openzoo 0.48.65 → 0.48.67

package/lib/anthropic.js

@@ -115,6 +115,12 @@ export function openAIToAnthropic(data, requestedModel) {
   const msg = choice.message ?? {};
   const content = [];
   if (msg.content) content.push({ type: 'text', text: msg.content });
+  // Same reason as the streaming path: a refusal arrives with `content: null`
+  // and the sentence in `refusal`, so dropping it yields an empty message that
+  // reads as a broken model rather than as the answer it is.
+  else if (typeof msg.refusal === 'string' && msg.refusal.length) {
+    content.push({ type: 'text', text: msg.refusal });
+  }
   for (const t of msg.tool_calls ?? []) {
     let input = {};
     try { input = JSON.parse(t.function?.arguments || '{}'); } catch { input = {}; }
@@ -240,6 +246,48 @@ export function streamOpenAIToAnthropic(res, upstream, requestedModel, onReceipt
     if (!ch) return;
     start();
     const delta = ch.delta ?? {};
+    // A REASONING CHUNK IS NOT A SILENT CHUNK.
+    //
+    // Only `delta.content` produced output here, so a reasoning model emitted
+    // NOTHING on the socket for the whole time it was thinking. fable reasons
+    // before it answers — measured at 53 completion tokens to say one word — so
+    // a real turn is minutes of zero bytes, and any hop in between is entitled
+    // to drop a connection that has gone idle.
+    //
+    // OBSERVED: a fable turn rendering as "I'll ground this in your own context
+    // first (holo), scout the monorepo layout, then f" and stopping mid-word.
+    // Claude Code 2.1.232 PRESERVES partial responses on a mid-stream drop
+    // instead of erroring, which is why this reads as "empty/truncated output"
+    // rather than as the connection failure it is.
+    //
+    // `ping` is in Anthropic's own SSE grammar and carries no content, so a
+    // client either ignores it or treats it as liveness — never as text.
+    if (!(typeof delta.content === 'string' && delta.content.length)
+        && !(delta.tool_calls ?? []).length
+        && !(typeof delta.refusal === 'string' && delta.refusal.length)) {
+      ev('ping', {});
+    }
+    // A REFUSAL IS A RESPONSE. SHOW IT.
+    //
+    // Upstream returns refusals out-of-band: `finish_reason: "content_filter"`,
+    // `content: null`, and the actual sentence in `refusal`. We only ever read
+    // `delta.content`, so the whole turn translated to a VALID, EMPTY stream —
+    // message_start, message_stop, output_tokens 0 — and rendered as silence.
+    //
+    // OBSERVED: a security audit of the user's OWN exploited protocol returning
+    // nothing at all, repeatedly, and reading as "fable is broken" when the
+    // provider had in fact answered: "This request triggered restrictions on
+    // violative cyber content and was blocked under Anthropic's Usage Policy."
+    // Being told you were refused is actionable — you can rephrase, or route to
+    // another model. Being shown an empty box is not.
+    if (typeof delta.refusal === 'string' && delta.refusal.length) {
+      if (textIndex < 0) {
+        textIndex = nextIndex++;
+        ev('content_block_start', { index: textIndex, content_block: { type: 'text', text: '' } });
+        open.add(textIndex);
+      }
+      ev('content_block_delta', { index: textIndex, delta: { type: 'text_delta', text: delta.refusal } });
+    }
     if (typeof delta.content === 'string' && delta.content.length) {
       if (textIndex < 0) {
         textIndex = nextIndex++;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openzoo",
-  "version": "0.48.65",
+  "version": "0.48.67",
   "description": "Local x402-paying proxy + MCP server for openzoo.fun — point any OpenAI-compatible harness (Cursor, Claude Code, aider, SDKs) at localhost and it pays per call from a local burner wallet. Solana and Base rails live; Robinhood experimental.",
   "license": "MIT",
   "type": "module",