@oh-my-pi/pi-coding-agent 15.12.3 → 15.13.0

package/dist/types/tts/models.d.ts

@@ -0,0 +1,82 @@
+import type { TinyModelDtype } from "../tiny/dtype";
+/**
+ * Voice exposed by a local TTS model. Kokoro ships a fixed catalog of named
+ * voices; a voice is just a stable id (e.g. `af_heart`) plus a display label.
+ * Selection is purely on-device — generating with a different voice needs no
+ * extra network fetch once the model weights are cached.
+ */
+export interface TtsLocalVoiceSpec {
+    id: string;
+    label: string;
+}
+/**
+ * A local (on-device, ONNX) text-to-speech model the worker can load. `repo` is
+ * the Hugging Face model id loaded through `kokoro-js`
+ * (`KokoroTTS.from_pretrained`), which runs on the same `@huggingface/transformers`
+ * + `onnxruntime` runtime as the rest of the tiny-model stack and bundles the
+ * misaki/espeak phonemizer Kokoro needs. `dtype` is the default ONNX precision
+ * (overridable via `providers.tinyModelDtype`/`PI_TINY_DTYPE`).
+ */
+export interface TtsLocalModelSpec {
+    key: string;
+    repo: string;
+    dtype: TinyModelDtype;
+    /** PCM sample rate the model emits; fallback only — the worker uses the value RawAudio reports. */
+    sampleRate: number;
+    label: string;
+    description: string;
+    /** First entry is the model's default voice. */
+    voices: readonly TtsLocalVoiceSpec[];
+}
+/**
+ * Curated Kokoro-82M voice catalog. Kokoro ships ~28 voices; we surface the
+ * higher-graded ones across American/British × female/male so the picker stays
+ * useful without listing every D/F-grade sample. `af_heart` (grade A) leads and
+ * is the default voice. Grades are Kokoro's own `overallGrade` ratings.
+ */
+export declare const KOKORO_VOICES: readonly TtsLocalVoiceSpec[];
+/** Default voice within the default model — Kokoro's flagship grade-A voice. */
+export declare const DEFAULT_TTS_VOICE = "af_heart";
+/** Default local TTS model used when `tts.localModel` is unset. */
+export declare const DEFAULT_TTS_LOCAL_MODEL_KEY = "kokoro";
+/**
+ * Local TTS model registry. Kokoro-82M is the on-device SoTA tiny TTS (tops the
+ * TTS Arena leaderboard); the `onnx-community` ONNX export runs through
+ * `kokoro-js` on the shared transformers.js/onnxruntime worker. q8 keeps the
+ * weights ~100 MB and CPU inference fast while preserving quality. One model
+ * spans every voice/accent — language selection is a voice choice, not a
+ * separate download.
+ */
+export declare const TTS_LOCAL_MODELS: readonly [{
+    readonly key: "kokoro";
+    readonly repo: "onnx-community/Kokoro-82M-v1.0-ONNX";
+    readonly dtype: "q8";
+    readonly sampleRate: 24000;
+    readonly label: "Kokoro-82M";
+    readonly description: "Kokoro-82M neural TTS — SoTA on-device quality, multi-voice, fully local";
+    readonly voices: readonly TtsLocalVoiceSpec[];
+}];
+export type TtsLocalModelKey = (typeof TTS_LOCAL_MODELS)[number]["key"];
+export declare const TTS_LOCAL_MODEL_VALUES: readonly ["kokoro"];
+export declare const TTS_LOCAL_MODEL_OPTIONS: readonly [{
+    readonly value: "kokoro";
+    readonly label: "Kokoro-82M";
+    readonly description: "Kokoro-82M neural TTS — SoTA on-device quality, multi-voice, fully local";
+}];
+/** Voice options for the `tts.localVoice` setting picker (default model's catalog). */
+export declare const TTS_LOCAL_VOICE_OPTIONS: ReadonlyArray<{
+    value: string;
+    label: string;
+}>;
+/** Accepted `tts.localVoice` values (default model's catalog) for schema validation. */
+export declare const TTS_LOCAL_VOICE_VALUES: readonly string[];
+export declare function getTtsLocalModelSpec(key: string): TtsLocalModelSpec | undefined;
+export declare function isTtsLocalModelKey(value: string): value is TtsLocalModelKey;
+/** Resolve a model key (or the default) to its Hugging Face repo id. */
+export declare function resolveTtsRepo(modelKey: string | undefined): string;
+/**
+ * Resolve a requested voice id to a concrete voice the model supports, falling
+ * back to the model's default voice (first entry) when the id is unknown or the
+ * legacy `"default"` sentinel. The returned id is always a valid Kokoro voice.
+ */
+export declare function resolveTtsVoice(modelKey: string | undefined, voice: string | undefined): string;

