pi-tsgw 0.5.0 → 0.6.0

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.
package/README.md CHANGED
@@ -89,7 +89,7 @@ pi install npm:pi-tsgw
89
89
  目录由各厂商的分片(`extensions/models/vendors/`)组装而成,涵盖 11 个厂商系列:
90
90
  DeepSeek、GLM(智谱)、MiMo(小米)、MiniMax、Kimi(Moonshot)、Qwen(阿里巴巴)、
91
91
  GPT(OpenAI)、Gemini(Google)、Claude(Anthropic)、LongCat 和 Grok(xAI)。
92
- 当前静态目录包含 36 个聊天模型 ID(含保留的别名);这不是实时可用模型数量。
92
+ 静态目录会持续调整,不代表网关实时可用模型数量。
93
93
  厂商文件包含文档链接和元数据注意事项。别名映射、输出限制和共享思考策略并未全部
94
94
  针对网关端点逐项验证;公开价格估算不构成计费保证。
95
95
  MiMo 使用 `mimo-v2.6-pro` 和 `mimo-v2.6-flash`;不再注册原有 V2.5 ID。
@@ -97,10 +97,33 @@ MiMo 使用 `mimo-v2.6-pro` 和 `mimo-v2.6-flash`;不再注册原有 V2.5 ID
97
97
 
98
98
  精选目录保留 DeepSeek `deepseek-flash` 和 V4 Pro;GLM 5.2+,以及明确例外
99
99
  `glm-4.5-air`;MiniMax M3+;Kimi K3;GPT 6+;以及 Gemini 3.5+ 和无版本号的
100
- `gemini-flash` / `gemini-pro` 别名。Qwen、MiMo、LongCat、Claude 和 Grok 保留现有注册。
101
- 已移除的 ID 及其专用请求策略不再注册;即使网关仍列出这些 ID,纳入规则也无法恢复它们。
100
+ `gemini-flash` / `gemini-pro` 别名。当前裁剪移除了 Qwen `qwen3.7-max`、`qwen3.6-plus`、
101
+ `qwen3.5-plus`,Claude `claude-fable-5`,以及 Grok `grok-4.20`、`grok-4.20-fast`、
102
+ `grok-4.3-low`、`grok-4.3-medium`、`grok-4.3-high`。Qwen 仅保留 `qwen3.7-plus` 及其
103
+ 完整请求策略,作为厂商最后一个接入参考;它仍须命中网关目录才会登记,并非可绕过网关
104
+ 交集的默认模型。Grok 保留 `grok-4.5` 并新增 `grok-4.7`;Claude 当前仅保留 Opus 和 Sonnet。
105
+ 其他已移除的 ID 及其专用请求策略不再注册;即使网关仍列出这些 ID,纳入规则也无法恢复它们。
102
106
  现有显式模型选择若引用了已移除的 ID,需要更新。
103
107
 
