pi-tsgw 0.0.11 → 0.0.12

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
@@ -6,8 +6,9 @@ openai-responses, anthropic-messages, google-generative-ai) and applies
6
6
  per-vendor request-body operations so thinking levels, reasoning formats, and
7
7
  optional built-in web search behave correctly on the gateway.
8
8
 
9
- No API keys, endpoints, or pricing are baked into this package: everything
10
- gateway-specific is configuration.
9
+ No private API keys or gateway endpoints are included. The static catalog
10
+ contains model metadata and public-price estimates; these are not your gateway's
11
+ billing rates.
11
12
 
12
13
  ## Install
13
14
 
@@ -36,7 +37,7 @@ Example:
36
37
  {
37
38
  "tsgw": {
38
39
  "baseUrl": "https://aih.example.net",
39
- "tsSearch": "off",
40
+ "tsSearch": "live",
40
41
  "traceHeaders": false,
41
42
  "excludeModels": ["gpt-*", "claude-*"],
42
43
  "includeModels": ["gpt-5.6-*"]
@@ -52,19 +53,23 @@ placeholder root is used that will not work.
52
53
 
53
54
  ### API key
54
55
 
55
- The extension never reads credential files. Pi resolves the key itself:
56
+ The extension delegates credential reading and resolution to Pi:
56
57
 
57
58
  1. **`/login`** — select the `tsgw` provider and enter the key; pi stores it
58
59
  in `~/.pi/agent/auth.json` (preferred)
59
60
  2. **`TSGW_API_KEY`** environment variable — Pi's fallback when nothing is
60
61
  stored
61
62
 
63
+ Startup discovery uses Pi's stored-credential API; later catalog refreshes use
64
+ the credential supplied by Pi. With environment-only authentication, the initial
65
+ catalog may fall back to static models until Pi refreshes it.
66
+
62
67
  ### Options
63
68
 
64
69
  | Setting (`settings.json` `tsgw.*`) | Default | Meaning |
65
70
  | --- | --- | --- |
66
71
  | `baseUrl` | placeholder | Gateway root (the `/v1` suffix is normalized away) |
67
- | `tsSearch` | `off` | `off`, `cached`, or `live` — appends the built-in `web_search` tool for models with built-in search (GPT / Grok) |
72
+ | `tsSearch` | `live` | GPT: `cached` or `live`; unset, legacy `off`, and invalid values fall back to `live`. Supported Grok models always use live search. |
68
73
  | `traceHeaders` | `false` | Adds `AH-Thread-Id` / `AH-Trace-Id` headers for gateway-side tracing |
69
74
  | `includeModels` | unset | Re-include override: matching model ids are kept even when also matched by `excludeModels` (exact id or `prefix-*` glob) |
70
75
  | `excludeModels` | unset | Blacklist: matching model ids are dropped unless pulled back by `includeModels` (exact id or `prefix-*` glob) |
@@ -72,42 +77,63 @@ The extension never reads credential files. Pi resolves the key itself:
72
77
  ## Model catalog
73
78
 
74
79
  The catalog is assembled from per-vendor slices (`extensions/models/vendors/`),
75
- currently covering 10 vendor families: DeepSeek, GLM (Zhipu), MiMo (Xiaomi),
76
- MiniMax, Kimi (Moonshot), Qwen (Alibaba), GPT (OpenAI), Gemini (Google),
77
- Claude (Anthropic), and LongCat — 59 chat models in total. Each vendor file
78
- carries the official documentation URL in its header comment. Pricing entries
79
- mirror the vendors' public list prices; replace them with your gateway's
80
- actual pricing if it differs.
80
+ covering 11 vendor families: DeepSeek, GLM (Zhipu), MiMo (Xiaomi), MiniMax,
81
+ Kimi (Moonshot), Qwen (Alibaba), GPT (OpenAI), Gemini (Google), Claude
82
+ (Anthropic), LongCat, and Grok (xAI). The current static catalog contains 54 chat
83
+ model IDs, including retained aliases; this is not a live availability count.
84
+ Vendor files carry documentation links and metadata caveats. Alias mappings,
85
+ output limits, and shared thinking policies are not all individually verified
86
+ against gateway endpoints; public-price estimates do not guarantee billing.
81
87
 
82
88
  Use broad `excludeModels` rules to drop model families and precise
83
89
  `includeModels` rules to pull selected models back. `includeModels` is not a
84
90
  global whitelist: models matching neither list remain available. To register
85
91
  only selected models, use `"excludeModels": ["*"]` and pull them back through
86
- `includeModels`.
92
+ `includeModels`. Both lists are case-sensitive and support only exact IDs or
93
+ trailing `*` prefix matching (not general globs); preserve `MiniMax-*` casing.
94
+ When a gateway catalog is available, filtering applies after intersecting it
95
+ with the static catalog. Include rules cannot add unknown or unavailable IDs.
87
96
 
88
97
  Opening `/model` refreshes the gateway catalog in the background, with a
89
98
  five-minute cache to avoid repeated requests. Run `/tsgw-refresh` to bypass the
90
99
  cache and force an immediate refresh without restarting or reloading Pi. If a
91
100
  refresh fails, the last successful model list remains active.
92
101
 
102
+ Memory and disk caches are scoped to the gateway root and credential using a
103
+ SHA-256 digest; plaintext credentials are not stored. The disk cache is a single
104
+ slot at `<agent-dir>/tsgw/models-cache.json`: changing scope may require a fresh
105
+ fetch. Legacy unscoped cache files are ignored and replaced after a successful
106
+ fetch. In-process request ordering prevents late results from replacing newer
107
+ cache writes; this is not a cross-process locking mechanism.
108
+
93
109
  ## Status bar and session controls
94
110
 
95
- Pi Web 0.11.0+ renders these extension statuses as clickable buttons:
111
+ Pi Web clients supporting `command:/…` statuses render these as clickable
112
+ buttons; the terminal CLI displays status text and accepts the same commands:
96
113
 
