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
@@ -1,152 +1,152 @@
1
- // workbuddy 流式 reasoning_content 整形:上游思考模型(如 glm-5.3-flash)把思考过程一词一 chunk
2
- // 地放进 delta.reasoning_content,客户端会把每个增量渲染成独立 "- Thought" 条目。
3
- // 旧版:连续 reasoning 增量缓冲到首个 content/finish 才一次性 flush → 思考期 10~30s 界面完全无变化(用户看到 workbuddy 后台流量走但 opencode 界面卡住)
4
- // 新版:增量聚合 + 阈值分片:缓冲 reasoning,达到 REASONING_CHUNK_SIZE(默认 100 字符)即分片 flush,
5
- // 期间仅首个 role 帧透传保持 TTFB,其余纯 reasoning 空帧吞掉(之前会透传 800+ 空 role 帧);
6
- // 尾段不足阈值的在 content/finish/流结束时兜底 flush。恒空 reasoning(deepseek-v4-flash)零改写。
7
- // 仅包 workbuddy 出口的 SSE 响应,不影响其他供应商与全局聚合器。
8
-
9
- const REASONING_CHUNK_SIZE = 150;
10
-
11
- export function reshapeWorkbuddySse(res) {
12
- try {
13
- const ct = res.headers?.get?.("content-type") || "";
14
- if (res.status !== 200 || !ct.includes("text/event-stream") || !res.body) return res;
15
- } catch { return res; }
16
-
17
- const reader = res.body.getReader();
18
- const decoder = new TextDecoder();
19
- const encoder = new TextEncoder();
20
- let buf = "";
21
- let reasoningBuf = "";
22
- let lastMeta = null;
23
- let closed = false;
24
- let hasSentRole = false;
25
- let lastFlushAt = Date.now();
26
-
27
- function flushReasoningInto(out) {
28
- if (!reasoningBuf) return;
29
- const meta = lastMeta || {};
30
- out.push(`data: ${JSON.stringify({
31
- id: meta.id || "chatcmpl-workbuddy-reshape",
32
- object: meta.object || "chat.completion.chunk",
33
- created: meta.created || Math.floor(Date.now() / 1000),
34
- model: meta.model || "",
35
- choices: [{ index: 0, delta: { reasoning_content: reasoningBuf }, finish_reason: null }],
36
- })}\n\n`);
37
- reasoningBuf = "";
38
- lastFlushAt = Date.now();
39
- }
40
-
41
- let blankStreak = false;
42
- // 处理一段完整行文本(以 \n 结尾),产出的行/帧推进 out(连续空行合并为单个事件分隔)
43
- function processInto(text, out) {
44
- for (const line of text.split("\n")) {
45
- if (!line) { if (!blankStreak) out.push("\n"); blankStreak = true; continue; }
46
- blankStreak = false;
47
- if (!line.startsWith("data:")) { out.push(line + "\n"); continue; }
48
- const payload = line.slice(5).trim();
49
- if (!payload || payload === "[DONE]") { out.push(line + "\n"); continue; }
50
- let obj;
51
- try { obj = JSON.parse(payload); } catch { out.push(line + "\n"); continue; }
52
- const ch = obj.choices?.[0];
53
- if (ch && typeof ch === "object") {
54
- if (obj.id) lastMeta = { id: obj.id, model: obj.model, object: obj.object, created: obj.created };
55
- const d = ch.delta && typeof ch.delta === "object" ? ch.delta : null;
56
- const rc = d ? d.reasoning_content : undefined;
57
- if (typeof rc === "string" && rc) {
58
- reasoningBuf += rc;
59
- d.reasoning_content = "";
60
- const isBoundary = (typeof d.content === "string" && d.content) || !!ch.finish_reason;
61
- // 首个 reasoning 帧的 role 透传一次,保 TTFB;后续纯 reasoning 的 role 重复帧直接吞掉
62
- if (!hasSentRole && d.role !== undefined) {
63
- hasSentRole = true;
64
- const rm = lastMeta || obj;
65
- out.push(`data: ${JSON.stringify({
66
- id: rm.id || obj.id || "chatcmpl-workbuddy-reshape",
67
- object: rm.object || obj.object || "chat.completion.chunk",
68
- created: rm.created || obj.created || Math.floor(Date.now() / 1000),
69
