@oh-my-pi/pi-coding-agent 15.7.1 → 15.7.3

package/dist/types/sdk.d.ts

@@ -13,12 +13,13 @@ 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 { type MnemopiSessionState } from "./mnemopi/state";
 import { AgentRegistry } from "./registry/agent-registry";
 import { AgentSession } from "./session/agent-session";
 import { AuthStorage } from "./session/auth-storage";
 import { SessionManager } from "./session/session-manager";
 import { type BuildSystemPromptResult } from "./system-prompt";
+import { type ConfiguredThinkingLevel } from "./thinking";
 import { BashTool, BUILTIN_TOOLS, createTools, EditTool, EvalTool, FindTool, HIDDEN_TOOLS, type LspStartupServerInfo, loadSshTool, ReadTool, ResolveTool, SearchTool, type Tool, type ToolSession, WebSearchTool, WriteTool } from "./tools";
 import { EventBus } from "./utils/event-bus";
 import { type WorkspaceTree } from "./workspace-tree";
@@ -39,7 +40,7 @@ export interface CreateAgentSessionOptions {
      * Used when model lookup is deferred because extension-provided models aren't registered yet. */
     modelPattern?: string;
     /** Thinking selector. Default: from settings, else unset */
-    thinkingLevel?: ThinkingLevel;
+    thinkingLevel?: ConfiguredThinkingLevel;
     /** Models available for cycling (Ctrl+P in interactive mode) */
     scopedModels?: Array<{
         model: Model;
@@ -98,8 +99,8 @@ 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;
+    /** Parent Mnemopi state to alias for subagent memory tools. */
+    parentMnemopiSessionState?: MnemopiSessionState;
     /** 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". */
@@ -175,6 +176,16 @@ export declare function discoverAuthStorage(agentDir?: string): Promise<AuthStor
  * Discover extensions from cwd.
  */
 export declare function discoverExtensions(cwd?: string): Promise<LoadExtensionsResult>;
+/**
+ * Load the discovered/configured extensions for a session — everything {@link
+ * createAgentSession} would load except the inline factory extensions it appends
+ * itself. Extracted so the CLI can resolve extension-registered flags (and thus
+ * classify `@file` arguments extension-aware) *before* a session — and its
+ * terminal breadcrumb — is created, then hand the result back through
+ * {@link CreateAgentSessionOptions.preloadedExtensions} so the work is not
+ * repeated. Keep this the single source of the discovery branch logic.
+ */
+export declare function loadSessionExtensions(options: Pick<CreateAgentSessionOptions, "disableExtensionDiscovery" | "additionalExtensionPaths">, cwd: string, settings: Settings, eventBus: EventBus): Promise<LoadExtensionsResult>;
 /**
  * Discover skills from cwd and agentDir.
  */

