fix: improve provider routing and assignments

.gitignore

@@ -26,6 +26,7 @@ product
 # misc
 .DS_Store
 *.pem
+.vscode
 
 # debug
 npm-debug.log*
@@ -84,4 +85,4 @@ graphify-out/*
 .script/
 .codegraph/
 .PR/
-.next-analyze/*
+.next-analyze/*

open-sse/executors/freebuff.js

@@ -75,13 +75,40 @@ function injectFreebuffMarker(body) {
   return { ...body, messages: [{ role: "system", content: FREEBUFF_SYSTEM_MARKER }, ...messages] };
 }
 
-// Freebuff root agent id per model (mirrors the CLI's FREEBUFF_ROOT_AGENT_ID_BY_MODEL).
+// The backend's foreign_toolset gate rejects any tool-calling request whose
+// toolset lacks the CLI's `end_turn` tool with a misleading 404 "No endpoints
+// found for {model}" (verified live 2026-08-14; the freebuff-proxy bridge
+// works around the same gate by injecting this definition). Every request that
+// declares tools must carry it or the router finds no serving endpoint.
+const END_TURN_TOOL = {
+  type: "function",
+  function: {
+    name: "end_turn",
+    description: "Signal the end of the current task.",
+    parameters: { type: "object", properties: {} },
+  },
+};
+
+function injectEndTurnTool(body) {
+  const tools = body?.tools;
+  if (!Array.isArray(tools) || tools.length === 0) return body;
+  const hasEndTurn = tools.some(
+    (t) => t?.function?.name === "end_turn",
+  );
+  if (hasEndTurn) return body;
+  return { ...body, tools: [...tools, END_TURN_TOOL] };
+}
+
+// Freebuff root agent id per model (mirrors the CLI's
+// FREEBUFF_CLI_BASE3_AGENT_ID_BY_MODEL — the CLI harness moved from base2 to
+// base3, and the backend can return 404 "No endpoints found" for the old
+// base2 roots during the transition).
 const FREE_ROOT_AGENT_BY_MODEL = {
-  "deepseek/deepseek-v4-flash": "base2-free-deepseek-flash",
-  "deepseek/deepseek-v4-pro": "base2-free-deepseek",
-  "mimo/mimo-v2.5": "base2-free-mimo",
-  "minimax/minimax-m3": "base2-free-minimax-m3",
-  "openai/gpt-5.6-luna": "base2-free-luna",
+  "deepseek/deepseek-v4-flash": "base3-free-deepseek-flash",
+  "deepseek/deepseek-v4-pro": "base3-free-deepseek",
+  "mimo/mimo-v2.5": "base3-free-mimo",
+  "minimax/minimax-m3": "base3-free-minimax-m3",
+  "openai/gpt-5.6-luna": "base3-free-luna",
 };
 
 // Per-token+model session cache (in-memory; keyed so multi-account setups
@@ -414,6 +441,23 @@ export class FreebuffExecutor extends BaseExecutor {
     return this.config.baseUrl;
   }
 
+  // The backend's model router answers 404 "No endpoints found for {model}"
+  // when a tool-calling request's toolset fails its foreign_toolset gate —
+  // normally prevented by injecting the CLI's `end_turn` tool (see
+  // injectEndTurnTool). If one still slips through, surface a helpful message
+  // instead of a bare 404, and let the standard cooldown pace retries.
+  async parseError(response, bodyText) {
+    const text = String(bodyText || "");
+    if (response?.status === 404 && /No endpoints found/i.test(text)) {
+      return {
+        status: 404,
+        message: `Freebuff upstream rejected the request (404: "${text.trim().slice(0, 90)}"). Tool-calling requests need the CLI's end_turn tool — retry; if it persists the Codebuff backend may be having trouble.`,
+        resetsAtMs: Date.now() + 120_000,
+      };
+    }
+    return super.parseError(response, bodyText);
+  }
+
   transformRequest(model, body, stream, credentials) {
     // Top-level wire shape — see header comment. `run_id` and
     // `freebuff_instance_id` are attached by execute() (they need the async
@@ -425,14 +469,16 @@ export class FreebuffExecutor extends BaseExecutor {
       cost_mode: "free",
     };
     body.provider = { allow_fallbacks: false };
-    // Freebuff agents (base2-free-*) own reasoning: the backend applies the
+    // Freebuff agents (base3-free-*) own reasoning: the backend applies the
     // agent's reasoningOptions.effort server-side, so a client-sent
     // reasoning_effort / reasoning.effort collides with that default →
     // 400 "both provided with conflicting values". Mirror the CLI: send none.
     delete body.reasoning_effort;
     delete body.reasoning;
     // Free-tier gate: first system message must open with the CLI marker.
-    return injectFreebuffMarker(body);
+    body = injectFreebuffMarker(body);
+    // Foreign-toolset gate: tool-calling requests must declare `end_turn`.
+    return injectEndTurnTool(body);
   }
 
   async execute({ model, body, stream, credentials, signal, log, proxyOptions = null }) {
@@ -638,6 +684,7 @@ export const __test__ = {
   resetSessionCache,
   rootAgentIdForModel,
   injectFreebuffMarker,
+  injectEndTurnTool,
   fetchWithNetworkRetry,
   FREEBUFF_SYSTEM_MARKER,
   SESSION_STALE_CODES,

open-sse/handlers/chatCore.js

@@ -309,8 +309,29 @@ export async function handleChatCore({ body, modelInfo, credentials, log, onCred
     proxyPoolId: psd?.proxyPoolId || psd?.connectionProxyPoolId || null,
   });
 
