@theokit/sdk 5.9.2 → 5.11.0

package/dist/{chunk-EVH3OUMY.js → chunk-74OA5DIM.js}

@@ -325,7 +325,10 @@ async function consumeAnthropicResponse(response) {
     toolCalls: extractAnthropicToolCalls(data),
     stopReason: mapAnthropicStopReason(data.stop_reason ?? null),
     inputTokens: data.usage?.input_tokens,
-    outputTokens: data.usage?.output_tokens
+    outputTokens: data.usage?.output_tokens,
+    // B-412 — the counters the Bedrock and Vertex Anthropic clients used to drop.
+    cacheReadTokens: data.usage?.cache_read_input_tokens,
+    cacheWriteTokens: data.usage?.cache_creation_input_tokens
   };
 }
 async function parseHttpErrorBody(response) {
@@ -538,13 +541,9 @@ var AnthropicClient = class {
     }
     const accumulator = new AnthropicStreamAccumulator();
     for await (const record of parseSseStream(response.body, signal)) {
-      if (record.event === "ping" || record.event === "message_start") continue;
-      let parsed;
-      try {
-        parsed = JSON.parse(record.data);
-      } catch {
-        continue;
-      }
+      if (record.event === "ping") continue;
+      const parsed = parseAnthropicFrame(record);
+      if (parsed === void 0) continue;
       const events = accumulator.consume(parsed);
       for (const event of events) yield event;
     }
@@ -556,6 +555,30 @@ var AnthropicClient = class {
     return accumulator.finish();
   }
 };
+var CONTENT_FRAMES = /* @__PURE__ */ new Set([
+  "content_block_start",
+  "content_block_delta",
+  "message_delta"
+]);
+function parseAnthropicFrame(record) {
+  try {
+    return JSON.parse(record.data);
+  } catch (parseErr) {
+    const event = record.event;
+    const size = `${record.data.length} chars`;
+    if (CONTENT_FRAMES.has(event)) {
+      throw new NetworkError(
+        `Anthropic SSE ${event} frame (${size}) is not valid JSON; the turn would lose content`,
+        { code: "stream_malformed_frame", cause: parseErr }
+      );
+    }
+    diag(
+      `[theokit-sdk] skipped an Anthropic SSE ${event} frame (${size}) that is not valid JSON
+`
+    );
+    return void 0;
+  }
+}
 var AnthropicStreamAccumulator = class {
   text = "";
   stopReason = "end_turn";
@@ -590,6 +613,11 @@ var AnthropicStreamAccumulator = class {
     if (event.type === "message_delta") {
       this.handleMessageDelta(event);
     }
+    if (event.type === "message_start") {
+      const usage = event.message?.usage;
+      if (typeof usage?.input_tokens === "number") this.inputTokens = usage.input_tokens;
+      this.readCacheCounters(usage);
+    }
     return [];
   }
   handleToolStart(start) {
@@ -642,14 +670,26 @@ var AnthropicStreamAccumulator = class {
       this.sawStopReason = true;
     }
     this.stopReason = mapAnthropicStopReason(md.delta.stop_reason);
-    if (md.usage?.input_tokens !== void 0) this.inputTokens = md.usage.input_tokens;
-    if (md.usage?.output_tokens !== void 0) this.outputTokens = md.usage.output_tokens;
-    if (md.usage?.cache_creation_input_tokens !== void 0 && md.usage.cache_creation_input_tokens > 0) {
-      this.cacheWriteTokens = md.usage.cache_creation_input_tokens;
-    }
-    if (md.usage?.cache_read_input_tokens !== void 0 && md.usage.cache_read_input_tokens > 0) {
-      this.cacheReadTokens = md.usage.cache_read_input_tokens;
+    const input = md.usage?.input_tokens;
+    if (typeof input === "number" && (input > 0 || this.inputTokens === void 0)) {
+      this.inputTokens = input;
     }
+    if (md.usage?.output_tokens !== void 0) this.outputTokens = md.usage.output_tokens;
+    this.readCacheCounters(md.usage);
+  }
+  /**
+   * T3.8 / B-412 / B-475 — the cache counters, from `message_start` and then `message_delta`. A
+   * value above 0 replaces the earlier one. A zero, a null or an absent field leaves it in place,
+   * so a `message_delta` that zero-fills its usage does not erase what `message_start` reported;
+   * a counter no event reports above 0 stays unreported (usage-accumulator strips zero entries).