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

@@ -14,8 +14,9 @@
  */
 import type { InMemorySnapshotStore } from "@oh-my-pi/hashline";
 import { type Agent, type AgentEvent, type AgentMessage, type AgentState, type AgentTool, ThinkingLevel } from "@oh-my-pi/pi-agent-core";
-import { type CompactionResult } from "@oh-my-pi/pi-agent-core/compaction";
-import type { AssistantMessage, Effort, ImageContent, Message, MessageAttribution, Model, ProviderSessionState, ServiceTier, SimpleStreamOptions, TextContent, ToolChoice, UsageReport } from "@oh-my-pi/pi-ai";
+import { type CompactionResult, type ShakeConfig } from "@oh-my-pi/pi-agent-core/compaction";
+import type { AssistantMessage, ImageContent, Message, MessageAttribution, Model, ProviderSessionState, ServiceTier, SimpleStreamOptions, TextContent, ToolChoice, UsageReport } from "@oh-my-pi/pi-ai";
+import { Effort } from "@oh-my-pi/pi-ai";
 import { type AsyncJob, type AsyncJobDeliveryState, AsyncJobManager } from "../async";
 import type { Rule } from "../capability/rule";
 import { type ModelRegistry } from "../config/model-registry";
@@ -35,26 +36,28 @@ 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 MnemopiSessionState } from "../mnemopi/state";
 import type { PlanModeState } from "../plan-mode/state";
 import { type AgentRegistry } from "../registry/agent-registry";
 import { type SecretObfuscator } from "../secrets/obfuscator";
+import { type ConfiguredThinkingLevel } from "../thinking";
 import { type DiscoverableTool, type DiscoverableToolSearchIndex } from "../tool-discovery/tool-index";
 import type { CheckpointState } from "../tools/checkpoint";
 import { type TodoItem, type TodoPhase } from "../tools/todo-write";
 import type { ClientBridge } from "./client-bridge";
 import { type CustomMessage } from "./messages";
 import type { BranchSummaryEntry, NewSessionOptions, SessionContext, SessionManager } from "./session-manager";
+import type { ShakeMode, ShakeResult } from "./shake-types";
 import { ToolChoiceQueue } from "./tool-choice-queue";
 import { YieldQueue } from "./yield-queue";
 /** Session-specific events that extend the core AgentEvent */
 export type AgentSessionEvent = AgentEvent | {
     type: "auto_compaction_start";
     reason: "threshold" | "overflow" | "idle" | "incomplete";
-    action: "context-full" | "handoff";
+    action: "context-full" | "handoff" | "shake" | "shake-summary";
 } | {
     type: "auto_compaction_end";
-    action: "context-full" | "handoff";
+    action: "context-full" | "handoff" | "shake" | "shake-summary";
     result: CompactionResult | undefined;
     aborted: boolean;
     willRetry: boolean;
@@ -102,6 +105,10 @@ export type AgentSessionEvent = AgentEvent | {
 } | {
     type: "thinking_level_changed";
     thinkingLevel: ThinkingLevel | undefined;
+    /** The user-configured selector when it differs from the effective level (e.g. `auto`). */
+    configured?: ConfiguredThinkingLevel;
+    /** The level `auto` resolved to this turn, once classified. */
+    resolved?: Effort;
 } | {
     type: "goal_updated";
     goal: Goal | null;
@@ -115,6 +122,7 @@ export interface AsyncJobSnapshot {
     recent: AsyncJobSnapshotItem[];
     delivery: AsyncJobDeliveryState;
 }
+export type { ShakeMode, ShakeResult };
 export interface AgentSessionConfig {
     agent: Agent;
     sessionManager: SessionManager;
@@ -125,7 +133,7 @@ export interface AgentSessionConfig {
         thinkingLevel?: ThinkingLevel;
     }>;
     /** Initial session thinking selector. */
-    thinkingLevel?: ThinkingLevel;
+    thinkingLevel?: ConfiguredThinkingLevel;
     /** Prompt templates for expansion */
     promptTemplates?: PromptTemplate[];
     /** File-based slash commands for expansion */
@@ -309,7 +317,7 @@ export declare class AgentSession {
     get providerSessionState(): Map<string, ProviderSessionState>;
     getHindsightSessionState(): HindsightSessionState | undefined;
     setHindsightSessionState(state: HindsightSessionState | undefined): HindsightSessionState | undefined;
-    getMnemosyneSessionState(): MnemosyneSessionState | undefined;
+    getMnemopiSessionState(): MnemopiSessionState | undefined;
     /** TTSR manager for time-traveling stream rules */
     get ttsrManager(): TtsrManager | undefined;
     /** Whether a TTSR abort is pending (stream was aborted to inject rules) */
@@ -364,8 +372,14 @@ export declare class AgentSession {
     get state(): AgentState;
     /** Current model (may be undefined if not yet selected) */
     get model(): Model | undefined;
-    /** Current thinking level */
+    /** Effective thinking level applied to the agent (the resolved level when `auto`). */
     get thinkingLevel(): ThinkingLevel | undefined;
+    /** The selector the user configured: `auto` when auto mode is active, else the effective level. */
+    configuredThinkingLevel(): ConfiguredThinkingLevel | undefined;
