fix(inbound-meta): preserve reply-context body tails

Preserve actionable tail content in long reply-context bodies before they enter prompt JSON or inline reply context formatting. - Apply UTF-16-safe head+tail truncation to ReplyChain JSON bodies and fallback ReplyToBody JSON blocks. - Use the same body-aware truncation for Telegram inline ReplyToBody fallback and chat_window message bodies, so those paths cannot suppress the JSON fallback and still lose the tail. - Adds regression coverage for ReplyChain, fallback ReplyToBody, Telegram inline ReplyToBody, chat_window reply targets, and emoji-heavy heads. Verification: - node scripts/run-vitest.mjs src/auto-reply/reply/inbound-meta.test.ts - node_modules/.bin/oxfmt --check --threads=1 src/auto-reply/reply/inbound-meta.ts src/auto-reply/reply/inbound-meta.test.ts - node scripts/run-oxlint.mjs src/auto-reply/reply/inbound-meta.ts src/auto-reply/reply/inbound-meta.test.ts - .agents/skills/autoreview/scripts/autoreview --mode branch --base origin/main - Testbox-through-Crabbox check:changed: provider=blacksmith-testbox id=tbx_01ktgthbb5xa9d5ap58h4134s0 exit=0 - PR CI: 158/158 completed, no failures, MERGEABLE/CLEAN Fixes #91042
2026-07-21 10:16:44 +00:00 · 2026-06-07 19:45:00 +09:00
parent 9bafa2a2b6
commit 3753c5e2c8
2 changed files with 154 additions and 5 deletions
@@ -598,6 +598,29 @@ describe("buildInboundUserContextPrefix", () => {
    expect(text).not.toContain("Reply target of current user message");
  });

+  it("preserves Telegram inline ReplyToBody tail content", () => {
+    const head = "BEGIN. ".repeat(300);
+    const tail = " TELEGRAM_INLINE_TAIL";
+    const longBody = head + tail;
+    expect(longBody.length).toBeGreaterThan(2_000);
+
+    const text = buildInboundUserContextPrefix({
+      Provider: "telegram",
+      Surface: "telegram",
+      OriginatingChannel: "telegram",
+      ChatType: "group",
+      MessageSid: "34974",
+      ReplyToId: "34971",
+      ReplyToBody: longBody,
+      SenderName: "obviyus",
+    } as TemplateContext);
+
+    expect(text).toContain("TELEGRAM_INLINE_TAIL");
+    expect(text).toContain("…[omitted]…");
+    expect(text).not.toContain("…[truncated]");
+    expect(text).not.toContain("Reply target of current user message");
+  });
+
  it("keeps Telegram current-message quote even when context already includes the target", () => {
    const text = buildInboundUserContextPrefix({
      Provider: "telegram",
@@ -948,6 +971,91 @@ describe("buildInboundUserContextPrefix", () => {
    expect(text).toContain('"body": "');
  });

