From d15cd18fa1f3e07e69df9a63fd017ab412a68c86 Mon Sep 17 00:00:00 2001
From: Brooklyn Nicholson <brooklyn.bb.nicholson@gmail.com>
Date: Wed, 19 Aug 2026 22:04:47 -0500
Subject: [PATCH] feat(desktop): hide the Thinking toggle where a disable is
 rejected
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The picker offered an off switch for every reasoning model, including routes
whose upstream answers a disable with HTTP 400 — so "thinking off" was a
control that could not work. Carry the catalog's mandatory verdict through
model.options as can_disable_reasoning and hide the toggle when it is false.

Effort levels are left alone. The catalog's supported_efforts under-reports
what the Portal serves (z-ai/glm-5.3 publishes max, high, low yet honors
minimal at its lowest thinking), so filtering the scale by it would hide
levels that work.
---
 .../src/app/shell/model-catalog-menu.tsx      |   1 +
 .../src/app/shell/model-edit-submenu.tsx      |   8 +-
 apps/desktop/src/types/hermes.ts              |   4 +
 hermes_cli/inventory.py                       |  56 ++++++-
 .../test_inventory_reasoning_caps.py          | 141 ++++++++++++++++++
 5 files changed, 205 insertions(+), 5 deletions(-)
 create mode 100644 tests/hermes_cli/test_inventory_reasoning_caps.py

diff --git a/apps/desktop/src/app/shell/model-catalog-menu.tsx b/apps/desktop/src/app/shell/model-catalog-menu.tsx
index 2513725fd9..1dfd70f15b 100644
--- a/apps/desktop/src/app/shell/model-catalog-menu.tsx
+++ b/apps/desktop/src/app/shell/model-catalog-menu.tsx
@@ -458,6 +458,7 @@ export function ModelCatalogMenu({
                           ) : null}
                         </DropdownMenuSubTrigger>
                         <ModelEditSubmenu
+                          canDisableReasoning={caps?.can_disable_reasoning}
                           defaultEffort={defaultEffort}
                           effort={effEffort}
                           fastControl={fastControl}
diff --git a/apps/desktop/src/app/shell/model-edit-submenu.tsx b/apps/desktop/src/app/shell/model-edit-submenu.tsx
index 1d5b5dc598..68b7281345 100644
--- a/apps/desktop/src/app/shell/model-edit-submenu.tsx
+++ b/apps/desktop/src/app/shell/model-edit-submenu.tsx
@@ -59,6 +59,10 @@ export function resolveFastControl(
 }
 
 interface ModelEditSubmenuProps {
+  /** Whether this model can turn thinking off. False on reasoning-mandatory
+   *  routes, whose upstream rejects a disable — the toggle is hidden rather
+   *  than offered as a control that silently does nothing. */
+  canDisableReasoning?: boolean
   /** The profile's configured default effort — what an unset row inherits.
    *  Passed in (not read from a store) so this submenu stays pure. */
   defaultEffort: string
@@ -98,6 +102,7 @@ export function ModelEditSubmenu(props: ModelEditSubmenuProps) {
 }
 
 function ModelEditSubmenuBody({
+  canDisableReasoning,
   defaultEffort,
   effort,
   fastControl,
@@ -111,6 +116,7 @@ function ModelEditSubmenuBody({
 
   const effortValue = resolveReasoningEffort(effort, defaultEffort)
   const thinkingOn = isThinkingEnabled(effort, defaultEffort)
+  const showThinkingToggle = reasoning && canDisableReasoning !== false
 
   const setFast = (enabled: boolean) => {
     if (fastControl.kind === 'variant') {
@@ -139,7 +145,7 @@ function ModelEditSubmenuBody({
   ) : (
     <>
       <DropdownMenuLabel className={dropdownMenuSectionLabel}>{copy.options}</DropdownMenuLabel>
-      {reasoning ? (
+      {showThinkingToggle ? (
         <DropdownMenuItem className={dropdownMenuRow} onSelect={event => event.preventDefault()}>
           {copy.thinking}
           <Switch
diff --git a/apps/desktop/src/types/hermes.ts b/apps/desktop/src/types/hermes.ts
index dc5e3d24a4..783ee945e5 100644
--- a/apps/desktop/src/types/hermes.ts
+++ b/apps/desktop/src/types/hermes.ts
@@ -427,6 +427,10 @@ export interface ModelOptionProvider {
 }
 
 export interface ModelCapabilities {
+  /** False when the route rejects a reasoning disable ("mandatory" in the
+   *  provider catalog), so the Thinking toggle must not be offered. Absent
+   *  when the catalog doesn't say. */
+  can_disable_reasoning?: boolean
   fast: boolean
   reasoning: boolean
 }
diff --git a/hermes_cli/inventory.py b/hermes_cli/inventory.py
index 7d1b14992a..47ca84a607 100644
--- a/hermes_cli/inventory.py
+++ b/hermes_cli/inventory.py
@@ -34,7 +34,7 @@ Substrate facts (verified May 2026):
 from __future__ import annotations
 
 from dataclasses import dataclass, replace
-from typing import Optional
+from typing import Any, Optional
 
