@kingdee-kcode/coding-agent 16.1.9

package/dist/types/tiny/dtype.d.ts

@@ -0,0 +1,85 @@
+import type { DataType } from "@huggingface/transformers";
+/** ONNX quantization / precision for local tiny models (transformers.js `dtype`). */
+export type TinyModelDtype = DataType;
+/**
+ * Validate and canonicalize a `PI_TINY_DTYPE` value. Returns `undefined` when
+ * unset/blank so callers fall back to the per-model spec dtype, and throws on an
+ * unrecognized value so a misconfiguration fails loudly instead of silently
+ * loading a different precision than requested.
+ */
+export declare function normalizeTinyModelDtype(value: string | undefined): TinyModelDtype | undefined;
+/**
+ * Resolve the `PI_TINY_DTYPE` override. `undefined` means "use the per-model spec
+ * dtype" (currently `q4` for every shipped model); a concrete value overrides the
+ * precision for whichever local tiny model loads.
+ */
+export declare function resolveTinyModelDtypeOverride(value?: string | undefined): TinyModelDtype | undefined;
+/** Sentinel `providers.tinyModelDtype` value meaning "use each model's shipped dtype". */
+export declare const TINY_MODEL_DTYPE_DEFAULT = "default";
+/** Accepted values for the `providers.tinyModelDtype` setting (validation + UI). */
+export declare const TINY_MODEL_DTYPE_SETTING_VALUES: readonly ["default", "q4", "q4f16", "q8", "fp16", "fp32", "int8", "uint8", "bnb4", "q2", "q2f16", "q1", "q1f16", "auto"];
+/** Submenu metadata for the `providers.tinyModelDtype` setting. */
+export declare const TINY_MODEL_DTYPE_SETTING_OPTIONS: readonly [{
+    readonly value: "default";
+    readonly label: "Default";
+    readonly description: "Each model's shipped dtype (currently q4)";
+}, {
+    readonly value: "q4";
+    readonly label: "q4";
+    readonly description: "4-bit weights; smallest and fastest";
+}, {
+    readonly value: "q4f16";
+    readonly label: "q4f16";
+    readonly description: "4-bit weights with fp16 activations";
+}, {
+    readonly value: "q8";
+    readonly label: "q8";
+    readonly description: "8-bit quantization";
+}, {
+    readonly value: "fp16";
+    readonly label: "fp16";
+    readonly description: "16-bit float; higher fidelity, larger";
+}, {
+    readonly value: "fp32";
+    readonly label: "fp32";
+    readonly description: "Full precision; largest and slowest";
+}, {
+    readonly value: "int8";
+    readonly label: "int8";
+    readonly description: "Signed 8-bit integer";
+}, {
+    readonly value: "uint8";
+    readonly label: "uint8";
+    readonly description: "Unsigned 8-bit integer";
+}, {
+    readonly value: "bnb4";
+    readonly label: "bnb4";
+    readonly description: "bitsandbytes 4-bit";
+}, {
+    readonly value: "q2";
+    readonly label: "q2";
+    readonly description: "2-bit weights";
+}, {
+    readonly value: "q2f16";
+    readonly label: "q2f16";
+    readonly description: "2-bit weights with fp16 activations";
+}, {
+    readonly value: "q1";
+    readonly label: "q1";
+    readonly description: "1-bit weights";
+}, {
+    readonly value: "q1f16";
+    readonly label: "q1f16";
+    readonly description: "1-bit weights with fp16 activations";
+}, {
+    readonly value: "auto";
+    readonly label: "Auto";
+    readonly description: "Let transformers.js choose per device";
+}];
+/**
+ * Map a `providers.tinyModelDtype` setting value onto a `PI_TINY_DTYPE` env value
+ * for the worker. Returns `undefined` for the default sentinel so the worker keeps
+ * each model's shipped dtype; the worker still validates the forwarded value via
+ * {@link normalizeTinyModelDtype}.
+ */
+export declare function tinyModelDtypeSettingToEnv(value: string | undefined): string | undefined;

