@konneal/engine 0.2.153 → 0.2.155

package/dist/answerGuards.d.ts

@@ -0,0 +1,4 @@
+export declare function toolNarration(t: string): boolean;
+export declare const TOOL_NARRATION_NOTICE = "Correction notice: your draft narrated an internal tool call ('the \u2026 tool returned \u2026') instead of answering. Rewrite: answer the question directly from the retrieved passages, cite the clauses, and never mention tools, queries, wire formats, or JSON.";
+export declare function intentPreamble(t: string): boolean;
+export declare const INTENT_PREAMBLE_NOTICE = "Correction notice: your draft announced a plan ('I'll look up \u2026') instead of answering. Do not narrate intent \u2014 answer the question NOW from the retrieved passages: give the values, clause references and citations the question asks for.";

package/dist/{ask-UIIE5LD2.js → ask-MRVXR7CY.js}

@@ -1,6 +1,6 @@
 import {
   handleAsk
-} from "./chunk-I6HU3EM6.js";
+} from "./chunk-TXF5VMNM.js";
 import "./chunk-ECOYO7EX.js";
 import "./chunk-VYEMFSU5.js";
 import "./chunk-G2RE4XRV.js";

package/dist/{chunk-I6HU3EM6.js → chunk-TXF5VMNM.js}

@@ -1637,7 +1637,7 @@ async function extractNameplate(ai, model, image) {
   return null;
 }
 
-// workers/worker_public/src/ask.ts
+// workers/worker_public/src/answerGuards.ts
 function toolNarration(t) {
   return /^(?:the\s+)?[a-z][a-z_.]*\.[a-z_]+ tool returned\b/i.test(t.slice(0, 200)) || /tool returned,? for the query\b/i.test(t.slice(0, 300));
 }
@@ -1649,6 +1649,8 @@ function intentPreamble(t) {
   return !/\[\[u:|\[\^?\d+\]|\b\d+\.\d+\b/.test(s);
 }
 var INTENT_PREAMBLE_NOTICE = "Correction notice: your draft announced a plan ('I'll look up \u2026') instead of answering. Do not narrate intent \u2014 answer the question NOW from the retrieved passages: give the values, clause references and citations the question asks for.";
+
+// workers/worker_public/src/ask.ts
 function userImageDataUrl(body) {
   const img = body?.image;
   if (img == null) return null;

package/dist/worker_public/src/index.js

@@ -4,7 +4,7 @@ import {
   handleMemories,
   scoreFaithfulness,
   scoreJudge
-} from "../../chunk-I6HU3EM6.js";
+} from "../../chunk-TXF5VMNM.js";
 import "../../chunk-ECOYO7EX.js";
 import {
   TOOLS_REGISTRY,
@@ -411,7 +411,11 @@ ${c.text.slice(0, 1500)}` }
           context = typeof raw === "string" ? raw.trim().replace(/^["\']|[\"']$/g, "").slice(0, 400) : "";
           if (!context) {
             console.log("enrich raw keys:", Object.keys(res ?? {}), "sample:", JSON.stringify(res).slice(0, 300));
-            return { id: c.id, ok: false, error: "empty enrichment" };
+            if (contextOnly || abMode) return { id: c.id, ok: false, error: "empty enrichment" };
+            const rawText = typeof c.metadata.chunk_text === "string" && c.metadata.chunk_text ? c.metadata.chunk_text : c.text;
+            const rawVector = await embed(portModelRunner(env), MODELS.embed, rawText.slice(0, 6e3));
+            await env.VECTORIZE.upsert([{ id: c.id, values: rawVector, metadata: { ...c.metadata, chunk_text: rawText } }]);
+            return { id: c.id, ok: true, raw: true, cached: false, context: "" };
           }
           if (res?.usage) {
             usage.prompt_tokens += Number(res.usage.prompt_tokens ?? 0);
@@ -1107,7 +1111,7 @@ async function callToolOnce(env, ctx, _req, tier, key, name, args, t0) {
       const r = await runTool({ ...env, DB: env.DB }, { name, args }, "mcp");
       payload = r ? { tool: r.name, query: r.query, result: r.output } : { error: { message: `tool ${name} returned nothing for the given arguments` } };
     } else {
-      const res = name === "ask" ? await (await import("../../ask-UIIE5LD2.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-GAU5SEZI.js")).handleSearch(env, ctx, inner, tier, key);
+      const res = name === "ask" ? await (await import("../../ask-MRVXR7CY.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-GAU5SEZI.js")).handleSearch(env, ctx, inner, tier, key);
       payload = await res.json().catch(() => ({ error: { message: "tool transport failed", status: res.status } }));
     }
     telemetry(env, ctx, tier, `mcp:${name}`, null, !payload?.error, JSON.stringify(payload).length, await sha256Hex(`${name}:${JSON.stringify(args ?? {})}`), void 0, void 0, { durationMs: Date.now() - t0 });

