@exulu/backend 3.7.4 → 4.0.0

package/dist/{chunk-27K2CO47.js → chunk-QMN6MVHQ.js}

@@ -171,7 +171,7 @@ async function validatePythonEnvironment(packageRoot, checkPackages = true) {
     };
   }
   if (checkPackages) {
-    const criticalPackages = ["docling", "transformers"];
+    const criticalPackages = ["pypdf", "transformers"];
     const missingPackages = [];
     for (const pkg of criticalPackages) {
       try {

package/dist/{chunk-AWMU6QXB.js → chunk-RBEWHG7I.js}

@@ -1716,7 +1716,7 @@ var ExuluTool = class _ExuluTool {
     if (!agent) {
       throw new Error("Agent not found.");
     }
-    const { convertExuluToolsToAiSdkTools: convertExuluToolsToAiSdkTools2 } = await import("./convert-exulu-tools-to-ai-sdk-tools-XNQ6Q3X6.js");
+    const { convertExuluToolsToAiSdkTools: convertExuluToolsToAiSdkTools2 } = await import("./convert-exulu-tools-to-ai-sdk-tools-6RU4IZMI.js");
     const tools = await convertExuluToolsToAiSdkTools2(
       [this],
       [],
@@ -2068,7 +2068,19 @@ var tuningSchema = z3.object({
   pinBoost: z3.number().min(0).max(1).default(0.15),
   identifierBoost: z3.number().min(0).max(1).default(0.15),
   pageWindow: z3.number().int().min(0).default(1),
-  maxQueriesPerContext: z3.number().int().positive().default(5)
+  maxQueriesPerContext: z3.number().int().positive().default(5),
+  /**
+   * Orchestration engine. "v1" is the sequential flow every agent ran before 2026-09;
+   * "v2" merges the phase-1 LLM hops and runs identifier pins in parallel. Per agent,
+   * so a candidate agent can run v2 while the production agent stays on v1.
+   */
+  engine: z3.enum(["v1", "v2"]).default("v1"),
+  /** v2 sub-features; each can be switched off on its own to bisect a regression. */
+  v2: z3.object({
+    mergedMemoryCall: z3.boolean().default(true),
+    mergedRoutingCall: z3.boolean().default(true),
+    parallelPins: z3.boolean().default(true)
+  }).default({ mergedMemoryCall: true, mergedRoutingCall: true, parallelPins: true })
 });
 var boolVal = (v) => v === true || v === "true" || v === 1;
 var strVal = (v, fallback) => typeof v === "string" && v.length > 0 ? v : fallback;
@@ -2291,6 +2303,17 @@ async function microCall(args) {
   );
 }
 
+// ee/agentic-retrieval/pipeline/timing.ts
+async function withTiming(sink, key, work, now = Date.now) {
+  if (!sink) return work();
+  const started = now();
+  try {
+    return await work();
+  } finally {
+    sink[key] = Math.max(0, now() - started);
+  }
+}
+
 // ee/agentic-retrieval/pipeline/prefilter.ts
 import Fuse from "fuse.js";
 import { z as z4 } from "zod";
@@ -2606,6 +2629,49 @@ Only set pageNumber when the user names a specific page.
 Return hasFilenameHint/hasPageHint false (and omit the hint fields) when neither
 is present. When in doubt, return false.
 `;
+async function runMergedRoutingCall({
+  model,
+  question,
+  knownIdentifiers,
+  enabledContexts,
+  routingRules,
+  extraInstructions
+}) {
+  const kbListing = enabledContexts.map((c) => `- ${c.id}: ${c.name}${c.description ? " \u2014 " + c.description : ""}`).join("\n");
+  const rulesLines = routingRules.map((r) => `- ${r.id} (${r.label}): ${r.description}`).join("\n");
+  const system = `You analyse the user's request and answer ${routingRules.length ? "three" : "two"} questions in one go.
