Merge pull request #6996 from odysseus-dev/test/ci-python-baseline-repair

test(caldav): restore redirect positive-control coverage

routes/cookbook_helpers.py

@@ -1030,6 +1030,166 @@ def _append_llama_cpp_linux_accel_build_lines(runner_lines: list[str]) -> None:
     runner_lines.append('  fi  # end _odysseus_have_prebuilt guard')
 
 
+ODYSSEUS_LLAMA_SHIM_MARKER = "# Auto-generated by Odysseus Cookbook: a `llama-server` lookalike"
+
+
+def _append_llama_server_resolution_definition_lines(runner_lines: list[str]) -> None:
+    """Define the helper shell function that discovers, reconciles and prioritizes llama-server."""
+    marker = f"'{_bash_squote(ODYSSEUS_LLAMA_SHIM_MARKER)}'"
+    runner_lines.append('_odysseus_resolve_llama_server() {')
+    runner_lines.append('  _odysseus_llama_bin=""')
+    runner_lines.append(f'  local _marker={marker}')
+    runner_lines.append('  local _saved_ifs="$IFS"')
+    runner_lines.append("  IFS=':'")
+    runner_lines.append('  local -a _parts')
+    runner_lines.append('  read -r -a _parts <<< "$PATH"')
+    runner_lines.append('  IFS="$_saved_ifs"')
+    runner_lines.append('  local _dir _cand')
+    runner_lines.append('  for _dir in "${_parts[@]}"; do')
+    runner_lines.append('    [ -n "$_dir" ] || continue')
+    runner_lines.append('    _cand="$_dir/llama-server"')
+    runner_lines.append('    [ -x "$_cand" ] || continue')
+    runner_lines.append('    if [ -f "$_cand" ] && [ ! -L "$_cand" ] && grep -Fxq "$_marker" "$_cand" 2>/dev/null; then')
+    runner_lines.append('      continue')
+    runner_lines.append('    fi')
+    runner_lines.append('    if "$_cand" --version >/dev/null 2>&1; then')
+    runner_lines.append('      _odysseus_llama_bin="$_cand"')
+    runner_lines.append('      break')
+    runner_lines.append('    fi')
+    runner_lines.append('  done')
+    runner_lines.append('  if [ -f "$HOME/bin/llama-server" ] && [ ! -L "$HOME/bin/llama-server" ] && grep -Fxq "$_marker" "$HOME/bin/llama-server" 2>/dev/null; then')
+    runner_lines.append('    if [ -n "$_odysseus_llama_bin" ]; then')
+    runner_lines.append('      echo "[odysseus] Native llama-server detected — retiring Odysseus compatibility shim in ~/bin"')
+    runner_lines.append('      rm -f "$HOME/bin/llama-server"')
+    runner_lines.append('      hash -r')
+    runner_lines.append('    elif python3 -c "import llama_cpp.server" 2>/dev/null; then')
+    runner_lines.append('      _odysseus_llama_bin="$HOME/bin/llama-server"')
+    runner_lines.append('    else')
+    runner_lines.append('      echo "[odysseus] Removing stale Odysseus llama-server shim (python3 -m llama_cpp.server unavailable)"')
+    runner_lines.append('      rm -f "$HOME/bin/llama-server"')
+    runner_lines.append('      hash -r')
+    runner_lines.append('    fi')
+    runner_lines.append('  fi')
+    runner_lines.append('  if [ -n "$_odysseus_llama_bin" ]; then')
+    runner_lines.append('    local _cand_dir')
+    runner_lines.append('    _cand_dir="$(dirname "$_odysseus_llama_bin")"')
+    runner_lines.append('    if [ "${PATH%%:*}" != "$_cand_dir" ]; then')
+    runner_lines.append('      export PATH="$_cand_dir:$PATH"')
+    runner_lines.append('    fi')
+    runner_lines.append('    hash -r')
+    runner_lines.append('    return 0')
+    runner_lines.append('  fi')
+    runner_lines.append('  return 1')
+    runner_lines.append('}')
+
+
+def _append_llama_server_shim_definition_lines(runner_lines: list[str]) -> None:
+    """Define the helper shell function that generates the Odysseus compatibility shim."""
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    marker = f"'{_bash_squote(ODYSSEUS_LLAMA_SHIM_MARKER)}'"
+    runner_lines.append('_odysseus_create_llama_shim() {')
+    runner_lines.append('  local _target="$HOME/bin/llama-server"')
+    runner_lines.append(f'  local _marker={marker}')
+    runner_lines.append('  if [ -e "$_target" ] || [ -L "$_target" ]; then')
+    runner_lines.append('    if [ -f "$_target" ] && [ ! -L "$_target" ] && grep -Fxq "$_marker" "$_target" 2>/dev/null; then')
+    runner_lines.append('      : # existing Odysseus shim — safe to replace')
+    runner_lines.append('    else')
+    runner_lines.append('      echo "[odysseus] WARNING: $_target already exists and is not an Odysseus-managed shim; leaving untouched." >&2')
+    runner_lines.append('      return 1')
+    runner_lines.append('    fi')
+    runner_lines.append('  fi')
+    runner_lines.append('  mkdir -p "$HOME/bin" || return 1')
+    runner_lines.append('  local _tmp')
+    runner_lines.append('  _tmp="$(mktemp "$HOME/bin/.llama-server.odysseus.tmp.XXXXXX")" || {')
+    runner_lines.append('    echo "[odysseus] ERROR: unable to create secure temporary llama-server shim in $HOME/bin" >&2')
+    runner_lines.append('    return 1')
+    runner_lines.append('  }')
+    runner_lines.append("  if ! cat > \"$_tmp\" <<'_ODY_LLAMA_SHIM_EOF'")
+    runner_lines.append('#!/usr/bin/env bash')
+    runner_lines.append(ODYSSEUS_LLAMA_SHIM_MARKER)
+    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('  then')
+    runner_lines.append('    rm -f -- "$_tmp"')
+    runner_lines.append('    return 1')
+    runner_lines.append('  fi')
+    runner_lines.append('  chmod +x "$_tmp" || { rm -f -- "$_tmp"; return 1; }')
+    runner_lines.append('  mv -f -- "$_tmp" "$_target" || { rm -f -- "$_tmp"; return 1; }')
+    runner_lines.append('  hash -r')
+    runner_lines.append('  echo "[odysseus] Created llama-server shim → python -m llama_cpp.server (no native binary needed)"')
+    runner_lines.append('}')
+
+
+def _append_llama_server_shim_reconciliation_lines(runner_lines: list[str]) -> None:
+    """Reconcile an existing ~/bin/llama-server if it is owned by Odysseus Cookbook.
+
+    Invariants:
+    - Never touch an arbitrary user-owned executable or symlink.
+    - If ~/bin/llama-server is an Odysseus generated shim:
+        - If a native llama-server is present and healthy elsewhere on PATH, remove the shim
+          so the native binary is used.
+        - If python3 -m llama_cpp.server is unavailable, remove the stale shim
+          and hash -r so the fallback/build path is triggered.
+        - If no healthy native server exists and llama_cpp.server is working, preserve it.
+    - If a healthy native candidate exists on PATH (even later in PATH), prioritize it and
+      ensure the launched executable is the validated candidate.
+    """
+    runner_lines.append('[ -r ~/.config/odysseus-llama-cpp-env ] && . ~/.config/odysseus-llama-cpp-env')
+    runner_lines.append('if ! declare -F _odysseus_resolve_llama_server >/dev/null 2>&1; then')
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    runner_lines.append('fi')
+    runner_lines.append('_odysseus_resolve_llama_server || true')
+
+
+def _append_llama_server_finalization_lines(runner_lines: list[str], platform_hint: str = "") -> None:
+    """Reconcile post-install launcher availability and enforce fail-closed invariant.
+
+    Guarantees that if python3 -m llama_cpp.server is usable but the `llama-server`
+    command is unresolvable, the compatibility launcher is created. If no usable
+    launcher exists after all build/install attempts, sets ODYSSEUS_PREFLIGHT_EXIT=127.
+    """
+    err_msg = (
+        f"ERROR: llama.cpp serving is not available on {platform_hint} after install attempts."
+        if platform_hint
+        else "ERROR: llama.cpp serving is not available after install/build attempts."
+    )
+    runner_lines.append('if ! declare -F _odysseus_resolve_llama_server >/dev/null 2>&1; then')
+    _append_llama_server_resolution_definition_lines(runner_lines)
+    runner_lines.append('fi')
+    runner_lines.append('if [ -z "$_odysseus_llama_bin" ] && python3 -c "import llama_cpp.server" 2>/dev/null; then')
+    runner_lines.append('  _odysseus_create_llama_shim')
+    runner_lines.append('  _odysseus_resolve_llama_server || true')
+    runner_lines.append('fi')
+    runner_lines.append('if [ -z "$_odysseus_llama_bin" ]; then')
+    runner_lines.append(f'  echo "{err_msg}"')
+    runner_lines.append('  ODYSSEUS_PREFLIGHT_EXIT=127')
+    runner_lines.append('fi')
+
+
 def _llama_cpp_rebuild_cmd(update_source: bool = False) -> str:
     """Shell command that clears the Cookbook-managed llama.cpp build.
 

routes/cookbook_routes.py

@@ -57,9 +57,11 @@ from routes.cookbook_helpers import (
     _pip_install_no_cache, _user_shell_path_bootstrap, _venv_safe_local_pip_install_cmd,
     _diagnose_serve_output, run_ssh_command_async,
     _ollama_bind_from_cmd, _pip_install_fallback_chain, _pip_install_no_cache,
-    _user_shell_path_bootstrap, _venv_safe_local_pip_install_cmd,
     _append_pip_install_runner_lines, _pip_install_command_without_break_system_packages,
     _normalize_llama_cpp_python_cache_types,
+    _append_llama_server_shim_definition_lines,
+    _append_llama_server_shim_reconciliation_lines,
+    _append_llama_server_finalization_lines,
     ModelDownloadRequest, ServeRequest,
 )
 
@@ -2161,7 +2163,7 @@ def setup_cookbook_routes() -> APIRouter:
                 ps_lines.append('}')
             elif "llama_cpp" in req.cmd or "llama-server" in req.cmd:
                 ps_lines.append('# Auto-install llama-cpp-python if missing')
-                ps_lines.append('try { python -c "import llama_cpp" 2>$null } catch {}')
+                ps_lines.append('try { python -c "import llama_cpp.server" 2>$null } catch {}')
                 ps_lines.append('if ($LASTEXITCODE -ne 0) {')
                 ps_lines.append('  Write-Host "Installing llama-cpp-python..."')
                 ps_lines.append('  python -m pip install llama-cpp-python[server]')
@@ -2247,14 +2249,17 @@ def setup_cookbook_routes() -> APIRouter:
                 # ollama is found (otherwise macOS falls back to a slow source build).
                 # /opt/homebrew = Apple Silicon, /usr/local = Intel; harmless on Linux.
                 runner_lines.append('export PATH="$HOME/.local/bin:$HOME/bin:$HOME/llama.cpp/build/bin:/opt/homebrew/bin:/usr/local/bin:$PATH"')
+                _append_llama_server_shim_definition_lines(runner_lines)
+                _append_llama_server_shim_reconciliation_lines(runner_lines)
                 runner_lines.append('if [ -d /data/data/com.termux ]; then')
                 runner_lines.append('  # Termux: no native build — use the Python bindings (CPU).')
-                runner_lines.append('  if ! python3 -c "import llama_cpp" 2>/dev/null; then')
+                runner_lines.append('  if ! python3 -c "import llama_cpp.server" 2>/dev/null; then')
                 runner_lines.append('    pkg install -y cmake 2>/dev/null')
                 runner_lines.append('    pip install numpy diskcache jinja2 2>/dev/null')
                 runner_lines.append('    CMAKE_ARGS="-DGGML_BLAS=OFF -DGGML_LLAMAFILE=OFF" pip install \'llama-cpp-python[server]\' --no-build-isolation --no-cache-dir 2>&1 || true')
                 runner_lines.append('  fi')
-                runner_lines.append('elif ! command -v llama-server &>/dev/null; then')
+                _append_llama_server_finalization_lines(runner_lines, platform_hint="Termux")
+                runner_lines.append('elif [ -z "$_odysseus_llama_bin" ]; then')
                 runner_lines.append('  echo "Native llama-server not found — building from source (one-time, may take a few minutes)..."')
                 runner_lines.append('  mkdir -p ~/bin')
                 runner_lines.append('  cd ~ && [ -d llama.cpp ] || git clone --depth 1 https://github.com/ggml-org/llama.cpp')
@@ -2271,8 +2276,7 @@ def setup_cookbook_routes() -> APIRouter:
                 # would reuse the bad settings and fail again. CMAKE_BUILD_TYPE is
                 # explicit so the binary is optimized (Metal auto-enables on macOS).
                 runner_lines.append('    cd ~/llama.cpp && rm -rf build && cmake -B build -DCMAKE_BUILD_TYPE=Release \\')
-                runner_lines.append('      && cmake --build build -j"$NPROC" --target llama-server \\')
-                runner_lines.append('      && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
+                runner_lines.append('      && cmake --build build -j"$NPROC" --target llama-server')
                 runner_lines.append('  else')
                 _append_llama_cpp_linux_accel_build_lines(runner_lines)
                 runner_lines.append('  fi')
@@ -2296,58 +2300,13 @@ def setup_cookbook_routes() -> APIRouter:
                 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('  _odysseus_resolve_llama_server || true')
+                # If the native build failed, fall back to the Python bindings.
+                runner_lines.append('  if [ -z "$_odysseus_llama_bin" ] && ! python3 -c "import llama_cpp.server" 2>/dev/null; then')
                 runner_lines.append('    echo "llama-server build failed — installing Python bindings as fallback..."')
                 runner_lines.append(f"    {_pip_install_fallback_chain('llama-cpp-python[server]', python_cmd='pip')} || true")
                 runner_lines.append('  fi')
-                runner_lines.append('  if ! command -v llama-server &>/dev/null && ! python3 -c "import llama_cpp" 2>/dev/null; then')
-                runner_lines.append('    echo "ERROR: llama.cpp serving is not available after install/build attempts."')
-                runner_lines.append('    ODYSSEUS_PREFLIGHT_EXIT=127')
-                runner_lines.append('  fi')
+                _append_llama_server_finalization_lines(runner_lines)
                 runner_lines.append('fi')
             elif re.search(r"\bollama\s+serve\b", req.cmd):
                 handled_ollama_serve = True

routes/session_routes.py

@@ -1315,7 +1315,7 @@ def setup_session_routes(
     def auto_sort_sessions(request: Request, skip_llm: bool = False):
         """Use AI to categorize all sessions into folders.
 
