@danceiny/gotry 0.0.1-rc.9 → 0.2.0-rc.28

package/dist/scripts/price-drift-watch.js

@@ -0,0 +1,447 @@
+/**
+ * LLM 官方定价漂移监测(issue #49 长效机制)。
+ *
+ * 设计纪律(与 ADR-11 / issue #49 一致):
+ *   - 永不自动 apply 价格变更到 llm-price-table.json——价格调整必须走 PR,
+ *     由人 review;founder 2026-08-26 实测「自动 apply 容易把 down 5% 的官方新价错认为我的 bug」;
+ *   - 本脚本只产出 PR-就绪的 Markdown diff,供人粘贴到 PR 描述;
+ *   - 联网取数(--fetch 模式)有超时 + 离线 SKIP,默认行为零网络,适配 CI 纪律;
+ *   - 解析规则保守:任何一行解析失败 → 整家 SKIP + warn,不写未知数据;
+ *   - 比较口径:与 fixture 缓存的上次官方价对照——若 fixture 不存在,fetch 模式写入新 fixture,
+ *     diff 模式直接报"首次运行,无 baseline"。
+ *
+ * 运行(在 ts/ 下):
+ *   npx tsx scripts/price-drift-watch.ts              # 离线模式:对照 fixture 算 diff(零网络)
+ *   npx tsx scripts/price-drift-watch.ts --fetch      # 联网模式:fetch 官方页 + 写 fixture + 算 diff
+ *   npx tsx scripts/price-drift-watch.ts --format json
+ *
+ * 输出:stdout Markdown(默认)/JSON(--format json),心形:
+ *   {
+ *     schema_version: 'gotry_llm_price_drift_v1',
+ *     observed_at: '<ISO>',
+ *     providers: {
+ *       deepseek: { status: 'unchanged' | 'drift' | 'skipped_no_baseline' | 'skipped_parse_error' | 'skipped_no_network', drift: [...] },
+ *       minimax: { ... },
+ *       openai:  { ... },
+ *       anthropic: { ... }
+ *     },
+ *     suggested_pr_paragraph: '<可直接粘贴到 PR 描述的一段>'
+ *   }
+ *
+ * 心智模型:
+ *   - provider="deepseek": fixture 文件 data/llm-price-baseline-deepseek.json,
+ *     上次 fetch 的官方价(rows 数组 + as_of);若不存在 → "skipped_no_baseline",
+ *     提示:首次跑 --fetch 一次入 fixture。
+ *   - provider="minimax": 同上,data/llm-price-baseline-minimax.json。
+ *   - provider="openai": 同上,data/llm-price-baseline-openai.json。
+ *   - provider="anthropic": 同上,data/llm-price-baseline-anthropic.json。
+ *   - 已入价表的 provider(当前为 deepseek/minimax)才有 baseline 期望;
+ *     未入价表的 provider(openai/anthropic)fetch 模式采得就当 narrative 记录,
+ *     仍不进 llm-price-table.json(那是未来单独的 PR)。
+ */
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';
+const FETCH_TIMEOUT_MS = 8000;
+const DEFAULT_PROVIDERS = [
+    { id: 'deepseek', source_url: 'https://api-docs.deepseek.com/quick_start/pricing', baseline_path: 'data/llm-price-baseline-deepseek.json', parse: parseDeepseekHtml },
+    { id: 'minimax', source_url: 'https://platform.minimax.io/docs/guides/pricing-paygo', baseline_path: 'data/llm-price-baseline-minimax.json', parse: parseMinimaxHtml },
+    { id: 'openai', source_url: 'https://openai.com/api/pricing/', baseline_path: 'data/llm-price-baseline-openai.json', snapshot_path: 'data/llm-pricing-snapshot-openai.html', parse: parseOpenaiHtml },
+    { id: 'anthropic', source_url: 'https://www.anthropic.com/pricing', baseline_path: 'data/llm-price-baseline-anthropic.json', snapshot_path: 'data/llm-pricing-snapshot-anthropic.html', parse: parseAnthropicHtml },
+];
+async function fetchWithTimeout(url, timeoutMs) {
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(), timeoutMs);
+    try {
+        const res = await fetch(url, { signal: controller.signal, headers: { 'User-Agent': 'GoTry-price-drift-watch/1 (issue #49)' } });
+        if (!res.ok)
+            throw new Error(`HTTP ${res.status}`);
+        return await res.text();
+    }
+    finally {
+        clearTimeout(timer);
+    }
+}
+/** DeepSeek 官方定价页是简单 HTML 表格;粗解析:美元数字 $/M token 表行。
+ *  失败时不抛:返回 null,watcher 标 skipped_parse_error。 */
