pi-tsgw 0.0.5 → 0.0.7

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
@@ -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
@@ -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
- /** 模型缓存文件:~/.pi/tsgw/models-cache.json(CONFIG_DIR_NAME 跟随 piConfig 覆盖)。 */
125
+ /** 模型缓存文件:<agent-dir>/tsgw/models-cache.json。 */
128
126
  function gatewayModelCachePath(): string {
129
- return join(homedir(), CONFIG_DIR_NAME, "tsgw", "models-cache.json");
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: gatewayModelCachePath(),
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 (error) {
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
- const intersected = gatewayIds
173
- ? staticModels.filter((model) => gatewayHasModel(gatewayIds, model.id))
174
- : staticModels;
175
- return { models: filterModels(intersected, filter), gatewayIds };
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 timedOut
117
- ? { ok: false, reason: "timeout" }
118
- : { ok: false, reason: "network", detail: errorDetail(error) };
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 timedOut
130
- ? { ok: false, reason: "timeout" }
131
- : { ok: false, reason: "invalid-json", detail: errorDetail(error) };
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
- let request = inFlightRequests.get(key);
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(key, request);
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(key) === request) inFlightRequests.delete(key);
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,
@@ -4,6 +4,7 @@ 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
8
  */
8
9
  export function anthropicModels(root: string): ProviderModelConfig[] {
9
10
  const anthropic = `${root}/anthropic`;
@@ -27,8 +28,8 @@ export function anthropicModels(root: string): ProviderModelConfig[] {
27
28
  compat: ANTHROPIC_COMPAT,
28
29
  },
29
30
  {
30
- id: "claude-opus-4-8",
31
- name: "Claude Opus 4.8",
31
+ id: "claude-opus-5",
32
+ name: "Claude Opus 5",
32
33
  api: "anthropic-messages",
33
34
  baseUrl: anthropic,
34
35
  reasoning: true,
@@ -36,158 +37,15 @@ export function anthropicModels(root: string): ProviderModelConfig[] {
36
37
  low: "low",
37
38
  medium: "medium",
38
39
  high: "high",
40
+ xhigh: "xhigh",
39
41
  max: "max",
40
42
  },
41
43
  input: ["text", "image"],
42
- cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
43
- contextWindow: 200000,
44
- maxTokens: 128000,
45
- compat: ANTHROPIC_COMPAT,
46
- },
47
- {
48
- id: "claude-sonnet",
49
- name: "Claude Sonnet 4.6",
50
- api: "anthropic-messages",
51
- baseUrl: anthropic,
52
- reasoning: true,
53
- thinkingLevelMap: {
54
- low: "low",
55
- medium: "medium",
56
- high: "high",
57
- max: "max",
58
- },
59
- input: ["text", "image"],
60
- cost: { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3.75 },
61
- contextWindow: 200000,
62
- maxTokens: 64000,
63
- compat: ANTHROPIC_COMPAT,
64
- },
65
- {
66
- id: "claude-opus-4",
67
- name: "Claude Opus 4",
68
- api: "anthropic-messages",
69
- baseUrl: anthropic,
70
- reasoning: true,
71
- thinkingLevelMap: {
72
- low: "low",
73
- medium: "medium",
74
- high: "high",
75
- max: "max",
76
- },
77
- input: ["text", "image"],
78
- cost: { input: 15, output: 75, cacheRead: 1.5, cacheWrite: 18.75 },
79
- contextWindow: 200000,
80
- maxTokens: 32000,
81
- compat: ANTHROPIC_COMPAT,
82
- },
83
- {
84
- id: "claude-opus-4.1",
85
- name: "Claude Opus 4.1",
86
- api: "anthropic-messages",
87
- baseUrl: anthropic,
88
- reasoning: true,
89
- thinkingLevelMap: {
