@anuma/sdk 6.15.0 → 6.15.2

package/dist/server/index.cjs

@@ -16830,14 +16830,14 @@ function resolveChunkText(content, chunk, chunksInRow) {
   if (covered !== content.length) return content;
   return content.slice(start, end);
 }
-async function readJsonField(model, column, walletAddress) {
-  const raw = model._getRaw(column);
-  if (!raw) return void 0;
-  if (walletAddress && isEncrypted(raw)) {
-    return await decryptJsonField(raw, walletAddress);
+async function readRawJsonField(raw, column, walletAddress) {
+  const value = raw[column];
+  if (!value) return void 0;
+  if (walletAddress && isEncrypted(value)) {
+    return await decryptJsonField(value, walletAddress);
   }
   try {
-    return JSON.parse(raw);
+    return JSON.parse(value);
   } catch {
     return void 0;
   }
@@ -16850,12 +16850,12 @@ async function searchMessagesOp(ctx, queryVector, options) {
     activeConversations.map((c) => String(c._getRaw("conversation_id")))
   );
   const queryConditions = conversationId ? [import_watermelondb14.Q.where("conversation_id", conversationId)] : [];
-  const messages = await ctx.messagesCollection.query(...queryConditions).fetch();
+  const messages = await ctx.messagesCollection.query(...queryConditions).unsafeFetchRaw();
   const candidates = [];
   for (const message of messages) {
-    const msgConvId = String(message._getRaw("conversation_id"));
+    const msgConvId = String(message.conversation_id);
     if (!activeConversationIds.has(msgConvId)) continue;
-    const messageVector = await readJsonField(message, "vector", ctx.walletAddress);
+    const messageVector = await readRawJsonField(message, "vector", ctx.walletAddress);
     if (!messageVector || messageVector.length === 0) continue;
     const similarity = cosineSimilarity(queryVector, messageVector);
     if (similarity >= minSimilarity) {
@@ -16865,7 +16865,7 @@ async function searchMessagesOp(ctx, queryVector, options) {
   const topK = candidates.sort((a, b) => b.similarity - a.similarity).slice(0, limit);
   return Promise.all(
     topK.map(async ({ message, similarity }) => {
-      const stored = await messageToStored(
+      const stored = await messageRawToStored(
         message,
         ctx.walletAddress,
         ctx.signMessage,
@@ -16889,21 +16889,21 @@ async function searchChunksOp(ctx, queryVector, options) {
     activeConversations.map((c) => String(c._getRaw("conversation_id")))
   );
   const queryConditions = conversationId ? [import_watermelondb14.Q.where("conversation_id", conversationId)] : [];
-  const messages = await ctx.messagesCollection.query(...queryConditions).fetch();
+  const messages = await ctx.messagesCollection.query(...queryConditions).unsafeFetchRaw();
   const candidates = [];
   let staleSkipped = 0;
   let unreadableVectors = 0;
   for (const message of messages) {
-    const msgConvId = String(message._getRaw("conversation_id"));
+    const msgConvId = String(message.conversation_id);
     if (!activeConversationIds.has(msgConvId)) continue;
     if (embeddingModel) {
-      const storedModel = message._getRaw("embedding_model");
+      const storedModel = message.embedding_model;
       if (storedModel !== null && storedModel !== void 0 && storedModel !== embeddingModel) {
         staleSkipped++;
         continue;
       }
     }
-    const version = Number(message._getRaw("updated_at")) || 0;
+    const version = Number(message.updated_at) || 0;
     const cached = chunkCache?.get(message.id);
     if (cached && cached.version === version) {
       if (cached.chunks.length > 0) {
@@ -16927,7 +16927,7 @@ async function searchChunksOp(ctx, queryVector, options) {
       }
       continue;
     }
-    const chunks = await readJsonField(message, "chunks", ctx.walletAddress);
