config(llm): reprioritize models and switch to gemini-2.5-flash-lite across profiles

docker-compose.yml

@@ -151,6 +151,9 @@ services:
     volumes:
       - ./.env:/app/.env:ro
       - ./llm_config.json:/app/llm_config.json:ro
+      - ./llm_config.premium.json:/app/llm_config.premium.json:ro
+      - ./llm_config.frontier.json:/app/llm_config.frontier.json:ro
+      - ./llm_config.custom.json:/app/llm_config.custom.json:ro
       - ./run:/app/run
     restart: unless-stopped
     develop:
@@ -166,6 +169,15 @@ services:
         - action: sync
           path: ./llm_config.json
           target: /app/llm_config.json
+        - action: sync
+          path: ./llm_config.premium.json
+          target: /app/llm_config.premium.json
+        - action: sync
+          path: ./llm_config.frontier.json
+          target: /app/llm_config.frontier.json
+        - action: sync
+          path: ./llm_config.custom.json
+          target: /app/llm_config.custom.json
         - action: sync
           path: ./.env
           target: /app/.env
@@ -196,6 +208,13 @@ services:
       PLANEXE_FRONTEND_MULTIUSER_ADMIN_PASSWORD: ${PLANEXE_FRONTEND_MULTIUSER_ADMIN_PASSWORD:-admin}
     ports:
       - "${PLANEXE_FRONTEND_MULTIUSER_PORT:-5001}:5000"
+    volumes:
+      - ./.env:/app/.env:ro
+      - ./llm_config.json:/app/llm_config.json:ro
+      - ./llm_config.premium.json:/app/llm_config.premium.json:ro
+      - ./llm_config.frontier.json:/app/llm_config.frontier.json:ro
+      - ./llm_config.custom.json:/app/llm_config.custom.json:ro
+      - ./run:/app/run
     healthcheck:
       test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:5000/healthcheck').read()"]
       interval: 10s

docs/llm_config.md

@@ -1,16 +1,72 @@
 ---
-title: LLM config (llm_config.json)
+title: LLM config profiles
 ---
 
-# LLM config (llm_config.json)
+# LLM config profiles
 
-This file defines which LLM providers and models PlanExe can use. Each top‑level key is a model id used in the UI and pipeline.
+PlanExe supports **4 model profiles**:
 
-`llm_config.json` lives in the PlanExe repo root and is read at runtime. Environment variables are substituted from `.env`.
+- `baseline`
+- `premium`
+- `frontier`
+- `custom`
+
+Each profile maps to a separate config file:
+
+- `baseline` → `llm_config.json`
+- `premium` → `llm_config.premium.json`
+- `frontier` → `llm_config.frontier.json`
+- `custom` → `llm_config.custom.json` (or `PLANEXE_LLM_CONFIG_CUSTOM_FILENAME`)
+
+If the selected profile file is missing or invalid, PlanExe safely falls back to `llm_config.json`.
+
+---
+
+## How profile selection works
+
+### Runtime env var
+
+Set:
+
+- `PLANEXE_MODEL_PROFILE=baseline|premium|frontier|custom`
+
+This is passed end-to-end in worker execution paths (frontend/API/task parameters → worker pipeline).
+
+### Request/task parameter
+
+Task producers (web frontend, MCP) can include:
+
+- `model_profile`
+
+Invalid values are normalized to `baseline`.
+
+---
+
+## Strict filename validation
+
+Config filenames are strictly validated:
+
+- must be a **filename only** (no `/`, `\\`, absolute path)
+- must match: `llm_config*.json`
+
+This prevents path traversal and unsafe file selection.
+
+Legacy override `PLANEXE_LLM_CONFIG_NAME` is still supported for backward compatibility, but profile-based selection is preferred.
+
+---
+
+## Provider-priority ordering per profile
+
+Within each profile config file, priority is defined per model entry:
+
+- lower `priority` value = tried first
+- higher `priority` value = fallback order
+
+`auto` mode uses this profile-specific priority ordering.
 