97
114
  - **`TSGW · N`** (`/tsgw`): one menu for **refreshing the model catalog** and
98
115
  **viewing diagnostics**. The count is the filtered registered catalog, not a
99
116
  count of healthy inference endpoints. 🟢 means a fresh directory result;
100
117
  🟡 means a static/expired directory or a failed refresh; ⚠️ means missing
101
118
  credentials or gateway configuration; ⏳ means a refresh is running.
102
- - **Built-in search** (`/tsgw-search`): choose a mode for the current session.
103
- GPT offers `off / cached / live`; supported Grok models offer `off / live`
104
- because the existing Grok request policy does not distinguish cache-only
105
- search. Unsupported TSGW models are labeled accordingly; this button is
106
- hidden when another provider is selected.
107
-
108
- You can also type `/tsgw-search off`, `/tsgw-search cached`, or
109
- `/tsgw-search live`. These controls do not modify `settings.json` or persist
110
- across a reload or a new session; `tsgw.tsSearch` remains the startup default.
119
+ - **Built-in search** (`/tsgw-search`): GPT offers only `cached / live`
120
+ for the current session, defaulting to `live`. Supported Grok models use
121
+ fixed live search: the command only shows a notice, without a picker or
122
+ changing the GPT session mode. Unsupported TSGW models are labeled
123
+ accordingly; this button is hidden when another provider is selected.
124
+
125
+ For GPT, you can also type `/tsgw-search cached` or `/tsgw-search live`;
126
+ `off` is no longer accepted. These controls do not modify `settings.json` or
127
+ persist across a reload or a new session; `tsgw.tsSearch` remains the startup
128
+ default. Existing `cached` configuration is preserved; unset, legacy `off`,
129
+ and invalid configuration now default to `live`.
130
+
131
+ Injection remains limited to the existing model/protocol support list. GPT
132
+ Responses appends a `web_search` tool (`external_web_access` is false for
133
+ `cached`, true for `live`); Grok Completions appends
134
+ `search_parameters: { mode: "on" }`. Existing search tools/fields are never
135
+ overwritten. This setting does not control Qwen's vendor policy, external
136
+ tools, or thinking mode and is not a global network-access switch.
111
137
 
112
138
  The standalone `ts_search` tool is no longer bundled. Use a separately configured
113
139
  search tool such as `ts_oht.search` instead. Built-in model search remains
@@ -126,16 +152,26 @@ statuses until the session runtime has started.
126
152
  ## Development
127
153
 
128
154
  ```bash
129
- npm install
155
+ npm ci
130
156
  npm test
157
+ npm exec -- tsc -p tsconfig.json
158
+ npm pack --dry-run --json
131
159
  ```
132
160
 
133
- Tests run against a fake pi host (`FakePi` + `FakeContext`) and isolate the
134
- config directory through `PI_CODING_AGENT_DIR`, so no pi installation is
135
- needed to run them.
161
+ Pi is installed as a development dependency; the lockfile defines the test
162
+ baseline. Tests isolate configuration and credentials using temporary agent
163
+ directories. Unit tests use a fake host; host integration tests load the extension
164
+ through the real Pi SDK with simulated HTTP responses, without production
165
+ requests. They do not verify upstream inference, native search results, or Pi
166
+ Web rendering. Run the full suite after upgrading Pi.
136
167
 
137
168
  ## Technical details
138
169
 
139
- Per-vendor thinking profiles (the request-body matrix), the built-in web
140
- search injection semantics, and the pi lifecycle compatibility notes live in
141
- [AGENTS.md](AGENTS.md).
170
+ - [Vendor profiles](https://github.com/tossp/pi-tsgw/tree/main/extensions/models/vendors): model metadata, compatibility flags, and thinking mappings.
171
+ - [Request operations](https://github.com/tossp/pi-tsgw/blob/main/extensions/models/operations.ts): copy-on-write dispatch; DeepSeek/Claude response handling stays with Pi's adapters.
172
+ - [Built-in search](https://github.com/tossp/pi-tsgw/blob/main/extensions/models/web-search.ts): exact model/protocol support and append-only injection.
173
+ - [Host integration](https://github.com/tossp/pi-tsgw/blob/main/extensions/index.ts): configuration, refreshes, commands, and lifecycle snapshots. Late request callbacks must not access stale session contexts.
174
+
175
+ These source links describe the repository's current main branch; an installed
176
+ release may differ. Development rules in `AGENTS.md` are repository-only and
177
+ are not part of the npm package.
@@ -200,7 +200,7 @@ export default async function registerTsgw(pi: ExtensionAPI): Promise<void> {
200
200
  const defaultSearchMode: WebSearchMode =
201
201
  settings.tsSearch === "cached" || settings.tsSearch === "live"
202
202
  ? settings.tsSearch
203
- : "off";
203
+ : "live";
204
204
  let tsSearchMode = defaultSearchMode;
205
205
  const traceEnabled = settings.traceHeaders === true;
206
206
  const modelFilter: ModelFilter = {
@@ -383,8 +383,11 @@ export default async function registerTsgw(pi: ExtensionAPI): Promise<void> {
383
383
  ctx.ui.notify("当前模型不支持插件内置联网。", "info");
384
384
  return;
385
385
  }
386
- const modes: WebSearchMode[] = support === "gpt"
387
- ? ["off", "cached", "live"] : ["off", "live"];
386
+ if (support === "grok") {
387
+ ctx.ui.notify("当前 Grok 模型内置联网固定为实时;不修改 GPT 会话模式。", "info");
388
+ return;
389
+ }
390
+ const modes: WebSearchMode[] = ["cached", "live"];
388
391
  const labels = modes.map((mode) => searchLabel(mode, support));
389
392
  const input = args.trim();
390
393
  const choice = input || (ctx.hasUI
@@ -396,7 +399,7 @@ export default async function registerTsgw(pi: ExtensionAPI): Promise<void> {
396
399
  }
397
400
  const mode = modes.find((mode, index) => mode === choice || labels[index] === choice);
398
401
  if (!mode) {
399
- ctx.ui.notify("不支持的模式;GPT: off/cached/live,Grok: off/live。", "warning");
402
+ ctx.ui.notify("不支持的模式;GPT 仅支持 cached/live,Grok 固定实时。", "warning");
400
403
  return;
401
404
  }
402
405
  tsSearchMode = mode;
@@ -5,6 +5,12 @@
5
5
  * 与 operations 调度都从这里引用,避免 vendors ↔ operations 循环依赖。
6
6
  */
7
7
 
