fix(model): require confirmation for expensive model selections

Rebased onto current main and re-ported across the restructured
surfaces: model flows now thread confirm_provider/base_url/api_key
through hermes_cli/model_setup_flows.py, the Discord picker lives in
plugins/platforms/discord/adapter.py, and the web dashboard picker
applies chat-mode switches via config.set so the expensive-model
confirmation can ride the response.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

hermes_cli/auth.py

@@ -6175,6 +6175,40 @@ def _reset_config_provider() -> Path:
     return config_path
 
 
+def _confirm_expensive_model_selection(
+    model_id: str,
+    *,
+    provider: str = "",
+    base_url: str = "",
+    api_key: str = "",
+) -> bool:
+    """Prompt before saving a model whose known pricing exceeds guardrails."""
+    try:
+        from hermes_cli.model_cost_guard import expensive_model_warning
+
+        warning = expensive_model_warning(
+            model_id,
+            provider=provider,
+            base_url=base_url,
+            api_key=api_key,
+        )
+    except Exception:
+        warning = None
+    if warning is None:
+        return True
+
+    print()
+    print("=" * 72)
+    print(warning.message)
+    print("=" * 72)
+    try:
+        response = input("Switch anyway? [y/N]: ").strip().lower()
+    except (KeyboardInterrupt, EOFError):
+        print()
+        return False
+    return response in {"y", "yes"}
+
+
 def _prompt_model_selection(
     model_ids: List[str],
     current_model: str = "",
@@ -6182,6 +6216,9 @@ def _prompt_model_selection(
     unavailable_models: Optional[List[str]] = None,
     portal_url: str = "",
     unavailable_message: str = "",
+    confirm_provider: str = "",
+    confirm_base_url: str = "",
+    confirm_api_key: str = "",
 ) -> Optional[str]:
     """Interactive model selection. Puts current_model first with a marker. Returns chosen model ID or None.
 
@@ -6195,6 +6232,18 @@ def _prompt_model_selection(
 
     _unavailable = unavailable_models or []
 
+    def _confirmed_selection(mid: str) -> Optional[str]:
+        if not mid:
+            return None
+        if confirm_provider and not _confirm_expensive_model_selection(
+            mid,
+            provider=confirm_provider,
+            base_url=confirm_base_url,
+            api_key=confirm_api_key,
+        ):
+            return None
+        return mid
+
     # Reorder: current model first, then the rest (deduplicated)
     ordered = []
     if current_model and current_model in model_ids:
@@ -6310,13 +6359,13 @@ def _prompt_model_selection(
             return None
         print()
         if idx < len(ordered):
-            return ordered[idx]
+            return _confirmed_selection(ordered[idx])
         elif idx == len(ordered):
             try:
                 custom = input("Enter model name: ").strip()
             except (EOFError, KeyboardInterrupt):
                 return None
-            return custom if custom else None
+            return _confirmed_selection(custom) if custom else None
         return None
     except (ImportError, NotImplementedError, OSError, subprocess.SubprocessError):
         pass
@@ -6348,10 +6397,10 @@ def _prompt_model_selection(
                 return None
             idx = int(choice)
             if 1 <= idx <= n:
-                return ordered[idx - 1]
+                return _confirmed_selection(ordered[idx - 1])
             elif idx == n + 1:
                 custom = input("Enter model name: ").strip()
-                return custom if custom else None
+                return _confirmed_selection(custom) if custom else None
             elif idx == n + 2:
                 return None
             print(f"Please enter 1-{n + 2}")
@@ -7730,6 +7779,9 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
                     unavailable_models=unavailable_models,
                     portal_url=_portal,
                     unavailable_message=unavailable_message,
+                    confirm_provider="nous",
+                    confirm_base_url=inference_base_url,
+                    confirm_api_key=runtime_key,
                 )
             elif unavailable_models:
                 _url = (_portal or DEFAULT_NOUS_PORTAL_URL).rstrip("/")

hermes_cli/main.py

@@ -3269,6 +3269,7 @@ def _aux_flow_provider_model(
             model_list,
             current_model=current_model,
             pricing=pricing,
+            confirm_provider=provider_slug,
         )
         if selected is None:
             print("No change.")

