pi-tsgw 0.0.6 → 0.0.8
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 +5 -0
- package/extensions/index.ts +108 -12
- package/extensions/models/gateway-catalog.ts +26 -10
- package/extensions/models/operations.ts +0 -2
- package/extensions/models/vendors/deepseek.ts +43 -39
- package/extensions/models/vendors/minimax.ts +11 -11
- package/extensions/models/vendors/openai.ts +7 -160
- package/extensions/models/web-search.ts +11 -5
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -85,6 +85,11 @@ global whitelist: models matching neither list remain available. To register
|
|
|
85
85
|
only selected models, use `"excludeModels": ["*"]` and pull them back through
|
|
86
86
|
`includeModels`.
|
|
87
87
|
|
|
88
|
+
Opening `/model` refreshes the gateway catalog in the background, with a
|
|
89
|
+
five-minute cache to avoid repeated requests. Run `/tsgw-refresh` to bypass the
|
|
90
|
+
cache and force an immediate refresh without restarting or reloading Pi. If a
|
|
91
|
+
refresh fails, the last successful model list remains active.
|
|
92
|
+
|
|
88
93
|
## Development
|
|
89
94
|
|
|
90
95
|
```bash
|
package/extensions/index.ts
CHANGED
|
@@ -1,11 +1,9 @@
|
|
|
1
1
|
import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
|
|
2
2
|
import {
|
|
3
|
-
CONFIG_DIR_NAME,
|
|
4
3
|
getAgentDir,
|
|
5
4
|
readStoredCredential,
|
|
6
5
|
} from "@earendil-works/pi-coding-agent";
|
|
7
6
|
import { readFileSync } from "node:fs";
|
|
8
|
-
import { homedir } from "node:os";
|
|
9
7
|
import { join } from "node:path";
|
|
10
8
|
import { randomUUID } from "node:crypto";
|
|
11
9
|
import {
|
|
@@ -124,9 +122,21 @@ function gatewayHasModel(gatewayIds: ReadonlySet<string>, id: string): boolean {
|
|
|
124
122
|
return false;
|
|
125
123
|
}
|
|
126
124
|
|
|
127
|
-
/**
|
|
125
|
+
/** 模型缓存文件:<agent-dir>/tsgw/models-cache.json。 */
|
|
128
126
|
function gatewayModelCachePath(): string {
|
|
129
|
-
return join(
|
|
127
|
+
return join(getAgentDir(), "tsgw", "models-cache.json");
|
|
128
|
+
}
|
|
129
|
+
|
|
130
|
+
function effectiveModelsForGateway(
|
|
131
|
+
root: string,
|
|
132
|
+
filter: ModelFilter | undefined,
|
|
133
|
+
gatewayIds: ReadonlySet<string> | undefined,
|
|
134
|
+
): ReturnType<typeof modelsForRoot> {
|
|
135
|
+
const staticModels = modelsForRoot(root);
|
|
136
|
+
const intersected = gatewayIds
|
|
137
|
+
? staticModels.filter((model) => gatewayHasModel(gatewayIds, model.id))
|
|
138
|
+
: staticModels;
|
|
139
|
+
return filterModels(intersected, filter);
|
|
130
140
|
}
|
|
131
141
|
|
|
132
142
|
/**
|
|
@@ -136,12 +146,11 @@ function gatewayModelCachePath(): string {
|
|
|
136
146
|
async function resolveEffectiveModels(
|
|
137
147
|
root: string,
|
|
138
148
|
filter: ModelFilter | undefined,
|
|
149
|
+
cacheFilePath: string,
|
|
139
150
|
): Promise<{
|
|
140
151
|
models: ReturnType<typeof modelsForRoot>;
|
|
141
152
|
gatewayIds: ReadonlySet<string> | undefined;
|
|
142
153
|
}> {
|
|
143
|
-
const staticModels = modelsForRoot(root);
|
|
144
|
-
|
|
145
154
|
// 凭据解析失败不阻塞:无凭据时直接跳过网关层。
|
|
146
155
|
let apiKey = "";
|
|
147
156
|
try {
|
|
@@ -155,24 +164,24 @@ async function resolveEffectiveModels(
|
|
|
155
164
|
if (apiKey) {
|
|
156
165
|
try {
|
|
157
166
|
const gateway = await getGatewayModelIds(root, apiKey, {
|
|
158
|
-
cacheFilePath
|
|
167
|
+
cacheFilePath,
|
|
159
168
|
});
|
|
160
169
|
if (gateway.ok) gatewayIds = new Set(gateway.ids);
|
|
161
170
|
else
|
|
162
171
|
console.warn(
|
|
163
172
|