 ---
 
-## File structure
+## File format (same for all profile files)
 
 ```json
 {
@@ -24,8 +80,6 @@ This file defines which LLM providers and models PlanExe can use. Each top‑lev
       "api_key": "${OPENROUTER_API_KEY}",
       "temperature": 0.1,
       "timeout": 60.0,
-      "is_function_calling_model": false,
-      "is_chat_model": true,
       "max_tokens": 8192,
       "max_retries": 5
     }
@@ -35,41 +89,11 @@ This file defines which LLM providers and models PlanExe can use. Each top‑lev
 
 ---
 
-## Top-level fields
+## Backward compatibility
 
-- **comment**: Plain‑text description for humans. Optional.
-- **priority**: Lower number = higher priority when `auto` is selected. Optional.
-- **luigi_workers**: Number of Luigi workers used for this model. Use `1` for local models (Ollama/LM Studio).
-- **class**: Provider class name (e.g., `OpenRouter`, `OpenAI`, `Ollama`, `LMStudio`, `OpenAILike`).
-- **arguments**: Provider‑specific settings passed to the LLM client.
-
----
-
-## Common arguments
-
-These keys are common across most providers:
-
-- **model** / **model_name**: Provider model identifier.
-- **api_key**: API key reference (usually `${ENV_VAR}`).
-- **base_url** / **api_base**: Override the provider base URL.
-- **temperature**: Controls randomness. Lower is more deterministic.
-- **timeout** / **request_timeout**: Max time per request in seconds.
-- **max_tokens** / **max_completion_tokens**: Output token limit (provider specific).
-- **max_retries**: Retry count on transient errors.
-- **is_function_calling_model**: Whether the model supports structured/tool output.
-- **is_chat_model**: Whether the model uses chat format.
-
----
-
-## Choosing values
-
-- Use **luigi_workers = 1** for local models (Ollama / LM Studio).
-- Use **luigi_workers > 1** for cloud models if you want parallel tasks.
-- Keep **timeout** higher for slower models.
-
----
+When no profile is provided, PlanExe defaults to:
 
-## Notes
+- `baseline`
+- `llm_config.json`
 
-- If `llm_config.json` is missing, PlanExe logs a warning and proceeds with defaults.
-- Changes to `llm_config.json` require a container restart (or rebuild if baked into the image).
+So existing deployments continue to work without changes.

frontend_multi_user/Dockerfile

@@ -13,6 +13,7 @@ WORKDIR /app
 COPY worker_plan/worker_plan_api /app/worker_plan_api
 COPY database_api /app/database_api
 COPY frontend_multi_user /app/frontend_multi_user
+COPY llm_config*.json /app/
 
 # Install dependencies from frontend_multi_user pyproject
 RUN set -eux; \

frontend_multi_user/src/app.py

@@ -53,6 +53,7 @@
 
 from worker_plan_api.planexe_dotenv import DotEnvKeyEnum, PlanExeDotEnv
 from worker_plan_api.planexe_config import PlanExeConfig
+from worker_plan_api.model_profile import ModelProfileEnum, normalize_model_profile
 
 RUN_DIR = "run"
 
@@ -122,6 +123,43 @@ def wrapper(*args, **kwargs):
         return view(*args, **kwargs)
     return wrapper
 
+
+def _profile_model_name_map() -> Dict[str, list[str]]:
+    profile_to_models: Dict[str, list[str]] = {}
+    for profile in ModelProfileEnum:
+        config = PlanExeConfig.load(model_profile_override=profile)
+        config_path = config.llm_config_json_path
+        if config_path is None:
+            profile_to_models[profile.value] = []
+            continue
+        try:
+            with config_path.open("r", encoding="utf-8") as fh:
+                model_map = json.load(fh)
+        except Exception:
+            profile_to_models[profile.value] = []
+            continue
+        if not isinstance(model_map, dict):
+            profile_to_models[profile.value] = []
+            continue
+
+        def sort_key(item: tuple[str, dict]) -> tuple[int, str]:
+            data = item[1] if isinstance(item[1], dict) else {}
+            priority = data.get("priority")
+            if not isinstance(priority, int):
+                priority = 999999
+            return priority, item[0]
+
+        names: list[str] = []
+        for model_id, model_data in sorted(model_map.items(), key=sort_key):
+            model_name = model_id
+            if isinstance(model_data, dict):
+                args = model_data.get("arguments")
+                if isinstance(args, dict) and isinstance(args.get("model"), str):
+                    model_name = args["model"]
+            names.append(model_name)
+        profile_to_models[profile.value] = names
+    return profile_to_models
+
 class MyFlaskApp:
     def __init__(self):
         logger.info(f"MyFlaskApp.__init__. Starting...")
@@ -1887,6 +1925,7 @@ def index():
                 nonce=nonce,
                 user_id=user_id,
                 example_prompts=example_prompts,
+                model_profile_models_json=json.dumps(_profile_model_name_map()),
             )
 
         @self.app.route('/healthcheck')
@@ -2401,6 +2440,12 @@ def run():
             if len(parameters) == 0:
                 parameters = None
 
+            # Normalize model profile to a known value with backward-compatible baseline default.
+            if not isinstance(parameters, dict):
+                parameters = {}
+            raw_profile = parameters.get("model_profile")
+            parameters["model_profile"] = normalize_model_profile(raw_profile).value
+
             # Get length of prompt_param in bytes and in characters