+
+A. DOCUMENT / PAGE REFERENCE (docPage):
+${buildDocPagePrompt(knownIdentifiers)}
+
+B. EXPLICIT KNOWLEDGE BASE REQUEST (explicitlyRequestedKnowledgeBases): check if the user has EXPLICITLY asked you to search in one or multiple of the following knowledge bases:
+${kbListing}
+EXPLICIT means the user names a knowledge base or clearly commands searching a specific source (e.g. "search in the tickets", "look this up in the manuals KB"). A question that merely CONCERNS a topic related to a knowledge base's name or contents (e.g. asking about software changes, norms, or a product) is NOT an explicit request. When in doubt, return an empty array. If explicit, return the knowledge base ids.
+` + (routingRules.length ? `
+C. CLASSIFICATION (classification): classify the request into exactly one of these categories:
+${rulesLines}` + (extraInstructions ? `
+<instructions>
+${extraInstructions}
+</instructions>` : "") : "");
+  const ids = enabledContexts.map((c) => c.id);
+  const schema = z5.object({
+    docPage: z5.object({
+      hasFilenameHint: z5.boolean(),
+      filenameHints: z5.array(z5.string()).optional(),
+      hasPageHint: z5.boolean(),
+      pageNumber: z5.number().int().nullable().optional()
+    }),
+    explicitlyRequestedKnowledgeBases: z5.array(z5.enum(ids)),
+    ...routingRules.length ? {
+      classification: z5.object({
+        ruleId: z5.enum(routingRules.map((r) => r.id)),
+        reason: z5.string()
+      })
+    } : {}
+  });
+  const { output } = await microCall({ model, system, messages: [{ role: "user", content: question }], schema });
+  return output;
+}
 async function runRoutingPhase(opts) {
   const {
     question,
@@ -2615,6 +2681,7 @@ async function runRoutingPhase(opts) {
     preselectedItems,
     extraInstructions,
     knownIdentifiers = [],
+    mergedCall = false,
     model
   } = opts;
   try {
@@ -2634,7 +2701,23 @@ async function runRoutingPhase(opts) {
 ` + enabledContexts.map((c) => `- ${c.id}: ${c.name}${c.description ? " \u2014 " + c.description : ""}`).join("\n") + `
 EXPLICIT means the user names a knowledge base or clearly commands searching a specific source (e.g. "search in the tickets", "look this up in the manuals KB"). A question that merely CONCERNS a topic related to a knowledge base's name or contents (e.g. asking about software changes, norms, or a product) is NOT an explicit request \u2014 classification routing handles those. When in doubt, return an empty array.
 If explicit, return the knowledge base ids. If not, return an empty array.`;
-    const [docPageRaw, explicitKBRaw] = await Promise.all([
+    const wantsClassification = routingRules.length > 0 && preselectedItems.size === 0;
+    const merged = mergedCall ? await withTiming(
+      opts.timings,
+      "routing.mergedMs",
+      () => runMergedRoutingCall({
+        model,
+        question,
+        knownIdentifiers,
+        enabledContexts,
+        routingRules: wantsClassification ? routingRules : [],
+        extraInstructions
+      }).catch((err) => {
+        console.warn("[EXULU pipeline] merged routing call failed \u2014 falling back to the v1 hops.", err);
+        return null;
+      })
+    ) : null;
+    const [docPageRaw, explicitKBRaw] = merged ? [{ output: merged.docPage }, { output: { explicitlyRequestedKnowledgeBases: merged.explicitlyRequestedKnowledgeBases } }] : await withTiming(opts.timings, "routing.detectMs", () => Promise.all([
       (async () => {
         try {
           return await microCall({
@@ -2676,7 +2759,7 @@ If explicit, return the knowledge base ids. If not, return an empty array.`;
           return { output: { explicitlyRequestedKnowledgeBases: [] } };
         }
       })()
-    ]);
+    ]));
     const userPinnedItemIdsByContext = /* @__PURE__ */ new Map();
     let userRequestedPage = null;
     if (docPageRaw.output.hasFilenameHint && docPageRaw.output.filenameHints?.length) {
@@ -2754,7 +2837,7 @@ ${extraInstructions}
 </instructions>`;
       }
       try {
