fix(cookbook): harden local llama-server launch commands

Author: zoomdbz

routes/cookbook_routes.py

@@ -1386,6 +1386,10 @@ async def model_serve(request: Request, req: ServeRequest):
             # shell resolves the bundled python3/hf, mirroring the download flow.
             if not remote:
                 runner_lines.append(_local_tooling_path_export(sys.executable))
+                if local_windows:
+                    # Detached Git Bash runs do not always inherit recently edited
+                    # user PATH entries from the already-running Odysseus process.
+                    runner_lines.append('export PATH="$HOME/bin:$HOME/llama.cpp/build-cuda/bin/Release:$HOME/llama.cpp/build/bin/Release:$HOME/llama.cpp/build/bin/Debug:$HOME/llama.cpp/build/bin:$PATH"')
             runner_lines.append("export FLASHINFER_DISABLE_VERSION_CHECK=1")
             if req.hf_token:
                 runner_lines.append(f"export HF_TOKEN='{_bash_squote(req.hf_token)}'")

static/js/cookbook.js

@@ -2111,14 +2111,14 @@ function _renderRecipes() {
     const isLocal = !s.host || s.host.toLowerCase() === 'local';
     if (isLocal) {
       s.host = '';
-      s.platform = _envState.hostPlatform || _envState.platform || '';
+      s.platform = _envState.hostPlatform || '';
       if (_localSeen) return false;
       _localSeen = true;
     }
     return true;
   });
   if (!_localSeen) {
-    _es.servers.unshift({ host: '', env: _es.env || 'none', envPath: _es.envPath || '', modelDir: '~/.cache/huggingface/hub', platform: _envState.hostPlatform || _envState.platform || '' });
+    _es.servers.unshift({ host: '', env: _es.env || 'none', envPath: _es.envPath || '', modelDir: '~/.cache/huggingface/hub', platform: _envState.hostPlatform || '' });
   }
   if (_es.remoteHost && !_es.servers.some(s => s.host === _es.remoteHost)) {
     _es.servers.push({ host: _es.remoteHost, env: _es.env || 'none', envPath: _es.envPath || '', modelDir: '~/.cache/huggingface/hub' });

static/js/cookbookRunning.js

@@ -1522,7 +1522,7 @@ export async function _launchServeTask(shortName, repo, cmd, fields, hostOverrid
   const _host = (hostOverride !== undefined) ? (hostOverride || '') : (_envState.remoteHost || '');
   const _hsrv = _serverByVal(_envState.remoteServerKey || _host)
     || _envState.servers.find(s => s.host === _host) || {};
-  const _hplatform = _host ? (_hsrv.platform || '') : (_envState.platform || '');
+  const _hplatform = _host ? (_hsrv.platform || '') : (_envState.hostPlatform || '');
 
   // Replace any serve already targeting this same host:port — you can't run two
   // servers on one port, so re-serving (or retrying) should stop & remove the

static/js/cookbookServe.js

@@ -119,7 +119,7 @@ function _selectedServeTarget(panel) {
     port: host ? (server?.port || _getPort(host) || '') : '',
     env: server?.env || '',
     venv,
-    platform: server?.platform || _envState.platform || '',
+    platform: host ? (server?.platform || '') : (_envState.hostPlatform || ''),
     label,
   };
 }
@@ -240,6 +240,12 @@ function _runnableGgufFiles(model) {
   return primary.length ? primary : files;
 }
 