 
 # ─── Public types ───────────────────────────────────────────────────────
@@ -402,14 +402,51 @@ def format_aux_picker_entries(
     return entries
 
 
+def _reasoning_catalog_reader(slug: str):
+    """Per-model reasoning-capability reader for aggregators that publish one.
+
+    Cache-only — building the picker payload must never block on HTTP. A cold
+    cache warms in the background so the next open is accurate; until then the
+    model reports no restriction and the UI offers the full scale.
+    """
+    try:
+        from hermes_cli.models import (
+            nous_model_reasoning_capabilities,
+            openrouter_model_reasoning_capabilities,
+            warm_nous_reasoning_caps_async,
+            warm_openrouter_reasoning_caps_async,
+        )
+    except Exception:
+        return None
+
+    if slug == "nous":
+        warm_nous_reasoning_caps_async()
+        return nous_model_reasoning_capabilities
+    if slug == "openrouter":
+        warm_openrouter_reasoning_caps_async()
+        return openrouter_model_reasoning_capabilities
+    return None
+
+
 def _apply_capabilities(rows: list[dict]) -> None:
-    """Attach a ``{model: {fast, reasoning}}`` map to each provider row.
+    """Attach a ``{model: {fast, reasoning, ...}}`` map to each provider row.
 
     `fast` mirrors ``model_supports_fast_mode`` (the same gate the runtime
     enforces). `reasoning` comes from the models.dev catalog when known and
     defaults to True otherwise — the effort dial is broadly accepted and a
     no-op on models that ignore it, whereas hiding it from a capable-but-
     uncatalogued model is the worse failure.
+
+    Aggregators that publish per-model reasoning detail add
+    `can_disable_reasoning`, False on reasoning-mandatory routes whose upstream
+    answers a disable with HTTP 400. Omitted when the catalog doesn't say,
+    which the UI reads as "no restriction known".
+
+    The catalog's `supported_efforts` list is deliberately NOT forwarded: it
+    under-reports. The Portal accepts and honors levels a route doesn't
+    advertise (``z-ai/glm-5.3`` publishes ``max, high, low`` yet serves
+    ``minimal`` at its lowest thinking), so filtering the picker by that list
+    would hide levels that demonstrably work.
     """
     from hermes_cli.models import model_supports_fast_mode
 
@@ -420,7 +457,8 @@ def _apply_capabilities(rows: list[dict]) -> None:
 
     for row in rows:
         slug = row.get("slug") or ""
-        caps: dict[str, dict[str, bool]] = {}
+        caps: dict[str, dict[str, Any]] = {}
+        read_reasoning_catalog = _reasoning_catalog_reader(slug.lower())
 
         for model in row.get("models") or []:
             reasoning = True
@@ -432,11 +470,21 @@ def _apply_capabilities(rows: list[dict]) -> None:
                 except Exception:
                     reasoning = True
 
-            caps[model] = {
+            entry: dict[str, Any] = {
                 "fast": bool(model_supports_fast_mode(model)),
                 "reasoning": reasoning,
             }
 
+            if reasoning and read_reasoning_catalog is not None:
+                try:
+                    detail = read_reasoning_catalog(model)
+                except Exception:
+                    detail = None
+                if detail:
+                    entry["can_disable_reasoning"] = not detail.get("mandatory")
+
+            caps[model] = entry
+
         row["capabilities"] = caps
 
