mocode-ai 1.2.6 → 1.2.8

package/dist/skills/skill-improve.js

@@ -0,0 +1,221 @@
+// skill description 进化循环(自进化 Phase 1b)。
+//
+// 协议移植自 skill-creator 的 run_loop.py(调研见 docs/skill-self-evolution-research.md Part 1.2):
+//  - 分层切 train/holdout(默认 0.4)防过拟合
+//  - 每轮:LLM 读失败案例 → 提案新 description → 评测 → 接受门
+//  - 接受门(保守双条件):train 严格提升 且 holdout 不降;holdout 为空时退化为全集严格提升
+//  - 拒绝缓冲:被拒的候选记下来,后续提案明确告知「这些已试过且更差」,防重复打转
+//  - 迭代上限(默认 5);全程记录 history,返回最佳 description
+//
+// 安全边界(调研 Part 3):
+//  - 默认 dry-run(只打印不落盘);--apply 才写回 SKILL.md,且写前再显式确认一次
+//  - builtin skill 无磁盘载体,拒绝(入口在 skill-eval.loadSkillForEval)
+//  - 进化对象仅 description 一行;正文/脚本/执行面字段不在自动编辑范围
+//
+// 依赖:skill-eval.ts(评测引擎 + 纯函数)+ llm.chat(提案,单轮无工具)。
+import { readFileSync, writeFileSync } from 'node:fs';
+import { chat } from '../llm/index.js';
+import { runTriggerEval, splitTriggerSet, applyDescription, MAX_DESCRIPTION_CHARS, } from './skill-eval.js';
+/** 提案 LLM 的系统指令:确定性模板(不用 t(),保证跨语言可比、可单测断言)。 */
+export function buildImprovePrompt(skill, currentDescription, failingCases, rejected) {
+    const body = (skill.body ?? '').slice(0, 1500);
+    const failLines = failingCases
+        .map((c) => {
+        const expect = c.should_trigger ? 'SHOULD trigger' : 'should NOT trigger';
+        return `- "${c.query}" → ${expect}, actual trigger rate ${c.trigger_rate.toFixed(2)}`;
+    })
+        .join('\n');
+    const rejectedBlock = rejected.length
+        ? `\nThese descriptions were already tried and performed WORSE. Do not propose any of them or near-duplicates:\n${rejected
+            .map((d) => `- ${d}`)
+            .join('\n')}\n`
+        : '';
+    return [
+        `You are optimizing the frontmatter "description" field of an agent skill named "${skill.name}".`,
+        'The description is the ONLY text the model sees when deciding whether to trigger the skill.',
+        `The skill body (what it actually does) begins with:\n---\n${body}\n---`,
+        `Current description:\n${currentDescription}`,
+        'Test results with the current description (failures to fix):',
+        failLines,
+        rejectedBlock,
+        'Propose ONE improved description that makes the skill trigger for the SHOULD-trigger cases and stay silent for the should-NOT-trigger cases.',
+        'Rules:',
+        '- Output ONLY the description text, nothing else (no quotes, no "description:", no markdown).',
+        '- One line only, 100-400 characters.',
+        '- Write it as a routing rule: state WHAT it does and WHEN to use it (concrete triggers/phrases), not as documentation.',
+        '- Preserve the language of the current description (English stays English, Chinese stays Chinese).',
+    ].join('\n');
+}
+/**
+ * 从 LLM 输出里提取候选 description(纯函数):
+ * 去首尾空白 → 若以 "description:" 开头则剥掉 → 只取第一行 → 去配对引号 → 校验长度。
+ * 非法(空 / 超长)返 null。
+ */
+export function extractDescriptionProposal(text, maxChars = MAX_DESCRIPTION_CHARS) {
+    let t = (text ?? '').trim();
+    if (!t)
+        return null;
+    t = t.replace(/^description\s*[:：]\s*/i, '');
+    let first = t.split(/\r?\n/)[0].trim();
+    if (!first)
+        return null;
+    if (first.length >= 2 && (first[0] === '"' || first[0] === "'") && first[first.length - 1] === first[0]) {
+        first = first.slice(1, -1).trim();
+    }
+    if (!first || first.length > maxChars)
+        return null;
+    return first;
+}
+/** 接受门(纯函数):train 严格提升 且 holdout 不降;holdout 为空时全集严格提升。 */
