@corbat-tech/coco 2.8.0 → 2.8.2

package/dist/cli/index.js

@@ -1012,22 +1012,44 @@ var init_anthropic = __esm({
       async *stream(messages, options) {
         this.ensureInitialized();
         try {
-          const stream = await this.client.messages.stream({
-            model: options?.model ?? this.config.model ?? DEFAULT_MODEL,
-            max_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
-            temperature: options?.temperature ?? this.config.temperature ?? 0,
-            system: this.extractSystem(messages, options?.system),
-            messages: this.convertMessages(messages)
-          });
-          for await (const event of stream) {
-            if (event.type === "content_block_delta") {
-              const delta = event.delta;
-              if (delta.type === "text_delta" && delta.text) {
-                yield { type: "text", text: delta.text };
+          const stream = await this.client.messages.stream(
+            {
+              model: options?.model ?? this.config.model ?? DEFAULT_MODEL,
+              max_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
+              temperature: options?.temperature ?? this.config.temperature ?? 0,
+              system: this.extractSystem(messages, options?.system),
+              messages: this.convertMessages(messages)
+            },
+            { signal: options?.signal }
+          );
+          const streamTimeout = this.config.timeout ?? 12e4;
+          let lastActivityTime = Date.now();
+          const checkTimeout = () => {
+            if (Date.now() - lastActivityTime > streamTimeout) {
+              throw new Error(`Stream timeout: No response from LLM for ${streamTimeout / 1e3}s`);
+            }
+          };
+          const timeoutInterval = setInterval(checkTimeout, 5e3);
+          try {
+            let streamStopReason;
+            for await (const event of stream) {
+              lastActivityTime = Date.now();
+              if (event.type === "content_block_delta") {
+                const delta = event.delta;
+                if (delta.type === "text_delta" && delta.text) {
+                  yield { type: "text", text: delta.text };
+                }
+              } else if (event.type === "message_delta") {
+                const delta = event.delta;
+                if (delta.stop_reason) {
+                  streamStopReason = this.mapStopReason(delta.stop_reason);
+                }
               }
             }
+            yield { type: "done", stopReason: streamStopReason };
+          } finally {
+            clearInterval(timeoutInterval);
           }
-          yield { type: "done" };
         } catch (error) {
           throw this.handleError(error);
         }
@@ -1038,90 +1060,112 @@ var init_anthropic = __esm({
       async *streamWithTools(messages, options) {
         this.ensureInitialized();
         try {
-          const stream = await this.client.messages.stream({
-            model: options?.model ?? this.config.model ?? DEFAULT_MODEL,
-            max_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
-            temperature: options?.temperature ?? this.config.temperature ?? 0,
-            system: this.extractSystem(messages, options?.system),
-            messages: this.convertMessages(messages),
-            tools: this.convertTools(options.tools),
-            tool_choice: options.toolChoice ? this.convertToolChoice(options.toolChoice) : void 0
-          });
+          const stream = await this.client.messages.stream(
+            {
+              model: options?.model ?? this.config.model ?? DEFAULT_MODEL,
+              max_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
+              temperature: options?.temperature ?? this.config.temperature ?? 0,
+              system: this.extractSystem(messages, options?.system),
+              messages: this.convertMessages(messages),
+              tools: this.convertTools(options.tools),
+              tool_choice: options.toolChoice ? this.convertToolChoice(options.toolChoice) : void 0
+            },
+            { signal: options?.signal }
+          );
           let currentToolCall = null;
           let currentToolInputJson = "";
-          for await (const event of stream) {
-            if (event.type === "content_block_start") {
-              const contentBlock = event.content_block;
-              if (contentBlock.type === "tool_use") {
+          const streamTimeout = this.config.timeout ?? 12e4;
+          let lastActivityTime = Date.now();
+          const checkTimeout = () => {
+            if (Date.now() - lastActivityTime > streamTimeout) {
+              throw new Error(`Stream timeout: No response from LLM for ${streamTimeout / 1e3}s`);
