mocode-ai 1.2.7 → 1.2.8

package/dist/context/budget.js

@@ -24,7 +24,11 @@ export const DEFAULT_BUDGET_POLICY = {
         reserve: 0.05,
     },
     hotTurnWindow: 4,
-    compactKeepRatio: 0.40,
+    // 压缩保留区:自动 15% / 强压 5%,且绝对上限 48k——摘要+最近几轮足够续工,
+    // 旧 0.40 在大窗口下导致"160k 压到 100k"的无效压缩。
+    compactKeepRatio: 0.15,
+    compactForceKeepRatio: 0.05,
+    compactKeepMaxTokens: 48000,
     pressureTriggerRatio: 0.80,
     schedulerTargetRatio: 0.80,
     estimateSafetyFactor: 1.05,

package/dist/llm/index.js

@@ -356,11 +356,16 @@ async function chatOnce(messages, handlers, signal, toolsOverride) {
     const stream = await create({
         model: getActiveModel(),
         messages: normalizeImageDetail(messages),
-        tools: activeTools,
         stream: true,
-        // 显式声明允许一次响应携带多个 tool_call(OpenAI 兼容协议标准字段)。
-        // 不设置时依赖各家后端的默认值,某些第三方网关/模型在缺省时会退化为串行单步调用。
-        ...(activeTools.length > 0 ? { parallel_tool_calls: true } : {}),
+        // 空工具表(如摘要请求)不发 tools 字段——部分网关拒绝空数组。
+        ...(activeTools.length > 0
+            ? {
+                tools: activeTools,
+                // 显式声明允许一次响应携带多个 tool_call(OpenAI 兼容协议标准字段)。
+                // 不设置时依赖各家后端的默认值,某些第三方网关/模型在缺省时会退化为串行单步调用。
+                parallel_tool_calls: true,
+            }
+            : {}),
         ...(config.maxTokens ? { max_tokens: config.maxTokens } : {}),
         ...(config.includeUsage ? { stream_options: { include_usage: true } } : {}),
     }, signal ? { signal } : undefined);

