pi-codebuddy-sdk 0.4.13 → 0.4.15

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-codebuddy-sdk",
-  "version": "0.4.13",
+  "version": "0.4.15",
   "description": "Pi extension that uses CodeBuddy (via Agent SDK) as a model provider and adds an AskCodebuddy tool.",
   "author": "slahser",
   "license": "MIT",
@@ -32,7 +32,7 @@
     "typecheck": "tsc --noEmit"
   },
   "dependencies": {
-    "@tencent-ai/agent-sdk": "^0.3.265",
+    "@tencent-ai/agent-sdk": "^0.3.267",
     "change-case": "^5.4.4",
     "typebox": "^1.3.34"
   },
@@ -45,7 +45,7 @@
     "@earendil-works/pi-ai": "^0.87.1",
     "@earendil-works/pi-coding-agent": "^0.87.1",
     "@earendil-works/pi-tui": "^0.87.1",
-    "@types/node": "^26.6.2",
+    "@types/node": "^26.6.3",
     "tsx": "^4.23.15",
     "typescript": "^7.0.2"
   },

package/src/index.ts

@@ -20,6 +20,7 @@ import { loadConfig, type Config } from "./config.js";
 import { jsonSchemaToZodShape } from "./typebox-to-zod.js";
 import { buildActionSummary, type ToolCallState } from "./askcodebuddy-ui.js";
 import { withSdkGate } from "./sdk-gate.js";
+import { closeQueryTransport, endQuery } from "./query-teardown.js";
 
 // Compat (#2): use factory if available (pi-ai ≥0.66), else fall back to constructor (gsd-pi etc.)
 const _piAi = piAi as any;
@@ -496,7 +497,7 @@ async function runIsolatedSummary(
 		stream.end();
 	} finally {
 		options?.signal?.removeEventListener("abort", onAbort);
-		try { sdkQuery?.interrupt(); } catch {}
+		endQuery(sdkQuery, "compact-summary", { log: debug });
 	}
 }
 
@@ -1667,7 +1668,7 @@ async function promptAndWait(
 		return { responseText, stopReason };
 	} finally {
 		signal?.removeEventListener("abort", onAbort);
-		void sdkQuery.interrupt().catch(() => {});
+		endQuery(sdkQuery, "askCodebuddy", { log: debug });
 	}
 }
 
@@ -1687,8 +1688,9 @@ let discoverInFlight: Promise<void> | null = null;
 
 async function discoverModels(pi: ExtensionAPI): Promise<void> {
 	await withSdkGate(async () => {
+		let q: ReturnType<typeof startQuery> | undefined;
 		try {
-			const q = startQuery({ prompt: " ", options: { maxTurns: 0, permissionMode: "bypassPermissions", tools: [] } });
+			q = startQuery({ prompt: " ", options: { maxTurns: 0, permissionMode: "bypassPermissions", tools: [] } });
 			const supported = await q.supportedModels();
 			await q.return().catch(() => {});
 			if (!supported.length) return;
@@ -1740,6 +1742,12 @@ async function discoverModels(pi: ExtensionAPI): Promise<void> {
 			} else {
 				debug("discoverModels: failed, using fallback models", err);
 			}
+		} finally {
+			// supportedModels() never iterates the query, so the SDK's cleanup()
+			// (the only place the transport is closed) never runs here — close it
+			// explicitly or the codebuddy-headless child leaks for the host's
+			// lifetime. Idempotent, so a completed iterator is a harmless no-op.
+			if (q) closeQueryTransport(q, "discoverModels", { log: debug });
 		}
 	});
 }

package/src/query-teardown.ts

@@ -0,0 +1,84 @@
+// Guarantee the CodeBuddy CLI subprocess is reaped once a query is finished.
+//
+// The @tencent-ai/agent-sdk `Query` closes its transport (which ends stdin and
+// terminates the spawned `codebuddy-headless.js` child) only from its private
+// `cleanup()`, and `cleanup()` runs only in `createIterator()`'s `finally` —
+// i.e. only if the query was actually iterated to completion. `Query.return()`
+// just sends an `interrupt` control frame and, per its own comment, "does not
+// close the transport".
+//
+// Any path that never iterates the query therefore leaks the child for the
+// lifetime of the host process. `discoverModels()` is the clearest offender: it
+// issues a single `supportedModels()` control request and abandons the query,
+// so its iterator (and `cleanup()`) never runs. Long-running sessions pile up
+// stopped CLIs and eventually exhaust the CLI's local port pool (EADDRINUSE on
+// 127.0.0.1:6252x/6253x).
+//
+// We reach for the transport the same way the issue #10 teardown guard in
+// index.ts already does: `Transport.close()` is public (transport/index.d.ts),
+// but `Query.transport` and `Query.cleanup()` are private (query.d.ts).
+// `close()` is idempotent (`if (this.closed) return;`), so calling it after a
+// query that already completed on its own is a harmless no-op.
+
+type ClosableTransport = { close?: () => void; process?: { pid?: number } };
+type TeardownQuery = { transport?: ClosableTransport | null };
+
+export type TeardownOptions = {
+	/** Force-kill the child if it is still alive this long after close(). */
+	forceKillAfterMs?: number;
+	/** Optional debug sink (wired to index.ts `debug()` at the call sites). */
+	log?: (msg: string) => void;
+};
+
+const DEFAULT_FORCE_KILL_MS = 2000;
+
+/**
+ * Close the query's transport and, as a backstop for a CLI that ignores
+ * SIGTERM, force-kill its child process if it is still alive shortly after.
+ * A no-op for anything that does not look like a closable query.
+ */
+export function closeQueryTransport(q: unknown, label: string, options: TeardownOptions = {}): void {
+	const transport = (q as TeardownQuery | null | undefined)?.transport;
+	if (!transport || typeof transport.close !== "function") return;
+
+	// Capture the pid before close() nulls out `transport.process`.
+	const pid = typeof transport.process?.pid === "number" ? transport.process.pid : undefined;
+
+	try {
+		transport.close();
+		options.log?.(`teardown(${label}): transport closed${pid !== undefined ? ` (pid ${pid})` : ""}`);
+	} catch (err) {
+		options.log?.(`teardown(${label}): transport close failed: ${err instanceof Error ? err.message : String(err)}`);
+	}
+
+	if (pid === undefined) return;
+
+	// Backstop: if the CLI did not exit on the transport's own terminate(),
+	// force-kill it so we never leak a stopped child.
+	const wait = options.forceKillAfterMs ?? DEFAULT_FORCE_KILL_MS;
+	const timer = setTimeout(() => {
+		try {
+			process.kill(pid, 0); // throws ESRCH once the child is gone
+		} catch {
+			return;
+		}
+		try {
+			process.kill(pid, "SIGKILL");
+			options.log?.(`teardown(${label}): CLI pid ${pid} survived transport close, sent SIGKILL`);
+		} catch {
+			// Raced with its own exit — nothing to do.
+		}
+	}, wait);
+	// Never let this safety net hold the process/event loop open.
+	timer.unref?.();
+}
+
+/** Stop the in-flight turn, then guarantee the CLI child process exits. */
+export function endQuery(q: unknown, label: string, options: TeardownOptions = {}): void {
+	try {
+		void (q as { interrupt?: () => Promise<unknown> } | null | undefined)?.interrupt?.()?.catch?.(() => {});
+	} catch {
+		// interrupt() failing on a half-dead query is expected; teardown still runs.
+	}
+	closeQueryTransport(q, label, options);
+}