@juspay/neurolink 9.79.3 → 9.80.1

package/dist/lib/providers/googleVertex.js

@@ -44,6 +44,144 @@ const hasAnthropicSupport = () => {
     // Actual availability is checked at runtime when creating the client
     return true;
 };
+// Parse stored rows into ordered segments, grouping tool_call/tool_result by (turn, step).
+function collectHistorySegments(conversationMessages) {
+    const segments = [];
+    const stepMap = new Map();
+    let turnCounter = 0;
+    const getStep = (stepIndex) => {
+        const key = `${turnCounter}:${stepIndex ?? "x"}`;
+        let step = stepMap.get(key);
+        if (!step) {
+            step = { type: "tool_step", callParts: [], resultParts: [] };
+            stepMap.set(key, step);
+            segments.push(step);
+        }
+        return step;
+    };
+    for (const msg of conversationMessages) {
+        if (msg.role === "tool_call") {
+            getStep(msg.metadata?.stepIndex).callParts.push({
+                name: msg.tool || "unknown",
+                input: msg.args || {},
+            });
+            continue;
+        }
+        if (msg.role === "tool_result") {
+            getStep(msg.metadata?.stepIndex).resultParts.push(msg.content && msg.content.length > 0 ? msg.content : "(no output)");
+            continue;
+        }
+        const role = msg.role === "assistant"
+            ? "assistant"
+            : msg.role === "user"
+                ? "user"
+                : null;
+        if (!role || !msg.content || msg.content.trim().length === 0) {
+            continue;
+        }
+        // New turn → fresh step namespace for any following tool rows.
+        turnCounter++;
+        segments.push({ type: "regular", role, parts: [msg.content] });
+    }
+    return segments;
+}
+// Append blocks to the trailing turn if it shares the role, else start a new turn.
+function pushTurn(messages, role, blocks) {
+    const last = messages[messages.length - 1];
+    if (last && last.role === role && Array.isArray(last.content)) {
+        last.content.push(...blocks);
+    }
+    else {
+        messages.push({ role, content: [...blocks] });
+    }
+}
+// Pair a tool step's calls/results by position; unbalanced extras are dropped
+// because Anthropic 400s on an orphaned tool_use/tool_result reference.
+function pairToolStep(step, ordinal) {
+    const calls = step.callParts;
+    const resultStrings = step.resultParts;
+    const paired = Math.min(calls.length, resultStrings.length);
+    if (paired !== calls.length || paired !== resultStrings.length) {
+        logger.debug("[GoogleVertex] Unbalanced tool step in Claude history replay", {
+            calls: calls.length,
+            results: resultStrings.length,
+            paired,
+        });
+    }
+    const uses = [];
+    const results = [];
+    for (let i = 0; i < paired; i++) {
+        const id = `hist_${ordinal}_${i}`;
+        uses.push({
+            type: "tool_use",
+            id,
+            name: calls[i].name,
+            input: calls[i].input,
+        });
+        results.push({
+            type: "tool_result",
+            tool_use_id: id,
+            content: resultStrings[i],
+        });
+    }
+    return { uses, results };
+}
+// Drop leading turns until the first is a real user turn — Anthropic requires it
+// (covers a leading assistant/tool step and an orphaned tool_result-only turn).
+function trimLeadingNonUser(messages) {
+    const isToolResultOnly = (m) => Array.isArray(m.content) &&
+        m.content.length > 0 &&
+        m.content.every((block) => block.type === "tool_result");
+    while (messages.length > 0 &&
+        (messages[0].role !== "user" || isToolResultOnly(messages[0]))) {
+        messages.shift();
+    }
+}
+// Rebuild Anthropic history with paired tool_use/tool_result turns from stored rows.