+function _projectorGgufFiles(model) {
+  return _ggufFilesForModel(model)
+    .filter(f => (f.role || '') === 'projector' || /(^|\/)mmproj[^/]*\.gguf$/i.test(f.rel_path || f.name || ''))
+    .sort((a, b) => String(a.rel_path || a.name || '').localeCompare(String(b.rel_path || b.name || '')));
+}
+
 function _ggufFileLabel(file) {
   const base = (file.name || file.rel_path || '').split('/').pop();
   const size = _formatGgufSize(file.size_bytes);
@@ -621,6 +627,7 @@ function _rerenderCachedModels() {
           : `This model's download isn't complete yet (${esc(m.size || 'partial')}). The serve will start but is likely to crash on a missing shard. Wait for the download to finish, or relaunch after it's done.`;
         panelHtml += `<div class="hwfit-serve-warn" style="margin:0 0 8px;padding:6px 10px;border-radius:5px;font-size:11px;background:color-mix(in srgb, var(--color-warning, #f0ad4e) 14%, transparent);border:1px solid color-mix(in srgb, var(--color-warning, #f0ad4e) 40%, transparent);color:var(--color-warning, #f0ad4e);display:flex;gap:6px;align-items:flex-start;line-height:1.4;"><span aria-hidden="true">⚠</span><span>${_warnText}</span></div>`;
       }
+      panelHtml += `<div class="hwfit-serve-vision-warn" style="display:none;margin:0 0 8px;padding:6px 10px;border-radius:5px;font-size:11px;background:color-mix(in srgb, var(--color-warning, #f0ad4e) 14%, transparent);border:1px solid color-mix(in srgb, var(--color-warning, #f0ad4e) 40%, transparent);color:var(--color-warning, #f0ad4e);gap:6px;align-items:flex-start;line-height:1.4;"><span aria-hidden="true">⚠</span><span>Vision is enabled, but no mmproj GGUF projector was found in the cached model scan. Download an mmproj-*.gguf for this model, then refresh the cached model list before launching.</span></div>`;
       // Row 1: Backend + Server + Env
       panelHtml += `<div class="hwfit-serve-row">`;
       const backendOpts = _backendChoices.map(([v,l]) => `<option value="${v}"${defaultBackend===v?' selected':''}>${l}</option>`).join('');
@@ -896,18 +903,22 @@ function _rerenderCachedModels() {
             : m.is_local_dir && m.path
             ? `$({ find ${_ldir} -name '*-00001-of-*.gguf' 2>/dev/null | sort; find ${_ldir} -name '*.gguf' 2>/dev/null | sort; } | head -1)`
             : `$({ find ${dir} -name '*-00001-of-*.gguf' 2>/dev/null | sort; find ${dir} -name '*.gguf' 2>/dev/null | sort; } | head -1)`;
-          // Vision: auto-find the mmproj (CLIP/projector) file in the same dir.
-          // Resolved at runtime so the toggle just works if an mmproj-*.gguf is
-          // present (downloaded alongside the model). Empty if none → cmd omits it.
-          const _vsearchdir = (m.is_local_dir && m.path) ? _ldir : dir;
-          f._mmproj_path = `$(find ${_vsearchdir} -iname 'mmproj*.gguf' 2>/dev/null | sort | head -1)`;
+          // Vision: use the scanned projector (CLIP/mmproj) file when present.
+          // Keeping this as a printf path avoids generating a command substitution
+          // that the backend serve-command validator must reject as unsafe.
+          const selectedProjector = _projectorGgufFiles(m)[0];
+          f._mmproj_path = selectedProjector ? _selectedGgufExpr(m, repo, selectedProjector.rel_path) : '';
         }
         if (f.reasoning_parser) {
           const _rpEl2 = panel.querySelector('[data-field="reasoning_parser"]');
           f._reasoning_parser_value = _rpEl2?.dataset?.parser || 'qwen3';
         }
         let cmd = _buildServeCmd(f, serveModel, backend);
         if (f.extra && f.extra.trim()) cmd += ' ' + f.extra.trim();
+        const missingVisionProjector = backend === 'llamacpp' && !!f.vision && !f._mmproj_path;
+        panel._visionMissingProjector = missingVisionProjector;
+        const _visionWarn = panel.querySelector('.hwfit-serve-vision-warn');
+        if (_visionWarn) _visionWarn.style.display = missingVisionProjector ? 'flex' : 'none';
         const _ce2 = panel.querySelector('.hwfit-serve-cmd'); _ce2.value = cmd; _ce2.style.height = 'auto'; _ce2.style.height = _ce2.scrollHeight + 'px';
         panel._cmd = cmd;
         panel._host = f.host || '';
@@ -1982,12 +1993,16 @@ function _rerenderCachedModels() {
         });
         serveState.backend = serveState.backend || (_detectBackend(m).backend) || 'vllm';
         const launchTarget = _selectedServeTarget(panel);
+        if (serveState.backend === 'llamacpp' && serveState.vision && !/(?:^|\s)(?:--mmproj|--clip_model_path)\b/.test(launchCmd)) {
+          _restoreLaunchBtn();
+          uiModule.showToast('Vision is checked, but no mmproj projector is in the launch command. Refresh cached models after downloading mmproj, or add --mmproj manually.', 8000);
+          return;
+        }
         if (serveState.backend === 'diffusers' && _remoteWindowsDiffusersUnsupported(launchTarget)) {
           _restoreLaunchBtn();
           uiModule.showToast('Diffusers serving is not supported on remote Windows servers yet. Use local Windows or a Linux server.', 9000);
           return;
         }
-
         // Pre-launch: check our own task list for a serve already running
         // on this host. Offer to stop+launch as the default action — the
         // SSH-based port probe below is more thorough but it can miss