+  const proxyScope = `${provider}::${model}`;
   let proxyOptions = buildProxyOptions(credentials?.providerSpecificData || {});
 
+  if (provider === "freebuff" && credentials?.providerSpecificData?.noFitPool === true) {
+    const error = new Error(`Freebuff has no healthy proxy pool for ${model}; all assigned pools are cooling down after limited-IP errors.`);
+    error.status = 503;
+    error.poolScoped = { poolId: null, scope: proxyScope, reason: "no_fit_pool" };
+    trackPendingRequest(model, provider, connectionId, false, true);
+    return createErrorResult(503, error.message);
+  }
+
+  if (
+    provider === "freebuff" &&
+    !proxyOptions.proxyPoolId &&
+    !proxyOptions.vercelRelayUrl &&
+    !(proxyOptions.connectionProxyEnabled && proxyOptions.connectionProxyUrl)
+  ) {
+    const error = new Error(`Freebuff requires a configured proxy pool for ${model}; direct egress is disabled to prevent limited-IP rate limits.`);
+    error.status = 503;
+    trackPendingRequest(model, provider, connectionId, false, true);
+    return createErrorResult(503, error.message);
+  }
+
   if (proxyOptions.vercelRelayUrl) {
     const connectionName = credentials?.connectionName || credentials?.connectionId || "unknown";
     const poolId = proxyOptions.proxyPoolId || "none";
@@ -342,7 +363,6 @@ export async function handleChatCore({ body, modelInfo, credentials, log, onCred
   // another pool instead of failing the account. Covers both thrown errors
   // (executor.execute) and non-ok responses declared poolScoped via
   // parseError — poolId/scope are completed here from proxyOptions.
-  const proxyScope = `${provider}::${model}`;
   let parsedNonOk = null;
 
   const tryNextPool = async (poolScoped, reasonMsg) => {

open-sse/services/proxyPoolFitness.js

@@ -17,14 +17,49 @@
 
 const FITNESS_STATE_KEY = "__9routerPoolFitness__";
 const fitness = (globalThis[FITNESS_STATE_KEY] ??= new Map()); // poolId -> Map<scope, { until, reason }>
+let persistTimer = null;
+let hydratePromise = null;
 
 export const POOL_UNFIT_MS = 5 * 60 * 1000;
 
+function schedulePersist() {
+  if (persistTimer) return;
+  persistTimer = setTimeout(async () => {
+    persistTimer = null;
+    try {
+      const { updateSettings } = await import("@/lib/db/repos/settingsRepo.js");
+      await updateSettings({ proxyPoolFitness: poolFitnessSnapshot() });
+    } catch {
+      // Fitness is advisory; persistence failures must not block requests.
+    }
+  }, 25);
+  if (persistTimer.unref) persistTimer.unref();
+}
+
+export function hydratePoolFitness(snapshot = {}) {
+  for (const [poolId, byScope] of Object.entries(snapshot || {})) {
+    const entries = Object.entries(byScope || {}).filter(([, entry]) => Number(entry?.until) > Date.now());
+    if (entries.length) fitness.set(poolId, new Map(entries));
+  }
+}
+
+export async function ensurePoolFitnessHydrated() {
+  if (!hydratePromise) {
+    hydratePromise = import("@/lib/db/repos/settingsRepo.js")
+      .then(({ getSettings }) => getSettings())
+      .then((settings) => hydratePoolFitness(settings.proxyPoolFitness || {}))
+      .catch(() => {})
+      .then(() => undefined);
+  }
+  return hydratePromise;
+}
+
 export function markPoolUnfit(poolId, scope, until = Date.now() + POOL_UNFIT_MS, reason = "") {
   if (!poolId || !scope) return;
   const byScope = fitness.get(poolId) || new Map();
   byScope.set(scope, { until, reason });
   fitness.set(poolId, byScope);
+  schedulePersist();
 }
 
 export function clearPoolUnfit(poolId, scope) {
@@ -32,6 +67,7 @@ export function clearPoolUnfit(poolId, scope) {
   if (!byScope) return;
   byScope.delete(scope);
   if (byScope.size === 0) fitness.delete(poolId);
+  schedulePersist();
 }
 
 // "provider::model" -> "provider::*" (null when the scope has no provider part)
@@ -77,9 +113,11 @@ export function clearAllPoolUnfit(provider = null) {
       }
       if (byScope.size === 0) fitness.delete(poolId);
     }
+    schedulePersist();
     return;
   }
   fitness.clear();
+  schedulePersist();
 }
 
 // Snapshot of live (non-expired) marks — expired entries are pruned here so
@@ -87,6 +125,8 @@ export function clearAllPoolUnfit(provider = null) {
 // Test helper: drop all marks (module state is globalThis-backed).
 export function resetPoolFitness() {
   fitness.clear();
+  hydratePromise = Promise.resolve();
+  schedulePersist();
 }
 
 // Sweep all expired marks. Returns how many scope entries were removed.
@@ -101,6 +141,7 @@ export function pruneExpired(now = Date.now()) {
     }
     if (byScope.size === 0) fitness.delete(poolId);
   }
+  if (removed) schedulePersist();
   return removed;
 }
 

src/app/(dashboard)/dashboard/providers/[id]/ConnectionRow.js

@@ -6,7 +6,7 @@ import PropTypes from "prop-types";
 import { Badge, Toggle, Tooltip } from "@/shared/components";
 import CooldownTimer from "./CooldownTimer";
 
