feat(adminapi): role-specific presets + dedicated OR slot per role

Replaces the single catch-all "workhorse" OR slot with one slot per
routing role (simple-or / default-or / complex-or / compaction-or) so
each can be reassigned independently from the admin UI without side
effects on the other roles.

Also adds a "Suggest for <role>" button to the routing table. Clicking
it applies a role-specific preset to the model catalog: transparent
include/exclude filters + a sort strategy — no opaque quality score,
no curated shortlist, just filters the user can read and relax.

Presets

  simple     tools, multilingual Russian-capable families, ≤ $1/M prompt,
             no :free / -coder / -vl variants. Sort: prompt price asc.
  default    same but ≤ $2/M prompt.
  complex    adds reasoning requirement, ≤ $5/M prompt. Claude via
             bridge remains the recommended option when configured.
  multimodal vision + tools, multilingual, ≤ $2/M prompt. Note: native
             audio transcription still uses direct Gemini outside OR.
  compaction ctx ≥ 32k, multilingual, sorted by COMPLETION price —
             summaries are output-heavy, so ranking by prompt price
             misrepresents the actual cost driver.
  classifier multilingual, ≤ $0.1/M prompt. Local Ollama stays primary.

The "fallback" role intentionally has no preset: it should point to a
DIRECT-provider slot (Gemini in our config) so an OpenRouter outage
doesn't take the whole bot down. Browsing OR candidates here would
be misleading.

Filter definitions live in internal/adminapi/recommend.go with a
conservative multilingual allowlist (Qwen3, Qwen Plus/Max/Turbo, GLM
4.5+, Kimi, Gemini 2.5/3.x, Mistral Large/Medium, Grok 3/4, DeepSeek).
English-primary small variants (nvidia, cohere-r-small) and specialised
-coder/-vl fine-tunes are excluded from general roles so the top of the
list doesn't fill up with things that don't speak Russian well.

UI

The models section is now one htmx-swappable div (banner + filters +
table). When a preset is active, a banner at the top names the role,
shows the preset's human-readable description, and offers a Clear
button to return to the raw catalog.

Config

config.yaml (dev + deploy) splits "workhorse" into four OR slots so the
model id per slot can diverge over time:
  simple-or     qwen/qwen3.5-flash-02-23         — $0.065, 1M ctx, V T R
  default-or    qwen/qwen3.5-flash-02-23         — same initial model
  complex-or    qwen/qwen3-235b-a22b-thinking... — fallback when bridge off
  compaction-or qwen/qwen3.5-9b                  — 262k ctx, cheap output

routing.{simple,default,complex,compaction} point at the new slots;
fallback stays on direct Gemini flash-lite.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: Alexey Panfilov

config/config.yaml

@@ -9,28 +9,35 @@ telegram:
 # Known providers: openrouter, gemini, ollama, claude-bridge, local, hf-tei, openai.
 # The key "embedding" is reserved for the MCP embedding provider.
 models:
-  # --- OpenRouter — one API key, many models. Add as many entries as you need. ---
-  workhorse:
+  # --- OpenRouter slots — one per role so each can be reassigned independently
+  #     from the admin UI. Initial models are suggestions; override via UI. ---
+  simple-or:
     provider: openrouter
-    model: ${OPENROUTER_MODEL:-deepseek/deepseek-chat-v3.1}
+    model: qwen/qwen3.5-flash-02-23        # V T R, 1M ctx, $0.065/$0.26
+    api_key: ${OPENROUTER_API_KEY}
+    max_tokens: 2048
+    base_url: https://openrouter.ai/api/v1
+
+  default-or:
+    provider: openrouter
+    model: qwen/qwen3.5-flash-02-23        # same as simple initially; diverge in UI
     api_key: ${OPENROUTER_API_KEY}
     max_tokens: 4096
     base_url: https://openrouter.ai/api/v1
-    vision: true
-
-  # Example: dedicate a separate OpenRouter model to a routing role.
-  # complex-or:
-  #   provider: openrouter
-  #   model: anthropic/claude-sonnet-4.5
-  #   api_key: ${OPENROUTER_API_KEY}
-  #   max_tokens: 8192
-  #   base_url: https://openrouter.ai/api/v1
-  # cheap-or:
-  #   provider: openrouter
-  #   model: google/gemini-2.5-flash-lite
-  #   api_key: ${OPENROUTER_API_KEY}
-  #   max_tokens: 2048
-  #   base_url: https://openrouter.ai/api/v1
+
+  complex-or:                              # alt to claude-bridge if bridge is off
+    provider: openrouter
+    model: qwen/qwen3-235b-a22b-thinking-2507
+    api_key: ${OPENROUTER_API_KEY}
+    max_tokens: 8192
+    base_url: https://openrouter.ai/api/v1
+
+  compaction-or:
+    provider: openrouter
+    model: qwen/qwen3.5-9b                 # min workable, 262k ctx, $0.10/$0.15
+    api_key: ${OPENROUTER_API_KEY}
+    max_tokens: 2048
+    base_url: https://openrouter.ai/api/v1
 
