@oh-my-pi/pi-coding-agent 15.5.15 → 15.7.0

package/dist/types/modes/theme/shimmer.d.ts

@@ -20,6 +20,8 @@ export interface ShimmerSegment {
     palette?: ShimmerPalette;
 }
 export declare const DEFAULT_SHIMMER_PALETTE: ShimmerPalette;
+/** Whether shimmer animations are active (any mode other than `disabled`). */
+export declare function shimmerEnabled(): boolean;
 /**
  * Apply a shimmer sweep across one or more segments, treating them as a
  * single continuous string for band positioning. Each segment can supply

package/dist/types/modes/theme/theme.d.ts

@@ -27,6 +27,17 @@ export declare class Theme {
     inverse(text: string): string;
     getFgAnsi(color: ThemeColor): string;
     getBgAnsi(color: ThemeBg): string;
+    /**
+     * Foreground ANSI for text drawn **on top of** `fillColor` used as a solid
+     * background (e.g. a powerline chip). Picks near-black or near-white by the
+     * fill's perceived luminance (Rec. 601 luma) so the label stays legible on
+     * both bright and dark fills, across light and dark themes.
+     *
+     * Reads the RGB out of the already-resolved truecolor escape; when the fill
+     * is encoded as a 256-palette index (limited terminals) the RGB is
+     * unavailable, so it falls back to the theme `text` color.
+     */
+    getContrastFgAnsi(fillColor: ThemeColor): string;
     getColorMode(): ColorMode;
     getThinkingBorderColor(level: ThinkingLevel | Effort): (str: string) => string;
     getBashModeBorderColor(): (str: string) => string;

package/dist/types/modes/types.d.ts

@@ -47,6 +47,9 @@ export type TodoPhase = {
     name: string;
     tasks: TodoItem[];
 };
+export interface InteractiveModeInitOptions {
+    suppressWelcomeIntro?: boolean;
+}
 export interface InteractiveModeContext {
     ui: TUI;
     chatContainer: Container;
@@ -111,7 +114,8 @@ export interface InteractiveModeContext {
     skillCommands: Map<string, string>;
     oauthManualInput: OAuthManualInputManager;
     todoPhases: TodoPhase[];
-    init(): Promise<void>;
+    init(options?: InteractiveModeInitOptions): Promise<void>;
+    playWelcomeIntro(): void;
     shutdown(): Promise<void>;
     checkShutdownRequested(): Promise<void>;
     setToolUIContext(uiContext: ExtensionUIContext, hasUI: boolean): void;

package/dist/types/modes/ultrathink.d.ts

@@ -4,7 +4,7 @@ export declare const ULTRATHINK_NOTICE: string;
 export declare function containsUltrathink(text: string): boolean;
 /**
  * Rainbow-highlight every standalone "ultrathink" in `text` for editor display.
- * Adds only zero-width SGR escapes — the visible width is unchanged — and returns
- * the input untouched when the keyword is absent.
+ * Sweeps red→violet (hue 0..330), stopping short of the wrap back to red so the
+ * gradient resolves smoothly regardless of casing or match length.
  */
-export declare function highlightUltrathink(text: string): string;
+export declare const highlightUltrathink: (text: string) => string;

package/dist/types/modes/utils/keybinding-matchers.d.ts

@@ -6,5 +6,10 @@
  * keybindings registered. In that case, fall back to raw Escape matching.
  */
 export declare function matchesAppInterrupt(data: string): boolean;
+/** Match the generic selector cancel keybinding. */
 export declare function matchesSelectCancel(data: string): boolean;
+/** Match the generic selector up-navigation keybinding. */
+export declare function matchesSelectUp(data: string): boolean;
+/** Match the generic selector down-navigation keybinding. */
+export declare function matchesSelectDown(data: string): boolean;
 export declare function matchesAppExternalEditor(data: string): boolean;

package/dist/types/sdk.d.ts

@@ -13,6 +13,7 @@ import { type FileSlashCommand } from "./extensibility/slash-commands";
 import type { HindsightSessionState } from "./hindsight/state";
 import { type LocalProtocolOptions } from "./internal-urls";
 import { MCPManager, type MCPToolsLoadResult } from "./mcp";
+import { type MnemosyneSessionState } from "./mnemosyne/state";
 import { AgentRegistry } from "./registry/agent-registry";
 import { AgentSession } from "./session/agent-session";
 import { AuthStorage } from "./session/auth-storage";
@@ -97,6 +98,8 @@ export interface CreateAgentSessionOptions {
     taskDepth?: number;
     /** Parent Hindsight state to alias for subagent memory tools. */
     parentHindsightSessionState?: HindsightSessionState;
+    /** Parent Mnemosyne state to alias for subagent memory tools. */
+    parentMnemosyneSessionState?: MnemosyneSessionState;
     /** Pre-allocated agent identity for IRC routing. Default: "0-Main" for top-level, parentTaskPrefix-derived for sub. */
     agentId?: string;
     /** Display name for the agent in IRC. Default: "main" or "sub". */

