@draht/ai 2026.10.20 → 2026.10.23

package/dist/api/openai-completions.js

@@ -1,11 +1,11 @@
 import OpenAI from "openai";
 import { calculateCost, clampThinkingLevel } from "../models.js";
-import { getDrahtUserAgent } from "../utils/draht-user-agent.js";
 import { formatProviderError, normalizeProviderError } from "../utils/error-body.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import { shortHash } from "../utils/hash.js";
 import { headersToRecord } from "../utils/headers.js";
-import { parseStreamingJson } from "../utils/json-parse.js";
+import { getIdentityHeaders } from "../utils/identity.js";
+import { parseStreamingJson, parseToolCallArguments } from "../utils/json-parse.js";
 import { getProviderEnvValue } from "../utils/provider-env.js";
 import { retryProviderRequest } from "../utils/provider-retry.js";
 import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
@@ -14,7 +14,7 @@ import { getDeclaredTools, resolveTranscript, resolveTranscriptTools, } from "..
 import { appendGrammarToolInputJsonDelta, createGrammarToolInputProperties, getGrammarToolInput, getJsonSchemaToolParameters, resolveGrammarConstrainedSampling, resolveJsonSchemaStrictSampling, } from "./constrained-sampling.js";
 import { buildCopilotDynamicHeaders, hasCopilotVisionInput } from "./github-copilot-headers.js";
 import { clampOpenAIPromptCacheKey } from "./openai-prompt-cache.js";
-import { buildBaseOptions, clampThinkingBudgetToAnswerRoom, thinkingBudgetForLevel } from "./simple-options.js";
+import { buildBaseOptions, clampThinkingBudgetToAnswerRoom, resolveSamplingParams, thinkingBudgetForLevel, } from "./simple-options.js";
 import { transformMessages } from "./transform-messages.js";
 /**
  * Check if conversation messages contain tool calls or tool results.
@@ -183,7 +183,7 @@ export const stream = (model, context, options) => {
             const grammarToolInputProperties = createGrammarToolInputProperties(getDeclaredTools(normalizedContext.messages), compat.supportsOpenAIGrammarTools);
             const cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env);
             const cacheSessionId = cacheRetention === "none" ? undefined : options?.sessionId;
-            const client = createClient(model, normalizedContext, apiKey, options?.headers, options?.fetch, cacheSessionId, compat);
+            const client = createClient(model, normalizedContext, apiKey, options?.headers, options?.identity, options?.fetch, cacheSessionId, compat);
             let params = buildParams(model, normalizedContext, options, compat, cacheRetention, grammarToolInputProperties);
             const nextParams = await options?.onPayload?.(params, model);
             if (nextParams !== undefined) {
@@ -258,7 +258,7 @@ export const stream = (model, context, options) => {
                         }
                     }
                     else {
-                        block.arguments = parseStreamingJson(block.partialArgs);
+                        Object.assign(block, parseToolCallArguments(block.partialArgs));
                     }
                     // Finalize in-place and strip the scratch buffers so replay only
                     // carries parsed arguments.
@@ -381,18 +381,8 @@ export const stream = (model, context, options) => {
                     hasFinishReason = true;
                 }
                 if (choice.delta) {
-                    if (choice.delta.content !== null &&
-                        choice.delta.content !== undefined &&
-                        choice.delta.content.length > 0) {
-                        const block = ensureTextBlock();
-                        block.text += choice.delta.content;
-                        stream.push({
-                            type: "text_delta",
-                            contentIndex: getContentIndex(block),
-                            delta: choice.delta.content,
-                            partial: output,
-                        });
-                    }
+                    // Handle reasoning before content: some proxies (e.g. LiteLLM) aggregate reasoning and
+                    // answer into one delta, and the thinking block must precede the text block.
                     // Some endpoints return reasoning in reasoning_content (llama.cpp),
                     // or reasoning (other openai compatible endpoints)
                     // Use the first non-empty reasoning field to avoid duplication
@@ -423,6 +413,31 @@ export const stream = (model, context, options) => {
                             });
                         }
                     }
+                    const reasoningDetails = choice.delta.reasoning_details;
+                    if (Array.isArray(reasoningDetails)) {
+                        for (const detail of reasoningDetails) {
+                            if (!isOpenAIReasoningDetail(detail))
+                                continue;
+                            ensureThinkingBlock("");
+                            streamedReasoningDetails ??= [];
+                            // Keep provider replay data in the existing signature slot. OpenRouter streams
+                            // reasoning_details as deltas: consecutive text/summary deltas are merged into
+                            // logical entries, while encrypted entries remain opaque and discrete.