108
+ 新增 `grok-4.7`、`gemini-3.5-flash-lite`、`gemini-3.8-flash-high` 和 `longcat-2.5-preview`:
109
+
110
+ | ID | 协议与端点后缀 | 思考映射 | 2026-10-17 最小真实验收 |
111
+ | --- | --- | --- | --- |
112
+ | `grok-4.7` | Responses,`/v1` | low/medium/high/xhigh;不支持 off,max→xhigh | 文本通过;函数调用未通过;原生搜索无事件/来源,故不加入搜索支持名单 |
113
+ | `gemini-3.5-flash-lite` | Completions,`/v1` | minimal/low/medium/high;off→minimal,xhigh/max→high | 文本、函数调用与默认工具选择下的结果往返通过;显式 tool_choice:none 的结果轮曾超时 |
114
+ | `gemini-3.8-flash-high` | Google,`/gemini/v1beta` | low/medium/high;off/minimal→low,xhigh/max→high | 文本、函数调用与工具结果往返通过 |
115
+ | `longcat-2.5-preview` | Completions,`/v1` | off 关闭,其他档开启;无已确认深度映射 | 上游额度耗尽,推理及工具验收阻塞 |
116
+
117
+ 上述验收仅覆盖短文本与最小工具夹具,不代表全部模态、上下文容量或七档已逐项验证。
118
+ 两款 Gemini 的旧 `/gemini` 路径均不可用;仅新型号按实测选择协议与版本路径,不修改旧型号。
119
+ 3.8 的网关 ID 后缀 `high` 不作为已确认的固定档位契约;规格/价格继承官方基础型号。
120
+ 3.8 采用截至 2026-12-31 的介绍价,2027-01-01 起应复核更新;Grok 4.7 采用输入少于
121
+ 200K 的公开价,达到 200K 后输入、缓存与输出均加倍。LongCat 使用国内限时价并按
122
+ 固定 0.1475 USD/CNY 折算;公开价格只用于估算,不代表网关实际计费。
123
+ Grok 4.7 的官方文本输出不设独立上限,目录以 500K 上下文作为输出容量上界,
124
+ 不代表可同时填满输入与输出。LongCat 2.5 输出上限按逐型号参数页采用 256K;
125
+ 官方快速开始仍写 128K,图像更新说明也与纯文本 API schema 存在差异,当前保守登记文本输入。
126
+
104
127
  Kimi 提供 `kimi-k3` 和网关变体 `kimi-k3-256k`,后者显示为
105
128
  **Kimi K3 (256K)**,上下文窗口为 262,144 个 token。其余元数据、价格估算和思考策略
106
129
  继承自 K3;131,072 个 token 的输出设置与上下文大小彼此独立。
@@ -4,29 +4,11 @@ import { ANTHROPIC_COMPAT } from "./_protocols.ts";
4
4
  /**
5
5
  * Anthropic Claude 系列(anthropic-messages 协议)。
6
6
  * 官方文档:https://docs.claude.com/en/docs/about-claude/models/
7
- * 规格/定价以网关目录为准(fable/opus/sonnet 三档)。
7
+ * 规格/定价以网关目录为准(opus/sonnet 两档)。
8
8
  */
9
9
  export function anthropicModels(root: string): ProviderModelConfig[] {
10
10
  const anthropic = `${root}/anthropic`;
11
11
  return [
12
- {
13
- id: "claude-fable-5",
14
- name: "Claude Fable 5",
15
- api: "anthropic-messages",
16
- baseUrl: anthropic,
17
- reasoning: true,
18
- thinkingLevelMap: {
19
- low: "low",
20
- medium: "medium",
21
- high: "high",
22
- max: "max",
23
- },
24
- input: ["text", "image"],
25
- cost: { input: 10, output: 50, cacheRead: 1, cacheWrite: 12.3 },
26
- contextWindow: 1000000,
27
- maxTokens: 128000,
28
- compat: ANTHROPIC_COMPAT,
29
- },
30
12
  {
31
13
  id: "claude-opus-5",
32
14
  name: "Claude Opus 5",
@@ -1,11 +1,27 @@
1
1
  import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
3
+ import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
4
 
4
5
  /**
5
- * Google Gemini 系列(google-generative-ai 协议)。
6
+ * Google Gemini 系列(原生 Google 协议;3.5 Flash-Lite 使用网关 Completions 路由)。
6
7
  * 官方文档:https://ai.google.dev/gemini-api/docs
7
8
  * GenerateContent 档位/预算:https://docs.cloud.google.com/vertex-ai/generative-ai/docs/thinking
8
9
  * gemini-flash 别名保留预算策略,其他保留型号使用档位策略。网关别名的路由仍需网关侧验证。
10
+ *
11
+ * 新增型号依据(2026-10-17 核查):
12
+ * - gemini-3.5-flash-lite:模型页 input 1,048,576 / output 65,536;思考档
13
+ * minimal/low/medium/high(默认 minimal,无 off),见
14
+ * https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite 与
15
+ * https://ai.google.dev/gemini-api/docs/thinking 的逐型号表。
16
+ * - gemini-3.8-flash-high:网关精确 ID。官方无该后缀型号,规格(input
17
+ * 1,048,576 / output 65,536;思考 low/medium/high,默认 medium,minimal
18
+ * 会报错)继承基础型号 gemini-3.8-flash,实际别名路由未验证;不登记
19
+ * 裸 ID gemini-3.8-flash,也不把 high 后缀当作固定思考档位。
20
+ * https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash
21
+ * https://ai.google.dev/gemini-api/docs/whats-new-gemini-3.8
22
+ * - 3.5 Flash-Lite 与 3.8 起废弃采样参数(temperature/top_p/top_k 被忽略,
23
+ * 未来版本报错)并改用 thinking_level 枚举,见
24
+ * https://ai.google.dev/gemini-api/docs/whats-new-gemini-3.6。
9
25
  */
10
26
  export function geminiModels(root: string): ProviderModelConfig[] {
11
27
  const gemini = `${root}/gemini`;
@@ -68,6 +84,40 @@ export function geminiModels(root: string): ProviderModelConfig[] {
68
84
  contextWindow: 1048576,
69
85
  maxTokens: 65536,
70
86
  },
87
+ {
88
+ id: "gemini-3.5-flash-lite",
89
+ name: "Gemini 3.5 Flash-Lite",
90
+ // 2026-10-17:原生 Google 路由 404;此 Completions 路由文本/函数通过,
91
+ // 默认工具选择下结果往返通过;显式 tool_choice:none 的结果轮曾超时。
92
+ api: "openai-completions",
93
+ baseUrl: `${root}/v1`,
94
+ compat: OPENAI_COMPLETIONS_COMPAT,
95
+ thinkingLevelMap: { off: null, minimal: "minimal", low: "low", medium: "medium", high: "high", xhigh: "high", max: "high" },
96
+ reasoning: true,
97
+ input: ["text", "image"],
98
+ // 标准层公开价(2026-10-17):输入 $0.30、输出 $2.50(含思考);
99
+ // 缓存读取 $0.03,存储 $1/1M tokens/小时按小时计费,无法映射为单价,置 0。
100
+ cost: { input: 0.3, output: 2.5, cacheRead: 0.03, cacheWrite: 0 },
101
+ contextWindow: 1048576,
102
+ maxTokens: 65536,
103
+ },
104
+ {
105
+ id: "gemini-3.8-flash-high",
106
+ name: "Gemini 3.8 Flash",
107
+ api: "google-generative-ai",
108
+ // 宿主指定 baseUrl 时不追加版本路径;此路由已通过文本/工具往返。
109
+ baseUrl: `${gemini}/v1beta`,
110
+ thinkingLevelMap: { off: null, minimal: "low", low: "low", medium: "medium", high: "high", xhigh: "high", max: "high" },
111
+ reasoning: true,
112
+ input: ["text", "image"],
113
+ // 网关精确 ID,规格与价格继承官方基础型号 gemini-3.8-flash(别名路由
114
+ // 未验证)。标准层介绍价:输入 $0.75、输出 $3.75(含思考)、缓存读取
115
+ // $0.075,优惠截止 2026-12-31;2027-01-01 起为 $1.50/$7.50/$0.15。
116
+ // 缓存存储 $0.50/1M tokens/小时(介绍价)按小时计费,无法映射为单价,置 0。
117
+ cost: { input: 0.75, output: 3.75, cacheRead: 0.075, cacheWrite: 0 },
118
+ contextWindow: 1048576,
119
+ maxTokens: 65536,
120
+ },
71
121
  ];
72
122
  }
73
123
 