   # --- Gemini (direct Google API) — fallback + multimodal/transcription. ---
   gemini-flash-lite:
@@ -68,15 +75,15 @@ models:
     max_tokens: 120
 
 routing:
-  simple: workhorse            # level 1: simple/cheap tasks
-  default: workhorse           # level 2: moderate tasks (agentic loop)
-  complex: claude-bridge       # level 3: complex reasoning via Claude bridge
-  fallback: gemini-flash-lite
-  multimodal: gemini-flash     # vision/audio — also used for voice transcription
-  compaction: workhorse
-  classifier: classifier       # rates complexity 1/2/3
-  classifier_timeout: 15       # seconds
-  classifier_min_length: 0     # 0 = always classify; >0 = min chars; <0 = disabled
+  simple: simple-or
+  default: default-or
+  complex: claude-bridge         # via Anthropic Max subscription (host bridge)
+  compaction: compaction-or
+  fallback: gemini-flash-lite    # DIRECT Google — different vendor, survives OR outage
+  multimodal: gemini-flash       # DIRECT Google — native input_audio for voice
+  classifier: classifier         # local Ollama
+  classifier_timeout: 15
+  classifier_min_length: 0       # 0 = always; >0 = min chars; <0 = disabled
 
 tool_filter:
   top_k: 20

internal/adminapi/adminapi_test.go

@@ -41,7 +41,7 @@ func TestTemplatesParse(t *testing.T) {
 	cases := map[string]any{
 		viewIndex:       data,
 		viewRouting:     data.Routing, // routing view takes uiRouting directly
-		viewModelsTable: data,
+		viewModelsBrowser: data,
 	}
 	for v, d := range cases {
 		t.Run(v, func(t *testing.T) {

internal/adminapi/handlers.go

@@ -13,9 +13,10 @@ import (
 // --- View data ---
 
 type uiRole struct {
-	Name    string // e.g. "default", "reasoner"
-	Current string // slot name the role currently points at
-	ModelID string // underlying model id of that slot (if OR-backed); empty otherwise
+	Name      string // e.g. "default", "complex"
+	Current   string // slot name the role currently points at
+	ModelID   string // underlying model id of that slot (if OR-backed); empty otherwise
+	HasPreset bool   // true when a "Suggest" preset is defined for this role
 }
 
 type uiSlot struct {
@@ -40,11 +41,13 @@ type uiRouting struct {
 }
 
 type uiFilters struct {
-	Search    string
-	Free      bool
-	Vision    bool
-	Tools     bool
-	Reasoning bool
+	Search            string
+	Free              bool
+	Vision            bool
+	Tools             bool
+	Reasoning         bool
+	ActivePreset      string // role name when a preset is applied; empty otherwise
+	PresetDescription string // human-readable summary for the banner
 }
 
 type indexData struct {
@@ -68,7 +71,7 @@ func (s *Server) handleIndex(w http.ResponseWriter, r *http.Request) {
 func (s *Server) handleModels(w http.ResponseWriter, r *http.Request) {
 	data := s.buildIndexData(r) // reuses model filtering + slots
 	w.Header().Set("Content-Type", "text/html; charset=utf-8")
-	if err := render(w, viewModelsTable, data); err != nil {
+	if err := render(w, viewModelsBrowser, data); err != nil {
 		s.logger.Error("render models_table", "err", err)
 		http.Error(w, "render error", http.StatusInternalServerError)
 	}
@@ -180,19 +183,45 @@ func (s *Server) handleRefresh(w http.ResponseWriter, r *http.Request) {
 	// Re-render just the tbody.
 	data := s.buildIndexData(r)
 	w.Header().Set("Content-Type", "text/html; charset=utf-8")
-	if err := render(w, viewModelsTable, data); err != nil {
+	if err := render(w, viewModelsBrowser, data); err != nil {
 		s.logger.Error("render models after refresh", "err", err)
 	}
 }
 
