@prestyj/ai 5.28.1 → 5.29.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/index.cjs CHANGED
@@ -35,14 +35,17 @@ __export(index_exports, {
35
35
  ProviderError: () => ProviderError,
36
36
  REDACTION_MARKER: () => REDACTED,
37
37
  StreamResult: () => StreamResult,
38
+ TOOL_CALL_NAME_RULES: () => TOOL_CALL_NAME_RULES,
38
39
  clampProviderContextImages: () => clampProviderContextImages,
39
40
  classifyProviderError: () => classifyProviderError,
41
+ dropInvalidToolCalls: () => dropInvalidToolCalls,
40
42
  environmentSecrets: () => environmentSecrets,
41
43
  formatError: () => formatError,
42
44
  formatErrorForDisplay: () => formatErrorForDisplay,
43
45
  hasLoneSurrogate: () => hasLoneSurrogate,
44
46
  isHardBillingMessage: () => isHardBillingMessage,
45
47
  isUsageLimitError: () => isUsageLimitError,
48
+ isValidToolCallName: () => isValidToolCallName,
46
49
  localWireModelId: () => localWireModelId,
47
50
  palsuAssistantMessage: () => palsuAssistantMessage,
48
51
  palsuText: () => palsuText,
@@ -61,7 +64,8 @@ __export(index_exports, {
61
64
  stream: () => stream,
62
65
  toAnthropicMessages: () => toAnthropicMessages,
63
66
  toOpenAIMessages: () => toOpenAIMessages,
64
- toWellFormedText: () => toWellFormedText
67
+ toWellFormedText: () => toWellFormedText,
68
+ toolCallNameRuleFor: () => toolCallNameRuleFor
65
69
  });
66
70
  module.exports = __toCommonJS(index_exports);
67
71
 
@@ -515,7 +519,11 @@ function zodToJsonSchema(schema) {
515
519
  return normalized;
516
520
  }
517
521
  function resolveToolSchema(tool) {
518
- return tool.rawInputSchema ?? zodToJsonSchema(tool.parameters);
522
+ const schema = tool.rawInputSchema ?? zodToJsonSchema(tool.parameters);
523
+ if (schema.type === "object" && schema.properties === void 0) {
524
+ return { ...schema, properties: {} };
525
+ }
526
+ return schema;
519
527
  }
520
528
  function normalizeRootForAnthropic(schema) {
521
529
  const branches = schema.oneOf ?? schema.anyOf;
@@ -722,6 +730,10 @@ var ANTHROPIC_INPUT_BLOCK_TYPES = /* @__PURE__ */ new Set([
722
730
  "connector_text",
723
731
  "container_upload",
724
732
  "document",
733
+ // Server-side refusal fallback marker. Only reaches the wire when the request
734
+ // carries the server-side-fallback beta; otherwise applyServerFallbackReplay
735
+ // strips it first (see toAnthropicMessages' `fallbackBlocks` option).
736
+ "fallback",
725
737
  "image",
726
738
  "mid_conv_system",
727
739
  "redacted_thinking",
@@ -743,6 +755,42 @@ function isPositionSensitiveThinking(part) {
743
755
  if (part.type === "thinking") return hasValidThinkingSignature(part);
744
756
  return isRawThinking(part);
745
757
  }
758
+ function isServerFallbackBlock(part) {
759
+ return part.type === "raw" && part.data.type === "fallback";
760
+ }
761
+ function applyServerFallbackReplay(content, keepMarkers, droppedToolCallIds) {
762
+ let lastFallbackIdx = -1;
763
+ content.forEach((part, idx) => {
764
+ if (isServerFallbackBlock(part)) lastFallbackIdx = idx;
765
+ });
766
+ if (lastFallbackIdx === -1) return content;
767
+ const resultIds = /* @__PURE__ */ new Set();
768
+ for (const part of content) {
769
+ if (part.type === "server_tool_result") resultIds.add(part.toolUseId);
770
+ else if (part.type === "raw" && typeof part.data.tool_use_id === "string")