+function parseDeepseekHtml(html) {
+    // DeepSeek 表格行形如:
+    //   <td>deepseek-v4-flash</td><td>0.44</td><td>0.014</td><td>1.32</td>  (cache_miss/cache_hit/output per 1M)
+    const rows = [];
+    const rowRe = /<tr[^>]*>([\s\S]*?)<\/tr>/g;
+    const cellRe = /<td[^>]*>([\s\S]*?)<\/td>/g;
+    let tr;
+    while ((tr = rowRe.exec(html)) !== null) {
+        const cells = [];
+        let m;
+        while ((m = cellRe.exec(tr[1])) !== null)
+            cells.push(m[1].replace(/<[^>]+>/g, '').trim());
+        if (cells.length < 4)
+            continue;
+        const model = cells[0];
+        if (!/^deepseek-v[34]/.test(model))
+            continue;
+        const miss = Number(cells[1]);
+        const hit = Number(cells[2]);
+        const out = Number(cells[3]);
+        if (!Number.isFinite(miss) || !Number.isFinite(out))
+            continue;
+        rows.push({ model, input_cache_miss_usd_per_1m: miss, input_cache_hit_usd_per_1m: Number.isFinite(hit) ? hit : null, output_usd_per_1m: out });
+    }
+    return rows.length > 0 ? rows : null;
+}
+/** MiniMax Pay-as-you-go 页面是 markdown 化 HTML 文档,粗解析:model + $/M tokens 行。 */
+function parseMinimaxHtml(html) {
+    const rows = [];
+    // 行形如: **MiniMax-M3** … $0.30 / M tokens (Input, ≤512k) … $1.20 / M tokens (Output, ≤512k) … $0.06 / M tokens (Cache Read)
+    const lineRe = /<p[^>]*>([\s\S]*?)<\/p>/g;
+    let p;
+    const seen = new Set();
+    while ((p = lineRe.exec(html)) !== null) {
+        const text = p[1].replace(/<[^>]+>/g, '').replace(/&nbsp;/g, ' ').replace(/\s+/g, ' ').trim();
+        const modelMatch = text.match(/(MiniMax-M2(?:\.1)?(?:-highspeed)?|MiniMax-M3)/);
+        if (!modelMatch)
+            continue;
+        const model = modelMatch[1];
+        // 取输入/输出/缓存读取三个数;按出现顺序:Input, Output, Cache Read
+        const numbers = Array.from(text.matchAll(/\$([0-9]+(?:\.[0-9]+)?)\s*\/\s*M\s*tokens?/g)).map(m => Number(m[1]));
+        if (numbers.length < 2)
+            continue;
+        // 输入/输出 两个必备;Cache Read 可选
+        const [miss, out] = numbers;
+        const hit = numbers.length >= 3 ? numbers[2] : null;
+        if (seen.has(model))
+            continue;
+        seen.add(model);
+        rows.push({ model, input_cache_miss_usd_per_1m: miss, input_cache_hit_usd_per_1m: hit, output_usd_per_1m: out });
+    }
+    return rows.length > 0 ? rows : null;
+}
+/** OpenAI 公开页是 React SPA(服务端只返壳 JSON 与内嵌 <script>)。
+ *  解析策略:`data/llm-pricing-snapshot-openai.html` snapshot 优先(入 git,CI 零联网);
+ *  `--fetch` 模式覆盖 snapshot;解析取自 Next.js 嵌入的 `__NEXT_DATA__` JSON dump
+ *  (其 `props.pageProps` 含 model price 数组),失败回退到段落文字粗解析。 */
+export function parseOpenaiHtml(html) {
+    const fromNextData = parseOpenaiNextData(html);
+    if (fromNextData)
+        return fromNextData;
+    return parseOpenaiParagraphFallback(html);
+}
+function parseOpenaiNextData(html) {
+    const m = html.match(/<script[^>]*id="__NEXT_DATA__"[^>]*>([\s\S]*?)<\/script>/);
+    if (!m)
+        return null;
+    try {
+        const data = JSON.parse(m[1]);
+        const models = data?.props?.pageProps?.models;
+        if (!Array.isArray(models) || models.length === 0)
+            return null;
+        const rows = [];
+        for (const m of models) {
+            const name = String(m['name'] ?? m['model'] ?? '').trim();
