n9router 0.4.49 → 0.4.51

package/app/open-sse/providers/registry/windsurf.js

@@ -0,0 +1,143 @@
+// Windsurf provider registry — Firebase+Codeium+Devin auth chain.
+// Chat = Codeium gRPC-web protobuf:
+//   POST {base}  Content-Type: application/grpc-web+proto
+//   Service: exa.language_server_pb.LanguageServerService / GetChatMessage
+export default {
+  id: "windsurf",
+  alias: "ws",
+  uiAlias: "ws",
+  display: {
+    name: "Windsurf",
+    icon: "surfing",
+    color: "#14B8A6",
+    website: "https://windsurf.com",
+    notice: { signupUrl: "https://windsurf.com" },
+  },
+  category: "oauth",
+  authType: "oauth",
+  hasOAuth: true,
+  authModes: ["oauth", "apikey"],
+
+  transport: {
+    baseUrl: "https://server.codeium.com/exa.language_server_pb.LanguageServerService/GetChatMessage",
+    format: "openai",
+    headers: {
+      "Content-Type": "application/grpc-web+proto",
+      "Accept": "application/grpc-web+proto",
+      "X-Grpc-Web": "1",
+    },
+    // apiKey (sk-ws-... or Firebase-derived) as Bearer + in protobuf Metadata.api_key.
+    auth: { combined: true, header: "Authorization", scheme: "Bearer" },
+  },
+
+  // Auth chain (4 terminal paths, all yield apiKey):
+  //  1) OAuth web → Firebase JWT → POST register.windsurf.com/.../RegisterUser {firebase_id_token} → {apiKey, apiServerUrl, name}
+  //  2) sk-ws-... direct API key (apiKey used as metadata.apiKey on GetUserStatus)
+  //  3) Firebase JWT (eyJ...) → same RegisterUser exchange as #1
+  //  4) Devin auth1_... → self-serve chain → ide_token used as apiKey on server.self-serve.windsurf.com
+  oauth: {
+    clientId: "3GUryQ7ldAeKEuD2obYnppsnmj58eP5u",
+    firebaseApiKey: "AIzaSyDsOl-1XpT5err0Tcn0TFFod1H8gVGIycY",
+    firebaseSignInUrl: "https://identitytoolkit.googleapis.com/v1/accounts:signInWithPassword",
+    registerUrl: "https://register.windsurf.com/exa.seat_management_pb.SeatManagementService/RegisterUser",
+    apiServerUrl: "https://server.codeium.com",
+    auth1ApiServerUrl: "https://server.self-serve.windsurf.com",
+    platform: "windsurf",
+    // Quota (Connect RPC, protobuf): POST windsurf.com/_backend/.../GetPlanStatus,
+    // headers Content-Type:application/proto + Connect-Protocol-Version:1 + X-Auth-Token:<session>,
+    // body = field1:session_token, field2:varint 1.
+    quotaUrl: "https://windsurf.com/_backend/exa.seat_management_pb.SeatManagementService/GetPlanStatus",
+  },
+
+  // Catalog verified against model_configs_v2.bin from Devin CLI (2026.5.x).
+  // Dot-notation ids; the executor MODEL_ALIAS_MAP maps these to Windsurf modelUid.
+  // contextLength dropped — 9router schema uses id+name only.
+  models: [
+    // Cognition / SWE
+    { id: "swe-1.6-fast", name: "SWE-1.6 Fast" },
+    { id: "swe-1.6", name: "SWE-1.6" },
+    { id: "swe-1.5-fast", name: "SWE-1.5 Fast" },
+    { id: "swe-1.5", name: "SWE-1.5" },
+    // Claude Opus 4.7 — effort-tiered
+    { id: "claude-opus-4.7-max", name: "Claude Opus 4.7 Max" },
+    { id: "claude-opus-4.7-xhigh", name: "Claude Opus 4.7 XHigh" },
+    { id: "claude-opus-4.7-high", name: "Claude Opus 4.7 High" },
+    { id: "claude-opus-4.7-medium", name: "Claude Opus 4.7 Medium" },
+    { id: "claude-opus-4.7-low", name: "Claude Opus 4.7 Low" },
+    { id: "claude-opus-4.7-review", name: "Claude Opus 4.7 Review" },
+    // Claude Sonnet/Opus 4.6
+    { id: "claude-sonnet-4.6-thinking-1m", name: "Claude Sonnet 4.6 Thinking 1M" },
+    { id: "claude-sonnet-4.6-1m", name: "Claude Sonnet 4.6 1M" },
+    { id: "claude-sonnet-4.6-thinking", name: "Claude Sonnet 4.6 Thinking" },
+    { id: "claude-sonnet-4.6", name: "Claude Sonnet 4.6" },
+    { id: "claude-opus-4.6-thinking", name: "Claude Opus 4.6 Thinking" },
