@oh-my-pi/pi-coding-agent 14.2.1 → 14.3.0

package/src/web/search/providers/codex.ts

@@ -106,6 +106,125 @@ function isImagePlaceholderAnswer(text: string): boolean {
 	return text.trim().toLowerCase() === "(see attached image)";
 }
 
+function addSource(sources: SearchSource[], source: SearchSource): void {
+	if (!sources.some(existing => existing.url === source.url)) {
+		sources.push(source);
+	}
+}
+
+function countCharacter(text: string, target: string): number {
+	let count = 0;
+	for (const char of text) {
+		if (char === target) {
+			count += 1;
+		}
+	}
+	return count;
+}
+
+/**
+ * Strips prose punctuation and unmatched closing delimiters from extracted URLs.
+ * Codex often returns links in markdown or sentence text without structured annotations.
+ */
+function normalizeExtractedUrl(candidate: string): string | null {
+	let url = candidate.trim();
+
+	while (url.length > 0) {
+		const lastCharacter = url.at(-1);
+		if (!lastCharacter) break;
+		if (/[.,!?;:'"]/u.test(lastCharacter)) {
+			url = url.slice(0, -1);
+			continue;
+		}
+		if (lastCharacter === ")" && countCharacter(url, ")") > countCharacter(url, "(")) {
+			url = url.slice(0, -1);
+			continue;
+		}
+		if (lastCharacter === "]" && countCharacter(url, "]") > countCharacter(url, "[")) {
+			url = url.slice(0, -1);
+			continue;
+		}
+		if (lastCharacter === "}" && countCharacter(url, "}") > countCharacter(url, "{")) {
+			url = url.slice(0, -1);
+			continue;
+		}
+		break;
+	}
+
+	if (!/^https?:\/\//.test(url)) {
+		return null;
+	}
+
+	try {
+		return new URL(url).toString();
+	} catch {
+		return null;
+	}
+}
+
+function findMarkdownLinkUrlEnd(text: string, openParenIndex: number): number | null {
+	let depth = 0;
+
+	for (let index = openParenIndex; index < text.length; index += 1) {
+		const character = text[index];
+		if (!character || character === "\n") {
+			return null;
+		}
+		if (character === "(") {
+			depth += 1;
+			continue;
+		}
+		if (character !== ")") {
+			continue;
+		}
+		depth -= 1;
+		if (depth === 0) {
+			return index;
+		}
+		if (depth < 0) {
+			return null;
+		}
+	}
+
+	return null;
+}
+
+/**
+ * Extracts citation sources from markdown links and bare URLs in the answer text.
+ * Used as a fallback when the Codex response omits `url_citation` annotations.
+ */
+function extractTextSources(text: string): SearchSource[] {
+	const sources: SearchSource[] = [];
+
+	for (let index = 0; index < text.length; index += 1) {
+		if (text[index] !== "[") {
+			continue;
+		}
+		const titleEnd = text.indexOf("]", index + 1);
+		if (titleEnd === -1 || text[titleEnd + 1] !== "(") {
+			continue;
+		}
+		const urlEnd = findMarkdownLinkUrlEnd(text, titleEnd + 1);
+		if (urlEnd === null) {
+			continue;
+		}
+		const title = text.slice(index + 1, titleEnd).trim();
+		const url = normalizeExtractedUrl(text.slice(titleEnd + 2, urlEnd));
+		if (url) {
+			addSource(sources, { title: title || url, url });
+		}
+		index = urlEnd;
+	}
+
+	for (const match of text.matchAll(/https?:\/\/\S+/g)) {
+		const url = normalizeExtractedUrl(match[0] ?? "");
+		if (!url) continue;
+		addSource(sources, { title: url, url });
+	}
+
+	return sources;
+}
+
 /**
  * Extracts account ID from a Codex access token.
  * @param accessToken - JWT access token
@@ -211,6 +330,7 @@ async function callCodexSearch(
 				search_context_size: options.searchContextSize ?? "high",
 			},
 		],
+		tool_choice: { type: "web_search" },
 		instructions: options.systemPrompt ?? DEFAULT_INSTRUCTIONS,
 	};
 
@@ -262,12 +382,7 @@ async function callCodexSearch(
 							for (const annotation of part.annotations) {
 								if (annotation.type === "url_citation" && annotation.url) {
 									// Deduplicate by URL
-									if (!sources.some(s => s.url === annotation.url)) {
-										sources.push({
-											title: annotation.title ?? annotation.url,
-											url: annotation.url,
-										});
-									}
+									addSource(sources, { title: annotation.title ?? annotation.url, url: annotation.url });
 								}
 							}
 						}
@@ -317,6 +432,14 @@ async function callCodexSearch(
 				? streamedAnswer
 				: finalAnswer;
 
+	// Fallback: when Codex omits url_citation annotations, scrape markdown links
+	// and bare URLs from the synthesized answer so callers still receive sources.
+	if (sources.length === 0 && answer.length > 0) {
+		for (const source of extractTextSources(answer)) {
+			addSource(sources, source);
+		}
+	}
+
 	return {
 		answer,
 		sources,