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

@@ -0,0 +1,212 @@
+/** Default session-title model: the online pi/smol path (no local download / on-device inference). */
+export declare const ONLINE_TINY_TITLE_MODEL_KEY = "online";
+/** Local model the `tiny-models` CLI downloads when none is named. Not the session-title default — that is {@link ONLINE_TINY_TITLE_MODEL_KEY}. */
+export declare const DEFAULT_TINY_TITLE_LOCAL_MODEL_KEY = "lfm2-700m";
+export interface TinyTitleLocalModelSpec {
+    key: string;
+    repo: string;
+    dtype: "q4";
+    label: string;
+    description: string;
+    contextNote: string;
+    /** Model family emits hidden reasoning unless the chat template disables it. */
+    reasoning?: boolean;
+}
+export declare const TINY_TITLE_LOCAL_MODELS: readonly [{
+    readonly key: "lfm2-350m";
+    readonly repo: "onnx-community/LFM2-350M-ONNX";
+    readonly dtype: "q4";
+    readonly label: "LFM2 350M";
+    readonly description: "Recommended local model; best speed/quality balance, about 212 MB cached.";
+    readonly contextNote: "Best local default from the title-generation spike.";
+}, {
+    readonly key: "qwen3-0.6b";
+    readonly repo: "onnx-community/Qwen3-0.6B-ONNX";
+    readonly dtype: "q4";
+    readonly label: "Qwen3 0.6B";
+    readonly description: "Most robust local option; slower first load, about 500 MB cached.";
+    readonly contextNote: "Use when title quality matters more than local startup cost.";
+    readonly reasoning: true;
+}, {
+    readonly key: "gemma-270m";
+    readonly repo: "onnx-community/gemma-3-270m-it-ONNX";
+    readonly dtype: "q4";
+    readonly label: "Gemma 270M";
+    readonly description: "Smallest viable local option; lower quality, lowest cache footprint.";
+    readonly contextNote: "Use on constrained machines that still need local titles.";
+}, {
+    readonly key: "qwen2.5-0.5b";
+    readonly repo: "onnx-community/Qwen2.5-0.5B-Instruct";
+    readonly dtype: "q4";
+    readonly label: "Qwen2.5 0.5B";
+    readonly description: "Balanced local fallback; moderate quality and cache footprint.";
+    readonly contextNote: "Useful when Qwen3 is too heavy but Gemma quality is insufficient.";
+}, {
+    readonly key: "lfm2-700m";
+    readonly repo: "onnx-community/LFM2-700M-ONNX";
+    readonly dtype: "q4";
+    readonly label: "LFM2 700M";
+    readonly description: "Highest-quality local option; larger and slower than LFM2 350M.";
+    readonly contextNote: "Use when local title quality is preferred over startup cost.";
+}];
+export declare const TINY_TITLE_MODEL_VALUES: readonly ["online", "lfm2-350m", "qwen3-0.6b", "gemma-270m", "qwen2.5-0.5b", "lfm2-700m"];
+export type TinyTitleModelKey = (typeof TINY_TITLE_MODEL_VALUES)[number];
+export type TinyTitleLocalModelKey = (typeof TINY_TITLE_LOCAL_MODELS)[number]["key"];
+export declare const TINY_TITLE_MODEL_OPTIONS: ({
+    value: "online";
+    label: string;
+    description: string;
+} | {
+    value: "gemma-270m" | "lfm2-350m" | "lfm2-700m" | "qwen2.5-0.5b" | "qwen3-0.6b";
+    label: "Gemma 270M" | "LFM2 350M" | "LFM2 700M" | "Qwen2.5 0.5B" | "Qwen3 0.6B";
+    description: "Balanced local fallback; moderate quality and cache footprint." | "Highest-quality local option; larger and slower than LFM2 350M." | "Most robust local option; slower first load, about 500 MB cached." | "Recommended local model; best speed/quality balance, about 212 MB cached." | "Smallest viable local option; lower quality, lowest cache footprint.";
+})[];
+export declare function isTinyTitleLocalModelKey(value: string): value is TinyTitleLocalModelKey;
+export declare function getTinyTitleModelSpec(key: TinyTitleLocalModelKey): (typeof TINY_TITLE_LOCAL_MODELS)[number];
+/** Default memory model: the online path (the configured smol / remote LLM; no local download). */
+export declare const ONLINE_MEMORY_MODEL_KEY = "online";
+/** Recommended local model for memory tasks when none is named. */
+export declare const DEFAULT_MEMORY_LOCAL_MODEL_KEY = "qwen3-1.7b";
+/**