hermes_cli/model_cost_guard.py

@@ -0,0 +1,134 @@
+"""Expensive-model confirmation helpers for model selection surfaces."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from decimal import Decimal, InvalidOperation
+from typing import Optional
+
+from agent.models_dev import ModelInfo
+
+
+INPUT_COST_WARNING_THRESHOLD = Decimal("20")
+OUTPUT_COST_WARNING_THRESHOLD = Decimal("100")
+GPT55_PRO_OPENROUTER_ID = "openai/gpt-5.5-pro"
+GPT55_SUGGESTION = "did you mean to select openai/gpt-5.5?"
+
+
+@dataclass(frozen=True)
+class ExpensiveModelWarning:
+    """Confirmation payload for models above Hermes' cost guardrail."""
+
+    model: str
+    provider: str
+    input_cost_per_million: Optional[Decimal]
+    output_cost_per_million: Optional[Decimal]
+    source: str
+    message: str
+
+
+def _to_decimal(value: object) -> Optional[Decimal]:
+    if value is None:
+        return None
+    try:
+        return Decimal(str(value))
+    except (InvalidOperation, ValueError):
+        return None
+
+
+def _format_money(value: Optional[Decimal]) -> str:
+    if value is None:
+        return "unknown"
+    return f"${value:.2f}/M"
+
+
+def _pricing_from_model_info(
+    model_info: Optional[ModelInfo],
+) -> tuple[Optional[Decimal], Optional[Decimal], str]:
+    if model_info is None or not model_info.has_cost_data():
+        return None, None, ""
+    return (
+        _to_decimal(model_info.cost_input),
+        _to_decimal(model_info.cost_output),
+        "models.dev",
+    )
+
+
+def expensive_model_warning(
+    model_name: str,
+    *,
+    provider: Optional[str] = None,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+    model_info: Optional[ModelInfo] = None,
+) -> Optional[ExpensiveModelWarning]:
+    """Return a warning payload when known pricing exceeds safety thresholds.
+
+    The guard only triggers when pricing is known. Callers should use this after
+    model resolution so aliases and provider-specific model IDs have settled.
+    """
+    model = (model_name or "").strip()
+    if not model:
+        return None
+
+    input_cost, output_cost, source = _pricing_from_model_info(model_info)
+    if input_cost is None and output_cost is None and provider:
+        try:
+            from agent.models_dev import get_model_info
+
+            input_cost, output_cost, source = _pricing_from_model_info(
+                get_model_info(provider, model)
+            )
+        except Exception:
+            pass
+    if input_cost is None and output_cost is None:
+        try:
+            from agent.usage_pricing import get_pricing_entry
+
+            entry = get_pricing_entry(
+                model,
+                provider=provider,
+                base_url=base_url,
+                api_key=api_key,
+            )
+        except Exception:
+            entry = None
+        if entry is not None:
+            input_cost = entry.input_cost_per_million
+            output_cost = entry.output_cost_per_million
+            source = entry.source
+
+    over_input = (
+        input_cost is not None and input_cost > INPUT_COST_WARNING_THRESHOLD
+    )
+    over_output = (
+        output_cost is not None and output_cost > OUTPUT_COST_WARNING_THRESHOLD
+    )
+    if not over_input and not over_output:
+        return None
+
+    lines = [
+        "!!! EXPENSIVE MODEL WARNING !!!",
+        "",
+        f"{model} has known pricing above Hermes' safety threshold.",
+        f"Input tokens: {_format_money(input_cost)}",
+        f"Output tokens: {_format_money(output_cost)}",
+        (
+            "Threshold: more than $20/M input tokens or more than "
+            "$100/M output tokens."
+        ),
+    ]
+    if source:
+        lines.append(f"Pricing source: {source}.")
+    if model.lower() == GPT55_PRO_OPENROUTER_ID:
+        lines.append(GPT55_SUGGESTION)
+    lines.append("Confirm only if you intend to use this model.")
+
+    return ExpensiveModelWarning(
+        model=model,
+        provider=(provider or "").strip(),
+        input_cost_per_million=input_cost,
+        output_cost_per_million=output_cost,
+        source=source or "unknown",
+        message="\n".join(lines),
+    )

hermes_cli/model_setup_flows.py

@@ -102,7 +102,12 @@ def _model_flow_openrouter(config, current_model=""):
     pricing = get_pricing_for_provider("openrouter", force_refresh=True)
 
