@oh-my-pi/pi-coding-agent 15.3.2 → 15.4.1

package/src/sdk.ts

@@ -51,6 +51,7 @@ import "./discovery";
 import { resolveConfigValue } from "./config/resolve-config-value";
 import { initializeWithSettings } from "./discovery";
 import { disposeAllKernelSessions, disposeKernelSessionsByOwner } from "./eval/py/executor";
+import { defaultEvalSessionId } from "./eval/session-id";
 import { TtsrManager } from "./export/ttsr";
 import {
 	type CustomCommandsLoadResult,
@@ -84,11 +85,7 @@ import type { HindsightSessionState } from "./hindsight/state";
 import { LocalProtocolHandler, type LocalProtocolOptions } from "./internal-urls";
 import { LSP_STARTUP_EVENT_CHANNEL, type LspStartupEvent } from "./lsp/startup-events";
 import { discoverAndLoadMCPTools, MCPManager, type MCPToolsLoadResult } from "./mcp";
-import {
-	collectDiscoverableMCPTools,
-	formatDiscoverableMCPToolServerSummary,
-	selectDiscoverableMCPToolNamesByServer,
-} from "./mcp/discoverable-tool-metadata";
+
 import { resolveMemoryBackend } from "./memory-backend";
 import asyncResultTemplate from "./prompts/tools/async-result.md" with { type: "text" };
 import { AgentRegistry, MAIN_AGENT_ID } from "./registry/agent-registry";
@@ -117,6 +114,9 @@ import { parseThinkingLevel, resolveThinkingLevelForModel, toReasoningEffort } f
 import {
 	collectDiscoverableTools,
 	type DiscoverableTool,
+	filterBySource,
+	formatDiscoverableToolServerSummary,
+	selectDiscoverableToolNamesByServer,
 	summarizeDiscoverableTools,
 } from "./tool-discovery/tool-index";
 import {
@@ -318,6 +318,8 @@ export interface CreateAgentSessionOptions {
 	agentRegistry?: AgentRegistry;
 	/** Parent task ID prefix for nested artifact naming (e.g., "6-Extensions") */
 	parentTaskPrefix?: string;
+	/** Inherited eval executor session id for subagents sharing parent eval state. */
+	parentEvalSessionId?: string;
 