+export function buildAnthropicHistoryMessages(conversationMessages) {
+    const messages = [];
+    let stepOrdinal = 0;
+    for (const seg of collectHistorySegments(conversationMessages)) {
+        if (seg.type === "regular") {
+            pushTurn(messages, seg.role === "assistant" ? "assistant" : "user", [
+                { type: "text", text: String(seg.parts[0] ?? "") },
+            ]);
+            continue;
+        }
+        const { uses, results } = pairToolStep(seg, stepOrdinal);
+        if (uses.length === 0) {
+            continue;
+        }
+        stepOrdinal++;
+        pushTurn(messages, "assistant", uses);
+        pushTurn(messages, "user", results);
+    }
+    trimLeadingNonUser(messages);
+    if (logger.shouldLog("debug")) {
+        const blocks = messages.flatMap((m) => Array.isArray(m.content) ? m.content : []);
+        logger.debug("[GoogleVertex] Replayed Claude history with tool turns", {
+            historyMessages: messages.length,
+            toolUseBlocks: blocks.filter((b) => b.type === "tool_use").length,
+            toolResultBlocks: blocks.filter((b) => b.type === "tool_result").length,
+        });
+    }
+    return messages;
+}
+// Append the live user turn, merging into a trailing user turn to avoid back-to-back user messages.
+export function appendUserMessage(messages, content) {
+    const last = messages[messages.length - 1];
+    if (last && last.role === "user") {
+        const head = Array.isArray(last.content)
+            ? last.content
+            : [{ type: "text", text: last.content }];
+        const tail = Array.isArray(content)
+            ? content
+            : [{ type: "text", text: content }];
+        last.content = [...head, ...tail];
+        return;
+    }
+    messages.push({ role: "user", content });
+}
 /**
  * Recursively strip JSON-schema fields that Vertex Gemini's function-call AND
  * `responseSchema` (structured output) validators reject with 400
@@ -2287,18 +2425,14 @@ export class GoogleVertexProvider extends BaseProvider {
         });
         // Build messages from input
         const messages = [];
-        // Add conversation history if present.
-        //
-        // Intentionally text-only. Anthropic's API rejects messages where a
-        // tool_use_id reference appears without its matching tool_use in the
-        // same turn — so synthesising tool_use / tool_result blocks from
-        // stored ChatMessages risks emitting orphaned references that fail
-        // validation. Tool rows are still persisted to Redis (chat-history
-        // UI renders them) but they don't re-enter the model's context on
-        // subsequent turns.
+        // Replay conversationMessages (with tool turns), else the legacy text-only conversationHistory.
         if (options.conversationMessages &&
             options.conversationMessages.length > 0) {
-            for (const msg of options.conversationMessages) {
+            messages.push(...buildAnthropicHistoryMessages(options.conversationMessages));
+        }
+        else if (options.conversationHistory &&
+            options.conversationHistory.length > 0) {
+            for (const msg of options.conversationHistory) {
                 if (msg.role === "user" || msg.role === "assistant") {
                     messages.push({
                         role: msg.role,
@@ -2430,13 +2564,10 @@ export class GoogleVertexProvider extends BaseProvider {
             type: "text",
             text: multimodalInput.text,
         });
-        // Add the user message with appropriate content format
-        messages.push({
-            role: "user",
-            content: userContentParts.length === 1 && userContentParts[0].type === "text"
-                ? multimodalInput.text
-                : userContentParts,
-        });
+        // Append the live user turn (merges into a trailing user turn if present).
+        appendUserMessage(messages, userContentParts.length === 1 && userContentParts[0].type === "text"
+            ? multimodalInput.text
+            : userContentParts);
         // Convert tools to Anthropic format if present
         let tools;
         const executeMap = new Map();
@@ -3056,20 +3187,14 @@ export class GoogleVertexProvider extends BaseProvider {
         // Build messages from input
         const messages = [];
         const inputText = options.prompt || options.input?.text || "Please respond.";
-        // Add conversation history if present. Prefer `conversationMessages`
-        // (what NeuroLink core injects today via MessageBuilder) and fall back
-        // to the legacy `conversationHistory` field for callers that still use
-        // the older surface. The Vertex Claude STREAM path already follows this
-        // priority — keeping the GENERATE path on `conversationHistory` only
-        // would silently drop multi-turn context for memory/loop sessions.
-        // Intentionally text-only: see the stream sibling for the rationale —
-        // synthesising tool_use / tool_result blocks from stored ChatMessages
-        // risks emitting orphaned references that Anthropic's API rejects.
