Refactor

cli/cli.js

@@ -61,6 +61,21 @@ const INSTALL_CMD_LATEST = `npm i -g ${APP_NAME}@latest --prefer-online`;
 
 const DEFAULT_PORT = 20128;
 const DEFAULT_HOST = "0.0.0.0";
+
+// First non-internal IPv4 — the address remote peers actually reach when bound to 0.0.0.0.
+function getLanIp() {
+  for (const ifaces of Object.values(os.networkInterfaces())) {
+    for (const i of ifaces || []) {
+      if (i.family === "IPv4" && !i.internal) return i.address;
+    }
+  }
+  return null;
+}
+
+// Local URL stays "localhost"; warn separately when bound to all interfaces (network-exposed).
+function getDisplayHost() {
+  return host === DEFAULT_HOST ? "localhost" : host;
+}
 const MAX_PORT_ATTEMPTS = 10;
 // Identifiers for killAllAppProcesses - only kill 9router specifically
 const PROCESS_IDENTIFIERS = [
@@ -501,7 +516,7 @@ async function showInterfaceMenu(latestVersion) {
 
   clearScreen();
 
-  const displayHost = host === DEFAULT_HOST ? "localhost" : host;
+  const displayHost = getDisplayHost();
 
   // Detect tunnel/local mode for server URL display
   let serverUrl;
@@ -542,8 +557,13 @@ const MAX_RESTARTS = 2;
 const RESTART_RESET_MS = 30000; // Reset counter if alive > 30s
 
 function startServer(latestVersion) {
-  const displayHost = host === DEFAULT_HOST ? "localhost" : host;
+  const displayHost = getDisplayHost();
   const url = `http://${displayHost}:${port}/dashboard`;
+  // Surface real network exposure when bound to all interfaces (default 0.0.0.0).
+  if (host === DEFAULT_HOST) {
+    const lanIp = getLanIp();
+    if (lanIp) console.log(`\x1b[33m⚠ Network-exposed: reachable at http://${lanIp}:${port} (bound 0.0.0.0). Use --host 127.0.0.1 for local-only.\x1b[0m`);
+  }
 
   let restartCount = 0;
   let serverStartTime = Date.now();

open-sse/config/mediaConfig.js

@@ -0,0 +1,27 @@
+// Central config for remote-media fetching security limits.
+
+// Max bytes accepted from a remote image fetch (reject larger to prevent memory DoS).
+export const MAX_IMAGE_BYTES = 10 * 1024 * 1024; // 10MB
+
+// Fetch timeout for remote media.
+export const FETCH_TIMEOUT_MS = 10000;
+
+// Magic-byte signatures -> mime. Each entry: { sig:[bytes], offset, mime }.
+// offset>0 for containers where the signature is not at byte 0 (e.g. webp).
+export const IMAGE_SIGNATURES = [
+  { sig: [0x89, 0x50, 0x4e, 0x47], offset: 0, mime: "image/png" },
+  { sig: [0xff, 0xd8, 0xff], offset: 0, mime: "image/jpeg" },
+  { sig: [0x47, 0x49, 0x46, 0x38], offset: 0, mime: "image/gif" },
+  { sig: [0x52, 0x49, 0x46, 0x46], offset: 0, mime: "image/webp", verifyWebp: true },
+  { sig: [0x42, 0x4d], offset: 0, mime: "image/bmp" },
+];
+
+// Hostnames/IPs that must never be fetched (SSRF guard for loopback + cloud metadata).
+export const BLOCKED_HOSTS = new Set([
+  "localhost",
+  "127.0.0.1",
+  "0.0.0.0",
+  "::1",
+  "169.254.169.254", // AWS/GCP/Azure IMDS
+  "metadata.google.internal",
+]);

open-sse/handlers/chatCore.js

@@ -21,6 +21,9 @@ import { detectClientTool, isNativePassthrough } from "../utils/clientDetector.j
 import { dedupeTools } from "../utils/toolDeduper.js";
 import { injectCaveman } from "../rtk/caveman.js";
 import { compressMessages, formatRtkLog } from "../rtk/index.js";
+import { getCapabilitiesForModel } from "../providers/capabilities.js";
+import { stripUnsupportedModalities } from "../translator/concerns/modality.js";
+import { prefetchRemoteImages } from "../translator/concerns/prefetch.js";
 
 /**
  * Core chat handler - shared between SSE and Worker
@@ -91,6 +94,19 @@ export async function handleChatCore({ body, modelInfo, credentials, log, onCred
   // Expose raw client headers to translators/executors for session-id resolution
   if (credentials) credentials.rawHeaders = clientRawRequest?.headers || {};
 
+  // Auto-strip media blocks the model can't read (vision/audio/pdf) before translation.
+  if (!passthrough) {
+    const caps = getCapabilitiesForModel(provider, model);
+    if (stripUnsupportedModalities(body, sourceFormat, caps)) {
+      log?.debug?.("MODALITY", `stripped unsupported media for ${provider}/${model}`);
+    }
+    // Convert remote image URLs to base64 for targets that can't fetch URLs.
+    try {
+      const n = await prefetchRemoteImages(body, sourceFormat, targetFormat, { signal: undefined });
+      if (n > 0) log?.debug?.("MODALITY", `prefetched ${n} remote image(s) for ${targetFormat}`);
+    } catch (e) { log?.warn?.("MODALITY", `image prefetch failed: ${e.message}`); }
+  }
+
   let translatedBody;
   let toolNameMap;
   if (passthrough) {

open-sse/providers/capabilities.js

@@ -41,6 +41,11 @@ export const DEFAULT_CAPABILITIES = {
   search: false,        // built-in web search tool / grounding
   tools: true,          // function / tool calling
   reasoning: false,     // thinking / reasoning
+  // thinking wire format (only meaningful when reasoning:true). null → derive from transport.format.
+  // enum: openai|claude-adaptive|claude-budget|gemini-level|gemini-budget|zai|qwen|deepseek|kimi|minimax|hunyuan|step
+  thinkingFormat: null,
+  thinkingCanDisable: true,  // false → model cannot turn thinking off (clamp to min instead of disable)
+  thinkingRange: null,       // { min, max } for budget formats; null = no clamp
   // limits (tokens)
   contextWindow: 200000,
   maxOutput: 64000,
@@ -51,22 +56,22 @@ export const DEFAULT_CAPABILITIES = {
  * otherwise mis-match. Only declare deltas vs DEFAULT.
  */
 export const MODEL_CAPABILITIES = {
-  // Claude 4.6/4.7 have 1M context (override generic claude pattern at 200k)
-  "claude-opus-4.6":   { vision: true, reasoning: true, search: true, contextWindow: 1000000, maxOutput: 128000 },
-  "claude-opus-4.7":   { vision: true, reasoning: true, search: true, contextWindow: 1000000, maxOutput: 128000 },
-  "claude-opus-4-6":   { vision: true, reasoning: true, search: true, contextWindow: 1000000, maxOutput: 128000 },
-  "claude-sonnet-4.6": { vision: true, reasoning: true, search: true, contextWindow: 1000000, maxOutput: 64000 },
-  "claude-sonnet-4-6": { vision: true, reasoning: true, search: true, contextWindow: 1000000, maxOutput: 64000 },
+  // Claude 4.6/4.7 have 1M context + adaptive thinking (override generic claude pattern)
+  "claude-opus-4.6":   { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive", contextWindow: 1000000, maxOutput: 128000 },
+  "claude-opus-4.7":   { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive", contextWindow: 1000000, maxOutput: 128000 },
+  "claude-opus-4-6":   { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive", contextWindow: 1000000, maxOutput: 128000 },
+  "claude-sonnet-4.6": { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive", contextWindow: 1000000, maxOutput: 64000 },
+  "claude-sonnet-4-6": { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive", contextWindow: 1000000, maxOutput: 64000 },
 
   // Gemini image-gen / OpenAI image / xai image variants
   "gpt-image-1":       { imageOutput: true, tools: false },
 
   // GLM vision variant (text GLM has no vision)
-  "glm-4.6v":          { vision: true, reasoning: true, contextWindow: 128000 },
+  "glm-4.6v":          { vision: true, reasoning: true, thinkingFormat: "zai", contextWindow: 128000 },
 
   // Qwen plain coder/text (no vision) — registry "vision-model" / "coder-model" aliases
-  "vision-model":      { vision: true, reasoning: true, contextWindow: 1000000 },
-  "coder-model":       { reasoning: true, contextWindow: 1000000 },
+  "vision-model":      { vision: true, reasoning: true, thinkingFormat: "qwen", contextWindow: 1000000 },
+  "coder-model":       { reasoning: true, thinkingFormat: "qwen", contextWindow: 1000000 },
 };
 
 /**
@@ -81,19 +86,25 @@ export const PROVIDER_CAPABILITIES = {};
  * a broad family pattern swallowing an exception (e.g. glm-4.6v vs glm-5).
  */
 export const PATTERN_CAPABILITIES = [
-  // ── Claude (4.x+ = vision + thinking + web search) ───────────────
-  { pattern: "*claude*opus*",   caps: { vision: true, reasoning: true, search: true } },
-  { pattern: "*claude*sonnet*", caps: { vision: true, reasoning: true, search: true } },
-  { pattern: "*claude*haiku*",  caps: { vision: true, reasoning: true, search: true } },
-  { pattern: "*claude*fable*",  caps: { vision: true, reasoning: true, search: true, contextWindow: 1000000, maxOutput: 128000 } },
-  { pattern: "*claude*mythos*", caps: { vision: true, reasoning: true, search: true, contextWindow: 1000000, maxOutput: 128000 } },
+  // ── Claude (4.6+ = adaptive thinking; older/haiku = budget) ──────
+  { pattern: "*claude*opus-4.6*",   caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive" } },
+  { pattern: "*claude*opus-4.7*",   caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive" } },
+  { pattern: "*claude*opus-4.8*",   caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive" } },
+  { pattern: "*claude*sonnet-4.6*", caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive" } },
+  { pattern: "*claude*sonnet-4.7*", caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-adaptive" } },
+  { pattern: "*claude*haiku*",  caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-budget" } },
+  { pattern: "*claude*opus*",   caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-budget" } },
+  { pattern: "*claude*sonnet*", caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-budget" } },
+  { pattern: "*claude*fable*",  caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-budget", contextWindow: 1000000, maxOutput: 128000 } },
+  { pattern: "*claude*mythos*", caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-budget", contextWindow: 1000000, maxOutput: 128000 } },
   { pattern: "*claude-3*",      caps: { vision: true } },
-  { pattern: "*claude*",        caps: { vision: true, reasoning: true, search: true } },
+  { pattern: "*claude*",        caps: { vision: true, reasoning: true, search: true, thinkingFormat: "claude-budget" } },
 
   // ── Gemini (all 2.0+ multimodal + google_search grounding, 1M ctx) ─
   { pattern: "*gemini*image*",  caps: { vision: true, imageOutput: true, contextWindow: 1048576 } },
-  { pattern: "*gemini-3*pro*",  caps: { vision: true, audioInput: true, videoInput: true, reasoning: true, search: true, contextWindow: 1048576, maxOutput: 65535 } },
-  { pattern: "*gemini-3*",      caps: { vision: true, audioInput: true, videoInput: true, search: true, contextWindow: 1048576, maxOutput: 65536 } },
+  { pattern: "*gemini-3*pro*",  caps: { vision: true, audioInput: true, videoInput: true, reasoning: true, search: true, thinkingFormat: "gemini-level", thinkingCanDisable: false, contextWindow: 1048576, maxOutput: 65535 } },
+  { pattern: "*gemini-3*",      caps: { vision: true, audioInput: true, videoInput: true, reasoning: true, search: true, thinkingFormat: "gemini-level", thinkingCanDisable: false, contextWindow: 1048576, maxOutput: 65536 } },
+  { pattern: "*gemini-2.5*",    caps: { vision: true, audioInput: true, videoInput: true, reasoning: true, search: true, thinkingFormat: "gemini-budget", thinkingRange: { min: 0, max: 24576 }, contextWindow: 1048576, maxOutput: 65536 } },
   { pattern: "*gemini-2*",      caps: { vision: true, audioInput: true, videoInput: true, search: true, contextWindow: 1048576, maxOutput: 65536 } },
   { pattern: "*gemini*",        caps: { vision: true, search: true, contextWindow: 1048576 } },
   { pattern: "*gemma*",         caps: { vision: true, contextWindow: 128000 } },
@@ -101,58 +112,59 @@ export const PATTERN_CAPABILITIES = [
 
   // ── OpenAI GPT-5.x (vision + thinking + web search) ──────────────
   { pattern: "*gpt-5*image*",   caps: { imageOutput: true } },
-  { pattern: "*gpt-5*codex*",   caps: { reasoning: true, search: true, contextWindow: 400000, maxOutput: 128000 } },
-  { pattern: "*gpt-5*",         caps: { vision: true, reasoning: true, search: true, contextWindow: 400000, maxOutput: 128000 } },
+  { pattern: "*gpt-5*codex*",   caps: { reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 400000, maxOutput: 128000 } },
+  { pattern: "*gpt-5*",         caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 400000, maxOutput: 128000 } },
   { pattern: "*gpt-4o*",        caps: { vision: true, search: true, contextWindow: 128000, maxOutput: 16384 } },
   { pattern: "*gpt-4.1*",       caps: { vision: true, contextWindow: 1000000, maxOutput: 32768 } },
   { pattern: "*gpt-4-turbo*",   caps: { vision: true, contextWindow: 128000 } },
   { pattern: "*gpt-4*",         caps: { contextWindow: 128000 } },
   { pattern: "*gpt-3.5*",       caps: { contextWindow: 16385, maxOutput: 4096 } },
-  { pattern: "*gpt-oss*",       caps: { reasoning: true, contextWindow: 128000 } },
+  { pattern: "*gpt-oss*",       caps: { reasoning: true, thinkingFormat: "openai", contextWindow: 128000 } },
 
   // ── OpenAI o-series (reasoning, vision) ──────────────────────────
-  { pattern: "*o1-mini*",       caps: { reasoning: true, contextWindow: 128000 } },
-  { pattern: "*o1*",            caps: { vision: true, reasoning: true, contextWindow: 200000, maxOutput: 100000 } },
-  { pattern: "*o3*",            caps: { vision: true, reasoning: true, contextWindow: 200000, maxOutput: 100000 } },
-  { pattern: "*o4*",            caps: { vision: true, reasoning: true, contextWindow: 200000, maxOutput: 100000 } },
+  { pattern: "*o1-mini*",       caps: { reasoning: true, thinkingFormat: "openai", contextWindow: 128000 } },
+  { pattern: "*o1*",            caps: { vision: true, reasoning: true, thinkingFormat: "openai", contextWindow: 200000, maxOutput: 100000 } },
+  { pattern: "*o3*",            caps: { vision: true, reasoning: true, thinkingFormat: "openai", contextWindow: 200000, maxOutput: 100000 } },
+  { pattern: "*o4*",            caps: { vision: true, reasoning: true, thinkingFormat: "openai", contextWindow: 200000, maxOutput: 100000 } },
 
   // ── Grok (vision + Live Search) ──────────────────────────────────
   { pattern: "*grok*image*",    caps: { imageOutput: true } },
-  { pattern: "*grok-code*",     caps: { reasoning: true, contextWindow: 256000 } },
-  { pattern: "*grok-4*",        caps: { vision: true, reasoning: true, search: true, contextWindow: 256000 } },
-  { pattern: "*grok-3*",        caps: { vision: true, reasoning: true, search: true, contextWindow: 131072 } },
-  { pattern: "*grok*",          caps: { vision: true, reasoning: true, search: true, contextWindow: 256000 } },
+  { pattern: "*grok-code*",     caps: { reasoning: true, thinkingFormat: "openai", contextWindow: 256000 } },
+  { pattern: "*grok-4*",        caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 256000 } },
+  { pattern: "*grok-3*",        caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 131072 } },
+  { pattern: "*grok*",          caps: { vision: true, reasoning: true, search: true, thinkingFormat: "openai", contextWindow: 256000 } },
 
-  // ── Qwen (VL = vision; max/plus = vision+1M; coder/text last) ─────
-  { pattern: "*qwen*vl*",       caps: { vision: true, reasoning: true, contextWindow: 262144 } },
-  { pattern: "*qwen*max*",      caps: { vision: true, reasoning: true, contextWindow: 1000000, maxOutput: 65536 } },
-  { pattern: "*qwen*plus*",     caps: { vision: true, reasoning: true, contextWindow: 1000000, maxOutput: 65536 } },
-  { pattern: "*qwen*235b*",     caps: { reasoning: true, contextWindow: 262144 } },
-  { pattern: "*qwen*coder*",    caps: { reasoning: true, contextWindow: 1000000 } },
-  { pattern: "*qwq*",           caps: { reasoning: true, contextWindow: 131072 } },
-  { pattern: "*qwen*",          caps: { reasoning: true, contextWindow: 262144 } },
+  // ── Qwen (enable_thinking + thinking_budget; QwQ = thinking-only) ─
+  { pattern: "*qwen*vl*",       caps: { vision: true, reasoning: true, thinkingFormat: "qwen", contextWindow: 262144 } },
+  { pattern: "*qwen*max*",      caps: { vision: true, reasoning: true, thinkingFormat: "qwen", contextWindow: 1000000, maxOutput: 65536 } },
+  { pattern: "*qwen*plus*",     caps: { vision: true, reasoning: true, thinkingFormat: "qwen", contextWindow: 1000000, maxOutput: 65536 } },
+  { pattern: "*qwen*235b*",     caps: { reasoning: true, thinkingFormat: "qwen", contextWindow: 262144 } },
+  { pattern: "*qwen*coder*",    caps: { reasoning: true, thinkingFormat: "qwen", contextWindow: 1000000 } },
+  { pattern: "*qwq*",           caps: { reasoning: true, thinkingFormat: "qwen", thinkingCanDisable: false, contextWindow: 131072 } },
+  { pattern: "*qwen*",          caps: { reasoning: true, thinkingFormat: "qwen", contextWindow: 262144 } },
 
-  // ── Kimi (K2.x = vision + thinking, 262K) ────────────────────────
-  { pattern: "*kimi*k2*",       caps: { vision: true, reasoning: true, contextWindow: 262144, maxOutput: 262144 } },
-  { pattern: "*kimi*",          caps: { reasoning: true, contextWindow: 262144 } },
+  // ── Kimi (enabled→reasoning_effort; K2.7-code cannot disable) ─────
+  { pattern: "*kimi*k2.7*code*", caps: { vision: true, reasoning: true, thinkingFormat: "kimi", thinkingCanDisable: false, contextWindow: 262144, maxOutput: 262144 } },
+  { pattern: "*kimi*k2*",       caps: { vision: true, reasoning: true, thinkingFormat: "kimi", contextWindow: 262144, maxOutput: 262144 } },
+  { pattern: "*kimi*",          caps: { reasoning: true, thinkingFormat: "kimi", contextWindow: 262144 } },
 
-  // ── GLM (4.6V vision handled by exact id; text GLM = reasoning) ───
-  { pattern: "*glm-5*",         caps: { reasoning: true, contextWindow: 200000, maxOutput: 128000 } },
-  { pattern: "*glm-4.7*",       caps: { reasoning: true, contextWindow: 200000, maxOutput: 128000 } },
-  { pattern: "*glm-4*",         caps: { reasoning: true, contextWindow: 200000 } },
-  { pattern: "*glm*",           caps: { reasoning: true, contextWindow: 200000 } },
+  // ── GLM / Z.ai (thinking.enabled; disable via enable_thinking:false) ─
+  { pattern: "*glm-5*",         caps: { reasoning: true, thinkingFormat: "zai", contextWindow: 200000, maxOutput: 128000 } },
+  { pattern: "*glm-4.7*",       caps: { reasoning: true, thinkingFormat: "zai", contextWindow: 200000, maxOutput: 128000 } },
+  { pattern: "*glm-4*",         caps: { reasoning: true, thinkingFormat: "zai", contextWindow: 200000 } },
+  { pattern: "*glm*",           caps: { reasoning: true, thinkingFormat: "zai", contextWindow: 200000 } },
 
