fix(auto-reply): restore prompt cache stability by moving per-turn ids to user context (#20597)

Merged via /review-pr -> /prepare-pr -> /merge-pr.

Prepared head SHA: 175919afb6
Co-authored-by: anisoptera <768771+anisoptera@users.noreply.github.com>
Co-authored-by: mbelinky <132747814+mbelinky@users.noreply.github.com>
Reviewed-by: @mbelinky

CHANGELOG.md

@@ -36,6 +36,7 @@ Docs: https://docs.openclaw.ai
 - Gateway/Daemon: forward `TMPDIR` into installed service environments so macOS LaunchAgent gateway runs can open SQLite temp/journal files reliably instead of failing with `SQLITE_CANTOPEN`. (#20512) Thanks @Clawborn.
 - Agents/Billing: include the active model that produced a billing error in user-facing billing messages (for example, `OpenAI (gpt-5.3)`) across payload, failover, and lifecycle error paths, so users can identify exactly which key needs credits. (#20510) Thanks @echoVic.
 - Gateway/TUI: honor `agents.defaults.blockStreamingDefault` for `chat.send` by removing the hardcoded block-streaming disable override, so replies can use configured block-mode delivery. (#19693) Thanks @neipor.
+- Auto-reply/Prompt caching: restore prefix-cache stability by keeping inbound system metadata session-stable and moving per-message IDs (`message_id`, `message_id_full`, `reply_to_id`, `sender_id`) into untrusted conversation context. (#20597) Thanks @anisoptera.
 - UI/Sessions: accept the canonical main session-key alias in Chat UI flows so main-session routing stays consistent. (#20311) Thanks @mbelinky.
 - OpenClawKit/Protocol: preserve JSON boolean literals (`true`/`false`) when bridging through `AnyCodable` so Apple client RPC params no longer re-encode booleans as `1`/`0`. Thanks @mbelinky.
 - Commands/Doctor: skip embedding-provider warnings when `memory.backend` is `qmd`, because QMD manages embeddings internally and does not require `memorySearch` providers. (#17263) Thanks @miloudbelarebia.

src/auto-reply/reply/inbound-meta.test.ts

@@ -19,7 +19,7 @@ function parseConversationInfoPayload(text: string): Record<string, unknown> {
 }
 
 describe("buildInboundMetaSystemPrompt", () => {
-  it("includes trusted message and routing ids for tool actions", () => {
+  it("includes session-stable routing fields", () => {
     const prompt = buildInboundMetaSystemPrompt({
       MessageSid: "123",
       MessageSidFull: "123",
@@ -33,41 +33,28 @@ describe("buildInboundMetaSystemPrompt", () => {
 
     const payload = parseInboundMetaPayload(prompt);
     expect(payload["schema"]).toBe("openclaw.inbound_meta.v1");
-    expect(payload["message_id"]).toBe("123");
-    expect(payload["message_id_full"]).toBeUndefined();
-    expect(payload["reply_to_id"]).toBe("99");
     expect(payload["chat_id"]).toBe("telegram:5494292670");
     expect(payload["channel"]).toBe("telegram");
   });
 
-  it("includes sender_id when provided", () => {
+  it("does not include per-turn message identifiers (cache stability)", () => {
     const prompt = buildInboundMetaSystemPrompt({
-      MessageSid: "456",
+      MessageSid: "123",
+      MessageSidFull: "123",
+      ReplyToId: "99",
       SenderId: "289522496",
-      OriginatingTo: "telegram:-1001249586642",
+      OriginatingTo: "telegram:5494292670",
       OriginatingChannel: "telegram",
       Provider: "telegram",
       Surface: "telegram",
-      ChatType: "group",
+      ChatType: "direct",
     } as TemplateContext);
 
     const payload = parseInboundMetaPayload(prompt);
-    expect(payload["sender_id"]).toBe("289522496");
-  });
-
-  it("trims sender_id before storing", () => {
-    const prompt = buildInboundMetaSystemPrompt({
-      MessageSid: "457",
-      SenderId: "  289522496  ",
-      OriginatingTo: "telegram:-1001249586642",
-      OriginatingChannel: "telegram",
-      Provider: "telegram",
-      Surface: "telegram",
-      ChatType: "group",
-    } as TemplateContext);
-
-    const payload = parseInboundMetaPayload(prompt);
-    expect(payload["sender_id"]).toBe("289522496");
+    expect(payload["message_id"]).toBeUndefined();
+    expect(payload["message_id_full"]).toBeUndefined();
+    expect(payload["reply_to_id"]).toBeUndefined();
+    expect(payload["sender_id"]).toBeUndefined();
   });
 
   it("omits sender_id when blank", () => {
@@ -84,36 +71,6 @@ describe("buildInboundMetaSystemPrompt", () => {
     const payload = parseInboundMetaPayload(prompt);
     expect(payload["sender_id"]).toBeUndefined();
   });
-
-  it("omits sender_id when not provided", () => {
-    const prompt = buildInboundMetaSystemPrompt({
-      MessageSid: "789",
-      OriginatingTo: "telegram:5494292670",
-      OriginatingChannel: "telegram",
-      Provider: "telegram",
-      Surface: "telegram",
-      ChatType: "direct",
-    } as TemplateContext);
-
-    const payload = parseInboundMetaPayload(prompt);
-    expect(payload["sender_id"]).toBeUndefined();
-  });
-
-  it("keeps message_id_full only when it differs from message_id", () => {
-    const prompt = buildInboundMetaSystemPrompt({
-      MessageSid: "short-id",
-      MessageSidFull: "full-provider-message-id",
-      OriginatingTo: "channel:C1",
-      OriginatingChannel: "slack",
-      Provider: "slack",
-      Surface: "slack",
-      ChatType: "group",
-    } as TemplateContext);
-
-    const payload = parseInboundMetaPayload(prompt);
-    expect(payload["message_id"]).toBe("short-id");
-    expect(payload["message_id_full"]).toBe("full-provider-message-id");
-  });
 });
 
 describe("buildInboundUserContextPrefix", () => {
@@ -156,6 +113,63 @@ describe("buildInboundUserContextPrefix", () => {
     expect(conversationInfo["message_id"]).toBe("msg-123");
   });
 