8
+ import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
9
+
10
+ // Pi 1.x also includes image/classifier configs; this catalog is chat-only.
11
+ // Structural extraction keeps compatibility with the older chat-only type.
12
+ export type ChatModelConfig = Extract<ProviderModelConfig, { reasoning: boolean }>;
13
+
8
14
  export type ThinkingLevel =
9
15
  | "off"
10
16
  | "minimal"
@@ -13,7 +19,7 @@ export type ThinkingLevel =
13
19
  | "high"
14
20
  | "xhigh"
15
21
  | "max";
16
- export type WebSearchMode = "off" | "cached" | "live";
22
+ export type WebSearchMode = "cached" | "live";
17
23
 
18
24
  export interface ModelOperationContext {
19
25
  provider: string;
@@ -28,7 +34,7 @@ type GoogleThinkingField =
28
34
  | "includeThoughts"
29
35
  | "thinkingLevel"
30
36
  | "thinkingBudget";
31
- type ThinkingUpdate = Partial<Pick<Payload, "type" | "clear_thinking">>;
37
+ type ThinkingUpdate = Partial<Pick<Payload, "type" | "clear_thinking" | "keep">>;
32
38
  type GoogleThinkingUpdate = Partial<Pick<Payload, GoogleThinkingField>>;
33
39
 
34
40
  export function isPlainObject(value: unknown): value is Payload {
@@ -109,7 +115,7 @@ export class PayloadWriter {
109
115
  }
110
116
  }
111
117
 
112
- const THINKING_FIELDS = ["type", "clear_thinking"] as const;
118
+ const THINKING_FIELDS = ["type", "clear_thinking", "keep"] as const;
113
119
  const GOOGLE_THINKING_FIELDS = [
114
120
  "includeThoughts",
115
121
  "thinkingLevel",
@@ -1,4 +1,4 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "./_tools.ts";
2
2
  import { anthropicModels } from "./vendors/anthropic.ts";
3
3
  import { deepseekModels } from "./vendors/deepseek.ts";
4
4
  import { geminiModels } from "./vendors/gemini.ts";
@@ -1,4 +1,4 @@
1
- import { randomUUID } from "node:crypto";
1
+ import { createHash, randomUUID } from "node:crypto";
2
2
  import {
3
3
  mkdirSync,
4
4
  readFileSync,
@@ -6,7 +6,7 @@ import {
6
6
  unlinkSync,
7
7
  writeFileSync,
8
8
  } from "node:fs";
9
- import { dirname } from "node:path";
9
+ import { dirname, resolve } from "node:path";
10
10
 
11
11
  export const DEFAULT_GATEWAY_MODEL_TIMEOUT_MS = 5_000;
12
12
  export const DEFAULT_GATEWAY_MODEL_CACHE_TTL_MS = 5 * 60 * 1_000;
@@ -126,6 +126,8 @@ export async function fetchGatewayModelIds(
126
126
  return { ok: false, reason: "network", detail: errorDetail(error) };
127
127
  }
128
128
 
129
+ if (timedOut) return { ok: false, reason: "timeout" };
130
+ if (opts.signal?.aborted) return { ok: false, reason: "aborted" };
129
131
  if (!response.ok) {
130
132
  return { ok: false, reason: "http", status: response.status };
131
133
  }
@@ -139,6 +141,8 @@ export async function fetchGatewayModelIds(
139
141
  return { ok: false, reason: "invalid-json", detail: errorDetail(error) };
140
142
  }
141
143
 
144
+ if (timedOut) return { ok: false, reason: "timeout" };
145
+ if (opts.signal?.aborted) return { ok: false, reason: "aborted" };
142
146
  if (!payload || typeof payload !== "object")
143
147
  return { ok: false, reason: "invalid-data" };
144
148
  const data = (payload as { data?: unknown }).data;
@@ -176,17 +180,27 @@ export function createModelCache(
176
180
  };
177
181
  }
178
182
 