771
+ resultIds.add(part.data.tool_use_id);
772
+ }
773
+ const out = [];
774
+ content.forEach((part, idx) => {
775
+ if (isServerFallbackBlock(part)) {
776
+ if (keepMarkers) out.push(part);
777
+ return;
778
+ }
779
+ if (idx < lastFallbackIdx) {
780
+ if (part.type === "thinking" || isRawThinking(part)) return;
781
+ if (part.type === "raw" && part.data.type === "connector_text") return;
782
+ if (part.type === "tool_call") {
783
+ droppedToolCallIds?.add(part.id);
784
+ return;
785
+ }
786
+ if (part.type === "server_tool_call" && !resultIds.has(part.id)) return;
787
+ if (part.type === "raw" && part.data.type === "server_tool_use" && !resultIds.has(String(part.data.id)))
788
+ return;
789
+ }
790
+ out.push(part);
791
+ });
792
+ return out.every(isServerFallbackBlock) ? [] : out;
793
+ }
746
794
  function toAnthropicAssistantPart(part, idMap) {
747
795
  if (part.type === "text") return { type: "text", text: part.text };
748
796
  if (part.type === "thinking") {
@@ -810,10 +858,17 @@ function countContextImages(messages) {
810
858
  }
811
859
  return count;
812
860
  }
861
+ var IMAGE_DROP_BATCH = 30;
862
+ function providerImageDropCount(imageCount2, budget) {
863
+ const overflow = imageCount2 - budget;
864
+ if (overflow <= 0) return 0;
865
+ const batch = Math.max(1, Math.min(IMAGE_DROP_BATCH, Math.floor(budget / 3)));
866
+ return Math.min(imageCount2, Math.ceil(overflow / batch) * batch);
867
+ }
813
868
  function clampProviderContextImages(messages, provider, supportsImages) {
814
869
  if (supportsImages === false) return messages;
815
870
  const budget = PROVIDER_IMAGE_BUDGETS[provider] ?? 5;
816
- let remainingToRemove = countContextImages(messages) - budget;
871
+ let remainingToRemove = providerImageDropCount(countContextImages(messages), budget);
817
872
  if (remainingToRemove <= 0) return messages;
818
873
  return messages.map((message) => {
819
874
  if (message.role === "user" && Array.isArray(message.content)) {
@@ -847,6 +902,144 @@ function clampProviderContextImages(messages, provider, supportsImages) {
847
902
  return message;
848
903
  });
849
904
  }
905
+ var TOOL_CALL_NAME_RULES = {
906
+ // Anthropic Messages API: tool / tool_use name `^[a-zA-Z0-9_-]{1,128}$`.
907
+ anthropic: { id: "anthropic", maxLength: 128, pattern: /^[a-zA-Z0-9_-]{1,128}$/ },
908
+ // OpenAI Chat Completions: function name a-z A-Z 0-9 _ -, max length 64.
909
+ "openai-chat": { id: "openai-chat", maxLength: 64, pattern: /^[a-zA-Z0-9_-]{1,64}$/ },
910
+ // OpenAI Responses (Codex): same charset; replayed `input[N].name` over 128
911
+ // chars is rejected with `string_above_max_length`.
912
+ "openai-responses": {
913
+ id: "openai-responses",
914
+ maxLength: 128,
915
+ pattern: /^[a-zA-Z0-9_-]{1,128}$/
916
+ },
917
+ // Gemini FunctionDeclaration name: starts with a letter or underscore, then
918
+ // a-z A-Z 0-9 _ . : -, max length 128.
919
+ gemini: { id: "gemini", maxLength: 128, pattern: /^[a-zA-Z_][a-zA-Z0-9_.:-]{0,127}$/ },
920
+ // OpenAI-compatible third parties (GLM, Kimi, DeepSeek, OpenRouter, xAI,
921
+ // local servers, MiniMax, …) don't share one documented charset, so only
922
+ // reject what no declared tool name can contain: blank, >128 chars, or any
923
+ // whitespace / control character (the signature of invocation text).
