@open-multi-agent/core 1.4.0

package/dist/llm/azure-openai.js

@@ -0,0 +1,267 @@
+/**
+ * @fileoverview Azure OpenAI adapter implementing {@link LLMAdapter}.
+ *
+ * Azure OpenAI uses regional deployment endpoints and API versioning that differ
+ * from standard OpenAI:
+ *
+ * - Endpoint: `https://{resource-name}.openai.azure.com`
+ * - API version: Query parameter (e.g., `?api-version=2024-10-21`)
+ * - Model/Deployment: Users deploy models with custom names; the `model` field
+ *   in agent config should contain the Azure deployment name, not the underlying
+ *   model name (e.g., `model: 'my-gpt4-deployment'`)
+ *
+ * The OpenAI SDK provides an `AzureOpenAI` client class that handles these
+ * Azure-specific requirements. This adapter uses that client while reusing all
+ * message conversion logic from `openai-common.ts`.
+ *
+ * Environment variable resolution order:
+ *   1. Constructor arguments
+ *   2. `AZURE_OPENAI_API_KEY` environment variable
+ *   3. `AZURE_OPENAI_ENDPOINT` environment variable
+ *   4. `AZURE_OPENAI_API_VERSION` environment variable (defaults to '2024-10-21')
+ *   5. `AZURE_OPENAI_DEPLOYMENT` as an optional fallback when `model` is blank
+ *
+ * Note: Azure introduced a next-generation v1 API (August 2025) that uses the standard
+ * OpenAI() client with baseURL set to `{endpoint}/openai/v1/` and requires no api-version.
+ * That path is not yet supported by this adapter. To use it, pass `provider: 'openai'`
+ * with `baseURL: 'https://{resource}.openai.azure.com/openai/v1/'` in your agent config.
+ *
+ * @example
+ * ```ts
+ * import { AzureOpenAIAdapter } from './azure-openai.js'
+ *
+ * const adapter = new AzureOpenAIAdapter()
+ * const response = await adapter.chat(messages, {
+ *   model: 'my-gpt4-deployment',  // Azure deployment name, not 'gpt-4'
+ *   maxTokens: 1024,
+ * })
+ * ```
+ */
+import { AzureOpenAI } from 'openai';
+import { toOpenAITool, fromOpenAICompletion, normalizeFinishReason, buildOpenAIMessageList, } from './openai-common.js';
+import { extractToolCallsFromText } from '../tool/text-tool-extractor.js';
+// ---------------------------------------------------------------------------
+// Adapter implementation
+// ---------------------------------------------------------------------------
+const DEFAULT_AZURE_OPENAI_API_VERSION = '2024-10-21';
+function resolveAzureDeploymentName(model) {
+    const explicitModel = model.trim();
+    if (explicitModel.length > 0)
+        return explicitModel;
+    const fallbackDeployment = process.env['AZURE_OPENAI_DEPLOYMENT']?.trim();
+    if (fallbackDeployment !== undefined && fallbackDeployment.length > 0) {
+        return fallbackDeployment;
+    }
+    throw new Error('Azure OpenAI deployment is required. Set agent model to your deployment name, or set AZURE_OPENAI_DEPLOYMENT.');
+}
+/**
+ * LLM adapter backed by Azure OpenAI Chat Completions API.
+ *
+ * Thread-safe — a single instance may be shared across concurrent agent runs.
+ */
+export class AzureOpenAIAdapter {
+    name = 'azure-openai';
+    #client;
+    /**
+     * @param apiKey - Azure OpenAI API key (falls back to AZURE_OPENAI_API_KEY env var)
+     * @param endpoint - Azure endpoint URL (falls back to AZURE_OPENAI_ENDPOINT env var)
+     * @param apiVersion - API version string (falls back to AZURE_OPENAI_API_VERSION, defaults to '2024-10-21')
+     */
+    constructor(apiKey, endpoint, apiVersion) {
+        this.#client = new AzureOpenAI({
+            apiKey: apiKey ?? process.env['AZURE_OPENAI_API_KEY'],
+            endpoint: endpoint ?? process.env['AZURE_OPENAI_ENDPOINT'],
+            apiVersion: apiVersion ?? process.env['AZURE_OPENAI_API_VERSION'] ?? DEFAULT_AZURE_OPENAI_API_VERSION,
+        });
+    }
+    // -------------------------------------------------------------------------
+    // chat()
+    // -------------------------------------------------------------------------
+    /**
+     * Send a synchronous (non-streaming) chat request and return the complete
+     * {@link LLMResponse}.
+     *
+     * Throws an `AzureOpenAI.APIError` on non-2xx responses. Callers should catch and
+     * handle these (e.g. rate limits, context length exceeded, deployment not found).
+     */
+    async chat(messages, options) {
+        const deploymentName = resolveAzureDeploymentName(options.model);
+        const openAIMessages = buildOpenAIMessageList(messages, options.systemPrompt);
+        const completion = await this.#client.chat.completions.create({
+            // Sampling params first so extraBody can override them. Structural
+            // fields (model/messages/tools/stream) come after extraBody so users
+            // cannot accidentally clobber them via extraBody.
+            // `top_k` / `min_p` deliberately omitted — vLLM-only, not accepted
+            // by Azure OpenAI's hosted endpoint.
