@yandy0725/pi-memory 1.2.0 → 1.3.0

package/README.md

@@ -44,11 +44,16 @@ Create `memory.json` in the agent directory (`~/.pi/agent/memory.json`) or the p
   "memoryDir": "~/.pi/memory",
   "memIndexMaxLines": 200,
   "memIndexMaxBytes": 25600,
+  "defaults": {
+    "model": "deepseek/deepseek-v4-flash",
+    "sessionPersistence": { "enabled": false }
+  },
   "dream": {
     "nudgeAfterSessions": 5,
     "nudgeAfterHours": 24,
     "model": "auto",
-    "thinkLevel": "high"
+    "thinkLevel": "high",
+    "sessionPersistence": { "enabled": false }
   },
   "sessionSearch": {
     "maxSessions": 10,
@@ -60,13 +65,15 @@ Create `memory.json` in the agent directory (`~/.pi/agent/memory.json`) or the p
     "thinkLevel": "off",
     "maxFiles": 5,
     "maxTopicBytes": 4096,
-    "maxInjectionBytes": 20480
+    "maxInjectionBytes": 20480,
+    "sessionPersistence": { "enabled": false }
   },
   "extractMemories": {
     "enabled": true,
     "model": "auto",
     "thinkLevel": "high",
-    "maxContextTokens": 2000
+    "maxContextTokens": 2000,
+    "sessionPersistence": { "enabled": false }
   }
 }
 ```
@@ -77,21 +84,28 @@ Create `memory.json` in the agent directory (`~/.pi/agent/memory.json`) or the p
 | `memoryDir` | `~/.pi/memory` | Root directory for all memory data |
 | `memIndexMaxLines` | `200` | Max lines in `MEMORY.md` before capacity errors |
 | `memIndexMaxBytes` | `25600` | Max bytes in `MEMORY.md` before capacity errors |
+| `defaults.model` | — | Shared model fallback for all sub-tasks. Per-task `model` overrides |
+| `defaults.sessionPersistence.enabled` | `false` | Shared session persistence fallback (default: in-memory). Per-task overrides |
+| `defaults.sessionPersistence.sessionDir` | `memoryDir/sessions/` | Custom session directory for persisted headless agent sessions |
 | `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.model` | — | Model for dream consolidation (`"provider/id"`). Falls back to `defaults.model` → parent model |
 | `dream.thinkLevel` | `"high"` | Thinking effort for dream subagent: `"off"`, `"minimal"`, `"low"`, `"medium"`, `"high"`, `"xhigh"` |
+| `dream.sessionPersistence.enabled` | `false` | Persist dream agent sessions to disk (debug/audit). Falls back to `defaults.sessionPersistence.enabled` |
+| `dream.sessionPersistence.sessionDir` | `memoryDir/sessions/` | Custom session directory for dream sessions |
 | `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.model` | — | ⭐ Model for side-query relevance selection. Falls back to `defaults.model` → parent model |
 | `autoSurfacing.thinkLevel` | `"off"` | ⭐ Thinking effort for side-query (recommended: `"off"` for lightweight selection) |
+| `autoSurfacing.sessionPersistence.enabled` | `false` | Persist side-query agent sessions to disk. Falls back to `defaults.sessionPersistence.enabled` |
 | `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.model` | — | ⭐ Model for the extraction subagent. Falls back to `defaults.model` → parent model |
