@proofhound/optimization-strategy 0.1.6

package/dist/error-pattern-analysis/prompts.js

@@ -0,0 +1,659 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.getSystemPrompts = exports.getOptimizationTipNames = exports.SYSTEM_PROMPTS_EN = exports.SYSTEM_PROMPTS = exports.SUMMARIZE_SYSTEM_PROMPT = exports.PROMPT_FILES = exports.OPTIMIZATION_TIP_NAMES_EN = exports.OPTIMIZATION_TIP_NAMES = exports.OPTIMIZATION_TIPS = exports.GENERATE_SYSTEM_PROMPT = exports.ANALYZE_REGRESSION_SYSTEM_PROMPT = exports.ANALYZE_CONFUSION_SYSTEM_PROMPT = void 0;
+exports.formatRoundHistory = formatRoundHistory;
+exports.formatToolboxSwitchHint = formatToolboxSwitchHint;
+exports.fitRoundHistoryToBudget = fitRoundHistoryToBudget;
+exports.buildAnalyzeConfusionMessages = buildAnalyzeConfusionMessages;
+exports.buildAnalyzeRegressionMessages = buildAnalyzeRegressionMessages;
+exports.buildSummarizeMessages = buildSummarizeMessages;
+exports.buildGenerateMessages = buildGenerateMessages;
+exports.extractVariableNames = extractVariableNames;
+// LLM prompt construction — system prompts all come from prompts/*.md (loaded by prompts/loader.ts)
+// This file is only responsible for assembling the user prompt (including dynamic variable substitution).
+const goals_1 = require("../loop/goals");
+const shared_1 = require("@proofhound/shared");
+const loader_1 = require("./prompts/loader");
+const token_budget_1 = require("./token-budget");
+// Re-export system prompts + the SYSTEM_PROMPTS map (keeps backward-compatible API)
+var loader_2 = require("./prompts/loader");
+Object.defineProperty(exports, "ANALYZE_CONFUSION_SYSTEM_PROMPT", { enumerable: true, get: function () { return loader_2.ANALYZE_CONFUSION_SYSTEM_PROMPT; } });
+Object.defineProperty(exports, "ANALYZE_REGRESSION_SYSTEM_PROMPT", { enumerable: true, get: function () { return loader_2.ANALYZE_REGRESSION_SYSTEM_PROMPT; } });
+Object.defineProperty(exports, "GENERATE_SYSTEM_PROMPT", { enumerable: true, get: function () { return loader_2.GENERATE_SYSTEM_PROMPT; } });
+Object.defineProperty(exports, "OPTIMIZATION_TIPS", { enumerable: true, get: function () { return loader_2.OPTIMIZATION_TIPS; } });
+Object.defineProperty(exports, "OPTIMIZATION_TIP_NAMES", { enumerable: true, get: function () { return loader_2.OPTIMIZATION_TIP_NAMES; } });
+Object.defineProperty(exports, "OPTIMIZATION_TIP_NAMES_EN", { enumerable: true, get: function () { return loader_2.OPTIMIZATION_TIP_NAMES_EN; } });
+Object.defineProperty(exports, "PROMPT_FILES", { enumerable: true, get: function () { return loader_2.PROMPT_FILES; } });
+Object.defineProperty(exports, "SUMMARIZE_SYSTEM_PROMPT", { enumerable: true, get: function () { return loader_2.SUMMARIZE_SYSTEM_PROMPT; } });
+Object.defineProperty(exports, "SYSTEM_PROMPTS", { enumerable: true, get: function () { return loader_2.SYSTEM_PROMPTS; } });
+Object.defineProperty(exports, "SYSTEM_PROMPTS_EN", { enumerable: true, get: function () { return loader_2.SYSTEM_PROMPTS_EN; } });
+Object.defineProperty(exports, "getOptimizationTipNames", { enumerable: true, get: function () { return loader_2.getOptimizationTipNames; } });
+Object.defineProperty(exports, "getSystemPrompts", { enumerable: true, get: function () { return loader_2.getSystemPrompts; } });
+// =========================
+// Common fragments
+// =========================
+function scopeLabel(scope, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (language === 'en-US')
+        return scope.kind === 'overall' ? 'overall' : `class "${scope.label}"`;
+    return scope.kind === 'overall' ? '整体' : `分类「${scope.label}」`;