`TSGW: gateway model list unavailable (${gateway.reason ?? "unknown"}); using static catalog.`,
|
|
164
173
|
);
|
|
165
|
-
} catch
|
|
174
|
+
} catch {
|
|
166
175
|
console.warn(
|
|
167
176
|
`TSGW: gateway model list load failed; using static catalog.`,
|
|
168
177
|
);
|
|
169
178
|
}
|
|
170
179
|
}
|
|
171
180
|
|
|
172
|
-
|
|
173
|
-
|
|
174
|
-
|
|
175
|
-
|
|
181
|
+
return {
|
|
182
|
+
models: effectiveModelsForGateway(root, filter, gatewayIds),
|
|
183
|
+
gatewayIds,
|
|
184
|
+
};
|
|
176
185
|
}
|
|
177
186
|
|
|
178
187
|
export default async function registerTsgw(pi: ExtensionAPI): Promise<void> {
|
|
@@ -187,6 +196,7 @@ export default async function registerTsgw(pi: ExtensionAPI): Promise<void> {
|
|
|
187
196
|
include: settings.includeModels,
|
|
188
197
|
exclude: settings.excludeModels,
|
|
189
198
|
};
|
|
199
|
+
const cacheFilePath = gatewayModelCachePath();
|
|
190
200
|
let requestState: RequestState | undefined;
|
|
191
201
|
|
|
192
202
|
const refreshRequestState = (ctx: {
|
|
@@ -200,13 +210,99 @@ export default async function registerTsgw(pi: ExtensionAPI): Promise<void> {
|
|
|
200
210
|
const { models, gatewayIds } = await resolveEffectiveModels(
|
|
201
211
|
root,
|
|
202
212
|
modelFilter,
|
|
213
|
+
cacheFilePath,
|
|
203
214
|
);
|
|
215
|
+
let latestModels = models;
|
|
204
216
|
pi.registerProvider(PROVIDER_ID, {
|
|
205
217
|
name: "TSGW",
|
|
206
218
|
baseUrl: `${root}/v1`,
|
|
207
219
|
api: "openai-completions",
|
|
208
220
|
apiKey: "$TSGW_API_KEY",
|
|
209
221
|
models,
|
|
222
|
+
async refreshModels(context) {
|
|
223
|
+
if (!context.allowNetwork) return latestModels;
|
|
224
|
+
context.signal.throwIfAborted();
|
|
225
|
+
const credential = context.credential;
|
|
226
|
+
if (credential?.type !== "api_key" || !credential.key)
|
|
227
|
+
throw new Error("TSGW API key is unavailable for model refresh.");
|
|
228
|
+
|
|
229
|
+
const gateway = await getGatewayModelIds(root, credential.key, {
|
|
230
|
+
cacheFilePath,
|
|
231
|
+
force: context.force === true,
|
|
232
|
+
signal: context.signal,
|
|
233
|
+
});
|
|
234
|
+
context.signal.throwIfAborted();
|
|
235
|
+
if (!gateway.ok)
|
|
236
|
+
throw new Error(
|
|
237
|
+
`TSGW gateway model refresh failed (${gateway.reason}).`,
|
|
238
|
+
);
|
|
239
|
+
if (gateway.stale)
|
|
240
|
+
throw new Error(
|
|
241
|
+
`TSGW gateway model refresh failed (${gateway.fallbackReason ?? "unknown"}); keeping the last successful catalog.`,
|
|
242
|
+
);
|
|
243
|
+
|
|
244
|
+
latestModels = effectiveModelsForGateway(
|
|
245
|
+
root,
|
|
246
|
+
modelFilter,
|
|
247
|
+
new Set(gateway.ids),
|
|
248
|
+
);
|
|
249
|
+
return latestModels;
|
|
250
|
+
},
|
|
251
|
+
});
|
|
252
|
+
|
|
253
|
+
pi.registerCommand("tsgw-refresh", {
|
|
254
|
+
description: "Force-refresh the TSGW gateway model catalog",
|
|
255
|
+
handler: async (_args, ctx) => {
|
|
256
|
+
await ctx.waitForIdle();
|
|
257
|
+
if (!ctx.modelRegistry.getProviderAuthStatus(PROVIDER_ID).configured) {
|
|
258
|
+
ctx.ui.notify(
|
|
259
|
+
"TSGW model refresh skipped: configure an API key with /login first.",
|
|
260
|
+
"warning",
|
|
261
|
+
);
|
|
262
|
+
return;
|
|
263
|
+
}
|
|
264
|
+
const count = (): number =>
|
|
265
|
+
ctx.modelRegistry
|
|
266
|
+
.getAvailable()
|
|
267
|
+
.filter((model) => model.provider === PROVIDER_ID).length;
|
|
268
|
+
const before = count();
|
|
269
|
+
try {
|
|
270
|
+
// Pi 0.84+ returns cancellation/errors; older supported Pi versions
|
|
271
|
+
// resolve void, so treat missing result metadata as success.