+ * Local models for Mnemopi memory tasks (fact extraction + consolidation).
+ * These are larger (1B-1.7B) than the title models: structured extraction and
+ * faithful summarization need more capacity than 3-6 word titles. All q4.
+ * Ranking/recipe rationale lives in docs/local-models.md.
+ */
+export declare const TINY_MEMORY_LOCAL_MODELS: readonly [{
+    readonly key: "qwen3-1.7b";
+    readonly repo: "onnx-community/Qwen3-1.7B-ONNX";
+    readonly dtype: "q4";
+    readonly label: "Qwen3 1.7B";
+    readonly description: "Recommended; most disciplined extraction (ignores chit-chat), good consolidation, about 1.1 GB cached.";
+    readonly contextNote: "Best single-model pick for memory from the local experiment.";
+    readonly reasoning: true;
+}, {
+    readonly key: "gemma-3-1b";
+    readonly repo: "onnx-community/gemma-3-1b-it-ONNX";
+    readonly dtype: "q4";
+    readonly label: "Gemma 3 1B";
+    readonly description: "Best consolidation/dedup; lighter footprint, but leaks small talk during extraction.";
+    readonly contextNote: "Use when consolidation quality and size matter most.";
+}, {
+    readonly key: "qwen2.5-1.5b";
+    readonly repo: "onnx-community/Qwen2.5-1.5B-Instruct";
+    readonly dtype: "q4";
+    readonly label: "Qwen2.5 1.5B";
+    readonly description: "Best extraction granularity (atomic facts); weaker consolidation.";
+    readonly contextNote: "Use when fine-grained, deduplicatable facts matter more than summaries.";
+}, {
+    readonly key: "lfm2-1.2b";
+    readonly repo: "onnx-community/LFM2-1.2B-ONNX";
+    readonly dtype: "q4";
+    readonly label: "LFM2 1.2B";
+    readonly description: "Fastest load; solid all-rounder, slightly noisier extraction labels.";
+    readonly contextNote: "Use when local startup cost is the priority.";
+}];
+export declare const TINY_MEMORY_MODEL_VALUES: readonly ["online", "qwen3-1.7b", "gemma-3-1b", "qwen2.5-1.5b", "lfm2-1.2b"];
+export type TinyMemoryModelKey = (typeof TINY_MEMORY_MODEL_VALUES)[number];
+export type TinyMemoryLocalModelKey = (typeof TINY_MEMORY_LOCAL_MODELS)[number]["key"];
+export declare const TINY_MEMORY_MODEL_OPTIONS: ({
+    value: "online";
+    label: string;
+    description: string;
+} | {
+    value: "gemma-3-1b" | "lfm2-1.2b" | "qwen2.5-1.5b" | "qwen3-1.7b";
+    label: "Gemma 3 1B" | "LFM2 1.2B" | "Qwen2.5 1.5B" | "Qwen3 1.7B";
+    description: "Best consolidation/dedup; lighter footprint, but leaks small talk during extraction." | "Best extraction granularity (atomic facts); weaker consolidation." | "Fastest load; solid all-rounder, slightly noisier extraction labels." | "Recommended; most disciplined extraction (ignores chit-chat), good consolidation, about 1.1 GB cached.";
+})[];
+export declare function isTinyMemoryLocalModelKey(value: string): value is TinyMemoryLocalModelKey;
+export declare function getTinyMemoryModelSpec(key: TinyMemoryLocalModelKey): (typeof TINY_MEMORY_LOCAL_MODELS)[number];
+/** Return whether a memory local model may emit reasoning tokens before answers. */
+export declare function isTinyMemoryReasoningModelKey(key: TinyMemoryLocalModelKey): boolean;
+/** Any local model key (title or memory), used by the shared inference worker. */
+export type TinyLocalModelKey = TinyTitleLocalModelKey | TinyMemoryLocalModelKey;
+/** Resolve a local model spec by key across both the title and memory registries. */
+export declare function getTinyLocalModelSpec(key: string): TinyTitleLocalModelSpec | undefined;
+export declare function isTinyLocalModelKey(value: string): value is TinyLocalModelKey;
+/** Combined local model registry (title + memory) for the shared tiny-models CLI. */
+export declare const TINY_LOCAL_MODELS: readonly [{
+    readonly key: "lfm2-350m";
+    readonly repo: "onnx-community/LFM2-350M-ONNX";
+    readonly dtype: "q4";
+    readonly label: "LFM2 350M";
+    readonly description: "Recommended local model; best speed/quality balance, about 212 MB cached.";
+    readonly contextNote: "Best local default from the title-generation spike.";
+}, {
+    readonly key: "qwen3-0.6b";