package/dist/types/tts/player.d.ts

@@ -0,0 +1,32 @@
+export interface PlayerCommand {
+    cmd: string;
+    args: string[];
+}
+/** Injection seam for {@link playerCommandsFor} — defaults to real PATH/tools lookups. */
+export interface PlayerLookup {
+    which?: (bin: string) => string | null;
+    ffmpeg?: () => string | null;
+}
+/**
+ * Build the ordered list of playback commands to try for `filePath` on the
+ * given platform. Pure + injectable so the selection logic is testable without
+ * spawning anything.
+ *
+ * - darwin: `afplay` (always present on macOS).
+ * - win32: PowerShell `Media.SoundPlayer.PlaySync()` (no extra deps).
+ * - linux/other POSIX: `paplay` (PulseAudio) → `aplay` (ALSA) → the bundled
+ *   static `ffmpeg` (`-f pulse` then `-f alsa`). Empty result means nothing is
+ *   available and the caller should surface an install hint.
+ */
+export declare function playerCommandsFor(platform: NodeJS.Platform, filePath: string, lookup?: PlayerLookup): PlayerCommand[];
+export interface PlayAudioOptions {
+    signal?: AbortSignal;
+}
+/**
+ * Play `filePath` through the speakers, trying each candidate command in order
+ * and returning on the first clean exit. Throws an actionable Error if no
+ * player exists or every candidate fails (with the collected stderr).
+ */
+export declare function playAudioFile(filePath: string, options?: PlayAudioOptions): Promise<void>;
+/** Best-effort temp-file cleanup used by callers after playback. */
+export declare function removeTempFile(filePath: string): Promise<void>;

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

@@ -0,0 +1,6 @@
+export declare const KOKORO_PACKAGE = "kokoro-js";
+export declare const KOKORO_VERSION = "1.2.1";
+export declare const ONNXRUNTIME_NODE_PACKAGE = "onnxruntime-node";
+export declare const ONNXRUNTIME_NODE_VERSION = "1.26.0";
+export declare function getTtsRuntimeDir(): string;
+export declare function isTtsRuntimeCached(): Promise<boolean>;

