mslxdff 0.1.161 → 0.1.163

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 (116) hide show
  1. package/bin/mslxdff.js +4 -4
  2. package/package.json +1 -1
  3. package/src/auto.js +254 -254
  4. package/src/bench/cline-bench.js +42 -42
  5. package/src/bench/probe.js +70 -70
  6. package/src/bench/report.js +162 -162
  7. package/src/bench/runner.js +77 -77
  8. package/src/bench/via-probe.js +124 -124
  9. package/src/bench/via-routes.js +87 -87
  10. package/src/bench/workbuddy-bench.js +54 -54
  11. package/src/chat/engine.js +160 -160
  12. package/src/chat/gateway.js +163 -163
  13. package/src/chat/orchestrator.js +234 -234
  14. package/src/chat/prompt.js +70 -70
  15. package/src/chat/repl.js +88 -88
  16. package/src/chat/terminal.js +135 -135
  17. package/src/chat/tools.js +306 -306
  18. package/src/chat-pipeline/index.js +127 -126
  19. package/src/chat-pipeline/policy.js +76 -76
  20. package/src/cli/commands/group.js +249 -249
  21. package/src/cli/commands/model/picks.js +50 -50
  22. package/src/cli/commands/provider/bench-via.js +247 -247
  23. package/src/cli/commands/provider/bench.js +141 -141
  24. package/src/cli/commands/provider/index.js +126 -126
  25. package/src/cli/commands/provider/models.js +142 -142
  26. package/src/cli/commands/provider/qwenwork-login.js +119 -119
  27. package/src/cli/commands/stats.js +21 -9
  28. package/src/cli/commands/sync.js +232 -232
  29. package/src/cli/format.js +1 -0
  30. package/src/cli/help.js +1 -1
  31. package/src/cli/status.js +279 -279
  32. package/src/daemon.js +101 -101
  33. package/src/model-capabilities/enrich.js +86 -86
  34. package/src/model-capabilities/index.js +183 -183
  35. package/src/model-capabilities/parse.js +70 -70
  36. package/src/model-trace.js +4 -3
  37. package/src/models.js +225 -225
  38. package/src/providers/cline/auth.js +228 -228
  39. package/src/providers/cline/chat.js +307 -307
  40. package/src/providers/cline.js +2 -2
  41. package/src/providers/dispatcher.js +26 -14
  42. package/src/providers/globalqwenwork/cosy.js +123 -123
  43. package/src/providers/globalqwenwork/crypto.js +218 -218
  44. package/src/providers/globalqwenwork/http.js +20 -20
  45. package/src/providers/globalqwenwork/rsa.js +54 -54
  46. package/src/providers/globalqwenwork/sse.js +268 -268
  47. package/src/providers/globalqwenwork/stream.js +130 -130
  48. package/src/providers/globalqwenwork/upstream.js +122 -122
  49. package/src/providers/globalworkbuddy/account-store.js +198 -0
  50. package/src/providers/globalworkbuddy/auth.js +97 -0
  51. package/src/providers/globalworkbuddy/catalog.js +138 -0
  52. package/src/providers/globalworkbuddy/constants.js +73 -0
  53. package/src/providers/globalworkbuddy/device-auth.js +120 -0
  54. package/src/providers/globalworkbuddy/errors.js +98 -0
  55. package/src/providers/globalworkbuddy/headers.js +89 -0
  56. package/src/providers/globalworkbuddy/models.js +96 -0
  57. package/src/providers/globalworkbuddy/payload.js +80 -0
  58. package/src/providers/keyring.js +60 -60
  59. package/src/providers/qoder/aggregate.js +12 -3
  60. package/src/providers/qoder/chat.js +200 -183
  61. package/src/providers/qoder/index.js +249 -230
  62. package/src/providers/qoder/models.js +7 -3
  63. package/src/providers/qoder/payload.js +7 -1
  64. package/src/providers/qoder/request.js +2 -1
  65. package/src/providers/qoder/sse.js +145 -103