+   * The `typeof` check is there because the wire is JSON: a null arrives although the type
+   * declares none.
+   */
+  readCacheCounters(usage) {
+    const write = usage?.cache_creation_input_tokens;
+    if (typeof write === "number" && write > 0) this.cacheWriteTokens = write;
+    const read = usage?.cache_read_input_tokens;
+    if (typeof read === "number" && read > 0) this.cacheReadTokens = read;
   }
   /** M2 #61 — whether the terminal `message_delta` (stop_reason) was seen. */
   get finishReasonSeen() {
@@ -1228,6 +1268,117 @@ function toOllamaTools(tools) {
   }));
 }
 
+// src/internal/budget/normalize-usage.ts
+function int(v) {
+  if (typeof v === "number") return Number.isFinite(v) ? Math.max(0, Math.trunc(v)) : 0;
+  if (typeof v === "string") {
+    const n = Number.parseInt(v, 10);
+    return Number.isFinite(n) ? Math.max(0, n) : 0;
+  }
+  return 0;
+}
+function buildTotal(buckets) {
+  return buckets.inputTokens + buckets.outputTokens + buckets.cacheReadTokens + buckets.cacheWriteTokens;
+}
+function omitUndefined(usage) {
+  return {
+    inputTokens: usage.inputTokens,
+    outputTokens: usage.outputTokens,
+    ...usage.cacheReadTokens > 0 ? { cacheReadTokens: usage.cacheReadTokens } : {},
+    ...usage.cacheWriteTokens > 0 ? { cacheWriteTokens: usage.cacheWriteTokens } : {},
+    ...usage.reasoningTokens > 0 ? { reasoningTokens: usage.reasoningTokens } : {},
+    totalTokens: usage.totalTokens
+  };
+}
+function inferApiMode(provider) {
+  const p = provider.toLowerCase();
+  if (p === "anthropic" || p === "claude" || p === "bedrock_anthropic") {
+    return "anthropic_messages";
+  }
+  if (p === "openai-codex" || p === "codex") return "openai_responses";
+  return "openai_chat_completions";
+}
+function normalizeUsage(rawUsage, opts) {
+  if (rawUsage === null || rawUsage === void 0) {
+    return { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
+  }
+  if (typeof rawUsage !== "object") {
+    return { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
+  }
+  const raw = rawUsage;
+  const mode = opts.apiMode ?? inferApiMode(opts.provider);
+  if (mode === "anthropic_messages") return normalizeAnthropic(raw);
+  if (mode === "openai_responses") return normalizeOpenAIResponses(raw);
+  return normalizeOpenAIChat(raw);
+}
+function normalizeAnthropic(raw) {
+  const inputTokens = int(raw.input_tokens);
+  const outputTokens = int(raw.output_tokens);
+  const cacheReadTokens = int(raw.cache_read_input_tokens);
+  const cacheWriteTokens = int(raw.cache_creation_input_tokens);
+  const usage = {
+    inputTokens,
+    outputTokens,
+    cacheReadTokens,
+    cacheWriteTokens,
+    reasoningTokens: 0,
+    totalTokens: buildTotal({ inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens })
+  };
+  return omitUndefined(usage);
+}
+function normalizeOpenAIResponses(raw) {
+  const inputTotal = int(raw.input_tokens);
+  const outputTokens = int(raw.output_tokens);
+  const inputDetails = raw.input_tokens_details ?? {};
+  const outputDetails = raw.output_tokens_details ?? {};
+  const cacheReadTokens = int(inputDetails.cached_tokens);
+  const cacheWriteTokens = int(inputDetails.cache_creation_tokens);
+  const reasoningTokens = int(outputDetails.reasoning_tokens);
+  const inputTokens = Math.max(0, inputTotal - cacheReadTokens - cacheWriteTokens);
+  const usage = {
+    inputTokens,
+    outputTokens,
+    cacheReadTokens,
+    cacheWriteTokens,
+    reasoningTokens,
+    totalTokens: buildTotal({ inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens })
+  };
+  return omitUndefined(usage);
+}
+function sent(v) {
+  return v !== void 0 && v !== null;
+}
+function splitChatCompletionsPrompt(raw) {
+  const details = raw.prompt_tokens_details ?? {};
+  const bucket = (nested, topLevel) => sent(nested) || sent(topLevel) ? int(nested) || int(topLevel) : void 0;