     selected = _prompt_model_selection(
-        openrouter_models, current_model=current_model, pricing=pricing
+        openrouter_models,
+        current_model=current_model,
+        pricing=pricing,
+        confirm_provider="openrouter",
+        confirm_base_url=OPENROUTER_BASE_URL,
+        confirm_api_key=_resolved or existing_key,
     )
     if selected:
         _save_model_choice(selected)
@@ -311,6 +316,9 @@ def _model_flow_nous(config, current_model="", args=None):
         unavailable_models=unavailable_models,
         portal_url=_nous_portal_url,
         unavailable_message=unavailable_message,
+        confirm_provider="nous",
+        confirm_base_url=creds.get("base_url", ""),
+        confirm_api_key=creds.get("api_key", ""),
     )
     if selected:
         _save_model_choice(selected)
@@ -416,7 +424,13 @@ def _model_flow_openai_codex(config, current_model=""):
 
     codex_models = get_codex_model_ids(access_token=_codex_token)
 
-    selected = _prompt_model_selection(codex_models, current_model=current_model)
+    selected = _prompt_model_selection(
+        codex_models,
+        current_model=current_model,
+        confirm_provider="openai-codex",
+        confirm_base_url=DEFAULT_CODEX_BASE_URL,
+        confirm_api_key=_codex_token or "",
+    )
     if selected:
         _save_model_choice(selected)
         _update_config_for_provider("openai-codex", DEFAULT_CODEX_BASE_URL)
@@ -546,7 +560,12 @@ def _model_flow_qwen_oauth(_config, current_model=""):
         models = list(_DEFAULT_QWEN_PORTAL_MODELS)
 
     default = current_model or (models[0] if models else "qwen3-coder-plus")
-    selected = _prompt_model_selection(models, current_model=default)
+    selected = _prompt_model_selection(
+        models,
+        current_model=default,
+        confirm_provider="qwen-oauth",
+        confirm_base_url=DEFAULT_QWEN_BASE_URL,
+    )
     if selected:
         _save_model_choice(selected)
         _update_config_for_provider("qwen-oauth", DEFAULT_QWEN_BASE_URL)
@@ -595,7 +614,12 @@ def _model_flow_minimax_oauth(config, current_model="", args=None):
     from hermes_cli.models import _PROVIDER_MODELS
 
     model_ids = _PROVIDER_MODELS.get("minimax-oauth", [])
-    selected = _prompt_model_selection(model_ids, current_model)
+    selected = _prompt_model_selection(
+        model_ids,
+        current_model,
+        confirm_provider="minimax-oauth",
+        confirm_base_url=creds["base_url"],
+    )
     if not selected:
         return
     _save_model_choice(selected)
@@ -664,7 +688,12 @@ def _model_flow_google_gemini_cli(_config, current_model=""):
 
     models = list(_PROVIDER_MODELS.get("google-gemini-cli") or [])
     default = current_model or (models[0] if models else "gemini-3-flash-preview")
-    selected = _prompt_model_selection(models, current_model=default)
+    selected = _prompt_model_selection(
+        models,
+        current_model=default,
+        confirm_provider="google-gemini-cli",
+        confirm_base_url=DEFAULT_GEMINI_CLOUDCODE_BASE_URL,
+    )
     if selected:
         _save_model_choice(selected)
         _update_config_for_provider(
@@ -1589,7 +1618,11 @@ def _model_flow_copilot(config, current_model=""):
 
     if model_list:
         selected = _prompt_model_selection(
-            model_list, current_model=normalized_current_model
+            model_list,
+            current_model=normalized_current_model,
+            confirm_provider=provider_id,
+            confirm_base_url=effective_base,
+            confirm_api_key=api_key,
         )
     else:
         try:
@@ -1727,6 +1760,9 @@ def _model_flow_copilot_acp(config, current_model=""):
         selected = _prompt_model_selection(
             model_list,
             current_model=normalized_current_model,
+            confirm_provider=provider_id,
+            confirm_base_url=effective_base,
+            confirm_api_key=catalog_api_key,
         )
     else:
         try:
@@ -1831,7 +1867,13 @@ def _model_flow_kimi(config, current_model=""):
         model_list = _PROVIDER_MODELS.get("moonshot", [])
 
     if model_list:
-        selected = _prompt_model_selection(model_list, current_model=current_model)
+        selected = _prompt_model_selection(
+            model_list,
+            current_model=current_model,
+            confirm_provider=provider_id,
+            confirm_base_url=effective_base,
+            confirm_api_key=existing_key,
+        )
     else:
         try:
             selected = input("Enter model name: ").strip()
@@ -1939,7 +1981,13 @@ def _model_flow_stepfun(config, current_model=""):
             )
 
