feat: OpenAI-compatible orchestrator + backend auto-routing

- openai_orchestrator.py — new ReAct tool loop engine for any
  OpenAI-compatible endpoint (OpenRouter, Open WebUI, Ollama, LiteLLM);
  model handles both tool loop and final response, no Claude handoff needed
- tools/__init__.py — auto-derive OpenAI JSON Schema from existing Gemini
  FunctionDeclarations so tool definitions have a single source of truth
- routers/orchestrator.py — route to openai_orchestrator when model registry
  "orchestrator" role resolves to a local_openai type host
- routers/chat.py — pass role to _backend_label(); fix fallback_used logic
  (only meaningful for explicit backend overrides, not auto-routing)
- static/app.js — add null/"auto" to backend cycle; fetch local model hint
  without overriding the auto default on page load
- model_registry.py — _normalize() back-fills host_type on old registry files
- requirements.txt — add openai>=1.0.0
- ARCH__BACKENDS.md — document OpenAI-compat backend and routing logic

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cortex/model_registry.py

@@ -122,13 +122,20 @@ def _empty() -> dict:
     return {"version": 1, "hosts": [], "models": [], "roles": {}}
 
 
+def _normalize(data: dict) -> dict:
+    """Back-fill any missing fields introduced by schema additions."""
+    for h in data.get("hosts", []):
+        h.setdefault("host_type", "openwebui")
+    return data
+
+
 def _load(username: str) -> dict:
     path = _registry_path(username)
     if path.exists():
         try:
             data = json.loads(path.read_text())
             if isinstance(data, dict) and "version" in data:
-                return data
+                return _normalize(data)
         except (json.JSONDecodeError, OSError):
             logger.warning("model_registry.json for %s is unreadable — starting fresh", username)
         return _empty()

