dsh-lcx-codex 0.6.0 → 0.7.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -22,29 +22,48 @@ function normalizeCacheRetention(value) {
22
22
  ? normalized
23
23
  : "short";
24
24
  }
25
+ function isOpenAIDirectSignInTarget(model) {
26
+ return (isRecord(model) &&
27
+ model.provider === "openai" &&
28
+ model.baseUrl === "https://api.openai.com/v1");
29
+ }
30
+ /**
31
+ * Matches Pi's `isChatGPTSignIn`: direct `openai` against `https://api.openai.com/v1` with a non-`sk-` token.
32
+ */
33
+ export function isChatGPTSignInCredential(model, apiKey) {
34
+ if (!isRecord(model))
35
+ return false;
36
+ return (isOpenAIDirectSignInTarget(model) &&
37
+ apiKey !== undefined &&
38
+ !apiKey.startsWith("sk-"));
39
+ }
25
40
  /**
26
41
  * The effort Pi 1.3.0 sends as `configuration_update` input items instead of `reasoning.effort`, or undefined
27
42
  * to send it as `reasoning.effort` (Pi's `getConfigurationUpdateEffort`): only for a model whose compat states
28
43
  * `supportsMidConvoEffort`, only with a requested level, and never for a Sign in with ChatGPT key, which rejects
29
44
  * those items. The key is read only when that last rule can apply.
30
45
  */
31
- export async function configurationUpdateEffort(model, reasoningEffort, apiKey) {
32
- if (model.compat?.supportsMidConvoEffort !== true || reasoningEffort === undefined || reasoningEffort === "off")
46
+ export async function configurationUpdateEffort(model, reasoningEffort, apiKey, options) {
47
+ if (options?.gatewayCompatibilityFallback === true ||
48
+ model.compat?.supportsMidConvoEffort !== true ||
49
+ reasoningEffort === undefined ||
50
+ reasoningEffort === "off")
33
51
  return undefined;
34
52
  const requested = String(reasoningEffort);
35
53
  const mapped = model.thinkingLevelMap?.[requested];
36
54
  // A level the model maps to null is unsupported; DSH refuses it before dispatch, so it never travels here.
37
55
  if (mapped === null)
38
56
  return undefined;
39
- if (model.provider === "openai" && model.baseUrl === "https://api.openai.com/v1" && !(await apiKey()).startsWith("sk-"))
57
+ if (isOpenAIDirectSignInTarget(model) && isChatGPTSignInCredential(model, await apiKey()))
40
58
  return undefined;
41
59
  return mapped ?? requested;
42
60
  }
43
61
  /**
44
62
  * Pi-parity generation controls for OpenAI Responses.
45
63
  */
46
- export function responsesGenerationEnvelope({ model = "unknown", reasoningEffort, temperature, maxTokens, configurationUpdateEffort, includeDefaultReasoning = false, } = {}) {
64
+ export function responsesGenerationEnvelope({ model = "unknown", reasoningEffort, temperature, maxTokens, configurationUpdateEffort, gatewayCompatibilityFallback = false, omitUnsupportedFields = false, apiKey, includeDefaultReasoning = false, } = {}) {
47
65
  const descriptor = modelRecord(model);
66
+ const suppressChatGptUnsupported = omitUnsupportedFields === true || isChatGPTSignInCredential(descriptor, apiKey);
48
67
  const result = {};
49
68
  if (descriptor.reasoning !== false) {
50
69
  if (descriptor.provider === "xai" ||
@@ -59,7 +78,11 @@ export function responsesGenerationEnvelope({ model = "unknown", reasoningEffort
59
78
  if (wire !== null) {
60
79
  // With mid-conversation effort the request carries a fixed effort, so an effort change keeps the
61
80
  // cached prefix; the level itself is the last `configuration_update` input item.