package/dist/types/session/agent-session.d.ts

@@ -35,6 +35,7 @@ import { type FileSlashCommand } from "../extensibility/slash-commands";
 import { GoalRuntime } from "../goals/runtime";
 import type { Goal, GoalModeState } from "../goals/state";
 import type { HindsightSessionState } from "../hindsight/state";
+import { type MnemosyneSessionState } from "../mnemosyne/state";
 import type { PlanModeState } from "../plan-mode/state";
 import { type AgentRegistry } from "../registry/agent-registry";
 import { type SecretObfuscator } from "../secrets/obfuscator";
@@ -243,6 +244,20 @@ export interface RoleModelCycleResult {
     thinkingLevel: ThinkingLevel | undefined;
     role: string;
 }
+/** A configured role resolved to a concrete model, used by role cycling and
+ *  the plan-approval model slider. */
+export interface ResolvedRoleModel {
+    role: string;
+    model: Model;
+    thinkingLevel?: ThinkingLevel;
+    explicitThinkingLevel: boolean;
+}
+/** The set of resolvable role models plus the index of the currently active
+ *  one within {@link ResolvedRoleModel.role} order. */
+export interface RoleModelCycle {
+    models: ResolvedRoleModel[];
+    currentIndex: number;
+}
 /** Session statistics for /session command */
 export interface SessionStats {
     sessionFile: string | undefined;
@@ -294,6 +309,7 @@ export declare class AgentSession {
     get providerSessionState(): Map<string, ProviderSessionState>;
     getHindsightSessionState(): HindsightSessionState | undefined;
     setHindsightSessionState(state: HindsightSessionState | undefined): HindsightSessionState | undefined;
+    getMnemosyneSessionState(): MnemosyneSessionState | undefined;
     /** TTSR manager for time-traveling stream rules */
     get ttsrManager(): TtsrManager | undefined;
     /** Whether a TTSR abort is pending (stream was aborted to inject rules) */
@@ -607,6 +623,23 @@ export declare class AgentSession {
      * @returns The new model info, or undefined if only one model available
      */
     cycleModel(direction?: "forward" | "backward"): Promise<ModelCycleResult | undefined>;
+    /**
+     * Resolve the configured role models in the given order plus the index of
+     * the currently active one. Roles that have no configured model, or whose
+     * configured model is not currently available, are skipped. The `default`
+     * role falls back to the active model when no explicit assignment exists.
+     *
+     * Returns `undefined` only when there is no current model or no available
+     * models at all; an empty `models` array is never returned (callers should
+     * still guard on `models.length`).
+     */
+    getRoleModelCycle(roleOrder: readonly string[]): RoleModelCycle | undefined;
+    /**
+     * Apply a resolved role model as the active model, persisting the choice to
+     * settings under its role. Mirrors the non-temporary branch of
+     * {@link cycleRoleModels} and is shared with the plan-approval model slider.
+     */
+    applyRoleModel(entry: ResolvedRoleModel): Promise<void>;
     /**
      * Cycle through configured role models in a fixed order.
      * Skips missing roles.

package/dist/types/system-prompt.d.ts

@@ -82,6 +82,8 @@ export interface BuildSystemPromptOptions {
     secretsEnabled?: boolean;
     /** Pre-loaded workspace tree (skips discovery if provided). May be a Promise to allow early kick-off. */
     workspaceTree?: WorkspaceTree | Promise<WorkspaceTree>;
+    /** Whether the local memory://root summary is active. */
+    memoryRootEnabled?: boolean;
 }
 /** Result of building provider-facing system prompt messages. */
 export interface BuildSystemPromptResult {

package/dist/types/task/executor.d.ts

@@ -11,6 +11,7 @@ import { type Skill } from "../extensibility/skills";
 import type { HindsightSessionState } from "../hindsight/state";
 import type { LocalProtocolOptions } from "../internal-urls";
 import type { MCPManager } from "../mcp/manager";
+import type { MnemosyneSessionState } from "../mnemosyne/state";
 import type { ArtifactManager } from "../session/artifacts";
 import type { AuthStorage } from "../session/auth-storage";
 import type { ContextFileEntry } from "../tools";
@@ -64,6 +65,7 @@ export interface ExecutorOptions {
      */
     parentArtifactManager?: ArtifactManager;
     parentHindsightSessionState?: HindsightSessionState;
+    parentMnemosyneSessionState?: MnemosyneSessionState;
     /** Parent agent's eval executor session id. Subagents reuse it so eval state is shared. */
     parentEvalSessionId?: string;
     /**

package/dist/types/task/render.d.ts

@@ -5,7 +5,11 @@ import type { TaskParams, TaskToolDetails } from "./types";
 /**
  * Render the tool call arguments.
  */
-export declare function renderCall(args: TaskParams, _options: RenderResultOptions, theme: Theme): Component;
+export declare function renderCall(args: TaskParams, options: RenderResultOptions & {
+    renderContext?: {
+        hasResult?: boolean;
+    };
+}, theme: Theme): Component;
 /**
  * Render the tool result.
  */