924
+ generic: { id: "generic", maxLength: 128 }
925
+ };
926
+ var TOOL_NAME_FORBIDDEN_CHARS = /[\s\p{Cc}]/u;
927
+ function isValidToolCallName(name, rule) {
928
+ if (typeof name !== "string" || name.trim().length === 0) return false;
929
+ if (name.length > rule.maxLength) return false;
930
+ if (TOOL_NAME_FORBIDDEN_CHARS.test(name)) return false;
931
+ return rule.pattern ? rule.pattern.test(name) : true;
932
+ }
933
+ function isDefaultOrHost(baseUrl, host) {
934
+ if (!baseUrl) return true;
935
+ try {
936
+ return new URL(baseUrl).hostname === host;
937
+ } catch {
938
+ return false;
939
+ }
940
+ }
941
+ function toolCallNameRuleFor(provider, options) {
942
+ switch (provider) {
943
+ case "anthropic":
944
+ return TOOL_CALL_NAME_RULES.anthropic;
945
+ case "openai":
946
+ if (options?.accountId) return TOOL_CALL_NAME_RULES["openai-responses"];
947
+ return isDefaultOrHost(options?.baseUrl, "api.openai.com") ? TOOL_CALL_NAME_RULES["openai-chat"] : TOOL_CALL_NAME_RULES.generic;
948
+ case "gemini":
949
+ return TOOL_CALL_NAME_RULES.gemini;
950
+ default:
951
+ return TOOL_CALL_NAME_RULES.generic;
952
+ }
953
+ }
954
+ function isValidToolCallId(id) {
955
+ return typeof id === "string" && id.trim().length > 0;
956
+ }
957
+ function isRawReasoning(part) {
958
+ return part.type === "raw" && part.data.type === "reasoning";
959
+ }
960
+ function isDanglingReasoning(parts, idx) {
961
+ for (let i = idx + 1; i < parts.length; i++) {
962
+ const next = parts[i];
963
+ if (isRawReasoning(next)) return true;
964
+ if (next.type === "text" || next.type === "tool_call") return false;
965
+ }
966
+ return true;
967
+ }
968
+ function hasReplayableAssistantContent(parts) {
969
+ return parts.some((part) => {
970
+ if (part.type === "text") return part.text.length > 0;
971
+ if (part.type === "thinking") return false;
972
+ if (part.type === "raw") {
973
+ const t = part.data.type;
974
+ return !(t === "thinking" || t === "redacted_thinking" || t === "reasoning" || t === "fallback");
975
+ }
976
+ return true;
977
+ });
978
+ }
979
+ function dropInvalidToolCalls(messages, rule) {
980
+ const isInvalid = (part) => part.type === "tool_call" && (!isValidToolCallName(part.name, rule) || !isValidToolCallId(part.id));
981
+ const hasInvalid = messages.some(
982
+ (m) => m.role === "assistant" && Array.isArray(m.content) && m.content.some(isInvalid)
983
+ );
984
+ if (!hasInvalid) return messages;
985
+ const out = [];
986
+ let pending = /* @__PURE__ */ new Map();
987
+ let prunedAssistant = null;
988
+ for (const msg of messages) {
989
+ if (msg.role === "user" || msg.role === "system") {
990
+ if (msg.role === "user") pending = /* @__PURE__ */ new Map();
991
+ out.push(msg);
992
+ continue;
993
+ }
994
+ if (msg.role === "assistant") {
995
+ pending = /* @__PURE__ */ new Map();
996
+ if (typeof msg.content === "string") {
997
+ out.push(msg);
998
+ continue;
999
+ }
1000
+ const original = msg.content;
1001
+ let dropped = false;
1002
+ const kept = [];
1003
+ for (const part of original) {
1004
+ if (part.type === "tool_call") {
1005
+ const keep = !isInvalid(part);
1006
+ const queue = pending.get(part.id) ?? [];
1007
+ queue.push(keep);
1008
+ pending.set(part.id, queue);
1009
+ if (!keep) {
1010
+ dropped = true;
1011
+ continue;
1012
+ }
1013
+ }
1014
+ kept.push(part);
1015
+ }
1016
+ if (!dropped) {
1017
+ out.push(msg);
1018
+ continue;
1019
+ }
1020
+ const content = kept.filter(
1021
+ (part, idx) => !isRawReasoning(part) || !isDanglingReasoning(kept, idx) || isDanglingReasoning(original, original.indexOf(part))
1022
+ );
1023
+ if (hasReplayableAssistantContent(content)) {
1024
+ const pruned = { ...msg, content };
1025
+ out.push(pruned);
1026
+ prunedAssistant = pruned;
1027
+ }
1028
+ continue;
1029
+ }
1030
+ let changed = false;
1031
