@memtensor/memos-local-openclaw-plugin 0.1.2 → 0.1.4

package/src/ingest/providers/gemini.ts

@@ -1,6 +1,133 @@
 import type { SummarizerConfig, Logger } from "../../types";
 
-const SYSTEM_PROMPT = `Summarize the text in ONE concise sentence (max 60 tokens). Preserve exact names, commands, error codes. No bullet points, no preamble — output only the sentence.`;
+const SYSTEM_PROMPT = `Summarize the text in ONE concise sentence (max 120 characters). IMPORTANT: Use the SAME language as the input text — if the input is Chinese, write Chinese; if English, write English. Preserve exact names, commands, error codes. No bullet points, no preamble — output only the sentence.`;
+
+const TASK_SUMMARY_PROMPT = `You create a DETAILED task summary from a multi-turn conversation. This summary will be the ONLY record of this conversation, so it must preserve ALL important information.
+
+CRITICAL LANGUAGE RULE: You MUST write in the SAME language as the user's messages. Chinese input → Chinese output. English input → English output. NEVER mix languages.
+
+Output EXACTLY this structure:
+
+📌 Title
+A short, descriptive title (10-30 characters). Like a chat group name.
+
+🎯 Goal
+One sentence: what the user wanted to accomplish.
+
+📋 Key Steps
+- Describe each meaningful step in detail
+- Include the ACTUAL content produced: code snippets, commands, config blocks, formulas, key paragraphs
+- For code: include the function signature and core logic (up to ~30 lines per block), use fenced code blocks
+- For configs: include the actual config values and structure
+- For lists/instructions: include the actual items, not just "provided a list"
+- Merge only truly trivial back-and-forth (like "ok" / "sure")
+- Do NOT over-summarize: "provided a function" is BAD; show the actual function
+
+✅ Result
+What was the final outcome? Include the final version of any code/config/content produced.
+
+💡 Key Details
+- Decisions made, trade-offs discussed, caveats noted, alternative approaches mentioned
+- Specific values: numbers, versions, thresholds, URLs, file paths, model names
+- Omit this section only if there truly are no noteworthy details
+
+RULES:
+- This summary is a KNOWLEDGE BASE ENTRY, not a brief note. Be thorough.
+- PRESERVE verbatim: code, commands, URLs, file paths, error messages, config values, version numbers, names, amounts
+- DISCARD only: greetings, filler, the assistant explaining what it will do before doing it
+- Replace secrets (API keys, tokens, passwords) with [REDACTED]
+- Target length: 30-50% of the original conversation length. Longer conversations need longer summaries.
+- Output summary only, no preamble.`;
+
+export async function summarizeTaskGemini(
+  text: string,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<string> {
+  const model = cfg.model ?? "gemini-1.5-flash";
+  const endpoint =
+    cfg.endpoint ??
+    `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent`;
+
+  const url = `${endpoint}?key=${cfg.apiKey}`;
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    ...cfg.headers,
+  };
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      systemInstruction: { parts: [{ text: TASK_SUMMARY_PROMPT }] },
+      contents: [{ parts: [{ text }] }],
+      generationConfig: { temperature: cfg.temperature ?? 0.1, maxOutputTokens: 4096 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Gemini task-summarize failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { candidates: Array<{ content: { parts: Array<{ text: string }> } }> };
+  return json.candidates?.[0]?.content?.parts?.[0]?.text?.trim() ?? "";
+}
+
+const TOPIC_JUDGE_PROMPT = `You are a conversation topic boundary detector. Given a summary of the CURRENT conversation and a NEW user message, determine if the new message starts a DIFFERENT topic/task.
+
+Answer ONLY "NEW" or "SAME".
+
+Rules:
+- "NEW" = the new message is about a completely different subject, project, or task
+- "SAME" = the new message continues, follows up on, or is closely related to the current topic
+- Follow-up questions, clarifications, refinements, bug fixes, or next steps on the same task = SAME
+- Greetings or meta-questions like "你好" or "谢谢" without new substance = SAME
+- A clearly unrelated request (e.g., current topic is deployment, new message asks about cooking) = NEW
+
+Output exactly one word: NEW or SAME`;
+
+export async function judgeNewTopicGemini(
+  currentContext: string,
+  newMessage: string,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<boolean> {
+  const model = cfg.model ?? "gemini-1.5-flash";
+  const endpoint =
+    cfg.endpoint ??
+    `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent`;
+
+  const url = `${endpoint}?key=${cfg.apiKey}`;
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    ...cfg.headers,
+  };
+
+  const userContent = `CURRENT CONVERSATION SUMMARY:\n${currentContext}\n\nNEW USER MESSAGE:\n${newMessage}`;
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      systemInstruction: { parts: [{ text: TOPIC_JUDGE_PROMPT }] },
+      contents: [{ parts: [{ text: userContent }] }],
+      generationConfig: { temperature: 0, maxOutputTokens: 10 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Gemini topic-judge failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { candidates: Array<{ content: { parts: Array<{ text: string }> } }> };
+  const answer = json.candidates?.[0]?.content?.parts?.[0]?.text?.trim().toUpperCase() ?? "";
+  log.debug(`Topic judge result: "${answer}"`);
+  return answer.startsWith("NEW");
+}
 
