Files
StepForge/core/text-intel.js
T
TylerandClaude Sonnet 4.6 8d645c0aca fix smart title generation from browser window context
The title engine was falling back to "Screen capture" for all browser
captures because it treated the entire browser window title as noise.

Changes:
- stripBrowserNameSuffix: removes "- Google Chrome" / "| Firefox" etc.
  from the end of window titles, leaving just the page title.
  "oracle - Google Search - Google Chrome" → "oracle - Google Search"
  "Oracle | Cloud Applications - Google Chrome" → page title only

- extractSearchQuery: detects "[query] - Google Search" / Bing / etc.
  patterns after stripping the browser suffix, and formats the result
  as "Search for oracle".

- buildCaptureTitle: uses stripped page title + search detection before
  falling through to the "Screen capture" fallback.

- pickBestOcrPhrase: considers the full OCR line (≤80 chars) as a
  candidate with a +35 completeness bonus before splitting on | or ·.
  This preserves "Oracle | Cloud Applications and Cloud Platform" as a
  single phrase instead of breaking it into fragments.

- candidateWords: filters out standalone punctuation tokens (|, ·, •)
  so they don't inflate word-count penalties for compound brand names.

- verbForElementRole: hyperlinks and links now produce "Select" instead
  of "Click"; search box / search field produces "Search for".

Five new unit tests cover: browser title stripping, search query
extraction, full pipe-separated link text, link and search box verbs.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
2026-06-24 09:44:36 -05:00

648 lines
22 KiB
JavaScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
'use strict';
const { deepClone, htmlToText, escapeHtml } = require('./util');
const { sanitizeHtml } = require('./sanitize');
const {
TEXTBLOCK_LEVELS,
TEXTBLOCK_POSITIONS,
normalizeTextBlock,
normalizeCodeBlock,
normalizeTableBlock,
} = require('./schema');
const DEFAULT_CAPTURE_TITLES = {
fullscreen: 'Screen capture',
window: 'Window capture',
region: 'Region capture',
};
const AI_LEVEL_ALIASES = new Map([
['note', 'info'],
['info', 'info'],
['tip', 'success'],
['success', 'success'],
['warning', 'warn'],
['warn', 'warn'],
['important', 'error'],
['error', 'error'],
]);
const GENERIC_OCR_PHRASES = new Set([
'button',
'click',
'double click',
'menu',
'item',
'field',
'text field',
'search',
'submit',
'cancel',
'ok',
'open',
'select',
'enter',
'type',
]);
const BROWSER_NAME_PHRASES = new Set([
'google chrome',
'chrome',
'chromium',
'microsoft edge',
'edge',
'brave',
'firefox',
'safari',
'opera',
'vivaldi',
]);
// Known search engine page title suffixes (what appears after the query in the window title).
const SEARCH_ENGINE_PAGE_NAMES = new Set([
'google search',
'google',
'bing',
'duckduckgo',
'yahoo search',
'yahoo',
'startpage',
'ecosia',
'brave search',
]);
const ACTION_PREFIXES = [
'click',
'select',
'open',
'choose',
'enter',
'type',
'search',
'switch to',
'go to',
'navigate to',
'toggle',
'turn on',
'turn off',
'enable',
'disable',
'pick',
'focus',
'launch',
'activate',
];
function normalizeWhitespace(text) {
return String(text == null ? '' : text)
.replace(/\s+/g, ' ')
.trim();
}
function titleCaseWord(word) {
if (!word) return word;
if (/^[A-Z0-9]{2,}$/.test(word)) return word;
if (/^\d+$/.test(word)) return word;
return word[0].toUpperCase() + word.slice(1).toLowerCase();
}
function displayText(text) {
const clean = normalizeWhitespace(text)
.replace(/^[\s"'`([{<]+|[\s"'`)}\]>.,;:!?]+$/g, '')
.trim();
if (!clean) return '';
if (clean === clean.toUpperCase()) {
return clean.split(/\s+/).map(titleCaseWord).join(' ');
}
return clean.replace(/\s+/g, ' ');
}
function sentenceCase(text) {
const clean = displayText(text);
if (!clean) return '';
return clean.charAt(0).toUpperCase() + clean.slice(1);
}