package/package.json

@@ -27,7 +27,7 @@
     "playwright": "^1.62.1",
     "yaml": "^2.9.1"
   },
-  "version": "0.2.153",
+  "version": "0.2.155",
   "description": "The Konneal engine: the publisher-agnostic build pipeline and API plane for standards intelligence (retrieval, answer contract, verdicts, evaluation).",
   "license": "BSD-3-Clause",
   "type": "module",

package/workers/worker_public/src/admin.ts

@@ -78,7 +78,20 @@ export async function handleEnrich(env: Env, ctx: ExecutionContext, req: Request
           context = typeof raw === "string" ? raw.trim().replace(/^["\']|[\"']$/g, "").slice(0, 400) : "";
           if (!context) {
             console.log("enrich raw keys:", Object.keys(res ?? {}), "sample:", JSON.stringify(res).slice(0, 300));
-            return { id: c.id, ok: false, error: "empty enrichment" };
+            // The deterministic-empty class (observed 2026-10-20 on B 18's
+            // thin terms sections and two R 134-1 clauses — empty under
+            // BOTH the default and the alternate enrichment model): the
+            // chunk's own text is real corpus content, so the production
+            // flow serves it RAW rather than leaving a hole in the index
+            // — embed the original text without the context band, upsert
+            // without the ctx marker; a later forced re-enrich overwrites
+            // the vector in place. mode:"context" and "ab" keep failing:
+            // their builders want the signal, not a side-effecting write.
+            if (contextOnly || abMode) return { id: c.id, ok: false, error: "empty enrichment" };
+            const rawText = typeof c.metadata.chunk_text === "string" && c.metadata.chunk_text ? c.metadata.chunk_text : c.text;
+            const rawVector = await embed(portModelRunner(env), MODELS.embed, rawText.slice(0, 6000));
+            await env.VECTORIZE.upsert([{ id: c.id, values: rawVector, metadata: { ...c.metadata, chunk_text: rawText } }]);
+            return { id: c.id, ok: true, raw: true, cached: false, context: "" };
           }
           if (res?.usage) {
             usage.prompt_tokens += Number(res.usage.prompt_tokens ?? 0);