+ const results = msg.content.filter((result) => {
1032
+ const queue = pending.get(result.toolCallId);
1033
+ const keep = queue && queue.length > 0 ? queue.shift() : true;
1034
+ if (!keep) changed = true;
1035
+ return keep;
1036
+ });
1037
+ if (!changed) out.push(msg);
1038
+ else if (results.length > 0) out.push({ ...msg, content: results });
1039
+ }
1040
+ if (prunedAssistant && out[out.length - 1] === prunedAssistant) out.pop();
1041
+ return out;
1042
+ }
850
1043
  var NON_VISION_USER_IMAGE_PLACEHOLDER = "(image omitted: model does not support images)";
851
1044
  var NON_VISION_TOOL_IMAGE_PLACEHOLDER = "(tool image omitted: model does not support images)";
852
1045
  var NON_VIDEO_USER_PLACEHOLDER = "(video omitted: model does not support video)";
@@ -953,10 +1146,12 @@ function remapAnthropicToolCallId(id, idMap) {
953
1146
  idMap.set(id, mapped);
954
1147
  return mapped;
955
1148
  }
956
- function toAnthropicMessages(messages, cacheControl) {
1149
+ function toAnthropicMessages(messages, cacheControl, options) {
957
1150
  let systemText;
958
1151
  const out = [];
959
1152
  const idMap = /* @__PURE__ */ new Map();
1153
+ const keepFallbackBlocks = options?.fallbackBlocks === true;
1154
+ const droppedToolCallIds = /* @__PURE__ */ new Set();
960
1155
  const trajectoryStartIdx = messages.reduce(
961
1156
  (last, m, i) => m.role === "user" ? i : last,
962
1157
  -1
@@ -1002,17 +1197,23 @@ function toAnthropicMessages(messages, cacheControl) {
1002
1197
  }
1003
1198
  if (msg.role === "assistant") {
1004
1199
  if (typeof msg.content === "string" && msg.content === "") continue;
1005
- const content = typeof msg.content === "string" ? msg.content : toAnthropicAssistantContent(msg.content, msgIdx > trajectoryStartIdx, idMap);
1200
+ const content = typeof msg.content === "string" ? msg.content : toAnthropicAssistantContent(
1201
+ applyServerFallbackReplay(msg.content, keepFallbackBlocks, droppedToolCallIds),
1202
+ msgIdx > trajectoryStartIdx,
1203
+ idMap
1204
+ );
1006
1205
  if (Array.isArray(content) && content.length === 0) continue;
1007
1206
  out.push({ role: "assistant", content });
1008
1207
  continue;
1009
1208
  }
1010
1209
  if (msg.role === "tool") {
1210
+ const results = droppedToolCallIds.size ? msg.content.filter((r) => !droppedToolCallIds.has(r.toolCallId)) : msg.content;
1211
+ if (results.length === 0) continue;
1011
1212
  out.push({
1012
1213
  role: "user",
1013
1214
  // Cast covers the video block (used by the Anthropic-compatible MiniMax
1014
1215
  // API), which isn't in the first-party Anthropic tool_result types.
1015
- content: msg.content.map((result) => ({
1216
+ content: results.map((result) => ({
1016
1217
  type: "tool_result",
1017
1218
  tool_use_id: remapAnthropicToolCallId(result.toolCallId, idMap),
1018
1219
  content: toAnthropicToolResultContent(result.content),
@@ -1356,6 +1557,48 @@ function fineGrainedToolStreamingEnabled() {
1356
1557
  const v = raw.trim().toLowerCase();
1357
1558
  return v === "1" || v === "true" || v === "yes" || v === "on";
1358
1559
  }
1560
+ var SERVER_FALLBACK_BETA = "server-side-fallback-2026-07-01";
1561
+ var anthropicServerFallback = {
1562
+ disabled: /* @__PURE__ */ new Set(),
1563
+ reset() {
1564
+ this.disabled.clear();
1565
+ }
1566
+ };
1567
+ function serverFallbackKey(options) {
1568
+ const auth = options.apiKey?.startsWith("sk-ant-oat") ? "oauth" : "key";
1569
+ return `${options.baseUrl ?? process.env.ANTHROPIC_BASE_URL ?? ""}|${auth}`;
1570
+ }
1571
+ function isDirectAnthropicApi(baseUrl) {
1572
+ const effective = baseUrl ?? process.env.ANTHROPIC_BASE_URL;
1573
+ if (!effective) return true;
1574
+ try {
1575
+ const url = new URL(effective);
1576
+ return url.protocol === "https:" && url.hostname === "api.anthropic.com";
1577
+ } catch {
1578
+ return false;