 	/** Session manager. Default: session stored under the configured agentDir sessions root */
 	sessionManager?: SessionManager;
@@ -1177,6 +1179,8 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			taskDepth: options.taskDepth ?? 0,
 			getSessionFile: () => sessionManager.getSessionFile() ?? null,
 			getEvalKernelOwnerId: () => evalKernelOwnerId,
+			getEvalSessionId: () =>
+				session?.getEvalSessionId() ?? options.parentEvalSessionId ?? defaultEvalSessionId(toolSession),
 			assertEvalExecutionAllowed: () => session?.assertEvalExecutionAllowed(),
 			trackEvalExecution: (execution, abortController) =>
 				session ? session.trackEvalExecution(execution, abortController) : execution,
@@ -1196,8 +1200,6 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			getTodoPhases: () => session.getTodoPhases(),
 			setTodoPhases: phases => session.setTodoPhases(phases),
 			isMCPDiscoveryEnabled: () => session.isMCPDiscoveryEnabled(),
-			getDiscoverableMCPTools: () => session.getDiscoverableMCPTools(),
-			getDiscoverableMCPSearchIndex: () => session.getDiscoverableMCPSearchIndex(),
 			getSelectedMCPToolNames: () => session.getSelectedMCPToolNames(),
 			activateDiscoveredMCPTools: toolNames => session.activateDiscoveredMCPTools(toolNames),
 			// Generic tool discovery (unified — covers built-in + MCP + extension)
@@ -1581,7 +1583,9 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			tools: Map<string, AgentTool>,
 		): Promise<BuildSystemPromptResult> => {
 			toolContextStore.setToolNames(toolNames);
-			const discoverableMCPTools = mcpDiscoveryEnabled ? collectDiscoverableMCPTools(tools.values()) : [];
+			const discoverableMCPTools: DiscoverableTool[] = mcpDiscoveryEnabled
+				? filterBySource(collectDiscoverableTools(tools.values()), "mcp")
+				: [];
 			const activeToolNames = new Set(toolNames);
 			const discoverableBuiltinTools: DiscoverableTool[] =
 				effectiveDiscoveryMode === "all"
@@ -1592,18 +1596,7 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 							{ source: "builtin" },
 						)
 					: [];
-			const discoverableToolsForDesc: DiscoverableTool[] = [
-				...discoverableBuiltinTools,
-				...discoverableMCPTools.map(t => ({
-					name: t.name,
-					label: t.label,
-					summary: t.description,
-					source: "mcp" as const,
-					serverName: t.serverName,
-					mcpToolName: t.mcpToolName,
-					schemaKeys: t.schemaKeys,
-				})),
-			];
+			const discoverableToolsForDesc: DiscoverableTool[] = [...discoverableBuiltinTools, ...discoverableMCPTools];
 			const discoverableToolSummary = summarizeDiscoverableTools(discoverableToolsForDesc);
 			const hasDiscoverableTools =
 				mcpDiscoveryEnabled && toolNames.includes("search_tool_bm25") && discoverableToolsForDesc.length > 0;
@@ -1647,7 +1640,7 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 				repeatToolDescriptions,
 				intentField,
 				mcpDiscoveryMode: hasDiscoverableTools,
-				mcpDiscoveryServerSummaries: discoverableToolSummary.servers.map(formatDiscoverableMCPToolServerSummary),
+				mcpDiscoveryServerSummaries: discoverableToolSummary.servers.map(formatDiscoverableToolServerSummary),
 				eagerTasks,
 				secretsEnabled,
 				workspaceTree: workspaceTreePromise,
@@ -1693,8 +1686,8 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			(settings.get("mcp.discoveryDefaultServers") ?? []).map(serverName => serverName.trim()).filter(Boolean),
 		);
 		const discoveryDefaultServerToolNames = mcpDiscoveryEnabled
-			? selectDiscoverableMCPToolNamesByServer(
-					collectDiscoverableMCPTools(toolRegistry.values()),
+			? selectDiscoverableToolNamesByServer(
+					filterBySource(collectDiscoverableTools(toolRegistry.values()), "mcp"),
 					discoveryDefaultServers,
 				)
 			: [];
@@ -1994,6 +1987,7 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			agentId: resolvedAgentId,
 			agentRegistry,
 			providerSessionId: options.providerSessionId,
+			parentEvalSessionId: options.parentEvalSessionId,
 		});
 		hasSession = true;
 		if (asyncJobManager) {

package/src/session/agent-session.ts

@@ -108,6 +108,7 @@ import {
 	executePython as executePythonCommand,
 	type PythonResult,
 } from "../eval/py/executor";
+import { defaultEvalSessionId } from "../eval/session-id";
 import { type BashResult, executeBash as executeBashCommand } from "../exec/bash-executor";
 import { exportSessionToHtml } from "../export/html";
 import type { TtsrManager, TtsrMatchContext } from "../export/ttsr";
@@ -141,14 +142,6 @@ import { GoalRuntime } from "../goals/runtime";
 import type { Goal, GoalModeState } from "../goals/state";
 import type { HindsightSessionState } from "../hindsight/state";
 import { type LocalProtocolOptions, resolveLocalUrlToPath } from "../internal-urls";
-import {
-	buildDiscoverableMCPSearchIndex,
-	collectDiscoverableMCPTools,
-	type DiscoverableMCPSearchIndex,
-	type DiscoverableMCPTool,
-	isMCPToolName,
-	selectDiscoverableMCPToolNamesByServer,
-} from "../mcp/discoverable-tool-metadata";
 import { resolveMemoryBackend } from "../memory-backend";
 import { getCurrentThemeName, theme } from "../modes/theme/theme";
 import type { PlanModeState } from "../plan-mode/state";
@@ -171,6 +164,9 @@ import {
 	collectDiscoverableTools,
 	type DiscoverableTool,
 	type DiscoverableToolSearchIndex,
+	filterBySource,
+	isMCPToolName,
+	selectDiscoverableToolNamesByServer,
 } from "../tool-discovery/tool-index";
 import { assertEditableFile } from "../tools/auto-generated-guard";
 import type { CheckpointState } from "../tools/checkpoint";
@@ -312,6 +308,8 @@ export interface AgentSessionConfig {
 	ttsrManager?: TtsrManager;
 	/** Secret obfuscator for deobfuscating streaming edit content */
 	obfuscator?: SecretObfuscator;
+	/** Inherited eval executor session id from a parent agent. */
+	parentEvalSessionId?: string;
 	/** Logical owner for retained Python kernels created by this session. */
 	evalKernelOwnerId?: string;
 	/**
@@ -808,6 +806,7 @@ export class AgentSession {
 	// Python execution state
 	#evalAbortControllers = new Set<AbortController>();
 	#evalKernelOwnerId: string;
+	#parentEvalSessionId: string | undefined;
 	/**
 	 * AsyncJobManager owned by this session (top-level only). Subagents leave
 	 * this undefined and **MUST NOT** dispose the global instance on teardown.
@@ -865,8 +864,7 @@ export class AgentSession {
 	 */
 	#lastAppliedToolSignature: string | undefined;
 	#mcpDiscoveryEnabled = false;
-	#discoverableMCPTools = new Map<string, DiscoverableMCPTool>();
-	#discoverableMCPSearchIndex: DiscoverableMCPSearchIndex | null = null;
+	#discoverableMCPTools = new Map<string, DiscoverableTool>();
 	#selectedMCPToolNames = new Set<string>();
 	// Generic tool discovery (covers built-in + MCP + extension when tools.discoveryMode === "all")
 	#discoverableToolSearchIndex: DiscoverableToolSearchIndex | null = null;
@@ -997,6 +995,7 @@ export class AgentSession {
 		this.settings = config.settings;
 		// Power assertions are taken per turn (see #beginInFlight); nothing acquired here.
 		this.#evalKernelOwnerId = config.evalKernelOwnerId ?? `agent-session:${Snowflake.next()}`;
+		this.#parentEvalSessionId = config.parentEvalSessionId;
 		this.#ownedAsyncJobManager = config.ownedAsyncJobManager;
 		this.#scopedModels = config.scopedModels ?? [];
 		this.#thinkingLevel = config.thinkingLevel;
@@ -2876,11 +2875,12 @@ export class AgentSession {
 		return this.#retryAttempt;
 	}
 