-        Phase 1 deletes empty/throwaway sessions and Phase 2 asks the LLM
+        Phase 1 deletes empty/Incognito sessions and Phase 2 asks the LLM
         to assign folders. When `skip_llm=true` the endpoint returns
         after Phase 1 — used by the "Tidy (no AI)" UI affordance so
         users can clean junk without spending tokens.
@@ -1326,21 +1326,11 @@ def setup_session_routes(
         single_user_mode = not user and _auth_disabled()
         user_sessions = session_manager.get_sessions_for_user(user)
 
-        # Delete empty and throwaway sessions before sorting
+        # Saved conversations are never disposable based on their title or prompt.
         from core.database import ChatMessage as DbMsg
         db = SessionLocal()
         deleted_empty = 0
         deleted_throwaway = 0
-        # Names that indicate a throwaway/test session (case-insensitive exact or prefix match)
-        _THROWAWAY_NAMES = {
-            "test", "testing", "asdf", "asd", "hello", "hi", "hey",
-            "yo", "sup", "hola", "hii", "hiii", "heyo",
-            "foo", "bar", "baz", "tmp", "temp", "scratch", "untitled",
-            "new chat", "delete", "remove", "junk", "trash", "xxx",
-            "abc", "qwerty", "blah", "stuff", "whatever", "idk",
-            "ok", "lol", "bruh", "hmm", "hm", "meh",
-        }
-        _THROWAWAY_MAX_MESSAGES = 4  # only delete if <= this many messages
         try:
             rows_q = db.query(DbSession).filter(DbSession.archived == False)
             if user:
@@ -1349,15 +1339,11 @@ def setup_session_routes(
                 rows_q = rows_q.filter(DbSession.owner == user)
             rows = rows_q.limit(2000).all()
             folder_map = {r.id: r.folder for r in rows}
-            # Precompute per-session message counts in TWO aggregate queries
+            # Precompute per-session message counts in one aggregate query
             # instead of 1–3 queries PER session — with many chats the per-row
             # loop was doing thousands of round-trips and blowing the timeout.
             from sqlalchemy import func as _sa_func
             _counts = dict(db.query(DbMsg.session_id, _sa_func.count(DbMsg.id)).group_by(DbMsg.session_id).all())
-            _asst_counts = dict(
-                db.query(DbMsg.session_id, _sa_func.count(DbMsg.id))
-                .filter(DbMsg.role == "assistant").group_by(DbMsg.session_id).all()
-            )
             cleanup_now = utcnow_naive()
             for row in rows:
                 # Never delete important sessions
@@ -1367,9 +1353,10 @@ def setup_session_routes(
                 if (row.name or "").strip() == "Incognito":
                     should_delete = True
                     deleted_throwaway += 1
-                    db.delete(row)
                     if hasattr(session_manager, 'delete_session'):
                         session_manager.delete_session(row.id)
+                    else:
+                        db.delete(row)
                     continue
                 if is_session_recently_active(row, now=cleanup_now):
                     continue
@@ -1378,31 +1365,11 @@ def setup_session_routes(
                 if msg_count == 0:
                     should_delete = True
                     deleted_empty += 1
-                elif msg_count <= _THROWAWAY_MAX_MESSAGES:
-                    name = (row.name or "").strip().lower()
-                    # Check first user message content (AI renames sessions, so
-                    # "hi" becomes "Casual Greeting Exchange" — name alone won't match)
-                    first_msg = db.query(DbMsg.content).filter(
-                        DbMsg.session_id == row.id, DbMsg.role == "user"
-                    ).order_by(DbMsg.timestamp).first()
-                    first_text = (first_msg[0] or "").strip().lower() if first_msg else ""
-                    # Count assistant messages — if user sent something but AI never replied, it's dead
-                    assistant_count = _asst_counts.get(row.id, 0)
-                    if name in _THROWAWAY_NAMES or name.startswith("chat:") or first_text in _THROWAWAY_NAMES:
-                        should_delete = True
-                        deleted_throwaway += 1
-                    # Single user message with no AI response = dead session
-                    elif msg_count == 1 and assistant_count == 0:
-                        should_delete = True
-                        deleted_throwaway += 1
-                    # Short phrase (1-3 words) with no real AI conversation (<=2 msgs)
-                    elif msg_count <= 2 and first_text and len(first_text.split()) <= 3 and len(first_text) <= 40:
-                        should_delete = True
-                        deleted_throwaway += 1
                 if should_delete:
-                    db.delete(row)
                     if hasattr(session_manager, 'delete_session'):
                         session_manager.delete_session(row.id)
+                    else:
+                        db.delete(row)
             if deleted_empty or deleted_throwaway:
                 db.commit()
                 logger.info(f"Auto-sort: deleted {deleted_empty} empty + {deleted_throwaway} throwaway sessions")

src/mcp_manager.py

@@ -10,6 +10,7 @@ import logging
 import os
 import re
 import asyncio 
+import contextlib
 from typing import Any, Dict, List, Optional, Set, Tuple
 from src.database import McpServer, SessionLocal
 
@@ -383,12 +384,21 @@ class McpManager:
 
             provider = build_provider(server_id, url, on_redirect=_on_redirect)
             stack = AsyncExitStack()
+            try:
                 transport = await stack.enter_async_context(streamablehttp_client(url, auth=provider))
                 read_stream, write_stream, _get_session_id = transport
                 session = await stack.enter_async_context(ClientSession(read_stream, write_stream))
                 await session.initialize()
 
                 tools_result = await session.list_tools()
+            except BaseException:
+                # Close the half-open transport here, in the task that entered
+                # it. Left for the garbage collector, its AnyIO cancel scope is
+                # exited from the wrong task and the event loop keeps spinning
+                # (#5518). Its own teardown error would only mask the real one.
+                with contextlib.suppress(Exception):
+                    await stack.aclose()
+                raise
             tools = []
             for tool in tools_result.tools:
                 tools.append({

src/session_actions.py

@@ -12,16 +12,6 @@ from datetime import datetime, timedelta, timezone
 
 logger = logging.getLogger(__name__)
 
-# Names that indicate a throwaway/test session
-_THROWAWAY_NAMES = {
-    "test", "testing", "asdf", "asd", "hello", "hi", "hey",
-    "yo", "sup", "hola", "hii", "hiii", "heyo",
-    "foo", "bar", "baz", "tmp", "temp", "scratch", "untitled",
-    "new chat", "delete", "remove", "junk", "trash", "xxx",
-    "abc", "qwerty", "blah", "stuff", "whatever", "idk",
-    "ok", "lol", "bruh", "hmm", "hm", "meh",
-}
-_THROWAWAY_MAX_MESSAGES = 4
 _FRESH_EMPTY_SESSION_GRACE = timedelta(minutes=10)
 _FRESH_SESSION_GRACE = _FRESH_EMPTY_SESSION_GRACE
 
@@ -54,14 +44,15 @@ def is_session_recently_active(row, now=None, grace=_FRESH_SESSION_GRACE) -> boo
 
 
 async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bool = True) -> str:
-    """Run session cleanup + (optional) AI folder sort for the given owner.
+    """Remove empty/Incognito sessions and optionally organize saved chats.
 
     Args:
         owner: user whose sessions to process
-        skip_llm: when True, do only Phase 1 (delete empty/throwaway sessions);
+        skip_llm: when True, do only Phase 1 (delete empty/Incognito sessions);
             skip Phase 2 (AI folder assignment). Used by the built-in daily
             background sweep so it never burns LLM tokens.
-        delete_throwaway: when False, only empty/incognito sessions are deleted.
+        delete_throwaway: retained for caller compatibility. Ordinary sessions
+            with saved messages are always preserved, regardless of this flag.
 
     Returns a human-readable summary of what was done.
     """
@@ -71,7 +62,7 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
 
     db = SessionLocal()
     try:
-        # ── Phase 1: Delete empty/throwaway sessions ──
+        # ── Phase 1: Delete only empty and explicitly Incognito sessions ──
         deleted_empty = 0
         deleted_throwaway = 0
 
@@ -93,47 +84,13 @@ async def run_auto_sort(owner: str, skip_llm: bool = False, delete_throwaway: bo
             if is_session_recently_active(row, now=cleanup_now):
                 continue
 
-            msg_count = db.query(DbMsg.id).filter(
+            has_messages = db.query(DbMsg.id).filter(
                 DbMsg.session_id == row.id
-            ).limit(_THROWAWAY_MAX_MESSAGES + 1).count()
-            should_delete = False
-
-            if msg_count == 0:
+            ).first() is not None
+            if not has_messages:
                 if is_fresh:
                     continue
-                should_delete = True
                 deleted_empty += 1
-            elif delete_throwaway and msg_count <= _THROWAWAY_MAX_MESSAGES:
-                name = (row.name or "").strip().lower()
-                first_msg = db.query(DbMsg.content).filter(
-                    DbMsg.session_id == row.id, DbMsg.role == "user"
-                ).order_by(DbMsg.timestamp).first()
-                first_text = (first_msg[0] or "").strip().lower() if first_msg else ""
-                assistant_count = db.query(DbMsg.id).filter(
-                    DbMsg.session_id == row.id, DbMsg.role == "assistant"
-                ).limit(1).count()
-
-                if name in _THROWAWAY_NAMES or name.startswith("chat:") or first_text in _THROWAWAY_NAMES:
-                    should_delete = True
-                    deleted_throwaway += 1
-                elif msg_count == 1 and assistant_count == 0:
-                    should_delete = True
-                    deleted_throwaway += 1
-                elif msg_count <= 4 and first_text and len(first_text.split()) <= 8 and len(first_text) <= 80:
-                    # Short trivial chats — e.g. "write hi to a friend" → "Hi!"
-                    should_delete = True
-                    deleted_throwaway += 1
-                else:
-                    # Aggressive: total message text under 250 chars combined = trivial
-                    msg_rows = db.query(DbMsg.content).filter(
-                        DbMsg.session_id == row.id
-                    ).all()
-                    total_chars = sum(len(m[0] or "") for m in msg_rows)
-                    if total_chars <= 250:
-                        should_delete = True
-                        deleted_throwaway += 1
-
-            if should_delete:
                 db.delete(row)
 
         if deleted_empty or deleted_throwaway:

static/js/galleryEditor.js

@@ -5759,20 +5759,7 @@ function _updateBrushCursor(e) {
     state.cursorEl.style.height = diameterY + 'px';
   }
   state.cursorEl.classList.toggle('ge-brush-cursor--pixel', pixelBrush);
-  // Updating an inline style on every move invalidates document-wide :has()
-  // selectors. A paused animation keeps the same transform in the rendering
-  // pipeline without mutating the cursor's style attribute on every frame.
-  const cursorTransform = `translate3d(${clientX - diameterX / 2}px, ${clientY - diameterY / 2}px, 0)`;
-  if (!state.cursorEl._positionAnimation || state.cursorEl._positionAnimation.playState === 'idle') {
-    state.cursorEl._positionAnimation = state.cursorEl.animate(
-      [{ transform: cursorTransform }, { transform: cursorTransform }],
-      { duration: 1, fill: 'both' },
-    );
-    state.cursorEl._positionAnimation.pause();
-    state.cursorEl._positionAnimation.currentTime = 0;
-  } else {
-    state.cursorEl._positionAnimation.effect.setKeyframes([{ transform: cursorTransform }, { transform: cursorTransform }]);
-  }
+  state.cursorEl.style.transform = `translate3d(${clientX - diameterX / 2}px, ${clientY - diameterY / 2}px, 0)`;
   if (state.cursorEl.style.display !== 'block') {
     state.cursorEl.style.display = 'block';
   }
@@ -14074,20 +14061,7 @@ export function closeEditor(options = {}) {
     const h = state.editorDocClickHandlers.pop();
     try { document.removeEventListener('click', h); } catch {}
   }
-  if (_cursorRaf) { cancelAnimationFrame(_cursorRaf); _cursorRaf = null; }
-  _pendingCursorEvent = null;
-  _lastCursorRect = null;
-  state._lastCursorDiameterX = null;
-  state._lastCursorDiameterY = null;
-  state._lastCursorBorderColor = null;
-  if (state.cursorEl) {
-    if (state.cursorEl._positionAnimation) {
-      try { state.cursorEl._positionAnimation.cancel(); } catch {}
-      state.cursorEl._positionAnimation = null;
-    }
-    state.cursorEl.remove();
-    state.cursorEl = null;
-  }
+  if (state.cursorEl) { state.cursorEl.remove(); state.cursorEl = null; }
   try { document.getElementById('ge-cursor-layer')?.remove(); } catch {}
   // Tear down all floating popups + the dock so closing the editor
   // doesn't leave stale chips/panels behind on top of the gallery.

