@lazyingart/agintiflow 0.20.232 → 0.20.233

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lazyingart/agintiflow",
-  "version": "0.20.232",
+  "version": "0.20.233",
   "type": "module",
   "description": "AgInTiFlow is a project-aware agent workspace for hybrid wet-dry R&D, hardware-aware intelligence, software automation, and industrial workflows.",
   "license": "Apache-2.0",

package/scripts/smoke-document-artifact-quality.js

@@ -42,14 +42,19 @@ const currentDocument = evaluateCurrentStateText({
 });
 assert.equal(currentDocument.ok, true, "authoritative current-state prose was rejected");
 
-function page(words, { height = 842, startY = 60, endY = 700, heading = "Section" } = {}) {
+function page(
+  words,
+  { height = 842, startY = 60, endY = 700, heading = "Section", wordHeight = 10 } = {}
+) {
   const items = [];
   const count = Math.max(1, words);
   for (let index = 0; index < count; index += 1) {
     const ratio = count === 1 ? 0 : index / (count - 1);
     const y = startY + (endY - startY) * ratio;
     const text = index === 0 ? heading : `word${index}`;
-    items.push(`<word xMin="60" yMin="${y.toFixed(2)}" xMax="100" yMax="${(y + 10).toFixed(2)}">${text}</word>`);
+    items.push(
+      `<word xMin="60" yMin="${y.toFixed(2)}" xMax="100" yMax="${(y + wordHeight).toFixed(2)}">${text}</word>`
+    );
   }
   items.push(`<word xMin="250" yMin="810" xMax="300" yMax="820">footer</word>`);
   return `<page width="595" height="${height}">${items.join("")}</page>`;
@@ -67,4 +72,18 @@ const intentionalAppendix = evaluatePdfPageBalance(
 );
 assert.equal(intentionalAppendix.ok, true, "an intentional sparse appendix page was rejected");
 
+const undersizedOnePage = evaluatePdfPageBalance(
+  `<doc>${page(340, { wordHeight: 8.3 })}</doc>`
+);
+assert.equal(undersizedOnePage.ok, false, "undersized one-page typography was accepted");
+assert(
+  undersizedOnePage.defects.some((item) => item.code === "undersized-document-text"),
+  "undersized typography did not produce a reader-facing defect"
+);
+
+const readableOnePage = evaluatePdfPageBalance(
+  `<doc>${page(260, { wordHeight: 9.4 })}</doc>`
+);
+assert.equal(readableOnePage.ok, true, "readable one-page typography was rejected");
+
 console.log("document artifact quality smoke test passed");

package/src/document-artifact-quality.js

@@ -29,6 +29,7 @@ const INTENTIONAL_SPARSE_PAGE_PATTERN =
   /^(?:appendix|approval|approvals|acknowledgements?|back cover|contact|notes|references|sign[- ]?off|signatures?)\b/i;
 const HISTORICAL_TRANSITION_PATTERN =
   /\b(?:formerly|no longer|previously|replac(?:ed|ing)|superseded|used to be)\b/i;
+const MIN_READABLE_MEDIAN_WORD_HEIGHT_PT = 8.8;
 
 function portablePath(value = "") {
   return String(value || "").replace(/\\/g, "/");
@@ -175,14 +176,27 @@ export function evaluatePdfPageBalance(bboxXml = "") {
     const yValues = contentWords.flatMap((word) => [word.yMin, word.yMax]).filter(Number.isFinite);
     const usableHeight = Math.max(1, Number(page.height || 0) - 120);
     const occupiedHeight = yValues.length ? Math.max(...yValues) - Math.min(...yValues) : 0;
+    const wordHeights = contentWords
+      .map((word) => word.yMax - word.yMin)
+      .filter((value) => Number.isFinite(value) && value > 0);
     return {
       page: index + 1,
       wordCount: contentWords.length,
       occupiedRatio: occupiedHeight / usableHeight,
+      medianWordHeight: median(wordHeights),
       leadingText: contentWords.slice(0, 16).map((word) => word.text).join(" ").trim(),
     };
   });
   const defects = [];
+  const documentMedianWordHeight = median(
+    metrics.flatMap((item) => Array(item.wordCount).fill(item.medianWordHeight))
+  );
+  if (documentMedianWordHeight > 0 && documentMedianWordHeight < MIN_READABLE_MEDIAN_WORD_HEIGHT_PT) {
+    defects.push({
+      code: "undersized-document-text",
+      message: `The document's median rendered word height is ${documentMedianWordHeight.toFixed(1)} pt, below the ${MIN_READABLE_MEDIAN_WORD_HEIGHT_PT.toFixed(1)} pt readability floor. Keep normal body type and rebalance coherent sections across pages instead of shrinking text to force a page count.`,
+    });
+  }
   if (metrics.length > 1) {
     const priorWordCounts = metrics.slice(0, -1).map((item) => item.wordCount).filter((value) => value > 0);
     const last = metrics.at(-1);
@@ -204,6 +218,7 @@ export function evaluatePdfPageBalance(bboxXml = "") {
     ok: pages.length > 0 && defects.length === 0,
     checked: pages.length > 0,
     pages: metrics,
+    documentMedianWordHeight,
     defects,
   };
 }

