dsh-plugin-om 0.0.55 → 0.0.57

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (72) hide show
  1. package/README.md +74 -18
  2. package/dist/client/OmCompactionCard.d.ts.map +1 -1
  3. package/dist/client/OmEvalModelPanel.d.ts +22 -0
  4. package/dist/client/OmEvalModelPanel.d.ts.map +1 -0
  5. package/dist/client/OmRecompressMenu.d.ts +3 -3
  6. package/dist/client/OmRecompressMenu.d.ts.map +1 -1
  7. package/dist/client/compaction-eval-jump.d.ts +20 -0
  8. package/dist/client/compaction-eval-jump.d.ts.map +1 -0
  9. package/dist/client/definition.d.ts +2 -2
  10. package/dist/client/definition.d.ts.map +1 -1
  11. package/dist/client/eval-model-panel.d.ts +57 -0
  12. package/dist/client/eval-model-panel.d.ts.map +1 -0
  13. package/dist/client/eval-model-register.d.ts +10 -0
  14. package/dist/client/eval-model-register.d.ts.map +1 -0
  15. package/dist/client/index.d.ts +2 -1
  16. package/dist/client/index.d.ts.map +1 -1
  17. package/dist/client/locales.d.ts +13 -2
  18. package/dist/client/locales.d.ts.map +1 -1
  19. package/dist/client/om-config-controller.d.ts +4 -1
  20. package/dist/client/om-config-controller.d.ts.map +1 -1
  21. package/dist/client/recompress-run.d.ts +23 -2
  22. package/dist/client/recompress-run.d.ts.map +1 -1
  23. package/dist/client.js +675 -116
  24. package/dist/client.js.map +1 -1
  25. package/dist/compaction-eval-command.d.ts +54 -0
  26. package/dist/compaction-eval-command.d.ts.map +1 -0
  27. package/dist/compaction-log.d.ts +31 -9
  28. package/dist/compaction-log.d.ts.map +1 -1
  29. package/dist/compress-loop.d.ts +21 -10
  30. package/dist/compress-loop.d.ts.map +1 -1
  31. package/dist/compress-tools.d.ts +16 -8
  32. package/dist/compress-tools.d.ts.map +1 -1
  33. package/dist/compress-view.d.ts +1 -3
  34. package/dist/compress-view.d.ts.map +1 -1
  35. package/dist/compress.d.ts +117 -8
  36. package/dist/compress.d.ts.map +1 -1
  37. package/dist/config.d.ts +4 -0
  38. package/dist/config.d.ts.map +1 -1
  39. package/dist/constants.d.ts +3 -1
  40. package/dist/constants.d.ts.map +1 -1
  41. package/dist/eval-session-text.d.ts +28 -0
  42. package/dist/eval-session-text.d.ts.map +1 -0
  43. package/dist/index.d.ts +4 -1
  44. package/dist/index.d.ts.map +1 -1
  45. package/dist/index.mjs +1392 -480
  46. package/dist/log-index.d.ts +2 -2
  47. package/dist/log-index.d.ts.map +1 -1
  48. package/dist/om-event.d.ts +26 -2
  49. package/dist/om-event.d.ts.map +1 -1
  50. package/dist/precompress.d.ts +100 -0
  51. package/dist/precompress.d.ts.map +1 -0
  52. package/dist/recall-xml.d.ts +2 -1
  53. package/dist/recall-xml.d.ts.map +1 -1
  54. package/dist/recompress-gate.d.ts +11 -0
  55. package/dist/recompress-gate.d.ts.map +1 -0
  56. package/dist/recompress.d.ts +20 -4
  57. package/dist/recompress.d.ts.map +1 -1
  58. package/dist/skill/compaction-eval/SKILL.md +72 -0
  59. package/dist/skill/compaction-eval/references/facts-prompt.md +42 -0
  60. package/dist/skill/compaction-eval/references/judge-prompt.md +48 -0
  61. package/dist/skill/compaction-eval/references/rules.md +56 -0
  62. package/dist/skill/compaction-eval/scripts/aggregate.mjs +372 -0