|
|
272
|
+
const result = (await ctx.modelRegistry.refresh({
|
|
273
|
+
providers: [PROVIDER_ID],
|
|
274
|
+
force: true,
|
|
275
|
+
})) as
|
|
276
|
+
| {
|
|
277
|
+
aborted?: boolean;
|
|
278
|
+
errors?: ReadonlyMap<string, Error>;
|
|
279
|
+
}
|
|
280
|
+
| undefined;
|
|
281
|
+
const after = count();
|
|
282
|
+
const error = result?.errors?.get(PROVIDER_ID);
|
|
283
|
+
if (result?.aborted) {
|
|
284
|
+
ctx.ui.notify(
|
|
285
|
+
`TSGW model refresh cancelled (${before} → ${after}).`,
|
|
286
|
+
"warning",
|
|
287
|
+
);
|
|
288
|
+
} else if (error) {
|
|
289
|
+
ctx.ui.notify(
|
|
290
|
+
`TSGW model refresh failed (${before} → ${after}): ${error.message}`,
|
|
291
|
+
"error",
|
|
292
|
+
);
|
|
293
|
+
} else {
|
|
294
|
+
ctx.ui.notify(
|
|
295
|
+
`TSGW models refreshed (${before} → ${after}).`,
|
|
296
|
+
"info",
|
|
297
|
+
);
|
|
298
|
+
}
|
|
299
|
+
} catch (error) {
|
|
300
|
+
ctx.ui.notify(
|
|
301
|
+
`TSGW model refresh failed (${before} → ${count()}): ${error instanceof Error ? error.message : String(error)}`,
|
|
302
|
+
"error",
|
|
303
|
+
);
|
|
304
|
+
}
|
|
305
|
+
},
|
|
210
306
|
});
|
|
211
307
|
|
|
212
308
|
// ts_search 条件注册:固定白名单后端与可用模型列表有交集才注册,
|
|
@@ -14,6 +14,7 @@ export const DEFAULT_GATEWAY_MODEL_CACHE_TTL_MS = 5 * 60 * 1_000;
|
|
|
14
14
|
export type GatewayModelFailureReason =
|
|
15
15
|
| "http"
|
|
16
16
|
| "timeout"
|
|
17
|
+
| "aborted"
|
|
17
18
|
| "network"
|
|
18
19
|
| "invalid-json"
|
|
19
20
|
| "invalid-data"
|
|
@@ -27,10 +28,12 @@ export type GatewayFetch = (
|
|
|
27
28
|
export interface FetchGatewayModelIdsOptions {
|
|
28
29
|
timeoutMs?: number;
|
|
29
30
|
fetcher?: GatewayFetch;
|
|
31
|
+
signal?: AbortSignal;
|
|
30
32
|
}
|
|
31
33
|
|
|
32
34
|
export interface GetGatewayModelIdsOptions extends FetchGatewayModelIdsOptions {
|
|
33
35
|
cacheFilePath?: string;
|
|
36
|
+
force?: boolean;
|
|
34
37
|
}
|
|
35
38
|
|
|
36
39
|
export type FetchGatewayModelIdsResult =
|
|
@@ -94,8 +97,12 @@ export async function fetchGatewayModelIds(
|
|
|
94
97
|
): Promise<FetchGatewayModelIdsResult> {
|
|
95
98
|
const timeoutMs = opts.timeoutMs ?? DEFAULT_GATEWAY_MODEL_TIMEOUT_MS;
|
|
96
99
|
const fetcher = opts.fetcher ?? globalThis.fetch;
|
|
100
|
+
if (opts.signal?.aborted) return { ok: false, reason: "aborted" };
|
|
101
|
+
|
|
97
102
|
const controller = new AbortController();
|
|
98
103
|
let timedOut = false;
|
|
104
|
+
const abortFromCaller = (): void => controller.abort(opts.signal?.reason);
|
|
105
|
+
opts.signal?.addEventListener("abort", abortFromCaller, { once: true });
|
|
99
106
|
const timeout = setTimeout(
|
|
100
107
|
() => {
|
|
101
108
|
timedOut = true;
|
|
@@ -113,9 +120,9 @@ export async function fetchGatewayModelIds(
|
|
|
113
120
|
signal: controller.signal,
|
|
114
121
|
});
|
|
115
122
|
} catch (error) {
|
|
116
|
-
return
|
|
117
|
-
|
|
118
|
-
|
|
123
|
+
if (timedOut) return { ok: false, reason: "timeout" };
|
|
124
|
+
if (opts.signal?.aborted) return { ok: false, reason: "aborted" };
|
|
125
|
+
return { ok: false, reason: "network", detail: errorDetail(error) };
|
|
119
126
|
}
|
|
120
127
|
|
|
121
128
|
if (!response.ok) {
|
|
@@ -126,9 +133,9 @@ export async function fetchGatewayModelIds(
|
|
|
126
133
|
try {
|
|
127
134
|
payload = await response.json();
|
|
128
135
|
} catch (error) {
|
|
129
|
-
return
|
|
130
|
-
|
|
131
|
-
|
|
136
|
+
if (timedOut) return { ok: false, reason: "timeout" };
|
|
137
|
+
if (opts.signal?.aborted) return { ok: false, reason: "aborted" };
|
|
138
|
+
return { ok: false, reason: "invalid-json", detail: errorDetail(error) };
|
|
132
139
|
}
|
|
133
140
|
|
|
134
141
|
if (!payload || typeof payload !== "object")
|
|
@@ -140,6 +147,7 @@ export async function fetchGatewayModelIds(
|
|
|
140
147
|
return { ok: true, ids };
|
|
141
148
|
} finally {
|
|
142
149
|
clearTimeout(timeout);
|
|
150
|
+
opts.signal?.removeEventListener("abort", abortFromCaller);
|
|
143
151
|
}
|
|
144
152
|
}
|
|
145
153
|
|
|
@@ -269,7 +277,7 @@ export async function getGatewayModelIds(
|
|
|
269
277
|
): Promise<GetGatewayModelIdsResult> {
|
|
270
278
|
const key = baseUrl.replace(/\/+$/, "");
|
|
271
279
|
const cache = cacheFor(key);
|
|
272
|
-
if (cache.isFresh(now)) {
|
|
280
|
+
if (!opts.force && cache.isFresh(now)) {
|
|
273
281
|
const entry = cache.get();
|
|
274
282
|
if (entry) return { ok: true, ids: entry.ids, cached: true, stale: false };
|
|
275
283
|
}
|
|
@@ -278,6 +286,7 @@ export async function getGatewayModelIds(
|
|
|
278
286
|
? readGatewayModelCacheEntry(opts.cacheFilePath)
|
|
279
287
|
: undefined;
|
|
280
288
|
if (
|
|
289
|
+
!opts.force &&
|
|
281
290
|
diskEntry &&
|
|
282
291
|
isCacheEntryFresh(diskEntry, now, DEFAULT_GATEWAY_MODEL_CACHE_TTL_MS)
|
|
283
292
|
) {
|
|
@@ -290,20 +299,27 @@ export async function getGatewayModelIds(
|
|
|
290
299
|
};
|
|
291
300
|
}
|
|
292
301
|
|
|
293
|
-
|
|
302
|
+
const requestKey = `${key}\0${opts.force ? "force" : "normal"}`;
|
|
303
|
+
// A shared request cannot safely inherit one caller's AbortSignal: Pi may
|
|
304
|
+
// supersede that refresh while a newer caller is still waiting. Only
|
|
305
|
+
// signal-free calls share in-flight work; abortable calls own their fetch.
