@sunerpy/opencode-kiro-auth 0.4.2 → 0.4.3

package/README.md

@@ -192,6 +192,12 @@ Details and the full JSON example: [docs/MODELS.md](docs/MODELS.md).
 > `kiro.json` instead. See
 > [Reasoning effort](docs/CONFIGURATION.md#reasoning-effort) for details.
 
+> **Note:** Reasoning-capable Kiro models (Claude Opus 4.x and other
+> reasoning-capable models) stream their chain-of-thought as a separate
+> event, which the plugin surfaces as OpenCode's own reasoning block — shown
+> as "Thought: `<duration>`" above the final reply. No config needed. See
+> [Reasoning display](docs/CONFIGURATION.md#reasoning-display) for details.
+
 ## Troubleshooting
 
 Common issues — 403/AccessDeniedException with IAM Identity Center, "No

package/dist/core/request/response-handler.js

@@ -84,12 +84,16 @@ export class ResponseHandler {
     async handleSdkNonStreaming(sdkResponse, model, conversationId) {
         // For non-streaming SDK responses, collect all events
         let content = '';
+        let reasoningContent = '';
         const toolCalls = [];
         let inputTokens = 0;
         let outputTokens = 0;
         const eventStream = sdkResponse.generateAssistantResponseResponse;
         if (eventStream) {
             for await (const event of eventStream) {
+                if (event.reasoningContentEvent?.text) {
+                    reasoningContent += event.reasoningContentEvent.text;
+                }
                 if (event.assistantResponseEvent?.content) {
                     content += event.assistantResponseEvent.content;
                 }
@@ -102,6 +106,10 @@ export class ResponseHandler {
                 }
             }
         }
+        const message = { role: 'assistant', content };
+        if (reasoningContent) {
+            message.reasoning_content = reasoningContent;
+        }
         const oai = {
             id: conversationId,
             object: 'chat.completion',
@@ -110,7 +118,7 @@ export class ResponseHandler {
             choices: [
                 {
                     index: 0,
-                    message: { role: 'assistant', content },
+                    message,
                     finish_reason: toolCalls.length > 0 ? 'tool_calls' : 'stop'
                 }
             ],

package/dist/plugin/streaming/sdk-stream-transformer.js

@@ -24,16 +24,53 @@ export async function* transformSdkStream(sdkResponse, model, conversationId) {
     let contextUsagePercentage = null;
     const toolCalls = [];
     let currentToolCall = null;
+    // Probe (opus-4.8): reasoning streams via `reasoningContentEvent{text,signature}` as a
+    // contiguous run BEFORE `assistantResponseEvent.content`; no `<thinking>` tags emitted.
+    // signature is metadata and ignored for rendering.
+    let reasoningStarted = false;
+    let reasoningClosed = false;
     const eventStream = sdkResponse.generateAssistantResponseResponse;
     if (!eventStream) {
         throw new Error('SDK response has no event stream');
     }
     try {
         for await (const event of eventStream) {
+            if (event.reasoningContentEvent?.text) {
+                const reasoningText = event.reasoningContentEvent.text;
+                if (reasoningClosed) {
+                    // Defensive, normally unreached (probe: reasoning is contiguous-before-text).
+                    // The stopped thinking index cannot be reused, so open a fresh one.
+                    streamState.thinkingBlockIndex = null;
+                    reasoningClosed = false;
+                }
+                reasoningStarted = true;
+                for (const ev of createThinkingDeltaEvents(reasoningText, streamState)) {
+                    const _c = convertToOpenAI(ev, conversationId, model);
+                    if (_c !== null)
+                        yield _c;
+                }
+                continue;
+            }
             if (event.assistantResponseEvent?.content) {
                 const text = event.assistantResponseEvent.content;
                 totalContent += text;
                 textOnlyContent += text;
+                if (reasoningStarted && !reasoningClosed) {
+                    for (const ev of stopBlock(streamState.thinkingBlockIndex, streamState)) {
+                        const _c = convertToOpenAI(ev, conversationId, model);
+                        if (_c !== null)
+                            yield _c;
+                    }
+                    reasoningClosed = true;
+                }
+                if (reasoningStarted) {
+                    for (const ev of createTextDeltaEvents(text, streamState)) {
+                        const _c = convertToOpenAI(ev, conversationId, model);
+                        if (_c !== null)
+                            yield _c;
+                    }
+                    continue;
+                }
                 if (!thinkingRequested) {
                     for (const ev of createTextDeltaEvents(text, streamState)) {
                         {
@@ -151,6 +188,15 @@ export async function* transformSdkStream(sdkResponse, model, conversationId) {
             toolCalls.push(currentToolCall);
             currentToolCall = null;
         }
+        // Reasoning-only responses (reasoning but no reply text): close the thinking block.
+        if (reasoningStarted && !reasoningClosed) {
+            for (const ev of stopBlock(streamState.thinkingBlockIndex, streamState)) {
+                const _c = convertToOpenAI(ev, conversationId, model);
+                if (_c !== null)
+                    yield _c;
+            }
+            reasoningClosed = true;
+        }
         if (thinkingRequested && streamState.buffer) {
             if (streamState.inThinking) {
                 for (const ev of createThinkingDeltaEvents(streamState.buffer, streamState)) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sunerpy/opencode-kiro-auth",
-  "version": "0.4.2",
+  "version": "0.4.3",
   "description": "OpenCode plugin for AWS Kiro (CodeWhisperer) providing access to Claude models",
   "type": "module",
   "main": "dist/index.js",