- model: rm.model || obj.model || "",
70
- choices: [{ index: 0, delta: { role: d.role }, finish_reason: null }],
71
- })}\n\n`);
72
- } else if (d.role !== undefined) {
73
- // 后续重复 role 帧不再透传,避免 800+ 空帧洪泛
74
- // 标记已发送过,避免后续非 reasoning 帧再误判为首次
75
- hasSentRole = true;
76
- }
77
- // 阈值分片:攒够 100 字符即向前吐一块,期间界面可见增量而非全程静默
78
- // 时间阈值:若距上次 flush 已超 700ms 且已攒 20+ 字符,也吐一块(避免小尾巴长时间不更新)
79
- const now = Date.now();
80
- const shouldChunk = reasoningBuf.length >= REASONING_CHUNK_SIZE
81
- || (reasoningBuf.length >= 50 && now - lastFlushAt > 1500);
82
- if (shouldChunk) flushReasoningInto(out);
83
- if (isBoundary) {
84
- // 兜底:content/finish 前把剩余思考一次性吐出,再透传本帧
85
- flushReasoningInto(out);
86
- const hasPassthrough = isBoundary || (Array.isArray(d.tool_calls) && d.tool_calls.length);
87
- if (hasPassthrough) out.push(`data: ${JSON.stringify(obj)}\n\n`);
88
- continue;
89
- }
90
- // 纯 reasoning 碎片帧已缓冲/或已分片吐出,原空帧吞掉
91
- continue;
92
- }
93
- // 非 reasoning 帧:若有 content/finish 先兜底 flush
94
- if ((typeof d?.content === "string" && d.content) || ch.finish_reason) flushReasoningInto(out);
95
- if (!hasSentRole && d?.role !== undefined) {
96
- hasSentRole = true;
97
- } else if (hasSentRole && d?.role !== undefined) {
98
- // 去重:首个 role 后续重复 role 去掉,避免下游每帧都带 role
99
- try { const clone = JSON.parse(payload); delete clone.choices[0].delta.role; out.push(`data: ${JSON.stringify(clone)}\n\n`); continue; } catch { /* fallback透传 */ }
100
- }
101
- }
102
- out.push(line + "\n");
103
- }
104
- }
105
-
106
- const body = new ReadableStream({
107
- async pull(controller) {
108
- if (closed) { try { controller.close(); } catch {} return; }
109
- try {
110
- // 循环读源,直到"有输出 / 流结束 / 出错"才返回。
111
- // WHATWG 流规范:pull 的 promise 解析后若未 enqueue,规范不会再自动调度 pull
112
- // (仅当解析期间有新读请求触发 pullAgain 才重调)——消费者只挂一个待处理 read 时,
113
- // 任何一次"只吞帧不出数据"的拉取都会让整流永久停摆。workbuddy 思考流按帧到达
114
- // (role 帧后每个 reasoning 碎片都不足 150 字符阈值)时必现,即"首帧后卡死"根因。
115
- for (;;) {
116
- const { done, value } = await reader.read();
117
- if (done) {
118
- closed = true;
119
- const out = [];
120
- if (buf) { processInto(buf + "\n", out); buf = ""; }
121
- flushReasoningInto(out);
122
- if (out.length) controller.enqueue(encoder.encode(out.join("")));
123
- controller.close();
124
- return;
125
- }
126
- buf += decoder.decode(value, { stream: true });
127
- const idx = buf.lastIndexOf("\n");
128
- if (idx < 0) continue;
129
- const complete = buf.slice(0, idx + 1);
130
- buf = buf.slice(idx + 1);
131
- const out = [];
132
- processInto(complete, out);
133
- if (out.length) { controller.enqueue(encoder.encode(out.join(""))); return; }
134
- }
135
- } catch {
136
- closed = true;
137
- try { controller.close(); } catch {}
138
- }
139
- },
140
- cancel() {
141
- closed = true;
142
- try {
143
- const p = reader.cancel();
144
- if (p && typeof p.catch === "function") p.catch(() => {});
145
- } catch {}
146
- },
147
- });
148
-
149
- const out = new Response(body, { status: res.status, statusText: res.statusText, headers: res.headers });
150
- try { out._t = res._t; } catch {}
151
- return out;
152
- }
1
+ // workbuddy 流式 reasoning_content 整形:上游思考模型(如 glm-5.3-flash)把思考过程一词一 chunk
2