-  // ── DeepSeek (NO vision; v4 = 1M ctx; r1/reasoner = thinking) ─────
-  { pattern: "*deepseek-v4*",   caps: { reasoning: true, contextWindow: 1000000, maxOutput: 384000 } },
-  { pattern: "*reasoner*",      caps: { reasoning: true, contextWindow: 128000 } },
-  { pattern: "*deepseek-r*",    caps: { reasoning: true, contextWindow: 128000 } },
+  // ── DeepSeek (thinking.enabled + reasoning_effort; r1 = thinking-only) ─
+  { pattern: "*deepseek-v4*",   caps: { reasoning: true, thinkingFormat: "deepseek", contextWindow: 1000000, maxOutput: 384000 } },
+  { pattern: "*reasoner*",      caps: { reasoning: true, thinkingFormat: "deepseek", thinkingCanDisable: false, contextWindow: 128000 } },
+  { pattern: "*deepseek-r*",    caps: { reasoning: true, thinkingFormat: "deepseek", thinkingCanDisable: false, contextWindow: 128000 } },
   { pattern: "*deepseek*",      caps: { contextWindow: 128000 } },
 
-  // ── MiniMax (M3 = 1M/512K; M2.x = 200K) ──────────────────────────
+  // ── MiniMax (M3 = adaptive; M2.x cannot disable) ─────────────────
   { pattern: "*minimax*image*", caps: { imageOutput: true } },
-  { pattern: "*minimax-m3*",    caps: { reasoning: true, contextWindow: 1048576, maxOutput: 512000 } },
-  { pattern: "*minimax-m2.7*",  caps: { reasoning: true, contextWindow: 204800, maxOutput: 131072 } },
-  { pattern: "*minimax*",       caps: { reasoning: true, contextWindow: 200000, maxOutput: 131072 } },
+  { pattern: "*minimax-m3*",    caps: { reasoning: true, thinkingFormat: "minimax", contextWindow: 1048576, maxOutput: 512000 } },
+  { pattern: "*minimax-m2.7*",  caps: { reasoning: true, thinkingFormat: "minimax", thinkingCanDisable: false, contextWindow: 204800, maxOutput: 131072 } },
+  { pattern: "*minimax*",       caps: { reasoning: true, thinkingFormat: "minimax", thinkingCanDisable: false, contextWindow: 200000, maxOutput: 131072 } },
 
   // ── Xiaomi MiMo (vision, 1M / 262K ctx) ──────────────────────────
   { pattern: "*mimo*v2.5*",     caps: { vision: true, contextWindow: 1048576, maxOutput: 131072 } },
@@ -178,9 +190,9 @@ export const PATTERN_CAPABILITIES = [
   { pattern: "*perplexity*",    caps: { search: true, contextWindow: 128000 } },
 
   // ── Others ───────────────────────────────────────────────────────
-  { pattern: "*hunyuan*",       caps: { reasoning: true, contextWindow: 262144, maxOutput: 262144 } },
-  { pattern: "hy3*",            caps: { reasoning: true, contextWindow: 262144, maxOutput: 262144 } },
-  { pattern: "*step-*",         caps: { reasoning: true, contextWindow: 128000 } },
+  { pattern: "*hunyuan*",       caps: { reasoning: true, thinkingFormat: "hunyuan", contextWindow: 262144, maxOutput: 262144 } },
+  { pattern: "hy3*",            caps: { reasoning: true, thinkingFormat: "hunyuan", contextWindow: 262144, maxOutput: 262144 } },
+  { pattern: "*step-*",         caps: { reasoning: true, thinkingFormat: "step", contextWindow: 128000 } },
   { pattern: "*nemotron*",      caps: { reasoning: true, contextWindow: 128000 } },
   { pattern: "*ling-*",         caps: { reasoning: true, contextWindow: 128000 } },
 ];

open-sse/providers/registry/blackbox.js

@@ -19,6 +19,7 @@ export default {
   category: "apikey",
   transport: {
     baseUrl: "https://api.blackbox.ai/chat/completions",
+    thinkingFormat: "openai",
   },
   models: [
     { id: "gpt-4o", name: "GPT-4o" },

open-sse/providers/registry/cloudflare-ai.js

@@ -22,6 +22,7 @@ export default {
   hasProviderSpecificData: true,
   transport: {
     baseUrl: "https://api.cloudflare.com/client/v4/accounts/{accountId}/ai/v1/chat/completions",
+    thinkingFormat: "openai",
   },
   models: [
     { id: "@cf/meta/llama-3.2-1b-instruct", name: "Llama 3.2 1B Instruct" },

open-sse/providers/registry/codex.js

@@ -40,6 +40,7 @@ export default {
     },
     usage: {
       url: "https://chatgpt.com/backend-api/wham/usage",
+      resetCreditsConsumeUrl: "https://chatgpt.com/backend-api/wham/rate-limit-reset-credits/consume",
     },
   },
   models: [

open-sse/providers/registry/gemini.js

@@ -40,8 +40,6 @@ export default {
     { id: "gemini-2.5-pro", name: "Gemini 2.5 Pro" },
     { id: "gemini-2.5-flash", name: "Gemini 2.5 Flash" },
     { id: "gemini-2.5-flash-lite", name: "Gemini 2.5 Flash Lite" },
-    { id: "gemini-2.0-flash", name: "Gemini 2.0 Flash" },
-    { id: "gemini-2.0-flash-lite", name: "Gemini 2.0 Flash Lite" },
     { id: "gemma-4-31b-it", name: "Gemma 4 31B IT" },
     { id: "gemini-embedding-2-preview", name: "Gemini Embedding 2 Preview", kind: "embedding" },
     { id: "gemini-embedding-001", name: "Gemini Embedding 001", kind: "embedding" },

open-sse/providers/registry/iflow.js

@@ -15,6 +15,7 @@ export default {
   category: "oauth",
   transport: {
     baseUrl: "https://apis.iflow.cn/v1/chat/completions",
+    thinkingFormat: "openai",
     headers: {
       "User-Agent": "iFlow-Cli",
     },

open-sse/providers/registry/openrouter.js

@@ -17,6 +17,7 @@ export default {
   category: "freeTier",
   transport: {
     baseUrl: "https://openrouter.ai/api/v1/chat/completions",
+    thinkingFormat: "openai",
     headers: {
       "HTTP-Referer": "https://endpoint-proxy.local",
       "X-Title": "Endpoint Proxy",

open-sse/providers/registry/siliconflow.js

@@ -16,6 +16,7 @@ export default {
   transport: {
     baseUrl: "https://api.siliconflow.com/v1/chat/completions",
     validateUrl: "https://api.siliconflow.com/v1/models",
+    thinkingFormat: "openai",
   },
   models: [
     { id: "deepseek-ai/DeepSeek-V4-Pro", name: "DeepSeek V4 Pro" },

open-sse/providers/registry/vercel-ai-gateway.js

@@ -20,6 +20,7 @@ export default {
   category: "apikey",
   transport: {
     baseUrl: "https://ai-gateway.vercel.sh/v1/chat/completions",
+    thinkingFormat: "openai",
     retry: {
       "429": 2,
     },

open-sse/rtk/constants.js

@@ -19,8 +19,12 @@ export const STATUS_MAX_UNTRACKED = 10;        // config::limits().status_max_un
 export const LS_EXT_SUMMARY_TOP = 5;           // top-N extensions in summary
 export const LS_NOISE_DIRS = [
   "node_modules", ".git", "target", "__pycache__",
-  ".next", "dist", "build", ".venv", "venv",
-  ".cache", ".idea", ".vscode", ".DS_Store"
+  ".next", "dist", "build", ".cache", ".turbo",
+  ".vercel", ".pytest_cache", ".mypy_cache", ".tox",
+  ".venv", "venv",
+  "env", // Python legacy virtualenv; .env (dotenv) intentionally excluded
+  "coverage", ".nyc_output", ".DS_Store", "Thumbs.db",
+  ".idea", ".vscode", ".vs", "*.egg-info", ".eggs"
 ];
 
 // tree filter_tree_output cap (no rust cap, we add one to be safe)

open-sse/rtk/filters/find.js

@@ -31,16 +31,15 @@ export function find(input) {
   const showDirs = dirs.slice(0, FIND_TOTAL_DIR_MAX);
   for (const dir of showDirs) {
     const files = byDir.get(dir);
-    out += `${dir}/ (${files.length}):\n`;
+    out += `${dir}/  (${files.length})\n`;
     const showFiles = files.slice(0, FIND_PER_DIR_MAX);
     for (const f of showFiles) out += `  ${f}\n`;
     if (files.length > FIND_PER_DIR_MAX) {
       out += `  +${files.length - FIND_PER_DIR_MAX}\n`;
     }
-    out += "\n";
   }
   if (dirs.length > FIND_TOTAL_DIR_MAX) {
-    out += `+${dirs.length - FIND_TOTAL_DIR_MAX} more dirs\n`;
+    out += `\n+${dirs.length - FIND_TOTAL_DIR_MAX} more dirs\n`;
   }
 
   return out;

open-sse/services/combo.js

@@ -4,6 +4,34 @@
 
 import { checkFallbackError, formatRetryAfter } from "./accountFallback.js";
 import { unavailableResponse } from "../utils/error.js";
+import { getCapabilitiesForModel } from "../providers/capabilities.js";
+
+// Hard capabilities = input modalities; missing one drops request data (e.g. image
+// stripped). Must be prioritized. Soft (e.g. search) only degrades a feature.
+const HARD_CAPS = new Set(["vision", "pdf", "audioInput", "videoInput"]);
+
+// Reorder combo models by capability fit. Stable; never drops a model (fallback intact).
+// Tier 0: satisfies all hard + all soft. Tier 1: all hard only. Tier 2: rest.
+export function reorderByCapabilities(models, required) {
+  if (!required || required.size === 0 || !Array.isArray(models) || models.length <= 1) return models;
+  const hard = [...required].filter((c) => HARD_CAPS.has(c));
+  const soft = [...required].filter((c) => !HARD_CAPS.has(c));
+
+  const tierOf = (m) => {
+    const slash = typeof m === "string" ? m.indexOf("/") : -1;
+    const provider = slash > 0 ? m.slice(0, slash) : "";
+    const model = slash > 0 ? m.slice(slash + 1) : m;
+    const caps = getCapabilitiesForModel(provider, model);
+    if (!hard.every((c) => caps[c] === true)) return 2;
+    return soft.every((c) => caps[c] === true) ? 0 : 1;
+  };
+
+  // Stable sort by tier (Array.prototype.sort is stable in modern engines).
+  return models
+    .map((m, i) => ({ m, i, t: tierOf(m) }))
+    .sort((a, b) => a.t - b.t || a.i - b.i)
+    .map((x) => x.m);
+}
 
 /**
  * Track rotation state per combo (for round-robin strategy)
@@ -11,6 +39,53 @@ import { unavailableResponse } from "../utils/error.js";
  */
 const comboRotationState = new Map();
 
+// Last array item whose role is "user" (current turn), or the last item when no
+// role is present. History media (older turns) must not pin the combo to a vision
+// model — those get stripped + placeholdered downstream instead.
+function lastUserItem(arr) {
+  if (!Array.isArray(arr) || arr.length === 0) return null;
+  for (let i = arr.length - 1; i >= 0; i--) {
+    if (!arr[i]?.role || arr[i].role === "user") return arr[i];
+  }
+  return arr[arr.length - 1];
+}
+
+// Detect which capabilities a request needs. Modalities (vision/pdf) are scanned
+// only on the current user turn; "search" is request-wide (lives in tools).
+// Returns a Set of: "vision" | "pdf" | "search".
+export function detectRequiredCapabilities(body) {
+  const required = new Set();
+  if (!body || typeof body !== "object") return required;
+
+  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");
+    // 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");
+  };
+
+  const scanContent = (content) => {
+    if (Array.isArray(content)) for (const b of content) scanBlock(b);
+  };
+
+  // Modalities: current user turn only (last item across each known shape).
+  const lastMsg = lastUserItem(body.messages);     // openai / claude
+  if (lastMsg) scanContent(lastMsg.content);
+  const lastInput = lastUserItem(body.input);      // responses
+  if (lastInput) scanContent(lastInput.content);
+  const contents = body.contents || body.request?.contents; // gemini / antigravity
+  const lastContent = lastUserItem(contents);
+  if (lastContent) scanContent(lastContent.parts);
+
+  // search: temporarily disabled in auto-switch (feature not wired yet).
+
+  return required;
+}
+
 function normalizeStickyLimit(stickyLimit) {
   const parsed = Number.parseInt(stickyLimit, 10);
   return Number.isFinite(parsed) && parsed > 0 ? parsed : 1;
@@ -105,9 +180,21 @@ export function getComboModelsFromData(modelStr, combosData) {
  * @param {number|string} [options.comboStickyLimit=1] - Requests per combo model before switching
  * @returns {Promise<Response>}
  */
-export async function handleComboChat({ body, models, handleSingleModel, log, comboName, comboStrategy, comboStickyLimit = 1 }) {
+export async function handleComboChat({ body, models, handleSingleModel, log, comboName, comboStrategy, comboStickyLimit = 1, autoSwitch = true }) {
   // Apply rotation strategy if enabled
-  const rotatedModels = getRotatedModels(models, comboName, comboStrategy, comboStickyLimit);
+  let rotatedModels = getRotatedModels(models, comboName, comboStrategy, comboStickyLimit);
+
+  // Auto-switch: float models that satisfy the request's required capabilities to the front.
+  if (autoSwitch) {
+    const required = detectRequiredCapabilities(body);
+    if (required.size > 0) {
+      const reordered = reorderByCapabilities(rotatedModels, required);
+      if (reordered[0] !== rotatedModels[0]) {
+        log.info("COMBO", `auto-switch for [${[...required].join(",")}] → ${reordered[0]}`);
+      }
+      rotatedModels = reordered;
+    }
+  }
   
   let lastError = null;
   let earliestRetryAfter = null;

open-sse/services/usage.js

@@ -5,7 +5,9 @@
 import { getGitHubUsage } from "./usage/github.js";
 import { getGeminiUsage, getAntigravityUsage } from "./usage/google.js";
 import { getClaudeUsage } from "./usage/claude.js";
-import { getCodexUsage } from "./usage/codex.js";
+import { getCodexUsage, consumeCodexRateLimitResetCredit } from "./usage/codex.js";
+
+export { consumeCodexRateLimitResetCredit };
 import { getKiroUsage } from "./usage/kiro.js";
 import { getMiniMaxUsage } from "./usage/minimax.js";
 import {

open-sse/services/usage/codex.js

@@ -8,6 +8,7 @@ import { U, parseResetTime, toFiniteNumber } from "./shared.js";
 // Codex (OpenAI) API config
 const CODEX_CONFIG = {
   usageUrl: U("codex").url,
+  resetCreditsConsumeUrl: U("codex").resetCreditsConsumeUrl,
 };
 
 function getCodexRateLimitBody(snapshot) {
@@ -82,6 +83,7 @@ export async function getCodexUsage(accessToken, proxyOptions = null) {
     const data = await response.json();
     const normalRateLimit = data.rate_limit || data.rate_limits || data.rate_limits_by_limit_id?.codex || {};
     const reviewRateLimit = getCodexReviewRateLimit(data);
+    const availableResetCredits = Math.max(0, toFiniteNumber(data.rate_limit_reset_credits?.available_count, 0));
     const quotas = {};
 
     appendCodexQuotaWindows(quotas, "", normalRateLimit);
@@ -91,9 +93,53 @@ export async function getCodexUsage(accessToken, proxyOptions = null) {
       plan: data.plan_type || data.summary?.plan || "unknown",
       limitReached: getCodexRateLimitBody(normalRateLimit)?.limit_reached || false,
       reviewLimitReached: getCodexRateLimitBody(reviewRateLimit)?.limit_reached || false,
+      resetCredits: { availableCount: availableResetCredits },
       quotas,
     };
   } catch (error) {
     throw new Error(`Failed to fetch Codex usage: ${error.message}`);
   }
 }
+
+// Consume one Codex rate-limit reset credit (irreversible, spends 1 credit)
+export async function consumeCodexRateLimitResetCredit(accessToken, redeemRequestId, proxyOptions = null) {
+  if (!accessToken) {
+    throw new Error("No Codex access token available. Please re-authorize the connection.");
+  }
+  if (!redeemRequestId || typeof redeemRequestId !== "string") {
+    throw new Error("A redeem request id is required to consume a Codex reset credit.");
+  }
+
+  let response;
+  let data = null;
+  try {
+    response = await proxyAwareFetch(CODEX_CONFIG.resetCreditsConsumeUrl, {
+      method: "POST",
+      headers: {
+        "Authorization": `Bearer ${accessToken}`,
+        "Accept": "application/json",
+        "Content-Type": "application/json",
+      },
+      body: JSON.stringify({ redeem_request_id: redeemRequestId }),
+    }, proxyOptions);
+
+    const text = await response.text();
+    data = text ? JSON.parse(text) : null;
+  } catch (error) {
+    throw new Error(`Failed to consume Codex reset credit: ${error.message}`);
+  }
+
+  const code = data?.code || null;
+  const windowsReset = toFiniteNumber(data?.windows_reset, 0);
+  const success = response.ok && (code === "reset" || windowsReset > 0);
+
+  return {
+    ok: success,
+    noCredit: response.ok && code === "no_credit",
+    status: response.status,
+    code,
+    windowsReset,
+    message: data?.message || null,
+    raw: data,
+  };
+}

open-sse/translator/concerns/image.js

@@ -12,34 +12,100 @@ export function parseDataUri(url) {
   return m ? { mimeType: m[1], base64: m[2] } : null;
 }
 
+import { lookup } from "node:dns/promises";
+import { MAX_IMAGE_BYTES, FETCH_TIMEOUT_MS, IMAGE_SIGNATURES, BLOCKED_HOSTS } from "../../config/mediaConfig.js";
+
+// True if an IPv4/IPv6 address is private/reserved (SSRF target).
+function isPrivateIp(ip) {
+  if (!ip) return true;
+  // IPv6 loopback / unique-local / link-local
+  if (ip === "::1" || ip.startsWith("fc") || ip.startsWith("fd") || ip.startsWith("fe80")) return true;
+  // IPv4-mapped IPv6 (::ffff:a.b.c.d) -> extract tail
+  const v4 = ip.includes(".") ? ip.split(":").pop() : ip;
+  const parts = v4.split(".").map((n) => Number.parseInt(n, 10));
+  if (parts.length !== 4 || parts.some((n) => Number.isNaN(n))) return ip.includes(":") ? false : true;
+  const [a, b] = parts;
+  if (a === 10 || a === 127 || a === 0) return true;
+  if (a === 172 && b >= 16 && b <= 31) return true;
+  if (a === 192 && b === 168) return true;
+  if (a === 169 && b === 254) return true; // link-local + cloud metadata
+  if (a === 100 && b >= 64 && b <= 127) return true; // CGNAT
+  return false;
+}
+
+// Resolve host and reject if it points at a private/blocked address (SSRF guard).
+async function assertPublicHost(hostname) {
+  if (!hostname || BLOCKED_HOSTS.has(hostname.toLowerCase())) return false;
+  try {
+    const { address } = await lookup(hostname);
+    return !isPrivateIp(address);
+  } catch {
+    return false;
+  }
+}
+
+// Verify buffer magic bytes match a known image signature; return its mime or null.
+function detectImageMime(buf) {
+  for (const { sig, offset, mime, verifyWebp } of IMAGE_SIGNATURES) {
+    if (buf.length < offset + sig.length) continue;
+    let match = true;
+    for (let i = 0; i < sig.length; i++) {
+      if (buf[offset + i] !== sig[i]) { match = false; break; }
+    }
+    if (!match) continue;
+    // WEBP: RIFF....WEBP — bytes 8..11 must be "WEBP".
+    if (verifyWebp && !(buf.length >= 12 && buf[8] === 0x57 && buf[9] === 0x45 && buf[10] === 0x42 && buf[11] === 0x50)) continue;
+    return mime;
+  }
+  return null;
+}
+
 /**
  * Fetch a remote image URL and return it as a base64 data URI.
- * Used when upstream providers (Codex, etc.) require inline base64 images
- * instead of remote URLs they cannot fetch.
- * Returns null if fetch fails.
+ * Hardened against SSRF (private/metadata IPs), memory DoS (size cap),
+ * and disguised non-image payloads (magic-byte verification).
+ * Returns null on any failure or rejection.
  *
  * @param {string} imageUrl - HTTP(S) URL of the image
- * @param {object} options - { signal, timeoutMs }
+ * @param {object} options - { signal, timeoutMs, maxBytes }
  * @returns {Promise<{url: string, mimeType: string}|null>}
  */
 export async function fetchImageAsBase64(imageUrl, options = {}) {
-  const { signal, timeoutMs = 10000 } = options;
+  const { signal, timeoutMs = FETCH_TIMEOUT_MS, maxBytes = MAX_IMAGE_BYTES } = options;
   if (!imageUrl || (!imageUrl.startsWith("http://") && !imageUrl.startsWith("https://"))) {
     return null;
   }
 