-export default function ConnectionRow({ connection, proxyPools, isOAuth, isFirst, isLast, onMoveUp, onMoveDown, onToggleActive, onUpdateProxy, onEdit, onDelete, oneByOneStatus = null, autoPing = null }) {
+export default function ConnectionRow({ connection, proxyPools, isOAuth, isFirst, isLast, onMoveUp, onMoveDown, onToggleActive, onUpdateProxy, onEdit, onDelete, oneByOneStatus = null, autoPing = null, modelAssignmentOptions = null, onModelAssignmentChange = null, strictModelAssignment = false }) {
   const [showProxyDropdown, setShowProxyDropdown] = useState(false);
   const [updatingProxy, setUpdatingProxy] = useState(false);
   const [selectedProxyIds, setSelectedProxyIds] = useState([]);
@@ -19,13 +19,14 @@ export default function ConnectionRow({ connection, proxyPools, isOAuth, isFirst
     const legacyProxyPoolId = connection.providerSpecificData?.proxyPoolId;
     
     // Migrate legacy single proxy to array format
-    if (legacyProxyPoolId && proxyPoolIds.length === 0) {
-      setSelectedProxyIds([legacyProxyPoolId]);
-    } else {
-      setSelectedProxyIds(proxyPoolIds);
-    }
-    
-    setRotationStrategy(connection.providerSpecificData?.proxyRotationStrategy || "none");
+    queueMicrotask(() => {
+      if (legacyProxyPoolId && proxyPoolIds.length === 0) {
+        setSelectedProxyIds([legacyProxyPoolId]);
+      } else {
+        setSelectedProxyIds(proxyPoolIds);
+      }
+      setRotationStrategy(connection.providerSpecificData?.proxyRotationStrategy || "none");
+    });
   }, [connection]);
 
   const proxyPoolMap = new Map((proxyPools || []).map((pool) => [pool.id, pool]));
@@ -201,7 +202,7 @@ export default function ConnectionRow({ connection, proxyPools, isOAuth, isFirst
     return () => {
       if (interval) clearInterval(interval);
     };
-  }, [modelLockUntil]);
+  }, [connection, modelLockUntil]);
 
   // Determine effective status (override unavailable if cooldown expired)
   const effectiveStatus = (connection.testStatus === "unavailable" && !isCooldown)
@@ -283,6 +284,20 @@ export default function ConnectionRow({ connection, proxyPools, isOAuth, isFirst
               </Badge>
             )}
           </div>
+           {modelAssignmentOptions && onModelAssignmentChange && (
+             <select
+               value={connection.providerSpecificData?.assignedModel || connection.providerSpecificData?.freebuffModel || ""}
+              onChange={(e) => onModelAssignmentChange(e.target.value)}
+              disabled={!strictModelAssignment}
+               className="mt-2 max-w-full rounded-md border border-border bg-background px-2 py-1 text-[11px] text-text-main"
+               title="Model assignment"
+             >
+               <option value="">Unassigned</option>
+               {modelAssignmentOptions.map((model) => (
+                <option key={model.id} value={model.id}>{model.name || model.id}</option>
+              ))}
+            </select>
+          )}
           {hasAnyProxy && (
             <div className="mt-1 flex items-center gap-2 flex-wrap">
               <span className="max-w-full truncate text-[11px] text-text-muted sm:max-w-[420px]" title={proxyDisplayText}>
@@ -495,6 +510,9 @@ ConnectionRow.propTypes = {
   onUpdateProxy: PropTypes.func,
   onEdit: PropTypes.func.isRequired,
   onDelete: PropTypes.func.isRequired,
+  modelAssignmentOptions: PropTypes.arrayOf(PropTypes.shape({ id: PropTypes.string.isRequired, name: PropTypes.string })),
+  onModelAssignmentChange: PropTypes.func,
+  strictModelAssignment: PropTypes.bool,
   oneByOneStatus: PropTypes.shape({
     state: PropTypes.string,
     error: PropTypes.string,

src/app/(dashboard)/dashboard/providers/[id]/page.js

@@ -64,6 +64,7 @@ export default function ProviderDetailPage() {
   const [bulkUpdatingProxy, setBulkUpdatingProxy] = useState(false);
   const [providerStrategy, setProviderStrategy] = useState(null);
   const [providerStickyLimit, setProviderStickyLimit] = useState("");
+  const [strictModelAssignment, setStrictModelAssignment] = useState(false);
   const [thinkingMode, setThinkingMode] = useState("auto");
   const [autoPing, setAutoPing] = useState({ enabled: false, connections: {} });
   const [suggestedModels, setSuggestedModels] = useState([]);
@@ -170,6 +171,21 @@ export default function ProviderDetailPage() {
     return levels && levels.includes(thinkingMode) ? thinkingMode : null;
   };
   const providerStorageAlias = isCompatible ? providerId : providerAlias;
+  const assignmentModels = (() => {
+    const byId = new Map();
+    const add = (model) => {
+      if (model?.id && !byId.has(model.id)) byId.set(model.id, model);
+    };
+    models.forEach(add);
+    kiloFreeModels.forEach(add);
+    customModels.forEach((model) => {
+      if (model.providerAlias === providerStorageAlias && (model.kind || model.type || "llm") === "llm") {
+        add(model);
+      }
+    });
+    const disabled = new Set(disabledModelIds);
+    return [...byId.values()].filter((model) => !disabled.has(model.id));
+  })();
   // Union of levels across this provider's reasoning models — drives the level picker options.
   // Include custom models too (e.g. manually added gpt-5.6-sol → max).
   const providerThinkingLevels = (() => {
@@ -313,6 +329,7 @@ export default function ProviderDetailPage() {
       const override = (settingsData.providerStrategies || {})[providerId] || {};
       setProviderStrategy(override.fallbackStrategy || null);
       setProviderStickyLimit(override.stickyRoundRobinLimit != null ? String(override.stickyRoundRobinLimit) : "1");
+      setStrictModelAssignment(override.strictModelAssignment === true);
       // Load per-provider thinking config
       const thinkingCfg = (settingsData.providerThinking || {})[providerId] || {};
       setThinkingMode(thinkingCfg.mode || "auto");
@@ -368,9 +385,13 @@ export default function ProviderDetailPage() {
       const settingsData = settingsRes.ok ? await settingsRes.json() : {};
       const current = settingsData.providerStrategies || {};
 
