@veryfront/ext-llm-openai 0.1.1223 → 0.1.1225

package/esm/openai-tool-input.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-tool-input.d.ts","sourceRoot":"","sources":["../src/openai-tool-input.ts"],"names":[],"mappings":"AAAA,eAAO,MAAM,qCAAqC,UAAY,CAAC;AAC/D,eAAO,MAAM,yCAAyC,OAAQ,CAAC;AAE/D,MAAM,MAAM,8BAA8B,GAAG;IAC3C,KAAK,EAAE,MAAM,CAAC;IACd,SAAS,EAAE,MAAM,CAAC;CACnB,CAAC;AAIF,wBAAgB,8BAA8B,CAC5C,MAAM,EAAE,8BAA8B,EACtC,MAAM,EAAE,MAAM,EAAE,EAChB,QAAQ,EAAE,MAAM,GACf,OAAO,GAAG,WAAW,GAAG,SAAS,CAgBnC;AAED,wBAAgB,6BAA6B,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE,GAAG,MAAM,CAE/E;AAED,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAOvD"}
+{"version":3,"file":"openai-tool-input.d.ts","sourceRoot":"","sources":["../src/openai-tool-input.ts"],"names":[],"mappings":"AAAA,eAAO,MAAM,qCAAqC,UAAY,CAAC;AAC/D,eAAO,MAAM,yCAAyC,OAAQ,CAAC;AAE/D,MAAM,MAAM,8BAA8B,GAAG;IAC3C,KAAK,EAAE,MAAM,CAAC;IACd,SAAS,EAAE,MAAM,CAAC;CACnB,CAAC;AAIF,wBAAgB,8BAA8B,CAC5C,MAAM,EAAE,8BAA8B,EACtC,MAAM,EAAE,MAAM,EAAE,EAChB,QAAQ,EAAE,MAAM,GACf,OAAO,GAAG,WAAW,GAAG,SAAS,CA+BnC;AAED,wBAAgB,6BAA6B,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE,GAAG,MAAM,CAE/E;AAED,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAOvD"}

package/esm/openai-tool-input.js

@@ -2,18 +2,32 @@ export const MAX_OPENAI_STREAM_TOOL_ARGUMENT_BYTES = 1_048_576;
 export const MAX_OPENAI_STREAM_TOOL_ARGUMENT_FRAGMENTS = 4_096;
 const TOOL_ARGUMENT_ENCODER = new TextEncoder();
 export function appendOpenAIStreamToolArgument(budget, chunks, fragment) {
-    if (budget.fragments >= MAX_OPENAI_STREAM_TOOL_ARGUMENT_FRAGMENTS) {
-        return "fragments";
-    }
     const fragmentBytes = TOOL_ARGUMENT_ENCODER.encode(fragment).byteLength;
+    // Zero-byte fragments never advance the byte budget, so they are the only
+    // ones that can arrive without bound -- and they are what the fragment cap
+    // exists to stop. Both provider streams exercise it exactly that way, by
+    // flooding empty deltas.
+    //
+    // Counting non-empty fragments here too made the cap bind long before the
+    // byte budget: at typical delta sizes, 4096 fragments is roughly 2-8% of the
+    // 1 MiB a tool call is nominally allowed, which left the byte limit
+    // unreachable. Fragment count is chosen by the provider's tokenizer rather
+    // than by the caller, so the same tool call passed or failed depending on how
+    // the stream happened to be chunked.
+    if (fragmentBytes === 0) {
+        if (budget.fragments >= MAX_OPENAI_STREAM_TOOL_ARGUMENT_FRAGMENTS) {
+            return "fragments";
+        }
+        budget.fragments++;
+        return undefined;
+    }
     if (fragmentBytes > MAX_OPENAI_STREAM_TOOL_ARGUMENT_BYTES - budget.bytes) {
         return "bytes";
     }
-    budget.fragments++;
+    // Content is bounded by bytes: every fragment kept here is at least one byte,
+    // so the array cannot outgrow the byte budget.
     budget.bytes += fragmentBytes;
-    if (fragment.length > 0) {
-        chunks.push(fragment);
-    }
+    chunks.push(fragment);
     return undefined;
 }
 export function joinOpenAIStreamToolArguments(chunks) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@veryfront/ext-llm-openai",
-  "version": "0.1.1223",
+  "version": "0.1.1225",
   "description": "Veryfront first-party extension package for ext-llm-openai",
   "keywords": [
     "veryfront",
@@ -45,7 +45,7 @@
     "capabilities": []
   },
   "peerDependencies": {
-    "veryfront": "^0.1.1223"
+    "veryfront": "^0.1.1225"
   },
   "type": "module",
   "types": "./esm/index.d.ts",