+  const cacheReadTokens = bucket(details.cached_tokens, raw.cache_read_input_tokens);
+  const cacheWriteTokens = bucket(details.cache_write_tokens, raw.cache_creation_input_tokens);
+  const cached = (cacheReadTokens ?? 0) + (cacheWriteTokens ?? 0);
+  return {
+    ...sent(raw.prompt_tokens) ? { inputTokens: Math.max(0, int(raw.prompt_tokens) - cached) } : {},
+    ...cacheReadTokens !== void 0 ? { cacheReadTokens } : {},
+    ...cacheWriteTokens !== void 0 ? { cacheWriteTokens } : {}
+  };
+}
+function normalizeOpenAIChat(raw) {
+  const outputTokens = int(raw.completion_tokens);
+  const completionDetails = raw.completion_tokens_details ?? {};
+  const split = splitChatCompletionsPrompt(raw);
+  const inputTokens = split.inputTokens ?? 0;
+  const cacheReadTokens = split.cacheReadTokens ?? 0;
+  const cacheWriteTokens = split.cacheWriteTokens ?? 0;
+  const reasoningTokens = int(completionDetails.reasoning_tokens);
+  const usage = {
+    inputTokens,
+    outputTokens,
+    cacheReadTokens,
+    cacheWriteTokens,
+    reasoningTokens,
+    totalTokens: buildTotal({ inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens })
+  };
+  return omitUndefined(usage);
+}
+
 // src/internal/base-url.ts
 function getConfiguredBaseUrl() {
   const value = process.env.THEOKIT_API_BASE_URL;
@@ -1504,6 +1655,16 @@ function assistantMessage(message) {
   return result;
 }
 
+// src/internal/llm/openrouter-system.ts
+function encodeOpenRouterSystemContent(system) {
+  if (!Array.isArray(system) || !system.some((block) => block.cacheable === true)) {
+    return void 0;
+  }
+  return system.map(
+    (block) => block.cacheable === true ? { type: "text", text: block.text, cache_control: { type: "ephemeral" } } : { type: "text", text: block.text }
+  );
+}
+
 // src/internal/llm/openai.ts
 function deriveChatPath(baseUrl) {
   try {
@@ -1646,11 +1807,8 @@ var OpenAIStreamAccumulator = class {
   allowedToolNames;
   text = "";
   stopReason = "end_turn";
-  inputTokens;
-  outputTokens;
-  cacheReadTokens;
-  cacheWriteTokens;
-  reasoningTokens;
+  /** The stream's `usage` object, a later chunk replacing an earlier one per top-level field. */
+  usage = {};
   toolCalls = /* @__PURE__ */ new Map();
   /** R7: present only when recovery is enabled AND the request declares tools — holds suspected
    *  leaked-dialect content back from the `text_delta` stream. `undefined` ⇒ stream immediately. */
@@ -1684,18 +1842,7 @@ var OpenAIStreamAccumulator = class {
     return { type: "reasoning_delta", text: reasoning };
   }
   applyUsage(usage) {
-    if (usage?.prompt_tokens !== void 0) this.inputTokens = usage.prompt_tokens;
-    if (usage?.completion_tokens !== void 0) this.outputTokens = usage.completion_tokens;
-    const cachedDetail = usage?.prompt_tokens_details?.cached_tokens;
-    if (cachedDetail !== void 0) this.cacheReadTokens = cachedDetail;
-    const reasoningDetail = usage?.completion_tokens_details?.reasoning_tokens;
-    if (reasoningDetail !== void 0) this.reasoningTokens = reasoningDetail;
-    if (usage?.cache_read_input_tokens !== void 0) {
-      this.cacheReadTokens = usage.cache_read_input_tokens;
-    }
-    if (usage?.cache_creation_input_tokens !== void 0) {
-      this.cacheWriteTokens = usage.cache_creation_input_tokens;
-    }
+    if (usage !== void 0) this.usage = { ...this.usage, ...usage };
   }
   applyContentDelta(content) {
     if (typeof content !== "string" || content.length === 0) return void 0;
@@ -1766,11 +1913,10 @@ var OpenAIStreamAccumulator = class {
       stopReason,
       text,
       toolCalls,
-      inputTokens: this.inputTokens,
-      outputTokens: this.outputTokens,
-      cacheReadTokens: this.cacheReadTokens,
-      cacheWriteTokens: this.cacheWriteTokens,
-      reasoningTokens: this.reasoningTokens
+      // D376: the prompt split (cache buckets out of inputTokens) has one owner, normalize-usage.ts.