package/dist/session/compact.js

@@ -1,6 +1,6 @@
 import { chat, chatTools, correctTokenEstimate, estimatePromptTokens, estimateTokens, } from '../llm/index.js';
 import { config } from '../config/index.js';
-import { MAX_HISTORY_RESULT, MAX_MEMORY_RESULT, MAX_OLD_TOOL_STUB, MAX_SKILL_RESULT } from '../tools/constants.js';
+import { MAX_HISTORY_RESULT, MAX_MEMORY_RESULT, MAX_OLD_TOOL_STUB, MAX_SKILL_RESULT, SUMMARY_MSG_MAX_CHARS, SUMMARY_OUTPUT_MAX_CHARS, SUMMARY_TRANSCRIPT_WINDOW_RATIO, } from '../tools/constants.js';
 import { ui } from '../ui/theme.js';
 import { Spinner } from '../ui/spinner.js';
 import * as layout from '../ui/layout.js';
@@ -221,45 +221,105 @@ function microcompactGroup(g) {
     return done;
 }
 // ── 默认摘要器:复用 chat(),空 handlers 不打印 ──────────────────────────
-async function defaultSummarize(older, focus) {
-    // 摘要前剥离多模态 user 消息里的图片(base64 会撑爆摘要 prompt;image 对摘要无信息量)。
-    const stripped = older.map(stripImagesForSummary);
-    let transcript = stripped
+/** 单条消息封顶:中间段省略、保头 + 尾(头有路径/意图,尾有结论/报错),总长 ≤ max。 */
+function capMessageText(text, max) {
+    if (text.length <= max)
+        return text;
+    const removed = text.length - max;
+    const marker = `\n…[中间 ${removed} 字符已省略]…\n`;
+    let remain = max - marker.length;
+    if (remain <= 0)
+        return text.slice(0, Math.max(0, max)) + marker;
+    const head = Math.ceil(remain * 0.6);
+    return text.slice(0, head) + marker + text.slice(text.length - (remain - head));
+}
+/** tool_call 参数封顶:只保头(路径/目标等关键信息在开头,长正文对摘要无信息量)。 */
+function capArgsHead(args, max) {
+    if (args.length <= max)
+        return args;
+    return args.slice(0, max) + ` …[另有 ${args.length - max} 字符已省略]`;
+}
+/** 摘要输出硬上限:按最后换行处裁断(不留半句),防模型不听话产出巨长摘要撑大 history。 */
+function capSummaryOutput(summary) {
+    if (summary.length <= SUMMARY_OUTPUT_MAX_CHARS)
+        return summary;
+    const cut = summary.lastIndexOf('\n', SUMMARY_OUTPUT_MAX_CHARS);
+    const at = cut > SUMMARY_OUTPUT_MAX_CHARS * 0.5 ? cut : SUMMARY_OUTPUT_MAX_CHARS;
+    return summary.slice(0, at) + '\n[摘要输出超长,已截断]';
+}
+/** 按封顶配额拼转录。scale ∈ (0,1] 等比缩小各角色配额(总预算不足时用)。 */
+function buildTranscript(stripped, scale) {
+    const caps = SUMMARY_MSG_MAX_CHARS;
+    const capFor = (role) => Math.max(200, Math.floor(scale *
+        (role === 'user'
+            ? caps.user
+            : role === 'assistant'
+                ? caps.assistant
+                : role === 'tool'
+                    ? caps.tool
+                    : caps.other)));
+    return stripped
         .map((m) => {
-        const role = m.role;
-        let line = `${role}: ${toText(m.content)}`;
+        const cap = capFor(m.role);
+        let line = `${m.role}: ${capMessageText(toText(m.content), cap)}`;
         const tcs = m.tool_calls;
         if (tcs) {
             for (const tc of tcs) {
-                line += `\n  [tool_call ${tc?.function?.name}] ${tc?.function?.arguments ?? ''}`;
+                const args = tc?.function?.arguments ?? '';
+                line += `\n  [tool_call ${tc?.function?.name}] ${capArgsHead(args, Math.max(200, Math.floor(scale * caps.tool)))}`;
             }
         }
         return line;
     })
         .join('\n');
-    // 防摘要提示本身溢出:超 60% 窗口就先中截到 50%
-    if (estimateTokens(transcript) >
-        Math.floor(config.contextWindowTokens * 0.6)) {
-        transcript = truncateMid(transcript, Math.floor(config.contextWindowTokens * 0.5));
+}
+async function defaultSummarize(older, focus) {
+    // 摘要前剥离多模态 user 消息里的图片(base64 会撑爆摘要 prompt;image 对摘要无信息量)。
+    const stripped = older.map(stripImagesForSummary);
+    // 触发压缩时旧区约占窗口 50-60%,摘要器共享同一窗口——逐条封顶后旧区原文即可装入,
+    // 不需要"先压一遍再摘要"。封顶策略:先按满额封顶;总量仍超窗口 55% 预算时等比缩小
+    // 配额重拼(优先保条数/每轮都有代表,其次保单条长度);再超才整段中截兜底(罕见)。
+    let transcript = buildTranscript(stripped, 1);
+    const tokenBudget = Math.floor(config.contextWindowTokens * SUMMARY_TRANSCRIPT_WINDOW_RATIO);
+    let tokens = estimateTokens(transcript);
+    if (tokens > tokenBudget) {
+        transcript = buildTranscript(stripped, tokenBudget / tokens);
+        tokens = estimateTokens(transcript);
+        if (tokens > Math.floor(config.contextWindowTokens * 0.6)) {
+            transcript = truncateMid(transcript, Math.floor(config.contextWindowTokens * 0.5));
+        }
     }
     const sysMsg = {
         role: 'system',
-        content: 'You are a session summarizer. Output only the summary body, max 300 words, preserving: the user\'s core request; files read/written/modified and key changes; source artifact IDs/hashes; key commands run and their result highlights; decisions made; current task progress and next step; open questions. Do not recap every detail.',
+        content: 'You are an aggressive session compressor writing a handoff note for an agent that lost its context. ' +
+            'The agent will continue the task with ONLY your summary plus a few most-recent messages, so your summary is the sole memory of everything older.\n' +
+            'Output ONLY the summary body in this exact structure (omit empty sections):\n' +
+            '## Objective — the user\'s core request(s); cover EVERY distinct user request in the transcript, in order, noting which are completed vs pending.\n' +
