@oh-my-pi/pi-coding-agent 17.3.3 → 17.3.5

package/dist/types/tools/read-pdf.d.ts

@@ -0,0 +1,15 @@
+import type { ToolSession } from "../sdk.js";
+import type { ScreenshotResult } from "./browser/tab-protocol.js";
+/** A legacy PDF image-member path interpreted as a page screenshot request. */
+export interface PdfImageReadTarget {
+    /** PDF path before the member delimiter. */
+    pdfPath: string;
+    /** Original member text after the delimiter. */
+    member: string;
+    /** One-indexed page inferred from names such as `p2-img0.png`; defaults to page 1. */
+    page: number;
+}
+/** Parse a former PDF image-member path as a Chromium page screenshot request. */
+export declare function splitPdfImageReadPath(readPath: string): PdfImageReadTarget | null;
+/** Render one PDF page through the browser tool's shared headless Chromium. */
+export declare function renderPdfPageScreenshot(session: ToolSession, absolutePdfPath: string, page: number, signal?: AbortSignal): Promise<ScreenshotResult>;

package/dist/types/utils/external-editor.d.ts

@@ -18,6 +18,13 @@ export interface OpenInEditorOptions {
     /** Keep the file's trailing newline instead of trimming it from the returned text. */
     trimTrailingNewline?: boolean;
 }
+/** Subprocess argv and Windows quoting mode used to launch an external editor. */
+export interface EditorSpawnCommand {
+    cmd: string[];
+    windowsVerbatimArguments: boolean;
+}
+/** Resolves shell argv without letting the host runtime re-quote the editor command. */
+export declare function resolveEditorSpawnCommand(editorCmd: string, tmpFile: string, platform?: NodeJS.Platform): EditorSpawnCommand;
 /**
  * Opens `content` in the user's external editor and returns the edited text.
  * Returns `null` if the editor exits with a non-zero code.

package/dist/types/utils/markit.d.ts

@@ -18,10 +18,9 @@ export interface MarkitConversionResult {
 }
 export interface MarkitFileConversionOptions {
     /**
-     * Directory the PDF converter writes extracted images/diagrams into. When
-     * set, each embedded image is rendered to `<id>.png` and referenced by path
-     * in the markdown; when unset, markit emits an `<!-- image: <id> ... -->`
-     * placeholder comment instead.
+     * Directory converters may use for extracted image or diagram files. Since
+     * those files are conversion side effects, conversions using this option
+     * bypass the markdown cache.
      */
     imageDir?: string;
 }