package/dist/types/tts/streaming-player.d.ts

@@ -0,0 +1,41 @@
+import { type PlayerCommand } from "./player";
+/** Output gain applied while ducked (the user is speaking over the assistant). */
+export declare const DUCK_GAIN = 0.25;
+/** Injection seam for {@link streamingPlayerCommandsFor} — defaults to real PATH/tools lookups. */
+export interface StreamingPlayerLookup {
+    which?: (bin: string) => string | null;
+    ffmpeg?: () => string | null;
+}
+/**
+ * Ordered candidate commands for a persistent raw-PCM player on `platform`: each
+ * reads 32-bit-float little-endian mono PCM at `sampleRate` from stdin (`pipe:0`)
+ * and plays it to the default output device. An empty list means no streaming
+ * backend is available and the caller should fall back to per-file playback.
+ *
+ * - darwin: none; `afplay` is file-only, so macOS uses the interruptible
+ *   per-file fallback.
+ * - linux/other POSIX: `ffmpeg` (`-f pulse` then `-f alsa`) → `paplay`/`aplay`
+ *   raw fallbacks.
+ * - win32: none (PowerShell `SoundPlayer` is file-only).
+ */
+export declare function streamingPlayerCommandsFor(platform: NodeJS.Platform, sampleRate: number, lookup?: StreamingPlayerLookup): PlayerCommand[];
+/**
+ * Single-session gapless player. Lifecycle: {@link start} once, {@link write}
+ * chunks in order, then {@link end} to drain or {@link stop} to abort. Not
+ * reusable after stop/end — create a new instance per utterance.
+ */
+export declare class StreamingAudioPlayer {
+    #private;
+    /** Pick a backend and begin draining. Idempotent; the first call's rate wins. */
+    start(sampleRate: number): void;
+    /** Queue a mono float32 PCM chunk for playback in arrival order. */
+    write(pcm: Float32Array): void;
+    /** Scale subsequent output (1 = normal, <1 = ducked). Applies within {@link LEAD_SECONDS}. */
+    setGain(gain: number): void;
+    /** Close the input; resolves once all queued audio has finished playing. */
+    end(): Promise<void>;
+    /** Stop immediately: kill the player, drop everything still queued. */
+    stop(): void;
+}
+/** Factory the vocalizer calls; a function so tests can stub it without spawning a player. */
+export declare function createStreamingPlayer(): StreamingAudioPlayer;

package/dist/types/tts/tts-client.d.ts

@@ -0,0 +1,93 @@
+import type { Subprocess } from "bun";
+import type { TtsProgressEvent, TtsWorkerInbound, TtsWorkerOutbound } from "./tts-protocol";
+/** Decoded PCM returned by a local synthesis request. */
+export interface TtsAudio {
+    pcm: Float32Array;
+    sampleRate: number;
+}
+/**
+ * Abstraction over the TTS subprocess. The runtime implementation is a Bun child
+ * process so `onnxruntime-node`'s NAPI finalizer never runs inside the main agent
+ * address space — that destructor segfaults Bun during shutdown (issue #1606).
+ */
+interface WorkerHandle {
+    send(message: TtsWorkerInbound): void;
+    onMessage(handler: (message: TtsWorkerOutbound) => void): () => void;
+    onError(handler: (error: Error) => void): () => void;
+    /** Re-reference the subprocess so a pending request keeps the parent event loop alive. */
+    ref(): void;
+    /** Drop the reference once the worker is idle so it never blocks process exit. */
+    unref(): void;
+    terminate(): Promise<void>;
+}
+export interface TtsSynthesizeOptions {
+    voice?: string;
+    signal?: AbortSignal;
+}
+export interface TtsDownloadOptions {
+    signal?: AbortSignal;
+    onProgress?: (event: TtsProgressEvent) => void;
+}
+export interface TtsStreamOptions {
+    voice?: string;
+    signal?: AbortSignal;
+}
+/** One synthesized sentence of a streaming session, in emission order. */
+export interface TtsAudioChunk {
+    index: number;
+    text: string;
+    pcm: Float32Array;
+    sampleRate: number;
+}
+/**
+ * A live streaming-synthesis session. Feed text incrementally with {@link push}
+ * and close the input with {@link end}; `chunks` yields each synthesized
+ * sentence's audio as soon as it is ready, then completes once the worker
+ * finishes draining the closed input.
+ */
+export interface TtsStreamHandle {
+    push(text: string): void;
+    end(): void;
+    chunks: AsyncIterableIterator<TtsAudioChunk>;
+}
+/**
+ * Hidden subcommand on the main CLI that boots the TTS worker in the spawned
+ * subprocess. Kept in sync with the dispatch in `cli.ts` (Main-owned).
+ */
+export declare const TTS_WORKER_ARG = "__omp_tts_worker";
+interface SpawnedSubprocess {
+    proc: Subprocess<"ignore", "ignore", "ignore">;
+    inbound: Set<(message: TtsWorkerOutbound) => void>;
+    errors: Set<(error: Error) => void>;
+    /** Flipped to `true` right before the deliberate SIGKILL so `onExit` can tell it apart from a crash. */
+    intentionalExit: {
+        value: boolean;
+    };
+}
+/**
+ * Spawn the TTS worker as a subprocess. Exported for tests and the smoke probe;
+ * production callers go through {@link spawnTtsWorker}.
+ */
+export declare function createTtsSubprocess(): SpawnedSubprocess;
+export declare class TtsClient {
+    #private;
+    constructor(spawnWorker?: () => WorkerHandle);
+    onProgress(listener: (event: TtsProgressEvent) => void): () => void;
+    synthesize(modelKey: string, text: string, options?: TtsSynthesizeOptions): Promise<TtsAudio | null>;
+    /**
+     * Open a streaming-synthesis session. Text is fed incrementally through the
+     * returned handle's `push`/`end`; audio is emitted one synthesized sentence at
+     * a time via `chunks`, so playback can begin before the full text is known.
+     * Returns an inert handle (immediately-ended `chunks`) for unknown models or
+     * an already-aborted signal, and fails the iterator if the worker cannot spawn.
+     */
+    synthesizeStream(modelKey: string, options?: TtsStreamOptions): TtsStreamHandle;
+    downloadModel(modelKey: string, options?: TtsDownloadOptions): Promise<boolean>;
+    terminate(): Promise<void>;
+}
+export declare const ttsClient: TtsClient;
+export declare function shutdownTtsClient(): Promise<void>;
+export declare function smokeTestTtsWorker({ timeoutMs, }?: {
+    timeoutMs?: number;
+}): Promise<void>;
+export {};