 | `extractMemories.thinkLevel` | `"high"` | ⭐ Thinking effort for extraction: `"off"` / `"minimal"` / `"low"` / `"medium"` / `"high"` / `"xhigh"` |
+| `extractMemories.sessionPersistence.enabled` | `false` | Persist extract agent sessions to disk. Falls back to `defaults.sessionPersistence.enabled` |
 | `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

@@ -44,11 +44,16 @@ pi install npm:@yandy0725/pi-memory
   "memoryDir": "~/.pi/memory",
   "memIndexMaxLines": 200,
   "memIndexMaxBytes": 25600,
+  "defaults": {
+    "model": "deepseek/deepseek-v4-flash",
+    "sessionPersistence": { "enabled": false }
+  },
   "dream": {
     "nudgeAfterSessions": 5,
     "nudgeAfterHours": 24,
     "model": "auto",
-    "thinkLevel": "high"
+    "thinkLevel": "high",
+    "sessionPersistence": { "enabled": false }
   },
   "sessionSearch": {
     "maxSessions": 10,
@@ -60,13 +65,15 @@ pi install npm:@yandy0725/pi-memory
     "thinkLevel": "off",
     "maxFiles": 5,
     "maxTopicBytes": 4096,
-    "maxInjectionBytes": 20480
+    "maxInjectionBytes": 20480,
+    "sessionPersistence": { "enabled": false }
   },
   "extractMemories": {
     "enabled": true,
     "model": "auto",
     "thinkLevel": "high",
-    "maxContextTokens": 2000
+    "maxContextTokens": 2000,
+    "sessionPersistence": { "enabled": false }
   }
 }
 ```
@@ -77,21 +84,28 @@ pi install npm:@yandy0725/pi-memory
 | `memoryDir` | `~/.pi/memory` | 所有记忆数据的根目录 |
 | `memIndexMaxLines` | `200` | `MEMORY.md` 最大行数 |
 | `memIndexMaxBytes` | `25600` | `MEMORY.md` 最大字节数 |
+| `defaults.model` | — | 所有子任务共享的模型回退值。per-task `model` 会覆盖 |
+| `defaults.sessionPersistence.enabled` | `false` | 共享的 session 持久化回退（默认不持久化）。per-task 可覆盖 |
+| `defaults.sessionPersistence.sessionDir` | `memoryDir/sessions/` | 自定义持久化目录 |
 | `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.model` | — | 整理使用的模型（`"provider/id"`）。回退链：per-task → `defaults.model` → 父模型 |
 | `dream.thinkLevel` | `"high"` | 整理子 agent 的思考深度：`"off"` / `"minimal"` / `"low"` / `"medium"` / `"high"` / `"xhigh"` |
+| `dream.sessionPersistence.enabled` | `false` | 持久化 dream agent 的 session 到磁盘（调试/审计）。回退到 `defaults.sessionPersistence.enabled` |
+| `dream.sessionPersistence.sessionDir` | `memoryDir/sessions/` | dream session 自定义目录 |
 | `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.model` | — | ⭐ side-query 相关性选择模型。回退链：per-task → `defaults.model` → 父模型 |
 | `autoSurfacing.thinkLevel` | `"off"` | ⭐ side-query 思考深度（推荐 `"off"`，轻量选择任务） |
+| `autoSurfacing.sessionPersistence.enabled` | `false` | 持久化 side-query agent 的 session。回退到 `defaults.sessionPersistence.enabled` |
 | `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.model` | — | ⭐ 提取子 agent 使用的模型。回退链：per-task → `defaults.model` → 父模型 |
 | `extractMemories.thinkLevel` | `"high"` | ⭐ 提取的思考深度：`"off"` / `"minimal"` / `"low"` / `"medium"` / `"high"` / `"xhigh"` |
+| `extractMemories.sessionPersistence.enabled` | `false` | 持久化 extract agent 的 session。回退到 `defaults.sessionPersistence.enabled` |
 | `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

@@ -2,12 +2,11 @@ import { readdir, readFile } from "node:fs/promises";
 import { join } from "node:path";
 import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
 import { SessionManager } from "@earendil-works/pi-coding-agent";
-import { loadConfig, type MemoryConfig } from "./src/config";
+import { loadConfig, type MemoryConfig, type SessionPersistenceConfig } from "./src/config";
 import { runDream } from "./src/dream";
 import { runExtract } from "./src/extract";
 import {
 	buildInjection,
-	buildSurfacingPrompt,
 	injectSurfacedContent,
 	loadIndexSnapshot,
 	runSideQuery,
@@ -18,6 +17,14 @@ import { readDreamMeta, shouldNudge, writeDreamMeta } from "./src/nudge";
 import { resolveMemoryDir } from "./src/paths";
 import { searchSessions } from "./src/session-search";
 
+function resolveDefault(cfg: MemoryConfig, task: "dream" | "autoSurfacing" | "extractMemories", key: "model"): string | undefined;
+function resolveDefault(cfg: MemoryConfig, task: "dream" | "autoSurfacing" | "extractMemories", key: "sessionPersistence"): SessionPersistenceConfig | undefined;
+function resolveDefault(cfg: MemoryConfig, task: "dream" | "autoSurfacing" | "extractMemories", key: "model" | "sessionPersistence"): string | SessionPersistenceConfig | undefined {
+	const perTask = cfg[task][key];
+	if (perTask !== undefined) return perTask;
+	return cfg.defaults?.[key];
+}
+
 export default function (pi: ExtensionAPI) {
 	let memoryDir: string | null = null;
 	let config: MemoryConfig | null = null;
@@ -55,7 +62,7 @@ export default function (pi: ExtensionAPI) {
 				if (ok) {
 					// Fire-and-forget: does not block session_start. The headless
 					// dream agent runs independently; completion notifies the user.
-					const dreamModel = config.dream.model;
+					const dreamModel = resolveDefault(config, "dream", "model");
 					const dreamThinkLevel = config.dream.thinkLevel;
 					const dir = memoryDir;
 					ctx.ui.setStatus("dream", "Consolidating memory...");
@@ -65,6 +72,7 @@ export default function (pi: ExtensionAPI) {
 						memoryDir: dir,
 						modelRegistry: ctx.modelRegistry,
 						parentModel: ctx.model,
+						sessionPersistence: resolveDefault(config, "dream", "sessionPersistence"),
 					})
 						.then(async (summary) => {
 							await writeDreamMeta(dir, sessions);
@@ -98,17 +106,17 @@ export default function (pi: ExtensionAPI) {
 				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,
+						event.prompt.slice(0, 4000),
+						injectedTopics,
 						autoSurfacing.maxFiles,
 						autoSurfacing.thinkLevel,
-						autoSurfacing.model,
+						resolveDefault(config, "autoSurfacing", "model"),
 						ctx.modelRegistry,
 						ctx.model,
 						memoryDir,
-						injectedTopics,
+						resolveDefault(config, "autoSurfacing", "sessionPersistence"),
 					);
 					if (selected.length > 0) {
 						const content = await injectSurfacedContent(
@@ -143,11 +151,12 @@ export default function (pi: ExtensionAPI) {
 		if (!extractConfig?.enabled) return;
 		if (!event.messages || event.messages.length === 0) return;
 		runExtract({
-			model: extractConfig.model,
+			model: resolveDefault(config, "extractMemories", "model"),
 			thinkLevel: extractConfig.thinkLevel,
 			memoryDir,
 			modelRegistry: ctx.modelRegistry,
 			parentModel: ctx.model,
+			sessionPersistence: resolveDefault(config, "extractMemories", "sessionPersistence"),
 			messages: event.messages.map((m) => ({
 				// biome-ignore lint/suspicious/noExplicitAny: pi event message union type
 				role: String((m as any).role ?? ""),
@@ -208,11 +217,12 @@ export default function (pi: ExtensionAPI) {
 			const dir = memoryDir;
 			ctx.ui.setStatus("dream", "Consolidating memory...");
 			runDream({
-				model: config.dream.model,
+				model: resolveDefault(config, "dream", "model"),
 				thinkLevel: config.dream.thinkLevel,
 				memoryDir,
 				modelRegistry: ctx.modelRegistry,
 				parentModel: ctx.model,
+				sessionPersistence: resolveDefault(config, "dream", "sessionPersistence"),
 			})
 				.then(async (summary) => {
 					const sessions = (await SessionManager.list(ctx.cwd)).length;

package/package.json

@@ -3,7 +3,7 @@
 	"publishConfig": {
 		"access": "public"
 	},
-	"version": "1.2.0",
+	"version": "1.3.0",
 	"description": "File-system driven persistent memory layer for pi coding agent",
 	"license": "MIT",
 	"repository": {

package/src/agent-runner.ts

@@ -1,4 +1,5 @@
 import type { Model } from "@earendil-works/pi-ai";
+import { join } from "node:path";
 import {
 	type AgentSession,
 	type AgentSessionEvent,
@@ -9,7 +10,7 @@ import {
 	SettingsManager,
 } from "@earendil-works/pi-coding-agent";
 import { MEMORY_AGENT_TOOLS } from "./agent-config";
-import type { ThinkLevel } from "./config";
+import type { SessionPersistenceConfig, ThinkLevel } from "./config";
 import { resolveModel } from "./model-resolver";
 
 export interface HeadlessAgentOpts {
@@ -22,13 +23,16 @@ export interface HeadlessAgentOpts {
 	maxTurns?: number;
 	signal?: AbortSignal;
 	timeoutMs?: number;
+	/** Session persistence config. When enabled, sessions are written to disk. */
+	sessionPersistence?: SessionPersistenceConfig;
 }
 
 const GRACE_TURNS = 1;
 
 /**
- * Run a headless memory-agent sub-session: create an in-memory, resource-free
- * session, drive the turn loop, collect the assistant response text, and dispose.
+ * Run a headless memory-agent sub-session: create a session (in-memory by default,
+ * persisted to disk when sessionPersistence.enabled is true), drive the turn loop,
+ * collect the assistant response text, and dispose.
  *
  * Does NOT call bindExtensions — no extension hooks fire in the sub-session,
  * so pi-memory's own before_agent_start cannot recurse.
@@ -53,25 +57,31 @@ export async function runHeadlessAgent(opts: HeadlessAgentOpts): Promise<string>
 	});
 	await loader.reload();
 
-	// 3. Forward abort signal BEFORE createAgentSession (handle early abort)
-	let session: AgentSession | undefined;
-	const onAbort = (): void => {
-		void session?.abort();
-	};
-	opts.signal?.addEventListener("abort", onAbort, { once: true });
+	// 3. Create session (in-memory or persisted based on config)
+	const sessionManager = opts.sessionPersistence?.enabled
+		? SessionManager.create(
+				opts.cwd,
+				opts.sessionPersistence.sessionDir ?? join(opts.cwd, "sessions"),
+			)
+		: SessionManager.inMemory(opts.cwd);
 
-	// 4. Create the in-memory session (no bindExtensions)
 	const created = await createAgentSession({
 		cwd: opts.cwd,
 		tools: [...MEMORY_AGENT_TOOLS],
 		model: resolvedModel as any,
 		thinkingLevel: opts.thinkLevel as any,
 		modelRegistry: opts.modelRegistry,
-		sessionManager: SessionManager.inMemory(opts.cwd),
+		sessionManager,
 		settingsManager,
 		resourceLoader: loader,
 	});
-	session = created.session as AgentSession;
+
+	// 4. Forward abort signal after session exists (avoid listener leak if creation throws)
+	let session: AgentSession | undefined = created.session as AgentSession;
+	const onAbort = (): void => {
+		void session?.abort();
+	};
+	opts.signal?.addEventListener("abort", onAbort, { once: true });
 
 	// 5. Collect response text + enforce turn limits
 	let text = "";

package/src/config.ts

@@ -5,6 +5,20 @@ import { CONFIG_DIR_NAME, getAgentDir } from "@earendil-works/pi-coding-agent";
 
 export type ThinkLevel = "off" | "minimal" | "low" | "medium" | "high" | "xhigh";
 
+/** Session persistence configuration for headless memory-agent sub-sessions. */
+export interface SessionPersistenceConfig {
+	/** Enable disk persistence (default: false = in-memory). */
+	enabled: boolean;
+	/** Custom session directory. Defaults to memoryDir/sessions/. */
+	sessionDir?: string;
+}
+
+/** Shared defaults that per-task configs inherit. Per-task fields override these. */
+export interface DefaultsConfig {
+	model?: string;
+	sessionPersistence?: SessionPersistenceConfig;
+}
+
 export interface AutoSurfacingConfig {
 	enabled: boolean;
 	model?: string;
@@ -23,13 +37,25 @@ export interface ExtractMemoriesConfig {
 
 export interface MemoryConfig {
 	enabled: boolean;
+	/** Shared defaults for model and sessionPersistence. Per-task configs override. */
+	defaults?: DefaultsConfig;
 	memoryDir: string;
 	memIndexMaxLines: number;
 	memIndexMaxBytes: number;
-	dream: { nudgeAfterSessions: number; nudgeAfterHours: number; model?: string; thinkLevel: ThinkLevel };
+	dream: {
+		nudgeAfterSessions: number;
+		nudgeAfterHours: number;
+		model?: string;
+		thinkLevel: ThinkLevel;
+		sessionPersistence?: SessionPersistenceConfig;
+	};
 	sessionSearch: { maxSessions: number; maxMatches: number };
-	autoSurfacing: AutoSurfacingConfig;
-	extractMemories: ExtractMemoriesConfig;
+	autoSurfacing: AutoSurfacingConfig & {
+		sessionPersistence?: SessionPersistenceConfig;
+	};
+	extractMemories: ExtractMemoriesConfig & {
+		sessionPersistence?: SessionPersistenceConfig;
+	};
 }
 
 export const DEFAULT_CONFIG: MemoryConfig = {

package/src/dream.ts

@@ -1,7 +1,7 @@
 import type { Model } from "@earendil-works/pi-ai";
 import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
 import { runHeadlessAgent } from "./agent-runner";
-import type { ThinkLevel } from "./config";
+import type { SessionPersistenceConfig, ThinkLevel } from "./config";
 
 /** Build dream consolidation task. (unchanged) */
 export function buildDreamTask(memoryDir: string, maxLines: number): string {
@@ -56,6 +56,7 @@ export interface RunDreamOpts {
 	memoryDir: string;
 	modelRegistry: ModelRegistry;
 	parentModel?: Model<any>;
+	sessionPersistence?: SessionPersistenceConfig;
 }
 
