@juspay/neurolink 10.12.8 → 10.12.9

package/dist/lib/utils/json/coerce.js

@@ -23,6 +23,48 @@ import { nextBalancedJsonSpan } from "./extract.js";
 function hasSafeParse(schema) {
     return typeof schema.safeParse === "function";
 }
+/**
+ * Does `value` satisfy `schema`? A schema we cannot validate with (absent, or
+ * without a Zod-style `safeParse`) accepts everything — callers use this as a
+ * gate, not as proof, so an unknown schema must never block a value.
+ *
+ * Exists so the consumers of coerceJsonToSchema can refuse to publish a
+ * recovered value as `structuredData` when the caller's schema rejects it —
+ * e.g. a raw *string* under an object schema, which is the shape a truncated
+ * response degrades to.
+ */
+export function schemaAccepts(schema, value) {
+    if (!schema || !hasSafeParse(schema)) {
+        return true;
+    }
+    return schema.safeParse(value).success;
+}
+/**
+ * Recover a JSON *scalar* root (string/number/boolean) from model text. The
+ * object/array case is handled by `coerceJsonToSchema`; this covers the
+ * residual scalar root after that path returns null. Encapsulates the JSON
+ * parsing, the empty-string normalization, and the `schemaAccepts` gate so the
+ * same policy cannot drift between consumers (`neurolink.recoverStructuredData`
+ * and `GenerationHandler.coerceTextMode`).
+ */
+export function recoverScalarRoot(text, schema) {
+    try {
+        const scalar = JSON.parse(text);
+        if (scalar === "") {
+            return { kind: "empty" };
+        }
+        if (scalar === null || scalar === undefined) {
+            return { kind: "nullish" };
+        }
+        if (schemaAccepts(schema, scalar)) {
+            return { kind: "accepted", value: scalar };
+        }
+        return { kind: "rejected", value: scalar };
+    }
+    catch {
+        return { kind: "not-json" };
+    }
+}
 /**
  * Parse `candidate` as JSON, repairing common escaping mistakes on failure.
  * Returns the parsed value plus whether jsonrepair had to alter the text.
@@ -51,6 +93,34 @@ function parseOrRepair(candidate) {
 }
 /** Bounds the recursive nested-string unwrap against pathological inputs. */
 const MAX_NESTED_UNWRAP_DEPTH = 6;
+/** Bounds the structural back-off when salvaging a truncated root object. */
+const MAX_SALVAGE_TRIMS = 8;
+/**
+ * Last-resort recovery for output cut off mid-JSON: walk the unclosed root span
+ * backwards to successively earlier structural boundaries (`,` / `}` / `]`) and
+ * re-attempt parse+repair at each one. Truncation can land somewhere jsonrepair
+ * cannot close on its own (inside an escape, a half-written key, a dangling
+ * separator); dropping back to the last completed field gives it a repairable
+ * prefix. Returns a PARTIAL object — the fields the model did finish — so a
+ * truncated response degrades to an incomplete object rather than to raw text.
+ */
+function salvageTruncatedRoot(span) {
+    let candidate = span;
+    for (let i = 0; i < MAX_SALVAGE_TRIMS && candidate.length > 1; i++) {
+        const cut = Math.max(candidate.lastIndexOf(","), candidate.lastIndexOf("}"), candidate.lastIndexOf("]"));
+        if (cut <= 0) {
+            break;
+        }
+        candidate = candidate.slice(0, cut);
+        const outcome = parseOrRepair(candidate);
+        if (outcome !== undefined &&
+            outcome.value !== null &&
+            typeof outcome.value === "object") {
+            return outcome.value;
+        }
+    }
+    return undefined;
+}
 /**
  * Recursively replace any string-valued field whose content is itself a JSON