tests/e2e/photo-editor/brush-cursor.spec.js

@@ -46,475 +46,3 @@ test('brush cursor matches the rendered brush diameter and replaces the crosshai
   await expect.poll(async () => (await cursor.boundingBox()).width)
     .toBeCloseTo(measurements.expectedWidth * 2, 1);
 });
-
-
-test('brush cursor follows the pointer before, during, and after a stroke and cancels animations on close', async ({ page }) => {
-  await openBlankEditor(page, { width: 400, height: 300 });
-  await selectTool(page, 'brush');
-  const canvas = await page.locator('.ge-main-canvas').boundingBox();
-  const cursor = page.locator('.ge-brush-cursor');
-  const checkCenter = async (x, y) => {
-    await expect.poll(async () => {
-      const box = await cursor.boundingBox();
-      return box ? Math.max(Math.abs(box.x + box.width / 2 - x), Math.abs(box.y + box.height / 2 - y)) : Infinity;
-    }).toBeLessThan(1);
-  };
-  const a = { x: canvas.x + canvas.width * 0.3, y: canvas.y + canvas.height * 0.4 };
-  const b = { x: canvas.x + canvas.width * 0.7, y: canvas.y + canvas.height * 0.6 };
-  await page.mouse.move(a.x, a.y);
-  await checkCenter(a.x, a.y);
-  await page.mouse.down();
-  await page.mouse.move(b.x, b.y, { steps: 8 });
-  await checkCenter(b.x, b.y);
-  await page.mouse.up();
-  await page.mouse.move(a.x, a.y);
-  await checkCenter(a.x, a.y);
-
-  // 1. Ensure at least one animation exists before checking cancellation
-  // 2. Verify the actual cursor-position animation is paused while active
-  const activeAnimation = await page.evaluate(() => {
-    const cursorEl = document.querySelector('.ge-brush-cursor');
-    const anims = cursorEl ? cursorEl.getAnimations() : [];
-    return {
-      count: anims.length,
-      hasPositionAnim: Boolean(cursorEl?._positionAnimation),
-      positionAnimPaused: cursorEl?._positionAnimation?.playState === 'paused',
-      positionAnimCurrentTime: cursorEl?._positionAnimation?.currentTime,
-      allPaused: anims.length > 0 && anims.every(a => a.playState === 'paused'),
-    };
-  });
-  expect(activeAnimation.count).toBeGreaterThanOrEqual(1);
-  expect(activeAnimation.hasPositionAnim).toBe(true);
-  expect(activeAnimation.positionAnimPaused).toBe(true);
-  expect(activeAnimation.positionAnimCurrentTime).toBe(0);
-  expect(activeAnimation.allPaused).toBe(true);
-
-  // 3. Verify it is canceled when the editor closes
-  // 4. Ensure temporary test-only references are cleaned up
-  const closeOutcome = await page.evaluate(async () => {
-    const cursorEl = document.querySelector('.ge-brush-cursor');
-    const anims = cursorEl ? cursorEl.getAnimations() : [];
-    const editor = await window.__photoEditorImport('/static/js/galleryEditor.js');
-    editor.closeEditor({ force: true });
-    return {
-      trackedCount: anims.length,
-      allIdle: anims.length > 0 && anims.every(a => a.playState === 'idle'),
-      states: anims.map(a => a.playState),
-    };
-  });
-  await expect(cursor).toHaveCount(0);
-  expect(closeOutcome.trackedCount).toBeGreaterThanOrEqual(1);
-  expect(closeOutcome.allIdle).toBe(true);
-  expect(closeOutcome.states.every(s => s === 'idle')).toBe(true);
-
-  // Verify no temporary references linger on window
-  const windowClean = await page.evaluate(() => {
-    return !('__cursorAnimationsForTest' in window) && !('__styleTransformMutations' in window);
-  });
-  expect(windowClean).toBe(true);
-});
-
-
-test('steady pointer movement does not perform repeated inline style-transform writes', async ({ page }) => {
-  await openBlankEditor(page, { width: 400, height: 300 });
-  await selectTool(page, 'brush');
-
-  const canvas = await page.locator('.ge-main-canvas').boundingBox();
-  const cursor = page.locator('.ge-brush-cursor');
-
-  // Warm up the cursor first: move pointer onto canvas
-  const startX = canvas.x + canvas.width * 0.25;
-  const startY = canvas.y + canvas.height * 0.25;
-  await page.mouse.move(startX, startY);
-  await expect(cursor).toBeVisible();
-
-  // Observe style attribute mutations during steady pointer movement.
-  // Brush size, color, and mode remain unchanged throughout.
-  await page.evaluate(() => {
-    const el = document.querySelector('.ge-brush-cursor');
-    window.__styleMutationRecords = [];
-    const observer = new MutationObserver((mutations) => {
-      for (const m of mutations) {
-        if (m.type === 'attributes' && m.attributeName === 'style') {
-          const oldStyle = m.oldValue || '';
-          const currentStyle = el.getAttribute('style') || '';
-          // Track mutations where style-transform was written inline
-          if (el.style.transform || currentStyle.includes('transform:') || oldStyle.includes('transform:')) {
-            window.__styleMutationRecords.push({
-              oldStyle,
-              currentStyle,
-              inlineTransform: el.style.transform,
-            });
-          }
-        }
-      }
-    });
-    observer.observe(el, { attributes: true, attributeFilter: ['style'], attributeOldValue: true });
-    window.__styleMutationObserver = observer;
-  });
-
-  // Preserve actual mouse interactions: steady movement across 10 steps
-  const endX = canvas.x + canvas.width * 0.75;
-  const endY = canvas.y + canvas.height * 0.75;
-  await page.mouse.move(endX, endY, { steps: 10 });
-
-  // Confirm the cursor accurately tracked the pointer
-  await expect.poll(async () => {
-    const box = await cursor.boundingBox();
-    return box ? Math.max(Math.abs(box.x + box.width / 2 - endX), Math.abs(box.y + box.height / 2 - endY)) : Infinity;
-  }).toBeLessThan(1);
-
-  // Collect observation results and clean up observer and globals
-  const { mutationCount, inlineTransform, animCount } = await page.evaluate(() => {
-    window.__styleMutationObserver?.disconnect();
-    delete window.__styleMutationObserver;
-    const records = window.__styleMutationRecords || [];
-    delete window.__styleMutationRecords;
-    const el = document.querySelector('.ge-brush-cursor');
-    return {
-      mutationCount: records.length,
-      inlineTransform: el?.style.transform || '',
-      animCount: el?.getAnimations?.().length || 0,
-    };
-  });
-
-  // The optimized implementation uses paused Web Animations keyframes, avoiding repeated
-  // inline style-transform writes that trigger expensive :has() document invalidations.
-  expect(mutationCount).toBe(0);
-  expect(inlineTransform).toBe('');
-  expect(animCount).toBeGreaterThanOrEqual(1);
-});
-
-
-test('cross-tool cursor positioning, sizing, and styling across brush, eraser, pixel, lasso, inpaint, liquify, and clone', async ({ page }) => {
-  await openBlankEditor(page, { width: 400, height: 300 });
-  const canvas = await page.locator('.ge-main-canvas').boundingBox();
-  const cursor = page.locator('.ge-brush-cursor');
-
-  const checkCenter = async (x, y) => {
-    await expect.poll(async () => {
-      const box = await cursor.boundingBox();
-      return box ? Math.max(Math.abs(box.x + box.width / 2 - x), Math.abs(box.y + box.height / 2 - y)) : Infinity;
-    }).toBeLessThan(1);
-  };
-
-  const midX = canvas.x + canvas.width / 2;
-  const midY = canvas.y + canvas.height / 2;
-
-  // 1. Brush tool
-  await selectTool(page, 'brush');
-  await page.mouse.move(midX, midY);
-  await expect(cursor).toBeVisible();
-  await checkCenter(midX, midY);
-  let props = await page.evaluate(() => {
-    const el = document.querySelector('.ge-brush-cursor');
-    return {
-      isPixel: el.classList.contains('ge-brush-cursor--pixel'),
-      display: el.style.display,
-      borderStyle: el.style.borderStyle,
-    };
-  });
-  expect(props.isPixel).toBe(false);
-  expect(props.display).toBe('block');
-
-  // 2. Eraser tool
-  await selectTool(page, 'eraser');
-  await page.mouse.move(midX + 10, midY + 10);
-  await checkCenter(midX + 10, midY + 10);
-  props = await page.evaluate(() => {
-    const el = document.querySelector('.ge-brush-cursor');
-    return {
-      borderColor: el.style.borderColor,
-      isPixel: el.classList.contains('ge-brush-cursor--pixel'),
-    };
-  });
-  expect(props.borderColor).toBe('rgba(255, 255, 255, 0.6)');
-  expect(props.isPixel).toBe(false);
-
-  // 3. Pixel brush (via tool / preset)
-  await page.evaluate(async () => {
-    const { state } = await import('/static/js/editor/state.js');
-    state.tool = 'pixel';
-    state.brushPixel = true;
-    state.brushSize = 16;
-  });
-  await page.mouse.move(midX - 10, midY - 10);
-  await checkCenter(midX - 10, midY - 10);
-  props = await page.evaluate(() => {
-    const el = document.querySelector('.ge-brush-cursor');
-    return {
-      isPixel: el.classList.contains('ge-brush-cursor--pixel'),
-      width: parseFloat(el.style.width),
-    };
-  });
-  expect(props.isPixel).toBe(true);
-
-  // 4. Lasso tool
-  await page.evaluate(async () => {
-    const { state } = await import('/static/js/editor/state.js');
-    state.tool = 'lasso';
-    state.brushPixel = false;
-  });
-  await page.mouse.move(midX + 20, midY - 10);
-  await checkCenter(midX + 20, midY - 10);
-  props = await page.evaluate(() => {
-    const el = document.querySelector('.ge-brush-cursor');
-    return {
-      borderColor: el.style.borderColor,
-      background: el.style.background,
-    };
-  });
-  expect(props.borderColor).toBe('rgba(255, 255, 255, 0.85)');
-  expect(props.background).toBe('rgba(0, 0, 0, 0.15)');
-
-  // 5. Inpaint tool (paint mode and erase mode)
-  await page.evaluate(async () => {
-    const { state } = await import('/static/js/editor/state.js');
-    state.tool = 'inpaint';
-    state.inpaintEraseMode = false;
-  });
-  await page.mouse.move(midX, midY);
-  await checkCenter(midX, midY);
-  props = await page.evaluate(() => {
-    const el = document.querySelector('.ge-brush-cursor');
-    return {
-      borderColor: el.style.borderColor,
-      background: el.style.background,
-      borderStyle: el.style.borderStyle,
-    };
-  });
-  expect(props.borderColor).toBe('rgba(255, 80, 80, 0.8)');
-  expect(props.borderStyle).toBe('solid');
-
-  // Inpaint erase mode
-  await page.evaluate(async () => {
-    const { state } = await import('/static/js/editor/state.js');
-    state.inpaintEraseMode = true;
-  });
-  await page.mouse.move(midX + 5, midY + 5);
-  props = await page.evaluate(() => {
-    const el = document.querySelector('.ge-brush-cursor');
-    return {
-      borderColor: el.style.borderColor,
-      borderStyle: el.style.borderStyle,
-    };
-  });
-  expect(props.borderColor).toBe('rgba(255, 255, 255, 0.9)');
-  expect(props.borderStyle).toBe('dashed');
-
-  // 6. Liquify tool
-  await page.evaluate(async () => {
-    const { state } = await import('/static/js/editor/state.js');
-    state.tool = 'liquify';
-    state.liquifySize = 100;
-  });
-  await page.mouse.move(midX - 15, midY + 15);
-  await checkCenter(midX - 15, midY + 15);
-  props = await page.evaluate(() => {
-    const el = document.querySelector('.ge-brush-cursor');
-    return {
-      width: parseFloat(el.style.width),
-      height: parseFloat(el.style.height),
-      display: el.style.display,
-    };
-  });
-  expect(props.display).toBe('block');
-  expect(props.width).toBeGreaterThan(0);
-
-  // 7. Clone tool
-  await page.evaluate(async () => {
-    const { state } = await import('/static/js/editor/state.js');
-    state.tool = 'clone';
-    state.brushSize = 50;
-  });
-  await page.mouse.move(midX, midY);
-  await checkCenter(midX, midY);
-  await expect(cursor).toBeVisible();
-
-  // Clean close
-  await page.evaluate(async () => {
-    const editor = await window.__photoEditorImport('/static/js/galleryEditor.js');
-    editor.closeEditor({ force: true });
-  });
-  await expect(cursor).toHaveCount(0);
-});
-
-
-test('canvas boundaries, pointer exit and re-entry, and viewport resize', async ({ page }) => {
-  await openBlankEditor(page, { width: 400, height: 300 });
-  await selectTool(page, 'brush');
-
-  const canvas = await page.locator('.ge-main-canvas').boundingBox();
-  const cursor = page.locator('.ge-brush-cursor');
-
-  const checkCenter = async (x, y) => {
-    await expect.poll(async () => {
-      const box = await cursor.boundingBox();
-      return box ? Math.max(Math.abs(box.x + box.width / 2 - x), Math.abs(box.y + box.height / 2 - y)) : Infinity;
-    }).toBeLessThan(1);
-  };
-
-  // 1. Enter canvas
-  const insideX = canvas.x + canvas.width * 0.5;
-  const insideY = canvas.y + canvas.height * 0.5;
-  await page.mouse.move(insideX, insideY);
-  await expect(cursor).toBeVisible();
-  await checkCenter(insideX, insideY);
-
-  // 2. Exit canvas boundary: cursor should hide
-  const outsideX = canvas.x - 50;
-  const outsideY = canvas.y - 50;
-  await page.mouse.move(outsideX, outsideY);
-  await expect(cursor).toBeHidden();
-
-  // 3. Re-enter canvas: cursor should restore visibility and position
-  await page.mouse.move(insideX, insideY);
-  await expect(cursor).toBeVisible();
-  await checkCenter(insideX, insideY);
-
-  // 4. Viewport resize
-  await page.setViewportSize({ width: 1280, height: 800 });
-  await page.waitForTimeout(100);
-  const newCanvas = await page.locator('.ge-main-canvas').boundingBox();
-  const newInsideX = newCanvas.x + newCanvas.width * 0.5;
-  const newInsideY = newCanvas.y + newCanvas.height * 0.5;
-  await page.mouse.move(newInsideX, newInsideY);
-  await expect(cursor).toBeVisible();
-  await checkCenter(newInsideX, newInsideY);
-
-  // Clean close
-  await page.evaluate(async () => {
-    const editor = await window.__photoEditorImport('/static/js/galleryEditor.js');
-    editor.closeEditor({ force: true });
-  });
-  await expect(cursor).toHaveCount(0);
-});
-
-
-for (const view of ['2x zoom', 'fit-to-view']) {
-  test(`zoom and fit-to-view cursor scaling and positioning: ${view}`, async ({ page }) => {
-    const size = view === '2x zoom' ? { width: 400, height: 300 } : { width: 1600, height: 1200 };
-    await openBlankEditor(page, size);
-    await selectTool(page, 'brush');
-
-    await page.evaluate(async () => {
-      const { state } = await import('/static/js/editor/state.js');
-      state.brushSize = 50;
-      // The zoom-in control applies 1.25x through the real zoom handler.
-      state.zoom = 1.6;
-      state.zoomViewMode = 'scale';
-    });
-    await page.locator('#ge-zoom-in').click();
-    await expect.poll(() => page.evaluate(async () => {
-      const { state } = await import('/static/js/editor/state.js');
-      return { zoom: state.zoom, mode: state.zoomViewMode };
-    })).toEqual({ zoom: 2, mode: 'scale' });
-
-    if (view === 'fit-to-view') {
-      await page.locator('#ge-zoom-100').click();
-      const fitted = await page.evaluate(async () => {
-        const { state } = await import('/static/js/editor/state.js');
-        const area = document.querySelector('.ge-canvas-area');
-        const pad = state._footerDocked ? 0 : 20;
-        const rulerPad = state.rulersVisible ? 18 : 0;
-        return {
-          zoom: state.zoom,
-          mode: state.zoomViewMode,
-          expectedZoom: Math.min(1,
-            (area.clientWidth - 2 * pad - rulerPad) / state.imgWidth,
-            (area.clientHeight - 2 * pad - rulerPad) / state.imgHeight),
-        };
-      });
-      expect(fitted.mode).toBe('fit');
-      expect(fitted.zoom).toBeGreaterThan(0);
-      expect(fitted.zoom).toBeLessThan(1);
-      expect(fitted.zoom).toBeCloseTo(fitted.expectedZoom, 6);
-    }
-
-    const canvas = page.locator('.ge-main-canvas');
-    const canvasBox = await canvas.boundingBox();
-    const cursor = page.locator('.ge-brush-cursor');
-    if (view === '2x zoom') {
-      expect(canvasBox.width).toBeCloseTo(size.width * 2, 1);
-      expect(canvasBox.height).toBeCloseTo(size.height * 2, 1);
-    }
-    const targetX = canvasBox.x + canvasBox.width * 0.4;
-    const targetY = canvasBox.y + canvasBox.height * 0.4;
-    await page.mouse.move(targetX, targetY);
-    await expect(cursor).toBeVisible();
-
-    // Pointer coordinates can be rounded to device pixels; keep the existing
-    // subpixel alignment tolerance separate from the stricter diameter check.
-    await expect.poll(async () => {
-      const box = await cursor.boundingBox();
-      return box ? Math.max(
-        Math.abs(box.x + box.width / 2 - targetX),
-        Math.abs(box.y + box.height / 2 - targetY),
-      ) : Infinity;
-    }).toBeLessThan(1);
-    await expect.poll(() => page.evaluate(() => {
-      const canvasEl = document.querySelector('.ge-main-canvas');
-      const canvasRect = canvasEl.getBoundingClientRect();
-      const cursorRect = document.querySelector('.ge-brush-cursor').getBoundingClientRect();
-      return Math.max(
-        Math.abs(cursorRect.width - 50 * canvasRect.width / canvasEl.width),
-        Math.abs(cursorRect.height - 50 * canvasRect.height / canvasEl.height),
-      );
-    })).toBeLessThan(0.1);
-
-    await page.evaluate(async () => {
-      const editor = await window.__photoEditorImport('/static/js/galleryEditor.js');
-      editor.closeEditor({ force: true });
-    });
-    await expect(cursor).toHaveCount(0);
-  });
-}
-
-
-test('memory and lifecycle: repeated editor open/close cycles retain no detached cursors or active animations', async ({ page }) => {
-  for (let cycle = 0; cycle < 5; cycle++) {
-    await openBlankEditor(page, { width: 400, height: 300 });
-    await selectTool(page, 'brush');
-
-    const canvas = await page.locator('.ge-main-canvas').boundingBox();
-    await page.mouse.move(canvas.x + canvas.width * 0.5, canvas.y + canvas.height * 0.5);
-    const cursor = page.locator('.ge-brush-cursor');
-    await expect(cursor).toBeVisible();
-
-    // Verify cursor layer and animation are active
-    const cycleState = await page.evaluate(() => {
-      const el = document.querySelector('.ge-brush-cursor');
-      return {
-        cursorCount: document.querySelectorAll('.ge-brush-cursor').length,
-        layerCount: document.querySelectorAll('#ge-cursor-layer').length,
-        hasAnim: Boolean(el?._positionAnimation),
-        animPaused: el?._positionAnimation?.playState === 'paused',
-      };
-    });
-    expect(cycleState.cursorCount).toBe(1);
-    expect(cycleState.layerCount).toBe(1);
-    expect(cycleState.animPaused).toBe(true);
-
-    // Close editor
-    await page.evaluate(async () => {
-      const editor = await window.__photoEditorImport('/static/js/galleryEditor.js');
-      editor.closeEditor({ force: true });
-    });
-
-    // Verify complete cleanup after close
-    await expect(cursor).toHaveCount(0);
-    const postClose = await page.evaluate(() => {
-      return {
-        cursorCount: document.querySelectorAll('.ge-brush-cursor').length,
-        layerCount: document.querySelectorAll('#ge-cursor-layer').length,
-        activeAnims: document.getAnimations().filter(a => {
-          const target = a.effect?.target;
-          return target && (target.classList?.contains('ge-brush-cursor') || target.id === 'ge-cursor-layer');
-        }).length,
-      };
-    });
-    expect(postClose.cursorCount).toBe(0);
-    expect(postClose.layerCount).toBe(0);
-    expect(postClose.activeAnims).toBe(0);
-  }
-});

