refactor(quality-score): update provider prestige logic and adjust top-K size for improved candidate selection

surfsense_backend/app/services/quality_score.py

@@ -26,8 +26,10 @@ from __future__ import annotations
 # Tunables (constants, not flags)
 # ---------------------------------------------------------------------------
 
-# Top-K size for deterministic spread inside the locked tier.
-_QUALITY_TOP_K: int = 5
+# Top-K size for deterministic spread inside the locked tier. Auto only
+# ever pins one of the K best-scoring candidates, so this is the effective
+# "shortlist" size regardless of how large the OpenRouter catalogue is.
+_QUALITY_TOP_K: int = 4
 
 # Hard health gate: any cfg whose best non-null uptime is below this %
 # is excluded from Auto-mode selection entirely.
@@ -74,39 +76,20 @@ _TINY_LEGACY_PENALTY_PATTERNS: tuple[str, ...] = (
 # ---------------------------------------------------------------------------
 
 # OpenRouter-side provider slug (the prefix before ``/`` in the model id).
-# Tiers are coarse: frontier labs > strong open / fast-moving labs >
-# specialist labs > everything else.
+# Deliberately frontier-labs-only: anything not listed falls through to the
+# neutral default (15) so smaller labs can't out-prestige the big four and
+# crowd into the Auto top-K shortlist.
 PROVIDER_PRESTIGE_OR: dict[str, int] = {
-    # Frontier labs
     "openai": 50,
     "anthropic": 50,
     "google": 50,
     "x-ai": 50,
-    # Strong open / fast-moving labs
-    "deepseek": 38,
-    "qwen": 38,
-    "meta-llama": 38,
-    "mistralai": 38,
-    "cohere": 38,
-    "nvidia": 38,
-    "alibaba": 38,
-    # Specialist / regional / strong second-tier
-    "microsoft": 28,
-    "01-ai": 28,
-    "minimax": 28,
-    "moonshot": 28,
-    "z-ai": 28,
-    "nousresearch": 28,
-    "ai21": 28,
-    "perplexity": 28,
-    # Smaller / niche providers
-    "liquid": 18,
-    "cognitivecomputations": 18,
-    "venice": 18,
-    "inflection": 18,
 }
 
 # YAML provider field (the upstream API shape the operator selected).
+# Frontier-labs-only, mirroring PROVIDER_PRESTIGE_OR; unlisted providers get
+# the neutral default (15). ``ollama_chat`` / ``custom`` stay listed because
+# they sit *below* the default — removing them would raise their score.
 PROVIDER_PRESTIGE_YAML: dict[str, int] = {
     "azure": 50,
     "openai": 50,
@@ -114,16 +97,6 @@ PROVIDER_PRESTIGE_YAML: dict[str, int] = {
     "gemini": 50,
     "vertex_ai": 50,
     "xai": 50,
-    "mistral": 38,
-    "deepseek": 38,
-    "cohere": 38,
-    "groq": 30,
-    "together_ai": 28,
-    "fireworks_ai": 28,
-    "perplexity": 28,
-    "bedrock": 28,
-    "openrouter": 25,
-    "requesty": 25,
     "ollama_chat": 12,
     "custom": 12,
 }

surfsense_web/lib/query-client/query-client.provider.check.tsx

@@ -0,0 +1,40 @@
+/**
+ * Self-check: the app must run on ONE jotai store.
+ *
+ * Regression guard for the "screenshot never sent" bug: wrapping the tree in
+ * jotai-tanstack-query's QueryClientAtomProvider mounted a second jotai
+ * <Provider> store, so hook writes (pending screenshot data URLs, mentions)
+ * were invisible to the chat stream engine's getDefaultStore() reads and
+ * user_images arrived empty at the backend.
+ *
+ * Run: pnpm exec tsx lib/query-client/query-client.provider.check.tsx
+ */
+import assert from "node:assert";
+import { getDefaultStore, useStore } from "jotai";
+import { queryClientAtom } from "jotai-tanstack-query";
+import { renderToString } from "react-dom/server";
+import { ReactQueryClientProvider } from "./query-client.provider";
+
+let seenStore: unknown;
+function Probe() {
+	seenStore = useStore();
+	return null;
+}
+
+renderToString(
+	<ReactQueryClientProvider>
+		<Probe />
+	</ReactQueryClientProvider>
+);
+
+assert.strictEqual(
+	seenStore,
+	getDefaultStore(),
+	"jotai hooks inside ReactQueryClientProvider must use the default store " +
+		"(a nested jotai <Provider> would hide composer atom writes from the chat engine)"
+);
+assert.ok(
+	getDefaultStore().get(queryClientAtom),
+	"queryClientAtom must be hydrated into the default store"
+);
+console.log("OK: single jotai store, queryClientAtom hydrated");

surfsense_web/lib/query-client/query-client.provider.tsx

@@ -1,5 +1,7 @@
 "use client";
-import { QueryClientAtomProvider } from "jotai-tanstack-query/react";
+import { QueryClientProvider } from "@tanstack/react-query";
+import { getDefaultStore } from "jotai";
+import { queryClientAtom } from "jotai-tanstack-query";
 import dynamic from "next/dynamic";
 import { queryClient } from "./client";
 
@@ -8,11 +10,20 @@ const ReactQueryDevtools = dynamic(
 	{ ssr: false }
 );
 
+// Do NOT use jotai-tanstack-query's QueryClientAtomProvider here: it mounts
+// its own jotai <Provider>, which creates a SECOND jotai store for the React
+// tree. Hook writes (e.g. pending screenshot data URLs, @-mentions) then land
+// in that provider store while the chat stream engine reads via
+// getDefaultStore(), so user_images silently arrived empty at the backend.
+// Instead, keep the whole app on the one default store and hydrate the
+// query-client atom into it directly.
+getDefaultStore().set(queryClientAtom, queryClient);
+
 export function ReactQueryClientProvider({ children }: { children: React.ReactNode }) {
 	return (
-		<QueryClientAtomProvider client={queryClient}>
+		<QueryClientProvider client={queryClient}>
 			{children}
 			{process.env.NODE_ENV === "development" && <ReactQueryDevtools initialIsOpen={false} />}
-		</QueryClientAtomProvider>
+		</QueryClientProvider>
 	);
 }