90
- low: "low",
91
- medium: "medium",
92
- high: "high",
93
- max: "max",
94
- },
95
- input: ["text", "image"],
96
- cost: { input: 15, output: 75, cacheRead: 1.5, cacheWrite: 18.75 },
97
- contextWindow: 200000,
98
- maxTokens: 32000,
99
- compat: ANTHROPIC_COMPAT,
100
- },
101
- {
102
- id: "claude-opus-4.5",
103
- name: "Claude Opus 4.5",
104
- api: "anthropic-messages",
105
- baseUrl: anthropic,
106
- reasoning: true,
107
- thinkingLevelMap: {
108
- low: "low",
109
- medium: "medium",
110
- high: "high",
111
- max: "max",
112
- },
113
- input: ["text", "image"],
114
- cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
115
- contextWindow: 200000,
116
- maxTokens: 64000,
117
- compat: ANTHROPIC_COMPAT,
118
- },
119
- {
120
- id: "claude-opus-4.6",
121
- name: "Claude Opus 4.6",
122
- api: "anthropic-messages",
123
- baseUrl: anthropic,
124
- reasoning: true,
125
- thinkingLevelMap: {
126
- low: "low",
127
- medium: "medium",
128
- high: "high",
129
- max: "max",
130
- },
131
- input: ["text", "image"],
132
- cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
133
- contextWindow: 1000000,
134
- maxTokens: 128000,
135
- compat: ANTHROPIC_COMPAT,
136
- },
137
- {
138
- id: "claude-opus-4.7",
139
- name: "Claude Opus 4.7",
140
- api: "anthropic-messages",
141
- baseUrl: anthropic,
142
- reasoning: true,
143
- thinkingLevelMap: {
144
- low: "low",
145
- medium: "medium",
146
- high: "high",
147
- max: "max",
148
- },
149
- input: ["text", "image"],
150
- cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
44
+ cost: { input: 10, output: 50, cacheRead: 1, cacheWrite: 12.5 },
151
45
  contextWindow: 1000000,
152
46
  maxTokens: 128000,
153
47
  compat: ANTHROPIC_COMPAT,
154
48
  },
155
- {
156
- id: "claude-sonnet-4",
157
- name: "Claude Sonnet 4",
158
- api: "anthropic-messages",
159
- baseUrl: anthropic,
160
- reasoning: true,
161
- thinkingLevelMap: {
162
- low: "low",
163
- medium: "medium",
164
- high: "high",
165
- max: "max",
166
- },
167
- input: ["text", "image"],
168
- cost: { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3.75 },
169
- contextWindow: 200000,
170
- maxTokens: 64000,
171
- compat: ANTHROPIC_COMPAT,
172
- },
173
- {
174
- id: "claude-sonnet-4.5",
175
- name: "Claude Sonnet 4.5",
176
- api: "anthropic-messages",
177
- baseUrl: anthropic,
178
- reasoning: true,
179
- thinkingLevelMap: {
180
- low: "low",
181
- medium: "medium",
182
- high: "high",
183
- max: "max",
184
- },
185
- input: ["text", "image"],
186
- cost: { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3.75 },
187
- contextWindow: 200000,
188
- maxTokens: 64000,
189
- compat: ANTHROPIC_COMPAT,
190
- },
191
49
  {
192
50
  id: "claude-sonnet-5",
193
51
  name: "Claude Sonnet 5",
@@ -198,11 +56,11 @@ export function anthropicModels(root: string): ProviderModelConfig[] {
198
56
  low: "low",
199
57
  medium: "medium",
200
58
  high: "high",
59
+ xhigh: "xhigh",
201
60
  max: "max",
202
61
  },
203
62
  input: ["text", "image"],
204
- // 官方 introductory 定价(至 2026-08-31),其后为 $3/$15。
205
- cost: { input: 2, output: 10, cacheRead: 0.2, cacheWrite: 2.5 },
63
+ cost: { input: 6, output: 30, cacheRead: 0.6, cacheWrite: 7.5 },
206
64
  contextWindow: 1000000,
207
65
  maxTokens: 128000,
208
66
  compat: ANTHROPIC_COMPAT,
@@ -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 在 openai-completions 协议基础上,要求 DeepSeek 特有 thinking 格式,
6
- // 且多轮(尤其工具调用)必须保留 reasoning_content。
7
- const DEEPSEEK_COMPAT = {
8
- ...OPENAI_COMPLETIONS_COMPAT,
9
- thinkingFormat: "deepseek" as const,