+    /** True when `auto` thinking mode is active. */
+    get isAutoThinking(): boolean;
+    /** The level `auto` resolved to for the current turn (undefined until classified). */
+    autoResolvedThinkingLevel(): Effort | undefined;
     get serviceTier(): ServiceTier | undefined;
     /** Whether agent is currently streaming a response */
     get isStreaming(): boolean;
@@ -603,12 +617,13 @@ export declare class AgentSession {
     fork(): Promise<boolean>;
     /**
      * Set model directly.
-     * Validates API key, saves to session and settings.
+     * Validates API key and saves to the active session. Persists settings only when requested.
      * @throws Error if no API key available for the model
      */
     setModel(model: Model, role?: string, options?: {
         selector?: string;
         thinkingLevel?: ThinkingLevel;
+        persist?: boolean;
     }): Promise<void>;
     /**
      * Set model temporarily (for this session only).
@@ -635,34 +650,33 @@ export declare class AgentSession {
      */
     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.
+     * Apply a resolved role model as the active model without changing global
+     * settings. Shared with role cycling and the plan-approval model slider.
      */
     applyRoleModel(entry: ResolvedRoleModel): Promise<void>;
     /**
      * Cycle through configured role models in a fixed order.
-     * Skips missing roles.
+     * Skips missing roles and changes only the active session model.
      * @param roleOrder - Order of roles to cycle through (e.g., ["slow", "default", "smol"])
-     * @param options - Optional settings: `temporary` to not persist to settings
+     * @param direction - "forward" (default) or "backward"
      */
-    cycleRoleModels(roleOrder: readonly string[], options?: {
-        temporary?: boolean;
-    }): Promise<RoleModelCycleResult | undefined>;
+    cycleRoleModels(roleOrder: readonly string[], direction?: "forward" | "backward"): Promise<RoleModelCycleResult | undefined>;
     /**
      * Get all available models with valid API keys.
      */
     getAvailableModels(): Model[];
     /**
-     * Set thinking level.
-     * Saves the effective metadata-clamped level to session and settings only if it changes.
+     * Set the thinking level. `auto` enables per-turn classification; the selector
+     * itself is never written to the session log, but resolved concrete levels are
+     * persisted when real user turns are classified so resumed sessions keep the
+     * last resolved effort instead of reverting to pending auto.
      */
-    setThinkingLevel(level: ThinkingLevel | undefined, persist?: boolean): void;
+    setThinkingLevel(level: ConfiguredThinkingLevel | undefined, persist?: boolean): void;
     /**
-     * Cycle to next thinking level.
-     * @returns New level, or undefined if model doesn't support thinking
+     * Cycle to next thinking level: off → auto → minimal..xhigh → off.
+     * @returns New selector, or undefined if model doesn't support thinking
      */
-    cycleThinkingLevel(): ThinkingLevel | undefined;
+    cycleThinkingLevel(): ConfiguredThinkingLevel | undefined;
     /**
      * True when *any* fast-mode-granting service tier is configured, regardless
      * of whether the active model's provider actually realizes it. Used by the
@@ -717,6 +731,27 @@ export declare class AgentSession {
     dropImages(): Promise<{
         removed: number;
     }>;
+    /**
+     * Surgically reduce context by dropping heavy content ("shake").
+     *
+     * - `images` delegates to {@link dropImages}.
+     * - `elide` replaces whole tool-call results and large fenced/XML blocks
+     *   with short placeholders that embed an `artifact://` recovery link.
+     * - `summary` extractively compresses the same regions with the configured
+     *   local on-device model (`providers.shakeSummaryModel`), falling back to
+     *   the elide placeholder per region (or wholesale when the local model is
+     *   unavailable). Never calls a remote/cloud LLM.
+     *
+     * Mutates the branch in place, persists via `rewriteEntries`, replays the
+     * rebuilt context through the agent, and tears down provider sessions that
+     * cache message identity — same rewrite contract as {@link dropImages}.
+     *
+     * No-op (zero counts) when nothing is eligible.
+     */
+    shake(mode: ShakeMode, opts?: {
+        config?: ShakeConfig;
+        signal?: AbortSignal;
+    }): Promise<ShakeResult>;
     /**
      * Manually compact the session context.
      * Aborts current agent operation first.
@@ -867,6 +902,7 @@ export declare class AgentSession {
         promptText: string;
         onTextDelta?: (delta: string) => void;
         signal?: AbortSignal;
+        dedupeReply?: boolean;
     }): Promise<{
         replyText: string;
         assistantMessage: AssistantMessage;