@@ -100,10 +150,26 @@ function applyGemini35Flash(writer: PayloadWriter, level: ThinkingLevel): void {
100
150
  writer.setGoogleThinking({ includeThoughts: level !== "off", thinkingLevel }, ["thinkingBudget"]);
101
151
  }
102
152
 
153
+ // 3.5 Flash-Lite 官方档位同为 minimal/low/medium/high 且默认 minimal,无 off:
154
+ // 网关 Completions 使用 reasoning_effort;off 就近 minimal,xhigh/max 收敛 high。
155
+ function applyGemini35FlashLite(writer: PayloadWriter, level: ThinkingLevel): void {
156
+ writer.remove("thinking");
157
+ const effort = level === "off" ? "minimal" : level === "xhigh" || level === "max" ? "high" : level;
158
+ writer.set("reasoning_effort", effort);
159
+ }
160
+
161
+ // 3.8 Flash 仅支持 low/medium/high(minimal 会报错,无 off);off/minimal
162
+ // 就近落到 LOW,xhigh/max 收敛到 HIGH;后缀 high 不锁定档位。
163
+ function applyGemini38Flash(writer: PayloadWriter, level: ThinkingLevel): void {
164
+ applyGeminiPro(writer, level);
165
+ }
166
+
103
167
  export const geminiThinking: Record<string, ThinkingApplier> = {
104
168
  "gemini-flash": (w, c) => applyGeminiFlash(w, c.thinkingLevel),
105
169
  "gemini-3.5-flash": (w, c) => applyGemini35Flash(w, c.thinkingLevel),
106
170
  "gemini-3.5-flash-low": (w, c) => applyGemini35Flash(w, c.thinkingLevel),
107
171
  "gemini-3.5-flash-extra-low": (w, c) => applyGemini35Flash(w, c.thinkingLevel),
172
+ "gemini-3.5-flash-lite": (w, c) => applyGemini35FlashLite(w, c.thinkingLevel),
108
173
  "gemini-pro": (w, c) => applyGeminiPro(w, c.thinkingLevel),
174
+ "gemini-3.8-flash-high": (w, c) => applyGemini38Flash(w, c.thinkingLevel),
109
175
  };
@@ -1,28 +1,30 @@
1
1
  import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
- import { type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
3
+ import { isPlainObject, type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
4
4
 