-        const { output: classified } = await microCall({
+        const { output: classified } = merged?.classification ? { output: merged.classification } : await withTiming(opts.timings, "routing.classifyMs", () => microCall({
           model,
           system: classifyPrompt,
           messages: [{ role: "user", content: question }],
@@ -2762,7 +2845,7 @@ ${extraInstructions}
             ruleId: z5.enum(ruleIds),
             reason: z5.string()
           })
-        });
+        }));
         const matchedRule = routingRules.find((r) => r.id === classified.ruleId);
         if (matchedRule) {
           const main = matchedRule.main.filter((id) => enabledIds.has(id));
@@ -2929,7 +3012,8 @@ async function recallMemoryByKeywords({
   importantKeyword,
   user,
   role,
-  memoryContext
+  memoryContext,
+  timings
 }) {
   const allKeywords = [
     ...new Set(
@@ -2946,7 +3030,7 @@ async function recallMemoryByKeywords({
     ...new Set(allKeywords.flatMap(deriveKeywordVariants).map(stripSeparators))
   ].filter((v) => v.length >= 4);
   if (!allVariants.length) return [];
-  const items = await loadMemoryItems(memoryContext);
+  const items = await withTiming(timings, "memory.keywordRecall.itemsMs", () => loadMemoryItems(memoryContext));
   const scored = [];
   for (const item of items) {
     const haystack = stripSeparators(
@@ -2968,14 +3052,14 @@ async function recallMemoryByKeywords({
     "[EXULU pipeline] keyword-triggered memory matches:",
     topMatches.map((s) => `${s.name} (hits=${s.hits}, important=${s.importantHit})`)
   );
-  const chunks = await singleSearch({
+  const chunks = await withTiming(timings, "memory.keywordRecall.searchMs", () => singleSearch({
     query: allKeywords.join(", "),
-    config: { method: "hybridSearch", cutoffs: void 0, limit: 50 },
+    config: { method: "tsvector", cutoffs: void 0, limit: 50 },
     user,
     role,
     pinnedItemIds: topMatches.map((s) => s.id),
     context: memoryContext
-  });
+  }));
   return chunks;
 }
 function neutralResult(question, keywords, importantKeyword, steps = []) {
@@ -2989,7 +3073,103 @@ function neutralResult(question, keywords, importantKeyword, steps = []) {
     steps
   };
 }
+async function runMergedMemoryCall({
+  model,
+  retrievedMemory,
+  question,
+  keywords,
+  importantKeyword,
+  memoryConfig,
+  glossary
+}) {
+  const glossaryText = glossary.length > 0 ? `
+The organization's documents use the following abbreviations/terms:
+${glossary.map((g) => `${g.term} : ${g.meaning}`).join("\n")}` : "";
+  const system = `
+    You review the shared company memory for the user's question and answer FOUR questions in one go.
+
+    1. RELEVANCE (relevantChunkIds): return the chunk_ids of chunks containing information relevant to the
+    question, or an empty array. Be generous: include chunks that are topically related, share key
+    terminology, describe the same symptom from a different angle, or could plausibly help diagnose the
+    issue \u2014 even if they don't answer the question directly. Memory entries are deliberately broad,
+    hand-curated hints written by domain experts; the user's wording will rarely match the memory verbatim.
+    When in doubt, include the chunk.
+
+    2. OVERRIDE (override): decide whether ONE of the relevant chunks should become the AUTHORITATIVE basis
+    of the answer, taking precedence over the official documentation even if the documents state something
+    different. This is a deliberately STRICT check. Set overrides=true ONLY if a single chunk, on its own,
+    contains a DIRECT and SUFFICIENT answer to exactly what the user asked. Being topically related, sharing
+    terminology, describing the same component, or only partially addressing the question is NOT sufficient:
+    then set overrides=false. When in doubt, set overrides=false. Memory entries may capture field experience
+    that the manuals get wrong, so a confident, direct match is meant to win over the documents.