+            max_tokens: options.maxTokens,
+            temperature: options.temperature,
+            frequency_penalty: options.frequencyPenalty,
+            presence_penalty: options.presencePenalty,
+            top_p: options.topP,
+            parallel_tool_calls: options.parallelToolCalls,
+            reasoning_effort: options.thinking?.effort,
+            ...options.extraBody,
+            model: deploymentName,
+            messages: openAIMessages,
+            tools: options.tools ? options.tools.map(toOpenAITool) : undefined,
+            stream: false,
+        }, {
+            signal: options.abortSignal,
+        });
+        const toolNames = options.tools?.map(t => t.name);
+        return fromOpenAICompletion(completion, toolNames);
+    }
+    // -------------------------------------------------------------------------
+    // stream()
+    // -------------------------------------------------------------------------
+    /**
+     * Send a streaming chat request and yield {@link StreamEvent}s incrementally.
+     *
+     * Sequence guarantees match {@link OpenAIAdapter.stream}:
+     * - Zero or more `text` events
+     * - Zero or more `tool_use` events (emitted once per tool call, after
+     *   arguments have been fully assembled)
+     * - Exactly one terminal event: `done` or `error`
+     */
+    async *stream(messages, options) {
+        const deploymentName = resolveAzureDeploymentName(options.model);
+        const openAIMessages = buildOpenAIMessageList(messages, options.systemPrompt);
+        // We request usage in the final chunk so we can include it in the `done` event.
+        const streamResponse = await this.#client.chat.completions.create({
+            // See chat() above for the field-ordering rationale and the
+            // `top_k` / `min_p` exclusion.
+            max_tokens: options.maxTokens,
+            temperature: options.temperature,
+            frequency_penalty: options.frequencyPenalty,
+            presence_penalty: options.presencePenalty,
+            top_p: options.topP,
+            parallel_tool_calls: options.parallelToolCalls,
+            reasoning_effort: options.thinking?.effort,
+            ...options.extraBody,
+            model: deploymentName,
+            messages: openAIMessages,
+            tools: options.tools ? options.tools.map(toOpenAITool) : undefined,
+            stream: true,
+            stream_options: { include_usage: true },
+        }, {
+            signal: options.abortSignal,
+        });
+        // Accumulate state across chunks.
+        let completionId = '';
+        let completionModel = '';
+        let finalFinishReason = 'stop';
+        let inputTokens = 0;
+        let outputTokens = 0;
+        // tool_calls are streamed piecemeal; key = tool call index
+        const toolCallBuffers = new Map();
+        // Full text accumulator for the `done` response.
+        let fullText = '';
+        try {
+            for await (const chunk of streamResponse) {
+                completionId = chunk.id;
+                completionModel = chunk.model;
+                // Usage is only populated in the final chunk when stream_options.include_usage is set.
+                if (chunk.usage !== null && chunk.usage !== undefined) {
+                    inputTokens = chunk.usage.prompt_tokens;
+                    outputTokens = chunk.usage.completion_tokens;
+                }
+                const choice = chunk.choices[0];
+                if (choice === undefined)
+                    continue;
+                const delta = choice.delta;
+                // --- text delta ---
+                if (delta.content !== null && delta.content !== undefined) {
+                    fullText += delta.content;
+                    const textEvent = { type: 'text', data: delta.content };
+                    yield textEvent;
+                }
+                // --- tool call delta ---
+                for (const toolCallDelta of delta.tool_calls ?? []) {
+                    const idx = toolCallDelta.index;
+                    if (!toolCallBuffers.has(idx)) {
+                        toolCallBuffers.set(idx, {
+                            id: toolCallDelta.id ?? '',
+                            name: toolCallDelta.function?.name ?? '',
+                            argsJson: '',
+                        });
+                    }
+                    const buf = toolCallBuffers.get(idx);
+                    // buf is guaranteed to exist: we just set it above.
+                    if (buf !== undefined) {
+                        if (toolCallDelta.id)
+                            buf.id = toolCallDelta.id;
+                        if (toolCallDelta.function?.name)
+                            buf.name = toolCallDelta.function.name;
+                        if (toolCallDelta.function?.arguments) {
+                            buf.argsJson += toolCallDelta.function.arguments;
+                        }
+                    }
+                }
+                if (choice.finish_reason !== null && choice.finish_reason !== undefined) {
+                    finalFinishReason = choice.finish_reason;
+                }
+            }
+            // Emit accumulated tool_use events after the stream ends.
+            const finalToolUseBlocks = [];
+            for (const buf of toolCallBuffers.values()) {
+                let parsedInput = {};
+                try {
+                    const parsed = JSON.parse(buf.argsJson);
+                    if (parsed !== null && typeof parsed === 'object' && !Array.isArray(parsed)) {
+                        parsedInput = parsed;
+                    }
+                }
+                catch {
+                    // Malformed JSON — surface as empty object.
+                }
+                const toolUseBlock = {
+                    type: 'tool_use',
+                    id: buf.id,
+                    name: buf.name,
+                    input: parsedInput,
+                };
+                finalToolUseBlocks.push(toolUseBlock);
+                const toolUseEvent = { type: 'tool_use', data: toolUseBlock };
+                yield toolUseEvent;
+            }
+            // Build the complete content array for the done response.
+            const doneContent = [];
+            if (fullText.length > 0) {
+                const textBlock = { type: 'text', text: fullText };
+                doneContent.push(textBlock);
+            }
+            doneContent.push(...finalToolUseBlocks);
+            // Fallback: extract tool calls from text when streaming produced no
+            // native tool_calls (same logic as fromOpenAICompletion).
