@yandy0725/pi-memory 1.0.0 → 1.1.0

package/README.md

@@ -47,7 +47,8 @@ Create `memory.json` in the agent directory (`~/.pi/agent/memory.json`) or the p
   "dream": {
     "nudgeAfterSessions": 5,
     "nudgeAfterHours": 24,
-    "model": "auto"
+    "model": "auto",
+    "thinkLevel": "high"
   },
   "sessionSearch": {
     "maxSessions": 10,
@@ -56,6 +57,7 @@ Create `memory.json` in the agent directory (`~/.pi/agent/memory.json`) or the p
   "autoSurfacing": {
     "enabled": true,
     "model": "auto",
+    "thinkLevel": "off",
     "maxFiles": 5,
     "maxTopicBytes": 4096,
     "maxInjectionBytes": 20480
@@ -63,6 +65,7 @@ Create `memory.json` in the agent directory (`~/.pi/agent/memory.json`) or the p
   "extractMemories": {
     "enabled": true,
     "model": "auto",
+    "thinkLevel": "high",
     "maxContextTokens": 2000
   }
 }
@@ -77,15 +80,18 @@ Create `memory.json` in the agent directory (`~/.pi/agent/memory.json`) or the p
 | `dream.nudgeAfterSessions` | `5` | Sessions since last dream before nudge is shown |
 | `dream.nudgeAfterHours` | `24` | Hours since last dream before nudge is shown |
 | `dream.model` | `"auto"` | Model for dream consolidation (`"auto"` = same as current, or `"provider/id"`) |
+| `dream.thinkLevel` | `"high"` | Thinking effort for dream subagent: `"off"`, `"minimal"`, `"low"`, `"medium"`, `"high"`, `"xhigh"` |
 | `sessionSearch.maxSessions` | `10` | Max sessions to scan when searching history |
 | `sessionSearch.maxMatches` | `5` | Max matches to return from history search |
 | `autoSurfacing.enabled` | `true` | ⭐ Enable per-turn topic file auto-injection |
 | `autoSurfacing.model` | `"auto"` | ⭐ Model for side-query relevance selection |
+| `autoSurfacing.thinkLevel` | `"off"` | ⭐ Thinking effort for side-query (recommended: `"off"` for lightweight selection) |
 | `autoSurfacing.maxFiles` | `5` | ⭐ Max topic files to inject per turn |
 | `autoSurfacing.maxTopicBytes` | `4096` | ⭐ Max bytes per injected topic file (truncated) |
 | `autoSurfacing.maxInjectionBytes` | `20480` | ⭐ Max total bytes of injected content per turn |
 | `extractMemories.enabled` | `true` | ⭐ Enable per-turn memory extraction |
 | `extractMemories.model` | `"auto"` | ⭐ Model for the extraction subagent |
+| `extractMemories.thinkLevel` | `"high"` | ⭐ Thinking effort for extraction: `"off"` / `"minimal"` / `"low"` / `"medium"` / `"high"` / `"xhigh"` |
 | `extractMemories.maxContextTokens` | `2000` | ⭐ Max tokens of conversation to analyze |
 
 Project-level config (`.pi/memory.json`) is only loaded when the project is trusted.