+ // 地放进 delta.reasoning_content,客户端会把每个增量渲染成独立 "- Thought" 条目。
3
+ // 旧版:连续 reasoning 增量缓冲到首个 content/finish 才一次性 flush → 思考期 10~30s 界面完全无变化(用户看到 workbuddy 后台流量走但 opencode 界面卡住)
4
+ // 新版:增量聚合 + 阈值分片:缓冲 reasoning,达到 REASONING_CHUNK_SIZE(默认 100 字符)即分片 flush,
5
+ // 期间仅首个 role 帧透传保持 TTFB,其余纯 reasoning 空帧吞掉(之前会透传 800+ 空 role 帧);
6
+ // 尾段不足阈值的在 content/finish/流结束时兜底 flush。恒空 reasoning(deepseek-v4-flash)零改写。
7
+ // 仅包 workbuddy 出口的 SSE 响应,不影响其他供应商与全局聚合器。
8
+
9
+ const REASONING_CHUNK_SIZE = 150;
10
+
11
+ export function reshapeWorkbuddySse(res) {
12
+ try {
13
+ const ct = res.headers?.get?.("content-type") || "";
14
+ if (res.status !== 200 || !ct.includes("text/event-stream") || !res.body) return res;
15
+ } catch { return res; }
16
+
17
+ const reader = res.body.getReader();
18
+ const decoder = new TextDecoder();
19
+ const encoder = new TextEncoder();
20
+ let buf = "";
21
+ let reasoningBuf = "";
22
+ let lastMeta = null;
23
+ let closed = false;
24
+ let hasSentRole = false;
25
+ let lastFlushAt = Date.now();
26
+
27
+ function flushReasoningInto(out) {
28
+ if (!reasoningBuf) return;
29
+ const meta = lastMeta || {};
30
+ out.push(`data: ${JSON.stringify({
31
+ id: meta.id || "chatcmpl-workbuddy-reshape",
32
+ object: meta.object || "chat.completion.chunk",
33
+ created: meta.created || Math.floor(Date.now() / 1000),
34
+ model: meta.model || "",
35
+ choices: [{ index: 0, delta: { reasoning_content: reasoningBuf }, finish_reason: null }],
36
+ })}\n\n`);
37
+ reasoningBuf = "";
38
+ lastFlushAt = Date.now();
39
+ }
40
+
41
+ let blankStreak = false;
42
+ // 处理一段完整行文本(以 \n 结尾),产出的行/帧推进 out(连续空行合并为单个事件分隔)
43
+ function processInto(text, out) {
44
+ for (const line of text.split("\n")) {
45
+ if (!line) { if (!blankStreak) out.push("\n"); blankStreak = true; continue; }
46
+ blankStreak = false;
47
+ if (!line.startsWith("data:")) { out.push(line + "\n"); continue; }
48
+ const payload = line.slice(5).trim();
49
+ if (!payload || payload === "[DONE]") { out.push(line + "\n"); continue; }
50
+ let obj;
51
+ try { obj = JSON.parse(payload); } catch { out.push(line + "\n"); continue; }
52
+ const ch = obj.choices?.[0];
53
+ if (ch && typeof ch === "object") {
54
+ if (obj.id) lastMeta = { id: obj.id, model: obj.model, object: obj.object, created: obj.created };
55
+ const d = ch.delta && typeof ch.delta === "object" ? ch.delta : null;
56
+ const rc = d ? d.reasoning_content : undefined;
57
+ if (typeof rc === "string" && rc) {
58
+ reasoningBuf += rc;
59
+ d.reasoning_content = "";
60
+ const isBoundary = (typeof d.content === "string" && d.content) || !!ch.finish_reason;
61
+ // 首个 reasoning 帧的 role 透传一次,保 TTFB;后续纯 reasoning 的 role 重复帧直接吞掉
62
+ if (!hasSentRole && d.role !== undefined) {
63
+ hasSentRole = true;
64
+ const rm = lastMeta || obj;
65
+ out.push(`data: ${JSON.stringify({
66
+ id: rm.id || obj.id || "chatcmpl-workbuddy-reshape",
67
+ object: rm.object || obj.object || "chat.completion.chunk",
68
+ created: rm.created || obj.created || Math.floor(Date.now() / 1000),
69
+ model: rm.model || obj.model || "",
70
+ choices: [{ index: 0, delta: { role: d.role }, finish_reason: null }],
71
+ })}\n\n`);
72
+ } else if (d.role !== undefined) {
73