183
+ /** Hash both identity components; no plaintext credentials in cache keys or files. */
184
+ export function gatewayModelCacheScope(baseUrl: string, apiKey: string): string {
185
+ return createHash("sha256")
186
+ .update(JSON.stringify([baseUrl.replace(/\/+$/, ""), apiKey]))
187
+ .digest("hex");
188
+ }
189
+
179
190
  function readGatewayModelCacheEntry(
180
191
  filePath: string,
192
+ scope: string,
181
193
  ): ModelCacheEntry | undefined {
182
194
  try {
183
195
  const payload: unknown = JSON.parse(readFileSync(filePath, "utf8"));
184
196
  if (!payload || typeof payload !== "object") return undefined;
185
- const { storedAt, ids } = payload as {
197
+ const { storedAt, ids, scope: storedScope } = payload as {
198
+ scope?: unknown;
186
199
  storedAt?: unknown;
187
200
  ids?: unknown;
188
201
  };
189
202
  if (
203
+ !scope || storedScope !== scope ||
190
204
  typeof storedAt !== "number" ||
191
205
  !Number.isFinite(storedAt) ||
192
206
  storedAt < 0 ||
@@ -214,6 +228,7 @@ function isCacheEntryFresh(
214
228
  /** Atomically persist gateway model IDs. Write failures are non-fatal. */
215
229
  export function saveGatewayModelCache(
216
230
  filePath: string,
231
+ scope: string,
217
232
  ids: readonly string[],
218
233
  now = Date.now(),
219
234
  ): void {
@@ -222,7 +237,7 @@ export function saveGatewayModelCache(
222
237
  mkdirSync(dirname(filePath), { recursive: true });
223
238
  writeFileSync(
224
239
  temporaryPath,
225
- JSON.stringify({ storedAt: now, ids: [...ids] }),
240
+ JSON.stringify({ scope, storedAt: now, ids: [...ids] }),
226
241
  { encoding: "utf8", mode: 0o600 },
227
242
  );
228
243
  renameSync(temporaryPath, filePath);
@@ -238,12 +253,13 @@ export function saveGatewayModelCache(
238
253
  /** Load a persisted gateway model list. Missing or invalid files are ignored. */
239
254
  export function loadGatewayModelCache(
240
255
  filePath: string,
256
+ scope: string,
241
257
  now = Date.now(),
242
258
  ttlMs = DEFAULT_GATEWAY_MODEL_CACHE_TTL_MS,
243
259
  ): LoadedGatewayModelCache | undefined {
244
260
  if (!Number.isFinite(ttlMs) || ttlMs < 0)
245
261
  throw new RangeError("ttlMs must be a non-negative finite number");
246
- const entry = readGatewayModelCacheEntry(filePath);
262
+ const entry = readGatewayModelCacheEntry(filePath, scope);
247
263
  if (!entry) return undefined;
248
264
  return {
249
265
  ids: [...entry.ids],
@@ -252,10 +268,18 @@ export function loadGatewayModelCache(
252
268
  }
253
269
 
254
270
  const gatewayCaches = new Map<string, ModelCache>();
255
- const inFlightRequests = new Map<string, Promise<FetchGatewayModelIdsResult>>();
271
+ interface PendingRequest {
272
+ promise: Promise<FetchGatewayModelIdsResult>;
273
+ revision: number;
274
+ storedAt: number;
275
+ }
276
+ const inFlightRequests = new Map<string, PendingRequest>();
277
+ const latestRequests = new Map<string, number>();
278
+ // Disk is single-slot per normalized path, independently of the memory scope.
279
+ const latestDiskRequests = new Map<string, number>();
280
+ let revision = 0;
256
281
 
257
- function cacheFor(baseUrl: string): ModelCache {
258
- const key = baseUrl.replace(/\/+$/, "");
282
+ function cacheFor(key: string): ModelCache {
259
283
  let cache = gatewayCaches.get(key);
260
284
  if (!cache) {
261
285
  cache = createModelCache();
@@ -276,7 +300,8 @@ export async function getGatewayModelIds(
276
300
  opts: GetGatewayModelIdsOptions = {},
277
301
  now = Date.now(),
278
302
  ): Promise<GetGatewayModelIdsResult> {
279
- const key = baseUrl.replace(/\/+$/, "");
303
+ if (opts.signal?.aborted) return { ok: false, reason: "aborted" };
304
+ const key = gatewayModelCacheScope(baseUrl, apiKey);
280
305
  const cache = cacheFor(key);
281
306
  if (!opts.force && cache.isFresh(now)) {
282
307
  const entry = cache.get();
@@ -291,7 +316,7 @@ export async function getGatewayModelIds(
291
316
  }
292
317
 
293
318
  const diskEntry = opts.cacheFilePath
294
- ? readGatewayModelCacheEntry(opts.cacheFilePath)
319
+ ? readGatewayModelCacheEntry(opts.cacheFilePath, key)
295
320
  : undefined;
296
321
  if (
297
322
  !opts.force &&
@@ -315,32 +340,53 @@ export async function getGatewayModelIds(
315
340
  const shareRequest = opts.signal === undefined;
316
341
  let request = shareRequest ? inFlightRequests.get(requestKey) : undefined;
317
342
  if (!request) {
318
- request = fetchGatewayModelIds(key, apiKey, {
319
- fetcher: opts.fetcher,
320
- timeoutMs: opts.timeoutMs,
321
- signal: opts.signal,
322
- });
343
+ const requestRevision = ++revision;
344
+ latestRequests.set(key, requestRevision);
345
+ request = {
346
+ revision: requestRevision,
347
+ storedAt: now,
348
+ promise: fetchGatewayModelIds(baseUrl, apiKey, {
349
+ fetcher: opts.fetcher,
350
+ timeoutMs: opts.timeoutMs,
351
+ signal: opts.signal,
352
+ }),
353
+ };
323
354
  if (shareRequest) inFlightRequests.set(requestKey, request);
324
355
  }
325
356
 
357
+ const diskPath = opts.cacheFilePath ? resolve(opts.cacheFilePath) : undefined;
358
+ if (diskPath) {
359
+ // Register every subscriber's path, including shared requests with different
360
+ // destinations. Joining an older request must not roll back a newer fence.
361
+ latestDiskRequests.set(diskPath, Math.max(
362
+ latestDiskRequests.get(diskPath) ?? 0, request.revision,
363
+ ));
364
+ }
365
+
326
366
  let fetched: FetchGatewayModelIdsResult;
327
367
  try {
328
- fetched = await request;
368
+ fetched = await request.promise;
329
369
  } finally {
330
370
  if (shareRequest && inFlightRequests.get(requestKey) === request)
331
371
  inFlightRequests.delete(requestKey);
332
372
  }
333
373
 
