@purista/harness-azure-foundry 1.2.1 → 1.2.2

package/dist/index.js

@@ -50,6 +50,7 @@ class AzureFoundryModelProvider extends BaseModelProvider {
         req.signal.throwIfAborted();
         let usage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
         let finishReason = 'stop';
+        const toolState = new Map();
         for await (const event of streamChat(this.client, req, false)) {
             req.signal.throwIfAborted();
             const data = parseStreamData(event, req, 'textStream');
@@ -59,16 +60,20 @@ class AzureFoundryModelProvider extends BaseModelProvider {
                 if (choice.delta?.content) {
                     yield { kind: 'delta', text: choice.delta.content };
                 }
-                const toolCalls = extractToolCalls(choice.delta?.tool_calls, req, 'textStream');
-                for (const call of toolCalls ?? []) {
-                    yield { kind: 'tool_call', call };
+                if (choice.delta?.tool_calls) {
+                    accumulateToolCallDeltas(toolState, choice.delta.tool_calls);
+                }
+                if (choice.finish_reason) {
+                    finishReason = toFinishReason(choice.finish_reason);
                 }
-                finishReason = toFinishReason(choice.finish_reason ?? finishReason);
             }
             if (data.usage) {
                 usage = toUsage(data.usage.prompt_tokens, data.usage.completion_tokens, data.usage.total_tokens);
             }
         }
+        for (const call of finalizeStreamToolCalls(toolState, req, 'textStream')) {
+            yield { kind: 'tool_call', call };
+        }
         yield { kind: 'finish', usage, finishReason };
     }
     async doObject(req) {
@@ -91,6 +96,7 @@ class AzureFoundryModelProvider extends BaseModelProvider {
         let partial = '';
         let usage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
         let finishReason = 'stop';
+        const toolState = new Map();
         for await (const event of streamChat(this.client, req, true)) {
             req.signal.throwIfAborted();
             const data = parseStreamData(event, req, 'objectStream');
@@ -101,16 +107,20 @@ class AzureFoundryModelProvider extends BaseModelProvider {
                     partial += choice.delta.content;
                     yield { kind: 'partial', partial: safePartialJson(partial) };
                 }
-                const toolCalls = extractToolCalls(choice.delta?.tool_calls, req, 'objectStream');
-                for (const call of toolCalls ?? []) {
-                    yield { kind: 'tool_call', call };
+                if (choice.delta?.tool_calls) {
+                    accumulateToolCallDeltas(toolState, choice.delta.tool_calls);
+                }
+                if (choice.finish_reason) {
+                    finishReason = toFinishReason(choice.finish_reason);
                 }
-                finishReason = toFinishReason(choice.finish_reason ?? finishReason);
             }
             if (data.usage) {
                 usage = toUsage(data.usage.prompt_tokens, data.usage.completion_tokens, data.usage.total_tokens);
             }
         }
+        for (const call of finalizeStreamToolCalls(toolState, req, 'objectStream')) {
+            yield { kind: 'tool_call', call };
+        }
         const object = parseJson(partial || '{}', req, 'objectStream');
         yield { kind: 'finish', object, usage, finishReason };
     }
@@ -163,11 +173,14 @@ async function postChat(client, req, stream) {
             model: req.model,
             messages: toAzureMessages(req.messages),
             stream,
+            // Only emits a usage event during streaming when this is set.
+            ...(stream ? { stream_options: { include_usage: true } } : {}),
             tools: toTools(req.tools),
             temperature: req.call?.temperature ?? req.defaults?.temperature,
             max_tokens: req.call?.maxTokens ?? req.defaults?.maxTokens,
             top_p: req.call?.topP ?? req.defaults?.topP,
             stop: req.call?.stopSequences ?? req.defaults?.stopSequences,
+            ...(req.tools && (req.call?.parallelToolCalls ?? req.defaults?.parallelToolCalls) !== undefined ? { parallel_tool_calls: req.call?.parallelToolCalls ?? req.defaults?.parallelToolCalls } : {}),
             response_format: toResponseFormat(req),
             ...bodyOptions
         },
@@ -179,7 +192,7 @@ async function* streamChat(client, req, objectMode) {
     const response = await postChat(client, req, true);
     const nodeResponse = typeof response.asNodeStream === 'function' ? await response.asNodeStream() : response;
     if (nodeResponse.status && nodeResponse.status !== '200' && nodeResponse.status !== 200) {
-        throw nodeResponse.body?.error ?? new Error('Azure AI Foundry streaming request failed.');
+        throw azureFailure(nodeResponse, 'Azure AI Foundry streaming request failed.');
     }
     if (nodeResponse.body?.[Symbol.asyncIterator]) {
         const sses = createSseStream(nodeResponse.body);
@@ -196,10 +209,25 @@ async function* streamChat(client, req, objectMode) {
 }
 function ensureOk(response) {
     if (response.status && response.status !== '200' && response.status !== 200) {
-        throw response.body?.error ?? new Error('Azure AI Foundry request failed.');
+        throw azureFailure(response, 'Azure AI Foundry request failed.');
     }
     return response.body ?? response;
 }
+/**
+ * Build an error that preserves the HTTP status (and body/headers) so the base
+ * provider's `normalizeError` can classify retriability (429/5xx) instead of
+ * misclassifying every failure as a non-retriable network error.
+ */
+function azureFailure(response, fallbackMessage) {
+    const status = Number(response?.status);
+    const body = response?.body;
+    const message = (body?.error?.message ?? body?.message);
+    return Object.assign(new Error(message ?? fallbackMessage), {
+        ...(Number.isFinite(status) ? { status } : {}),
+        ...(body?.error ? { error: body.error } : body !== undefined ? { body } : {}),
+        ...(response?.headers ? { headers: response.headers } : {})
+    });
+}
 function toAzureMessages(messages) {
     return messages.map((message) => {
         if (message.role === 'assistant' && message.toolCalls && message.toolCalls.length > 0) {
@@ -271,6 +299,29 @@ function extractToolCalls(toolCalls, req, method) {
         arguments: parseJson(call.function.arguments ?? '{}', req, method)
     }));
 }
+function accumulateToolCallDeltas(state, deltas) {
+    for (const delta of deltas) {
+        const index = typeof delta?.index === 'number' ? delta.index : 0;
+        const existing = state.get(index) ?? { args: '' };
+        if (delta?.id)
+            existing.id = String(delta.id);
+        if (delta?.function?.name)
+            existing.name = String(delta.function.name);
+        if (typeof delta?.function?.arguments === 'string')
+            existing.args += delta.function.arguments;
+        state.set(index, existing);
+    }
+}
+function finalizeStreamToolCalls(state, req, method) {
+    return [...state.entries()]
+        .sort((a, b) => a[0] - b[0])
+        .filter(([, call]) => call.id && call.name)
+        .map(([, call]) => ({
+        id: call.id,
+        name: call.name,
+        arguments: parseJson(call.args || '{}', req, method)
+    }));
+}
 function parseStreamData(event, req, method) {
     if (event === '[DONE]')
         return undefined;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@purista/harness-azure-foundry",
-  "version": "1.2.1",
+  "version": "1.2.2",
   "description": "Azure AI Foundry model provider adapter for @purista/harness.",
   "type": "module",
   "main": "./dist/index.js",