62
- result.reasoning = { effort: configurationUpdateEffort === undefined ? wire : CONFIGURATION_UPDATE_REQUEST_EFFORT, summary: "auto" };
81
+ const useConfigurationUpdate = !gatewayCompatibilityFallback && configurationUpdateEffort !== undefined;
82
+ result.reasoning = {
83
+ effort: useConfigurationUpdate ? CONFIGURATION_UPDATE_REQUEST_EFFORT : wire,
84
+ summary: "auto",
85
+ };
63
86
  result.include = ["reasoning.encrypted_content"];
64
87
  }
65
88
  }
@@ -74,37 +97,46 @@ export function responsesGenerationEnvelope({ model = "unknown", reasoningEffort
74
97
  throw Object.assign(new Error("Responses temperature must be finite"), {
75
98
  code: "LCX_RESPONSES_INVALID_INPUT",
76
99
  });
77
- result.temperature = Number(temperature);
100
+ if (!suppressChatGptUnsupported)
101
+ result.temperature = Number(temperature);
78
102
  }
79
103
  if (maxTokens !== undefined) {
80
104
  if (!Number.isSafeInteger(maxTokens) || Number(maxTokens) <= 0)
81
105
  throw Object.assign(new Error("Responses maxTokens must be a positive safe integer"), { code: "LCX_RESPONSES_INVALID_INPUT" });
82
- if (descriptor.compat?.supportsMaxOutputTokens !== false)
106
+ if (!suppressChatGptUnsupported && descriptor.compat?.supportsMaxOutputTokens !== false)
83
107
  result.max_output_tokens = Math.max(OPENAI_RESPONSES_MIN_OUTPUT_TOKENS, Number(maxTokens));
84
108
  }
85
109
  return result;
86
110
  }
87
111
  /**
88
- * Build the shared LCX-owned request envelope while retaining Pi 0.86.0 Responses semantics.
112
+ * Build the shared LCX-owned request envelope while retaining Pi Responses semantics.
89
113
  * Production LCX ordinary/compact/replay construction places the DSH system prompt in canonical input.
90
114
  */
91
- export function buildResponsesBody({ model, input, tools, sessionId, promptCacheKey, promptCacheRetention, cacheRetention, reasoningEffort, temperature, maxTokens, configurationUpdateEffort, samplingParams, }) {
115
+ export function buildResponsesBody({ model, input, tools, sessionId, promptCacheKey, promptCacheRetention, cacheRetention, reasoningEffort, temperature, maxTokens, configurationUpdateEffort, gatewayCompatibilityFallback, omitUnsupportedFields, apiKey, samplingParams, }) {
92
116
  if (!Array.isArray(input))
93
117
  throw Object.assign(new Error("Responses input must be an array"), {
94
118
  code: "LCX_RESPONSES_INVALID_INPUT",
95
119
  });
96
120
  const descriptor = modelRecord(model);
121
+ const suppressChatGptUnsupported = omitUnsupportedFields === true || isChatGPTSignInCredential(descriptor, apiKey);
97
122
  const retention = normalizeCacheRetention(cacheRetention);
98
123
  const compat = descriptor.compat ?? {};
99
- const cacheKey = retention === "none"
124
+ const supportsLongCache = compat.supportsLongCacheRetention !== false;
125
+ const sendPromptCacheKey = retention !== "none" && compat.promptCacheKeyMode !== "disabled";
126
+ const cacheKey = !sendPromptCacheKey
100
127
  ? undefined
101
128
  : (promptCacheKey ?? clampOpenAIPromptCacheKey(sessionId));
102
- const longRetention = retention === "long" && compat.supportsExplicitPromptCacheMode !== true
103
- ? promptCacheRetention ?? (compat.supportsLongCacheRetention === true ? "24h" : undefined)
129
+ const longRetention = !suppressChatGptUnsupported && retention === "long" && compat.supportsExplicitPromptCacheMode !== true
130
+ ? promptCacheRetention ?? (supportsLongCache ? "24h" : undefined)
104
131
  : undefined;
105
- // Pi 0.86.0 uses explicit 30m only for long retention; short is implicit.
106
- const currentCache = retention === "long" && compat.supportsLongCacheRetention === true && compat.supportsExplicitPromptCacheMode === true;
107
- const explicitCache = retention === "none" && compat.supportsExplicitPromptCacheMode === true;
132
+ // Pi uses explicit 30m only for long retention; short is implicit.