  66. package/src/providers/qoder/stream.js +16 -4
  67. package/src/providers/qwenwork/account-store.js +133 -133
  68. package/src/providers/qwenwork/constants.js +67 -67
  69. package/src/providers/qwenwork/cosy.js +120 -120
  70. package/src/providers/qwenwork/crypto.js +218 -218
  71. package/src/providers/qwenwork/http.js +20 -20
  72. package/src/providers/qwenwork/index.js +327 -327
  73. package/src/providers/qwenwork/payload.js +142 -142
  74. package/src/providers/qwenwork/rsa.js +54 -54
  75. package/src/providers/qwenwork/sse.js +268 -268
  76. package/src/providers/qwenwork/stream.js +130 -130
  77. package/src/providers/qwenwork/upstream.js +120 -120
  78. package/src/providers/qwenwork.js +1 -1
  79. package/src/providers/registry.js +74 -74
  80. package/src/providers/workbuddy/chat.js +248 -248
  81. package/src/providers/workbuddy/reshape.js +152 -152
  82. package/src/providers/workbuddy.js +2 -2
  83. package/src/reasoning.js +32 -32
  84. package/src/routes/chat/exhausted-handler.js +13 -2
  85. package/src/routes/chat/gateway.js +46 -46
  86. package/src/routes/chat/relay-pipeline.js +292 -276
  87. package/src/routes/chat/via-route-handler.js +146 -146
  88. package/src/routes/hedge.js +255 -255
  89. package/src/routes/models-route.js +167 -167
  90. package/src/routes/peers.js +273 -273
  91. package/src/routes/stream-scan.js +22 -5
  92. package/src/routes/stream.js +385 -386
  93. package/src/runtime/bootstrap.js +45 -45
  94. package/src/runtime/provider-gate.js +34 -34
  95. package/src/runtime/providers-setup.js +165 -165
  96. package/src/runtime/server-lifecycle.js +5 -6
  97. package/src/server.js +69 -64
  98. package/src/state/schemas/allowlist.js +92 -92
  99. package/src/sync-opencode.js +280 -280
  100. package/src/talk-full-redact.js +129 -0
  101. package/src/talk-full.js +130 -0
  102. package/src/talk-log.js +42 -15
  103. package/src/transport/index.js +244 -244
  104. package/src/transport/pool.js +56 -56
  105. package/src/transport/retry.js +24 -24
  106. package/src/transport/sse.js +93 -93
  107. package/src/upstream-engine/sdk/responses.js +5 -0
  108. package/src/upstream-probe/display.js +52 -52
  109. package/src/upstream-probe/probe.js +49 -49
  110. package/src/upstream-probe/rotate.js +110 -110
  111. package/src/upstream-probe/start.js +45 -45
  112. package/src/upstream.js +289 -289
  113. package/src/usage/report.js +14 -8
  114. package/docs/cli_help_mini.md +0 -135
  115. package/src/providers/AGENTS.md +0 -46
  116. package/src/routes/AGENTS.md +0 -37
@@ -0,0 +1,129 @@
1
+ // talk-full 的纯函数核(redact):写前处理链 + 会话标识归一,无 IO、无 env 读取(阈值由调用方传入),可独立单测。
2
+ // 处理链顺序冻结:结构摘要化 → 递归字段黑名单抹凭据 → JSON.stringify → 整串过 maskText 正则 → 单条超限降级。
3
+ import { createHash } from "node:crypto";
4
+ import { maskText } from "./talk-log.js";
5
+
6
+ const NO_CAP = Number.MAX_SAFE_INTEGER; // 只借 maskText 的正则,不套它的字节截断(截断会把 JSON 语法层切坏)
7
+ const REDACTED = "[已脱敏]";
8
+ // 凭据键名黑名单(键名归一化后比对):命中即整值抹掉,同层其它字段原样保留。
9
+ const DENY = new Set(["apikey", "accesstoken", "refreshtoken", "token", "secret", "password", "clientsecret", "cookie", "authorization", "xapikey"]);
10
+ // 字符串叶同款键值正则:覆盖 tool 结果 / tool_calls.arguments 这类「JSON 文本塞在字符串里」的凭据。
11
+ const LEAF_CRED = /"(apiKey|api_key|accessToken|refreshToken|token|secret|password|client_secret|cookie|authorization|x-api-key)"(\s*:\s*)"[^"]*"/gi;
12
+ const SKIP_PARAM = new Set(["messages", "tools", "functions", "model"]);
