
     assert gosu_capture < path_export < setup_call
     assert python_capture < path_export < setup_call

tests/test_fenced_example_not_executed_for_native_models.py

@@ -221,6 +221,60 @@ def test_skip_fenced_still_recovers_xml_invoke_markup():
     assert "latest python release" in blocks[0].content
 
 
+def test_stepfun_native_tool_tokens_are_executed_even_when_fenced_fallback_is_skipped():
+    leaked = (
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>web_search<｜tool▁sep｜>"
+        '{"query":"Sweden news today"}'
+        "<｜tool▁call▁end｜>"
+        "<｜tool▁calls▁end｜>"
+    )
+    blocks = parse_tool_blocks(leaked, skip_fenced=True)
+    assert len(blocks) == 1
+    assert blocks[0].tool_type == "web_search"
+    assert "Sweden news today" in blocks[0].content
+    assert strip_tool_blocks(leaked, skip_fenced=True) == ""
+
+
+def test_stepfun_native_tool_tokens_accept_plain_web_query():
+    leaked = (
+        "<｜tool▁call▁begin｜>web_search<｜tool▁sep｜>"
+        "Sweden news today"
+        "<｜tool▁call▁end｜>"
+    )
+    blocks = parse_tool_blocks(leaked, skip_fenced=True)
+    assert len(blocks) == 1
+    assert blocks[0].tool_type == "web_search"
+    assert "Sweden news today" in blocks[0].content
+
+
+def test_skip_fenced_still_recovers_direct_xml_tool_markup():
+    leaked = (
+        "I'll search now.\n"
+        "<tool_call><web_search>News in Sweden today 2026-06-22</web_search></tool_call>"
+    )
+    blocks = parse_tool_blocks(leaked, skip_fenced=True)
+    assert len(blocks) == 1
+    assert blocks[0].tool_type == "web_search"
+    assert "News in Sweden today 2026-06-22" in blocks[0].content
+    assert strip_tool_blocks(leaked, skip_fenced=True) == "I'll search now."
+
+
+def test_skip_fenced_recovers_direct_xml_tool_markup_with_unclosed_wrapper():
+    leaked = (
+        "I'll search now.\n"
+        "<tool_call>\n"
+        "<web_search>\n"
+        "Sweden news today 2026-06-22\n"
+        "</web_search>"
+    )
+    blocks = parse_tool_blocks(leaked, skip_fenced=True)
+    assert len(blocks) == 1
+    assert blocks[0].tool_type == "web_search"
+    assert "Sweden news today 2026-06-22" in blocks[0].content
+    assert strip_tool_blocks(leaked, skip_fenced=True) == "I'll search now."
+
+
 def test_skip_fenced_still_recovers_dsml_markup():
     dsml = (
         "Let me search for that.\n"

tests/test_gpu_compose_standalone.py

@@ -124,9 +124,9 @@ def test_nvidia_odysseus_adds_only_overlay(base):
         {"driver": "nvidia", "count": "all", "capabilities": ["gpu"]}
     ]
 
-    # No AMD-only keys leaked in.
+    # Base Docker socket group is preserved; no AMD-only keys leaked in.
     assert "devices" not in svc
-    assert "group_add" not in svc
+    assert svc["group_add"] == base_svc["group_add"]
 
 
 def test_amd_odysseus_adds_only_overlay(base):
@@ -137,11 +137,10 @@ def test_amd_odysseus_adds_only_overlay(base):
     # Environment is unchanged from base for AMD.
     assert svc["environment"] == base_svc["environment"]
 
-    # devices and group_add are new and match the overlay exactly.
+    # devices are new; group_add preserves the base Docker group and appends AMD groups.
     assert "devices" not in base_svc
-    assert "group_add" not in base_svc
     assert svc["devices"] == ["/dev/kfd", "/dev/dri"]
-    assert svc["group_add"] == ["video", "${RENDER_GID:-render}"]
+    assert svc["group_add"] == base_svc["group_add"] + ["video", "${RENDER_GID:-render}"]
 