+}
+function fmtNum(n, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (n === null || n === undefined || Number.isNaN(n))
+        return language === 'en-US' ? '(missing)' : '（缺失）';
+    return n.toFixed(4);
+}
+function signedGap(observed, target, op) {
+    // Normalize the gap by op direction: positive = exceeded the goal; negative = how much more is needed
+    // >= / > : gap = observed - target (larger is better)
+    // <= : gap = target - observed (smaller is better; displayed as "how much further it needs to drop")
+    const delta = op === '<=' ? target - observed : observed - target;
+    const sign = delta >= 0 ? '+' : '';
+    return `${sign}${delta.toFixed(4)}`;
+}
+// "Optimization goal vs current actual" comparison table — display goals and actual values side by side
+function formatGoalsWithProgress(goals, metrics, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (goals.length === 0)
+        return language === 'en-US' ? '(no optimization goals declared)' : '（未声明优化目标）';
+    return goals
+        .map((g) => {
+        const observed = (0, goals_1.readMetric)(metrics, g);
+        const achieved = observed !== null && (0, goals_1.compare)(observed, g.op, g.value);
+        const gapStr = observed === null ? '?' : signedGap(observed, g.value, g.op);
+        const status = achieved
+            ? language === 'en-US'
+                ? 'achieved'
+                : '✅ 已达成'
+            : language === 'en-US'
+                ? 'not achieved'
+                : '❌ 未达成';
+        if (language === 'en-US') {
+            return `- \`${g.metric}\` for ${scopeLabel(g.scope, language)}: target \`${g.op} ${g.value}\`; observed \`${fmtNum(observed, language)}\`; gap \`${gapStr}\`; ${status}`;
+        }
+        return `- ${scopeLabel(g.scope, language)} 的 \`${g.metric}\`：目标 \`${g.op} ${g.value}\`；当前实际 \`${fmtNum(observed, language)}\`；差距 \`${gapStr}\`；${status}`;
+    })
+        .join('\n');
+}
+// "Full metrics for the scope of interest" — only display all metrics within the scopes the goals cover (overall / specific classes),
+// classes not covered are not displayed (saves tokens + lets the LLM focus)
+function formatRelevantMetrics(goals, metrics, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (goals.length === 0) {
+        // When there are no goals, degrade to displaying overall (avoiding a completely empty user prompt)
+        return [
+            language === 'en-US' ? '### Overall' : '### 整体',
+            '```json',
+            JSON.stringify(metrics.overall, null, 2),
+            '```',
+        ].join('\n');
+    }
+    const involvedClasses = new Set();
+    let needsOverall = false;
+    for (const g of goals) {
+        if (g.scope.kind === 'overall')
+            needsOverall = true;
+        else
+            involvedClasses.add(g.scope.label);
+    }
+    const sections = [];
+    if (needsOverall) {
+        sections.push(language === 'en-US' ? '### Overall' : '### 整体', '```json', JSON.stringify(metrics.overall ?? {}, null, 2), '```');
+    }
+    for (const cls of involvedClasses) {
+        const slice = metrics.perClass?.[cls];
+        sections.push(language === 'en-US' ? `### Class "${cls}"` : `### 分类「${cls}」`, '```json', JSON.stringify(slice ?? {}, null, 2), '```');
+    }
+    return sections.join('\n');
+}
+function formatVariableList(vars, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (vars.length === 0) {
+        return language === 'en-US'
+            ? '(the current prompt references no dataset field variables)'
+            : '（当前 prompt 不引用任何数据集字段变量）';