  * object/array with the parsed value. Models sometimes double-encode a NESTED
@@ -135,30 +205,54 @@ export function coerceJsonToSchema(text, schema) {
     //  3. first "{" or "[" to end of text (TRUNCATED output —
     //     finishReason=length — where the closing bracket was cut off;
     //     jsonrepair closes it)
-    // `truncated` marks the first-open-to-end candidate: it is only reachable
-    // when no balanced span and no first-to-last span matched, i.e. there was no
-    // closing bracket at all — the signature of token-truncated output.
+    // `truncated` marks a candidate that could only come from output cut short:
+    // the first-open-to-end span, and — when the ROOT bracket never closes —
+    // every other candidate too, since all of them are then partial views of an
+    // unfinished document.
+    //
+    // `rootAligned` marks candidates that start at the document's first opening
+    // bracket, i.e. at the real root. Candidates are ordered root-aligned FIRST.
+    // On truncated output the root brace never balances, so the balanced scan
+    // walks on and matches brackets that live INSIDE a string value (`[step 1]`
+    // in a shell script, say) — a syntactically fine but semantically bogus
+    // array. Preferring the root keeps the partial real object ahead of that.
     const candidates = [];
+    const openIndexes = [text.indexOf("{"), text.indexOf("[")].filter((i) => i >= 0);
+    const firstOpen = openIndexes.length > 0 ? Math.min(...openIndexes) : -1;
+    let rootClosed = false;
     let searchFrom = 0;
     for (;;) {
         const found = nextBalancedJsonSpan(text, searchFrom);
         if (!found) {
             break;
         }
-        candidates.push({ text: found.span, truncated: false });
+        const start = found.end - found.span.length;
+        if (start === firstOpen) {
+            rootClosed = true;
+        }
+        candidates.push({
+            text: found.span,
+            truncated: false,
+            rootAligned: start === firstOpen,
+        });
         searchFrom = found.end;
     }
-    const openIndexes = [text.indexOf("{"), text.indexOf("[")].filter((i) => i >= 0);
-    const firstOpen = openIndexes.length > 0 ? Math.min(...openIndexes) : -1;
+    // No balanced span starts at the root bracket → the document is unclosed.
+    const rootUnclosed = firstOpen >= 0 && !rootClosed;
     const lastClose = Math.max(text.lastIndexOf("}"), text.lastIndexOf("]"));
     if (firstOpen >= 0 && lastClose > firstOpen) {
         candidates.push({
             text: text.slice(firstOpen, lastClose + 1),
-            truncated: false,
+            truncated: rootUnclosed,
+            rootAligned: true,
         });
     }
     if (firstOpen >= 0) {
-        candidates.push({ text: text.slice(firstOpen), truncated: true });
+        candidates.push({
+            text: text.slice(firstOpen),
+            truncated: true,
+            rootAligned: true,
+        });
     }
     // JSON-string-literal wrapper: some providers double-encode and return the
     // object as a JSON *string* (e.g. `"{\"k\":1}"`). Unwrap one layer and add
@@ -174,7 +268,11 @@ export function coerceJsonToSchema(text, schema) {
                     if (!innerSpan) {
                         break;
                     }
-                    candidates.push({ text: innerSpan.span, truncated: false });
+                    candidates.push({
+                        text: innerSpan.span,
+                        truncated: false,
+                        rootAligned: false,
+                    });