cortex/openai_orchestrator.py

@@ -0,0 +1,196 @@
+"""
+OpenAI-compatible orchestrator engine.
+
+Implements the same ReAct tool loop as orchestrator_engine.py but uses the
+OpenAI tool calling format, which works with any OpenAI-compatible endpoint:
+OpenRouter, LiteLLM, Open WebUI, Ollama (tool-capable models), etc.
+
+The model both runs the tool loop AND writes the final user-facing response —
+no separate handoff step needed when a single capable model handles everything.
+
+Flow:
+  1. POST to {api_url}/chat/completions with tools + user message
+  2. If finish_reason == "tool_calls": execute tools, feed results back, repeat
+  3. If finish_reason == "stop": final assistant message is the user-facing response
+
+Used when the "orchestrator" role in the model registry resolves to a local_openai
+type model. The Gemini engine (orchestrator_engine.py) is used otherwise.
+"""
+
+import asyncio
+import json
+import logging
+
+from openai import AsyncOpenAI
+
+from config import settings
+from orchestrator_engine import OrchestratorResult
+from tools import OPENAI_TOOL_SCHEMAS, call_tool
+
+logger = logging.getLogger(__name__)
+
+# Appended to the persona system prompt so the model knows it has tools.
+# Kept brief — capable models handle tool use without much coaching.
+_TOOL_INSTRUCTION = (
+    "\n\nYou have access to tools. Use them when you need current information, "
+    "need to read files, or need to take actions on the user's behalf. "
+    "Respond naturally after gathering what you need."
+)
+
+
+async def run(
+    task: str,
+    system_prompt: str = "",
+    session_messages: list[dict] | None = None,
+    model_cfg: dict | None = None,
+    respond_with_final: bool = True,
+) -> OrchestratorResult:
+    """
+    Run a tool-enabled task using an OpenAI-compatible API.
+
+    Args:
+        task:               The user's request (plain text)
+        system_prompt:      Persona system prompt from context_loader (passed through)
+        session_messages:   Recent conversation history for session continuity
+        model_cfg:          Resolved model config from model_registry (local_openai type)
+        respond_with_final: If False, return just the tool-loop summary without a
+                            full persona-voiced response (faster; for cron/background)
+
+    Returns:
+        OrchestratorResult — same shape as the Gemini engine for drop-in compatibility
+    """
+    if not model_cfg:
+        raise RuntimeError("model_cfg is required for the OpenAI orchestrator")
+
+    api_url   = model_cfg.get("api_url", "")
+    api_key   = model_cfg.get("api_key", "") or "none"
+    model_name = model_cfg.get("model_name", "")
+
+    if not api_url or not model_name:
+        raise RuntimeError(
+            f"model_cfg missing api_url or model_name: {model_cfg.get('label', model_cfg)}"
+        )
+
+    client = AsyncOpenAI(base_url=api_url, api_key=api_key)
+
+    # System prompt: persona context + brief tool instruction
+    sys_content = (system_prompt or "") + _TOOL_INSTRUCTION
+
+    # Build messages: [system, ...recent_session, current_task]
+    messages: list[dict] = [{"role": "system", "content": sys_content}]
+    if session_messages:
+        messages.extend(session_messages[-6:])   # last 3 turns for context
+    messages.append({"role": "user", "content": task})
+
+    tool_call_log: list[dict] = []
+    final_response = ""
+
+    for round_num in range(settings.orchestrator_max_rounds):
+        logger.info("OpenAI orchestrator round %d / %d  model=%s",
+                    round_num + 1, settings.orchestrator_max_rounds, model_name)
+
+        response = await client.chat.completions.create(
+            model=model_name,
+            messages=messages,
+            tools=OPENAI_TOOL_SCHEMAS,
+            tool_choice="auto",
+        )
+
+        choice = response.choices[0]
+        msg    = choice.message
+
+        # Append the assistant turn (MUST include tool_calls if present so the
+        # next request is valid — OpenAI requires the full history to be consistent)
+        assistant_msg: dict = {"role": "assistant"}
+        if msg.content:
+            assistant_msg["content"] = msg.content
+        if msg.tool_calls:
+            assistant_msg["tool_calls"] = [
+                {
+                    "id":   tc.id,
+                    "type": "function",
+                    "function": {
+                        "name":      tc.function.name,
+                        "arguments": tc.function.arguments,
+                    },
+                }
+                for tc in msg.tool_calls
+            ]
+        messages.append(assistant_msg)
+
+        if choice.finish_reason == "tool_calls" and msg.tool_calls:
+            # Execute all tool calls in parallel, then feed results back
+            tool_tasks = [
+                _execute_tool(tc.function.name, tc.function.arguments)
+                for tc in msg.tool_calls
+            ]
+            results = await asyncio.gather(*tool_tasks, return_exceptions=True)
+
+            for tc, result in zip(msg.tool_calls, results):
+                result_str = (
+                    str(result)
+                    if not isinstance(result, Exception)
+                    else f"Tool error: {result}"
+                )
+                logger.info("Tool %s → %d chars", tc.function.name, len(result_str))
+
+                try:
+                    args_parsed = json.loads(tc.function.arguments)
+                except json.JSONDecodeError:
+                    args_parsed = {"raw": tc.function.arguments}
+
+                tool_call_log.append({
+                    "tool":   tc.function.name,
+                    "args":   args_parsed,
+                    "result": result_str,
+                })
+
+                # Tool result message — tools array must be re-sent on every request
+                messages.append({
+                    "role":         "tool",
+                    "tool_call_id": tc.id,
+                    "content":      result_str,
+                })
+
+        else:
+            # finish_reason == "stop" (or no tool_calls) — model is done
+            final_response = msg.content or ""
+            logger.info(
+                "OpenAI orchestrator done after %d round(s). Tools used: %d",
+                round_num + 1, len(tool_call_log),
+            )
+            break
+
+    else:
+        # Hit the round limit
+        logger.warning("OpenAI orchestrator hit max rounds (%d)", settings.orchestrator_max_rounds)
+        final_response = (
+            f"Reached the tool iteration limit ({settings.orchestrator_max_rounds} rounds). "
+            "Here is what was gathered:\n\n"
+            + "\n\n".join(
+                f"**{t['tool']}**: {t['result'][:500]}" for t in tool_call_log
+            )
+        )
+
+    model_label = model_cfg.get("label") or model_name
+    logger.info("OpenAI orchestrator complete — model=%s tools=%d", model_label, len(tool_call_log))
+
+    return OrchestratorResult(
+        response=final_response,
+        tool_calls=tool_call_log,
+        backend="local",
+        gemini_summary=final_response,  # reused for UI display; same content in single-model mode
+    )
+
+
+async def _execute_tool(name: str, arguments_json: str) -> str:
+    """Parse tool arguments and execute, returning a string result."""
+    try:
+        args = json.loads(arguments_json)
+    except json.JSONDecodeError:
+        args = {}
+    try:
+        return await call_tool(name, args)
+    except Exception as e:
+        logger.warning("Tool %s failed: %s", name, e)
+        return f"Tool error: {e}"