  63. package/dist/skill/compaction-eval/scripts/compress.mjs +15675 -0
  64. package/dist/skill/compaction-eval/scripts/merge.mjs +287 -0
  65. package/dist/skill/compaction-eval/scripts/prepare.mjs +1620 -0
  66. package/dist/skill/compaction-eval/scripts/view.mjs +15983 -0
  67. package/dist/system-reminder.d.ts.map +1 -1
  68. package/dist/thinking-binding.d.ts +28 -0
  69. package/dist/thinking-binding.d.ts.map +1 -0
  70. package/dist/types.d.ts +9 -6
  71. package/dist/types.d.ts.map +1 -1
  72. package/package.json +10 -3
@@ -0,0 +1,372 @@
1
+ import { existsSync, readFileSync, writeFileSync } from "node:fs";
2
+ import { basename, join, resolve } from "node:path";
3
+ import { fileURLToPath } from "node:url";
4
+ import { createHash } from "node:crypto";
5
+ //#region .agents/skills/compaction-eval/scripts/judging.ts
6
+ /**
7
+ * facts 与评审文件的共享类型、判定词与校验逻辑,供 aggregate.ts 与 merge.ts 使用。
8
+ */
9
+ /** 事实判定词到得分。 */
10
+ const VERDICT = {
11
+ 保留: 1,
12
+ 部分: .5,
13
+ 丢失: 0,
14
+ 歪曲: -1
15
+ };
16
+ /** 无关信息判定词到得分。 */
17
+ const D_VERDICT = {
18
+ 已省略: 1,
19
+ 已概括: .7,
20
+ 详细保留: 0
21
+ };
22
+ /** 事实分类权重。 */
23
+ const W = {
24
+ A: 3,
25
+ B: 2,
26
+ C: 1
27
+ };
28
+ /**
29
+ * 校验单个评审内容:事实与无关信息全覆盖、无重复、无未知 id、判定词合法。
30
+ * 返回全部问题(带评审文件名前缀),为空即通过。
31
+ */
32
+ function judgeProblems(judge, fj, name) {
33
+ const problems = [];
34
+ for (const [items, words, label, all] of [[
35
+ judge.facts,
36
+ VERDICT,
37
+ "事实",
38
+ fj.facts
39
+ ], [
40
+ judge.irrelevant,
41
+ D_VERDICT,
42
+ "无关信息",
43
+ fj.irrelevant
44
+ ]]) {
45
+ const byId = /* @__PURE__ */ new Map();
46
+ for (const x of items ?? []) if (byId.has(x.id)) problems.push(`${name} ${label} ${x.id} 重复判定`);
47
+ else byId.set(x.id, x);
48
+ for (const f of all) {
49
+ const x = byId.get(f.id);
50
+ if (x === void 0) problems.push(`${name} 未覆盖${label} ${f.id}`);
51
+ else if (!(x.verdict in words)) problems.push(`${name} ${label} ${f.id} 判定词非法:${x.verdict}`);
52
+ }
53
+ for (const x of items ?? []) if (!all.some((f) => f.id === x.id)) problems.push(`${name} ${label} ${x.id} 不在清单中`);
54
+ }
55
+ return problems;
56
+ }
57
+ /** facts.json 的规范哈希:按稳定序列化计算,供评审文件记录与 aggregate 比对。 */
58
+ function factsHash(fj) {
59
+ const canon = JSON.stringify({
60
+ facts: fj.facts.map((f) => ({
61
+ id: f.id,
62
+ fact: f.fact,
63
+ category: f.category,
64
+ position: f.position,
65
+ source_indices: f.source_indices,
66
+ check_hint: f.check_hint,
67
+ evidence_after: f.evidence_after ?? ""
68
+ })),
69
+ irrelevant: fj.irrelevant.map((d) => ({
70
+ id: d.id,
71
+ fact: d.fact,
72
+ is_credential: d.is_credential ?? false
73
+ }))
74
+ });
75
+ return createHash("sha256").update(canon, "utf8").digest("hex");
76
+ }
77
+ //#endregion
78
+ //#region .agents/skills/compaction-eval/scripts/paths.ts
79
+ /** 用法行中的脚本名:取实际入口文件名,源码版为 .ts,发布版为 .mjs。 */
80
+ function selfName(fallback) {
81
+ const entry = process.argv[1];
82
+ return entry === void 0 ? fallback : basename(entry);
83
+ }
84
+ //#endregion
85
+ //#region .agents/skills/compaction-eval/scripts/aggregate.ts
86
+ /**
87
+ * 汇总盲评结果与日志指标,输出 scores.json 与 scores.md。
88
+ *
89
+ * 用法:node aggregate.ts <评估目录> [--reference original|provider/model[:effort]]...
90
+ * 输入:meta.json、_blind_key.json、facts.json、judge-<V>-1.json、judge-<V>-2.json、V*.xml、source.xml。
91
+ * 维度:关键信息保留 70%、无关信息压缩 20%、精简度 5%、模型耗时 5%。
92
+ * --reference 可重复,选中的候选作为参考:照常算保留率与无关信息压缩,
93
+ * 精简度与耗时按参与候选的最短产物、最快耗时折算并封顶 100,不参与排名与汇总。
94
+ * 诊断输出不计分:分段保留率、分段压缩率、后续依据保留率、缓存 token、评审分歧。
95
+ */
96
+ /** 综合分维度权重。 */
97
+ const WEIGHTS = {
98
+ retention: .7,
99
+ conciseness: .05,
100
+ irrelevant: .2,
101
+ time: .05
102
+ };
103
+ /** 每条无依据断言扣除的分子分。 */
104
+ const CLAIM_PENALTY = 2;
105
+ /** 每个变体的评审数量。 */
106
+ const JUDGE_COUNT = 2;
107
+ /** 位置分段。 */
108
+ const POS = [
109
+ "head",
110
+ "mid",
111
+ "tail"
112
+ ];
113
+ /** 事实分类。 */
114
+ const CATS = [
115
+ "A",
116
+ "B",
117
+ "C"
118
+ ];
119
+ /** 保留一位小数。 */
120
+ const r1 = (x) => Math.round(x * 10) / 10;
121
+ /** 保留两位小数。 */
122
+ const r2 = (x) => Math.round(x * 100) / 100;
123
+ /** index 所在分段。 */
124
+ function bucketOf(i, bk) {
125
+ return POS.find((p) => bk[p][0] <= i && i <= bk[p][1]) ?? "tail";
126
+ }
127
+ /** assistant 条目的 [起, 止, CDATA 字符数];toolcall 条目计 args 与 result。 */
128
+ function entrySpans(xml) {
129
+ const out = [];
130
+ for (const m of xml.matchAll(/<assistant([^>]*)>([\s\S]*?)<\/assistant>/g)) {
131
+ const attrs = m[1] ?? "";
132
+ const body = m[2] ?? "";
133
+ const se = /start="(\d+)" end="(\d+)"/.exec(attrs);
134
+ const idx = /index="(\d+)"/.exec(attrs);
135
+ const lo = Number(se?.[1] ?? idx?.[1]);
136
+ const hi = Number(se?.[2] ?? idx?.[1]);
137
+ const chars = [...body.matchAll(/<!\[CDATA\[([\s\S]*?)\]\]>/g)].reduce((s, c) => s + (c[1]?.length ?? 0), 0);
138
+ out.push([
139
+ lo,
140
+ hi,
141
+ chars
142
+ ]);
143
+ }
144
+ return out;
145
+ }
146
+ /** 按分段统计字符数,区间条目按覆盖的 assistant index 均摊。 */
147
+ function positionChars(xml, bk, assistantIdx) {
148
+ const acc = {
149
+ head: 0,
150
+ mid: 0,
151
+ tail: 0
152
+ };
153
+ for (const [lo, hi, chars] of entrySpans(xml)) {
154
+ const covered = Array.from({ length: hi - lo + 1 }, (_, k) => lo + k).filter((i) => assistantIdx.has(i));
155
+ const idx = covered.length ? covered : [lo];
156
+ for (const i of idx) acc[bucketOf(i, bk)] += chars / idx.length;
157
+ }
158
+ return {
159
+ head: Math.round(acc.head),
160
+ mid: Math.round(acc.mid),
161
+ tail: Math.round(acc.tail)
162
+ };
163
+ }
164
+ /** 加权得分率(%)。 */
165
+ function weightedRatio(items) {
166
+ const den = items.reduce((s, [c]) => s + W[c], 0);
167
+ return den ? r1(items.reduce((s, [c, v]) => s + W[c] * v, 0) / den * 100) : 0;
168
+ }
169
+ /** 读取评估目录下的 JSON 文件。 */
170
+ function readJson(dir, name) {
171
+ return JSON.parse(readFileSync(join(dir, name), "utf8"));
