@konneal/engine 0.2.150 → 0.2.152

package/dist/{ask-KVGDQTF7.js → ask-LP6M4QNS.js}

@@ -1,6 +1,6 @@
 import {
   handleAsk
-} from "./chunk-CPPIKASJ.js";
+} from "./chunk-VSK2RGUF.js";
 import "./chunk-ECOYO7EX.js";
 import "./chunk-VYEMFSU5.js";
 import "./chunk-G2RE4XRV.js";

package/dist/{chunk-CPPIKASJ.js → chunk-VSK2RGUF.js}

@@ -1642,6 +1642,13 @@ 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));
 }
 var 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.";
+function intentPreamble(t) {
+  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);
+}
+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.";
 function userImageDataUrl(body) {
   const img = body?.image;
   if (img == null) return null;
@@ -2623,12 +2630,12 @@ Answer account questions from these records ONLY: name the record when you use i
               }
               if (full.length < 80) continue;
               headSettled = true;
-              if (toolNarration(full)) {
-                console.log("tool-narration(stream): regenerating once");
+              if (toolNarration(full) || intentPreamble(full)) {
+                console.log(toolNarration(full) ? "tool-narration(stream): regenerating once" : "intent-preamble(stream): regenerating once");
                 send({ type: "retry" });
                 const regen = await generateStream(env, model, [
                   ...messages,
-                  { role: "system", content: TOOL_NARRATION_NOTICE }
+                  { role: "system", content: toolNarration(full) ? TOOL_NARRATION_NOTICE : INTENT_PREAMBLE_NOTICE }
                 ], routeEffort(routed.route));
                 if (regen) {
                   full = "";
@@ -2645,7 +2652,7 @@ Answer account questions from these records ONLY: name the record when you use i
             }
             if (!headSettled) {
               headSettled = true;
-              if (!toolNarration(full)) send({ type: "token", v: full });
+              if (!toolNarration(full) && !intentPreamble(full)) send({ type: "token", v: full });
             }
           } catch {
           }
@@ -2776,17 +2783,17 @@ Answer account questions from these records ONLY: name the record when you use i
       );
       return nums.filter((n) => tableNums.has(n)).length >= 2;
     })();
-    if (anchors.violations.length > 0 || retyped || unreferenced || toolNarration(answer)) {
-      console.log("contract check:", anchors.violations.length, "anchor violations; tableRetyped:", retyped, "; tableDataUnreferenced:", unreferenced, "; toolNarration:", toolNarration(answer), "\u2014 regenerating");
+    if (anchors.violations.length > 0 || retyped || unreferenced || toolNarration(answer) || intentPreamble(answer)) {
+      console.log("contract check:", anchors.violations.length, "anchor violations; tableRetyped:", retyped, "; tableDataUnreferenced:", unreferenced, "; toolNarration:", toolNarration(answer), "; intentPreamble:", intentPreamble(answer), "\u2014 regenerating");
       const tableUnitId = unreferenced ? used.find((h) => h.metadata.unit_id && h.metadata.block === "table")?.metadata.unit_id : void 0;
-      const note = toolNarration(answer) ? TOOL_NARRATION_NOTICE : retyped || unreferenced ? `Correction notice: your draft reproduced a table as markdown or presented a served table's data without its reference. Rewrite the answer: describe the table in prose, cite the clause, and write the reference token [[u:${tableUnitId ?? "<unit id>"}]] exactly where the table belongs. Do not render any table as markdown.` : ANCHOR_CORRECTION_NOTE;
+      const note = toolNarration(answer) ? TOOL_NARRATION_NOTICE : intentPreamble(answer) ? INTENT_PREAMBLE_NOTICE : retyped || unreferenced ? `Correction notice: your draft reproduced a table as markdown or presented a served table's data without its reference. Rewrite the answer: describe the table in prose, cite the clause, and write the reference token [[u:${tableUnitId ?? "<unit id>"}]] exactly where the table belongs. Do not render any table as markdown.` : ANCHOR_CORRECTION_NOTE;
       generateRetries += 1;
       const corrected = await generateOnce(env, model, [...messages, { role: "system", content: note }], routeEffort(routed.route));
       if (corrected) {
         const correctedAnswer = canonicalRefusal(corrected, refusalScope(answerCites(corrected)));
         const retryAnchors = checkQuoteAnchors(correctedAnswer, used.map((h) => h.text));
         const retryRetyped = tableRetyped(correctedAnswer, hasTableUnit);
-        if (retryAnchors.violations.length < anchors.violations.length || !retryRetyped && retyped || unreferenced && correctedAnswer.includes("[[u:") || !toolNarration(correctedAnswer)) {
+        if (retryAnchors.violations.length < anchors.violations.length || !retryRetyped && retyped || unreferenced && correctedAnswer.includes("[[u:") || !toolNarration(correctedAnswer) || !intentPreamble(correctedAnswer)) {
           answer = correctedAnswer;
         }
       }

