@remnic/core 9.3.613 → 9.3.614

package/src/extraction-judge-chain.test.ts

@@ -0,0 +1,55 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import { parseConfig } from "./config.js";
+import { judgeFactDurability } from "./extraction-judge.js";
+import type { FallbackLlmClient } from "./fallback-llm.js";
+
+// Regression for cursor review on #1425: the extraction judge must route gateway
+// fallback calls through the SAME precedence as ExtractionEngine.withGatewayAgent
+// — taskModelChain wins over gatewayAgentId — so judge-gated extractions use the
+// configured task chain instead of silently using the persona/default chain.
+
+function captureLlm(captured: Array<Record<string, unknown>>): FallbackLlmClient {
+  return {
+    isAvailable: () => true,
+    chatCompletion: async (_messages: unknown, options: Record<string, unknown>) => {
+      captured.push(options);
+      // A parseable batch verdict so the judge doesn't error after the call.
+      return { content: JSON.stringify([{ index: 0, durable: true, reason: "ok" }]) };
+    },
+  } as unknown as FallbackLlmClient;
+}
+
+// "preference" is not auto-approved (only "correction"/"principle" are), so a
+// non-critical preference candidate reaches the LLM path.
+const candidate = { text: "user prefers dark mode", category: "preference", confidence: 0.5 };
+
+test("judge forwards taskModelChain to the fallback LLM in gateway mode", async () => {
+  const config = parseConfig({
+    modelSource: "gateway",
+    gatewayAgentId: "persona-agent",
+    taskModelChain: { primary: "zai/glm-4.7-flash", fallbacks: ["fireworks/x/glm-5p1"] },
+  });
+  const captured: Array<Record<string, unknown>> = [];
+
+  await judgeFactDurability([candidate], config, null, captureLlm(captured), new Map(), new Map());
+
+  assert.equal(captured.length, 1, "judge should call the fallback LLM once");
+  assert.deepEqual(captured[0]?.modelChain, {
+    primary: "zai/glm-4.7-flash",
+    fallbacks: ["fireworks/x/glm-5p1"],
+  });
+  assert.equal(captured[0]?.agentId, undefined, "taskModelChain takes precedence over gatewayAgentId");
+});
+
+test("judge falls back to gatewayAgentId when no taskModelChain is set", async () => {
+  const config = parseConfig({ modelSource: "gateway", gatewayAgentId: "persona-agent" });
+  const captured: Array<Record<string, unknown>> = [];
+
+  await judgeFactDurability([candidate], config, null, captureLlm(captured), new Map(), new Map());
+
+  assert.equal(captured.length, 1);
+  assert.equal(captured[0]?.agentId, "persona-agent");
+  assert.equal(captured[0]?.modelChain, undefined);
+});

package/src/extraction-judge.ts

@@ -17,7 +17,7 @@ import { createHash } from "node:crypto";
 import { log } from "./logger.js";
 import type { PluginConfig, ImportanceLevel } from "./types.js";
 import type { LocalLlmClient } from "./local-llm.js";
-import type { FallbackLlmClient } from "./fallback-llm.js";
+import { type FallbackLlmClient, gatewayTaskChainOptions } from "./fallback-llm.js";
 import { extractJsonCandidates } from "./json-extract.js";
 import { normalizeProcedureSteps } from "./procedural/procedure-types.js";
 
@@ -656,13 +656,11 @@ async function callJudgeLlm(
   // routing preference.
   const skipLocal = config.modelSource === "gateway";
 
-  // Resolve the gateway agent ID so the fallback LLM routes through the
-  // correct agent persona's model chain — identical to the pattern used
-  // by ExtractionEngine.withGatewayAgent().
-  const agentId =
-    config.modelSource === "gateway"
-      ? (config.gatewayAgentId || undefined)
-      : undefined;
+  // Route judge-gated extractions through the SAME shared resolution as every
+  // other background task (taskModelChain > gatewayAgentId in gateway mode), so
+  // the judge never silently falls back to the persona/default chain when a
+  // task chain is configured (gotcha #22, #39). Issue #1365 / PR #1425.
+  const gatewayChain = gatewayTaskChainOptions(config);
 
   // Try local LLM first (unless modelSource says gateway)
   if (localLlm && !skipLocal) {
@@ -695,7 +693,7 @@ async function callJudgeLlm(
           maxTokens: 2048,
           timeoutMs: 1500,
           ...(modelOverride ? { model: modelOverride } : {}),
-          ...(agentId ? { agentId } : {}),
+          ...gatewayChain,
         },
       );
       if (result?.content) {

package/src/extraction.ts

@@ -2,7 +2,7 @@ import OpenAI from "openai";
 import { log } from "./logger.js";
 import { delinearize } from "./delinearize.js";
 import { LocalLlmClient } from "./local-llm.js";
-import { FallbackLlmClient, fallbackLlmRuntimeContextFromConfig } from "./fallback-llm.js";
+import { FallbackLlmClient, fallbackLlmRuntimeContextFromConfig, gatewayTaskChainOptions } from "./fallback-llm.js";
 import {
   ExtractionResultSchema,
   ConsolidationResultSchema,
@@ -123,7 +123,16 @@ export class ExtractionEngine {
     if (config.modelSource === "gateway") {
       log.debug(
         `extraction engine: gateway model source active; extraction uses the gateway chain as its primary path` +
-          (config.gatewayAgentId ? ` (agent: ${config.gatewayAgentId})` : " (defaults)"),
+          (config.taskModelChain ? " (taskModelChain)" :
+            config.gatewayAgentId ? ` (agent: ${config.gatewayAgentId})` : " (defaults)"),
+      );
+    } else if (config.taskModelChain) {
+      // taskModelChain resolves through gateway providers, so it only applies
+      // under modelSource: "gateway". Warn rather than silently ignore it so a
+      // misconfigured plugin-mode setup is visible. Issue #1365 / PR #1370.
+      log.warn(
+        `taskModelChain is set but modelSource is "${config.modelSource}"; the chain is ignored. ` +
+          `Set modelSource: "gateway" to use it for extraction/consolidation/summarization.`,
       );
     }
   }
@@ -157,8 +166,9 @@ export class ExtractionEngine {
    */
   private withGatewayAgent(options: import("./fallback-llm.js").FallbackLlmOptions): import("./fallback-llm.js").FallbackLlmOptions {
     if (!this.useGatewayModelSource) return options;
-    const agentId = this.config.gatewayAgentId || undefined;
-    return agentId ? { ...options, agentId } : options;
+    // Shared resolution (taskModelChain > gatewayAgentId) so every background
+    // task routes identically (gotcha #22). Issue #1365.
+    return { ...options, ...gatewayTaskChainOptions(this.config) };
   }
 
   private emit(event: LlmTraceEvent): void {
@@ -1098,7 +1108,8 @@ export class ExtractionEngine {
     if (this.useGatewayModelSource) {
       log.debug(
         `extraction: using gateway model chain as primary path` +
-          (this.config.gatewayAgentId ? ` (agent: ${this.config.gatewayAgentId})` : " (defaults)"),
+          (this.config.taskModelChain ? " (taskModelChain)" :
+            this.config.gatewayAgentId ? ` (agent: ${this.config.gatewayAgentId})` : " (defaults)"),
       );
     } else {
       log.info("extraction: falling back to gateway default AI");