+  let url;
+  try { url = new URL(imageUrl); } catch { return null; }
+  if (!(await assertPublicHost(url.hostname))) return null;
+
   const controller = new AbortController();
   const timeout = signal ? null : setTimeout(() => controller.abort(), timeoutMs);
   const fetchSignal = signal || controller.signal;
 
   try {
-    const response = await fetch(imageUrl, { signal: fetchSignal });
-    if (!response.ok) return null;
+    // redirect:"manual" prevents a public URL redirecting to a private one (SSRF bypass).
+    const response = await fetch(imageUrl, { signal: fetchSignal, redirect: "manual" });
+    if (!response.ok || !response.body) return null;
 
-    const mimeType = response.headers.get("Content-Type") || "image/jpeg";
-    const arrayBuffer = await response.arrayBuffer();
-    const base64 = Buffer.from(arrayBuffer).toString("base64");
-    return { url: `data:${mimeType};base64,${base64}`, mimeType };
+    // Stream-read with a hard byte cap to avoid loading huge payloads into memory.
+    const reader = response.body.getReader();
+    const chunks = [];
+    let total = 0;
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      total += value.length;
+      if (total > maxBytes) { try { await reader.cancel(); } catch { /* ignore */ } return null; }
+      chunks.push(value);
+    }
+
+    const buf = Buffer.concat(chunks.map((c) => Buffer.from(c)));
+    const mimeType = detectImageMime(buf);
+    if (!mimeType) return null; // not a recognized image — reject disguised payloads
+
+    return { url: `data:${mimeType};base64,${buf.toString("base64")}`, mimeType };
   } catch {
     return null;
   } finally {

open-sse/translator/concerns/modality.js

@@ -0,0 +1,155 @@
+// Strip multimodal content blocks a model cannot read, BEFORE translation.
+// Driven by getCapabilitiesForModel: vision/audioInput/pdf. Replaces removed
+// media with a short text placeholder so messages never become empty.
+import { FORMATS } from "../formats.js";
+
+// Placeholder text inserted where a media block was removed.
+// Current turn: explain the active model can't read what the user just sent.
+const PLACEHOLDER_CURRENT = {
+  vision: "[image omitted: model has no vision support]",
+  audioInput: "[audio omitted: model has no audio support]",
+  pdf: "[file omitted: model has no document support]",
+};
+// Earlier turns: neutral (a combo may route to a different model each turn).
+const PLACEHOLDER_PREV = {
+  vision: "[Previous image omitted from context.]",
+  audioInput: "[Previous audio omitted from context.]",
+  pdf: "[Previous file omitted from context.]",
+};
+const ph = (cap, isLast) => (isLast ? PLACEHOLDER_CURRENT : PLACEHOLDER_PREV)[cap];
+
+// Map gemini inlineData/fileData mime prefix -> capability it requires.
+function capForMime(mime) {
+  if (typeof mime !== "string") return null;
+  if (mime.startsWith("image/")) return "vision";
+  if (mime.startsWith("audio/")) return "audioInput";
+  if (mime === "application/pdf") return "pdf";
+  return null;
+}
+
+// OpenAI chat content block -> required capability (null = plain text/other, keep).
+function capForOpenAIBlock(block) {
+  const t = block?.type;
+  if (t === "image_url" || t === "image") return "vision";
+  if (t === "input_audio" || t === "audio_url") return "audioInput";
+  if (t === "file") return "pdf";
+  return null;
+}
+
+// Claude content block -> required capability.
+function capForClaudeBlock(block) {
+  const t = block?.type;
+  if (t === "image") return "vision";
+  if (t === "document") return "pdf";
+  return null;
+}
+
+// Filter an array of content blocks; drop unsupported, inject one placeholder per kind.
+// isLast = block belongs to the current user turn (picks the explanatory placeholder).
+function filterBlocks(blocks, capOf, caps, removed, isLast) {
+  const out = [];
+  for (const block of blocks) {
+    const cap = capOf(block);
+    if (cap && caps[cap] === false) { removed.add(cap); continue; }
+    out.push(block);
+  }
+  for (const cap of removed) out.push({ type: "text", text: ph(cap, isLast) });
+  return out;
+}
+
+// OpenAI / OpenAI-compatible chat messages[].content[].
+function stripOpenAI(body, caps) {
+  if (!Array.isArray(body.messages)) return;
+  const last = body.messages.length - 1;
+  body.messages.forEach((msg, i) => {
+    if (!Array.isArray(msg.content)) return;
+    const removed = new Set();
+    msg.content = filterBlocks(msg.content, capForOpenAIBlock, caps, removed, i === last);
+  });
+}
+
+// Claude messages[].content[].
+function stripClaude(body, caps) {
+  if (!Array.isArray(body.messages)) return;
+  const last = body.messages.length - 1;
+  body.messages.forEach((msg, i) => {
+    if (!Array.isArray(msg.content)) return;
+    const removed = new Set();
+    msg.content = filterBlocks(msg.content, capForClaudeBlock, caps, removed, i === last);
+  });
+}
+
+// OpenAI Responses input[].content[] (input_image / input_file).
+function stripResponses(body, caps) {
+  if (!Array.isArray(body.input)) return;
+  const last = body.input.length - 1;
+  body.input.forEach((item, i) => {
+    if (!Array.isArray(item.content)) return;
+    const removed = new Set();
+    item.content = item.content.filter((b) => {
+      const cap = b?.type === "input_image" ? "vision" : b?.type === "input_file" ? "pdf" : null;
+      if (cap && caps[cap] === false) { removed.add(cap); return false; }
+      return true;
+    });
+    for (const cap of removed) item.content.push({ type: "input_text", text: ph(cap, i === last) });
+  });
+}
+
+// Gemini / gemini-cli contents[].parts[] (inlineData / fileData by mime).
+function stripGeminiParts(contents, caps) {
+  if (!Array.isArray(contents)) return;
+  const last = contents.length - 1;
+  contents.forEach((c, i) => {
+    if (!Array.isArray(c.parts)) return;
+    const removed = new Set();
+    c.parts = c.parts.filter((p) => {
+      const mime = p?.inlineData?.mimeType || p?.fileData?.mimeType;
+      const cap = capForMime(mime);
+      if (cap && caps[cap] === false) { removed.add(cap); return false; }
+      return true;
+    });
+    for (const cap of removed) c.parts.push({ text: ph(cap, i === last) });
+  });
+}
+
+/**
+ * Remove media blocks the model can't read, in-place on the source-format body.
+ * @param {object} body - request body (source format)
+ * @param {string} sourceFormat - one of FORMATS
+ * @param {object} caps - capabilities from getCapabilitiesForModel
+ * @returns {boolean} true if anything was stripped-eligible (cap false for some modality)
+ */
+export function stripUnsupportedModalities(body, sourceFormat, caps) {
+  if (!body || !caps) return false;
+  // Fast exit: model supports everything we'd strip.
+  if (caps.vision !== false && caps.audioInput !== false && caps.pdf !== false) return false;
+
+  switch (sourceFormat) {
+    case FORMATS.OPENAI:
+    case FORMATS.OLLAMA:
+    case FORMATS.KIRO:
+    case FORMATS.CURSOR:
+    case FORMATS.COMMANDCODE:
+      stripOpenAI(body, caps);
+      break;
+    case FORMATS.CLAUDE:
+      stripClaude(body, caps);
+      break;
+    case FORMATS.OPENAI_RESPONSES:
+    case FORMATS.OPENAI_RESPONSE:
+    case FORMATS.CODEX:
+      stripResponses(body, caps);
+      break;
+    case FORMATS.GEMINI:
+    case FORMATS.GEMINI_CLI:
+    case FORMATS.VERTEX:
+      stripGeminiParts(body.contents, caps);
+      break;
+    case FORMATS.ANTIGRAVITY:
+      stripGeminiParts(body?.request?.contents, caps);
+      break;
+    default:
+      stripOpenAI(body, caps);
+  }
+  return true;
+}

open-sse/translator/concerns/prefetch.js

@@ -0,0 +1,96 @@
+// Pre-fetch remote image URLs into base64 BEFORE translation, for target
+// formats whose upstream providers cannot fetch remote URLs themselves
+// (they require inline base64). Runs on the source-format body.
+import { FORMATS } from "../formats.js";
+import { fetchImageAsBase64, parseDataUri } from "./image.js";
+
+// Targets that require inline base64 images (cannot accept remote URLs).
+const TARGETS_NEED_BASE64 = new Set([
+  FORMATS.GEMINI, FORMATS.GEMINI_CLI, FORMATS.VERTEX,
+  FORMATS.ANTIGRAVITY, FORMATS.OLLAMA, FORMATS.KIRO,
+]);
+
+function isRemoteUrl(url) {
+  return typeof url === "string" && (url.startsWith("http://") || url.startsWith("https://"));
+}
+
+// Collect {get,set} accessors for every remote image URL in a source body.
+function collectImageRefs(body, sourceFormat) {
+  const refs = [];
+  const pushOpenAI = (messages) => {
+    for (const msg of messages || []) {
+      if (!Array.isArray(msg.content)) continue;
+      for (const block of msg.content) {
+        if (block?.type === "image_url") {
+          const url = typeof block.image_url === "string" ? block.image_url : block.image_url?.url;
+          if (isRemoteUrl(url)) refs.push({ get: () => url, set: (v) => {
+            if (typeof block.image_url === "string") block.image_url = v; else block.image_url.url = v;
+          } });
+        }
+      }
+    }
+  };
+  const pushGemini = (contents) => {
+    for (const c of contents || []) {
+      for (const p of c.parts || []) {
+        const uri = p?.fileData?.fileUri;
+        if (isRemoteUrl(uri)) refs.push({ get: () => uri, part: p });
+      }
+    }
+  };
+
+  switch (sourceFormat) {
+    case FORMATS.OPENAI:
+    case FORMATS.OLLAMA:
+    case FORMATS.KIRO:
+    case FORMATS.CURSOR:
+    case FORMATS.COMMANDCODE:
+      pushOpenAI(body.messages);
+      break;
+    case FORMATS.CLAUDE:
+      for (const msg of body.messages || []) {
+        if (!Array.isArray(msg.content)) continue;
+        for (const block of msg.content) {
+          if (block?.type === "image" && block.source?.type === "url" && isRemoteUrl(block.source.url)) {
+            refs.push({ get: () => block.source.url, claudeBlock: block });
+          }
+        }
+      }
+      break;
+    case FORMATS.GEMINI:
+    case FORMATS.GEMINI_CLI:
+    case FORMATS.VERTEX:
+      pushGemini(body.contents);
+      break;
+    case FORMATS.ANTIGRAVITY:
+      pushGemini(body?.request?.contents);
+      break;
+    default:
+      pushOpenAI(body.messages);
+  }
+  return refs;
+}
+
+/**
+ * Replace remote image URLs with base64 data when the target needs inline data.
+ * No-op when target accepts remote URLs (e.g. openai, claude) or body has none.
+ * @returns {Promise<number>} count of images converted
+ */
+export async function prefetchRemoteImages(body, sourceFormat, targetFormat, options = {}) {
+  if (!body || !TARGETS_NEED_BASE64.has(targetFormat)) return 0;
+  const refs = collectImageRefs(body, sourceFormat);
+  if (!refs.length) return 0;
+
+  let converted = 0;
+  for (const ref of refs) {
+    const url = ref.get();
+    if (parseDataUri(url)) continue; // already inline
+    const fetched = await fetchImageAsBase64(url, options);
+    if (!fetched) continue;
+    if (ref.set) ref.set(fetched.url);
+    else if (ref.part) { delete ref.part.fileData; ref.part.inlineData = { mimeType: fetched.mimeType, data: fetched.url.split(",")[1] }; }
+    else if (ref.claudeBlock) ref.claudeBlock.source = { type: "base64", media_type: fetched.mimeType, data: fetched.url.split(",")[1] };
+    converted++;
+  }
+  return converted;
+}

open-sse/translator/concerns/reasoning.js

@@ -6,3 +6,19 @@ export function reasoningDelta(text, withRole = false) {
     ? { role: ROLE.ASSISTANT, reasoning_content: text }
     : { reasoning_content: text };
 }
+
+// Extract reasoning text from a streamed OpenAI-compatible delta across vendor shapes:
+//   - reasoning_content (GLM, Qwen, DeepSeek, Kimi, Step, Hunyuan)
+//   - reasoning (some compat layers)
+//   - reasoning_details[] (MiniMax reasoning_split=true): [{ text|content }]
+// Returns concatenated reasoning string, or "" when none.
+export function extractReasoningText(delta) {
+  if (!delta || typeof delta !== "object") return "";
+  if (typeof delta.reasoning_content === "string" && delta.reasoning_content) return delta.reasoning_content;
+  if (typeof delta.reasoning === "string" && delta.reasoning) return delta.reasoning;
+  const details = delta.reasoning_details;
+  if (Array.isArray(details)) {
+    return details.map((d) => (typeof d === "string" ? d : d?.text || d?.content || "")).join("");
+  }
+  return "";
+}

open-sse/translator/concerns/thinking.js

@@ -1,26 +1,50 @@
 // Concern: reasoning_effort ↔ provider-native thinking config.
-// Each provider expresses "how much to think" differently — centralize the maps here.
-// Streaming reasoning delta shape lives in reasoning.js; this file is request-side config only.
+// Central source of truth for level↔budget maps (web-standard values).
+// Provider-specific application lives in thinkingUnified.js; this file is maps-only.
 
-// OpenAI reasoning_effort → Claude thinking.budget_tokens
-const EFFORT_TO_BUDGET = { none: 0, low: 4096, medium: 8192, high: 16384, xhigh: 32768 };
+// Discrete effort levels, ordered low→high.
+export const EFFORT_LEVELS = ["minimal", "low", "medium", "high", "xhigh", "max"];
 
-// Returns budget_tokens for a reasoning_effort, or undefined if unknown effort.
-// 0 means "no thinking" (caller skips enabling); undefined means "effort not recognized".
+// Web-standard level → budget_tokens (Anthropic/Gemini docs).
+export const LEVEL_TO_BUDGET = {
+  none: 0,
+  minimal: 512,
+  low: 1024,
+  medium: 8192,
+  high: 24576,
+  xhigh: 32768,
+  max: 128000,
+};
+
+// Returns budget_tokens for an effort level, or undefined if unknown.
+// 0 means "no thinking"; undefined means "effort not recognized".
 export function effortToBudget(effort) {
   if (!effort) return undefined;
-  return EFFORT_TO_BUDGET[String(effort).toLowerCase()];
+  return LEVEL_TO_BUDGET[String(effort).toLowerCase()];
 }
 
 // OpenAI reasoning_effort → Gemini thinkingLevel (gemini-3 enum: minimal|low|medium|high).
-// Gemini 3 cannot fully disable thinking; "none"/"off" map to "minimal" (closest to off).
+// Gemini 3 cannot fully disable thinking; "none"/"off" map to "minimal".
 export function effortToThinkingLevel(effort) {
   const e = String(effort).toLowerCase().trim();
-  return e === "none" || e === "off" ? "minimal" : e;
+  if (e === "none" || e === "off") return "minimal";
+  if (e === "xhigh" || e === "max") return "high";
+  return e;
+}
+
+// Numeric budget → nearest discrete level (reverse map via thresholds).
+// Returns null when budget <= 0 (no reasoning).
+export function budgetToLevel(budget) {
+  const b = Number(budget);
+  if (!b || b <= 0) return null;
+  if (b <= 768) return "minimal";
+  if (b <= 4096) return "low";
+  if (b <= 16384) return "medium";
+  if (b <= 28672) return "high";
+  return "xhigh";
 }
 
 // Gemini thinkingBudget (numeric) → OpenAI reasoning_effort (antigravity reverse map).
-// Returns null when budget <= 0 (no reasoning).
 export function budgetToEffort(budget) {
   if (!budget || budget <= 0) return null;
   if (budget <= 2048) return "low";