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-completions",
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: DEEPSEEK_COMPAT,
48
+ compat: DEEPSEEK_RESPONSES_COMPAT,
31
49
  thinkingLevelMap: {
32
- minimal: null,
33
- low: null,
34
- medium: null,
50
+ off: "none",
51
+ minimal: "low",
52
+ low: "low",
53
+ medium: "high",
35
54
  high: "high",
36
- xhigh: "max",
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-completions",
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: DEEPSEEK_COMPAT,
74
+ compat: DEEPSEEK_RESPONSES_COMPAT,
56
75
  thinkingLevelMap: {
57
- minimal: null,
58
- low: null,
59
- medium: null,
76
+ off: "none",
77
+ minimal: "low",
78
+ low: "low",
79
+ medium: "high",
60
80
  high: "high",
61
- xhigh: "max",
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
- };
@@ -9,8 +9,8 @@ export function openaiModels(root: string): ProviderModelConfig[] {
9
9
  const v1 = `${root}/v1`;
10
10
  return [
11
11
  {
12
- id: "gpt-5.6-sol",
13
- name: "GPT 5.6 Sol",
12
+ id: "gpt-6-astra",
13
+ name: "GPT-6 Astra",
14
14
  api: "openai-responses",
15
15
  baseUrl: v1,
16
16
  reasoning: true,
@@ -22,13 +22,13 @@ export function openaiModels(root: string): ProviderModelConfig[] {
22
22
  max: "max",
23
23
  },
24
24
  input: ["text", "image"],
25
- cost: { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 6.25 },
26
- contextWindow: 372000,
25
+ cost: { input: 10, output: 50, cacheRead: 1, cacheWrite: 12.5 },
26
+ contextWindow: 1050000,
27
27
  maxTokens: 128000,
28
28
  },
29
29
  {
30
- id: "gpt-5.6-terra",
31
- name: "GPT 5.6 Terra",
30
+ id: "gpt-5.6-sol",
31
+ name: "GPT 5.6 Sol",
32
32
  api: "openai-responses",
33
33
  baseUrl: v1,
34
34
  reasoning: true,
@@ -40,13 +40,13 @@ export function openaiModels(root: string): ProviderModelConfig[] {
40
40
  max: "max",
41
41
  },
42
42
  input: ["text", "image"],
43
- cost: { input: 2.5, output: 15, cacheRead: 0.25, cacheWrite: 3.125 },
43
+ cost: { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 6.25 },
44
44
  contextWindow: 372000,
45
45
  maxTokens: 128000,
46
46
  },
47
47
  {
48
- id: "gpt-5.6-luna",
49
- name: "GPT 5.6 Luna",
48
+ id: "gpt-5.6-terra",
49
+ name: "GPT 5.6 Terra",
50
50
  api: "openai-responses",
51
51
  baseUrl: v1,
52
52
  reasoning: true,
@@ -58,138 +58,13 @@ export function openaiModels(root: string): ProviderModelConfig[] {
58
58
  max: "max",
59
59
  },
60
60
  input: ["text", "image"],
61
- cost: { input: 1, output: 6, cacheRead: 0.01, cacheWrite: 1.25 },
61
+ cost: { input: 2.5, output: 15, cacheRead: 0.25, cacheWrite: 3.125 },
62
62
  contextWindow: 372000,
63
63
  maxTokens: 128000,
64
64
  },
65
65
  {
66
- id: "gpt-5.5",
67
- name: "GPT-5.5",
68
- api: "openai-responses",
69
- baseUrl: v1,
70
- reasoning: true,
71
- thinkingLevelMap: {
72
- low: "low",
73
- medium: "medium",
74
- high: "high",
75
- xhigh: "xhigh",
76
- },
77
- input: ["text", "image"],
78
- cost: { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 0 },
79
- contextWindow: 272000,
80
- maxTokens: 128000,
81
- },
82
- {
83
- id: "gpt-5.3-codex-spark",
84
- name: "GPT-5.3 Codex Spark",
85
- api: "openai-responses",
86
- baseUrl: v1,
87
- reasoning: true,
88
- thinkingLevelMap: {
89
- low: "low",
90
- medium: "medium",
91
- high: "high",
92
- xhigh: "xhigh",
93
- },
94
- input: ["text"],
95
- cost: { input: 1.75, output: 14, cacheRead: 0.175, cacheWrite: 0 },
96
- contextWindow: 128000,
97
- maxTokens: 32000,
98
- },
99
- {
100
- id: "gpt-5.6",
101
- name: "GPT-5.6",
102