+    }
+    return vars.map((v) => `- \`{{${v}}}\``).join('\n');
+}
+function formatAnalysisOnlyFields(fields, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (!fields || fields.length === 0)
+        return language === 'en-US' ? '(none)' : '（无）';
+    return fields.map((f) => `- \`${f}\``).join('\n');
+}
+// =========================
+// Cross-round history section rendering — see docs/specs/25-optimizations.md §11.3 "cross-round history injection"
+// Primary metric taken from goals[0] (consistent with the deltaFromPrev calculation); when there are no goals, degrade to metrics.overall.accuracy
+// =========================
+function fmtDelta(d) {
+    if (d === null)
+        return 'Δ -- ';
+    const sign = d >= 0 ? '+' : '';
+    return `Δ ${sign}${d.toFixed(4)}`;
+}
+function readPrimaryMetric(metrics, goals) {
+    const primary = goals[0];
+    if (primary) {
+        const v = (0, goals_1.readMetric)(metrics, primary);
+        return v;
+    }
+    const v = metrics.overall?.accuracy;
+    return typeof v === 'number' ? v : null;
+}
+function primaryMetricName(goals) {
+    return goals[0]?.metric ?? 'accuracy';
+}
+function formatRoundHistory(history, goals, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (history.length === 0)
+        return '';
+    const metricName = primaryMetricName(goals);
+    const head = language === 'en-US'
+        ? [
+            `## Optimization History (${history.length} rounds, chronological; use it to avoid repeated failed directions)`,
+            `> Reading guide: Δ is the change in the primary metric (${metricName}) from the previous round; ★ marks the current best round.`,
+            '',
+        ]
+        : [
+            `## 历史优化轨迹（共 ${history.length} 轮，按时间正序；用于避免重复无效尝试）`,
+            `> 解读：Δ 是与上一轮主指标 (${metricName}) 的差值；★ 标记当前 best 轮；★+ 表示该 best 同时被采用为下一轮 base`,
+            '',
+        ];
+    const lines = history.map((entry) => {
+        const primary = readPrimaryMetric(entry.metrics, goals);
+        const primaryStr = primary === null ? (language === 'en-US' ? '(missing)' : '（缺失）') : primary.toFixed(4);
+        const bestMark = entry.isBest ? '★' : '';
+        const changeIds = entry.appliedChanges.map((c) => c.changeId).filter(Boolean);
+        const changeIdsStr = changeIds.length > 0 ? `[${changeIds.join(', ')}]` : language === 'en-US' ? '(none)' : '（无）';
+        const summary = entry.changeSummary.trim().length > 0
+            ? entry.changeSummary.trim()
+            : language === 'en-US'
+                ? '(not provided)'
+                : '（未提供）';
+        const tips = entry.appliedTips.filter((t) => t.trim().length > 0);
+        const tipsStr = tips.length > 0 ? `[${tips.join(', ')}]` : language === 'en-US' ? '(not declared)' : '（未声明）';
+        const prefix = language === 'en-US'
+            ? `- Round ${entry.roundIndex} (${fmtDelta(entry.deltaFromPrev)}) ${bestMark} ${metricName} ${primaryStr}`
+            : `- 第 ${entry.roundIndex} 轮 (${fmtDelta(entry.deltaFromPrev)}) ${bestMark} ${metricName} ${primaryStr}`;
+        return [
+            prefix,
+            `  - changeSummary: ${summary}`,
+            `  - appliedChanges: ${changeIdsStr}`,
+            `  - appliedTips: ${tipsStr}`,
+        ].join('\n');
+    });
+    return [...head, ...lines].join('\n');
+}
+// Toolbox-rotation-hint section rendering — injected by the caller into the generate user prompt when !isBest for ≥ 2 consecutive rounds.
+// A soft hint: list already-tried techniques + the full toolbox + suggested wording; does not force the LLM to switch.