package/dist/types/vibe/runtime.d.ts

@@ -102,15 +102,15 @@ export declare class VibeSessionRegistry {
     /** Reset the global registry. Test-only. */
     static resetGlobalForTests(): void;
     /**
-     * Insert a bare worker record without the spawn/job machinery. Test-only —
-     * lets {@link aggregateVibeWorkerTokensPerSecond} be exercised against a
-     * fake roster + AgentRegistry session without driving a real turn.
+     * Insert a bare worker record without the spawn machinery. Test-only —
+     * lets focused runtime tests attach an optional synthetic in-flight job.
      */
     registerRecordForTests(record: {
         id: string;
         cli?: VibeCli;
         ownerId: string;
         state?: VibeSessionState;
+        jobId?: string;
     }): void;
     /** Override the teardown grace period for deterministic lifecycle tests. */
     setTeardownGraceForTesting(timeoutMs: number): void;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-coding-agent",
-	"version": "17.3.3",
+	"version": "17.3.5",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -41,8 +41,6 @@
 		"format-prompts": "bun scripts/format-prompts.ts",
 		"gen:tool-views": "bun --cwd=../collab-web run gen:tool-views",
 		"gen:bundle": "bun scripts/bundle-dist.ts",
-		"gen:mupdf": "bun scripts/embed-mupdf-wasm.ts --generate",
-		"gen:mupdf:reset": "bun scripts/embed-mupdf-wasm.ts --reset",
 		"gen:native": "bun --cwd=../natives run gen:native",
 		"gen:native:reset": "bun --cwd=../natives run gen:native:reset",
 		"prepack": "bun run gen:tool-views && bun run gen:bundle",
@@ -50,18 +48,18 @@
 	},
 	"dependencies": {
 		"@babel/parser": "^7.29.7",
-		"@oh-my-pi/hashline": "17.3.3",
-		"@oh-my-pi/omp-stats": "17.3.3",
-		"@oh-my-pi/omptype": "17.3.3",
-		"@oh-my-pi/pi-agent-core": "17.3.3",
-		"@oh-my-pi/pi-ai": "17.3.3",
-		"@oh-my-pi/pi-catalog": "17.3.3",
-		"@oh-my-pi/pi-mnemopi": "17.3.3",
-		"@oh-my-pi/pi-natives": "17.3.3",
-		"@oh-my-pi/pi-tui": "17.3.3",
-		"@oh-my-pi/pi-utils": "17.3.3",
-		"@oh-my-pi/pi-wire": "17.3.3",
-		"@oh-my-pi/snapcompact": "17.3.3",
+		"@oh-my-pi/hashline": "17.3.5",
+		"@oh-my-pi/omp-stats": "17.3.5",
+		"@oh-my-pi/omptype": "17.3.5",
+		"@oh-my-pi/pi-agent-core": "17.3.5",
+		"@oh-my-pi/pi-ai": "17.3.5",
+		"@oh-my-pi/pi-catalog": "17.3.5",
+		"@oh-my-pi/pi-mnemopi": "17.3.5",
+		"@oh-my-pi/pi-natives": "17.3.5",
+		"@oh-my-pi/pi-tui": "17.3.5",
+		"@oh-my-pi/pi-utils": "17.3.5",
+		"@oh-my-pi/pi-wire": "17.3.5",
+		"@oh-my-pi/snapcompact": "17.3.5",
 		"@opentelemetry/api": "^1.9.1",
 		"@opentelemetry/api-logs": "^0.220.0",
 		"@opentelemetry/context-async-hooks": "^2.9.0",
@@ -73,7 +71,6 @@
 		"@opentelemetry/sdk-metrics": "^2.9.0",
 		"@opentelemetry/sdk-trace-base": "^2.9.0",
 		"@opentelemetry/sdk-trace-node": "^2.9.0",
-		"mupdf": "^1.28.0",
 		"puppeteer-core": "25.3.0"
 	},
 	"optionalDependencies": {

package/scripts/build-binary.ts

@@ -88,7 +88,6 @@ async function main(): Promise<void> {
 			["bun", "--cwd=../natives", "run", "gen:native"],
 			crossBuild ? { ...Bun.env, TARGET_PLATFORM: crossBuild.platform, TARGET_ARCH: crossBuild.arch } : Bun.env,
 		);
-		await runCommand(["bun", "run", "gen:mupdf"]);
 		try {
 			await compileCodingAgent({
 				repoRoot,
@@ -104,7 +103,6 @@ async function main(): Promise<void> {
 				await runCommand(["codesign", "--force", "--sign", "-", outputPath]);
 			}
 		} finally {
-			await runCommand(["bun", "run", "gen:mupdf:reset"]);
 			await runCommand(["bun", "--cwd=../natives", "run", "gen:native:reset"]);
 		}
 	} finally {

package/scripts/bundle-dist.ts

@@ -15,7 +15,6 @@ const legacyHtmlExportAssetPattern = /^(?:template-[^.]+\.(?:css|html|js)|tool-v
 // `omp-legacy-pi-modules` exists only in compiled binaries via the build plugin;
 // the npm bundle never executes that `isCompiledBinary()` branch.
 const ALWAYS_EXTERNAL = [
-	"mupdf",
 	"@oh-my-pi/pi-natives",
 	"@huggingface/transformers",
 	"fastembed",

package/src/auto-thinking/classifier.ts

@@ -14,7 +14,7 @@
  * Throws on any failure (no model, no key, unparseable output, abort/timeout);
  * the caller falls back to a concrete level and continues the turn.
  */
-import { type AssistantMessage, completeSimple, Effort, type Model } from "@oh-my-pi/pi-ai";
+import { type AssistantMessage, completeSimple, Effort, type Model, retryTransientCompletion } from "@oh-my-pi/pi-ai";
 import { getSupportedEfforts } from "@oh-my-pi/pi-catalog/model-thinking";
 import { prompt } from "@oh-my-pi/pi-utils";
 