+    ${memoryConfig.override ? "" : "(Override is disabled for this agent: return overrides=false.)"}
+
+    3. FILE PRIORITIZATION (filePrioritization): set shouldPrioritizeFiles=true only if a relevant memory
+    entry explicitly says to look in, prioritize, prefer, or always search a particular document, file, or
+    file family (for example "When asked about X, always search in Y-Dateien first"). General background
+    facts, glossaries, or synonyms are NOT a file prioritization instruction. When true, return
+    fileNameHints exactly as referenced in the memory, bare names without folder paths.
+    ${memoryConfig.filePrioritization ? "" : "(File prioritization is disabled for this agent: return false.)"}
+
+    4. QUERY AUGMENTATION (augmentation): if, and only if, the relevant memory (or the glossary below)
+    contains synonyms or similar terms for what the user asked, return the user question and keywords
+    updated to include those synonyms \u2014 always keeping the original wording as well. Otherwise return the
+    original question, keywords and important keyword unchanged.
+    ${memoryConfig.queryAugmentation ? "" : "(Query augmentation is disabled for this agent: return the originals.)"}
+
+    <memory_chunks>
+    ${retrievedMemory.map((chunk) => `- ${chunk.chunk_id}: ${chunk.item_name} - ${chunk.chunk_content}`).join("\n")}
+    </memory_chunks>
+    ${glossaryText}
+  `;
+  const { output } = await microCall({
+    model,
+    system,
+    messages: [
+      {
+        role: "user",
+        content: `
+        <user_question>${question}</user_question>
+        <relevant_keywords>${keywords.join(", ")}</relevant_keywords>
+        <important_keyword>${importantKeyword}</important_keyword>
+        `
+      }
+    ],
+    schema: z6.object({
+      relevantChunkIds: z6.array(z6.string()).describe("chunk_ids (UUIDs at the start of each bullet) of relevant chunks; empty array if none."),
+      override: z6.object({
+        overrides: z6.boolean().describe("True ONLY if a chunk directly and sufficiently answers the question."),
+        confidence: z6.enum(["high", "medium", "low"]),
+        authoritativeChunkIds: z6.array(z6.string()).describe("chunk_ids that directly answer the question; empty if overrides is false."),
+        reason: z6.string().describe("One short sentence.")
+      }),
+      filePrioritization: z6.object({
+        shouldPrioritizeFiles: z6.boolean(),
+        fileNameHints: z6.array(z6.string()).optional()
+      }),
+      augmentation: z6.object({
+        updatedUserQuestion: z6.string(),
+        updatedRelevantKeywords: z6.array(z6.string()),
+        updatedImportantKeyword: z6.string()
+      })
+    })
+  });
+  return output;
+}
+function mergedFollowups(merged, memoryConfig, hasAugmentationContent, question, importantKeyword) {
+  const overrideResult = {
+    output: memoryConfig.override ? merged.override : { overrides: false, confidence: "low", authoritativeChunkIds: [], reason: "" }
+  };
+  const fileResult = {
+    output: memoryConfig.filePrioritization ? merged.filePrioritization : { shouldPrioritizeFiles: false, fileNameHints: [] }
+  };
+  const queryResult = {
+    output: memoryConfig.queryAugmentation && hasAugmentationContent ? merged.augmentation : { updatedUserQuestion: question, updatedRelevantKeywords: [], updatedImportantKeyword: importantKeyword }
+  };
+  return [overrideResult, fileResult, queryResult];
+}
 async function runMemoryPhase({
+  timings,
   memoryChunks,
   memoryContext,
   question,
@@ -3000,7 +3180,8 @@ async function runMemoryPhase({
   model,
   memoryConfig,
   glossary,
-  documentContexts
+  documentContexts,
+  mergedCall = false
 }) {
   try {
     if (!memoryConfig.enabled || memoryChunks.length === 0 && !memoryContext) {
@@ -3010,13 +3191,14 @@ async function runMemoryPhase({
     let retrieved_memory = [...memoryChunks];
     if (memoryContext) {
       try {
-        const keywordMatched = await recallMemoryByKeywords({
+        const keywordMatched = await withTiming(timings, "memory.keywordRecallMs", () => recallMemoryByKeywords({
