maxpool 1.5.29 → 1.5.31

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "maxpool",
-  "version": "1.5.29",
+  "version": "1.5.31",
   "description": "Multi-account Claude Code proxy with adaptive, rate-aware load balancing across Claude accounts",
   "type": "module",
   "main": "src/index.js",

package/src/account-manager.js

@@ -1609,6 +1609,13 @@ export class AccountManager {
   _isRequestCompatible(account, profile, requestInfo = {}) {
     if (!this._matchesProfile(account, profile)) return false;
 
+    // Kimi (Moonshot) 400s on images Anthropic/GLM accept ("failed to decode
+    // image") — and a provider 400 is TERMINAL (not retried to another account), so
+    // an image request that fell back to Kimi FAILS the whole request. Keep image
+    // requests off Kimi; they still route to GLM (handles images) + OAuth. If those
+    // are all unavailable the request HOLDS/queues (recoverable) rather than 400ing.
+    if (requestInfo.hasImage && account.provider === 'kimi') return false;
+
     const { incompatible, homeProvider } = this._effectiveIncompatible(requestInfo);
     const policy = this._crossProviderFallbackPolicy();
 

package/src/server.js

@@ -123,7 +123,7 @@ export function createProxyServer(accountManager, config, hooks = {}) {
       if (req.method === 'POST' && req.url === '/v1/oauth/token') {
         const reqId = ++requestCounter;
         const ctx = { account: '(oauth relay)', status: null };
-        const accepted = hooks.onRequestStart?.(reqId, { method: req.method, path: req.url });
+        const accepted = hooks.onRequestStart?.(reqId, { method: req.method, path: req.url, sessionKey: headerValue(req.headers, 'x-maxpool-session') });
         if (accepted === false) {
           res.writeHead(503, {
             'Content-Type': 'application/json',
@@ -156,7 +156,10 @@ export function createProxyServer(accountManager, config, hooks = {}) {
 
       // Track request
       const reqId = ++requestCounter;
-      const accepted = hooks.onRequestStart?.(reqId, { method: req.method, path: req.url });
+      // sessionKey (from the per-terminal x-maxpool-session header) lets the TUI show
+      // how many distinct client sessions the in-flight requests span — so "N in-flight"
+      // isn't misread as "N sessions" (subagents of one terminal share one session key).
+      const accepted = hooks.onRequestStart?.(reqId, { method: req.method, path: req.url, sessionKey: headerValue(req.headers, 'x-maxpool-session') });
       if (accepted === false) {
         res.writeHead(503, {
           'Content-Type': 'application/json',
@@ -1593,6 +1596,9 @@ function describeRequest(req, body) {
     const origin = detectTranscriptOrigin(json);
     if (origin.anthropicIncompatible) info.anthropicIncompatible = true;
     if (origin.homeProvider) info.homeProvider = origin.homeProvider;
+    // Image content (incl. tool_result-nested screenshots) keeps a request off Kimi
+    // (Moonshot 400s on some images GLM/Anthropic accept, and a 400 is terminal).
+    if (containsImageBlock(json.messages)) info.hasImage = true;
   } catch {
     // Non-JSON requests are rare; body size still gives a useful load signal.
   }
@@ -1679,6 +1685,21 @@ function containsThinkingBlock(value) {
   return false;
 }
 
+// True if any message carries an `{type:'image'}` content block at ANY nesting
+// depth — including images nested inside a `tool_result` block (the primary
+// Claude Code image path: a browser/Playwright tool returning a screenshot). A
+// shallow 2-level scan misses those. Used to keep image requests off Kimi, whose
+// decoder 400s on images Anthropic/GLM accept.
+function containsImageBlock(value) {
+  if (!value) return false;
+  if (Array.isArray(value)) return value.some(containsImageBlock);
+  if (typeof value !== 'object') return false;
+  if (value.type === 'image') return true;
+  if (value.content && containsImageBlock(value.content)) return true;
+  if (value.messages && containsImageBlock(value.messages)) return true;
+  return false;
+}
+
 function getMaxpoolProfile(headers) {
   // headerValue() handles the x-teamclaude-* legacy fallback.
   const profile = String(headerValue(headers, 'x-maxpool-profile') || 'claude').trim().toLowerCase();

package/src/tui.js

@@ -982,8 +982,15 @@ export class TUI {
 
     // ── Activity header
     lines.push('');
+    // "N active" reads as "N sessions" — but this is in-flight REQUESTS (Claude Code
+    // fans out many parallel subagent requests per session). Label it "in-flight" and
+    // show the distinct SESSION count (subagents of one terminal share one session key)
+    // so a busy 1-2 sessions firing dozens of requests isn't mistaken for a leak.
     const ac = this.active.size;
-    const acTag = ac > 0 ? `  ${cyan(ac + ' active')}` : '';
+    const sessions = new Set([...this.active.values()].map(r => r.sessionKey).filter(Boolean)).size;
+    const acTag = ac > 0
+      ? `  ${cyan(ac + ' in-flight')}${sessions ? dim(` · ${sessions} session${sessions === 1 ? '' : 's'}`) : ''}`
+      : '';
     const aHdr = ` Activity${acTag} `;
     lines.push(aHdr + dim('─'.repeat(Math.max(1, W - vw(aHdr)))));