port: sync from Mike/pygienium@4e56b46d

src/agent-runner.ts

@@ -246,7 +246,7 @@ async function runSessionToCompletion(
 	session: AgentSession,
 	opts: AgentTaskOptions,
 ): Promise<AgentRunResult> {
-	const acc: SessionEventAccumulator = { text: "" };
+	const acc: SessionEventAccumulator = { text: "", sawMessage: false };
 	try {
 		const unsubscribe = session.subscribe((event: AgentSessionEvent) => {
 			applySessionEvent(acc, event, opts.onEvent);
@@ -259,16 +259,15 @@ async function runSessionToCompletion(
 		unsubscribe();
 		// Surface session errors that didn't throw but left no useful output.
 		// A session ending with stopReason "error" and no text means the model
-		// call failed silently — treat that as a failed run, not ok:true.
-		if (acc.errorMessage) {
-			return { ok: false, text: acc.text, error: acc.errorMessage };
-		}
-		if (!acc.text.trim() && acc.stopReason === "error") {
-			return {
-				ok: false,
-				text: acc.text,
-				error: "sub-agent session ended in error with no output.",
-			};
+		// call failed silently — treat that as a failed run, not ok:true. A
+		// session that settled with NO text and NO observed events means the
+		// model never produced anything at all (dead provider stream, failed
+		// start) — reporting that as a successful analysis would let an empty
+		// scan masquerade as a clean one, and the check's verify hook would
+		// fail only later with a confusing "artifact missing" error.
+		const sessionError = emptySessionError(acc);
+		if (sessionError) {
+			return { ok: false, text: acc.text, error: sessionError };
 		}
 		return { ok: true, text: acc.text };
 	} catch (err) {
@@ -290,12 +289,34 @@ async function runSessionToCompletion(
 export interface SessionEventAccumulator {
 	/** Joined assistant text seen so far (text_delta stream). */
 	text: string;
+	/** Whether any assistant-message or tool event was observed at all. */
+	sawMessage: boolean;
 	/** stopReason of the final assistant message, when reported. */
 	stopReason?: string;
 	/** errorMessage of the final assistant message, when reported. */
 	errorMessage?: string;
 }
 
+/**
+ * Classify a settled session's capture: `undefined` when the result is a
+ * legitimate (possibly empty-text) outcome, else the error that should fail
+ * the run. Kept pure so the decision is unit-testable without a session.
+ */
+export function emptySessionError(
+	acc: SessionEventAccumulator,
+): string | undefined {
+	if (acc.errorMessage) return acc.errorMessage;
+	if (!acc.text.trim()) {
+		if (acc.stopReason === "error") {
+			return "sub-agent session ended in error with no output.";
+		}
+		if (!acc.sawMessage) {
+			return "sub-agent session settled with no output — no assistant message or tool activity was observed. Check model/provider connectivity, then resume.";
+		}
+	}
+	return undefined;
+}
+
 /**
  * Interpret one session event into the running accumulator and forward the
  * stream-driving events to the chat.
@@ -316,6 +337,16 @@ export function applySessionEvent(
 ): void {
 	if (!event) return;
 	try {
+		// Any message or tool event means the session actually ran — a settled
+		// capture with none of these is a dead session, not an empty scan.
+		if (
+			event.type === "message_update" ||
+			event.type === "message_end" ||
+			event.type === "tool_execution_start" ||
+			event.type === "tool_execution_end"
+		) {
+			acc.sawMessage = true;
+		}
 		if (
 			event.type === "message_update" &&
 			event.assistantMessageEvent?.type === "text_delta"

src/checks/scope.ts

@@ -56,6 +56,13 @@ export const SCOPE_EXCLUDE_DIRS: ReadonlySet<string> = new Set([
 	".nuxt",
 	".turbo",
 	".svelte-kit",
+	// Framework/deploy build output: Nitro, Vercel, Netlify artifact dirs.
+	// Generated bundles dominate a naive marker/stub scan (freno-dev alone
+	// flagged 119 of 124 candidates inside `.output/`/`.vercel/` minified
+	// bundles) and must never feed the pre-scan or the agent's file walk.
+	".output",
+	".vercel",
+	".netlify",
 	"__pycache__",
 	".venv",
 	"venv",

src/checks/todos.ts

@@ -124,6 +124,21 @@ const MAX_CANDIDATES = 500;
 const FALLBACK_CAP = 16;
 /** Candidate list embedded in the live prompt is truncated at this many. */
 const PROMPT_CAP = 40;
+/**
+ * Max characters of a candidate's code line embedded in the task prompt.
+ * Generated/bundled single lines can be hundreds of KB (e.g. minified assets
+ * sneaking past scope); embedding them wholesale balloons the task to
+ * megabytes and chokes the sub-agent. `path:line` plus a truncated prefix is
+ * enough to classify — the agent can read the file for full context.
+ */
+const CODE_DISPLAY_CAP = 160;
+
+/** Truncate a candidate's code line for prompt embedding. */
+function displayCode(code: string): string {
+	return code.length > CODE_DISPLAY_CAP
+		? `${code.slice(0, CODE_DISPLAY_CAP)}…`
+		: code;
+}
 
 /** Extract the declared function name from a header line, when present. */
 function headerName(line: string): string | undefined {
@@ -369,7 +384,7 @@ function renderFindings(
 		parts.push(`## ${title}`);
 		list.slice(0, FALLBACK_CAP).forEach((c, i) => {
 			parts.push(
-				`### ${i + 1}. ${relative(cwd, c.path)}:${c.line} — ${c.code} | snippet: ${c.snippet} | context: ${c.context || relative(cwd, c.path)}`,
+				`### ${i + 1}. ${relative(cwd, c.path)}:${c.line} — ${displayCode(c.code)} | snippet: ${c.snippet} | context: ${c.context || relative(cwd, c.path)}`,
 			);
 		});
 		if (list.length > FALLBACK_CAP) {
@@ -434,7 +449,7 @@ export async function buildTodosScanTask(
 
 	const candidateList = candidates
 		.slice(0, PROMPT_CAP)
-		.map((c) => ` - ${relative(cwd, c.path)}:${c.line} [${c.kind}] ${c.code}`);
+		.map((c) => ` - ${relative(cwd, c.path)}:${c.line} [${c.kind}] ${displayCode(c.code)}`);
 	if (candidates.length > PROMPT_CAP) {
 		candidateList.push(
 			` - … and ${candidates.length - PROMPT_CAP} more (truncated for brevity)`,