+            if (!name)
+                continue;
+            // 启发式:per-token 数字通常极小(<1),per-1M 数字通常 ≥ 0.1;前者要 ×1_000_000
+            const rawMiss = Number(m['input_cost_per_token'] ?? m['input_per_million'] ?? m['input']);
+            const rawOut = Number(m['output_cost_per_token'] ?? m['output_per_million'] ?? m['output']);
+            const rawHit = Number(m['cached_input_cost_per_token'] ?? m['cached_input_per_million'] ?? m['cache_read']);
+            if (!Number.isFinite(rawMiss) || !Number.isFinite(rawOut))
+                continue;
+            rows.push({
+                model: name,
+                input_cache_miss_usd_per_1m: rawMiss < 1 ? rawMiss * 1_000_000 : rawMiss,
+                input_cache_hit_usd_per_1m: Number.isFinite(rawHit) ? (rawHit < 1 ? rawHit * 1_000_000 : rawHit) : null,
+                output_usd_per_1m: rawOut < 1 ? rawOut * 1_000_000 : rawOut,
+            });
+        }
+        return rows.length > 0 ? rows : null;
+    }
+    catch {
+        return null;
+    }
+}
+function parseOpenaiParagraphFallback(html) {
+    // 兜底:段落文字形如 "gpt-4o … $2.50 / 1M input … $10.00 / 1M output"
+    const rows = [];
+    const lineRe = /<p[^>]*>([\s\S]*?)<\/p>|<li[^>]*>([\s\S]*?)<\/li>/g;
+    let m;
+    while ((m = lineRe.exec(html)) !== null) {
+        const text = (m[1] ?? m[2] ?? '').replace(/<[^>]+>/g, '').replace(/&nbsp;/g, ' ').replace(/\s+/g, ' ').trim();
+        const modelMatch = text.match(/(gpt-4o(?:-mini)?|gpt-4\.1(?:-mini)?|gpt-5(?:-mini)?|o[1-9](?:-mini|-pro)?)/);
+        if (!modelMatch)
+            continue;
+        const numbers = Array.from(text.matchAll(/\$([0-9]+(?:\.[0-9]+)?)\s*\/\s*(?:1M|M)/g)).map(x => Number(x[1]));
+        if (numbers.length < 2)
+            continue;
+        const [miss, out] = numbers;
+        rows.push({ model: modelMatch[1], input_cache_miss_usd_per_1m: miss, input_cache_hit_usd_per_1m: null, output_usd_per_1m: out });
+    }
+    return rows.length > 0 ? rows : null;
+}
+/** Anthropic pricing 页同样 SPA,内嵌 `__NEXT_DATA__` 形态可能变;双路径同 OpenAI。 */
+function parseAnthropicHtml(html) {
+    const fromNextData = parseAnthropicNextData(html);
+    if (fromNextData)
+        return fromNextData;
+    return parseAnthropicParagraphFallback(html);
+}
+function parseAnthropicNextData(html) {
+    const m = html.match(/<script[^>]*id="__NEXT_DATA__"[^>]*>([\s\S]*?)<\/script>/);
+    if (!m)
+        return null;
+    try {
+        const data = JSON.parse(m[1]);
+        const pricing = data?.props?.pageProps?.pricing;
+        if (!Array.isArray(pricing) || pricing.length === 0)
+            return null;
+        const rows = [];
+        for (const p of pricing) {
+            const name = String(p['model'] ?? p['name'] ?? '').trim();
+            if (!name)
+                continue;
+            const miss = Number(p['input_cost_per_mtok'] ?? p['input']);
+            const out = Number(p['output_cost_per_mtok'] ?? p['output']);
+            const hit = Number(p['cache_read_cost_per_mtok'] ?? p['cache_read']);
+            if (!Number.isFinite(miss) || !Number.isFinite(out))
+                continue;
+            rows.push({ model: name, input_cache_miss_usd_per_1m: miss, input_cache_hit_usd_per_1m: Number.isFinite(hit) ? hit : null, output_usd_per_1m: out });
+        }
+        return rows.length > 0 ? rows : null;
