@nexus-cortex/core 4.124.43 → 4.124.45

package/dist/orchestrator/APIClient.js

@@ -19,6 +19,8 @@ import { GoogleGenAI } from '@google/genai';
 import { v4 as uuidv4 } from 'uuid';
 import { assertCompletionBody } from './completionBodyGuard.js';
 import { parseHFCompletion, normalizeToolCallArguments } from '../models/hfSpace/normalize.js';
+import { isResponsesInlineRemindersEnabled, inlineRemindersForResponses, isResponsesStopReasonEnabled, isAnthropicHistoryCacheEnabled, applyAnthropicHistoryCache, countAnthropicBreakpoints, } from './transportFixes.js';
+import { isGeminiSystemFixEnabled, isOpenAICacheKeyEnabled, isStreamUsageEnabled, applyAnthropicCacheLevers, streamFinalUsage, } from './transportCacheFixes.js';
 import { anthropicCredentialService, CredentialError } from '../config/AnthropicCredentialService.js';
 /**
  * Anthropic families on the adaptive-thinking-only request surface — sending
@@ -353,6 +355,15 @@ export class APIClient {
             }
             this.applyToolChoice(anthropicRequest, request, 'messages'); // §13-B1 forced tool_choice
         }
+        // R218 (DARK, CORTEX_ANTHROPIC_HISTORY_CACHE=on): rolling message-level breakpoints so the conversation
+        // history is cached, within the 4-breakpoint cap (system + last tool counted). Anthropic only.
+        if (enableCaching && isAnthropicHistoryCacheEnabled()) {
+            anthropicRequest.messages = applyAnthropicHistoryCache(anthropicRequest.messages, countAnthropicBreakpoints({ system: anthropicRequest.system, tools: anthropicRequest.tools })).messages;
+        }
+        // R225 (DARK): CORTEX_ANTHROPIC_AUTO_CACHE top-level automatic cache_control (free slot only) +
+        // CORTEX_ANTHROPIC_CACHE_TTL=1h on every marker. Anthropic only; no-op with both flags off.
+        if (enableCaching)
+            applyAnthropicCacheLevers(anthropicRequest);
         // Phase 2.8: Enable extended thinking for Claude 4+ models with reasoning support
         // Requires beta header: anthropic-beta: interleaved-thinking-2025-05-14
         // Skip if disableThinking option is set (used for continuation requests)
@@ -778,6 +789,16 @@ export class APIClient {
             chatRequest.tools = request.tools;
             this.applyToolChoice(chatRequest, request, 'chat/completions'); // §13-B1 forced tool_choice
         }
+        // R225 (DARK, CORTEX_OPENAI_CACHE_KEY): OpenAI routes same-key requests to the same prompt cache.
+        // OpenAI provider only (other chat/completions providers may reject the unknown field).
+        if (isOpenAICacheKeyEnabled() && modelConfig.provider === 'openai' && request.conversationId) {
+            chatRequest.prompt_cache_key = request.conversationId;
+        }
+        // R224a (DARK, CORTEX_STREAM_USAGE): ask for the final usage chunk on streaming requests (otherwise the
+        // reassembled message reports zeros). Body-level, after the prefix — cache-neutral.
+        if (opts.stream && isStreamUsageEnabled()) {
+            chatRequest.stream_options = { include_usage: true };
+        }
         this.logChatCompletionsPayload(chatRequest, request.modelId, opts.stream);
         return { client, chatRequest };
     }
@@ -982,7 +1003,12 @@ export class APIClient {
         //
         // Extract <system-reminder> content from user messages into `instructions` parameter.
         // This lets the Responses API cache system context separately from conversation content.
-        const { instructions, cleanedItems: inputItems } = this.extractSystemRemindersForResponsesAPI(request.messages);
+        // R215 (DARK, CORTEX_RESPONSES_INLINE_REMINDERS=on): keep reminders IN PLACE as user input_text items
+        // (xAI never receives `instructions`; OpenAI lost their position after the tool output). R63 system
+        // delivery below is unchanged — `instructions` then carries only the static system prompt.
+        const { instructions, cleanedItems: inputItems } = isResponsesInlineRemindersEnabled()
+            ? { instructions: undefined, cleanedItems: inlineRemindersForResponses(request.messages) }
+            : this.extractSystemRemindersForResponsesAPI(request.messages);
         // Check reasoning support
         const supportsReasoning = modelConfig.reasoning?.supported;
         const { disableThinking, reasoningEffort, ...apiParameters } = request.parameters;
@@ -1043,6 +1069,10 @@ export class APIClient {
             // opt-in ZDR mode: "DESIGN NOTE — stateless ZDR mode" in
             // ResponsesAPIAdapter.ts (adapters/).
         }
+        // R225 (DARK, CORTEX_OPENAI_CACHE_KEY): the same session-scoped cache routing key for OpenAI Responses.