open-sse/translator/concerns/thinkingUnified.js

@@ -0,0 +1,257 @@
+// Unified thinking normalization: extract client intent → apply provider-native format.
+// Config-driven: thinking format/limits come from capabilities.js + registry transport,
+// never hardcoded per-model here. See .docs/thinking/plan.md MATRIX VI-A.
+
+import { getCapabilitiesForModel } from "../../providers/capabilities.js";
+import { PROVIDERS } from "../../providers/index.js";
+import { LEVEL_TO_BUDGET, budgetToLevel, effortToBudget } from "./thinking.js";
+
+// Map a target wire-format to its native thinking format (when capability has none).
+const FORMAT_TO_NATIVE = {
+  openai: "openai",
+  "openai-responses": "openai",
+  "openai-response": "openai",
+  codex: "openai",
+  claude: "claude-budget",
+  gemini: "gemini-budget",
+  "gemini-cli": "gemini-budget",
+  vertex: "gemini-budget",
+  antigravity: "gemini-budget",
+  kiro: "kiro",
+};
+
+// Parse model-name suffix "model(value)" → { cleanModel, override }.
+// value: level name (high) | number (8192) | auto | none. null override when absent.
+export function parseSuffix(model) {
+  if (typeof model !== "string") return { cleanModel: model, override: null };
+  const m = model.match(/^(.*)\(([^()]+)\)\s*$/);
+  if (!m) return { cleanModel: model, override: null };
+  const cleanModel = m[1].trim();
+  const raw = m[2].trim().toLowerCase();
+  if (raw === "none" || raw === "off") return { cleanModel, override: { mode: "none" } };
+  if (raw === "auto") return { cleanModel, override: { mode: "auto" } };
+  if (/^\d+$/.test(raw)) return { cleanModel, override: { mode: "budget", budget: Number(raw) } };
+  if (LEVEL_TO_BUDGET[raw] !== undefined) return { cleanModel, override: { mode: "level", level: raw } };
+  return { cleanModel, override: null };
+}
+
+// Extract unified thinking intent from a request body (post-translation, mixed shapes).
+// Returns { mode, budget?, level? } or null when no thinking intent present.
+export function extractThinking(body) {
+  if (!body || typeof body !== "object") return null;
+
+  // Claude shape
+  const t = body.thinking;
+  if (t && typeof t === "object") {
+    if (t.type === "disabled") return { mode: "none" };
+    if (t.type === "adaptive" || t.type === "enabled") {
+      const budget = Number(t.budget_tokens);
+      if (Number.isFinite(budget) && budget > 0) return { mode: "budget", budget };
+      return { mode: "auto" };
+    }
+  }
+
+  // OpenAI chat / Responses shape
+  const effort = body.reasoning_effort ?? (typeof body.reasoning === "object" ? body.reasoning?.effort : null);
+  if (typeof effort === "string" && effort) {
+    const e = effort.toLowerCase();
+    if (e === "none" || e === "off") return { mode: "none" };
+    if (e === "auto") return { mode: "auto" };
+    return { mode: "level", level: e };
+  }
+
+  // Gemini shape (top-level, generationConfig, or request envelope)
+  const tc = body.thinkingConfig || body.generationConfig?.thinkingConfig || body.request?.generationConfig?.thinkingConfig;
+  if (tc && typeof tc === "object") {
+    if (typeof tc.thinkingLevel === "string") return { mode: "level", level: tc.thinkingLevel.toLowerCase() };
+    const tb = Number(tc.thinkingBudget);
+    if (Number.isFinite(tb)) {
+      if (tb === 0) return { mode: "none" };
+      if (tb < 0) return { mode: "auto" };
+      return { mode: "budget", budget: tb };
+    }
+  }
+
+  // Qwen shape
+  if (body.enable_thinking === false) return { mode: "none" };
+  if (body.enable_thinking === true) {
+    const tb = Number(body.thinking_budget);
+    if (Number.isFinite(tb) && tb > 0) return { mode: "budget", budget: tb };
+    return { mode: "auto" };
+  }
+
+  return null;
+}
+
+// Capture thinking intent from a body. Alias of extractThinking, named for clarity
+// at the call-site where intent is snapshotted before format translation.
+export const captureThinking = extractThinking;
+
+// Resolve thinking format: provider override > capability > derive(targetFormat).
+function resolveFormat(targetFormat, model, provider) {
+  const providerFmt = provider ? PROVIDERS[provider]?.thinkingFormat : null;
+  if (providerFmt) return providerFmt;
+  const caps = getCapabilitiesForModel(provider, model);
+  if (caps.thinkingFormat) return caps.thinkingFormat;
+  return FORMAT_TO_NATIVE[targetFormat] || "openai";
+}
+
+// Convert unified config to a budget number (for budget-based formats).
+function toBudget(cfg, range) {
+  let budget;
+  if (cfg.mode === "budget") budget = cfg.budget;
+  else if (cfg.mode === "level") budget = effortToBudget(cfg.level);
+  else if (cfg.mode === "auto") return -1;
+  if (!Number.isFinite(budget)) return undefined;
+  if (range) {
+    if (range.min != null && budget < range.min) budget = range.min;
+    if (range.max != null && budget > range.max) budget = range.max;
+  }
+  return budget;
+}
+
+// Convert unified config to a discrete level string.
+function toLevel(cfg) {
+  if (cfg.mode === "level") return cfg.level;
+  if (cfg.mode === "budget") return budgetToLevel(cfg.budget) || "medium";
+  if (cfg.mode === "auto") return "auto";
+  return null;
+}
+
+// Gemini nests thinkingConfig under generationConfig. gemini-cli / antigravity wrap
+// the whole request in a { request: { generationConfig } } envelope — target the
+// envelope's generationConfig when present, else the top-level one.
+function setGeminiThinking(body, tc) {
+  const gc = body.request?.generationConfig
+    ? body.request.generationConfig
+    : (body.generationConfig && typeof body.generationConfig === "object"
+        ? body.generationConfig
+        : (body.generationConfig = {}));
+  gc.thinkingConfig = tc;
+}
+
+// Strip every known thinking field from a body (used before re-applying / when unsupported).
+function stripAll(body) {
+  delete body.thinking;
+  delete body.reasoning_effort;
+  delete body.reasoning;
+  delete body.thinkingConfig;
+  delete body.enable_thinking;
+  delete body.thinking_budget;
+  delete body.output_config;
+  if (body.generationConfig) delete body.generationConfig.thinkingConfig;
+  if (body.request?.generationConfig) delete body.request.generationConfig.thinkingConfig;
+}
+
+// Apply unified thinking config to body in the resolved provider-native format.
+function applyFormat(fmt, body, cfg, caps) {
+  const none = cfg.mode === "none";
+  const canDisable = caps.thinkingCanDisable !== false;
+  // Model cannot disable thinking → clamp "none" to minimal effort instead.
+  const eff = none && !canDisable ? { mode: "level", level: "minimal" } : cfg;
+
+  switch (fmt) {
+    case "openai": {
+      if (none && canDisable) { body.reasoning_effort = "none"; break; }
+      const level = toLevel(eff);
+      if (level) body.reasoning_effort = level;
+      break;
+    }
+    case "claude-adaptive": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      const level = toLevel(eff);
+      body.output_config = { effort: level === "xhigh" ? "high" : level };
+      break;
+    }
+    case "claude-budget": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      const budget = toBudget(eff, caps.thinkingRange);
+      body.thinking = budget === -1 ? { type: "enabled" } : { type: "enabled", budget_tokens: budget || 8192 };
+      break;
+    }
+    case "gemini-level": {
+      const level = none ? "minimal" : (toLevel(eff) || "high");
+      setGeminiThinking(body, { thinkingLevel: level, includeThoughts: level !== "minimal" });
+      break;
+    }
+    case "gemini-budget": {
+      if (none && canDisable) { setGeminiThinking(body, { thinkingBudget: 0, includeThoughts: false }); break; }
+      const budget = toBudget(eff, caps.thinkingRange);
+      setGeminiThinking(body, { thinkingBudget: budget ?? -1, includeThoughts: true });
+      break;
+    }
+    case "zai": {
+      // Z.ai ignores thinking.disabled → must use enable_thinking:false to turn off.
+      if (none && canDisable) { body.enable_thinking = false; delete body.thinking; break; }
+      body.thinking = { type: "enabled" };
+      break;
+    }
+    case "qwen": {
+      if (none && canDisable) { body.enable_thinking = false; break; }
+      body.enable_thinking = true;
+      const budget = toBudget(eff, caps.thinkingRange);
+      if (Number.isFinite(budget) && budget > 0) body.thinking_budget = budget;
+      break;
+    }
+    case "deepseek": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      body.thinking = { type: "enabled" };
+      // DeepSeek: low/medium→high, xhigh/max→max.
+      const level = toLevel(eff);
+      body.reasoning_effort = level === "xhigh" || level === "max" ? "max" : "high";
+      break;
+    }
+    case "kimi": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      const level = toLevel(eff);
+      if (level) body.reasoning_effort = level === "max" ? "high" : level;
+      break;
+    }
+    case "minimax": {
+      // M3 adaptive; M2.x cannot disable (handled via canDisable clamp).
+      body.thinking = { type: none && canDisable ? "disabled" : "adaptive" };
+      break;
+    }
+    case "hunyuan": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      const budget = toBudget(eff, caps.thinkingRange);
+      body.thinking = budget === -1 ? { type: "enabled" } : { type: "enabled", budget_tokens: budget || 8192 };
+      break;
+    }
+    case "step": {
+      if (none && canDisable) break;
+      const level = toLevel(eff);
+      if (level) body.reasoning_effort = level === "xhigh" || level === "max" ? "high" : level;
+      break;
+    }
+    case "kiro":
+      // Kiro thinking handled via system-tag injection in openai-to-kiro.js; no body field here.
+      break;
+    default:
+      break;
+  }
+}
+
+// Public entry: normalize thinking for the resolved target format.
+// Mutates and returns body. No-op when model has no reasoning capability.
+// `intent` is a pre-captured config (from captureThinking on the original body);
+// falls back to extracting from the current body when omitted.
+export function applyThinking(targetFormat, model, body, provider = null, intent = undefined) {
+  if (!body || typeof body !== "object") return body;
+
+  const { cleanModel, override } = parseSuffix(model);
+  const cfg = override || intent || extractThinking(body);
+  const caps = getCapabilitiesForModel(provider, cleanModel);
+
+  // Model cannot reason → strip any stray thinking fields.
+  if (!caps.reasoning) {
+    stripAll(body);
+    return body;
+  }
+  if (!cfg) return body;
+
+  const fmt = resolveFormat(targetFormat, cleanModel, provider);
+  stripAll(body);
+  applyFormat(fmt, body, cfg, caps);
+  return body;
+}

open-sse/translator/formats/gemini.js

@@ -77,6 +77,14 @@ export function convertOpenAIContentToParts(content) {
             inlineData: { mime_type: mimeType, data: data }
           });
         }
+      } else if (item.type === OPENAI_BLOCK.FILE && item.file?.file_data?.startsWith("data:")) {
+        const url = item.file.file_data;
+        const commaIndex = url.indexOf(",");
+        if (commaIndex !== -1) {
+          const mimeType = url.substring(5, commaIndex).split(";")[0];
+          const data = url.substring(commaIndex + 1);
+          parts.push({ inlineData: { mime_type: mimeType, data: data } });
+        }
       }
     }
   }

open-sse/translator/index.js

@@ -4,6 +4,7 @@ import { prepareClaudeRequest } from "./formats/claude.js";
 import { cloakClaudeTools } from "../utils/claudeCloaking.js";
 import { filterToOpenAIFormat } from "./formats/openai.js";
 import { normalizeThinkingConfig } from "../services/provider.js";
+import { applyThinking, captureThinking } from "./concerns/thinkingUnified.js";
 import { AntigravityExecutor } from "../executors/antigravity.js";
 import { PROVIDERS } from "../providers/index.js";
 
@@ -63,6 +64,10 @@ export function translateRequest(sourceFormat, targetFormat, model, body, stream
   // Fix missing tool responses (insert empty tool_result if needed)
   fixMissingToolResponses(result);
 
+  // Capture thinking intent from the original (pre-translation) body, before any
+  // format conversion strips/renames the fields. Applied after translation.
+  const thinkingIntent = captureThinking(result);
+
   // If same format, skip translation steps
   if (sourceFormat !== targetFormat) {
     // Step 1: source -> openai (if source is not openai)
@@ -84,6 +89,9 @@ export function translateRequest(sourceFormat, targetFormat, model, body, stream
     }
   }
 
+  // Normalize thinking to the target provider-native format (config-driven, capability-aware)
+  applyThinking(targetFormat, model, result, provider, thinkingIntent);
+
   // Always normalize to clean OpenAI format when target is OpenAI
   // This handles hybrid requests (e.g., OpenAI messages + Claude tools)
   if (targetFormat === FORMATS.OPENAI) {

open-sse/translator/request/openai-to-claude.js

@@ -4,7 +4,6 @@ import { CLAUDE_SYSTEM_PROMPT } from "../../config/appConstants.js";
 import { adjustMaxTokens } from "../formats/maxTokens.js";
 import { safeParseJSON } from "../concerns/json.js";
 import { parseDataUri } from "../concerns/image.js";
-import { effortToBudget } from "../concerns/thinking.js";
 import { extractTextContent } from "../formats/gemini.js";
 import { ROLE, OPENAI_BLOCK, CLAUDE_BLOCK } from "../schema/index.js";
 
@@ -175,25 +174,7 @@ Respond ONLY with the JSON object, no other text.`);
     result.tool_choice = convertOpenAIToolChoice(body.tool_choice);
   }
 
-  // Thinking configuration
-  if (body.thinking) {
-    result.thinking = {
-      type: body.thinking.type || "enabled",
-      ...(body.thinking.budget_tokens && { budget_tokens: body.thinking.budget_tokens }),
-      ...(body.thinking.max_tokens && { max_tokens: body.thinking.max_tokens })
-    };
-  }
-
-  // Map OpenAI reasoning_effort → Claude thinking.budget_tokens
-  // When client sends reasoning_effort (OpenAI format) but no explicit thinking block,
-  // translate to Claude's native format.
-  if (body.reasoning_effort && !result.thinking) {
-    const budget = effortToBudget(body.reasoning_effort);
-    if (budget) {
-      result.thinking = { type: "enabled", budget_tokens: budget };
-    }
-    // budget === 0 (none) or undefined (unknown) → no thinking
-  }
+  // Thinking is normalized centrally by applyThinking (thinkingUnified.js) after translation.
 
   // Attach toolNameMap to result for response translation
   if (toolNameMap.size > 0) {
@@ -245,6 +226,16 @@ function getContentBlocksFromMessage(msg, toolNameMap = new Map()) {
           }
         } else if (part.type === OPENAI_BLOCK.IMAGE && part.source) {
           blocks.push({ type: CLAUDE_BLOCK.IMAGE, source: part.source });
+        } else if (part.type === OPENAI_BLOCK.FILE && part.file) {
+          // OpenAI file block -> Claude document (PDF only; Claude rejects other mimes).
+          const fileData = part.file.file_data;
+          const parsed = parseDataUri(fileData);
+          if (parsed && parsed.mimeType === "application/pdf") {
+            blocks.push({
+              type: CLAUDE_BLOCK.DOCUMENT,
+              source: { type: "base64", media_type: parsed.mimeType, data: parsed.base64 }
+            });
+          }
         }
       }
     }

open-sse/translator/request/openai-to-gemini.js

@@ -3,8 +3,6 @@ import { FORMATS } from "../formats.js";
 import { DEFAULT_THINKING_AG_SIGNATURE, DEFAULT_THINKING_GEMINI_CLI_SIGNATURE } from "../../config/defaultThinkingSignature.js";
 import { ANTIGRAVITY_DEFAULT_SYSTEM } from "../../config/appConstants.js";
 import { openaiToClaudeRequestForAntigravity } from "./openai-to-claude.js";
-import { effortToThinkingLevel } from "../concerns/thinking.js";
-
 function generateUUID() {
   return crypto.randomUUID();
 }
@@ -230,23 +228,7 @@ export function openaiToGeminiRequest(model, body, stream) {
 // OpenAI -> Gemini CLI (Cloud Code Assist)
 export function openaiToGeminiCLIRequest(model, body, stream) {
   const gemini = openaiToGeminiBase(model, body, stream, DEFAULT_THINKING_GEMINI_CLI_SIGNATURE);
-  const isClaude = model.toLowerCase().includes("claude");
-
-  // Map reasoning effort → thinkingConfig.thinkingLevel (gemini-3 enum: minimal|low|medium|high)
-  // Gemini 3 cannot fully disable thinking; "none"/"off" map to "minimal" (closest to no-thinking)
-  // Accept both OpenAI chat (reasoning_effort) and Responses (reasoning.effort) shapes
-  const reasoningEffort = body.reasoning_effort ?? body.reasoning?.effort;
-  if (reasoningEffort) {
-    const level = effortToThinkingLevel(reasoningEffort);
-    gemini.generationConfig.thinkingConfig = { thinkingLevel: level, includeThoughts: level !== "minimal" };
-  }
-
-  // Claude-format thinking: disabled → minimal, enabled → high
-  if (body.thinking?.type === "disabled") {
-    gemini.generationConfig.thinkingConfig = { thinkingLevel: "minimal", includeThoughts: false };
-  } else if (body.thinking?.type === "enabled") {
-    gemini.generationConfig.thinkingConfig = { thinkingLevel: "high", includeThoughts: true };
-  }
+  // Thinking is normalized centrally by applyThinking (thinkingUnified.js) after translation.
 
   // Clean schema for tools
   if (gemini.tools?.[0]?.functionDeclarations) {

open-sse/translator/response/openai-responses.js

@@ -7,7 +7,7 @@ import { FORMATS } from "../formats.js";
 import { buildChunk } from "../concerns/chunk.js";
 import { buildUsage } from "../concerns/usage.js";
 import { fallbackToolCallId } from "../concerns/toolCall.js";
-import { reasoningDelta } from "../concerns/reasoning.js";
+import { reasoningDelta, extractReasoningText } from "../concerns/reasoning.js";
 import { ROLE, OPENAI_BLOCK, RESPONSES_ITEM, OPENAI_FINISH, MODEL_FALLBACK } from "../schema/index.js";
 
 /**
@@ -62,10 +62,11 @@ export function openaiToOpenAIResponsesResponse(chunk, state) {
     });
   }
 
-  // Handle reasoning_content
-  if (delta.reasoning_content) {
+  // Handle reasoning across vendor shapes (reasoning_content / reasoning / reasoning_details)
+  const reasoningText = extractReasoningText(delta);
+  if (reasoningText) {
     startReasoning(state, emit, idx);
-    emitReasoningDelta(state, emit, delta.reasoning_content);
+    emitReasoningDelta(state, emit, reasoningText);
   }
 
   // Handle text content

open-sse/translator/response/openai-to-claude.js

@@ -2,6 +2,7 @@ import { register } from "../index.js";
 import { FORMATS } from "../formats.js";
 import { ROLE, CLAUDE_BLOCK, MODEL_FALLBACK } from "../schema/index.js";
 import { fromOpenAIFinish } from "../concerns/finishReason.js";
+import { extractReasoningText } from "../concerns/reasoning.js";
 
 // Legacy "proxy_" prefix used by older request translators. Response strips it
 // defensively so tool names from such turns resolve back (e.g. proxy_Read → Read
@@ -134,8 +135,8 @@ export function openaiToClaudeResponse(chunk, state) {
     });
   }
 
-  // Handle reasoning_content (thinking) - GLM, DeepSeek, etc.
-  const reasoningContent = delta?.reasoning_content || delta?.reasoning;
+  // Handle reasoning (thinking) across vendor shapes - GLM/DeepSeek/Qwen/MiniMax/etc.
+  const reasoningContent = extractReasoningText(delta);
   if (reasoningContent) {
     stopTextBlock(state, results);
 

open-sse/translator/schema/blocks.js

@@ -7,6 +7,7 @@ export const OPENAI_BLOCK = {
   IMAGE: "image",
   INPUT_AUDIO: "input_audio",
   AUDIO_URL: "audio_url",
+  FILE: "file",
   FUNCTION: "function",
 };
 
@@ -14,6 +15,7 @@ export const OPENAI_BLOCK = {
 export const CLAUDE_BLOCK = {
   TEXT: "text",
   IMAGE: "image",
+  DOCUMENT: "document",
   TOOL_USE: "tool_use",
   TOOL_RESULT: "tool_result",
   THINKING: "thinking",
@@ -34,7 +36,7 @@ export const RESPONSES_ITEM = {
 
 // Valid OpenAI block types (used by filterToOpenAIFormat).
 export const VALID_OPENAI_CONTENT_TYPES = [
-  OPENAI_BLOCK.TEXT, OPENAI_BLOCK.IMAGE_URL, OPENAI_BLOCK.IMAGE, OPENAI_BLOCK.INPUT_AUDIO, OPENAI_BLOCK.AUDIO_URL,
+  OPENAI_BLOCK.TEXT, OPENAI_BLOCK.IMAGE_URL, OPENAI_BLOCK.IMAGE, OPENAI_BLOCK.INPUT_AUDIO, OPENAI_BLOCK.AUDIO_URL, OPENAI_BLOCK.FILE,
 ];
 export const VALID_OPENAI_MESSAGE_TYPES = [
   OPENAI_BLOCK.TEXT, OPENAI_BLOCK.IMAGE_URL, OPENAI_BLOCK.IMAGE, "tool_calls", CLAUDE_BLOCK.TOOL_RESULT,

open-sse/utils/claudeCloaking.js

@@ -40,8 +40,11 @@ export function cloakClaudeTools(body) {
   const clientToolNames = new Set();
   const clientDeclarations = [];
 
-  // All client tools get renamed with suffix
+  // All client tools get renamed with suffix.
+  // Built-in server tools (web_search_20250305, etc.) carry a `type` and require
+  // an exact reserved `name` — never suffix those or Claude rejects the request.
   for (const tool of tools) {
+    if (tool.type) { clientDeclarations.push(tool); continue; }
     const suffixed = suffix(tool.name);
     toolNameMap.set(suffixed, tool.name);
     clientToolNames.add(tool.name);

open-sse/utils/streamHandler.js

@@ -128,7 +128,10 @@ export function createDisconnectAwareStream(transformStream, streamController, o
         controller.enqueue(value);
       } catch (error) {
         const wasConnected = streamController.isConnected();
-        streamController.handleError(error);
+        // Controller already closed = downstream ended; not an upstream error, skip noisy log.
+        const msg0 = error?.message || "";
+        const isControllerClosed = msg0.includes("already closed") || msg0.includes("Invalid state");
+        if (!isControllerClosed) streamController.handleError(error);
         reader.cancel().catch(() => {});
         writer.abort().catch(() => {});
 

package.json

@@ -7,7 +7,7 @@
     "dev": "next dev --webpack --port 20127",
     "build": "next build --webpack",
     "start": "next start",
-    "dev:bun": "bun --bun next dev --webpack --port 20128",
+    "dev:bun": "bun --bun next dev --webpack --port 20127",
     "build:bun": "bun --bun next build --webpack",
     "start:bun": "bun ./.next/standalone/server.js",
     "cli:pack": "npm --prefix cli run pack:cli",

scripts/test-combo-autoswitch.mjs

@@ -0,0 +1,84 @@
+// Live test: combo capacity display + auto-switch routing.
+// Sends text / image / search requests to a combo and reports which member ran.
+//   node scripts/test-combo-autoswitch.mjs
+const BASE = process.env.BASE_URL || "http://localhost:20127";
+const KEY = process.env.API_KEY || "sk-6581be4f05a82b6b-uxy6jn-c8190ea8";
+const COMBO = process.env.COMBO || "haha";
+
+// 16x16 PNG (valid image so vision providers accept it).
+const PNG = "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABAAAAAQCAIAAACQkWg2AAAAFklEQVR4nGO4I2JDEmIY1TCqYfhqAAAeBCwQ8YdREQAAAABJRU5ErkJggg==";
+
+function memberFromModel(model) {
+  // Response model usually = upstream id; map back to a combo member by substring.
+  return model || "(none)";
+}
+
+async function send(label, content, extra = {}) {
+  const body = {
+    model: COMBO,
+    stream: false,
+    max_tokens: 64,
+    messages: [{ role: "user", content }],
+    ...extra,
+  };
+  const t0 = Date.now();
+  let res, json, text;
+  try {
+    res = await fetch(`${BASE}/v1/chat/completions`, {
+      method: "POST",
+      headers: { "Content-Type": "application/json", Authorization: `Bearer ${KEY}` },
+      body: JSON.stringify(body),
+    });
+    text = await res.text();
+    try { json = JSON.parse(text); } catch { /* keep text */ }
+  } catch (e) {
+    console.log(`\n[${label}] NETWORK ERROR: ${e.message}`);
+    return;
+  }
+  const ms = Date.now() - t0;
+  const model = json?.model || "(no model field)";
+  const ok = res.ok;
+  const snippet = (json?.choices?.[0]?.message?.content || text || "").slice(0, 80).replace(/\n/g, " ");
+  console.log(`\n[${label}] ${ok ? "OK" : "FAIL"} ${res.status} (${ms}ms)`);
+  console.log(`  model executed: ${memberFromModel(model)}`);
+  if (!ok) console.log(`  error: ${(json?.error?.message || text || "").slice(0, 160)}`);
+  else console.log(`  reply: ${snippet}`);
+}
+
+async function showCaps() {
+  try {
+    const r = await fetch(`${BASE}/api/models`, { headers: { Authorization: `Bearer ${KEY}` } });
+    if (!r.ok) { console.log("(/api/models needs dashboard auth, skipping caps table)"); return; }
+    const { models } = await r.json();
+    const map = {};
+    for (const m of models || []) if (m.caps) map[m.fullModel] = m.caps;
+    console.log("Capacity of combo members (vision/search):");
+    for (const m of (process.env.MEMBERS || "").split(",").filter(Boolean)) {
+      const c = map[m] || {};
+      console.log(`  ${m}: vision=${!!c.vision} search=${!!c.search}`);
+    }
+  } catch { /* ignore */ }
+}
+
+(async () => {
+  console.log(`Testing combo "${COMBO}" @ ${BASE}\n${"=".repeat(50)}`);
+  await showCaps();
+
+  // 1. Text-only: round-robin order (no capability requirement).
+  await send("text-only #1", "Say hello in one word.");
+  await send("text-only #2", "Say hi in one word.");
+
+  // 2. Image: should auto-switch to a vision-capable member.
+  await send("image (needs vision)", [
+    { type: "text", text: "What color is this image? One word." },
+    { type: "image_url", image_url: { url: PNG } },
+  ]);
+
+  // 3. Search: should auto-switch to a search-capable member.
+  // Claude built-in web search requires a versioned tool type.
+  await send("search (needs search)", "What is the latest news today?", {
+    tools: [{ type: "web_search_20250305", name: "web_search" }],
+  });
+
+  console.log(`\n${"=".repeat(50)}\nDone. Compare 'model executed' across cases to verify auto-switch.`);
+})();