+ // 后续重复 role 帧不再透传,避免 800+ 空帧洪泛
74
+ // 标记已发送过,避免后续非 reasoning 帧再误判为首次
75
+ hasSentRole = true;
76
+ }
77
+ // 阈值分片:攒够 100 字符即向前吐一块,期间界面可见增量而非全程静默
78
+ // 时间阈值:若距上次 flush 已超 700ms 且已攒 20+ 字符,也吐一块(避免小尾巴长时间不更新)
79
+ const now = Date.now();
80
+ const shouldChunk = reasoningBuf.length >= REASONING_CHUNK_SIZE
81
+ || (reasoningBuf.length >= 50 && now - lastFlushAt > 1500);
82
+ if (shouldChunk) flushReasoningInto(out);
83
+ if (isBoundary) {
84
+ // 兜底:content/finish 前把剩余思考一次性吐出,再透传本帧
85
+ flushReasoningInto(out);
86
+ const hasPassthrough = isBoundary || (Array.isArray(d.tool_calls) && d.tool_calls.length);
87
+ if (hasPassthrough) out.push(`data: ${JSON.stringify(obj)}\n\n`);
88
+ continue;
89
+ }
90
+ // 纯 reasoning 碎片帧已缓冲/或已分片吐出,原空帧吞掉
91
+ continue;
92
+ }
93
+ // 非 reasoning 帧:若有 content/finish 先兜底 flush
94
+ if ((typeof d?.content === "string" && d.content) || ch.finish_reason) flushReasoningInto(out);
95
+ if (!hasSentRole && d?.role !== undefined) {
96
+ hasSentRole = true;
97
+ } else if (hasSentRole && d?.role !== undefined) {
98
+ // 去重:首个 role 后续重复 role 去掉,避免下游每帧都带 role
99
+ try { const clone = JSON.parse(payload); delete clone.choices[0].delta.role; out.push(`data: ${JSON.stringify(clone)}\n\n`); continue; } catch { /* fallback透传 */ }
100
+ }
101
+ }
102
+ out.push(line + "\n");
103
+ }
104
+ }
105
+
106
+ const body = new ReadableStream({
107
+ async pull(controller) {
108
+ if (closed) { try { controller.close(); } catch {} return; }
109
+ try {
110
+ // 循环读源,直到"有输出 / 流结束 / 出错"才返回。
111
+ // WHATWG 流规范:pull 的 promise 解析后若未 enqueue,规范不会再自动调度 pull
112
+ // (仅当解析期间有新读请求触发 pullAgain 才重调)——消费者只挂一个待处理 read 时,
113
+ // 任何一次"只吞帧不出数据"的拉取都会让整流永久停摆。workbuddy 思考流按帧到达
114
+ // (role 帧后每个 reasoning 碎片都不足 150 字符阈值)时必现,即"首帧后卡死"根因。
115
+ for (;;) {
116
+ const { done, value } = await reader.read();
117
+ if (done) {
118
+ closed = true;
119
+ const out = [];
120
+ if (buf) { processInto(buf + "\n", out); buf = ""; }
121
+ flushReasoningInto(out);
122
+ if (out.length) controller.enqueue(encoder.encode(out.join("")));
123
+ controller.close();
124
+ return;
125
+ }
126
+ buf += decoder.decode(value, { stream: true });
127
+ const idx = buf.lastIndexOf("\n");
128
+ if (idx < 0) continue;
129
+ const complete = buf.slice(0, idx + 1);
130
+ buf = buf.slice(idx + 1);
131
+ const out = [];
132
+ processInto(complete, out);
133
+ if (out.length) { controller.enqueue(encoder.encode(out.join(""))); return; }
134
+ }
135
+ } catch {
136
+ closed = true;
137
+ try { controller.close(); } catch {}
138
+ }
139
+ },
140
+ cancel() {
141
+ closed = true;
142
+ try {
143
+ const p = reader.cancel();
144
+ if (p && typeof p.catch === "function") p.catch(() => {});
145
+ } catch {}
146
+ },
147
+ });
148
+
149
+ const out = new Response(body, { status: res.status, statusText: res.statusText, headers: res.headers });
150
+ try { out._t = res._t; } catch {}
151
+ return out;
152
+ }
@@ -1,2 +1,2 @@
1
- // Shim — 保持历史导入路径兼容,真实实现下沉至 ./workbuddy/ 深模块
2
- export * from "./workbuddy/index.js";
1
+ // Shim — 保持历史导入路径兼容,真实实现下沉至 ./workbuddy/ 深模块
2
+ export * from "./workbuddy/index.js";
package/src/reasoning.js CHANGED
@@ -1,33 +1,33 @@
1
- const PLACEHOLDER = " ";
2
-
3
