@oh-my-pi/pi-coding-agent 16.3.6 → 16.3.7

package/dist/types/web/search/providers/base.d.ts

@@ -1,5 +1,5 @@
 import type { AuthStorage, FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchProviderId, SearchResponse } from "../types";
+import type { SearchProviderId, SearchResponse } from "../types.js";
 /**
  * Shared web search parameters passed to providers.
  *

package/dist/types/web/search/providers/brave.d.ts

@@ -5,9 +5,9 @@
  * SearchResponse shape used by the web search tool.
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 export interface BraveSearchParams {
     query: string;
     num_results?: number;

package/dist/types/web/search/providers/codex.d.ts

@@ -1,7 +1,7 @@
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 export interface CodexSearchParams {
     signal?: AbortSignal;
     fetch?: FetchImpl;

package/dist/types/web/search/providers/duckduckgo.d.ts

@@ -1,7 +1,7 @@
 import type { AuthStorage } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 /** Execute a DuckDuckGo web search via the no-JS HTML frontend. */
 export declare function searchDuckDuckGo(params: SearchParams): Promise<SearchResponse>;
 /** Search provider for DuckDuckGo (no API key required). */

package/dist/types/web/search/providers/exa.d.ts

@@ -7,9 +7,9 @@
  * them into a combined `answer` string on the SearchResponse.
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 /** Reset Exa request pacing state for isolated provider tests. */
 export declare function resetExaSearchThrottleForTest(): void;
 type ExaSearchType = "neural" | "fast" | "auto" | "deep";

package/dist/types/web/search/providers/firecrawl.d.ts

@@ -5,9 +5,9 @@
  * SearchResponse shape used by the web search tool.
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 export interface FirecrawlSearchParams {
     query: string;
     num_results?: number;

package/dist/types/web/search/providers/gemini.d.ts

@@ -9,9 +9,9 @@
  * endpoint.
  */
 import { type AuthStorage, type FetchImpl, type OAuthAccess } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 declare const GEMINI_PROVIDERS: readonly ["google-gemini-cli", "google-antigravity"];
 type GeminiProviderId = (typeof GEMINI_PROVIDERS)[number];
 interface GeminiToolParams {

package/dist/types/web/search/providers/jina.d.ts

@@ -5,9 +5,9 @@
  * cleaned content.
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 type SearchParamsWithFetch = SearchParams & {
     fetch?: FetchImpl;
 };

package/dist/types/web/search/providers/kagi.d.ts

@@ -4,9 +4,9 @@
  * Thin wrapper that adapts shared Kagi API utilities to SearchResponse shape.
  */
 import type { AuthStorage, FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 type SearchParamsWithFetch = SearchParams & {
     fetch?: FetchImpl;
 };

package/dist/types/web/search/providers/kimi.d.ts

@@ -5,9 +5,9 @@
  * Endpoint: POST https://api.kimi.com/coding/v1/search
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 type SearchParamsWithFetch = SearchParams & {
     fetch?: FetchImpl;
 };

package/dist/types/web/search/providers/parallel.d.ts

@@ -1,7 +1,7 @@
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 export declare function searchParallel(params: {
     query: string;
     num_results?: number;

package/dist/types/web/search/providers/perplexity.d.ts

@@ -8,9 +8,9 @@
  * - Anonymous via `www.perplexity.ai/rest/sse/perplexity_ask`
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 export interface PerplexitySearchParams {
     signal?: AbortSignal;
     query: string;

package/dist/types/web/search/providers/searxng.d.ts

@@ -25,9 +25,9 @@
  * Reference: https://docs.searxng.org/dev/search_api.html
  */
 import type { AuthStorage, FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 /** Execute SearXNG web search. */
 export declare function searchSearXNG(params: {
     query: string;

package/dist/types/web/search/providers/synthetic.d.ts

@@ -5,9 +5,9 @@
  * Endpoint: POST https://api.synthetic.new/v2/search
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 type SearchParamsWithFetch = SearchParams & {
     fetch?: FetchImpl;
 };

package/dist/types/web/search/providers/tavily.d.ts

@@ -5,9 +5,9 @@
  * optional synthesized answer.
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 export interface TavilySearchParams {
     query: string;
     num_results?: number;