 // --- Data builders ---
 
 func (s *Server) buildIndexData(r *http.Request) indexData {
-	// Tool calling is required for the agentic loop, so default the filter ON
-	// when the page is first loaded (no query string yet). Once the user
-	// submits the filters form, whatever they sent wins — including unchecking
-	// Tools to browse the full catalog.
 	q := r.URL.Query()
+	preset := q.Get("preset")
+
+	var allCaps map[string]llm.Capabilities
+	if s.capStore != nil {
+		ctx, cancel := context.WithTimeout(r.Context(), 5*time.Second)
+		defer cancel()
+		allCaps, _ = s.capStore.GetAllCapabilities(ctx, "openrouter")
+	}
+
+	// Preset path — pre-filter + pre-sort via the role's preset. Checkbox
+	// filters are ignored on this path: the preset is a complete override.
+	if p, ok := rolePresets[preset]; ok {
+		return indexData{
+			Routing: s.buildRouting(),
+			Slots:   s.openRouterSlots(),
+			Models:  applyPreset(allCaps, preset),
+			Filters: uiFilters{
+				ActivePreset:      preset,
+				PresetDescription: p.Description,
+				// reflect preset intent in the checkboxes so user can tell what's on
+				Tools:     requiresTools(preset),
+				Vision:    preset == "multimodal",
+				Reasoning: preset == "complex",
+			},
+		}
+	}
+
+	// Manual filter path — Tools defaults ON on initial page load (empty
+	// query string), since the agentic loop requires tool calling. Once the
+	// user submits the filters form, whatever they send wins.
 	formSubmitted := len(q) > 0
 	f := uiFilters{
 		Search:    strings.ToLower(strings.TrimSpace(q.Get("q"))),
@@ -202,13 +231,6 @@ func (s *Server) buildIndexData(r *http.Request) indexData {
 		Reasoning: q.Get("reasoning") != "",
 	}
 
-	var allCaps map[string]llm.Capabilities
-	if s.capStore != nil {
-		ctx, cancel := context.WithTimeout(r.Context(), 5*time.Second)
-		defer cancel()
-		allCaps, _ = s.capStore.GetAllCapabilities(ctx, "openrouter")
-	}
-
 	models := make([]uiModel, 0, len(allCaps))
 	for id, c := range allCaps {
 		free := c.Free()
@@ -257,6 +279,15 @@ func (s *Server) buildIndexData(r *http.Request) indexData {
 	}
 }
 
+// requiresTools returns true when the preset's filter requires tool-calling.
+func requiresTools(preset string) bool {
+	switch preset {
+	case "simple", "default", "complex", "multimodal":
+		return true
+	}
+	return false
+}
+
 func (s *Server) buildRouting() uiRouting {
 	cfg := s.router.GetConfig()
 	allSlots := s.router.ProviderNames()
@@ -277,12 +308,14 @@ func (s *Server) buildRouting() uiRouting {
 		"compaction": cfg.Compaction,
 	}
 	order := []string{"simple", "default", "complex", "multimodal", "fallback", "classifier", "compaction"}
+	presets := presetRoles()
 	roles := make([]uiRole, 0, len(order))
 	for _, r := range order {
 		roles = append(roles, uiRole{
-			Name:    r,
-			Current: roleCur[r],
-			ModelID: orSlots[roleCur[r]],
+			Name:      r,
+			Current:   roleCur[r],
+			ModelID:   orSlots[roleCur[r]],
+			HasPreset: presets[r],
 		})
 	}
 	return uiRouting{Roles: roles, AllSlots: allSlots}