package/dist/mcp-proto.d.ts

@@ -15,6 +15,9 @@ export interface McpTool {
     };
 }
 export declare function buildTools(): McpTool[];
+export declare const TOOL_RESULT_CHAR_CAP = 24000;
+export declare const TRUNCATION_MARK = "--- TRUNCATED ---";
+export declare function capToolResult(payload: unknown, cap?: number): unknown;
 export type McpResult = {
     ok: true;
     result: unknown;

package/dist/worker_public/src/index.js

@@ -4,7 +4,7 @@ import {
   handleMemories,
   scoreFaithfulness,
   scoreJudge
-} from "../../chunk-CPPIKASJ.js";
+} from "../../chunk-VSK2RGUF.js";
 import "../../chunk-ECOYO7EX.js";
 import {
   TOOLS_REGISTRY,
@@ -1021,6 +1021,34 @@ function buildTools() {
     }))
   ];
 }
+var TOOL_RESULT_CHAR_CAP = 24e3;
+var TRUNCATION_MARK = "--- TRUNCATED ---";
+var jsonLen = (x) => JSON.stringify(x)?.length ?? 0;
+function clipStrings(value, budget) {
+  if (typeof value === "string") return value.length <= budget ? value : `${value.slice(0, budget)}
+${TRUNCATION_MARK}`;
+  if (Array.isArray(value)) return value.map((v) => clipStrings(v, budget));
+  if (value && typeof value === "object") return Object.fromEntries(Object.entries(value).map(([k, v]) => [k, clipStrings(v, budget)]));
+  return value;
+}
+function capToolResult(payload, cap = TOOL_RESULT_CHAR_CAP) {
+  if (jsonLen(payload) <= cap) return payload;
+  let lo = 200;
+  let hi = cap;
+  let best = clipStrings(payload, lo);
+  while (lo <= hi) {
+    const mid = lo + hi >> 1;
+    const candidate = clipStrings(payload, mid);
+    if (jsonLen(candidate) <= cap) {
+      best = candidate;
+      lo = mid + 1;
+    } else {
+      hi = mid - 1;
+    }
+  }
+  if (jsonLen(best) <= cap) return best;
+  return { truncation: `the tool result exceeded the ${cap}-character budget and was withheld` };
+}
 function dispatch(method, params, callTool) {
   switch (method) {
     case "initialize":
@@ -1036,7 +1064,7 @@ function dispatch(method, params, callTool) {
         return Promise.resolve({ ok: false, code: -32602, message: `unknown tool: ${name}` });
       }
       return callTool(registryName(name), params?.arguments ?? {}).then(
-        (payload) => ({ ok: true, result: { content: [{ type: "text", text: JSON.stringify(payload) }] } })
+        (payload) => ({ ok: true, result: { content: [{ type: "text", text: JSON.stringify(capToolResult(payload)) }] } })
       );
     }
     default:
@@ -1079,7 +1107,7 @@ 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 ${name} returned nothing for the given arguments` } };
     } else {
-      const res = name === "ask" ? await (await import("../../ask-KVGDQTF7.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-GAU5SEZI.js")).handleSearch(env, ctx, inner, tier, key);
+      const res = name === "ask" ? await (await import("../../ask-LP6M4QNS.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-GAU5SEZI.js")).handleSearch(env, ctx, inner, tier, key);
       payload = await res.json().catch(() => ({ error: { message: "tool transport failed", status: res.status } }));
     }
     telemetry(env, ctx, tier, `mcp:${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": "0.2.150",
+  "version": "0.2.152",
   "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/ask.ts