           keywords,
           importantKeyword,
           user,
           role,
-          memoryContext
-        });
+          memoryContext,
+          timings
+        }));
         if (keywordMatched.length > 0) {
           const seen = new Set(retrieved_memory.map((c) => c.chunk_id));
           const additions = keywordMatched.filter((c) => !seen.has(c.chunk_id));
@@ -3041,8 +3223,24 @@ async function runMemoryPhase({
     </memory_chunks>
     `;
     let relevantMemoryChunks = [];
+    let mergedOutput;
     try {
-      const { output: output_relevant_memory } = await microCall({
+      if (mergedCall) {
+        mergedOutput = await withTiming(
+          timings,
+          "memory.mergedMs",
+          () => runMergedMemoryCall({
+            model,
+            retrievedMemory: retrieved_memory,
+            question,
+            keywords,
+            importantKeyword,
+            memoryConfig,
+            glossary
+          })
+        );
+      }
+      const { output: output_relevant_memory } = mergedOutput ? { output: { relevantChunkIds: mergedOutput.relevantChunkIds } } : await withTiming(timings, "memory.relevanceMs", () => microCall({
         model,
         system: CHECK_MEMORIES_FOR_RELEVANT_INFORMATION,
         messages: [
@@ -3060,7 +3258,7 @@ async function runMemoryPhase({
             "The chunk_ids (UUIDs at the start of each bullet) of chunks containing information relevant to the user's question. Empty array if none are relevant."
           )
         })
-      });
+      }));
       const ids = new Set(output_relevant_memory?.relevantChunkIds ?? []);
       relevantMemoryChunks = ids.size === 0 ? [] : retrieved_memory.filter((c) => ids.has(c.chunk_id));
     } catch (e) {
@@ -3150,7 +3348,7 @@ ${glossary.map((g) => `${g.term} : ${g.meaning}`).join("\n")}` : "";
 
       Otherwise, return the original user question, relevant keywords and important keyword.
       `;
-      const [overrideResult, fileResult, queryResult] = await Promise.all([
+      const [overrideResult, fileResult, queryResult] = mergedOutput ? mergedFollowups(mergedOutput, memoryConfig, hasAugmentationContent, question, importantKeyword) : await withTiming(timings, "memory.followupsMs", () => Promise.all([
         // Override check: strict gate to decide if memory should be authoritative
         memoryConfig.override ? microCall({
           model,
@@ -3231,7 +3429,7 @@ ${glossary.map((g) => `${g.term} : ${g.meaning}`).join("\n")}` : "";
             updatedImportantKeyword: importantKeyword
           }
         })
-      ]);
+      ]));
       const overrideIds = new Set(overrideResult.output?.authoritativeChunkIds ?? []);
       const authoritativeChunks = memoryChunksForAnswer.filter(
         (c) => c.chunk_id && overrideIds.has(c.chunk_id)
@@ -3407,7 +3605,7 @@ async function searchContexts(opts) {
     skipPrefilter
   } = opts;
   const chunkArrays = await Promise.all(
-    contextIds.map(async (ctxId) => {
+    contextIds.map((ctxId) => withTiming(opts.timings, `${opts.timingPrefix ?? "search"}.${ctxId}Ms`, async () => {
       try {
         const ctx = contextsById.get(ctxId);
         if (!ctx) return [];
@@ -3445,13 +3643,13 @@ async function searchContexts(opts) {
           if (multiQuery) {
             let hydePassage = null;
             if (hyde) {
-              hydePassage = await generateHydePassage({
+              hydePassage = await withTiming(opts.timings, `${opts.timingPrefix ?? "search"}.${ctxId}.hydeMs`, () => generateHydePassage({
                 originalQuestion: question,
                 relevantKeywords: keywords,
                 importantKeyword,
                 styleHint,
                 model
-              });
+              }));
             }
             const candidates = [
               question,
@@ -3466,14 +3664,14 @@ async function searchContexts(opts) {
         }
         if (kind === "conversations") {
           if (keywordPrefilter && pinnedItemIds.length === 0) {
-            const prefiltered = await fuzzyPrefilter({
