@miphamai/cli 0.7.3 → 0.7.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miphamai/cli",
-  "version": "0.7.3",
+  "version": "0.7.5",
   "description": "Mipham Code — Multi-model open-core intelligent coding terminal by MiphamAI",
   "keywords": [
     "ai",

package/src/core/engine.ts

@@ -300,7 +300,7 @@ export class QueryEngine {
   }
 
   private async *continueWithTools(signal?: AbortSignal): AsyncGenerator<StreamChunk> {
-    const MAX_TURNS = 10
+    const MAX_TURNS = 20
     const toolDefs = this.getToolDefinitions()
 
     for (let turn = 0; turn < MAX_TURNS; turn++) {
@@ -381,11 +381,34 @@ export class QueryEngine {
         this.context.addMessage(msg)
       }
 
-      // Safety: prevent silent truncation when max turns reached
+      // Safety: when max turns reached with pending tools, ask model to summarize
       if (turn === MAX_TURNS - 1 && toolUses.length > 0) {
-        yield {
-          type: 'error',
-          error: `Max tool-calling turns (${MAX_TURNS}) reached. Some tool calls were not executed.`,
+        this.context.addMessage({
+          role: 'user',
+          content:
+            `You've reached the maximum of ${MAX_TURNS} tool-calling rounds. ` +
+            `${toolUses.length} tool call(s) were not executed. ` +
+            'Please summarize what you found so far and any next steps the user should take.',
+        })
+        // Give model one final chance to respond with a summary
+        try {
+          const finalSystemPrompt = this.context.getSystemPrompt()
+          const finalMessages = this.context.getMessages()
+          for await (const chunk of this.registry.chat({
+            model: this.registry.getActiveModel(),
+            messages: finalMessages,
+            systemPrompt: finalSystemPrompt,
+            tools: undefined, // no tools — force text-only summary
+            signal,
+          })) {
+            yield chunk
+            if (chunk.type === 'error') return
+          }
+        } catch {
+          yield {
+            type: 'error',
+            error: `Max tool-calling turns (${MAX_TURNS}) reached. Some tool calls were not executed.`,
+          }
         }
         return
       }

package/src/providers/anthropic.ts

@@ -101,8 +101,26 @@ export class AnthropicProvider implements ProviderInstance {
     const decoder = new TextDecoder()
     let buffer = ''
 
+    // Streaming read timeout: if no data arrives for 90s, abort to prevent UI freeze.
+    const STREAM_READ_TIMEOUT_MS = 90_000
+
     while (true) {
-      const { done, value } = await reader.read()
+      let readResult: Awaited<ReturnType<typeof reader.read>>
+      try {
+        readResult = await Promise.race([
+          reader.read(),
+          new Promise<never>((_, reject) =>
+            setTimeout(
+              () => reject(new Error('Stream read timeout — no data for 90s')),
+              STREAM_READ_TIMEOUT_MS,
+            ),
+          ),
+        ])
+      } catch (err) {
+        yield { type: 'error', error: `Stream stalled: ${String(err)}` }
+        return
+      }
+      const { done, value } = readResult
       if (done) break
 
       buffer += decoder.decode(value, { stream: true })

package/src/providers/openai-compat.ts

@@ -48,8 +48,27 @@ export class OpenAICompatProvider implements ProviderInstance {
     const pendingToolCalls = new Map<number, { id: string; name: string; arguments: string }>()
     let reasoningContent = ''
 
+    // Streaming read timeout: if no data arrives for 90s, abort to prevent UI freeze.
+    // DeepSeek V4 thinking mode can take 30-60s between chunks — 90s is a generous ceiling.
+    const STREAM_READ_TIMEOUT_MS = 90_000
+
     while (true) {
-      const { done, value } = await reader.read()
+      let readResult: Awaited<ReturnType<typeof reader.read>>
+      try {
+        readResult = await Promise.race([
+          reader.read(),
+          new Promise<never>((_, reject) =>
+            setTimeout(
+              () => reject(new Error('Stream read timeout — no data for 90s')),
+              STREAM_READ_TIMEOUT_MS,
+            ),
+          ),
+        ])
+      } catch (err) {
+        yield { type: 'error', error: `Stream stalled: ${String(err)}` }
+        return
+      }
+      const { done, value } = readResult
       if (done) break
 
       buffer += decoder.decode(value, { stream: true })