cortex/requirements.txt

@@ -19,5 +19,8 @@ python-multipart>=0.0.9   # required by FastAPI for Form() data
 # Async HTTP client — used for local OpenAI-compatible backend (Open WebUI / Ollama)
 httpx>=0.27.0
 
+# OpenAI-compatible client — tool calling for OpenRouter / LiteLLM / any OAI-compat host
+openai>=1.0.0
+
 # anthropic SDK not needed — using claude CLI subprocess for auth
 # anthropic>=0.40.0

cortex/routers/chat.py

@@ -18,14 +18,14 @@ import event_bus
 router = APIRouter()
 
 
-def _backend_label(backend: str, username: str) -> str:
+def _backend_label(backend: str, username: str, role: str = "chat") -> str:
     """Human-readable label for the model that handled a request."""
     if backend == "claude":
         return "Claude"
     if backend == "gemini":
         return "Gemini"
     if backend == "local":
-        cfg = model_registry.get_best_local_model(username)
+        cfg = model_registry.get_best_local_model(username, role)
         if cfg:
             return cfg.get("label") or cfg.get("model_name") or "Local"
         return "Local"
@@ -113,14 +113,16 @@ async def _stream_chat(req: ChatRequest):
             if not req.off_record:
                 log_turn(session_id, req.message, response_text)
 
-            requested = req.model or settings.primary_backend
+            # fallback_used only makes sense for explicit backend selections.
+            # In auto mode (req.model is None), just report what responded.
+            fallback_used = bool(req.model and actual_backend != req.model)
             payload = {
                 "type": "response",
                 "response": response_text,
                 "session_id": session_id,
                 "backend": actual_backend,
-                "backend_label": _backend_label(actual_backend, user),
-                "fallback_used": actual_backend != requested,
+                "backend_label": _backend_label(actual_backend, user, role="chat"),
+                "fallback_used": fallback_used,
             }
             yield f"data: {json.dumps(payload)}\n\n"
 

cortex/routers/orchestrator.py

@@ -22,7 +22,9 @@ from auth_utils import get_user_gemini_key
 from config import settings
 from context_loader import load_context
 from persona import set_context, validate as validate_persona
+import model_registry
 import orchestrator_engine
+import openai_orchestrator
 
 logger = logging.getLogger(__name__)
 router = APIRouter(prefix="/orchestrate", tags=["orchestrator"])
@@ -157,13 +159,25 @@ async def _run_job(job_id: str, req: OrchestrateRequest, user: str) -> None:
         history = load_session(session_id)
         session_messages = history or None
 
-        result = await orchestrator_engine.run(
-            task=req.task,
-            system_prompt=system_prompt,
-            session_messages=session_messages,
-            respond_with_claude=req.respond_with_claude,
-            gemini_api_key=get_user_gemini_key(user),
-        )
+        # Choose engine based on the orchestrator role in the model registry
+        orch_model = model_registry.get_model_for_role(user, "orchestrator")
+
+        if orch_model and orch_model.get("type") == "local_openai":
+            result = await openai_orchestrator.run(
+                task=req.task,
+                system_prompt=system_prompt,
+                session_messages=session_messages,
+                model_cfg=orch_model,
+                respond_with_final=req.respond_with_claude,
+            )
+        else:
+            result = await orchestrator_engine.run(
+                task=req.task,
+                system_prompt=system_prompt,
+                session_messages=session_messages,
+                respond_with_claude=req.respond_with_claude,
+                gemini_api_key=get_user_gemini_key(user),
+            )
 