- api: "openai-responses",
103
- baseUrl: v1,
104
- reasoning: true,
105
- thinkingLevelMap: {
106
- low: "low",
107
- medium: "medium",
108
- high: "high",
109
- xhigh: "xhigh",
110
- max: "max",
111
- },
112
- input: ["text", "image"],
113
- cost: { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 6.25 },
114
- contextWindow: 1050000,
115
- maxTokens: 128000,
116
- },
117
- {
118
- id: "gpt-5.4",
119
- name: "GPT-5.4",
120
- api: "openai-responses",
121
- baseUrl: v1,
122
- reasoning: true,
123
- thinkingLevelMap: {
124
- low: "low",
125
- medium: "medium",
126
- high: "high",
127
- xhigh: "xhigh",
128
- max: "max",
129
- },
130
- input: ["text", "image"],
131
- cost: { input: 2.5, output: 15, cacheRead: 0.25, cacheWrite: 0 },
132
- contextWindow: 1050000,
133
- maxTokens: 128000,
134
- },
135
- {
136
- id: "gpt-5.4-mini",
137
- name: "GPT-5.4 mini",
138
- api: "openai-responses",
139
- baseUrl: v1,
140
- reasoning: true,
141
- thinkingLevelMap: {
142
- low: "low",
143
- medium: "medium",
144
- high: "high",
145
- xhigh: "xhigh",
146
- max: "max",
147
- },
148
- input: ["text", "image"],
149
- cost: { input: 0.75, output: 4.5, cacheRead: 0.075, cacheWrite: 0 },
150
- contextWindow: 400000,
151
- maxTokens: 128000,
152
- },
153
- {
154
- id: "gpt-5.2",
155
- name: "GPT-5.2",
156
- api: "openai-responses",
157
- baseUrl: v1,
158
- reasoning: true,
159
- thinkingLevelMap: {
160
- low: "low",
161
- medium: "medium",
162
- high: "high",
163
- xhigh: "xhigh",
164
- max: "max",
165
- },
166
- input: ["text", "image"],
167
- cost: { input: 1.75, output: 14, cacheRead: 0.175, cacheWrite: 0 },
168
- contextWindow: 400000,
169
- maxTokens: 128000,
170
- },
171
- {
172
- id: "gpt-5.2-pro",
173
- name: "GPT-5.2 Pro",
174
- api: "openai-responses",
175
- baseUrl: v1,
176
- reasoning: true,
177
- thinkingLevelMap: {
178
- low: "low",
179
- medium: "medium",
180
- high: "high",
181
- xhigh: "xhigh",
182
- max: "max",
183
- },
184
- input: ["text", "image"],
185
- // 官方未公布缓存价,置 0。
186
- cost: { input: 21, output: 168, cacheRead: 0, cacheWrite: 0 },
187
- contextWindow: 400000,
188
- maxTokens: 128000,
189
- },
190
- {
191
- id: "gpt-5.1",
192
- name: "GPT-5.1",
66
+ id: "gpt-5.6-luna",
67
+ name: "GPT 5.6 Luna",
193
68
  api: "openai-responses",
194
69
  baseUrl: v1,
195
70
  reasoning: true,
@@ -201,13 +76,13 @@ export function openaiModels(root: string): ProviderModelConfig[] {
201
76
  max: "max",
202
77
  },
203
78
  input: ["text", "image"],
204
- cost: { input: 1.25, output: 10, cacheRead: 0.125, cacheWrite: 0 },
205
- contextWindow: 400000,
79
+ cost: { input: 1, output: 6, cacheRead: 0.01, cacheWrite: 1.25 },
80
+ contextWindow: 372000,
206
81
  maxTokens: 128000,
207
82
  },
208
83
  {
209
- id: "gpt-5",
210
- name: "GPT-5",
84
+ id: "gpt-5.5",
85
+ name: "GPT-5.5",
211
86
  api: "openai-responses",
212
87
  baseUrl: v1,
213
88
  reasoning: true,
@@ -216,11 +91,10 @@ export function openaiModels(root: string): ProviderModelConfig[] {
216
91
  medium: "medium",
217
92
  high: "high",
218
93
  xhigh: "xhigh",
219
- max: "max",
220
94
  },
221
95
  input: ["text", "image"],
222
- cost: { input: 1.25, output: 10, cacheRead: 0.125, cacheWrite: 0 },
223
- contextWindow: 400000,
96
+ cost: { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 0 },
97
+ contextWindow: 272000,
224
98
  maxTokens: 128000,
225
99
  },
226
100
  ];
@@ -234,17 +108,10 @@ function applyOpenAIResponses(writer: PayloadWriter, modelId: string): void {
234
108
  writer.setTextVerbosity(verbosity);
235
109
  }
236
110
 