+            if (finalToolUseBlocks.length === 0 && fullText.length > 0 && options.tools) {
+                const toolNames = options.tools.map(t => t.name);
+                const extracted = extractToolCallsFromText(fullText, toolNames);
+                if (extracted.length > 0) {
+                    doneContent.push(...extracted);
+                    for (const block of extracted) {
+                        yield { type: 'tool_use', data: block };
+                    }
+                }
+            }
+            const hasToolUseBlocks = doneContent.some(b => b.type === 'tool_use');
+            const resolvedStopReason = hasToolUseBlocks && finalFinishReason === 'stop'
+                ? 'tool_use'
+                : normalizeFinishReason(finalFinishReason);
+            const finalResponse = {
+                id: completionId,
+                content: doneContent,
+                model: completionModel,
+                stop_reason: resolvedStopReason,
+                usage: { input_tokens: inputTokens, output_tokens: outputTokens },
+            };
+            const doneEvent = { type: 'done', data: finalResponse };
+            yield doneEvent;
+        }
+        catch (err) {
+            const error = err instanceof Error ? err : new Error(String(err));
+            const errorEvent = { type: 'error', data: error };
+            yield errorEvent;
+        }
+    }
+}
+//# sourceMappingURL=azure-openai.js.map

package/dist/llm/azure-openai.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"azure-openai.js","sourceRoot":"","sources":["../../src/llm/azure-openai.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAsCG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,QAAQ,CAAA;AAiBpC,OAAO,EACL,YAAY,EACZ,oBAAoB,EACpB,qBAAqB,EACrB,sBAAsB,GACvB,MAAM,oBAAoB,CAAA;AAC3B,OAAO,EAAE,wBAAwB,EAAE,MAAM,gCAAgC,CAAA;AAEzE,8EAA8E;AAC9E,yBAAyB;AACzB,8EAA8E;AAE9E,MAAM,gCAAgC,GAAG,YAAY,CAAA;AAErD,SAAS,0BAA0B,CAAC,KAAa;IAC/C,MAAM,aAAa,GAAG,KAAK,CAAC,IAAI,EAAE,CAAA;IAClC,IAAI,aAAa,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,aAAa,CAAA;IAElD,MAAM,kBAAkB,GAAG,OAAO,CAAC,GAAG,CAAC,yBAAyB,CAAC,EAAE,IAAI,EAAE,CAAA;IACzE,IAAI,kBAAkB,KAAK,SAAS,IAAI,kBAAkB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACtE,OAAO,kBAAkB,CAAA;IAC3B,CAAC;IAED,MAAM,IAAI,KAAK,CACb,+GAA+G,CAChH,CAAA;AACH,CAAC;AAED;;;;GAIG;AACH,MAAM,OAAO,kBAAkB;IACpB,IAAI,GAAW,cAAc,CAAA;IAE7B,OAAO,CAAa;IAE7B;;;;OAIG;IACH,YAAY,MAAe,EAAE,QAAiB,EAAE,UAAmB;QACjE,IAAI,CAAC,OAAO,GAAG,IAAI,WAAW,CAAC;YAC7B,MAAM,EAAE,MAAM,IAAI,OAAO,CAAC,GAAG,CAAC,sBAAsB,CAAC;YACrD,QAAQ,EAAE,QAAQ,IAAI,OAAO,CAAC,GAAG,CAAC,uBAAuB,CAAC;YAC1D,UAAU,EAAE,UAAU,IAAI,OAAO,CAAC,GAAG,CAAC,0BAA0B,CAAC,IAAI,gCAAgC;SACtG,CAAC,CAAA;IACJ,CAAC;IAED,4EAA4E;IAC5E,SAAS;IACT,4EAA4E;IAE5E;;;;;;OAMG;IACH,KAAK,CAAC,IAAI,CAAC,QAAsB,EAAE,OAAuB;QACxD,MAAM,cAAc,GAAG,0BAA0B,CAAC,OAAO,CAAC,KAAK,CAAC,CAAA;QAChE,MAAM,cAAc,GAAG,sBAAsB,CAAC,QAAQ,EAAE,OAAO,CAAC,YAAY,CAAC,CAAA;QAE7E,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,WAAW,CAAC,MAAM,CAC3D;YACE,mEAAmE;YACnE,qEAAqE;YACrE,kDAAkD;YAClD,mEAAmE;YACnE,qCAAqC;YACrC,UAAU,EAAE,OAAO,CAAC,SAAS;YAC7B,WAAW,EAAE,OAAO,CAAC,WAAW;YAChC,iBAAiB,EAAE,OAAO,CAAC,gBAAgB;YAC3C,gBAAgB,EAAE,OAAO,CAAC,eAAe;YACzC,KAAK,EAAE,OAAO,CAAC,IAAI;YACnB,mBAAmB,EAAE,OAAO,CAAC,iBAAiB;YAC9C,gBAAgB,EAAE,OAAO,CAAC,QAAQ,EAAE,MAAM;YAC1C,GAAG,OAAO,CAAC,SAAS;YACpB,KAAK,EAAE,cAAc;YACrB,QAAQ,EAAE,cAAc;YACxB,KAAK,EAAE,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,SAAS;YAClE,MAAM,EAAE,KAAK;SACd,EACD;YACE,MAAM,EAAE,OAAO,CAAC,WAAW;SAC5B,CACF,CAAA;QAED,MAAM,SAAS,GAAG,OAAO,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAA;QACjD,OAAO,oBAAoB,CAAC,UAAU,EAAE,SAAS,CAAC,CAAA;IACpD,CAAC;IAED,4EAA4E;IAC5E,WAAW;IACX,4EAA4E;IAE5E;;;;;;;;OAQG;IACH,KAAK,CAAC,CAAC,MAAM,CACX,QAAsB,EACtB,OAAyB;QAEzB,MAAM,cAAc,GAAG,0BAA0B,CAAC,OAAO,CAAC,KAAK,CAAC,CAAA;QAChE,MAAM,cAAc,GAAG,sBAAsB,CAAC,QAAQ,EAAE,OAAO,CAAC,YAAY,CAAC,CAAA;QAE7E,gFAAgF;QAChF,MAAM,cAAc,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,WAAW,CAAC,MAAM,CAC/D;YACE,4DAA4D;YAC5D,+BAA+B;YAC/B,UAAU,EAAE,OAAO,CAAC,SAAS;YAC7B,WAAW,EAAE,OAAO,CAAC,WAAW;YAChC,iBAAiB,EAAE,OAAO,CAAC,gBAAgB;YAC3C,gBAAgB,EAAE,OAAO,CAAC,eAAe;YACzC,KAAK,EAAE,OAAO,CAAC,IAAI;YACnB,mBAAmB,EAAE,OAAO,CAAC,iBAAiB;YAC9C,gBAAgB,EAAE,OAAO,CAAC,QAAQ,EAAE,MAAM;YAC1C,GAAG,OAAO,CAAC,SAAS;YACpB,KAAK,EAAE,cAAc;YACrB,QAAQ,EAAE,cAAc;YACxB,KAAK,EAAE,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,SAAS;YAClE,MAAM,EAAE,IAAI;YACZ,cAAc,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE;SACxC,EACD;YACE,MAAM,EAAE,OAAO,CAAC,WAAW;SAC5B,CACF,CAAA;QAED,kCAAkC;QAClC,IAAI,YAAY,GAAG,EAAE,CAAA;QACrB,IAAI,eAAe,GAAG,EAAE,CAAA;QACxB,IAAI,iBAAiB,GAAW,MAAM,CAAA;QACtC,IAAI,WAAW,GAAG,CAAC,CAAA;QACnB,IAAI,YAAY,GAAG,CAAC,CAAA;QAEpB,2DAA2D;QAC3D,MAAM,eAAe,GAAG,IAAI,GAAG,EAG5B,CAAA;QAEH,iDAAiD;QACjD,IAAI,QAAQ,GAAG,EAAE,CAAA;QAEjB,IAAI,CAAC;YACH,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,cAAc,EAAE,CAAC;gBACzC,YAAY,GAAG,KAAK,CAAC,EAAE,CAAA;gBACvB,eAAe,GAAG,KAAK,CAAC,KAAK,CAAA;gBAE7B,uFAAuF;gBACvF,IAAI,KAAK,CAAC,KAAK,KAAK,IAAI,IAAI,KAAK,CAAC,KAAK,KAAK,SAAS,EAAE,CAAC;oBACtD,WAAW,GAAG,KAAK,CAAC,KAAK,CAAC,aAAa,CAAA;oBACvC,YAAY,GAAG,KAAK,CAAC,KAAK,CAAC,iBAAiB,CAAA;gBAC9C,CAAC;gBAED,MAAM,MAAM,GAA2C,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAA;gBACvE,IAAI,MAAM,KAAK,SAAS;oBAAE,SAAQ;gBAElC,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,CAAA;gBAE1B,qBAAqB;gBACrB,IAAI,KAAK,CAAC,OAAO,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,KAAK,SAAS,EAAE,CAAC;oBAC1D,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAA;oBACzB,MAAM,SAAS,GAAgB,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,CAAC,OAAO,EAAE,CAAA;oBACpE,MAAM,SAAS,CAAA;gBACjB,CAAC;gBAED,0BAA0B;gBAC1B,KAAK,MAAM,aAAa,IAAI,KAAK,CAAC,UAAU,IAAI,EAAE,EAAE,CAAC;oBACnD,MAAM,GAAG,GAAG,aAAa,CAAC,KAAK,CAAA;oBAE/B,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC;wBAC9B,eAAe,CAAC,GAAG,CAAC,GAAG,EAAE;4BACvB,EAAE,EAAE,aAAa,CAAC,EAAE,IAAI,EAAE;4BAC1B,IAAI,EAAE,aAAa,CAAC,QAAQ,EAAE,IAAI,IAAI,EAAE;4BACxC,QAAQ,EAAE,EAAE;yBACb,CAAC,CAAA;oBACJ,CAAC;oBAED,MAAM,GAAG,GAAG,eAAe,CAAC,GAAG,CAAC,GAAG,CAAC,CAAA;oBACpC,oDAAoD;oBACpD,IAAI,GAAG,KAAK,SAAS,EAAE,CAAC;wBACtB,IAAI,aAAa,CAAC,EAAE;4BAAE,GAAG,CAAC,EAAE,GAAG,aAAa,CAAC,EAAE,CAAA;wBAC/C,IAAI,aAAa,CAAC,QAAQ,EAAE,IAAI;4BAAE,GAAG,CAAC,IAAI,GAAG,aAAa,CAAC,QAAQ,CAAC,IAAI,CAAA;wBACxE,IAAI,aAAa,CAAC,QAAQ,EAAE,SAAS,EAAE,CAAC;4BACtC,GAAG,CAAC,QAAQ,IAAI,aAAa,CAAC,QAAQ,CAAC,SAAS,CAAA;wBAClD,CAAC;oBACH,CAAC;gBACH,CAAC;gBAED,IAAI,MAAM,CAAC,aAAa,KAAK,IAAI,IAAI,MAAM,CAAC,aAAa,KAAK,SAAS,EAAE,CAAC;oBACxE,iBAAiB,GAAG,MAAM,CAAC,aAAa,CAAA;gBAC1C,CAAC;YACH,CAAC;YAED,0DAA0D;YAC1D,MAAM,kBAAkB,GAAmB,EAAE,CAAA;YAC7C,KAAK,MAAM,GAAG,IAAI,eAAe,CAAC,MAAM,EAAE,EAAE,CAAC;gBAC3C,IAAI,WAAW,GAA4B,EAAE,CAAA;gBAC7C,IAAI,CAAC;oBACH,MAAM,MAAM,GAAY,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAA;oBAChD,IAAI,MAAM,KAAK,IAAI,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;wBAC5E,WAAW,GAAG,MAAiC,CAAA;oBACjD,CAAC;gBACH,CAAC;gBAAC,MAAM,CAAC;oBACP,4CAA4C;gBAC9C,CAAC;gBAED,MAAM,YAAY,GAAiB;oBACjC,IAAI,EAAE,UAAU;oBAChB,EAAE,EAAE,GAAG,CAAC,EAAE;oBACV,IAAI,EAAE,GAAG,CAAC,IAAI;oBACd,KAAK,EAAE,WAAW;iBACnB,CAAA;gBACD,kBAAkB,CAAC,IAAI,CAAC,YAAY,CAAC,CAAA;gBACrC,MAAM,YAAY,GAAgB,EAAE,IAAI,EAAE,UAAU,EAAE,IAAI,EAAE,YAAY,EAAE,CAAA;gBAC1E,MAAM,YAAY,CAAA;YACpB,CAAC;YAED,0DAA0D;YAC1D,MAAM,WAAW,GAAmB,EAAE,CAAA;YACtC,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACxB,MAAM,SAAS,GAAc,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAA;gBAC7D,WAAW,CAAC,IAAI,CAAC,SAAS,CAAC,CAAA;YAC7B,CAAC;YACD,WAAW,CAAC,IAAI,CAAC,GAAG,kBAAkB,CAAC,CAAA;YAEvC,oEAAoE;YACpE,0DAA0D;YAC1D,IAAI,kBAAkB,CAAC,MAAM,KAAK,CAAC,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC,IAAI,OAAO,CAAC,KAAK,EAAE,CAAC;gBAC5E,MAAM,SAAS,GAAG,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAA;gBAChD,MAAM,SAAS,GAAG,wBAAwB,CAAC,QAAQ,EAAE,SAAS,CAAC,CAAA;gBAC/D,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;oBACzB,WAAW,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,CAAA;oBAC9B,KAAK,MAAM,KAAK,IAAI,SAAS,EAAE,CAAC;wBAC9B,MAAM,EAAE,IAAI,EAAE,UAAU,EAAE,IAAI,EAAE,KAAK,EAAwB,CAAA;oBAC/D,CAAC;gBACH,CAAC;YACH,CAAC;YAED,MAAM,gBAAgB,GAAG,WAAW,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,UAAU,CAAC,CAAA;YACrE,MAAM,kBAAkB,GAAG,gBAAgB,IAAI,iBAAiB,KAAK,MAAM;gBACzE,CAAC,CAAC,UAAU;gBACZ,CAAC,CAAC,qBAAqB,CAAC,iBAAiB,CAAC,CAAA;YAE5C,MAAM,aAAa,GAAgB;gBACjC,EAAE,EAAE,YAAY;gBAChB,OAAO,EAAE,WAAW;gBACpB,KAAK,EAAE,eAAe;gBACtB,WAAW,EAAE,kBAAkB;gBAC/B,KAAK,EAAE,EAAE,YAAY,EAAE,WAAW,EAAE,aAAa,EAAE,YAAY,EAAE;aAClE,CAAA;YAED,MAAM,SAAS,GAAgB,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,aAAa,EAAE,CAAA;YACpE,MAAM,SAAS,CAAA;QACjB,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,KAAK,GAAG,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAA;YACjE,MAAM,UAAU,GAAgB,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,CAAA;YAC9D,MAAM,UAAU,CAAA;QAClB,CAAC;IACH,CAAC;CACF"}

package/dist/llm/bedrock.d.ts

@@ -0,0 +1,41 @@
+/**
+ * @fileoverview AWS Bedrock LLM adapter implementing {@link LLMAdapter}.
+ *
+ * Uses the Converse / ConverseStream APIs from `@aws-sdk/client-bedrock-runtime`
+ * (the unified Anthropic-shaped schema) so the same adapter works across Claude,
+ * Llama, Mistral, Cohere, and Titan model families without per-model shims.
+ *
+ * Region resolution order:
+ *   1. `region` constructor argument
+ *   2. `AWS_REGION` environment variable
+ *   3. `'us-east-1'` (hard fallback)
+ *
+ * AWS credentials are resolved via the SDK default provider chain:
+ * env vars → shared config file → EC2/ECS/Lambda IAM role.
