feat: hide assistant card names in extraction transcript, split sectioned recentMessages into 2 system messages

2026-05-15 22:30:38 +08:00 · 2026-04-12 13:33:39 +08:00
parent 433e62e084
commit f40b03c306
8 changed files with 232 additions and 46 deletions
--- a/maintenance/extraction-context.js
+++ b/maintenance/extraction-context.js
@@ -289,6 +289,32 @@ function resolveSpeakerName(message = {}, role = "assistant", names = {}) {
  return role || "assistant";
 }

+function shouldHideSpeakerLabel(message = {}, role = "assistant", names = {}) {
+  if (message?.hideSpeakerLabel === true) {
+    return true;
+  }
+  if (message?.hideSpeakerLabel === false) {
+    return false;
+  }
+  if (role !== "assistant") {
+    return false;
+  }
+  if (String(message?.source || "").trim() === "worldInfo-atDepth") {
+    return false;
+  }
+  const explicitSpeaker = String(
+    message?.speaker ?? message?.name ?? message?.displayName ?? "",
+  ).trim();
+  if (!explicitSpeaker) {
+    return true;
+  }
+  const activeCharName = String(names?.charName || "").trim();
+  if (!activeCharName) {
+    return false;
+  }
+  return explicitSpeaker === activeCharName;
+}
+
 function normalizeExtractionMessage(message = {}, index = 0, names = {}) {
  const role = normalizeRole(
    message?.role ?? (message?.is_user === true ? "user" : "assistant"),
@@ -296,6 +322,7 @@ function normalizeExtractionMessage(message = {}, index = 0, names = {}) {
  const content = String(resolveMessageContent(message) || "").trim();
  const rawContent = String(resolveMessageRawContent(message) || content).trim();
  const speaker = resolveSpeakerName(message, role, names);
+  const hideSpeakerLabel = shouldHideSpeakerLabel(message, role, names);
  const seq = Number.isFinite(Number(message?.seq)) ? Number(message.seq) : null;

  return {
@@ -304,6 +331,7 @@ function normalizeExtractionMessage(message = {}, index = 0, names = {}) {
    role,
    speaker,
    name: speaker,
+    hideSpeakerLabel,
    content,
    rawContent,
    sourceType: role === "user" ? "user_input" : "ai_output",
@@ -347,7 +375,8 @@ export function formatExtractionTranscript(messages = []) {
      : `#${index + 1}`;
    const role = normalizeRole(message?.role || "assistant");
    const speaker = String(message?.speaker || message?.name || "").trim();
-    const speakerLabel = speaker ? `|${speaker}` : "";
+    const speakerLabel =
+      message?.hideSpeakerLabel === true || !speaker ? "" : `|${speaker}`;
    const line = `${seqLabel} [${role}${speakerLabel}]: ${String(message?.content || "")}`;
    if (String(line || "").trim()) {
      lines.push(line);
--- a/maintenance/extractor.js
+++ b/maintenance/extractor.js
@@ -873,6 +873,7 @@ export async function extractMemories({
        content: message?.content,
        speaker: message?.speaker,
        name: message?.name,
+        hideSpeakerLabel: message?.hideSpeakerLabel === true,
        isContextOnly: message?.isContextOnly === true,
      }))
    : [];
--- a/prompting/prompt-builder.js
+++ b/prompting/prompt-builder.js
@@ -296,6 +296,7 @@ function getPromptMessageLikeDescriptor(value) {
      role: role === "user" ? "user" : "assistant",
      seq: getOptionalFiniteNumber(value.seq),
      speaker,
+      hideSpeakerLabel: value?.hideSpeakerLabel === true,
      isContextOnly:
        typeof value.isContextOnly === "boolean" ? value.isContextOnly : null,
    };
@@ -310,6 +311,7 @@ function getPromptMessageLikeDescriptor(value) {
      role: value.is_user === true ? "user" : "assistant",
      seq: getOptionalFiniteNumber(value.seq),
      speaker,
+      hideSpeakerLabel: value?.hideSpeakerLabel === true,
      isContextOnly:
        typeof value.isContextOnly === "boolean" ? value.isContextOnly : null,
    };
@@ -326,6 +328,11 @@ function isPromptMessageArray(value) {
  );
 }