+    }
+    catch {
+        return null;
+    }
+}
+function parseAnthropicParagraphFallback(html) {
+    const rows = [];
+    const lineRe = /<p[^>]*>([\s\S]*?)<\/p>|<li[^>]*>([\s\S]*?)<\/li>/g;
+    let m;
+    while ((m = lineRe.exec(html)) !== null) {
+        const text = (m[1] ?? m[2] ?? '').replace(/<[^>]+>/g, '').replace(/&nbsp;/g, ' ').replace(/\s+/g, ' ').trim();
+        const modelMatch = text.match(/(claude-(?:opus|sonnet|haiku)-[0-9](?:\.[0-9])?(?:-[0-9]+k)?)/);
+        if (!modelMatch)
+            continue;
+        const numbers = Array.from(text.matchAll(/\$([0-9]+(?:\.[0-9]+)?)\s*\/\s*(?:1M|M|MT)/g)).map(x => Number(x[1]));
+        if (numbers.length < 2)
+            continue;
+        const [miss, out] = numbers;
+        rows.push({ model: modelMatch[1], input_cache_miss_usd_per_1m: miss, input_cache_hit_usd_per_1m: null, output_usd_per_1m: out });
+    }
+    return rows.length > 0 ? rows : null;
+}
+function diffRows(baseline, current) {
+    const baseMap = new Map(baseline.map(r => [r.model, r]));
+    const currMap = new Map(current.map(r => [r.model, r]));
+    const drift = [];
+    for (const [model, curr] of currMap) {
+        const base = baseMap.get(model);
+        if (!base) {
+            drift.push({ model, field: 'input_cache_miss_usd_per_1m', from: null, to: curr.input_cache_miss_usd_per_1m, direction: 'new' });
+            continue;
+        }
+        if (base.input_cache_miss_usd_per_1m !== curr.input_cache_miss_usd_per_1m) {
+            drift.push({ model, field: 'input_cache_miss_usd_per_1m', from: base.input_cache_miss_usd_per_1m, to: curr.input_cache_miss_usd_per_1m, direction: curr.input_cache_miss_usd_per_1m > base.input_cache_miss_usd_per_1m ? 'up' : 'down' });
+        }
+        if (base.input_cache_hit_usd_per_1m !== curr.input_cache_hit_usd_per_1m) {
+            drift.push({ model, field: 'input_cache_hit_usd_per_1m', from: base.input_cache_hit_usd_per_1m, to: curr.input_cache_hit_usd_per_1m, direction: (curr.input_cache_hit_usd_per_1m ?? 0) > (base.input_cache_hit_usd_per_1m ?? 0) ? 'up' : 'down' });
+        }
+        if (base.output_usd_per_1m !== curr.output_usd_per_1m) {
+            drift.push({ model, field: 'output_usd_per_1m', from: base.output_usd_per_1m, to: curr.output_usd_per_1m, direction: curr.output_usd_per_1m > base.output_usd_per_1m ? 'up' : 'down' });
+        }
+    }
+    for (const [model, base] of baseMap) {
+        if (!currMap.has(model)) {
+            drift.push({ model, field: 'input_cache_miss_usd_per_1m', from: base.input_cache_miss_usd_per_1m, to: null, direction: 'removed' });
+        }
+    }
+    return drift;
+}
+function loadBaseline(path) {
+    if (!existsSync(path))
+        return null;
+    try {
+        const raw = JSON.parse(readFileSync(path, 'utf8'));
+        if (raw['schema_version'] !== 'gotry_llm_price_baseline_v1')
+            throw new Error(`baseline ${path} schema_version must be gotry_llm_price_baseline_v1, got ${String(raw['schema_version'])}`);
+        const rows = raw['rows'];
+        if (!Array.isArray(rows))
+            throw new Error(`baseline ${path} rows must be array`);
+        return raw;
+    }
+    catch (error) {
+        return null;
+    }
+}
+/** 读 snapshot HTML 文件;不存在或空 → null。snapshot 由 `--fetch` 模式首次写盘(founder 手动 PR)。 */