|
|
306
|
+
const shareRequest = opts.signal === undefined;
|
|
307
|
+
let request = shareRequest ? inFlightRequests.get(requestKey) : undefined;
|
|
294
308
|
if (!request) {
|
|
295
309
|
request = fetchGatewayModelIds(key, apiKey, {
|
|
296
310
|
fetcher: opts.fetcher,
|
|
297
311
|
timeoutMs: opts.timeoutMs,
|
|
312
|
+
signal: opts.signal,
|
|
298
313
|
});
|
|
299
|
-
inFlightRequests.set(
|
|
314
|
+
if (shareRequest) inFlightRequests.set(requestKey, request);
|
|
300
315
|
}
|
|
301
316
|
|
|
302
317
|
let fetched: FetchGatewayModelIdsResult;
|
|
303
318
|
try {
|
|
304
319
|
fetched = await request;
|
|
305
320
|
} finally {
|
|
306
|
-
if (inFlightRequests.get(
|
|
321
|
+
if (shareRequest && inFlightRequests.get(requestKey) === request)
|
|
322
|
+
inFlightRequests.delete(requestKey);
|
|
307
323
|
}
|
|
308
324
|
|
|
309
325
|
if (fetched.ok) {
|
|
@@ -12,7 +12,6 @@ import {
|
|
|
12
12
|
type ThinkingApplier,
|
|
13
13
|
} from "./_tools.ts";
|
|
14
14
|
import { applyBuiltinSearchTool } from "./web-search.ts";
|
|
15
|
-
import { deepseekThinking } from "./vendors/deepseek.ts";
|
|
16
15
|
import { geminiThinking } from "./vendors/gemini.ts";
|
|
17
16
|
import { glmThinking } from "./vendors/glm.ts";
|
|
18
17
|
import { grokThinking } from "./vendors/grok.ts";
|
|
@@ -27,7 +26,6 @@ import { qwenThinking } from "./vendors/qwen.ts";
|
|
|
27
26
|
const TSGW_PROVIDER = "tsgw";
|
|
28
27
|
|
|
29
28
|
const THINKING_STRATEGIES: Record<string, ThinkingApplier> = {
|
|
30
|
-
...deepseekThinking,
|
|
31
29
|
...glmThinking,
|
|
32
30
|
...mimoThinking,
|
|
33
31
|
...minimaxThinking,
|
|
@@ -1,46 +1,65 @@
|
|
|
1
1
|
import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
|
|
2
|
-
import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
|
|
3
|
-
import { isMaximum, type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
|
|
4
2
|
|
|
5
|
-
// DeepSeek
|
|
6
|
-
//
|
|
7
|
-
const
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
requiresReasoningContentOnAssistantMessages: true,
|
|
3
|
+
// DeepSeek Responses 将 developer role 当作 user,且不支持 OpenAI 的
|
|
4
|
+
// prompt_cache_retention;其余 reasoning 字段交给 Pi 的 Responses adapter。
|
|
5
|
+
const DEEPSEEK_RESPONSES_COMPAT = {
|
|
6
|
+
supportsDeveloperRole: false,
|
|
7
|
+
supportsLongCacheRetention: false,
|
|
11
8
|
};
|
|
12
9
|
|
|
13
10
|
/**
|
|
14
11
|
* DeepSeek(深度求索)V4 系列。
|
|
15
|
-
* 官方文档:https://api-docs.deepseek.com/
|
|
12
|
+
* 官方文档:https://api-docs.deepseek.com/guides/responses_api
|
|
16
13
|
*/
|
|
17
14
|
export function deepseekModels(root: string): ProviderModelConfig[] {
|
|
18
15
|
const v1 = `${root}/v1`;
|
|
19
16
|
return [
|
|
17
|
+
{
|
|
18
|
+
id: "deepseek-flash",
|
|
19
|
+
name: "DeepSeek V4.1 Flash",
|
|
20
|
+
api: "openai-responses",
|
|
21
|
+
baseUrl: v1,
|
|
22
|
+
reasoning: true,
|
|
23
|
+
input: ["text", "image"],
|
|
24
|
+
cost: { input: 0.1475, output: 0.295, cacheRead: 0.00295, cacheWrite: 0 },
|
|
25
|
+
contextWindow: 1000000,
|
|
26
|
+
maxTokens: 384000,
|
|
27
|
+
compat: DEEPSEEK_RESPONSES_COMPAT,
|
|
28
|
+
thinkingLevelMap: {
|
|
29
|
+
off: "none",
|
|
30
|
+
minimal: "low",
|
|
31
|
+
low: "low",
|
|
32
|
+
medium: "high",
|
|
33
|
+
high: "high",
|
|
34
|
+
xhigh: "high",
|
|
35
|
+