1579
+ }
1580
+ }
1581
+ function isServerFallbackRejection(err) {
1582
+ const status = err?.status;
1583
+ if (status !== 400) return false;
1584
+ const message = err instanceof Error ? err.message : String(err);
1585
+ return /fallbacks|server-side-fallback/i.test(message);
1586
+ }
1587
+ function sumUsageIterations(usage) {
1588
+ const iterations = usage?.iterations;
1589
+ if (!Array.isArray(iterations) || iterations.length === 0) return null;
1590
+ const num = (v) => typeof v === "number" && Number.isFinite(v) ? v : 0;
1591
+ const totals = { inputTokens: 0, outputTokens: 0, cacheRead: 0, cacheWrite: 0 };
1592
+ for (const it of iterations) {
1593
+ if (!it || typeof it !== "object") continue;
1594
+ const rec = it;
1595
+ totals.inputTokens += num(rec.input_tokens);
1596
+ totals.outputTokens += num(rec.output_tokens);
1597
+ totals.cacheRead += num(rec.cache_read_input_tokens);
1598
+ totals.cacheWrite += num(rec.cache_creation_input_tokens);
1599
+ }
1600
+ return totals;
1601
+ }
1359
1602
  function createClient(options) {
1360
1603
  const isOAuth = options.apiKey?.startsWith("sk-ant-oat");
1361
1604
  const userAgent = isOAuth ? options.userAgent ?? "claude-cli/2.1.75 (external, cli)" : "";
@@ -1450,12 +1693,16 @@ function streamAnthropic(options) {
1450
1693
  async function* runStream(options) {
1451
1694
  const client = createClient(options);
1452
1695
  const isOAuth = options.apiKey?.startsWith("sk-ant-oat");
1453
- const useStreaming = options.streaming !== false;
1696
+ const useStreaming = options.streaming !== false && !options.prewarm;
1454
1697
  const cacheControl = toAnthropicCacheControl(options.cacheRetention, options.baseUrl);
1455
1698
  const supportsFirstPartyToolExtras = !options.baseUrl || options.baseUrl.includes("api.anthropic.com");
1456
1699
  const downgradedImages = downgradeUnsupportedImages(options.messages, options.supportsImages);
1457
1700
  const downgradedMessages = downgradeUnsupportedVideos(downgradedImages, options.supportsVideo);
1458
- const { system: rawSystem, messages } = toAnthropicMessages(downgradedMessages, cacheControl);
1701
+ const fallbackKey = serverFallbackKey(options);
1702
+ const useServerFallback = options.provider !== "minimax" && isDirectAnthropicApi(options.baseUrl) && !anthropicServerFallback.disabled.has(fallbackKey);
1703
+ const { system: rawSystem, messages } = toAnthropicMessages(downgradedMessages, cacheControl, {
1704
+ fallbackBlocks: useServerFallback
1705
+ });
1459
1706
  const system = isOAuth ? [
1460
1707
  {
1461
1708
  type: "text",
@@ -1474,6 +1721,17 @@ async function* runStream(options) {
1474
1721
  outputConfig = t.outputConfig;
1475
1722
  }
1476
1723
  }
1724
+ if (options.prewarm) {
1725
+ const budget = thinking?.budget_tokens;
1726
+ if (budget != null && budget >= 1) {
1727
+ return {
1728
+ message: { role: "assistant", content: [] },
1729
+ stopReason: "end_turn",
1730
+ usage: { inputTokens: 0, outputTokens: 0 }
1731
+ };
1732
+ }
1733
+ maxTokens = 1;
1734
+ }
1477
1735
  const params = {
1478
1736
  model: options.model,
1479
1737
  max_tokens: maxTokens,
@@ -1514,6 +1772,7 @@ async function* runStream(options) {
1514
1772
  ];
1515
1773
  return contextEdits.length ? { context_management: { edits: contextEdits } } : {};
1516
1774
  })(),
1775
+ ...useServerFallback ? { fallbacks: "default" } : {},
1517
1776
  stream: useStreaming
1518
1777
  };
1519
1778
  const hasAdaptiveThinking = isAdaptiveThinkingModel(options.model);
@@ -1530,18 +1789,39 @@ async function* runStream(options) {
1530
1789
  // it Anthropic silently ignores ttl:"1h" and falls back to the 5-min default,
1531
1790
  // so a pre-warmed cache expires before the user's first turn. cacheControl.ttl
1532
1791
  // is only "1h" on the first-party endpoint (see toAnthropicCacheControl).