-      // Build override: null strategy means remove override, use global
-      const override = {};
+      // Preserve Freebuff-only settings while changing the shared strategy.
+      const override = { ...(current[providerId] || {}) };
       if (strategy) override.fallbackStrategy = strategy;
+      else {
+        delete override.fallbackStrategy;
+        delete override.stickyRoundRobinLimit;
+      }
       if (strategy === "round-robin" && stickyLimit !== "") {
         override.stickyRoundRobinLimit = Number(stickyLimit) || 3;
       }
@@ -392,6 +413,44 @@ export default function ProviderDetailPage() {
     }
   };
 
+  const handleStrictAssignmentToggle = async (enabled) => {
+    setStrictModelAssignment(enabled);
+    try {
+      const settingsRes = await fetch("/api/settings", { cache: "no-store" });
+      const settingsData = settingsRes.ok ? await settingsRes.json() : {};
+      const current = settingsData.providerStrategies || {};
+      await fetch("/api/settings", {
+        method: "PATCH",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({
+          providerStrategies: {
+            ...current,
+            [providerId]: { ...(current[providerId] || {}), strictModelAssignment: enabled },
+          },
+        }),
+      });
+    } catch (error) {
+      console.log("Error saving Freebuff strict assignment:", error);
+    }
+  };
+
+  const handleModelAssignment = async (connectionId, assignedModel) => {
+    try {
+      const res = await fetch(`/api/providers/${connectionId}`, {
+        method: "PUT",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ providerSpecificData: { assignedModel: assignedModel || null } }),
+      });
+      if (res.ok) {
+        setConnections((prev) => prev.map((c) => c.id === connectionId
+          ? { ...c, providerSpecificData: { ...(c.providerSpecificData || {}), assignedModel: assignedModel || null } }
+          : c));
+      }
+    } catch (error) {
+      console.log("Error saving Freebuff model assignment:", error);
+    }
+  };
+
   const handleRoundRobinToggle = (enabled) => {
     const strategy = enabled ? "round-robin" : null;
     const sticky = enabled ? (providerStickyLimit || "1") : providerStickyLimit;
@@ -452,10 +511,12 @@ export default function ProviderDetailPage() {
   };
 
   useEffect(() => {
-    fetchConnections();
-    fetchAliases();
-    fetchCustomModels();
-    fetchDisabledModels();
+    Promise.resolve().then(() => {
+      fetchConnections();
+      fetchAliases();
+      fetchCustomModels();
+      fetchDisabledModels();
+    });
   }, [fetchConnections, fetchAliases, fetchCustomModels, fetchDisabledModels]);
 
   // Cursor's model availability is account-specific and changes frequently.
@@ -463,13 +524,13 @@ export default function ProviderDetailPage() {
   // registry remains the fallback while the request is pending or unavailable.
   useEffect(() => {
     if (providerId !== "cursor") {
-      setLiveModels([]);
+      queueMicrotask(() => setLiveModels([]));
       return;
     }
 
     const connection = connections.find((item) => item.isActive !== false);
     if (!connection?.id) {
-      setLiveModels([]);
+      queueMicrotask(() => setLiveModels([]));
       return;
     }
 
@@ -867,7 +928,9 @@ export default function ProviderDetailPage() {
   };
 
   useEffect(() => {
-    setSelectedConnectionIds((prev) => prev.filter((id) => connections.some((conn) => conn.id === id)));
+    queueMicrotask(() => {
+      setSelectedConnectionIds((prev) => prev.filter((id) => connections.some((conn) => conn.id === id)));
+    });
   }, [connections]);
 
   const selectedProxySummary = (() => {
@@ -1014,6 +1077,9 @@ export default function ProviderDetailPage() {
                 }}
                 onDelete={() => handleDelete(conn.id)}
                 oneByOneStatus={oneByOneResults[conn.id] || null}
+                modelAssignmentOptions={assignmentModels}
+                onModelAssignmentChange={(model) => handleModelAssignment(conn.id, model)}
+                strictModelAssignment={strictModelAssignment}
               />
             </div>
           </div>
@@ -1511,6 +1577,13 @@ export default function ProviderDetailPage() {
                   </div>
                 )}
               </div>
+              <div className="flex flex-wrap items-center gap-2 border-t border-black/[0.03] pt-2 dark:border-white/[0.03]">
+                <div>
+                  <span className="text-xs text-text-muted font-medium">Strict Model Assignment</span>
+                  <p className="text-[10px] text-text-muted">Only assigned accounts can serve each model for this provider.</p>
+                </div>
+                <Toggle checked={strictModelAssignment} onChange={handleStrictAssignmentToggle} />
+              </div>
             </div>
           </div>
 

src/app/(dashboard)/dashboard/usage/components/ProviderLimits/utils.js

@@ -522,6 +522,23 @@ export function parseQuotaData(provider, data) {
         }
         break;
 
