pi-freeflow 1.0.6 → 1.0.8

package/README.md

@@ -137,17 +137,6 @@ export default {
       status: response.status,
       headers: response.headers,
     });
-  }
-};
-```
-
-After deploying, register your Cloudflare Worker URL in OMP:
-```text
-/freeflow use https://your-worker-name.your-subdomain.workers.dev
-```
-
----
-
 ### 2. Vercel Relay
 Deploy using the built-in deploy command:
 ```text
@@ -159,13 +148,20 @@ Deploy using the built-in deploy command:
 
 ## 🛠️ Diagnostics & Troubleshooting
 
+### Custom Port Configuration
+By default, `pi-freeflow` uses a **Single Shared Port** (`18080`). All concurrent sub-agents automatically share the same master proxy port without spawning redundant servers.
+
+To change the base port:
+```env
+FREEFLOW_PORT=19000
+```
+
 View live debug logs:
 ```bash
 cat ~/.pi/agent/pi-freeflow.log | tail -n 30
 # or inside OMP:
 /freeflow logs
 ```
-
 ---
 
 ## 📄 License

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);
@@ -1267,23 +1290,44 @@ function startProxy(
 	});
 }
 
-// ── Main extension ─────────────────────────────────────────────────
-export default async function (pi: ExtensionAPI) {
-	log("info", "extension loading...");
-	let server: http.Server;
-	let actualPort: number;
+// Probe whether an existing freeflow proxy is already running on a port
+async function isProxyAlive(port: number): Promise<boolean> {
 	try {
-		const r = await startProxy();
-		server = r.server;
-		actualPort = r.port;
+		const res = await fetch(`http://${HOST}:${port}/v1/models`, {
+			signal: AbortSignal.timeout(500),
+		});
+		const ct = res.headers.get("content-type") || "";
+		return res.ok && ct.includes("application/json");
 	} catch {
-		log(
-			"error",
-			"extension inactive — could not bind proxy port. resolve the port conflict and restart pi.",
-		);
-		return;
+		return false;
 	}
+}
 
+// ── Main extension ─────────────────────────────────────────────────
+export default async function (pi: ExtensionAPI) {
+	log("info", "extension loading...");
+	let server: http.Server | null = null;
+	let actualPort = PORT;
+
+	// Single-Port Shared Pattern: If proxy is already running on PORT (e.g. parent session),
+	// subagents reuse http://127.0.0.1:18080 directly without spawning redundant servers!
+	const alreadyRunning = await isProxyAlive(PORT);
+	if (alreadyRunning) {
+		log("info", `reusing existing freeflow proxy on http://${HOST}:${PORT}`);
+		actualPort = PORT;
+	} else {
+		try {
+			const r = await startProxy();
+			server = r.server;
+			actualPort = r.port;
+		} catch {
+			log(
+				"error",
+				"extension inactive — could not bind proxy port. resolve the port conflict and restart pi.",
+			);
+			return;
+		}
+	}
 	// Health check opencode models
 	log("info", `checking ${KNOWN_MODELS.length} opencode model(s)...`);
 	const opencodeChecks = await Promise.all(
@@ -1591,7 +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);
@@ -1600,7 +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 ||
@@ -1609,7 +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": "1.0.6",
+	"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",