+ *
+ * @example
+ * ```ts
+ * import { BedrockAdapter } from './bedrock.js'
+ *
+ * const adapter = new BedrockAdapter('us-east-1')
+ * const response = await adapter.chat(messages, {
+ *   model: 'anthropic.claude-3-5-haiku-20241022-v1:0',
+ *   maxTokens: 1024,
+ * })
+ * ```
+ */
+import type { ContentBlock, ImageBlock, LLMAdapter, LLMChatOptions, LLMMessage, LLMResponse, LLMStreamOptions, LLMToolDef, StreamEvent, TextBlock, ToolResultBlock, ToolUseBlock } from '../types.js';
+/**
+ * LLM adapter backed by AWS Bedrock Converse / ConverseStream APIs.
+ *
+ * Thread-safe — a single instance may be shared across concurrent agent runs.
+ */
+export declare class BedrockAdapter implements LLMAdapter {
+    #private;
+    readonly name = "bedrock";
+    constructor(region?: string);
+    chat(messages: LLMMessage[], options: LLMChatOptions): Promise<LLMResponse>;
+    stream(messages: LLMMessage[], options: LLMStreamOptions): AsyncIterable<StreamEvent>;
+}
+export type { ContentBlock, ImageBlock, LLMAdapter, LLMChatOptions, LLMMessage, LLMResponse, LLMStreamOptions, LLMToolDef, StreamEvent, TextBlock, ToolResultBlock, ToolUseBlock };
+//# sourceMappingURL=bedrock.d.ts.map

package/dist/llm/bedrock.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"bedrock.d.ts","sourceRoot":"","sources":["../../src/llm/bedrock.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAeH,OAAO,KAAK,EACV,YAAY,EACZ,UAAU,EACV,UAAU,EACV,cAAc,EACd,UAAU,EACV,WAAW,EACX,gBAAgB,EAChB,UAAU,EAEV,WAAW,EACX,SAAS,EACT,eAAe,EACf,YAAY,EACb,MAAM,aAAa,CAAA;AA2JpB;;;;GAIG;AACH,qBAAa,cAAe,YAAW,UAAU;;IAC/C,QAAQ,CAAC,IAAI,aAAY;gBAIb,MAAM,CAAC,EAAE,MAAM;IASrB,IAAI,CAAC,QAAQ,EAAE,UAAU,EAAE,EAAE,OAAO,EAAE,cAAc,GAAG,OAAO,CAAC,WAAW,CAAC;IA4C1E,MAAM,CAAC,QAAQ,EAAE,UAAU,EAAE,EAAE,OAAO,EAAE,gBAAgB,GAAG,aAAa,CAAC,WAAW,CAAC;CAkI7F;AAED,YAAY,EAAE,YAAY,EAAE,UAAU,EAAE,UAAU,EAAE,cAAc,EAAE,UAAU,EAAE,WAAW,EAAE,gBAAgB,EAAE,UAAU,EAAE,WAAW,EAAE,SAAS,EAAE,eAAe,EAAE,YAAY,EAAE,CAAA"}

package/dist/llm/bedrock.js

@@ -0,0 +1,345 @@
+/**
+ * @fileoverview AWS Bedrock LLM adapter implementing {@link LLMAdapter}.
+ *
+ * Uses the Converse / ConverseStream APIs from `@aws-sdk/client-bedrock-runtime`
+ * (the unified Anthropic-shaped schema) so the same adapter works across Claude,
+ * Llama, Mistral, Cohere, and Titan model families without per-model shims.
+ *
+ * Region resolution order:
+ *   1. `region` constructor argument
+ *   2. `AWS_REGION` environment variable
+ *   3. `'us-east-1'` (hard fallback)
+ *
+ * AWS credentials are resolved via the SDK default provider chain:
+ * env vars → shared config file → EC2/ECS/Lambda IAM role.
+ *
+ * @example
+ * ```ts
+ * import { BedrockAdapter } from './bedrock.js'
+ *
+ * const adapter = new BedrockAdapter('us-east-1')
+ * const response = await adapter.chat(messages, {
+ *   model: 'anthropic.claude-3-5-haiku-20241022-v1:0',
+ *   maxTokens: 1024,
+ * })
+ * ```
+ */
+import { BedrockRuntimeClient, ConverseCommand, ConverseStreamCommand, } from '@aws-sdk/client-bedrock-runtime';
+// ---------------------------------------------------------------------------
+// Internal helpers
+// ---------------------------------------------------------------------------
+const MEDIA_TYPE_TO_FORMAT = {
+    'image/jpeg': 'jpeg',
+    'image/png': 'png',
+    'image/gif': 'gif',
+    'image/webp': 'webp',
+};
+function base64ToUint8Array(b64) {
+    return Buffer.from(b64, 'base64');
+}
+/**
+ * Convert a single framework {@link ContentBlock} into a Bedrock
+ * {@link BedrockContentBlock} for the messages array.
+ *
+ * Reasoning blocks are dropped on input — see the `case 'reasoning'` arm below
+ * for the round-trip limitation that forces this.
+ */
+function toBedrockContentBlock(block) {
+    switch (block.type) {
+        case 'text':
+            return { text: block.text };
+        case 'tool_use':
+            // DocumentType is not publicly exported from the SDK; cast the whole block.