+  it("includes message_id_full when it differs from message_id", () => {
+    const text = buildInboundUserContextPrefix({
+      ChatType: "group",
+      MessageSid: "short-id",
+      MessageSidFull: "full-provider-message-id",
+    } as TemplateContext);
+
+    const conversationInfo = parseConversationInfoPayload(text);
+    expect(conversationInfo["message_id"]).toBe("short-id");
+    expect(conversationInfo["message_id_full"]).toBe("full-provider-message-id");
+  });
+
+  it("omits message_id_full when it matches message_id", () => {
+    const text = buildInboundUserContextPrefix({
+      ChatType: "direct",
+      MessageSid: "same-id",
+      MessageSidFull: "same-id",
+    } as TemplateContext);
+
+    const conversationInfo = parseConversationInfoPayload(text);
+    expect(conversationInfo["message_id"]).toBe("same-id");
+    expect(conversationInfo["message_id_full"]).toBeUndefined();
+  });
+
+  it("includes reply_to_id in conversation info", () => {
+    const text = buildInboundUserContextPrefix({
+      ChatType: "direct",
+      MessageSid: "msg-200",
+      ReplyToId: "msg-199",
+    } as TemplateContext);
+
+    const conversationInfo = parseConversationInfoPayload(text);
+    expect(conversationInfo["reply_to_id"]).toBe("msg-199");
+  });
+
+  it("includes sender_id in conversation info", () => {
+    const text = buildInboundUserContextPrefix({
+      ChatType: "group",
+      MessageSid: "msg-456",
+      SenderId: "289522496",
+    } as TemplateContext);
+
+    const conversationInfo = parseConversationInfoPayload(text);
+    expect(conversationInfo["sender_id"]).toBe("289522496");
+  });
+
+  it("trims sender_id in conversation info", () => {
+    const text = buildInboundUserContextPrefix({
+      ChatType: "direct",
+      MessageSid: "msg-457",
+      SenderId: "  289522496  ",
+    } as TemplateContext);
+
+    const conversationInfo = parseConversationInfoPayload(text);
+    expect(conversationInfo["sender_id"]).toBe("289522496");
+  });
+
   it("falls back to SenderId when sender phone is missing", () => {
     const text = buildInboundUserContextPrefix({
       ChatType: "direct",

src/auto-reply/reply/inbound-meta.ts

@@ -13,20 +13,15 @@ function safeTrim(value: unknown): string | undefined {
 export function buildInboundMetaSystemPrompt(ctx: TemplateContext): string {
   const chatType = normalizeChatType(ctx.ChatType);
   const isDirect = !chatType || chatType === "direct";
-  const messageId = safeTrim(ctx.MessageSid);
-  const messageIdFull = safeTrim(ctx.MessageSidFull);
-  const replyToId = safeTrim(ctx.ReplyToId);
-  const chatId = safeTrim(ctx.OriginatingTo);
 
   // Keep system metadata strictly free of attacker-controlled strings (sender names, group subjects, etc.).
   // Those belong in the user-role "untrusted context" blocks.
+  // Per-message identifiers (message_id, reply_to_id, sender_id) are also excluded here: they change
+  // on every turn and would bust prefix-based prompt caches on local model providers. They are
+  // included in the user-role conversation info block via buildInboundUserContextPrefix() instead.
   const payload = {
     schema: "openclaw.inbound_meta.v1",
-    message_id: messageId,
-    message_id_full: messageIdFull && messageIdFull !== messageId ? messageIdFull : undefined,
-    sender_id: safeTrim(ctx.SenderId),
-    chat_id: chatId,
-    reply_to_id: replyToId,
+    chat_id: safeTrim(ctx.OriginatingTo),
     channel: safeTrim(ctx.OriginatingChannel) ?? safeTrim(ctx.Surface) ?? safeTrim(ctx.Provider),
     provider: safeTrim(ctx.Provider),
     surface: safeTrim(ctx.Surface),
@@ -60,8 +55,13 @@ export function buildInboundUserContextPrefix(ctx: TemplateContext): string {
   const chatType = normalizeChatType(ctx.ChatType);
   const isDirect = !chatType || chatType === "direct";
 
+  const messageId = safeTrim(ctx.MessageSid);
+  const messageIdFull = safeTrim(ctx.MessageSidFull);
   const conversationInfo = {
-    message_id: safeTrim(ctx.MessageSid),
+    message_id: messageId,
+    message_id_full: messageIdFull && messageIdFull !== messageId ? messageIdFull : undefined,
+    reply_to_id: safeTrim(ctx.ReplyToId),
+    sender_id: safeTrim(ctx.SenderId),
     conversation_label: isDirect ? undefined : safeTrim(ctx.ConversationLabel),
     sender: safeTrim(ctx.SenderE164) ?? safeTrim(ctx.SenderId) ?? safeTrim(ctx.SenderUsername),
     group_subject: safeTrim(ctx.GroupSubject),