                     innerFrom = innerSpan.end;
                 }
             }
@@ -183,11 +281,20 @@ export function coerceJsonToSchema(text, schema) {
             // not a string literal — ignore
         }
     }
+    // Stable partition: root-aligned candidates first. Only the ORDER changes —
+    // no candidate is dropped — so the multi-object "prefer the most complete
+    // schema-valid candidate" selection below is unaffected (it is order
+    // independent), while the unschema'd first-parseable-wins path and the
+    // `firstValid` fallback now favour the document's real root.
+    const ordered = [
+        ...candidates.filter((c) => c.rootAligned),
+        ...candidates.filter((c) => !c.rootAligned),
+    ];
     let firstValid;
     const schemaValid = [];
     const hasSchema = !!(schema && hasSafeParse(schema));
     const seen = new Set();
-    for (const candidate of candidates) {
+    for (const candidate of ordered) {
         if (seen.has(candidate.text)) {
             continue;
         }
@@ -202,6 +309,7 @@ export function coerceJsonToSchema(text, schema) {
             value: outcome.value,
             repaired: outcome.repaired,
             truncated: candidate.truncated,
+            rootAligned: candidate.rootAligned,
         };
         if (firstValid === undefined) {
             firstValid = record;
@@ -229,6 +337,7 @@ export function coerceJsonToSchema(text, schema) {
                     value: unwrapped.value,
                     repaired: true,
                     truncated: candidate.truncated,
+                    rootAligned: candidate.rootAligned,
                 });
             }
             else if (
@@ -247,6 +356,7 @@ export function coerceJsonToSchema(text, schema) {
                     value: outcome.value[0],
                     repaired: true,
                     truncated: candidate.truncated,
+                    rootAligned: candidate.rootAligned,
                 });
             }
         }
@@ -261,7 +371,34 @@ export function coerceJsonToSchema(text, schema) {
             ? cur
             : best)
         : undefined;
-    const chosen = schemaMatch ?? firstValid;
+    let chosen = schemaMatch ?? firstValid;
+    // Salvage the root when the document is unclosed AND nothing trustworthy came
+    // out of it: either nothing parsed at all, or the only parseable candidate is
+    // a bracket pair scraped from INSIDE a string value (not root-aligned, not
+    // schema-valid) — the `["step 1"]`-from-a-shell-script case. Backing off to
+    // the last completed field yields a PARTIAL OBJECT flagged `truncated`,
+    // instead of a bogus value or raw text degrading to a string.
+    if (rootUnclosed &&
+        schemaMatch === undefined &&
+        (chosen === undefined || !chosen.rootAligned)) {
+        const salvaged = salvageTruncatedRoot(text.slice(firstOpen));
+        if (salvaged !== undefined) {
+            logger.debug("[coerceJsonToSchema] salvaged a partial truncated object", {
+                textLength: text.length,
+            });
+            chosen = {
+                value: salvaged,
+                repaired: true,
+                truncated: true,
+                rootAligned: true,
+            };
+        }
+        else if (chosen !== undefined) {
+            // Nothing salvageable, but we still know the output was truncated —
+            // never report a candidate scraped from a truncated document as complete.
+            chosen = { ...chosen, truncated: true };
+        }
+    }
     if (chosen === undefined) {
         return null;
     }