+function loadSnapshot(path) {
+    if (!existsSync(path))
+        return null;
+    try {
+        const text = readFileSync(path, 'utf8');
+        return text.length > 0 ? text : null;
+    }
+    catch {
+        return null;
+    }
+}
+function writeSnapshot(path, html) {
+    writeFileSync(path, html, 'utf8');
+}
+function writeBaseline(path, rows, sourceUrl) {
+    const payload = {
+        schema_version: 'gotry_llm_price_baseline_v1',
+        as_of: new Date().toISOString(),
+        source_url: sourceUrl,
+        rows,
+    };
+    writeFileSync(path, JSON.stringify(payload, null, 2) + '\n', 'utf8');
+}
+export async function runPriceDriftWatch(options) {
+    const observedAt = options.observedAt().toISOString();
+    const providers = {};
+    const providerSpecs = options.providers ?? DEFAULT_PROVIDERS;
+    const fetcher = options.fetcher ?? ((url) => fetchWithTimeout(url, FETCH_TIMEOUT_MS));
+    for (const spec of providerSpecs) {
+        const baseline = loadBaseline(spec.baseline_path);
+        const snapshot = spec.snapshot_path ? loadSnapshot(spec.snapshot_path) : null;
+        if (options.mode === 'offline') {
+            // 离线模式两种来源:
+            //   (a) 有 snapshot → 走 snapshot.html parse 出 current,与 baseline diff(零联网,CI 心形)
+            //   (b) 无 snapshot → 只能报 baseline 状态(无东西可 diff,标 unchanged + 提示)
+            if (!baseline) {
+                providers[spec.id] = { status: 'skipped_no_baseline', source_url: spec.source_url, reason: `no baseline fixture at ${spec.baseline_path}; run --fetch once to seed` };
+                continue;
+            }
+            if (snapshot === null) {
+                if (spec.snapshot_path) {
+                    providers[spec.id] = { status: 'skipped_no_snapshot', source_url: spec.source_url, reason: `no snapshot at ${spec.snapshot_path}; run --fetch once to seed (or: founder manually save official pricing page HTML as the snapshot)` };
+                }
+                else {
+                    providers[spec.id] = { status: 'unchanged', source_url: spec.source_url, reason: 'offline mode; baseline present, run --fetch to compare against current official prices' };
+                }
+                continue;
+            }
+            const currentFromSnapshot = spec.parse(snapshot);
+            if (!currentFromSnapshot) {
+                providers[spec.id] = { status: 'skipped_parse_error', source_url: spec.source_url, reason: 'snapshot present but parser failed to extract rows; refresh snapshot or fix parser' };
+                continue;
+            }
+            const driftFromSnapshot = diffRows(baseline.rows, currentFromSnapshot) ?? [];
+            if (driftFromSnapshot.length === 0) {
+                providers[spec.id] = { status: 'unchanged', source_url: spec.source_url, reason: 'offline; baseline == snapshot-derived current' };
+            }
+            else {
+                providers[spec.id] = { status: 'drift', source_url: spec.source_url, fetched_at: observedAt, drift: driftFromSnapshot, reason: 'offline diff baseline vs snapshot' };
+            }
+            continue;
+        }
+        // fetch mode
+        let html;
+        try {
+            html = await fetcher(spec.source_url);
+        }
+        catch (error) {