@@ -60,14 +60,24 @@ function difficultySystemPromptFor(ceiling: Effort): string {
 
 /** Local classifiers occasionally need more room for chat-template boilerplate. */
 const LOCAL_ANSWER_MAX_TOKENS = 16;
+/** On-device reasoning classifiers need room for the bucket keyword after the `<think>` preamble. */
+const LOCAL_REASONING_MAX_TOKENS = 1024;
 /**
- * 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).
+ * Online classifier budget. Sized against two independent constraints:
+ *   - Backends that ignore `disableReasoning` still emit a thinking preamble
+ *     (e.g. Qwen3 via llama.cpp catalogued `reasoning: false` but still thinking;
+ *     Anthropic via LiteLLM/Vertex, whose `openai-completions` route downgrades a
+ *     disabled request to the lowest reasoning effort instead of turning thinking
+ *     off). The classifier keyword must have room to land after that preamble
+ *     (issue #4355).
+ *   - Anthropic-dialect proxies reject `max_tokens <= thinking.budget_tokens`. The
+ *     pinned lowest effort maps to at least Anthropic's 1024-token minimum budget,
+ *     so the cap MUST comfortably exceed 1024 or every classifier call 400s with
+ *     `max_tokens must be greater than thinking.budget_tokens` (issue #8610).
+ * `maxTokens` is a hard cap — non-thinking completions still return in a handful
+ * of tokens.
  */
-const REASONING_SAFE_MAX_TOKENS = 1024;
+const ONLINE_REASONING_SAFE_MAX_TOKENS = 4096;
 
 export interface ClassifyDifficultyDeps {
 	settings: Settings;
@@ -113,21 +123,25 @@ async function classifyOnline(input: string, deps: ClassifyDifficultyDeps, ceili
 	}
 	// Resolve metadata after getApiKey so the session-sticky credential is recorded first.
 	const metadata = deps.metadataResolver?.(model.provider);
-	const maxTokens = REASONING_SAFE_MAX_TOKENS;
-
-	const response = await completeSimple(
-		model,
-		{
-			systemPrompt: [difficultySystemPromptFor(ceiling)],
-			messages: [{ role: "user", content: input, timestamp: Date.now() }],
-		},
-		{
-			apiKey: deps.registry.resolver(model, deps.sessionId),
-			maxTokens,
-			disableReasoning: true,
-			metadata,
-			signal: deps.signal,
-		},
+	const maxTokens = ONLINE_REASONING_SAFE_MAX_TOKENS;
+
+	const response = await retryTransientCompletion(
+		() =>
+			completeSimple(
+				model,
+				{
+					systemPrompt: [difficultySystemPromptFor(ceiling)],
+					messages: [{ role: "user", content: input, timestamp: Date.now() }],
+				},
+				{
+					apiKey: deps.registry.resolver(model, deps.sessionId),
+					maxTokens,
+					disableReasoning: true,
+					metadata,
+					signal: deps.signal,
+				},
+			),
+		{ signal: deps.signal },
 	);
 
 	if (response.stopReason === "error") {
@@ -147,7 +161,7 @@ async function classifyLocal(input: string, modelKey: string, deps: ClassifyDiff
 		throw new Error(`auto-thinking: unsupported local classifier model: ${modelKey}`);
 	}
 	const maxTokens = isTinyMemoryReasoningModelKey(modelKey)
-		? Math.max(LOCAL_ANSWER_MAX_TOKENS, REASONING_SAFE_MAX_TOKENS)
+		? Math.max(LOCAL_ANSWER_MAX_TOKENS, LOCAL_REASONING_MAX_TOKENS)
 		: LOCAL_ANSWER_MAX_TOKENS;
 	const builtPrompt = prompt.render(difficultyLocalPrompt, { prompt: input });
 	const text = await tinyModelClient.complete(modelKey, builtPrompt, {