374
+ // Cancellation never falls back or writes, even if a fetcher ignores its signal.
375
+ if (opts.signal?.aborted) return { ok: false, reason: "aborted" };
334
376
  if (fetched.ok) {
335
- cache.set(fetched.ids, now);
336
- if (opts.cacheFilePath)
337
- saveGatewayModelCache(opts.cacheFilePath, fetched.ids, now);
377
+ // Latest-started wins even if the newer request fails or aborts.
378
+ // Superseded successes still return to their callers without cache writes.
379
+ if (latestRequests.get(key) === request.revision) {
380
+ cache.set(fetched.ids, request.storedAt);
381
+ if (diskPath && latestDiskRequests.get(diskPath) === request.revision)
382
+ saveGatewayModelCache(diskPath, key, fetched.ids, request.storedAt);
383
+ }
338
384
  return {
339
385
  ok: true,
340
386
  ids: [...fetched.ids],
341
387
  cached: false,
342
388
  stale: false,
343
- storedAt: now,
389
+ storedAt: request.storedAt,
344
390
  };
345
391
  }
346
392
 
@@ -1,4 +1,4 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
 
3
3
  /**
4
4
  * Wire 协议公共配置。
@@ -1,4 +1,4 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { ANTHROPIC_COMPAT } from "./_protocols.ts";
3
3
 
4
4
  /**
@@ -1,4 +1,4 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
 
3
3
  // DeepSeek Responses 将 developer role 当作 user,且不支持 OpenAI 的
4
4
  // prompt_cache_retention;其余 reasoning 字段交给 Pi 的 Responses adapter。
@@ -1,9 +1,11 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
2
- import { isAtLeastHigh, type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
+ import { type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
3
3
 
4
4
  /**
5
5
  * Google Gemini 系列(google-generative-ai 协议)。
6
6
  * 官方文档:https://ai.google.dev/gemini-api/docs
7
+ * GenerateContent 档位/预算:https://docs.cloud.google.com/vertex-ai/generative-ai/docs/thinking
8
+ * 2.5 使用 budget;3.x 使用 level。网关别名的路由仍需网关侧验证。
7
9
  */
8
10
  export function geminiModels(root: string): ProviderModelConfig[] {
9
11
  const gemini = `${root}/gemini`;
@@ -135,14 +137,31 @@ function applyGeminiPro(writer: PayloadWriter, level: ThinkingLevel): void {
135
137
  writer.setGoogleThinking({ includeThoughts: level !== "off", thinkingLevel }, ["thinkingBudget"]);
136
138
  }
137
139
 
140
+ // 2.5 Pro 不可关闭思考,官方 budget 范围 128..32768。
141
+ // 128/16000/32768 是项目档位映射选择,并非官方档位表;off 仅隐藏摘要并取最小预算。
142
+ function applyGemini25Pro(writer: PayloadWriter, level: ThinkingLevel): void {
143
+ writer.remove("reasoning_effort");
144
+ const thinkingBudget = level === "off" || level === "minimal" || level === "low"
145
+ ? 128 : level === "max" ? 32768 : 16000;
146
+ writer.setGoogleThinking({ includeThoughts: level !== "off", thinkingBudget }, ["thinkingLevel"]);
147
+ }
148
+
149
+ // 3.5 Flash 支持 MINIMAL/LOW/MEDIUM/HIGH,不使用 2.5 的 token budget。
150
+ function applyGemini35Flash(writer: PayloadWriter, level: ThinkingLevel): void {
151
+ writer.remove("reasoning_effort");
152
+ const thinkingLevel = level === "off" || level === "minimal" ? "MINIMAL"
153
+ : level === "low" ? "LOW" : level === "medium" ? "MEDIUM" : "HIGH";
154
+ writer.setGoogleThinking({ includeThoughts: level !== "off", thinkingLevel }, ["thinkingBudget"]);
155
+ }
156
+
138
157
  export const geminiThinking: Record<string, ThinkingApplier> = {
139
158
  "gemini-flash": (w, c) => applyGeminiFlash(w, c.thinkingLevel),
140
- "gemini-3.5-flash": (w, c) => applyGeminiFlash(w, c.thinkingLevel),
141
- "gemini-3.5-flash-low": (w, c) => applyGeminiFlash(w, c.thinkingLevel),
142
- "gemini-3.5-flash-extra-low": (w, c) => applyGeminiFlash(w, c.thinkingLevel),
159
+ "gemini-3.5-flash": (w, c) => applyGemini35Flash(w, c.thinkingLevel),
160
+ "gemini-3.5-flash-low": (w, c) => applyGemini35Flash(w, c.thinkingLevel),
161
+ "gemini-3.5-flash-extra-low": (w, c) => applyGemini35Flash(w, c.thinkingLevel),
143
162
  "gemini-2.5-flash": (w, c) => applyGeminiFlash(w, c.thinkingLevel),
144
163
  "gemini-2.5-flash-lite": (w, c) => applyGeminiFlash(w, c.thinkingLevel),
145
164
  "gemini-pro": (w, c) => applyGeminiPro(w, c.thinkingLevel),
146
165
  "gemini-3.1-pro-low": (w, c) => applyGeminiPro(w, c.thinkingLevel),
147
- "gemini-2.5-pro": (w, c) => applyGeminiPro(w, c.thinkingLevel),
166
+ "gemini-2.5-pro": (w, c) => applyGemini25Pro(w, c.thinkingLevel),
148
167
  };
@@ -1,10 +1,12 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
3
  import { isMaximum, type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
4
4
 
5
5
  /**
6
6
  * 智谱 AI(Zhipu)GLM 系列。
7
- * 官方文档:https://open.bigmodel.cn/(国际站 docs.z.ai)
7
+ * 国内官方依据:
8
+ * https://docs.bigmodel.cn/cn/guide/capabilities/thinking (effort 仅 5.2+)
9
+ * https://docs.bigmodel.cn/cn/guide/capabilities/thinking-mode (保留式思考)
8
10
  */
9
11
  export function glmModels(root: string): ProviderModelConfig[] {
10
12
  const v1 = `${root}/v1`;
@@ -130,28 +132,29 @@ export function glmModels(root: string): ProviderModelConfig[] {
130
132
  }
131
133
 