tests/test_agent_editor_routes.py

@@ -1929,23 +1929,60 @@ def test_place_layer_reports_false_on_queued_response(monkeypatch):
     assert placed is False
 
 
-def test_bridge_await_cancellation_handled_cleanly():
+def test_bridge_await_cancellation_handled_cleanly(monkeypatch):
     import asyncio
+    from contextlib import suppress
     from httpx import ASGITransport, AsyncClient
     app = _bridge_app()
 
     async def scenario():
+        entered_wait = asyncio.Event()
+        real_wait_for = asyncio.wait_for
+        waits = []
+
+        async def observe_wait(awaitable, timeout):
+            # Observe the real wait; do not replace its result or cancellation.
+            waits.append(awaitable)
+            entered_wait.set()
+            return await real_wait_for(awaitable, timeout)
+
+        monkeypatch.setattr(eaar.asyncio, "wait_for", observe_wait)
         async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
             cmd_resp = await c.post("/api/editor/bridge/command", json={"session": "cancel-sess", "op": "document.info", "wait": False})
+            assert cmd_resp.status_code == 200
+            assert cmd_resp.json()["completed"] is False
             cid = cmd_resp.json()["queued"][0]
 
             await_task = asyncio.create_task(c.post("/api/editor/bridge/await", json={"id": cid, "timeout": 30}))
-            await asyncio.sleep(0.05)
-            await_task.cancel()
             try:
+                await real_wait_for(entered_wait.wait(), 5)
+                assert len(waits) == 1
+                assert not waits[0].done()
+                assert not await_task.done(), "await returned without a queued result"
+                await_task.cancel()
+                response = await real_wait_for(await_task, 5)
+                assert response.status_code == 200
+                assert response.json() == {
+                    "ok": False, "error": "cancelled", "status": "cancelled",
+                    "completed": False, "id": cid,
+                }
+            finally:
+                if not await_task.done():
+                    await_task.cancel()
+                    with suppress(asyncio.CancelledError):
                         await await_task
-            except asyncio.CancelledError:
-                pass
+
+            # Cancelling the waiter must not cancel the queued command itself.
+            payload = {"width": 137, "height": 251}
+            delivered = await c.post("/api/editor/bridge/result", json={"id": cid, "ok": True, "result": payload})
+            assert delivered.status_code == 200
+            completed = await c.post("/api/editor/bridge/await", json={"id": cid, "timeout": 1})
+            assert completed.status_code == 200
+            assert completed.json()["status"] == "completed"
+            assert completed.json()["completed"] is True
+            assert completed.json()["ok"] is True
+            assert completed.json()["result"] == payload
+            assert len(waits) == 2
 
     asyncio.run(scenario())
 

tests/test_auth_config_lock_concurrency.py

@@ -206,35 +206,50 @@ class TestDiskConsistency:
     @pytest.mark.slow
     def test_file_always_valid_json_during_concurrent_ops(self, tmp_path):
         mgr = _fresh_auth_manager(tmp_path)
-        mgr.create_user("admin", "adminpw", is_admin=True)
+        assert mgr.create_user("admin", "adminpw", is_admin=True)
 
         stop_event = threading.Event()
+        reader_ready = threading.Event()
         corruption_found = []
+        observations = {"reads": 0, "with_writes": 0}
 
         def reader():
             while not stop_event.is_set():
                 try:
                     with open(mgr.auth_path, "r") as f:
                         content = f.read()
-                    json.loads(content)
+                    data = json.loads(content)
+                    observations["reads"] += 1
+                    if any(name.startswith("stress") for name in data["users"]):
+                        observations["with_writes"] += 1
+                    reader_ready.set()
                 except json.JSONDecodeError as e:
                     corruption_found.append(str(e))
                     break
-                except FileNotFoundError:
-                    pass
                 time.sleep(0.001)
 
         def writer():
+            assert reader_ready.wait(5), "reader never observed the initial file"
+            results = []
             for i in range(50):
-                mgr.create_user(f"stress{i}", f"pw{i}")
+                results.append(mgr.create_user(f"stress{i}", f"pw{i}"))
+            return results
 