237
- export const openaiThinking: Record<string, ThinkingApplier> = {
111
+ export const openaiThinking = {
112
+ "gpt-6-astra": (w, c) => applyOpenAIResponses(w, c.modelId),
238
113
  "gpt-5.6-sol": (w, c) => applyOpenAIResponses(w, c.modelId),
239
114
  "gpt-5.6-terra": (w, c) => applyOpenAIResponses(w, c.modelId),
240
115
  "gpt-5.6-luna": (w, c) => applyOpenAIResponses(w, c.modelId),
241
116
  "gpt-5.5": (w, c) => applyOpenAIResponses(w, c.modelId),
242
- "gpt-5.3-codex-spark": (w, c) => applyOpenAIResponses(w, c.modelId),
243
- "gpt-5.6": (w, c) => applyOpenAIResponses(w, c.modelId),
244
- "gpt-5.4": (w, c) => applyOpenAIResponses(w, c.modelId),
245
- "gpt-5.4-mini": (w, c) => applyOpenAIResponses(w, c.modelId),
246
- "gpt-5.2": (w, c) => applyOpenAIResponses(w, c.modelId),
247
- "gpt-5.2-pro": (w, c) => applyOpenAIResponses(w, c.modelId),
248
- "gpt-5.1": (w, c) => applyOpenAIResponses(w, c.modelId),
249
- "gpt-5": (w, c) => applyOpenAIResponses(w, c.modelId),
250
- };
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 { isPlainObject, PayloadWriter, type WebSearchMode } from "./_tools.ts";
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: Record<string, unknown>): boolean {
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: unknown,
85
+ export function applyBuiltinSearchTool<T>(
86
+ payload: T,
81
87
  modelId: string,
82
88
  api: string,
83
89
  mode: WebSearchMode,
84
- ): unknown {
90
+ ): T | Payload {
85
91
  if (!isPlainObject(payload) || mode === "off") return payload;
86
92
  const writer = new PayloadWriter(payload);
87
93
 