function isPathOrUrlLike(text) {
return /^(?:https?:\/\/|file:\/\/|about:blank|chrome:\/\/|edge:\/\/|moz-extension:\/\/|view-source:|localhost(?:[:/]|$)|www\.)/i.test(text) ||
/[A-Za-z]:\\/.test(text) ||
/\/(?:[^/\s]+\/){2,}/.test(text) ||
/\\/.test(text);
}
function isBrowserNoise(text) {
const clean = normalizeWhitespace(text).toLowerCase();
if (!clean) return true;
if (BROWSER_NAME_PHRASES.has(clean)) return true;
if (isPathOrUrlLike(clean)) return true;
let foundBrowserName = false;
for (const name of BROWSER_NAME_PHRASES) {
if (clean.includes(name)) {
foundBrowserName = true;
break;
}
}
return foundBrowserName && /[\s|•·*]{2,}|[-–—]|\/|\\/.test(clean);
}
function isUsefulTitleCandidate(text, { source = 'ocr' } = {}) {
const clean = displayText(text);
if (!clean) return false;
const lower = clean.toLowerCase();
if (GENERIC_OCR_PHRASES.has(lower)) return false;
if (BROWSER_NAME_PHRASES.has(lower)) return false;
if (isPathOrUrlLike(clean)) return false;
if ((source === 'window' || source === 'app') && isBrowserNoise(clean)) return false;
if (/^[\p{P}\p{S}0-9]+$/u.test(clean)) return false;
return true;
}
function splitTitleFragments(text) {
const clean = normalizeWhitespace(text);
if (!clean) return [];
return clean
.split(/\s*(?:\*\*+|[|•·]+|::|\/+|\\+|\s[-–—]\s|\s{2,})\s*/g)
.map((part) => displayText(part))
.filter(Boolean);
}
function candidateWords(text) {
const clean = normalizeWhitespace(text);
if (!clean) return [];
// Exclude standalone punctuation tokens (e.g. "|" in "Oracle | Cloud...") from word count.
return clean.split(/\s+/).filter((w) => /[a-zA-Z0-9]/.test(w));
}
// Remove trailing "- Google Chrome", "| Firefox", etc. from a browser window title,
// leaving just the page title portion: "oracle - Google Search - Google Chrome" → "oracle - Google Search".
function stripBrowserNameSuffix(text) {
let clean = normalizeWhitespace(text);
for (const name of BROWSER_NAME_PHRASES) {
const escaped = name.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
clean = clean.replace(new RegExp(`\\s*[-–—·|•]\\s*${escaped}\\s*$`, 'i'), '').trim();
}
return clean;
}
// Detect "[query] - Google Search" or "[query] - Bing" patterns in a (already-stripped) page title.
// Returns the query word(s) if found, otherwise ''.
function extractSearchQuery(pageTitle) {
const frags = splitTitleFragments(pageTitle);
if (frags.length < 2) return '';
const last = frags[frags.length - 1].toLowerCase();
if (SEARCH_ENGINE_PAGE_NAMES.has(last)) {
const query = frags[0];
if (query && isUsefulTitleCandidate(query, { source: 'ocr' })) return query;
}
return '';
}
function scoreCandidate(text, { source = 'ocr' } = {}) {
const clean = displayText(text);
if (!clean) return -Infinity;
const words = candidateWords(clean);
if (!words.length) return -Infinity;
let score = 0;
score += source === 'ocr' ? 140 : source === 'element' ? 95 : source === 'window' ? 35 : source === 'app' ? 25 : 90;
score += Math.min(words.length, 5) * 10;
score -= Math.max(0, words.length - 5) * 11;
score -= Math.max(0, clean.length - 42) * 0.8;
if (GENERIC_OCR_PHRASES.has(clean.toLowerCase())) score -= 50;
if (BROWSER_NAME_PHRASES.has(clean.toLowerCase())) score -= 80;
if (isBrowserNoise(clean)) score -= 60;
if (clean.length <= 24) score += 10;
if (/^(click|select|open|choose|enter|type|search|switch to|go to|navigate to|toggle|turn on|turn off|enable|disable|pick|focus|launch|activate)\b/i.test(clean)) score += 12;
if (/^[\p{P}\p{S}0-9]+$/u.test(clean)) score -= 100;
return score;
}
function pickBestOcrPhrase(ocrText) {
const text = normalizeWhitespace(ocrText);
if (!text) return '';
let best = '';
let bestScore = -Infinity;
for (const rawLine of text.split(/\n+/)) {
const line = normalizeWhitespace(rawLine);
if (!line) continue;
// For short lines (link text, button labels) try the FULL line first before splitting.