     # No NVIDIA-only keys leaked in.
     assert "deploy" not in svc

tests/test_tool_support_heuristic.py

@@ -18,7 +18,7 @@ def _compute_is_api_model(model: str, endpoint_url: str, endpoint_supports=None)
     model_supports_tools = any(kw in model_lc for kw in (
         "gpt-4", "gpt-5", "gpt-o", "claude", "gemini", "gemma",
         "qwen3", "qwen2.5", "mixtral", "mistral", "llama-3.1", "llama-3.2",
-        "llama-3.3", "llama-4",
+        "llama-3.3", "llama-4", "llama3.1", "llama3.2", "llama3.3", "llama4",
         "minimax", "kimi", "yi-", "phi-3", "phi-4", "command-r",
         "glm-4", "internlm", "hermes",
         "deepseek-v", "deepseek-chat",

tests/test_upload_multifile.py

@@ -19,7 +19,12 @@ from pathlib import Path
 
 import pytest
 from fastapi import APIRouter
+from sqlalchemy import create_engine
+from sqlalchemy.orm import sessionmaker
+from sqlalchemy.pool import NullPool
 
+import core.database as cdb
+from core.database import GalleryImage
 from src.upload_handler import count_recent_uploads, UploadHandler
 import routes.upload_routes as up
 
@@ -82,6 +87,10 @@ def _files(n):
     return [types.SimpleNamespace(filename=f"f{i}.txt") for i in range(n)]
 
 
+def _image_upload(name="photo.png", content=b"not really png but enough for route metadata"):
+    return types.SimpleNamespace(filename=name, file=io.BytesIO(content))
+
+
 @pytest.fixture(autouse=True)
 def _reset_router(monkeypatch):
     # Module-level router accumulates routes across setup calls; reset it.
@@ -163,3 +172,64 @@ def test_six_file_batch_is_not_rate_limited(tmp_path):
         assert meta and meta.get("id")
         saved += 1
     assert saved == 6
+
+
+async def test_chat_image_upload_is_added_to_gallery(tmp_path, monkeypatch):
+    engine = create_engine(
+        f"sqlite:///{tmp_path / 'gallery.db'}",
+        connect_args={"check_same_thread": False},
+        poolclass=NullPool,
+    )
+    cdb.Base.metadata.create_all(engine)
+    TestingSession = sessionmaker(bind=engine, autoflush=False, autocommit=False)
+    gallery_dir = tmp_path / "generated_images"
+
+    monkeypatch.setattr(up, "SessionLocal", TestingSession)
+    monkeypatch.setattr(up, "GENERATED_IMAGES_DIR", str(gallery_dir))
+
+    h = UploadHandler(base_dir=str(tmp_path), upload_dir=str(tmp_path / "uploads"))
+    up.setup_upload_routes(h)
+    endpoint = _endpoint(up.router)
+
+    result = await endpoint(_request(user="alice"), [_image_upload()])
+    uploaded = result["files"][0]
+
+    assert uploaded["gallery_id"]
+    db = TestingSession()
+    try:
+        image = db.query(GalleryImage).filter(GalleryImage.id == uploaded["gallery_id"]).one()
+        assert image.owner == "alice"
+        assert image.model == "chat-upload"
+        assert image.prompt == "photo.png"
+        assert image.file_hash == uploaded["hash"]
+        assert (gallery_dir / image.filename).exists()
+    finally:
+        db.close()
+
+
+async def test_non_image_chat_upload_is_not_added_to_gallery(tmp_path, monkeypatch):
+    engine = create_engine(
+        f"sqlite:///{tmp_path / 'gallery.db'}",
+        connect_args={"check_same_thread": False},
+        poolclass=NullPool,
+    )
+    cdb.Base.metadata.create_all(engine)
+    TestingSession = sessionmaker(bind=engine, autoflush=False, autocommit=False)
+    monkeypatch.setattr(up, "SessionLocal", TestingSession)
+    monkeypatch.setattr(up, "GENERATED_IMAGES_DIR", str(tmp_path / "generated_images"))
+
+    h = UploadHandler(base_dir=str(tmp_path), upload_dir=str(tmp_path / "uploads"))
+    up.setup_upload_routes(h)
+    endpoint = _endpoint(up.router)
+
+    result = await endpoint(_request(user="alice"), [types.SimpleNamespace(
+        filename="notes.txt",
+        file=io.BytesIO(b"plain text upload"),
+    )])
+
+    assert "gallery_id" not in result["files"][0]
+    db = TestingSession()
+    try:
+        assert db.query(GalleryImage).count() == 0
+    finally:
+        db.close()