+    readonly repo: "onnx-community/Qwen3-0.6B-ONNX";
+    readonly dtype: "q4";
+    readonly label: "Qwen3 0.6B";
+    readonly description: "Most robust local option; slower first load, about 500 MB cached.";
+    readonly contextNote: "Use when title quality matters more than local startup cost.";
+    readonly reasoning: true;
+}, {
+    readonly key: "gemma-270m";
+    readonly repo: "onnx-community/gemma-3-270m-it-ONNX";
+    readonly dtype: "q4";
+    readonly label: "Gemma 270M";
+    readonly description: "Smallest viable local option; lower quality, lowest cache footprint.";
+    readonly contextNote: "Use on constrained machines that still need local titles.";
+}, {
+    readonly key: "qwen2.5-0.5b";
+    readonly repo: "onnx-community/Qwen2.5-0.5B-Instruct";
+    readonly dtype: "q4";
+    readonly label: "Qwen2.5 0.5B";
+    readonly description: "Balanced local fallback; moderate quality and cache footprint.";
+    readonly contextNote: "Useful when Qwen3 is too heavy but Gemma quality is insufficient.";
+}, {
+    readonly key: "lfm2-700m";
+    readonly repo: "onnx-community/LFM2-700M-ONNX";
+    readonly dtype: "q4";
+    readonly label: "LFM2 700M";
+    readonly description: "Highest-quality local option; larger and slower than LFM2 350M.";
+    readonly contextNote: "Use when local title quality is preferred over startup cost.";
+}, {
+    readonly key: "qwen3-1.7b";
+    readonly repo: "onnx-community/Qwen3-1.7B-ONNX";
+    readonly dtype: "q4";
+    readonly label: "Qwen3 1.7B";
+    readonly description: "Recommended; most disciplined extraction (ignores chit-chat), good consolidation, about 1.1 GB cached.";
+    readonly contextNote: "Best single-model pick for memory from the local experiment.";
+    readonly reasoning: true;
+}, {
+    readonly key: "gemma-3-1b";
+    readonly repo: "onnx-community/gemma-3-1b-it-ONNX";
+    readonly dtype: "q4";
+    readonly label: "Gemma 3 1B";
+    readonly description: "Best consolidation/dedup; lighter footprint, but leaks small talk during extraction.";
+    readonly contextNote: "Use when consolidation quality and size matter most.";
+}, {
+    readonly key: "qwen2.5-1.5b";
+    readonly repo: "onnx-community/Qwen2.5-1.5B-Instruct";
+    readonly dtype: "q4";
+    readonly label: "Qwen2.5 1.5B";
+    readonly description: "Best extraction granularity (atomic facts); weaker consolidation.";
+    readonly contextNote: "Use when fine-grained, deduplicatable facts matter more than summaries.";
+}, {
+    readonly key: "lfm2-1.2b";
+    readonly repo: "onnx-community/LFM2-1.2B-ONNX";
+    readonly dtype: "q4";
+    readonly label: "LFM2 1.2B";
+    readonly description: "Fastest load; solid all-rounder, slightly noisier extraction labels.";
+    readonly contextNote: "Use when local startup cost is the priority.";
+}];
+/**
+ * Difficulty-classifier model for the `auto` thinking level. Defaults to the
+ * online smol path; the local options reuse the memory-model registry because
+ * the shared worker's `complete()` only accepts memory local keys, and the
+ * 1B+ memory models classify coding difficulty far more reliably than the
+ * sub-1B title models.
+ */
+export declare const ONLINE_AUTO_THINKING_MODEL_KEY = "online";
+export declare const AUTO_THINKING_MODEL_VALUES: readonly ["online", "qwen3-1.7b", "gemma-3-1b", "qwen2.5-1.5b", "lfm2-1.2b"];
+export type AutoThinkingModelKey = TinyMemoryModelKey;
+export declare const AUTO_THINKING_MODEL_OPTIONS: ({
+    value: "online";
+    label: string;
+    description: string;
+} | {
+    value: "gemma-3-1b" | "lfm2-1.2b" | "qwen2.5-1.5b" | "qwen3-1.7b";
+    label: "Gemma 3 1B" | "LFM2 1.2B" | "Qwen2.5 1.5B" | "Qwen3 1.7B";
+    description: "Best consolidation/dedup; lighter footprint, but leaks small talk during extraction." | "Best extraction granularity (atomic facts); weaker consolidation." | "Fastest load; solid all-rounder, slightly noisier extraction labels." | "Recommended; most disciplined extraction (ignores chit-chat), good consolidation, about 1.1 GB cached.";
+})[];