-        const historyMessages = options.conversationMessages && options.conversationMessages.length > 0
-            ? options.conversationMessages
-            : options.conversationHistory;
-        if (historyMessages && historyMessages.length > 0) {
-            for (const msg of historyMessages) {
+        // Replay conversationMessages (with tool turns), else the legacy text-only conversationHistory.
+        if (options.conversationMessages &&
+            options.conversationMessages.length > 0) {
+            messages.push(...buildAnthropicHistoryMessages(options.conversationMessages));
+        }
+        else if (options.conversationHistory &&
+            options.conversationHistory.length > 0) {
+            for (const msg of options.conversationHistory) {
                 if (msg.role === "user" || msg.role === "assistant") {
                     messages.push({
                         role: msg.role,
@@ -3201,13 +3326,10 @@ export class GoogleVertexProvider extends BaseProvider {
             type: "text",
             text: inputText,
         });
-        // Add the user message with appropriate content format
-        messages.push({
-            role: "user",
-            content: userContentParts.length === 1 && userContentParts[0].type === "text"
-                ? inputText
-                : userContentParts,
-        });
+        // Append the live user turn (merges into a trailing user turn if present).
+        appendUserMessage(messages, userContentParts.length === 1 && userContentParts[0].type === "text"
+            ? inputText
+            : userContentParts);
         // Convert tools to Anthropic format if present
         let tools;
         const executeMap = new Map();

package/dist/lib/routing/classifierRouter.d.ts

@@ -0,0 +1,52 @@
+/**
+ * ClassifierRouter — the single "classify → pick model + tools → run" engine.
+ *
+ * Given a request snapshot it (1) classifies difficulty (heuristic or a cheap
+ * LLM), (2) selects the best provider/model from the host-declared base pool —
+ * cheaper/faster for easy tiers, more capable for hard tiers — enriching pool
+ * members with real cost/quality metadata from the model registry, and
+ * (3) narrows the tool set via per-difficulty directives or classifier hints.
+ *
+ * Pure of provider imports (mirrors ModelPool): the LLM caller is injected.
+ * The only data dependency is the read-only `ModelResolver` registry lookup,
+ * used purely for metadata enrichment and never for provider construction.
+ */
+import type { ClassifierRouterConfig, ClassifierRouterDecision, ClassifierRouterDeps, ClassifierRouterInput } from "../types/index.js";
+export declare class ClassifierRouter {
+    private readonly config;
+    private readonly deps;
+    private readonly metaCache;
+    constructor(config: ClassifierRouterConfig, deps?: ClassifierRouterDeps);
+    /**
+     * Classify the request and produce a combined model + tool decision, or
+     * `null` when nothing should change. Never throws (fails open).
+     */
+    route(input: ClassifierRouterInput): Promise<ClassifierRouterDecision | null>;
+    /** Run the configured strategy; LLM falls back to heuristic on failure. */
+    private classify;
+    /**
+     * Build LLM-facing model descriptors + an id→member map so the classifier can
+     * pick a model directly. Works for ANY pool, registry-backed or not.
+     */
+    private buildCandidates;
+    /**
+     * Resolve the ranked model list for a decision. Prefers the LLM's direct pick
+     * (`selectedModelId`); otherwise falls back to difficulty-based tierMap /
+     * metadata scoring.
+     */
+    private selectModels;
+    /** Candidate members for a difficulty: explicit tierMap or eligible pool. */
+    private candidatesFor;
+    /** Drop members that cannot satisfy the required capabilities (lenient). */
+    private filterByCapabilities;
+    /** Rank candidates best-first for the difficulty's strategy. */
+    private rank;
+    /** Tool narrowing: per-difficulty directive, then classifier hints. */
+    private selectTools;
+    /**
+     * Resolve cost/quality/capabilities for a member. Declared values win;
+     * gaps are filled from the model registry (by model name/alias) when known.
+     * Results are cached per (provider/model) for the router's lifetime.