-        reader_thread = threading.Thread(target=reader)
-        writer_thread = threading.Thread(target=writer)
-
-        reader_thread.start()
-        writer_thread.start()
-        writer_thread.join()
+        # Futures propagate reader/writer exceptions to the test thread.
+        with ThreadPoolExecutor(max_workers=2) as pool:
+            reading = pool.submit(reader)
+            writing = pool.submit(writer)
+            try:
+                results = writing.result(timeout=120)
+            finally:
                 stop_event.set()
-        reader_thread.join()
+            reading.result(timeout=5)
 
         assert not corruption_found, f"Corrupt JSON detected: {corruption_found[0]}"
+        assert len(results) == 50 and all(results), "writer operations did not succeed"
+        assert observations["reads"] > 0
+        assert observations["with_writes"] > 0, "reader saw no persisted writer activity"
+        expected = {"admin"} | {f"stress{i}" for i in range(50)}
+        with open(mgr.auth_path) as persisted:
+            assert set(json.load(persisted)["users"]) == expected
+        assert set(_fresh_auth_manager(tmp_path).users) == expected

tests/test_caldav_redirect_hardening.py

@@ -13,6 +13,7 @@ import socketserver
 import threading
 
 import pytest
+import requests
 
 from src import caldav_sync, caldav_writeback
 
@@ -26,12 +27,13 @@ def test_build_dav_client_disables_redirects():
 
 def test_dav_client_does_not_follow_redirect_to_internal_host():
     """End-to-end through the real DAVClient: a 302 toward an internal host
-    must NOT be followed. Without the fix the sink is contacted (SSRF); with it
-    the redirect is refused and the sink is never reached."""
+    must NOT be followed. Exercise the session limit separately because some
+    caldav versions independently refuse cross-origin PROPFIND redirects."""
     pytest.importorskip("caldav")
 
     sink_hits: list[str] = []
     public_methods: list[str] = []
+    same_origin_hits: list[str] = []
 
     class _Internal(http.server.BaseHTTPRequestHandler):
         # Stand-in for an internal service the attacker redirects toward.
@@ -48,9 +50,15 @@ def test_dav_client_does_not_follow_redirect_to_internal_host():
     class _Public(http.server.BaseHTTPRequestHandler):
         # The "validated" public CalDAV server that redirects everything inward.
         def do_GET(self):  # noqa: N802
+            if self.path == "/landing":
+                same_origin_hits.append(self.path)
+                self.send_response(200)
+                self.end_headers()
+                return
             public_methods.append(self.command)
             self.send_response(302)
-            self.send_header("Location", f"http://127.0.0.1:{internal_port}/leak")
+            location = "/landing" if self.path == "/same" else f"http://127.0.0.1:{internal_port}/leak"
+            self.send_header("Location", location)
             self.end_headers()
 
         do_PROPFIND = do_GET
@@ -62,8 +70,10 @@ def test_dav_client_does_not_follow_redirect_to_internal_host():
     internal_port = internal.server_address[1]
     public = socketserver.TCPServer(("127.0.0.1", 0), _Public)
     public_port = public.server_address[1]
-    threading.Thread(target=internal.serve_forever, daemon=True).start()
-    threading.Thread(target=public.serve_forever, daemon=True).start()
+    internal_thread = threading.Thread(target=internal.serve_forever, daemon=True)
+    public_thread = threading.Thread(target=public.serve_forever, daemon=True)
+    internal_thread.start()
+    public_thread.start()
     try:
         public_url = f"http://127.0.0.1:{public_port}/dav"
         client = caldav_sync._build_dav_client(public_url, "u", "p")
@@ -82,17 +92,35 @@ def test_dav_client_does_not_follow_redirect_to_internal_host():
         assert public_methods in (["PROPFIND"], ["PROPFIND", "GET"]), public_methods
         assert sink_hits == [], "redirect toward an internal host must not be followed"
 
-        # Positive control: demonstrate that the sink would see the redirect
-        # if the session's redirect protection were removed.
+        # GET redirects are handled by the HTTP session even in caldav 3.7.
+        # This makes the cross-port SSRF assertion sensitive to our own limit.
+        with pytest.raises(Exception, match="Exceeded 0 redirects"):
+            client.request(public_url, "GET", "")
+        assert sink_hits == [], "GET redirect toward the internal sink must be blocked"
+
+        # requests can reach the cross-port sink when redirects are enabled,
+        # even if DAVClient independently rejects cross-origin PROPFIND redirects.
+        with requests.Session() as session:
+            response = session.get(public_url, timeout=5)
+            assert response.status_code == 207
+        assert sink_hits == ["/leak"], "the cross-port redirect must reach the sink"
+
+        # GET follows redirects at the HTTP session layer across these caldav
+        # versions. Check the application's limit on a same-origin redirect too.
+        same_url = f"http://127.0.0.1:{public_port}/same"
+        with pytest.raises(Exception, match="Exceeded 0 redirects"):
+            client.request(same_url, "GET", "")
+        assert same_origin_hits == []
         client.session.max_redirects = 30
-        try:
-            client.request(public_url, "PROPFIND", "")
-        except Exception:
-            pass  # The sink's response is intentionally not a valid DAV body.
-        assert sink_hits == ["/leak"], "the test must detect a followed redirect"
+        client.request(same_url, "GET", "")
+        assert same_origin_hits == ["/landing"]
     finally:
         internal.shutdown()
         public.shutdown()
+        internal_thread.join()
+        public_thread.join()
+        internal.server_close()
+        public.server_close()
 
 
 def test_sync_and_writeback_construct_clients_through_the_helper():

tests/test_chat_tidy_content_preservation.py

@@ -0,0 +1,148 @@
+"""Both chat tidy entry points retain saved conversations and still remove empties."""
+import asyncio
+from datetime import datetime, timedelta, timezone
+
+import pytest
+from fastapi import APIRouter, FastAPI
+from fastapi.testclient import TestClient
+from sqlalchemy import create_engine
+from sqlalchemy.orm import sessionmaker
+
+import core.database as database
+import core.session_manager as manager_module
+import routes.session_routes as routes
+import src.session_actions as actions
+
+
+@pytest.fixture
+def tidy_api(tmp_path, monkeypatch):
+    engine = create_engine(
+        f"sqlite:///{(tmp_path / 'tidy.db').as_posix()}",
+        connect_args={"check_same_thread": False},
+    )
+    database.Base.metadata.create_all(engine)
+    factory = sessionmaker(bind=engine, autoflush=False)
+    monkeypatch.setattr(database, "SessionLocal", factory)
+    monkeypatch.setattr(manager_module, "SessionLocal", factory)
+    monkeypatch.setattr(routes, "SessionLocal", factory)
+    monkeypatch.setattr(routes, "router", APIRouter(prefix="/api"))
+    monkeypatch.setattr(routes, "effective_user", lambda request: "alice")
+    manager = manager_module.SessionManager.__new__(manager_module.SessionManager)
+    manager.sessions = {}
+    manager.upload_handler = None
+    monkeypatch.setattr(manager, "get_sessions_for_user", lambda owner: {})
+    app = FastAPI()
+    app.include_router(routes.setup_session_routes(manager, {}))
+    with TestClient(app) as client:
+        yield client, factory, manager
+    engine.dispose()
+
+
+def run_tidy(client, entrypoint):
+    if entrypoint == "action":
+        return asyncio.run(actions.run_auto_sort("alice", skip_llm=True))
+    response = client.post("/api/sessions/auto-sort?skip_llm=true")
+    assert response.status_code == 200, response.text
+    return response.json()
+
+
+@pytest.mark.parametrize("entrypoint", ["route", "action"])
+@pytest.mark.parametrize("title,prompt", [
+    ("New chat", "Please analyze the following contract and retain all details. " * 20),
+    ("Contract analysis", "Review contract"),
+    ("Chat: analysis", "Explain these results and retain the full answer"),
+    ("Test", "Here is a large input that should be analyzed and retained"),
+], ids=["generic-title", "short-prompt", "title-prefix", "test-title"])
+def test_tidy_preserves_nontrivial_saved_answer(tidy_api, title, prompt, entrypoint):
+    client, factory, manager = tidy_api
+    old = datetime.now(timezone.utc).replace(tzinfo=None) - timedelta(hours=2)
+    answer = "Clause analysis and recommended changes.\n" * 2000
+    with factory() as db:
+        db.add(database.Session(
+            id="valuable-chat", name=title, owner="alice", endpoint_url="", model="",
+            archived=False, is_important=False, message_count=2,
+            created_at=old, updated_at=old, last_accessed=old, last_message_at=old,
+        ))
+        db.commit()
+        db.add_all([
+            database.ChatMessage(id="prompt", session_id="valuable-chat", role="user",
+                                 content=prompt, timestamp=old),
+            database.ChatMessage(id="answer", session_id="valuable-chat", role="assistant",
+                                 content=answer, timestamp=old + timedelta(seconds=1)),
+        ])
+        db.commit()
+    run_tidy(client, entrypoint)
+    with factory() as db:
+        remaining = db.query(database.ChatMessage).filter_by(session_id="valuable-chat").count()
+        survives = db.get(database.Session, "valuable-chat") is not None
+    assert survives and remaining == 2, "Title/prompt heuristics must not erase saved answers"
+    with factory() as db:
+        assert db.get(database.ChatMessage, "answer").content == answer
+
+
+@pytest.mark.parametrize("entrypoint", ["route", "action"])
+@pytest.mark.parametrize("kind", ["unanswered", "greeting", "empty", "fresh-empty", "starred-empty", "archived", "other-owner", "incognito"])
+def test_tidy_cleanup_boundaries(tidy_api, entrypoint, kind):
+    client, factory, manager = tidy_api
+    now = datetime.now(timezone.utc).replace(tzinfo=None)
+    timestamp = now if kind == "fresh-empty" else now - timedelta(hours=2)
+    message = kind in {"unanswered", "greeting", "archived", "other-owner", "incognito"}
+    with factory() as db:
+        db.add(database.Session(
+            id="boundary-chat", name="Incognito" if kind == "incognito" else "New chat",
+            owner="bob" if kind == "other-owner" else "alice", endpoint_url="", model="",
+            archived=kind == "archived", is_important=kind == "starred-empty",
+            # Deliberately inaccurate metadata: cleanup must use message rows.
+            message_count=0 if message else 40,
+            created_at=timestamp, updated_at=timestamp, last_accessed=timestamp, last_message_at=timestamp,
+        ))
+        db.commit()
+        if message:
+            db.add(database.ChatMessage(id="boundary-message", session_id="boundary-chat", role="user",
+                                        content="hi" if kind == "greeting" else "Pending question " * 100,
+                                        timestamp=timestamp))
+            db.commit()
+    run_tidy(client, entrypoint)
+    deleted = kind in {"empty", "incognito"}
+    with factory() as db:
+        assert (db.get(database.Session, "boundary-chat") is None) == deleted
+        assert db.query(database.ChatMessage).filter_by(session_id="boundary-chat").count() == int(message and not deleted)
+
+
+@pytest.mark.parametrize("entrypoint", ["route", "action"])
+def test_tidy_keeps_folder_assignment_working(tidy_api, monkeypatch, entrypoint):
+    from types import SimpleNamespace
+    import src.llm_core as llm
+    import src.task_endpoint as endpoints
+    client, factory, manager = tidy_api
+    old = datetime.now(timezone.utc).replace(tzinfo=None) - timedelta(hours=2)
+    for sid in ["chat-one-long-id", "chat-two-long-id"]:
+        with factory() as db:
+            db.add(database.Session(id=sid, name="New chat", owner="alice", endpoint_url="", model="",
+                                    archived=False, created_at=old, updated_at=old,
+                                    last_accessed=old, last_message_at=old))
+            db.commit()
+            db.add(database.ChatMessage(id=f"{sid}-message", session_id=sid, role="user",
+                                        content="Saved question", timestamp=old))
+            db.commit()
+        manager.sessions[sid] = SimpleNamespace(id=sid, name="New chat", archived=False,
+                                                updated_at=old.isoformat(), created_at=old.isoformat())
+    monkeypatch.setattr(manager, "get_sessions_for_user", lambda owner: manager.sessions)
+    monkeypatch.setattr(endpoints, "resolve_task_endpoint", lambda *args, **kwargs: ("http://model.invalid", "test", {}))
+    response_text = '{"folders":{"Projects":["chat-one","chat-two"]}}'
+    monkeypatch.setattr(llm, "llm_call", lambda *args, **kwargs: response_text)
+
+    async def async_response(*args, **kwargs):
+        return response_text
+
+    monkeypatch.setattr(llm, "llm_call_async", async_response)
+    if entrypoint == "action":
+        result = asyncio.run(actions.run_auto_sort("alice"))
+        assert "Sorted 2 sessions" in result
+    else:
+        response = client.post("/api/sessions/auto-sort")
+        assert response.status_code == 200, response.text
+        assert response.json()["updated"] == 2
+    with factory() as db:
+        assert db.query(database.ChatMessage).count() == 2
+        assert {row.folder for row in db.query(database.Session).all()} == {"Projects"}