+        if (isOpenAICacheKeyEnabled() && modelConfig.provider?.toLowerCase() === 'openai' && request.conversationId) {
+            responsesRequest.prompt_cache_key = request.conversationId;
+        }
         // Set system context as instructions (enables provider-side caching).
         // R63: request.systemMessage (the R28 static system prompt) was previously
         // dropped on this path — only <system-reminder> text mined from user
@@ -1154,6 +1184,12 @@ export class APIClient {
         if (request.tools && request.tools.length > 0) {
             return this.sendGenerateContentHTTP(request, modelConfig, apiKey);
         }
+        // R229 (DARK, CORTEX_GEMINI_SYSTEM_FIX): the @google/generative-ai SDK (v0.2.1, v1 endpoint) has no
+        // systemInstruction support, so a request carrying a system prompt goes through the v1beta REST builder,
+        // which sends it as system_instruction.
+        if (isGeminiSystemFixEnabled() && request.systemMessage) {
+            return this.sendGenerateContentHTTP(request, modelConfig, apiKey);
+        }
         // For non-tool requests, use SDK as before
         if (!this.googleClient) {
             this.googleClient = new GoogleGenerativeAI(apiKey);
@@ -1201,6 +1237,11 @@ export class APIClient {
             contents: request.messages,
             generation_config: generationConfig // Use the nested object directly
         };
+        // R229 (DARK, CORTEX_GEMINI_SYSTEM_FIX): deliver the system prompt (the streaming HTTP builder always did;
+        // this non-streaming builder never did). Same snake_case shape as streamGenerateContentHTTP.
+        if (isGeminiSystemFixEnabled() && request.systemMessage) {
+            requestBody.system_instruction = { parts: [{ text: request.systemMessage }] };
+        }
         // Add tools if present
         if (request.tools && request.tools.length > 0) {
             // REST API uses snake_case "function_declarations", not camelCase "functionDeclarations"
@@ -1338,6 +1379,13 @@ export class APIClient {
                 }
                 this.applyToolChoice(anthropicRequest, request, 'messages'); // §13-B1 forced tool_choice
             }
+            // R218 (DARK): rolling message-level cache breakpoints — mirrors sendAnthropicMessagesAPI.
+            if (enableCaching && isAnthropicHistoryCacheEnabled()) {
+                anthropicRequest.messages = applyAnthropicHistoryCache(anthropicRequest.messages, countAnthropicBreakpoints({ system: anthropicRequest.system, tools: anthropicRequest.tools })).messages;
+            }
+            // R225 (DARK): automatic cache_control + 1h TTL — mirrors sendAnthropicMessagesAPI.
+            if (enableCaching)
+                applyAnthropicCacheLevers(anthropicRequest);
             // Phase 2.8: Enable extended thinking for Claude 4+ models with reasoning support
             // Requires beta header: anthropic-beta: interleaved-thinking-2025-05-14
             // Skip if disableThinking option is set (used for continuation requests)
@@ -1966,6 +2014,9 @@ export class APIClient {
         // R153 (2026-09-16): keep the provider's LAST non-null finish_reason so a `length` cutoff survives reassembly
         // (the empty-response classifier keys `truncated` on it; the old hardcode reported every text turn as `stop`).
         let lastFinishReason = null;
+        // R224a (DARK, CORTEX_STREAM_USAGE): the provider's final usage chunk (choices: []) when requested.
+        const captureStreamUsage = isStreamUsageEnabled();
+        let streamUsage = null;
         // Capture this context before entering generator
         const self = this;
         // Create async generator that both yields chunks AND accumulates final message
@@ -1979,7 +2030,9 @@ export class APIClient {
                 if (!firstChunk)
                     firstChunk = chunk;
                 chunkCount++;
-                const delta = chunk.choices[0]?.delta;
+                if (captureStreamUsage && chunk?.usage)
+                    streamUsage = chunk.usage;
+                const delta = chunk.choices?.[0]?.delta;
                 // Debug: Log first few chunks to see what fields are present
                 if ((process.env.DEBUG === 'true' || process.env.DEBUG_THINKING === 'true') && chunkCount <= 3) {
                     console.log(`[DEBUG APIClient] Chunk ${chunkCount} delta keys:`, delta ? Object.keys(delta) : 'null');
@@ -2030,7 +2083,7 @@ export class APIClient {
                     }
                 }
                 // Check if finish_reason indicates tool_calls completion (or length-truncated tool call)
-                const finishReason = chunk.choices[0]?.finish_reason;
+                const finishReason = chunk.choices?.[0]?.finish_reason;
                 if (finishReason)
                     lastFinishReason = String(finishReason);
                 if ((finishReason === 'tool_calls' || finishReason === 'length') && accumulatedToolCalls.size > 0) {
@@ -2096,11 +2149,7 @@ export class APIClient {
                         // ('length' = output cap hit) instead of a synthetic 'stop'.