package/README.zh.md

@@ -47,7 +47,8 @@ pi install npm:@yandy0725/pi-memory
   "dream": {
     "nudgeAfterSessions": 5,
     "nudgeAfterHours": 24,
-    "model": "auto"
+    "model": "auto",
+    "thinkLevel": "high"
   },
   "sessionSearch": {
     "maxSessions": 10,
@@ -56,6 +57,7 @@ pi install npm:@yandy0725/pi-memory
   "autoSurfacing": {
     "enabled": true,
     "model": "auto",
+    "thinkLevel": "off",
     "maxFiles": 5,
     "maxTopicBytes": 4096,
     "maxInjectionBytes": 20480
@@ -63,6 +65,7 @@ pi install npm:@yandy0725/pi-memory
   "extractMemories": {
     "enabled": true,
     "model": "auto",
+    "thinkLevel": "high",
     "maxContextTokens": 2000
   }
 }
@@ -77,15 +80,18 @@ pi install npm:@yandy0725/pi-memory
 | `dream.nudgeAfterSessions` | `5` | 触发提醒需经过的会话数 |
 | `dream.nudgeAfterHours` | `24` | 触发提醒需经过的小时数 |
 | `dream.model` | `"auto"` | 整理使用的模型（`"auto"` = 当前模型，或 `"provider/id"`） |
+| `dream.thinkLevel` | `"high"` | 整理子 agent 的思考深度：`"off"` / `"minimal"` / `"low"` / `"medium"` / `"high"` / `"xhigh"` |
 | `sessionSearch.maxSessions` | `10` | 搜索历史时最多扫描的会话数 |
 | `sessionSearch.maxMatches` | `5` | 历史搜索最多返回的匹配数 |
 | `autoSurfacing.enabled` | `true` | ⭐ 启用 per-turn topic 文件自动注入 |
 | `autoSurfacing.model` | `"auto"` | ⭐ side-query 相关性选择模型 |
+| `autoSurfacing.thinkLevel` | `"off"` | ⭐ side-query 思考深度（推荐 `"off"`，轻量选择任务） |
 | `autoSurfacing.maxFiles` | `5` | ⭐ 每轮最多注入的 topic 文件数 |
 | `autoSurfacing.maxTopicBytes` | `4096` | ⭐ 单个注入 topic 文件最大字节数（截断） |
 | `autoSurfacing.maxInjectionBytes` | `20480` | ⭐ 每轮注入内容总字节数上限 |
 | `extractMemories.enabled` | `true` | ⭐ 启用 per-turn 记忆自动提取 |
 | `extractMemories.model` | `"auto"` | ⭐ 提取子 agent 使用的模型 |
+| `extractMemories.thinkLevel` | `"high"` | ⭐ 提取的思考深度：`"off"` / `"minimal"` / `"low"` / `"medium"` / `"high"` / `"xhigh"` |
 | `extractMemories.maxContextTokens` | `2000` | ⭐ 分析对话的最大 token 数 |
 
 项目级配置（`.pi/memory.json`）仅在项目受信任时加载。

package/index.ts

@@ -61,12 +61,14 @@ export default function (pi: ExtensionAPI) {
 					// session_start handlers (including pi-subagents') have completed.
 					// Does not block session_start.
 					const dreamModel = config.dream.model;
+					const dreamThinkLevel = config.dream.thinkLevel;
 					const dir = memoryDir;
 					ctx.ui.setStatus("dream", "Consolidating memory...");
 					setTimeout(async () => {
 						try {
 							const summary = await runDream({
 								model: dreamModel,
+								thinkLevel: dreamThinkLevel,
 								memoryDir: dir,
 								events: pi.events,
 							});
@@ -87,17 +89,23 @@ export default function (pi: ExtensionAPI) {
 	pi.on("before_agent_start", async (event, ctx) => {
 		if (!config?.enabled || !indexSnapshot || !memoryDir) return;
 