package/dist/types/tiny/text.d.ts

@@ -0,0 +1,36 @@
+export declare const MAX_TITLE_INPUT_CHARS = 2000;
+export declare function truncateTitleInput(message: string): string;
+/**
+ * Strip fenced code blocks from a message before titling.
+ *
+ * Small title models latch onto literal text inside code blocks — e.g. a pasted
+ * UI mockup containing "Welcome to Claude Code v2.1.158" yields that string as
+ * the title instead of the surrounding intent. Removing fenced blocks leaves the
+ * prose that actually describes the task. Inline code (single backticks) is kept
+ * — it is short, high-signal context like `/login`.
+ *
+ * Falls back to the original message when stripping leaves too little to title
+ * (a message that is essentially just a code block).
+ */
+export declare function stripCodeBlocks(message: string): string;
+/** Prepare a raw user message for titling: drop code blocks, then bound length. */
+export declare function prepareTitleInput(message: string): string;
+export declare function formatTitleUserMessage(message: string): string;
+/**
+ * True when a first user message is too low-signal to title (greeting, ack,
+ * bare number, or empty once code/punctuation/emoji are stripped).
+ *
+ * Deterministic pre-filter: the default tiny title model (~350M local) cannot
+ * reliably follow a "respond with none" instruction and tends to hallucinate a
+ * title for trivial input, so we never ask it — the caller defers titling to
+ * the next message instead.
+ */
+export declare function isLowSignalTitleInput(message: string): boolean;
+/**
+ * Sentinel a capable title model may emit when a message carries no concrete
+ * task. Treated as "no title yet" so the caller can defer titling. Backstop for
+ * the deterministic {@link isLowSignalTitleInput} filter; kept in sync with the
+ * `none` instruction in `prompts/system/title-system.md`.
+ */
+export declare const NO_TITLE_SENTINEL = "none";
+export declare function normalizeGeneratedTitle(value: string | null | undefined): string | null;