@@ -65,6 +65,22 @@ function toolNarration(t: string): boolean {
 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
  *  question text still drives retrieval; the image is CONTEXT for the
@@ -1570,12 +1586,12 @@ let convEntities: Array<{ entity: string; kind: string }> = [];
               if (headSettled) { send({ type: "token", v: tok }); continue; }
               if (full.length < 80) continue;
               headSettled = true;
-              if (toolNarration(full)) {
-                console.log("tool-narration(stream): regenerating once");
+              if (toolNarration(full) || intentPreamble(full)) {
+                console.log(toolNarration(full) ? "tool-narration(stream): regenerating once" : "intent-preamble(stream): regenerating once");
                 send({ type: "retry" });
                 const regen = await generateStream(env, model, [
                   ...messages,
-                  { role: "system", content: TOOL_NARRATION_NOTICE },
+                  { role: "system", content: toolNarration(full) ? TOOL_NARRATION_NOTICE : INTENT_PREAMBLE_NOTICE },
                 ], routeEffort(routed.route));
                 if (regen) {
                   full = "";
@@ -1594,7 +1610,7 @@ let convEntities: Array<{ entity: string; kind: string }> = [];
             }
             if (!headSettled) {
               headSettled = true;
-              if (!toolNarration(full)) send({ type: "token", v: full });
+              if (!toolNarration(full) && !intentPreamble(full)) send({ type: "token", v: full });
             }
           } catch {
             // stream ended prematurely — deliver what we have
@@ -1776,8 +1792,8 @@ let convEntities: Array<{ entity: string; kind: string }> = [];
       );
       return nums.filter((n) => tableNums.has(n)).length >= 2;
     })();
-    if (anchors.violations.length > 0 || retyped || unreferenced || toolNarration(answer)) {
-      console.log("contract check:", anchors.violations.length, "anchor violations; tableRetyped:", retyped, "; tableDataUnreferenced:", unreferenced, "; toolNarration:", toolNarration(answer), "— regenerating");
+    if (anchors.violations.length > 0 || retyped || unreferenced || toolNarration(answer) || intentPreamble(answer)) {
+      console.log("contract check:", anchors.violations.length, "anchor violations; tableRetyped:", retyped, "; tableDataUnreferenced:", unreferenced, "; toolNarration:", toolNarration(answer), "; intentPreamble:", intentPreamble(answer), "— regenerating");
       // name the EXACT unit the token must reference — a generic note
       // leaves the model guessing which id to write
       const tableUnitId = unreferenced
@@ -1785,6 +1801,8 @@ let convEntities: Array<{ entity: string; kind: string }> = [];
         : undefined;
       const note = toolNarration(answer)
         ? TOOL_NARRATION_NOTICE
+        : intentPreamble(answer)
+        ? INTENT_PREAMBLE_NOTICE
         : retyped || unreferenced
         ? `Correction notice: your draft reproduced a table as markdown or presented a served table's data without its reference. Rewrite the answer: describe the table in prose, cite the clause, and write the reference token [[u:${tableUnitId ?? "<unit id>"}]] exactly where the table belongs. Do not render any table as markdown.`
         : ANCHOR_CORRECTION_NOTE;
@@ -1794,7 +1812,7 @@ let convEntities: Array<{ entity: string; kind: string }> = [];
         const correctedAnswer = canonicalRefusal(corrected, refusalScope(answerCites(corrected)));
         const retryAnchors = checkQuoteAnchors(correctedAnswer, used.map((h: Hit) => h.text));
         const retryRetyped = tableRetyped(correctedAnswer, hasTableUnit);
-        if (retryAnchors.violations.length < anchors.violations.length || (!retryRetyped && retyped) || (unreferenced && correctedAnswer.includes("[[u:")) || !toolNarration(correctedAnswer)) {
+        if (retryAnchors.violations.length < anchors.violations.length || (!retryRetyped && retyped) || (unreferenced && correctedAnswer.includes("[[u:")) || !toolNarration(correctedAnswer) || !intentPreamble(correctedAnswer)) {
           answer = correctedAnswer;
         }
       }

package/workers/worker_public/src/mcp-proto.ts

@@ -94,6 +94,44 @@ export function buildTools(): McpTool[] {
   ];
 }
 
+// The tool-result budget (the cloudflare/mcp pattern, 2026-10-19): an
+// MCP client spends its own context reading tool results, so an
+// unbounded payload (a broad db.find over the register, a large
+// retrieve) can eat an agent's window. An oversized result comes back
+// size-bounded and still valid JSON: short fields (names, numbers,
+// metadata) stay whole, long strings clip, every cut is marked.
+export const TOOL_RESULT_CHAR_CAP = 24000; // ≈6k tokens
+export const TRUNCATION_MARK = "--- TRUNCATED ---";
+
+const jsonLen = (x: unknown): number => JSON.stringify(x)?.length ?? 0;
+
+function clipStrings(value: unknown, budget: number): unknown {
+  if (typeof value === "string") return value.length <= budget ? value : `${value.slice(0, budget)}\n${TRUNCATION_MARK}`;
+  if (Array.isArray(value)) return value.map((v) => clipStrings(v, budget));
+  if (value && typeof value === "object") return Object.fromEntries(Object.entries(value).map(([k, v]) => [k, clipStrings(v, budget)]));
+  return value;
+}
+
+export function capToolResult(payload: unknown, cap: number = TOOL_RESULT_CHAR_CAP): unknown {
+  if (jsonLen(payload) <= cap) return payload;
+  // the per-string budget shrinks until the whole payload fits
+  let lo = 200;
+  let hi = cap;
+  let best = clipStrings(payload, lo);
+  while (lo <= hi) {
+    const mid = (lo + hi) >> 1;
+    const candidate = clipStrings(payload, mid);
+    if (jsonLen(candidate) <= cap) {
+      best = candidate;
+      lo = mid + 1;
+    } else {
+      hi = mid - 1;
+    }
+  }
+  if (jsonLen(best) <= cap) return best;
+  return { truncation: `the tool result exceeded the ${cap}-character budget and was withheld` };
+}
+
 export type McpResult =
   | { ok: true; result: unknown }
   | { ok: true; accepted: true }
@@ -120,7 +158,7 @@ export function dispatch(
         return Promise.resolve({ ok: false, code: -32602, message: `unknown tool: ${name}` });
       }
       return callTool(registryName(name), params?.arguments ?? {}).then(
-        (payload) => ({ ok: true, result: { content: [{ type: "text", text: JSON.stringify(payload) }] } }) as McpResult,
+        (payload) => ({ ok: true, result: { content: [{ type: "text", text: JSON.stringify(capToolResult(payload)) }] } }) as McpResult,
       );
     }
     default: