fix: filter subagent runtime wrappers from auto-capture

src/auto-capture-cleanup.ts

@@ -11,6 +11,10 @@ const AUTO_CAPTURE_SESSION_RESET_PREFIX =
   "A new session was started via /new or /reset. Execute your Session Startup sequence now";
 const AUTO_CAPTURE_ADDRESSING_PREFIX_RE = /^(?:<@!?[0-9]+>|@[A-Za-z0-9_.-]+)\s*/;
 const AUTO_CAPTURE_SYSTEM_EVENT_LINE_RE = /^System:\s*\[[^\n]*?\]\s*Exec\s+(?:completed|failed|started)\b.*$/gim;
+const AUTO_CAPTURE_RUNTIME_WRAPPER_LINE_RE = /^\[(?:Subagent Context|Subagent Task)\]\s*/i;
+const AUTO_CAPTURE_RUNTIME_WRAPPER_PREFIX_RE = /^\[(?:Subagent Context|Subagent Task)\]/i;
+const AUTO_CAPTURE_RUNTIME_WRAPPER_BOILERPLATE_RE =
+  /(?:You are running as a subagent\b.*?(?:$|(?<=\.)\s+)|Results auto-announce to your requester\.?\s*|do not busy-poll for status\.?\s*|Reply with a brief acknowledgment only\.?\s*|Do not use any memory tools\.?\s*)/gi;
 
 function escapeRegExp(value: string): string {
   return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
@@ -62,6 +66,60 @@ function stripAutoCaptureAddressingPrefix(text: string): string {
   return text.replace(AUTO_CAPTURE_ADDRESSING_PREFIX_RE, "").trim();
 }
 
+function stripRuntimeWrapperBoilerplate(text: string): string {
+  return text
+    .replace(AUTO_CAPTURE_RUNTIME_WRAPPER_BOILERPLATE_RE, "")
+    .replace(/\s{2,}/g, " ")
+    .trim();
+}
+
+function stripRuntimeWrapperLine(line: string): string {
+  const trimmed = line.trim();
+  if (!AUTO_CAPTURE_RUNTIME_WRAPPER_PREFIX_RE.test(trimmed)) {
+    return line;
+  }
+
+  const remainder = trimmed.replace(AUTO_CAPTURE_RUNTIME_WRAPPER_LINE_RE, "").trim();
+  if (!remainder) {
+    return "";
+  }
+
+  return stripRuntimeWrapperBoilerplate(remainder);
+}
+
+function stripLeadingRuntimeWrappers(text: string): string {
+  const trimmed = text.trim();
+  if (!trimmed) {
+    return trimmed;
+  }
+
+  const lines = trimmed.split("\n");
+  const cleanedLines: string[] = [];
+  let strippingLeadIn = true;
+
+  for (const line of lines) {
+    const current = line.trim();
+
+    if (strippingLeadIn && current === "") {
+      continue;
+    }
+
+    if (strippingLeadIn && AUTO_CAPTURE_RUNTIME_WRAPPER_PREFIX_RE.test(current)) {
+      const cleaned = stripRuntimeWrapperLine(current);
+      if (cleaned) {
+        cleanedLines.push(cleaned);
+        strippingLeadIn = false;
+      }
+      continue;
+    }
+
+    strippingLeadIn = false;
+    cleanedLines.push(line);
+  }
+
+  return cleanedLines.join("\n").trim();
+}
+
 export function stripAutoCaptureInjectedPrefix(role: string, text: string): string {
   if (role !== "user") {
     return text.trim();
@@ -76,6 +134,7 @@ export function stripAutoCaptureInjectedPrefix(role: string, text: string): stri
   normalized = stripAutoCaptureSessionResetPrefix(normalized);
   normalized = stripLeadingInboundMetadata(normalized);
   normalized = stripAutoCaptureAddressingPrefix(normalized);
+  normalized = stripLeadingRuntimeWrappers(normalized);
   normalized = stripLeadingInboundMetadata(normalized);
   normalized = normalized.replace(/\n{3,}/g, "\n\n");
   return normalized.trim();

src/extraction-prompts.ts

@@ -32,6 +32,7 @@ ${conversationText}
 - Temporary information: One-time questions or conversations
 - Vague information: "User has questions about a feature" (no specific details)
 - Tool output, error logs, or boilerplate
+- Runtime scaffolding or orchestration wrappers such as "[Subagent Context]", "[Subagent Task]", bootstrap wrappers, task envelopes, or agent instructions — these are execution metadata, NEVER store them as memories
 - Recall queries / meta-questions: "Do you remember X?", "你还记得X吗?", "你知道我喜欢什么吗" — these are retrieval requests, NOT new information to store
 - Degraded or incomplete references: If the user mentions something vaguely ("that thing I said"), do NOT invent details or create a hollow memory
 

src/smart-extractor.ts

@@ -70,8 +70,19 @@ import { batchDedup } from "./batch-dedup.js";
  * - Standalone JSON blocks containing message_id/sender_id fields
  */
 export function stripEnvelopeMetadata(text: string): string {
-  // 1. Strip "System: [timestamp] Channel..." lines
+  // 0. Strip runtime orchestration wrappers that should never become memories
+  //    (sub-agent task scaffolding is execution metadata, not conversation content).
   let cleaned = text.replace(
+    /^\[(?:Subagent Context|Subagent Task)\]\s*(?:You are running as a subagent.*?(?:$|(?<=\.)\s+)|Results auto-announce to your requester\.?\s*|do not busy-poll for status\.?\s*|Reply with a brief acknowledgment only\.?\s*|Do not use any memory tools\.?\s*)?/gim,
+    "",
+  );
+  cleaned = cleaned.replace(
+    /^(?:Results auto-announce to your requester\.?|do not busy-poll for status\.?|Do not use any memory tools\.?)\s*$/gim,
+    "",
+  );
+
+  // 1. Strip "System: [timestamp] Channel..." lines
+  cleaned = cleaned.replace(
     /^System:\s*\[[\d\-: +GMT]+\]\s+\S+\[.*?\].*$/gm,
     "",
   );

test/auto-capture-cleanup.test.mjs

@@ -0,0 +1,50 @@
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+import jitiFactory from "jiti";
+
+const jiti = jitiFactory(import.meta.url, { interopDefault: true });
+
+const {
+  normalizeAutoCaptureText,
+  stripAutoCaptureInjectedPrefix,
+} = jiti("../src/auto-capture-cleanup.ts");
+
+describe("auto-capture cleanup", () => {
+  it("preserves real content when wrapper lines are mixed with facts in the same payload", () => {
+    const input = [
+      "[Subagent Context] You are running as a subagent (depth 1/1). Results auto-announce to your requester.",
+      "[Subagent Task] Reply with a brief acknowledgment only. Facts for automatic memory extraction quality test: 1) Shen prefers concise blunt status updates. 2) Project Orion deploy window is Friday 21:00 Asia/Shanghai. 3) If a database migration touches billing tables, require a dry run first. Do not use any memory tools.",
+    ].join("\n");
+
+    const result = normalizeAutoCaptureText("user", input);
+    assert.equal(
+      result,
+      "Facts for automatic memory extraction quality test: 1) Shen prefers concise blunt status updates. 2) Project Orion deploy window is Friday 21:00 Asia/Shanghai. 3) If a database migration touches billing tables, require a dry run first.",
+    );
+  });
+
+  it("drops wrapper-only payloads", () => {
+    const input = [
+      "[Subagent Context] You are running as a subagent (depth 1/1). Results auto-announce to your requester.",
+      "[Subagent Task] Reply with a brief acknowledgment only.",
+    ].join("\n");
+
+    assert.equal(normalizeAutoCaptureText("user", input), null);
+  });
+
+  it("strips inbound metadata before preserving the remaining content", () => {
+    const input = [
+      "Conversation info (untrusted metadata):",
+      "```json",
+      '{"message_id":"om_123","sender_id":"ou_456"}',
+      "```",
+      "",
+      "[Subagent Task] Reply with a brief acknowledgment only. Actual user content starts here.",
+    ].join("\n");
+
+    assert.equal(
+      stripAutoCaptureInjectedPrefix("user", input),
+      "Actual user content starts here.",
+    );
+  });
+});

test/strip-envelope-metadata.test.mjs

@@ -118,6 +118,30 @@ describe("stripEnvelopeMetadata", () => {
   // -----------------------------------------------------------------------
   // Edge cases
   // -----------------------------------------------------------------------
+  it("strips subagent runtime wrapper lines but preserves real conversation that follows", () => {
+    const input = [
+      "[Subagent Context] You are running as a subagent (depth 1/1). Results auto-announce to your requester.",
+      "[Subagent Task] Reply with a brief acknowledgment only.",
+      "Actual user content starts here.",
+    ].join("\n");
+
+    const result = stripEnvelopeMetadata(input);
+    assert.equal(result, "Actual user content starts here.");
+  });
+
+  it("strips multiline wrapper continuation text but preserves following conversation", () => {
+    const input = [
+      "[Subagent Context] You are running as a subagent (depth 1/1).",
+      "Results auto-announce to your requester.",
+      "[Subagent Task] Reply with a brief acknowledgment only.",
+      "Do not use any memory tools.",
+      "Actual user content starts here.",
+    ].join("\n");
+
+    const result = stripEnvelopeMetadata(input);
+    assert.equal(result, "Actual user content starts here.");
+  });
+
   it("handles Telegram-style envelope headers", () => {
     const input = [
       "System: [2026-03-18 14:21:36 GMT+8] Telegram[bot123] DM | user_456 [msg:12345]",