tests/test_cookbook_cpu_only_serve.py

@@ -67,13 +67,27 @@ def test_local_windows_platform_comes_from_backend_host_state():
     assert "hostOrTask === 'local'" in text
     assert "if (hostOrTask === 'local') return _envState.hostPlatform || '';" in text
     assert "return _envState.hostPlatform || _envState.platform || ''" not in text
+    assert "s.platform = _envState.hostPlatform || '';" in text
+    assert "platform: _envState.hostPlatform || ''" in text
+    assert "s.platform = _envState.hostPlatform || _envState.platform || '';" not in text
+    assert "platform: _envState.hostPlatform || _envState.platform || ''" not in text
     assert 'return "windows" if IS_WINDOWS else ""' in routes
     assert 'env["hostPlatform"] = _client_host_platform()' in routes
     assert "return _state_for_client({})" in routes
     assert 'env.pop("hostPlatform", None)' in routes
     assert "delete env.hostPlatform;" in running
 
 
+def test_local_serve_payload_ignores_stale_env_platform():
+    serve = SERVE_SRC.read_text(encoding="utf-8")
+    running = (SRC.parent / "cookbookRunning.js").read_text(encoding="utf-8")
+
+    assert "platform: host ? (server?.platform || '') : (_envState.hostPlatform || '')," in serve
+    assert "platform: server?.platform || _envState.platform || ''" not in serve
+    assert "const _hplatform = _host ? (_hsrv.platform || '') : (_envState.hostPlatform || '');" in running
+    assert "const _hplatform = _host ? (_hsrv.platform || '') : (_envState.platform || '');" not in running
+
+
 def test_local_windows_llamacpp_prefers_native_llama_server():
     text = SRC.read_text(encoding="utf-8")
     helpers = (ROOT / "routes/cookbook_helpers.py").read_text(encoding="utf-8")
@@ -90,4 +104,28 @@ def test_local_windows_llama_server_skips_source_bootstrap():
     routes = (ROOT / "routes/cookbook_routes.py").read_text(encoding="utf-8")
 