+            providers[spec.id] = { status: 'skipped_no_network', source_url: spec.source_url, reason: `fetch failed: ${error.message}` };
+            continue;
+        }
+        // snapshot 写盘:OpenAI/Anthropic SPA 的真实状态在 snapshot 文件里(founder 手动 PR)
+        if (spec.snapshot_path)
+            writeSnapshot(spec.snapshot_path, html);
+        const current = spec.parse(html);
+        if (!current) {
+            providers[spec.id] = { status: 'skipped_parse_error', source_url: spec.source_url, reason: 'parser did not extract any rows; snapshot was written anyway (manual review of fixture + parser required)' };
+            continue;
+        }
+        if (!baseline) {
+            // 首次 fetch 写 fixture
+            writeBaseline(spec.baseline_path, current, spec.source_url);
+            providers[spec.id] = { status: 'fetched_no_table_entry', source_url: spec.source_url, fetched_at: observedAt, reason: 'first fetch; baseline fixture written; no current llm-price-table.json entry (add a PR if you want this provider in nightly-evidence cost math)' };
+            continue;
+        }
+        const drift = diffRows(baseline.rows, current) ?? [];
+        if (drift.length === 0) {
+            providers[spec.id] = { status: 'unchanged', source_url: spec.source_url, fetched_at: observedAt };
+        }
+        else {
+            providers[spec.id] = { status: 'drift', source_url: spec.source_url, fetched_at: observedAt, drift };
+        }
+    }
+    const suggestedPrParagraph = renderSuggestedPr(providers, observedAt);
+    return {
+        schema_version: 'gotry_llm_price_drift_v1',
+        observed_at: observedAt,
+        mode: options.mode,
+        providers,
+        suggested_pr_paragraph: suggestedPrParagraph,
+    };
+}
+function renderSuggestedPr(providers, observedAt) {
+    const lines = [`## LLM 价格漂移监测(${observedAt})`];
+    for (const [id, report] of Object.entries(providers)) {
+        if (report.status === 'drift' && report.drift && report.drift.length > 0) {
+            const driftItems = report.drift;
+            lines.push(`- **${id}**(${report.source_url}):${driftItems.length} 项变动`);
+            for (const d of driftItems) {
+                lines.push(`  - \`${d.model}\` ${d.field}: ${d.from ?? '∅'} → ${d.to ?? '∅'} (${d.direction})`);
+            }
+            lines.push('  - 建议:按 ADR-11 复核,人工提 PR 改 `data/llm-price-table.json`。');
+        }
+        else if (report.status === 'unchanged') {
+            lines.push(`- **${id}**:无变动`);
+        }
+        else if (report.status === 'fetched_no_table_entry') {
+            lines.push(`- **${id}**:首次 fetch,baseline 已写入(${report.reason ?? ''})`);
+        }
+        else {
+            lines.push(`- **${id}** [${report.status}]:${report.reason ?? ''}`);
+        }
+    }
+    lines.push('');
+    lines.push('纪律:本监测只产 diff,不自动改价表;价格调整必须人工 PR,符合 ADR-11 「peak only-high-not-low」。');
+    return lines.join('\n');
+}
+function formatMarkdown(report) {
+    const lines = [`# LLM 价格漂移监测(${report.observed_at},mode=${report.mode})`, ''];
+    for (const [id, r] of Object.entries(report.providers)) {
+        lines.push(`## ${id} [${r.status}]`);
+        lines.push(`source: ${r.source_url}`);
+        if (r.reason)
+            lines.push(`reason: ${r.reason}`);
+        if (r.drift && r.drift.length > 0) {
+            lines.push('| model | field | from | to | direction |');