+const EXTRACTION_CONTEXT_REVIEW_HEADER =
+  "--- 以下是上下文回顾（已提取过），仅供理解剧情 ---";
+const EXTRACTION_TARGET_CONTENT_HEADER =
+  "--- 以下是本次需要提取记忆的新对话内容 ---";
+
 function getPromptMessageContextGroup(value) {
  const descriptor = getPromptMessageLikeDescriptor(value);
  if (!descriptor || typeof descriptor.isContextOnly !== "boolean") {
@@ -334,6 +341,16 @@ function getPromptMessageContextGroup(value) {
  return descriptor.isContextOnly ? "context" : "target";
 }

+function getPromptMessageContextHeader(group = "") {
+  if (group === "context") {
+    return EXTRACTION_CONTEXT_REVIEW_HEADER;
+  }
+  if (group === "target") {
+    return EXTRACTION_TARGET_CONTENT_HEADER;
+  }
+  return "";
+}
+
 function formatPromptMessageTranscript(value) {
  const entries = Array.isArray(value) ? value : [value];
  const hasContextMessages = entries.some(
@@ -353,16 +370,12 @@ function formatPromptMessageTranscript(value) {
    }
    const group = getPromptMessageContextGroup(entry);
    if (hasContextMessages && hasTargetMessages && group && group !== activeGroup) {
-      lines.push(
-        group === "context"
-          ? "--- 以下是上下文回顾（已提取过），仅供理解剧情 ---"
-          : "--- 以下是本次需要提取记忆的新对话内容 ---",
-      );
+      lines.push(getPromptMessageContextHeader(group));
      activeGroup = group;
    }
    const seqLabel =
      descriptor.seq != null ? `#${descriptor.seq}` : `#${index + 1}`;
-    const speakerLabel = descriptor.speaker
+    const speakerLabel = !descriptor.hideSpeakerLabel && descriptor.speaker
      ? `|${descriptor.speaker}`
      : "";
    lines.push(`${seqLabel} [${descriptor.role}${speakerLabel}]: ${descriptor.content}`);
@@ -1910,6 +1923,87 @@ function clonePayloadMessage(message = {}) {
  });
 }

+function splitSectionedTranscriptPayloadMessage(message = {}) {
+  const normalizedRole = normalizeRole(message?.role);
+  const sourceKey = String(message?.sourceKey || "").trim();
+  const content = String(message?.content || "").trim();
+  if (
+    normalizedRole !== "system" ||
+    !["recentMessages", "dialogueText"].includes(sourceKey) ||
+    !content.includes(EXTRACTION_CONTEXT_REVIEW_HEADER) ||
+    !content.includes(EXTRACTION_TARGET_CONTENT_HEADER)
+  ) {
+    return [message];
+  }
+
+  const headerMatches = [];
+  let searchIndex = 0;
+  while (searchIndex < content.length) {
+    const contextIndex = content.indexOf(
+      EXTRACTION_CONTEXT_REVIEW_HEADER,
+      searchIndex,
+    );
+    const targetIndex = content.indexOf(
+      EXTRACTION_TARGET_CONTENT_HEADER,
+      searchIndex,
+    );
+    let nextIndex = -1;
+    let nextHeader = "";
+    if (contextIndex >= 0 && (targetIndex < 0 || contextIndex <= targetIndex)) {
+      nextIndex = contextIndex;
+      nextHeader = EXTRACTION_CONTEXT_REVIEW_HEADER;
+    } else if (targetIndex >= 0) {
+      nextIndex = targetIndex;
+      nextHeader = EXTRACTION_TARGET_CONTENT_HEADER;
+    }
+    if (nextIndex < 0 || !nextHeader) {
+      break;
+    }
+    headerMatches.push({
+      index: nextIndex,
+      header: nextHeader,
+    });
+    searchIndex = nextIndex + nextHeader.length;
+  }
+
+  if (headerMatches.length < 2 || headerMatches[0].index !== 0) {
+    return [message];
+  }
+
+  const { role: _role, content: _content, ...sharedMeta } = message;
+  const splitMessages = [];
+
+  for (let index = 0; index < headerMatches.length; index += 1) {
+    const current = headerMatches[index];
+    const next = headerMatches[index + 1];
+    const sectionBody = content
+      .slice(current.index + current.header.length, next ? next.index : content.length)
+      .trim();
+    const transcriptSection =
+      current.header === EXTRACTION_CONTEXT_REVIEW_HEADER ? "context" : "target";
+    splitMessages.push(
+      createExecutionMessage(
+        "system",
+        sectionBody ? `${current.header}\n\n${sectionBody}` : current.header,
+        {
+          ...sharedMeta,
+          sourceKey,
+          transcriptSection,
+          transcriptSectionPart: "section",
+        },
+      ),
+    );
+  }
+
+  return splitMessages.filter(Boolean);
+}
+
+function expandSectionedTranscriptPayloadMessages(messages = []) {
+  return (Array.isArray(messages) ? messages : []).flatMap((message) =>
+    splitSectionedTranscriptPayloadMessage(message),
+  );
+}
+
 function collectPayloadUserMessageTexts(messages = []) {
  return (Array.isArray(messages) ? messages : [])
    .filter((message) => String(message?.role || "").trim().toLowerCase() === "user")
@@ -2008,8 +2102,11 @@ export function buildTaskLlmPayload(promptBuild = null, fallbackUserPrompt = "")
        !(isCustomFilter && messageUsesWorldInfoContent(message)),
    },
  );
+  const expandedExecutionMessages = expandSectionedTranscriptPayloadMessages(
+    executionMessages,
+  );

