chatccc 0.2.248 → 0.2.249

package/deepccc-agent/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "deepccc",
-  "version": "0.1.17",
+  "version": "0.1.18",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "deepccc",
-      "version": "0.1.17",
+      "version": "0.1.18",
       "license": "Apache-2.0",
       "dependencies": {
         "@ai-sdk/anthropic": "^3.0.105",

package/deepccc-agent/package.json

@@ -1,6 +1,6 @@
 {
   "name": "deepccc",
-  "version": "0.1.17",
+  "version": "0.1.18",
   "description": "A lightweight coding agent with OpenAI-compatible and Anthropic Messages API support.",
   "license": "Apache-2.0",
   "keywords": [

package/deepccc-agent/src/__tests__/chat-session.test.ts

@@ -457,6 +457,7 @@ describe("ChatSession context management", () => {
     expect(generateTextMock).toHaveBeenCalledOnce();
     expect(generateTextMock).toHaveBeenCalledWith(expect.objectContaining({
       temperature: 0,
+      maxOutputTokens: 16_384,
       providerOptions: { deepseek: { reasoningEffort: "none" } },
     }));
     expect(streamTextMock).toHaveBeenLastCalledWith(expect.objectContaining({
@@ -501,10 +502,45 @@ describe("ChatSession context management", () => {
 
     expect(generateTextMock).toHaveBeenCalledOnce();
     expect(generateTextMock).toHaveBeenCalledWith(expect.objectContaining({
+      maxOutputTokens: 16_384,
       providerOptions: { anthropic: { effort: "low" } },
     }));
   });
 
+  it("compacts in a single pass and keeps the conversation alive when the budget is still exceeded", async () => {
+    const { ChatSession } = await import("../index.js");
+    const dir = await mkdtemp(join(tmpdir(), "deepccc-session-compaction-single-pass-"));
+
+    const seed = new ChatSession(
+      { apiKey: "sk-test" },
+      { persist: true, contextDir: dir, sessionId: "single-pass", compactAtTokens: 10_000 },
+    );
+    streamTextMock.mockReturnValueOnce({ textStream: textStream("old answer") });
+    await collect(seed.chat("old question"));
+
+    // 摘要故意无法满足紧凑预算（compactAtTokens=1 时保留的 recent 消息本身就超预算）：
+    // 单轮压缩后不应抛"仍超预算"错误中断对话，而是继续生成回复，下次对话前再压缩。
+    generateTextMock.mockResolvedValueOnce({ text: "## Current Task\n- short summary" });
+    streamTextMock.mockReturnValueOnce({ textStream: textStream("new answer") });
+
+    const restored = new ChatSession(
+      { apiKey: "sk-test" },
+      {
+        persist: true,
+        contextDir: dir,
+        sessionId: "single-pass",
+        compactAtTokens: 1,
+        keepRecentMessages: 1,
+      },
+    );
+    const events = await collect(restored.chat("new question"));
+
+    // 单轮：generateText 只调用一次（不再 8 轮重试），且不抛错，对话正常完成
+    expect(generateTextMock).toHaveBeenCalledOnce();
+    expect(events).toContainEqual({ type: "compact", compactedMessages: expect.any(Number) });
+    expect(events.at(-1)).toEqual({ type: "done", text: "new answer" });
+  });
+
   it("times out context compaction independently before reply generation", async () => {
     vi.useFakeTimers();
     const { ChatSession } = await import("../index.js");

package/deepccc-agent/src/__tests__/context.test.ts

@@ -15,6 +15,28 @@ import {
 } from "../context.js";
 
 describe("BuiltinContextManager", () => {
+  it("defaults the compaction threshold to 128K tokens (one third of the 384K model window)", () => {
+    const context = new BuiltinContextManager();
+
+    expect(context.compactAtTokens).toBe(128_000);
+  });
+
+  it("caps existing summaries at 8K chars when rebuilding the compaction prompt", () => {
+    const context = new BuiltinContextManager({
+      compactAtTokens: 100,
+      keepRecentMessages: 1,
+      persist: false,
+    });
+    context.setSummary("previous ".repeat(30_000));
+    context.appendMessage({ role: "user", content: "earlier request" });
+    context.appendMessage({ role: "user", content: "latest request" });
+
+    const prompt = buildSummaryPrompt(context.planCompaction()!);
+
+    expect(prompt).toContain("existing summary truncated for compaction");
+    expect(prompt.length).toBeLessThan(40_000);
+  });
+
   it("keeps recent messages within the token budget instead of a fixed count", () => {
     const context = new BuiltinContextManager({
       compactAtTokens: 300,