+export function shouldAcceptCandidate(base, cand) {
+    const trainUp = cand.trainPass > base.trainPass;
+    if (!trainUp) {
+        return { accept: false, reason: `train 未提升 (${cand.trainPass}/${cand.trainTotal} ≤ ${base.trainPass}/${base.trainTotal})` };
+    }
+    if (base.holdoutPass === null || cand.holdoutPass === null) {
+        return { accept: true, reason: `train 提升且无 holdout(${base.trainPass}→${cand.trainPass})` };
+    }
+    if (cand.holdoutPass < base.holdoutPass) {
+        return { accept: false, reason: `holdout 回退 (${cand.holdoutPass} < ${base.holdoutPass}),过拟合 train` };
+    }
+    return { accept: true, reason: `train ${base.trainPass}→${cand.trainPass},holdout ${base.holdoutPass}→${cand.holdoutPass}(不降)` };
+}
+/** 把报告按 query 集合拆成 train / holdout 的分项(纯函数)。 */
+function partitionReport(report, trainQueries) {
+    const train = report.results.filter((r) => trainQueries.has(r.query));
+    const holdout = report.results.filter((r) => !trainQueries.has(r.query));
+    return {
+        trainPass: train.filter((r) => r.pass).length,
+        trainTotal: train.length,
+        holdoutPass: holdout.length ? holdout.filter((r) => r.pass).length : null,
+        holdoutTotal: holdout.length,
+    };
+}
+function failingCases(report) {
+    return report.results
+        .filter((r) => !r.pass)
+        .map((r) => ({ query: r.query, should_trigger: r.should_trigger, trigger_rate: r.trigger_rate }));
+}
+/** 调 LLM 提案一次(单轮、无工具、超时 abort)。失败抛错由调用方捕获。 */
+async function proposeDescription(prompt, timeoutMs, signal) {
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(), timeoutMs);
+    // AbortSignal.any 需 Node ≥20;engines>=18,手动桥接外部 signal。
+    const externalAbort = () => controller.abort();
+    if (signal) {
+        if (signal.aborted)
+            controller.abort();
+        else
+            signal.addEventListener('abort', externalAbort, { once: true });
+    }
+    try {
+        const messages = [{ role: 'user', content: prompt }];
+        const res = await chat(messages, {}, controller.signal);
+        return extractDescriptionProposal(res.content ?? '');
+    }
+    finally {
+        clearTimeout(timer);
+        signal?.removeEventListener('abort', externalAbort);
+    }
+}
+/**
+ * 进化主循环。evalSet 必须非空(调用方保证;无 evals 的 skill 在 CLI 层拒绝——
+ * 「没验证门的 skill 禁止自动落盘」,调研共识 1)。
+ */
+export async function runImproveLoop(skill, evalSet, opts = {}) {
+    const maxIterations = opts.maxIterations ?? 5;
+    const runsPerQuery = opts.runsPerQuery ?? 3;
+    const threshold = opts.threshold ?? 0.5;
+    const { train, holdout } = splitTriggerSet(evalSet, opts.holdout ?? 0.4, opts.seed ?? 42);
+    if (train.length === 0)
+        throw new Error('train 集为空:eval 集太小,无法进化(至少需要 2 条且覆盖触发/不触发)');
+    const trainQueries = new Set(train.map((t) => t.query));
+    const evalOpts = { runsPerQuery, threshold, timeoutMs: opts.timeoutMs, signal: opts.signal };
+    let baseDescription = skill.description;
+    let baseReport = await runTriggerEval(skill, baseDescription, evalSet, runsPerQuery, threshold, evalOpts);
+    let baseParts = partitionReport(baseReport, trainQueries);
+    const iterations = [];
+    const rejected = [];
+    let bestDescription = baseDescription;
+    let bestKey = [baseParts.trainPass, baseParts.holdoutPass ?? 0];
+    const earlyExit = () => baseParts.trainTotal > 0 &&
+        baseParts.trainPass === baseParts.trainTotal &&