+    const chunks = await readRawJsonField(message, "chunks", ctx.walletAddress);
     if (chunks && chunks.length > 0) {
       const vectors = chunks.map(
         (chunk) => decodeChunkVector(chunk.vector, () => {
@@ -16948,7 +16948,7 @@ async function searchChunksOp(ctx, queryVector, options) {
         }
       }
     } else {
-      const messageVector = await readJsonField(message, "vector", ctx.walletAddress);
+      const messageVector = await readRawJsonField(message, "vector", ctx.walletAddress);
       if (!messageVector || messageVector.length === 0) continue;
       const fallback = Float32Array.from(messageVector);
       chunkCache?.set(message.id, { version, chunks: [], fallback });
@@ -16974,7 +16974,7 @@ async function searchChunksOp(ctx, queryVector, options) {
   const storedById = new Map(
     await Promise.all(
       Array.from(uniqueMessages, async ([id, m]) => {
-        const stored = await messageToStored(
+        const stored = await messageRawToStored(
           m,
           ctx.walletAddress,
           ctx.signMessage,
@@ -16992,8 +16992,8 @@ async function searchChunksOp(ctx, queryVector, options) {
   await Promise.all(
     Array.from(needsChunkText, async (id) => {
       const m = uniqueMessages.get(id);
-      const parsed = await readJsonField(m, "chunks", ctx.walletAddress);
-      const version = Number(m._getRaw("updated_at")) || 0;
+      const parsed = await readRawJsonField(m, "chunks", ctx.walletAddress);
+      const version = Number(m.updated_at) || 0;
       if (parsed) chunkStateByMsgId.set(id, { version, chunks: parsed });
     })
   );
@@ -20611,6 +20611,9 @@ function supportsResponseFormat(model, variant = "json_object") {
   const allow = variant === "json_schema" ? RESPONSE_SCHEMA_OK : RESPONSE_FORMAT_OK;
   return model.split("/").some((seg) => allow.has(seg));
 }
+function requiresResponsesTransport(model) {
+  return model.split("/").some((seg) => RESPONSES_ONLY_MODEL_PREFIXES.some((prefix) => seg.startsWith(prefix)));
+}
 function responsesExtra(extra) {
   if (!extra) return {};
   const out = {};
@@ -20946,7 +20949,7 @@ function extractResponsesContent(body) {
   const text12 = root.output.filter((item) => asRecord(item)?.type === "message").flatMap((item) => Array.isArray(item.content) ? item.content : []).map((part) => asRecord(part)?.text).filter((t) => typeof t === "string").join("");
   return text12;
 }
-var RESPONSE_FORMAT_OK, RESPONSE_SCHEMA_OK, JSON_CONTRACT_REMINDER, CHAT_TO_RESPONSES_FIELDS, FOREIGN_ENDPOINT_SUFFIX, RETRYABLE_HTTP;
+var RESPONSE_FORMAT_OK, RESPONSE_SCHEMA_OK, JSON_CONTRACT_REMINDER, RESPONSES_ONLY_MODEL_PREFIXES, CHAT_TO_RESPONSES_FIELDS, FOREIGN_ENDPOINT_SUFFIX, RETRYABLE_HTTP;
 var init_portalLlm = __esm({
   "src/lib/memory/portalLlm.ts"() {
     "use strict";
@@ -20958,6 +20961,7 @@ var init_portalLlm = __esm({
     RESPONSE_FORMAT_OK = /* @__PURE__ */ new Set(["openai", "inclusionai", "deepseek"]);
     RESPONSE_SCHEMA_OK = /* @__PURE__ */ new Set(["openai"]);
     JSON_CONTRACT_REMINDER = "Output ONLY the strict JSON object requested. No prose, no explanation, no markdown.";
+    RESPONSES_ONLY_MODEL_PREFIXES = ["gpt-5.6"];
     CHAT_TO_RESPONSES_FIELDS = {
       max_completion_tokens: "max_output_tokens"
     };
@@ -30733,6 +30737,9 @@ var DEFAULT_BASE_URL2 = BASE_URL;
 var DEFAULT_MODEL4 = "anthropic/claude-sonnet-4-6";
 var DEFAULT_MAX_TOKENS = 4096;
 var REQUEST_TIMEOUT_MS = 6e4;
+var MIN_RETRY_BUDGET_MS = 2e3;
+var MIN_RESPONSES_OUTPUT_TOKENS = 2048;
+var SCHEMA_FALLBACK_SKIP_STATUSES = /* @__PURE__ */ new Set([401, 403, 404, 408, 409, 413, 425, 429]);
 var DEFAULT_SYSTEM_PROMPT = `You are a personal assistant with access to the user's memory. Answer the user's question using the supplied memories as evidence.