13
+
14
+ export const sha1 = (s) => createHash("sha1").update(String(s)).digest("hex");
15
+ const normKey = (k) => String(k).toLowerCase().replace(/[^a-z0-9]/g, "");
16
+ const clip = (v, n) => { const s = String(v ?? ""); return s.length > n ? `${s.slice(0, n)}…[已截断]` : s; };
17
+ // 循环引用 / BigInt 兜底:诊断材料宁可标「已循环」也不能抛出去影响请求。
18
+ export function stringifySafe(v) {
19
+ const seen = new WeakSet();
20
+ return JSON.stringify(v, (k, x) => {
21
+ if (typeof x === "bigint") return String(x);
22
+ if (x && typeof x === "object") { if (seen.has(x)) return "[已循环]"; seen.add(x); }
23
+ return x;
24
+ });
25
+ }
26
+
27
+ const imgTag = (p) => `[图片 ${String(p?.image_url?.url ?? p?.image_url ?? "").slice(0, 60)}]`;
28
+ /** 非文本部件只留形不留料(沿用 talk-log.js 的 60 字口径);全程造新对象,不改调用方的 body。 */
29
+ function summarizeParts(v, d = 0) {
30
+ if (d > 24) return v;
31
+ if (Array.isArray(v)) return v.map((x) => (x && typeof x === "object" && (x.image_url !== undefined || x.type === "image_url")) ? imgTag(x) : summarizeParts(x, d + 1));
32
+ if (v && typeof v === "object") { const o = {}; for (const [k, x] of Object.entries(v)) o[k] = summarizeParts(x, d + 1); return o; }
33
+ return v;
34
+ }
35
+ /** 递归字段黑名单:命中键名整值抹掉;字符串叶另走键值正则。 */
36
+ function redactDeep(v, d = 0) {
37
+ if (d > 24) return REDACTED;
38
+ if (Array.isArray(v)) return v.map((x) => redactDeep(x, d + 1));
39
+ if (v && typeof v === "object") { const o = {}; for (const [k, x] of Object.entries(v)) o[k] = DENY.has(normKey(k)) ? REDACTED : redactDeep(x, d + 1); return o; }
40
+ if (typeof v === "string") return v.replace(LEAF_CRED, '"$1"$2"[已脱敏]"');
41
+ return v;
42
+ }
43
+ /** 结构层安全的兜底:逐字符串叶过 maskText 正则(整串跑法偶尔会吃掉 JSON 闭引号)。 */
44
+ function maskLeaves(v) {
45
+ if (Array.isArray(v)) return v.map(maskLeaves);
46
+ if (v && typeof v === "object") { const o = {}; for (const [k, x] of Object.entries(v)) o[k] = maskLeaves(x); return o; }
47
+ return typeof v === "string" ? maskText(v, NO_CAP) : v;
48
+ }
49
+ /** 序列化 + 整串脱敏;结果不是合法 JSON 就退回逐叶脱敏 —— 落盘行必须可独立解析。 */
50
+ export function lineOf(rec) {
51
+ const whole = maskText(stringifySafe(rec), NO_CAP);
52
+ try { JSON.parse(whole); return whole; } catch { return stringifySafe(maskLeaves(rec)); }
53
+ }
54
+ /** 超长字符串叶(含超限标量 params,如 completions 的巨型 prompt):地板兜底,逐层截短。 */
55
+ function clipLongStrings(v, max = 200) {
56
+ if (Array.isArray(v)) return v.map((x) => clipLongStrings(x, max));
57
+ if (v && typeof v === "object") { const o = {}; for (const [k, x] of Object.entries(v)) o[k] = clipLongStrings(x, max); return o; }
58
+ return typeof v === "string" && v.length > max ? `${v.slice(0, max)}…[已截断]` : v;
59
+ }
60
+ /** 三级钳制:二级摘要自身也超限时(极端 message 条数)再削 —— 先丢 head 只留 role/chars/hash,再折半只留末尾 N 条。 */
61
+ function clampSummaries(rec, limitBytes, line) {
62
+ if (Buffer.byteLength(line, "utf8") <= limitBytes) return line;
63
+ const rows = (Array.isArray(rec.request.messages) ? rec.request.messages : []).map(({ head, ...rest }) => rest); // ① 逐条退化:只留 role + chars + hash(无 head)
64
+ rec.meta.summaryTrimmed = true; // 读侧:摘要被削过(② 末尾 N 条 + messagesOmitted 读数)
65
+ let n = rows.length;
66
+ for (let i = 0; i < 64; i++) { // 折半削减,上限 64 轮 = 有界(绝不死循环)
67
+ rec.request.messages = n > 0 ? rows.slice(-n) : [];
68
+ rec.meta.messagesOmitted = rows.length - rec.request.messages.length;
69
+ line = lineOf(rec);
70
+ if (Buffer.byteLength(line, "utf8") <= limitBytes) return line;