max: "max",
|
|
36
|
+
},
|
|
37
|
+
},
|
|
20
38
|
{
|
|
21
39
|
id: "deepseek-v4-flash",
|
|
22
|
-
name: "DeepSeek V4 Flash",
|
|
23
|
-
api: "openai-
|
|
40
|
+
name: "DeepSeek V4 Flash (Legacy Alias)",
|
|
41
|
+
api: "openai-responses",
|
|
24
42
|
baseUrl: v1,
|
|
25
43
|
reasoning: true,
|
|
26
|
-
input: ["text"],
|
|
44
|
+
input: ["text", "image"],
|
|
27
45
|
cost: { input: 0.1475, output: 0.295, cacheRead: 0.00295, cacheWrite: 0 },
|
|
28
46
|
contextWindow: 1000000,
|
|
29
47
|
maxTokens: 384000,
|
|
30
|
-
compat:
|
|
48
|
+
compat: DEEPSEEK_RESPONSES_COMPAT,
|
|
31
49
|
thinkingLevelMap: {
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
|
|
50
|
+
off: "none",
|
|
51
|
+
minimal: "low",
|
|
52
|
+
low: "low",
|
|
53
|
+
medium: "high",
|
|
35
54
|
high: "high",
|
|
36
|
-
xhigh: "
|
|
55
|
+
xhigh: "high",
|
|
37
56
|
max: "max",
|
|
38
57
|
},
|
|
39
58
|
},
|
|
40
59
|
{
|
|
41
60
|
id: "deepseek-v4-pro",
|
|
42
61
|
name: "DeepSeek V4 Pro",
|
|
43
|
-
api: "openai-
|
|
62
|
+
api: "openai-responses",
|
|
44
63
|
baseUrl: v1,
|
|
45
64
|
reasoning: true,
|
|
46
65
|
input: ["text"],
|
|
@@ -52,31 +71,16 @@ export function deepseekModels(root: string): ProviderModelConfig[] {
|
|
|
52
71
|
},
|
|
53
72
|
contextWindow: 1000000,
|
|
54
73
|
maxTokens: 384000,
|
|
55
|
-
compat:
|
|
74
|
+
compat: DEEPSEEK_RESPONSES_COMPAT,
|
|
56
75
|
thinkingLevelMap: {
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
|
|
76
|
+
off: "none",
|
|
77
|
+
minimal: "low",
|
|
78
|
+
low: "low",
|
|
79
|
+
medium: "high",
|
|
60
80
|
high: "high",
|
|
61
|
-
xhigh: "
|
|
81
|
+
xhigh: "high",
|
|
62
82
|
max: "max",
|
|
63
83
|
},
|
|
64
84
|
},
|
|
65
85
|
];
|
|
66
86
|
}
|
|
67
|
-
|
|
68
|
-
// DeepSeek 思维链:`thinking.type` + `reasoning_effort`(high/max 两档)。
|
|
69
|
-
function applyDeepSeek(writer: PayloadWriter, level: ThinkingLevel): void {
|
|
70
|
-
writer.remove("reasoning_effort");
|
|
71
|
-
if (level === "off") {
|
|
72
|
-
writer.setThinking({ type: "disabled" });
|
|
73
|
-
return;
|
|
74
|
-
}
|
|
75
|
-
writer.setThinking({ type: "enabled" });
|
|
76
|
-
writer.set("reasoning_effort", isMaximum(level) ? "max" : "high");
|
|
77
|
-
}
|
|
78
|
-
|
|
79
|
-
export const deepseekThinking: Record<string, ThinkingApplier> = {
|
|
80
|
-
"deepseek-v4-flash": (w, c) => applyDeepSeek(w, c.thinkingLevel),
|
|
81
|
-
"deepseek-v4-pro": (w, c) => applyDeepSeek(w, c.thinkingLevel),
|
|
82
|
-
};
|
|
@@ -10,8 +10,8 @@ export function minimaxModels(root: string): ProviderModelConfig[] {
|
|
|
10
10
|
const v1 = `${root}/v1`;
|
|
11
11
|
return [
|
|
12
12
|
{
|
|
13
|
-
id: "
|
|
14
|
-
name: "
|
|
13
|
+
id: "MiniMax-M3",
|
|
14
|
+
name: "MiniMax-M3",
|
|
15
15
|
api: "openai-completions",
|
|
16
16
|
baseUrl: v1,
|
|
17
17
|
reasoning: true,
|
|
@@ -27,7 +27,7 @@ export function minimaxModels(root: string): ProviderModelConfig[] {
|
|
|
27
27
|
compat: OPENAI_COMPLETIONS_COMPAT,
|
|
28
28
|
},
|
|
29
29
|
{
|
|
30
|
-
id: "
|
|
30
|
+
id: "MiniMax-M2.7",
|
|
31
31
|
name: "MiniMax-M2.7",
|
|
32
32
|
api: "openai-completions",
|
|
33
33
|
baseUrl: v1,
|
|
@@ -40,7 +40,7 @@ export function minimaxModels(root: string): ProviderModelConfig[] {
|
|
|
40
40
|
compat: OPENAI_COMPLETIONS_COMPAT,
|
|
41
41
|
},
|
|