 
 Rules:
@@ -30764,14 +30771,9 @@ ${m.content}`).join("\n\n");
   const baseUrl = options.baseUrl ?? DEFAULT_BASE_URL2;
   const model = options.llmModel ?? DEFAULT_MODEL4;
   const wantsStructured = !!options.responseSchema;
-  const sendResponseFormat = wantsStructured && supportsResponseFormat(model, "json_schema");
+  const useResponsesTransport = requiresResponsesTransport(model);
+  const sendResponseFormat = wantsStructured && !useResponsesTransport && supportsResponseFormat(model, "json_schema");
   const basePrompt = options.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;
-  const systemPrompt = wantsStructured && !sendResponseFormat ? `${basePrompt}
-
-Respond with ONLY a single JSON object conforming to this JSON Schema, with no prose, comments, or code fences:
-${JSON.stringify(
-    options.responseSchema
-  )}` : basePrompt;
   const userMessage = [
     `Question:
 ${trimmed}`,
@@ -30781,85 +30783,121 @@ ${evidence}`
   ].join("\n\n");
   const maxTokens = options.maxTokens ?? DEFAULT_MAX_TOKENS;
   const fetchImpl = options.fetchFn ?? fetch;
+  const buildBody = (useResponseFormat) => {
+    const systemPrompt = wantsStructured && !useResponseFormat ? `${basePrompt}
+
+Respond with ONLY a single JSON object conforming to this JSON Schema, with no prose, comments, or code fences:
+${JSON.stringify(
+      options.responseSchema
+    )}` : basePrompt;
+    const messages = [
+      { role: "system", content: systemPrompt },
+      { role: "user", content: userMessage }
+    ];
+    if (useResponsesTransport) {
+      return JSON.stringify({
+        model,
+        // Same system+user pair under the Responses-API field names: `input`
+        // for the turns, `max_output_tokens` for the cap. Sending the
+        // chat-completions spelling here is silently ignored, which caps the
+        // answer at the portal default.
+        input: messages,
+        max_output_tokens: Math.max(maxTokens, MIN_RESPONSES_OUTPUT_TOKENS)
+      });
+    }
+    return JSON.stringify({
+      model,
+      // Modern OpenAI field; the portal reads only `max_completion_tokens`
+      // (the deprecated `max_tokens` is silently ignored → falls back to the
+      // portal's default output cap and truncates the answer).
+      max_completion_tokens: maxTokens,
+      messages,
+      ...useResponseFormat && {
+        response_format: {
+          type: "json_schema",
+          json_schema: { name: "reflect_output", schema: options.responseSchema }
+        }
+      }
+    });
+  };
   const log = getLogger();
   const authHeaders = await resolvePortalAuthHeaders(options, "memory/reflect");
   if (authHeaders === null) return baseResult;
-  const controller = new AbortController();
   const deadline = Date.now() + REQUEST_TIMEOUT_MS;
   const remaining = () => Math.max(0, deadline - Date.now());
-  const timer = setTimeout(() => controller.abort(), remaining());
-  let response;
-  try {
-    response = await fetchImpl(`${baseUrl}/api/v1/chat/completions`, {
-      method: "POST",
-      // No task type unless the caller named one — an undeclared task adds no
-      // header at all, which is what keeps user-facing reflect() unlabelled.
-      headers: {
-        ...authHeaders,
-        ...taskTypeHeader(options.taskType),
-        "Content-Type": "application/json"
-      },
-      body: JSON.stringify({
-        model,
-        // Modern OpenAI field; the portal reads only `max_completion_tokens`
-        // (the deprecated `max_tokens` is silently ignored → falls back to the
-        // portal's default output cap and truncates the answer).
-        max_completion_tokens: maxTokens,
-        messages: [
-          { role: "system", content: systemPrompt },
-          { role: "user", content: userMessage }
-        ],
-        ...sendResponseFormat && {
-          response_format: {
-            type: "json_schema",
-            json_schema: { name: "reflect_output", schema: options.responseSchema }
-          }
-        }
-      }),
-      signal: controller.signal
-    });
-  } catch (err) {
+  const endpoint = useResponsesTransport ? "/api/v1/responses" : "/api/v1/chat/completions";
+  const sendOnce = async (useResponseFormat) => {
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(), remaining());
+    let response;
+    try {
+      response = await fetchImpl(`${baseUrl}${endpoint}`, {
+        method: "POST",
+        // No task type unless the caller named one — an undeclared task adds no
+        // header at all, which is what keeps user-facing reflect() unlabelled.