71
+ if (n === 0) break;
72
+ n = n === 1 ? 0 : Math.floor(n / 2);
73
+ }
74
+ return lineOf(Object.assign(rec, clipLongStrings(rec))); // ③ 地板兜底:messages 清空仍超限(超长标量 params)→ 长字符串叶截短
75
+ }
76
+ /** 超限降级的结构摘要:每条只留 role + 字符数 + sha1尾8 + 首500字。 */
77
+ function summarizeMessages(msgs) {
78
+ return (Array.isArray(msgs) ? msgs : []).map((m, i) => {
79
+ const t = typeof m?.content === "string" ? m.content : stringifySafe(m?.content ?? "");
80
+ return { i, role: String(m?.role || "?"), chars: t.length, hash: sha1(t).slice(-8), head: t.slice(0, 500) };
81
+ });
82
+ }
83
+ /** 单条超 limitBytes 时原地降级(记录不得丢),返回最终行。二级降级只给极端体量兜底。 */
84
+ export function composeLine(rec, limitBytes) {
85
+ let line = lineOf(rec);
86
+ if (Buffer.byteLength(line, "utf8") <= limitBytes) return line;
87
+ rec.request.messages = summarizeMessages(rec.request.messages);
88
+ rec.meta.truncated = true;
89
+ line = lineOf(rec);
90
+ if (Buffer.byteLength(line, "utf8") <= limitBytes) return line;
91
+ rec.request.tools = null;
92
+ rec.response = { reasoning: clip(rec.response.reasoning, 2000), content: clip(rec.response.content, 2000), toolCalls: [] };
93
+ return clampSummaries(rec, limitBytes, lineOf(rec)); // 二级之后仍超限 → 三级钳制,保证最终行 ≤ limitBytes
94
+ }
95
+
96
+ /** 采样参数:body 里除 messages/tools/functions/model 外的标量键逐个收(对象键不 dump,避免整身重复)。 */
97
+ export function scalarParams(body) {
98
+ const o = {};
99
+ for (const [k, v] of Object.entries(body && typeof body === "object" ? body : {})) {
100
+ if (SKIP_PARAM.has(k)) continue;
101
+ if (v === null || typeof v === "string" || typeof v === "number" || typeof v === "boolean") o[k] = v;
102
+ }
103
+ return o;
104
+ }
105
+ /** request/response 两侧的落盘形状(已结构摘要化 + 黑名单脱敏)。 */
106
+ export function shapeForCapture({ body, model, msgs, reasoning, content, toolCalls }) {
107
+ return redactDeep(summarizeParts({
108
+ request: { model: body?.model ?? model ?? null, messages: msgs, tools: body?.tools ?? body?.functions ?? null, params: scalarParams(body) },
109
+ response: { reasoning, content, toolCalls },
110
+ }));
111
+ }
112
+
113
+ // 会话标识(Session hint):客户端头 > 首条 system+user 派生 > 进程级兜底;落盘一律归一化。
114
+ export const PROCESS_SESSION = `ses_${sha1(`${process.pid}:${Math.random()}`)}`;
115
+ /** 形态同 upstream.js:17 的兜底公式,自成一份(不 import 未合并的会话身份变更)。 */
116
+ function sessionFromMessages(msgs) {
117
+ const pick = (role) => {
118
+ const m = (msgs || []).find((x) => x?.role === role);
119
+ if (!m) return "";
120
+ return typeof m.content === "string" ? m.content : stringifySafe(m.content ?? "");
121
+ };
122
+ const seed = `${pick("system")}|${pick("user")}`.slice(0, 4000);
123
+ return seed === "|" ? null : `ses_${sha1(seed)}`;
124
+ }
125
+ /** 归一化短标识:sha1尾12 + "-" + 原值前 8 字(同一原值→同一标识,跨轮稳定、换模型不漂)。 */
126
+ export function normSessionKey(raw, msgs) {
127
+ const v = String(raw ?? "").trim() || sessionFromMessages(msgs) || PROCESS_SESSION;
128
+ return `${sha1(v).slice(-12)}-${v.slice(0, 8)}`;
129
+ }
@@ -0,0 +1,130 @@
1
+ // agent 回路全量捕获(talk/full lane):每轮 agent↔模型的完整请求/响应以 JSONL 只读语料落盘,供人调试 agent 与 skill。
2
+ // 与 talk-log.js 分工:那边「人读环形稿、只留最后一问」,这里「机读全量、一条自足」—— 缺省开(opt-out),只有显式关闭词 MSLXDFF_TALK_FULL=0/off/false/no/disable 才停写。
3
+ // 纯旁路:不改响应字节/状态码/闸门与空轮判据,异常一律吞掉返回 false;写盘走模块内单 promise 链串行 appendFile(design D11)。
4
+ // 处理链(写前,顺序冻结)在 ./talk-full-redact.js;执行时机在写队列任务内、不在请求收场路径(P1-C)。
5
+ import { chmodSync, mkdirSync } from "node:fs";
6