+// See docs/specs/25 §11.3 "toolbox rotation hint"
+function formatToolboxSwitchHint(recentlyUsedTips, allTipNames, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    const usedSet = new Set(recentlyUsedTips.filter((t) => t.trim().length > 0));
+    const usedList = usedSet.size > 0
+        ? Array.from(usedSet)
+            .map((t) => `\`${t}\``)
+            .join(', ')
+        : language === 'en-US'
+            ? '(no historical technique names recognized)'
+            : '（无法识别历史技巧名）';
+    const toolboxList = allTipNames.map((t) => `\`${t}\``).join(', ');
+    if (language === 'en-US') {
+        return [
+            '## Toolbox Rotation Hint',
+            '> The last 2 new versions did not improve the historical best metric. Prefer techniques not already tried.',
+            '',
+            `- Techniques tried in the last 2 rounds (appliedTips): ${usedList}`,
+            `- All toolbox techniques: ${toolboxList}`,
+            "- Suggestion: include at least one toolbox item outside the tried list in this round's `appliedTips`, unless the evidence bundle clearly requires staying with the same direction.",
+        ].join('\n');
+    }
+    return [
+        '## 工具箱轮换提示',
+        '> 近 2 轮的新版本均未刷新历史最佳指标，可能在某种技巧组合上原地转圈。请优先尝试**未使用过**的优化技巧。',
+        '',
+        `- 近 2 轮已尝试技巧（appliedTips）：${usedList}`,
+        `- 工具箱全部技巧（见 optimization-tips.md）：${toolboxList}`,
+        '- 建议：本轮 `appliedTips` 中至少包含一个上面"已尝试"清单外的工具箱条目；若 evidenceBundle 确无对应方向证据，可在 `changeSummary` 说明仍沿用原方向的理由。',
+    ].join('\n');
+}
+// Common helper: append the optimization-history section into the user prompt array
+// Call site: after the '' separator following the goal-vs-actual section, spread ...renderRoundHistorySection(history, goals)
+// Returns [formatted, ''] to keep the existing section separator style; when history is empty, returns [] to render nothing (backward compatible with first round)
+function renderRoundHistorySection(history, goals, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (!history || history.length === 0)
+        return [];
+    return [formatRoundHistory(history, goals, language), ''];
+}
+// Toolbox-rotation-hint assembly — the caller spreads this after the optimization-history section in the generate user prompt
+// When hint is undefined, returns [] and renders nothing (streak < 2 / first-round scenario)
+function renderToolboxSwitchHintSection(hint, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (!hint)
+        return [];
+    return [formatToolboxSwitchHint(hint.recentlyUsedTips, hint.allTipNames, language), ''];
+}
+// =========================
+// Cross-round history token-budget degradation — see docs/specs/25-optimizations.md §11.3
+// L0 full → L1 early rounds: changeSummary truncated to 200 chars + appliedChanges only keep changeId →
+// L2 early rounds: changeSummary truncated to 50 chars + appliedChanges cleared →
+// L3 only the most recent 1 round contains changeSummary / appliedChanges; the rest are cleared
+// Estimation calibration goes through formatRoundHistory + estimateMessagesTokens (same function as the caller probe, to avoid drift)
+// =========================
+const HISTORY_RECENT_KEEP = 3;
+const HISTORY_L1_CHANGE_SUMMARY_CHARS = 200;
+const HISTORY_L2_CHANGE_SUMMARY_CHARS = 50;
+function estimateRoundHistoryTokens(history, goals, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (history.length === 0)
+        return 0;
+    return (0, token_budget_1.estimateMessagesTokens)('', formatRoundHistory(history, goals, language), 0).inputTokens;
+}
+function fitRoundHistoryToBudget(history, budgetTokens, goals, language = shared_1.DEFAULT_PROMPT_LANGUAGE) {
+    if (!history || history.length === 0) {
+        return {
+            fitted: history,
+            level: 0,
+            truncated: false,
+            entryCount: 0,
+            budgetTokens,
+            estimatedTokens: 0,
+        };
+    }
+    const l0Tokens = estimateRoundHistoryTokens(history, goals, language);
+    if (l0Tokens <= budgetTokens) {
+        return {
+            fitted: history,
+            level: 0,
+            truncated: false,
+            entryCount: history.length,
+            budgetTokens,
+            estimatedTokens: l0Tokens,
+        };