// This preserves "Oracle | Cloud Applications and Cloud Platform" instead of splitting on |.
// Full-line bonus (+35) nudges it ahead of its own fragments.
const candidates = line.length <= 80
? [[line, 35], ...splitTitleFragments(line).map((f) => [f, 0])]
: splitTitleFragments(line).map((f) => [f, 0]);
for (const [part, bonus] of candidates) {
if (!isUsefulTitleCandidate(part, { source: 'ocr' })) continue;
const score = scoreCandidate(part, { source: 'ocr' }) + bonus;
if (score > bestScore) {
best = part;
bestScore = score;
}
}
}
return best;
}
function isShortUiLabel(text) {
const words = candidateWords(text);
return words.length > 0 && words.length <= 2 && text.length <= 24;
}
function isDirectiveTitle(text) {
const clean = displayText(text);
if (!clean) return false;
const lower = clean.toLowerCase();
return ACTION_PREFIXES.some((prefix) => lower.startsWith(prefix));
}
function verbForElementRole(role) {
const clean = normalizeWhitespace(role).toLowerCase();
if (!clean) return null;
if (/(tab|menu item|menuitem|option|list item|tree item|radio button|dropdown list|combo box option|hyperlink|link)/.test(clean)) {
return 'Select';
}
if (/(search box|searchbox|search field|search bar|search input)/.test(clean)) {
return 'Search for';
}
if (/(button|check box|checkbox|toggle button|switch|item|command)/.test(clean)) {
return 'Click';
}
if (/(text field|edit|combo box|textbox|text box|input|field)/.test(clean)) {
return 'Click';
}
return null;
}
function formatCaptureTitle(text, { source = 'ocr', metadata = {} } = {}) {
const clean = displayText(text);
if (!clean) return '';
if (isDirectiveTitle(clean)) {
return sentenceCase(clean);
}
const roleVerb = (source === 'ocr' || source === 'element') ? verbForElementRole(metadata.elementRole) : null;
if (roleVerb) {
return `${roleVerb} ${sentenceCase(clean)}`;
}
if (source === 'window' || source === 'app') {
return `Open ${sentenceCase(clean)}`;
}
if (source === 'ocr' || source === 'element') {
return isShortUiLabel(clean) ? `Click ${sentenceCase(clean)}` : sentenceCase(clean);
}
return sentenceCase(clean);
}
function pickBestTitleFragment(text, { source = 'window', metadata = {} } = {}) {
const fragments = splitTitleFragments(text).filter((line) => isUsefulTitleCandidate(line, { source }));
if (!fragments.length) return '';
let best = '';
let bestScore = -Infinity;
for (const part of fragments) {
const score = scoreCandidate(part, { source });
if (score > bestScore) {
best = part;
bestScore = score;
}
}
return best ? formatCaptureTitle(best, { source, metadata }) : '';
}
function buildCaptureTitle({ mode = 'fullscreen', metadata = {}, ocrText = '' } = {}) {
const ocrPhrase = pickBestOcrPhrase(ocrText);
if (ocrPhrase) return formatCaptureTitle(ocrPhrase, { source: 'ocr', metadata });
const elementPhrase = pickBestTitleFragment(metadata.elementLabel, { source: 'element', metadata });
if (elementPhrase) return elementPhrase;
// Strip the browser name suffix before processing window titles so that
// "oracle - Google Search - Google Chrome" becomes "oracle - Google Search"
// and "Oracle | Cloud Applications - Google Chrome" becomes just the page title.
const rawWindowTitle = metadata.windowTitle || '';
const strippedWindowTitle = stripBrowserNameSuffix(rawWindowTitle);
if (strippedWindowTitle) {
// Detect "[query] - Google Search" → "Search for oracle"
const searchQuery = extractSearchQuery(strippedWindowTitle);
if (searchQuery) return `Search for ${sentenceCase(searchQuery)}`;
// Use the page title (now free of browser name noise).