+ import { appendFile, rename, rm, stat } from "node:fs/promises";
7
+ import { dirname, join } from "node:path";
8
+ import { logDir } from "./logs.js";
9
+ import { talkLogName } from "./talk-log.js";
10
+ import { fmtShanghaiYMDHMS } from "./time.js";
11
+ import { composeLine, normSessionKey, shapeForCapture, stringifySafe } from "./talk-full-redact.js";
12
+
13
+ const MB = 1024 * 1024;
14
+ function envInt(name, dflt) { const n = Number(process.env[name]); return Number.isFinite(n) && n > 0 ? Math.floor(n) : dflt; }
15
+ const maxLineBytes = () => envInt("MSLXDFF_TALK_FULL_MAX_LINE_MB", 16) * MB; // 单条超限降级阈值
16
+ const maxFileBytes = () => envInt("MSLXDFF_TALK_FULL_MAX_MB", 50) * MB; // 单文件轮转阈值(按字节不按时间)
17
+ const keepCount = () => { const n = Number(process.env.MSLXDFF_TALK_FULL_KEEP); return Number.isInteger(n) && n >= 0 ? n : 3; };
18
+
19
+ /** 缺省开(opt-out,ADR-0046):只有显式关闭词 0/off/false/no/disable(trim+lowercase)才关;未设、1、true、on、yes 与任意其它值一律开 —— 写错的值不得静默变关。 */
20
+ export function agentLoopEnabled() {
21
+ const v = String(process.env.MSLXDFF_TALK_FULL ?? "").trim().toLowerCase();
22
+ return !(v === "0" || v === "off" || v === "false" || v === "no" || v === "disable");
23
+ }
24
+
25
+ /** 响应正文桶上限的单一真相:MSLXDFF_TALK_CAP_CHARS 正数覆盖 > 缺省开 200 万 > 显式关 40 万(两档数字逐字不变,档位随 agentLoopEnabled 连动,避免两处真相)。 */
26
+ export function talkCapChars() {
27
+ const n = Number(process.env.MSLXDFF_TALK_CAP_CHARS);
28
+ if (Number.isFinite(n) && n > 0) return Math.floor(n);
29
+ return agentLoopEnabled() ? 2_000_000 : 400_000;
30
+ }
31
+
32
+ export function agentLoopDir() { return join(logDir(), "talk", "full"); }
33
+ /** 与 talk.log 命名同源:<dir>/<供应商>-<模型>-talkfull.jsonl(talkLogName 的 -talk.log 换成 -talkfull.jsonl)。 */
34
+ export function agentLoopFile(model) { return join(agentLoopDir(), `${talkLogName(model).replace(/(?:-talk)?\.log$/, "")}-talkfull.jsonl`); }
35
+
36
+ // ── 落盘:模块内单队列(写与轮转同队、不抢跑)+ 按字节 rename 轮转 ──────────
37
+ let queue = Promise.resolve();
38
+ const readyDirs = new Set();
39
+ const chmodded = new Set();
40
+ function enqueue(task) { queue = queue.then(task).catch(() => {}); return queue; }
41
+ async function rotate(file, keep) {
42
+ await rm(`${file}.${keep}`, { force: true }).catch(() => {});
43
+ for (let i = keep - 1; i >= 1; i--) await rename(`${file}.${i}`, `${file}.${i + 1}`).catch(() => {});
44
+ await rename(file, `${file}.1`).catch(() => {});
45
+ }
46
+ /** 单次写尝试:仅缓存未命中才 mkdir;appendFile 失败返回 false(不抛),由 queueWrite 决定是否重试一次。 */
47
+ async function writeOnce(file, line) {
48
+ const dir = dirname(file);
49
+ if (!readyDirs.has(dir)) { try { mkdirSync(dir, { recursive: true }); readyDirs.add(dir); } catch { return false; } } // 目录只在受理后出现
50
+ const st = await stat(file).catch(() => null);
51
+ if (st && st.size > maxFileBytes()) {
52
+ const keep = keepCount();
53
+ if (keep < 1) await rm(file, { force: true }).catch(() => {});
54
+ else await rotate(file, keep);
55
+ }
56
+ try { await appendFile(file, `${line}\n`, { mode: 0o600 }); } catch { return false; }
57
+ if (!chmodded.has(file)) { chmodded.add(file); try { chmodSync(file, 0o600); } catch {} } // mode 只在新建生效,存量补一次
58
+ return true;
59
+ }
60
+ // 读数口径单一真相(design D3):本次尝试墙钟优先;起点不可得时回退 relay 内部 totalMs(此时 elapsedMs 与 relayMs 同值,不伪造)。
61
+ function elapsedOf(attempt, total) {
62
+ const a = Number(attempt);
63
+ if (Number.isFinite(a)) return a;
64
+ const t = Number(total);
65
+ return Number.isFinite(t) ? t : null;