+  it("preserves tail content in ReplyChain body via head+tail truncation", () => {
+    const head = "BEGIN. ".repeat(300);
+    const tail = " IMPORTANT_TAIL_SENTINEL";
+    const longBody = head + tail;
+    expect(longBody.length).toBeGreaterThan(2_000);
+
+    const text = buildInboundUserContextPrefix({
+      ChatType: "group",
+      ReplyChain: [{ body: longBody, sender: "Alice" }],
+    } as TemplateContext);
+
+    const [reply] = parseReplyChainPayload(text);
+    expect(reply?.["body"]).toContain("IMPORTANT_TAIL_SENTINEL");
+    expect(reply?.["body"]).toContain("…[omitted]…");
+    expect(reply?.["body"]).not.toContain("…[truncated]");
+  });
+
+  it("preserves tail content in fallback ReplyToBody via head+tail truncation", () => {
+    const head = "BEGIN. ".repeat(300);
+    const tail = " REPLY_TAIL_SENTINEL";
+    const longBody = head + tail;
+    expect(longBody.length).toBeGreaterThan(2_000);
+
+    const text = buildInboundUserContextPrefix({
+      ChatType: "group",
+      ReplyToBody: longBody,
+    } as TemplateContext);
+
+    const reply = parseReplyPayload(text);
+    expect(reply["body"]).toContain("REPLY_TAIL_SENTINEL");
+    expect(reply["body"]).toContain("…[omitted]…");
+    expect(reply["body"]).not.toContain("…[truncated]");
+  });
+
+  it("preserves fallback ReplyToBody tail when the head is emoji-heavy", () => {
+    const head = "😀".repeat(1_200);
+    const tail = " TAIL_AFTER_EMOJI_HEAD";
+    const longBody = head + tail;
+    expect(longBody.length).toBeGreaterThan(2_000);
+
+    const text = buildInboundUserContextPrefix({
+      ReplyToSender: "Quoter",
+      ReplyToBody: longBody,
+    } as TemplateContext);
+
+    const reply = parseReplyPayload(text);
+    expect(reply["body"]).toContain("TAIL_AFTER_EMOJI_HEAD");
+    expect(reply["body"]).toContain("…[omitted]…");
+    expect(reply["body"]).not.toContain("…[truncated]");
+  });
+
+  it("preserves chat window reply-target body tail content", () => {
+    const head = "BEGIN. ".repeat(300);
+    const tail = " CHAT_WINDOW_TAIL";
+    const longBody = head + tail;
+    expect(longBody.length).toBeGreaterThan(2_000);
+
+    const text = buildInboundUserContextPrefix({
+      ChatType: "group",
+      ReplyToId: "msg-1",
+      UntrustedStructuredContext: [
+        {
+          label: "Conversation context",
+          type: "chat_window",
+          payload: {
+            relation: "around_reply_target",
+            messages: [
+              {
+                message_id: "msg-1",
+                sender: "Avery",
+                body: longBody,
+                is_reply_target: true,
+              },
+            ],
+          },
+        },
+      ],
+    } as TemplateContext);
+
+    expect(text).toContain("CHAT_WINDOW_TAIL");
+    expect(text).toContain("…[omitted]…");
+    expect(text).not.toContain("…[truncated]");
+    expect(text).not.toContain("Reply target of current user message");
+  });
+
  it("caps serialized inbound history to the most recent bounded tail", () => {
    const text = buildInboundUserContextPrefix({
      ChatType: "group",
@@ -6,7 +6,7 @@ import { getLoadedChannelPluginById } from "../../channels/plugins/registry-load
 import type { ChannelPlugin } from "../../channels/plugins/types.plugin.js";
 import { normalizeAnyChannelId } from "../../channels/registry.js";
 import { resolveSenderLabel } from "../../channels/sender-label.js";
-import { truncateUtf16Safe } from "../../utils.js";
+import { sliceUtf16Safe, truncateUtf16Safe } from "../../utils.js";
 import type { EnvelopeFormatOptions } from "../envelope.js";
 import { formatEnvelopeTimestamp } from "../envelope.js";
 import type { SourceReplyDeliveryMode } from "../get-reply-options.types.js";
@@ -65,6 +65,34 @@ function truncateUntrustedJsonString(value: string): string {
  return `${truncateUtf16Safe(value, Math.max(0, MAX_UNTRUSTED_JSON_STRING_CHARS - 14)).trimEnd()}…[truncated]`;
 }