+            }
+          };
+          const timeoutInterval = setInterval(checkTimeout, 5e3);
+          try {
+            let streamStopReason;
+            for await (const event of stream) {
+              lastActivityTime = Date.now();
+              if (event.type === "message_delta") {
+                const delta = event.delta;
+                if (delta.stop_reason) {
+                  streamStopReason = this.mapStopReason(delta.stop_reason);
+                }
+              } else if (event.type === "content_block_start") {
+                const contentBlock = event.content_block;
+                if (contentBlock.type === "tool_use") {
+                  if (currentToolCall) {
+                    getLogger().warn(
+                      `[Anthropic] content_block_stop missing for tool '${currentToolCall.name}' \u2014 finalizing early to prevent data bleed.`
+                    );
+                    try {
+                      currentToolCall.input = currentToolInputJson ? JSON.parse(currentToolInputJson) : {};
+                    } catch {
+                      currentToolCall.input = {};
+                    }
+                    yield {
+                      type: "tool_use_end",
+                      toolCall: { ...currentToolCall }
+                    };
+                  }
+                  currentToolCall = {
+                    id: contentBlock.id,
+                    name: contentBlock.name
+                  };
+                  currentToolInputJson = "";
+                  yield {
+                    type: "tool_use_start",
+                    toolCall: { ...currentToolCall }
+                  };
+                }
+              } else if (event.type === "content_block_delta") {
+                const delta = event.delta;
+                if (delta.type === "text_delta" && delta.text) {
+                  yield { type: "text", text: delta.text };
+                } else if (delta.type === "input_json_delta" && delta.partial_json) {
+                  currentToolInputJson += delta.partial_json;
+                  yield {
+                    type: "tool_use_delta",
+                    toolCall: {
+                      ...currentToolCall
+                    },
+                    text: delta.partial_json
+                  };
+                }
+              } else if (event.type === "content_block_stop") {
                 if (currentToolCall) {
-                  getLogger().warn(
-                    `[Anthropic] content_block_stop missing for tool '${currentToolCall.name}' \u2014 finalizing early to prevent data bleed.`
-                  );
                   try {
                     currentToolCall.input = currentToolInputJson ? JSON.parse(currentToolInputJson) : {};
                   } catch {
-                    currentToolCall.input = {};
+                    let repaired = false;
+                    if (currentToolInputJson) {
+                      try {
+                        currentToolCall.input = JSON.parse(jsonrepair(currentToolInputJson));
+                        repaired = true;
+                        getLogger().debug(`Repaired JSON for tool ${currentToolCall.name}`);
+                      } catch {
+                      }
+                    }
+                    if (!repaired) {
+                      getLogger().warn(
+                        `Failed to parse tool call arguments for ${currentToolCall.name}: ${currentToolInputJson?.slice(0, 300)}`
+                      );
+                      currentToolCall.input = {};
+                    }
                   }
                   yield {
                     type: "tool_use_end",
                     toolCall: { ...currentToolCall }
                   };
+                  currentToolCall = null;
+                  currentToolInputJson = "";
                 }
-                currentToolCall = {
-                  id: contentBlock.id,
-                  name: contentBlock.name
-                };
-                currentToolInputJson = "";
-                yield {
-                  type: "tool_use_start",
-                  toolCall: { ...currentToolCall }
-                };
-              }
-            } else if (event.type === "content_block_delta") {
-              const delta = event.delta;
-              if (delta.type === "text_delta" && delta.text) {
-                yield { type: "text", text: delta.text };
-              } else if (delta.type === "input_json_delta" && delta.partial_json) {
-                currentToolInputJson += delta.partial_json;
-                yield {
-                  type: "tool_use_delta",
-                  toolCall: {