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

@@ -306,6 +306,24 @@ export declare class SessionManager {
     getCwd(): string;
     /** Get usage statistics across all assistant messages in the session. */
     getUsageStatistics(): UsageStatistics;
+    /**
+     * Open a new per-turn budget window: snapshot the cumulative output baseline,
+     * reset the eval-subagent counter, and set the (optional) ceiling. Called once
+     * per real user message; `total` is null when no `+Nk` directive was present.
+     */
+    beginTurnBudget(total: number | null, hard: boolean): void;
+    /** Record output tokens consumed by an eval-spawned subagent in the current turn. */
+    recordEvalSubagentOutput(output: number): void;
+    /**
+     * Current turn budget for the eval `budget` helper: the ceiling (null = none),
+     * output tokens spent this turn (main loop + eval-spawned subagents, no
+     * double-count), and whether the ceiling is hard.
+     */
+    getTurnBudget(): {
+        total: number | null;
+        spent: number;
+        hard: boolean;
+    };
     getSessionDir(): string;
     getSessionId(): string;
     getSessionFile(): string | undefined;

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

@@ -65,6 +65,12 @@ export declare class MemorySessionStorage implements SessionStorage {
     ensureDirSync(_dir: string): void;
     existsSync(path: string): boolean;
     writeTextSync(path: string, content: string): void;
+    /**
+     * Internal O(1) append used by {@link MemorySessionStorageWriter}. Lazily
+     * creates the entry. External callers should go through `openWriter()`
+     * rather than touching the mirror directly.
+     */
+    appendChunkSync(path: string, chunk: string): void;
     readTextSync(path: string): string;
     statSync(path: string): SessionStorageStat;
     listFilesSync(dir: string, pattern: string): string[];

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

@@ -0,0 +1,24 @@
+/**
+ * Public shape of the `shake` operation, kept in a dependency-free leaf module
+ * so slash-command registries and controllers can import `formatShakeSummary`
+ * without pulling in the heavy `agent-session` module graph (which would form
+ * an import cycle through the slash-command registry).
+ */
+/** Mode selector for `AgentSession.shake`. */
+export type ShakeMode = "elide" | "summary" | "images";
+/** Outcome of an `AgentSession.shake` run. */
+export interface ShakeResult {
+    mode: ShakeMode;
+    /** Whole tool-call results dropped/compressed. */
+    toolResultsDropped: number;
+    /** Large fenced/XML blocks dropped/compressed. */
+    blocksDropped: number;
+    /** Image blocks removed (images mode only). */
+    imagesDropped?: number;
+    /** Estimated context tokens reclaimed. */
+    tokensFreed: number;
+    /** Session artifact holding the dropped originals, when persisted. */
+    artifactId?: string;
+}
+/** One-line operator summary of a {@link ShakeResult} (shared by TUI + ACP). */
+export declare function formatShakeSummary(result: ShakeResult): string;

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

@@ -11,7 +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 { MnemopiSessionState } from "../mnemopi/state";
 import type { ArtifactManager } from "../session/artifacts";
 import type { AuthStorage } from "../session/auth-storage";
 import type { ContextFileEntry } from "../tools";
@@ -65,7 +65,7 @@ export interface ExecutorOptions {
      */
     parentArtifactManager?: ArtifactManager;
     parentHindsightSessionState?: HindsightSessionState;
-    parentMnemosyneSessionState?: MnemosyneSessionState;
+    parentMnemopiSessionState?: MnemopiSessionState;
     /** Parent agent's eval executor session id. Subagents reuse it so eval state is shared. */
     parentEvalSessionId?: string;
     /**