package/workers/worker_public/src/answerGuards.ts

@@ -0,0 +1,36 @@
+// The answer-contract's text guards, dependency-free so plain node can
+// load them for unit tests (ask.ts imports .md prompts that only the
+// bundler resolves — the guards were module-local there and rode e2e
+// alone). Both lanes share them: the JSON lane's contract retry and the
+// streamed lane's 80-char head ruling.
+
+// The tool-narration leak (2026-10-14, the R 134-1 probe): the answer
+// model sometimes NARRATES the agent loop's call in prose — 'The
+// db.find tool returned, for the query "…": {…json…}' — instead of
+// answering. The TOOL-line scrub cannot catch prose. The JSON lane
+// regenerates through the contract retry; the streamed lane buffers
+// the head and sends the retry sentinel (the blind-answer guard's
+// client contract) before streaming the fresh answer. One bounded
+// restart per lane.
+export function toolNarration(t: string): boolean {
+  return /^(?:the\s+)?[a-z][a-z_.]*\.[a-z_]+ tool returned\b/i.test(t.slice(0, 200)) ||
+    /tool returned,? for the query\b/i.test(t.slice(0, 300));
+}
+export const TOOL_NARRATION_NOTICE =
+  "Correction notice: your draft narrated an internal tool call ('the … tool returned …') instead of answering. Rewrite: answer the question directly from the retrieved passages, cite the clauses, and never mention tools, queries, wire formats, or JSON.";
+
+// The intent preamble (2026-10-19, the ~200-word technical-file
+// question): the draft is a PLAN, not an answer — "I'll look up the R
+// 76-1 MPE clause and classification details in the corpus." — shipped
+// while ten citations rode the same response. Narrow by construction:
+// it must OPEN with a first-person intent verb AND carry no substance —
+// short, no clause citation, no normative numbers (a real answer's
+// opening names values or clauses).
+export function intentPreamble(t: string): boolean {
+  const s = (t ?? "").trim();
+  if (s.length > 240) return false;
+  if (!/^(?:i'll|i will|let me|let's|i'm going to|i am going to)\b/i.test(s)) return false;
+  return !/\[\[u:|\[\^?\d+\]|\b\d+\.\d+\b/.test(s);
+}
+export const INTENT_PREAMBLE_NOTICE =
+  "Correction notice: your draft announced a plan ('I'll look up …') instead of answering. Do not narrate intent — answer the question NOW from the retrieved passages: give the values, clause references and citations the question asks for.";

package/workers/worker_public/src/ask.ts

@@ -49,37 +49,8 @@ import { TOOL_DECLARATION, parseToolCall, runTool, toolNote } from "./tools";
 import { operationsLane } from "./operations.ts";
 import { extractNameplate, nameplateRegisterQuery } from "./nameplate.ts";
 import { P } from "./profile.ts";
+import { toolNarration, intentPreamble, TOOL_NARRATION_NOTICE, INTENT_PREAMBLE_NOTICE } from "./answerGuards.ts";
 
-// the tool-narration leak (2026-10-14, the R 134-1 probe): the answer
-// model sometimes NARRATES the agent loop's call in prose — 'The
-// db.find tool returned, for the query "…": {…json…}' — instead of
-// answering. The TOOL-line scrub cannot catch prose. The JSON lane
-// regenerates through the contract retry; the streamed lane buffers
-// the head and sends the retry sentinel (the blind-answer guard's
-// client contract) before streaming the fresh answer. One bounded
-// restart per lane.
-function toolNarration(t: string): boolean {
-  return /^(?:the\s+)?[a-z][a-z_.]*\.[a-z_]+ tool returned\b/i.test(t.slice(0, 200)) ||
-    /tool returned,? for the query\b/i.test(t.slice(0, 300));
-}
-const TOOL_NARRATION_NOTICE =
-  "Correction notice: your draft narrated an internal tool call ('the … tool returned …') instead of answering. Rewrite: answer the question directly from the retrieved passages, cite the clauses, and never mention tools, queries, wire formats, or JSON.";
-
-// The intent preamble (2026-10-19, the ~200-word technical-file
-// question): the draft is a PLAN, not an answer — "I'll look up the R
-// 76-1 MPE clause and classification details in the corpus." — shipped
-// while ten citations rode the same response. Narrow by construction:
-// it must OPEN with a first-person intent verb AND carry no substance —
-// short, no clause citation, no normative numbers (a real answer's
-// opening names values or clauses).
-function intentPreamble(t: string): boolean {
-  const s = (t ?? "").trim();
-  if (s.length > 240) return false;
-  if (!/^(?:i'll|i will|let me|let's|i'm going to|i am going to)\b/i.test(s)) return false;
-  return !/\[\[u:|\[\^?\d+\]|\b\d+\.\d+\b/.test(s);
-}
-const INTENT_PREAMBLE_NOTICE =
-  "Correction notice: your draft announced a plan ('I'll look up …') instead of answering. Do not narrate intent — answer the question NOW from the retrieved passages: give the values, clause references and citations the question asks for.";
 
 /** User-uploaded image for multimodal questions: a data URL
  *  (data:image/(png|jpeg|webp|gif);base64,…) up to 6 MB of payload. The