     if model_list:
-        selected = _prompt_model_selection(model_list, current_model=current_model)
+        selected = _prompt_model_selection(
+            model_list,
+            current_model=current_model,
+            confirm_provider=provider_id,
+            confirm_base_url=effective_base,
+            confirm_api_key=existing_key,
+        )
     else:
         try:
             selected = input("Model name: ").strip()
@@ -2015,7 +2063,13 @@ def _model_flow_bedrock_api_key(config, region, current_model=""):
     print(f"  Showing {len(model_list)} curated models")
 
     if model_list:
-        selected = _prompt_model_selection(model_list, current_model=current_model)
+        selected = _prompt_model_selection(
+            model_list,
+            current_model=current_model,
+            confirm_provider="custom",
+            confirm_base_url=mantle_base_url,
+            confirm_api_key=existing_key,
+        )
     else:
         try:
             selected = input("  Model ID: ").strip()
@@ -2204,7 +2258,12 @@ def _model_flow_bedrock(config, current_model=""):
 
     # 4. Model selection
     if model_list:
-        selected = _prompt_model_selection(model_list, current_model=current_model)
+        selected = _prompt_model_selection(
+            model_list,
+            current_model=current_model,
+            confirm_provider="bedrock",
+            confirm_base_url=f"https://bedrock-runtime.{region}.amazonaws.com",
+        )
     else:
         try:
             selected = input("  Model ID: ").strip()
@@ -2488,7 +2547,13 @@ def _model_flow_api_key_provider(config, provider_id, current_model=""):
         model_list = list(dict.fromkeys(mid for mid in model_list if mid))
 
     if model_list:
-        selected = _prompt_model_selection(model_list, current_model=current_model)
+        selected = _prompt_model_selection(
+            model_list,
+            current_model=current_model,
+            confirm_provider=provider_id,
+            confirm_base_url=effective_base,
+            confirm_api_key=existing_key,
+        )
     else:
         try:
             selected = input("Model name: ").strip()
@@ -2638,7 +2703,11 @@ def _model_flow_anthropic(config, current_model=""):
     # Model selection
     model_list = _PROVIDER_MODELS.get("anthropic", [])
     if model_list:
-        selected = _prompt_model_selection(model_list, current_model=current_model)
+        selected = _prompt_model_selection(
+            model_list,
+            current_model=current_model,
+            confirm_provider="anthropic",
+        )
     else:
         try:
             selected = input("Model name (e.g., claude-sonnet-4-20250514): ").strip()

hermes_cli/web_server.py

@@ -657,21 +657,6 @@ class AudioTranscriptionRequest(BaseModel):
     mime_type: Optional[str] = None
 
 
-class ModelAssignment(BaseModel):
-    """Payload for POST /api/model/set — assign a provider/model to a slot.
-
-    scope="main"        → writes model.provider + model.default
-    scope="auxiliary"   → writes auxiliary.<task>.provider + auxiliary.<task>.model
-    scope="auxiliary" with task=""  → applied to every auxiliary.* slot
-    scope="auxiliary" with task="__reset__"  → resets every slot to provider="auto"
-    """
-
-    scope: str
-    provider: str
-    model: str
-    task: str = ""
-
-
 _AUDIO_MIME_EXTENSIONS: Dict[str, str] = {
     "audio/aac": ".aac",
     "audio/flac": ".flac",
@@ -713,6 +698,7 @@ class ModelAssignment(BaseModel):
     # reads model.base_url from config (it ignores OPENAI_BASE_URL), so this is
     # the path that actually wires a local endpoint into resolution.
     base_url: str = ""
+    confirm_expensive_model: bool = False
 
 
 def _apply_main_model_assignment(
@@ -2470,6 +2456,27 @@ async def set_model_assignment(body: ModelAssignment):
     try:
         cfg = load_config()
 
+        if model and not body.confirm_expensive_model:
+            try:
+                from hermes_cli.model_cost_guard import expensive_model_warning
+
+                warning = expensive_model_warning(
+                    model,
+                    provider=provider,
+                    base_url=base_url,
+                )
+            except Exception:
+                warning = None
+            if warning is not None:
+                return {
+                    "ok": False,
+                    "scope": scope,
+                    "provider": provider,
+                    "model": model,
+                    "confirm_required": True,
+                    "confirm_message": warning.message,
+                }
+
         if scope == "main":
             if not provider or not model:
                 raise HTTPException(status_code=400, detail="provider and model required for main")