42
42
|
{
|
|
43
|
-
id: "
|
|
43
|
+
id: "MiniMax-M2.5",
|
|
44
44
|
name: "MiniMax-M2.5",
|
|
45
45
|
api: "openai-completions",
|
|
46
46
|
baseUrl: v1,
|
|
@@ -53,7 +53,7 @@ export function minimaxModels(root: string): ProviderModelConfig[] {
|
|
|
53
53
|
compat: OPENAI_COMPLETIONS_COMPAT,
|
|
54
54
|
},
|
|
55
55
|
{
|
|
56
|
-
id: "
|
|
56
|
+
id: "MiniMax-M2.1",
|
|
57
57
|
name: "MiniMax-M2.1",
|
|
58
58
|
api: "openai-completions",
|
|
59
59
|
baseUrl: v1,
|
|
@@ -76,9 +76,9 @@ function applyMiniMax(writer: PayloadWriter, level: ThinkingLevel): void {
|
|
|
76
76
|
writer.setThinking({ type: level === "off" ? "disabled" : "adaptive" });
|
|
77
77
|
}
|
|
78
78
|
|
|
79
|
-
export const minimaxThinking
|
|
80
|
-
"
|
|
81
|
-
"
|
|
82
|
-
"
|
|
83
|
-
"
|
|
84
|
-
}
|
|
79
|
+
export const minimaxThinking = {
|
|
80
|
+
"MiniMax-M3": (w, c) => applyMiniMax(w, c.thinkingLevel),
|
|
81
|
+
"MiniMax-M2.7": (w, c) => applyMiniMax(w, c.thinkingLevel),
|
|
82
|
+
"MiniMax-M2.5": (w, c) => applyMiniMax(w, c.thinkingLevel),
|
|
83
|
+
"MiniMax-M2.1": (w, c) => applyMiniMax(w, c.thinkingLevel),
|
|
84
|
+
} satisfies Record<string, ThinkingApplier>;
|
|
@@ -9,9 +9,8 @@ export function openaiModels(root: string): ProviderModelConfig[] {
|
|
|
9
9
|
const v1 = `${root}/v1`;
|
|
10
10
|
return [
|
|
11
11
|
{
|
|
12
|
-
|
|
13
|
-
|
|
14
|
-
name: "日站专用",
|
|
12
|
+
id: "gpt-6-astra",
|
|
13
|
+
name: "GPT-6 Astra",
|
|
15
14
|
api: "openai-responses",
|
|
16
15
|
baseUrl: v1,
|
|
17
16
|
reasoning: true,
|
|
@@ -23,8 +22,8 @@ export function openaiModels(root: string): ProviderModelConfig[] {
|
|
|
23
22
|
max: "max",
|
|
24
23
|
},
|
|
25
24
|
input: ["text", "image"],
|
|
26
|
-
cost: { input:
|
|
27
|
-
contextWindow:
|
|
25
|
+
cost: { input: 10, output: 50, cacheRead: 1, cacheWrite: 12.5 },
|
|
26
|
+
contextWindow: 1050000,
|
|
28
27
|
maxTokens: 128000,
|
|
29
28
|
},
|
|
30
29
|
{
|
|
@@ -98,150 +97,6 @@ export function openaiModels(root: string): ProviderModelConfig[] {
|
|
|
98
97
|
contextWindow: 272000,
|
|
99
98
|
maxTokens: 128000,
|
|
100
99
|
},
|
|
101
|
-
{
|
|
102
|
-
id: "gpt-5.3-codex-spark",
|
|
103
|
-
name: "GPT-5.3 Codex Spark",
|
|
104
|
-
api: "openai-responses",
|
|
105
|
-
baseUrl: v1,
|
|
106
|
-
reasoning: true,
|
|
107
|
-
thinkingLevelMap: {
|
|
108
|
-
low: "low",
|
|
109
|
-
medium: "medium",
|
|
110
|
-
high: "high",
|
|
111
|
-
xhigh: "xhigh",
|
|
112
|
-
},
|
|
113
|
-
input: ["text"],
|
|
114
|
-
cost: { input: 1.75, output: 14, cacheRead: 0.175, cacheWrite: 0 },
|
|
115
|
-
contextWindow: 128000,
|
|
116
|
-
maxTokens: 32000,
|
|
117
|
-
},
|
|
118
|
-
{
|
|
119
|
-
id: "gpt-5.6",
|
|
120
|
-
name: "GPT-5.6",
|
|
121
|
-
api: "openai-responses",
|
|
122
|
-
baseUrl: v1,
|
|
123
|
-
reasoning: true,
|
|
124
|
-
thinkingLevelMap: {
|
|
125
|
-
low: "low",
|
|
126
|
-
medium: "medium",
|
|
127
|
-
high: "high",
|
|
128
|
-
xhigh: "xhigh",
|
|
129
|
-
max: "max",
|
|
130
|
-
},
|
|
131
|
-
input: ["text", "image"],
|
|
132
|
-
cost: { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 6.25 },
|
|
133
|
-
contextWindow: 1050000,
|
|
134
|
-
maxTokens: 128000,
|
|
135
|
-
},
|
|
136
|
-
{
|
|
137
|
-
id: "gpt-5.4",
|
|
138
|
-
name: "GPT-5.4",
|
|
139
|
-
api: "openai-responses",