const windowPhrase = pickBestTitleFragment(strippedWindowTitle, { source: 'window', metadata });
if (windowPhrase) return windowPhrase;
}
const appPhrase = pickBestTitleFragment(metadata.appName, { source: 'app', metadata });
if (appPhrase) return appPhrase;
return DEFAULT_CAPTURE_TITLES[mode] || 'Capture';
}
function plainTextToHtml(text) {
const trimmed = normalizeWhitespace(text);
if (!trimmed) return '';
return trimmed
.split(/\n{2,}/)
.map((para) => `<p>${escapeHtml(para).replace(/\n/g, '<br>')}</p>`)
.join('');
}
function normalizeOllamaHost(host) {
const raw = normalizeWhitespace(host);
if (!raw) return '';
if (/^https?:\/\//i.test(raw)) return raw.replace(/\/+$/, '');
return `http://${raw.replace(/\/+$/, '')}`;
}
function normalizeAiLevel(level) {
const key = normalizeWhitespace(level).toLowerCase();
return AI_LEVEL_ALIASES.get(key) || (TEXTBLOCK_LEVELS.includes(key) ? key : 'info');
}
function normalizeAiPosition(position) {
const key = normalizeWhitespace(position).toLowerCase();
return TEXTBLOCK_POSITIONS.includes(key) ? key : 'after-description';
}
function normalizeAiBlock(block) {
if (!block || typeof block !== 'object') return null;
const kind = normalizeWhitespace(block.kind).toLowerCase();
if (kind === 'text') {
const normalized = normalizeTextBlock({
id: block.id,
order: Number.isFinite(block.order) ? block.order : null,
position: normalizeAiPosition(block.position),
level: normalizeAiLevel(block.level),
title: displayText(block.title),
descriptionHtml: plainTextToHtml(block.body ?? block.description ?? block.text ?? ''),
}, Number.isFinite(block.order) ? block.order : null);
return { ...normalized, kind: 'text' };
}
if (kind === 'code') {
return {
...normalizeCodeBlock({
id: block.id,
order: Number.isFinite(block.order) ? block.order : null,
language: displayText(block.language).toLowerCase(),
code: String(block.code ?? ''),
}, Number.isFinite(block.order) ? block.order : null),
kind: 'code',
};
}
if (kind === 'table') {
const rows = Array.isArray(block.rows)
? block.rows.map((row) => (Array.isArray(row) ? row.map((cell) => displayText(cell)) : []))
: [];
return {
...normalizeTableBlock({
id: block.id,
order: Number.isFinite(block.order) ? block.order : null,
rows,
}, Number.isFinite(block.order) ? block.order : null),
kind: 'table',
};
}
return null;
}
function normalizeAiPatch(raw) {
let data = raw;
if (typeof raw === 'string') {
const trimmed = raw.trim().replace(/^```(?:json)?\s*|\s*```$/g, '');
const start = trimmed.indexOf('{');
const end = trimmed.lastIndexOf('}');
const jsonText = start >= 0 && end > start ? trimmed.slice(start, end + 1) : trimmed;
data = JSON.parse(jsonText);
}
if (!data || typeof data !== 'object' || Array.isArray(data)) {
throw new Error('AI response must be a JSON object');
}
const out = {
title: displayText(data.title),
descriptionHtml: plainTextToHtml(data.description ?? data.descriptionText ?? ''),
blocks: Array.isArray(data.blocks)
? data.blocks.map((block) => normalizeAiBlock(block)).filter(Boolean)
: [],
};
return out;
}
function summarizeBlocks(step = {}) {
const parts = [];
for (const block of step.textBlocks || []) {
const body = htmlToText(block.descriptionHtml || '');
parts.push(`- Text (${block.level || 'info'}, ${block.position || 'after-description'}): ${block.title || ''}${body ? ` — ${body}` : ''}`.trim());
}
for (const block of step.codeBlocks || []) {
const code = String(block.code || '').trim();
parts.push(`- Code (${block.language || 'plain'}):\n${code || '(empty)'}`);
}
for (const block of step.tableBlocks || []) {
const rows = Array.isArray(block.rows) ? block.rows.length : 0;
const cols = rows > 0 && Array.isArray(block.rows[0]) ? block.rows[0].length : 0;
parts.push(`- Table (${rows}x${cols})`);
}
return parts.length ? parts.join('\n') : '(none)';
}
const DEFAULT_PLACEHOLDER_TITLES = new Set(
Object.values(DEFAULT_CAPTURE_TITLES).concat(['Capture', 'Untitled step']),
);
function isPlaceholderTitle(title) {
return !title || DEFAULT_PLACEHOLDER_TITLES.has(title);
}
function summarizeStepForAi(step = {}) {
const titleLine = isPlaceholderTitle(step.title)
? 'Step title: (not set — generate a specific action title from the capture context)'