package/dist/types/web/search/providers/tinyfish.d.ts

@@ -5,9 +5,9 @@
  * SearchResponse shape used by the web search tool.
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 export interface TinyFishSearchParams {
     query: string;
     num_results?: number;

package/dist/types/web/search/providers/utils.d.ts

@@ -1,5 +1,5 @@
-import type { AgentStorage } from "../../../session/agent-storage";
-import { SearchProviderError, type SearchProviderId, type SearchSource } from "../../../web/search/types";
+import type { AgentStorage } from "../../../session/agent-storage.js";
+import { SearchProviderError, type SearchProviderId, type SearchSource } from "../../../web/search/types.js";
 /**
  * Search for an API credential by checking an env-derived key first,
  * then falling back to agent.db stored credentials for the given providers.

package/dist/types/web/search/providers/xai.d.ts

@@ -1,7 +1,7 @@
 import { type AuthStorage } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 /** Execute xAI Responses API web search. */
 export declare function searchXAI(params: SearchParams): Promise<SearchResponse>;
 /** Search provider for xAI web search. */

package/dist/types/web/search/providers/zai.d.ts

@@ -5,9 +5,9 @@
  * the unified SearchResponse shape used by the web search tool.
  */
 import { type AuthStorage, type FetchImpl } from "@oh-my-pi/pi-ai";
-import type { SearchResponse } from "../../../web/search/types";
-import type { SearchParams } from "./base";
-import { SearchProvider } from "./base";
+import type { SearchResponse } from "../../../web/search/types.js";
+import type { SearchParams } from "./base.js";
+import { SearchProvider } from "./base.js";
 export interface ZaiSearchParams {
     query: string;
     num_results?: number;

package/dist/types/web/search/render.d.ts

@@ -4,9 +4,9 @@
  * Tree-based rendering with collapsed/expanded states for web search results.
  */
 import type { Component } from "@oh-my-pi/pi-tui";
-import type { RenderResultOptions } from "../../extensibility/custom-tools/types";
-import { type Theme } from "../../modes/theme/theme";
-import type { SearchResponse } from "./types";
+import type { RenderResultOptions } from "../../extensibility/custom-tools/types.js";
+import { type Theme } from "../../modes/theme/theme.js";
+import type { SearchResponse } from "./types.js";
 export interface SearchRenderDetails {
     response: SearchResponse;
     error?: string;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-coding-agent",
-	"version": "16.3.6",
+	"version": "16.3.7",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -56,17 +56,17 @@
 		"@agentclientprotocol/sdk": "0.25.0",
 		"@babel/parser": "^7.29.7",
 		"@mozilla/readability": "^0.6.0",
-		"@oh-my-pi/hashline": "16.3.6",
-		"@oh-my-pi/omp-stats": "16.3.6",
-		"@oh-my-pi/pi-agent-core": "16.3.6",
-		"@oh-my-pi/pi-ai": "16.3.6",
-		"@oh-my-pi/pi-catalog": "16.3.6",
-		"@oh-my-pi/pi-mnemopi": "16.3.6",
-		"@oh-my-pi/pi-natives": "16.3.6",
-		"@oh-my-pi/pi-tui": "16.3.6",
-		"@oh-my-pi/pi-utils": "16.3.6",
-		"@oh-my-pi/pi-wire": "16.3.6",
-		"@oh-my-pi/snapcompact": "16.3.6",
+		"@oh-my-pi/hashline": "16.3.7",
+		"@oh-my-pi/omp-stats": "16.3.7",
+		"@oh-my-pi/pi-agent-core": "16.3.7",
+		"@oh-my-pi/pi-ai": "16.3.7",
+		"@oh-my-pi/pi-catalog": "16.3.7",
+		"@oh-my-pi/pi-mnemopi": "16.3.7",
+		"@oh-my-pi/pi-natives": "16.3.7",
+		"@oh-my-pi/pi-tui": "16.3.7",
+		"@oh-my-pi/pi-utils": "16.3.7",
+		"@oh-my-pi/pi-wire": "16.3.7",
+		"@oh-my-pi/snapcompact": "16.3.7",
 		"@opentelemetry/api": "^1.9.1",
 		"@opentelemetry/context-async-hooks": "^2.7.1",
 		"@opentelemetry/exporter-trace-otlp-proto": "^0.218.0",