+    }
+    const earlyCount = Math.max(0, history.length - HISTORY_RECENT_KEEP);
+    // L1: compress changeSummary in early rounds + slim appliedChanges
+    const l1 = history.map((entry, i) => {
+        if (i >= earlyCount)
+            return entry;
+        return {
+            ...entry,
+            changeSummary: (0, token_budget_1.truncateLongText)(entry.changeSummary, HISTORY_L1_CHANGE_SUMMARY_CHARS),
+            appliedChanges: entry.appliedChanges.map((c) => ({ changeId: c.changeId })),
+        };
+    });
+    const l1Tokens = estimateRoundHistoryTokens(l1, goals, language);
+    if (l1Tokens <= budgetTokens) {
+        return {
+            fitted: l1,
+            level: 1,
+            truncated: true,
+            entryCount: l1.length,
+            budgetTokens,
+            estimatedTokens: l1Tokens,
+        };
+    }
+    // L2: further truncate changeSummary in early rounds + clear appliedChanges
+    const l2 = history.map((entry, i) => {
+        if (i >= earlyCount)
+            return entry;
+        return {
+            ...entry,
+            changeSummary: (0, token_budget_1.truncateLongText)(entry.changeSummary, HISTORY_L2_CHANGE_SUMMARY_CHARS),
+            appliedChanges: [],
+        };
+    });
+    const l2Tokens = estimateRoundHistoryTokens(l2, goals, language);
+    if (l2Tokens <= budgetTokens) {
+        return {
+            fitted: l2,
+            level: 2,
+            truncated: true,
+            entryCount: l2.length,
+            budgetTokens,
+            estimatedTokens: l2Tokens,
+        };
+    }
+    // L3: only the most recent 1 round contains changeSummary / appliedChanges; the rest are cleared (but metrics + delta are kept)
+    const l3 = history.map((entry, i) => {
+        if (i === history.length - 1)
+            return entry;
+        return { ...entry, changeSummary: '', appliedChanges: [] };
+    });
+    const l3Tokens = estimateRoundHistoryTokens(l3, goals, language);
+    return {
+        fitted: l3,
+        level: 3,
+        truncated: true,
+        entryCount: l3.length,
+        budgetTokens,
+        estimatedTokens: l3Tokens,
+    };
+}
+function buildAnalyzeConfusionMessages(args) {
+    const { pair, currentVersion, metrics, goals, fieldWhitelist, roundHistory } = args;
+    const language = args.promptLanguage ?? shared_1.DEFAULT_PROMPT_LANGUAGE;
+    const system = (0, loader_1.getSystemPrompts)(language).analyzeConfusion;
+    const user = language === 'en-US'
+        ? [
+            `## Confusion Pair: \`${pair.expected}\` -> \`${pair.predicted}\` (${pair.count} failed samples; showing ${pair.samples.length})`,
+            '',
+            '## Current Prompt',
+            '```',
+            currentVersion.body,
+            '```',
+            '',
+            '## promptVariables (available, immutable)',
+            formatVariableList(fieldWhitelist.promptVariables, language),
+            '',
+            '## analysisOnlyFields (read-only; forbidden in the final prompt)',
+            formatAnalysisOnlyFields(fieldWhitelist.analysisOnlyFields, language),
+            '',
+            '## Optimization Goals vs Current Metrics',
+            formatGoalsWithProgress(goals, metrics, language),
+            '',
+            ...renderRoundHistorySection(roundHistory, goals, language),
+            '## Relevant Metrics',
+            formatRelevantMetrics(goals, metrics, language),
+            '',
+            `## Failed Samples (${pair.samples.length})`,
+            '```json',
+            JSON.stringify(pair.samples, null, 2),
+            '```',
+            '',
+            'Output JSON according to the system instructions.',
+        ].join('\n')
+        : [
+            `## 本批混淆对：\`${pair.expected}\` → \`${pair.predicted}\`（共 ${pair.count} 条失败样本，本批展示 ${pair.samples.length} 条）`,
+            '',
+            '## 当前提示词全文',
+            '```',
+            currentVersion.body,
+            '```',
+            '',
+            '## promptVariables（可用、不可改）',
+            formatVariableList(fieldWhitelist.promptVariables, language),
+            '',
+            '## analysisOnlyFields（仅可阅读、严禁出现在最终 prompt 中）',
+            formatAnalysisOnlyFields(fieldWhitelist.analysisOnlyFields, language),