+      case "freebuff":
+        // Session quotas keyed by model id — label rows with the friendly
+        // displayName (from the registry) and keep modelKey for ordering.
+        if (data.quotas) {
+          Object.entries(data.quotas).forEach(([modelKey, quota]) => {
+            normalizedQuotas.push({
+              name: quota.displayName || modelKey,
+              modelKey,
+              used: quota.used || 0,
+              total: quota.total || 0,
+              resetAt: quota.resetAt || null,
+              recurring: quota.recurring !== false,
+            });
+          });
+        }
+        break;
+
       case "ollama":
         // Session (5h) / Weekly (7d) usage % from ollama.com/api/usage.
         // remainingPercentage only — no absolute remaining (UI treats remaining as %).

src/app/api/providers/[id]/test/testUtils.js

@@ -102,9 +102,24 @@ const OAUTH_TEST_CONFIG = {
     },
     refreshable: false,
   },
+  freebuff: {
+    // The session endpoint doubles as the auth probe: GET never claims a
+    // session (POST would burn 1.0 unit of the daily quota). Mirrors the usage
+    // handler: 401 = bad token, 403 = region/account gate (token still valid),
+    // 404 = no session row yet (pre-join, token valid). No refresh path —
+    // when the authToken dies the user re-logs in.
+    url: "https://www.codebuff.com/api/v1/freebuff/session",
+    method: "GET",
+    authHeader: "Authorization",
+    authPrefix: "Bearer ",
+    extraHeaders: { Accept: "application/json", "User-Agent": "codebuff-cli/0.0.138" },
+    acceptStatuses: [403, 404],
+    softFailMessage: {
+      403: "Connected, but Freebuff is gated (403) — country blocked or account banned.",
+    },
+  },
   // Grok CLI / Grok Build — probe /v1/user (no inference quota). Headers mirror official CLI.
-  "grok-cli": {
-    url: PROVIDERS["grok-cli"]?.userUrl || "https://cli-chat-proxy.grok.com/v1/user",
+  "grok-cli": {    url: PROVIDERS["grok-cli"]?.userUrl || "https://cli-chat-proxy.grok.com/v1/user",
     method: "GET",
     authHeader: "Authorization",
     authPrefix: "Bearer ",

src/app/api/proxy-pools/[id]/fitness/clear/route.js

@@ -1,10 +1,11 @@
 import { NextResponse } from "next/server";
-import { clearPoolUnfit } from "open-sse/services/proxyPoolFitness.js";
+import { clearPoolUnfit, ensurePoolFitnessHydrated } from "open-sse/services/proxyPoolFitness.js";
 
 // POST /api/proxy-pools/[id]/fitness/clear
 // Body: { scope: "provider::model" } — clears the mark for this pool + scope.
 export async function POST(request, { params }) {
   try {
+    await ensurePoolFitnessHydrated();
     const { id } = await params;
     let body = {};
     try {
@@ -22,4 +23,4 @@ export async function POST(request, { params }) {
     console.log("Error clearing pool fitness:", error);
     return NextResponse.json({ error: "Failed to clear pool fitness" }, { status: 500 });
   }
-}
+}

src/app/api/proxy-pools/fitness/clear-all/route.js

@@ -1,11 +1,12 @@
 import { NextResponse } from "next/server";
-import { clearAllPoolUnfit } from "open-sse/services/proxyPoolFitness.js";
+import { clearAllPoolUnfit, ensurePoolFitnessHydrated } from "open-sse/services/proxyPoolFitness.js";
 
 // POST /api/proxy-pools/fitness/clear-all
 // Body: { provider?: string } — clears every mark, or only marks scoped to the
 // given provider ("provider::*") when provided.
 export async function POST(request) {
   try {
+    await ensurePoolFitnessHydrated();
     let body = {};
     try {
       body = await request.json();
@@ -19,4 +20,4 @@ export async function POST(request) {
     console.log("Error clearing proxy fitness:", error);
     return NextResponse.json({ error: "Failed to clear proxy fitness" }, { status: 500 });
   }
-}
+}

src/app/api/proxy-pools/fitness/route.js

@@ -1,13 +1,14 @@
 import { NextResponse } from "next/server";
-import { poolFitnessSnapshot } from "open-sse/services/proxyPoolFitness.js";
+import { ensurePoolFitnessHydrated, poolFitnessSnapshot } from "open-sse/services/proxyPoolFitness.js";
 
 // GET /api/proxy-pools/fitness — in-memory snapshot of pool fitness marks.
 // Returns { pools: { [poolId]: { [scope]: { until, reason } } } }.
 export async function GET() {
   try {
+    await ensurePoolFitnessHydrated();
     return NextResponse.json({ pools: poolFitnessSnapshot() });
   } catch (error) {
     console.log("Error reading proxy fitness:", error);
     return NextResponse.json({ error: "Failed to read proxy fitness" }, { status: 500 });
   }
-}
+}

src/lib/network/connectionProxy.js

@@ -1,5 +1,5 @@
 import { getProxyPoolById } from "@/models";
-import { fitPoolIds } from "open-sse/services/proxyPoolFitness.js";
+import { ensurePoolFitnessHydrated, fitPoolIds } from "open-sse/services/proxyPoolFitness.js";
 