package/src/advisor/__tests__/advisor.test.ts

@@ -164,6 +164,65 @@ describe("advisor", () => {
 			expect(md).toContain("[irc]");
 			expect(md).not.toContain("<primary-context");
 		});
+
+		it("omits hidden non-primary custom messages while keeping visible custom messages", () => {
+			const hiddenPrelude = {
+				role: "custom",
+				customType: "eager-todo-prelude",
+				content: "<system-reminder>Task delegation is enabled",
+				display: false,
+				timestamp: 1,
+			} as AgentMessage;
+			const hiddenHookMessage = {
+				role: "hookMessage",
+				customType: "hidden-hook-reminder",
+				content: "Hidden hook reminder should never reach advisor history",
+				display: false,
+				timestamp: 2,
+			} as AgentMessage;
+			const visibleCustom = {
+				role: "custom",
+				customType: "visible-status",
+				content: "Visible custom update",
+				display: true,
+				timestamp: 3,
+			} as AgentMessage;
+
+			const md = formatSessionHistoryMarkdown([hiddenPrelude, hiddenHookMessage, visibleCustom], {
+				expandPrimaryContext: true,
+			});
+
+			expect(md).toContain("[visible-status] Visible custom update");
+			expect(md).not.toContain("eager-todo-prelude");
+			expect(md).not.toContain("system-reminder");
+			expect(md).not.toContain("Task delegation");
+			expect(md).not.toContain("hidden-hook-reminder");
+			expect(md).not.toContain("Hidden hook reminder");
+		});
+
+		it("keeps hidden image descriptions because they are the text transcript for attached images", () => {
+			const imageDescription = {
+				role: "custom",
+				customType: "image-attachment-description",
+				content: [{ type: "text", text: '<image path="local://session/cat.png">cat on a keyboard</image>' }],
+				display: false,
+				timestamp: 1,
+			} as AgentMessage;
+			const hiddenPrelude = {
+				role: "custom",
+				customType: "eager-todo-prelude",
+				content: "<system-reminder>Task delegation is enabled",
+				display: false,
+				timestamp: 2,
+			} as AgentMessage;
+
+			const md = formatSessionHistoryMarkdown([imageDescription, hiddenPrelude], { expandPrimaryContext: true });
+
+			expect(md).toContain("[image-attachment-description]");
+			expect(md).toContain("cat on a keyboard");
+			expect(md).not.toContain("eager-todo-prelude");
+			expect(md).not.toContain("Task delegation");
+		});
 	});
 
 	describe("formatSessionHistoryMarkdown expandEditDiffs", () => {

package/src/auto-thinking/classifier.ts

@@ -35,13 +35,14 @@ const DIFFICULTY_SYSTEM_PROMPT = prompt.render(difficultySystemPrompt);
 const MAX_INPUT_CHARS = 6000;
 const HEAD_CHARS = 4000;
 const TAIL_CHARS = 2000;
-/** The online answer is a single word; keep budgets tiny for non-reasoning backends. */
-const ANSWER_MAX_TOKENS = 8;
 /** Local classifiers occasionally need more room for chat-template boilerplate. */
 const LOCAL_ANSWER_MAX_TOKENS = 16;
 /**
- * Reasoning backends ignore `disableReasoning` on some providers, so reserve
- * enough output room for the keyword to still land after unavoidable thinking.
+ * Online classifier budget. Sized to survive backends that ignore
+ * `disableReasoning` (e.g. Qwen3 via llama.cpp catalogued `reasoning: false`
+ * but still emitting thinking): the classifier keyword needs to land after any
+ * unavoidable thinking preamble. `maxTokens` is a hard cap — non-thinking
+ * completions still return in a handful of tokens (issue #4355).
  */
 const REASONING_SAFE_MAX_TOKENS = 1024;
 