+    { id: "claude-opus-4.6", name: "Claude Opus 4.6" },
+    // Claude 4.5
+    { id: "claude-opus-4.5-thinking", name: "Claude Opus 4.5 Thinking" },
+    { id: "claude-opus-4.5", name: "Claude Opus 4.5" },
+    { id: "claude-sonnet-4.5-thinking", name: "Claude Sonnet 4.5 Thinking" },
+    { id: "claude-sonnet-4.5", name: "Claude Sonnet 4.5" },
+    { id: "claude-haiku-4.5", name: "Claude Haiku 4.5" },
+    // GPT-5.5 — effort-tiered
+    { id: "gpt-5.5-xhigh-fast", name: "GPT-5.5 XHigh Fast" },
+    { id: "gpt-5.5-xhigh", name: "GPT-5.5 XHigh" },
+    { id: "gpt-5.5-high-fast", name: "GPT-5.5 High Fast" },
+    { id: "gpt-5.5-high", name: "GPT-5.5 High" },
+    { id: "gpt-5.5-medium-fast", name: "GPT-5.5 Medium Fast" },
+    { id: "gpt-5.5-medium", name: "GPT-5.5 Medium" },
+    { id: "gpt-5.5-low-fast", name: "GPT-5.5 Low Fast" },
+    { id: "gpt-5.5-low", name: "GPT-5.5 Low" },
+    { id: "gpt-5.5-none-fast", name: "GPT-5.5 None Fast" },
+    { id: "gpt-5.5-none", name: "GPT-5.5 None" },
+    // GPT-5.4 — effort-tiered
+    { id: "gpt-5.4-xhigh-fast", name: "GPT-5.4 XHigh Fast" },
+    { id: "gpt-5.4-xhigh", name: "GPT-5.4 XHigh" },
+    { id: "gpt-5.4-high-fast", name: "GPT-5.4 High Fast" },
+    { id: "gpt-5.4-high", name: "GPT-5.4 High" },
+    { id: "gpt-5.4-medium-fast", name: "GPT-5.4 Medium Fast" },
+    { id: "gpt-5.4-medium", name: "GPT-5.4 Medium" },
+    { id: "gpt-5.4-low-fast", name: "GPT-5.4 Low Fast" },
+    { id: "gpt-5.4-low", name: "GPT-5.4 Low" },
+    { id: "gpt-5.4-none-fast", name: "GPT-5.4 None Fast" },
+    { id: "gpt-5.4-none", name: "GPT-5.4 None" },
+    { id: "gpt-5.4-mini-xhigh", name: "GPT-5.4 Mini XHigh" },
+    { id: "gpt-5.4-mini-high", name: "GPT-5.4 Mini High" },
+    { id: "gpt-5.4-mini-medium", name: "GPT-5.4 Mini Medium" },
+    { id: "gpt-5.4-mini-low", name: "GPT-5.4 Mini Low" },
+    // GPT-5.3 Codex
+    { id: "gpt-5.3-codex-xhigh-fast", name: "GPT-5.3 Codex XHigh Fast" },
+    { id: "gpt-5.3-codex-xhigh", name: "GPT-5.3 Codex XHigh" },
+    { id: "gpt-5.3-codex-high-fast", name: "GPT-5.3 Codex High Fast" },
+    { id: "gpt-5.3-codex-high", name: "GPT-5.3 Codex High" },
+    { id: "gpt-5.3-codex-medium-fast", name: "GPT-5.3 Codex Medium Fast" },
+    { id: "gpt-5.3-codex-medium", name: "GPT-5.3 Codex Medium" },
+    { id: "gpt-5.3-codex-low-fast", name: "GPT-5.3 Codex Low Fast" },
+    { id: "gpt-5.3-codex-low", name: "GPT-5.3 Codex Low" },
+    // GPT-5.2 / 5
+    { id: "gpt-5.2-xhigh", name: "GPT-5.2 XHigh" },
+    { id: "gpt-5.2-high", name: "GPT-5.2 High" },
+    { id: "gpt-5.2-medium", name: "GPT-5.2 Medium" },
+    { id: "gpt-5.2-low", name: "GPT-5.2 Low" },
+    { id: "gpt-5.2-none", name: "GPT-5.2 None" },
+    { id: "gpt-5", name: "GPT-5" },
+    // GPT-4.1 / 4o
+    { id: "gpt-4.1", name: "GPT-4.1" },
+    { id: "gpt-4.1-mini", name: "GPT-4.1 Mini" },
+    { id: "gpt-4.1-nano", name: "GPT-4.1 Nano" },
+    { id: "gpt-4o", name: "GPT-4o" },
+    { id: "gpt-4o-mini", name: "GPT-4o Mini" },
+    // Gemini
+    { id: "gemini-3.1-pro-high", name: "Gemini 3.1 Pro High" },
+    { id: "gemini-3.1-pro-low", name: "Gemini 3.1 Pro Low" },
+    { id: "gemini-3.0-flash-high", name: "Gemini 3 Flash High" },
+    { id: "gemini-3.0-flash-medium", name: "Gemini 3 Flash Medium" },
+    { id: "gemini-3.0-flash-low", name: "Gemini 3 Flash Low" },
+    { id: "gemini-3.0-flash-minimal", name: "Gemini 3 Flash Minimal" },
+    { id: "gemini-2.5-pro", name: "Gemini 2.5 Pro" },
+    // Others
+    { id: "deepseek-v4", name: "DeepSeek V4" },
+    { id: "kimi-k2.6", name: "Kimi K2.6" },
+    { id: "kimi-k2.5", name: "Kimi K2.5" },
+    { id: "glm-5.1", name: "GLM-5.1" },
+  ],
+};