+            lines.push('|---|---|---|---|---|');
+            for (const d of r.drift)
+                lines.push(`| \`${d.model}\` | ${d.field} | ${d.from ?? '∅'} | ${d.to ?? '∅'} | ${d.direction} |`);
+        }
+        lines.push('');
+    }
+    lines.push('---');
+    lines.push('');
+    lines.push(report.suggested_pr_paragraph);
+    return lines.join('\n');
+}
+function arg(name) {
+    const index = process.argv.indexOf(name);
+    return index >= 0 ? process.argv[index + 1] : undefined;
+}
+async function main() {
+    const asJson = (arg('--format') ?? 'markdown') === 'json';
+    const mode = process.argv.includes('--fetch') ? 'fetch' : 'offline';
+    mkdirSync('gotry-state', { recursive: true });
+    try {
+        const report = await runPriceDriftWatch({ mode, observedAt: () => new Date() });
+        if (asJson)
+            console.log(JSON.stringify(report, null, 2));
+        else
+            console.log(formatMarkdown(report));
+    }
+    catch (error) {
+        const payload = { state: 'error', error: error.message };
+        if (asJson)
+            console.log(JSON.stringify(payload, null, 2));
+        else
+            console.error(`price-drift-watch error: ${error.message}`);
+        process.exitCode = 1;
+    }
+}
+if (process.argv[1]?.endsWith('price-drift-watch.ts')) {
+    await main();
+}

package/dist/scripts/probe-poi-tests.js

@@ -1,88 +1,96 @@
+/**
+ * probePoi 单测(D-7a 债的产品层测试,不动求解器)
+ *
+ * ⚠️  本单测是**稳定性回归**,不验证精确字符串。
+ * 原因: tsx + Node 24 experimental-strip-types 在中文 unicode character class
+ * 上有微小行为差(regex char class 贪婪+unicode 编码在 tsx 与 node 直接调
+ * 时输出不一致)。probePoi 函数本身在 e2e 路径(datasources.runTurn)实
+ * 测可工作(commit a5ca01e 已验证);精确字符串覆盖在那里。
+ *
+ * 单测覆盖:不抛错 + 5 类触发各返回非 null/合理 + 2 类不触发返 null。
+ * 运行: cd ts && npx tsx scripts/probe-poi-tests.ts
+ */
 import assert from 'node:assert/strict';
 import { probePoi } from '../src/loop.js';
 function pass(name, body) {
     body();
     console.log(`  ${name} OK`);
 }
-pass('1. 不抛错 + 空串/纯空白返 null', ()=>{
-    assert.doesNotThrow(()=>probePoi(''));
+pass('1. 不抛错 + 空串/纯空白返 null', () => {
+    assert.doesNotThrow(() => probePoi(''));
     assert.equal(probePoi(''), null);
     assert.equal(probePoi('   '), null);
 });
-pass('2. 短纯地名(<24 字符非问句)触发', ()=>{
-    for (const t of [
-        '普吉',
-        '清迈',
-        '大理',
-        '北京 酒店',
-        '丽江市'
-    ]){
+pass('2. 短纯地名(<24 字符非问句)触发', () => {
+    for (const t of ['普吉', '清迈', '大理', '北京 酒店', '丽江市']) {
         const r = probePoi(t);
-        if (r === null) throw new Error(`${JSON.stringify(t)} 应触发,得 null`);
-        if (typeof r !== 'string') throw new Error(`返类型错: ${typeof r}`);
-        if (r.length < 2 || r.length > 24) throw new Error(`长度越界: ${r.length}`);
+        if (r === null)
+            throw new Error(`${JSON.stringify(t)} 应触发,得 null`);
+        if (typeof r !== 'string')
+            throw new Error(`返类型错: ${typeof r}`);
+        if (r.length < 2 || r.length > 24)
+            throw new Error(`长度越界: ${r.length}`);
     }
 });
-pass('3. 结构触发 — 含"查/搜/找/看看/推荐/告诉我/检索"', ()=>{
-    for (const t of [
-        '查一下 大理',
-        '搜 普吉',
-        '找清迈',
-        '看看曼谷',
-        '推荐 沙巴'
-    ]){
+pass('3. 结构触发 — 含"查/搜/找/看看/推荐/告诉我/检索"', () => {
+    for (const t of ['查一下 大理', '搜 普吉', '找清迈', '看看曼谷', '推荐 沙巴']) {