5
5
  /**
6
- * xAI Grok 系列(openai-completions 协议)。
6
+ * xAI Grok 系列:4.5 使用 Completions,4.7 使用 Responses。
7
7
  * 官方文档:https://docs.x.ai/developers/models
8
8
  * 定价(docs.x.ai/developers/pricing,提示词 <200k 档):
9
- * grok-4.20 / grok-4.3:每百万词元输入 $1.25 / 缓存命中 $0.20 / 输出 $2.50
10
9
  * grok-4.5:每百万词元输入 $2.00 / 缓存命中 $0.30 / 输出 $6.00
11
10
  */
12
11
  export function grokModels(root: string): ProviderModelConfig[] {
13
12
  const v1 = `${root}/v1`;
14
13
  return [
15
14
  {
16
- id: "grok-4.20",
17
- name: "Grok 4.20",
18
- api: "openai-completions",
15
+ id: "grok-4.7",
16
+ name: "Grok 4.7",
17
+ api: "openai-responses",
19
18
  baseUrl: v1,
20
19
  reasoning: true,
21
20
  input: ["text", "image"],
22
- cost: { input: 1.25, output: 2.5, cacheRead: 0.2, cacheWrite: 0 },
23
- contextWindow: 1000000,
24
- maxTokens: 65536,
25
- compat: OPENAI_COMPLETIONS_COMPAT,
21
+ // 2026-10-17:https://docs.x.ai/developers/models/grok-4.7
22
+ // <200K 官方参考价;≥200K 输入/缓存/输出均 ×2,cacheWrite=0 表示未单列。
23
+ cost: { input: 2, output: 6, cacheRead: 0.5, cacheWrite: 0 },
24
+ contextWindow: 500000,
25
+ // 官方不设独立文本输出上限;此值取上下文容量上界,不宣称可同时填满输入与输出。
26
+ maxTokens: 500000,
27
+ thinkingLevelMap: { off: null, minimal: "low", low: "low", medium: "medium", high: "high", xhigh: "xhigh", max: "xhigh" },
26
28
  },
27
29
  {
28
30
  id: "grok-4.5",
@@ -36,65 +38,11 @@ export function grokModels(root: string): ProviderModelConfig[] {
36
38
  maxTokens: 65536,
37
39
  compat: OPENAI_COMPLETIONS_COMPAT,
38
40
  },
39
- {
40
- id: "grok-4.3-low",
41
- name: "Grok 4.3",
42
- api: "openai-completions",
43
- baseUrl: v1,
44
- reasoning: true,
45
- input: ["text", "image"],
46
- // 网关思考档位别名,规格继承 grok-4.3。
47
- cost: { input: 1.25, output: 2.5, cacheRead: 0.2, cacheWrite: 0 },
48
- contextWindow: 1000000,
49
- maxTokens: 65536,
50
- compat: OPENAI_COMPLETIONS_COMPAT,
51
- },
52
- {
53
- id: "grok-4.3-medium",
54
- name: "Grok 4.3",
55
- api: "openai-completions",
56
- baseUrl: v1,
57
- reasoning: true,
58
- input: ["text", "image"],
59
- // 网关思考档位别名,规格继承 grok-4.3。
60
- cost: { input: 1.25, output: 2.5, cacheRead: 0.2, cacheWrite: 0 },
61
- contextWindow: 1000000,
62
- maxTokens: 65536,
63
- compat: OPENAI_COMPLETIONS_COMPAT,
64
- },
65
- {
66
- id: "grok-4.3-high",
67
- name: "Grok 4.3",
68
- api: "openai-completions",
69
- baseUrl: v1,
70
- reasoning: true,
71
- input: ["text", "image"],
72
- // 网关思考档位别名,规格继承 grok-4.3。
73
- cost: { input: 1.25, output: 2.5, cacheRead: 0.2, cacheWrite: 0 },
74
- contextWindow: 1000000,
75
- maxTokens: 65536,
76
- compat: OPENAI_COMPLETIONS_COMPAT,
77
- },
78
- {
79
- id: "grok-4.20-fast",
80
- name: "Grok 4.20 Fast",
81
- api: "openai-completions",
82
- baseUrl: v1,
83
- reasoning: true,
84
- input: ["text", "image"],
85
- // 网关快速档别名,规格继承 grok-4.20。
86
- cost: { input: 1.25, output: 2.5, cacheRead: 0.2, cacheWrite: 0 },
87
- contextWindow: 1000000,
88
- maxTokens: 65536,
89
- compat: OPENAI_COMPLETIONS_COMPAT,
90
- },
91
41
  ];
92
42
  }
93
43
 
94
44
  // 官方:https://docs.x.ai/developers/model-capabilities/text/reasoning
95
45
  // grok-4.5 支持 low/medium/high,不支持禁用 reasoning,off 映射到 low。
96
- // 4.20 / 4.3 网关别名暂保留既有映射;上述官方指南没有确认其思考深度支持,
97
- // 不应将同族回归测试视为这些别名已通过端点验证。
98
46
  function applyGrok(writer: PayloadWriter, level: ThinkingLevel): void {
99
47
  writer.remove("thinking");
100
48
  if (level === "off") {
@@ -110,11 +58,19 @@ function applyGrok(writer: PayloadWriter, level: ThinkingLevel): void {
110
58
  writer.set("reasoning_effort", effort);
111
59
  }
112
60
 
