@askalf/dario 3.10.1 → 3.10.3

package/dist/cc-template.js

@@ -291,21 +291,23 @@ export function buildCCRequest(clientBody, billingTag, cache1h, identity, opts =
             }
         }
     }
-    // ── Drop trailing empty/assistant turns ──
+    // ── Drop trailing empty turns ──
     // An assistant turn that was thinking-only before the strip above becomes
     // content: []. Forwarding that shape makes Anthropic interpret the request
     // as a prefill ("continue from this assistant text"), which Opus 4.6 under
     // adaptive thinking + the claude-code beta refuses with:
     //   "This model does not support assistant message prefill. The
     //    conversation must end with a user message."
-    // Clients that preserve full thinking in history (OpenClaw, Hermes) hit
-    // this any time a prior turn was interrupted mid-thinking. Drop trailing
-    // assistant/empty turns so the request ends on a user message.
+    // Drop ONLY empty trailing turns. Do not pop trailing assistant turns that
+    // still carry text or tool_use content — v3.10.1 popped any trailing
+    // assistant and that caused a runaway loop in OpenClaw (#37): the client
+    // appended its assistant reply locally, dario stripped it from the next
+    // request, the model regenerated the same reply, dario stripped that, and
+    // the loop never terminated (133 POSTs from a single user prompt).
     while (messages.length > 0) {
         const last = messages[messages.length - 1];
         const contentEmpty = Array.isArray(last.content) && last.content.length === 0;
-        const isTrailingAssistant = last.role === 'assistant';
-        if (contentEmpty || isTrailingAssistant) {
+        if (contentEmpty) {
             messages.pop();
             continue;
         }

package/dist/proxy.js

@@ -1013,14 +1013,37 @@ export async function startProxy(opts = {}) {
                 }
             }
             requestCount++;
-            // Log billing classification on first request or in verbose mode
+            // Log billing classification on first request or in verbose mode.
+            //
+            // Anthropic is inconsistent about returning rate-limit headers:
+            // - Non-200 responses (429, 500, early aborts) often omit them entirely.
+            // - The overage-utilization header is omitted when there is no overage
+            //   bucket configured or when the subscription claim covers the request
+            //   — in that case "overage" is effectively 0%, not unknown.
+            // Pre-fix we logged `overage: ?` on every five_hour request that had no
+            // overage configured, which looked like a broken parser (see #37 log
+            // dump). Fix: treat missing overage header as 0% when the claim is
+            // five_hour / five_hour_fallback (the subscription covered it), and fall
+            // back to `n/a` in the genuinely-unknown case.
             const billingClaim = upstream.headers.get('anthropic-ratelimit-unified-representative-claim');
             const overageUtil = upstream.headers.get('anthropic-ratelimit-unified-overage-utilization');
             if (requestCount === 1 || verbose) {
                 if (billingClaim) {
-                    const overagePct = overageUtil ? `${Math.round(parseFloat(overageUtil) * 100)}%` : '?';
+                    let overagePct;
+                    if (overageUtil !== null) {
+                        overagePct = `${Math.round(parseFloat(overageUtil) * 100)}%`;
+                    }
+                    else if (billingClaim === 'five_hour' || billingClaim === 'five_hour_fallback') {
+                        overagePct = '0%';
+                    }
+                    else {
+                        overagePct = 'n/a';
+                    }
                     console.log(`[dario] #${requestCount} billing: ${billingClaim} (overage: ${overagePct})`);
                 }
+                else if (verbose) {
+                    console.log(`[dario] #${requestCount} billing: headers absent (status=${upstream.status})`);
+                }
             }
             res.writeHead(upstream.status, responseHeaders);
             if (isStream && upstream.body) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@askalf/dario",
-  "version": "3.10.1",
+  "version": "3.10.3",
   "description": "A local LLM router. One endpoint, every provider — Claude subscriptions, OpenAI, OpenRouter, Groq, local LiteLLM, any OpenAI-compat endpoint — your tools don't need to change.",
   "type": "module",
   "bin": {