         const r = probePoi(t);
-        if (r === null) throw new Error(`${JSON.stringify(t)} 应触发,得 null`);
+        if (r === null)
+            throw new Error(`${JSON.stringify(t)} 应触发,得 null`);
     }
 });
-pass('4. 内容触发 — 含"酒店/民宿/客栈/有什么/玩什么/有哪些"', ()=>{
-    for (const t of [
-        '大理酒店',
-        '清迈民宿',
-        '曼谷有什么',
-        '丽江有哪些'
-    ]){
+pass('4. 内容触发 — 含"酒店/民宿/客栈/有什么/玩什么/有哪些"', () => {
+    for (const t of ['大理酒店', '清迈民宿', '曼谷有什么', '丽江有哪些']) {
         const r = probePoi(t);
-        if (r === null) throw new Error(`${JSON.stringify(t)} 应触发,得 null`);
+        if (r === null)
+            throw new Error(`${JSON.stringify(t)} 应触发,得 null`);
     }
 });
-pass('5. 不触发 — 长问句(>24 字符 触发上限 isShort 拒)', ()=>{
+pass('5. 不触发 — 长问句(>24 字符 触发上限 isShort 拒)', () => {
+    // 短查询直通(<=24)是设计意图,产品行为快照:任何短查询都返回原句做关键词。
+    // 因此"3.4 k"等英文短查询也返,不期望 null。
+    // 长问句(>24 字符)才是不触发场景。
     for (const t of [
-        'lorem ipsum dolor sit amet consect',
-        '我想查询从北京到上海的高铁票价格'
-    ]){
+        'lorem ipsum dolor sit amet consect', // 35 字符
+        '我想查询从北京到上海的高铁票价格', // 16 字符 ≤ 24 短查询直通(返原句),产品行为
+    ]) {
         const r = probePoi(t);
         if (t.length > 24) {
-            if (r !== null) throw new Error(`>24 字符 ${JSON.stringify(t)} 应不触发,得 ${JSON.stringify(r)}`);
-        } else {
-            if (r === null) throw new Error(`<=24 字符 ${JSON.stringify(t)} 应触发,得 null`);
+            if (r !== null)
+                throw new Error(`>24 字符 ${JSON.stringify(t)} 应不触发,得 ${JSON.stringify(r)}`);
+        }
+        else {
+            if (r === null)
+                throw new Error(`<=24 字符 ${JSON.stringify(t)} 应触发,得 null`);
         }
     }
 });
-pass('6. 不抛错 + 返回类型是 string 或 null(边界)', ()=>{
+pass('6. 不抛错 + 返回类型是 string 或 null(边界)', () => {
     for (const t of [
         'a'.repeat(200),
         'a?'.repeat(50),
         '!@#',
         '中文' + 'a'.repeat(100),
-        '  前后空格  '
-    ]){
+        '  前后空格  ',
+    ]) {
         const r = probePoi(t);
         if (r !== null && typeof r !== 'string') {
             throw new Error(`返类型错: ${typeof r}`);
         }
     }
 });
-pass('7. 金标准对话噪音回归(2026-08-28 巡检:访谈答案/多段首查不再误触发)', ()=>{
+pass('7. 金标准对话噪音回归(2026-08-28 巡检:访谈答案/多段首查不再误触发)', () => {
+    // 旧版 ≤24 直通把短访谈答案整句当关键词 → 垃圾 hbcli 调用 + 证据噪音
     assert.equal(probePoi('我的工作时间是UTC+4的早上10点到下午7点'), null, '访谈答案(>12 字含陈述动词)不触发');
     assert.equal(probePoi('明天有空'), null, '短句含陈述动词不触发');
+    // 多段行程首轮:动词宾语为句号/宾语为垃圾后缀 → 不触发(宁可不探,不垃圾探)
     assert.equal(probePoi('7.17周五22:40落地深圳,7.18早上去香港办银行开户&保险签约;争取7.18当天飞泰国普吉岛……8.10周一凌晨从深圳起飞,周一上班前到迪拜。请给我做机票和酒店的行程规划和推荐。'), null, '多段首查无 POI 信号不触发');
+    // 住宿名词后段优先:订了酒店:名称 → 抓名称而非「我订了」
     const booking = probePoi('我订了酒店:The Title East Wing Rawai,7.18入住 7.23 退房');
-    if (!booking || !booking.includes('Title')) throw new Error(`订酒店应抓名称段,得 ${JSON.stringify(booking)}`);
+    if (!booking || !booking.includes('Title'))
+        throw new Error(`订酒店应抓名称段,得 ${JSON.stringify(booking)}`);
+    assert.equal(booking, 'The Title East Wing Rawai', '真实酒店名须原样回显(旧 24 字上界会截成 …Rawa 且无省略标记)');
     console.log('  7a. 访谈答案/多段首查 → null OK');
     console.log('  7b. 订酒店抓名称段(The Title…) OK');
 });
 console.log('\nprobePoi TESTS: 6 类 OK(不崩 + 5 类触发 + 1 类不触发返 null)');
-
-
-//# sourceURL=/Users/bytedance/work/gotry/ts/scripts/probe-poi-tests.ts