@@ -1,11 +1,6 @@
1
1
  /**
2
- * 独立搜索工具(ts_search):供所有模型直接调用的统一搜索入口。
3
- *
4
- * 与模型模块的内置查询注入(models/web-search.ts)分离:
5
- * - 内置查询注入:GPT/Grok 模型请求时自动附加搜索能力
6
- * - 本工具:任何模型主动调用,内部通过固定的 GPT / Grok 后端联网搜索
7
- *
8
- * 工具只接收查询;搜索后端由扩展维护,避免调用方猜测或误选模型。
2
+ * 独立搜索工具(ts_search):任何模型均可通过配置或内置的 GPT / Grok
3
+ * 后端执行联网搜索。模型模块自身的内置查询注入在 models/web-search.ts。
9
4
  */
10
5
 
11
6
  import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
@@ -13,22 +8,24 @@ import { Type } from "typebox";
13
8
 
14
9
  export interface TsSearchOptions {
15
10
  baseUrl: string;
16
- /** 凭据 provider(默认 tsgw,走 Pi 凭据机制)。 */
17
11
  apiKeyProvider?: string;
18
- /** @deprecated 固定白名单已内置;仅兼容旧 index.ts,值会被忽略。 */
19
12
  searchModels?: readonly string[];
20
13
  }
21
14
 
22
- export const DEFAULT_SEARCH_MODELS = ["gpt-5.6-luna", "grok-chat-fast"] as const;
15
+ export const DEFAULT_SEARCH_MODELS = [
16
+ "gpt-5.6-luna",
17
+ "grok-chat-fast",
18
+ ] as const;
23
19
  const CHAT_COMPLETIONS_PATH = "v1/chat/completions";
24
-
20
+ const MAX_ATTEMPTS = 3;
21
+ const REQUEST_TIMEOUT_MS = 45_000;
22
+ const RETRY_DELAY_MS = 1_000;
25
23
  const SEARCH_PROMPT = [
26
24
  "Answer the user's question using the latest available information from the web.",
27
25
  "Be concise.",
28
26
  "If sources are available, include only the most relevant source URLs needed to support the answer; avoid link dumping.",
29
27
  "If no live web information is available, reply exactly: NO_LIVE_WEB.",
30
28
  ].join(" ");
31
-
32
29
  type SearchFamily = "gpt" | "grok";
33
30
 
34
31
  export function detectSearchFamily(
@@ -40,8 +37,28 @@ export function detectSearchFamily(
40
37
  return "unsupported";
41
38
  }
42
39
 
43
- export function resolveSearchModels(): string[] {
44
- return [...DEFAULT_SEARCH_MODELS];
40
+ export function resolveSearchModels(
41
+ options: Pick<TsSearchOptions, "searchModels"> = {},
42
+ ): string[] {
43
+ if (!options.searchModels) return [...DEFAULT_SEARCH_MODELS];
44
+
45
+ const models: string[] = [];
46
+ const seen = new Set<string>();
47
+ for (const value of options.searchModels as readonly unknown[]) {
48
+ const model = typeof value === "string" ? value.trim() : "";
49
+ if (!model || detectSearchFamily(model) === "unsupported") {
50
+ console.warn(
51
+ `TSGW: ignoring invalid ts_search model ${JSON.stringify(value)}; expected gpt-* or grok-*`,
52
+ );
53
+ continue;
54
+ }
55
+ if (!seen.has(model)) {
56
+ seen.add(model);
57
+ models.push(model);
58
+ }
59
+ }
60
+
61
+ return models.length > 0 ? models : [...DEFAULT_SEARCH_MODELS];
45
62
  }
46
63
 
47
64
  export function buildSearchUrl(baseUrl: string): string {
@@ -146,7 +163,7 @@ export function buildSearchPayload(
146
163
  return { model, messages, search_parameters: { mode: "on" } };
147
164
  }
148
165
 
149
- interface RouteResult {
166
+ export interface RouteResult {
150
167
  ok: boolean;
151
168
  model: string;
152
169
  family: SearchFamily;
@@ -156,78 +173,195 @@ interface RouteResult {
156
173
  error: string;
157
174
  }
158
175
 