-	#collectDiscoverableMCPToolsFromRegistry(): Map<string, DiscoverableMCPTool> {
-		return new Map(collectDiscoverableMCPTools(this.#toolRegistry.values()).map(tool => [tool.name, tool] as const));
+	#collectDiscoverableMCPToolsFromRegistry(): Map<string, DiscoverableTool> {
+		const mcpTools = filterBySource(collectDiscoverableTools(this.#toolRegistry.values()), "mcp");
+		return new Map(mcpTools.map(tool => [tool.name, tool] as const));
 	}
 
-	#setDiscoverableMCPTools(discoverableMCPTools: Map<string, DiscoverableMCPTool>): void {
+	#setDiscoverableMCPTools(discoverableMCPTools: Map<string, DiscoverableTool>): void {
 		this.#discoverableMCPTools = discoverableMCPTools;
 		this.#invalidateDiscoveryCaches();
 	}
@@ -2889,7 +2889,6 @@ export class AgentSession {
 	 *  affect which tools should be discoverable: registry mutations (refreshMCPTools,
 	 *  refreshRpcHostTools) or active-tool mutations (#applyActiveToolsByName). */
 	#invalidateDiscoveryCaches(): void {
-		this.#discoverableMCPSearchIndex = null;
 		this.#discoverableToolSearchIndex = null;
 	}
 
@@ -2900,7 +2899,7 @@ export class AgentSession {
 	#getConfiguredDefaultSelectedMCPToolNames(): string[] {
 		return this.#filterSelectableMCPToolNames([
 			...this.#defaultSelectedMCPToolNames,
-			...selectDiscoverableMCPToolNamesByServer(
+			...selectDiscoverableToolNamesByServer(
 				this.#discoverableMCPTools.values(),
 				this.#defaultSelectedMCPServerNames,
 			),
@@ -2993,28 +2992,6 @@ export class AgentSession {
 		return this.#mcpDiscoveryEnabled;
 	}
 