66
+ }
67
+
68
+ // 失败分支重试一次(正常路径零额外成本):readyDirs 缓存过期(用户手工 rm -r / 移走 talk/full)时
69
+ // 先清掉目录缓存重走 mkdir + chmod,仍失败才彻底放弃(继续吞,契约:捕获失败绝不打扰请求)。
70
+ // 重活全在已排队的异步任务里做(P1-C):整 body 序列化算 requestBytes + composeLine(递归黑名单串化 → 整串 6 条正则
71
+ // → 整记录 JSON.parse 回验 → 超限时最多 64 趟三级钳制重序列化)。留在同步段等于每条请求收场都付这份 CPU
72
+ // (4MB body 实测 ~21ms、16MB 级可到百毫秒事件循环停顿),与 design「纯旁路不占请求路径」相抵。
73
+ // 落盘逐字不变的前提:① rec 里派生自 body 的结构(request/response)已由 shapeForCapture 在同步段深拷定稿,
74
+ // 队列任务只读 body 算长度、不再读它的结构(recordAgentLoop 之后无任何生产调用方改写 body:空转抬额走的
75
+ // withRaisedMaxTokens 也是产新对象);② 键序按字面量插入序冻结 → requestBytes 先占位再回填,事后补键会被
76
+ // JSON.stringify 排到 meta 末尾;③ 队列内异常照旧由 enqueue 的 .catch(() => {}) 吞掉(受理即返回 true,与写盘失败同语义);
77
+ // ④ limitBytes 与 agentLoopFile 仍在受理时刻算:env 阈值若挪到队列里读,测试/关停路径的 env 还原会把降级档位读错。
78
+ function queueWrite(file, rec, body, limitBytes) {
79
+ return enqueue(async () => {
80
+ rec.meta.requestBytes = Buffer.byteLength(stringifySafe(body ?? {}), "utf8");
81
+ const line = composeLine(rec, limitBytes);
82
+ if (await writeOnce(file, line)) return;
83
+ readyDirs.delete(dirname(file)); chmodded.delete(file);
84
+ await writeOnce(file, line);
85
+ });
86
+ }
87
+ /** 排空写队列:测试断言前与进程退出前调用。 */
88
+ export async function flushAgentLoop() {
89
+ for (let i = 0; i < 1000; i++) { const q = queue; await q; if (q === queue) return; }
90
+ }
91
+
92
+ /**
93
+ * 一条上游尝试 = 一行自包含 JSONL。判据顺序即契约:
94
+ * ① 开关 ② hops>0(组员转发的对话不落盘,spec「只捕获本机自发流量」)③ 两侧全空且 status≠200 ④ 受理。
95
+ * 任何异常吞掉返回 false,MUST NOT 抛出影响请求。
96
+ */
97
+ export function recordAgentLoop({ reqId, model, via, hops, body, out, sessionKey, clientIp, echo, attemptMs, relayMs } = {}) {
98
+ if (!agentLoopEnabled()) return false;
99
+ try {
100
+ if (Number(hops) > 0) return false;
101
+ const talk = out?.detail?.talk || {};
102
+ const msgs = Array.isArray(body?.messages) ? body.messages : [];
103
+ const reasoning = (talk.reasoning || []).join("");
104
+ const content = (talk.content || []).join("");
105
+ const toolCalls = Array.isArray(talk.tools) ? talk.tools : [];
106
+ const status = out?.status == null || !Number.isFinite(Number(out.status)) ? null : Number(out.status);
107
+ if (!msgs.length && !reasoning && !content && !toolCalls.length && status !== 200) return false;
108
+ const at = Date.now();
109
+ const rec = {
110
+ ts: at, time: fmtShanghaiYMDHMS(new Date(at)),
111
+ reqId: reqId ?? null, sessionKey: normSessionKey(sessionKey, msgs), clientIp: clientIp ?? null,
112
+ model: String(model || "unknown"), via: via ?? null,
113
+ hops: hops == null || hops === "" || !Number.isFinite(Number(hops)) ? null : Number(hops),
114
+ status, stream: body?.stream ? 1 : 0, elapsedMs: elapsedOf(attemptMs, out?.totalMs),
115
+ usage: out?.detail?.usage ?? null, finishReason: out?.detail?.sawFinishReason ?? null,
116
+ upstream: echo?.upstream ?? null, account: echo?.account ?? null, pick: echo?.pick ?? null,
117
+ ...shapeForCapture({ body, model, msgs, reasoning, content, toolCalls }),
118
+ meta: {
119
+ requestBytes: 0, // 占位只冻键序(P1-C):真值在 queueWrite 的队列任务里按整 body 序列化回填
120
+ responseChars: Number.isFinite(Number(talk.n)) ? Number(talk.n) : content.length, // 桶的 n 显式写进读数
121
+ reasoningChars: reasoning.length, // 思考与正文分列,不互相吞并
122
+ truncated: Boolean(talk.capped), // 撞顶必须看得见
123
+ cap: talkCapChars(),
124
+ relayMs: elapsedOf(relayMs, out?.totalMs), // relay 内部计时恒写:与 elapsedMs(尝试墙钟)分列,「上游多久 vs 写给客户端多久」可各自判读
125
+ },
126
+ };