package/dist/types/tts/tts-protocol.d.ts

@@ -0,0 +1,95 @@
+import type { TtsLocalModelKey } from "./models";
+export type TtsProgressStatus = "initiate" | "download" | "progress" | "progress_total" | "done" | "ready" | "error";
+export interface TtsProgressFileState {
+    loaded: number;
+    total: number;
+}
+export interface TtsProgressEvent {
+    modelKey: TtsLocalModelKey;
+    status: TtsProgressStatus;
+    name?: string;
+    file?: string;
+    progress?: number;
+    loaded?: number;
+    total?: number;
+    files?: Record<string, TtsProgressFileState>;
+    task?: string;
+    model?: string;
+}
+export type TtsWorkerInbound = {
+    type: "ping";
+    id: string;
+} | {
+    type: "synthesize";
+    id: string;
+    modelKey: TtsLocalModelKey;
+    text: string;
+    voice?: string;
+} | {
+    type: "download";
+    id: string;
+    modelKey: TtsLocalModelKey;
+} | {
+    type: "stream-start";
+    id: string;
+    modelKey: TtsLocalModelKey;
+    voice?: string;
+} | {
+    type: "stream-push";
+    id: string;
+    text: string;
+} | {
+    type: "stream-end";
+    id: string;
+} | {
+    type: "stream-cancel";
+    id: string;
+};
+export type TtsWorkerOutbound = {
+    type: "pong";
+    id: string;
+} | {
+    type: "audio";
+    id: string;
+    pcm: Float32Array;
+    sampleRate: number;
+} | {
+    type: "downloaded";
+    id: string;
+} | {
+    type: "error";
+    id: string;
+    error: string;
+} | {
+    type: "progress";
+    id: string;
+    event: TtsProgressEvent;
+} | {
+    type: "log";
+    level: "debug" | "warn" | "error";
+    msg: string;
+    meta?: Record<string, unknown>;
+} | {
+    type: "audio-chunk";
+    id: string;
+    index: number;
+    text: string;
+    pcm: Float32Array;
+    sampleRate: number;
+} | {
+    type: "stream-done";
+    id: string;
+};
+/**
+ * Wire transport between the parent (`TtsClient`) and the local TTS subprocess.
+ * The parent owns the subprocess lifecycle (graceful work, hard SIGKILL on
+ * shutdown); the protocol carries no explicit close handshake — once the parent
+ * decides to terminate, it signals the OS to reap the child so
+ * `onnxruntime-node`'s NAPI finalizer never runs in the main agent address
+ * space (it segfaults Bun on shutdown — issue #1606). See `tts-client.ts` for
+ * the spawn/kill glue.
+ */
+export interface TtsTransport {
+    send(message: TtsWorkerOutbound): void;
+    onMessage(handler: (message: TtsWorkerInbound) => void): () => void;
+}