@@ -85,7 +86,7 @@ async function classifyOnline(input: string, deps: ClassifyDifficultyDeps): Prom
 	}
 	// Resolve metadata after getApiKey so the session-sticky credential is recorded first.
 	const metadata = deps.metadataResolver?.(model.provider);
-	const maxTokens = model.reasoning ? Math.max(ANSWER_MAX_TOKENS, REASONING_SAFE_MAX_TOKENS) : ANSWER_MAX_TOKENS;
+	const maxTokens = REASONING_SAFE_MAX_TOKENS;
 
 	const response = await completeSimple(
 		model,

package/src/collab/relay-client.ts

@@ -21,6 +21,9 @@ const BACKOFF_BASE_MS = 1_000;
 const BACKOFF_MAX_MS = 30_000;
 /** Max enveloped frames buffered while a reconnect is pending; overflow is dropped. */
 const MAX_PENDING_SENDS = 256;
+const WS_BACKPRESSURE_THRESHOLD = 64 * 1024;
+const WS_BACKPRESSURE_DRAIN_THRESHOLD = 32 * 1024;
+const WS_BACKPRESSURE_DRAIN_RETRY_MS = 25;
 
 export interface CollabSocketOptions {
 	/** wss://host[:port]/r/<roomId> — no query string. */
@@ -40,6 +43,7 @@ export class CollabSocket {
 	readonly #opts: CollabSocketOptions;
 	#ws: WebSocket | null = null;
 	#retryTimer: NodeJS.Timeout | undefined;
+	#backpressureDrainTimer: NodeJS.Timeout | undefined;
 	#attempt = 0;
 	/** Terminal state: intentional close or fatal failure. Cleared by connect(). */
 	#closed = false;
@@ -72,28 +76,84 @@ export class CollabSocket {
 					logger.debug("collab: dropping frame, socket closed", { t: frame.t });
 					return;
 				}
+				const openWs = this.#ws;
+				if (openWs && openWs.readyState === WebSocket.OPEN) this.#drainPendingSends(openWs);
 				const sealed = await seal(this.#opts.key, frame);
 				const envelope = packEnvelope(targetPeer, sealed);
 				const ws = this.#ws;
 				if (ws && ws.readyState === WebSocket.OPEN) {
+					if (this.#pendingSends.length > 0) {
+						this.#enqueuePendingSend(envelope, frame.t);
+						if (ws.bufferedAmount < WS_BACKPRESSURE_DRAIN_THRESHOLD) {
+							this.#drainPendingSends(ws);
+						} else {
+							this.#scheduleBackpressureDrain(ws);
+						}
+						return;
+					}
+					if (ws.bufferedAmount >= WS_BACKPRESSURE_THRESHOLD) {
+						this.#enqueuePendingSend(envelope, frame.t);
+						this.#scheduleBackpressureDrain(ws);
+						return;
+					}
 					ws.send(envelope);
 					return;
 				}
-				if (this.#pendingSends.length >= MAX_PENDING_SENDS) {
-					logger.debug("collab: dropping frame, reconnect buffer full", { t: frame.t });
-					return;
-				}
-				this.#pendingSends.push(envelope);
+				this.#enqueuePendingSend(envelope, frame.t);
 			})
 			.catch((err: unknown) => {
 				logger.debug("collab: send failed", { error: String(err) });
 			});
 	}
 