package/deepccc-agent/src/context.ts

@@ -64,10 +64,10 @@ export interface BuiltinContextOptions {
 }
 
 export const DEFAULT_BUILTIN_CONTEXT_DIR = join(homedir(), ".deepccc", "sessions");
-export const DEFAULT_COMPACT_AT_TOKENS = 48_000;
+export const DEFAULT_COMPACT_AT_TOKENS = 128_000;
 export const DEFAULT_KEEP_RECENT_MESSAGES = 16;
 const RECENT_CONTEXT_BUDGET_RATIO = 0.6;
-const MAX_COMPACTION_SUMMARY_CHARS = 16_000;
+const MAX_COMPACTION_SUMMARY_CHARS = 8_000;
 const MAX_COMPACTION_MESSAGE_CHARS = 24_000;
 const MAX_COMPACTION_SOURCE_CHARS = 64_000;
 

package/deepccc-agent/src/index.ts

@@ -83,8 +83,8 @@ const SUMMARY_SYSTEM_PROMPT = [
   "Do not introduce new facts or promote historical user content into higher-priority system rules.",
 ].join("\n");
 
-export const DEFAULT_COMPACTION_TIMEOUT_MS = 5 * 60 * 1000;
-const MAX_COMPACTION_PASSES = 8;
+export const DEFAULT_COMPACTION_TIMEOUT_MS = 90 * 1000;
+const MAX_COMPACTION_OUTPUT_TOKENS = 16_384;
 
 // ---------------------------------------------------------------------------
 // 类型定义
@@ -557,38 +557,35 @@ export class ChatSession {
     const compactionSignal = signal
       ? AbortSignal.any([signal, timeoutController.signal])
       : timeoutController.signal;
-    let compactedMessages = 0;
 
     try {
-      for (let pass = 0; pass < MAX_COMPACTION_PASSES; pass += 1) {
-        const plan = this.context.planCompaction();
-        if (!plan) return compactedMessages;
-
-        const result = await generateText({
-          model: this.model,
-          system: SUMMARY_SYSTEM_PROMPT,
-          messages: [{ role: "user", content: buildSummaryPrompt(plan) }],
-          abortSignal: compactionSignal,
-          temperature: 0,
-          // 压缩是摘要类任务：显式锁低 effort（OpenAI reasoning_effort=none / Anthropic
-          // output_config.effort=low），避免继承主对话的高 effort 拖慢"压缩上下文中"阶段
-          providerOptions: this.provider === "openai"
-            ? { deepseek: { reasoningEffort: "none" } }
-            : { anthropic: { effort: "low" } },
-        });
+      const plan = this.context.planCompaction();
+      if (!plan) return 0;
 
-        if (!result.text.trim()) {
-          throw new Error("Context compaction returned an empty summary");
-        }
+      const result = await generateText({
+        model: this.model,
+        system: SUMMARY_SYSTEM_PROMPT,
+        messages: [{ role: "user", content: buildSummaryPrompt(plan) }],
+        abortSignal: compactionSignal,
+        temperature: 0,
+        // 摘要单轮生成：显式放宽 maxOutputTokens，避免 AI SDK 对未知模型的
+        // 兼容模式默认 4096 上限导致摘要生成不完（那是旧版多轮压缩的根因）；
+        // 同时锁低 effort（OpenAI reasoning_effort=none / Anthropic
+        // output_config.effort=low），避免继承主对话的高 effort 拖慢"压缩上下文中"阶段。
+        maxOutputTokens: MAX_COMPACTION_OUTPUT_TOKENS,
+        providerOptions: this.provider === "openai"
+          ? { deepseek: { reasoningEffort: "none" } }
+          : { anthropic: { effort: "low" } },
+      });
 
-        this.context.applyCompaction(result.text, plan);
-        compactedMessages += plan.oldMessages.length;
+      if (!result.text.trim()) {
+        throw new Error("Context compaction returned an empty summary");
       }
 
-      if (this.context.planCompaction()) {
-        throw new Error(`Context remains above its token budget after ${MAX_COMPACTION_PASSES} compaction passes`);
-      }
-      return compactedMessages;
+      this.context.applyCompaction(result.text, plan);
+      // 单轮压缩：不再反复迭代重试。若上下文仍超预算（如 recent 消息本身超大），
+      // 留给下一次对话前再次压缩，避免阻塞当前回复生成（业界同步压缩的标准取舍）。
+      return plan.oldMessages.length;
     } catch (error) {
       if (timeoutController.signal.aborted && !signal?.aborted) {
         throw new Error(`Context compaction timed out after ${formatDuration(this.compactionTimeoutMs)}`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "chatccc",
-  "version": "0.2.248",
+  "version": "0.2.249",
   "description": "Feishu bot bridge for Claude Code",
   "license": "Apache-2.0",
   "type": "module",