+            const prefiltered = await withTiming(opts.timings, `${opts.timingPrefix ?? "search"}.${ctxId}.prefilterMs`, () => fuzzyPrefilter({
               cacheKey: `conversations:${ctxId}`,
               relevantKeywords: keywords,
               importantKeyword,
               context: ctx,
               fields: ["name", "id", "external_id", "description"],
               normalize: (i) => [i.name, i.description].filter(Boolean).join(": ")
-            });
+            }));
             pinnedItemIds = prefiltered.map((r) => r.id);
           }
           const keywordQuery = keywords.length ? keywords.join(" ") + " " + importantKeyword : question;
@@ -3488,7 +3686,7 @@ async function searchContexts(opts) {
         console.warn(`[EXULU pipeline] searchContexts failed for context "${ctxId}":`, err);
         return [];
       }
-    })
+    }))
   );
   return { chunks: chunkArrays.flat() };
 }
@@ -3633,6 +3831,27 @@ async function rerankResults(opts) {
   return { limited_results, sorted_reranked_results: sorted, rerank_score_max_genuine };
 }
 
+// ee/agentic-retrieval/pipeline/pin-rerun.ts
+function needsPinRerun(originalQuestion, updatedQuestion) {
+  if (originalQuestion === updatedQuestion) return false;
+  const before = designations(originalQuestion);
+  for (const token of designations(updatedQuestion)) {
+    if (!before.has(token)) return true;
+  }
+  return false;
+}
+function designations(text) {
+  const out = /* @__PURE__ */ new Set();
+  for (const raw of text.split(/[\s,;:()\[\]"'?!]+/)) {
+    const token = raw.replace(/^[^\p{L}\p{N}]+|[^\p{L}\p{N}]+$/gu, "");
+    if (!token) continue;
+    const hasDigit = /\p{N}/u.test(token);
+    const capitals = (token.match(/\p{Lu}/gu) ?? []).length;
+    if (hasDigit || capitals >= 2) out.add(token.toLowerCase());
+  }
+  return out;
+}
+
 // ee/agentic-retrieval/pipeline/index.ts
 function addChunks(result, chunks) {
   const seen = new Set(result.chunks.map((c) => c.chunk_id));
@@ -3798,6 +4017,14 @@ function createAgenticRetrievalTool(opts) {
         usage: [],
         totalTokens: 0
       };
+      const t0 = Date.now();
+      const timings = {};
+      let tPhase = t0;
+      const lap = (name) => {
+        const now = Date.now();
+        timings[name] = (timings[name] ?? 0) + (now - tPhase);
+        tPhase = now;
+      };
       try {
         let enabledContexts = contexts.filter(
           (ctx) => cfg.knowledgeBases[ctx.id]?.enabled !== false
@@ -3870,8 +4097,19 @@ function createAgenticRetrievalTool(opts) {
           resolvedProject && projectScope?.customInstructions ? `Instructions for the attached project "${projectScope.name}":
 ${projectScope.customInstructions}` : ""
         ].filter(Boolean).join("\n");
-        const [memResult, routResult] = await Promise.all([
-          runMemoryPhase({
+        const engineV2 = cfg.tuning.engine === "v2";
+        const v2 = cfg.tuning.v2;
+        const pinsFor = (question) => resolveIdentifierPins({
+          question,
+          identifierSets: cfg.vocabulary.identifiers,
+          contextsById,
+          kbKindById,
+          model: utilityModel
+        });
+        const [memResult, routResult, parallelPins] = await Promise.all([
+          withTiming(timings, "memoryMs", () => runMemoryPhase({
+            timings,
+            mergedCall: engineV2 && v2.mergedMemoryCall,
             memoryChunks: memoryItems ?? [],
             memoryContext,
             question: userQuery,
@@ -3883,8 +4121,10 @@ ${projectScope.customInstructions}` : ""
             memoryConfig: cfg.memory,
             glossary: cfg.vocabulary.glossary,
             documentContexts
-          }),
-          runRoutingPhase({
+          })),
+          withTiming(timings, "routingMs", () => runRoutingPhase({
+            timings,
+            mergedCall: engineV2 && v2.mergedRoutingCall,
             question: userQuery,
             enabledContexts,
             documentContexts,
@@ -3895,8 +4135,12 @@ ${projectScope.customInstructions}` : ""
             // detector must never treat these as filename hints.