 export async function summarizeGemini(
   text: string,

package/src/ingest/providers/index.ts

@@ -1,8 +1,8 @@
 import type { SummarizerConfig, Logger } from "../../types";
-import { summarizeOpenAI } from "./openai";
-import { summarizeAnthropic } from "./anthropic";
-import { summarizeGemini } from "./gemini";
-import { summarizeBedrock } from "./bedrock";
+import { summarizeOpenAI, summarizeTaskOpenAI, judgeNewTopicOpenAI } from "./openai";
+import { summarizeAnthropic, summarizeTaskAnthropic, judgeNewTopicAnthropic } from "./anthropic";
+import { summarizeGemini, summarizeTaskGemini, judgeNewTopicGemini } from "./gemini";
+import { summarizeBedrock, summarizeTaskBedrock, judgeNewTopicBedrock } from "./bedrock";
 
 export class Summarizer {
   constructor(
@@ -23,6 +23,19 @@ export class Summarizer {
     }
   }
 
+  async summarizeTask(text: string): Promise<string> {
+    if (!this.cfg) {
+      return taskFallback(text);
+    }
+
+    try {
+      return await this.callTaskProvider(text);
+    } catch (err) {
+      this.log.warn(`Task summarizer failed, using fallback: ${err}`);
+      return taskFallback(text);
+    }
+  }
+
   private async callProvider(text: string): Promise<string> {
     const cfg = this.cfg!;
     switch (cfg.provider) {
@@ -41,12 +54,65 @@ export class Summarizer {
         throw new Error(`Unknown summarizer provider: ${cfg.provider}`);
     }
   }
+
+  /**
+   * Ask the LLM whether the new message starts a different topic from the current conversation.
+   * Returns true if it's a new topic, false if it continues the current one.
+   * Returns null if no summarizer is configured (caller should fall back to heuristic).
+   */
+  async judgeNewTopic(currentContext: string, newMessage: string): Promise<boolean | null> {
+    if (!this.cfg) return null;
+
+    try {
+      return await this.callTopicJudge(currentContext, newMessage);
+    } catch (err) {
+      this.log.warn(`Topic judge failed: ${err}`);
+      return null;
+    }
+  }
+
+  private async callTopicJudge(currentContext: string, newMessage: string): Promise<boolean> {
+    const cfg = this.cfg!;
+    switch (cfg.provider) {
+      case "openai":
+      case "openai_compatible":
+      case "azure_openai":
+        return judgeNewTopicOpenAI(currentContext, newMessage, cfg, this.log);
+      case "anthropic":
+        return judgeNewTopicAnthropic(currentContext, newMessage, cfg, this.log);
+      case "gemini":
+        return judgeNewTopicGemini(currentContext, newMessage, cfg, this.log);
+      case "bedrock":
+        return judgeNewTopicBedrock(currentContext, newMessage, cfg, this.log);
+      default:
+        throw new Error(`Unknown summarizer provider: ${cfg.provider}`);
+    }
+  }
+
+  private async callTaskProvider(text: string): Promise<string> {
+    const cfg = this.cfg!;
+    switch (cfg.provider) {
+      case "openai":
+      case "openai_compatible":
+      case "azure_openai":
+        return summarizeTaskOpenAI(text, cfg, this.log);
+      case "anthropic":
+        return summarizeTaskAnthropic(text, cfg, this.log);
+      case "gemini":
+        return summarizeTaskGemini(text, cfg, this.log);
+      case "bedrock":
+        return summarizeTaskBedrock(text, cfg, this.log);
+      default:
+        throw new Error(`Unknown summarizer provider: ${cfg.provider}`);
+    }
+  }
+}
+
+function taskFallback(text: string): string {
+  const lines = text.split("\n").filter((l) => l.trim().length > 10);
+  return lines.slice(0, 30).join("\n").slice(0, 2000);
 }
 
-/**
- * Rule-based fallback: produce a single short sentence from the first
- * meaningful line, appending any key entities found in the text.
- */
 function ruleFallback(text: string): string {
   const lines = text.split("\n").filter((l) => l.trim().length > 10);
   const first = (lines[0] ?? text).trim();

package/src/ingest/providers/openai.ts

@@ -1,6 +1,80 @@
 import type { SummarizerConfig, Logger } from "../../types";
 