package/dist/neurolink.d.ts

@@ -775,6 +775,21 @@ export declare class NeuroLink {
     private applyClassifierRouting;
     private prepareGenerateAugmentations;
     private buildGenerateTextOptions;
+    /**
+     * Provider-agnostic JSON recovery for schema requests. Structured-output
+     * enforcement makes valid JSON the overwhelming case; for every other
+     * provider path — including generate() overrides (Vertex, Anthropic,
+     * Bedrock, Google AI Studio) — object/array roots are recovered here via
+     * balanced-scan + jsonrepair and scalar JSON roots via plain JSON.parse,
+     * with the parsed value exposed as `structuredData`. If nothing JSON-shaped
+     * is recoverable (pure prose), the raw text is returned, `structuredData`
+     * stays undefined, and a WARN makes the case observable.
+     *
+     * Mutates `textResult` in place, and must run BEFORE the end-of-generation
+     * emits so event consumers see the same content/structuredData the caller
+     * receives.
+     */
+    private recoverStructuredData;
     private finalizeGenerateRequestResult;
     private emitGenerateErrorEvent;
     /**

package/dist/neurolink.js

@@ -74,7 +74,7 @@ import { CircuitBreaker, ERROR_CODES, ErrorFactory, isAbortError, isRetriableErr
 import { hasLifecycleErrorFired, markLifecycleErrorFired, } from "./utils/lifecycleCallbacks.js";
 import { resolveLifecycleTimeoutMs } from "./utils/lifecycleTimeout.js";
 import { cloneOptionsForCallIsolation } from "./utils/cloneOptions.js";
-import { coerceJsonToSchema } from "./utils/json/coerce.js";
+import { coerceJsonToSchema, recoverScalarRoot, schemaAccepts, } from "./utils/json/coerce.js";
 // Factory processing imports
 import { createCleanStreamOptions, enhanceTextGenerationOptions, processFactoryOptions, processStreamingFactoryOptions, validateFactoryConfig, } from "./utils/factoryProcessing.js";
 import { logger, mcpLogger } from "./utils/logger.js";
@@ -4058,52 +4058,85 @@ Current user's request: ${currentInput}`;
         }
         return textOptions;
     }
-    finalizeGenerateRequestResult(params) {
-        const { generateSpan, options, textOptions, textResult, factoryResult, originalPrompt, startTime, } = params;
-        // Provider-agnostic JSON coercion for schema requests. Structured-output
-        // enforcement makes valid JSON the overwhelming case; for every other
-        // provider path — including generate() overrides (Vertex, Anthropic,
-        // Bedrock, Google AI Studio) — object/array roots are recovered here via
-        // balanced-scan + jsonrepair and scalar JSON roots via plain JSON.parse,
-        // with the parsed value exposed as `structuredData`. If nothing
-        // JSON-shaped is recoverable (pure prose), the raw text is returned,
-        // `structuredData` stays undefined, and a WARN makes the case observable.
-        // Runs BEFORE the end-of-generation emits below so event consumers see
-        // the same coerced content/structuredData the caller receives.
-        if (textOptions.schema &&
-            textResult.structuredData === undefined &&
-            typeof textResult.content === "string") {
-            const coerced = coerceJsonToSchema(textResult.content, textOptions.schema);
-            if (coerced) {
-                textResult.content = coerced.content;
-                textResult.structuredData = coerced.structuredData;
-                if (coerced.repaired) {
-                    textResult.jsonRepaired = true;
-                }
-                if (coerced.truncated) {
-                    textResult.jsonTruncated = true;
-                }
+    /**
+     * Provider-agnostic JSON recovery for schema requests. Structured-output
+     * enforcement makes valid JSON the overwhelming case; for every other
+     * provider path — including generate() overrides (Vertex, Anthropic,
+     * Bedrock, Google AI Studio) — object/array roots are recovered here via
+     * balanced-scan + jsonrepair and scalar JSON roots via plain JSON.parse,
+     * with the parsed value exposed as `structuredData`. If nothing JSON-shaped
+     * is recoverable (pure prose), the raw text is returned, `structuredData`
+     * stays undefined, and a WARN makes the case observable.
+     *
+     * Mutates `textResult` in place, and must run BEFORE the end-of-generation
+     * emits so event consumers see the same content/structuredData the caller
+     * receives.
+     */
+    recoverStructuredData(textResult, schema) {
+        // A provider path that produced its own `structuredData` normally owns it.
+        // The one exception is a STRING the caller's schema rejects: that is the