133
+ const currentCache = !suppressChatGptUnsupported &&
134
+ retention === "long" &&
135
+ supportsLongCache &&
136
+ compat.supportsExplicitPromptCacheMode === true;
137
+ const explicitCache = !suppressChatGptUnsupported &&
138
+ retention === "none" &&
139
+ compat.supportsExplicitPromptCacheMode === true;
108
140
  const promptCacheOptions = currentCache
109
141
  ? { ttl: "30m" }
110
142
  : explicitCache
@@ -128,6 +160,8 @@ export function buildResponsesBody({ model, input, tools, sessionId, promptCache
128
160
  temperature,
129
161
  maxTokens,
130
162
  configurationUpdateEffort,
163
+ gatewayCompatibilityFallback,
164
+ omitUnsupportedFields: suppressChatGptUnsupported,
131
165
  includeDefaultReasoning: true,
132
166
  }),
133
167
  };
@@ -3,7 +3,8 @@ import { abortError } from "./abort-error.js";
3
3
  import { IMAGE_OFFLOAD_REQUIRED_CODE, ToolCallId, isContextWindowExceededError } from "@deepseek-ai/dsh-llm";
4
4
  import { fetchSseWithRetry, readSseEvents } from "./transport.js";
5
5
  import { sanitizeContextDetails } from "./search-accounting.js";
6
- import { createGrokNativeReplayEnvelope, grokPendingSourcesStart, sanitizeGrokVisibleText, } from "./grok-native-search.js";
6
+ import { baseURLFingerprint } from "./route.js";
7
+ import { createGrokNativeReplayEnvelope, grokNativeReplayRouteFingerprint, grokPendingSourcesStart, sanitizeGrokVisibleText, } from "./grok-native-search.js";
7
8
  import { codedError, isPositiveFinite, isRecord } from "./util.js";
8
9
  import { DEFAULTS } from "./defaults.js";
9
10
  /** Durable LCX-only pressure evidence; canonical usage stays on Pi/DSH billing. */
@@ -183,7 +184,11 @@ export function managedFailure(error, signal, upstream) {
183
184
  /time(?:d)?\s*out|timeout/iu.test(String(signalReason?.message ?? "")));
184
185
  const upstreamClass = signal?.aborted ? undefined : classifyUpstreamFailure(upstream);
185
186
  let code = "RESPONSES_ERROR";
186
- if (upstreamClass?.definite === true)
187
+ if (sourceCodes.has("REQUEST_HEADERS"))
188
+ code = "REQUEST_HEADERS";
189
+ else if (sourceCodes.has("MALFORMED_RESPONSE"))
190
+ code = "MALFORMED_RESPONSE";
191
+ else if (upstreamClass?.definite === true)
187
192
  code = upstreamClass.code;
188
193
  else if (signalTimedOut ||
189
194
  facts.status === 408 ||
@@ -246,6 +251,8 @@ export function managedFailure(error, signal, upstream) {
246
251
  if (code === "RESPONSES_ERROR" && upstreamClass !== undefined)
247
252
  code = upstreamClass.code;
248
253
  const messages = {
254
+ MALFORMED_RESPONSE: "Responses provider completed a tool call with malformed arguments",
255
+ REQUEST_HEADERS: error instanceof Error ? error.message : "Request header preparation failed",
249
256
  ABORTED: "Responses request was aborted",
250
257
  AUTH: "Responses request was rejected by authentication",
251
258
  INSUFFICIENT_QUOTA: "服务商额度不足或未满足请求预留额度,请检查中转额度(并非 API 密钥无效) / Provider quota is insufficient for this request",
@@ -1113,7 +1120,7 @@ function rawArguments(value) {
1113
1120
  return "{}";
1114
1121
  }
1115
1122
  }
1116
- function replayState(message, nativeOutput, nativeReplayRoute, nativeServerSearchEchoCallIds = []) {
1123
+ function replayState(message, nativeOutput, nativeReplayRoute, nativeServerSearchEchoCallIds = [], baseURL, sessionId, headerAuthorityFingerprint) {
1117
1124
  if (!isRecord(message))
1118
1125
  return undefined;
1119
1126
  const content = Array.isArray(message.content)
@@ -1125,6 +1132,24 @@ function replayState(message, nativeOutput, nativeReplayRoute, nativeServerSearc
1125
1132
  if (!provider || !model || !api)
1126
1133
  return undefined;
1127
1134
  const grokNative = createGrokNativeReplayEnvelope(nativeOutput, nativeReplayRoute, nativeServerSearchEchoCallIds);
1135
+ const normalizedBaseUrl = typeof baseURL === "string" ? baseURL.trim() : "";
1136
+ const normalizedSessionId = typeof sessionId === "string" ? sessionId : "";
1137
+ const lcxReplay = normalizedBaseUrl &&
1138
+ normalizedSessionId &&
1139
+ typeof headerAuthorityFingerprint === "string" &&
1140
+ headerAuthorityFingerprint.length > 0
1141
+ ? {
1142
+ version: 1,
1143
+ baseURLFingerprint: baseURLFingerprint(normalizedBaseUrl),
1144
+ sourceSessionId: normalizedSessionId,
1145
+ headerAuthorityFingerprint,
1146
+ ...(nativeReplayRoute
1147
+ ? {
1148
+ grokRouteAuthorityFingerprint: grokNativeReplayRouteFingerprint(nativeReplayRoute),
1149
+ }
1150
+ : {}),
1151
+ }
1152
+ : undefined;
1128
1153
  return {
1129
1154
  response: {
1130
1155
  kind: "pi-ai",
@@ -1171,6 +1196,7 @@ function replayState(message, nativeOutput, nativeReplayRoute, nativeServerSearc
1171
1196
  : {}),
1172
1197
  };
1173
1198
  }),
1199
+ ...(lcxReplay === undefined ? {} : { lcxReplay }),
1174
1200
  ...(grokNative === undefined ? {} : { grokNative }),
1175
1201
  };
1176
1202
  }
@@ -1198,8 +1224,9 @@ function successfulFinish(message) {
1198
1224
  /**
1199
1225
  * Pi event vocabulary -> DSH StreamChunk. This remains thin and provider-neutral.