         # Save the turn to the session store so it survives a page refresh
         history.append({"role": "user", "content": req.task})

cortex/static/app.js

@@ -84,7 +84,7 @@
         if (helpLink) helpLink.href = `/help?persona=${encodeURIComponent(CORTEX_PERSONA)}`;
 
         let sessionId        = null;
-        let primaryBackend   = 'claude';
+        let primaryBackend   = null;   // null = auto / role-based routing
         let activeController = null;
         let currentHistory   = [];  // mirrors backend session [{role, content}, ...]
         let talkThinkingDiv  = null; // pending "thinking…" bubble for live Talk updates
@@ -340,23 +340,30 @@
         }
 
         // ── Backend toggle ───────────────────────────────────────────
+        // null = "auto" — uses role-based routing from model registry
+        // 'claude' / 'gemini' / 'local' = explicit override
 
-        fetch('/backend').then(r => r.json()).then(d => setBackendUI(d));
+        // On load only fetch local_model hint; don't override primaryBackend default (null)
+        fetch('/backend').then(r => r.json()).then(d => {
+            if (backendModelHint && d.local_model) {
+                // Pre-fill hint in case user is already in local mode
+                backendModelHint.textContent = d.local_model.label || d.local_model.model_name;
+            }
+        });
 
-        const BACKEND_CYCLE = ['claude', 'gemini', 'local'];
+        const BACKEND_CYCLE = [null, 'claude', 'gemini', 'local'];
         const BACKEND_CLASS = { claude: '', gemini: 'mem-on', local: 'local-on' };
         const backendModelHint = document.getElementById('backend-model-hint');
 
-        function setBackendUI(d) {
-            const backend = d.primary || d;  // accept full response obj or bare string
+        function setBackendUI(backend, localModel) {
             primaryBackend = backend;
-            backendToggle.textContent = backend;
-            const extra = BACKEND_CLASS[backend] || '';
+            backendToggle.textContent = backend === null ? 'auto' : backend;
+            const extra = backend === null ? '' : (BACKEND_CLASS[backend] || '');
             backendToggle.className = 'ctx-btn' + (extra ? ' ' + extra : '');
 
             if (backendModelHint) {
-                if (backend === 'local' && d.local_model) {
-                    backendModelHint.textContent = d.local_model.label || d.local_model.model_name;
+                if (backend === 'local' && localModel) {
+                    backendModelHint.textContent = localModel.label || localModel.model_name;
                     backendModelHint.style.display = '';
                 } else {
                     backendModelHint.textContent = '';
@@ -365,17 +372,26 @@
             }
         }
 
+        // Initialize to auto mode
+        setBackendUI(null, null);
+
         backendToggle.addEventListener('click', async () => {
             const idx = BACKEND_CYCLE.indexOf(primaryBackend);
             const next = BACKEND_CYCLE[(idx + 1) % BACKEND_CYCLE.length];
-            const res = await fetch('/backend', {
-                method: 'POST',
-                headers: { 'Content-Type': 'application/json' },
-                body: JSON.stringify({ primary: next }),
-            });
-            const d = await res.json();
-            setBackendUI(d);
-            addMessage('system', `Backend: ${d.primary} (fallback: ${d.fallback})`);
+            if (next === null) {
+                // Auto: role-based routing — no server call needed
+                setBackendUI(null, null);
+                addMessage('system', 'Backend: auto (role-based routing)');
+            } else {
+                const res = await fetch('/backend', {
+                    method: 'POST',
+                    headers: { 'Content-Type': 'application/json' },
+                    body: JSON.stringify({ primary: next }),
+                });
+                const d = await res.json();
+                setBackendUI(next, d.local_model);
+                addMessage('system', `Backend: ${next} (fallback: ${d.fallback})`);
+            }
         });
 
         // ── Sessions panel ───────────────────────────────────────────
@@ -917,42 +933,15 @@
             if (activeController) activeController.abort();
         });
 