+     */
+    private metaFor;
+}

package/dist/lib/routing/classifierRouter.js

@@ -0,0 +1,269 @@
+/**
+ * ClassifierRouter — the single "classify → pick model + tools → run" engine.
+ *
+ * Given a request snapshot it (1) classifies difficulty (heuristic or a cheap
+ * LLM), (2) selects the best provider/model from the host-declared base pool —
+ * cheaper/faster for easy tiers, more capable for hard tiers — enriching pool
+ * members with real cost/quality metadata from the model registry, and
+ * (3) narrows the tool set via per-difficulty directives or classifier hints.
+ *
+ * Pure of provider imports (mirrors ModelPool): the LLM caller is injected.
+ * The only data dependency is the read-only `ModelResolver` registry lookup,
+ * used purely for metadata enrichment and never for provider construction.
+ */
+import { classifyHeuristic, classifyLlm } from "./classifierStrategies.js";
+import { ModelResolver } from "../models/modelResolver.js";
+/** Maps a registry quality enum to a comparable numeric score. */
+const QUALITY_SCORE = { high: 3, medium: 2, low: 1 };
+/** How each difficulty ranks candidate members. */
+const DIFFICULTY_RANK_MODE = {
+    trivial: "cost-asc",
+    simple: "cost-asc",
+    moderate: "balanced",
+    hard: "quality-desc",
+    expert: "quality-desc",
+};
+/** Neutral default for an unknown numeric metric (keeps ordering stable). */
+const NEUTRAL = 0.5;
+/**
+ * Cap on the per-(provider,model) metadata cache so it can't grow unbounded in
+ * long-lived processes with large or dynamic pools. FIFO eviction.
+ */
+const MAX_META_CACHE_ENTRIES = 1000;
+export class ClassifierRouter {
+    config;
+    deps;
+    metaCache = new Map();
+    constructor(config, deps = {}) {
+        this.config = config;
+        this.deps = deps;
+    }
+    /**
+     * Classify the request and produce a combined model + tool decision, or
+     * `null` when nothing should change. Never throws (fails open).
+     */
+    async route(input) {
+        try {
+            // For the LLM strategy, hand the pool to the classifier so it can select
+            // a model directly — the generic path for custom/registry-less models.
+            const useLlm = this.config.classifier === "llm" && !!this.deps.generate;
+            const built = useLlm ? this.buildCandidates() : undefined;
+            const decision = await this.classify(input, built?.descriptors);
+            const ranked = this.selectModels(decision, built?.byId);
+            const primary = ranked[0];
+            const { toolFilter, excludeTools } = this.selectTools(decision);
+            if (!primary && !toolFilter && !excludeTools) {
+                return null;
+            }
+            return {
+                provider: primary?.provider,
+                model: primary?.model,
+                region: primary?.region,
+                difficulty: decision.difficulty,
+                modelFallbacks: ranked.slice(1),
+                toolFilter,
+                excludeTools,
+                reason: decision.reason,
+            };
+        }
+        catch (err) {
+            this.deps.logger?.warn?.("[ClassifierRouter] route failed — no-op", {
+                error: err instanceof Error ? err.message : String(err),
+            });
+            return null;
+        }
+    }
+    /** Run the configured strategy; LLM falls back to heuristic on failure. */
+    async classify(input, candidates) {
+        if (this.config.classifier === "llm" && this.deps.generate) {
+            try {
+                return await classifyLlm(input, this.deps.generate, this.config.classifierModel, this.config.timeoutMs, candidates);
+            }
+            catch (err) {
+                this.deps.logger?.warn?.("[ClassifierRouter] LLM classify failed — using heuristic", { error: err instanceof Error ? err.message : String(err) });
+            }
+        }
+        return classifyHeuristic(input);
+    }
+    /**
+     * Build LLM-facing model descriptors + an id→member map so the classifier can
+     * pick a model directly. Works for ANY pool, registry-backed or not.