package/src/cli/models-cli.ts

@@ -304,7 +304,7 @@ export async function runModelsListing(options: RunModelsListingOptions): Promis
 		cwd,
 		eventBus,
 		disableExtensionDiscovery ? undefined : disabledExtensionIds,
-		{ ambient: !disableExtensionDiscovery },
+		{ ambient: !disableExtensionDiscovery, includeAmbientHooks: false },
 	);
 	const extensionRunner =
 		extensionsResult.extensions.length > 0

package/src/cli/read-cli.ts

@@ -10,6 +10,7 @@ import chalk from "@oh-my-pi/pi-utils/chalk";
 import { Settings } from "../config/settings";
 import { extractUriScheme } from "../internal-urls/parse";
 import { InternalUrlRouter } from "../internal-urls/router";
+import { closeDaemonClients } from "../launch/client";
 import { discoverAndLoadMCPTools } from "../mcp/loader";
 import { MCPManager } from "../mcp/manager";
 import { discoverAuthStorage } from "../session/auth-broker-config";
@@ -93,6 +94,7 @@ export async function runReadCommand(cmd: ReadCommandArgs): Promise<void> {
 			if (MCPManager.instance() === mcpManager) MCPManager.setInstance(undefined);
 		}
 		authStorage?.close();
+		await closeDaemonClients();
 	}
 
 	if (failed) process.exit(1);

package/src/commit/analysis/conventional.ts

@@ -1,6 +1,6 @@
 import type { ThinkingLevel } from "@oh-my-pi/pi-agent-core";
 import type { Api, ApiKey, Model } from "@oh-my-pi/pi-ai";
-import { completeSimple } from "@oh-my-pi/pi-ai";
+import { completeSimple, retryTransientCompletion } from "@oh-my-pi/pi-ai";
 import { prompt } from "@oh-my-pi/pi-utils";
 import analysisSystemPrompt from "../../commit/prompts/analysis-system.md" with { type: "text" };
 import analysisUserPrompt from "../../commit/prompts/analysis-user.md" with { type: "text" };
@@ -50,15 +50,21 @@ export async function generateConventionalAnalysis({
 		diff,
 	});
 
-	const response = await completeSimple(
-		model,
-		{
-			systemPrompt: [prompt.render(analysisSystemPrompt)],
-			messages: [{ role: "user", content: userContent, timestamp: Date.now() }],
-			tools: [ConventionalAnalysisTool],
-		},
-		{ apiKey, maxTokens: 2400, reasoning: toReasoningEffort(thinkingLevel) },
+	const response = await retryTransientCompletion(() =>
+		completeSimple(
+			model,
+			{
+				systemPrompt: [prompt.render(analysisSystemPrompt)],
+				messages: [{ role: "user", content: userContent, timestamp: Date.now() }],
+				tools: [ConventionalAnalysisTool],
+			},
+			{ apiKey, maxTokens: 2400, reasoning: toReasoningEffort(thinkingLevel) },
+		),
 	);
 
+	if (response.stopReason === "error") {
+		throw new Error(response.errorMessage ?? "provider error");
+	}
+
 	return parseConventionalAnalysisResponse(response, ConventionalAnalysisTool);
 }

package/src/commit/analysis/summary.ts

@@ -1,7 +1,7 @@
 import { type } from "@oh-my-pi/omptype";
 import type { ThinkingLevel } from "@oh-my-pi/pi-agent-core";
 import type { Api, ApiKey, AssistantMessage, Model } from "@oh-my-pi/pi-ai";
-import { completeSimple, validateToolCall } from "@oh-my-pi/pi-ai";
+import { completeSimple, retryTransientCompletion, validateToolCall } from "@oh-my-pi/pi-ai";
 import { prompt } from "@oh-my-pi/pi-utils";
 import summarySystemPrompt from "../../commit/prompts/summary-system.md" with { type: "text" };
 import summaryUserPrompt from "../../commit/prompts/summary-user.md" with { type: "text" };
@@ -52,16 +52,22 @@ export async function generateSummary({
 		stat,
 	});
 