tests/test_cookbook_dependency_completion_regression.py

@@ -1,4 +1,14 @@
+import os
+import subprocess
 from pathlib import Path
+import pytest
+from core.platform_compat import find_bash, git_bash_path
+from routes.cookbook_helpers import (
+    ODYSSEUS_LLAMA_SHIM_MARKER,
+    _append_llama_server_shim_definition_lines,
+    _append_llama_server_shim_reconciliation_lines,
+    _append_llama_server_finalization_lines,
+)
 
 from tests.helpers.stylesheets import app_css
 
@@ -615,3 +625,730 @@ def test_local_dependency_probe_refreshes_user_site_visibility():
     # deps (realesrgan) probe as not-installed until a restart. See #4810.
     assert "if user_site and os.path.isdir(user_site):" in source
     assert "site.addsitedir(user_site)" in source
+
+
+def test_llama_runner_path_preserves_standard_user_bin_order():
+    """The runner PATH must preserve standard user $HOME/bin precedence while
+    relying on ownership-aware reconciliation to retire Odysseus shims when
+    native builds exist."""
+    source = _read("routes/cookbook_routes.py")
+
+    assert (
+        'export PATH="$HOME/.local/bin:$HOME/bin:$HOME/llama.cpp/build/bin:/opt/homebrew/bin:/usr/local/bin:$PATH"'
+        in source
+    )
+
+
+def _run_shell(tmp_path, lines, cwd=None, path=None, extra_env=None):
+    from core.platform_compat import find_bash, git_bash_path
+    bash = find_bash() or "bash"
+    env = dict(os.environ, HOME=git_bash_path(tmp_path))
+    if path is not None:
+        env["PATH"] = path
+    if extra_env:
+        env.update(extra_env)
+    return subprocess.run(
+        [bash, "-c", "\n".join(lines)],
+        cwd=str(cwd) if cwd else None,
+        env=env,
+        capture_output=True,
+        text=True,
+    )
+
+
+def _write_odysseus_shim(path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(
+        f"#!/usr/bin/env bash\n{ODYSSEUS_LLAMA_SHIM_MARKER}\nexec python3 -m llama_cpp.server \"$@\"\n",
+        encoding="utf-8",
+    )
+    path.chmod(0o755)
+
+
+def _write_mock_native(path: Path, succeeds: bool = True) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    if succeeds:
+        content = '#!/usr/bin/env bash\nif [ "$1" = "--version" ]; then echo "version: 0.1.0-dev"; exit 0; fi\necho "RUN_NATIVE: $@"\nexit 0\n'
+    else:
+        content = '#!/usr/bin/env bash\necho "error: missing libllama.so" >&2\nexit 127\n'
+    path.write_text(content, encoding="utf-8")
+    path.chmod(0o755)
+
+
+def _write_mock_python3(path: Path, succeeds: bool = True) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(f"#!/usr/bin/env bash\nexit {0 if succeeds else 1}\n", encoding="utf-8")
+    path.chmod(0o755)
+
+
+def test_odysseus_create_llama_shim_never_overwrites_user_regular_file(tmp_path):
+    """Creation boundary: never overwrite a non-executable user-owned regular file."""
+    user_file = tmp_path / "bin/llama-server"
+    user_file.parent.mkdir(parents=True, exist_ok=True)
+    original_content = "user configuration or binary data without marker"
+    user_file.write_text(original_content, encoding="utf-8")
+
+    lines = []
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode != 0, "Creation should return non-zero when target is a user file"
+    assert user_file.read_text(encoding="utf-8") == original_content, "User file must remain byte-for-byte unchanged"
+
+
+def test_odysseus_create_llama_shim_never_overwrites_user_symlink(tmp_path):
+    """Creation boundary: never follow or overwrite a user-owned symlink."""
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir(parents=True, exist_ok=True)
+    target_file = tmp_path / "user_target.txt"
+    target_file.write_text("precious user target content", encoding="utf-8")
+    user_symlink = bin_dir / "llama-server"
+    user_symlink.symlink_to(target_file)
+
+    lines = []
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode != 0, "Creation should return non-zero when target is a user symlink"
+    assert user_symlink.is_symlink(), "Symlink must remain intact"
+    assert target_file.read_text(encoding="utf-8") == "precious user target content", "Target file must not be modified"
+
+
+def test_odysseus_create_llama_shim_never_overwrites_dangling_symlink(tmp_path):
+    """Creation boundary: never overwrite a dangling user-owned symlink."""
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir(parents=True, exist_ok=True)
+    dangling_symlink = bin_dir / "llama-server"
+    dangling_symlink.symlink_to(tmp_path / "nonexistent_target")
+
+    lines = []
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode != 0, "Creation should return non-zero when target is a dangling symlink"
+    assert dangling_symlink.is_symlink(), "Dangling symlink must remain intact"
+
+
+def test_odysseus_create_llama_shim_safely_replaces_existing_odysseus_shim(tmp_path):
+    """Creation boundary: safely replace an existing Odysseus-owned compatibility shim."""
+    shim_file = tmp_path / "bin/llama-server"
+    shim_file.parent.mkdir(parents=True, exist_ok=True)
+    shim_file.write_text(f"#!/bin/bash\n{ODYSSEUS_LLAMA_SHIM_MARKER}\n# old shim body\n", encoding="utf-8")
+
+    lines = []
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    new_content = shim_file.read_text(encoding="utf-8")
+    assert ODYSSEUS_LLAMA_SHIM_MARKER in new_content
+    assert "exec python3 -m llama_cpp.server" in new_content
+
+
+def test_odysseus_create_llama_shim_secure_temp_prevents_symlink_attack(tmp_path):
+    """Creation boundary: secure mktemp prevents predictable $$ temp symlink attacks.
+
+    Given:
+    - A victim file at ~/victim containing DO_NOT_TOUCH
+    - A pre-created symlink at ~/bin/.llama-server.odysseus.tmp.$$ pointing to ~/victim
+    - _odysseus_create_llama_shim executed within that exact same bash PID ($$)
+
+    Prove:
+    - ~/victim remains unmodified
+    - Pre-existing crafted symlink is NOT followed or overwritten
+    - Legitimate shim is successfully created at ~/bin/llama-server via secure temp path
+    """
+    lines = [
+        'victim="$HOME/victim"',
+        'printf "DO_NOT_TOUCH\\n" > "$victim"',
+        'mkdir -p "$HOME/bin"',
+        'old_tmp="$HOME/bin/.llama-server.odysseus.tmp.$$";',
+        'ln -s "$victim" "$old_tmp"',
+    ]
+    _append_llama_server_shim_definition_lines(lines)
+    lines.append('_odysseus_create_llama_shim')
+    lines.append('res=$?')
+    lines.append('echo "CREATION_RES=$res"')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert "CREATION_RES=0" in res.stdout
+
+    victim_file = tmp_path / "victim"
+    assert victim_file.exists()
+    assert victim_file.read_text(encoding="utf-8") == "DO_NOT_TOUCH\n", "Victim file must remain untouched"
+
+    target_shim = tmp_path / "bin/llama-server"
+    assert target_shim.exists()
+    assert ODYSSEUS_LLAMA_SHIM_MARKER in target_shim.read_text(encoding="utf-8")
+    assert target_shim.stat().st_mode & 0o111 != 0, "Shim must be executable"
+
+
+def test_stale_odysseus_shim_removed_when_python_server_unusable(tmp_path):
+    """Scenario 1: Stale Odysseus shim + missing Python server -> owned shim removed."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Stale Odysseus shim must be removed when python server is unavailable"
+
+
+def test_healthy_odysseus_shim_retained_when_no_native_server(tmp_path):
+    """Scenario 2: Healthy Odysseus shim + no native server -> shim preserved."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines, path=f"{git_bash_path(mock_py_dir)}:{os.environ.get('PATH', '')}")
+    assert res.returncode == 0
+    assert shim_file.exists(), "Healthy Odysseus shim must be preserved when no native server is present"
+
+
+def test_reconciliation_does_not_retire_healthy_shim_for_broken_native_binary(tmp_path):
+    """Reconciliation health check: a broken native candidate executable must NOT retire
+    a healthy Odysseus Python shim."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    _write_mock_native(tmp_path / "llama.cpp/build/bin/llama-server", succeeds=False)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = [f'export PATH="$HOME/llama.cpp/build/bin:{git_bash_path(mock_py_dir)}:$PATH"']
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Healthy Odysseus shim must NOT be retired when native binary is broken"
+
+
+def test_odysseus_shim_retired_when_native_server_healthy(tmp_path):
+    """Scenario 3: Odysseus shim + healthy native server available -> native server wins."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    _write_mock_native(tmp_path / "llama.cpp/build/bin/llama-server", succeeds=True)
+
+    lines = ['export PATH="$HOME/bin:$HOME/llama.cpp/build/bin:$PATH"']
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Odysseus shim must be retired when healthy native server is available"
+
+
+def test_user_owned_bin_llama_server_never_touched(tmp_path):
+    """Scenario 4: Arbitrary user ~/bin/llama-server -> untouched and not deprioritized."""
+    user_bin = tmp_path / "bin/llama-server"
+    user_bin.parent.mkdir(parents=True, exist_ok=True)
+    user_bin.write_text("#!/bin/bash\n# Custom user build script\nexec /opt/custom/llama-server\n", encoding="utf-8")
+    user_bin.chmod(0o755)
+
+    _write_mock_native(tmp_path / "llama.cpp/build/bin/llama-server", succeeds=True)
+
+    lines = ['export PATH="$HOME/bin:$HOME/llama.cpp/build/bin:$PATH"']
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert user_bin.exists(), "User-owned ~/bin/llama-server must never be deleted"
+    assert "Custom user build script" in user_bin.read_text(encoding="utf-8")
+
+
+def test_reconciliation_ignores_native_binary_outside_effective_path(tmp_path):
+    """Native candidate outside effective PATH must NOT be considered."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    # Native binary placed in /usr_custom/bin which is NOT in PATH
+    _write_mock_native(tmp_path / "usr_custom/bin/llama-server", succeeds=True)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    # PATH only contains ~/bin and mock_py_dir — usr_custom is omitted
+    path = f"{git_bash_path(tmp_path / 'bin')}:{git_bash_path(mock_py_dir)}"
+    res = _run_shell(tmp_path, lines, path=path)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Shim must NOT be retired by native binary outside effective PATH"
+
+
+def test_reconciliation_skips_home_bin_with_or_without_trailing_slash(tmp_path):
+    """Reconciliation must exclude $HOME/bin and $HOME/bin/ from native candidate discovery."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    # PATH has ~/bin/ with trailing slash
+    path = f"{git_bash_path(tmp_path / 'bin')}/:{git_bash_path(mock_py_dir)}"
+    res = _run_shell(tmp_path, lines, path=path)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Shim must NOT rediscover itself when PATH has trailing slash in $HOME/bin/"
+
+
+def test_reconciliation_does_not_discover_cwd_executable_when_path_omits_dot(tmp_path):
+    """Reconciliation must NOT synthesize an empty/trailing PATH element that searches cwd.
+
+    Given:
+    - Current working directory contains an executable `llama-server`
+    - Healthy Odysseus Python shim in ~/bin/llama-server
+    - PATH does NOT contain '.' or cwd
+    - No other healthy native binary on PATH
+
+    Prove:
+    - Reconciliation does not discover ./llama-server
+    - The healthy Odysseus shim in ~/bin/llama-server is preserved
+    """
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    cwd_dir = tmp_path / "workdir"
+    _write_mock_native(cwd_dir / "llama-server", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    path = f"{git_bash_path(tmp_path / 'bin')}:{git_bash_path(mock_py_dir)}:/usr/bin:/bin"
+    res = _run_shell(tmp_path, lines, cwd=cwd_dir, path=path)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Healthy Odysseus shim must NOT be retired by finding ./llama-server in cwd"
+
+
+def test_reconciliation_discovers_healthy_native_on_explicit_other_path_entry(tmp_path):
+    """Reconciliation must discover healthy native binaries placed on explicit PATH entries."""
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+    custom_opt_bin = tmp_path / "custom_opt/bin"
+    _write_mock_native(custom_opt_bin / "llama-server", succeeds=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    path = f"{git_bash_path(tmp_path / 'bin')}:{git_bash_path(custom_opt_bin)}:/usr/bin:/bin"
+    res = _run_shell(tmp_path, lines, path=path)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Odysseus shim must be retired when healthy native server is on explicit PATH entry"
+
+
+def test_reconciliation_safely_iterates_path_with_glob_metacharacters(tmp_path):
+    """Reconciliation must not perform pathname expansion / globbing on PATH components.
+
+    Given:
+    - A directory path containing shell glob characters: ~/custom[dir]/bin
+    - A native binary ~/custom[dir]/bin/llama-server
+    - A working directory containing matching glob entries
+    - An existing Odysseus shim in ~/bin/llama-server
+
+    Prove:
+    - Reconciliation correctly parses the literal path without expanding [dir]
+    - Discovers the healthy native server and retires the Odysseus shim
+    """
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    glob_dir = tmp_path / "custom[dir]/bin"
+    _write_mock_native(glob_dir / "llama-server", succeeds=True)
+
+    cwd_dir = tmp_path / "cwd"
+    cwd_dir.mkdir(parents=True, exist_ok=True)
+    (cwd_dir / "customd").mkdir(parents=True, exist_ok=True)
+    (cwd_dir / "customi").mkdir(parents=True, exist_ok=True)
+
+    lines = []
+    _append_llama_server_shim_reconciliation_lines(lines)
+
+    path = f"{git_bash_path(tmp_path / 'bin')}:{git_bash_path(glob_dir)}:/usr/bin:/bin"
+    res = _run_shell(tmp_path, lines, cwd=cwd_dir, path=path)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Reconciliation must discover literal bracketed PATH without globbing"
+
+
+def test_fallback_install_success_creates_usable_launcher(tmp_path):
+    """Scenario 5: No native + no Python server -> fallback install succeeds
+    -> usable compatibility launcher exists and resolves command -v llama-server."""
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = [f'export PATH="$HOME/bin:{git_bash_path(mock_py_dir)}:$PATH"']
+    _append_llama_server_shim_definition_lines(lines)
+    _append_llama_server_finalization_lines(lines)
+    lines.append('command -v llama-server')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    created_shim = tmp_path / "bin/llama-server"
+    assert created_shim.exists()
+    assert ODYSSEUS_LLAMA_SHIM_MARKER in created_shim.read_text(encoding="utf-8")
+    assert str(tmp_path / "bin/llama-server") in res.stdout
+
+
+def test_failed_fallback_sets_preflight_exit_127(tmp_path):
+    """Scenario 6: No native + failed fallback installation -> ODYSSEUS_PREFLIGHT_EXIT=127."""
+    lines = [
+        'ODYSSEUS_PREFLIGHT_EXIT=""',
+        'export PATH="$HOME/bin:$PATH"',
+    ]
+    _append_llama_server_shim_definition_lines(lines)
+    _append_llama_server_finalization_lines(lines)
+    lines.append('echo "EXIT=$ODYSSEUS_PREFLIGHT_EXIT"')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert "EXIT=127" in res.stdout
+
+
+def test_python_server_readiness_requires_llama_cpp_server_module():
+    """Scenario 7 & 8: Python server readiness requires `import llama_cpp.server`,
+    while GPU offload capability checks still use base `import llama_cpp`."""
+    routes_source = _read("routes/cookbook_routes.py")
+    helpers_source = _read("routes/cookbook_helpers.py")
+
+    # Windows fallback readiness check requires .server
+    assert 'try { python -c "import llama_cpp.server" 2>$null } catch {}' in routes_source
+    # Termux fallback readiness check requires .server
+    assert 'if ! python3 -c "import llama_cpp.server" 2>/dev/null; then' in routes_source
+    # Build-failure fallback pip check requires .server
+    assert 'if [ -z "$_odysseus_llama_bin" ] && ! python3 -c "import llama_cpp.server" 2>/dev/null; then' in routes_source
+    # Shim reconciliation and post-install finalization require .server
+    assert 'elif python3 -c "import llama_cpp.server" 2>/dev/null; then' in helpers_source
+    assert 'if [ -z "$_odysseus_llama_bin" ] && python3 -c "import llama_cpp.server" 2>/dev/null; then' in helpers_source
+
+    # GPU offload support checks still intentionally inspect base llama_cpp
+    assert 'llama_cpp.llama_supports_gpu_offload()' in routes_source
+
+
+def test_darwin_native_build_does_not_force_write_bin_symlink():
+    """Darwin source build must not overwrite ~/bin/llama-server with a forced symlink."""
+    source = _read("routes/cookbook_routes.py")
+
+    assert "&& ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server" not in source
+    assert 'cmake --build build -j"$NPROC" --target llama-server' in source
+
+
+@pytest.mark.asyncio
+async def test_generated_runner_preflight_retires_stale_shim_before_backend_selection(monkeypatch, tmp_path):
+    """Integration regression: The actual generated Cookbook runner must execute
+    ownership reconciliation BEFORE evaluating the backend selection (`if Termux ... elif ! command -v llama-server`).
+
+    Given:
+    - Executable stale Odysseus shim in ~/bin/llama-server (with canonical marker)
+    - Healthy native llama-server in ~/llama.cpp/build/bin/llama-server
+    - ~/bin before ~/llama.cpp/build/bin on PATH
+
+    Prove:
+    1. Generated runner contains reconciliation shell lines strictly before the backend branching
+    2. Executing the generated runner script retires the stale ~/bin/llama-server shim
+    3. Resolves to the healthy native binary without attempting source build or pip fallback.
+    """
+    from unittest.mock import AsyncMock, MagicMock
+    from routes.cookbook_helpers import ServeRequest
+    import routes.cookbook_routes as cookbook_routes
+    from starlette.requests import Request
+
+    fake_home = tmp_path / "fake_home"
+    stale_shim = fake_home / "bin/llama-server"
+    _write_odysseus_shim(stale_shim)
+    native_bin = fake_home / "llama.cpp/build/bin/llama-server"
+    _write_mock_native(native_bin, succeeds=True)
+
+    # Mock environment to generate the real runner file with isolated DB
+    tmux_log_dir = tmp_path / "tmux_logs"
+    tmux_log_dir.mkdir(parents=True)
+    monkeypatch.setattr(cookbook_routes, "TMUX_LOG_DIR", tmux_log_dir)
+    monkeypatch.setattr(cookbook_routes, "require_admin", lambda request: None)
+    monkeypatch.setattr(cookbook_routes, "_binary_available", AsyncMock(return_value=True))
+    monkeypatch.setattr(cookbook_routes, "load_stored_hf_token", lambda **kwargs: "")
+    monkeypatch.setattr("core.database.SessionLocal", MagicMock())
+
+    class _Proc:
+        returncode = 0
+        stderr = None
+        async def wait(self):
+            return 0
+    monkeypatch.setattr(cookbook_routes.asyncio, "create_subprocess_shell", AsyncMock(return_value=_Proc()))
+
+    router = cookbook_routes.setup_cookbook_routes()
+    endpoint = None
+    for route in router.routes:
+        if route.path == "/api/model/serve" and "POST" in route.methods:
+            endpoint = route.endpoint
+            break
+    assert endpoint is not None
+
+    from types import SimpleNamespace
+    http_req = Request({
+        "type": "http",
+        "method": "POST",
+        "path": "/api/model/serve",
+        "headers": [],
+        "app": SimpleNamespace(state=SimpleNamespace(auth_manager=SimpleNamespace(
+            is_configured=True, is_admin=lambda user: user == "admin"))),
+    })
+    http_req.state.current_user = "admin"
+    http_req.state.local_model_authority = True
+
+    req = ServeRequest(
+        repo_id="test/repo",
+        cmd="llama-server -m /dummy/model.gguf --port 8085",
+        remote_host=None,
+    )
+
+    resp = await endpoint(http_req, req)
+    assert resp["ok"] is True
+    session_id = resp["session_id"]
+    runner_file = tmux_log_dir / f"{session_id}_run.sh"
+    assert runner_file.exists()
+    runner_script = runner_file.read_text(encoding="utf-8")
+
+    # Assert structural ordering in the generated script:
+    # Reconciliation must appear strictly BEFORE the Termux/backend check
+    recon_idx = runner_script.find("Native llama-server detected — retiring Odysseus compatibility shim in ~/bin")
+    termux_idx = runner_script.find("if [ -d /data/data/com.termux ]; then")
+    backend_idx = runner_script.find('elif [ -z "$_odysseus_llama_bin" ]; then')
+    assert recon_idx != -1, "Reconciliation block missing from generated runner"
+    assert termux_idx != -1, "Termux check missing from generated runner"
+    assert backend_idx != -1, "Backend selection missing from generated runner"
+    assert recon_idx < termux_idx < backend_idx, "Reconciliation must execute strictly before backend selection"
+
+    # Execute the generated runner preflight in a real subshell
+    res = _run_shell(fake_home, [str(runner_file)], path=f"{git_bash_path(fake_home / 'bin')}:{git_bash_path(fake_home / 'llama.cpp/build/bin')}:/usr/bin:/bin")
+    assert res.returncode == 0
+    assert not stale_shim.exists(), "Stale ~/bin/llama-server must have been retired by the generated preflight"
+    assert "RUN_NATIVE: -m /dummy/model.gguf --port 8085" in res.stdout
+    assert "Native llama-server not found — building from source" not in res.stdout
+
+
+def test_broken_first_native_candidate_healthy_second_candidate_executes_healthy(tmp_path):
+    """Regression finding P2: When PATH contains an earlier broken candidate followed by a
+    healthy candidate, reconciliation and launch must select and execute the healthy candidate."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken_bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    healthy_dir = tmp_path / "healthy_bin"
+    _write_mock_native(healthy_dir / "llama-server", succeeds=True)
+
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    lines = [
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(healthy_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_reconciliation_lines(lines)
+    lines.append('echo "RESOLVED=$_odysseus_llama_bin"')
+    lines.append('command -v llama-server')
+    lines.append('llama-server -m /dummy/model.gguf --port 8080')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Stale Odysseus shim must be retired"
+    assert f"RESOLVED={git_bash_path(healthy_dir / 'llama-server')}" in res.stdout
+    assert str(healthy_dir / "llama-server") in res.stdout
+    assert "RUN_NATIVE: -m /dummy/model.gguf --port 8080" in res.stdout
+    assert "error: missing libllama.so" not in res.stdout
+
+
+def test_paths_with_spaces_broken_first_healthy_second_launch(tmp_path):
+    """Regression: Paths containing spaces must be safely resolved, prioritized and launched."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken path with spaces/bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    healthy_dir = tmp_path / "healthy path with spaces/bin"
+    _write_mock_native(healthy_dir / "llama-server", succeeds=True)
+
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    model_dir = tmp_path / "models with spaces"
+    model_dir.mkdir(parents=True, exist_ok=True)
+    model_file = model_dir / "my model.gguf"
+    model_file.write_text("dummy", encoding="utf-8")
+
+    lines = [
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(healthy_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_reconciliation_lines(lines)
+    lines.append(f'llama-server -m "{git_bash_path(model_file)}" --port 8080')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert not shim_file.exists(), "Stale shim must be retired"
+    assert f"RUN_NATIVE: -m {git_bash_path(model_file)} --port 8080" in res.stdout
+    assert "error: missing libllama.so" not in res.stdout
+
+
+def test_command_hash_invalidation_switches_from_cached_broken_to_healthy(tmp_path):
+    """Regression: Command cache invalidation (`hash -r`) must prevent bash from reusing a
+    previously cached lookup to an earlier broken executable."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken_bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    healthy_dir = tmp_path / "healthy_bin"
+    _write_mock_native(healthy_dir / "llama-server", succeeds=True)
+
+    lines = [
+        # Populate bash hash cache with broken binary
+        f'export PATH="{git_bash_path(broken_dir)}:/usr/bin:/bin"',
+        'llama-server >/dev/null 2>&1 || true',
+        'hash',
+        # Now add healthy candidate later in PATH and run reconciliation
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(healthy_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_reconciliation_lines(lines)
+    lines.append('llama-server -m /dummy/test.gguf')
+    lines.append('hash')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert "RUN_NATIVE: -m /dummy/test.gguf" in res.stdout
+    # Verify hash table ended with healthy binary, not broken
+    assert str(healthy_dir / "llama-server") in res.stdout
+
+
+def test_broken_native_candidate_falls_back_to_healthy_python_compatibility_shim(tmp_path):
+    """Regression: A broken native candidate must not shadow or retire a healthy Python shim."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken_bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    shim_file = tmp_path / "bin/llama-server"
+    _write_odysseus_shim(shim_file)
+
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=True)
+
+    lines = [
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(mock_py_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_reconciliation_lines(lines)
+    lines.append('echo "RESOLVED=$_odysseus_llama_bin"')
+    lines.append('command -v llama-server')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert shim_file.exists(), "Healthy Python compatibility shim must be preserved"
+    assert f"RESOLVED={git_bash_path(shim_file)}" in res.stdout
+    assert str(shim_file) in res.stdout
+
+
+def test_broken_native_and_missing_python_backend_sets_preflight_127(tmp_path):
+    """Regression: When the only native candidates are broken and Python server is unavailable,
+    finalization must fail closed with ODYSSEUS_PREFLIGHT_EXIT=127 rather than claiming readiness."""
+    from core.platform_compat import git_bash_path
+    broken_dir = tmp_path / "broken_bin"
+    _write_mock_native(broken_dir / "llama-server", succeeds=False)
+
+    mock_py_dir = tmp_path / "mock_bin"
+    _write_mock_python3(mock_py_dir / "python3", succeeds=False)
+
+    lines = [
+        'ODYSSEUS_PREFLIGHT_EXIT=""',
+        f'export PATH="$HOME/bin:{git_bash_path(broken_dir)}:{git_bash_path(mock_py_dir)}:/usr/bin:/bin"',
+    ]
+    _append_llama_server_shim_definition_lines(lines)
+    _append_llama_server_shim_reconciliation_lines(lines)
+    _append_llama_server_finalization_lines(lines)
+    lines.append('echo "EXIT=$ODYSSEUS_PREFLIGHT_EXIT"')
+
+    res = _run_shell(tmp_path, lines)
+    assert res.returncode == 0
+    assert "EXIT=127" in res.stdout
+
+
+@pytest.mark.asyncio
+async def test_generated_runner_broken_first_healthy_second_e2e(monkeypatch, tmp_path):
+    """Integration regression: The full generated Cookbook runner script must launch the healthy
+    native candidate even when an earlier broken candidate exists on PATH."""
+    from unittest.mock import AsyncMock, MagicMock
+    from types import SimpleNamespace
+    from routes.cookbook_helpers import ServeRequest
+    import routes.cookbook_routes as cookbook_routes
+    from starlette.requests import Request
+    from core.platform_compat import git_bash_path
+
+    fake_home = tmp_path / "fake_home"
+    stale_shim = fake_home / "bin/llama-server"
+    _write_odysseus_shim(stale_shim)
+
+    broken_bin = fake_home / "broken/bin/llama-server"
+    _write_mock_native(broken_bin, succeeds=False)
+
+    healthy_bin = fake_home / "healthy/bin/llama-server"
+    _write_mock_native(healthy_bin, succeeds=True)
+
+    tmux_log_dir = tmp_path / "tmux_logs"
+    tmux_log_dir.mkdir(parents=True)
+    monkeypatch.setattr(cookbook_routes, "TMUX_LOG_DIR", tmux_log_dir)
+    monkeypatch.setattr(cookbook_routes, "require_admin", lambda request: None)
+    monkeypatch.setattr(cookbook_routes, "_binary_available", AsyncMock(return_value=True))
+    monkeypatch.setattr(cookbook_routes, "load_stored_hf_token", lambda **kwargs: "")
+    monkeypatch.setattr("core.database.SessionLocal", MagicMock())
+
+    class _Proc:
+        returncode = 0
+        stderr = None
+        async def wait(self):
+            return 0
+    monkeypatch.setattr(cookbook_routes.asyncio, "create_subprocess_shell", AsyncMock(return_value=_Proc()))
+
+    router = cookbook_routes.setup_cookbook_routes()
+    endpoint = None
+    for route in router.routes:
+        if route.path == "/api/model/serve" and "POST" in route.methods:
+            endpoint = route.endpoint
+            break
+    assert endpoint is not None
+
+    http_req = Request({
+        "type": "http",
+        "method": "POST",
+        "path": "/api/model/serve",
+        "headers": [],
+        "app": SimpleNamespace(state=SimpleNamespace(auth_manager=SimpleNamespace(
+            is_configured=True, is_admin=lambda user: user == "admin"))),
+    })
+    http_req.state.current_user = "admin"
+    http_req.state.local_model_authority = True
+
+    req = ServeRequest(
+        repo_id="test/repo",
+        cmd="llama-server -m /dummy/model.gguf --port 8090",
+        remote_host=None,
+    )
+
+    resp = await endpoint(http_req, req)
+    assert resp["ok"] is True
+    session_id = resp["session_id"]
+    runner_file = tmux_log_dir / f"{session_id}_run.sh"
+    assert runner_file.exists()
+
+    # Broken is earlier on PATH than healthy
+    res = _run_shell(
+        fake_home,
+        [str(runner_file)],
+        path=f"{git_bash_path(fake_home / 'bin')}:{git_bash_path(broken_bin.parent)}:{git_bash_path(healthy_bin.parent)}:/usr/bin:/bin",
+    )
+    assert res.returncode == 0
+    assert not stale_shim.exists(), "Stale shim must have been retired"
+    assert "RUN_NATIVE: -m /dummy/model.gguf --port 8090" in res.stdout
+    assert "error: missing libllama.so" not in res.stdout
+    assert "Native llama-server not found — building from source" not in res.stdout