             knownIdentifiers: cfg.vocabulary.identifiers.flatMap((i) => i.examples),
             model: utilityModel
-          })
+          })),
+          // engine v2: identifier pins depend only on the question, so they run alongside
+          // memory and routing instead of after them (re-run below if memory rewrote the question).
+          engineV2 && v2.parallelPins ? withTiming(timings, "pinsParallelMs", () => pinsFor(userQuery)) : Promise.resolve(null)
         ]);
+        lap("memoryRoutingMs");
         for (const step of [...memResult.steps, ...routResult.steps]) {
           result.steps.push({
             stepNumber: 1,
@@ -3946,13 +4190,8 @@ ${projectScope.customInstructions}` : ""
           memoryPinnedItemIdsByContext,
           memoryOverride
         } = memResult;
-        const { pinsByContext: identifierPinsByContext, exactPinsByContext, steps: pinSteps } = await resolveIdentifierPins({
-          question: updatedQuestion,
-          identifierSets: cfg.vocabulary.identifiers,
-          contextsById,
-          kbKindById,
-          model: utilityModel
-        });
+        const { pinsByContext: identifierPinsByContext, exactPinsByContext, steps: pinSteps } = parallelPins && !needsPinRerun(userQuery, updatedQuestion) ? parallelPins : await pinsFor(updatedQuestion);
+        lap("pinsMs");
         for (const step of pinSteps) {
           result.steps.push({
             stepNumber: 1,
@@ -3982,6 +4221,8 @@ ${projectScope.customInstructions}` : ""
             rewrites: cfg.vocabulary.rewrites,
             styleHint: cfg.vocabulary.styleHint,
             maxQueries: cfg.tuning.maxQueriesPerContext,
+            timings,
+            timingPrefix: "search.main",
             skipPrefilter: false
           }),
           fallbackContextsToSearch.length > 0 && !hasExplicitDocAndPage ? searchContexts({
@@ -4002,9 +4243,12 @@ ${projectScope.customInstructions}` : ""
             rewrites: cfg.vocabulary.rewrites,
             styleHint: cfg.vocabulary.styleHint,
             maxQueries: cfg.tuning.maxQueriesPerContext,
+            timings,
+            timingPrefix: "search.fallback",
             skipPrefilter: true
           }) : Promise.resolve({ chunks: [] })
         ]);
+        lap("searchMs");
         const pinnedItemIds = /* @__PURE__ */ new Set([
           ...(function* () {
             for (const s of memoryPinnedItemIdsByContext.values()) yield* s;
@@ -4067,6 +4311,7 @@ ${projectScope.customInstructions}` : ""
           chunks: [],
           tokens: 0
         });
+        lap("rerankMs");
         addChunks(result, mainRerank.limited_results);
         yield { result: serializeOutput(result) };
         const literalLookupSatisfied = hasExplicitDocAndPage && mainRerank.limited_results.length > 0 && mainRerank.limited_results.some((r) => {
@@ -4115,6 +4360,7 @@ ${projectScope.customInstructions}` : ""
           });
           result.reasoning.push({ text: "Fallback results reranked", tools: [] });