-	const response = await completeSimple(
-		model,
-		{
-			systemPrompt: [systemPrompt],
-			messages: [{ role: "user", content: userPrompt, timestamp: Date.now() }],
-			tools: [SummaryTool],
-		},
-		{ apiKey, maxTokens: 200, reasoning: toReasoningEffort(thinkingLevel) },
+	const response = await retryTransientCompletion(() =>
+		completeSimple(
+			model,
+			{
+				systemPrompt: [systemPrompt],
+				messages: [{ role: "user", content: userPrompt, timestamp: Date.now() }],
+				tools: [SummaryTool],
+			},
+			{ apiKey, maxTokens: 200, reasoning: toReasoningEffort(thinkingLevel) },
+		),
 	);
 
+	if (response.stopReason === "error") {
+		throw new Error(response.errorMessage ?? "provider error");
+	}
+
 	return parseSummaryFromResponse(response, commitType, scope);
 }
 

package/src/commit/changelog/generate.ts

@@ -1,7 +1,7 @@
 import { type } from "@oh-my-pi/omptype";
 import type { ThinkingLevel } from "@oh-my-pi/pi-agent-core";
 import type { Api, ApiKey, AssistantMessage, Model } from "@oh-my-pi/pi-ai";
-import { completeSimple, validateToolCall } from "@oh-my-pi/pi-ai";
+import { completeSimple, retryTransientCompletion, validateToolCall } from "@oh-my-pi/pi-ai";
 import { prompt } from "@oh-my-pi/pi-utils";
 import changelogSystemPrompt from "../../commit/prompts/changelog-system.md" with { type: "text" };
 import changelogUserPrompt from "../../commit/prompts/changelog-user.md" with { type: "text" };
@@ -55,16 +55,22 @@ export async function generateChangelogEntries({
 		stat,
 		diff,
 	});
-	const response = await completeSimple(
-		model,
-		{
-			systemPrompt: [prompt.render(changelogSystemPrompt)],
-			messages: [{ role: "user", content: userContent, timestamp: Date.now() }],
-			tools: [changelogTool],
-		},
-		{ apiKey, maxTokens: 1200, reasoning: toReasoningEffort(thinkingLevel) },
+	const response = await retryTransientCompletion(() =>
+		completeSimple(
+			model,
+			{
+				systemPrompt: [prompt.render(changelogSystemPrompt)],
+				messages: [{ role: "user", content: userContent, timestamp: Date.now() }],
+				tools: [changelogTool],
+			},
+			{ apiKey, maxTokens: 1200, reasoning: toReasoningEffort(thinkingLevel) },
+		),
 	);
 
+	if (response.stopReason === "error") {
+		throw new Error(response.errorMessage ?? "provider error");
+	}
+
 	const parsed = parseChangelogResponse(response);
 	return { entries: dedupeEntries(parsed.entries) };
 }

package/src/commit/map-reduce/map-phase.ts

@@ -1,6 +1,6 @@
 import type { ThinkingLevel } from "@oh-my-pi/pi-agent-core";
 import type { Api, ApiKey, AssistantMessage, Message, Model } from "@oh-my-pi/pi-ai";
-import { completeSimple } from "@oh-my-pi/pi-ai";
+import { completeSimple, retryTransientCompletion } from "@oh-my-pi/pi-ai";
 import { prompt } from "@oh-my-pi/pi-utils";
 import fileObserverSystemPrompt from "../../commit/prompts/file-observer-system.md" with { type: "text" };
 import fileObserverUserPrompt from "../../commit/prompts/file-observer-user.md" with { type: "text" };
@@ -66,7 +66,7 @@ export async function runMapPhase({
 			messages: [{ role: "user", content: userContent, timestamp: Date.now() }] as Message[],
 		};
 