-                    ...currentToolCall
-                  },
-                  text: delta.partial_json
-                };
-              }
-            } else if (event.type === "content_block_stop") {
-              if (currentToolCall) {
-                try {
-                  currentToolCall.input = currentToolInputJson ? JSON.parse(currentToolInputJson) : {};
-                } catch {
-                  let repaired = false;
-                  if (currentToolInputJson) {
-                    try {
-                      currentToolCall.input = JSON.parse(jsonrepair(currentToolInputJson));
-                      repaired = true;
-                      getLogger().debug(`Repaired JSON for tool ${currentToolCall.name}`);
-                    } catch {
-                    }
-                  }
-                  if (!repaired) {
-                    getLogger().warn(
-                      `Failed to parse tool call arguments for ${currentToolCall.name}: ${currentToolInputJson?.slice(0, 300)}`
-                    );
-                    currentToolCall.input = {};
-                  }
-                }
-                yield {
-                  type: "tool_use_end",
-                  toolCall: { ...currentToolCall }
-                };
-                currentToolCall = null;
-                currentToolInputJson = "";
               }
             }
+            yield { type: "done", stopReason: streamStopReason };
+          } finally {
+            clearInterval(timeoutInterval);
           }
-          yield { type: "done" };
         } catch (error) {
           throw this.handleError(error);
         }
@@ -1621,13 +1665,18 @@ var init_openai = __esm({
             stream: true,
             ...supportsTemp && { temperature: options?.temperature ?? this.config.temperature ?? 0 }
           });
+          let streamStopReason;
           for await (const chunk of stream) {
             const delta = chunk.choices[0]?.delta;
             if (delta?.content) {
               yield { type: "text", text: delta.content };
             }
+            const finishReason = chunk.choices[0]?.finish_reason;
+            if (finishReason) {
+              streamStopReason = this.mapFinishReason(finishReason);
+            }
           }
-          yield { type: "done" };
+          yield { type: "done", stopReason: streamStopReason };
         } catch (error) {
           throw this.handleError(error);
         }
@@ -1692,6 +1741,7 @@ var init_openai = __esm({
             return input;
           };
           try {
+            let streamStopReason;
             for await (const chunk of stream) {
               const delta = chunk.choices[0]?.delta;
               if (delta?.content || delta?.tool_calls) {
@@ -1738,6 +1788,9 @@ var init_openai = __esm({
                 }
               }
               const finishReason = chunk.choices[0]?.finish_reason;
+              if (finishReason) {
+                streamStopReason = this.mapFinishReason(finishReason);
+              }
               if (finishReason && toolCallBuilders.size > 0) {
                 for (const [, builder] of toolCallBuilders) {
                   yield {
@@ -1762,7 +1815,7 @@ var init_openai = __esm({
                 }
               };
             }
-            yield { type: "done" };
+            yield { type: "done", stopReason: streamStopReason };
           } finally {
             clearInterval(timeoutInterval);
           }
@@ -3569,7 +3622,7 @@ var init_codex = __esm({
             }
           }
         }
-        yield { type: "done" };
+        yield { type: "done", stopReason: response.stopReason };
       }
       /**
        * Stream a chat response with tool use
@@ -3737,13 +3790,18 @@ var init_gemini = __esm({
           const { history, lastMessage } = this.convertMessages(messages);
           const chat = model.startChat({ history });
           const result = await chat.sendMessageStream(lastMessage);
+          let streamStopReason;
           for await (const chunk of result.stream) {
             const text13 = chunk.text();
             if (text13) {
               yield { type: "text", text: text13 };
             }
+            const finishReason = chunk.candidates?.[0]?.finishReason;
+            if (finishReason) {
+              streamStopReason = this.mapFinishReason(finishReason);
+            }
           }
-          yield { type: "done" };
+          yield { type: "done", stopReason: streamStopReason };
         } catch (error) {
           throw this.handleError(error);
         }
@@ -3778,11 +3836,16 @@ var init_gemini = __esm({
           const chat = model.startChat({ history });