61
+ // Responses 保留宿主的 encrypted reasoning/include 与其他 reasoning 字段,
62
+ // 仅归一化不受支持的档位;off 不关闭思考,遗留会话回退 low。
63
+ function applyGrok47(writer: PayloadWriter, level: ThinkingLevel): void {
64
+ writer.remove("thinking");
65
+ writer.remove("reasoning_effort");
66
+ const current = writer.get("reasoning");
67
+ if (current !== undefined && !isPlainObject(current)) return;
68
+ const reasoning = isPlainObject(current) ? current : {};
69
+ const effort = level === "off" || level === "minimal" ? "low" : level === "max" ? "xhigh" : level;
70
+ if (reasoning.effort !== effort) writer.set("reasoning", { ...reasoning, effort });
71
+ }
72
+
113
73
  export const grokThinking: Record<string, ThinkingApplier> = {
114
- "grok-4.20": (w, c) => applyGrok(w, c.thinkingLevel),
74
+ "grok-4.7": (w, c) => applyGrok47(w, c.thinkingLevel),
115
75
  "grok-4.5": (w, c) => applyGrok(w, c.thinkingLevel),
116
- "grok-4.3-low": (w, c) => applyGrok(w, c.thinkingLevel),
117
- "grok-4.3-medium": (w, c) => applyGrok(w, c.thinkingLevel),
118
- "grok-4.3-high": (w, c) => applyGrok(w, c.thinkingLevel),
119
- "grok-4.20-fast": (w, c) => applyGrok(w, c.thinkingLevel),
120
76
  };
@@ -3,8 +3,10 @@ import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
3
  import { applyEnabledThinking, type ThinkingApplier } from "../_tools.ts";
4
4
 
5
5
  /**
6
- * LongCat 2.0(龙猫)。
7
- * 官方文档:https://longcat.chat/platform/docs/zh/
6
+ * 美团 LongCat(龙猫)系列。
7
+ * 官方中文文档(2026-10-17 核实):https://longcat.ai/platform/docs/zh/
8
+ * 网关目录 ID 为小写 longcat-*;上游 API 模型名保留大小写
9
+ * (LongCat-2.5-Preview / LongCat-2.0)。
8
10
  */
9
11
  export function longcatModels(root: string): ProviderModelConfig[] {
10
12
  const v1 = `${root}/v1`;
@@ -21,10 +23,41 @@ export function longcatModels(root: string): ProviderModelConfig[] {
21
23
  maxTokens: 131072,
22
24
  compat: OPENAI_COMPLETIONS_COMPAT,
23
25
  },
26
+ {
27
+ // LongCat-2.5-Preview(2026-09-25 上线)。依据官方中文文档:
28
+ // - 输入模态:模型详情 /api/model 示例 input_modalities 为 ["text"]、
29
+ // modality "text->text",OpenAI/Anthropic 两套消息 schema 的 content
30
+ // 均为“纯文本字符串”;更新日志宣称新增图片理解能力,但当前 API
31
+ // 契约未提供图像入参,按 API 契约登记为纯文本输入(冲突见更新日志)。
32
+ // - 定价(国内 chat 版,限时折扣价)/pricing/longcat-2.5:
33
+ // 输入未命中缓存 ¥2、命中缓存 ¥0.04、输出 ¥8 每百万词元,无缓存写计费;
34
+ // 按既有固定 0.1475 USD/CNY 折算为 0.295 / 0.0059 / 1.18 USD。
35
+ // - 参数 /api/model-params:上下文 1,048,576、max_tokens 上限
36
+ // 262,144(256K)。快速开始页仍写最大输出 128K 且两型号相同,
37
+ // 与逐型号对照的参数参考页冲突;以参数参考页为准(thinking_budget
38
+ // 上限同为 262,144 可互证)。
39
+ id: "longcat-2.5-preview",
40
+ name: "LongCat-2.5-Preview",
41
+ api: "openai-completions",
42
+ baseUrl: v1,
43
+ reasoning: true,
44
+ input: ["text"],
45
+ cost: { input: 0.295, output: 1.18, cacheRead: 0.0059, cacheWrite: 0 },
46
+ contextWindow: 1048576,
47
+ maxTokens: 262144,
48
+ // 官方协议字段为 max_tokens,由宿主 compat 映射,不自行改写 token 上限。
49
+ compat: { ...OPENAI_COMPLETIONS_COMPAT, maxTokensField: "max_tokens" },
50
+ },
24
51
  ];
25
52
  }
26
53
 