package/app/open-sse/providers/registry/xai.js

@@ -32,9 +32,13 @@ export default {
     { id: "grok-code-fast-1", name: "Grok Code Fast" },
     { id: "grok-3", name: "Grok 3" },
     { id: "grok-2-image-1212", name: "Grok 2 Image", params: ["n","response_format"], kind: "image" },
+    { id: "grok-imagine-video", name: "Grok Imagine Video", params: ["duration","aspect_ratio","resolution"], kind: "video" },
   ],
-  serviceKinds: ["llm","imageToText","webSearch","image"],
+  serviceKinds: ["llm","imageToText","webSearch","image","video"],
   imageConfig: { baseUrl: "https://api.x.ai/v1/images/generations", bodyFields: ["model","prompt","n","response_format"] },
+  // Async video jobs (POST returns { request_id }, GET polls until done/failed).
+  // Docs: https://docs.x.ai/developers/rest-api-reference/inference/videos
+  videoConfig: { baseUrl: "https://api.x.ai/v1/videos" },
   searchViaChat: {
     defaultModel: "grok-4.20-reasoning",
     endpoint: "https://api.x.ai/v1/responses",

package/app/open-sse/providers/registry/xiaomi-mimo.js

@@ -15,10 +15,11 @@ export default {
     textIcon: "XM",
     website: "https://xiaomimimo.com",
     notice: {
-      apiKeyUrl: "https://xiaomimimo.com",
+      apiKeyUrl: "https://platform.xiaomimimo.com/console/api-keys",
     },
   },
   category: "apikey",
+  serviceKinds: ["llm", "tts"],
   transport: {
     baseUrl: "https://api.xiaomimimo.com/v1/chat/completions",
     validateUrl: "https://api.xiaomimimo.com/v1/models",
@@ -42,5 +43,12 @@ export default {
     { id: "mimo-v2.5", name: "MiMo V2.5" },
     { id: "mimo-v2-omni", name: "MiMo V2 Omni" },
     { id: "mimo-v2-flash", name: "MiMo V2 Flash" },
+    { id: "mimo-v2.5-tts", name: "MiMo V2.5 TTS", kind: "tts" },
   ],
+  ttsConfig: {
+    baseUrl: "https://api.xiaomimimo.com/v1/chat/completions",
+    authType: "apikey",
+    authHeader: "bearer",
+    format: "xiaomi-mimo-tts",
+  },
 };