-	/** @deprecated Use {@link getDiscoverableTools} with `{ source: "mcp" }` instead.
-	 *  Preserves the legacy `description`-bearing MCP shape for back-compat callers. */
-	getDiscoverableMCPTools(): DiscoverableMCPTool[] {
-		return Array.from(this.#discoverableMCPTools.values()).map(t => ({
-			name: t.name,
-			label: t.label,
-			description: t.description,
-			serverName: t.serverName,
-			mcpToolName: t.mcpToolName,
-			schemaKeys: t.schemaKeys,
-		}));
-	}
-
-	/** @deprecated Use {@link getDiscoverableToolSearchIndex} instead.
-	 *  Returns the legacy MCP search index whose documents expose `tool.description`. */
-	getDiscoverableMCPSearchIndex(): DiscoverableMCPSearchIndex {
-		if (!this.#discoverableMCPSearchIndex) {
-			this.#discoverableMCPSearchIndex = buildDiscoverableMCPSearchIndex(this.#discoverableMCPTools.values());
-		}
-		return this.#discoverableMCPSearchIndex;
-	}
-
 	getSelectedMCPToolNames(): string[] {
 		if (!this.#mcpDiscoveryEnabled) {
 			return this.getActiveToolNames().filter(name => isMCPToolName(name) && this.#toolRegistry.has(name));
@@ -3062,17 +3039,7 @@ export class AgentSession {
 		// For "mcp-only" mode we only return MCP tools.
 		const mode = this.#resolveEffectiveDiscoveryMode();
 		const activeNames = new Set(this.getActiveToolNames());
-		const mcpTools: DiscoverableTool[] = Array.from(this.#discoverableMCPTools.values())
-			.filter(t => !activeNames.has(t.name))
-			.map(t => ({
-				name: t.name,
-				label: t.label,
-				summary: t.description,
-				source: "mcp" as const,
-				serverName: t.serverName,
-				mcpToolName: t.mcpToolName,
-				schemaKeys: t.schemaKeys,
-			}));
+		const mcpTools = Array.from(this.#discoverableMCPTools.values()).filter(t => !activeNames.has(t.name));
 		const builtinTools: DiscoverableTool[] = mode === "all" ? this.#collectDiscoverableBuiltinTools() : [];
 		const allTools = [...builtinTools, ...mcpTools];
 		return filter?.source ? allTools.filter(t => t.source === filter.source) : allTools;
@@ -3693,6 +3660,13 @@ export class AgentSession {
 	get sessionId(): string {
 		return this.#providerSessionId ?? this.sessionManager.getSessionId();
 	}
+	getEvalSessionId(): string | null {
+		if (this.#parentEvalSessionId !== undefined) return this.#parentEvalSessionId;
+		return defaultEvalSessionId({
+			cwd: this.sessionManager.getCwd(),
+			getSessionFile: () => this.sessionManager.getSessionFile() ?? null,
+		});
+	}
 
 	/** Current session display name, if set */
 	get sessionName(): string | undefined {
@@ -4225,7 +4199,6 @@ export class AgentSession {
 				void this.dispose();
 				process.exit(0);
 			},
-			hasQueuedMessages: () => this.queuedMessageCount > 0,
 			getContextUsage: () => this.getContextUsage(),
 			waitForIdle: () => this.waitForIdle(),
 			newSession: async options => {
@@ -4558,11 +4531,11 @@ export class AgentSession {
 	}
 
 	/**
-	 * Send a user message to the agent. Always triggers a turn.
-	 * When the agent is streaming, use deliverAs to specify how to queue the message.
+	 * Send a user message to the agent.
+	 * When deliverAs is set, queue the message instead of starting a new turn.
 	 *
 	 * @param content User message content (string or content array)
-	 * @param options.deliverAs Delivery mode when streaming: "steer" or "followUp"
+	 * @param options.deliverAs Delivery mode: "steer" or "followUp"
 	 */
 	async sendUserMessage(
 		content: string | (TextContent | ImageContent)[],
@@ -4588,10 +4561,18 @@ export class AgentSession {
 			if (images.length === 0) images = undefined;
 		}
 