+        // raw completion leaking through as structured output (the shape a
+        // truncated response degrades to), so re-run recovery over the text rather
+        // than handing back a value the declared schema forbids.
+        const structuredIsRejectedString = typeof textResult.structuredData === "string" &&
+            !schemaAccepts(schema, textResult.structuredData);
+        if (!schema ||
+            (textResult.structuredData !== undefined &&
+                !structuredIsRejectedString) ||
+            typeof textResult.content !== "string") {
+            return;
+        }
+        if (structuredIsRejectedString) {
+            textResult.structuredData = undefined;
+        }
+        const coerced = coerceJsonToSchema(textResult.content, schema);
+        if (coerced) {
+            textResult.content = coerced.content;
+            textResult.structuredData = coerced.structuredData;
+            if (coerced.repaired) {
+                textResult.jsonRepaired = true;
             }
-            else {
-                try {
-                    const scalar = JSON.parse(textResult.content);
-                    if (scalar === "") {
-                        // A JSON-encoded empty string is an EMPTY completion, not a
-                        // recovered scalar — normalize to a true empty so callers'
-                        // empty-response handling fires instead of a literal '""'
-                        // reaching the user. `structuredData` stays undefined.
-                        textResult.content = "";
-                        logger.warn("[NeuroLink] schema requested but the model returned an empty JSON string; normalizing to empty content", { provider: textResult.provider, model: textResult.model });
-                    }
-                    else if (scalar !== null && scalar !== undefined) {
-                        textResult.structuredData = scalar;
-                    }
-                }
-                catch {
-                    logger.warn("[NeuroLink] schema requested but no JSON could be recovered from model output; returning raw text", { provider: textResult.provider, model: textResult.model });
-                }
+            if (coerced.truncated) {
+                textResult.jsonTruncated = true;
             }
+            return;
         }
+        const scalar = recoverScalarRoot(textResult.content, schema);
+        switch (scalar.kind) {
+            case "empty":
+                // A JSON-encoded empty string is an EMPTY completion, not a recovered
+                // scalar — normalize to a true empty so callers' empty-response
+                // handling fires instead of a literal '""' reaching the user.
+                // `structuredData` stays undefined.
+                textResult.content = "";
+                logger.warn("[NeuroLink] schema requested but the model returned an empty JSON string; normalizing to empty content", { provider: textResult.provider, model: textResult.model });
+                break;
+            case "accepted":
+                // Only publish a scalar root the caller's schema actually accepts.
+                // Under an OBJECT schema a recovered string is the raw completion in
+                // disguise — the shape a truncated response degrades to — and exposing
+                // it hands the caller a `structuredData` that violates the schema they
+                // passed.
+                textResult.structuredData = scalar.value;
+                break;
+            case "rejected":
+                logger.warn("[NeuroLink] recovered a JSON scalar the requested schema rejects; leaving structuredData unset", {
+                    provider: textResult.provider,
+                    model: textResult.model,
+                    scalarType: typeof scalar.value,
+                });
+                break;
+            case "nullish":
+                // JSON null/undefined — no structured value to publish.
+                break;
+            case "not-json":
+                logger.warn("[NeuroLink] schema requested but no JSON could be recovered from model output; returning raw text", { provider: textResult.provider, model: textResult.model });
+                break;
+        }
+    }
+    finalizeGenerateRequestResult(params) {
+        const { generateSpan, options, textOptions, textResult, factoryResult, originalPrompt, startTime, } = params;
+        this.recoverStructuredData(textResult, textOptions.schema);
         // Surface truncation when a schema was requested: either the provider
         // reported finishReason="length" or the recovered JSON came from an
         // unclosed span. Either way `structuredData` may be incomplete — warn at