package/app/open-sse/providers/registry/zed.js

@@ -0,0 +1,71 @@
+// Zed provider — RSA keypair callback auth (NOT standard OAuth).
+export default {
+  id: "zed",
+  priority: 10,
+  alias: "zd",
+  uiAlias: "zd",
+  hidden: true,
+  display: {
+    name: "Zed",
+    icon: "code",
+    color: "#A855F7",
+    website: "https://zed.dev",
+    notice: {
+      signupUrl: "https://zed.dev/native_app_signin",
+    },
+  },
+  category: "oauth",
+  authType: "oauth",
+  hasOAuth: true,
+
+  transport: {
+    // Zed hosted LLM aggregator: cloud.zed.dev/completions is a
+    // multi-format proxy fronting Anthropic/OpenAI/Google/xAI depending on the model.
+    // Wire protocol = NDJSON/SSE-ish stream authenticated with a short-lived LLM bearer
+    // token exchanged from the RSA-decrypted access_token (see open-sse/shared/zedAuth).
+    baseUrl: "https://cloud.zed.dev/completions",
+    format: "openai",
+    forceStream: true,
+    headers: {
+      "content-type": "application/json",
+    },
+    // Auth scheme is non-standard: "Authorization: <user_id> <access_token>" plus a duplicate
+    // x-zed-cloud-token header (verified in zed_account.rs build_authorization_header +
+    // cloud fetch). Executor builds both; scheme here is a marker for config-driven tooling.
+    auth: {
+      combined: true,
+      header: "Authorization",
+      scheme: "<user_id> <access_token>", // placeholder — real value built in executor
+    },
+    usage: {
+      url: "https://cloud.zed.dev/client/users/me", // verified in zed_account.rs
+    },
+    // Live catalog discovery — Zed's hosted model list changes frequently and is fetched
+    // per-connection rather than hardcoded.
+    modelsUrl: "https://cloud.zed.dev/models",
+  },
+
+  // Empty static catalog + passthrough: Zed fronts a rotating set of upstream models
+  // (Claude/GPT/Gemini/Grok). Resolved live via modelsUrl; any client-sent model id is
+  // forwarded as-is rather than validated against a frozen list.
+  models: [],
+  passthroughModels: true,
+
+  oauth: {
+    // Zed auth flow is RSA-based, NOT OAuth2/PKCE:
+    //   1. App generates RSA-2048 keypair locally (PKCS#1 DER, URL-safe base64).
+    //   2. Bind random TCP port on 127.0.0.1.
+    //   3. Open https://zed.dev/native_app_signin?native_app_port={port}&native_app_public_key={pub}.
+    //   4. After login, browser redirects http://127.0.0.1:{port}/?user_id=...&access_token=...
+    //      where access_token = base64(RSA-encrypted plaintext token).
+    //   5. Decrypt with private key (OAEP-SHA256, fallback PKCS1v15). Store user_id + plaintext token.
+    // No clientId/clientSecret/tokenUrl/refreshUrl — long-lived access_token, no refresh.
+    authorizeUrl: "https://zed.dev/native_app_signin",
+    platform: "zed",
+    rsaKeyExchange: true, // new flag: signals frontend/router this flow needs local RSA + TCP listener.
+  },
+
+  features: {
+    usage: true,
+  },
+};

package/app/open-sse/providers/shared.js

@@ -47,6 +47,26 @@ export const CLAUDE_CLI_SPOOF_HEADERS = {
   "X-Stainless-Timeout": "600"
 };
 