159
- async function executeSearchRoute(
176
+ export interface SearchRetryOptions {
177
+ delayMs?: number;
178
+ requestTimeoutMs?: number;
179
+ fetch?: typeof fetch;
180
+ }
181
+
182
+ interface SearchAttemptResult {
183
+ result: RouteResult;
184
+ retryable: boolean;
185
+ }
186
+
187
+ function failedAttempt(
188
+ model: string,
189
+ family: SearchFamily,
190
+ error: string,
191
+ retryable: boolean,
192
+ requestId = "",
193
+ ): SearchAttemptResult {
194
+ return {
195
+ result: { ...buildFailure(model, family, error), requestId },
196
+ retryable,
197
+ };
198
+ }
199
+
200
+ function requestSignal(
201
+ signal: AbortSignal | undefined,
202
+ timeoutMs: number,
203
+ ): AbortSignal {
204
+ const timeout = AbortSignal.timeout(timeoutMs);
205
+ return signal ? AbortSignal.any([signal, timeout]) : timeout;
206
+ }
207
+
208
+ function waitForRetry(
209
+ delayMs: number,
210
+ signal: AbortSignal | undefined,
211
+ ): Promise<void> {
212
+ if (delayMs <= 0) return Promise.resolve();
213
+ if (signal?.aborted) return Promise.reject(signal.reason);
214
+
215
+ return new Promise((resolve, reject) => {
216
+ const onAbort = (): void => {
217
+ clearTimeout(timer);
218
+ reject(signal?.reason);
219
+ };
220
+ const timer = setTimeout(() => {
221
+ signal?.removeEventListener("abort", onAbort);
222
+ resolve();
223
+ }, delayMs);
224
+ signal?.addEventListener("abort", onAbort, { once: true });
225
+ });
226
+ }
227
+
228
+ async function executeSearchAttempt(
160
229
  model: string,
161
230
  query: string,
162
231
  apiKey: string,
163
232
  baseUrl: string,
164
233
  signal: AbortSignal | undefined,
165
- ): Promise<RouteResult> {
234
+ fetchImpl: typeof fetch,
235
+ requestTimeoutMs: number,
236
+ ): Promise<SearchAttemptResult> {
166
237
  const family = detectSearchFamily(model);
167
- if (family === "unsupported") {
168
- return buildFailure(model, "gpt", `不支持的模型系列:${model}`);
169
- }
238
+ if (family === "unsupported")
239
+ return failedAttempt(model, "gpt", `不支持的模型系列:${model}`, false);
240
+ if (signal?.aborted)
241
+ return failedAttempt(model, family, "请求已取消。", false);
170
242
 
171
243
  let response: Response;
172
244
  try {
173
- response = await fetch(buildSearchUrl(baseUrl), {
245
+ response = await fetchImpl(buildSearchUrl(baseUrl), {
174
246
  method: "POST",
175
247
  headers: {
176
248
  authorization: `Bearer ${apiKey}`,
177
249
  "content-type": "application/json",
178
250
  },
179
251
  body: JSON.stringify(buildSearchPayload(model, family, query)),
180
- signal,
252
+ signal: requestSignal(signal, requestTimeoutMs),
181
253
  });
182
254
  } catch (error) {
183
- return buildFailure(
255
+ return failedAttempt(
184
256
  model,
185
257
  family,
186
- `请求失败:${error instanceof Error ? error.message : String(error)}`,
258
+ signal?.aborted
259
+ ? "请求已取消。"
260
+ : `请求失败:${error instanceof Error ? error.message : String(error)}`,
261
+ !signal?.aborted,
187
262
  );
188
263
  }
189
264
 
190
265
  const requestId = response.headers.get("ah-request-id") ?? "";
191
- const bodyText = await response.text();
192
- if (!response.ok) {
193
- return {
194
- ...buildFailure(
195
- model,
196
- family,
197
- `网关请求失败:HTTP ${response.status} ${response.statusText}\n${bodyText}`,
198
- ),
266
+ let bodyText: string;
267
+ try {
268
+ bodyText = await response.text();
269
+ } catch (error) {
270
+ return failedAttempt(
271
+ model,
272
+ family,
273
+ signal?.aborted
274
+ ? "请求已取消。"
275
+ : `读取网关响应失败:${error instanceof Error ? error.message : String(error)}`,
276
+ !signal?.aborted,
199
277
  requestId,
200
- };
278
+ );
201
279
  }
280
+ if (!response.ok)
281
+ return failedAttempt(
282
+ model,
283
+ family,
284
+ `网关请求失败:HTTP ${response.status} ${response.statusText}\n${bodyText}`,
285
+ response.status === 429 || response.status >= 500,
286
+ requestId,
287
+ );
202
288
 