1200
1226
  */
1201
- async function* toDshChunks(events, signal, wireMeta, nativeReplayRoute) {
1227
+ async function* toDshChunks(events, signal, wireMeta, nativeReplayRoute, baseURL, sessionId, headerAuthorityFingerprint) {
1202
1228
  const toolIds = new Map();
1229
+ let argumentsError = false;
1203
1230
  for await (const event of events) {
1204
1231
  switch (event.type) {
1205
1232
  case "start":
@@ -1282,6 +1309,10 @@ async function* toDshChunks(events, signal, wireMeta, nativeReplayRoute) {
1282
1309
  const call = isRecord(event.toolCall)
1283
1310
  ? event.toolCall
1284
1311
  : {};
1312
+ if (call.argumentsError !== undefined) {
1313
+ argumentsError = true;
1314
+ break;
1315
+ }
1285
1316
  yield {
1286
1317
  type: "block-end",
1287
1318
  index: Number(event.contentIndex),
@@ -1299,10 +1330,12 @@ async function* toDshChunks(events, signal, wireMeta, nativeReplayRoute) {
1299
1330
  ? event.message
1300
1331
  : {};
1301
1332
  yield { type: "usage", usage: dshUsage(message.usage) };
1333
+ if (argumentsError && message.stopReason !== "length")
1334
+ throw codedError("Responses provider completed a tool call with malformed arguments", "MALFORMED_RESPONSE");
1302
1335
  const reason = successfulFinish(message);
1303
1336
  const replay = reason.kind === "error"
1304
1337
  ? undefined
1305
- : replayState(message, wireMeta?.nativeOutput, nativeReplayRoute, wireMeta?.nativeServerSearchEchoCallIds);
1338
+ : replayState(message, wireMeta?.nativeOutput, nativeReplayRoute, wireMeta?.nativeServerSearchEchoCallIds, baseURL, sessionId, headerAuthorityFingerprint);
1306
1339
  yield {
1307
1340
  type: "finish",
1308
1341
  reason,
@@ -1352,7 +1385,7 @@ async function* toDshChunks(events, signal, wireMeta, nativeReplayRoute) {
1352
1385
  * Send one LCX-owned OpenAI Responses request. Ordinary and replay use one provider attempt;
1353
1386
  * the DSH agent recovery layer remains the visible retry owner.
1354
1387
  */
1355
- export async function* streamResponsesRequest({ baseURL, provider, model, piModel, body, grammarToolInputProperties, headers, signal, timeoutMs, applyDefaultTimeout = true, streamIdleTimeoutMs, maxAttempts = 1, maxResponseBytes, serverToolTypes, isServerToolItem, declaredToolNames, nativeReplayRoute, providerThinkingLevel, onServerToolUsage, onFailure, }) {
1388
+ export async function* streamResponsesRequest({ baseURL, provider, model, piModel, body, grammarToolInputProperties, headers, signal, timeoutMs, applyDefaultTimeout = true, streamIdleTimeoutMs, maxAttempts = 1, maxResponseBytes, serverToolTypes, isServerToolItem, declaredToolNames, sessionId, headerAuthorityFingerprint, nativeReplayRoute, providerThinkingLevel, onServerToolUsage, onFailure, }) {
1356
1389
  let contentEmitted = false;
1357
1390
  let failureReported = false;
1358
1391
  /** One redacted report per request; user aborts are not failures. */
@@ -1436,7 +1469,7 @@ export async function* streamResponsesRequest({ baseURL, provider, model, piMode
1436
1469
  piEvents.end();
1437
1470
  }
1438
1471
  })();
1439
- const chunks = toDshChunks(piEvents, signal, wireMeta, nativeReplayRoute)[Symbol.asyncIterator]();
1472
+ const chunks = toDshChunks(piEvents, signal, wireMeta, nativeReplayRoute, baseURL, sessionId, headerAuthorityFingerprint)[Symbol.asyncIterator]();
1440
1473
  let exhausted = false;
1441
1474
  try {
1442
1475
  while (true) {