+        (baseParts.holdoutPass === null || (baseParts.holdoutTotal > 0 && baseParts.holdoutPass === baseParts.holdoutTotal));
+    if (earlyExit()) {
+        opts.onIteration?.(0, maxIterations, '当前 description 在 train + holdout 全通过,无需进化。');
+    }
+    for (let iter = 1; iter <= maxIterations && !earlyExit(); iter++) {
+        opts.onIteration?.(iter, maxIterations, `第 ${iter}/${maxIterations} 轮:提案新 description…`);
+        const prompt = buildImprovePrompt(skill, baseDescription, failingCases(baseReport), rejected);
+        let candidate;
+        try {
+            candidate = await proposeDescription(prompt, opts.proposeTimeoutMs ?? 120_000, opts.signal);
+        }
+        catch (e) {
+            const msg = e instanceof Error ? e.message : String(e);
+            opts.onIteration?.(iter, maxIterations, `提案失败(${msg.slice(0, 120)}),终止。`);
+            break;
+        }
+        if (!candidate || candidate === baseDescription) {
+            opts.onIteration?.(iter, maxIterations, '提案无效(空/超长/与当前相同),终止。');
+            break;
+        }
+        const candReport = await runTriggerEval(skill, candidate, evalSet, runsPerQuery, threshold, evalOpts);
+        const candParts = partitionReport(candReport, trainQueries);
+        const gate = shouldAcceptCandidate(baseParts, candParts);
+        const iteration = {
+            iteration: iter,
+            description: candidate,
+            trainPass: candParts.trainPass,
+            trainTotal: candParts.trainTotal,
+            holdoutPass: candParts.holdoutPass,
+            holdoutTotal: candParts.holdoutTotal,
+            accepted: gate.accept,
+            reason: gate.reason,
+        };
+        iterations.push(iteration);
+        opts.onIteration?.(iter, maxIterations, `${gate.accept ? '✓ 接受' : '· 拒绝'}: train ${candParts.trainPass}/${candParts.trainTotal}, holdout ${candParts.holdoutPass ?? '-'}/${candParts.holdoutTotal} — ${gate.reason}`);
+        if (gate.accept) {
+            baseDescription = candidate;
+            baseReport = candReport;
+            baseParts = candParts;
+            const key = [candParts.trainPass, candParts.holdoutPass ?? 0];
+            if (key[0] > bestKey[0] || (key[0] === bestKey[0] && key[1] > bestKey[1])) {
+                bestKey = key;
+                bestDescription = candidate;
+            }
+        }
+        else {
+            rejected.push(candidate);
+            if (rejected.length > 5)
+                rejected.shift();
+        }
+    }
+    return {
+        skill: skill.name,
+        originalDescription: skill.description,
+        bestDescription,
+        improved: bestDescription !== skill.description,
+        iterations,
+    };
+}
+/**
+ * 把进化结果写回 SKILL.md 的 description(唯一的落盘点;CLI --apply 时才调)。
+ * 返回写后的文件内容;frontmatter 无法安全替换时抛错(不落盘)。
+ */
+export function applyImprovedDescription(skill, description) {
+    if (skill.dir === 'builtin')
+        throw new Error('内置 skill 不可修改');
+    if (description.length > MAX_DESCRIPTION_CHARS)
+        throw new Error(`description 超长(>${MAX_DESCRIPTION_CHARS})`);
+    const original = readFileSync(skill.skillMdPath, 'utf8');
+    const next = applyDescription(original, description);
+    if (next === null) {
+        throw new Error('无法定位 SKILL.md frontmatter 的 description 字段,拒绝写回(请手工编辑)');
+    }
+    if (next === original) {
+        throw new Error('写回无变化(description 与当前一致)');
+    }
+    writeFileSync(skill.skillMdPath, next, 'utf8');
+    return next;
+}