src/app/(dashboard)/dashboard/cli-tools/components/ClaudeToolCard.js

@@ -347,10 +347,10 @@ export default function ClaudeToolCard({
                   <label className="flex items-center gap-1.5 cursor-pointer select-none">
                     <input type="checkbox" checked={ccFilterNaming} onChange={handleCcFilterNamingToggle} className="w-3.5 h-3.5 accent-primary cursor-pointer" />
                     <span className="text-xs text-text-muted">Filter naming requests</span>
-                  </label>
                     <Tooltip text="Intercepts Claude Code's topic-naming requests and returns a fake response locally, saving API tokens.">
                       <span className="material-symbols-outlined text-text-muted text-[14px] cursor-help">info</span>
                     </Tooltip>
+                  </label>
                 </div>
               </div>
 

src/app/(dashboard)/dashboard/combos/page.js

@@ -5,7 +5,7 @@ import { DndContext, closestCenter, KeyboardSensor, PointerSensor, useSensor, us
 import { arrayMove, SortableContext, sortableKeyboardCoordinates, useSortable, verticalListSortingStrategy } from "@dnd-kit/sortable";
 import { CSS } from "@dnd-kit/utilities";
 import { restrictToVerticalAxis, restrictToParentElement } from "@dnd-kit/modifiers";
-import { Card, Button, Modal, Input, CardSkeleton, ModelSelectModal, Toggle, ConfirmModal } from "@/shared/components";
+import { Card, Button, Modal, Input, CardSkeleton, ModelSelectModal, Toggle, ConfirmModal, CapacityBadges } from "@/shared/components";
 import { useCopyToClipboard } from "@/shared/hooks/useCopyToClipboard";
 import { isOpenAICompatibleProvider, isAnthropicCompatibleProvider } from "@/shared/constants/providers";
 
@@ -19,6 +19,7 @@ export default function CombosPage() {
   const [editingCombo, setEditingCombo] = useState(null);
   const [activeProviders, setActiveProviders] = useState([]);
   const [comboStrategies, setComboStrategies] = useState({});
+  const [modelCaps, setModelCaps] = useState({});
   const [confirmState, setConfirmState] = useState(null);
   const { copied, copy } = useCopyToClipboard();
 
@@ -28,10 +29,11 @@ export default function CombosPage() {
 
   const fetchData = async () => {
     try {
-      const [combosRes, providersRes, settingsRes] = await Promise.all([
+      const [combosRes, providersRes, settingsRes, modelsRes] = await Promise.all([
         fetch("/api/combos"),
         fetch("/api/providers"),
         fetch("/api/settings"),
+        fetch("/api/models"),
       ]);
       const combosData = await combosRes.json();
       const providersData = await providersRes.json();
@@ -42,6 +44,13 @@ export default function CombosPage() {
       if (providersRes.ok) {
         setActiveProviders(providersData.connections || []);
       }
+      if (modelsRes.ok) {
+        const md = await modelsRes.json();
+        // Build fullModel -> caps map for badge lookup
+        const map = {};
+        for (const m of md.models || []) if (m.caps) map[m.fullModel] = m.caps;
+        setModelCaps(map);
+      }
       setComboStrategies(settingsData.comboStrategies || {});
     } catch (error) {
       console.log("Error fetching data:", error);
@@ -143,7 +152,7 @@ export default function CombosPage() {
         <div className="min-w-0">
           <h1 className="text-2xl font-semibold">Combos</h1>
           <p className="text-sm text-text-muted mt-1">
-            Create model combos with fallback support
+            Create model combos with fallback support — auto-adapts per request: routes images to vision models and web search to search-capable models.
           </p>
         </div>
         <Button icon="add" onClick={() => setShowCreateModal(true)} className="w-full sm:w-auto">
@@ -171,6 +180,7 @@ export default function CombosPage() {
             <ComboCard
               key={combo.id}
               combo={combo}
+              modelCaps={modelCaps}
               copied={copied}
               onCopy={copy}
               onEdit={() => setEditingCombo(combo)}
@@ -214,7 +224,7 @@ export default function CombosPage() {
   );
 }
 
-function ComboCard({ combo, copied, onCopy, onEdit, onDelete, roundRobinEnabled, onToggleRoundRobin }) {
+function ComboCard({ combo, modelCaps = {}, copied, onCopy, onEdit, onDelete, roundRobinEnabled, onToggleRoundRobin }) {
   return (
     <Card padding="sm" className="group">
       <div className="flex min-w-0 flex-col gap-3 sm:flex-row sm:items-center sm:justify-between">
@@ -229,8 +239,9 @@ function ComboCard({ combo, copied, onCopy, onEdit, onDelete, roundRobinEnabled,
                 <span className="text-xs text-text-muted italic">No models</span>
               ) : (
                 combo.models.slice(0, 3).map((model, index) => (
-                  <code key={index} className="max-w-full truncate rounded bg-black/5 px-1.5 py-0.5 font-mono text-[10px] text-text-muted dark:bg-white/5 sm:max-w-[220px]">
-                    {model}
+                  <code key={index} className="inline-flex items-center gap-1 rounded bg-black/5 px-1.5 py-0.5 font-mono text-xs text-text-muted dark:bg-white/5">
+                    <span>{model}</span>
+                    <CapacityBadges caps={modelCaps[model]} />
                   </code>
                 ))
               )}

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

@@ -1,6 +1,7 @@
 import PropTypes from "prop-types";
+import { CapacityBadges } from "@/shared/components";
 
-export default function ModelRow({ model, fullModel, alias, copied, onCopy, testStatus, isCustom, isFree, onDeleteAlias, onTest, isTesting, onDisable }) {
+export default function ModelRow({ model, fullModel, alias, copied, onCopy, testStatus, isCustom, isFree, onDeleteAlias, onTest, isTesting, onDisable, caps }) {
   const borderColor = testStatus === "ok"
     ? "border-green-500/40"
     : testStatus === "error"
@@ -24,7 +25,10 @@ export default function ModelRow({ model, fullModel, alias, copied, onCopy, test
         </span>
         <div className="flex min-w-0 flex-1 flex-col gap-1">
           <code className="max-w-[72vw] truncate rounded bg-sidebar px-1.5 py-0.5 font-mono text-xs text-text-muted sm:max-w-[360px]">{fullModel}</code>
-          {model.name && <span className="truncate pl-1 text-[9px] italic text-text-muted/70">{model.name}</span>}
+          <span className="flex min-w-0 items-center text-[9px] gap-1 pl-1">
+            {model.name && <span className="truncate text-[9px] italic text-text-muted/70">{model.name}</span>}
+            <CapacityBadges caps={caps} colorOverride="text-text-muted/70" size={12} />
+          </span>
         </div>
         {onTest && (
           <div className="relative shrink-0 group/btn">
@@ -92,4 +96,5 @@ ModelRow.propTypes = {
   onTest: PropTypes.func,
   isTesting: PropTypes.bool,
   onDisable: PropTypes.func,
+  caps: PropTypes.object,
 };

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

@@ -8,6 +8,7 @@ import { Card, Button, Badge, Input, Modal, CardSkeleton, OAuthModal, KiroOAuthW
 import { OAUTH_PROVIDERS, APIKEY_PROVIDERS, FREE_PROVIDERS, FREE_TIER_PROVIDERS, WEB_COOKIE_PROVIDERS, getProviderAlias, isOpenAICompatibleProvider, isAnthropicCompatibleProvider, AI_PROVIDERS, THINKING_CONFIG } from "@/shared/constants/providers";
 import { getModelsByProviderId, getModelKind } from "@/shared/constants/models";
 import { useCopyToClipboard } from "@/shared/hooks/useCopyToClipboard";
+import { useModelCaps } from "@/shared/hooks/useModelCaps";
 import { translate } from "@/i18n/runtime";
 import { fetchSuggestedModels } from "@/shared/utils/providerModelsFetcher";
 import ModelRow from "./ModelRow";
@@ -29,6 +30,7 @@ export default function ProviderDetailPage() {
   const params = useParams();
   const router = useRouter();
   const providerId = params.id;
+  const { getCaps } = useModelCaps();
   const [connections, setConnections] = useState([]);
   const [loading, setLoading] = useState(true);
   const [providerNode, setProviderNode] = useState(null);
@@ -953,6 +955,7 @@ export default function ProviderDetailPage() {
             isTesting={testingModelId === model.id}
             isCustom
             isFree={false}
+            caps={getCaps(`${providerId}/${model.id}`)}
           />
         ))}
 
@@ -977,6 +980,7 @@ export default function ProviderDetailPage() {
               isTesting={testingModelId === model.id}
               isFree={model.isFree}
               onDisable={() => handleDisableModel(model.id)}
+              caps={getCaps(`${providerId}/${model.id}`)}
             />
           );
         })}

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

@@ -4,6 +4,7 @@ import { useState, useEffect, useCallback, useRef, useMemo } from "react";
 import ProviderIcon from "@/shared/components/ProviderIcon";
 import QuotaTable from "./QuotaTable";
 import Toggle from "@/shared/components/Toggle";
+import Tooltip from "@/shared/components/Tooltip";
 import {
   parseQuotaData,
   calculatePercentage,
@@ -34,9 +35,15 @@ import {
   QUOTA_SORT_OPTIONS,
 } from "./utils";
 import Card from "@/shared/components/Card";
-import { EditConnectionModal } from "@/shared/components";
+import { ConfirmModal, EditConnectionModal } from "@/shared/components";
 import { USAGE_SUPPORTED_PROVIDERS } from "@/shared/constants/providers";
 
+function getCodexResetCreditCount(quota) {
+  const value = quota?.raw?.resetCredits?.availableCount;
+  const count = typeof value === "number" ? value : Number(value);
+  return Number.isFinite(count) ? Math.max(0, count) : 0;
+}
+
 export default function ProviderLimits() {
   const [connections, setConnections] = useState([]);
   const [quotaData, setQuotaData] = useState({});
@@ -50,6 +57,8 @@ export default function ProviderLimits() {
   const [connectionsLoading, setConnectionsLoading] = useState(true);
   const [deletingId, setDeletingId] = useState(null);
   const [togglingId, setTogglingId] = useState(null);
+  const [resettingLimitId, setResettingLimitId] = useState(null);
+  const [resetConfirmState, setResetConfirmState] = useState(null);
   const [showEditModal, setShowEditModal] = useState(false);
   const [selectedConnection, setSelectedConnection] = useState(null);
   const [proxyPools, setProxyPools] = useState([]);
@@ -207,6 +216,32 @@ export default function ProviderLimits() {
     [fetchQuota],
   );
 
+  const handleResetCodexLimit = useCallback(
+    async (connectionId, provider) => {
+      if (provider !== "codex" || resettingLimitId) return;
+
+      setResettingLimitId(connectionId);
+      setErrors((prev) => ({ ...prev, [connectionId]: null }));
+
+      try {
+        const response = await fetch(`/api/usage/${connectionId}/codex-reset-credits`, { method: "POST" });
+        const result = await response.json().catch(() => ({}));
+
+        if (!response.ok) {
+          throw new Error(result.message || result.error || result.code || "Failed to reset Codex limit");
+        }
+
+        await fetchQuota(connectionId, provider);
+        setLastUpdated(new Date());
+      } catch (error) {
+        setErrors((prev) => ({ ...prev, [connectionId]: error.message || "Failed to reset Codex limit" }));
+      } finally {
+        setResettingLimitId(null);
+      }
+    },
+    [fetchQuota, resettingLimitId],
+  );
+
   const handleDeleteConnection = useCallback(
     async (id) => {
       if (!confirm("Delete this connection?")) return;
@@ -791,7 +826,10 @@ export default function ProviderLimits() {
 
           // Use table layout for all providers
           const isInactive = conn.isActive === false;
-          const rowBusy = deletingId === conn.id || togglingId === conn.id;
+          const isCodex = conn.provider === "codex";
+          const resetCreditCount = getCodexResetCreditCount(quota);
+          const isResettingLimit = resettingLimitId === conn.id;
+          const rowBusy = deletingId === conn.id || togglingId === conn.id || isResettingLimit;
 
           return (
             <Card
@@ -822,17 +860,51 @@ export default function ProviderLimits() {
                           {getConnectionLabel(conn)}
                         </p>
                       ) : null}
+                      {isCodex && (
+                        <p className="text-[11px] text-text-muted truncate">
+                          Reset eligible: {resetCreditCount}
+                        </p>
+                      )}
                     </div>
                   </div>
 
                   <div className="flex items-center gap-1 shrink-0">
+                    {isCodex && (
+                      <Tooltip text={`Codex reset credits remaining: ${resetCreditCount}`}>
+                        <div
+                          className={`hidden h-8 items-center gap-1 rounded-lg border px-2 text-[11px] sm:flex ${
+                            resetCreditCount > 0
+                              ? "border-primary/30 bg-primary/5 text-primary"
+                              : "border-black/10 bg-black/[0.02] text-text-muted dark:border-white/10 dark:bg-white/[0.03]"
+                          }`}
+                        >
+                          <span className="material-symbols-outlined text-[14px]">restart_alt</span>
+                          <span className="tabular-nums">{resetCreditCount}</span>
+                        </div>
+                      </Tooltip>
+                    )}
+                    {isCodex && resetCreditCount > 0 && (
+                      <Tooltip text={`Use one Codex reset credit. Available: ${resetCreditCount}`}>
+                        <button
+                          type="button"
+                          onClick={() => setResetConfirmState({ connection: conn, resetCreditCount })}
+                          disabled={isLoading || rowBusy}
+                          className="flex h-8 items-center gap-1 rounded-lg border border-primary/30 px-2 text-[11px] text-primary transition-colors hover:bg-primary/10 disabled:opacity-50"
+                        >
+                          <span className={`material-symbols-outlined text-[15px] ${isResettingLimit ? "animate-spin" : ""}`}>
+                            {isResettingLimit ? "progress_activity" : "bolt"}
+                          </span>
+                          <span className="hidden lg:inline">Reset limit</span>
+                        </button>
+                      </Tooltip>
+                    )}
+                    <Tooltip text="Refresh quota">
                       <button
                         type="button"
                         onClick={() => refreshProvider(conn.id, conn.provider)}
                         disabled={isLoading || rowBusy}
                         aria-label="Refresh quota"
-                      className="p-1.5 rounded-lg hover:bg-black/5 dark:hover:bg-white/5 transition-colors disabled:opacity-50"
-                      title="Refresh quota"
+                        className="flex h-8 w-8 items-center justify-center rounded-lg hover:bg-black/5 dark:hover:bg-white/5 transition-colors disabled:opacity-50"
                       >
                         <span
                           className={`material-symbols-outlined text-[18px] text-text-muted ${isLoading ? "animate-spin" : ""}`}
@@ -840,6 +912,8 @@ export default function ProviderLimits() {
                           refresh
                         </span>
                       </button>
+                    </Tooltip>
+                    <Tooltip text="Edit connection">
                       <button
                         type="button"
                         onClick={() => {
@@ -848,20 +922,20 @@ export default function ProviderLimits() {
                         }}
                         disabled={rowBusy}
                         aria-label="Edit connection"
-                      className="p-1.5 rounded-lg hover:bg-black/5 dark:hover:bg-white/5 text-text-muted hover:text-primary transition-colors disabled:opacity-50"
-                      title="Edit connection"
+                        className="flex h-8 w-8 items-center justify-center rounded-lg hover:bg-black/5 dark:hover:bg-white/5 text-text-muted hover:text-primary transition-colors disabled:opacity-50"
                       >
                         <span className="material-symbols-outlined text-[18px]">
                           edit
                         </span>
                       </button>
+                    </Tooltip>
+                    <Tooltip text="Delete connection">
                       <button
                         type="button"
                         onClick={() => handleDeleteConnection(conn.id)}
                         disabled={rowBusy}
                         aria-label="Delete connection"
-                      className="p-1.5 rounded-lg hover:bg-red-500/10 text-red-500 transition-colors disabled:opacity-50"
-                      title="Delete connection"
+                        className="flex h-8 w-8 items-center justify-center rounded-lg hover:bg-red-500/10 text-red-500 transition-colors disabled:opacity-50"
                       >
                         <span
                           className={`material-symbols-outlined text-[18px] ${deletingId === conn.id ? "animate-pulse" : ""}`}
@@ -869,6 +943,7 @@ export default function ProviderLimits() {
                           delete
                         </span>
                       </button>
+                    </Tooltip>
                     <div
                       className="inline-flex items-center pl-0.5"
                       title={
@@ -1047,6 +1122,25 @@ export default function ProviderLimits() {
           </div>
         </div>
 
+      <ConfirmModal
+        isOpen={Boolean(resetConfirmState)}
+        onClose={() => {
+          if (!resettingLimitId) setResetConfirmState(null);
+        }}
+        onConfirm={async () => {
+          const connection = resetConfirmState?.connection;
+          if (!connection) return;
+          await handleResetCodexLimit(connection.id, connection.provider);
+          setResetConfirmState(null);
+        }}
+        title="Reset Codex limit?"
+        message={`Use 1 Codex reset credit for ${getConnectionLabel(resetConfirmState?.connection || {}) || "this account"}. This cannot be undone. Remaining credits: ${resetConfirmState?.resetCreditCount ?? 0}.`}
+        confirmText="Reset limit"
+        cancelText="Cancel"
+        variant="danger"
+        loading={Boolean(resettingLimitId)}
+      />
+
       <EditConnectionModal
         isOpen={showEditModal}
         connection={selectedConnection}

src/app/api/models/route.js

@@ -3,6 +3,7 @@ import { getModelAliases, setModelAlias } from "@/models";
 import { getDisabledModels } from "@/lib/disabledModelsDb";
 import { AI_MODELS } from "@/shared/constants/config";
 import { getProviderAlias } from "@/shared/constants/providers";
+import { getCapabilitiesForModel } from "open-sse/providers/capabilities.js";
 