-  const hasUserMessage = executionMessages.some(
+  const hasUserMessage = expandedExecutionMessages.some(
    (message) => message.role === "user",
  );
  if (!hasUserMessage && rawExecutionMessages.length > 0) {
@@ -2028,7 +2125,7 @@ export function buildTaskLlmPayload(promptBuild = null, fallbackUserPrompt = "")
        `after recreate=${userBlocksAfterRaw.length}, ` +
        `after sanitize=${userBlocksAfterSanitize.length}, ` +
        `blockedContents count=${blockedContents.length}, ` +
-        `total executionMessages=${executionMessages.length}`,
+        `total executionMessages=${expandedExecutionMessages.length}`,
    );
    if (userBlocksBefore.length > 0) {
      for (const block of userBlocksBefore) {
@@ -2046,17 +2143,19 @@ export function buildTaskLlmPayload(promptBuild = null, fallbackUserPrompt = "")
    }
  }
  const additionalMessages =
-    executionMessages.length > 0
+    expandedExecutionMessages.length > 0
      ? []
-      : sanitizePromptMessages(
-          settings,
-          taskType,
-          rawPrivateTaskMessages,
-          {
-            blockedContents,
-            applySanitizer: (message) =>
-              !(isCustomFilter && messageUsesWorldInfoContent(message)),
-          },
+      : expandSectionedTranscriptPayloadMessages(
+          sanitizePromptMessages(
+            settings,
+            taskType,
+            rawPrivateTaskMessages,
+            {
+              blockedContents,
+              applySanitizer: (message) =>
+                !(isCustomFilter && messageUsesWorldInfoContent(message)),
+            },
+          ),
        );
  const hasAdditionalUserMessage = additionalMessages.some(
    (message) => message.role === "user",
@@ -2078,9 +2177,11 @@ export function buildTaskLlmPayload(promptBuild = null, fallbackUserPrompt = "")

  return {
    systemPrompt:
-      executionMessages.length > 0 ? "" : String(promptBuild?.systemPrompt || ""),
+      expandedExecutionMessages.length > 0
+        ? ""
+        : String(promptBuild?.systemPrompt || ""),
    userPrompt: fallbackUserPromptResult.text,
-    promptMessages: executionMessages,
+    promptMessages: expandedExecutionMessages,
    additionalMessages,
    fallbackUserPromptSource: fallbackUserPromptResult.source,
    fallbackUserPromptApplied: Boolean(fallbackUserPromptResult.text),
--- a/tests/extraction-context-only-flag.mjs
+++ b/tests/extraction-context-only-flag.mjs
@@ -76,9 +76,23 @@ const chat = [
 {
  const mixed = [
    { seq: 1, role: "user", content: "context user", speaker: "A", isContextOnly: true },
-    { seq: 2, role: "assistant", content: "context ai", speaker: "B", isContextOnly: true },
+    {
+      seq: 2,
+      role: "assistant",
+      content: "context ai",
+      speaker: "B",
+      hideSpeakerLabel: true,
+      isContextOnly: true,
+    },
    { seq: 3, role: "user", content: "target user", speaker: "A", isContextOnly: false },
-    { seq: 4, role: "assistant", content: "target ai", speaker: "B", isContextOnly: false },
+    {
+      seq: 4,
+      role: "assistant",
+      content: "target ai",
+      speaker: "B",
+      hideSpeakerLabel: true,
+      isContextOnly: false,
+    },
  ];
  const transcript = formatExtractionTranscript(mixed);
  assert.match(transcript, /已提取过/, "transcript should contain context review header");
@@ -89,6 +103,8 @@ const chat = [
  );
  assert.match(transcript, /#1.*context user/, "context message should appear");
  assert.match(transcript, /#3.*target user/, "target message should appear");
+  assert.match(transcript, /#2 \[assistant\]: context ai/, "assistant card name should be hidden");
+  assert.doesNotMatch(transcript, /#2 \[assistant\|B\]:/, "assistant card name should not be rendered");
  console.log("  ✓ formatExtractionTranscript: section dividers for mixed context/target");
 }