127
+ queueWrite(agentLoopFile(rec.model), rec, body, maxLineBytes()); // 阈值与目标文件在受理时刻定档,序列化/降级/回验全在队列内
128
+ return true;
129
+ } catch { return false; }
130
+ }
package/src/talk-log.js CHANGED
@@ -43,12 +43,14 @@ const MASKS = [
43
43
  [/\b(token|accessToken|refreshToken|api[_-]?key|apiKey|secret|password|cookie|client_secret|authorization)\s*[:=]\s*[\w./+-]{12,}/gi, "$1=[已脱敏]"],
44
44
  ];
45
45
 
46
- /** 脱敏 + 按字节封顶截断(中文 3 字节/字,逐字累加才不会超预算)。 */
47
- export function maskText(v, cap = 100000) {
46
+ const CUT_MARK = "\n…[已截断]";
47
+
48
+ /** 脱敏 + 按字节封顶截断(中文 3 字节/字,逐字累加才不会超预算),并报告这一刀是否真的落下。 */
49
+ function maskCut(v, cap = 100000) {
48
50
  let s = String(v ?? "");
49
51
  for (const [re, repl] of MASKS) s = s.replace(re, repl);
50
- if (Buffer.byteLength(s, "utf8") <= cap) return s;
51
- const tail = Buffer.from("\n…[已截断]", "utf8");
52
+ if (Buffer.byteLength(s, "utf8") <= cap) return { text: s, capped: false };
53
+ const tail = Buffer.from(CUT_MARK, "utf8");
52
54
  const budget = Math.max(0, cap - tail.length);
53
55
  let bytes = 0, i = 0;
54
56
  for (const ch of s) {
@@ -57,9 +59,12 @@ export function maskText(v, cap = 100000) {
57
59
  bytes += n;
58
60
  i += ch.length;
59
61
  }
60
- return s.slice(0, i) + "\n…[已截断]";
62
+ return { text: s.slice(0, i) + CUT_MARK, capped: true };
61
63
  }
62
64
 
65
+ /** 脱敏 + 按字节封顶截断(对外口径与改前逐字一致;要知道是否砍过用 maskCut/pushBlock)。 */
66
+ export function maskText(v, cap = 100000) { return maskCut(v, cap).text; }
67
+
63
68
  function textOf(content) {
64
69
  if (typeof content === "string") return content;
65
70
  if (Array.isArray(content)) return content.map((p) => p?.text || (p?.image_url ? `[图片 ${String(p.image_url.url).slice(0, 60)}]` : "")).filter(Boolean).join("\n");
@@ -78,26 +83,44 @@ export function lastUserQuestion(body) {
78
83
 
79
84
  const CUT_IN = "----8<----", CUT_OUT = "---->8----";
80
85
 
86
+ // 标签行的「N 字」在字节封顶之后只是**已落盘**的字数;不带封顶标记,就会被读成「模型只说了这么多」。
81
87
  function pushBlock(out, label, text, cap) {
82
- const m = maskText(text, cap);
83
- if (!m.trim()) return;
84
- out.push(`[${label} · ${m.length} 字]`, CUT_IN, m, CUT_OUT, "");
88
+ const { text: m, capped } = maskCut(text, cap);
89
+ if (!m.trim()) return false;
90
+ out.push(`[${label} · ${m.length} 字${capped ? " · 已封顶" : ""}]`, CUT_IN, m, CUT_OUT, "");
91
+ return capped;
85
92
  }
86
93
 
87
- /** 一条对话 = 一个可读块:头部一行元信息(ts= 是环形淘汰的唯一依据),正文三段。 */
88
- export function formatTalkEntry({ reqId, model, via, hops, status, stream, elapsedMs, question, thinking, answer, tools, usage, upstream, account, pick, ts } = {}) {
94
+ // 撞捕获上限时紧跟头部行的人类可读提醒:不必先认识 truncated=1 这个键名也能一眼看出「可能不完整」。
95
+ const TRUNC_NOTE = "[!] 本轮已撞捕获上限(truncated=1):思考/正文后半段未落盘,内容可能不完整 —— 别当成模型的完整回答";
96
+
97
+ /** 一条对话 = 一个可读块:头部一行元信息(ts= 是环形淘汰的唯一依据),正文三段。elapsed= 取本次尝试墙钟(design D3),finish=/relayMs= 为 ADR-0046 新读数。 */
98
+ export function formatTalkEntry({ reqId, model, via, hops, status, stream, elapsedMs, question, thinking, answer, tools, usage, upstream, account, pick, ts, talkCapped = false, talkChars, talkCap, finishReason, relayMs } = {}) {
99
+ const fval = finishReason || "-"; // 缺值占位(grill Q5):「没给 finish」不得被读成 stop;字段恒在场
89
100
  const now = Number.isFinite(ts) && ts > 0 ? ts : Date.now();
101
+ const trunc = talkCapped === true;
90
102
  const head = [
91
103
  `${MARK} ts=${now}`, `time=${fmtShanghaiYMDHMS(new Date(now))}`, `req=${reqId || "-"}`, `model=${model || "-"}`,
92
104
