@pentoshi/clai 3.8.41 → 3.8.42

package/dist/agent/runner.js

@@ -2358,20 +2358,20 @@ export async function runAgentTurn(prompt, options = {}) {
                                     }
                                 }
                             }
-                            if (!sawReasoning && /<think/i.test(token)) {
+                            if (!sawReasoning && /<think(?:ing)?\b/i.test(token)) {
                                 sawReasoning = true;
                                 inThinking = true;
                                 spinner.setLabel("thinking");
                                 if (!writesDirectly)
                                     emit({ type: "status", text: "thinking" });
                             }
-                            if (/<\/think>/i.test(token)) {
+                            if (/<\/think(?:ing)?>/i.test(token)) {
                                 inThinking = false;
                                 spinner.setLabel("generating response (0 tokens)");
                                 generatedTokens = 0;
                             }
                             if (inThinking) {
-                                const cleaned = token.replace(/<\/?think[^>]*>/gi, "");
+                                const cleaned = token.replace(/<\/?think(?:ing)?[^>]*>/gi, "");
                                 if (cleaned) {
                                     spinner.pushPreview(cleaned);
                                     const approx = cleaned.split(/\s+/).filter(Boolean).length;
@@ -2433,7 +2433,16 @@ export async function runAgentTurn(prompt, options = {}) {
                     (toolsAttached && !isTextOnlyModel(provider, model));
                 const assistantTextResult = rememberThinkingFromText(completion.text);
                 assistantText = assistantTextResult;
-                if (assistantText.hasThinking) {
+                // Only emit a thinking-block event when the classic renderer is
+                // active (writesDirectly / no deltaParser).  In TUI v2 the
+                // deltaParser already streamed thinking-delta events that created
+                // the thinking item in the correct transcript position; emitting
+                // a redundant thinking-block after tool-call events have cleared
+                // pendingThinkingId causes the reducer to append a *duplicate*
+                // thinking item at the end of the transcript — the root cause of
+                // the "ghost thinking blocks below the response" bug seen with
+                // models that use reasoning_content (e.g. Kimi K2-thinking).
+                if (assistantText.hasThinking && !deltaParser) {
                     writeThinkingBlock(assistantText.thinkContent);
                 }
                 // Native-first: prefer structured toolCalls from the provider.