+                            appendOpenAIReasoningDetail(streamedReasoningDetails, detail);
+                        }
+                    }
+                    if (choice.delta.content !== null &&
+                        choice.delta.content !== undefined &&
+                        choice.delta.content.length > 0) {
+                        const block = ensureTextBlock();
+                        block.text += choice.delta.content;
+                        stream.push({
+                            type: "text_delta",
+                            contentIndex: getContentIndex(block),
+                            delta: choice.delta.content,
+                            partial: output,
+                        });
+                    }
                     if (choice?.delta?.tool_calls) {
                         for (const toolCall of choice.delta.tool_calls) {
                             const block = ensureToolCallBlock(toolCall);
@@ -452,19 +467,6 @@ export const stream = (model, context, options) => {
                             });
                         }
                     }
-                    const reasoningDetails = choice.delta.reasoning_details;
-                    if (Array.isArray(reasoningDetails)) {
-                        for (const detail of reasoningDetails) {
-                            if (!isOpenAIReasoningDetail(detail))
-                                continue;
-                            ensureThinkingBlock("");
-                            streamedReasoningDetails ??= [];
-                            // Keep provider replay data in the existing signature slot. OpenRouter streams
-                            // reasoning_details as deltas: consecutive text/summary deltas are merged into
-                            // logical entries, while encrypted entries remain opaque and discrete.
-                            appendOpenAIReasoningDetail(streamedReasoningDetails, detail);
-                        }
-                    }
                 }
             }
             for (const block of blocks) {
@@ -529,8 +531,8 @@ export const streamSimple = (model, context, options) => {
         thinkingBudgets: options?.thinkingBudgets,
     });
 };
-function createClient(model, context, apiKey, optionsHeaders, fetch, sessionId, compat = getCompat(model)) {
-    const headers = { "User-Agent": getDrahtUserAgent(), ...model.headers };
+function createClient(model, context, apiKey, optionsHeaders, identity, fetch, sessionId, compat = getCompat(model)) {
+    const headers = { ...getIdentityHeaders(model, identity), ...model.headers };
     if (model.provider === "github-copilot") {
         const hasImages = hasCopilotVisionInput(context.messages);
         const copilotHeaders = buildCopilotDynamicHeaders({
@@ -564,7 +566,9 @@ function createClient(model, context, apiKey, optionsHeaders, fetch, sessionId,
     });
 }
 function buildParams(model, context, options, compat = getCompat(model), cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env), grammarToolInputProperties = createGrammarToolInputProperties(getDeclaredTools(context.messages), compat.supportsOpenAIGrammarTools)) {
-    const transcriptTools = resolveTranscriptTools(context.messages, compat.supportsMidConvoSystemMessages === true && compat.supportsMidConvoToolAdditions === true);
+    const transcriptTools = resolveTranscriptTools(context.messages, compat.supportsMidConvoSystemMessages === true && compat.supportsMidConvoToolAdditions === true
+        ? "additions"
+        : "none");
     const messages = convertMessages(model, context, compat, {
         grammarToolInputProperties,
     });
@@ -744,8 +748,11 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
             params.providerOptions = { gateway: gatewayOptions };
         }
     }
-    // Last so custom keys override the named request fields. Per-request keys override model defaults.
-    Object.assign(params, model.samplingParams, options?.samplingParams);
+    // Last so model and request sampling parameters override named request fields.
+    const samplingParams = resolveSamplingParams(model, options?.reasoningEffort ?? "off", options?.samplingParams);
+    if (samplingParams) {
+        Object.assign(params, samplingParams);
+    }
     return params;
 }
 function resolveThinkingTokenBudgetField(compat) {
@@ -892,9 +899,12 @@ export function convertMessages(model, context, compat, options) {
         return id;
     };
     const transformedMessages = transformMessages(normalizedContext.messages, model, (id) => normalizeToolCallId(id));
-    const transcriptTools = resolveTranscriptTools(normalizedContext.messages, compat.supportsMidConvoSystemMessages === true && compat.supportsMidConvoToolAdditions === true);
+    const transcriptTools = resolveTranscriptTools(normalizedContext.messages, compat.supportsMidConvoSystemMessages === true && compat.supportsMidConvoToolAdditions === true
+        ? "additions"
+        : "none");
     const instructionRole = model.reasoning && compat.supportsDeveloperRole ? "developer" : "system";
     let lastRole = null;
+    let systemIndex = 0;
     for (let i = 0; i < transformedMessages.length; i++) {
         const msg = transformedMessages[i];
         // Some providers don't allow user messages directly after tool results
@@ -906,7 +916,8 @@ export function convertMessages(model, context, compat, options) {
             });
         }
         if (msg.role === "system") {
-            const addedTools = i > 0 && transcriptTools.anchorsAdditions ? (msg.toolsAdded ?? []) : [];
+            // transformMessages keeps system messages in order, so their position matches the transcript.
+            const addedTools = systemIndex++ >= transcriptTools.foldedSystemMessages ? (msg.toolsAdded ?? []) : [];
             if (addedTools.length > 0) {
                 const kimiToolMessage = {
                     role: "system",
@@ -1071,7 +1082,7 @@ export function convertMessages(model, context, compat, options) {
                     .join("\n");
                 const hasImages = toolMsg.content.some((c) => c.type === "image");
                 // Always send tool result with text (or placeholder if only images)
-                const hasText = textResult.length > 0;
+                const hasText = textResult.trim().length > 0;
                 const toolResultText = hasText ? textResult : hasImages ? "(see attached image)" : "(no tool output)";
                 // Some providers require the 'name' field in tool results
                 const toolResultMsg = {