-        async function sendMessage() {
-            const text = inputEl.value.trim();
-            if (!text || activeController) return;
-
-            inputEl.value = '';
-            syncHeight();
-            sendBtn.style.display = 'none';
-            stopBtn.style.display = 'flex';
-            headerEmoji.classList.add('processing');
-
-            activeController = new AbortController();
-
-            const userHistIdx = currentHistory.length;
-            currentHistory.push({ role: 'user', content: text });
-            const userMsgDiv = addMessage('user', text);
-            attachHistoryControls(userMsgDiv, userHistIdx);
-            scrollToBottom();
-
-            const thinkingDiv = addMessage('assistant thinking', '✨ thinking…');
-
+        // ── Chat fetch + SSE handler ─────────────────────────────────
+        // Extracted so the retry button can call it without re-adding the
+        // user message to the DOM or currentHistory.
+        async function _doSend(payload, thinkingDiv) {
             try {
                 const res = await fetch('/chat', {
                     method: 'POST',
                     headers: { 'Content-Type': 'application/json' },
-                    body: JSON.stringify({
-                        message: text,
-                        session_id: sessionId,
-                        tier: currentTier,
-                        include_long: memLong,
-                        include_mid: memMid,
-                        include_short: memShort,
-                        off_record: current_mode === 'otr',
-                        model: primaryBackend,
-                        user: CORTEX_USER,
-                        persona: CORTEX_PERSONA,
-                    }),
+                    body: JSON.stringify(payload),
                     signal: activeController.signal,
                 });
 
@@ -1004,10 +993,77 @@
                     thinkingDiv.className = 'message system';
                     thinkingDiv.textContent = 'Stopped.';
                 } else {
+                    // Show error + retry button
                     thinkingDiv.className = 'message error';
-                    thinkingDiv.textContent = `Error: ${err.message}`;
+                    thinkingDiv.innerHTML = '';
+
+                    const errSpan = document.createElement('span');
+                    errSpan.textContent = `Error: ${err.message}`;
+                    thinkingDiv.appendChild(errSpan);
+
+                    const retryBtn = document.createElement('button');
+                    retryBtn.className = 'retry-btn';
+                    retryBtn.textContent = '↺ Retry';
+                    retryBtn.addEventListener('click', async () => {
+                        // Roll back the failed user push, re-push, and try again
+                        if (currentHistory.at(-1)?.role === 'user') currentHistory.pop();
+                        currentHistory.push({ role: 'user', content: payload.message });
+
+                        thinkingDiv.className = 'message assistant thinking';
+                        thinkingDiv.textContent = '✨ thinking…';
+
+                        activeController = new AbortController();
+                        sendBtn.style.display = 'none';
+                        stopBtn.style.display = 'flex';
+                        headerEmoji.classList.add('processing');
+
+                        await _doSend(payload, thinkingDiv);
+
+                        activeController = null;
+                        headerEmoji.classList.remove('processing');
+                        sendBtn.style.display = 'block';
+                        stopBtn.style.display = 'none';
+                        inputEl.focus();
+                    });
+                    thinkingDiv.appendChild(retryBtn);
                 }
             }
+        }
+
+        async function sendMessage() {
+            const text = inputEl.value.trim();
+            if (!text || activeController) return;
+
+            inputEl.value = '';
+            syncHeight();
+            sendBtn.style.display = 'none';
+            stopBtn.style.display = 'flex';
+            headerEmoji.classList.add('processing');
+
+            activeController = new AbortController();
+
+            const userHistIdx = currentHistory.length;
+            currentHistory.push({ role: 'user', content: text });
+            const userMsgDiv = addMessage('user', text);
+            attachHistoryControls(userMsgDiv, userHistIdx);
+            scrollToBottom();
+
+            const thinkingDiv = addMessage('assistant thinking', '✨ thinking…');
+
+            const payload = {
+                message: text,
+                session_id: sessionId,
+                tier: currentTier,
+                include_long: memLong,
+                include_mid: memMid,
+                include_short: memShort,
+                off_record: current_mode === 'otr',
+                model: primaryBackend,
+                user: CORTEX_USER,
+                persona: CORTEX_PERSONA,
+            };
+
+            await _doSend(payload, thinkingDiv);
 
             activeController = null;
             headerEmoji.classList.remove('processing');