package/dist/types/tts/tts-worker.d.ts

@@ -0,0 +1,2 @@
+import type { TtsTransport } from "./tts-protocol";
+export declare function startTtsWorker(transport: TtsTransport): void;

package/dist/types/tts/vocalizer.d.ts

@@ -0,0 +1,41 @@
+export interface VocalizerPlayer {
+    start(sampleRate: number): void;
+    write(pcm: Float32Array): void;
+    setGain(gain: number): void;
+    end(): Promise<void>;
+    stop(): void;
+}
+export declare class Vocalizer {
+    #private;
+    constructor(createPlayer?: () => VocalizerPlayer);
+    /**
+     * Stream a delta of assistant text into the engine. No-op when vocalization
+     * is disabled. The engine buffers the running text and emits audio for each
+     * complete sentence; the trailing partial is flushed by {@link flush}.
+     */
+    pushDelta(text: string): void;
+    /**
+     * Close the current input stream (call at message/turn end). The engine
+     * flushes its trailing partial as a final chunk; the player keeps draining
+     * queued audio until it completes.
+     */
+    flush(): void;
+    /**
+     * Speak a complete piece of text in one shot (ask questions, yield-mode final
+     * message): stream it in and immediately close the input. No-op when disabled.
+     */
+    speak(text: string): void;
+    /**
+     * Interrupt and drop the current session, killing in-flight playback and
+     * synthesis (new turn / user message / Esc interrupt). Audio stops at once.
+     */
+    clear(): void;
+    /** Lower the volume while the user is speaking (push-to-talk), so speech doesn't drown them out. */
+    duck(): void;
+    /** Restore full volume once the user stops speaking. */
+    unduck(): void;
+    /** Resolve once the playback chain has drained (tests / shutdown). */
+    idle(): Promise<void>;
+}
+/** Process-level vocalizer shared by the event controller and the ask tool. */
+export declare const vocalizer: Vocalizer;

package/dist/types/tts/wav.d.ts

@@ -0,0 +1,8 @@
+/**
+ * Assemble a mono PCM16 WAV byte buffer from Float32 PCM samples (the shape
+ * transformers.js `RawAudio` emits: normalized [-1, 1] amplitudes plus a sample
+ * rate). No external encoder is involved — we write a canonical 44-byte RIFF/
+ * WAVE header followed by little-endian signed 16-bit samples. Samples are
+ * clamped before quantization so out-of-range float values do not wrap.
+ */
+export declare function encodeWav(samples: Float32Array, sampleRate: number): Uint8Array;

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

@@ -13,9 +13,10 @@ export declare function copyToClipboard(text: string): Promise<void>;
  * Read an image from the system clipboard.
  *
  * Returns null on Termux (no image clipboard support) or when no display
- * server is available (headless/SSH without forwarding). Under WSL the
- * Windows clipboard is reached through `powershell.exe`, since WSLg's
- * Wayland clipboard does not carry image payloads through to `arboard`.
+ * server is available (headless/SSH without forwarding). Under native Windows
+ * and WSL, the Windows clipboard is also reached through `powershell.exe`
+ * because terminal clipboard paths can leave image payloads invisible to the
+ * native bridge.
  *
  * @returns PNG payload or null when no image is available.
  */