package/dist/types/thinking.d.ts

@@ -1,5 +1,5 @@
 import { type ResolvedThinkingLevel, ThinkingLevel } from "@oh-my-pi/pi-agent-core";
-import { type Effort, type Model } from "@oh-my-pi/pi-ai";
+import { Effort, type Model } from "@oh-my-pi/pi-ai";
 /**
  * Metadata used to render thinking selector values in the coding-agent UI.
  */
@@ -28,3 +28,41 @@ export declare function toReasoningEffort(level: ThinkingLevel | undefined): Eff
  * Resolves a selector against the current model while preserving explicit "off".
  */
 export declare function resolveThinkingLevelForModel(model: Model | undefined, level: ThinkingLevel | undefined): ResolvedThinkingLevel | undefined;
+/**
+ * Sentinel selector for the coding-agent "auto" thinking mode. Kept entirely
+ * inside the coding-agent layer: it is never an {@link Effort} or
+ * {@link ThinkingLevel}, so provider mapping/clamping keeps seeing concrete
+ * efforts. The session resolves `auto` to a concrete effort each turn.
+ */
+export declare const AUTO_THINKING: "auto";
+/** A thinking selector as configured by the user — a concrete level or `auto`. */
+export type ConfiguredThinkingLevel = ThinkingLevel | typeof AUTO_THINKING;
+/** Metadata used to render the `auto` selector value alongside concrete levels. */
+export interface ConfiguredThinkingLevelMetadata {
+    value: ConfiguredThinkingLevel;
+    label: string;
+    description: string;
+}
+/**
+ * Parses a configured thinking selector, accepting `auto` in addition to every
+ * value {@link parseThinkingLevel} accepts. {@link parseThinkingLevel} itself
+ * stays strict so model-suffix parsing (`model:high`) keeps rejecting `auto`.
+ */
+export declare function parseConfiguredThinkingLevel(value: string | null | undefined): ConfiguredThinkingLevel | undefined;
+/** Returns display metadata for a configured selector, including `auto`. */
+export declare function getConfiguredThinkingLevelMetadata(level: ConfiguredThinkingLevel): ConfiguredThinkingLevelMetadata;
+/**
+ * Resolves an auto-classified effort against the active model's supported
+ * range. Unlike {@link clampThinkingLevelForModel}, `auto` never resolves below
+ * {@link Effort.Low}: the eligible pool is the model's supported efforts at or
+ * above Low (falling back to the full supported set only when the model maxes
+ * out below Low). Within that pool the request snaps to the highest level not
+ * exceeding it, or the pool minimum when the request is below the pool.
+ */
+export declare function clampAutoThinkingEffort(model: Model | undefined, effort: Effort): Effort;
+/**
+ * The provisional concrete level shown while `auto` is configured but before a
+ * turn has been classified. Prefers the model's `defaultLevel`, otherwise High,
+ * clamped into the auto range. Returns `undefined` for non-reasoning models.
+ */
+export declare function resolveProvisionalAutoLevel(model: Model | undefined): Effort | undefined;

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

@@ -7,7 +7,7 @@ export interface TinyModelDevicePreference {
 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". */
+/** Sentinel `providers.tinyModelDevice` value meaning "use the built-in CPU 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"];
@@ -15,7 +15,7 @@ export declare const TINY_MODEL_DEVICE_SETTING_VALUES: readonly ["default", "gpu
 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 description: "CPU-only inference";
 }, {
     readonly value: "gpu";
     readonly label: "GPU";
@@ -72,7 +72,7 @@ export declare const TINY_MODEL_DEVICE_SETTING_OPTIONS: readonly [{
 /**
  * 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
+ * worker keeps its built-in CPU default; the worker still validates the
  * forwarded value via {@link normalizeTinyModelDevice}.
  */
 export declare function tinyModelDeviceSettingToEnv(value: string | undefined): string | undefined;