-		const response = await withRetry(
+		const response = await retryTransientCompletion(
 			() =>
 				completeSimple(model, request, {
 					apiKey,
@@ -74,8 +74,7 @@ export async function runMapPhase({
 					reasoning: toReasoningEffort(thinkingLevel),
 					signal: AbortSignal.timeout(timeoutMs),
 				}),
-			maxRetries,
-			retryBackoffMs,
+			{ maxAttempts: maxRetries, baseDelayMs: retryBackoffMs },
 		);
 
 		const observations = parseObservations(response);
@@ -176,18 +175,3 @@ async function runWithConcurrency<T, R>(
 	await Promise.all(runners);
 	return results;
 }
-
-async function withRetry<T>(fn: () => Promise<T>, attempts: number, backoffMs: number): Promise<T> {
-	let lastError: unknown;
-	for (let attempt = 0; attempt < attempts; attempt += 1) {
-		try {
-			return await fn();
-		} catch (error) {
-			lastError = error;
-			if (attempt < attempts - 1) {
-				await Bun.sleep(backoffMs * (attempt + 1));
-			}
-		}
-	}
-	throw lastError;
-}

package/src/commit/map-reduce/reduce-phase.ts

@@ -1,6 +1,6 @@
 import type { ThinkingLevel } from "@oh-my-pi/pi-agent-core";
 import type { Api, ApiKey, Model } from "@oh-my-pi/pi-ai";
-import { completeSimple } from "@oh-my-pi/pi-ai";
+import { completeSimple, retryTransientCompletion } from "@oh-my-pi/pi-ai";
 import { prompt } from "@oh-my-pi/pi-utils";
 import reduceSystemPrompt from "../../commit/prompts/reduce-system.md" with { type: "text" };
 import reduceUserPrompt from "../../commit/prompts/reduce-user.md" with { type: "text" };
@@ -35,15 +35,21 @@ export async function runReducePhase({
 		stat,
 		scope_candidates: scopeCandidates,
 	});
-	const response = await completeSimple(
-		model,
-		{
-			systemPrompt: [prompt.render(reduceSystemPrompt)],
-			messages: [{ role: "user", content: userContent, timestamp: Date.now() }],
-			tools: [ReduceTool],
-		},
-		{ apiKey, maxTokens: 2400, reasoning: toReasoningEffort(thinkingLevel) },
+	const response = await retryTransientCompletion(() =>
+		completeSimple(
+			model,
+			{
+				systemPrompt: [prompt.render(reduceSystemPrompt)],
+				messages: [{ role: "user", content: userContent, timestamp: Date.now() }],
+				tools: [ReduceTool],
+			},
+			{ apiKey, maxTokens: 2400, reasoning: toReasoningEffort(thinkingLevel) },
+		),
 	);
 
+	if (response.stopReason === "error") {
+		throw new Error(response.errorMessage ?? "provider error");
+	}
+
 	return parseConventionalAnalysisResponse(response, ReduceTool);
 }

package/src/config/model-discovery.ts

@@ -645,7 +645,7 @@ export async function discoverLlamaCppModels(
 					name: id,
 					api: providerConfig.api,
 					provider: providerConfig.provider,
-					baseUrl,
+					baseUrl: ensureLlamaCppV1BaseUrl(baseUrl),
 					reasoning: false,
 					input: item.input ?? serverMetadata?.input ?? ["text"],
 					imageInputDecoder: "stb",
@@ -1018,7 +1018,7 @@ export async function discoverProxyModels(
 	return discovered;
 }
 