+        headers: {
+          ...authHeaders,
+          ...taskTypeHeader(options.taskType),
+          "Content-Type": "application/json"
+        },
+        body: buildBody(useResponseFormat),
+        signal: controller.signal
+      });
+    } catch (err) {
+      clearTimeout(timer);
+      log.warn("[memory/reflect] portal request failed", {
+        err: err instanceof Error ? err.message : String(err),
+        baseUrl
+      });
+      return { kind: "error" };
+    }
     clearTimeout(timer);
-    log.warn("[memory/reflect] portal request failed", {
-      err: err instanceof Error ? err.message : String(err),
-      baseUrl
+    if (!response.ok) {
+      return { kind: "http", status: response.status, statusText: response.statusText };
+    }
+    const bodyTimer = setTimeout(() => controller.abort(), remaining());
+    try {
+      const body = await response.json();
+      clearTimeout(bodyTimer);
+      return { kind: "ok", body };
+    } catch (err) {
+      clearTimeout(bodyTimer);
+      log.warn("[memory/reflect] failed to parse portal response body", {
+        err: err instanceof Error ? err.message : String(err)
+      });
+      return { kind: "error" };
+    }
+  };
+  let attempt = await sendOnce(sendResponseFormat);
+  let schemaFallbackUsed = false;
+  const budgetLeftMs = remaining();
+  if (attempt.kind === "http" && sendResponseFormat && !SCHEMA_FALLBACK_SKIP_STATUSES.has(attempt.status) && budgetLeftMs >= MIN_RETRY_BUDGET_MS) {
+    log.warn("[memory/reflect] response_format json_schema rejected; retrying schema-in-prompt", {
+      status: attempt.status,
+      statusText: attempt.statusText,
+      model,
+      taskType: options.taskType,
+      budgetLeftMs
     });
-    return baseResult;
+    schemaFallbackUsed = true;
+    attempt = await sendOnce(false);
   }
-  clearTimeout(timer);
-  if (!response.ok) {
+  if (attempt.kind === "error") return baseResult;
+  if (attempt.kind === "http") {
     log.warn("[memory/reflect] portal returned non-OK", {
-      status: response.status,
-      statusText: response.statusText
+      status: attempt.status,
+      statusText: attempt.statusText,
+      // Distinguishes "the fallback ran and still failed" from "never tried".
+      schemaFallbackUsed
     });
     return baseResult;
   }
-  const bodyTimer = setTimeout(() => controller.abort(), remaining());
-  let body;
-  try {
-    body = await response.json();
-  } catch (err) {
-    clearTimeout(bodyTimer);
-    log.warn("[memory/reflect] failed to parse portal response body", {
-      err: err instanceof Error ? err.message : String(err)
-    });
-    return baseResult;
-  }
-  clearTimeout(bodyTimer);
-  return parseAnswer(body, baseResult, !!options.responseSchema);
+  return parseAnswer(attempt.body, baseResult, !!options.responseSchema);
 }
 function parseAnswer(body, base, parseSchema) {
   if (typeof body !== "object" || body === null) return base;
   const obj = body;
-  const choices = obj.choices;
-  let text12 = "";
-  if (Array.isArray(choices) && choices.length > 0) {
-    const message = choices[0].message;
-    if (typeof message === "object" && message !== null) {
-      const content = message.content;
-      if (typeof content === "string") text12 = content;
-    }
-  }
+  const text12 = extractCompletionContent(obj) ?? "";
   const usage = obj.usage;
   let structuredOutput;
   if (parseSchema && text12) {
@@ -30871,14 +30909,19 @@ function parseAnswer(body, base, parseSchema) {
       });
     }
   }
+  const promptTokens = usage?.prompt_tokens ?? usage?.input_tokens ?? 0;
+  const completionTokens = usage?.completion_tokens ?? usage?.output_tokens ?? 0;
   return {
     text: text12,
     ...structuredOutput !== void 0 && { structuredOutput },
     basedOn: base.basedOn,
     usage: {
-      promptTokens: usage?.prompt_tokens ?? 0,
-      completionTokens: usage?.completion_tokens ?? 0,
-      totalTokens: usage?.total_tokens ?? 0
+      promptTokens,
+      completionTokens,
+      // Derived when absent: the Responses API reports the two component counts
+      // and does not always send a total, which would otherwise report zero
+      // spend for a call that plainly cost something.