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

@@ -65,7 +65,7 @@ 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).
+ * 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.
@@ -113,6 +113,20 @@ export declare const TINY_MEMORY_MODEL_OPTIONS: ({
 })[];
 export declare function isTinyMemoryLocalModelKey(value: string): value is TinyMemoryLocalModelKey;
 export declare function getTinyMemoryModelSpec(key: TinyMemoryLocalModelKey): (typeof TINY_MEMORY_LOCAL_MODELS)[number];
+/**
+ * Shake-summary models. Shake's `summary` mode (and the `shake-summary`
+ * compaction strategy) compress heavy regions strictly on-device — there is no
+ * online/remote option, so this registry reuses the local memory models only.
+ */
+export declare const SHAKE_SUMMARY_MODEL_VALUES: readonly ["qwen3-1.7b", "gemma-3-1b", "qwen2.5-1.5b", "lfm2-1.2b"];
+export type ShakeSummaryModelKey = (typeof SHAKE_SUMMARY_MODEL_VALUES)[number];
+export declare const SHAKE_SUMMARY_MODEL_OPTIONS: {
+    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.";
+}[];
+/** Default shake-summary local model when none is named. */
+export declare const DEFAULT_SHAKE_SUMMARY_MODEL_KEY: ShakeSummaryModelKey;
 /** 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. */
@@ -183,3 +197,22 @@ export declare const TINY_LOCAL_MODELS: readonly [{
     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/title-protocol.d.ts

@@ -30,6 +30,10 @@ export type TinyTitleWorkerInbound = {
     modelKey: TinyLocalModelKey;
     prompt: string;
     maxTokens?: number;
+    /** Optional assistant-turn prefix appended after the generation prompt to pin output format. */
+    prefill?: string;
+    /** Optional literal stop string; generation halts once it appears in the decoded tail. */
+    stop?: string;
 } | {
     type: "download";
     id: string;

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

@@ -6,7 +6,9 @@ import type { Settings } from "../config/settings";
 import type { Skill } from "../extensibility/skills";
 import type { GoalModeState, GoalRuntime } from "../goals";
 import type { HindsightSessionState } from "../hindsight/state";
-import type { MnemosyneSessionState } from "../mnemosyne/state";
+import type { LocalProtocolOptions } from "../internal-urls";
+import type { MCPManager } from "../mcp";
+import type { MnemopiSessionState } from "../mnemopi/state";
 import type { PlanModeState } from "../plan-mode/state";
 import { type AgentRegistry } from "../registry/agent-registry";
 import type { ArtifactManager } from "../session/artifacts";
@@ -107,8 +109,8 @@ export interface ToolSession {
     getSessionId?: () => string | null;
     /** Get Hindsight runtime state for this agent session. */
     getHindsightSessionState?: () => HindsightSessionState | undefined;
-    /** Get Mnemosyne runtime state for this agent session. */
-    getMnemosyneSessionState?: () => MnemosyneSessionState | undefined;
+    /** Get Mnemopi runtime state for this agent session. */
+    getMnemopiSessionState?: () => MnemopiSessionState | undefined;
     /** Agent identity used for IRC routing. Returns the registry id (e.g. "0-Main", "0-AuthLoader"). */
     getAgentId?: () => string | null;
     /** Look up a registered tool by name (used by the eval js backend's tool bridge). */
@@ -136,6 +138,10 @@ export interface ToolSession {
     modelRegistry?: import("../config/model-registry").ModelRegistry;
     /** Agent output manager for unique agent:// IDs across task invocations */
     agentOutputManager?: AgentOutputManager;
+    /** MCP manager visible to subagents without relying on the process-global singleton. */
+    mcpManager?: MCPManager;
+    /** Local protocol root to propagate to nested subagents and eval-created agents. */
+    localProtocolOptions?: LocalProtocolOptions;
     /** Settings instance for passing to subagents */
     settings: Settings;
     /** Plan mode state (if active) */
@@ -144,6 +150,16 @@ export interface ToolSession {
     getGoalModeState?: () => GoalModeState | undefined;
     /** Goal runtime for the active agent session. */
     getGoalRuntime?: () => GoalRuntime | undefined;
+    /** Get cumulative session usage statistics (input/output tokens, cost). */
+    getUsageStatistics?: () => import("../session/session-manager").UsageStatistics;
+    /** Current per-turn token budget {total, spent, hard} for the eval `budget` helper. */
+    getTurnBudget?: () => {
+        total: number | null;
+        spent: number;
+        hard: boolean;
+    };
+    /** Record output tokens consumed by an eval-spawned subagent toward the current turn budget. */
+    recordEvalSubagentUsage?: (output: number) => void;
     /** Bridge to the connected client (e.g. ACP editor host). Tools should route fs/terminal/permission requests through this when available. */
     getClientBridge?: () => ClientBridge | undefined;
     /** Get compact conversation context for subagents (excludes tool results, system prompts) */