           addChunks(result, fallbackRerank.limited_results);
+          lap("fallbackRerankMs");
           yield { result: serializeOutput(result) };
         }
         if (memoryOverride.active) {
@@ -4137,9 +4383,19 @@ Verified answer:
           addChunks(result, memoryOverride.chunks);
           yield { result: serializeOutput(result) };
         }
+        timings.totalMs = Date.now() - t0;
+        result.timings = timings;
+        result.steps.push({
+          stepNumber: 1,
+          text: `Timing: memory+routing ${timings.memoryRoutingMs ?? 0}ms (memory ${timings.memoryMs ?? 0}ms, routing ${timings.routingMs ?? 0}ms), pins ${timings.pinsMs ?? 0}ms, search ${timings.searchMs ?? 0}ms, rerank ${timings.rerankMs ?? 0}ms` + (timings.fallbackRerankMs !== void 0 ? `, fallback rerank ${timings.fallbackRerankMs}ms` : "") + `, total ${timings.totalMs}ms`,
+          toolCalls: [],
+          chunks: [],
+          tokens: 0
+        });
         if (cfg.logging) {
-          console.log("[EXULU pipeline] final result:", JSON.stringify({ steps: result.steps.length, chunks: result.chunks.length }));
+          console.log("[EXULU pipeline] final result:", JSON.stringify({ steps: result.steps.length, chunks: result.chunks.length, timings }));
         }
+        yield { result: serializeOutput(result) };
         return { result: serializeOutput(result) };
       } catch (err) {
         console.warn("[EXULU pipeline] retrieval pipeline failed:", err);
@@ -6052,7 +6308,7 @@ async function downloadKeyIntoSandbox(opts) {
 }
 async function resolvePythonVenvPath() {
   try {
-    const { getPythonVenvPath } = await import("./python-setup-JZGHWQCG.js");
+    const { getPythonVenvPath } = await import("./python-setup-DRJ3QX5F.js");
     return getPythonVenvPath();
   } catch (err) {
     console.warn("[SKILLS] Could not resolve the Python venv for the session sandbox; skill scripts fall back to the system python.", err);
@@ -8248,6 +8504,7 @@ export {
   sanitizeToolName,
   KB_EDITOR_TOOL_ID,
   createKbEditorPickerTool,
+  isIgnoredArtifactPath,
   reportSystemDependencies,
   downloadKeyIntoSandbox,
   truncateToolOutput,
@@ -8263,6 +8520,7 @@ export {
   guardExtractedFileText,
   SCRUBBED_CREDENTIAL_TEXT,
   SCRUBBED_OAUTH_TEXT,
+  sessionFilePrefix,
   PreviewRenderError,
   getPdfPreviewBytes,
   imageAttachmentGuard,

package/dist/cli/start-whisper.js

@@ -2,7 +2,7 @@
 import "dotenv/config";
 import {
   getPackageRoot
-} from "../chunk-27K2CO47.js";
+} from "../chunk-QMN6MVHQ.js";
 
 // src/cli/start-whisper.ts
 import "dotenv/config";

package/dist/{convert-exulu-tools-to-ai-sdk-tools-XNQ6Q3X6.js → convert-exulu-tools-to-ai-sdk-tools-6RU4IZMI.js}

@@ -2,7 +2,7 @@ import "dotenv/config";
 import {
   convertExuluToolsToAiSdkTools,
   hydrateVariables
-} from "./chunk-AWMU6QXB.js";
+} from "./chunk-RBEWHG7I.js";
 import "./chunk-4PDWNVNT.js";
 export {
   convertExuluToolsToAiSdkTools,