+const ANTHROPIC_BETA_BASE = [
+  "claude-code-20250219",
+  "oauth-2025-04-20",
+  "interleaved-thinking-2025-05-14",
+  "context-management-2025-06-27",
+  "prompt-caching-scope-2026-01-05",
+  "structured-outputs-2025-12-15",
+  "fast-mode-2026-02-01",
+  "redact-thinking-2026-02-12",
+  "token-efficient-tools-2026-03-28",
+];
+const ANTHROPIC_BETA_HEAVY_AGENT = ["advanced-tool-use-2025-11-20", "effort-2025-11-24"];
+
+// Heavy-agent beta flags are gated to opus/sonnet — cheaper models don't need them.
+export function selectAnthropicBeta(model = "") {
+  const flags = [...ANTHROPIC_BETA_BASE];
+  if (/^claude-(opus|sonnet)/.test(model)) flags.push(...ANTHROPIC_BETA_HEAVY_AGENT);
+  return flags.join(",");
+}
+
 // Shared baseUrls
 export const KIMI_CODING_BASE_URL = "https://api.kimi.com/coding/v1/messages";
 
@@ -58,7 +78,7 @@ export const ANTHROPIC_COMPAT_BASE = "https://api.anthropic.com/v1";
 // Keep this static even when 9router runs on Linux: the provider profile is
 // intentionally matching the IDE client, not the server host.
 export const ANTIGRAVITY_IDE_VERSION = "2.1.1";
-export const ANTIGRAVITY_IDE_BASE_URL = "https://cloudcode-pa.googleapis.com";
+export const ANTIGRAVITY_IDE_BASE_URL = "https://daily-cloudcode-pa.googleapis.com";
 export const ANTIGRAVITY_IDE_USER_AGENT = `antigravity/ide/${ANTIGRAVITY_IDE_VERSION} darwin/arm64`;
 
 // Antigravity OAuth client credentials (public CLI client — duplicated in usage.js + src/lib/oauth)

package/app/open-sse/providers/thinkingLevels.js

@@ -2,6 +2,7 @@
 // Reuses capabilities.js (thinkingFormat/canDisable) so this file only maps format→levels (DRY).
 import { getCapabilitiesForModel } from "./capabilities.js";
 import { matchPattern } from "./pricing.js";
+import { resolveKiroEffortPath } from "../config/kiroConstants.js";
 
 // Shared level sets (deduped) — verified against provider docs + wire in thinkingUnified.applyFormat.
 const L = {
@@ -30,18 +31,24 @@ const FORMAT_LEVELS = {
   step: L.base,
 };
 
+const CODEX_GPT_5_6_LEVELS = ["none", "minimal", "low", "medium", "high", "xhigh", "max"];
+
 // Model-name pattern overrides (glob, first match wins) — more precise than format default.
 const PATTERN_THINKING = [
-  // gpt-5.6-sol accepts max (maps to xhigh on wire); live probe rejected ultra.
-  { pattern: "*gpt-5.6-sol*", levels: ["none", "minimal", "low", "medium", "high", "xhigh", "max"] },
+  { provider: "codex", pattern: "*gpt-5.6-sol*", levels: [...CODEX_GPT_5_6_LEVELS, "ultra"] },
+  { provider: "codex", pattern: "*gpt-5.6-terra*", levels: [...CODEX_GPT_5_6_LEVELS, "ultra"] },
+  { provider: "codex", pattern: "*gpt-5.6-luna*", levels: CODEX_GPT_5_6_LEVELS },
   { pattern: "*codex*", levels: ["low", "medium", "high", "xhigh"] }, // codex cannot disable thinking
 ];
 