package/dist/types/tiny/title-client.d.ts

@@ -0,0 +1,84 @@
+import type { Subprocess } from "bun";
+import type { TinyTitleProgressEvent, TinyTitleWorkerInbound, TinyTitleWorkerOutbound } from "./title-protocol";
+/**
+ * Abstraction over the tiny-model subprocess. Modelled as a worker interface
+ * so existing callers (titles, memory completions, downloads) compose the
+ * same way; 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 on Windows during shutdown
+ * (issue #1606).
+ */
+interface WorkerHandle {
+    send(message: TinyTitleWorkerInbound): void;
+    onMessage(handler: (message: TinyTitleWorkerOutbound) => void): () => void;
+    onError(handler: (error: Error) => void): () => void;
+    terminate(): Promise<void>;
+}
+export interface TinyTitleDownloadOptions {
+    signal?: AbortSignal;
+    onProgress?: (event: TinyTitleProgressEvent) => void;
+}
+/**
+ * Per-request controls for {@link TinyTitleClient.generate}.
+ *
+ * Carries the optional abort signal and title-system-prompt override used by
+ * callers that customize automatic session-title generation.
+ */
+export interface TinyTitleGenerateOptions {
+    signal?: AbortSignal;
+    systemPrompt?: string;
+}
+/**
+ * Hidden subcommand on the main CLI that boots the tiny-model worker in the
+ * spawned subprocess. Kept in sync with the dispatch in `cli.ts`.
+ */
+export declare const TINY_WORKER_ARG = "__omp_worker_tiny_inference";
+/**
+ * Decide which `PI_TINY_DEVICE` / `PI_TINY_DTYPE` vars to overlay onto the worker
+ * env. A present env var wins (left untouched); otherwise the mapped persisted
+ * setting is used. Returns only the keys to add — never the default sentinel.
+ * Pure for testability; see {@link tinyWorkerEnv} for the spawn-time glue.
+ * @internal
+ */
+export declare function tinyWorkerEnvOverlay(env: Record<string, string | undefined>, deviceSetting: string | undefined, dtypeSetting: string | undefined): Record<string, string>;
+interface SpawnedSubprocess {
+    proc: Subprocess<"ignore", "ignore", "ignore">;
+    inbound: Set<(message: TinyTitleWorkerOutbound) => void>;
+    errors: Set<(error: Error) => void>;
+    /**
+     * Flipped to `true` by {@link wrapSubprocess}'s `terminate()` right
+     * before it SIGKILLs the child so `onExit` can distinguish the
+     * expected hard-kill from a crash/OOM/external signal. Only the
+     * latter is surfaced as a worker error.
+     */
+    intentionalExit: {
+        value: boolean;
+    };
+}
+/**
+ * Spawn the tiny-model worker as a subprocess. Exported for tests and the
+ * smoke probe; production callers go through {@link spawnTinyTitleWorker}
+ * which wraps the result in a {@link WorkerHandle}.
+ */
+export declare function createTinyTitleSubprocess(): SpawnedSubprocess;
+export declare class TinyTitleClient {
+    #private;
+    constructor(spawnWorker?: () => WorkerHandle);
+    onProgress(listener: (event: TinyTitleProgressEvent) => void): () => void;
+    generate(modelKey: string, message: string, signal?: AbortSignal): Promise<string | null>;
+    generate(modelKey: string, message: string, options?: TinyTitleGenerateOptions): Promise<string | null>;
+    complete(modelKey: string, prompt: string, options?: {
+        maxTokens?: number;
+        signal?: AbortSignal;
+    }): Promise<string | null>;
+    downloadModel(modelKey: string, options?: TinyTitleDownloadOptions): Promise<boolean>;
+    terminate(): Promise<void>;
+}
+export declare const tinyTitleClient: TinyTitleClient;
+/** Alias for the shared tiny-model worker client (titles + memory completions). */
+export declare const tinyModelClient: TinyTitleClient;
+export declare function shutdownTinyTitleClient(): Promise<void>;
+export declare function smokeTestTinyTitleWorker({ timeoutMs, }?: {
+    timeoutMs?: number;
+}): Promise<void>;
+export {};