  via ? `via=${via}` : "", hops != null ? `hops=${hops}` : "", `status=${status ?? "-"}`,
93
105
  stream ? "stream=1" : "stream=0", Number.isFinite(elapsedMs) ? `elapsed=${Math.round(elapsedMs)}ms` : "",
106
+ // 尝试墙钟与 relay 内部计时同值时不重复打印(头行噪音克制,grill Q3);两值不同(hedge/缓冲重放)才另列 relayMs
107
+ Number.isFinite(relayMs) && Number.isFinite(elapsedMs) && Math.round(relayMs) !== Math.round(elapsedMs) ? `relayMs=${Math.round(relayMs)}ms` : "",
108
+ `finish=${fval}`,
94
109
  usage ? `usage(prompt=${usage.prompt_tokens ?? "-"},completion=${usage.completion_tokens ?? "-"})` : "",
95
110
  upstream ? `upstream=${upstream}` : "", account ? `account=${account}` : "", pick ? `pick=${pick}` : "",
111
+ // 截断读数只在真截断时出现(未截断不得新增噪音字段):cap= 上限、chars= 已捕获字符数
112
+ trunc ? `truncated=1 cap=${Number.isFinite(talkCap) && talkCap > 0 ? talkCap : "-"} chars=${Number.isFinite(talkChars) ? talkChars : "-"}` : "",
96
113
  ].filter(Boolean).join(" ");
97
- const out = [head, ""];
114
+ const out = [head];
115
+ if (trunc) out.push(TRUNC_NOTE);
116
+ out.push("");
98
117
  pushBlock(out, "我问", question, 20000);
99
118
  pushBlock(out, "思考", thinking, 100000);
100
- pushBlock(out, "回答", answer, 200000);
119
+ // 零正文不再静默跳块(spec「零输出轮必须可判读」/grill Q4):缺块与「正文该有却没落盘」同形,缺席必须是可读读数。
120
+ // 只落一行标注 + 空行(没有内容可围,不套 CUT 围栏);有真实正文的字符串照旧走 pushBlock,该段与改动前逐字节一致。
121
+ // 判据看 trim 不看长度(P1-B):" "/"\n" 进 pushBlock 也会被 !m.trim() 整块跳过 → 与「正文丢盘」再度同形,而「只调工具 + 吐个换行」是真实 agent 形态。
122
+ if (!String(answer ?? "").trim()) out.push(`[回答 · 0 字 · 本轮无正文(finish=${fval})]`, "");
123
+ else pushBlock(out, "回答", answer, 200000);
101
124
  const t = tools == null || tools === "" ? "" : String(tools);
102
125
  pushBlock(out, "工具调用", t, 20000);
103
126
  out.push("[END]", "");
@@ -177,8 +200,10 @@ export function trimTalkLog(file, nowTs = Date.now()) {
177
200
  /**
178
201
  * relay 落盘的唯一入口:把 relay() 累积的 detail.talk(思考/正文/工具分片)拼成一条对话。
179
202
  * 空轮(零正文零思考零工具)不占环形窗口 —— 那是 relay-pipeline 已经另行报错的场景。
203
+ * 桶的 capped/n/cap 必须一路传到 formatTalkEntry:撞顶的那条要能看出「后半截没了」。
204
+ * attemptMs(本次上游尝试墙钟)由调用点传入;未传(拿不到尝试起点)回退 out.totalMs —— 此时 elapsed 与 relayMs 同值(spec 降级 scenario)。
180
205
  */
181
- export function recordRelayTalk({ reqId, model, via, hops, body, out, echo = {} } = {}) {
206
+ export function recordRelayTalk({ reqId, model, via, hops, body, out, echo = {}, attemptMs, finishReason } = {}) {
182
207
  if (!talkLogEnabled()) return false;
183
208
  try {
184
209
  const talk = out?.detail?.talk;
@@ -190,9 +215,11 @@ export function recordRelayTalk({ reqId, model, via, hops, body, out, echo = {}
190
215
  if (!answer && !thinking && !tools) return false;
191
216
  const ts = Date.now();
192
217
  const block = formatTalkEntry({
193
- reqId, model, via, hops, status, stream: Boolean(body?.stream),
194
- elapsedMs: out?.totalMs, question: lastUserQuestion(body || {}),
218
+ reqId, model, via, hops, status, stream: Boolean(body?.stream), finishReason: finishReason ?? out?.detail?.sawFinishReason,
219
+ elapsedMs: Number.isFinite(Number(attemptMs)) ? Number(attemptMs) : out?.totalMs, // 新口径:本次尝试墙钟优先(design D3)
220
+ relayMs: out?.totalMs, question: lastUserQuestion(body || {}),
195
221
  thinking, answer, tools, usage: out?.detail?.usage, ...echo, ts,
222
+ talkCapped: talk?.capped === true, talkChars: talk?.n, talkCap: talk?.cap,
196
223
  });
197
224
  appendTalkEntry({ file: talkLogFile(model), block, atMs: ts });
198
225
  return true;