           const result = await chat.sendMessageStream(lastMessage);
           const emittedToolCalls = /* @__PURE__ */ new Set();
+          let streamStopReason;
           for await (const chunk of result.stream) {
             const text13 = chunk.text();
             if (text13) {
               yield { type: "text", text: text13 };
             }
+            const finishReason = chunk.candidates?.[0]?.finishReason;
+            if (finishReason) {
+              streamStopReason = this.mapFinishReason(finishReason);
+            }
             const candidate = chunk.candidates?.[0];
             if (candidate?.content?.parts) {
               for (const part of candidate.content.parts) {
@@ -3816,7 +3879,7 @@ var init_gemini = __esm({
               }
             }
           }
-          yield { type: "done" };
+          yield { type: "done", stopReason: streamStopReason };
         } catch (error) {
           throw this.handleError(error);
         }
@@ -6564,7 +6627,7 @@ CONVERSATION:
        * @param provider - The LLM provider to use for summarization
        * @returns Compacted messages with summary replacing older messages
        */
-      async compact(messages, provider) {
+      async compact(messages, provider, signal) {
         const conversationMessages = messages.filter((m) => m.role !== "system");
         if (conversationMessages.length <= this.config.preserveLastN) {
           return {
@@ -6596,7 +6659,7 @@ CONVERSATION:
         }
         const originalTokens = this.estimateTokens(messages, provider);
         const conversationText = this.formatMessagesForSummary(messagesToSummarize);
-        const summary = await this.generateSummary(conversationText, provider);
+        const summary = await this.generateSummary(conversationText, provider, signal);
         const systemMessages = messages.filter((m) => m.role === "system");
         const summaryMessage = {
           role: "user",
@@ -6650,16 +6713,30 @@ ${summary}
       /**
        * Generate a summary of the conversation using the LLM
        */
-      async generateSummary(conversationText, provider) {
+      async generateSummary(conversationText, provider, signal) {
+        if (signal?.aborted) return "[Compaction cancelled]";
         const prompt = COMPACTION_PROMPT + conversationText;
         try {
-          const response = await provider.chat([{ role: "user", content: prompt }], {
+          const chatPromise = provider.chat([{ role: "user", content: prompt }], {
             maxTokens: this.config.summaryMaxTokens,
             temperature: 0.3
             // Lower temperature for more consistent summaries
           });
+          if (signal) {
+            const abortPromise = new Promise((_, reject) => {
+              signal.addEventListener(
+                "abort",
+                () => reject(new DOMException("Aborted", "AbortError")),
+                { once: true }
+              );
+            });
+            const response2 = await Promise.race([chatPromise, abortPromise]);
+            return response2.content;
+          }
+          const response = await chatPromise;
           return response.content;
         } catch (error) {
+          if (error instanceof DOMException && error.name === "AbortError") throw error;
           const errorMessage = error instanceof Error ? error.message : String(error);
           return `[Summary generation failed: ${errorMessage}. Previous conversation had ${conversationText.length} characters.]`;
         }
@@ -6795,7 +6872,14 @@ function addMessage(session, message) {
   session.messages.push(message);
   const maxMessages = session.config.ui.maxHistorySize * 2;
   if (session.messages.length > maxMessages) {
-    session.messages = session.messages.slice(-session.config.ui.maxHistorySize);
+    let sliceStart = session.messages.length - session.config.ui.maxHistorySize;
+    while (sliceStart > 0 && sliceStart < session.messages.length) {
+      const msg = session.messages[sliceStart];
+      const isToolResult = Array.isArray(msg?.content) && msg.content.length > 0 && msg.content[0]?.type === "tool_result";
+      if (!isToolResult) break;
+      sliceStart--;
+    }
+    session.messages = session.messages.slice(sliceStart);
   }
 }
 function substituteDynamicContext(body, cwd) {
@@ -7062,7 +7146,7 @@ function updateContextTokens(session, provider) {
   }
   session.contextManager.setUsedTokens(totalTokens);
 }
-async function checkAndCompactContext(session, provider) {
+async function checkAndCompactContext(session, provider, signal) {
   if (!session.contextManager) {
     initializeContextManager(session, provider);
   }
@@ -7074,7 +7158,7 @@ async function checkAndCompactContext(session, provider) {