132
134
  // GLM 思维链:`tool_stream=true` 仅在流式请求时附加;高级档位加 `clear_thinking=false`。
133
- function applyGlm(writer: PayloadWriter, level: ThinkingLevel): void {
135
+ function applyGlm(writer: PayloadWriter, level: ThinkingLevel, effort = true, preserve = true): void {
136
+ if (!effort) writer.remove("reasoning_effort");
134
137
  if (writer.get("stream") === true) writer.set("tool_stream", true);
135
138
  if (level === "off") {
136
139
  writer.setThinking({ type: "disabled" });
137
- writer.set("reasoning_effort", "none");
140
+ if (effort) writer.set("reasoning_effort", "none");
138
141
  return;
139
142
  }
140
143
 
141
144
  writer.setThinking({ type: "enabled" });
142
145
  if (level === "high" || isMaximum(level)) {
143
- writer.setThinking({ clear_thinking: false });
144
- writer.set("reasoning_effort", isMaximum(level) ? "max" : "high");
146
+ if (preserve) writer.setThinking({ clear_thinking: false });
147
+ if (effort) writer.set("reasoning_effort", isMaximum(level) ? "max" : "high");
145
148
  }
146
149
  }
147
150
 
148
151
  export const glmThinking: Record<string, ThinkingApplier> = {
149
152
  "glm-5.2": (w, c) => applyGlm(w, c.thinkingLevel),
150
- "glm-5.1": (w, c) => applyGlm(w, c.thinkingLevel),
151
- "glm-5": (w, c) => applyGlm(w, c.thinkingLevel),
152
- "glm-5-turbo": (w, c) => applyGlm(w, c.thinkingLevel),
153
- "glm-4.7": (w, c) => applyGlm(w, c.thinkingLevel),
154
- "glm-4.6": (w, c) => applyGlm(w, c.thinkingLevel),
155
- "glm-4.5": (w, c) => applyGlm(w, c.thinkingLevel),
156
- "glm-4.5-air": (w, c) => applyGlm(w, c.thinkingLevel),
153
+ "glm-5.1": (w, c) => applyGlm(w, c.thinkingLevel, false),
154
+ "glm-5": (w, c) => applyGlm(w, c.thinkingLevel, false),
155
+ "glm-5-turbo": (w, c) => applyGlm(w, c.thinkingLevel, false),
156
+ "glm-4.7": (w, c) => applyGlm(w, c.thinkingLevel, false),
157
+ "glm-4.6": (w, c) => applyGlm(w, c.thinkingLevel, false, false),
158
+ "glm-4.5": (w, c) => applyGlm(w, c.thinkingLevel, false, false),
159
+ "glm-4.5-air": (w, c) => applyGlm(w, c.thinkingLevel, false, false),
157
160
  };
@@ -1,4 +1,4 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
3
  import { type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
4
4
 
@@ -91,8 +91,10 @@ export function grokModels(root: string): ProviderModelConfig[] {
91
91
  ];
92
92
  }
93
93
 
94
- // Grok 思维链:`reasoning_effort`(none/low/medium/high)。
95
- // 注意:grok-4.5 不支持禁用 reasoning(none),off 档映射到最小 effort "low"。
94
+ // 官方:https://docs.x.ai/developers/model-capabilities/text/reasoning
95
+ // 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 effort 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。
96
98
  function applyGrok(writer: PayloadWriter, level: ThinkingLevel): void {
97
99
  writer.remove("thinking");
98
100
  if (level === "off") {
@@ -1,10 +1,11 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
3
  import { applyEnabledThinking, type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
4
4
 