cortex/static/style.css

@@ -565,6 +565,26 @@
         }
         .model-tag.fallback { color: #f59e0b; }
 
+        /* Retry button — shown in error message bubbles */
+        .retry-btn {
+            display: inline-block;
+            margin-top: 0.6rem;
+            margin-left: 0.15rem;
+            padding: 0.25rem 0.7rem;
+            font-size: 0.78rem;
+            font-family: inherit;
+            background: transparent;
+            color: var(--error-text);
+            border: 1px solid var(--error-border);
+            border-radius: 4px;
+            cursor: pointer;
+            transition: background 0.15s, color 0.15s;
+        }
+        .retry-btn:hover {
+            background: var(--error-border);
+            color: #fff;
+        }
+
         /* Note messages */
         .message.note-private {
             align-self: flex-end;

cortex/tools/__init__.py

@@ -551,3 +551,61 @@ async def call_tool(name: str, args: dict) -> str:
     if fn is None:
         return f"Unknown tool: {name}"
     return await fn(**args)
+
+
+# ---------------------------------------------------------------------------
+# OpenAI JSON Schema format — auto-derived from the Gemini declarations above
+# so there is a single source of truth for tool definitions.
+# ---------------------------------------------------------------------------
+
+_GEMINI_TYPE_TO_JSON = {
+    "OBJECT":  "object",
+    "STRING":  "string",
+    "INTEGER": "integer",
+    "NUMBER":  "number",
+    "BOOLEAN": "boolean",
+    "ARRAY":   "array",
+}
+
+
+def _schema_to_json(schema) -> dict:
+    """Recursively convert a Gemini types.Schema to a JSON Schema dict."""
+    type_name = getattr(getattr(schema, "type", None), "name", "STRING")
+    result: dict = {"type": _GEMINI_TYPE_TO_JSON.get(type_name, "string")}
+
+    if getattr(schema, "description", None):
+        result["description"] = schema.description
+
+    props = getattr(schema, "properties", None) or {}
+    if result["type"] == "object":
+        result["properties"] = {k: _schema_to_json(v) for k, v in props.items()}
+
+    req = getattr(schema, "required", None)
+    if req:
+        result["required"] = list(req)
+
+    return result
+
+
+def _build_openai_tools() -> list[dict]:
+    """Convert TOOL_DECLARATIONS (Gemini format) to OpenAI tool schemas."""
+    out = []
+    for decl in TOOL_DECLARATIONS[0].function_declarations:
+        params = (
+            _schema_to_json(decl.parameters)
+            if decl.parameters
+            else {"type": "object", "properties": {}}
+        )
+        out.append({
+            "type": "function",
+            "function": {
+                "name": decl.name,
+                "description": decl.description or "",
+                "parameters": params,
+            },
+        })
+    return out
+
+
+# OpenAI-format tool list — pass to client.chat.completions.create(tools=...)
+OPENAI_TOOL_SCHEMAS: list[dict] = _build_openai_tools()

documentation/ARCH__BACKENDS.md

@@ -1,47 +1,130 @@
 # Architecture: LLM Backends
 
-> How Cortex talks to AI models.
-> Last updated: 2026-04-03
+> How Cortex selects and talks to AI models.
+> Last updated: 2026-04-06
 
 ---
 
-## Three Backends
+## Backends
 
-| Backend | Used for | Auth | Config |
+| Backend | Type | Auth | Notes |
 |---|---|---|---|
-| **Claude CLI** | Primary chat, all user-facing responses | OAuth token from `~/.claude/.credentials.json` | `DEFAULT_MODEL` in `.env` |
-| **Gemini CLI** | Fallback when Claude unavailable | Gemini CLI credentials | Auto-fallback |
-| **Local (Open WebUI)** | Private/offline tasks, cost-free use | API key per user in `local_llm.json` | `/settings/local` UI |
-
-The **Gemini API** (google-genai SDK) is also used — but only by the orchestrator tool loop, not as a general chat backend. See [`ARCH__FUTURE.md`](ARCH__FUTURE.md) for the orchestrator pattern.
+| **Claude CLI** | `claude_cli` | OAuth token from `~/.claude/.credentials.json` | Primary chat; model set via `DEFAULT_MODEL` in `.env` |
+| **Gemini CLI** | `gemini_cli` | Gemini CLI credentials | Fallback / explicit selection |
+| **Gemini API** | `gemini_api` | `GEMINI_API_KEY` in `.env` | Orchestrator tool loop only — not general chat |
+| **Local (OpenAI-compat)** | `local_openai` | API key per host in model registry | Open WebUI, Ollama, OpenRouter, LiteLLM, etc. |
 