package/dist/providers/anthropic/client.js

@@ -1206,6 +1206,9 @@ export class AnthropicProvider extends BaseProvider {
                 }
                 const content = [];
                 let finalResultText;
+                // Text emitted in forced-json mode, kept only as a fallback (see below).
+                const jsonModeText = [];
+                let jsonToolAnswered = false;
                 for (const block of response.content) {
                     if (block.type === "thinking") {
                         content.push({ type: "reasoning", text: block.thinking });
@@ -1213,13 +1216,17 @@ export class AnthropicProvider extends BaseProvider {
                     else if (block.type === "text") {
                         // In forced-json mode the payload arrives via the tool input, not
                         // text — pass text through only in normal mode.
-                        if (!jsonTool) {
+                        if (jsonTool) {
+                            jsonModeText.push(block.text);
+                        }
+                        else {
                             content.push({ type: "text", text: block.text });
                         }
                     }
                     else if (block.type === "tool_use") {
                         if (jsonTool && block.name === jsonTool) {
                             // Unwrap the synthetic tool call back into text JSON.
+                            jsonToolAnswered = true;
                             content.push({
                                 type: "text",
                                 text: stringifyToolInput(block.input),
@@ -1241,6 +1248,15 @@ export class AnthropicProvider extends BaseProvider {
                         }
                     }
                 }
+                // Forced-json mode normally drops text blocks because the payload rides
+                // in the synthetic tool's input. But when the response is cut short
+                // (stop_reason "max_tokens") the tool call can be missing entirely, and
+                // dropping the text would leave an EMPTY completion with nothing for
+                // coerceJsonToSchema to recover. Fall back to the text so a partial
+                // object can still be salvaged and flagged truncated.
+                if (jsonTool && !jsonToolAnswered && jsonModeText.length > 0) {
+                    content.push({ type: "text", text: jsonModeText.join("") });
+                }
                 // final_result is terminal — parity with the native Claude-on-Vertex
                 // and Gemini loops, which break out of the tool loop the moment it
                 // arrives. Reasoning blocks are kept; any prose preamble and any tool

package/dist/types/utilities.d.ts

@@ -279,3 +279,32 @@ export type JsonCoercionResult = {
      */
     truncated: boolean;
 };
+/**
+ * Decision returned by `recoverScalarRoot`. Each caller applies it to its own
+ * result shape and logger prefix, preserving its existing warning behaviour:
+ *
+ *   - `empty`     — the text is a JSON-encoded empty string (an EMPTY
+ *                   completion, not a recovered scalar). Callers normalize to a
+ *                   true empty.
+ *   - `accepted`  — a scalar root the caller's schema accepts; safe to publish
+ *                   as `structuredData`.
+ *   - `rejected`  — a scalar root the caller's schema rejects (e.g. a raw
+ *                   string under an object schema — the shape a truncated
+ *                   response degrades to). Do NOT publish it.
+ *   - `nullish`   — the text is the JSON literals `null`/`undefined`; there is
+ *                   no structured value to publish.
+ *   - `not-json`  — the text is not JSON at all.
+ */
+export type ScalarRecoveryDecision = {
+    kind: "empty";
+} | {
+    kind: "accepted";
+    value: unknown;
+} | {
+    kind: "rejected";
+    value: unknown;
+} | {
+    kind: "nullish";
+} | {
+    kind: "not-json";
+};

package/dist/utils/json/coerce.d.ts

@@ -1,4 +1,24 @@
-import type { JsonCoercionResult, ValidationSchema } from "../../types/index.js";
+import type { JsonCoercionResult, ScalarRecoveryDecision, ValidationSchema } from "../../types/index.js";
+/**
+ * Does `value` satisfy `schema`? A schema we cannot validate with (absent, or
+ * without a Zod-style `safeParse`) accepts everything — callers use this as a
+ * gate, not as proof, so an unknown schema must never block a value.
+ *
+ * Exists so the consumers of coerceJsonToSchema can refuse to publish a
+ * recovered value as `structuredData` when the caller's schema rejects it —
+ * e.g. a raw *string* under an object schema, which is the shape a truncated
+ * response degrades to.
+ */
+export declare function schemaAccepts(schema: ValidationSchema | undefined, value: unknown): boolean;
+/**
+ * Recover a JSON *scalar* root (string/number/boolean) from model text. The
+ * object/array case is handled by `coerceJsonToSchema`; this covers the
+ * residual scalar root after that path returns null. Encapsulates the JSON
+ * parsing, the empty-string normalization, and the `schemaAccepts` gate so the
+ * same policy cannot drift between consumers (`neurolink.recoverStructuredData`
+ * and `GenerationHandler.coerceTextMode`).
+ */
+export declare function recoverScalarRoot(text: string, schema: ValidationSchema | undefined): ScalarRecoveryDecision;
 /**
  * Try to produce canonical JSON from `text`. Returns null when no JSON object
  * could be recovered (caller should then keep the raw text).