1533
- ...cacheControl?.ttl === "1h" ? ["extended-cache-ttl-2025-04-11"] : []
1792
+ ...cacheControl?.ttl === "1h" ? ["extended-cache-ttl-2025-04-11"] : [],
1793
+ ...useServerFallback ? [SERVER_FALLBACK_BETA] : []
1534
1794
  ];
1535
- const requestOptions = {
1795
+ const toRequestOptions = (betas) => ({
1536
1796
  signal: options.signal ?? void 0,
1537
- ...betaHeaders.length ? { headers: { "anthropic-beta": betaHeaders.join(",") } } : {}
1797
+ ...betas.length ? { headers: { "anthropic-beta": betas.join(",") } } : {}
1798
+ });
1799
+ const requestOptions = toRequestOptions(betaHeaders);
1800
+ const send = async (create) => {
1801
+ try {
1802
+ return await create(params, requestOptions);
1803
+ } catch (err) {
1804
+ if (!useServerFallback || !isServerFallbackRejection(err)) throw err;
1805
+ anthropicServerFallback.disabled.add(fallbackKey);
1806
+ const { fallbacks: _dropped, ...rest } = params;
1807
+ const retryParams = {
1808
+ ...rest,
1809
+ messages: toAnthropicMessages(downgradedMessages, cacheControl, { fallbackBlocks: false }).messages
1810
+ };
1811
+ return create(
1812
+ retryParams,
1813
+ toRequestOptions(betaHeaders.filter((b) => b !== SERVER_FALLBACK_BETA))
1814
+ );
1815
+ }
1538
1816
  };
1539
1817
  if (!useStreaming) {
1540
1818
  const nonStreamingClient = client.withOptions({ timeout: NON_STREAMING_TIMEOUT_MS });
1541
1819
  try {
1542
- const message = await nonStreamingClient.messages.create(
1543
- { ...params, stream: false },
1544
- requestOptions
1820
+ const message = await send(
1821
+ (p, o) => nonStreamingClient.messages.create(
1822
+ { ...p, stream: false },
1823
+ o
1824
+ )
1545
1825
  );
1546
1826
  yield* synthesizeEventsFromMessage(message);
1547
1827
  return messageToResponse(message);
@@ -1559,9 +1839,8 @@ async function* runStream(options) {
1559
1839
  const keepalive = { type: "keepalive" };
1560
1840
  let receivedAnyEvent = false;
1561
1841
  try {
1562
- const stream2 = await client.messages.create(
1563
- params,
1564
- requestOptions
1842
+ const stream2 = await send(
1843
+ (p, o) => client.messages.create(p, o)
1565
1844
  );
1566
1845
  for await (const event of stream2) {
1567
1846
  receivedAnyEvent = true;
@@ -1739,6 +2018,13 @@ async function* runStream(options) {
1739
2018
  if (usage?.output_tokens != null) {
1740
2019
  outputTokens = usage.output_tokens;
1741
2020
  }
2021
+ const totals = sumUsageIterations(usage);
2022
+ if (totals) {
2023
+ inputTokens = totals.inputTokens;
2024
+ outputTokens = totals.outputTokens;
2025
+ cacheRead = totals.cacheRead;
2026
+ cacheWrite = totals.cacheWrite;
2027
+ }
1742
2028
  yield keepalive;
1743
2029
  break;
1744
2030
  }
@@ -1868,10 +2154,11 @@ function messageToResponse(message) {
1868
2154
  }
1869
2155
  }
1870
2156
  const usage = message.usage;
1871
- const inputTokens = usage.input_tokens ?? 0;
1872
- const outputTokens = usage.output_tokens ?? 0;
1873
- const cacheRead = usage.cache_read_input_tokens;
1874
- const cacheWrite = usage.cache_creation_input_tokens;
2157
+ const totals = sumUsageIterations(usage);
2158
+ const inputTokens = totals?.inputTokens ?? usage.input_tokens ?? 0;
2159
+ const outputTokens = totals?.outputTokens ?? usage.output_tokens ?? 0;
2160
+ const cacheRead = totals?.cacheRead ?? usage.cache_read_input_tokens;
2161
+ const cacheWrite = totals?.cacheWrite ?? usage.cache_creation_input_tokens;
1875
2162
  return {
1876
2163
  message: {
1877
2164
  role: "assistant",
@@ -3905,6 +4192,41 @@ function sanitizeMessagesForWire(messages) {
3905
4192
  return sanitized ?? messages;