 // Returns valid thinking levels for a model, or null when the model has no reasoning.
 export function getThinkingLevels(provider, model) {
+  if (provider === "kiro" && resolveKiroEffortPath(model) === null) return null;
   const caps = getCapabilitiesForModel(provider, model);
   if (!caps.reasoning) return null;
-  const hit = PATTERN_THINKING.find((p) => matchPattern(p.pattern, model));
+  const hit = PATTERN_THINKING.find((entry) =>
+    (!entry.provider || entry.provider === provider) && matchPattern(entry.pattern, model)
+  );
   let levels = hit?.levels || FORMAT_LEVELS[caps.thinkingFormat] || L.base;
   if (caps.thinkingCanDisable === false) levels = levels.filter((l) => l !== "none");
   return levels;

package/app/open-sse/rtk/headroom.js

@@ -25,9 +25,17 @@ function messagePayload(body) {
 
 function captureSizeSnapshot(body) {
   const messages = messagePayload(body);
+  const toolHistory = messages?.filter((message) =>
+    message?.role === "tool"
+    || message?.role === "function"
+    || message?.tool_calls?.length
+    || message?.content?.some?.((part) => part?.type === "tool_use" || part?.type === "tool_result")
+  ) || [];
   return {
     bodyBytes: jsonBytes(body),
     messageBytes: messages ? jsonBytes(messages) : 0,
+    toolSchemaBytes: jsonBytes(body?.tools || []),
+    toolHistoryBytes: jsonBytes(toolHistory),
   };
 }
 
@@ -336,7 +344,10 @@ export function formatHeadroomSizeLog(diagnostics) {
   const before = diagnostics?.before;
   const after = diagnostics?.after;
   if (!before || !after) return "";
-  return `body=${before.bodyBytes}B→${after.bodyBytes}B messages=${before.messageBytes}B→${after.messageBytes}B`;
+  const effective = before.bodyBytes > 0
+    ? (((before.bodyBytes - after.bodyBytes) / before.bodyBytes) * 100).toFixed(1)
+    : "0.0";
+  return `body=${before.bodyBytes}B→${after.bodyBytes}B messages=${before.messageBytes}B→${after.messageBytes}B tools=${before.toolSchemaBytes || 0}B→${after.toolSchemaBytes || 0}B toolHistory=${before.toolHistoryBytes || 0}B→${after.toolHistoryBytes || 0}B effective=${effective}%`;
 }
 
 export function isHeadroomPhantomSavings(stats, diagnostics, minShrinkRatio = 0.05) {