- const MODEL_RULES = [
4
- { match: (m) => /^kimi-/i.test(m || ""), scope: "toolCalls" },
5
- { match: (m) => /deepseek/i.test(m || ""), scope: "all" },
6
- ];
7
-
8
- export function normalizeModel(model) {
9
- return model.startsWith("oc/") ? model.slice(3) : model;
10
- }
11
-
12
- function shouldInject(message, scope) {
13
- if (message?.role !== "assistant") return false;
14
- const rc = message.reasoning_content;
15
- if (typeof rc === "string" && rc.length > 0) return false;
16
- if (scope === "toolCalls") {
17
- return Array.isArray(message.tool_calls) && message.tool_calls.length > 0;
18
- }
19
- return true;
20
- }
21
-
22
- function applyRule(body, rule) {
23
- if (!rule || !body?.messages) return body;
24
- const messages = body.messages.map((m) =>
25
- shouldInject(m, rule.scope) ? { ...m, reasoning_content: PLACEHOLDER } : m
26
- );
27
- return { ...body, messages };
28
- }
29
-
30
- export function injectReasoningContent(model, body) {
31
- const rule = MODEL_RULES.find((r) => r.match(model));
32
- return applyRule(body, rule);
1
+ const PLACEHOLDER = " ";
2
+
3
+ const MODEL_RULES = [
4
+ { match: (m) => /^kimi-/i.test(m || ""), scope: "toolCalls" },
5
+ { match: (m) => /deepseek/i.test(m || ""), scope: "all" },
6
+ ];
7
+
8
+ export function normalizeModel(model) {
9
+ return model.startsWith("oc/") ? model.slice(3) : model;
10
+ }
11
+
12
+ function shouldInject(message, scope) {
13
+ if (message?.role !== "assistant") return false;
14
+ const rc = message.reasoning_content;
15
+ if (typeof rc === "string" && rc.length > 0) return false;
16
+ if (scope === "toolCalls") {
17
+ return Array.isArray(message.tool_calls) && message.tool_calls.length > 0;
18
+ }
19
+ return true;
20
+ }
21
+
22
+ function applyRule(body, rule) {
23
+ if (!rule || !body?.messages) return body;
24
+ const messages = body.messages.map((m) =>
25
+ shouldInject(m, rule.scope) ? { ...m, reasoning_content: PLACEHOLDER } : m
26
+ );
27
+ return { ...body, messages };
28
+ }
29
+
30
+ export function injectReasoningContent(model, body) {
31
+ const rule = MODEL_RULES.find((r) => r.match(model));
32
+ return applyRule(body, rule);
33
33
  }
@@ -4,6 +4,7 @@ import { endHeldFailure } from "../stream-hold.js"; // 终局收场出口(502
4
4
  import { json } from "../helpers.js";
5
5
  import { performance } from "node:perf_hooks";
6
6
  import { recordRelayTalk } from "../../talk-log.js"; // 环形对话日志:exhausted 收尾直连 relay,落盘点必须自己补一份
7
+ import { recordAgentLoop } from "../../talk-full.js"; // agent 回路全量语料(talk/full):与 talk.log 并排各调一次,见 openspec add-agent-loop-capture D1
7
8
 
8
9
  // 最后一站的收场:relay 的「不写响应」路径(超时 / 空轮延后封口)都汇到这里,没有上层 failover 了,
9
10
  // 必须把响应「恰好一次」地关掉。形状由 endHeldFailure 按 headers 是否已 flush 决定
@@ -25,7 +26,13 @@ export async function handleExhaustedLocal({ res, body, lastErr, order, handlerC
25
26
  });
26
27
  evt("relay-done", { reqId: handlerCtx.reqId, model: lastErr.model, via: "local-exhausted", status: out.status, ttfMs: out.ttfMs, totalMs: out.totalMs, aborted: out.aborted, interrupted: out.interrupted ?? false, detail: out.detail ?? null });
27
28
  // 绕过 relay-pipeline 的最后一站:对话日志照常落盘(正文只进 talk/*.log,不进 events.log)
28