package/dist/types/tiny/title-protocol.d.ts

@@ -0,0 +1,78 @@
+import type { TinyLocalModelKey, TinyTitleLocalModelKey } from "./models";
+export type TinyTitleProgressStatus = "initiate" | "download" | "progress" | "progress_total" | "done" | "ready" | "error";
+export interface TinyTitleProgressFileState {
+    loaded: number;
+    total: number;
+}
+export interface TinyTitleProgressEvent {
+    modelKey: TinyLocalModelKey;
+    status: TinyTitleProgressStatus;
+    name?: string;
+    file?: string;
+    progress?: number;
+    loaded?: number;
+    total?: number;
+    files?: Record<string, TinyTitleProgressFileState>;
+    task?: string;
+    model?: string;
+}
+export type TinyTitleWorkerInbound = {
+    type: "ping";
+    id: string;
+} | {
+    type: "generate";
+    id: string;
+    modelKey: TinyTitleLocalModelKey;
+    message: string;
+    systemPrompt?: string;
+} | {
+    type: "complete";
+    id: string;
+    modelKey: TinyLocalModelKey;
+    prompt: string;
+    maxTokens?: number;
+} | {
+    type: "download";
+    id: string;
+    modelKey: TinyLocalModelKey;
+};
+export type TinyTitleWorkerOutbound = {
+    type: "pong";
+    id: string;
+} | {
+    type: "title";
+    id: string;
+    title: string | null;
+} | {
+    type: "completion";
+    id: string;
+    text: string | null;
+} | {
+    type: "downloaded";
+    id: string;
+} | {
+    type: "error";
+    id: string;
+    error: string;
+} | {
+    type: "progress";
+    id: string;
+    event: TinyTitleProgressEvent;
+} | {
+    type: "log";
+    level: "debug" | "warn" | "error";
+    msg: string;
+    meta?: Record<string, unknown>;
+};
+/**
+ * Wire transport between the parent (`TinyTitleClient`) and the tiny-model
+ * subprocess. The parent owns the subprocess lifecycle (graceful work, hard
+ * kill on shutdown); the protocol therefore 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 any
+ * shared address space. See `title-client.ts` for the spawn/kill glue.
+ */
+export interface TinyTitleTransport {
+    send(message: TinyTitleWorkerOutbound): void;
+    onMessage(handler: (message: TinyTitleWorkerInbound) => void): () => void;
+}

package/dist/types/tiny/worker.d.ts

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

package/dist/types/tool-discovery/mode.d.ts

@@ -0,0 +1,8 @@
+import type { Settings } from "../config/settings";
+import type { SettingValue } from "../config/settings-schema";
+export declare const TOOL_DISCOVERY_AUTO_THRESHOLD = 40;
+export declare const TOOL_DISCOVERY_SEARCH_TOOL_NAME = "search_tool_bm25";
+export type ToolDiscoveryModeSetting = SettingValue<"tools.discoveryMode">;
+export type EffectiveToolDiscoveryMode = Exclude<ToolDiscoveryModeSetting, "auto">;
+export declare function countToolsForAutoDiscovery(toolNames: Iterable<string>): number;
+export declare function resolveEffectiveToolDiscoveryMode(settings: Settings, toolCount: number): EffectiveToolDiscoveryMode;

