@juspay/neurolink 10.10.2 → 10.10.4

package/dist/utils/tokenEstimation.js

@@ -33,6 +33,12 @@ export const TOKENS_PER_MESSAGE = 4;
 export const TOKENS_PER_CONVERSATION = 24;
 /** Image token estimate (flat) */
 export const IMAGE_TOKEN_ESTIMATE = 1_024;
+/**
+ * Chars charged for a value that cannot be serialized for estimation (V8 max
+ * string length). Deliberately large: such a value is enormous by definition,
+ * and under-charging it would defeat the budget check it feeds.
+ */
+const OVERSIZED_VALUE_FALLBACK_CHARS = 200_000;
 /**
  * Per-provider token multipliers.
  * Applied on top of the base GPT-style character estimate.
@@ -80,9 +86,39 @@ export function estimateTokens(text, provider, isCode) {
     const safetyBuffer = baseTokens * TOKEN_SAFETY_MARGIN_ADDITIVE;
     return Math.ceil(providerAdjusted + safetyBuffer);
 }
+/**
+ * Serialize an arbitrary value for estimation. Never throws: `JSON.stringify`
+ * raises RangeError once a value exceeds V8's max string length, and a tool
+ * argument blob is exactly the shape that gets there. A payload that large is
+ * charged at the fallback size rather than aborting the estimate (and with it
+ * the whole turn).
+ */
+function serializeForEstimate(value) {
+    if (typeof value === "string") {
+        return value;
+    }
+    try {
+        return JSON.stringify(value) ?? "";
+    }
+    catch {
+        return "x".repeat(OVERSIZED_VALUE_FALLBACK_CHARS);
+    }
+}
 /**
  * Estimate token count for a single ChatMessage.
  * Includes message framing overhead.
+ *
+ * Counts `content` AND `args`. A `tool_call` is persisted with an EMPTY
+ * `content` and its entire payload in `args` (see flushPendingToolData),
+ * so a content-only estimate scored a 39 KB Write call at ~28 tokens against a
+ * real cost near 9,750 — the budget checker, the compaction trigger and the
+ * summarization threshold were all blind to the single largest source of
+ * context growth in an agentic session.
+ *
+ * `result` is deliberately NOT counted: `result.result` is re-hydrated FROM
+ * `content` at read time (redisConversationMemoryManager), so counting both
+ * double-bills the same bytes. Only `result.error`, which has no counterpart in
+ * `content`, is included.
  */
 export function estimateMessageTokens(message, provider) {
     let contentStr = "";
@@ -100,8 +136,16 @@ export function estimateMessageTokens(message, provider) {
             }
         }
     }
-    const contentTokens = estimateTokens(contentStr, provider);
-    return contentTokens + TOKENS_PER_MESSAGE;
+    let total = estimateTokens(contentStr, provider) + TOKENS_PER_MESSAGE;
+    const args = message.args;
+    if (args) {
+        total += estimateTokens(serializeForEstimate(args), provider);
+    }
+    const resultError = message.result?.error;
+    if (resultError) {
+        total += estimateTokens(serializeForEstimate(resultError), provider);
+    }
+    return total;
 }
 /**
  * Estimate total token count for an array of messages.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "10.10.2",
+  "version": "10.10.4",
   "packageManager": "pnpm@10.15.1",
   "description": "TypeScript AI SDK with 24+ LLM providers behind one consistent API. MCP-native (connect any MCP server), voice TTS/STT/realtime, RAG, agents, memory, context compaction. OpenAI · Anthropic · Gemini · Bedrock · Azure · Ollama · DeepSeek · NVIDIA NIM and more.",
   "author": {
@@ -83,6 +83,8 @@
     "test:litellm-parity": "npx tsx test/continuous-test-suite-litellm-parity.ts",
     "test:memory": "npx tsx test/continuous-test-suite-memory.ts",
     "test:tool-pairing": "npx tsx test/continuous-test-suite-tool-pairing.ts",
+    "test:token-accounting": "npx tsx test/continuous-test-suite-token-accounting.ts",
+    "test:step-guard": "npx tsx test/continuous-test-suite-step-guard.ts",
     "test:middleware": "npx tsx test/continuous-test-suite-middleware.ts",
     "test:observability": "npx tsx test/continuous-test-suite-observability.ts",
     "test:ppt": "npx tsx test/continuous-test-suite-ppt.ts",