+            '',
+            '## 优化目标 vs 当前实际',
+            formatGoalsWithProgress(goals, metrics, language),
+            '',
+            ...renderRoundHistorySection(roundHistory, goals, language),
+            '## 涉及范围的完整指标（仅展示与优化目标相关的范围）',
+            formatRelevantMetrics(goals, metrics, language),
+            '',
+            `## 本批失败样本（共 ${pair.samples.length} 条）`,
+            '```json',
+            JSON.stringify(pair.samples, null, 2),
+            '```',
+            '',
+            '请按 system 指令输出 JSON。',
+        ].join('\n');
+    return { system, user };
+}
+function buildAnalyzeRegressionMessages(args) {
+    const { group, currentVersion, previousVersion, metrics, goals, fieldWhitelist, roundHistory } = args;
+    const language = args.promptLanguage ?? shared_1.DEFAULT_PROMPT_LANGUAGE;
+    const system = (0, loader_1.getSystemPrompts)(language).analyzeRegression;
+    const previousPromptSection = language === 'en-US'
+        ? previousVersion
+            ? [
+                `## Previous Comparable Prompt (v${previousVersion.versionNumber}; for regression attribution only)`,
+                '```',
+                previousVersion.body,
+                '```',
+            ]
+            : ['## Previous Comparable Prompt', '(not provided; do not claim a specific prompt change caused regression)']
+        : previousVersion
+            ? [
+                `## 上一可比 prompt 模板（v${previousVersion.versionNumber}；仅用于回归归因）`,
+                '```',
+                previousVersion.body,
+                '```',
+            ]
+            : [
+                '## 上一可比 prompt 模板',
+                '（未提供上一版 prompt / diff；不得声称某段具体改动导致回归，只能基于样本描述当前 prompt 的风险倾向。）',
+            ];
+    const user = language === 'en-US'
+        ? [
+            `## Regression Samples (predicted=\`${group.predicted}\`, count ${group.count})`,
+            '',
+            ...previousPromptSection,
+            '',
+            '## Current Prompt',
+            '```',
+            currentVersion.body,
+            '```',
+            '',
+            '## promptVariables (available, immutable)',
+            formatVariableList(fieldWhitelist.promptVariables, language),
+            '',
+            '## analysisOnlyFields (read-only; forbidden in the final prompt)',
+            formatAnalysisOnlyFields(fieldWhitelist.analysisOnlyFields, language),
+            '',
+            '## Optimization Goals vs Current Metrics',
+            formatGoalsWithProgress(goals, metrics, language),
+            '',
+            ...renderRoundHistorySection(roundHistory, goals, language),
+            '## Relevant Metrics',
+            formatRelevantMetrics(goals, metrics, language),
+            '',
+            `## Regression Samples (${group.samples.length})`,
+            '```json',
+            JSON.stringify(group.samples, null, 2),
+            '```',
+            '',
+            'Output JSON according to the system instructions.',
+        ].join('\n')
+        : [
+            `## 本批回归样本（predicted=\`${group.predicted}\`，共 ${group.count} 条）`,
+            '',
+            ...previousPromptSection,
+            '',
+            '## 当前提示词全文',
+            '```',
+            currentVersion.body,
+            '```',
+            '',
+            '## promptVariables（可用、不可改）',
+            formatVariableList(fieldWhitelist.promptVariables, language),
+            '',
+            '## analysisOnlyFields（仅可阅读、严禁出现在最终 prompt 中）',
+            formatAnalysisOnlyFields(fieldWhitelist.analysisOnlyFields, language),
+            '',
+            '## 优化目标 vs 当前实际',
+            formatGoalsWithProgress(goals, metrics, language),
+            '',
+            ...renderRoundHistorySection(roundHistory, goals, language),
+            '## 涉及范围的完整指标（仅展示与优化目标相关的范围）',
+            formatRelevantMetrics(goals, metrics, language),
+            '',
+            `## 回归样本（共 ${group.samples.length} 条）`,
+            '```json',
+            JSON.stringify(group.samples, null, 2),
+            '```',
+            '',
+            '请按 system 指令输出 JSON。',
+        ].join('\n');
+    return { system, user };
+}
+function buildSummarizeMessages(args) {
+    const language = args.promptLanguage ?? shared_1.DEFAULT_PROMPT_LANGUAGE;