172
+ }
173
+ /** 读取并校验单个评审文件:必须存在、全覆盖、判定词合法、与当前 facts.json 同版。 */
174
+ function loadJudge(dir, name, fj) {
175
+ if (!existsSync(join(dir, name))) throw new Error(`缺少评审文件 ${name}`);
176
+ const j = readJson(dir, name);
177
+ const problems = judgeProblems(j, fj, name);
178
+ const sha = factsHash(fj);
179
+ if (j.facts_sha !== sha) problems.push(`${name} 基于旧版 facts.json(facts_sha 不符),修正清单后需重新评审`);
180
+ if (problems.length) throw new Error(problems.join("\n"));
181
+ return j;
182
+ }
183
+ /** 解析参考选择器,返回命中的变体号:original 选原始压缩,provider/model[:effort] 选模型。 */
184
+ function referenceVariants(key, sel) {
185
+ const entries = Object.entries(key);
186
+ if (sel === "original") return entries.filter(([, k]) => !k.recompress).map(([v]) => v);
187
+ const cut = sel.lastIndexOf(":");
188
+ const model = cut >= 0 ? sel.slice(0, cut) : sel;
189
+ const effort = cut >= 0 ? sel.slice(cut + 1) : void 0;
190
+ return entries.filter(([, k]) => k.model === model && (effort === void 0 || k.effort === effort)).map(([v]) => v);
191
+ }
192
+ /** 综合分均值与样本标准差;变体不足 2 个的模型不输出。 */
193
+ function modelSummary(variants) {
194
+ const groups = /* @__PURE__ */ new Map();
195
+ for (const r of Object.values(variants)) {
196
+ const g = groups.get(r.model) ?? [];
197
+ g.push(r.composite);
198
+ groups.set(r.model, g);
199
+ }
200
+ const out = [];
201
+ for (const [model, xs] of groups) {
202
+ if (xs.length < 2) continue;
203
+ const mean = xs.reduce((s, x) => s + x, 0) / xs.length;
204
+ const va = xs.reduce((s, x) => s + (x - mean) ** 2, 0) / (xs.length - 1);
205
+ out.push({
206
+ model,
207
+ n: xs.length,
208
+ mean: r1(mean),
209
+ std: r1(Math.sqrt(va))
210
+ });
211
+ }
212
+ return out.sort((a, b) => b.mean - a.mean);
213
+ }
214
+ /** 对全部变体评分;referenceSels 指定的候选作为参考,不参与排名与汇总。 */
215
+ function scoreAll(dir, referenceSels = []) {
216
+ const text = (name) => readFileSync(join(dir, name), "utf8");
217
+ const meta = readJson(dir, "meta.json");
218
+ const key = readJson(dir, "_blind_key.json");
219
+ const fj = readJson(dir, "facts.json");
220
+ const bk = meta.buckets;
221
+ const sourceXml = text("source.xml");
222
+ const assistantIdx = new Set(entrySpans(sourceXml).map(([lo]) => lo));
223
+ const source_chars = positionChars(sourceXml, bk, assistantIdx);
224
+ const variants = {};
225
+ const disagreements = [];
226
+ const refSet = /* @__PURE__ */ new Set();
227
+ for (const sel of referenceSels) {
228
+ const hit = referenceVariants(key, sel);
229
+ if (!hit.length) throw new Error(`参考选择器 ${sel} 未匹配任何候选`);
230
+ for (const v of hit) refSet.add(v);
231
+ }
232