+		if (options?.deliverAs === "followUp") {
+			await this.#queueFollowUp(text, images);
+			return;
+		}
+		if (options?.deliverAs === "steer") {
+			await this.#queueSteer(text, images);
+			return;
+		}
+
 		// Use prompt() with expandPromptTemplates: false to skip command handling and template expansion
 		await this.prompt(text, {
 			expandPromptTemplates: false,
-			streamingBehavior: options?.deliverAs,
 			images,
 		});
 	}
@@ -6781,10 +6762,11 @@ export class AgentSession {
 	#isTransientTransportErrorMessage(errorMessage: string): boolean {
 		// Match: overloaded_error, provider returned error, rate limit, 429, 500, 502, 503, 504,
 		// service unavailable, provider-suggested retry, network/connection/socket errors, fetch failed,
-		// terminated, retry delay exceeded
+		// terminated, retry delay exceeded, Bun HTTP/2 stream resets (RST_STREAM / REFUSED_STREAM /
+		// ENHANCE_YOUR_CALM, surfaced verbatim from src/http/h2_client/dispatch.zig)
 		return (
 			isUnexpectedSocketCloseMessage(errorMessage) ||
-			/overloaded|provider.?returned.?error|rate.?limit|too many requests|429|500|502|503|504|service.?unavailable|server.?error|internal.?error|retry your request|network.?error|connection.?error|connection.?refused|other side closed|fetch failed|upstream.?connect|reset before headers|socket hang up|timed? out|timeout|terminated|retry delay|stream stall|no error details in response/i.test(
+			/overloaded|provider.?returned.?error|rate.?limit|too many requests|429|500|502|503|504|service.?unavailable|server.?error|internal.?error|retry your request|network.?error|connection.?error|connection.?refused|other side closed|fetch failed|upstream.?connect|reset before headers|socket hang up|timed? out|timeout|terminated|retry delay|stream stall|no error details in response|HTTP2(?:StreamReset|RefusedStream|EnhanceYourCalm)/i.test(
 				errorMessage,
 			)
 		);
@@ -7421,9 +7403,13 @@ export class AgentSession {
 				}
 			}
 
-			// Use the same session ID as eval's Python backend for kernel sharing
-			const sessionFile = this.sessionManager.getSessionFile();
-			const sessionId = sessionFile ? `session:${sessionFile}:cwd:${cwd}` : `cwd:${cwd}`;
+			// Use the same session ID as eval's Python backend for kernel sharing.
+			const sessionId =
+				this.getEvalSessionId() ??
+				defaultEvalSessionId({
+					cwd,
+					getSessionFile: () => this.sessionManager.getSessionFile() ?? null,
+				});
 			const result = await executePythonCommand(code, {
 				cwd,
 				sessionId,
@@ -7562,11 +7548,11 @@ export class AgentSession {
 	 * Generate an ephemeral reply to a background message (e.g. an IRC ping from
 	 * another agent) using this session's current model + system prompt + history.
 	 *
-	 * The reply is computed via a side-channel `streamSimple` call (analogous to
-	 * `/btw`) so it never blocks on the recipient's in-flight tool calls.  After
-	 * the reply is generated, both the incoming question and the auto-reply are
-	 * queued for injection into the recipient's persisted history so the model
-	 * sees the exchange on its next turn.  Injection happens immediately when the
+	 * The incoming message is queued for injection into the recipient's persisted
+	 * history immediately so timeouts/abort still preserve delivery. The reply is
+	 * computed via a side-channel `streamSimple` call (analogous to `/btw`) so it
+	 * never blocks on the recipient's in-flight tool calls. When a reply is
+	 * generated, it is queued separately. Injection happens immediately when the
 	 * session is idle, otherwise it is deferred until streaming ends.
 	 */
 	async respondAsBackground(args: {
@@ -7595,8 +7581,8 @@ export class AgentSession {
 			timestamp: incomingTimestamp,
 		});
 
+		this.#queueBackgroundExchangeInjection([incomingRecord]);
 		if (!awaitReply) {
-			this.#queueBackgroundExchangeInjection([incomingRecord]);
 			return { replyText: null };
 		}
 
@@ -7626,7 +7612,7 @@ export class AgentSession {
 			kind: "reply",
 			timestamp: replyRecord.timestamp,
 		});
-		this.#queueBackgroundExchangeInjection([incomingRecord, replyRecord]);
+		this.#queueBackgroundExchangeInjection([replyRecord]);
 
 		return { replyText };
 	}
@@ -7708,10 +7694,17 @@ export class AgentSession {
 			// removes the surface entirely.
 			tools: [],
 		};
+		const cacheSessionId = this.sessionId;
 		const options = this.prepareSimpleStreamOptions(
 			{
 				apiKey,
-				sessionId: this.sessionId,
+				// Side-channel turns must not share OpenAI/Codex append-only
+				// conversation state with the main agent turn: IRC and /btw can run
+				// while the main turn is mid-tool-call. Keep the prompt-cache key
+				// stable, but give provider routing a unique request lineage.
+				sessionId: `${cacheSessionId}:side:${Snowflake.next()}`,
+				promptCacheKey: cacheSessionId,
+				preferWebsockets: false,
 				reasoning: toReasoningEffort(this.thinkingLevel),
 				hideThinkingSummary: this.agent.hideThinkingSummary,
 				serviceTier: this.serviceTier,