-		// Auto-surfacing: select relevant topic files via LLM side-query and inject as message
+		// Auto-surfacing: select relevant topic files via LLM side-query and inject as message.
+		// Skip for subagents: pi-subagents strips "subagent" from all children's
+		// tool sets, so its absence reliably identifies subagents. Without this guard,
+		// runSideQuery's subagent spawn → before_agent_start → re-enter here → OOM.
+		const agentTools = event.systemPromptOptions?.selectedTools;
+		const isSubagent = agentTools && !agentTools.includes("subagent");
+
 		const autoSurfacing = config.autoSurfacing;
 		// biome-ignore lint/suspicious/noExplicitAny: message injection result
 		let injectedMessage: any;
-		if (autoSurfacing?.enabled && event.prompt) {
+		if (autoSurfacing?.enabled && event.prompt && !isSubagent) {
 			try {
 				if (ctx.hasUI) ctx.ui.setStatus("surfacing", "Searching relevant memories…");
 				const manifest = await scanTopics(memoryDir);
 				if (manifest.length > 0) {
 					const queryPrompt = buildSurfacingPrompt(manifest, event.prompt.slice(0, 4000), injectedTopics);
-					const selected = await runSideQuery(queryPrompt, manifest, autoSurfacing.maxFiles, pi.events);
+					const selected = await runSideQuery(queryPrompt, manifest, autoSurfacing.maxFiles, autoSurfacing.thinkLevel, pi.events);
 					if (selected.length > 0) {
 						const content = await injectSurfacedContent(
 							memoryDir,
@@ -135,6 +143,7 @@ export default function (pi: ExtensionAPI) {
 		// Fire-and-forget: extract memories in background
 		runExtract({
 			model: extractConfig.model,
+			thinkLevel: extractConfig.thinkLevel,
 			memoryDir,
 			messages: event.messages.map((m) => ({
 				// biome-ignore lint/suspicious/noExplicitAny: pi event message union type
@@ -197,6 +206,7 @@ export default function (pi: ExtensionAPI) {
 			try {
 				const summary = await runDream({
 					model: config.dream.model,
+					thinkLevel: config.dream.thinkLevel,
 					memoryDir,
 					signal: ctx.signal,
 					events: pi.events,

package/package.json

@@ -3,7 +3,7 @@
 	"publishConfig": {
 		"access": "public"
 	},
-	"version": "1.0.0",
+	"version": "1.1.0",
 	"description": "File-system driven persistent memory layer for pi coding agent",
 	"license": "MIT",
 	"repository": {

package/src/config.ts

@@ -3,9 +3,12 @@ import { homedir } from "node:os";
 import { join } from "node:path";
 import { CONFIG_DIR_NAME, getAgentDir } from "@earendil-works/pi-coding-agent";
 
+export type ThinkLevel = "off" | "minimal" | "low" | "medium" | "high" | "xhigh";
+
 export interface AutoSurfacingConfig {
 	enabled: boolean;
 	model: string;
+	thinkLevel: ThinkLevel;
 	maxFiles: number;
 	maxTopicBytes: number;
 	maxInjectionBytes: number;
@@ -14,6 +17,7 @@ export interface AutoSurfacingConfig {
 export interface ExtractMemoriesConfig {
 	enabled: boolean;
 	model: string;
+	thinkLevel: ThinkLevel;
 	maxContextTokens: number;
 }
 
@@ -22,7 +26,7 @@ export interface MemoryConfig {
 	memoryDir: string;
 	memIndexMaxLines: number;
 	memIndexMaxBytes: number;
-	dream: { nudgeAfterSessions: number; nudgeAfterHours: number; model: string };
+	dream: { nudgeAfterSessions: number; nudgeAfterHours: number; model: string; thinkLevel: ThinkLevel };
 	sessionSearch: { maxSessions: number; maxMatches: number };
 	autoSurfacing: AutoSurfacingConfig;
 	extractMemories: ExtractMemoriesConfig;
@@ -33,11 +37,12 @@ export const DEFAULT_CONFIG: MemoryConfig = {
 	memoryDir: join(homedir(), CONFIG_DIR_NAME, "memory"),
 	memIndexMaxLines: 200,
 	memIndexMaxBytes: 25600,
-	dream: { nudgeAfterSessions: 5, nudgeAfterHours: 24, model: "auto" },
+	dream: { nudgeAfterSessions: 5, nudgeAfterHours: 24, model: "auto", thinkLevel: "high" },
 	sessionSearch: { maxSessions: 10, maxMatches: 5 },
 	autoSurfacing: {
 		enabled: true,
 		model: "auto",
+		thinkLevel: "off",
 		maxFiles: 5,
 		maxTopicBytes: 4096,
 		maxInjectionBytes: 20480,
@@ -45,6 +50,7 @@ export const DEFAULT_CONFIG: MemoryConfig = {
 	extractMemories: {
 		enabled: true,
 		model: "auto",
+		thinkLevel: "high",
 		maxContextTokens: 2000,
 	},
 };

package/src/dream.ts

@@ -1,4 +1,5 @@
 import { access } from "node:fs/promises";
+import type { ThinkLevel } from "./config";
 import { getSubagentsService, type SubagentsService, type WorkspaceProvider } from "@yandy0725/pi-subagents";
 