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

@@ -0,0 +1,78 @@
+import type { DeviceType } from "@huggingface/transformers";
+export type TinyModelDevice = DeviceType;
+export interface TinyModelDevicePreference {
+    device: TinyModelDevice;
+    raw: string | undefined;
+}
+export declare function normalizeTinyModelDevice(value: string | undefined): TinyModelDevice | undefined;
+export declare function resolveTinyModelDevicePreference(value?: string | undefined): TinyModelDevicePreference;
+export declare function tinyModelDeviceLoadOrder(preference: TinyModelDevicePreference): readonly TinyModelDevice[];
+/** Sentinel `providers.tinyModelDevice` value meaning "use the built-in platform default". */
+export declare const TINY_MODEL_DEVICE_DEFAULT = "default";
+/** Accepted values for the `providers.tinyModelDevice` setting (validation + UI). */
+export declare const TINY_MODEL_DEVICE_SETTING_VALUES: readonly ["default", "gpu", "cpu", "metal", "webgpu", "cuda", "dml", "coreml", "auto", "wasm", "webnn", "webnn-gpu", "webnn-cpu", "webnn-npu"];
+/** Submenu metadata for the `providers.tinyModelDevice` setting. */
+export declare const TINY_MODEL_DEVICE_SETTING_OPTIONS: readonly [{
+    readonly value: "default";
+    readonly label: "Default";
+    readonly description: "DirectML on Windows, CUDA on Linux x64, CPU elsewhere";
+}, {
+    readonly value: "gpu";
+    readonly label: "GPU";
+    readonly description: "Accelerated provider (WebGPU/Metal, CUDA, or DirectML)";
+}, {
+    readonly value: "cpu";
+    readonly label: "CPU";
+    readonly description: "CPU-only inference";
+}, {
+    readonly value: "metal";
+    readonly label: "Metal";
+    readonly description: "WebGPU alias for Apple GPUs";
+}, {
+    readonly value: "webgpu";
+    readonly label: "WebGPU";
+    readonly description: "WebGPU/Metal backend";
+}, {
+    readonly value: "cuda";
+    readonly label: "CUDA";
+    readonly description: "NVIDIA CUDA (Linux x64)";
+}, {
+    readonly value: "dml";
+    readonly label: "DirectML";
+    readonly description: "DirectML backend (Windows)";
+}, {
+    readonly value: "coreml";
+    readonly label: "CoreML";
+    readonly description: "Apple CoreML (opt-in; can fail to load)";
+}, {
+    readonly value: "auto";
+    readonly label: "Auto";
+    readonly description: "Let ONNX Runtime choose a provider";
+}, {
+    readonly value: "wasm";
+    readonly label: "WASM";
+    readonly description: "WebAssembly backend";
+}, {
+    readonly value: "webnn";
+    readonly label: "WebNN";
+    readonly description: "WebNN backend";
+}, {
+    readonly value: "webnn-gpu";
+    readonly label: "WebNN GPU";
+    readonly description: "WebNN GPU device";
+}, {
+    readonly value: "webnn-cpu";
+    readonly label: "WebNN CPU";
+    readonly description: "WebNN CPU device";
+}, {
+    readonly value: "webnn-npu";
+    readonly label: "WebNN NPU";
+    readonly description: "WebNN NPU device";
+}];
+/**
+ * Map a `providers.tinyModelDevice` setting value onto a `PI_TINY_DEVICE` env
+ * value for the worker. Returns `undefined` for the default sentinel so the
+ * worker keeps its built-in platform default; the worker still validates the
+ * forwarded value via {@link normalizeTinyModelDevice}.
+ */
+export declare function tinyModelDeviceSettingToEnv(value: string | undefined): string | undefined;

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 +1,185 @@
+/** 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;
+}
+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 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 Mnemosyne 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 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];
+/** 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 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 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.";
+}];

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

@@ -0,0 +1,19 @@
+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;
+export declare function normalizeGeneratedTitle(value: string | null | undefined): string | null;

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

@@ -0,0 +1,32 @@
+import type { TinyTitleProgressEvent } from "./title-protocol";
+export interface TinyTitleDownloadOptions {
+    signal?: AbortSignal;
+    onProgress?: (event: TinyTitleProgressEvent) => void;
+}
+/**
+ * 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>;
+export declare function createTinyTitleWorker(): Worker;
+export declare class TinyTitleClient {
+    #private;
+    onProgress(listener: (event: TinyTitleProgressEvent) => void): () => void;
+    generate(modelKey: string, message: string, signal?: AbortSignal): 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>;