package/dist/types/utils/image-loading.d.ts

@@ -1,7 +1,20 @@
-import type { ImageContent } from "@oh-my-pi/pi-ai";
+import type { ImageContent, Model } from "@oh-my-pi/pi-ai";
 import { type ImageResizeOptions } from "./image-resize";
 export declare const MAX_IMAGE_INPUT_BYTES: number;
 export declare const SUPPORTED_INPUT_IMAGE_MIME_TYPES: Set<string>;
+/**
+ * Ollama and its local-backend family decode image input through llama.cpp /
+ * `stb_image`, which is compiled without WebP support, so a WebP upload fails
+ * with an opaque HTTP 400. Detect those models so the resize pipeline encodes
+ * to PNG/JPEG instead — the automatic equivalent of `OMP_NO_WEBP=1`.
+ */
+export declare function modelLacksWebpSupport(model: Pick<Model, "provider" | "api"> | undefined): boolean;
+/**
+ * `true` when `model` cannot decode WebP, otherwise `undefined` so the
+ * `OMP_NO_WEBP` env fallback in {@link resizeImage} still applies. Feed straight
+ * into {@link ImageResizeOptions.excludeWebP}.
+ */
+export declare function webpExclusionForModel(model: Pick<Model, "provider" | "api"> | undefined): true | undefined;
 export interface LoadImageInputOptions {
     path: string;
     cwd: string;
@@ -9,6 +22,8 @@ export interface LoadImageInputOptions {
     maxBytes?: number;
     resolvedPath?: string;
     detectedMimeType?: string;
+    /** Force non-WebP output (e.g. for Ollama). Leave unset to honor `OMP_NO_WEBP`. */
+    excludeWebP?: boolean;
 }
 export interface LoadedImageInput {
     resolvedPath: string;
@@ -25,6 +40,8 @@ export declare class ImageInputTooLargeError extends Error {
 }
 export declare function ensureSupportedImageInput(image: ImageContent): Promise<ImageContent | null>;
 export interface NormalizeModelContextImagesOptions {
+    /** Model the images are bound for; used to derive encoder constraints (WebP exclusion for Ollama). */
+    model?: Model;
     resize?: ImageResizeOptions;
 }
 /**

package/dist/types/utils/thinking-display.d.ts

@@ -0,0 +1,17 @@
+import type { AssistantMessage } from "@oh-my-pi/pi-ai";
+type AssistantContentBlock = AssistantMessage["content"][number];
+type ThinkingBlock = Extract<AssistantContentBlock, {
+    type: "thinking";
+}>;
+/**
+ * Returns the operator-visible thinking text for a block.
+ *
+ * Some OpenAI-compatible reasoning gateways require a non-empty
+ * `reasoning_content` field on historical assistant tool-call turns even when
+ * the model did not emit any reasoning. The provider adapter uses a single dot
+ * as the wire-only placeholder those gateways accept; if that value is later
+ * replayed or echoed as a thinking block, it should not render as model thought.
+ */
+export declare function getVisibleThinkingText(block: ThinkingBlock): string;
+export declare function hasVisibleThinking(block: ThinkingBlock): boolean;
+export {};

package/dist/types/utils/tool-choice.d.ts

@@ -5,3 +5,11 @@ import type { Api, Model, ToolChoice } from "@oh-my-pi/pi-ai";
  * narrowing their request tool list before transport.
  */
 export declare function buildNamedToolChoice(toolName: string, model?: Model<Api>): ToolChoice | undefined;
+/**
+ * Whether the given tool choice can be satisfied by the active tool set for the
+ * upcoming turn. Non-named choices (`"none"`, `"required"`, etc.) do not name a
+ * specific tool and are therefore always active.
+ */
+export declare function isToolChoiceActive(toolChoice: ToolChoice | undefined, tools: readonly {
+    name: string;
+}[]): boolean;

package/dist/types/utils/tools-manager.d.ts

@@ -1,4 +1,5 @@
-export type ToolName = "sd" | "sg" | "yt-dlp" | "trafilatura";
+export declare function ffmpegAssetName(_version: string, plat: string, architecture: string): string | null;
+export type ToolName = "sd" | "sg" | "yt-dlp" | "trafilatura" | "ffmpeg";
 export declare function getToolPath(tool: ToolName): string | null;
 type EnsureToolOptions = {
     signal?: AbortSignal;

package/dist/types/utils/tools-manager.test.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/types/web/scrapers/github.d.ts