|
|
140
|
-
baseUrl: v1,
|
|
141
|
-
reasoning: true,
|
|
142
|
-
thinkingLevelMap: {
|
|
143
|
-
low: "low",
|
|
144
|
-
medium: "medium",
|
|
145
|
-
high: "high",
|
|
146
|
-
xhigh: "xhigh",
|
|
147
|
-
max: "max",
|
|
148
|
-
},
|
|
149
|
-
input: ["text", "image"],
|
|
150
|
-
cost: { input: 2.5, output: 15, cacheRead: 0.25, cacheWrite: 0 },
|
|
151
|
-
contextWindow: 1050000,
|
|
152
|
-
maxTokens: 128000,
|
|
153
|
-
},
|
|
154
|
-
{
|
|
155
|
-
id: "gpt-5.4-mini",
|
|
156
|
-
name: "GPT-5.4 mini",
|
|
157
|
-
api: "openai-responses",
|
|
158
|
-
baseUrl: v1,
|
|
159
|
-
reasoning: true,
|
|
160
|
-
thinkingLevelMap: {
|
|
161
|
-
low: "low",
|
|
162
|
-
medium: "medium",
|
|
163
|
-
high: "high",
|
|
164
|
-
xhigh: "xhigh",
|
|
165
|
-
max: "max",
|
|
166
|
-
},
|
|
167
|
-
input: ["text", "image"],
|
|
168
|
-
cost: { input: 0.75, output: 4.5, cacheRead: 0.075, cacheWrite: 0 },
|
|
169
|
-
contextWindow: 400000,
|
|
170
|
-
maxTokens: 128000,
|
|
171
|
-
},
|
|
172
|
-
{
|
|
173
|
-
id: "gpt-5.2",
|
|
174
|
-
name: "GPT-5.2",
|
|
175
|
-
api: "openai-responses",
|
|
176
|
-
baseUrl: v1,
|
|
177
|
-
reasoning: true,
|
|
178
|
-
thinkingLevelMap: {
|
|
179
|
-
low: "low",
|
|
180
|
-
medium: "medium",
|
|
181
|
-
high: "high",
|
|
182
|
-
xhigh: "xhigh",
|
|
183
|
-
max: "max",
|
|
184
|
-
},
|
|
185
|
-
input: ["text", "image"],
|
|
186
|
-
cost: { input: 1.75, output: 14, cacheRead: 0.175, cacheWrite: 0 },
|
|
187
|
-
contextWindow: 400000,
|
|
188
|
-
maxTokens: 128000,
|
|
189
|
-
},
|
|
190
|
-
{
|
|
191
|
-
id: "gpt-5.2-pro",
|
|
192
|
-
name: "GPT-5.2 Pro",
|
|
193
|
-
api: "openai-responses",
|
|
194
|
-
baseUrl: v1,
|
|
195
|
-
reasoning: true,
|
|
196
|
-
thinkingLevelMap: {
|
|
197
|
-
low: "low",
|
|
198
|
-
medium: "medium",
|
|
199
|
-
high: "high",
|
|
200
|
-
xhigh: "xhigh",
|
|
201
|
-
max: "max",
|
|
202
|
-
},
|
|
203
|
-
input: ["text", "image"],
|
|
204
|
-
// 官方未公布缓存价,置 0。
|
|
205
|
-
cost: { input: 21, output: 168, cacheRead: 0, cacheWrite: 0 },
|
|
206
|
-
contextWindow: 400000,
|
|
207
|
-
maxTokens: 128000,
|
|
208
|
-
},
|
|
209
|
-
{
|
|
210
|
-
id: "gpt-5.1",
|
|
211
|
-
name: "GPT-5.1",
|
|
212
|
-
api: "openai-responses",
|
|
213
|
-
baseUrl: v1,
|
|
214
|
-
reasoning: true,
|
|
215
|
-
thinkingLevelMap: {
|
|
216
|
-
low: "low",
|
|
217
|
-
medium: "medium",
|
|
218
|
-
high: "high",
|
|
219
|
-
xhigh: "xhigh",
|
|
220
|
-
max: "max",
|
|
221
|
-
},
|
|
222
|
-
input: ["text", "image"],
|
|
223
|
-
cost: { input: 1.25, output: 10, cacheRead: 0.125, cacheWrite: 0 },
|
|
224
|
-
contextWindow: 400000,
|
|
225
|
-
maxTokens: 128000,
|
|
226
|
-
},
|
|
227
|
-
{
|
|
228
|
-
id: "gpt-5",
|
|
229
|
-
name: "GPT-5",
|
|
230
|
-
api: "openai-responses",
|
|
231
|
-
baseUrl: v1,
|
|
232
|
-
reasoning: true,
|
|
233
|
-
thinkingLevelMap: {
|
|
234
|
-
low: "low",
|
|
235
|
-
medium: "medium",
|
|
236
|
-
high: "high",
|
|
237
|
-
xhigh: "xhigh",
|
|
238
|
-
max: "max",
|
|
239
|
-
},
|
|
240
|
-
input: ["text", "image"],
|
|
241
|
-
cost: { input: 1.25, output: 10, cacheRead: 0.125, cacheWrite: 0 },
|
|
242
|
-
contextWindow: 400000,
|
|
243
|
-
maxTokens: 128000,
|
|
244
|
-
},
|
|
245
100
|
];
|
|
246
101
|
}
|
|
247
102
|
|
|
@@ -253,18 +108,10 @@ function applyOpenAIResponses(writer: PayloadWriter, modelId: string): void {