-function normalizeLlamaCppBaseUrl(baseUrl?: string): string {
+export function normalizeLlamaCppBaseUrl(baseUrl?: string): string {
 	const defaultBaseUrl = "http://127.0.0.1:8080";
 	const raw = baseUrl || defaultBaseUrl;
 	try {
@@ -1033,7 +1033,7 @@ function normalizeLlamaCppBaseUrl(baseUrl?: string): string {
 // ensureLlamaCppV1BaseUrl appends the OpenAI-compatible `/v1` prefix a
 // chat-completions request needs; native discovery keeps the bare root, which
 // serves `/models` and `/props` but not `/chat/completions`.
-function ensureLlamaCppV1BaseUrl(baseUrl: string): string {
+export function ensureLlamaCppV1BaseUrl(baseUrl: string): string {
 	return baseUrl.endsWith("/v1") ? baseUrl : `${baseUrl}/v1`;
 }
 

package/src/config/model-registry.ts

@@ -61,9 +61,11 @@ import {
 	type DiscoveryProviderConfig,
 	discoverLlamaCppModelRuntimeMetadata,
 	discoverModelsByProviderType,
+	ensureLlamaCppV1BaseUrl,
 	getImplicitOllamaBaseUrl,
 	getOllamaContextLengthOverride,
 	normalizeLiteLLMDiscoveryBaseUrl,
+	normalizeLlamaCppBaseUrl,
 } from "./model-discovery";
 import {
 	AUTHORITATIVE_RUNTIME_CATALOG_PROVIDERS,
@@ -577,7 +579,7 @@ export class ModelRegistry {
 		const withConfigModels = this.#mergeCustomModels(resolvedDefaults, select(this.#customModelOverlays));
 		const combined = this.#mergeCustomModels(withConfigModels, select(this.#runtimeModelOverlays));
 		const withModelOverrides = this.#applyModelOverrides(collapseBuiltModelVariants(combined), this.#modelOverrides);
-		return this.#applyLlamaCppQwenThinkingToModels(this.#applyRuntimeProviderOverrides(withModelOverrides));
+		return this.#applyLlamaCppModelFixups(this.#applyRuntimeProviderOverrides(withModelOverrides));
 	}
 
 	#composeStaticModels(providerFilter?: ReadonlySet<string>): Model<Api>[] {
@@ -1064,9 +1066,7 @@ export class ModelRegistry {
 		const withConfigModels = this.#mergeCustomModels(resolved, this.#customModelOverlays);
 		const combined = this.#mergeCustomModels(withConfigModels, this.#runtimeModelOverlays);
 		const withModelOverrides = this.#applyModelOverrides(collapseBuiltModelVariants(combined), this.#modelOverrides);
-		this.#unprojectedModels = this.#applyLlamaCppQwenThinkingToModels(
-			this.#applyRuntimeProviderOverrides(withModelOverrides),
-		);
+		this.#unprojectedModels = this.#applyLlamaCppModelFixups(this.#applyRuntimeProviderOverrides(withModelOverrides));
 		this.#models = this.#applyRuntimeModelModifiers(this.#unprojectedModels);
 	}
 
@@ -1446,20 +1446,28 @@ export class ModelRegistry {
 		});
 	}
 
-	// #applyLlamaCppQwenThinkingToModels re-runs applyLlamaCppQwenThinking as the
-	// outermost transform for llama.cpp-provider models, after discovery merges,
-	// cache fallbacks, and provider/transport overrides have run. It is
-	// idempotent, so it restores the routed Qwen model's chat-completions api,
-	// `/v1` runtime base URL, and disable dialect even when a configured `baseUrl`
-	// override (which wins in mergeDiscoveredModel) or a fallback to a pre-fix
-	// cached row would otherwise leave the old spec in place.
-	#applyLlamaCppQwenThinkingToModels(models: Model<Api>[]): Model<Api>[] {
+	// #applyLlamaCppModelFixups is the outermost transform for llama.cpp-provider
+	// models, after discovery merges, cache fallbacks, and provider/transport
+	// overrides have run. It applies Qwen-specific fixes (api, reasoning, compat)
+	// and ensures all non-transport models have the `/v1` prefix in their baseUrl,
+	// even when a configured override or stale cache row would strip it.
+	#applyLlamaCppModelFixups(models: Model<Api>[]): Model<Api>[] {
 		const llamaCppProviders = new Set<string>();
 		for (const provider of this.#discoverableProviders) {
 			if (provider.discovery.type === "llama.cpp") llamaCppProviders.add(provider.provider);
 		}
 		if (llamaCppProviders.size === 0) return models;
-		return models.map(model => (llamaCppProviders.has(model.provider) ? applyLlamaCppQwenThinking(model) : model));
+		return models.map(model => {
+			if (!llamaCppProviders.has(model.provider)) return model;
+			const withFixups = applyLlamaCppQwenThinking(model);
+			if (!withFixups.transport && !withFixups.baseUrl.endsWith("/v1")) {
+				return buildModel({
+					...withFixups,
+					baseUrl: ensureLlamaCppV1BaseUrl(normalizeLlamaCppBaseUrl(withFixups.baseUrl)),
+				});
+			}
+			return withFixups;
+		});
 	}
 