- try { recordRelayTalk({ reqId: handlerCtx.reqId, model: lastErr.model, via: "local-exhausted", body, out }); } catch {}
29
+ // 本收尾点没有 attemptStartMs 可取(绕过 relay-pipeline 直连 relay)→ 不传 attemptMs,按回退路径走:
30
+ // elapsed=relay 内部计时(与 relayMs 同值,spec 降级 scenario);finishReason 传 out.detail.sawFinishReason 与 3.3 同源。
31
+ try { recordRelayTalk({ reqId: handlerCtx.reqId, model: lastErr.model, via: "local-exhausted", body, out, finishReason: out.detail?.sawFinishReason }); } catch {}
32
+ // 回路语料另起一条独立 try:talk.log 是人读环形稿、capture 是机读全量语料,两者生死独立。
33
+ // hops 走 handlerCtx.hops(serial-trial 以 {...handlerCtx} 展开传进来,与 reqId 同源);缺值时不兜 0 ——
34
+ // 交给 talk-full 落 null(「hops 未知」在语料里看得见,比伪造成 0 诚实),组员流量拦截靠真值。
35
+ try { recordAgentLoop({ reqId: handlerCtx.reqId, model: lastErr.model, via: "local-exhausted", hops: handlerCtx?.hops, body, out, sessionKey: handlerCtx?.sessionId, clientIp: handlerCtx?.clientIp }); } catch {} // attemptMs 缺省 → 回退 out.totalMs;缺 echo 的既有形状本期不补(tasks 3.4)
29
36
  evt("result", { reqId: handlerCtx.reqId, model: lastErr.model, status: out.status, via: "local", timing: lastErr.upstream._t ?? null, ttfMs: out.ttfMs, totalMs: out.totalMs, detail: out.detail ?? null });
30
37
  evt("client-response", { requested, actual: lastErr.model, via: "local", fallback: false, status: out.status, reqId: handlerCtx.reqId });
31
38
  // 最后一站:relay 超时/空轮留口路径都不写响应,这里没有上层,必须自己收尾
@@ -55,7 +62,11 @@ export async function handleExhaustedAll({ res, body, lastErr, order, requested,
55
62
  });
56
63
  evt("relay-done", { reqId: handlerCtx.reqId, model: lastErr.model, via: "local-final", status: out.status, ttfMs: out.ttfMs, totalMs: out.totalMs, aborted: out.aborted, interrupted: out.interrupted ?? false, detail: out.detail ?? null });
57
64
  // 同上:全链耗尽的最终转发也要留下问答原文
58
- try { recordRelayTalk({ reqId: handlerCtx.reqId, model: lastErr.model, via: "local-final", body, out }); } catch {}
65
+ // 同上:本收尾点没有 attemptStartMs 可取(relay 直连、无尝试起点)→ 不传 attemptMs,
66
+ // recordRelayTalk/recordAgentLoop 内部按回退路径走 elapsed=relay 计时(与 relayMs 同值,spec 降级 scenario);finishReason 同源传 out.detail.sawFinishReason。
67
+ try { recordRelayTalk({ reqId: handlerCtx.reqId, model: lastErr.model, via: "local-final", body, out, finishReason: out.detail?.sawFinishReason }); } catch {}
68
+ // 同上:全链耗尽的最终转发也要进语料(这条往往是「客户端只看到 502」时唯一的现场材料)
69
+ try { recordAgentLoop({ reqId: handlerCtx.reqId, model: lastErr.model, via: "local-final", hops: handlerCtx?.hops, body, out, sessionKey: handlerCtx?.sessionId, clientIp: handlerCtx?.clientIp }); } catch {} // attemptMs 缺省 → 回退 out.totalMs(见上行注释;缺 echo 的形状本期不补)
59
70
  evt("result", { reqId: handlerCtx.reqId, model: lastErr.model, status: out.status, via: "local", timing: lastErr.upstream._t ?? null, ttfMs: out.ttfMs, totalMs: out.totalMs, detail: out.detail ?? null });
60
71
  evt("client-response", { requested, actual: lastErr.model, via: "local", fallback: false, status: out.status, reqId: handlerCtx.reqId });
61
72
  // 最后一站:relay 超时/空轮留口路径不写响应,这里没有上层 failover,必须自己收尾
@@ -1,47 +1,47 @@
1
- import { clientIp, json, readBody, parseHops } from "../helpers.js";
2
- import { runHook } from "../../plugins.js";
3
- import { createChatPipeline } from "../../chat-pipeline/index.js";
4
-
5