 /** Run dream consolidation via a headless agent. Fire-and-forget (caller decides await). */
@@ -70,5 +71,6 @@ export async function runDream(opts: RunDreamOpts): Promise<string> {
 		thinkLevel: opts.thinkLevel,
 		maxTurns: undefined,
 		timeoutMs: 600_000,
+		sessionPersistence: opts.sessionPersistence,
 	});
 }

package/src/extract.ts

@@ -1,7 +1,7 @@
 import type { Model } from "@earendil-works/pi-ai";
 import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
 import { runHeadlessAgent } from "./agent-runner";
-import type { ThinkLevel } from "./config";
+import type { SessionPersistenceConfig, ThinkLevel } from "./config";
 
 export interface RunExtractOpts {
 	model?: string;
@@ -11,6 +11,7 @@ export interface RunExtractOpts {
 	maxContextTokens: number;
 	modelRegistry: ModelRegistry;
 	parentModel?: Model<any>;
+	sessionPersistence?: SessionPersistenceConfig;
 }
 
 /** Build extraction task prompt. (unchanged) */
@@ -90,6 +91,7 @@ export async function runExtract(opts: RunExtractOpts): Promise<void> {
 		thinkLevel: opts.thinkLevel,
 		maxTurns: 5,
 		timeoutMs: 120_000,
+		sessionPersistence: opts.sessionPersistence,
 	}).catch(() => {
 		/* silently ignore extract errors */
 	});

package/src/inject.ts

@@ -3,7 +3,7 @@ import { join } from "node:path";
 import type { Model } from "@earendil-works/pi-ai";
 import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
 import { runHeadlessAgent } from "./agent-runner";
-import type { ThinkLevel } from "./config";
+import type { SessionPersistenceConfig, ThinkLevel } from "./config";
 import { truncateForInjection } from "./index-file";
 import { parseFrontmatter } from "./topic-file";
 
@@ -60,25 +60,15 @@ export async function scanTopics(memoryDir: string): Promise<TopicManifest[]> {
 	return manifests.sort((a, b) => b.mtimeMs - a.mtimeMs);
 }
 
-export function buildSurfacingPrompt(
-	manifest: TopicManifest[],
-	userPrompt: string,
-	injectedTopics: Set<string>,
-): string {
+function buildSurfacingPrompt(manifest: TopicManifest[], userPrompt: string): string {
 	const lines = manifest.map((t) => {
-		const marker = injectedTopics.has(t.filename) ? " [already injected]" : "";
-		return `[${t.type}] ${t.filename} — ${t.description.slice(0, 80)}${marker}`;
+		return `[${t.type}] ${t.filename} — ${t.description.slice(0, 80)}`;
 	});
-	const alreadyInjected =
-		injectedTopics.size > 0
-			? `\n\nNote: ${injectedTopics.size} topic file(s) have already been injected in this session and are marked [already injected]. Prefer selecting uninjected topics.`
-			: "";
 
 	return [
 		"You are a memory relevance selector. Below is a list of memory topic files and a user message.",
 		"Select up to N topic files that are relevant to the user's current query.",
 		"Response format: JSON with a 'selected_files' array of filenames.",
-		alreadyInjected,
 		"",
 		"=== Topic Files ===",
 		...lines,
@@ -148,19 +138,21 @@ function parseSelectedFiles(result: string, candidates: TopicManifest[], maxFile
 /** Run a lightweight headless side-query to select relevant topic files.
  *  Returns [] on timeout/failure — no fallback. */
 export async function runSideQuery(
-	prompt: string,
 	manifest: TopicManifest[],
+	userPrompt: string,
+	injectedTopics: Set<string>,
 	maxFiles: number,
 	thinkLevel: ThinkLevel,
 	model: string | undefined,
 	modelRegistry: ModelRegistry,
 	parentModel: Model<any> | undefined,
 	memoryDir: string,
-	injectedTopics: Set<string>,
+	sessionPersistence?: SessionPersistenceConfig,
 ): Promise<string[]> {
 	const candidates = manifest.filter((t) => !injectedTopics.has(t.filename));
 	if (candidates.length === 0) return [];
-	const task = buildSideQueryTask(prompt, maxFiles);
+	const surfacingPrompt = buildSurfacingPrompt(candidates, userPrompt);
+	const task = buildSideQueryTask(surfacingPrompt, maxFiles);
 	try {
 		const result = await runHeadlessAgent({
 			task,
@@ -171,6 +163,7 @@ export async function runSideQuery(
 			thinkLevel,
 			maxTurns: 1,
 			timeoutMs: 30_000,
+			sessionPersistence,
 		});
 		return parseSelectedFiles(result, candidates, maxFiles);
 	} catch {