     preserveLastN: 4,
     summaryMaxTokens: 1e3
   });
-  const result = await compactor.compact(session.messages, provider);
+  const result = await compactor.compact(session.messages, provider, signal);
   if (result.wasCompacted) {
     const compactedNonSystem = result.messages.filter((m) => m.role !== "system");
     session.messages = compactedNonSystem;
@@ -44587,10 +44671,12 @@ async function executeAgentTurn(session, userMessage, provider, toolRegistry, op
     let responseContent = "";
     const collectedToolCalls = [];
     let thinkingEnded = false;
+    let lastStopReason;
     const toolCallBuilders = /* @__PURE__ */ new Map();
     for await (const chunk of provider.streamWithTools(messages, {
       tools,
-      maxTokens: session.config.provider.maxTokens
+      maxTokens: session.config.provider.maxTokens,
+      signal: options.signal
     })) {
       if (options.signal?.aborted) {
         break;
@@ -44640,6 +44726,9 @@ async function executeAgentTurn(session, userMessage, provider, toolRegistry, op
         }
       }
       if (chunk.type === "done") {
+        if (chunk.stopReason) {
+          lastStopReason = chunk.stopReason;
+        }
         if (!thinkingEnded) {
           options.onThinkingEnd?.();
           thinkingEnded = true;
@@ -44658,6 +44747,14 @@ async function executeAgentTurn(session, userMessage, provider, toolRegistry, op
       if (options.signal?.aborted) {
         return abortReturn();
       }
+      if (lastStopReason === "max_tokens" && responseContent) {
+        addMessage(session, { role: "assistant", content: responseContent });
+        addMessage(session, {
+          role: "user",
+          content: "[System: Your previous response was cut off due to the output token limit. Continue exactly where you left off.]"
+        });
+        continue;
+      }
       addMessage(session, { role: "assistant", content: responseContent });
       break;
     }
@@ -44864,7 +44961,8 @@ async function executeAgentTurn(session, userMessage, provider, toolRegistry, op
         const finalMessages = getConversationContext(session, toolRegistry);
         for await (const chunk of provider.streamWithTools(finalMessages, {
           tools: [],
-          maxTokens: session.config.provider.maxTokens
+          maxTokens: session.config.provider.maxTokens,
+          signal: options.signal
         })) {
           if (options.signal?.aborted) break;
           if (chunk.type === "text" && chunk.text) {
@@ -44878,6 +44976,14 @@ async function executeAgentTurn(session, userMessage, provider, toolRegistry, op
       break;
     }
   }
+  if (iteration >= maxIterations) {
+    const notice = `
+
+---
+_Reached the iteration limit (${maxIterations}). The task may be incomplete. You can say "continue" to resume._`;
+    finalContent += notice;
+    options.onStream?.({ type: "text", text: notice });
+  }
   options.onStream?.({ type: "done" });
   return {
     content: finalContent,
@@ -46059,16 +46165,33 @@ async function startRepl(options = {}) {
       const usageBefore = getContextUsagePercent(session);
       let usageForDisplay = usageBefore;
       try {
-        const compactionResult = await checkAndCompactContext(session, provider);
-        if (compactionResult?.wasCompacted) {
-          usageForDisplay = getContextUsagePercent(session);
-          console.log(
-            chalk25.dim(
-              `Context compacted (${usageBefore.toFixed(0)}% -> ${usageForDisplay.toFixed(0)}%)`
-            )
+        const compactAbort = new AbortController();
+        const compactTimeout = setTimeout(() => compactAbort.abort(), 3e4);
+        const compactSigint = () => compactAbort.abort();
+        process.once("SIGINT", compactSigint);
+        const compactSpinner = createSpinner("Compacting context");
+        compactSpinner.start();
+        try {
+          const compactionResult = await checkAndCompactContext(
+            session,
+            provider,
+            compactAbort.signal
           );
-          warned75 = false;
-          warned90 = false;
+          if (compactionResult?.wasCompacted) {
+            usageForDisplay = getContextUsagePercent(session);
+            compactSpinner.stop(
+              `Context compacted (${usageBefore.toFixed(0)}% \u2192 ${usageForDisplay.toFixed(0)}%)`
+            );
+            warned75 = false;
+            warned90 = false;
+          } else {
+            compactSpinner.clear();
+          }
+        } catch {
+          compactSpinner.clear();
+        } finally {
+          clearTimeout(compactTimeout);
+          process.off("SIGINT", compactSigint);
         }
       } catch {
       }