Open-source Grok Bot alternative for named AI teammates. Local-first desktop, web, and mobile clients with per-bot tools, instructions, and memory. Independent fork of T3 Code.
import { GENERAL_TOPIC, GOAL_TOPICS, type GoalSignal, type GoalTopicId } from "./goalPlanData";
import { englishOnboardingTranslate, type OnboardingTranslate } from "./onboardingTranslate";
function escapeForRegExp(value: string): string {
return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
}
/**
* Whole-word match so "post" does not fire on "postpone". Hyphens and
* apostrophes count as part of a word, which keeps "to-do" intact.
*/
function wordPattern(word: string): RegExp {
return new RegExp(`(?<![\\w-])${escapeForRegExp(word)}(?![\\w-])`, "u");
}
function countKeywords(answer: string, keywords: readonly string[]): number {
let matches = 0;
for (const keyword of keywords) {
if (wordPattern(keyword).test(answer)) matches += 1;
}
return matches;
}
/**
* Which plan the answer earns. Scored by how many of a topic's words the
* answer uses, so a sentence that leans one way gets that topic even when it
* brushes another. Ties go to the earlier topic, and an answer that matches
* nothing gets the general plan rather than a wrong guess.
*/
export function classifyDesktopOnboardingGoal(answer: string): GoalTopicId {
const normalized = answer.toLowerCase();
let best: GoalTopicId = "general";
let bestScore = 0;
for (const topic of GOAL_TOPICS) {
const score = countKeywords(normalized, topic.keywords);
if (score > bestScore) {
best = topic.id;
bestScore = score;
}
}
return best;
}
/** How many of the things the answer named are said back to it. */
const MAX_MENTIONS = 2;
/**
* The concrete things the answer named, in the order it named them. Reading
* order is what makes the echo feel like listening rather than like matching:
* someone who wrote "LinkedIn and X" is told "LinkedIn and X", not "X and
* LinkedIn" because of how the table happens to be sorted.
*/
function mentionsIn(
answer: string,
signals: readonly GoalSignal[],
t: OnboardingTranslate,
): readonly string[] {
const found: Array<{ readonly label: string; readonly at: number }> = [];
for (const signal of signals) {
const at = answer.search(wordPattern(signal.match));
const label = signal.label(t);
if (at === -1 || found.some((entry) => entry.label === label)) continue;
found.push({ label, at });
}
return found
.sort((left, right) => left.at - right.at)
.slice(0, MAX_MENTIONS)
.map((entry) => entry.label);
}
function joinMentions(mentions: readonly string[], t: OnboardingTranslate): string | null {
const [first, second] = mentions;
if (first === undefined) return null;
if (second === undefined) return first;
return t("{first} and {second}", { first, second });
}
export interface GoalPlan {
readonly topic: GoalTopicId;
/** What the bot says it will do, in the order it will do it. */
readonly steps: readonly [string, string, string];
/** Things the answer named, said back to it. Empty when it named none. */
readonly mentions: readonly string[];
/** The detail this work needs later, and when it will be asked for. */
readonly deferred: string;
}
/**
* The plan behind one answer. Deterministic and local: the topic decides the
* shape, and anything concrete the answer named leads the first step, so the
* plan reads as a response to what was written rather than as a template that
* happened to be picked.
*
* `t` translates the plan for display. The bot brief keeps the English default,
* because the plan it writes out is the user's first message, not interface copy.
*/
export function desktopOnboardingGoalPlan(
goal: string,
t: OnboardingTranslate = englishOnboardingTranslate,
): GoalPlan {
const answer = condense(goal).toLowerCase();
const id = classifyDesktopOnboardingGoal(answer);
const topic = GOAL_TOPICS.find((candidate) => candidate.id === id) ?? GENERAL_TOPIC;
const mentions = mentionsIn(answer, topic.signals, t);
const focus = joinMentions(mentions, t);
const lead = focus === null ? null : topic.lead.focused?.(t, focus);
return {
topic: id,
steps: [lead ?? topic.lead.plain(t), ...topic.rest(t)],
mentions,
deferred: topic.deferred(t),
};
}
/** One line per answer: the goal is quoted back to the user and to the bot. */
export function condense(value: string): string {
return value.trim().replace(/\s+/gu, " ");
}