             prompt_param_bytes = len(prompt_param.encode('utf-8'))
             prompt_param_characters = len(prompt_param)
@@ -2502,8 +2547,10 @@ def create_plan():
             parameters.pop('user_id', None)
             parameters.pop('nonce', None)
             parameters.pop('redirect_to_plan', None)
-            if len(parameters) == 0:
-                parameters = None
+
+            # Normalize model profile to a known value with backward-compatible baseline default.
+            raw_profile = parameters.get("model_profile")
+            parameters["model_profile"] = normalize_model_profile(raw_profile).value
 
             prompt_param_bytes = len(prompt_param.encode('utf-8'))
             prompt_param_characters = len(prompt_param)

frontend_multi_user/templates/demo_run.html

@@ -211,6 +211,7 @@ <h1>Demo Run</h1>
         <input type="hidden" name="nonce" value="{{ nonce }}">
         <!-- Values are submitted only when enabled (not disabled) -->
         <input type="hidden" name="speed_vs_detail" id="form-speed-vs-detail" value="ping_llm">
+        <input type="hidden" name="model_profile" id="form-model-profile" value="baseline">
         <input type="hidden" name="developer" id="form-developer" value="true">
     </form>
 
@@ -282,7 +283,7 @@ <h1>Demo Run</h1>
 
             if (methodSelect.value === 'GET') {
                 // GET method: build URL with query parameters
-                let url = `/run?prompt=${encodeURIComponent(promptValue)}&user_id={{ user_id }}&nonce={{ nonce }}&speed_vs_detail=${encodeURIComponent(speedVsDetailValue)}`;
+                let url = `/run?prompt=${encodeURIComponent(promptValue)}&user_id={{ user_id }}&nonce={{ nonce }}&speed_vs_detail=${encodeURIComponent(speedVsDetailValue)}&model_profile=baseline`;
                 if (developerChecked) {
                     url += '&developer';
                 }

frontend_multi_user/templates/index.html

@@ -468,6 +468,21 @@ <h2>Start a New Plan</h2>
         <form id="new-plan-form" method="POST" action="{{ url_for('create_plan') }}">
             <input type="hidden" name="csrf_token" value="{{ csrf_token() }}">
             <input type="hidden" name="speed_vs_detail" value="all_details_but_slow">
+            <label for="model-profile" style="display:block; margin-bottom:8px; font-size:0.9rem; color:#a0aec0;">Model profile</label>
+            <select id="model-profile" name="model_profile" style="margin-bottom:12px; width:100%; max-width:240px; padding:8px; border-radius:8px;">
+                <option value="baseline" selected>baseline (default balanced)</option>
+                <option value="premium">premium (higher-cost ordering)</option>
+                <option value="frontier">frontier (highest-capability ordering)</option>
+                <option value="custom">custom (your custom file)</option>
+            </select>
+            <div style="margin-bottom:12px; font-size:0.85rem; color:#6b7280; line-height:1.4;">
+                baseline -> <code>llm_config.json</code>,
+                premium -> <code>llm_config.premium.json</code>,
+                frontier -> <code>llm_config.frontier.json</code>,
+                custom -> <code>llm_config.custom.json</code> (or <code>PLANEXE_LLM_CONFIG_CUSTOM_FILENAME</code>).
+                The actual models are read from the selected file's priority order.
+            </div>
+            <div id="model-profile-models" style="margin-bottom:12px; font-size:0.85rem; color:#4b5563; line-height:1.4;"></div>
             <textarea name="prompt" id="plan-prompt" placeholder="Describe your project or idea in detail. The more context you provide, the better the plan will be." required></textarea>
             <div class="char-count" id="char-count">0 characters</div>
             <div class="new-plan-footer">
@@ -586,4 +601,32 @@ <h3>Avoid Surprises</h3>
     });
 </script>
 {% endif %}
+{% if user %}
+<script>
+    var profileToModels = {{ model_profile_models_json | safe }};
+    var profileSelect = document.getElementById('model-profile');
+    var profileModelsDiv = document.getElementById('model-profile-models');
+
+    function renderProfileModels() {
+        if (!profileSelect || !profileModelsDiv) {
+            return;
+        }
+        var profile = profileSelect.value || 'baseline';
+        var models = profileToModels[profile] || [];
+        if (models.length === 0) {
+            profileModelsDiv.innerHTML = '<strong>Models in ' + profile + ':</strong> none found';
+            return;
+        }
+        var lines = models.map(function(modelName) {
+            return '<li><code>' + modelName + '</code></li>';
+        }).join('');
+        profileModelsDiv.innerHTML = '<strong>Models in ' + profile + ':</strong><ul style="margin:6px 0 0 16px;">' + lines + '</ul>';
+    }
+
+    if (profileSelect) {
+        profileSelect.addEventListener('change', renderProfileModels);
+    }
+    renderProfileModels();
+</script>
+{% endif %}
 {% endblock %}

frontend_single_user/Dockerfile

@@ -19,7 +19,7 @@ RUN pip install --no-cache-dir --upgrade pip \
 # Copy application code and supporting files
 COPY worker_plan/worker_plan_api /app/worker_plan_api
 COPY frontend_single_user /app/frontend_single_user
-COPY llm_config.json /app/
+COPY llm_config*.json /app/
 
 # Default location for generated plans
 RUN mkdir -p /app/run