+            return {
+                toolUse: { toolUseId: block.id, name: block.name, input: block.input },
+            };
+        case 'tool_result': {
+            const rawContent = block.content;
+            const textContent = typeof rawContent === 'string' ? rawContent : JSON.stringify(rawContent);
+            return {
+                toolResult: {
+                    toolUseId: block.tool_use_id,
+                    content: [{ text: textContent }],
+                    status: block.is_error ? 'error' : 'success',
+                },
+            };
+        }
+        case 'image': {
+            const format = MEDIA_TYPE_TO_FORMAT[block.source.media_type] ?? 'png';
+            return {
+                image: {
+                    format,
+                    source: { bytes: base64ToUint8Array(block.source.data) },
+                },
+            };
+        }
+        case 'reasoning':
+            // Bedrock requires the original reasoning text plus its `signature`
+            // (and `redactedData` for redacted blocks) to be echoed back unchanged
+            // when continuing a multi-turn tool-use conversation with extended
+            // thinking. The IR now carries these fields on {@link ReasoningBlock}
+            // (added in #200), but wiring them through Bedrock's
+            // `reasoningContent.reasoningText.signature` shape has been deferred
+            // to a follow-up PR — until then, we drop reasoning on input.
+            return null;
+        default: {
+            const _exhaustive = block;
+            throw new Error(`Unhandled content block type: ${JSON.stringify(_exhaustive)}`);
+        }
+    }
+}
+/**
+ * Convert framework messages into Bedrock `Message[]`.
+ *
+ * System prompt is passed separately via `options.systemPrompt` and handled
+ * in `chat()`/`stream()` — it never appears as a message role in the framework.
+ * Reasoning blocks are silently dropped on input — see {@link toBedrockContentBlock}
+ * for why; multi-turn tool-use flows with extended thinking are tracked as
+ * a follow-up to #200's phase-1 IR additions.
+ */
+function toBedrockMessages(messages) {
+    const bedrockMessages = [];
+    for (const msg of messages) {
+        const content = [];
+        for (const block of msg.content) {
+            const converted = toBedrockContentBlock(block);
+            if (converted !== null)
+                content.push(converted);
+        }
+        if (content.length > 0) {
+            bedrockMessages.push({ role: msg.role, content });
+        }
+    }
+    return bedrockMessages;
+}
+function toBedrockTools(tools) {
+    // Tool is a discriminated union with a required $unknown variant; cast to satisfy it.
+    const bedrockTools = tools.map((t) => ({
+        toolSpec: {
+            name: t.name,
+            description: t.description,
+            inputSchema: { json: t.inputSchema },
+        },
+    }));
+    return { tools: bedrockTools };
+}
+/**
+ * Convert a Bedrock response {@link BedrockContentBlock} into a framework
+ * {@link ContentBlock}.
+ *
+ * `toolUse.input` arrives as a parsed object in chat() responses.
+ */
+function fromBedrockContentBlock(block) {
+    if (block.text !== undefined) {
+        const text = { type: 'text', text: block.text };
+        return text;
+    }
+    if (block.toolUse !== undefined) {
+        const toolUse = {
+            type: 'tool_use',
+            id: block.toolUse.toolUseId ?? '',
+            name: block.toolUse.name ?? '',
+            input: block.toolUse.input ?? {},
+        };
+        return toolUse;
+    }
+    if (block.reasoningContent !== undefined) {
+        const r = block.reasoningContent;
+        const reasoning = {
+            type: 'reasoning',
+            text: r.reasoningText?.text ?? '',
+        };
+        return reasoning;
+    }
+    return null;
+}
+function buildInferenceConfig(options) {
+    const cfg = {};
+    if (options.maxTokens !== undefined)
+        cfg.maxTokens = options.maxTokens ?? 4096;
+    if (options.temperature !== undefined)
+        cfg.temperature = options.temperature;
+    if (options.topP !== undefined)
+        cfg.topP = options.topP;
+    return Object.keys(cfg).length > 0 ? cfg : undefined;
+}
+// ---------------------------------------------------------------------------
+// Adapter implementation
+// ---------------------------------------------------------------------------
+/**
+ * LLM adapter backed by AWS Bedrock Converse / ConverseStream APIs.
+ *
+ * Thread-safe — a single instance may be shared across concurrent agent runs.
+ */
+export class BedrockAdapter {
+    name = 'bedrock';
+    #client;
+    constructor(region) {
+        const resolvedRegion = region ?? process.env['AWS_REGION'] ?? 'us-east-1';
+        this.#client = new BedrockRuntimeClient({ region: resolvedRegion });
+    }
+    // -------------------------------------------------------------------------
+    // chat()
+    // -------------------------------------------------------------------------
+    async chat(messages, options) {
+        const bedrockMessages = toBedrockMessages(messages);
+        const system = options.systemPrompt ? [{ text: options.systemPrompt }] : undefined;
+        const input = {
+            modelId: options.model,
+            messages: bedrockMessages,
+            system,
+            toolConfig: options.tools ? toBedrockTools(options.tools) : undefined,
+            inferenceConfig: buildInferenceConfig(options) ?? { maxTokens: 4096 },
+        };
+        if (options.topK !== undefined || options.extraBody) {
+            input.additionalModelRequestFields = {
+                ...(options.topK !== undefined ? { top_k: options.topK } : {}),
+                ...options.extraBody,
+            };
+        }
+        const response = await this.#client.send(new ConverseCommand(input), {
+            abortSignal: options.abortSignal,
+        });
+        const rawContent = response.output?.message?.content ?? [];
+        const content = rawContent
+            .map(fromBedrockContentBlock)
+            .filter((b) => b !== null);
+        return {
+            id: response.$metadata.requestId ?? '',
+            content,
+            model: options.model,
+            stop_reason: response.stopReason ?? 'end_turn',
+            usage: {
+                input_tokens: response.usage?.inputTokens ?? 0,
+                output_tokens: response.usage?.outputTokens ?? 0,
+            },
+        };
+    }
+    // -------------------------------------------------------------------------
+    // stream()
+    // -------------------------------------------------------------------------
+    async *stream(messages, options) {
+        const bedrockMessages = toBedrockMessages(messages);
+        const system = options.systemPrompt ? [{ text: options.systemPrompt }] : undefined;
+        const input = {
+            modelId: options.model,
+            messages: bedrockMessages,
+            system,
+            toolConfig: options.tools ? toBedrockTools(options.tools) : undefined,
+            inferenceConfig: buildInferenceConfig(options) ?? { maxTokens: 4096 },
+        };
+        if (options.topK !== undefined || options.extraBody) {
+            input.additionalModelRequestFields = {
+                ...(options.topK !== undefined ? { top_k: options.topK } : {}),
+                ...options.extraBody,
+            };
+        }
+        // Accumulate tool-use input JSON deltas; keyed by content block index.