27
- // OC/网关兼容策略:LongCat 的公开 HTTP thinking 结构定义未验证,用通用开关式。
54
+ // 2.5 的 /api/chat 与 /api/model-params 已确认 thinking.type enabled/disabled;
55
+ // 独立核对后可复用通用开关式,不支持 reasoning_effort。预算没有逐档依据,不臆造映射。
56
+ // 可关闭思考,故不设 off:null;采样温度由服务端按思考开关覆盖。
57
+
58
+ // OC/网关兼容策略:longcat-2.0 的公开 HTTP thinking 结构定义未验证,沿用通用开关式;
59
+ // longcat-2.5-preview 经独立核对后复用相同的开关形状(见上方依据)。
28
60
  export const longcatThinking: Record<string, ThinkingApplier> = {
29
61
  "longcat-2.0": (w, c) => applyEnabledThinking(w, c.thinkingLevel),
62
+ "longcat-2.5-preview": (w, c) => applyEnabledThinking(w, c.thinkingLevel),
30
63
  };
@@ -6,6 +6,7 @@ import { type PayloadWriter, type ThinkingApplier } from "../_tools.ts";
6
6
  * OpenAI GPT 系列(Responses 协议)。
7
7
  * 官方文档:https://platform.openai.com/docs/models
8
8
  * GPT-6.1 Sol:https://developers.openai.com/api/docs/models/gpt-6.1-sol
9
+ * 用户确认网关接入 Codex 对应型号,上下文为 500000;最大输出与价格保持原值。
9
10
  * 标准档 ≤272K:输入 $2 / 缓存命中 $0.1 / 缓存写入 $2.5 / 输出 $10(2026-09-30)。
10
11
  * GPT-6 Sol:https://developers.openai.com/api/docs/models/gpt-6-sol
11
12
  * GPT-6 Luna:https://developers.openai.com/api/docs/models/gpt-6-luna
@@ -53,7 +54,7 @@ export function openaiModels(root: string): ProviderModelConfig[] {
53
54
  },
54
55
  input: ["text", "image"],
55
56
  cost: { input: 2, output: 10, cacheRead: 0.1, cacheWrite: 2.5 },
56
- contextWindow: 1050000,
57
+ contextWindow: 500000,
57
58
  maxTokens: 128000,
58
59
  },
59
60
  {
@@ -4,6 +4,7 @@ import { isAtLeastHigh, type PayloadWriter, type ThinkingApplier, type ThinkingL
4
4
 
5
5
  /**
6
6
  * 阿里云通义(Alibaba Qwen)系列。
7
+ * 当前仅保留 qwen3.7-plus 及策略作为厂商接入参考;实际登记仍须命中网关目录。
7
8
  * 官方文档:https://help.aliyun.com/zh/model-studio/
8
9
  */
9
10
  export function qwenModels(root: string): ProviderModelConfig[] {
@@ -26,44 +27,6 @@ export function qwenModels(root: string): ProviderModelConfig[] {
26
27
  maxTokens: 64000,
27
28
  compat: OPENAI_COMPLETIONS_COMPAT,
28
29
  },
29
- {
30
- id: "qwen3.7-max",
31
- name: "Qwen3.7 Max",
32
- api: "openai-completions",
33
- baseUrl: v1,
34
- reasoning: true,
35
- input: ["text"],
36
- cost: { input: 1.77, output: 5.31, cacheRead: 0.354, cacheWrite: 2.2125 },
37
- contextWindow: 1000000,
38
- maxTokens: 64000,
39
- compat: OPENAI_COMPLETIONS_COMPAT,
40
- },
41
- {
42
- id: "qwen3.6-plus",
43
- name: "Qwen3.6 Plus",
44
- api: "openai-completions",
45
- baseUrl: v1,
46
- reasoning: true,
47
- input: ["text", "image"],
48
- // 阿里云百炼 CNY 定价按 0.1475 USD/CNY 换算(与 qwen3.7 系列一致)。
49
- cost: { input: 0.295, output: 1.77, cacheRead: 0.0295, cacheWrite: 0.36875 },
50
- contextWindow: 1000000,
51
- maxTokens: 65536,
52
- compat: OPENAI_COMPLETIONS_COMPAT,
53
- },
54
- {
55
- id: "qwen3.5-plus",
56
- name: "Qwen3.5 Plus",
57
- api: "openai-completions",
58
- baseUrl: v1,
59
- reasoning: true,
60
- input: ["text", "image"],
61
- // 阿里云百炼 CNY 定价按 0.1475 USD/CNY 换算(与 qwen3.7 系列一致)。
62
- cost: { input: 0.118, output: 0.708, cacheRead: 0.0118, cacheWrite: 0.1475 },
63
- contextWindow: 1000000,
64
- maxTokens: 65536,
65
- compat: OPENAI_COMPLETIONS_COMPAT,
66
- },
67
30
  ];
68
31
  }
69
32
 