+    const system = (0, loader_1.getSystemPrompts)(language).summarize;
+    const user = language === 'en-US'
+        ? [
+            '## Optimization Goals vs Current Metrics',
+            formatGoalsWithProgress(args.goals, args.metrics, language),
+            '',
+            ...renderRoundHistorySection(args.roundHistory, args.goals, language),
+            '## Relevant Metrics',
+            formatRelevantMetrics(args.goals, args.metrics, language),
+            '',
+            `## Child Analysis Results (${args.collectedBatches.length} batches)`,
+            '```json',
+            JSON.stringify(args.collectedBatches, null, 2),
+            '```',
+            '',
+            'Summarize these batches into final JSON according to the system instructions.',
+        ].join('\n')
+        : [
+            '## 优化目标 vs 当前实际',
+            formatGoalsWithProgress(args.goals, args.metrics, language),
+            '',
+            ...renderRoundHistorySection(args.roundHistory, args.goals, language),
+            '## 涉及范围的完整指标（仅展示与优化目标相关的范围）',
+            formatRelevantMetrics(args.goals, args.metrics, language),
+            '',
+            `## 子分析结果汇总（${args.collectedBatches.length} 个 batch）`,
+            '```json',
+            JSON.stringify(args.collectedBatches, null, 2),
+            '```',
+            '',
+            '请按 system 指令把这些 batch 汇总成最终 JSON。',
+        ].join('\n');
+    return { system, user };
+}
+function buildGenerateMessages(args) {
+    const { currentVersion, errorAnalysisText, analysisEvidenceBundle, metrics, goals, fieldWhitelist, optimizationHint, roundHistory, toolboxSwitchHint, } = args;
+    const language = args.promptLanguage ?? shared_1.DEFAULT_PROMPT_LANGUAGE;
+    const system = (0, loader_1.getSystemPrompts)(language).generate;
+    // The output schema is no longer re-stated by the LLM — the system auto-assembles the output-format section from the schema at runtime and appends it to the body tail.
+    // Here we stuff the "auto-assembled output format section" as-is into the user prompt purely to let the LLM see what the full prompt looks like,
+    // avoiding it reinventing the wheel (the 6th hard constraint in the system prompt already forbids newPromptBody from re-stating the output format).
+    const autoOutputFormat = (0, shared_1.buildOutputFormatInstruction)(currentVersion.outputSchema, { language });
+    const schemaSection = autoOutputFormat
+        ? language === 'en-US'
+            ? [
+                '## Runtime Output-Format Section (reference only; do not restate it in newPromptBody)',
+                '> The system appends this section to newPromptBody at runtime. The output format is determined only by output schema.',
+                '',
+                autoOutputFormat,
+            ]
+            : [
+                '## 运行时自动拼接的输出格式段（仅供参考；禁止在 newPromptBody 中复述任何输出格式 / JSON schema / 字段说明）',
+                '> 下面这段会由系统在运行时拼接到 newPromptBody 尾部，输出格式由 output schema 唯一决定，不需要你写。',
+                '',
+                autoOutputFormat,
+            ]
+        : [];
+    const judgmentSection = currentVersion.judgmentRules
+        ? [
+            language === 'en-US' ? '## Immutable Judgment Rules' : '## 不可改动的 judgment rules',
+            '```json',
+            JSON.stringify(currentVersion.judgmentRules, null, 2),
+            '```',
+        ]
+        : [];
+    // base's already-used ∩ whitelist = must be retained verbatim in newPromptBody (system hard constraint #1)
+    // Explicitly listing them is less likely to be missed by the LLM during a full-section rewrite than burying them in the text
+    const allowedSet = new Set(fieldWhitelist.promptVariables);
+    const requiredVariables = extractVariableNames(currentVersion.body).filter((v) => allowedSet.has(v));
+    const requiredVariablesSection = requiredVariables.length > 0
+        ? [
+            language === 'en-US'
+                ? '## Required Variable Placeholders (used by base; do not remove)'
+                : '## 必须保留的变量占位（base 已使用，禁止删除 — 硬约束 #1）',
+            language === 'en-US'
+                ? '> These placeholders must appear exactly in newPromptBody. They are the only runtime path from sample data into the business model.'