 ---
 
 ## Backend Selection
 
-User toggles backend in the UI: `claude → gemini → local` (cycles). The active backend is stored server-side; the UI reflects it with color coding (default / green / amber).
+### Default: Role-Based Routing (Auto)
 
-When local is active, the active model name appears below the toggle button.
+When no explicit backend is selected, Cortex routes to the model configured for the
+request's **role** in the user's model registry. Roles: `chat`, `orchestrator`, `distill`,
+`coder`, `research` (extensible via `DEFINED_ROLES` in `.env`).
 
-**Fallback chain** (automatic, on error):
+Resolution order for a role:
+1. User registry: `roles[role].primary → backup_1 → backup_2 → backup_3 → backup_4`
+2. `.env` role default: `ROLE_CHAT=claude_cli`, `ROLE_DISTILL=gemini_api`, etc.
+3. Hardcoded last-resort: `chat/distill/coder → claude_cli`, `orchestrator/research → gemini_api`
+
+### Explicit Override
+
+The UI backend toggle cycles: **auto → claude → gemini → local → auto**
+
+- **auto** (default): role-based routing as above; sends `model: null` to `/chat`
+- **claude / gemini / local**: bypasses role routing; forces that specific backend
+- When "local" is active, the configured model name appears below the toggle button
+
+**Fallback chain** (automatic, on any error):
 ```
 claude  → gemini
 gemini  → claude
 local   → claude
 ```
 
+Each response includes a model label (bottom-right of the message bubble) showing what
+actually responded. Amber label with `⚡` = fallback was used.
+
 Auth expiry on Claude triggers a UI banner + `claude_auth_expired` SSE event.
 
 ---
 
+## Model Registry
+
+Per-user configuration stored in `home/{user}/model_registry.json`.
+
+Hosts and models are managed at **Settings → Model Registry** (`/settings/local`).
+
+### Schema
+
+```json
+{
+  "version": 1,
+  "hosts": [
+    {
+      "id": "abc123",
+      "label": "Home ML Laptop",
+      "api_url": "http://192.168.x.x:3000",
+      "api_key": "sk-...",
+      "host_type": "openwebui"
+    }
+  ],
+  "models": [
+    {
+      "id": "def456",
+      "type": "local_openai",
+      "label": "Gemma Medium",
+      "model_name": "agent-support-gemma-medium",
+      "host_id": "abc123",
+      "context_k": 50,
+      "tags": ["chat", "fast"]
+    }
+  ],
+  "roles": {
+    "chat": {
+      "primary": "def456",
+      "backup_1": "claude_cli"
+    }
+  }
+}
+```
+
+### host_type
+
+Controls which API path layout is used:
+
+| `host_type` | Chat endpoint | Models endpoint | Use for |
+|---|---|---|---|
+| `openwebui` (default) | `POST {url}/api/chat/completions` | `GET {url}/api/models` | Open WebUI, Ollama |
+| `openai` | `POST {url}/chat/completions` | `GET {url}/models` | OpenRouter, LiteLLM, Anthropic-compat |
+
+Set `api_url` to the base path ending just before `/chat/completions`:
+- OpenRouter: `https://openrouter.ai/api/v1`
+- LiteLLM proxy: `http://host:port`
+
+### Built-in model IDs
+
+Always resolvable without a registry entry:
+
+| ID | Backend |
+|---|---|
+| `claude_cli` | Claude CLI subprocess |
+| `gemini_cli` | Gemini CLI subprocess |
+| `gemini_api` | Gemini API (SDK) — orchestrator only |
+
+---
+
 ## Claude Backend (`_claude()`)
 
 Runs `claude --print --no-session-persistence --output-format text` as a subprocess.
 