+      ...splitChatCompletionsPrompt(this.usage),
+      outputTokens: this.usage.completion_tokens,
+      reasoningTokens: this.usage.completion_tokens_details?.reasoning_tokens
     });
   }
 };
@@ -1791,10 +1937,10 @@ function applyToolsRequest(body, request) {
 }
 function buildOpenAIBody(request, providerName) {
   const messages = [];
+  const systemParts = providerName === "openrouter" ? encodeOpenRouterSystemContent(request.system) : void 0;
   const systemText = openAISystemText(request.system);
-  if (systemText.length > 0) {
-    messages.push({ role: "system", content: systemText });
-  }
+  if (systemParts !== void 0) messages.push({ role: "system", content: systemParts });
+  else if (systemText.length > 0) messages.push({ role: "system", content: systemText });
   for (const message of request.messages) {
     for (const out of toOpenAIMessages(message)) messages.push(out);
   }
@@ -3291,6 +3437,6 @@ async function autoCompactIfNeeded(opts) {
   }
 }
 
-export { COMPACT_SUMMARY_MARKER, COMPACT_USER_MESSAGE_MAX_TOKENS, CircuitBreaker, DEFAULT_PRE_COMPACT_TIMEOUT_MS, abortError, autoCompactIfNeeded, buildDefaultSummarizer, compactSessionTranscript, discoverProviderPlugins, getConfiguredBaseUrl, httpRequest, isCompactSummary, isFixtureApiKey, mapOllamaHttpError, mapOllamaTransportError, mergeExplicitApiKey, parseSseStream, presentProviderCredentialEnvVars, readErrorResponseBody, relayStream, renderToolResultContentText, resolveProviderChain, resolveRunProvider, resolveSummarizerRoute, shouldAutoCompact, shouldUseFixtureMode, shouldUseRealLocalRuntime, tryFirstEvent, validateApiKeyShape };
-//# sourceMappingURL=chunk-EVH3OUMY.js.map
-//# sourceMappingURL=chunk-EVH3OUMY.js.map
+export { COMPACT_SUMMARY_MARKER, COMPACT_USER_MESSAGE_MAX_TOKENS, CircuitBreaker, DEFAULT_PRE_COMPACT_TIMEOUT_MS, abortError, autoCompactIfNeeded, buildDefaultSummarizer, compactSessionTranscript, discoverProviderPlugins, getConfiguredBaseUrl, httpRequest, inferApiMode, isCompactSummary, isFixtureApiKey, mapOllamaHttpError, mapOllamaTransportError, mergeExplicitApiKey, normalizeUsage, parseSseStream, presentProviderCredentialEnvVars, readErrorResponseBody, relayStream, renderToolResultContentText, resolveProviderChain, resolveRunProvider, resolveSummarizerRoute, shouldAutoCompact, shouldUseFixtureMode, shouldUseRealLocalRuntime, tryFirstEvent, validateApiKeyShape };
+//# sourceMappingURL=chunk-74OA5DIM.js.map
+//# sourceMappingURL=chunk-74OA5DIM.js.map