package/app/open-sse/services/capacityAdapter.js

@@ -0,0 +1,173 @@
+/**
+ * Capacity Adapter — global fallback pools of models per input-modality capability
+ * (vision / pdf / audioInput / videoInput).
+ *
+ * The pool models are appended as extra fallback candidates behind whatever models
+ * were already going to be tried (a combo's members, or a single target model).
+ * combo.js's existing reorderByCapabilities then floats a capable pool model to the
+ * front only when none of the original models can handle the request — so this
+ * never overrides a combo that already has a member covering the capability.
+ */
+import { getCapabilitiesForModel } from "../providers/capabilities.js";
+
+const CAPABILITY_KEYS = ["vision", "pdf", "audioInput", "videoInput"];
+const HARD_CAPS = new Set(CAPABILITY_KEYS);
+const DEFAULT_FALLBACK_MODEL = "oc/mimo-v2.5-free";
+
+// Normalize a capability entry to { enabled, roundRobin, models }. Backward-compat:
+// accept the legacy array form [{model, enabled}] (treated as enabled, fallback).
+function normalizeCapEntry(entry) {
+  if (Array.isArray(entry)) {
+    return { enabled: true, roundRobin: false, models: entry.map((e) => e?.model || e).filter(Boolean) };
+  }
+  if (entry && typeof entry === "object") {
+    return {
+      enabled: entry.enabled !== false,
+      roundRobin: !!entry.roundRobin,
+      models: Array.isArray(entry.models) ? entry.models.filter(Boolean) : [],
+    };
+  }
+  return { enabled: false, roundRobin: false, models: [] };
+}
+
+// Resolve one capability's full config. Enabled pools with no models fall back
+// to DEFAULT_FALLBACK_MODEL so the toggle is never a no-op.
+export function getCapacityAdapterConfig(cap, settings) {
+  const entry = normalizeCapEntry(settings?.capacityAdapter?.[cap]);
+  if (entry.enabled && entry.models.length === 0) {
+    return { ...entry, models: [DEFAULT_FALLBACK_MODEL] };
+  }
+  return entry;
+}
+
+// Flatten enabled models across all capability pools, in priority order, deduped.
+export function getCapacityAdapterModels(settings) {
+  const seen = new Set();
+  const models = [];
+  for (const cap of CAPABILITY_KEYS) {
+    const { enabled, models: pool } = getCapacityAdapterConfig(cap, settings);
+    if (!enabled) continue;
+    for (const m of pool) {
+      if (!seen.has(m)) {
+        seen.add(m);
+        models.push(m);
+      }
+    }
+  }
+  return models;
+}
+
+// Strategy for a capability: "round-robin" when enabled+roundRobin, else "fallback".
+export function getCapacityAdapterStrategy(cap, settings) {
+  const { enabled, roundRobin } = getCapacityAdapterConfig(cap, settings);
+  return enabled && roundRobin ? "round-robin" : "fallback";
+}
+
+// Strategy from the request's required capabilities: picks the first capability
+// whose adapter pool is enabled and can satisfy a hard requirement.
+export function getActiveAdapterStrategy(requiredCapabilities, settings) {
+  const hard = [...(requiredCapabilities || [])].filter((c) => HARD_CAPS.has(c));
+  for (const cap of hard) {
+    const { enabled, models } = getCapacityAdapterConfig(cap, settings);
+    if (!enabled || models.length === 0) continue;
+    return getCapacityAdapterStrategy(cap, settings);
+  }
+  return "fallback";
+}
+
+function modelSatisfies(modelStr, requiredHard) {
+  const slash = modelStr.indexOf("/");
+  const provider = slash > 0 ? modelStr.slice(0, slash) : "";
+  const model = slash > 0 ? modelStr.slice(slash + 1) : modelStr;
+  const caps = getCapabilitiesForModel(provider, model);
+  return requiredHard.every((c) => caps[c] === true);
+}
+
+// Prepend capacity-adapter models as priority candidates when NONE of the
+// original models (combo members, or the single target model) can satisfy the
+// request's required capabilities. Adapter models go FIRST (priority); the
+// original models follow as fallback. Leaves `models` untouched when the
+// original list already covers it (combo.js's reorderByCapabilities handles
+// that case via autoSwitch).
+export function augmentModelsWithCapacityAdapter(models, requiredCapabilities, settings) {
+  const hard = [...(requiredCapabilities || [])].filter((c) => HARD_CAPS.has(c));
+  if (hard.length === 0 || !Array.isArray(models) || models.length === 0) return models;
+  if (models.some((m) => modelSatisfies(m, hard))) return models;
+
+  const pool = getCapacityAdapterModels(settings).filter((m) => !models.includes(m) && modelSatisfies(m, hard));
+  if (pool.length === 0) return models;
+  return [...pool, ...models];
+}
+
+const CHARS_PER_TOKEN = 4; // rough estimate; avoids pulling in a tokenizer dependency
+const HEAD_KEEP = 6;      // messages after system kept verbatim before dropping the middle
+
+function blockLength(content) {
+  if (typeof content === "string") return content.length;
+  if (Array.isArray(content)) {
+    return content.reduce((sum, b) => sum + (typeof b?.text === "string" ? b.text.length : 50), 0);
+  }
+  return 0;
+}
+
+// Trim history to fit a (possibly smaller) context window by dropping the MIDDLE.
+// Preserves: all system/instruction messages (head), and the trailing user run
+// carrying the media the switch happened for (tail). Older middle turns between
+// the head instructions and the current turn are dropped first.
+export function stripHistoryForContext(body, contextWindow) {
+  const key = Array.isArray(body.messages) ? "messages"
+    : Array.isArray(body.input) ? "input"
+    : Array.isArray(body.contents) ? "contents"
+    : null;
+  if (!key) return body;
+  const arr = body[key];
+  if (!arr || arr.length === 0) return body;
+
+  const isSystem = (r) => r === "system" || r === "developer";
+  const systemMsgs = arr.filter((m) => isSystem(m?.role));
+  const rest = arr.filter((m) => !isSystem(m?.role));
+  if (rest.length === 0) return body;
+
+  const isAssistant = (r) => r === "assistant" || r === "model";
+  let i = rest.length - 1;
+  while (i >= 0 && !isAssistant(rest[i]?.role)) i--;
+  const tail = rest.slice(i + 1);          // current user turn (has media) — always kept
+  const older = rest.slice(0, i + 1);      // everything before it
+  if (older.length === 0) return body;
+
+  const contentOf = (m) => m.content ?? m.parts;
+  // Cap at 80% of the adapter model's context window — leaves room for the response.
+  const budgetChars = (contextWindow || 200000) * 0.8 * CHARS_PER_TOKEN;
+
+  // Prefer keeping the first HEAD_KEEP messages (initial instructions/context) verbatim;
+  // only trim further if even that exceeds the adapter model's context window.
+  const headKept = older.slice(0, HEAD_KEEP);
+  let total = systemMsgs.concat(headKept, tail).reduce((s, m) => s + blockLength(contentOf(m)), 0);
+
+  // If head + tail overflow, drop head turns from the end (closest to middle) first.
+  let head = headKept;
+  while (total > budgetChars && head.length > 0) {
+    const dropped = head.pop();
+    total -= blockLength(contentOf(dropped));
+  }
+
+  if (head.length === older.length) return body;
+  return { ...body, [key]: [...systemMsgs, ...head, ...tail] };
+}
+
+// Wrap a handleSingleModel callback so calls to a capacity-adapter model strip
+// history to fit its context window first. No-op passthrough when the pool is empty.
+export function withCapacityAdapterStripping(handleSingleModel, adapterModels) {
+  const adapterSet = new Set(adapterModels);
+  if (adapterSet.size === 0) return handleSingleModel;
+  return (body, modelStr, ...rest) => {
+    if (adapterSet.has(modelStr)) {
+      const slash = modelStr.indexOf("/");
+      const provider = slash > 0 ? modelStr.slice(0, slash) : "";
+      const model = slash > 0 ? modelStr.slice(slash + 1) : modelStr;
+      const { contextWindow } = getCapabilitiesForModel(provider, model);
+      body = stripHistoryForContext(body, contextWindow);
+    }
+    return handleSingleModel(body, modelStr, ...rest);
+  };
+}