 - System prompt passed via `--system-prompt`
 - Conversation history formatted as `<conversation>` block
-- Token read live from `~/.claude/.credentials.json` on every call — never relies on the env var, which goes stale after `claude auth login`
-- Model override via `--model` flag (e.g. `claude-opus-4-6`)
+- Token read live from `~/.claude/.credentials.json` on every call — never relies on the
+  env var, which goes stale after `claude auth login`
+- Model override via `--model` flag when a specific `model_name` is configured in the registry
 
 Timeout: `TIMEOUT_CLAUDE=60` seconds (`.env`)
 
@@ -51,7 +134,7 @@ Timeout: `TIMEOUT_CLAUDE=60` seconds (`.env`)
 
 Runs `gemini --output-format text --extensions "" -p <prompt>` as a subprocess.
 
-- `--extensions ""` disables all MCP extensions — prevents child processes from keeping pipes open after responding
+- `--extensions ""` disables all MCP extensions — prevents child processes keeping pipes open
 - `start_new_session=True` puts the process in its own group for clean `os.killpg` on timeout
 - Output is cleaned to strip CLI noise lines (loading messages, retry notices, quota warnings)
 
@@ -61,46 +144,33 @@ Timeout: `TIMEOUT_GEMINI=120` seconds (`.env`)
 
 ## Local Backend (`_local()`)
 
-HTTP POST to Open WebUI's OpenAI-compatible endpoint: `{api_url}/api/chat/completions`.
+HTTP POST to an OpenAI-compatible endpoint. Model config is resolved via the model registry.
 
-Per-user config in `home/{user}/local_llm.json`:
-```json
-{
-  "hosts": [{"id": "...", "label": "scott_gaming", "api_url": "http://192.168.32.19:3000", "api_key": "sk-..."}],
-  "models": [{"id": "...", "host_id": "...", "label": "Gemma 4 Small", "model_name": "agent-support-gemma-small"}],
-  "active_model_id": "..."
-}
+```python
+# host_type "openwebui":  POST {api_url}/api/chat/completions
+# host_type "openai":     POST {api_url}/chat/completions
 ```
 
-Resolution order for active model:
-1. User's `active_model_id` in `local_llm.json`
-2. `.env` server defaults (`LOCAL_API_URL` / `LOCAL_MODEL`)
-3. Error — user is prompted to configure at `/settings/local`
-
 Timeout: `TIMEOUT_LOCAL=300` seconds (`.env`) — local models may need to load from disk.
 
-**Manage at:** `/settings/local` — supports multiple hosts and models per user, "Fetch from host" button to populate model list from the server.
+---
+
+## Distillation
+
+Memory distillation uses `role="distill"` for mid and long passes. Configure the distill
+model via the Model Registry → Role Assignments → Distill role.
+
+`.env` override: `ROLE_DISTILL=claude_cli` (default). Set to any built-in ID or leave blank
+to fall through to the hardcoded default (`claude_cli`).
 
 ---
 
-## Distillation Backends
+## Code locations
 
-Memory distillation runs on a schedule and uses the LLM for mid and long distill passes. By default uses the primary backend (`claude`). Override in `.env`:
-
-```
-DISTILL_BACKEND_MID=local   # saves API credits — Gemma handles summarization well
-DISTILL_BACKEND_LONG=       # empty = use primary (claude recommended for quality)
-```
-
----
-
-## Current Local Models (scott_gaming, 8 GB VRAM)
-
-| Model | Alias | Speed | Practical Context |
-|---|---|---|---|
-| Gemma 4 E4B | `agent-support-gemma-small` | ~25 t/s | **72k tokens** |
-| Gemma 4 26B A4B (MoE) | `agent-support-gemma-medium` | ~9 t/s | **50k tokens** |
-
-Both support OpenAI `tools` / `tool_choice` function calling — required for the local orchestrator.
-
-Full Open WebUI API reference: [`docs/OPEN_WEBUI_API.md`](../docs/OPEN_WEBUI_API.md)
+| File | Responsibility |
+|---|---|
+| `cortex/llm_client.py` | `complete()` — routing, dispatch, fallback |
+| `cortex/model_registry.py` | Per-user registry CRUD and resolution |
+| `cortex/routers/local_llm.py` | Settings UI routes + `/api/models/role` AJAX |
+| `cortex/routers/chat.py` | `_backend_label()`, `fallback_used` flag |
+| `cortex/config.py` | `ROLE_*` env defaults, `DEFINED_ROLES`, `PRIMARY_BACKEND` |