|
|
|
253
108
|
writer.setTextVerbosity(verbosity);
|
|
254
109
|
}
|
|
255
110
|
|
|
256
|
-
export const openaiThinking
|
|
257
|
-
"
|
|
111
|
+
export const openaiThinking = {
|
|
112
|
+
"gpt-6-astra": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
258
113
|
"gpt-5.6-sol": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
259
114
|
"gpt-5.6-terra": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
260
115
|
"gpt-5.6-luna": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
261
116
|
"gpt-5.5": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
262
|
-
|
|
263
|
-
"gpt-5.6": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
264
|
-
"gpt-5.4": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
265
|
-
"gpt-5.4-mini": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
266
|
-
"gpt-5.2": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
267
|
-
"gpt-5.2-pro": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
268
|
-
"gpt-5.1": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
269
|
-
"gpt-5": (w, c) => applyOpenAIResponses(w, c.modelId),
|
|
270
|
-
};
|
|
117
|
+
} satisfies Record<string, ThinkingApplier>;
|
|
@@ -12,13 +12,19 @@
|
|
|
12
12
|
* 独立的 ts_search 工具(所有模型可用)单独实现,与内置查询注入分离。
|
|
13
13
|
*/
|
|
14
14
|
|
|
15
|
-
import {
|
|
15
|
+
import {
|
|
16
|
+
isPlainObject,
|
|
17
|
+
type Payload,
|
|
18
|
+
PayloadWriter,
|
|
19
|
+
type WebSearchMode,
|
|
20
|
+
} from "./_tools.ts";
|
|
16
21
|
|
|
17
22
|
const OPENAI_RESPONSES = "openai-responses";
|
|
18
23
|
const OPENAI_COMPLETIONS = "openai-completions";
|
|
19
24
|
|
|
20
25
|
// 支持内置查询的模型名单(GPT Responses 系列)。
|
|
21
26
|
const BUILTIN_SEARCH_MODELS = new Set([
|
|
27
|
+
"gpt-6-astra",
|
|
22
28
|
"gpt-5.6-sol",
|
|
23
29
|
"gpt-5.6-terra",
|
|
24
30
|
"gpt-5.6-luna",
|
|
@@ -42,7 +48,7 @@ function isExistingWebSearchTool(tools: readonly unknown[]): boolean {
|
|
|
42
48
|
});
|
|
43
49
|
}
|
|
44
50
|
|
|
45
|
-
function hasGrokSearchFields(payload:
|
|
51
|
+
function hasGrokSearchFields(payload: Payload): boolean {
|
|
46
52
|
return (
|
|
47
53
|
Object.hasOwn(payload, "search_parameters") ||
|
|
48
54
|
Object.hasOwn(payload, "web_search_options")
|
|
@@ -76,12 +82,12 @@ function applyGrokSearch(writer: PayloadWriter): void {
|
|
|
76
82
|
* preserved. Pi drops requested sources, so GPT injection deliberately does
|
|
77
83
|
* not request them.
|
|
78
84
|
*/
|
|
79
|
-
export function applyBuiltinSearchTool(
|
|
80
|
-
payload:
|
|
85
|
+
export function applyBuiltinSearchTool<T>(
|
|
86
|
+
payload: T,
|
|
81
87
|
modelId: string,
|
|
82
88
|
api: string,
|
|
83
89
|
mode: WebSearchMode,
|
|
84
|
-
):
|
|
90
|
+
): T | Payload {
|
|
85
91
|
if (!isPlainObject(payload) || mode === "off") return payload;
|
|
86
92
|
const writer = new PayloadWriter(payload);
|
|
87
93
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "pi-tsgw",
|
|
3
|
-
"version": "0.0.
|
|
3
|
+
"version": "0.0.8",
|
|
4
4
|
"description": "Pi extension for the TOSSP AIH gateway: per-vendor model catalog and thinking operations for openai-completions, openai-responses, anthropic-messages, and google-generative-ai wire protocols.",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"keywords": [
|