 
diff --git a/tests/hermes_cli/test_inventory_reasoning_caps.py b/tests/hermes_cli/test_inventory_reasoning_caps.py
new file mode 100644
index 0000000000..dd35b48162
--- /dev/null
+++ b/tests/hermes_cli/test_inventory_reasoning_caps.py
@@ -0,0 +1,141 @@
+"""Tests for the reasoning detail inventory._apply_capabilities puts on the
+
+picker payload. The desktop model picker hides its Thinking toggle from this,
+so a route that can't disable reasoning must be describable here — otherwise
+the UI offers an off switch whose setting the upstream rejects.
+
+The catalog's `supported_efforts` is intentionally absent from the payload:
+the Portal honors levels a route doesn't advertise, so publishing it would
+invite a picker filter that hides working levels.
+"""
+
+import hermes_cli.inventory as inv
+import hermes_cli.models as models_mod
+
+
+def _patch_catalog(monkeypatch, caps_by_model, *, provider="nous"):
+    """Point the Nous/OpenRouter catalog readers at a fixed capability map."""
+    monkeypatch.setattr(models_mod, "model_supports_fast_mode", lambda model: False)
+    monkeypatch.setattr(
+        models_mod, "warm_nous_reasoning_caps_async", lambda: None, raising=False
+    )
+    monkeypatch.setattr(
+        models_mod, "warm_openrouter_reasoning_caps_async", lambda: None, raising=False
+    )
+    reader = f"{provider}_model_reasoning_capabilities"
+    monkeypatch.setattr(
+        models_mod, reader, lambda model, **kw: caps_by_model.get(model), raising=False
+    )
+
+
+def test_optional_reasoning_route_can_disable(monkeypatch):
+    """A route that accepts a disable says so."""
+    _patch_catalog(monkeypatch, {
+        "deepseek/deepseek-v4-pro": {
+            "supports_reasoning": True,
+            "supported_efforts": ["xhigh", "high"],
+            "mandatory": False,
+        },
+    })
+    rows = [{"slug": "nous", "models": ["deepseek/deepseek-v4-pro"]}]
+    inv._apply_capabilities(rows)
+
+    assert rows[0]["capabilities"]["deepseek/deepseek-v4-pro"]["can_disable_reasoning"] is True
+
+
+def test_advertised_efforts_never_reach_the_picker(monkeypatch):
+    """The catalog's level list stays off the wire even when it is published.
+
+    It under-reports what the Portal serves, so forwarding it would let the
+    picker hide levels that work. Only the disable verdict crosses.
+    """
+    _patch_catalog(monkeypatch, {
+        "deepseek/deepseek-v4-pro": {
+            "supports_reasoning": True,
+            "supported_efforts": ["xhigh", "high"],
+            "mandatory": False,
+        },
+    })
+    rows = [{"slug": "nous", "models": ["deepseek/deepseek-v4-pro"]}]
+    inv._apply_capabilities(rows)
+
+    assert "supported_efforts" not in rows[0]["capabilities"]["deepseek/deepseek-v4-pro"]
+
+
+def test_reasoning_mandatory_route_cannot_disable(monkeypatch):
+    """`mandatory` inverts into the flag the Thinking toggle keys off.
+
+    The Portal answers a disable on these routes with HTTP 400, so offering
+    the toggle would be offering a control that cannot work.
+    """
+    _patch_catalog(monkeypatch, {
+        "z-ai/glm-5.3": {
+            "supports_reasoning": True,
+            "supported_efforts": ["max", "high", "low"],
+            "mandatory": True,
+        },
+    })
+    rows = [{"slug": "nous", "models": ["z-ai/glm-5.3"]}]
+    inv._apply_capabilities(rows)
+
+    assert rows[0]["capabilities"]["z-ai/glm-5.3"]["can_disable_reasoning"] is False
+
+
+def test_unlisted_model_states_no_restriction(monkeypatch):
+    """A model the catalog doesn't cover omits both keys rather than guessing.
+
+    The UI reads "absent" as no known restriction and offers the full scale,
+    which is the right failure: over-offering beats hiding levels a model
+    actually accepts.
+    """
+    _patch_catalog(monkeypatch, {})
+    rows = [{"slug": "nous", "models": ["mystery/model"]}]
+    inv._apply_capabilities(rows)
+
+    caps = rows[0]["capabilities"]["mystery/model"]
+    assert "supported_efforts" not in caps
+    assert "can_disable_reasoning" not in caps
+    assert caps["reasoning"] is True
+
+
+def test_providers_without_a_reasoning_catalog_are_untouched(monkeypatch):
+    """Only aggregators that publish per-model detail gain the extra keys."""
+    _patch_catalog(monkeypatch, {
+        "gpt-5.6": {"supports_reasoning": True, "supported_efforts": ["high"], "mandatory": True},
+    })
+    rows = [{"slug": "openai-api", "models": ["gpt-5.6"]}]
+    inv._apply_capabilities(rows)
+
+    caps = rows[0]["capabilities"]["gpt-5.6"]
+    assert "supported_efforts" not in caps
+    assert "can_disable_reasoning" not in caps
+
+
+def test_openrouter_uses_its_own_catalog(monkeypatch):
+    """The reader is chosen per provider row, not hardcoded to one aggregator."""
+    _patch_catalog(
+        monkeypatch,
+        {"x-ai/grok-5": {"supports_reasoning": True, "mandatory": True}},
+        provider="openrouter",
+    )
+    rows = [{"slug": "openrouter", "models": ["x-ai/grok-5"]}]
+    inv._apply_capabilities(rows)
+
+    assert rows[0]["capabilities"]["x-ai/grok-5"]["can_disable_reasoning"] is False
+
+
+def test_catalog_failure_never_breaks_the_picker(monkeypatch):
+    """A raising catalog reader degrades to "unknown", not to a broken payload."""
+    monkeypatch.setattr(models_mod, "model_supports_fast_mode", lambda model: False)
+    monkeypatch.setattr(models_mod, "warm_nous_reasoning_caps_async", lambda: None, raising=False)
+
+    def _boom(model, **kw):
+        raise RuntimeError("catalog exploded")
+
+    monkeypatch.setattr(models_mod, "nous_model_reasoning_capabilities", _boom, raising=False)
+    rows = [{"slug": "nous", "models": ["deepseek/deepseek-v4-pro"]}]
+    inv._apply_capabilities(rows)
+
+    caps = rows[0]["capabilities"]["deepseek/deepseek-v4-pro"]
+    assert "supported_efforts" not in caps
+    assert caps["reasoning"] is True