package/dist/skills/stats.js

@@ -0,0 +1,102 @@
+// skill 使用台账(自进化 Phase 0)。
+// 为什么不在 trace.jsonl 上做:工具事件出于隐私只存参数指纹(sha256/keys,
+// trace-sanitize.ts 刻意不存值),拿不到 skill name。台账因此在工具层直接记录——
+// use_skill / run_skill 是唯一知道真实 skill name 的落点。
+//
+// 落盘:<cwd>/.mocode/skill-stats.jsonl(append-only JSONL,每行一次使用)。
+// 纯观测:任何写失败静默吞掉,绝不阻断 agent 主流程(风格对齐 session/trace.ts)。
+// 聚合是纯函数(aggregateSkillStats),吃记录数组吐按 skill 的计数,便于单测。
+import { appendFileSync, existsSync, mkdirSync, readFileSync } from 'node:fs';
+import path from 'node:path';
+/**
+ * 台账路径(项目级,与 sessions 同根;使用是项目上下文相关的,不写全局)。
+ * baseDir 可选:测试指向临时目录用;缺省 process.cwd()。
+ */
+export function skillStatsPath(baseDir = process.cwd()) {
+    return path.join(baseDir, '.mocode', 'skill-stats.jsonl');
+}
+/**
+ * 追加一条台账记录;任何失败静默(观测不得阻断主流程)。
+ * MOCODE_SKILL_EVAL=1(触发评测进程内设置)时跳过:评测里的人工构造调用
+ * 是测量手段不是真实使用,记入会污染自进化的输入信号。
+ */
+export function recordSkillUsage(rec, baseDir = process.cwd()) {
+    if (process.env.MOCODE_SKILL_EVAL === '1')
+        return;
+    try {
+        const p = skillStatsPath(baseDir);
+        const dir = path.dirname(p);
+        if (!existsSync(dir))
+            mkdirSync(dir, { recursive: true });
+        appendFileSync(p, JSON.stringify(rec) + '\n', 'utf8');
+    }
+    catch {
+        // 观测失败静默
+    }
+}
+/** 读取台账原始记录;文件不存在 / 单行损坏 → 跳过(不抛)。 */
+export function loadSkillUsage(baseDir = process.cwd()) {
+    try {
+        const content = readFileSync(skillStatsPath(baseDir), 'utf8');
+        const out = [];
+        for (const line of content.split('\n')) {
+            const s = line.trim();
+            if (!s)
+                continue;
+            try {
+                const v = JSON.parse(s);
+                if (v && typeof v.skill === 'string' && (v.kind === 'use' || v.kind === 'run')) {
+                    out.push(v);
+                }
+            }
+            catch {
+                // 损坏行跳过
+            }
+        }
+        return out;
+    }
+    catch {
+        return [];
+    }
+}
+/**
+ * 纯聚合:记录 → 按 skill 的计数视图。按 skill 名分组(大小写敏感),
+ * lastUsedAt 取 ts 字符串字典序最大(ISO 时间戳字典序 == 时间序)。
+ * 输入乱序也安全。空输入返回 []。
+ */
+export function aggregateSkillStats(records) {
+    const bySkill = new Map();
+    for (const r of records) {
+        const list = bySkill.get(r.skill);
+        if (list)
+            list.push(r);
+        else
+            bySkill.set(r.skill, [r]);
+    }
+    const out = [];
+    for (const [skill, list] of bySkill) {
+        const runs = list.filter((r) => r.kind === 'run');
+        const runSuccess = runs.filter((r) => r.status === 'success').length;
+        let lastFailure = null;
+        let lastUsedAt = '';
+        for (const r of list) {
+            if (typeof r.ts === 'string' && r.ts > lastUsedAt)
+                lastUsedAt = r.ts;
+            if (r.status !== 'success' && (!lastFailure || r.ts > lastFailure.ts)) {
+                lastFailure = { ts: r.ts, status: r.status, code: r.code };
+            }
+        }
+        out.push({
+            skill,
+            total: list.length,
+            uses: list.length - runs.length,
+            runs: runs.length,
+            runSuccessRate: runs.length ? runSuccess / runs.length : null,
+            lastFailure,
+            lastUsedAt,
+        });
+    }
+    // 按最近使用倒序,让 /skills 徽标与人工浏览都「最活跃的在前」。
+    out.sort((a, b) => (a.lastUsedAt < b.lastUsedAt ? 1 : a.lastUsedAt > b.lastUsedAt ? -1 : 0));
+    return out;
+}