package/dist/types/tools/memory-edit.d.ts

@@ -32,7 +32,7 @@ export declare class MemoryEditTool implements AgentTool<typeof memoryEditSchema
     }, z.core.$strip>;
     readonly strict = true;
     readonly loadMode = "discoverable";
-    readonly summary = "Update, forget, or invalidate Mnemosyne memories";
+    readonly summary = "Update, forget, or invalidate Mnemopi memories";
     constructor(session: ToolSession);
     static createIf(session: ToolSession): MemoryEditTool | null;
     execute(_id: string, params: MemoryEditParams): Promise<AgentToolResult>;

package/examples/sdk/09-api-keys-and-oauth.ts

@@ -26,7 +26,7 @@ console.log("Session with default auth storage and model registry");
 
 // Custom auth storage location
 const customAuthStorage = await AuthStorage.create("/tmp/my-app/agent.db");
-const customModelRegistry = new ModelRegistry(customAuthStorage, "/tmp/my-app/models.json");
+const customModelRegistry = await ModelRegistry.create(customAuthStorage, "/tmp/my-app/models.json");
 
 await createAgentSession({
 	sessionManager: SessionManager.inMemory(),
@@ -45,7 +45,7 @@ await createAgentSession({
 console.log("Session with runtime API key override");
 
 // No models.json - only built-in models
-const simpleRegistry = new ModelRegistry(authStorage); // null = no models.json
+const simpleRegistry = await ModelRegistry.create(authStorage); // null = no models.json
 await createAgentSession({
 	sessionManager: SessionManager.inMemory(),
 	authStorage,

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-coding-agent",
-	"version": "15.7.1",
+	"version": "15.7.3",
 	"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",
+		"test": "bun test --parallel",
 		"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",
@@ -47,14 +47,14 @@
 		"@agentclientprotocol/sdk": "0.22.1",
 		"@babel/parser": "^7.29.7",
 		"@mozilla/readability": "^0.6.0",
-		"@oh-my-pi/hashline": "15.7.1",
-		"@oh-my-pi/omp-stats": "15.7.1",
-		"@oh-my-pi/pi-agent-core": "15.7.1",
-		"@oh-my-pi/pi-ai": "15.7.1",
-		"@oh-my-pi/pi-mnemosyne": "15.7.1",
-		"@oh-my-pi/pi-natives": "15.7.1",
-		"@oh-my-pi/pi-tui": "15.7.1",
-		"@oh-my-pi/pi-utils": "15.7.1",
+		"@oh-my-pi/hashline": "15.7.3",
+		"@oh-my-pi/omp-stats": "15.7.3",
+		"@oh-my-pi/pi-agent-core": "15.7.3",
+		"@oh-my-pi/pi-ai": "15.7.3",
+		"@oh-my-pi/pi-mnemopi": "15.7.3",
+		"@oh-my-pi/pi-natives": "15.7.3",
+		"@oh-my-pi/pi-tui": "15.7.3",
+		"@oh-my-pi/pi-utils": "15.7.3",
 		"@puppeteer/browsers": "^3.0.4",
 		"@types/turndown": "5.0.6",
 		"@xterm/headless": "^6.0.0",