5
5
  /**
6
6
  * 月之暗面(Moonshot)Kimi 系列。
7
7
  * 官方文档:https://platform.moonshot.cn/docs
8
+ * K2.7 Code 参数:https://platform.moonshot.cn/docs/api/models-overview
8
9
  */
9
10
  export function kimiModels(root: string): ProviderModelConfig[] {
10
11
  const v1 = `${root}/v1`;
@@ -33,6 +34,7 @@ export function kimiModels(root: string): ProviderModelConfig[] {
33
34
  baseUrl: v1,
34
35
  reasoning: true,
35
36
  thinkingLevelMap: {
37
+ off: null,
36
38
  low: "low",
37
39
  medium: "medium",
38
40
  high: "high",
@@ -51,10 +53,11 @@ export function kimiModels(root: string): ProviderModelConfig[] {
51
53
  api: "openai-completions",
52
54
  baseUrl: v1,
53
55
  reasoning: true,
56
+ thinkingLevelMap: { off: null },
54
57
  input: ["text", "image"],
55
58
  cost: { input: 0.95, output: 4, cacheRead: 0.19, cacheWrite: 0 },
56
59
  contextWindow: 262144,
57
- // 官方未公布最大输出上限,保守取上下文窗口。
60
+ // 兼容性暂存值:独立输出上限未确认,并非官方确认的输出规格。
58
61
  maxTokens: 262144,
59
62
  compat: OPENAI_COMPLETIONS_COMPAT,
60
63
  },
@@ -64,10 +67,11 @@ export function kimiModels(root: string): ProviderModelConfig[] {
64
67
  api: "openai-completions",
65
68
  baseUrl: v1,
66
69
  reasoning: true,
70
+ thinkingLevelMap: { off: null },
67
71
  input: ["text", "image"],
68
72
  cost: { input: 1.9, output: 8, cacheRead: 0.38, cacheWrite: 0 },
69
73
  contextWindow: 262144,
70
- // 官方未公布最大输出上限,保守取上下文窗口。
74
+ // 兼容性暂存值:独立输出上限未确认,并非官方确认的输出规格。
71
75
  maxTokens: 262144,
72
76
  compat: OPENAI_COMPLETIONS_COMPAT,
73
77
  },
@@ -80,7 +84,7 @@ export function kimiModels(root: string): ProviderModelConfig[] {
80
84
  input: ["text", "image"],
81
85
  cost: { input: 0.95, output: 4, cacheRead: 0.16, cacheWrite: 0 },
82
86
  contextWindow: 262144,
83
- // 官方未公布最大输出上限,保守取上下文窗口。
87
+ // 兼容性暂存值:独立输出上限未确认,并非官方确认的输出规格。
84
88
  maxTokens: 262144,
85
89
  compat: OPENAI_COMPLETIONS_COMPAT,
86
90
  },
@@ -93,7 +97,7 @@ export function kimiModels(root: string): ProviderModelConfig[] {
93
97
  input: ["text", "image"],
94
98
  cost: { input: 0.6, output: 3, cacheRead: 0.1, cacheWrite: 0 },
95
99
  contextWindow: 262144,
96
- // 官方未公布最大输出上限,保守取上下文窗口。
100
+ // 兼容性暂存值:独立输出上限未确认,并非官方确认的输出规格。
97
101
  maxTokens: 262144,
98
102
  compat: OPENAI_COMPLETIONS_COMPAT,
99
103
  },
@@ -103,22 +107,25 @@ export function kimiModels(root: string): ProviderModelConfig[] {
103
107
  // Kimi K3 思维链:移除 `thinking` 字段,改用 `reasoning_effort` 表达深度。
104
108
  function applyKimiK3(writer: PayloadWriter, level: ThinkingLevel): void {
105
109
  writer.remove("thinking");
106
- if (level === "off") {
107
- writer.remove("reasoning_effort");
108
- return;
109
- }
110
- if (level === "minimal" || level === "low") {
110
+ // K3 始终思考,省略 effort 默认 max;直接或迟到的 off 请求降至 low。
111
+ if (level === "off" || level === "minimal" || level === "low") {
111
112
  writer.set("reasoning_effort", "low");
112
113
  return;
113
114
  }
114
115
  writer.set("reasoning_effort", level === "max" ? "max" : "high");
115
116
  }
116
117
 
118
+ // canonical Code 与 highspeed 为同模型:不可关闭,显式 thinking 仅接受此组合。
119
+ function applyKimiCode(writer: PayloadWriter): void {
120
+ writer.remove("reasoning_effort");
121
+ writer.setThinking({ type: "enabled", keep: "all" });
122
+ }
123
+
117
124
  export const kimiThinking: Record<string, ThinkingApplier> = {
118
125
  "kimi-for-coding": (w, c) => applyEnabledThinking(w, c.thinkingLevel),
119
126
  "kimi-k3": (w, c) => applyKimiK3(w, c.thinkingLevel),
120
- "kimi-k2.7-code": (w, c) => applyEnabledThinking(w, c.thinkingLevel),
121
- "kimi-k2.7-code-highspeed": (w, c) => applyEnabledThinking(w, c.thinkingLevel),
127
+ "kimi-k2.7-code": (w) => applyKimiCode(w),
128
+ "kimi-k2.7-code-highspeed": (w) => applyKimiCode(w),
122
129
  "kimi-k2.6": (w, c) => applyEnabledThinking(w, c.thinkingLevel),
123
130
  "kimi-k2.5": (w, c) => applyEnabledThinking(w, c.thinkingLevel),
124
131
  };
@@ -1,4 +1,4 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
3
  import { applyEnabledThinking, type ThinkingApplier } from "../_tools.ts";
4
4
 
@@ -1,4 +1,4 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
3
  import { applyEnabledThinking, type ThinkingApplier } from "../_tools.ts";
4
4
 
@@ -1,10 +1,12 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
3
  import { type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
4
4
 
5
5
  /**
6
6
  * MiniMax 系列。
7
- * 官方文档:https://platform.minimaxi.com/(国际站 platform.minimax.io)
7
+ * 国内官方依据:
8
+ * https://platform.minimaxi.com/docs/api-reference/text-chat (输出上限)
9
+ * https://platform.minimaxi.com/docs/api-reference/text-openai-api (M2.x 强制思考)
8
10
  */
9
11
  export function minimaxModels(root: string): ProviderModelConfig[] {
10
12
  const v1 = `${root}/v1`;
@@ -23,7 +25,7 @@ export function minimaxModels(root: string): ProviderModelConfig[] {
23
25
  cacheWrite: 0,
24
26
  },
25
27
  contextWindow: 1000000,
26
- maxTokens: 500000,
28
+ maxTokens: 524288,
27
29
  compat: OPENAI_COMPLETIONS_COMPAT,
28
30
  },
29
31
  {
@@ -32,10 +34,11 @@ export function minimaxModels(root: string): ProviderModelConfig[] {
32
34
  api: "openai-completions",
33
35
  baseUrl: v1,
34
36
  reasoning: true,
37
+ thinkingLevelMap: { off: null },
35
38
  input: ["text"],
36
39
  cost: { input: 0.3, output: 1.2, cacheRead: 0.06, cacheWrite: 0.375 },
37
40
  contextWindow: 204800,
38
- // 官方未公布最大输出上限,保守取上下文窗口。
41
+ // 官方 max_completion_tokens 上限,并非上下文估算。
39
42
  maxTokens: 204800,
40
43
  compat: OPENAI_COMPLETIONS_COMPAT,
41
44
  },
@@ -45,10 +48,11 @@ export function minimaxModels(root: string): ProviderModelConfig[] {
45
48
  api: "openai-completions",
46
49
  baseUrl: v1,
47
50
  reasoning: true,
51
+ thinkingLevelMap: { off: null },
48
52
  input: ["text"],
49
53
  cost: { input: 0.3, output: 1.2, cacheRead: 0.03, cacheWrite: 0.375 },
50
54
  contextWindow: 204800,
51
- // 官方未公布最大输出上限,保守取上下文窗口。
55
+ // 官方 max_completion_tokens 上限,并非上下文估算。
52
56
  maxTokens: 204800,
53
57
  compat: OPENAI_COMPLETIONS_COMPAT,
54
58
  },
@@ -58,10 +62,11 @@ export function minimaxModels(root: string): ProviderModelConfig[] {
58
62
  api: "openai-completions",
59
63
  baseUrl: v1,
60
64
  reasoning: true,
65
+ thinkingLevelMap: { off: null },
61
66
  input: ["text"],
62
67
  cost: { input: 0.3, output: 1.2, cacheRead: 0.03, cacheWrite: 0.375 },
63
68
  contextWindow: 204800,
64
- // 官方未公布最大输出上限,保守取上下文窗口。
69
+ // 官方 max_completion_tokens 上限,并非上下文估算。
65
70
  maxTokens: 204800,
66
71
  compat: OPENAI_COMPLETIONS_COMPAT,
67
72
  },
@@ -69,7 +74,7 @@ export function minimaxModels(root: string): ProviderModelConfig[] {
69
74
  }
70
75
 