package/src/session/agent-storage.ts

@@ -240,11 +240,10 @@ FROM model_usage_legacy
 	}
 
 	/**
-	 * Retrieves all settings from storage (legacy, for migration only).
-	 * Settings are now stored in config.yml. This method is only used
-	 * during migration from agent.db to config.yml.
+	 * Reads legacy settings persisted in the agent.db `settings` table.
+	 * The canonical settings store is `config.yml`; this accessor only
+	 * exists so the config loader can migrate values from older installs.
 	 * @returns Settings object, or null if no settings are stored
-	 * @deprecated Use config.yml instead. This is only for migration.
 	 */
 	getSettings(): Settings | null {
 		const rows = (this.#listSettingsStmt.all() as SettingsRow[]) ?? [];
@@ -263,16 +262,6 @@ FROM model_usage_legacy
 		return settings as Settings;
 	}
 
-	/**
-	 * @deprecated Settings are now stored in config.yml, not agent.db.
-	 * This method is kept for backward compatibility but does nothing.
-	 */
-	saveSettings(settings: Settings): void {
-		logger.warn("AgentStorage.saveSettings is deprecated - settings are now stored in config.yml", {
-			keys: Object.keys(settings),
-		});
-	}
-
 	/**
 	 * Records model usage, updating the last-used timestamp.
 	 * @param modelKey - Model key in "provider/modelId" format
@@ -313,6 +302,16 @@ FROM model_usage_legacy
 		return this.#authStore.listAuthCredentials().length > 0;
 	}
 
+	/**
+	 * Returns the underlying {@link AuthCredentialStore} so callers that need
+	 * the lower-level pi-ai abstraction (e.g. `findAnthropicAuth(store)`) can
+	 * reuse this storage's open database connection instead of opening their
+	 * own.
+	 */
+	get authStore(): AuthCredentialStore {
+		return this.#authStore;
+	}
+
 	/**
 	 * Lists auth credentials, optionally filtered by provider.
 	 * Only returns active (non-disabled) credentials by default.

package/src/slash-commands/acp-builtins.ts

@@ -1,9 +1,9 @@
 import type { AvailableCommand } from "@agentclientprotocol/sdk";
 import { BUILTIN_SLASH_COMMANDS_INTERNAL, lookupBuiltinSlashCommand } from "./builtin-registry";
 import { parseSlashCommand } from "./helpers/parse";
-import type { AcpBuiltinCommandRuntime, AcpBuiltinSlashCommandResult } from "./types";
+import type { AcpBuiltinSlashCommandResult, SlashCommandRuntime } from "./types";
 
-export type { AcpBuiltinCommandRuntime, AcpBuiltinSlashCommandResult } from "./types";
+export type { AcpBuiltinSlashCommandResult } from "./types";
 
 /**
  * Commands advertised to ACP clients. Entries without a text-mode `handle`
@@ -34,7 +34,7 @@ export const ACP_BUILTIN_SLASH_COMMANDS: AvailableCommand[] = BUILTIN_SLASH_COMM
  */
 export async function executeAcpBuiltinSlashCommand(
 	text: string,
-	runtime: AcpBuiltinCommandRuntime,
+	runtime: SlashCommandRuntime,
 ): Promise<AcpBuiltinSlashCommandResult> {
 	const parsed = parseSlashCommand(text);
 	if (!parsed) return false;

package/src/slash-commands/types.ts

@@ -116,11 +116,5 @@ export interface SlashCommandSpec extends BuiltinSlashCommand {
 	) => Promise<SlashCommandResult> | SlashCommandResult;
 }
 
-/**
- * @deprecated Use `SlashCommandRuntime` directly. Retained as an alias so
- * downstream code that imported the ACP-specific name keeps compiling.
- */
-export type AcpBuiltinCommandRuntime = SlashCommandRuntime;
-
 /** Result returned by `executeAcpBuiltinSlashCommand`. */
 export type AcpBuiltinSlashCommandResult = false | { consumed: true } | { prompt: string };