+	#enqueuePendingSend(envelope: Uint8Array, frameType: CollabFrame["t"]): void {
+		if (this.#pendingSends.length >= MAX_PENDING_SENDS) {
+			logger.debug("collab: dropping frame, reconnect buffer full", { t: frameType });
+			return;
+		}
+		this.#pendingSends.push(envelope);
+	}
+
+	#drainPendingSends(ws: WebSocket): void {
+		while (
+			this.#pendingSends.length > 0 &&
+			ws.readyState === WebSocket.OPEN &&
+			ws.bufferedAmount < WS_BACKPRESSURE_DRAIN_THRESHOLD
+		) {
+			const envelope = this.#pendingSends.shift();
+			if (!envelope) return;
+			ws.send(envelope);
+		}
+	}
+
+	#scheduleBackpressureDrain(ws: WebSocket): void {
+		if (this.#backpressureDrainTimer !== undefined) return;
+		this.#backpressureDrainTimer = setTimeout(() => {
+			this.#backpressureDrainTimer = undefined;
+			this.#sendChain = this.#sendChain
+				.then(async () => {
+					if (this.#closed || this.#ws !== ws || ws.readyState !== WebSocket.OPEN) return;
+					this.#drainPendingSends(ws);
+					if (this.#pendingSends.length > 0) this.#scheduleBackpressureDrain(ws);
+				})
+				.catch((err: unknown) => {
+					logger.debug("collab: backpressure drain failed", { error: String(err) });
+				});
+		}, WS_BACKPRESSURE_DRAIN_RETRY_MS);
+	}
+
+	#clearBackpressureDrain(): void {
+		if (this.#backpressureDrainTimer !== undefined) {
+			clearTimeout(this.#backpressureDrainTimer);
+			this.#backpressureDrainTimer = undefined;
+		}
+	}
+
 	/** Intentional close: clears any retry timer, suppresses reconnect. A later connect() starts fresh. */
 	close(): void {
 		const hadActivity = this.#ws !== null || this.#retryTimer !== undefined;
 		this.#clearRetry();
+		this.#clearBackpressureDrain();
 		const wasClosed = this.#closed;
 		this.#closed = true;
 		this.#pendingSends.length = 0;
@@ -110,14 +170,17 @@ export class CollabSocket {
 	}
 
 	#openSocket(): void {
+		this.#clearBackpressureDrain();
 		const ws = new WebSocket(`${this.#opts.wsUrl}?role=${this.#opts.role}`);
 		ws.binaryType = "arraybuffer";
 		this.#ws = ws;
 		ws.onopen = () => {
 			if (this.#ws !== ws) return;
 			this.#attempt = 0;
-			for (const envelope of this.#pendingSends) ws.send(envelope);
-			this.#pendingSends.length = 0;
+			if (this.#pendingSends.length > 0) {
+				this.#drainPendingSends(ws);
+				if (this.#pendingSends.length > 0) this.#scheduleBackpressureDrain(ws);
+			}
 			this.onOpen?.();
 		};
 		ws.onmessage = (event: MessageEvent) => {
@@ -129,6 +192,7 @@ export class CollabSocket {
 		};
 		ws.onclose = (event: CloseEvent) => {
 			if (this.#ws !== ws) return;
+			this.#clearBackpressureDrain();
 			this.#ws = null;
 			this.#handleClose(event.code, event.reason);
 		};
@@ -167,6 +231,7 @@ export class CollabSocket {
 
 	#handleClose(code: number, reason: string): void {
 		if (this.#closed) return;
+		this.#clearBackpressureDrain();
 		const fatalReason = FATAL_CLOSE_REASONS[code];
 		if (fatalReason !== undefined) {
 			this.#closed = true;
@@ -186,6 +251,7 @@ export class CollabSocket {
 		this.#pendingSends.length = 0;
 		const ws = this.#ws;
 		this.#ws = null;
+		this.#clearBackpressureDrain();
 		if (ws) {
 			try {
 				ws.close(1000);

package/src/config/model-registry.ts

@@ -48,6 +48,13 @@ const STARTUP_MODEL_CACHE_PROVIDER_IDS: readonly string[] = [
 // packages/ai/src/registry/lm-studio.ts, and packages/ai/src/registry/vllm.ts.
 const LOCAL_PROVIDER_PLACEHOLDERS = new Set<string>(["llama-cpp-local", "lm-studio-local", "vllm-local"]);
 
+/**
+ * Hard bound for extension-provided fetchDynamicModels to prevent indefinite hangs
+ * during runtime provider discovery. Uses a cancellable manual timer (not AbortSignal.timeout)
+ * so a successful fast path does not leave an armed timeout signal for concurrent GC.
+ */
+const RUNTIME_DYNAMIC_MODEL_FETCH_TIMEOUT_MS = 15_000;
+
 import type { ApiKeyResolver, FetchImpl } from "@oh-my-pi/pi-ai";
 import { registerOAuthProvider, unregisterOAuthProviders } from "@oh-my-pi/pi-ai/oauth";
 import type { OAuthCredentials, OAuthLoginCallbacks } from "@oh-my-pi/pi-ai/oauth/types";
@@ -81,6 +88,24 @@ function isDiscoveryBearerApiKey(apiKey: string | undefined | null): apiKey is s
 	return isAuthenticated(apiKey) && !LOCAL_PROVIDER_PLACEHOLDERS.has(apiKey);
 }
 