 /** Build dream consolidation task.
@@ -53,6 +54,7 @@ CRITICAL for hooks and descriptions:
 
 export interface RunDreamOpts {
 	model: string;
+	thinkLevel: ThinkLevel;
 	memoryDir: string;
 	signal?: AbortSignal;
 	// biome-ignore lint/suspicious/noExplicitAny: pi events API handler
@@ -88,7 +90,7 @@ export async function runDream(opts: RunDreamOpts): Promise<string> {
 	const agentId = service.spawn(
 		"memory-agent",
 		task,
-		model ? { model, thinkingLevel: "high" } : { thinkingLevel: "high" },
+		model ? { model, thinkingLevel: opts.thinkLevel } : { thinkingLevel: opts.thinkLevel },
 	);
 
 	// Wait for completion/failure via pi-subagents events (abort-while-queued

package/src/extract.ts

@@ -1,8 +1,10 @@
 import { access } from "node:fs/promises";
+import type { ThinkLevel } from "./config";
 import { getSubagentsService, type SubagentsService, type WorkspaceProvider } from "@yandy0725/pi-subagents";
 
 export interface RunExtractOpts {
 	model: string;
+	thinkLevel: ThinkLevel;
 	memoryDir: string;
 	messages: Array<{ role: string; content: string }>;
 	maxContextTokens: number;
@@ -84,6 +86,7 @@ export async function runExtract(opts: RunExtractOpts): Promise<void> {
 	if (!service) return; // silently skip if no subagent service
 
 	const model = opts.model === "auto" ? undefined : opts.model;
+	const thinkLevel = opts.thinkLevel;
 	const task = buildExtractTask(opts.memoryDir, opts.messages, opts.maxContextTokens);
 
 	const provider: WorkspaceProvider = {
@@ -104,7 +107,7 @@ export async function runExtract(opts: RunExtractOpts): Promise<void> {
 		"memory-agent",
 		task,
 		model
-			? { model, inheritContext: false, maxTurns: 5, thinkingLevel: "high" }
-			: { inheritContext: false, maxTurns: 5, thinkingLevel: "high" },
+			? { model, inheritContext: false, maxTurns: 5, thinkingLevel: thinkLevel }
+			: { inheritContext: false, maxTurns: 5, thinkingLevel: thinkLevel },
 	);
 }

package/src/inject.ts

@@ -117,10 +117,13 @@ export async function injectSurfacedContent(
 /** Run a lightweight LLM side-query to select relevant topic files.
  *  Uses pi-subagents spawn + Promise pattern (same as dream.ts).
  *  Falls back to keyword matching if subagent service is unavailable. */
+import type { ThinkLevel } from "./config";
+
 export async function runSideQuery(
 	prompt: string,
 	manifest: TopicManifest[],
 	maxFiles: number,
+	thinkLevel: ThinkLevel,
 	// biome-ignore lint/suspicious/noExplicitAny: pi events API handler
 	events?: { on(channel: string, handler: (data: any) => void): () => void },
 ): Promise<string[]> {
@@ -219,7 +222,7 @@ export async function runSideQuery(
 			const agentId = service.spawn("memory-agent", task, {
 				maxTurns: 1,
 				inheritContext: false,
-				thinkingLevel: "off",
+				thinkingLevel: thinkLevel,
 				foreground: true,
 				bypassQueue: true,
 			});