: `Step title: ${step.title}`;
const descText = htmlToText(step.descriptionHtml || '');
return [
titleLine,
`Step description: ${descText || '(empty)'}`,
`Step status: ${step.status || 'todo'}`,
`Blocks:\n${summarizeBlocks(step)}`,
].join('\n');
}
function summarizeGuideForAi(guide = {}) {
return [
`Guide title: ${guide.title || '(untitled)'}`,
`Guide description: ${htmlToText(guide.descriptionHtml || '') || '(empty)'}`,
].join('\n');
}
function hasRichCaptureContext(captureContext) {
if (!captureContext) return false;
const ocr = normalizeWhitespace(captureContext.ocrText || '');
const win = normalizeWhitespace(captureContext.windowTitle || '');
const app = normalizeWhitespace(captureContext.appName || '');
const element = normalizeWhitespace(captureContext.elementLabel || '');
// Any non-trivial context signal is enough — even just an app name.
return ocr.length > 3 || win.length > 2 || app.length > 1 || element.length > 1;
}
function buildAiPrompt({
target = 'all',
guide = null,
step = null,
captureContext = null,
block = null,
} = {}) {
const hasDraftTitle = step && !isPlaceholderTitle(step.title);
const hasDraftDesc = step && Boolean(htmlToText(step.descriptionHtml || ''));
const targetText = {
title: hasDraftTitle
? 'improve the user\'s draft step title — keep their intent, make it read like professional documentation'
: 'write a specific action title for this step using the capture context',
description: hasDraftDesc
? 'improve the user\'s draft description — keep their intent, make it read like professional documentation'
: 'write a 12 sentence description of what the user does in this step, using the capture context',
block: 'rewrite only the target block',
all: 'write the step title, description, and any useful blocks from the capture context',
}[target] || 'rewrite the step';
const richContext = hasRichCaptureContext(captureContext);
const allowedBlockNote = [
'Use block.kind = "text" with level in [info, warn, error, success] for note / warning / important / tip blocks.',
'Use block.kind = "code" for code snippets.',
'Use block.kind = "table" for tables, with rows as arrays of strings.',
'Use block.position values from [before-title, after-title, before-image, after-image, before-description, after-description].',
].join(' ');
const contextLines = captureContext ? [
captureContext.windowTitle ? `Active window: ${captureContext.windowTitle}` : null,
captureContext.appName ? `App: ${captureContext.appName}` : null,
captureContext.elementLabel ? `UI element: ${captureContext.elementLabel}${captureContext.elementRole ? ` (${captureContext.elementRole})` : ''}` : null,
captureContext.ocrText ? `OCR text near click:\n${captureContext.ocrText}` : null,
captureContext.titleCandidate ? `Suggested title: ${captureContext.titleCandidate}` : null,
].filter(Boolean) : [];
const prompt = [
'You write concise, action-focused step-by-step documentation for a desktop application guide.',
'Return JSON only. No markdown fences, no commentary, no extra keys outside the schema below.',
'Schema:',
'{',
' "title": string,',
' "description": string,',
' "blocks": [{',
' "kind": "text" | "code" | "table",',
' "position"?: "before-title" | "after-title" | "before-image" | "after-image" | "before-description" | "after-description",',
' "level"?: "info" | "warn" | "error" | "success",',
' "title"?: string,',
' "body"?: string,',
' "language"?: string,',
' "code"?: string,',
' "rows"?: string[][]',
' }]',
'}',
'',
`Target: ${targetText}.`,
allowedBlockNote,
'',
guide ? summarizeGuideForAi(guide) : 'Guide: (not provided)',
'',
step ? summarizeStepForAi(step) : 'Step: (not provided)',
'',
contextLines.length
? `Capture context:\n${contextLines.join('\n')}`
: 'Capture context: (not available)',
'',
block ? `Target block:\n${JSON.stringify(block, null, 2)}` : null,
'',
'Rules:',
'- Titles must be short imperative actions: "Click Save", "Select New document", "Open Settings".',
'- NEVER output "Screen capture", "Window capture", "Region capture", or "Capture" as a title — always produce something specific.',
hasDraftTitle
? '- The user has a draft title. Improve its wording without changing their intent.'