+const HEAD_TAIL_OMISSION_MARKER = "…[omitted]…";
+const HEAD_TAIL_MARKER_LENGTH = HEAD_TAIL_OMISSION_MARKER.length;
+const MIN_HEAD_TAIL_CHARS = 20;
+
+/**
+ * Applies head+tail truncation so the result is ≤ maxChars and the downstream
+ * {@link truncateUntrustedJsonString} (prefix-only 2000-char cap) is a no-op.
+ * Head and tail portions are sized to keep the body within
+ * {@link MAX_UNTRUSTED_JSON_STRING_CHARS}, preserving actionable tail content
+ * that prefix-only truncation would drop.
+ */
+function truncateBodyHeadTail(body: string, maxChars = MAX_UNTRUSTED_JSON_STRING_CHARS): string {
+  if (body.length <= maxChars) {
+    return body;
+  }
+  const available = maxChars - HEAD_TAIL_MARKER_LENGTH;
+  if (available < MIN_HEAD_TAIL_CHARS * 2) {
+    return `${truncateUtf16Safe(body, Math.max(0, maxChars - 14)).trimEnd()}…[truncated]`;
+  }
+  // Budget in UTF-16 code units because truncateUntrustedJsonString enforces
+  // that same cap after JSON serialization.
+  const headChars = Math.floor(available * 0.6);
+  const tailChars = available - headChars;
+  const head = truncateUtf16Safe(body, headChars);
+  const tail = sliceUtf16Safe(body, -tailChars);
+  return `${head}${HEAD_TAIL_OMISSION_MARKER}${tail}`;
+}
+
 function sanitizeUntrustedJsonValue(value: unknown): unknown {
  if (typeof value === "string") {
    return neutralizeMarkdownFences(truncateUntrustedJsonString(value));
@@ -100,6 +128,17 @@ function sanitizeTranscriptField(value: unknown): string | undefined {
    .trim();
 }

+function sanitizeTranscriptBody(value: unknown): string | undefined {
+  const body = sanitizePromptBody(value);
+  if (!body) {
+    return undefined;
+  }
+  const sanitized = neutralizeMarkdownFences(truncateBodyHeadTail(body))
+    .replace(/\s+/g, " ")
+    .trim();
+  return sanitized || undefined;
+}
+
 function formatUntrustedStructuredContextLabel(label: unknown): string {
  const normalized = normalizePromptMetadataString(label);
  return normalized
@@ -156,7 +195,7 @@ function formatChatWindowMessage(
  const replyToId = sanitizeTranscriptField(value["reply_to_id"]);
  const mediaType = sanitizeTranscriptField(value["media_type"]);
  const mediaRef = sanitizeTranscriptField(value["media_ref"]);
-  const body = sanitizeTranscriptField(value["body"]);
+  const body = sanitizeTranscriptBody(value["body"]);
  const details = [
    messageId ? `#${messageId}` : undefined,
    timestamp,
@@ -274,7 +313,8 @@ function buildReplyChainPayload(ctx: TemplateContext): Array<Record<string, unkn
    return [];
  }
  return ctx.ReplyChain.flatMap((entry) => {
-    const body = sanitizePromptBody(entry.body);
+    const rawBody = sanitizePromptBody(entry.body);
+    const body = rawBody ? truncateBodyHeadTail(rawBody) : rawBody;
    const mediaType = normalizePromptMetadataString(entry.mediaType);
    const mediaPath = normalizePromptMetadataString(entry.mediaPath);
    const mediaRef = normalizePromptMetadataString(entry.mediaRef);
@@ -312,7 +352,7 @@ function isTelegramInboundContext(ctx: TemplateContext): boolean {
 }

 function resolveInlineReplyQuote(ctx: TemplateContext): string | undefined {
-  return sanitizeTranscriptField(ctx.ReplyToQuoteText) ?? sanitizeTranscriptField(ctx.ReplyToBody);
+  return sanitizeTranscriptField(ctx.ReplyToQuoteText) ?? sanitizeTranscriptBody(ctx.ReplyToBody);
 }

 function formatTelegramCurrentMessageContext(ctx: TemplateContext): string | undefined {
@@ -552,7 +592,8 @@ export function buildInboundUserContextPrefix(
    );
  }

-  const replyToBody = sanitizePromptBody(ctx.ReplyToBody);
+  const rawReplyToBody = sanitizePromptBody(ctx.ReplyToBody);
+  const replyToBody = rawReplyToBody ? truncateBodyHeadTail(rawReplyToBody) : rawReplyToBody;
  if (replyChainPayload.length > 0 && !chatWindowCoversReplyContext && !currentMessageContext) {
    blocks.push(
      formatUntrustedJsonBlock(