@prestyj/ai 5.3.0 → 5.3.1

package/dist/index.cjs

@@ -1091,6 +1091,12 @@ function parseToolArguments(argsJson) {
 // src/providers/anthropic.ts
 var NON_STREAMING_TIMEOUT_MS = 60 * 60 * 1e3;
 var anthropicClientCache = /* @__PURE__ */ new Map();
+function fineGrainedToolStreamingEnabled() {
+  const raw = process.env.GG_FINE_GRAINED_TOOL_STREAMING ?? process.env.CLAUDE_CODE_ENABLE_FINE_GRAINED_TOOL_STREAMING;
+  if (!raw) return false;
+  const v = raw.trim().toLowerCase();
+  return v === "1" || v === "true" || v === "yes" || v === "on";
+}
 function createClient(options) {
   const isOAuth = options.apiKey?.startsWith("sk-ant-oat");
   const userAgent = isOAuth ? options.userAgent ?? "claude-cli/2.1.75 (external, cli)" : "";
@@ -1148,7 +1154,9 @@ async function prewarmAnthropicCache(options) {
     ] : system;
     const tools = options.tools?.length ? toAnthropicTools(options.tools, {
       cacheControl,
-      enableFineGrainedToolStreaming: true
+      // Keep the serialized tool bytes identical to runStream so the
+      // prewarmed prompt cache actually hits — both are gated by the flag.
+      enableFineGrainedToolStreaming: fineGrainedToolStreamingEnabled()
     }) : void 0;
     await client.messages.create(
       {
@@ -1228,7 +1236,7 @@ async function* runStream(options) {
         options.tools.filter((t) => !reservedServerNames.has(t.name)),
         {
           ...supportsFirstPartyToolExtras && cacheControl ? { cacheControl } : {},
-          ...supportsFirstPartyToolExtras ? { enableFineGrainedToolStreaming: true } : {}
+          ...supportsFirstPartyToolExtras && fineGrainedToolStreamingEnabled() ? { enableFineGrainedToolStreaming: true } : {}
         }
       ) : [];
       return {
@@ -1254,7 +1262,10 @@ async function* runStream(options) {
     ...isOAuth ? ["claude-code-20250219", "oauth-2025-04-20"] : [],
     ...options.compaction ? ["compact-2026-01-12"] : [],
     ...options.clearToolUses ? ["context-management-2025-06-27"] : [],
-    "fine-grained-tool-streaming-2025-05-14",
+    // Eager tool-input streaming beta — opt-in only (see
+    // fineGrainedToolStreamingEnabled). Off by default: the un-buffered stream
+    // truncates large tool payloads into malformed JSON → phantom empty calls.
+    ...fineGrainedToolStreamingEnabled() ? ["fine-grained-tool-streaming-2025-05-14"] : [],
     ...!hasAdaptiveThinking ? ["interleaved-thinking-2025-05-14"] : [],
     // The 1-h cache TTL (cacheRetention "long") is gated behind this beta. Without
     // it Anthropic silently ignores ttl:"1h" and falls back to the 5-min default,
@@ -1387,7 +1398,14 @@ async function* runStream(options) {
               try {
                 const parsed = JSON.parse(accum.argsJson);
                 args = isJsonObject(parsed) ? parsed : {};
-              } catch {
+              } catch (parseErr) {
+                const rawPartial = accum.argsJson;
+                const snippet = rawPartial.length > 200 ? `${rawPartial.slice(0, 200)}\u2026` : rawPartial;
+                throw new ProviderError(
+                  "anthropic",
+                  `Tool "${accum.toolName}" input JSON was truncated in the stream (${rawPartial.length} bytes): ${snippet}; ${parseErr.message}`,
+                  { cause: parseErr }
+                );
               }
             }
             const tc = {
@@ -1608,6 +1626,7 @@ function readUnifiedRateLimit(headers) {
   return { rejected: status === "rejected", ...resetsAt ? { resetsAt } : {} };
 }
 function toError(err) {
+  if (err instanceof ProviderError) return err;
   if (err instanceof import_sdk.default.APIError) {
     const errorBody = err.error;
     const nestedError = errorBody?.error;