+            '## Completed — what is already done: files created/modified (exact paths), key change per file, commands run and their outcomes (pass/fail, key numbers), decisions made and why.\n' +
+            '## In Progress — what is being worked on right now and exactly where it stopped (e.g. "edit applied to foo.ts, test not yet run").\n' +
+            '## Next Steps — the concrete next actions in order.\n' +
+            '## Key Facts — only what later steps cannot work without: exact paths, symbols/API shapes, artifact IDs/hashes, constraints, open questions, failed approaches and errors to avoid repeating.\n' +
+            'What counts as important (keep): user requests and constraints; final state of each modified file; conclusions and results, not the steps that led there; decisions with reasons; precise references (paths, symbols, hashes, commands) that later steps must cite; failures and what was tried, so mistakes are not repeated.\n' +
+            'What to drop: verbatim file contents and tool-output dumps, step-by-step recaps, exploration dead-ends that led nowhere, polite chatter, anything re-derivable by re-reading files.\n' +
+            'Rules: total ≤ 400 words. State conclusions and locations (path:line where useful), never paste content. ' +
+            'Never invent facts, paths, or results that are not in the transcript; if unsure whether something happened, omit it.',
     };
     const userMsg = {
         role: 'user',
         content: focus
-            ? `请将以下会话历史压缩成摘要,重点保留与「${focus}」相关的事实/决策/文件改动:\n\n${transcript}\n\n摘要:`
-            : `请将以下会话历史压缩成摘要:\n\n${transcript}\n\n摘要:`,
+            ? `请将以下会话历史压缩成交接摘要,重点保留与「${focus}」相关的事实/决策/文件改动:\n\n${transcript}\n\n摘要:`
+            : `请将以下会话历史压缩成交接摘要:\n\n${transcript}\n\n摘要:`,
     };
     const spinner = new Spinner((msg, frame) => layout.setStatus(msg, frame ?? undefined));
     spinner.start('压缩中');
     try {
-        const r = await chat([sysMsg, userMsg], {}); // 空 handlers:不打印、不外显流式
+        // 空 handlers:不打印、不外显流式;tools=[] 不带工具表——摘要纯文本任务,
+        // 全量工具 schema 白占几千 token 窗口,还诱导幻觉工具调用。
+        const r = await chat([sysMsg, userMsg], {}, undefined, []);
         // 推理模型可能只返 reasoning_content(content 为 null),或幻觉出 tool_calls → 视为失败
         if (r.toolCalls.length > 0 || !r.content)
             return null;
-        return r.content;
+        return capSummaryOutput(r.content);
     }
     finally {
         spinner.stop();
@@ -277,7 +337,12 @@ export async function compactHistory(history, opts) {
     state.lastEstimate = estimateBefore;
     const groups = groupFromEnd(history);
     // 保近期:按策略中的 token 比例累积(至少保 1 组),永不劈开 group。
-    const keepBudget = Math.floor(opts.window * DEFAULT_BUDGET_POLICY.compactKeepRatio);
+    // force(手动 /compact 默认、硬闸触发)用更激进的保留比例;再加绝对上限,
+    // 防大窗口(如 256k)下保留区按比例仍过大——压缩目标 = 摘要 + 最小续工上下文。
+    const keepRatio = opts.force
+        ? DEFAULT_BUDGET_POLICY.compactForceKeepRatio
+        : DEFAULT_BUDGET_POLICY.compactKeepRatio;
+    const keepBudget = Math.min(Math.floor(opts.window * keepRatio), DEFAULT_BUDGET_POLICY.compactKeepMaxTokens);
     const kept = [];
     let keptTokens = 0;
     for (let k = groups.length - 1; k >= 0; k--) {
@@ -360,18 +425,10 @@ export async function compactHistory(history, opts) {
         }
         return { ...noop, reason: 'noop-protected', protectedRatio };
     }
-    // 第一层:微压缩——旧区原地截短(保 tool_call_id,无 LLM 调用)
-    // 覆盖三类大字段,均只裁模型/工具产物,不动 user 原话与 system(摘要):
-    //   ① tool 结果 content;
-    //   ② 旧 assistant 的 tool_calls.arguments —— provenance stub:整体超长才进,大字段
-    //     替换为 "<N 字符,已省略>"(保 path + JSON 合法,见 stubToolCallArguments);
-    //   ③ 旧 assistant 正文 content(模型长解释,回看价值低)。
-    let microcompactDone = false;
-    for (const g of oldGroups) {
-        if (microcompactGroup(g))
-            microcompactDone = true;
-    }
-    // 第二层:摘要——把旧区(微压缩后)压成一条 system 摘要
+    // 主层:摘要——把旧区【原始内容】压成一条 system 摘要。
+    // 触发压缩时旧区约占窗口 50-60%,摘要请求与摘要器共享窗口,逐条封顶后即装得下;
+    // 不在这里预跑微压缩——预截断会让摘要模型看到 600 字符残片,等于弄瞎它,
+    // 且摘要成功后旧区整体丢弃,预截断本身也无收益。
     const older = flattenGroups(oldGroups);
     const summarizeFn = opts.summarize ?? defaultSummarize;
     let summary = null;
@@ -379,7 +436,7 @@ export async function compactHistory(history, opts) {
         summary = await summarizeFn(older, opts.focus);
     }
     catch {
-        summary = null; // 摘要失败 → 回退仅微压缩,不崩
+        summary = null; // 摘要失败 → 回退微压缩兜底,不崩
     }
     if (summary) {
         const artifactRefs = collectArtifactRefs(older);
@@ -417,7 +474,15 @@ export async function compactHistory(history, opts) {
             reason: 'summarize',
         };
     }
-    // 摘要失败:回退仅微压缩(tool content 已原地改),结构不动
+    // 摘要失败兜底:微压缩——旧区原地截短(保 tool_call_id,无 LLM 调用),结构不动。
+    // 覆盖三类大字段,均只裁模型/工具产物,不动 user 原话与 system:
+    //   ① tool 结果 content;② 旧 assistant 的 tool_calls.arguments(provenance stub,
+    //     整体超长才进,大字段替换为 "<N 字符,已省略>",保 path + JSON 合法);③ 旧 assistant 正文。
+    let microcompactDone = false;
+    for (const g of oldGroups) {
+        if (microcompactGroup(g))
+            microcompactDone = true;
+    }
     const estimateAfter = estimatePromptTokens(history, activeTools, state.correction);
     state.lastEstimate = estimateAfter;
     state.lastUsage = undefined; // token 数已变,旧 usage 失效