 // Safely normalize any value into a trimmed string.
 function normalizeString(value) {
@@ -30,8 +30,10 @@ export function pickProxyPoolId(poolIds, strategy, providerId, opts = {}) {
   if (strategy === "smart" && scope) eligible = fitPoolIds(eligible, scope);
 
   if (eligible.length === 0) {
-    // Every candidate is unfit/excluded — fall back to the first non-excluded
-    // pool rather than deadlocking; its egress may have recovered.
+    // Freebuff must never reuse a limited-IP egress. Other providers retain
+    // the previous fail-open behavior when every smart candidate is marked
+    // unfit; their executors may have their own pool fallback semantics.
+    if (providerId === "freebuff" && strategy === "smart") return null;
     eligible = poolIds.filter((id) => !(excludeIds || []).includes(id));
     if (eligible.length === 0) return null;
   }
@@ -88,6 +90,7 @@ export async function resolveConnectionProxyConfig(
   excludePoolIds = null
 ) {
   try {
+    await ensurePoolFitnessHydrated();
     // Handle new multi-proxy format
     const proxyPoolIds = providerSpecificData?.proxyPoolIds || [];
     const proxyRotationStrategy = providerSpecificData?.proxyRotationStrategy || "none";
@@ -97,6 +100,8 @@ export async function resolveConnectionProxyConfig(
     const proxyPoolIdRaw = legacyProxyPoolId === "__none__" ? "" : legacyProxyPoolId;
 
     const legacy = normalizeLegacyProxy(providerSpecificData);
+    const multiPoolScope = providerSpecificData?.proxyPoolScope || null;
+    let selectedPoolId = null;
 
     /**
      * -----------------------------
@@ -104,10 +109,9 @@ export async function resolveConnectionProxyConfig(
      * -----------------------------
      */
     if (proxyPoolIds.length > 0) {
-      const scope = providerSpecificData?.proxyPoolScope || null;
-      const selectedPoolId = pickProxyPoolId(proxyPoolIds, proxyRotationStrategy, connectionId, { scope, excludeIds: excludePoolIds });
+      selectedPoolId = pickProxyPoolId(proxyPoolIds, proxyRotationStrategy, connectionId, { scope: multiPoolScope, excludeIds: excludePoolIds });
       
-      if (selectedPoolId) {
+    if (selectedPoolId) {
         const proxyPool = await getProxyPoolById(selectedPoolId);
         const proxyUrl = normalizeString(proxyPool?.proxyUrl);
         const noProxy = normalizeString(proxyPool?.noProxy);
@@ -147,6 +151,22 @@ export async function resolveConnectionProxyConfig(
         }
       }
     }
+    if (
+      !selectedPoolId &&
+      proxyRotationStrategy === "smart" &&
+      multiPoolScope?.startsWith("freebuff::")
+    ) {
+      return {
+        source: "pool",
+        proxyPoolId: null,
+        proxyPool: null,
+        noFitPool: true,
+        connectionProxyEnabled: false,
+        connectionProxyUrl: "",
+        connectionNoProxy: "",
+        strictProxy: true,
+      };
+    }
 
     /**
      * -----------------------------

src/sse/services/auth.js

@@ -8,6 +8,18 @@ import * as log from "../utils/logger.js";
 // Mutex to prevent race conditions during account selection
 let selectionMutex = Promise.resolve();
 
+export function filterConnectionsForModel(providerId, connections, model, settings = {}) {
+  const override = (settings.providerStrategies || {})[providerId] || {};
+  if (override.strictModelAssignment !== true || !model) {
+    return connections;
+  }
+  return connections.filter((connection) => {
+    const assignedModel = connection.providerSpecificData?.assignedModel
+      || (providerId === "freebuff" ? connection.providerSpecificData?.freebuffModel : null);
+    return assignedModel === model;
+  });
+}
+
 const GITHUB_MONTHLY_USAGE_LIMIT = "you've reached your additional usage limit for your plan";
 
 function githubMonthlyResetMs(status, errorText, provider) {
@@ -81,7 +93,10 @@ export async function getProviderCredentials(provider, excludeConnectionIds = nu
       };
     }
 
-    const connections = await getProviderConnections({ provider: providerId, isActive: true });
+    let connections = await getProviderConnections({ provider: providerId, isActive: true });
+    const settings = await getSettings();
+    const providerOverride = (settings.providerStrategies || {})[providerId] || {};
+    connections = filterConnectionsForModel(providerId, connections, model, settings);
     log.debug("AUTH", `${provider} | total connections: ${connections.length}, excludeIds: ${excludeSet.size > 0 ? [...excludeSet].join(",") : "none"}, model: ${model || "any"}`);
 
     if (connections.length === 0) {
@@ -126,9 +141,7 @@ export async function getProviderCredentials(provider, excludeConnectionIds = nu
       return null;
     }
 