package/src/task/executor.ts

@@ -7,7 +7,6 @@
 import path from "node:path";
 import type { AgentEvent, AgentIdentity, AgentTelemetryConfig, ThinkingLevel } from "@oh-my-pi/pi-agent-core";
 import { recordHandoff, resolveTelemetry } from "@oh-my-pi/pi-agent-core";
-import { type JsonSchemaValidationIssue, validateJsonSchemaValue } from "@oh-my-pi/pi-ai/utils/schema";
 import { logger, prompt, untilAborted } from "@oh-my-pi/pi-utils";
 import { ModelRegistry } from "../config/model-registry";
 import { resolveModelOverrideWithAuthFallback } from "../config/model-resolver";
@@ -33,7 +32,10 @@ import { SKILL_PROMPT_MESSAGE_TYPE } from "../session/messages";
 import { SessionManager } from "../session/session-manager";
 import { truncateTail } from "../session/streaming-output";
 import type { ContextFileEntry } from "../tools";
-import { jtdToJsonSchema, normalizeSchema } from "../tools/jtd-to-json-schema";
+import { normalizeSchema } from "../tools/jtd-to-json-schema";
+import { buildOutputValidator, summarizeValidationFailure } from "../tools/output-schema-validator";
+
+import { type ReportFindingDetails, toReviewFinding } from "../tools/review";
 import { ToolAbortError } from "../tools/tool-errors";
 import type { EventBus } from "../utils/event-bus";
 import { buildNamedToolChoice } from "../utils/tool-choice";
@@ -183,6 +185,8 @@ export interface ExecutorOptions {
 	 */
 	parentArtifactManager?: ArtifactManager;
 	parentHindsightSessionState?: HindsightSessionState;
+	/** Parent agent's eval executor session id. Subagents reuse it so eval state is shared. */
+	parentEvalSessionId?: string;
 	/**
 	 * Parent agent's OpenTelemetry configuration. When defined, the subagent's
 	 * loop is started with the same tracer/hooks but its own agent identity
@@ -208,51 +212,6 @@ function parseStringifiedJson(value: unknown): unknown {
 	}
 }
 
-interface OutputValidator {
-	validate: (value: unknown) => { ok: true } | { ok: false; message: string; missingRequired: string[] };
-	requiredFields: string[];
-}
-
-function buildOutputValidator(schema: unknown): { validator?: OutputValidator; error?: string } {
-	const { normalized, error } = normalizeSchema(schema);
-	if (error) return { error };
-	if (normalized === undefined) return {};
-	const jsonSchema = jtdToJsonSchema(normalized);
-	const required = extractRequiredFields(jsonSchema);
-	return {
-		validator: {
-			requiredFields: required,
-			validate: value => {
-				const result = validateJsonSchemaValue(jsonSchema, value);
-				if (result.success) return { ok: true };
-				const missing = computeMissingRequired(required, value);
-				const message = formatValidationIssue(result.issues[0]) ?? "schema validation failed";
-				return { ok: false, message, missingRequired: missing };
-			},
-		},
-	};
-}
-
-function extractRequiredFields(jsonSchema: unknown): string[] {
-	if (!jsonSchema || typeof jsonSchema !== "object") return [];
-	const required = (jsonSchema as { required?: unknown }).required;
-	return Array.isArray(required) ? required.filter((k): k is string => typeof k === "string") : [];
-}
-
-function computeMissingRequired(required: readonly string[], value: unknown): string[] {
-	if (required.length === 0) return [];
-	if (value === null || value === undefined) return [...required];
-	if (typeof value !== "object" || Array.isArray(value)) return [];
-	const record = value as Record<string, unknown>;
-	return required.filter(key => !(key in record) || record[key] === undefined);
-}
-
-function formatValidationIssue(issue: JsonSchemaValidationIssue | undefined): string | undefined {
-	if (!issue) return undefined;
-	const path = issue.path.length > 0 ? issue.path.map(String).join(".") : "(root)";
-	return `${path}: ${issue.message}`;
-}
-
 function previewOffendingData(value: unknown, maxLength = 500): string {
 	let serialized: string;
 	try {
@@ -306,7 +265,7 @@ function resolveFallbackCompletion(rawOutput: string, outputSchema: unknown): {
 	if (candidate === undefined) return null;
 	const { validator, error } = buildOutputValidator(outputSchema);
 	if (error) return null;
-	if (validator && !validator.validate(candidate).ok) return null;
+	if (validator && !validator.validate(candidate).success) return null;
 	return { data: candidate };
 }
 