@@ -132,8 +148,20 @@ const chat = [
  const targetFiltered = result.filteredMessages.filter((m) => !m.isContextOnly);
  assert.equal(contextFiltered.length, 2, "context messages propagated through filtering");
  assert.equal(targetFiltered.length, 2, "target messages propagated through filtering");
+  assert.equal(
+    result.filteredMessages.find((m) => m.seq === 2)?.hideSpeakerLabel,
+    true,
+    "active character assistant label should be hidden",
+  );
+  assert.equal(
+    result.filteredMessages.find((m) => m.seq === 1)?.hideSpeakerLabel,
+    false,
+    "user label should remain visible",
+  );
  assert.match(result.filteredTranscript, /已提取过/, "transcript includes context header");
  assert.match(result.filteredTranscript, /本次需要提取/, "transcript includes target header");
+  assert.match(result.filteredTranscript, /#2 \[assistant\]: old answer/, "assistant transcript should hide character name");
+  assert.doesNotMatch(result.filteredTranscript, /#2 \[assistant\|B\]:/, "assistant transcript should not show character name");
  console.log("  ✓ buildExtractionInputContext: isContextOnly propagated to filteredMessages and transcript");
 }

--- a/tests/extractor-input-context.mjs
+++ b/tests/extractor-input-context.mjs
@@ -141,8 +141,9 @@ try {
    (message) => message.sourceKey === "recentMessages",
  );
  assert.ok(recentBlock);
-  assert.match(String(recentBlock?.content || ""), /#10 \[assistant\|艾琳\]: 继续说明/);
+  assert.match(String(recentBlock?.content || ""), /#10 \[assistant\]: 继续说明/);
  assert.match(String(recentBlock?.content || ""), /#11 \[user\|玩家\]: 用户输入/);
+  assert.doesNotMatch(String(recentBlock?.content || ""), /#10 \[assistant\|艾琳\]:/);
  assert.doesNotMatch(String(recentBlock?.content || ""), /隐式思维|<think>/);
 } finally {
  restore();
--- a/tests/extractor-phase3-layered-context.mjs
+++ b/tests/extractor-phase3-layered-context.mjs
@@ -224,16 +224,25 @@ function collectAllPromptContent(captured) {
    assert.equal(result.success, true);
    assert.ok(captured);