: '- No title yet. Use the capture context (OCR text, window, app) to write a specific action title.',
hasDraftDesc
? '- The user has a draft description. Polish it to read like professional documentation.'
: '- No description yet. Write 12 sentences describing exactly what the user does.',
'- Only include blocks that provide genuinely useful supplemental information (warnings, tips, code).',
richContext
? '- Use the OCR text, window title, app name, and element info to make the documentation specific.'
: '- Context is limited. Use the app name or window title if available; generate a reasonable action title.',
'- Do NOT generate blocks that describe the technical capture process or mention OCR.',
'- Do NOT invent details not supported by the capture context.',
'- If the target is one block, only rewrite that block.',
].filter((l) => l !== null).join('\n');
return {
systemPrompt: 'You are a technical documentation writer. Emit only valid JSON matching the schema. Never add commentary or markdown.',
prompt,
};
}
function applyAiPatchToStep(step, patch, { target = 'all', blockId = null } = {}) {
const next = deepClone(step);
if ((target === 'all' || target === 'title') && patch.title) {
next.title = displayText(patch.title);
}
if ((target === 'all' || target === 'description') && patch.descriptionHtml) {
next.descriptionHtml = sanitizeHtml(patch.descriptionHtml);
}
if (target === 'all' && Array.isArray(patch.blocks) && patch.blocks.length) {
const textBlocks = [];
const codeBlocks = [];
const tableBlocks = [];
let nextOrder = 1;
for (const block of patch.blocks) {
const clone = deepClone(block);
clone.order = nextOrder++;
if (clone.kind === 'text') textBlocks.push(clone);
else if (clone.kind === 'code') codeBlocks.push(clone);
else if (clone.kind === 'table') tableBlocks.push(clone);
}
next.textBlocks = textBlocks;
next.codeBlocks = codeBlocks;
next.tableBlocks = tableBlocks;
} else if (target === 'block' && blockId && Array.isArray(patch.blocks) && patch.blocks.length) {
const replacement = patch.blocks[0];
const textBlock = (next.textBlocks || []).find((block) => block.id === blockId);
const codeBlock = (next.codeBlocks || []).find((block) => block.id === blockId);
const tableBlock = (next.tableBlocks || []).find((block) => block.id === blockId);
if (textBlock && replacement.kind === 'text') {
if (replacement.position) textBlock.position = replacement.position;
if (replacement.level) textBlock.level = replacement.level;
if (replacement.title) textBlock.title = replacement.title;
if (replacement.descriptionHtml) textBlock.descriptionHtml = sanitizeHtml(replacement.descriptionHtml);
} else if (codeBlock && replacement.kind === 'code') {
if (replacement.language) codeBlock.language = replacement.language;
if (replacement.code) codeBlock.code = replacement.code;
} else if (tableBlock && replacement.kind === 'table') {
if (replacement.rows) tableBlock.rows = replacement.rows;
}
}
if (!next.image) {
const hasBody = Boolean(
next.title ||
htmlToText(next.descriptionHtml || '') ||
(next.textBlocks || []).length ||
(next.codeBlocks || []).length ||
(next.tableBlocks || []).length,
);
if (hasBody) next.kind = 'content';
}
return next;
}
module.exports = {
DEFAULT_CAPTURE_TITLES,
buildCaptureTitle,
plainTextToHtml,
normalizeOllamaHost,
normalizeAiPatch,
buildAiPrompt,
applyAiPatchToStep,
summarizeStepForAi,
summarizeGuideForAi,
displayText,
normalizeWhitespace,
scoreCandidate,
pickBestOcrPhrase,
};