pi-freeflow 1.0.7 → 1.0.8

package/extensions/index.ts

@@ -99,7 +99,11 @@ function loadRelayState(): RelayState {
 }
 function saveRelayState(s: RelayState): void {
 	try {
-		fs.writeFileSync(RELAY_STATE_FILE, JSON.stringify(s));
+		const dir = path.dirname(RELAY_STATE_FILE);
+		if (!fs.existsSync(dir)) fs.mkdirSync(dir, { recursive: true });
+		const tmpPath = `${RELAY_STATE_FILE}.${randomUUID()}.tmp`;
+		fs.writeFileSync(tmpPath, JSON.stringify(s, null, 2), "utf8");
+		fs.renameSync(tmpPath, RELAY_STATE_FILE);
 	} catch (e) {
 		log("warn", "could not persist relay state", { error: String(e) });
 	}
@@ -663,6 +667,7 @@ const ALLOWED_METHODS = new Set(["GET", "POST", "OPTIONS", "HEAD"]);
 const STRIP_HEADERS = new Set([
 	"authorization",
 	"host",
+	"content-length",
 	"x-forwarded-for",
 	"x-forwarded-host",
 	"x-forwarded-proto",
@@ -942,12 +947,16 @@ function normalizeRequestBody(
 		}
 	}
 
-	// 3. Max tokens clamp for Vercel relay
-	if (isRelay) {
-		const mt = body.max_tokens ?? body.maxTokens;
-		if (typeof mt === "number" && mt > RELAY_MAX_TOKENS) {
-			body.max_tokens = RELAY_MAX_TOKENS;
-		}
+	// 3. Max & Min tokens clamping (OpenCode Zen requires min 16; Vercel relay caps at 64k)
+	const mt = (body.max_tokens ?? body.maxTokens ?? body.max_output_tokens) as number | undefined;
+	if (typeof mt === "number") {
+		let clamped = mt;
+		if (clamped < 16) clamped = 16;
+		else if (isRelay && clamped > RELAY_MAX_TOKENS) clamped = RELAY_MAX_TOKENS;
+
+		if (body.max_tokens !== undefined) body.max_tokens = clamped;
+		if (body.maxTokens !== undefined) body.maxTokens = clamped;
+		if (body.max_output_tokens !== undefined) body.max_output_tokens = clamped;
 	}
 
 	return body;
@@ -1002,6 +1011,12 @@ function pipeUpstreamStream(
 	req.on("close", () => {
 		if (!nodeStream.destroyed) nodeStream.destroy();
 	});
+	res.on("close", () => {
+		if (!nodeStream.destroyed) nodeStream.destroy();
+	});
+	res.on("error", () => {
+		if (!nodeStream.destroyed) nodeStream.destroy();
+	});
 }
 
 // ── Start local proxy ──────────────────────────────────────────────
@@ -1083,13 +1098,15 @@ function startProxy(
 				if (isKilo && parsedBody) {
 					// KiloCode gateway routing (free models are keyless)
 					const isStream = parsedBody.stream === true;
+					const kiloBodyObj = structuredClone(parsedBody);
+					normalizeRequestBody(kiloBodyObj, true);
 					const response = await relayFetch(KILO_CHAT_URL, {
 						method: "POST",
 						headers: {
 							"Content-Type": "application/json",
 							Authorization: "Bearer kilo-free",
 						},
-						body: JSON.stringify(parsedBody),
+						body: JSON.stringify(kiloBodyObj),
 						signal: AbortSignal.timeout(300_000),
 					});
 					if (isStream && response.ok && response.body) {
@@ -1215,12 +1232,18 @@ function startProxy(
 							res.end();
 						}
 					});
-					proxy.setTimeout(30_000, () => {
+					proxy.setTimeout(300_000, () => {
 						proxy.destroy(new Error("timeout"));
 					});
 					req.on("aborted", () => {
 						if (!proxy.destroyed) proxy.destroy();
 					});
+					req.on("close", () => {
+						if (!proxy.destroyed) proxy.destroy();
+					});
+					res.on("close", () => {
+						if (!proxy.destroyed) proxy.destroy();
+					});
 					// ponytail: body already buffered in bodyChunks above for model routing;
 					// req is drained so pipe() would send an empty body → upstream hang → 502.
 					proxy.end(directBody);
@@ -1273,7 +1296,8 @@ async function isProxyAlive(port: number): Promise<boolean> {
 		const res = await fetch(`http://${HOST}:${port}/v1/models`, {
 			signal: AbortSignal.timeout(500),
 		});
-		return res.ok;
+		const ct = res.headers.get("content-type") || "";
+		return res.ok && ct.includes("application/json");
 	} catch {
 		return false;
 	}
@@ -1611,7 +1635,7 @@ export default async function (pi: ExtensionAPI) {
 	pi.registerCommand("bansos", commandSpec);
 
 	// Reload persisted state on session start/resume (env overrides still win).
-	pi.on("session_start", async (_event, ctx) => {
+	pi.on?.("session_start", async (_event, ctx) => {
 		relayState = resolveRelayState();
 		statusUi = ctx.ui;
 		ctx.ui?.setStatus?.("freeflow", undefined);
@@ -1620,7 +1644,7 @@ export default async function (pi: ExtensionAPI) {
 
 	// Pi normally pauses after threshold compaction. Queue a follow-up while the
 	// original run is still active so the core agent continues automatically.
-	pi.on("session_compact", (event, ctx) => {
+	pi.on?.("session_compact", (event, ctx) => {
 		if (
 			event.reason !== "threshold" ||
 			event.willRetry ||
@@ -1629,7 +1653,7 @@ export default async function (pi: ExtensionAPI) {
 		) {
 			return;
 		}
-		pi.sendUserMessage(
+		pi.sendUserMessage?.(
 			"Continue the current task from the compacted context. Do not wait for another user message; proceed with the next required step.",
 			{ deliverAs: "followUp" },
 		);

package/package.json

@@ -1,6 +1,7 @@
 {
 	"name": "pi-freeflow",
-	"version": "1.0.7",
+	"type": "module",
+	"version": "1.0.8",
 	"description": "Personal multi-cloud rolling fallback relay for OpenCode Zen and KiloCode models in OMP/Pi",
 	"keywords": [
 		"pi-package",