+      totalTokens: usage?.total_tokens ?? promptTokens + completionTokens
     }
   };
 }

package/dist/server/index.mjs

@@ -141,23 +141,25 @@ import {
   updateVaultMemoryOp,
   upsertMessageOp,
   vaultMemoryToStored
-} from "./chunk-2QFK4L7Y.mjs";
+} from "./chunk-K7H7ICPK.mjs";
 import {
   createStreamingDeAnonymizer,
   notifyConsolidationFallback,
   resolvePiiRedactor
-} from "./chunk-JPLFOSB2.mjs";
+} from "./chunk-4LPSEROP.mjs";
 import {
   INTERNAL_FLOW_MARKER,
   callPortalJsonCompletion,
+  extractCompletionContent,
   extractJsonCandidate,
   getLogger,
+  requiresResponsesTransport,
   resolvePortalAuthHeaders,
   supportsResponseFormat,
   taskTypeHeader,
   validateEndpointOverride,
   withInternalFlowMarker
-} from "./chunk-H4G3DLDJ.mjs";
+} from "./chunk-BSCEZOIW.mjs";
 import {
   BASE_URL,
   __commonJS,
@@ -5700,7 +5702,7 @@ async function tryConsolidate(trimmed, ctx, options, prepared) {
     content: m.content,
     similarity: m.similarity
   }));
-  const { consolidateMemory: doConsolidate } = await import("./consolidate-CMMDTZ37.mjs");
+  const { consolidateMemory: doConsolidate } = await import("./consolidate-MSAXASHL.mjs");
   const decision = await doConsolidate(trimmed, candidates, consolidateOptions);
   if (decision.action === "create") return null;
   if (decision.action === "supersede" && decision.content) {
@@ -6924,6 +6926,9 @@ var DEFAULT_BASE_URL = BASE_URL;
 var DEFAULT_MODEL3 = "anthropic/claude-sonnet-4-6";
 var DEFAULT_MAX_TOKENS = 4096;
 var REQUEST_TIMEOUT_MS = 6e4;
+var MIN_RETRY_BUDGET_MS = 2e3;
+var MIN_RESPONSES_OUTPUT_TOKENS = 2048;
+var SCHEMA_FALLBACK_SKIP_STATUSES = /* @__PURE__ */ new Set([401, 403, 404, 408, 409, 413, 425, 429]);
 var DEFAULT_SYSTEM_PROMPT = `You are a personal assistant with access to the user's memory. Answer the user's question using the supplied memories as evidence.
 
 Rules:
@@ -6955,14 +6960,9 @@ ${m.content}`).join("\n\n");
   const baseUrl = options.baseUrl ?? DEFAULT_BASE_URL;
   const model = options.llmModel ?? DEFAULT_MODEL3;
   const wantsStructured = !!options.responseSchema;
-  const sendResponseFormat = wantsStructured && supportsResponseFormat(model, "json_schema");
+  const useResponsesTransport = requiresResponsesTransport(model);
+  const sendResponseFormat = wantsStructured && !useResponsesTransport && supportsResponseFormat(model, "json_schema");
   const basePrompt = options.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;
-  const systemPrompt = wantsStructured && !sendResponseFormat ? `${basePrompt}
-
-Respond with ONLY a single JSON object conforming to this JSON Schema, with no prose, comments, or code fences:
-${JSON.stringify(
-    options.responseSchema
-  )}` : basePrompt;
   const userMessage = [
     `Question:
 ${trimmed}`,
@@ -6972,85 +6972,121 @@ ${evidence}`
   ].join("\n\n");
   const maxTokens = options.maxTokens ?? DEFAULT_MAX_TOKENS;
   const fetchImpl = options.fetchFn ?? fetch;
+  const buildBody = (useResponseFormat) => {
+    const systemPrompt = wantsStructured && !useResponseFormat ? `${basePrompt}
+
+Respond with ONLY a single JSON object conforming to this JSON Schema, with no prose, comments, or code fences:
+${JSON.stringify(
+      options.responseSchema
+    )}` : basePrompt;
+    const messages = [
+      { role: "system", content: systemPrompt },
+      { role: "user", content: userMessage }
+    ];
+    if (useResponsesTransport) {
+      return JSON.stringify({
+        model,
+        // Same system+user pair under the Responses-API field names: `input`
+        // for the turns, `max_output_tokens` for the cap. Sending the
+        // chat-completions spelling here is silently ignored, which caps the
+        // answer at the portal default.