 // GET /api/models - Get models with aliases
 export async function GET() {
@@ -18,10 +19,12 @@ export async function GET() {
       })
       .map((m) => {
         const fullModel = `${m.provider}/${m.model}`;
+        const c = getCapabilitiesForModel(m.provider, m.model);
         return {
           ...m,
           fullModel,
           alias: modelAliases[fullModel] || m.model,
+          caps: { vision: c.vision, search: c.search, reasoning: c.reasoning },
         };
       });
 

src/app/api/usage/[connectionId]/codex-reset-credits/route.js

@@ -0,0 +1,104 @@
+// Ensure proxyFetch is loaded to patch globalThis.fetch
+import "open-sse/index.js";
+
+import { getProviderConnectionById } from "@/lib/localDb";
+import { consumeCodexRateLimitResetCredit } from "open-sse/services/usage.js";
+import { resolveConnectionProxyConfig } from "@/lib/network/connectionProxy";
+import { refreshAndUpdateCredentials } from "../route.js";
+
+const AUTH_EXPIRED_PATTERNS = ["expired", "authentication", "unauthorized", "401", "re-authorize"];
+
+function isAuthExpiredResult(result) {
+  const values = [result?.message, result?.code, result?.raw?.detail, result?.raw?.error]
+    .filter(Boolean)
+    .map((value) => String(value).toLowerCase());
+  return values.some((value) => AUTH_EXPIRED_PATTERNS.some((pattern) => value.includes(pattern)));
+}
+
+function getResponseForConsumeResult(result, redeemRequestId) {
+  if (result.ok) {
+    return Response.json({
+      code: result.code,
+      reset: true,
+      windows_reset: result.windowsReset,
+      redeemRequestId,
+      credit: result.raw?.credit || null,
+    });
+  }
+
+  if (result.noCredit) {
+    return Response.json({
+      code: "no_credit",
+      reset: false,
+      windows_reset: result.windowsReset,
+      message: "No Codex reset credits available.",
+    }, { status: 409 });
+  }
+
+  return Response.json({
+    code: result.code || "unknown_response",
+    reset: false,
+    windows_reset: result.windowsReset,
+    message: result.message || "Codex reset credit consume returned an unexpected response.",
+  }, { status: result.status >= 400 && result.status < 500 ? result.status : 502 });
+}
+
+export async function POST(request, { params }) {
+  let connection;
+  try {
+    const { connectionId } = await params;
+    connection = await getProviderConnectionById(connectionId);
+    if (!connection) {
+      return Response.json({ error: "Connection not found" }, { status: 404 });
+    }
+
+    if (connection.provider !== "codex") {
+      return Response.json({ error: "Codex reset credits are only available for Codex connections." }, { status: 400 });
+    }
+
+    const isOAuth = connection.authType === "oauth";
+    const isAccessToken = connection.authType === "access_token";
+    if (!isOAuth && !isAccessToken) {
+      return Response.json({ error: "Codex reset credits require an OAuth or access-token connection." }, { status: 400 });
+    }
+
+    const proxyConfig = await resolveConnectionProxyConfig(connection.providerSpecificData);
+    const proxyOptions = {
+      connectionProxyEnabled: proxyConfig.connectionProxyEnabled === true,
+      connectionProxyUrl: proxyConfig.connectionProxyUrl || "",
+      connectionNoProxy: proxyConfig.connectionNoProxy || "",
+      vercelRelayUrl: proxyConfig.vercelRelayUrl || "",
+      strictProxy: false,
+    };
+
+    if (isOAuth) {
+      try {
+        const result = await refreshAndUpdateCredentials(connection, false, proxyOptions);
+        connection = result.connection;
+      } catch (refreshError) {
+        console.error("[Codex Reset Credits API] Credential refresh failed:", refreshError);
+        return Response.json({ error: `Credential refresh failed: ${refreshError.message}` }, { status: 401 });
+      }
+    }
+
+    // Server-generated redeem id prevents client-controlled replay
+    const redeemRequestId = crypto.randomUUID();
+    let consumeResult = await consumeCodexRateLimitResetCredit(connection.accessToken, redeemRequestId, proxyOptions);
+
+    if (isOAuth && isAuthExpiredResult(consumeResult) && connection.refreshToken) {
+      try {
+        const retryResult = await refreshAndUpdateCredentials(connection, true, proxyOptions);
+        connection = retryResult.connection;
+        consumeResult = await consumeCodexRateLimitResetCredit(connection.accessToken, redeemRequestId, proxyOptions);
+      } catch (retryError) {
+        console.warn(`[Codex Reset Credits] force refresh failed: ${retryError.message}`);
+      }
+    }
+
+    return getResponseForConsumeResult(consumeResult, redeemRequestId);
+  } catch (error) {
+    const provider = connection?.provider ?? "unknown";
+    console.warn(`[Codex Reset Credits] ${provider}: ${error.message}`);
+    return Response.json({ error: error.message }, { status: 500 });
+  }
+}

src/app/api/usage/[connectionId]/route.js

@@ -20,7 +20,7 @@ function isAuthExpiredMessage(usage) {
  * @param {boolean} force - Skip needsRefresh check and always attempt refresh
  * @returns Promise<{ connection, refreshed: boolean }>
  */
-async function refreshAndUpdateCredentials(connection, force = false, proxyOptions = null) {
+export async function refreshAndUpdateCredentials(connection, force = false, proxyOptions = null) {
   const executor = getExecutor(connection.provider);
 
   // Build credentials object from connection

src/app/globals.css

@@ -324,6 +324,20 @@ button {
   user-select: none;
 }
 
+/* Tailwind v4 dropped default button cursor — restore globally */
+button:not(:disabled),
+[role="button"]:not([aria-disabled="true"]),
+label[for],
+summary,
+select {
+  cursor: pointer;
+}
+
+button:disabled,
+[role="button"][aria-disabled="true"] {
+  cursor: not-allowed;
+}
+
 /* ============================================================
    Animations
    ============================================================ */

src/dashboardGuard.js

@@ -32,7 +32,7 @@ const PUBLIC_API_PATHS = [
 ];
 
 // Public top-level prefixes (LLM API endpoints with their own API key auth).
-const PUBLIC_PREFIXES = ["/v1", "/v1beta", "/api/v1", "/api/v1beta"];
+const PUBLIC_PREFIXES = ["/v1", "/v1beta", "/api/v1", "/api/v1beta", "/codex"];
 
 // Always require JWT token regardless of requireLogin setting
 const ALWAYS_PROTECTED = [
@@ -90,7 +90,14 @@ function isLoopbackHostname(h) {
 }
 
 export function isLocalRequest(request) {
-  if (!isLoopbackHostname(request.headers.get("host"))) return false;
+  // Trusted peer IP from TCP socket (custom-server.js); unspoofable. Primary anchor for "local".
+  const realIp = request.headers.get("x-9r-real-ip");
+  if (realIp) {
+    if (!isLoopbackHostname(realIp)) return false;
+  } else if (!isLoopbackHostname(request.headers.get("host"))) {
+    // Fallback for bare server.js (dev) without custom-server: legacy Host-based check.
+    return false;
+  }
   const origin = request.headers.get("origin");
   if (origin) {
     try {

src/shared/components/CapacityBadges.js

@@ -0,0 +1,28 @@
+"use client";
+
+import { CAPACITY_META } from "@/shared/constants/models";
+import Tooltip from "./Tooltip";
+
+// Render small icon badges for a model's capabilities (only those set true).
+// colorOverride: force a single color class for all badges (default: per-cap color).
+// size: icon font-size in px (default 16).
+export default function CapacityBadges({ caps, className = "", colorOverride, size = 16 }) {
+  if (!caps) return null;
+  const active = Object.keys(CAPACITY_META).filter((k) => caps[k]);
+  if (active.length === 0) return null;
+
+  return (
+    <span className={`inline-flex items-center gap-0.5 ${className}`}>
+      {active.map((k) => (
+        <Tooltip key={k} text={`${CAPACITY_META[k].label} — ${CAPACITY_META[k].desc}`}>
+          <span
+            className={`material-symbols-outlined leading-none cursor-help ${colorOverride || CAPACITY_META[k].color}`}
+            style={{ fontSize: `${size}px` }}
+          >
+            {CAPACITY_META[k].icon}
+          </span>
+        </Tooltip>
+      ))}
+    </span>
+  );
+}

src/shared/components/ModelSelectModal.js

@@ -4,6 +4,8 @@ import { useState, useMemo, useEffect } from "react";
 import PropTypes from "prop-types";
 import Modal from "./Modal";
 import ProviderIcon from "./ProviderIcon";
+import CapacityBadges from "./CapacityBadges";
+import { useModelCaps } from "@/shared/hooks/useModelCaps";
 import { getModelsByProviderId, getModelKind } from "@/shared/constants/models";
 import { OAUTH_PROVIDERS, APIKEY_PROVIDERS, FREE_PROVIDERS, FREE_TIER_PROVIDERS, AI_PROVIDERS, isOpenAICompatibleProvider, isAnthropicCompatibleProvider, getProviderAlias } from "@/shared/constants/providers";
 
@@ -40,6 +42,7 @@ export default function ModelSelectModal({
       return kinds.includes(kindFilter);
     });
   }, [activeProviders, kindFilter]);
+  const { getCaps } = useModelCaps();
   const [searchQuery, setSearchQuery] = useState("");
   const [combos, setCombos] = useState([]);
   const [providerNodes, setProviderNodes] = useState([]);
@@ -499,9 +502,13 @@ export default function ModelSelectModal({
                         <>
                           {model.name}
                           <span className="text-[9px] opacity-60 font-normal">custom</span>
+                          <CapacityBadges caps={getCaps(model.value)} />
                         </>
                       ) : (
-                        model.name
+                        <>
+                          {model.name}
+                          <CapacityBadges caps={getCaps(model.value)} />
+                        </>
                       )}
                     </span>
                   </button>

src/shared/components/Tooltip.js

@@ -12,10 +12,10 @@ export default function Tooltip({ text, children, position = "top", color }) {
   const bgClass = color ? "" : "bg-gray-900";
 
   return (
-    <div className="relative inline-flex group">
+    <div className="relative inline-flex group/tt">
       {children}
       <div
-        className={`pointer-events-none absolute ${posClass} z-50 w-max max-w-56 rounded px-2 py-1 text-[11px] leading-snug ${bgClass} text-white opacity-0 group-hover:opacity-100 transition-opacity duration-150 whitespace-normal`}
+        className={`pointer-events-none absolute ${posClass} z-50 w-max max-w-56 rounded px-2 py-1 text-[11px] leading-snug ${bgClass} text-white opacity-0 group-hover/tt:opacity-100 transition-opacity duration-150 whitespace-normal`}
         style={bgStyle}
       >
         {text}

src/shared/components/index.js

@@ -36,6 +36,7 @@ export { default as NoAuthProxyCard } from "./NoAuthProxyCard";
 export { default as SegmentedControl } from "./SegmentedControl";
 export { default as Tooltip } from "./Tooltip";
 export { default as ProviderInfoCard } from "./ProviderInfoCard";
+export { default as CapacityBadges } from "./CapacityBadges";
 
 // Layouts
 export * from "./layouts";

src/shared/constants/models.js

@@ -38,3 +38,10 @@ export const AI_MODELS = Object.entries(MODELS).flatMap(([alias, models]) =>
 );
 
 export const getModelKind = (m, fallback = null) => m?.kind || m?.type || fallback;
+
+// Capacity metadata for UI badges — icon + label + color per capability.
+export const CAPACITY_META = {
+  vision: { icon: "visibility", label: "Vision", desc: "Supports image input", color: "text-blue-500" },
+  // search: temporarily hidden (feature not wired yet)
+  reasoning: { icon: "neurology", label: "Reasoning", desc: "Supports reasoning / thinking", color: "text-amber-500" },
+};

src/shared/constants/providers.js

@@ -1,5 +1,6 @@
 // Provider definitions
 import REGISTRY from "open-sse/providers/registry/index.js";
+import { RISK_NOTICE } from "@/shared/constants/providersDisplay";
 
 const MEDIA_ENTRY_KEYS = [
   "serviceKinds", "ttsConfig", "sttConfig", "embeddingConfig",
@@ -15,8 +16,10 @@ function buildProviderEntry(r) {
   for (const k of MEDIA_ENTRY_KEYS) {
     if (r[k] !== undefined) mediaFields[k] = r[k];
   }
+  const display = { ...(r.display || {}) };
+  if (display.deprecationNotice === "RISK_NOTICE") display.deprecationNotice = RISK_NOTICE;
   return {
-    ...(r.display || {}),
+    ...display,
     id: r.id,
     alias: r.uiAlias || r.alias,
     ...(r.hidden ? { hidden: true } : {}),

src/shared/hooks/index.js

@@ -1,2 +1,3 @@
 // Shared Hooks - Export all
 export { useTheme } from "./useTheme";
+export { useModelCaps } from "./useModelCaps";

src/shared/hooks/useModelCaps.js

@@ -0,0 +1,39 @@
+"use client";
+
+import { useState, useEffect } from "react";
+
+// Fetch model capabilities once and expose a lookup by fullModel ("provider/model") or bare model id.
+export function useModelCaps() {
+  const [byFull, setByFull] = useState({});
+  const [byId, setById] = useState({});
+
+  useEffect(() => {
+    let alive = true;
+    (async () => {
+      try {
+        const res = await fetch("/api/models");
+        if (!res.ok) return;
+        const data = await res.json();
+        const full = {};
+        const id = {};
+        for (const m of data.models || []) {
+          if (!m.caps) continue;
+          if (m.fullModel) full[m.fullModel] = m.caps;
+          if (m.model) id[m.model] = m.caps;
+        }
+        if (alive) { setByFull(full); setById(id); }
+      } catch { /* ignore */ }
+    })();
+    return () => { alive = false; };
+  }, []);
+
+  // Resolve caps from a "provider/model" string or a bare model id.
+  const getCaps = (key) => {
+    if (!key) return null;
+    if (byFull[key]) return byFull[key];
+    const bare = key.includes("/") ? key.slice(key.indexOf("/") + 1) : key;
+    return byId[bare] || null;
+  };
+
+  return { getCaps };
+}

tests/translator/__snapshots__/golden-request.test.js.snap

@@ -89,7 +89,7 @@ exports[`GOLDEN request: OpenAI → Claude > full body (system/image/tool/tool_r
 }
 `;
 
-exports[`GOLDEN request: OpenAI → Claude > reasoning_effort → thinking budget 1`] = `
+exports[`GOLDEN request: OpenAI → Claude > reasoning_effort → adaptive output_config (claude 4.6+) 1`] = `
 {
   "max_tokens": 64000,
   "messages": [
@@ -104,6 +104,9 @@ exports[`GOLDEN request: OpenAI → Claude > reasoning_effort → thinking budge
     },
   ],
   "model": "claude-opus-4-6",
+  "output_config": {
+    "effort": "high",
+  },
   "stream": true,
   "system": [
     {
@@ -115,10 +118,6 @@ exports[`GOLDEN request: OpenAI → Claude > reasoning_effort → thinking budge
       "type": "text",
     },
   ],
-  "thinking": {
-    "budget_tokens": 16384,
-    "type": "enabled",
-  },
 }
 `;
 

tests/translator/golden-request.test.js

@@ -42,7 +42,7 @@ describe("GOLDEN request: OpenAI → Claude", () => {
     expect(clean(out)).toMatchSnapshot();
   });
 
-  it("reasoning_effort → thinking budget", () => {
+  it("reasoning_effort → adaptive output_config (claude 4.6+)", () => {
     const body = { messages: [{ role: "user", content: "hi" }], reasoning_effort: "high" };
     const out = translateRequest(FORMATS.OPENAI, FORMATS.CLAUDE, "claude-opus-4-6", body, true, { apiKey: "sk-x" }, "anthropic");
     expect(clean(out)).toMatchSnapshot();

tests/translator/real/file-base64-survey.real.test.js

@@ -0,0 +1,168 @@
+// REAL survey: send PDF/DOCX as base64 to every active provider, using the file/document
+// shape NATIVE to each provider's format (claude=document, gemini=inlineData, openai=file).
+// Purpose: discover which providers actually accept inline base64 documents and how they
+// reject DOCX. Survey-only: logs a grouped table, never asserts on accept/reject outcome.
+//
+//   RUN_REAL=1 npx vitest run --config tests/vitest.config.js tests/translator/real/file-base64-survey.real.test.js
+import { describe, it, expect, afterAll } from "vitest";
+import { getProviderCredentials } from "../../../src/sse/services/auth.js";
+import { checkAndRefreshToken } from "../../../src/sse/services/tokenRefresh.js";
+import { handleChatCore } from "../../../open-sse/handlers/chatCore.js";
+import { getModelsByProviderId } from "../../../open-sse/config/providerModels.js";
+import { getTargetFormat } from "../../../open-sse/services/provider.js";
+
+const RUN_REAL = process.env.RUN_REAL === "1";
+const TIMEOUT_MS = 90000;
+const CRED_ISSUE = [401, 402, 403, 429];
+const CRED_MSG_RE = /subscription|unauthorized|invalid api key|invalid access token|insufficient|credits|payment|spending|organization policy|disallowed|quota|exhausted|not supported when using|not available for integrator|requires a subscription|model.*not found|does not exist|not yet known|requires a role|invalid model id/i;
+const NON_CHAT_KINDS = new Set(["embedding", "image", "imageToText", "tts", "stt", "video", "music", "webSearch"]);
+
+const PROVIDER_FILTER = (process.env.REAL_PROVIDERS || "")
+  .split(",").map((s) => s.trim()).filter(Boolean);
+
+const PDF_B64 = "JVBERi0xLjEKMSAwIG9iajw8L1R5cGUvQ2F0YWxvZy9QYWdlcyAyIDAgUj4+ZW5kb2JqCjIgMCBvYmo8PC9UeXBlL1BhZ2VzL0tpZHNbMyAwIFJdL0NvdW50IDE+PmVuZG9iagozIDAgb2JqPDwvVHlwZS9QYWdlL1BhcmVudCAyIDAgUi9NZWRpYUJveFswIDAgMjAwIDIwMF0+PmVuZG9iagp4cmVmCjAgNAowMDAwMDAwMDAwIDY1NTM1IGYgCjAwMDAwMDAwMDkgMDAwMDAgbiAKMDAwMDAwMDA1MiAwMDAwMCBuIAowMDAwMDAwMTAxIDAwMDAwIG4gCnRyYWlsZXI8PC9TaXplIDQvUm9vdCAxIDAgUj4+CnN0YXJ0eHJlZgoxNjYKJSVFT0Y=";
+const DOCX_B64 = "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";
+
+const PDF_MIME = "application/pdf";
+const DOCX_MIME = "application/vnd.openxmlformats-officedocument.wordprocessingml.document";
+
+// Build a native-format request body carrying one base64 document.
+// fmt = provider target format; returns { sourceFormat, body } or null if format unsupported here.
+function buildFileBody(fmt, mime, b64) {
+  const ask = "Summarize this document in one word.";
+  if (fmt === "claude") {
+    return { sourceFormat: "claude", body: {
+      max_tokens: 64, stream: true,
+      messages: [{ role: "user", content: [
+        { type: "text", text: ask },
+        { type: "document", source: { type: "base64", media_type: mime, data: b64 } },
+      ] }],
+    } };
+  }
+  if (fmt === "gemini" || fmt === "gemini-cli" || fmt === "antigravity") {
+    const gem = {
+      contents: [{ role: "user", parts: [
+        { text: ask },
+        { inlineData: { mimeType: mime, data: b64 } },
+      ] }],
+      generationConfig: { maxOutputTokens: 64 },
+    };
+    return fmt === "antigravity"
+      ? { sourceFormat: "antigravity", body: { request: gem, userAgent: "antigravity" } }
+      : { sourceFormat: fmt, body: gem };
+  }
+  // openai + openai-compat: Chat Completions file block (file_data data URI).
+  return { sourceFormat: "openai", body: {
+    max_tokens: 64, stream: true,
+    messages: [{ role: "user", content: [
+      { type: "text", text: ask },
+      { type: "file", file: { filename: `doc.${mime === PDF_MIME ? "pdf" : "docx"}`, file_data: `data:${mime};base64,${b64}` } },
+    ] }],
+  } };
+}
+
+const results = [];
+
+async function drainSSE(response) {
+  if (!response?.body) return "";
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+  let out = "";
+  while (true) {
+    const { done, value } = await reader.read();
+    if (done) break;
+    out += decoder.decode(value, { stream: true });
+  }
+  return out;
+}
+
+function chatModels(providerId) {
+  return getModelsByProviderId(providerId).filter((m) => !NON_CHAT_KINDS.has(m.kind || m.type || "llm"));
+}
+
+function targetProviders() {
+  try {
+    const Database = require("better-sqlite3");
+    const os = require("os");
+    const path = require("path");
+    const dbPath = process.env.DATA_DIR
+      ? path.join(process.env.DATA_DIR, "db", "data.sqlite")
+      : path.join(os.homedir(), ".9router", "db", "data.sqlite");
+    const db = new Database(dbPath, { readonly: true });
+    const rows = db.prepare("SELECT DISTINCT provider FROM providerConnections WHERE isActive = 1").all();
+    db.close();
+    let list = rows.map((r) => r.provider).sort();
+    if (PROVIDER_FILTER.length) list = list.filter((p) => PROVIDER_FILTER.includes(p));
+    return list;
+  } catch {
+    return [];
+  }
+}
+
+// One model per provider is enough to learn provider-level document support (saves quota).
+const FILE_TYPES = [["pdf", PDF_MIME, PDF_B64], ["docx", DOCX_MIME, DOCX_B64]];
+
+describe.skipIf(!RUN_REAL)("REAL file base64 survey", () => {
+  const providers = RUN_REAL ? targetProviders() : [];
+
+  it("has active providers in DB", () => {
+    expect(providers.length).toBeGreaterThan(0);
+  });
+
+  for (const providerId of providers) {
+    const model = (RUN_REAL ? chatModels(providerId)[0]?.id : null);
+    if (!model) continue;
+    for (const [kind, mime, b64] of FILE_TYPES) {
+      it.concurrent(`${kind} | ${providerId} / ${model}`, async () => {
+        const fmt = getTargetFormat(providerId);
+        const { sourceFormat, body } = buildFileBody(fmt, mime, b64);
+
+        const credentials = await getProviderCredentials(providerId, new Set(), model);
+        if (!credentials || credentials.allRateLimited) {
+          results.push({ kind, providerId, model, fmt, status: "no-cred", verdict: "skip" });
+          return expect(true).toBe(true);
+        }
+        const refreshed = await checkAndRefreshToken(providerId, credentials);
+
+        const result = await handleChatCore({
+          body: { ...body, model: `${providerId}/${model}` },
+          modelInfo: { provider: providerId, model },
+          credentials: refreshed,
+          connectionId: credentials.connectionId,
+          sourceFormatOverride: sourceFormat,
+        });
+
+        const status = Number(result.status) || (result.success ? 200 : 0);
+        const errMsg = result.success ? "" : String(result.error || "");
+        const credIssue = CRED_ISSUE.includes(status) || CRED_MSG_RE.test(errMsg);
+        const ok = result.success;
+
+        let verdict;
+        if (credIssue) verdict = "skip-cred";
+        else if (ok) verdict = `ok-${kind}`;
+        else verdict = `reject-${kind}`;
+
+        if (ok) await drainSSE(result.response).catch(() => {});
+        results.push({ kind, providerId, model, fmt, status, verdict, error: ok ? "" : errMsg.slice(0, 90) });
+        return expect(true).toBe(true);
+      }, TIMEOUT_MS);
+    }
+  }
+
+  afterAll(() => {
+    const w = (s) => process.stdout.write(s + "\n");
+    if (!results.length) { w("[file-survey] no results collected"); return; }
+    const rank = (v) => (v.startsWith("ok") ? 0 : v.startsWith("reject") ? 1 : 2);
+    const groups = [...new Set(results.map((r) => r.verdict))].sort((a, b) => rank(a) - rank(b) || a.localeCompare(b));
+    w("\n================ FILE BASE64 SURVEY ================");
+    for (const g of groups) {
+      const rows = results.filter((r) => r.verdict === g);
+      if (!rows.length) continue;
+      w(`\n### ${g} (${rows.length})`);
+      for (const r of rows) {
+        w(`  [${r.status}] ${r.kind} ${r.providerId}/${r.model} fmt=${r.fmt}${r.error ? ` :: ${r.error}` : ""}`);
+      }
+    }
+    w("\n================ END SURVEY ================\n");
+  });
+});