package/src/task-profiles.js

@@ -276,7 +276,7 @@ export const TASK_PROFILES = {
     id: "word",
     label: "Word documents",
     prompt:
-      "Bias toward Word/docx/document workflows while still using writing, conversion, LaTeX, or scripts when useful. Preserve source material byte-for-byte and synthesize a reader-facing document from the authoritative current facts instead of concatenating notes, logs, schemas, task IDs, private paths, or delivery instructions. Reconcile conflicting or superseded source facts before drafting: values labeled corrected, replaced, earlier, old, prior, cancelled, or no longer selected identify what to discard, and those discarded literals must not appear in the final document unless the user explicitly requests history or the history is necessary to explain a live decision. Keep the reconciliation history private: a current-state handoff should say who owns the work, which value is approved, and which option is selected, not narrate who was replaced or which preliminary value is no longer used. Before finishing, extract the final text and search for every superseded literal found in the sources. Prefer mature editable-document tooling already available in the workspace, such as python-docx, pandoc, or LibreOffice, over hand-written OOXML; when direct OOXML is genuinely necessary, validate its package parts and openability. Keep one maintainable source of truth and a reproducible project-local build command. Verify the DOCX is structurally editable, compile the PDF, run pdftotext or an equivalent extraction check that rejects replacement characters and unexpected control glyphs, and verify PDF text bounding boxes remain inside a readable page margin. Render every PDF page to a separate image under an ignored build/verification directory. Inspect one rendered page per read_image call, never batch pages into one vision call, and repair orphaned headings, near-empty spill pages, awkward table or paragraph breaks, overlaps, clipping, excessive whitespace, weak hierarchy, and inconsistent number formatting before finishing. Prefer one balanced readable page over pushing one short section onto a mostly empty trailing page; preserve an intentionally sparse appendix, approval, signature, reference, or back-matter page when it serves a real purpose. Retain ignored verification renders as evidence; optional cleanup must never block completion. Use clear descriptive filenames, exclude caches and generated debris from commits, inspect git status/diff before committing, and report success only after the editable source, reader-facing current-state content, visual layout, searchable text, and requested artifacts all pass.",
+      "Bias toward Word/docx/document workflows while still using writing, conversion, LaTeX, or scripts when useful. Preserve source material byte-for-byte and synthesize a reader-facing document from the authoritative current facts instead of concatenating notes, logs, schemas, task IDs, private paths, or delivery instructions. Reconcile conflicting or superseded source facts before drafting: values labeled corrected, replaced, earlier, old, prior, cancelled, or no longer selected identify what to discard, and those discarded literals must not appear in the final document unless the user explicitly requests history or the history is necessary to explain a live decision. Keep the reconciliation history private: a current-state handoff should say who owns the work, which value is approved, and which option is selected, not narrate who was replaced or which preliminary value is no longer used. Before finishing, extract the final text and search for every superseded literal found in the sources. Prefer mature editable-document tooling already available in the workspace, such as python-docx, pandoc, or LibreOffice, over hand-written OOXML; when direct OOXML is genuinely necessary, validate its package parts and openability. Keep one maintainable source of truth and a reproducible project-local build command. Verify the DOCX is structurally editable, compile the PDF, run pdftotext or an equivalent extraction check that rejects replacement characters and unexpected control glyphs, and verify PDF text bounding boxes remain inside a readable page margin. Render every PDF page to a separate image under an ignored build/verification directory. Inspect one rendered page per read_image call, never batch pages into one vision call, and repair orphaned headings, near-empty spill pages, awkward table or paragraph breaks, overlaps, clipping, excessive whitespace, weak hierarchy, and inconsistent number formatting before finishing. Keep normal readable body type, line spacing, and margins; never shrink typography merely to force the document onto one page. When content spans pages, move or redistribute coherent sections so each page is useful instead of leaving a short spill page. Preserve an intentionally sparse appendix, approval, signature, reference, or back-matter page when it serves a real purpose. Retain ignored verification renders as evidence; optional cleanup must never block completion. Use clear descriptive filenames, exclude caches and generated debris from commits, inspect git status/diff before committing, and report success only after the editable source, reader-facing current-state content, visual layout, searchable text, and requested artifacts all pass.",
     tools: ["files", "shell", "canvas", "sandbox"],
   },
   latex: {