tests/test_delete_user_invalidates_token_cache.py

@@ -11,6 +11,8 @@ cache from the DB, where the rows are already gone, and the token is rejected.
 """
 import asyncio
 import types
+import pytest
+from fastapi import HTTPException
 
 from routes.auth_routes import setup_auth_routes, DeleteUserRequest
 
@@ -60,12 +62,9 @@ def test_refused_delete_does_not_invalidate_cache():
     invalidations = []
     router = setup_auth_routes(_auth_manager(delete_result=False))
     handler = _handler(router)
-    try:
+    with pytest.raises(HTTPException) as refused:
         asyncio.run(handler(DeleteUserRequest(username="admin"), _fake_request(invalidations)))
-        raised = False
-    except Exception:
-        raised = True
-    assert raised, "a refused delete should raise (HTTP 400)"
+    assert refused.value.status_code == 400
     assert invalidations == [], "a refused delete must not touch the token cache"
 
 

tests/test_email_list_stale_while_revalidate.py

@@ -56,14 +56,35 @@ def test_cold_start_with_no_snapshot_and_no_index_is_not_marked_stale(monkeypatc
     # falls through to the real (blocking) path, so the response is not a
     # stale-marked snapshot.
     endpoint = _list_endpoint()
+    calls = []
+
+    class EmptyMailbox:
+        def select(self, folder, readonly=False):
+            calls.append(("select", folder, readonly))
+            return "OK", [b"0"]
+
+        def uid(self, command, *args):
+            assert command == "SEARCH"
+            calls.append(("search", args))
+            return "OK", [b""]
+
+        def logout(self):
+            pass
+
+    def connect(*args, **kwargs):
+        calls.append(("connect", args, kwargs))
+        return EmptyMailbox()
+
+    monkeypatch.setattr(er, "_start_poller", lambda: None)
     monkeypatch.setattr(er, "_email_index_list", lambda *a, **k: ([], 0, None))
-    monkeypatch.setattr(er, "_imap_connect", lambda *a, **k: (_ for _ in ()).throw(
-        RuntimeError("offline test")))
-    try:
+    monkeypatch.setattr(er, "_imap_connect", connect)
     result = _call(endpoint)
-    except Exception:
-        return  # the cold path legitimately attempted real IMAP
+    assert [call[0] for call in calls] == ["connect", "select", "search"]
+    assert calls[1][2] is True
     assert isinstance(result, dict)
+    assert result["emails"] == []
+    assert result["total"] == 0
+    assert not result.get("error")
     assert not (result.get("sync") or {}).get("stale")
 
 

tests/test_mcp_manager.py

@@ -39,3 +39,74 @@ def test_http_transport_routes_to_start_http_connect():
         result = asyncio.run(mgr.connect_server("id1", "n", "http", url="https://x/mcp"))
     assert result == "ROUTED"
     m.assert_called_once()
+
+
+def _patch_http_transport(monkeypatch, initialize):
+    """Fake the Streamable HTTP transport and record, for each close, whether
+    it ran in the task that opened it (AnyIO requires that)."""
+    import contextlib
+    import mcp
+    import mcp.client.streamable_http as streamable_http
+    import src.mcp_oauth as mcp_oauth
+
+    closes = []
+
+    @contextlib.asynccontextmanager
+    async def fake_client(url, auth=None):
+        opener = asyncio.current_task()
+        try:
+            yield (object(), object(), lambda: None)
+        finally:
+            closes.append(asyncio.current_task() is opener)
+
+    class FakeSession:
+        def __init__(self, *a):
+            pass
+
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, *exc):
+            return False
+
+    FakeSession.initialize = initialize
+    monkeypatch.setattr(streamable_http, "streamablehttp_client", fake_client)
+    monkeypatch.setattr(mcp, "ClientSession", FakeSession)
+    monkeypatch.setattr(mcp_oauth, "build_provider", lambda *a, **k: None)
+    return closes
+
+
+def test_failed_http_connect_closes_transport_in_its_own_task(monkeypatch):
+    """#5518: a failed connect left its transport for the GC, which exited the
+    AnyIO cancel scope from the wrong task and kept the event loop spinning."""
+    async def initialize(self):
+        raise RuntimeError("Session terminated")
+
+    closes = _patch_http_transport(monkeypatch, initialize)
+    mgr = McpManager()
+
+    ok = asyncio.run(mgr._start_http_connect("s1", "remote", "http://x/mcp/", wait=2))
+
+    assert ok is False
+    assert mgr._connections["s1"]["status"] == "error"
+    assert "Session terminated" in mgr._connections["s1"]["error"]
+    assert closes == [True]
+    assert "s1" not in mgr._stacks
+
+
+def test_cancelled_http_connect_closes_transport(monkeypatch):
+    """Deleting a server mid-connect cancels the connect task; the transport
+    must still be closed rather than leaked."""
+    async def initialize(self):
+        await asyncio.sleep(3600)
+
+    closes = _patch_http_transport(monkeypatch, initialize)
+    mgr = McpManager()
+
+    async def run():
+        await mgr._start_http_connect("s1", "remote", "http://x/mcp/", wait=0.05)
+        await mgr.disconnect_server("s1")
+        await asyncio.sleep(0)
+
+    asyncio.run(run())
+    assert closes == [True]

tests/test_session_actions_cleanup.py

@@ -128,7 +128,7 @@ def test_auto_sort_keeps_fresh_session_while_first_response_is_pending(monkeypat
         db.close()
 
 
-def test_auto_sort_still_deletes_old_throwaway_sessions(monkeypatch):
+def test_auto_sort_keeps_old_nonempty_sessions_even_with_throwaway_title(monkeypatch):
     session_factory = _make_session_factory()
     _install_session_factory(monkeypatch, session_factory)
 
@@ -160,7 +160,8 @@ def test_auto_sort_still_deletes_old_throwaway_sessions(monkeypatch):
 
     db = session_factory()
     try:
-        assert db.query(DbSession).filter(DbSession.id == sid).first() is None
-        assert "Cleaned 1 sessions" in result
+        assert db.query(DbSession).filter(DbSession.id == sid).first() is not None
+        assert db.query(DbMessage).filter(DbMessage.session_id == sid).count() == 1
+        assert "Cleaned 0 sessions" in result
     finally:
         db.close()

tests/test_session_list_owner_scope.py

@@ -19,6 +19,7 @@ from core.database import ChatMessage as DbMessage
 from core.database import Session as DbSession
 
 _TMPDB = tempfile.NamedTemporaryFile(suffix=".db", delete=False)
+_TMPDB.close()  # SQLite owns its connections; release the Windows file handle.
 _ENGINE = create_engine(
     f"sqlite:///{_TMPDB.name}",
     connect_args={"check_same_thread": False},
@@ -230,12 +231,14 @@ def test_patch_session_updates_persisted_cwd(monkeypatch):
 
 def test_auto_sort_skip_llm_cleans_owner_stamped_sessions_when_auth_disabled(monkeypatch):
     import routes.session_routes as sr
+    from fastapi import APIRouter
     from unittest.mock import MagicMock
 
     _stub_multipart_if_missing(monkeypatch)
     monkeypatch.setenv("AUTH_ENABLED", "false")
     monkeypatch.setattr(sr, "SessionLocal", _TS)
     monkeypatch.setattr(sr, "effective_user", lambda request: None)
+    monkeypatch.setattr(sr, "router", APIRouter(prefix="/api"))
 
     sid = str(uuid.uuid4())
     old_time = cdb.utcnow_naive() - timedelta(hours=2)
@@ -250,19 +253,12 @@ def test_auto_sort_skip_llm_cleans_owner_stamped_sessions_when_auth_disabled(mon
             endpoint_url="http://localhost",
             model="gpt-4",
             archived=False,
-            message_count=1,
+            message_count=0,
             created_at=old_time,
             updated_at=old_time,
             last_message_at=old_time,
             last_accessed=old_time,
         ))
-        db.add(DbMessage(
-            id="m-" + uuid.uuid4().hex,
-            session_id=sid,
-            role="user",
-            content="hi",
-            timestamp=old_time,
-        ))
         db.commit()
     finally:
         db.close()
@@ -270,6 +266,13 @@ def test_auto_sort_skip_llm_cleans_owner_stamped_sessions_when_auth_disabled(mon
     session = MagicMock(id=sid, name="New chat", model="gpt-4", endpoint_url="http://localhost", rag=False, archived=False)
     sm = MagicMock()
     sm.get_sessions_for_user.return_value = {sid: session}
+
+    def delete_session(session_id):
+        with _TS() as db:
+            db.query(DbSession).filter(DbSession.id == session_id).delete()
+            db.commit()
+
+    sm.delete_session.side_effect = delete_session
     router = sr.setup_session_routes(sm, {})
     endpoint = next(r.endpoint for r in router.routes
                     if getattr(r, "path", "") == "/api/sessions/auto-sort"
@@ -277,7 +280,8 @@ def test_auto_sort_skip_llm_cleans_owner_stamped_sessions_when_auth_disabled(mon
 
     result = endpoint(request=MagicMock(), skip_llm=True)
 
-    assert result["deleted_throwaway"] == 1
+    assert result["deleted_empty"] == 1
+    assert result["deleted_throwaway"] == 0
     db = _TS()
     try:
         assert db.query(DbSession).filter(DbSession.id == sid).first() is None