     assert 'local_windows_llama_cmd = local_windows and ("llama_cpp" in req.cmd or "llama-server" in req.cmd)' in routes
-    assert 'if ("llama_cpp" in req.cmd or "llama-server" in req.cmd) and not local_windows_llama_cmd:' in routes
+    assert 'if ("llama_cpp" in req.cmd or "llama-server" in req.cmd) and not local_windows_llama_cmd:' in routes
+
+
+def test_local_windows_llama_server_path_includes_user_wrapper_and_cuda_builds():
+    routes = (ROOT / "routes/cookbook_routes.py").read_text(encoding="utf-8")
+
+    assert 'if local_windows:' in routes
+    assert (
+        'export PATH="$HOME/bin:$HOME/llama.cpp/build-cuda/bin/Release:'
+        '$HOME/llama.cpp/build/bin/Release:$HOME/llama.cpp/build/bin/Debug:'
+        '$HOME/llama.cpp/build/bin:$PATH"'
+    ) in routes
+
+
+def test_llamacpp_vision_uses_scanned_projector_instead_of_runtime_find():
+    text = SERVE_SRC.read_text(encoding="utf-8")
+
+    assert "function _projectorGgufFiles(model)" in text
+    assert "const selectedProjector = _projectorGgufFiles(m)[0];" in text
+    assert "f._mmproj_path = selectedProjector ? _selectedGgufExpr(m, repo, selectedProjector.rel_path) : '';" in text
+    assert "const missingVisionProjector = backend === 'llamacpp' && !!f.vision && !f._mmproj_path;" in text
+    assert "hwfit-serve-vision-warn" in text
+    assert "!/(?:^|\\s)(?:--mmproj|--clip_model_path)\\b/.test(launchCmd)" in text
+    assert "no mmproj projector is in the launch command" in text
+    assert "find ${_vsearchdir} -iname 'mmproj*.gguf'" not in text

tests/test_cookbook_helpers.py

@@ -419,8 +419,6 @@ def test_pip_install_attempt_failure_propagates_real_exit_code():
     """Run the generated snippet against a deliberately broken pip install
     to confirm the subshell exits with pip's non-zero status."""
     snippet = _pip_install_attempt("python3 -m pip install __nonexistent_package_12345__")
-    if sys.platform == "win32":
-        snippet = snippet.replace("$", "\\$")
     result = subprocess.run(
         ["bash", "-c", snippet],
         capture_output=True,
@@ -433,8 +431,6 @@ def test_pip_install_attempt_failure_propagates_real_exit_code():
 def test_pip_install_attempt_success_exits_zero():
     """When pip succeeds, the subshell should exit 0."""
     snippet = _pip_install_attempt("python3 -c 'pass'")
-    if sys.platform == "win32":
-        snippet = snippet.replace("$", "\\$")
     result = subprocess.run(
         ["bash", "-c", snippet],
         capture_output=True,
@@ -447,8 +443,6 @@ def test_pip_install_attempt_success_exits_zero():
 def test_pip_install_attempt_surfaces_stderr_on_failure():
     """On failure, the last 5 lines of pip output should appear in stdout."""
     snippet = _pip_install_attempt("python3 -m pip install __nonexistent_package_12345__")
-    if sys.platform == "win32":
-        snippet = snippet.replace("$", "\\$")
     result = subprocess.run(
         ["bash", "-c", snippet],
         capture_output=True,
@@ -557,6 +551,19 @@ def test_validate_serve_cmd_accepts_windows_printf_format():
     assert _validate_serve_cmd(cmd) == cmd
 
 
+def test_validate_serve_cmd_accepts_llama_mmproj_printf_format():
+    cmd = (
+        "CUDA_VISIBLE_DEVICES=0 llama-server --model "
+        "\"$(printf %s ${HOME}'/.cache/huggingface/hub/models--unsloth--Qwen3.6-35B-A3B-GGUF/snapshots/abc/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf')\" "
+        "--host 0.0.0.0 --port 8000 -ngl 99 -c 20000 "
+        "--cache-type-k q4_0 --cache-type-v q4_0 --mmproj "
+        "\"$(printf %s ${HOME}'/.cache/huggingface/hub/models--unsloth--Qwen3.6-35B-A3B-GGUF/snapshots/abc/mmproj-BF16.gguf')\" "
+        "--image-max-tokens 1024"
+    )
+
+    assert _validate_serve_cmd(cmd) == cmd
+
+
 def test_normalize_llama_cpp_python_cache_types_for_stale_client_cmd():
     cmd = (
         "python -m llama_cpp.server --model model.gguf --host 0.0.0.0 --port 8000 "