-    const settings = await getSettings();
     // Per-provider strategy overrides global setting
-    const providerOverride = (settings.providerStrategies || {})[providerId] || {};
     const strategy = providerOverride.fallbackStrategy || settings.fallbackStrategy || "fill-first";
 
     let connection;
@@ -211,6 +224,7 @@ export async function getProviderCredentials(provider, excludeConnectionIds = nu
         connectionProxyPoolId: resolvedProxy.proxyPoolId || null,
         vercelRelayUrl: resolvedProxy.vercelRelayUrl || "",
         proxyPoolId: resolvedProxy.proxyPoolId || null,
+        noFitPool: resolvedProxy.noFitPool === true,
         strictProxy: resolvedProxy.strictProxy === true,
       },
       connectionId: connection.id,

tests/unit/freebuff-model-assignment.test.js

@@ -0,0 +1,36 @@
+import { describe, expect, it } from "vitest";
+import { filterConnectionsForModel } from "../../src/sse/services/auth.js";
+
+const connections = [
+  { id: "flash-1", providerSpecificData: { freebuffModel: "deepseek/deepseek-v4-flash" } },
+  { id: "mimo-1", providerSpecificData: { freebuffModel: "mimo/mimo-v2.5" } },
+  { id: "unassigned", providerSpecificData: {} },
+];
+
+describe("Freebuff strict model assignment", () => {
+  it("keeps only accounts assigned to the requested model", () => {
+    const result = filterConnectionsForModel("freebuff", connections, "mimo/mimo-v2.5", {
+      providerStrategies: { freebuff: { strictModelAssignment: true } },
+    });
+
+    expect(result.map((connection) => connection.id)).toEqual(["mimo-1"]);
+  });
+
+  it("excludes unassigned accounts when strict mode is enabled", () => {
+    const result = filterConnectionsForModel("freebuff", connections, "deepseek/deepseek-v4-flash", {
+      providerStrategies: { freebuff: { strictModelAssignment: true } },
+    });
+
+    expect(result.map((connection) => connection.id)).toEqual(["flash-1"]);
+  });
+
+  it("preserves the existing pool when strict mode is disabled", () => {
+    expect(filterConnectionsForModel("freebuff", connections, "mimo/mimo-v2.5", {})).toBe(connections);
+  });
+
+  it("does not affect other providers when their toggle is off", () => {
+    expect(filterConnectionsForModel("codex", connections, "mimo/mimo-v2.5", {
+      providerStrategies: {},
+    })).toBe(connections);
+  });
+});

tests/unit/freebuff-provider.test.js

@@ -207,6 +207,43 @@ describe("freebuff executor wire shape", () => {
     const ex = new FreebuffExecutor();
     expect(ex.buildUrl()).toBe("https://www.codebuff.com/api/v1/chat/completions");
   });
+
+  it("injects the end_turn tool into any tool-calling request (backend foreign_toolset gate)", () => {
+    const ex = new FreebuffExecutor();
+    const body = {
+      model: "deepseek/deepseek-v4-flash",
+      messages: [{ role: "user", content: "hi" }],
+      tools: [{ type: "function", function: { name: "read_file", description: "read" } }],
+    };
+    const out = ex.transformRequest(body.model, body, true, { providerSpecificData: { fingerprintId: "fp-1" } });
+    const names = out.tools.map((t) => t.function.name);
+    expect(names).toContain("read_file");
+    expect(names).toContain("end_turn");
+    expect(out.tools[out.tools.length - 1].function).toMatchObject({
+      name: "end_turn",
+      description: "Signal the end of the current task.",
+    });
+  });
+
+  it("does not inject end_turn when the request has no tools", () => {
+    const ex = new FreebuffExecutor();
+    const body = { model: "deepseek/deepseek-v4-flash", messages: [{ role: "user", content: "hi" }] };
+    const out = ex.transformRequest(body.model, body, true, { providerSpecificData: { fingerprintId: "fp-1" } });
+    expect(out.tools).toBeUndefined();
+  });
+
+  it("does not duplicate end_turn when the caller already declared it", () => {
+    const ex = new FreebuffExecutor();
+    const endTurn = { type: "function", function: { name: "end_turn", description: "Signal the end of the current task.", parameters: { type: "object", properties: {} } } };
+    const body = {
+      model: "deepseek/deepseek-v4-flash",
+      messages: [{ role: "user", content: "hi" }],
+      tools: [endTurn],
+    };
+    const out = ex.transformRequest(body.model, body, true, { providerSpecificData: { fingerprintId: "fp-1" } });
+    expect(out.tools).toHaveLength(1);
+    expect(out.tools[0].function.name).toBe("end_turn");
+  });
 });
 
 describe("freebuff session pre-flight", () => {
@@ -301,11 +338,11 @@ describe("freebuff free-tier system marker", () => {
 
 describe("freebuff run registration", () => {
   it("maps freebuff models to their root free agent ids", () => {
-    expect(rootAgentIdForModel("deepseek/deepseek-v4-flash")).toBe("base2-free-deepseek-flash");
-    expect(rootAgentIdForModel("deepseek/deepseek-v4-pro")).toBe("base2-free-deepseek");
-    expect(rootAgentIdForModel("mimo/mimo-v2.5")).toBe("base2-free-mimo");
-    expect(rootAgentIdForModel("minimax/minimax-m3")).toBe("base2-free-minimax-m3");
-    expect(rootAgentIdForModel("openai/gpt-5.6-luna")).toBe("base2-free-luna");
+    expect(rootAgentIdForModel("deepseek/deepseek-v4-flash")).toBe("base3-free-deepseek-flash");
+    expect(rootAgentIdForModel("deepseek/deepseek-v4-pro")).toBe("base3-free-deepseek");
+    expect(rootAgentIdForModel("mimo/mimo-v2.5")).toBe("base3-free-mimo");
+    expect(rootAgentIdForModel("minimax/minimax-m3")).toBe("base3-free-minimax-m3");
+    expect(rootAgentIdForModel("openai/gpt-5.6-luna")).toBe("base3-free-luna");
     expect(rootAgentIdForModel("some/unknown-model")).toBe("base2-free");
   });
 