+        input: messages,
+        max_output_tokens: Math.max(maxTokens, MIN_RESPONSES_OUTPUT_TOKENS)
+      });
+    }
+    return JSON.stringify({
+      model,
+      // Modern OpenAI field; the portal reads only `max_completion_tokens`
+      // (the deprecated `max_tokens` is silently ignored → falls back to the
+      // portal's default output cap and truncates the answer).
+      max_completion_tokens: maxTokens,
+      messages,
+      ...useResponseFormat && {
+        response_format: {
+          type: "json_schema",
+          json_schema: { name: "reflect_output", schema: options.responseSchema }
+        }
+      }
+    });
+  };
   const log = getLogger();
   const authHeaders = await resolvePortalAuthHeaders(options, "memory/reflect");
   if (authHeaders === null) return baseResult;
-  const controller = new AbortController();
   const deadline = Date.now() + REQUEST_TIMEOUT_MS;
   const remaining = () => Math.max(0, deadline - Date.now());
-  const timer = setTimeout(() => controller.abort(), remaining());
-  let response;
-  try {
-    response = await fetchImpl(`${baseUrl}/api/v1/chat/completions`, {
-      method: "POST",
-      // No task type unless the caller named one — an undeclared task adds no
-      // header at all, which is what keeps user-facing reflect() unlabelled.
-      headers: {
-        ...authHeaders,
-        ...taskTypeHeader(options.taskType),
-        "Content-Type": "application/json"
-      },
-      body: JSON.stringify({
-        model,
-        // Modern OpenAI field; the portal reads only `max_completion_tokens`
-        // (the deprecated `max_tokens` is silently ignored → falls back to the
-        // portal's default output cap and truncates the answer).
-        max_completion_tokens: maxTokens,
-        messages: [
-          { role: "system", content: systemPrompt },
-          { role: "user", content: userMessage }
-        ],
-        ...sendResponseFormat && {
-          response_format: {
-            type: "json_schema",
-            json_schema: { name: "reflect_output", schema: options.responseSchema }
-          }
-        }
-      }),
-      signal: controller.signal
-    });
-  } catch (err) {
+  const endpoint = useResponsesTransport ? "/api/v1/responses" : "/api/v1/chat/completions";
+  const sendOnce = async (useResponseFormat) => {
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(), remaining());
+    let response;
+    try {
+      response = await fetchImpl(`${baseUrl}${endpoint}`, {
+        method: "POST",
+        // No task type unless the caller named one — an undeclared task adds no
+        // header at all, which is what keeps user-facing reflect() unlabelled.