@@ -1,6 +1,6 @@
 import type { SpecialHandler } from "./types";
 interface GitHubUrl {
-    type: "blob" | "tree" | "repo" | "issue" | "issues" | "pull" | "pulls" | "discussion" | "discussions" | "actions-run" | "actions-job" | "other";
+    type: "blob" | "tree" | "repo" | "commit" | "issue" | "issues" | "pull" | "pulls" | "discussion" | "discussions" | "actions-run" | "actions-job" | "other";
     owner: string;
     repo: string;
     ref?: string;

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

@@ -6,7 +6,7 @@
  */
 import type { AgentTool, AgentToolContext, AgentToolResult, AgentToolUpdateCallback } from "@oh-my-pi/pi-agent-core";
 import type { AuthStorage } from "@oh-my-pi/pi-ai";
-import * as z from "zod/v4";
+import { z } from "zod/v4";
 import type { CustomTool } from "../../extensibility/custom-tools/types";
 import type { ToolSession } from "../../tools";
 import { type SearchRenderDetails } from "./render";

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-coding-agent",
-	"version": "15.12.3",
+	"version": "15.13.0",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -35,7 +35,7 @@
 		"check": "biome check . && bun run check:types",
 		"check:types": "tsgo -p tsconfig.json --noEmit",
 		"lint": "biome lint .",
-		"test": "bun test --parallel",
+		"test": "bun test --parallel=2",
 		"fix": "biome check --write --unsafe . && bun run format-prompts && bun run generate-docs-index",
 		"fmt": "biome format --write . && bun run format-prompts",
 		"format-prompts": "bun scripts/format-prompts.ts",
@@ -44,20 +44,20 @@
 		"bench:guard": "bun scripts/bench-guard.ts"
 	},
 	"dependencies": {
-		"@agentclientprotocol/sdk": "0.22.1",
+		"@agentclientprotocol/sdk": "0.25.0",
 		"@babel/parser": "^7.29.7",
 		"@mozilla/readability": "^0.6.0",
-		"@oh-my-pi/hashline": "15.12.3",
-		"@oh-my-pi/omp-stats": "15.12.3",
-		"@oh-my-pi/pi-agent-core": "15.12.3",
-		"@oh-my-pi/pi-ai": "15.12.3",
-		"@oh-my-pi/pi-catalog": "15.12.3",
-		"@oh-my-pi/pi-mnemopi": "15.12.3",
-		"@oh-my-pi/pi-natives": "15.12.3",
-		"@oh-my-pi/pi-tui": "15.12.3",
-		"@oh-my-pi/pi-utils": "15.12.3",
-		"@oh-my-pi/pi-wire": "15.12.3",
-		"@oh-my-pi/snapcompact": "15.12.3",
+		"@oh-my-pi/hashline": "15.13.0",
+		"@oh-my-pi/omp-stats": "15.13.0",
+		"@oh-my-pi/pi-agent-core": "15.13.0",
+		"@oh-my-pi/pi-ai": "15.13.0",
+		"@oh-my-pi/pi-catalog": "15.13.0",
+		"@oh-my-pi/pi-mnemopi": "15.13.0",
+		"@oh-my-pi/pi-natives": "15.13.0",
+		"@oh-my-pi/pi-tui": "15.13.0",
+		"@oh-my-pi/pi-utils": "15.13.0",
+		"@oh-my-pi/pi-wire": "15.13.0",
+		"@oh-my-pi/snapcompact": "15.13.0",
 		"@opentelemetry/api": "^1.9.1",
 		"@opentelemetry/context-async-hooks": "^2.7.1",
 		"@opentelemetry/exporter-trace-otlp-proto": "^0.218.0",
@@ -77,10 +77,11 @@
 		"puppeteer-core": "^25.1.0",
 		"turndown": "7.2.4",
 		"turndown-plugin-gfm": "1.0.2",
-		"zod": "4.4.3"
+		"zod": "^4"
 	},
 	"optionalDependencies": {
-		"@huggingface/transformers": "^4.2.0"
+		"@huggingface/transformers": "^4.2.0",
+		"sherpa-onnx-node": "1.13.2"
 	},
 	"devDependencies": {
 		"@types/bun": "^1.3.14"