+        const toolBuffers = new Map();
+        // Accumulate reasoning text deltas; keyed by content block index. Each
+        // index becomes one ReasoningBlock in the final `done` payload, matching
+        // what `chat()` produces for the same response shape.
+        const reasoningBuffers = new Map();
+        // Accumulated content blocks for the done event.
+        const accumulatedContent = [];
+        let stopReason = 'end_turn';
+        let inputTokens = 0;
+        let outputTokens = 0;
+        const requestId = '';
+        try {
+            const response = await this.#client.send(new ConverseStreamCommand(input), {
+                abortSignal: options.abortSignal,
+            });
+            for await (const event of response.stream ?? []) {
+                if (event.contentBlockStart?.start?.toolUse) {
+                    const { toolUseId, name } = event.contentBlockStart.start.toolUse;
+                    const index = event.contentBlockStart.contentBlockIndex ?? 0;
+                    toolBuffers.set(index, { toolUseId: toolUseId ?? '', name: name ?? '', json: '' });
+                }
+                if (event.contentBlockDelta?.delta) {
+                    const delta = event.contentBlockDelta.delta;
+                    const index = event.contentBlockDelta.contentBlockIndex ?? 0;
+                    if (delta.text !== undefined) {
+                        const textEvent = { type: 'text', data: delta.text };
+                        yield textEvent;
+                        accumulatedContent.push({ type: 'text', text: delta.text });
+                    }
+                    else if (delta.toolUse?.input !== undefined) {
+                        const buf = toolBuffers.get(index);
+                        if (buf)
+                            buf.json += delta.toolUse.input;
+                    }
+                    else if (delta.reasoningContent?.text !== undefined) {
+                        const text = delta.reasoningContent.text;
+                        const reasoningEvent = { type: 'reasoning', data: text };
+                        yield reasoningEvent;
+                        const buf = reasoningBuffers.get(index) ?? { text: '' };
+                        buf.text += text;
+                        reasoningBuffers.set(index, buf);
+                    }
+                }
+                if (event.contentBlockStop !== undefined) {
+                    const index = event.contentBlockStop.contentBlockIndex ?? 0;
+                    const reasoningBuf = reasoningBuffers.get(index);
+                    if (reasoningBuf) {
+                        const reasoningBlock = { type: 'reasoning', text: reasoningBuf.text };
+                        accumulatedContent.push(reasoningBlock);
+                        reasoningBuffers.delete(index);
+                    }
+                    const buf = toolBuffers.get(index);
+                    if (buf) {
+                        let parsedInput = {};
+                        try {
+                            const parsed = JSON.parse(buf.json || '{}');
+                            if (parsed !== null && typeof parsed === 'object' && !Array.isArray(parsed)) {
+                                parsedInput = parsed;
+                            }
+                        }
+                        catch {
+                            // malformed JSON → empty object
+                        }
+                        const toolUseBlock = {
+                            type: 'tool_use',
+                            id: buf.toolUseId,
+                            name: buf.name,
+                            input: parsedInput,
+                        };
+                        accumulatedContent.push(toolUseBlock);
+                        const toolUseEvent = { type: 'tool_use', data: toolUseBlock };
+                        yield toolUseEvent;
+                        toolBuffers.delete(index);
+                    }
+                }
+                if (event.messageStop) {
+                    stopReason = event.messageStop.stopReason ?? 'end_turn';
+                }
+                if (event.metadata?.usage) {
+                    inputTokens = event.metadata.usage.inputTokens ?? 0;
+                    outputTokens = event.metadata.usage.outputTokens ?? 0;
+                }
+            }
+            // Safety net: if Bedrock ever omits a contentBlockStop for a reasoning
+            // block, flush whatever we buffered so the done payload still matches
+            // what chat() would have returned for the same response.
+            for (const [, buf] of reasoningBuffers) {
+                accumulatedContent.push({ type: 'reasoning', text: buf.text });
+            }
+            reasoningBuffers.clear();
+            const finalResponse = {
+                id: requestId,
+                content: accumulatedContent,
+                model: options.model,
+                stop_reason: stopReason,
+                usage: { input_tokens: inputTokens, output_tokens: outputTokens },
+            };
+            const doneEvent = { type: 'done', data: finalResponse };
+            yield doneEvent;
+        }
+        catch (err) {
+            const error = err instanceof Error ? err : new Error(String(err));
+            const errorEvent = { type: 'error', data: error };
+            yield errorEvent;
+        }
+    }
+}
+//# sourceMappingURL=bedrock.js.map