 	#mergeProviderOverride(baseOverride: ProviderOverride | undefined, override: ProviderOverride): ProviderOverride {
@@ -1667,17 +1675,22 @@ export class ModelRegistry {
 	 *
 	 * Side-effect-free and synchronous: a command-backed key (`!cmd`) counts as
 	 * configured by its presence alone — the program is NOT executed — and OAuth
-	 * tokens are NOT refreshed (`authStorage.hasAuth`). This is what keeps the
+	 * tokens are NOT refreshed (`authStorage.hasResolvableAuth`). This is what keeps the
 	 * model-switch pre-flight off the event loop's hot path; the real key
 	 * (command execution + OAuth refresh) is resolved lazily per request via
 	 * {@link ModelRegistry.resolver}.
+	 *
+	 * Cross-provider env aliases count here (`xai-oauth` can borrow `XAI_API_KEY`)
+	 * so an explicit `xai-oauth/…` selector does not fail with "No API key".
+	 * Default-model availability still uses {@link AuthStorage.hasAuth}, which
+	 * ignores that alias so SuperGrok is not auto-selected from a paid key.
 	 */
 	hasConfiguredAuth(model: Model<Api>): boolean {
 		const keyConfig = this.#customProviderApiKeys.get(model.provider);
 		return (
 			isCommandConfigValue(keyConfig) ||
 			this.#keylessProviders.has(model.provider) ||
-			this.authStorage.hasAuth(model.provider)
+			this.authStorage.hasResolvableAuth(model.provider)
 		);
 	}
 
@@ -2097,7 +2110,7 @@ export class ModelRegistry {
 				transportOverride,
 			);
 			this.#runtimeProviderOverrides.set(providerName, nextRuntimeOverride);
-			this.#unprojectedModels = this.#applyLlamaCppQwenThinkingToModels(
+			this.#unprojectedModels = this.#applyLlamaCppModelFixups(
 				this.#unprojectedModels.map(model => {
 					if (model.provider !== providerName) return model;
 					return this.#applyProviderTransportOverrideToModel(model, transportOverride);

package/src/config/settings-schema.ts

@@ -152,6 +152,7 @@ export const TAB_GROUPS: Record<SettingTab, readonly string[]> = {
 		"Output Limits",
 		"Execution",
 		"Discovery & MCP",
+		"Extensions",
 		"Developer",
 	],
 	tasks: ["Modes", "Subagents", "Isolation", "Commands & Skills"],
@@ -5494,6 +5495,18 @@ export const SETTINGS_SCHEMA = {
 
 	"commit.changelogMaxDiffChars": { type: "number", default: 120000 },
 
+	"extensionHandlers.toolCallTimeoutMs": {
+		type: "number",
+		default: 30_000,
+		ui: {
+			tab: "tools",
+			group: "Extensions",
+			label: "Tool Call Handler Timeout (ms)",
+			description:
+				"Positive finite active-work timeout for extension tool_call handlers; invalid values use 30000ms, and time awaiting OMP-owned dialogs does not count",
+		},
+	},
+
 	"dev.autoqa": {
 		type: "boolean",
 		default: true,