203
289
  try {
204
290
  const parsed: unknown = JSON.parse(bodyText);
205
291
  if (parsed && typeof parsed === "object") {
206
292
  const record = parsed as Record<string, unknown>;
207
293
  return {
208
- ok: true,
209
- model,
210
- family,
211
- answer: extractAnswer(record) || JSON.stringify(record),
212
- urls: Array.from(collectUrls(record)),
213
- requestId,
214
- error: "",
294
+ result: {
295
+ ok: true,
296
+ model,
297
+ family,
298
+ answer: extractAnswer(record) || JSON.stringify(record),
299
+ urls: Array.from(collectUrls(record)),
300
+ requestId,
301
+ error: "",
302
+ },
303
+ retryable: false,
215
304
  };
216
305
  }
217
306
  } catch {
218
307
  // 非 JSON 响应按纯文本处理。
219
308
  }
220
309
  return {
221
- ok: true,
222
- model,
223
- family,
224
- answer: bodyText || "网关返回空响应。",
225
- urls: Array.from(collectUrls(bodyText)),
226
- requestId,
227
- error: "",
310
+ result: {
311
+ ok: true,
312
+ model,
313
+ family,
314
+ answer: bodyText || "网关返回空响应。",
315
+ urls: Array.from(collectUrls(bodyText)),
316
+ requestId,
317
+ error: "",
318
+ },
319
+ retryable: false,
228
320
  };
229
321
  }
230
322
 
323
+ export async function retrySearchRoute(
324
+ model: string,
325
+ query: string,
326
+ apiKey: string,
327
+ baseUrl: string,
328
+ signal: AbortSignal | undefined,
329
+ options: SearchRetryOptions = {},
330
+ ): Promise<RouteResult> {
331
+ const delayMs = options.delayMs ?? RETRY_DELAY_MS;
332
+ const requestTimeoutMs = options.requestTimeoutMs ?? REQUEST_TIMEOUT_MS;
333
+ const fetchImpl = options.fetch ?? globalThis.fetch;
334
+ let lastResult = buildFailure(model, "gpt", "搜索请求未执行。");
335
+
336
+ for (let attempt = 0; attempt < MAX_ATTEMPTS; attempt += 1) {
337
+ const current = await executeSearchAttempt(
338
+ model,
339
+ query,
340
+ apiKey,
341
+ baseUrl,
342
+ signal,
343
+ fetchImpl,
344
+ requestTimeoutMs,
345
+ );
346
+ lastResult = current.result;
347
+ if (current.result.ok || !current.retryable || attempt === MAX_ATTEMPTS - 1)
348
+ return current.result;
349
+
350
+ try {
351
+ await waitForRetry(delayMs * 2 ** attempt, signal);
352
+ } catch {
353
+ const family = detectSearchFamily(model);
354
+ return buildFailure(
355
+ model,
356
+ family === "grok" ? "grok" : "gpt",
357
+ "请求已取消。",
358
+ );
359
+ }
360
+ }
361
+
362
+ return lastResult;
363
+ }
364
+
231
365
  function buildFailure(
232
366
  model: string,
233
367
  family: SearchFamily,
@@ -288,11 +422,11 @@ function renderMergedResult(results: RouteResult[]): string {
288
422
  return lines.join("\n");
289
423
  }
290
424
 
291
- /** 注册 ts_search 工具,固定并行请求白名单内的 GPT + Grok 后端。 */
292
425
  export function registerTsSearch(
293
426
  pi: ExtensionAPI,
294
427
  options: TsSearchOptions,
295
428
  ): void {
429
+ const models = resolveSearchModels(options);
296
430
  const argsSchema = Type.Object({
297
431
  query: Type.String({ minLength: 1, description: "搜索查询" }),
298
432
  });
@@ -340,10 +474,9 @@ export function registerTsSearch(
340
474
  };
341
475
  }
342
476
 
343
- const models = resolveSearchModels();
344
477
  const settled = await Promise.allSettled(
345
478
  models.map((model) =>
346
- executeSearchRoute(model, query, apiKey, options.baseUrl, signal),
479
+ retrySearchRoute(model, query, apiKey, options.baseUrl, signal),
347
480
  ),
348
481
  );
349
482
  const results = settled.map((item, index) => {
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "pi-tsgw",
3
- "version": "0.0.5",
3
+ "version": "0.0.7",
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": [