3906
4193
  }
3907
4194
 
4195
+ // src/utils/context-observation.ts
4196
+ function imageCount(message) {
4197
+ if (!message || !Array.isArray(message.content)) return 0;
4198
+ if (message.role === "user")
4199
+ return message.content.filter((part) => part.type === "image").length;
4200
+ if (message.role !== "tool") return 0;
4201
+ return message.content.reduce(
4202
+ (count, result) => count + (Array.isArray(result.content) ? result.content.filter((part) => part.type === "image").length : 0),
4203
+ 0
4204
+ );
4205
+ }
4206
+ function observePreparedContext(before, after, tools) {
4207
+ let imagesBefore = 0;
4208
+ let imagesAfter = 0;
4209
+ let firstImageDropMessage = null;
4210
+ for (let index = 0; index < before.length; index++) {
4211
+ const oldCount = imageCount(before[index]);
4212
+ const newCount = imageCount(after[index]);
4213
+ imagesBefore += oldCount;
4214
+ imagesAfter += newCount;
4215
+ if (firstImageDropMessage === null && newCount < oldCount) firstImageDropMessage = index;
4216
+ }
4217
+ return {
4218
+ messages: after,
4219
+ tools: tools.map((tool) => ({
4220
+ name: tool.name,
4221
+ description: tool.description,
4222
+ parameters: resolveToolSchema(tool)
4223
+ })),
4224
+ imagesBefore,
4225
+ imagesAfter,
4226
+ firstImageDropMessage
4227
+ };
4228
+ }
4229
+
3908
4230
  // src/stream.ts
3909
4231
  var GLM_CODING_BASE_URL = "https://api.z.ai/api/coding/paas/v4";
3910
4232
  var KIMI_CODE_USER_AGENT = `kimi-code-cli/${process.env.KIMI_CODE_VERSION ?? "1.0.11"}`;
@@ -4053,11 +4375,27 @@ function stream(options) {
4053
4375
  throw new VideoUnsupportedError();
4054
4376
  }
4055
4377
  const wireMessages = stripMessageProvenance(options.messages);
4056
- const messages = clampProviderContextImages(
4057
- sanitizeMessagesForWire(wireMessages),
4058
- options.provider,
4059
- options.supportsImages
4378
+ const sanitized = sanitizeMessagesForWire(wireMessages);
4379
+ const replayable = dropInvalidToolCalls(
4380
+ sanitized,
4381
+ toolCallNameRuleFor(options.provider, {
4382
+ accountId: options.accountId,
4383
+ baseUrl: options.baseUrl
4384
+ })
4060
4385
  );
4386
+ const messages = clampProviderContextImages(replayable, options.provider, options.supportsImages);
4387
+ if (options.onContextPrepared) {
4388
+ try {
4389
+ options.onContextPrepared(
4390
+ observePreparedContext(
4391
+ replayable === sanitized ? wireMessages : replayable,
4392
+ messages,
4393
+ options.tools ?? []
4394
+ )
4395
+ );
4396
+ } catch {
4397
+ }
4398
+ }
4061
4399
  return entry.stream(messages === options.messages ? options : { ...options, messages });
4062
4400
  }
4063
4401
  function stripMessageProvenance(messages) {
@@ -4469,14 +4807,17 @@ function registerPalsuProvider(config) {
4469
4807
  ProviderError,
4470
4808
  REDACTION_MARKER,
4471
4809
  StreamResult,
4810
+ TOOL_CALL_NAME_RULES,
4472
4811
  clampProviderContextImages,
4473
4812
  classifyProviderError,
4813
+ dropInvalidToolCalls,
4474
4814
  environmentSecrets,
4475
4815
  formatError,
4476
4816
  formatErrorForDisplay,
4477
4817
  hasLoneSurrogate,
4478
4818
  isHardBillingMessage,
4479
4819
  isUsageLimitError,
4820
+ isValidToolCallName,
4480
4821
  localWireModelId,
4481
4822
  palsuAssistantMessage,
4482
4823
  palsuText,
@@ -4495,6 +4836,7 @@ function registerPalsuProvider(config) {
4495
4836
  stream,
4496
4837
  toAnthropicMessages,
4497
4838
  toOpenAIMessages,
4498
- toWellFormedText
4839
+ toWellFormedText,
4840
+ toolCallNameRuleFor
4499
4841
  });
4500
4842
  //# sourceMappingURL=index.cjs.map