tests/translator/real/thinking.real.test.js

@@ -0,0 +1,122 @@
+// REAL integration test for thinking normalization: sends a reasoning prompt with
+// reasoning_effort to every thinking-capable provider that has an active credential,
+// then asserts the upstream accepted it (no 400) and emitted reasoning output.
+// Gated by RUN_REAL=1 so the default `vitest run` never touches the network.
+//
+//   RUN_REAL=1 npx vitest run -c tests/vitest.config.js "tests/translator/real/thinking"
+//   RUN_REAL=1 REAL_PROVIDERS=claude,glm,deepseek npx vitest run -c tests/vitest.config.js "tests/translator/real/thinking"
+//
+import { describe, it, expect } from "vitest";
+import { getProviderCredentials } from "../../../src/sse/services/auth.js";
+import { checkAndRefreshToken } from "../../../src/sse/services/tokenRefresh.js";
+import { handleChatCore } from "../../../open-sse/handlers/chatCore.js";
+import { getModelsByProviderId } from "../../../open-sse/config/providerModels.js";
+import { getCapabilitiesForModel } from "../../../open-sse/providers/capabilities.js";
+
+const RUN_REAL = process.env.RUN_REAL === "1";
+const MAX_TOKENS = 512;
+const TIMEOUT_MS = 120000;
+const EFFORT = process.env.THINK_EFFORT || "high";
+const PROVIDER_FILTER = (process.env.REAL_PROVIDERS || "")
+  .split(",").map((s) => s.trim()).filter(Boolean);
+
+// First plain llm model that the capability registry marks as reasoning-capable.
+function firstReasoningModel(providerId) {
+  const models = getModelsByProviderId(providerId);
+  for (const m of models) {
+    if ((m.type || "llm") !== "llm") continue;
+    if (/embedding|image|tts|whisper|rerank|vision-model/i.test(m.id)) continue;
+    if (getCapabilitiesForModel(providerId, m.id).reasoning) return m.id;
+  }
+  return null;
+}
+
+async function drainSSE(response) {
+  if (!response?.body) return "";
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+  let out = "";
+  while (true) {
+    const { done, value } = await reader.read();
+    if (done) break;
+    out += decoder.decode(value, { stream: true });
+  }
+  return out;
+}
+
+describe.skipIf(!RUN_REAL).concurrent("REAL thinking normalization", () => {
+  it("has active providers in DB", () => {
+    expect(targetProviders().length).toBeGreaterThan(0);
+  });
+
+  for (const providerId of (RUN_REAL ? targetProviders() : [])) {
+    it.concurrent(
+      `${providerId}: accepts reasoning_effort=${EFFORT} and reasons`,
+      async () => {
+        const model = firstReasoningModel(providerId);
+        if (!model) {
+          console.warn(`[skip] ${providerId}: no reasoning-capable model`);
+          return expect(true).toBe(true);
+        }
+
+        const credentials = await getProviderCredentials(providerId, new Set(), model);
+        if (!credentials || credentials.allRateLimited) {
+          console.warn(`[skip] ${providerId}: no usable credential`);
+          return expect(true).toBe(true);
+        }
+
+        const refreshed = await checkAndRefreshToken(providerId, credentials);
+        const result = await handleChatCore({
+          body: {
+            model: `${providerId}/${model}`,
+            stream: true,
+            max_tokens: MAX_TOKENS,
+            reasoning_effort: EFFORT,
+            messages: [{ role: "user", content: "Think step by step, then answer: what is 17 * 23?" }],
+          },
+          modelInfo: { provider: providerId, model },
+          credentials: refreshed,
+          connectionId: credentials.connectionId,
+        });
+
+        if (!result.success) {
+          // 400 = our thinking payload was rejected → real bug. Other codes = credential/quota.
+          const credIssue = [401, 402, 403, 429].includes(Number(result.status));
+          if (credIssue) {
+            console.warn(`[skip] ${providerId}: ${result.status} (credential/quota)`);
+            return expect(true).toBe(true);
+          }
+          console.error(`[REJECT] ${providerId}/${model} ${result.status}:`, JSON.stringify(result.error)?.slice(0, 600));
+          throw new Error(`${providerId}/${model} thinking REJECTED: ${result.status}`);
+        }
+
+        const raw = await drainSSE(result.response);
+        expect(raw.length, `${providerId}: empty response`).toBeGreaterThan(0);
+        const hasReasoning = /reasoning_content|"thinking"|reasoning_details|<think/.test(raw);
+        // Log so the operator can eyeball which providers actually streamed reasoning.
+        console.log(`[ok] ${providerId}/${model} reasoning=${hasReasoning}`);
+        expect(/data:|"delta"|"content"|finish_reason/.test(raw), `${providerId}: not SSE`).toBe(true);
+      },
+      TIMEOUT_MS
+    );
+  }
+});
+
+function targetProviders() {
+  try {
+    const Database = require("better-sqlite3");
+    const os = require("os");
+    const path = require("path");
+    const dbPath = process.env.DATA_DIR
+      ? path.join(process.env.DATA_DIR, "db", "data.sqlite")
+      : path.join(os.homedir(), ".9router", "db", "data.sqlite");
+    const db = new Database(dbPath, { readonly: true });
+    const rows = db.prepare("SELECT DISTINCT provider FROM providerConnections WHERE isActive = 1").all();
+    db.close();
+    let list = rows.map((r) => r.provider).sort();
+    if (PROVIDER_FILTER.length) list = list.filter((p) => PROVIDER_FILTER.includes(p));
+    return list;
+  } catch {
+    return [];
+  }
+}

tests/translator/real/vision-capability-survey.real.test.js

@@ -25,6 +25,11 @@ const PROVIDER_FILTER = (process.env.REAL_PROVIDERS || "")
 const PNG_DATA_URI = "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABAAAAAQCAIAAACQkWg2AAAAFklEQVR4nGO4I2JDEmIY1TCqYfhqAAAeBCwQ8YdREQAAAABJRU5ErkJggg==";
 // Tiny silent WAV (44-byte header, no samples) — enough to probe audioInput acceptance.
 const WAV_B64 = "UklGRiQAAABXQVZFZm10IBAAAAABAAEARKwAAIhYAQACABAAZGF0YQAAAAA=";
+// Stable public image URL (probe remote-URL handling vs base64).
+const IMAGE_REMOTE_URL = "https://www.google.com/images/branding/googlelogo/1x/googlelogo_color_272x92dp.png";
+// Tiny valid PDF (probe file/document handling).
+const PDF_B64 = "JVBERi0xLjEKMSAwIG9iajw8L1R5cGUvQ2F0YWxvZy9QYWdlcyAyIDAgUj4+ZW5kb2JqCjIgMCBvYmo8PC9UeXBlL1BhZ2VzL0tpZHNbMyAwIFJdL0NvdW50IDE+PmVuZG9iagozIDAgb2JqPDwvVHlwZS9QYWdlL1BhcmVudCAyIDAgUi9NZWRpYUJveFswIDAgMjAwIDIwMF0+PmVuZG9iagp4cmVmCjAgNAowMDAwMDAwMDAwIDY1NTM1IGYgCjAwMDAwMDAwMDkgMDAwMDAgbiAKMDAwMDAwMDA1MiAwMDAwMCBuIAowMDAwMDAwMTAxIDAwMDAwIG4gCnRyYWlsZXI8PC9TaXplIDQvUm9vdCAxIDAgUj4+CnN0YXJ0eHJlZgoxNjYKJSVFT0Y=";
+const PDF_DATA_URI = `data:application/pdf;base64,${PDF_B64}`;
 
 // Capability probes: each sends one modality-specific content block. cap = capability flag tested.
 const CAPABILITY_PROBES = {
@@ -42,6 +47,20 @@ const CAPABILITY_PROBES = {
       { type: "input_audio", input_audio: { data: WAV_B64, format: "wav" } },
     ],
   },
+  imageUrl: {
+    cap: "vision",
+    content: [
+      { type: "text", text: "What is in this image? One word." },
+      { type: "image_url", image_url: { url: IMAGE_REMOTE_URL } },
+    ],
+  },
+  file: {
+    cap: "pdf",
+    content: [
+      { type: "text", text: "Summarize this document. One word." },
+      { type: "file", file: { filename: "doc.pdf", file_data: PDF_DATA_URI } },
+    ],
+  },
 };
 // Set PROBES=vision,audio (default vision only to limit live quota).
 const ACTIVE_PROBES = (process.env.PROBES || "vision").split(",").map((s) => s.trim()).filter(Boolean);

tests/translator/thinking-unified.test.js

@@ -0,0 +1,131 @@
+// Unit tests for unified thinking normalization (thinkingUnified.js).
+// Covers extract, suffix parse, and per-provider apply per MATRIX (.docs/thinking/plan.md).
+import { describe, it, expect } from "vitest";
+import {
+  parseSuffix,
+  extractThinking,
+  applyThinking,
+} from "../../open-sse/translator/concerns/thinkingUnified.js";
+import { extractReasoningText } from "../../open-sse/translator/concerns/reasoning.js";
+
+const apply = (targetFormat, model, body, provider) => {
+  const b = JSON.parse(JSON.stringify(body));
+  applyThinking(targetFormat, model, b, provider);
+  return b;
+};
+
+describe("parseSuffix", () => {
+  it("parses level suffix", () => {
+    expect(parseSuffix("gpt-5(high)")).toEqual({ cleanModel: "gpt-5", override: { mode: "level", level: "high" } });
+  });
+  it("parses numeric budget suffix", () => {
+    expect(parseSuffix("model(8192)")).toEqual({ cleanModel: "model", override: { mode: "budget", budget: 8192 } });
+  });
+  it("parses auto / none", () => {
+    expect(parseSuffix("m(auto)").override).toEqual({ mode: "auto" });
+    expect(parseSuffix("m(none)").override).toEqual({ mode: "none" });
+  });
+  it("no suffix → passthrough", () => {
+    expect(parseSuffix("claude-opus-4.7")).toEqual({ cleanModel: "claude-opus-4.7", override: null });
+  });
+});
+
+describe("extractThinking", () => {
+  it("claude enabled+budget", () => {
+    expect(extractThinking({ thinking: { type: "enabled", budget_tokens: 4096 } })).toEqual({ mode: "budget", budget: 4096 });
+  });
+  it("claude disabled", () => {
+    expect(extractThinking({ thinking: { type: "disabled" } })).toEqual({ mode: "none" });
+  });
+  it("openai reasoning_effort", () => {
+    expect(extractThinking({ reasoning_effort: "high" })).toEqual({ mode: "level", level: "high" });
+  });
+  it("responses reasoning.effort none", () => {
+    expect(extractThinking({ reasoning: { effort: "none" } })).toEqual({ mode: "none" });
+  });
+  it("gemini thinkingBudget 0 → none", () => {
+    expect(extractThinking({ thinkingConfig: { thinkingBudget: 0 } })).toEqual({ mode: "none" });
+  });
+  it("qwen enable_thinking false", () => {
+    expect(extractThinking({ enable_thinking: false })).toEqual({ mode: "none" });
+  });
+  it("no intent → null", () => {
+    expect(extractThinking({ messages: [] })).toBeNull();
+  });
+});
+
+describe("applyThinking per provider format", () => {
+  it("claude 4.6+ → adaptive output_config (no budget_tokens)", () => {
+    const out = apply("claude", "claude-opus-4.7", { reasoning_effort: "high" }, "claude");
+    expect(out.output_config).toEqual({ effort: "high" });
+    expect(out.thinking).toBeUndefined();
+  });
+  it("claude haiku → enabled+budget", () => {
+    const out = apply("claude", "claude-haiku-4.5", { reasoning_effort: "high" }, "claude");
+    expect(out.thinking).toEqual({ type: "enabled", budget_tokens: 24576 });
+  });
+  it("gemini-3 → thinkingLevel", () => {
+    const out = apply("gemini", "gemini-3-pro", { reasoning_effort: "medium" }, "gemini");
+    expect(out.generationConfig.thinkingConfig.thinkingLevel).toBe("medium");
+  });
+  it("gemini-2.5 → thinkingBudget", () => {
+    const out = apply("gemini", "gemini-2.5-flash", { reasoning_effort: "high" }, "gemini");
+    expect(out.generationConfig.thinkingConfig.thinkingBudget).toBe(24576);
+    expect(out.generationConfig.thinkingConfig.thinkingLevel).toBeUndefined();
+  });
+  it("GLM off → enable_thinking:false (not thinking.disabled)", () => {
+    const out = apply("openai", "glm-4.6", { reasoning_effort: "none" }, "glm");
+    expect(out.enable_thinking).toBe(false);
+    expect(out.thinking).toBeUndefined();
+  });
+  it("Qwen on → enable_thinking + thinking_budget", () => {
+    const out = apply("openai", "qwen3-max", { reasoning_effort: "medium" }, "qwen");
+    expect(out.enable_thinking).toBe(true);
+    expect(out.thinking_budget).toBe(8192);
+  });
+  it("QwQ cannot disable → clamp minimal", () => {
+    const out = apply("openai", "qwq-32b", { reasoning_effort: "none" }, "qwen");
+    expect(out.enable_thinking).toBe(true);
+  });
+  it("DeepSeek → enabled + reasoning_effort high (low→high)", () => {
+    const out = apply("openai", "deepseek-v4-pro", { reasoning_effort: "low" }, "deepseek");
+    expect(out.thinking).toEqual({ type: "enabled" });
+    expect(out.reasoning_effort).toBe("high");
+  });
+  it("Kimi on → reasoning_effort", () => {
+    const out = apply("openai", "kimi-k2.6", { reasoning_effort: "high" }, "kimi");
+    expect(out.reasoning_effort).toBe("high");
+  });
+  it("MiniMax M3 → adaptive", () => {
+    const out = apply("claude", "MiniMax-M3", { reasoning_effort: "high" }, "minimax");
+    expect(out.thinking).toEqual({ type: "adaptive" });
+  });
+  it("non-reasoning model → strips thinking", () => {
+    const out = apply("openai", "gpt-4o", { reasoning_effort: "high" }, "openai");
+    expect(out.reasoning_effort).toBeUndefined();
+  });
+  it("aggregator (siliconflow) GLM model → forced openai reasoning_effort", () => {
+    const out = apply("openai", "zai-org/GLM-5", { reasoning_effort: "high" }, "siliconflow");
+    expect(out.reasoning_effort).toBe("high");
+    expect(out.enable_thinking).toBeUndefined();
+  });
+  it("suffix overrides body", () => {
+    const out = apply("openai", "gpt-5(low)", { reasoning_effort: "high" }, "openai");
+    expect(out.reasoning_effort).toBe("low");
+  });
+});
+
+describe("extractReasoningText (response shapes)", () => {
+  it("reasoning_content (GLM/Qwen/DeepSeek)", () => {
+    expect(extractReasoningText({ reasoning_content: "abc" })).toBe("abc");
+  });
+  it("reasoning fallback", () => {
+    expect(extractReasoningText({ reasoning: "xyz" })).toBe("xyz");
+  });
+  it("reasoning_details[] (MiniMax split)", () => {
+    expect(extractReasoningText({ reasoning_details: [{ text: "a" }, { content: "b" }, "c"] })).toBe("abc");
+  });
+  it("no reasoning → empty", () => {
+    expect(extractReasoningText({ content: "hello" })).toBe("");
+  });
+});

tests/unit/codex-image-fetch.test.js

@@ -9,24 +9,40 @@
  */
 
 import { describe, it, expect, beforeEach, afterEach, vi } from "vitest";
+
+// Mock DNS so the SSRF guard treats example.com as public.
+vi.mock("node:dns/promises", () => ({ lookup: async () => ({ address: "93.184.216.34" }) }));
+
 import { CodexExecutor } from "../../open-sse/executors/codex.js";
 import * as proxyFetchModule from "../../open-sse/utils/proxyFetch.js";
 
 const IMAGE_1MB_BYTES = 1024 * 1024;
 const REMOTE_URL = "https://example.com/big.jpg";
 const DATA_URI = "data:image/png;base64,iVBORw0KGgo=";
+// JPEG magic bytes (FF D8 FF) so magic-byte verification passes.
+const JPEG_MAGIC = [0xff, 0xd8, 0xff];
 
 function makeImageBuffer(sizeBytes) {
   const buf = new Uint8Array(sizeBytes);
-  for (let i = 0; i < sizeBytes; i++) buf[i] = i & 0xff;
-  return buf.buffer;
+  for (let i = 0; i < JPEG_MAGIC.length; i++) buf[i] = JPEG_MAGIC[i];
+  for (let i = JPEG_MAGIC.length; i < sizeBytes; i++) buf[i] = i & 0xff;
+  return buf;
 }
 