71
76
  // MiniMax 思维链:`reasoning_split=true`,档位映射 disabled/adaptive。
72
- // M2.x 沿用 M3 策略(同族,未逐一验证)。
77
+ // 国内 OpenAI SDK 文档明确区分:M3 可关闭;M2.x disabled 被接受但不生效。
73
78
  function applyMiniMax(writer: PayloadWriter, level: ThinkingLevel): void {
74
79
  writer.remove("reasoning_effort");
75
80
  writer.set("reasoning_split", true);
@@ -78,7 +83,7 @@ function applyMiniMax(writer: PayloadWriter, level: ThinkingLevel): void {
78
83
 
79
84
  export const minimaxThinking = {
80
85
  "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),
86
+ "MiniMax-M2.7": (w) => applyMiniMax(w, "high"),
87
+ "MiniMax-M2.5": (w) => applyMiniMax(w, "high"),
88
+ "MiniMax-M2.1": (w) => applyMiniMax(w, "high"),
84
89
  } satisfies Record<string, ThinkingApplier>;
@@ -1,4 +1,4 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { type PayloadWriter, type ThinkingApplier } from "../_tools.ts";
3
3
 
4
4
  /**
@@ -1,4 +1,4 @@
1
- import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
1
+ import type { ChatModelConfig as ProviderModelConfig } from "../_tools.ts";
2
2
  import { OPENAI_COMPLETIONS_COMPAT } from "./_protocols.ts";
3
3
  import { isAtLeastHigh, type PayloadWriter, type ThinkingApplier, type ThinkingLevel } from "../_tools.ts";
4
4
 
@@ -92,10 +92,10 @@ function applyGrokSearch(writer: PayloadWriter): void {
92
92
 
93
93
  /**
94
94
  * Append-only built-in search injection for models with native search.
95
- * Non-plain payloads, `off` mode, and unsupported models are returned
95
+ * Non-plain payloads and unsupported models are returned
96
96
  * unchanged. Existing function tools, `tool_choice`, and `include` are
97
- * preserved. Pi drops requested sources, so GPT injection deliberately does
98
- * not request them.
97
+ * preserved. Sources are not requested: this conservative policy originated
98
+ * with Pi 0.82.0; source rendering on newer hosts needs separate validation.
99
99
  */
100
100
  export function applyBuiltinSearchTool<T>(
101
101
  payload: T,
@@ -103,7 +103,7 @@ export function applyBuiltinSearchTool<T>(
103
103
  api: string,
104
104
  mode: WebSearchMode,
105
105
  ): T | Payload {
106
- if (!isPlainObject(payload) || mode === "off") return payload;
106
+ if (!isPlainObject(payload)) return payload;
107
107
  const writer = new PayloadWriter(payload);
108
108
  const support = builtinSearchSupport(modelId, api);
109
109
 
@@ -82,8 +82,7 @@ export function searchLabel(
82
82
  support: "gpt" | "grok" | undefined,
83
83
  ): string {
84
84
  if (support === undefined) return "⚪ 内置联网:不支持";
85
- if (mode === "off") return "⚪ 内置联网:关";
86
- // Grok 的 cached/live 均注入 mode: "on",不能声称是缓存查询。
85
+ // Grok 固定实时,不受 GPT 的 cached/live 会话模式影响。
87
86
  if (support === "grok" || mode === "live") return "🟢 内置联网:实时";
88
87
  return "🟡 内置联网:缓存";
89
88
  }
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "pi-tsgw",
3
- "version": "0.0.11",
3
+ "version": "0.0.12",
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": [
@@ -21,14 +21,14 @@
21
21
  ]
22
22
  },
23
23
  "scripts": {
24
- "test": "tsc -p tsconfig.test.json && node dist/extensions/index.test.js && node dist/extensions/status.test.js && node dist/extensions/status-integration.test.js && node dist/extensions/models/gateway-catalog-metadata.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"
24
+ "test": "tsc -p tsconfig.test.json && node dist/extensions/index.test.js && node dist/extensions/status.test.js && node dist/extensions/status-integration.test.js && node dist/extensions/models/gateway-catalog-metadata.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/host-integration.test.js"
25
25
  },
26
26
  "peerDependencies": {
27
27
  "@earendil-works/pi-coding-agent": ">=0.82.0",
28
28
  "typebox": "*"
29
29
  },
30
30
  "devDependencies": {
31
- "@earendil-works/pi-coding-agent": "^0.84.1",
31
+ "@earendil-works/pi-coding-agent": "^1.1.0",
32
32
  "typebox": "^1.3.11",
33
33
  "typescript": "^7.0.2"
34
34
  }