+     */
+    buildCandidates() {
+        const pool = this.config.pool ?? [];
+        const byId = new Map();
+        const used = new Set();
+        const descriptors = [];
+        pool.forEach((m, i) => {
+            let id = m.id ?? (m.model ? `${m.provider}/${m.model}` : m.provider);
+            if (used.has(id)) {
+                id = `${id}#${i}`;
+            }
+            used.add(id);
+            byId.set(id, m);
+            descriptors.push({
+                id,
+                provider: m.provider,
+                model: m.model,
+                description: m.description,
+                tiers: m.tiers,
+                capabilities: this.metaFor(m).capabilities,
+            });
+        });
+        return { descriptors, byId };
+    }
+    /**
+     * Resolve the ranked model list for a decision. Prefers the LLM's direct pick
+     * (`selectedModelId`); otherwise falls back to difficulty-based tierMap /
+     * metadata scoring.
+     */
+    selectModels(decision, byId) {
+        if (decision.selectedModelId && byId) {
+            const picked = byId.get(decision.selectedModelId);
+            if (picked) {
+                const rest = (this.config.pool ?? []).filter((m) => m !== picked);
+                const rankedRest = this.rank(this.filterByCapabilities(rest, decision.requiredCapabilities), decision.difficulty);
+                return [picked, ...rankedRest];
+            }
+        }
+        const eligible = this.candidatesFor(decision);
+        const capable = this.filterByCapabilities(eligible, decision.requiredCapabilities);
+        return this.rank(capable, decision.difficulty);
+    }
+    /** Candidate members for a difficulty: explicit tierMap or eligible pool. */
+    candidatesFor(decision) {
+        const tierMembers = this.config.tierMap?.[decision.difficulty];
+        if (tierMembers && tierMembers.length > 0) {
+            return tierMembers;
+        }
+        const pool = this.config.pool ?? [];
+        const eligible = pool.filter((m) => !m.tiers || m.tiers.includes(decision.difficulty));
+        return eligible.length > 0 ? eligible : pool;
+    }
+    /** Drop members that cannot satisfy the required capabilities (lenient). */
+    filterByCapabilities(members, required) {
+        if (!required || required.length === 0) {
+            return members;
+        }
+        const kept = members.filter((m) => {
+            const caps = this.metaFor(m).capabilities;
+            // Unknown capabilities → keep (don't starve the pool on missing metadata).
+            if (!caps || caps.length === 0) {
+                return true;
+            }
+            return required.every((c) => caps.includes(c));
+        });
+        return kept.length > 0 ? kept : members;
+    }
+    /** Rank candidates best-first for the difficulty's strategy. */
+    rank(members, difficulty) {
+        const mode = DIFFICULTY_RANK_MODE[difficulty];
+        const originalIndex = new Map(members.map((m, i) => [m, i]));
+        const num = (v) => (typeof v === "number" ? v : NEUTRAL);
+        return [...members].sort((a, b) => {
+            const ma = this.metaFor(a);
+            const mb = this.metaFor(b);
+            let delta;
+            if (mode === "cost-asc") {
+                delta = num(ma.cost) - num(mb.cost);
+            }
+            else if (mode === "quality-desc") {
+                delta = num(mb.quality) - num(ma.quality);
+            }
+            else {
+                // balanced: maximize quality-minus-cost
+                delta =
+                    num(mb.quality) - num(mb.cost) - (num(ma.quality) - num(ma.cost));
+            }
+            if (delta !== 0) {
+                return delta;
+            }
+            const weightDelta = (b.weight ?? 1) - (a.weight ?? 1);
+            if (weightDelta !== 0) {
+                return weightDelta;
+            }
+            // Stable: preserve declared pool order on a tie.
+            return (originalIndex.get(a) ?? 0) - (originalIndex.get(b) ?? 0);
+        });
+    }
+    /** Tool narrowing: per-difficulty directive, then classifier hints. */
+    selectTools(decision) {
+        const directive = this.config.toolDirectives?.[decision.difficulty];
+        let toolFilter = directive?.toolFilter
+            ? [...directive.toolFilter]
+            : undefined;
+        const excludeTools = directive?.excludeTools
+            ? [...directive.excludeTools]
+            : undefined;
+        // When no explicit allowlist is configured, honor the classifier's hint.