package/dist/types/tool-discovery/tool-index.d.ts

@@ -0,0 +1,57 @@
+import type { AgentTool } from "@oh-my-pi/pi-agent-core";
+export type DiscoverableToolSource = "builtin" | "mcp" | "extension" | "custom";
+export interface DiscoverableTool {
+    name: string;
+    label: string;
+    /** Short BM25 corpus entry; falls back to description first 200 chars */
+    summary: string;
+    source: DiscoverableToolSource;
+    /** MCP only */
+    serverName?: string;
+    /** MCP only */
+    mcpToolName?: string;
+    schemaKeys: string[];
+}
+export interface DiscoverableToolServerSummary {
+    name: string;
+    toolCount: number;
+}
+export interface DiscoverableToolSummary {
+    servers: DiscoverableToolServerSummary[];
+    toolCount: number;
+}
+export interface DiscoverableToolSearchDocument {
+    tool: DiscoverableTool;
+    termFrequencies: Map<string, number>;
+    length: number;
+}
+export interface DiscoverableToolSearchIndex {
+    documents: DiscoverableToolSearchDocument[];
+    averageLength: number;
+    documentFrequencies: Map<string, number>;
+}
+export interface DiscoverableToolSearchResult {
+    tool: DiscoverableTool;
+    score: number;
+}
+export declare function isMCPToolName(name: string): boolean;
+/**
+ * Convert a raw AgentTool into a DiscoverableTool generic descriptor.
+ * source: "mcp" if name starts with "mcp__", else "builtin" (caller may override).
+ */
+export declare function getDiscoverableTool(tool: AgentTool, overrides?: {
+    source?: DiscoverableToolSource;
+    summary?: string;
+}): DiscoverableTool | null;
+/** Collect all DiscoverableTools from a tool iterable. Skips tools that return null. */
+export declare function collectDiscoverableTools(tools: Iterable<AgentTool>, options?: {
+    source?: DiscoverableToolSource;
+    summaryMap?: Map<string, string>;
+}): DiscoverableTool[];
+/** Filter discoverable tools by source */
+export declare function filterBySource(tools: DiscoverableTool[], source: DiscoverableToolSource): DiscoverableTool[];
+export declare function formatDiscoverableToolServerSummary(server: DiscoverableToolServerSummary): string;
+export declare function selectDiscoverableToolNamesByServer(tools: Iterable<DiscoverableTool>, serverNames: ReadonlySet<string>): string[];
+export declare function summarizeDiscoverableTools(tools: DiscoverableTool[]): DiscoverableToolSummary;
+export declare function buildDiscoverableToolSearchIndex(tools: Iterable<DiscoverableTool>): DiscoverableToolSearchIndex;
+export declare function searchDiscoverableTools(index: DiscoverableToolSearchIndex, query: string, limit: number): DiscoverableToolSearchResult[];

package/dist/types/tools/__tests__/json-tree.test.d.ts

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

package/dist/types/tools/approval.d.ts

@@ -0,0 +1,46 @@
+/**
+ * Tool approval resolution.
+ *
+ * Approval policy is declared by each tool. This module only knows how to:
+ * - normalize user `tools.approval.<tool>: allow | deny | prompt` overrides,
+ * - compare a tool capability tier against the active approval mode,
+ * - format the generic approval prompt body.
+ */
+import type { AgentTool, ToolTier } from "@oh-my-pi/pi-agent-core";
+export type { ToolApproval, ToolApprovalDecision, ToolTier } from "@oh-my-pi/pi-agent-core";
+export type ApprovalPolicy = "allow" | "deny" | "prompt";
+export type ApprovalMode = "always-ask" | "write" | "yolo";
+type ApprovalSubject = Pick<AgentTool, "name" | "approval" | "formatApprovalDetails">;
+export interface ResolvedApproval {
+    policy: ApprovalPolicy;
+    tier: ToolTier;
+    reason?: string;
+    override: boolean;
+}
+/**
+ * Resolve approval policy for a tool call.
+ *
+ * Resolution order:
+ *  1. Tool `approval(args)` decision, defaulting to tier "exec" when omitted.
+ *  2. User per-tool override, if set and valid.
+ *  3. Active mode tier comparison.
+ *
+ * In yolo mode, override-based tool prompts are ignored; user `tools.approval`
+ * settings remain authoritative.
+ */
+export declare function resolveApproval(tool: ApprovalSubject, args: unknown, mode: ApprovalMode, userConfig?: Record<string, unknown>): ResolvedApproval;
+/**
+ * Check if a tool call requires user approval.
+ *
+ * @throws Error if policy is 'deny'
+ * @returns Object with required flag and optional reason for the prompt
+ */
+export declare function requiresApproval(tool: ApprovalSubject, args: unknown, mode: ApprovalMode, userConfig?: Record<string, unknown>): {
+    required: boolean;
+    reason?: string;
+};
+export declare function truncateForPrompt(value: string, maxChars?: number): string;
+/**
+ * Format the approval prompt body shown to the user.
+ */
+export declare function formatApprovalPrompt(tool: ApprovalSubject, args: unknown, reason?: string): string;