open-memex/dist/capture/keywords.js

Local-first memory layer and protocol for AI coding agents. Markdown source of truth, SQLite FTS5 index, zero cloud.

/**
 * D67: capture-body length bounds. The lower bound rejects fragments — a
 * trigger that yields "这个" or "OK" is a match on the trigger, not a memory,
 * and a 2-char memory is noise in every future session's injection. The floor
 * stays deliberately above 1: D67 considered lowering it and decided against
 * it. What changed is that a rejection is no longer silent (see KeywordDrop).
 */
export const MIN_CAPTURE_CHARS = 3;
export const MAX_CAPTURE_CHARS = 2000;
/**
 * Scan user text for memory-capture keyword triggers, reporting both what
 * would be captured and what matched but was rejected (D67) — a dropped match
 * that says nothing reads as a broken feature.
 * Each pattern must have a single capture group whose value becomes the memory body.
 * Personal patterns (cfg.keywordPersonalPatterns) force the personal scope.
 *
 * Personal patterns run first and claim their line: an explicit scope signal
 * (e.g. "记住我对花粉过敏") must not also fire the generic 记住 into the project scope.
 */
export function scanKeywords(userText, cfg) {
    const hits = [];
    const dropped = [];
    if (!cfg.keywordCaptureEnabled)
        return { hits, dropped };
    const lines = userText.split(/\r?\n/);
    const claimed = new Set();
    const scan = (patterns, personal, skipClaimed) => {
        for (const src of patterns) {
            let re;
            try {
                re = new RegExp(src, "i");
            }
            catch {
                continue;
            }
            for (let i = 0; i < lines.length; i++) {
                if (skipClaimed && claimed.has(i))
                    continue;
                const m = lines[i].match(re);
                if (!m || !m[1])
                    continue;
                const content = m[1].trim().replace(/[.!?]$/, "").trim();
                if (personal)
                    claimed.add(i);
                if (content.length >= MIN_CAPTURE_CHARS && content.length <= MAX_CAPTURE_CHARS) {
                    hits.push({ content, pattern: src, personal });
                    break; // one hit per pattern per message
                }
                // Rejected: record it once per pattern. The line stays claimed for
                // personal patterns even here - "记住我：OK" must not fall through to
                // the generic 记住 and be saved as "我：OK" (D67).
                if (!dropped.some((d) => d.pattern === src && d.content === content)) {
                    dropped.push({
                        content,
                        pattern: src,
                        personal,
                        reason: content.length < MIN_CAPTURE_CHARS ? "too-short" : "too-long",
                    });
                }
                break;
            }
        }
    };
    scan(cfg.keywordPersonalPatterns, true, false);
    scan(cfg.keywordPatterns, false, true);
    if (cfg.logLevel === "debug" && dropped.length > 0) {
        for (const d of dropped) {
            console.error(`[open-memex] keyword match not captured (${d.reason === "too-short" ? `body ${d.content.length} chars < ${MIN_CAPTURE_CHARS}` : `body ${d.content.length} chars > ${MAX_CAPTURE_CHARS}`}): "${d.content}"`);
        }
    }
    return { hits, dropped };
}
/** Capture-only view of {@link scanKeywords} (drops are logged, not returned). */
export function detectKeywords(userText, cfg) {
    return scanKeywords(userText, cfg).hits;
}