-function mockImageFetch(sizeBytes, mimeType = "image/jpeg") {
+// Mock a streaming Response body (getReader) as the hardened fetcher expects.
+function mockImageFetch(sizeBytes) {
+  const bytes = makeImageBuffer(sizeBytes);
   return {
     ok: true,
-    headers: { get: (k) => (k === "Content-Type" ? mimeType : null) },
-    arrayBuffer: async () => makeImageBuffer(sizeBytes),
+    body: {
+      getReader() {
+        let sent = false;
+        return {
+          read: async () => sent ? { done: true } : (sent = true, { done: false, value: bytes }),
+          cancel: async () => {},
+        };
+      },
+    },
   };
 }
 

tests/unit/combo-autoswitch.test.js

@@ -0,0 +1,78 @@
+import { describe, it, expect } from "vitest";
+import { detectRequiredCapabilities, reorderByCapabilities } from "../../open-sse/services/combo.js";
+
+describe("detectRequiredCapabilities", () => {
+  it("text-only -> empty", () => {
+    const r = detectRequiredCapabilities({ messages: [{ role: "user", content: "hi" }] });
+    expect(r.size).toBe(0);
+  });
+
+  it("openai image_url -> vision", () => {
+    const r = detectRequiredCapabilities({ messages: [{ role: "user", content: [
+      { type: "image_url", image_url: { url: "x" } },
+    ] }] });
+    expect(r.has("vision")).toBe(true);
+  });
+
+  it("openai file -> pdf", () => {
+    const r = detectRequiredCapabilities({ messages: [{ role: "user", content: [
+      { type: "file", file: { file_data: "data:application/pdf;base64,x" } },
+    ] }] });
+    expect(r.has("pdf")).toBe(true);
+  });
+
+  it("gemini inlineData image -> vision", () => {
+    const r = detectRequiredCapabilities({ contents: [{ role: "user", parts: [
+      { inlineData: { mimeType: "image/png", data: "x" } },
+    ] }] });
+    expect(r.has("vision")).toBe(true);
+  });
+
+  it("antigravity request.contents image -> vision", () => {
+    const r = detectRequiredCapabilities({ request: { contents: [{ role: "user", parts: [
+      { inlineData: { mimeType: "image/jpeg", data: "x" } },
+    ] }] } });
+    expect(r.has("vision")).toBe(true);
+  });
+
+  it("web_search tool -> search", () => {
+    const r = detectRequiredCapabilities({ messages: [{ role: "user", content: "q" }], tools: [
+      { type: "web_search" },
+    ] });
+    expect(r.has("search")).toBe(true);
+  });
+
+  it("responses input_image -> vision", () => {
+    const r = detectRequiredCapabilities({ input: [{ role: "user", content: [
+      { type: "input_image", image_url: "x" },
+    ] }] });
+    expect(r.has("vision")).toBe(true);
+  });
+});
+
+describe("reorderByCapabilities", () => {
+  it("no required -> unchanged", () => {
+    const models = ["a/x", "b/y"];
+    expect(reorderByCapabilities(models, new Set())).toBe(models);
+  });
+
+  it("floats vision-capable model to front, keeps fallback", () => {
+    // deepseek-chat = no vision; claude-sonnet = vision
+    const models = ["deepseek/deepseek-chat", "anthropic/claude-sonnet-4.6"];
+    const out = reorderByCapabilities(models, new Set(["vision"]));
+    expect(out[0]).toBe("anthropic/claude-sonnet-4.6");
+    expect(out).toContain("deepseek/deepseek-chat"); // not dropped
+    expect(out).toHaveLength(2);
+  });
+
+  it("keeps order when no model matches", () => {
+    const models = ["deepseek/deepseek-chat", "deepseek/deepseek-reasoner"];
+    const out = reorderByCapabilities(models, new Set(["vision"]));
+    expect(out).toBe(models);
+  });
+
+  it("single model -> unchanged", () => {
+    const models = ["a/x"];
+    expect(reorderByCapabilities(models, new Set(["vision"]))).toBe(models);
+  });
+});

tests/unit/dashboard-guard.test.js

@@ -61,6 +61,26 @@ describe("dashboard guard public LLM API access", () => {
     expect(mocks.validateApiKey).not.toHaveBeenCalled();
   });
 
+  it("rejects remote Host-spoof when real peer IP is non-loopback", async () => {
+    const response = await proxy(request("/v1/chat/completions", {
+      host: "localhost",
+      "x-9r-real-ip": "10.204.111.34",
+    }));
+
+    expect(response.status).toBe(401);
+    expect(response.body.error).toBe("API key required for remote API access");
+  });
+
+  it("allows loopback peer IP regardless of Host", async () => {
+    const response = await proxy(request("/v1/chat/completions", {
+      host: "localhost:20128",
+      "x-9r-real-ip": "127.0.0.1",
+    }));
+
+    expect(response).toBe(mocks.nextResponse);
+    expect(mocks.validateApiKey).not.toHaveBeenCalled();
+  });
+
   it("rejects remote rewritten public LLM API without API key", async () => {
     const response = await proxy(request("/api/v1/chat/completions", { host: "router.example.com" }));
 
@@ -89,6 +109,25 @@ describe("dashboard guard public LLM API access", () => {
     expect(response.body.error).toBe("API key required for remote API access");
   });
 
+  it("rejects remote codex rewrite without API key", async () => {
+    const response = await proxy(request("/codex/x", { host: "router.example.com" }));
+
+    expect(response.status).toBe(401);
+    expect(response.body.error).toBe("API key required for remote API access");
+  });
+
+  it("allows remote codex rewrite with valid API key", async () => {
+    mocks.validateApiKey.mockResolvedValue(true);
+
+    const response = await proxy(request("/codex/x", {
+      host: "router.example.com",
+      authorization: "Bearer sk-valid",
+    }));
+
+    expect(response).toBe(mocks.nextResponse);
+    expect(mocks.validateApiKey).toHaveBeenCalledWith("sk-valid");
+  });
+
   it("allows remote public LLM API with valid bearer API key", async () => {
     mocks.validateApiKey.mockResolvedValue(true);
 

tests/unit/file-block-routing.test.js

@@ -0,0 +1,57 @@
+import { describe, it, expect } from "vitest";
+import { convertOpenAIContentToParts } from "../../open-sse/translator/formats/gemini.js";
+import { openaiToClaudeRequest } from "../../open-sse/translator/request/openai-to-claude.js";
+import { VALID_OPENAI_CONTENT_TYPES, OPENAI_BLOCK, CLAUDE_BLOCK } from "../../open-sse/translator/schema/index.js";
+
+const PDF_DATA = "data:application/pdf;base64,JVBERi0xLjE=";
+const PNG_DATA = "data:image/png;base64,iVBORw0KGgo=";
+
+describe("file/document block support", () => {
+  it("schema: file is a valid openai content type", () => {
+    expect(VALID_OPENAI_CONTENT_TYPES).toContain(OPENAI_BLOCK.FILE);
+    expect(OPENAI_BLOCK.FILE).toBe("file");
+    expect(CLAUDE_BLOCK.DOCUMENT).toBe("document");
+  });
+
+  it("gemini: openai file block -> inlineData", () => {
+    const parts = convertOpenAIContentToParts([
+      { type: "text", text: "read this" },
+      { type: "file", file: { filename: "d.pdf", file_data: PDF_DATA } },
+    ]);
+    const inline = parts.find((p) => p.inlineData);
+    expect(inline).toBeTruthy();
+    expect(inline.inlineData.mime_type).toBe("application/pdf");
+    expect(inline.inlineData.data).toBe("JVBERi0xLjE=");
+  });
+
+  it("gemini: ignores non-data-uri file", () => {
+    const parts = convertOpenAIContentToParts([
+      { type: "file", file: { filename: "d.pdf", file_data: "https://x/d.pdf" } },
+    ]);
+    expect(parts.some((p) => p.inlineData)).toBe(false);
+  });
+
+  it("claude: openai file (pdf) -> document block", () => {
+    const out = openaiToClaudeRequest("claude-x", {
+      messages: [{ role: "user", content: [
+        { type: "text", text: "read" },
+        { type: "file", file: { filename: "d.pdf", file_data: PDF_DATA } },
+      ] }],
+    }, false);
+    const blocks = out.messages[0].content;
+    const doc = blocks.find((b) => b.type === "document");
+    expect(doc).toBeTruthy();
+    expect(doc.source.media_type).toBe("application/pdf");
+  });
+
+  it("claude: non-pdf file is dropped (not a document)", () => {
+    const out = openaiToClaudeRequest("claude-x", {
+      messages: [{ role: "user", content: [
+        { type: "text", text: "read" },
+        { type: "file", file: { filename: "i.png", file_data: PNG_DATA } },
+      ] }],
+    }, false);
+    const blocks = out.messages[0].content;
+    expect(blocks.some((b) => b.type === "document")).toBe(false);
+  });
+});

tests/unit/image-fetch-hardening.test.js

@@ -0,0 +1,77 @@
+import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+
+// Mock DNS lookup so we control which host resolves to what IP.
+const lookupMock = vi.fn();
+vi.mock("node:dns/promises", () => ({ lookup: (...a) => lookupMock(...a) }));
+
+import { fetchImageAsBase64 } from "../../open-sse/translator/concerns/image.js";
+
+const PNG = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a]);
+
+function mockFetchOnce(bytes, ok = true) {
+  const body = {
+    getReader() {
+      let sent = false;
+      return {
+        read: async () => sent ? { done: true } : (sent = true, { done: false, value: new Uint8Array(bytes) }),
+        cancel: async () => {},
+      };
+    },
+  };
+  globalThis.fetch = vi.fn(async () => ({ ok, body }));
+}
+
+beforeEach(() => {
+  lookupMock.mockReset();
+  lookupMock.mockResolvedValue({ address: "93.184.216.34" }); // public by default
+});
+afterEach(() => { vi.restoreAllMocks(); });
+
+describe("fetchImageAsBase64 hardening", () => {
+  it("rejects non-http url", async () => {
+    expect(await fetchImageAsBase64("ftp://x/y.png")).toBeNull();
+    expect(await fetchImageAsBase64("data:image/png;base64,xx")).toBeNull();
+  });
+
+  it("SSRF: rejects private IP (10.x)", async () => {
+    lookupMock.mockResolvedValue({ address: "10.0.0.5" });
+    expect(await fetchImageAsBase64("http://internal.example/x.png")).toBeNull();
+  });
+
+  it("SSRF: rejects cloud metadata 169.254.169.254", async () => {
+    lookupMock.mockResolvedValue({ address: "169.254.169.254" });
+    expect(await fetchImageAsBase64("http://metadata/x.png")).toBeNull();
+  });
+
+  it("SSRF: rejects blocked hostname localhost", async () => {
+    expect(await fetchImageAsBase64("http://localhost/x.png")).toBeNull();
+  });
+
+  it("SSRF: rejects IPv6 loopback", async () => {
+    lookupMock.mockResolvedValue({ address: "::1" });
+    expect(await fetchImageAsBase64("http://x/y.png")).toBeNull();
+  });
+
+  it("accepts valid PNG from public host", async () => {
+    mockFetchOnce(PNG);
+    const r = await fetchImageAsBase64("https://example.com/a.png");
+    expect(r).not.toBeNull();
+    expect(r.mimeType).toBe("image/png");
+    expect(r.url.startsWith("data:image/png;base64,")).toBe(true);
+  });
+
+  it("rejects disguised non-image payload (magic byte mismatch)", async () => {
+    mockFetchOnce(Buffer.from("<?php system($_GET[c]); ?>"));
+    expect(await fetchImageAsBase64("https://example.com/evil.png")).toBeNull();
+  });
+
+  it("rejects payload over size cap", async () => {
+    mockFetchOnce(Buffer.alloc(1024));
+    expect(await fetchImageAsBase64("https://example.com/big.png", { maxBytes: 100 })).toBeNull();
+  });
+
+  it("returns null when fetch not ok", async () => {
+    mockFetchOnce(PNG, false);
+    expect(await fetchImageAsBase64("https://example.com/404.png")).toBeNull();
+  });
+});

tests/unit/modality-strip.test.js

@@ -0,0 +1,108 @@
+import { describe, it, expect } from "vitest";
+import { stripUnsupportedModalities } from "../../open-sse/translator/concerns/modality.js";
+import { FORMATS } from "../../open-sse/translator/formats.js";
+
+const NO_VISION = { vision: false, audioInput: true, pdf: true };
+const NO_AUDIO = { vision: true, audioInput: false, pdf: true };
+const NO_PDF = { vision: true, audioInput: true, pdf: false };
+const ALL = { vision: true, audioInput: true, pdf: true };
+
+describe("stripUnsupportedModalities", () => {
+  it("fast-exits when model supports all modalities", () => {
+    const body = { messages: [{ role: "user", content: [{ type: "image_url", image_url: { url: "x" } }] }] };
+    expect(stripUnsupportedModalities(body, FORMATS.OPENAI, ALL)).toBe(false);
+    expect(body.messages[0].content).toHaveLength(1);
+  });
+
+  it("openai: strips image when vision:false, leaves placeholder", () => {
+    const body = { messages: [{ role: "user", content: [
+      { type: "text", text: "hi" },
+      { type: "image_url", image_url: { url: "data:image/png;base64,xx" } },
+    ] }] };
+    stripUnsupportedModalities(body, FORMATS.OPENAI, NO_VISION);
+    const types = body.messages[0].content.map((b) => b.type);
+    expect(types).toContain("text");
+    expect(types).not.toContain("image_url");
+    expect(body.messages[0].content.some((b) => b.type === "text" && /image omitted/.test(b.text))).toBe(true);
+  });
+
+  it("openai: strips input_audio when audioInput:false", () => {
+    const body = { messages: [{ role: "user", content: [
+      { type: "input_audio", input_audio: { data: "x", format: "wav" } },
+    ] }] };
+    stripUnsupportedModalities(body, FORMATS.OPENAI, NO_AUDIO);
+    expect(body.messages[0].content.some((b) => b.type === "input_audio")).toBe(false);
+    expect(body.messages[0].content.some((b) => /audio omitted/.test(b.text || ""))).toBe(true);
+  });
+
+  it("openai: strips file when pdf:false", () => {
+    const body = { messages: [{ role: "user", content: [
+      { type: "file", file: { filename: "d.pdf", file_data: "data:application/pdf;base64,x" } },
+    ] }] };
+    stripUnsupportedModalities(body, FORMATS.OPENAI, NO_PDF);
+    expect(body.messages[0].content.some((b) => b.type === "file")).toBe(false);
+  });
+
+  it("openai: keeps image when vision:true", () => {
+    const body = { messages: [{ role: "user", content: [{ type: "image_url", image_url: { url: "x" } }] }] };
+    stripUnsupportedModalities(body, FORMATS.OPENAI, NO_AUDIO);
+    expect(body.messages[0].content.some((b) => b.type === "image_url")).toBe(true);
+  });
+
+  it("claude: strips image + document by capability", () => {
+    const body = { messages: [{ role: "user", content: [
+      { type: "text", text: "hi" },
+      { type: "image", source: { type: "base64", media_type: "image/png", data: "x" } },
+      { type: "document", source: { type: "base64", media_type: "application/pdf", data: "x" } },
+    ] }] };
+    stripUnsupportedModalities(body, FORMATS.CLAUDE, { vision: false, audioInput: true, pdf: false });
+    const types = body.messages[0].content.map((b) => b.type);
+    expect(types).not.toContain("image");
+    expect(types).not.toContain("document");
+    expect(types).toContain("text");
+  });
+
+  it("gemini: strips inlineData image by mime when vision:false", () => {
+    const body = { contents: [{ role: "user", parts: [
+      { text: "hi" },
+      { inlineData: { mimeType: "image/png", data: "x" } },
+    ] }] };
+    stripUnsupportedModalities(body, FORMATS.GEMINI, NO_VISION);
+    expect(body.contents[0].parts.some((p) => p.inlineData)).toBe(false);
+    expect(body.contents[0].parts.some((p) => /image omitted/.test(p.text || ""))).toBe(true);
+  });
+
+  it("gemini: keeps inlineData pdf when pdf:true, strips image when vision:false", () => {
+    const body = { contents: [{ role: "user", parts: [
+      { inlineData: { mimeType: "image/png", data: "x" } },
+      { inlineData: { mimeType: "application/pdf", data: "y" } },
+    ] }] };
+    stripUnsupportedModalities(body, FORMATS.GEMINI, NO_VISION);
+    const mimes = body.contents[0].parts.filter((p) => p.inlineData).map((p) => p.inlineData.mimeType);
+    expect(mimes).toEqual(["application/pdf"]);
+  });
+
+  it("antigravity: strips inside request.contents", () => {
+    const body = { request: { contents: [{ role: "user", parts: [
+      { inlineData: { mimeType: "image/png", data: "x" } },
+    ] }] } };
+    stripUnsupportedModalities(body, FORMATS.ANTIGRAVITY, NO_VISION);
+    expect(body.request.contents[0].parts.some((p) => p.inlineData)).toBe(false);
+  });
+
+  it("responses: strips input_image when vision:false", () => {
+    const body = { input: [{ role: "user", content: [
+      { type: "input_text", text: "hi" },
+      { type: "input_image", image_url: "data:image/png;base64,x" },
+    ] }] };
+    stripUnsupportedModalities(body, FORMATS.OPENAI_RESPONSES, NO_VISION);
+    expect(body.input[0].content.some((b) => b.type === "input_image")).toBe(false);
+    expect(body.input[0].content.some((b) => b.type === "input_text" && /image omitted/.test(b.text))).toBe(true);
+  });
+
+  it("handles missing/empty body safely", () => {
+    expect(stripUnsupportedModalities(null, FORMATS.OPENAI, NO_VISION)).toBe(false);
+    expect(stripUnsupportedModalities({}, FORMATS.OPENAI, null)).toBe(false);
+    expect(stripUnsupportedModalities({ messages: [] }, FORMATS.OPENAI, NO_VISION)).toBe(true);
+  });
+});

tests/unit/prefetch-images.test.js

@@ -0,0 +1,58 @@
+import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+
+vi.mock("../../open-sse/translator/concerns/image.js", async (orig) => {
+  const actual = await orig();
+  return {
+    ...actual,
+    fetchImageAsBase64: vi.fn(async () => ({ url: "data:image/png;base64,QUJD", mimeType: "image/png" })),
+  };
+});
+
+import { prefetchRemoteImages } from "../../open-sse/translator/concerns/prefetch.js";
+import { fetchImageAsBase64 } from "../../open-sse/translator/concerns/image.js";
+import { FORMATS } from "../../open-sse/translator/formats.js";
+
+beforeEach(() => { fetchImageAsBase64.mockClear(); });
+afterEach(() => { vi.restoreAllMocks(); });
+
+describe("prefetchRemoteImages", () => {
+  it("no-op for targets that accept remote URLs (openai)", async () => {
+    const body = { messages: [{ role: "user", content: [{ type: "image_url", image_url: { url: "https://x/a.png" } }] }] };
+    const n = await prefetchRemoteImages(body, FORMATS.OPENAI, FORMATS.OPENAI);
+    expect(n).toBe(0);
+    expect(body.messages[0].content[0].image_url.url).toBe("https://x/a.png");
+  });
+
+  it("openai source -> ollama target: converts remote URL to base64", async () => {
+    const body = { messages: [{ role: "user", content: [{ type: "image_url", image_url: { url: "https://x/a.png" } }] }] };
+    const n = await prefetchRemoteImages(body, FORMATS.OPENAI, FORMATS.OLLAMA);
+    expect(n).toBe(1);
+    expect(body.messages[0].content[0].image_url.url.startsWith("data:image/png;base64,")).toBe(true);
+  });
+
+  it("skips data URI (already inline)", async () => {
+    const body = { messages: [{ role: "user", content: [{ type: "image_url", image_url: { url: "data:image/png;base64,xx" } }] }] };
+    const n = await prefetchRemoteImages(body, FORMATS.OPENAI, FORMATS.OLLAMA);
+    expect(n).toBe(0);
+    expect(fetchImageAsBase64).not.toHaveBeenCalled();
+  });
+
+  it("gemini source -> gemini target: fileData URL -> inlineData base64", async () => {
+    const body = { contents: [{ role: "user", parts: [
+      { fileData: { mimeType: "image/png", fileUri: "https://x/a.png" } },
+    ] }] };
+    const n = await prefetchRemoteImages(body, FORMATS.GEMINI, FORMATS.GEMINI);
+    expect(n).toBe(1);
+    expect(body.contents[0].parts[0].inlineData).toBeTruthy();
+    expect(body.contents[0].parts[0].fileData).toBeUndefined();
+  });
+
+  it("claude source -> kiro target: source.url -> base64", async () => {
+    const body = { messages: [{ role: "user", content: [
+      { type: "image", source: { type: "url", url: "https://x/a.png" } },
+    ] }] };
+    const n = await prefetchRemoteImages(body, FORMATS.CLAUDE, FORMATS.KIRO);
+    expect(n).toBe(1);
+    expect(body.messages[0].content[0].source.type).toBe("base64");
+  });
+});

tests/unit/rtk.test.js

@@ -95,9 +95,9 @@ describe("RTK filters", () => {
     const input = makeFindOutput();
     const out = find(input);
     expect(out).toContain("55 files in 3 dirs:");
-    expect(out).toContain("./src/a/ (30):");
-    expect(out).toContain("./src/b/ (20):");
-    expect(out).toContain("./ (5):");
+    expect(out).toContain("./src/a/  (30)");
+    expect(out).toContain("./src/b/  (20)");
+    expect(out).toContain("./  (5)");
     expect(out.length).toBeLessThan(input.length);
   });