+ if (refSet.size && refSet.size === Object.keys(key).length) throw new Error("全部候选都被设为参考,没有参与排名的候选");
233
+ for (const [v, k] of Object.entries(key).sort(([a], [b]) => a.localeCompare(b))) {
234
+ const js = Array.from({ length: JUDGE_COUNT }, (_, i) => loadJudge(dir, `judge-${v}-${i + 1}.json`, fj));
235
+ const verdicts = (j) => new Map(j.facts.map((x) => [x.id, x.verdict]));
236
+ const dverdicts = (j) => new Map(j.irrelevant.map((x) => [x.id, x.verdict]));
237
+ const vm = js.map(verdicts);
238
+ const dm = js.map(dverdicts);
239
+ /** 单条事实两名评审的平均判定值;分歧只记录一次。 */
240
+ const vals = /* @__PURE__ */ new Map();
241
+ for (const f of fj.facts) {
242
+ const vs = vm.map((m) => m.get(f.id) ?? "丢失");
243
+ if (vs[0] !== vs[1]) disagreements.push({
244
+ variant: v,
245
+ id: f.id,
246
+ verdicts: [vs[0] ?? "", vs[1] ?? ""]
247
+ });
248
+ vals.set(f.id, vs.reduce((s, x) => s + (VERDICT[x] ?? 0), 0) / vs.length);
249
+ }
250
+ const factValue = (id) => vals.get(id) ?? 0;
251
+ /** 分段或分类得分为空时记 null,区别于全部丢失的 0。 */
252
+ const sc = (fs) => fs.length ? weightedRatio(fs.map((f) => [f.category, factValue(f.id)])) : null;
253
+ const chars = positionChars(text(`${v}.xml`), bk, assistantIdx);
254
+ const unsupported = js.reduce((s, j) => s + (j.unsupported_claims?.length ?? 0), 0) / js.length;
255
+ const extra = js.reduce((s, j) => s + (j.extra_irrelevant?.length ?? 0), 0) / js.length;
256
+ const den = fj.facts.reduce((s, f) => s + W[f.category], 0);
257
+ const num = fj.facts.reduce((s, f) => s + W[f.category] * factValue(f.id), 0) - CLAIM_PENALTY * unsupported;
258
+ const irrelevant = fj.irrelevant.length ? r1(fj.irrelevant.reduce((s, x) => s + dm.reduce((t, m) => t + (D_VERDICT[m.get(x.id) ?? "详细保留"] ?? 0), 0) / dm.length, 0) / fj.irrelevant.length * 100) : 100;
259
+ const after = fj.facts.filter((f) => (f.evidence_after ?? "").trim() !== "");
260
+ variants[v] = {
261
+ model: `${k.model} ${k.effort}${k.recompress ? "" : "(原始)"}`,
262
+ reference: refSet.has(v),
263
+ retention: den ? r1(num / den * 100) : 0,
264
+ retention_after: after.length ? weightedRatio(after.map((f) => [f.category, factValue(f.id)])) : null,
265
+ after_facts: after.length,
266
+ by_category: Object.fromEntries(CATS.map((c) => [c, sc(fj.facts.filter((f) => f.category === c))])),
267
+ by_position: Object.fromEntries(POS.map((p) => [p, sc(fj.facts.filter((f) => f.position === p))])),
268
+ A_failed: fj.facts.filter((f) => f.category === "A" && factValue(f.id) < .5).map((f) => f.id),
269
+ irrelevant,
270
+ credential_leak: fj.irrelevant.some((x) => x.is_credential && dm.some((m) => m.get(x.id) === "详细保留")),
271