@@ -393,9 +352,10 @@ export function finalizeSubprocessOutput(args: FinalizeSubprocessOutputArgs): Fi
 					stderr = `schema_violation: invalid output schema: ${schemaError}`;
 					exitCode = 1;
 				} else {
-					const verdict = validator ? validator.validate(completeData) : { ok: true as const };
-					if (!verdict.ok) {
-						const outcome = buildSchemaViolationOutcome(verdict, completeData);
+					const result = validator?.validate(completeData) ?? { success: true as const };
+					if (!result.success) {
+						const summary = summarizeValidationFailure(result, completeData, validator?.requiredFields ?? []);
+						const outcome = buildSchemaViolationOutcome(summary, completeData);
 						rawOutput = outcome.rawOutput;
 						stderr = outcome.stderr;
 						exitCode = outcome.exitCode;
@@ -420,9 +380,10 @@ export function finalizeSubprocessOutput(args: FinalizeSubprocessOutputArgs): Fi
 		if (fallback) {
 			const completeData = normalizeCompleteData(fallback.data, reportFindings);
 			const { validator } = buildOutputValidator(outputSchema);
-			const verdict = validator ? validator.validate(completeData) : { ok: true as const };
-			if (!verdict.ok) {
-				const outcome = buildSchemaViolationOutcome(verdict, completeData);
+			const result = validator?.validate(completeData) ?? { success: true as const };
+			if (!result.success) {
+				const summary = summarizeValidationFailure(result, completeData, validator?.requiredFields ?? []);
+				const outcome = buildSchemaViolationOutcome(summary, completeData);
 				rawOutput = outcome.rawOutput;
 				stderr = outcome.stderr;
 				exitCode = outcome.exitCode;
@@ -1277,6 +1238,7 @@ export async function runSubprocess(options: ExecutorOptions): Promise<SingleRes
 					customTools: mcpProxyTools.length > 0 ? mcpProxyTools : undefined,
 					localProtocolOptions: options.localProtocolOptions,
 					telemetry: subagentTelemetry,
+					parentEvalSessionId: options.parentEvalSessionId,
 				}),
 			);
 
@@ -1324,22 +1286,25 @@ export async function runSubprocess(options: ExecutorOptions): Promise<SingleRes
 			}
 
 			const extensionRunner = session.extensionRunner;
+			const pendingExtensionMessages: Promise<void>[] = [];
 			if (extensionRunner) {
 				extensionRunner.initialize(
 					{
 						sendMessage: (message, options) => {
-							session.sendCustomMessage(message, options).catch(e => {
+							const sendPromise = session.sendCustomMessage(message, options).catch(e => {
 								logger.error("Extension sendMessage failed", {
 									error: e instanceof Error ? e.message : String(e),
 								});
 							});
+							pendingExtensionMessages.push(sendPromise);
 						},
 						sendUserMessage: (content, options) => {
-							session.sendUserMessage(content, options).catch(e => {
+							const sendPromise = session.sendUserMessage(content, options).catch(e => {
 								logger.error("Extension sendUserMessage failed", {
 									error: e instanceof Error ? e.message : String(e),
 								});
 							});
+							pendingExtensionMessages.push(sendPromise);
 						},
 						appendEntry: (customType, data) => {
 							session.sessionManager.appendCustomEntry(customType, data);
@@ -1375,6 +1340,9 @@ export async function runSubprocess(options: ExecutorOptions): Promise<SingleRes
 					logger.error("Extension error", { path: err.extensionPath, error: err.error });
 				});
 				await awaitAbortable(extensionRunner.emit({ type: "session_start" }));
+				while (pendingExtensionMessages.length > 0) {
+					await awaitAbortable(Promise.all(pendingExtensionMessages.splice(0)));
+				}
 			}
 
 			const MAX_YIELD_RETRIES = 3;
@@ -1547,7 +1515,8 @@ export async function runSubprocess(options: ExecutorOptions): Promise<SingleRes
 	// Use final output if available, otherwise accumulated output
 	let rawOutput = finalOutputChunks.length > 0 ? finalOutputChunks.join("") : outputChunks.join("");
 	const yieldItems = progress.extractedToolData?.yield as YieldItem[] | undefined;
-	const reportFindings = progress.extractedToolData?.report_finding as ReviewFinding[] | undefined;
+	const reportFindingDetails = progress.extractedToolData?.report_finding as ReportFindingDetails[] | undefined;
+	const reportFindings: ReviewFinding[] | undefined = reportFindingDetails?.map(toReviewFinding);
 	const finalized = finalizeSubprocessOutput({
 		rawOutput,
 		exitCode,