+        if ((!toolFilter || toolFilter.length === 0) &&
+            decision.suggestedTools &&
+            decision.suggestedTools.length > 0) {
+            toolFilter = [...decision.suggestedTools];
+        }
+        return { toolFilter, excludeTools };
+    }
+    /**
+     * Resolve cost/quality/capabilities for a member. Declared values win;
+     * gaps are filled from the model registry (by model name/alias) when known.
+     * Results are cached per (provider/model) for the router's lifetime.
+     */
+    metaFor(member) {
+        const key = `${member.provider}::${member.model ?? ""}`;
+        const cached = this.metaCache.get(key);
+        if (cached) {
+            return cached;
+        }
+        let cost = member.cost;
+        let quality = member.quality;
+        let capabilities = member.capabilities
+            ? [...member.capabilities]
+            : undefined;
+        const needsEnrichment = cost === undefined || quality === undefined || capabilities === undefined;
+        if (needsEnrichment && member.model) {
+            try {
+                const info = ModelResolver.resolveModel(member.model);
+                if (info) {
+                    if (cost === undefined) {
+                        cost = info.pricing.inputCostPer1K + info.pricing.outputCostPer1K;
+                    }
+                    if (quality === undefined) {
+                        quality = QUALITY_SCORE[info.performance.quality] ?? NEUTRAL;
+                    }
+                    if (capabilities === undefined) {
+                        const caps = [];
+                        if (info.capabilities.vision) {
+                            caps.push("vision");
+                        }
+                        if (info.capabilities.functionCalling) {
+                            caps.push("tools");
+                        }
+                        if (info.capabilities.reasoning) {
+                            caps.push("reasoning");
+                        }
+                        if (info.capabilities.codeGeneration) {
+                            caps.push("code");
+                        }
+                        if (info.capabilities.multimodal) {
+                            caps.push("multimodal");
+                        }
+                        capabilities = caps;
+                    }
+                }
+            }
+            catch {
+                // Enrichment is best-effort; ignore registry lookup failures.
+            }
+        }
+        const meta = { cost, quality, capabilities };
+        if (this.metaCache.size >= MAX_META_CACHE_ENTRIES) {
+            const oldest = this.metaCache.keys().next().value;
+            if (oldest !== undefined) {
+                this.metaCache.delete(oldest);
+            }
+        }
+        this.metaCache.set(key, meta);
+        return meta;
+    }
+}
+//# sourceMappingURL=classifierRouter.js.map

package/dist/lib/routing/classifierStrategies.d.ts

@@ -0,0 +1,23 @@
+/**
+ * Classification strategies for the ClassifierRouter.
+ *
+ * - `classifyHeuristic` — zero-cost, no LLM. Generalizes the existing binary
+ *   task classifier scorer (fast vs reasoning) into five difficulty tiers.
+ * - `classifyLlm` — runs a cheap "classifier model" via an injected generate
+ *   function, asking for a schema-constrained difficulty verdict. Falls back to
+ *   the heuristic on any failure.
+ */
+import type { ClassifierCandidate, ClassifierDecision, ClassifierDifficulty, ClassifierGenerateFn, ClassifierModelRef, ClassifierRouterInput } from "../types/index.js";
+/** Difficulty tiers, ordered easiest → hardest. */
+export declare const CLASSIFIER_DIFFICULTIES: ClassifierDifficulty[];
+/**
+ * Heuristic classifier — maps the binary fast/reasoning scores plus prompt
+ * length into one of five difficulty tiers. Deterministic and dependency-free.
+ */
+export declare function classifyHeuristic(input: ClassifierRouterInput): ClassifierDecision;
+/**
+ * LLM classifier — asks a cheap model for a schema-constrained verdict.
+ * Falls back to the heuristic if the model is unavailable, times out, or
+ * returns output that does not match the schema.
+ */
+export declare function classifyLlm(input: ClassifierRouterInput, generate: ClassifierGenerateFn, classifierModel?: ClassifierModelRef, timeoutMs?: number, candidates?: ClassifierCandidate[]): Promise<ClassifierDecision>;