+        headers: {
+          ...authHeaders,
+          ...taskTypeHeader(options.taskType),
+          "Content-Type": "application/json"
+        },
+        body: buildBody(useResponseFormat),
+        signal: controller.signal
+      });
+    } catch (err) {
+      clearTimeout(timer);
+      log.warn("[memory/reflect] portal request failed", {
+        err: err instanceof Error ? err.message : String(err),
+        baseUrl
+      });
+      return { kind: "error" };
+    }
     clearTimeout(timer);
-    log.warn("[memory/reflect] portal request failed", {
-      err: err instanceof Error ? err.message : String(err),
-      baseUrl
+    if (!response.ok) {
+      return { kind: "http", status: response.status, statusText: response.statusText };
+    }
+    const bodyTimer = setTimeout(() => controller.abort(), remaining());
+    try {
+      const body = await response.json();
+      clearTimeout(bodyTimer);
+      return { kind: "ok", body };
+    } catch (err) {
+      clearTimeout(bodyTimer);
+      log.warn("[memory/reflect] failed to parse portal response body", {
+        err: err instanceof Error ? err.message : String(err)
+      });
+      return { kind: "error" };
+    }
+  };
+  let attempt = await sendOnce(sendResponseFormat);
+  let schemaFallbackUsed = false;
+  const budgetLeftMs = remaining();
+  if (attempt.kind === "http" && sendResponseFormat && !SCHEMA_FALLBACK_SKIP_STATUSES.has(attempt.status) && budgetLeftMs >= MIN_RETRY_BUDGET_MS) {
+    log.warn("[memory/reflect] response_format json_schema rejected; retrying schema-in-prompt", {
+      status: attempt.status,
+      statusText: attempt.statusText,
+      model,
+      taskType: options.taskType,
+      budgetLeftMs
     });
-    return baseResult;
+    schemaFallbackUsed = true;
+    attempt = await sendOnce(false);
   }
-  clearTimeout(timer);
-  if (!response.ok) {
+  if (attempt.kind === "error") return baseResult;
+  if (attempt.kind === "http") {
     log.warn("[memory/reflect] portal returned non-OK", {
-      status: response.status,
-      statusText: response.statusText
+      status: attempt.status,
+      statusText: attempt.statusText,
+      // Distinguishes "the fallback ran and still failed" from "never tried".
+      schemaFallbackUsed
     });
     return baseResult;
   }
-  const bodyTimer = setTimeout(() => controller.abort(), remaining());
-  let body;
-  try {
-    body = await response.json();
-  } catch (err) {
-    clearTimeout(bodyTimer);
-    log.warn("[memory/reflect] failed to parse portal response body", {
-      err: err instanceof Error ? err.message : String(err)
-    });
-    return baseResult;
-  }
-  clearTimeout(bodyTimer);
-  return parseAnswer(body, baseResult, !!options.responseSchema);
+  return parseAnswer(attempt.body, baseResult, !!options.responseSchema);
 }
 function parseAnswer(body, base, parseSchema) {
   if (typeof body !== "object" || body === null) return base;
   const obj = body;
-  const choices = obj.choices;
-  let text12 = "";
-  if (Array.isArray(choices) && choices.length > 0) {
-    const message = choices[0].message;
-    if (typeof message === "object" && message !== null) {
-      const content = message.content;
-      if (typeof content === "string") text12 = content;
-    }
-  }
+  const text12 = extractCompletionContent(obj) ?? "";
   const usage = obj.usage;
   let structuredOutput;
   if (parseSchema && text12) {
@@ -7062,14 +7098,19 @@ function parseAnswer(body, base, parseSchema) {
       });
     }
   }
+  const promptTokens = usage?.prompt_tokens ?? usage?.input_tokens ?? 0;
+  const completionTokens = usage?.completion_tokens ?? usage?.output_tokens ?? 0;
   return {
     text: text12,
     ...structuredOutput !== void 0 && { structuredOutput },
     basedOn: base.basedOn,
     usage: {
-      promptTokens: usage?.prompt_tokens ?? 0,
-      completionTokens: usage?.completion_tokens ?? 0,
-      totalTokens: usage?.total_tokens ?? 0
+      promptTokens,
+      completionTokens,
+      // Derived when absent: the Responses API reports the two component counts
+      // and does not always send a total, which would otherwise report zero
+      // spend for a call that plainly cost something.
+      totalTokens: usage?.total_tokens ?? promptTokens + completionTokens
     }
   };
 }

package/dist/server/{recall-EODNEKIA.mjs → recall-YXJATOSD.mjs}

@@ -1,7 +1,7 @@
 import {
   recall
-} from "./chunk-2QFK4L7Y.mjs";
-import "./chunk-H4G3DLDJ.mjs";
+} from "./chunk-K7H7ICPK.mjs";
+import "./chunk-BSCEZOIW.mjs";
 import "./chunk-3RPGXXPZ.mjs";
 export {
   recall

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anuma/sdk",
-  "version": "6.15.0",
+  "version": "6.15.2",
   "description": "",
   "main": "./dist/index.cjs",
   "module": "./dist/index.mjs",