- /**
6
- * ChatGateway 薄适配层:仅负责 HTTP I/O(readBody + request:received hook)→ 委托 ChatPipeline.execute
7
- * Policy/Order/AutoRace/hedge/peer/broadband 全部在 chat-pipeline 深模块内。
8
- */
9
- export function createChatGateway(deps = {}) {
10
- const pipeline = createChatPipeline(deps);
11
-
12
- async function handle({ req, res }) {
13
- let body;
14
- try {
15
- body = await readBody(req);
16
- } catch {
17
- return json(res, 400, { error: "Invalid JSON body" });
18
- }
19
- if (deps.plugins?.length) {
20
- const rc = await runHook(deps.plugins, "request:received", {
21
- ip: clientIp(req),
22
- hops: parseHops(req.headers["x-mslxdff-hops"]),
23
- headers: { "content-type": req.headers["content-type"] },
24
- body,
25
- });
26
- for (const e of rc.errors) deps.logs?.appendEvent?.({ ts: Date.now(), type: "plugin-hook-error", hook: "request:received", plugin: e.plugin, error: e.error });
27
- const respond = rc.value?.respond;
28
- if (respond && typeof respond === "object") return json(res, respond.status || 200, respond.body ?? {});
29
- }
30
- req.body = body;
31
- try {
32
- await pipeline.execute({ req, res });
33
- } catch (err) {
34
- const msg = err?.message || String(err);
35
- if (!res.headersSent) return json(res, 502, { error: msg });
36
- try { res.end(); } catch {}
37
- }
38
- }
39
-
40
- return { handle };
41
- }
42
-
43
- // 薄适配:保持原 chatHandler 签名兼容
44
- export async function chatHandler(ctx) {
45
- const gw = createChatGateway(ctx);
46
- return gw.handle(ctx);
1
+ import { clientIp, json, readBody, parseHops } from "../helpers.js";
2
+ import { runHook } from "../../plugins.js";
3
+ import { createChatPipeline } from "../../chat-pipeline/index.js";
4
+
5
+ /**
6
+ * ChatGateway 薄适配层:仅负责 HTTP I/O(readBody + request:received hook)→ 委托 ChatPipeline.execute
7
+ * Policy/Order/AutoRace/hedge/peer/broadband 全部在 chat-pipeline 深模块内。
8
+ */
9
+ export function createChatGateway(deps = {}) {
10
+ const pipeline = createChatPipeline(deps);
11
+
12
+ async function handle({ req, res }) {
13
+ let body;
14
+ try {
15
+ body = await readBody(req);
16
+ } catch {
17
+ return json(res, 400, { error: "Invalid JSON body" });
18
+ }
19
+ if (deps.plugins?.length) {
20
+ const rc = await runHook(deps.plugins, "request:received", {
21
+ ip: clientIp(req),
22
+ hops: parseHops(req.headers["x-mslxdff-hops"]),
23
+ headers: { "content-type": req.headers["content-type"] },
24
+ body,
25
+ });
26
+ for (const e of rc.errors) deps.logs?.appendEvent?.({ ts: Date.now(), type: "plugin-hook-error", hook: "request:received", plugin: e.plugin, error: e.error });
27
+ const respond = rc.value?.respond;
28
+ if (respond && typeof respond === "object") return json(res, respond.status || 200, respond.body ?? {});
29
+ }
30
+ req.body = body;
31
+ try {
32
+ await pipeline.execute({ req, res });
33
+ } catch (err) {
34
+ const msg = err?.message || String(err);
35
+ if (!res.headersSent) return json(res, 502, { error: msg });
36
+ try { res.end(); } catch {}
37
+ }
38
+ }
39
+
40
+ return { handle };
41
+ }
42
+
43
+ // 薄适配:保持原 chatHandler 签名兼容
44
+ export async function chatHandler(ctx) {
45
+ const gw = createChatGateway(ctx);
46
+ return gw.handle(ctx);
47
47
  }