@@ -96,7 +59,4 @@ function applyQwen(writer: PayloadWriter, level: ThinkingLevel): void {
96
59
 
97
60
  export const qwenThinking: Record<string, ThinkingApplier> = {
98
61
  "qwen3.7-plus": (w, c) => applyQwen(w, c.thinkingLevel),
99
- "qwen3.7-max": (w, c) => applyQwen(w, c.thinkingLevel),
100
- "qwen3.6-plus": (w, c) => applyQwen(w, c.thinkingLevel),
101
- "qwen3.5-plus": (w, c) => applyQwen(w, c.thinkingLevel),
102
62
  };
@@ -6,7 +6,7 @@
6
6
  * `before_provider_request` 中应用。
7
7
  *
8
8
  * - GPT(openai-responses):追加 `tools: [{ type: "web_search", ... }]`
9
- * - Grok(openai-completions):追加 `search_parameters: { mode: "on" }`
9
+ * - Grok 4.5(Completions):追加 `search_parameters: { mode: "on" }`
10
10
  * (xAI 文档:https://docs.x.ai/developers/tools/web-search)
11
11
  *
12
12
  * 名单外主模型通过 tsgw_search 调用目录内支持模型;注入逻辑保持共用。
@@ -31,14 +31,7 @@ const BUILTIN_SEARCH_MODELS = new Set([
31
31
  ]);
32
32
 
33
33
  // 支持内置查询的 Grok 系列(openai-completions 协议,search_parameters 模式)。
34
- const GROK_SEARCH_MODELS = new Set([
35
- "grok-4.20",
36
- "grok-4.5",
37
- "grok-4.3-low",
38
- "grok-4.3-medium",
39
- "grok-4.3-high",
40
- "grok-4.20-fast",
41
- ]);
34
+ const GROK_SEARCH_MODELS = new Set(["grok-4.5"]);
42
35
 
43
36
  /** 精确判定模型/协议支持性,由请求注入与展示代码共用。 */
44
37
  export function builtinSearchSupport(
@@ -65,11 +58,6 @@ const DELEGATE_PREFERENCE = [
65
58
  "gpt-6-sol",
66
59
  "gpt-6.1-sol",
67
60
  "gpt-6-astra",
68
- "grok-4.20",
69
- "grok-4.20-fast",
70
- "grok-4.3-low",
71
- "grok-4.3-medium",
72
- "grok-4.3-high",
73
61
  "grok-4.5",
74
62
  ] as const;
75
63
 