-    const recentBlock = (Array.isArray(captured.promptMessages) ? captured.promptMessages : []).find(
+    const recentMessages = (Array.isArray(captured.promptMessages)
+      ? captured.promptMessages
+      : []
+    ).filter(
      (m) => m.sourceKey === "recentMessages",
    );
-    assert.ok(recentBlock, "recentMessages block should exist");
-    const recentContent = String(recentBlock?.content || "");
-    assert.match(recentContent, /以下是上下文回顾（已提取过），仅供理解剧情/);
-    assert.match(recentContent, /以下是本次需要提取记忆的新对话内容/);
+    assert.equal(recentMessages.length, 2, "recentMessages should split into 2 section system messages");
+    assert.equal(recentMessages[0]?.role, "system");
+    assert.equal(recentMessages[0]?.transcriptSection, "context");
+    assert.match(String(recentMessages[0]?.content || ""), /^--- 以下是上下文回顾（已提取过），仅供理解剧情 ---/);
+    assert.match(String(recentMessages[0]?.content || ""), /#10 \[user\|玩家\]: 第一轮消息/);
+    assert.equal(recentMessages[1]?.role, "system");
+    assert.equal(recentMessages[1]?.transcriptSection, "target");
+    assert.match(String(recentMessages[1]?.content || ""), /^--- 以下是本次需要提取记忆的新对话内容 ---/);
+    assert.match(String(recentMessages[1]?.content || ""), /#12 \[user\|玩家\]: 第二轮消息/);
    assert.ok(
-      recentContent.indexOf("已提取过") < recentContent.indexOf("本次需要提取"),
-      "context review should appear before extraction target section",
+      recentMessages[0].content.includes("已提取过") &&
+        recentMessages[1].content.includes("本次需要提取"),
+      "context and target sections should each be emitted as a single system message",
    );
  } finally {
    restore();
--- a/tests/extractor-phase5-context-fidelity.mjs
+++ b/tests/extractor-phase5-context-fidelity.mjs
@@ -286,7 +286,7 @@ try {
      ).find((message) => message.sourceKey === "recentMessages");
      assert.ok(recentBlock, "recentMessages block should exist");
      const recentContent = String(recentBlock?.content || "");
-      assert.match(recentContent, /#30 \[assistant\|艾琳\]: 艾琳说：去调查蓝钥匙。/);
+      assert.match(recentContent, /#30 \[assistant\]: 艾琳说：去调查蓝钥匙。/);
      assert.match(
        recentContent,
        /#31 \[assistant\|旁白\]: 旁白补充：<status mood='tense'>雨夜<\/status>巷子很安静。/,
@@ -351,7 +351,7 @@ try {
        : []
      ).find((message) => message.sourceKey === "recentMessages");
      assert.ok(recentBlock, "recentMessages block should still exist when worldbook is disabled");
-      assert.match(String(recentBlock?.content || ""), /#30 \[assistant\|艾琳\]: 艾琳说：去调查蓝钥匙。/);
+      assert.match(String(recentBlock?.content || ""), /#30 \[assistant\]: 艾琳说：去调查蓝钥匙。/);
    } finally {
      restore();
    }
--- a/tests/prompt-builder-mixed-transcript.mjs
+++ b/tests/prompt-builder-mixed-transcript.mjs
@@ -117,6 +117,7 @@ const promptBuild = await buildTaskPrompt(settings, "extract", {
      content: "继续说明",
      name: "艾琳",
      speaker: "艾琳",
+      hideSpeakerLabel: true,
      isContextOnly: true,
    },
    {
@@ -133,25 +134,41 @@ const promptBuild = await buildTaskPrompt(settings, "extract", {
  currentRange: "41 ~ 42",
 });
 const payload = buildTaskLlmPayload(promptBuild, "fallback-user");
-const recentBlock = payload.promptMessages.find(
+const recentMessages = payload.promptMessages.filter(
  (message) => message.sourceKey === "recentMessages",
 );
-assert.match(
-  String(recentBlock?.content || ""),
-  /以下是上下文回顾（已提取过），仅供理解剧情/,
+assert.deepEqual(
+  recentMessages.map((message) => ({
+    role: message.role,
+    sourceKey: message.sourceKey,
+    transcriptSection: message.transcriptSection,
+    transcriptSectionPart: message.transcriptSectionPart,
+  })),
+  [
+    {
+      role: "system",
+      sourceKey: "recentMessages",
+      transcriptSection: "context",
+      transcriptSectionPart: "section",
+    },
+    {
+      role: "system",
+      sourceKey: "recentMessages",
+      transcriptSection: "target",
+      transcriptSectionPart: "section",
+    },
+  ],
 );
-assert.match(
-  String(recentBlock?.content || ""),
-  /以下是本次需要提取记忆的新对话内容/,
-);
-assert.match(String(recentBlock?.content || ""), /#41 \[assistant\|艾琳\]: 助手已净化/);
-assert.match(String(recentBlock?.content || ""), /#42 \[user\|玩家\]: 用户已净化/);
+assert.match(String(recentMessages[0]?.content || ""), /^--- 以下是上下文回顾（已提取过），仅供理解剧情 ---/);
+assert.match(String(recentMessages[0]?.content || ""), /#41 \[assistant\]: 助手已净化/);
+assert.match(String(recentMessages[1]?.content || ""), /^--- 以下是本次需要提取记忆的新对话内容 ---/);
+assert.match(String(recentMessages[1]?.content || ""), /#42 \[user\|玩家\]: 用户已净化/);
 assert.doesNotMatch(
-  String(recentBlock?.content || ""),
-  /#41 \[assistant\|艾琳\]: 用户已净化/,
+  String(recentMessages[0]?.content || ""),
+  /#41 \[assistant\|艾琳\]:/,
 );
 assert.doesNotMatch(
-  String(recentBlock?.content || ""),
+  String(recentMessages[1]?.content || ""),
  /#42 \[user\|玩家\]: 助手已净化/,
 );