+ unsupported_claims: r1(unsupported),
272
+ extra_irrelevant: r1(extra),
273
+ chars,
274
+ product_chars: chars.head + chars.mid + chars.tail,
275
+ compression: Object.fromEntries(POS.map((p) => [p, source_chars[p] ? r2(chars[p] / source_chars[p]) : null])),
276
+ conciseness: 0,
277
+ time: 0,
278
+ composite: 0,
279
+ model_sec: k.model_sec,
280
+ wall_sec: k.wall_sec,
281
+ requests: k.requests,
282
+ tokens_in: k.tokens_in,
283
+ tokens_cache_read: k.tokens_cache_read,
284
+ tokens_cache_write: k.tokens_cache_write,
285
+ tokens_out: k.tokens_out
286
+ };
287
+ }
288
+ /** 参与排名的候选:精简度与耗时的基准只取自这些候选。 */
289
+ const pool = Object.values(variants).filter((r) => !r.reference);
290
+ const pmin = Math.min(...pool.map((r) => r.product_chars));
291
+ const tmin = Math.min(...pool.map((r) => r.model_sec));
292
+ for (const r of Object.values(variants)) {
293
+ r.conciseness = r.product_chars ? Math.min(100, r1(pmin / r.product_chars * 100)) : 0;
294
+ r.time = r.model_sec ? Math.min(100, r1(tmin / r.model_sec * 100)) : 0;
295
+ r.composite = r1(WEIGHTS.retention * r.retention + WEIGHTS.conciseness * r.conciseness + WEIGHTS.irrelevant * r.irrelevant + WEIGHTS.time * r.time);
296
+ }
297
+ return {
298
+ source_chars,
299
+ variants,
300
+ disagreements
301
+ };
302
+ }
303
+ /** 渲染 scores.md。 */
304
+ function renderMarkdown(dir, result) {
305
+ const meta = readJson(dir, "meta.json");
306
+ const facts = readJson(dir, "facts.json").facts;
307
+ const { variants, disagreements, source_chars } = result;
308
+ const all = Object.entries(variants);
309
+ const pool = all.filter(([, r]) => !r.reference);
310
+ const refs = all.filter(([, r]) => r.reference);
311
+ const phase = meta.phase ? ` ${meta.phase}` : "";
312
+ const header = "| 变体 | 模型 | 综合 | 保留 | 精简 | 无关压缩 | 耗时 | A 类 | 无依据断言 | 模型耗时 | 输入/输出 token |";
313
+ const rows = (xs) => xs.sort(([, a], [, b]) => b.composite - a.composite).map(([v, r]) => `| ${v} | ${r.model} | ${r.composite} | ${r.retention} | ${r.conciseness} | ${r.irrelevant} | ${r.time} | ${r.by_category.A ?? "-"} | ${r.unsupported_claims} | ${r.model_sec}s | ${r.tokens_in} / ${r.tokens_out} |`);
314
+ const lines = [
315
+ `# 压缩质量评分:${meta.session}${phase} [${meta.range[0]}..${meta.range[1]}]`,
316
+ "",
317
+ ...[header, "|---|---|---|---|---|---|---|---|---|---|---|"],
318
+ ...rows(pool)
319
+ ];
320
+ if (refs.length) lines.push("", "## 参考候选,不参与排名与汇总", "", header, "|---|---|---|---|---|---|---|---|---|---|---|", ...rows(refs));
321
+ const summary = modelSummary(Object.fromEntries(pool));
322
+ if (summary.length) {
323
+ lines.push("", "## 按模型汇总", "", "| 模型 | 变体数 | 综合均值 | 标准差 |", "|---|---|---|---|");