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "pi-tsgw",
3
- "version": "0.5.0",
3
+ "version": "0.6.0",
4
4
  "description": "面向 TOSSP AIH gateway 的 Pi 扩展:按厂商维护聊天与图片模型目录及思考策略,支持四种聊天协议与图片生成、编辑工具。",
5
5
  "type": "module",
6
6
  "keywords": [
@@ -24,7 +24,7 @@
24
24
  ]
25
25
  },
26
26
  "scripts": {
27
- "test": "tsc -p tsconfig.test.json && node dist/extensions/provider/index.test.js && node dist/extensions/provider/refresh-tool.test.js && node dist/extensions/provider/startup-catalog.test.js && node dist/extensions/prompts.test.js && node dist/extensions/provider/status.test.js && node dist/extensions/search/index.test.js && node dist/extensions/search/proxy.test.js && node dist/extensions/search/host-integration.test.js && node dist/extensions/provider/status-integration.test.js && node dist/extensions/models/gateway-catalog-metadata.test.js && node dist/extensions/models/gateway-catalog-policy.test.js && node dist/extensions/models/operations.test.js && node dist/extensions/models/catalog.test.js && node dist/extensions/models/web-search.test.js && node dist/extensions/models/gateway-catalog.test.js && node dist/extensions/models/gateway-catalog-isolation.test.js && node dist/extensions/models/vendor-strategies.test.js && node dist/extensions/images/models.test.js && node dist/extensions/images/parameters.test.js && node dist/extensions/images/input-budget.test.js && node dist/extensions/images/edit-request.test.js && node dist/extensions/images/image-data.test.js && node dist/extensions/images/response.test.js && node dist/extensions/images/stream-response.test.js && node dist/extensions/images/adapter.test.js && node dist/extensions/images/tool-files.test.js && node dist/extensions/images/tools.test.js && node dist/extensions/images/host-integration.test.js && node dist/extensions/images/tools-host.test.js && node dist/index.test.js && node dist/extensions/skills.test.js && node dist/extensions/skill-resources.test.js && node dist/extensions/upload/protocol.test.js && node dist/extensions/upload/schema.test.js && node dist/extensions/upload/uploader.test.js && node dist/extensions/upload/transport.test.js && node dist/extensions/upload/host-integration.test.js && node dist/extensions/upload/readiness-host.test.js",
27
+ "test": "tsc -p tsconfig.test.json && node dist/extensions/provider/index.test.js && node dist/extensions/provider/refresh-tool.test.js && node dist/extensions/provider/startup-catalog.test.js && node dist/extensions/prompts.test.js && node dist/extensions/provider/status.test.js && node dist/extensions/search/index.test.js && node dist/extensions/search/proxy.test.js && node dist/extensions/search/host-integration.test.js && node dist/extensions/provider/status-integration.test.js && node dist/extensions/models/gateway-catalog-metadata.test.js && node dist/extensions/models/gateway-catalog-policy.test.js && node dist/extensions/models/operations.test.js && node dist/extensions/models/catalog.test.js && node dist/extensions/models/web-search.test.js && node dist/extensions/models/gateway-catalog.test.js && node dist/extensions/models/gateway-catalog-isolation.test.js && node dist/extensions/models/vendor-strategies.test.js && node dist/extensions/models/grok-new.test.js && node dist/extensions/models/gemini-new.test.js && node dist/extensions/models/longcat-new.test.js && node dist/extensions/images/models.test.js && node dist/extensions/images/parameters.test.js && node dist/extensions/images/input-budget.test.js && node dist/extensions/images/edit-request.test.js && node dist/extensions/images/image-data.test.js && node dist/extensions/images/response.test.js && node dist/extensions/images/stream-response.test.js && node dist/extensions/images/adapter.test.js && node dist/extensions/images/tool-files.test.js && node dist/extensions/images/tools.test.js && node dist/extensions/images/host-integration.test.js && node dist/extensions/images/tools-host.test.js && node dist/index.test.js && node dist/extensions/skills.test.js && node dist/extensions/skill-resources.test.js && node dist/extensions/upload/protocol.test.js && node dist/extensions/upload/schema.test.js && node dist/extensions/upload/uploader.test.js && node dist/extensions/upload/transport.test.js && node dist/extensions/upload/host-integration.test.js && node dist/extensions/upload/readiness-host.test.js",
28
28
  "test:upload:live": "tsc -p tsconfig.test.json && node dist/extensions/upload/live.test.js"
29
29
  },
30
30
  "peerDependencies": {
@@ -36,10 +36,16 @@ compatibility: 需要宿主提供子代理工具及可用的 tsgw 模型。
36
36
 
37
37
  - † MiniMax 的 29 分为模型级参考,Pi high 映射为 adaptive;待评项不借用其他型号或档位的分数。
38
38
  - GLM 5.3 三型号始终思考,实际 effort 为 low/high/max;Pi medium 映射 high,xhigh 映射 max。FlashX 可作未定档备选,评分待评,输入/输出参考价为 0.295/1.0325。
39
+ - `gpt-6.1-sol` 按用户确认的网关 Codex 接入采用 500000 tokens 上下文;最大输出仍为 128000,不与上下文窗口混淆。
39
40
  - GPT 价格对应输入 ≤272K,超过后输入 ×2、输出 ×1.5;MiniMax 输入超过 512K 为 0.60/2.40。
40
41
  - DeepSeek 采用国内价(2026-10-16 核验,Blinko #604):输入 1–2 元、输出 4–8 元/百万 tokens,按固定 0.1475 USD/CNY 折算;区间为低谷/高峰价。静态目录采用区间高端,不自动按时段切换;高峰为北京时间工作日(不含中国法定节假日)9–12、14–18。
41
42
  - GLM 依据 Blinko #929,同样以固定 0.1475 USD/CNY 折算国内人民币标价。其余数值沿用既有选型资料。
42
43
 
44
+ 新增目录候选 `grok-4.7`、`gemini-3.5-flash-lite`、`gemini-3.8-flash-high`、
45
+ `longcat-2.5-preview` 暂不进入上表排序:Grok 工具完整往返尚未通过,
46
+ LongCat 受上游额度阻塞;两款 Gemini 仅通过最小文本/工具往返,尚无本表同口径评分。
47
+ 登记不等于适合长期工程委派,不借用同族评分。
48
+
43
49
  ## 必要边界
44
50
 
45
51
  - **MiniMax-M3** 只承担步骤明确的小任务。参考使用 Pi 的 high;非 off 档位均映射为 adaptive,不代表不同能力档。adaptive 不是 Pi 档位,此映射不适用于 M3.1 Flash Preview。