@@ -320,7 +357,7 @@ describe("freebuff run registration", () => {
     expect(opts.headers.Authorization).toBe("Bearer tok-1");
     const payload = JSON.parse(opts.body);
     expect(payload.action).toBe("START");
-    expect(payload.agentId).toBe("base2-free-deepseek-flash");
+    expect(payload.agentId).toBe("base3-free-deepseek-flash");
     expect(payload.ancestorRunIds).toEqual([]);
   });
 
@@ -586,3 +623,28 @@ describe("freebuff executor execute", () => {
     expect(JSON.parse(finishes[0][1].body).status).toBe("cancelled");
   });
 });
+
+describe("freebuff executor parseError", () => {
+  it("explains a 404 'No endpoints found' as the toolset gate, not a credential problem", async () => {
+    const ex = new FreebuffExecutor();
+    const res = jsonResponse(
+      { error: { message: "No endpoints found for deepseek/deepseek-v4-flash.", code: 404, type: null, param: null } },
+      { status: 404, ok: false },
+    );
+    const parsed = await ex.parseError(res, JSON.stringify({ error: { message: "No endpoints found for deepseek/deepseek-v4-flash.", code: 404 } }));
+
+    expect(parsed.status).toBe(404);
+    expect(parsed.message).toMatch(/end_turn/i);
+    expect(parsed.message).not.toMatch(/credential/i);
+    expect(parsed.resetsAtMs).toBeGreaterThan(Date.now());
+  });
+
+  it("passes other statuses through untouched", async () => {
+    const ex = new FreebuffExecutor();
+    const res = jsonResponse({ error: "bad" }, { status: 500, ok: false });
+    const parsed = await ex.parseError(res, JSON.stringify({ error: "bad" }));
+    expect(parsed.status).toBe(500);
+    expect(parsed.message).toContain("bad");
+    expect(parsed.resetsAtMs).toBeUndefined();
+  });
+});

tests/unit/grok-cli-oauth-probe.test.js

@@ -48,3 +48,31 @@ describe("classifyOAuthProbeResult (grok-cli)", () => {
     expect(r).toEqual({ valid: true, error: null, soft: false });
   });
 });
+
+describe("classifyOAuthProbeResult (freebuff)", () => {
+  const FREEBUFF_PROBE = {
+    acceptStatuses: [403, 404],
+    softFailMessage: { 403: "gated" },
+  };
+
+  it("treats 404 (no session row) as silent success", () => {
+    const r = classifyOAuthProbeResult({ ok: false, status: 404 }, FREEBUFF_PROBE, "");
+    expect(r).toEqual({ valid: true, error: null, soft: false });
+  });
+
+  it("treats 403 (region/account gate) as soft success", () => {
+    const r = classifyOAuthProbeResult(
+      { ok: false, status: 403 },
+      FREEBUFF_PROBE,
+      JSON.stringify({ status: "country_blocked", countryCode: "XX" }),
+    );
+    expect(r.valid).toBe(true);
+    expect(r.soft).toBe(true);
+    expect(r.error).toMatch(/gated/);
+  });
+
+  it("treats 401 as hard auth failure", () => {
+    const r = classifyOAuthProbeResult({ ok: false, status: 401 }, FREEBUFF_PROBE, "unauthorized");
+    expect(r).toEqual({ valid: false, error: "Token invalid or revoked", soft: false });
+  });
+});

tests/unit/proxy-pool-fitness.test.js

@@ -9,6 +9,7 @@ import {
   pruneExpired,
   resetPoolFitness,
 } from "open-sse/services/proxyPoolFitness.js";
+import { pickProxyPoolId } from "../../src/lib/network/connectionProxy.js";
 
 describe("proxy pool fitness registry", () => {
   beforeEach(() => resetPoolFitness());
@@ -39,6 +40,30 @@ describe("proxy pool fitness registry", () => {
     expect(snap.p1["fb::m2"]).toBeUndefined();
   });
 
+  it("does not reuse a pool when every smart candidate is unfit", () => {
+    markPoolUnfit("p1", "freebuff::openai/gpt-5.6-luna", Date.now() + 60_000, "limited_ip");
+    markPoolUnfit("p2", "freebuff::openai/gpt-5.6-luna", Date.now() + 60_000, "limited_ip");
+
+    expect(pickProxyPoolId(
+      ["p1", "p2"],
+      "smart",
+      "freebuff",
+      { scope: "freebuff::openai/gpt-5.6-luna" },
+    )).toBeNull();
+  });
+
+  it("preserves fail-open smart fallback for non-Freebuff providers", () => {
+    markPoolUnfit("p1", "opencode::sonnet-4.6", Date.now() + 60_000, "ip-limit");
+    markPoolUnfit("p2", "opencode::sonnet-4.6", Date.now() + 60_000, "ip-limit");
+
+    expect(pickProxyPoolId(
+      ["p1", "p2"],
+      "smart",
+      "opencode",
+      { scope: "opencode::sonnet-4.6" },
+    )).toBe("p1");
+  });
+
   it("clear per scope, clear-all per provider, clear-all global, pruneExpired", () => {
     markPoolUnfit("p1", "freebuff::m1", Date.now() + 60_000);
     markPoolUnfit("p1", "kiro::m3", Date.now() + 60_000);