package/app/open-sse/services/combo.js

@@ -106,15 +106,32 @@ export function detectRequiredCapabilities(body) {
   const required = new Set();
   if (!body || typeof body !== "object") return required;
 
+  const addByMime = (mime) => {
+    if (typeof mime !== "string") return;
+    if (mime.startsWith("image/")) required.add("vision");
+    else if (mime === "application/pdf") required.add("pdf");
+    else if (mime.startsWith("audio/")) required.add("audioInput");
+    else if (mime.startsWith("video/")) required.add("videoInput");
+  };
+
   const scanBlock = (b) => {
     if (!b || typeof b !== "object") return;
     const t = b.type;
     if (t === "image_url" || t === "image" || t === "input_image") required.add("vision");
-    if (t === "file" || t === "document" || t === "input_file") required.add("pdf");
+    if (t === "input_audio" || t === "audio_url" || t === "audio") required.add("audioInput");
+    if (t === "input_video" || t === "video_url" || t === "video") required.add("videoInput");
+    if (t === "file" || t === "document" || t === "input_file") {
+      // Infer modality from embedded mime when available; fall back to pdf for generic files.
+      let fmime = null;
+      if (b.input_audio?.format) fmime = `audio/${b.input_audio.format}`;
+      else if (b.file?.file_data) fmime = String(b.file.file_data).match(/^data:([^;,]+)/)?.[1];
+      else if (b.source?.media_type) fmime = b.source.media_type;
+      else if (b.source?.data) fmime = String(b.source.data).match(/^data:([^;,]+)/)?.[1];
+      if (fmime) addByMime(fmime);
+      else required.add("pdf");
+    }
     // gemini parts: inlineData/fileData carry a mime
-    const mime = b.inlineData?.mimeType || b.fileData?.mimeType;
-    if (typeof mime === "string" && mime.startsWith("image/")) required.add("vision");
-    if (mime === "application/pdf") required.add("pdf");
+    addByMime(b.inlineData?.mimeType || b.fileData?.mimeType);
   };
 
   const scanContent = (content) => {