-const SYSTEM_PROMPT = `Summarize the text in ONE concise sentence (max 60 tokens). Preserve exact names, commands, error codes. No bullet points, no preamble — output only the sentence.`;
+const SYSTEM_PROMPT = `Summarize the text in ONE concise sentence (max 120 characters). IMPORTANT: Use the SAME language as the input text — if the input is Chinese, write Chinese; if English, write English. Preserve exact names, commands, error codes. No bullet points, no preamble — output only the sentence.`;
+
+const TASK_SUMMARY_PROMPT = `You create a DETAILED task summary from a multi-turn conversation. This summary will be the ONLY record of this conversation, so it must preserve ALL important information.
+
+CRITICAL LANGUAGE RULE: You MUST write in the SAME language as the user's messages. Chinese input → Chinese output. English input → English output. NEVER mix languages.
+
+Output EXACTLY this structure:
+
+📌 Title
+A short, descriptive title (10-30 characters). Like a chat group name.
+
+🎯 Goal
+One sentence: what the user wanted to accomplish.
+
+📋 Key Steps
+- Describe each meaningful step in detail
+- Include the ACTUAL content produced: code snippets, commands, config blocks, formulas, key paragraphs
+- For code: include the function signature and core logic (up to ~30 lines per block), use fenced code blocks
+- For configs: include the actual config values and structure
+- For lists/instructions: include the actual items, not just "provided a list"
+- Merge only truly trivial back-and-forth (like "ok" / "sure")
+- Do NOT over-summarize: "provided a function" is BAD; show the actual function
+
+✅ Result
+What was the final outcome? Include the final version of any code/config/content produced.
+
+💡 Key Details
+- Decisions made, trade-offs discussed, caveats noted, alternative approaches mentioned
+- Specific values: numbers, versions, thresholds, URLs, file paths, model names
+- Omit this section only if there truly are no noteworthy details
+
+RULES:
+- This summary is a KNOWLEDGE BASE ENTRY, not a brief note. Be thorough.
+- PRESERVE verbatim: code, commands, URLs, file paths, error messages, config values, version numbers, names, amounts
+- DISCARD only: greetings, filler, the assistant explaining what it will do before doing it
+- Replace secrets (API keys, tokens, passwords) with [REDACTED]
+- Target length: 30-50% of the original conversation length. Longer conversations need longer summaries.
+- Output summary only, no preamble.`;
+
+export async function summarizeTaskOpenAI(
+  text: string,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<string> {
+  const endpoint = normalizeChatEndpoint(cfg.endpoint ?? "https://api.openai.com/v1/chat/completions");
+  const model = cfg.model ?? "gpt-4o-mini";
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    Authorization: `Bearer ${cfg.apiKey}`,
+    ...cfg.headers,
+  };
+
+  const resp = await fetch(endpoint, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      model,
+      temperature: cfg.temperature ?? 0.1,
+      max_tokens: 4096,
+      messages: [
+        { role: "system", content: TASK_SUMMARY_PROMPT },
+        { role: "user", content: text },
+      ],
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`OpenAI task-summarize failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { choices: Array<{ message: { content: string } }> };
+  return json.choices[0]?.message?.content?.trim() ?? "";
+}
 
 export async function summarizeOpenAI(
   text: string,
@@ -40,6 +114,61 @@ export async function summarizeOpenAI(
   return json.choices[0]?.message?.content?.trim() ?? "";
 }
 
+const TOPIC_JUDGE_PROMPT = `You are a conversation topic boundary detector. Given a summary of the CURRENT conversation and a NEW user message, determine if the new message starts a DIFFERENT topic/task.
+
+Answer ONLY "NEW" or "SAME".
+
+Rules:
+- "NEW" = the new message is about a completely different subject, project, or task
+- "SAME" = the new message continues, follows up on, or is closely related to the current topic
+- Follow-up questions, clarifications, refinements, bug fixes, or next steps on the same task = SAME
+- Greetings or meta-questions like "你好" or "谢谢" without new substance = SAME
+- A clearly unrelated request (e.g., current topic is deployment, new message asks about cooking) = NEW
+
+Output exactly one word: NEW or SAME`;
+
+export async function judgeNewTopicOpenAI(
+  currentContext: string,
+  newMessage: string,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<boolean> {
+  const endpoint = normalizeChatEndpoint(cfg.endpoint ?? "https://api.openai.com/v1/chat/completions");
+  const model = cfg.model ?? "gpt-4o-mini";
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    Authorization: `Bearer ${cfg.apiKey}`,
+    ...cfg.headers,
+  };
+
+  const userContent = `CURRENT CONVERSATION SUMMARY:\n${currentContext}\n\nNEW USER MESSAGE:\n${newMessage}`;
+
+  const resp = await fetch(endpoint, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      model,
+      temperature: 0,
+      max_tokens: 10,
+      messages: [
+        { role: "system", content: TOPIC_JUDGE_PROMPT },
+        { role: "user", content: userContent },
+      ],
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`OpenAI topic-judge failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { choices: Array<{ message: { content: string } }> };
+  const answer = json.choices[0]?.message?.content?.trim().toUpperCase() ?? "";
+  log.debug(`Topic judge result: "${answer}"`);
+  return answer.startsWith("NEW");
+}
+
 function normalizeChatEndpoint(url: string): string {
   const stripped = url.replace(/\/+$/, "");
   if (stripped.endsWith("/chat/completions")) return stripped;