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/dist/tools/registry.js

@@ -220,7 +220,11 @@ export async function executeToolOutcome(name, argsRaw, signal, opts) {
     }
     const validation = validateToolArguments(tool, parsed);
     if (!validation.valid) {
-        return terminalOutcome('error', validation.code, `错误:工具 ${name} 参数无效: ${validation.message}`, startedAt);
+        const hint = opts?.argumentErrorHint?.trim();
+        const message = hint
+            ? `错误:工具 ${name} 参数无效: ${validation.message}\n${hint}`
+            : `错误:工具 ${name} 参数无效: ${validation.message}`;
+        return terminalOutcome('error', validation.code, message, startedAt);
     }
     const args = parsed;
     const sandboxError = enforceSandbox(name, args);

package/dist/ui/layout.js

@@ -671,6 +671,17 @@ export function contentDeleteFrom(startIdx, n) {
 export function totalRows() {
     return content.totalRows();
 }
+/** 缓冲尾部(已提交行)是否已经是空白行(去掉 ANSI 后无可见字符)。
+ *  供 compact 等在 step 循环顶部写通知行前判断是否需要补空行分隔。 */
+export function isLastContentRowBlank() {
+    const committed = content.committedRows();
+    if (committed === 0)
+        return false;
+    const line = content.lineAt(committed - 1);
+    if (line === null)
+        return false;
+    return line.replace(/\x1b\[[0-9;]*m/g, '').trim().length === 0;
+}
 /** 正文→mutation 首摘要前，把尾部间距强制归一为一条视觉空行。 */
 export function normalizeMutationBoundary() {
     if (!active || !ui.isTTY)

package/dist/verification/prompt.js

@@ -0,0 +1,55 @@
+import path from 'node:path';
+import { discoverPackageValidationCommands } from './discovery.js';
+import { discoverProjectProfile } from './profile.js';
+/** Keep the prompt section small on large monorepos; the agent can still discover the rest. */
+const MAX_LISTED_PACKAGES = 8;
+function displayRoot(profile, packageProfile) {
+    const relative = path.relative(profile.root, packageProfile.root);
+    return relative === '' ? '.' : relative.split(path.sep).join('/');
+}
+function lineFor(profile, packageProfile) {
+    const commands = discoverPackageValidationCommands(profile, packageProfile);
+    if (commands.length === 0)
+        return null;
+    const cwd = displayRoot(profile, packageProfile);
+    const rendered = commands.map((item) => `\`${item.command}\``).join(', ');
+    return `- ${packageProfile.name} (cwd \`${cwd}\`): ${rendered}`;
+}
+/**
+ * Deterministic project validation map injected into the system prompt: which package owns which
+ * script, and the exact command plus cwd to run it. Commands are listed in increasing cost order
+ * (typecheck → build → test) and are never executed here — this is evidence, not a completion gate.
+ *
+ * Returns '' when no package exposes a validation script, or when discovery fails for any reason
+ * (missing/invalid manifest, unreadable workspace): prompt construction must never break.
+ */
+export function buildValidationCommandsSection(root = process.cwd()) {
+    try {
+        const profile = discoverProjectProfile(root);
+        const lines = [];
+        let omitted = 0;
+        for (const packageProfile of profile.packages) {
+            const line = lineFor(profile, packageProfile);
+            if (!line)
+                continue;
+            if (lines.length >= MAX_LISTED_PACKAGES)
+                omitted += 1;
+            else
+                lines.push(line);
+        }
+        if (lines.length === 0)
+            return '';
+        if (omitted > 0) {
+            lines.push(`- …${omitted} more package(s) with scripts: read their package.json when needed.`);
+        }
+        return [
+            '',
+            '## Validation commands (discovered from project manifests)',
+            'Listed in increasing cost order. Use them when a check is worth running; prefer the package that owns your change over repository-wide runs. Not a completion gate.',
+            ...lines,
+        ].join('\n');
+    }
+    catch {
+        return ''; // Discovery is best-effort: never let it break prompt construction.
+    }
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mocode-ai",
-  "version": "1.2.6",
+  "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": {
@@ -23,7 +23,7 @@
     "build": "tsc -p tsconfig.build.json",
     "test": "tsc -p tsconfig.test-build.json && node --test --experimental-test-isolation=none \"dist-tests/tests/*.test.js\"",
     "typecheck": "tsc --noEmit && tsc -p tests/tsconfig.json && tsc -p evals/tsconfig.json",
-    "eval:smoke": "tsx evals/smoke.ts && tsx evals/coding/smoke.ts",
+    "eval:smoke": "tsx evals/smoke.ts && tsx evals/coding/smoke.ts && tsx evals/work-discipline.ts",
     "eval:coding": "tsx evals/coding/runner.ts",
     "eval:coding:list": "tsx evals/coding/runner.ts --list",
     "prepare": "npm run build"