                         finish_reason: finalToolCalls.length > 0 ? 'tool_calls' : (lastFinishReason ?? 'stop')
                     }],
-                usage: {
-                    prompt_tokens: 0,
-                    completion_tokens: 0,
-                    total_tokens: 0
-                }
+                usage: streamFinalUsage(streamUsage)
             });
         }();
         return {
@@ -2125,7 +2174,10 @@ export class APIClient {
         // Re-wrapping them would strip tool results and function calls, breaking the tool loop.
         //
         // Extract <system-reminder> content from user messages into `instructions` parameter.
-        const { instructions, cleanedItems: inputItems } = this.extractSystemRemindersForResponsesAPI(request.messages);
+        // R215 (DARK): CORTEX_RESPONSES_INLINE_REMINDERS=on keeps them in place (mirrors sendResponsesAPI).
+        const { instructions, cleanedItems: inputItems } = isResponsesInlineRemindersEnabled()
+            ? { instructions: undefined, cleanedItems: inlineRemindersForResponses(request.messages) }
+            : this.extractSystemRemindersForResponsesAPI(request.messages);
         // Accumulate content from stream
         let fullContent = '';
         let fullReasoningContent = '';
@@ -2197,6 +2249,10 @@ export class APIClient {
                     responsesRequest.prompt_cache_key = request.conversationId;
                 }
             }
+            // R225 (DARK, CORTEX_OPENAI_CACHE_KEY): mirrors sendResponsesAPI.
+            if (isOpenAICacheKeyEnabled() && modelConfig.provider?.toLowerCase() === 'openai' && request.conversationId) {
+                responsesRequest.prompt_cache_key = request.conversationId;
+            }
             // Set system context as instructions (enables provider-side caching).
             // R63: merge the R28 static system prompt (previously dropped on this
             // path) ahead of reminders mined from user messages; XAI gets it as a
@@ -2427,7 +2483,12 @@ export class APIClient {
                     prompt_tokens: 0,
                     completion_tokens: 0,
                     total_tokens: 0
-                }
+                },
+                // R216 (DARK, CORTEX_RESPONSES_STOP_REASON=on): carry the terminal event's status +
+                // incomplete_details (response.completed / response.incomplete) so extractStopReason sees truncation.
+                ...(isResponsesStopReasonEnabled() && typeof finalResponse?.status === 'string'
+                    ? { status: finalResponse.status, incomplete_details: finalResponse.incomplete_details ?? null }
+                    : {})
             });
         }();
         return {
@@ -2452,6 +2513,11 @@ export class APIClient {
         if (request.tools && request.tools.length > 0) {
             return this.streamGenerateContentHTTP(request, modelConfig, apiKey);
         }
+        // R229 (DARK, CORTEX_GEMINI_SYSTEM_FIX): the old SDK cannot send a system prompt — use the v1beta REST
+        // streaming builder (already sends system_instruction) whenever one is present.
+        if (isGeminiSystemFixEnabled() && request.systemMessage) {
+            return this.streamGenerateContentHTTP(request, modelConfig, apiKey);
+        }
         // For non-tool requests, use SDK streaming
         if (!this.googleClient) {
             this.googleClient = new GoogleGenerativeAI(apiKey);
@@ -2820,7 +2886,14 @@ export class APIClient {
             config
         };
         if (request.systemMessage) {
-            sdkRequest.systemInstruction = request.systemMessage;
+            if (isGeminiSystemFixEnabled()) {
+                // R229 (DARK): @google/genai reads the system prompt ONLY from config.systemInstruction
+                // (GenerateContentParameters = {model, contents, config}); a top-level field is dropped.
+                config.systemInstruction = request.systemMessage;
+            }
+            else {
+                sdkRequest.systemInstruction = request.systemMessage;
+            }
         }
         if (process.env.DEBUG === 'true') {
             console.log(`[DEBUG APIClient SDK] Non-streaming request for model: ${modelId}`);
@@ -2917,9 +2990,15 @@ export class APIClient {
             contents: request.messages,
             config // Config contains both generation params and tools
         };
-        // Add system instruction if present (top level, not in config)
+        // Add system instruction if present. Legacy: top level (the SDK DROPS it — GenerateContentParameters =
+        // {model, contents, config}). R229 (DARK, CORTEX_GEMINI_SYSTEM_FIX): config.systemInstruction.
         if (request.systemMessage) {
-            sdkRequest.systemInstruction = request.systemMessage;
+            if (isGeminiSystemFixEnabled()) {
+                config.systemInstruction = request.systemMessage;
+            }
+            else {
+                sdkRequest.systemInstruction = request.systemMessage;
+            }
         }
         if (process.env.DEBUG === 'true') {
             console.log(`[DEBUG APIClient SDK] Using model: ${modelId}`);