+/**
+ * Wraps an extension-provided fetchDynamicModels call with a hard timeout.
+ * Uses a cancellable manual timer (not AbortSignal.timeout) so that a fast
+ * successful path does not leave an armed timeout signal for concurrent GC.
+ * The inner fetcher does not receive a signal (extension contract has none).
+ */
+async function withRuntimeDynamicModelsTimeout<T>(timeoutMs: number, run: () => Promise<T>): Promise<T> {
+	const { promise: timeoutPromise, reject: timeoutReject } = Promise.withResolvers<never>();
+	const timer = setTimeout(() => {
+		timeoutReject(new Error(`fetchDynamicModels timed out after ${timeoutMs}ms`));
+	}, timeoutMs);
+	try {
+		return await Promise.race([run(), timeoutPromise]);
+	} finally {
+		clearTimeout(timer);
+	}
+}
+
 /** Provider override config (baseUrl, headers, apiKey, compat, transport) without custom models */
 interface ProviderOverride {
 	baseUrl?: string;
@@ -240,6 +265,13 @@ interface CustomModelsResult {
 }
 
 const commandValueCache = new Map<string, string>();
+// Failed `!command` resolutions (non-zero exit, empty stdout) are negative-cached
+// with a TTL instead of forever: a transient failure (locked password manager,
+// network hiccup) must not disable the key until process restart, but re-running
+// the command on every resolution would restore the execSync storm this cache
+// exists to prevent. One probe per TTL window bounds both.
+const COMMAND_FAILURE_RETRY_MS = 30_000;
+const commandFailureRetryAt = new Map<string, number>();
 
 function isCommandConfigValue(valueConfig: string | undefined): valueConfig is string {
 	return valueConfig?.startsWith("!") === true;
@@ -248,13 +280,20 @@ function isCommandConfigValue(valueConfig: string | undefined): valueConfig is s
 function resolveCommandConfig(command: string): string | undefined {
 	const cached = commandValueCache.get(command);
 	if (cached !== undefined) return cached;
+	const retryAt = commandFailureRetryAt.get(command);
+	if (retryAt !== undefined && Date.now() < retryAt) return undefined;
 	try {
 		const stdout = execSync(command, { encoding: "utf8", timeout: 10_000, windowsHide: true });
 		const trimmed = stdout.trim();
-		if (trimmed.length === 0) return undefined;
+		if (trimmed.length === 0) {
+			commandFailureRetryAt.set(command, Date.now() + COMMAND_FAILURE_RETRY_MS);
+			return undefined;
+		}
+		commandFailureRetryAt.delete(command);
 		commandValueCache.set(command, trimmed);
 		return trimmed;
 	} catch {
+		commandFailureRetryAt.set(command, Date.now() + COMMAND_FAILURE_RETRY_MS);
 		return undefined;
 	}
 }
@@ -2134,7 +2173,9 @@ export class ModelRegistry {
 				fetchDynamicModels: async () => {
 					const apiKey = await this.#peekApiKeyForProvider(providerName);
 					const resolvedKey = isAuthenticated(apiKey) ? apiKey : undefined;
-					const modelDefs = await fetcher(resolvedKey);
+					const modelDefs = await withRuntimeDynamicModelsTimeout(RUNTIME_DYNAMIC_MODEL_FETCH_TIMEOUT_MS, () =>
+						fetcher(resolvedKey),
+					);
 					const results: Model<Api>[] = [];
 					for (const modelDef of modelDefs) {
 						const overlay = buildCustomModelOverlay(

package/src/config/model-resolver.ts

@@ -864,7 +864,7 @@ function getModelRoleAlias(value: string): ModelRole | undefined {
 
 function normalizeModelPatternList(value: string | string[] | undefined): string[] {
 	if (!value) return [];
-	const patterns = Array.isArray(value) ? value : value.split(",");
+	const patterns = Array.isArray(value) ? value.flatMap(pattern => pattern.split(",")) : value.split(",");
 	return patterns.map(pattern => pattern.trim()).filter(Boolean);
 }