+                : '> 下列占位**必须逐字、原样**出现在 newPromptBody 中（位置随意）。它们是运行时把样本数据注入业务模型的唯一通道；删掉它们模型推理时根本看不到样本，整批输出会立即塌缩到同一标签。',
+            '',
+            ...requiredVariables.map((v) => `- \`{{${v}}}\``),
+            '',
+        ]
+        : [];
+    const user = language === 'en-US'
+        ? [
+            `## Current Prompt Template (v${currentVersion.versionNumber})`,
+            '```',
+            currentVersion.body,
+            '```',
+            '',
+            ...requiredVariablesSection,
+            '## Structured Evidence Bundle (primary source)',
+            analysisEvidenceBundle
+                ? ['```json', JSON.stringify(analysisEvidenceBundle, null, 2), '```'].join('\n')
+                : '(no structured evidence bundle; use fallback summary below)',
+            '',
+            '## Error Analysis Fallback Summary',
+            errorAnalysisText,
+            '',
+            '## Optimization Goals vs Current Metrics',
+            formatGoalsWithProgress(goals, metrics, language),
+            '',
+            ...renderRoundHistorySection(roundHistory, goals, language),
+            ...renderToolboxSwitchHintSection(toolboxSwitchHint, language),
+            '## Relevant Metrics',
+            formatRelevantMetrics(goals, metrics, language),
+            '',
+            '## promptVariables (available, immutable)',
+            formatVariableList(fieldWhitelist.promptVariables, language),
+            '',
+            '## analysisOnlyFields (forbidden in the new prompt)',
+            formatAnalysisOnlyFields(fieldWhitelist.analysisOnlyFields, language),
+            '',
+            '## modifiableSections',
+            fieldWhitelist.modifiableSections && fieldWhitelist.modifiableSections.length > 0
+                ? fieldWhitelist.modifiableSections.map((s) => `- \`${s}\``).join('\n')
+                : '(not constrained; rewrite as needed while respecting all other constraints)',
+            '',
+            ...schemaSection,
+            '',
+            ...judgmentSection,
+            '',
+            '## User Generation Guidance',
+            optimizationHint && optimizationHint.trim().length > 0 ? optimizationHint : '(none)',
+            '',
+            'Output JSON according to the system instructions.',
+        ].join('\n')
+        : [
+            '## 当前 prompt 模板（v' + currentVersion.versionNumber + '）',
+            '```',
+            currentVersion.body,
+            '```',
+            '',
+            ...requiredVariablesSection,
+            '## 结构化错误证据包（来自 analyze / summarize 阶段，优先依据）',
+            analysisEvidenceBundle
+                ? ['```json', JSON.stringify(analysisEvidenceBundle, null, 2), '```'].join('\n')
+                : '（无结构化证据包，使用下方旧摘要 fallback）',
+            '',
+            '## 错误分析摘要 fallback（仅在证据包缺字段时参考）',
+            errorAnalysisText,
+            '',
+            '## 优化目标 vs 当前实际',
+            formatGoalsWithProgress(goals, metrics, language),
+            '',
+            ...renderRoundHistorySection(roundHistory, goals, language),
+            ...renderToolboxSwitchHintSection(toolboxSwitchHint, language),
+            '## 涉及范围的完整指标（仅展示与优化目标相关的范围）',
+            formatRelevantMetrics(goals, metrics, language),
+            '',
+            '## promptVariables（可用、不可改）',
+            formatVariableList(fieldWhitelist.promptVariables, language),
+            '',
+            '## analysisOnlyFields（严禁出现在新 prompt 中）',
+            formatAnalysisOnlyFields(fieldWhitelist.analysisOnlyFields, language),
+            '',
+            '## modifiableSections（仅可在这些段落内改）',
+            fieldWhitelist.modifiableSections && fieldWhitelist.modifiableSections.length > 0
+                ? fieldWhitelist.modifiableSections.map((s) => `- \`${s}\``).join('\n')
+                : '（未限定 — 可在不违反其它约束的前提下整体改写）',
+            '',
+            ...schemaSection,
+            '',
+            ...judgmentSection,
+            '',
+            '## 用户给的提示词生成指引',
+            optimizationHint && optimizationHint.trim().length > 0 ? optimizationHint : '（无）',
+            '',
+            '请按 system 指令输出 JSON。',
+        ].join('\n');
+    return { system, user };
+}
+// =========================
+// Variable name extraction — used to validate "the new prompt can only use a subset of promptVariables"
+// =========================
+function extractVariableNames(promptBody) {
+    const re = /\{\{\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*\}\}/g;
+    const set = new Set();
+    let m;
+    while ((m = re.exec(promptBody)) !== null) {
+        set.add(m[1]);
+    }
+    return [...set];
+}
+//# sourceMappingURL=prompts.js.map