package/dist/tools/constants.js

@@ -10,6 +10,20 @@ export const MAX_HISTORY_RESULT = 8000;
 export const MAX_SKILL_RESULT = 64000;
 /** 微压缩时旧工具结果截到的存根长度(字符)。 */
 export const MAX_OLD_TOOL_STUB = 600;
+// ── 上下文压缩摘要(见 session/compact.ts defaultSummarize)────────────────
+/** 摘要输入单条消息封顶:逐条封顶保证每轮对话都有代表,替代整段中截(会切掉中间整轮)。
+ *  转录总量超 SUMMARY_TRANSCRIPT_WINDOW_RATIO 时,这些封顶会被等比缩小重拼。 */
+export const SUMMARY_MSG_MAX_CHARS = {
+    user: 3000, // 用户原话价值最高,给大额度
+    assistant: 1600,
+    tool: 800,
+    other: 2000,
+};
+/** 摘要转录总预算占上下文窗口的比例:超预算时等比缩小单条封顶重拼——
+ *  优先保消息条数(每轮都有代表),其次才保单条长度。 */
+export const SUMMARY_TRANSCRIPT_WINDOW_RATIO = 0.55;
+/** 摘要输出硬上限(字符):模型不听话产出超长摘要时按段落边界裁,防摘要本身撑大 history。 */
+export const SUMMARY_OUTPUT_MAX_CHARS = 6000;
 // ── 记忆(Tier-2 JSONL 工具库)──────────────────────────────────────────────
 /** 单条记忆 body 上限(字符)。进 history 前 memory_search 结果另有 MAX_MEMORY_RESULT 兜底。 */
 export const MAX_MEMORY_ENTRY = 4000;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mocode-ai",
-  "version": "1.2.7",
+  "version": "1.2.8",
   "description": "终端编码 agent:LLM + tool-call 循环 + 流式输出(含思考)+ 16 个工具,接任意 OpenAI 兼容后端。",
   "type": "module",
   "bin": {