324
+ lines.push(...summary.map((s) => `| ${s.model} | ${s.n} | ${s.mean} | ${s.std} |`));
325
+ }
326
+ lines.push("", "## 诊断,不计分", "");
327
+ for (const [v, r] of all) lines.push(`- ${v} 分段保留 head ${r.by_position.head ?? "-"} / mid ${r.by_position.mid ?? "-"} / tail ${r.by_position.tail ?? "-"},分段压缩率 head ${r.compression.head ?? "-"} / mid ${r.compression.mid ?? "-"} / tail ${r.compression.tail ?? "-"},后续依据保留率 ${r.retention_after ?? "-"}(${r.after_facts} 条),清单外无关内容 ${r.extra_irrelevant} 条,缓存读 ${r.tokens_cache_read} / 写 ${r.tokens_cache_write}`);
328
+ lines.push(`- 原文分段字符 head ${source_chars.head} / mid ${source_chars.mid} / tail ${source_chars.tail}`);
329
+ if (disagreements.length) {
330
+ lines.push("", "## 两名评审判定不一致的条目", "");
331
+ lines.push(...disagreements.map((d) => `- ${d.variant} ${d.id}:${d.verdicts.join(" / ")}`));
332
+ }
333
+ const common = facts.filter((f) => f.category === "A" && pool.filter(([, r]) => r.A_failed.includes(f.id)).length * 2 >= pool.length);
334
+ const leaks = all.filter(([, r]) => r.credential_leak).map(([v]) => `- ${v}`);
335
+ lines.push("", "## 过半参与候选失败的 A 类事实,需回 source.xml 核实", "");
336
+ lines.push(...common.length ? common.map((f) => `- ${f.id} ${f.fact}`) : ["- 无"]);
337
+ lines.push("", "## 凭据泄漏", "", ...leaks.length ? leaks : ["- 无"]);
338
+ return `${lines.join("\n")}\n`;
339
+ }
340
+ /** 命令行入口。 */
341
+ function main() {
342
+ const usage = `用法:node ${selfName("aggregate.ts")} <评估目录> [--reference original|provider/model[:effort]]...`;
343
+ const args = process.argv.slice(2);
344
+ const refs = [];
345
+ const positional = [];
346
+ for (let i = 0; i < args.length; i++) {
347
+ const a = args[i] ?? "";
348
+ if (a === "--reference") {
349
+ const v = args[i + 1];
350
+ if (!v) {
351
+ console.error(usage);
352
+ process.exit(2);
353
+ }
354
+ refs.push(v);
355
+ i++;
356
+ } else if (a.startsWith("--reference=")) refs.push(a.slice(12));
357
+ else positional.push(a);
358
+ }
359
+ if (positional.length !== 1) {
360
+ console.error(usage);
361
+ process.exit(2);
362
+ }
363
+ const dir = positional[0] ?? ".";
364
+ const result = scoreAll(dir, refs);
365
+ writeFileSync(join(dir, "scores.json"), JSON.stringify(result, null, 1), "utf8");
366
+ const md = renderMarkdown(dir, result);
367
+ writeFileSync(join(dir, "scores.md"), md, "utf8");
368
+ console.log(md);
369
+ }
370
+ if (process.argv[1] && resolve(process.argv[1]) === fileURLToPath(import.meta.url)) main();
371
+ //#endregion
372
+ export { entrySpans, loadJudge, modelSummary, positionChars, referenceVariants, scoreAll, weightedRatio };