@corbat-tech/coco 2.15.0 → 2.16.0

package/dist/cli/index.js

@@ -4705,12 +4705,12 @@ var init_copilot2 = __esm({
     init_openai();
     init_copilot();
     CONTEXT_WINDOWS4 = {
-      // Claude models
-      "claude-sonnet-4.6": 2e5,
-      "claude-opus-4.6": 2e5,
-      "claude-sonnet-4.5": 2e5,
-      "claude-opus-4.5": 2e5,
-      "claude-haiku-4.5": 2e5,
+      // Claude models — Copilot API caps these at 168 000 (not 200 000 like Anthropic direct)
+      "claude-sonnet-4.6": 168e3,
+      "claude-opus-4.6": 168e3,
+      "claude-sonnet-4.5": 168e3,
+      "claude-opus-4.5": 168e3,
+      "claude-haiku-4.5": 168e3,
       // OpenAI models — chat/completions
       "gpt-4.1": 1048576,
       // OpenAI models — /responses API (Codex/GPT-5+)
@@ -7716,7 +7716,11 @@ var init_manager = __esm({
   "src/cli/repl/context/manager.ts"() {
     DEFAULT_CONTEXT_CONFIG = {
       maxTokens: 2e5,
-      compactionThreshold: 0.8,
+      // 75% leaves a comfortable headroom for the compaction summary call itself
+      // and the active tool-call accumulation within the current turn.
+      // Industry reference: OpenCode uses 75%, Claude Code effective ~83.5% but
+      // with an additional 33K output reserve. 75% is the recommended safe value.
+      compactionThreshold: 0.75,
       reservedTokens: 4096
     };
     ContextManager = class {
@@ -7832,35 +7836,59 @@ var init_manager = __esm({
 
 // src/cli/repl/context/compactor.ts
 function buildCompactionPrompt(focusTopic) {
-  let prompt = `Summarize the following conversation history concisely, preserving:
-1. Key decisions made
-2. Important code/file changes discussed (always include file paths)
-3. Current task context and goals
-4. Any errors or issues encountered
-5. Original user requests (verbatim if short)`;
+  let prompt = `This is a coding agent session that needs to be compacted due to context length.
+Create a structured summary that preserves everything the agent needs to continue working.
+
+## Required sections (use these exact headings):
+
+### Original Request
+State the user's original task or question verbatim (or paraphrase if very long).
+
+### Work Completed
+List every concrete action taken: files created/modified (with paths), commands run,
+bugs fixed, features implemented. Be specific \u2014 include file paths and function names.
+
+### Key Decisions
+Document architectural decisions, approaches chosen, and the reasoning behind them.
+
+### Current State
+Describe exactly where the work stands: what is done, what is in progress, what remains.
+
+### Files Touched
+List all file paths that were read, modified, or created during this session.
+
+### Errors & Resolutions
+Document any errors encountered and how they were resolved (or if still unresolved).
+
+### Next Steps
+If the task is incomplete, list the immediate next actions the agent should take.`;
   if (focusTopic) {
     prompt += `
 
-**IMPORTANT**: Preserve ALL details related to "${focusTopic}" \u2014 include specific code snippets, file paths, decisions, and context about this topic. You may be more concise about unrelated topics.`;
+**PRIORITY**: Preserve ALL details related to "${focusTopic}" \u2014 include specific code snippets, exact file paths, and full context. Be concise about unrelated topics.`;
   }
   prompt += `
 
-Keep the summary under 500 words. Format as bullet points.
+Keep the total summary under 600 words. Use bullet points within each section.
 
-CONVERSATION:
+SESSION HISTORY TO SUMMARIZE:
 `;
   return prompt;
 }
 function createContextCompactor(config) {
   return new ContextCompactor(config);
 }
-var DEFAULT_COMPACTOR_CONFIG, ContextCompactor;
+var DEFAULT_COMPACTOR_CONFIG, PRESERVED_RESULT_SOFT_CAP, PRESERVED_RESULT_SOFT_HEAD, PRESERVED_RESULT_SOFT_TAIL, HOT_TAIL_TOOL_PAIRS, ContextCompactor;
 var init_compactor = __esm({
   "src/cli/repl/context/compactor.ts"() {
     DEFAULT_COMPACTOR_CONFIG = {
-      preserveLastN: 4,
+      preserveLastN: 8,
       summaryMaxTokens: 1e3
     };
+    PRESERVED_RESULT_SOFT_CAP = 16e3;
+    PRESERVED_RESULT_SOFT_HEAD = 13e3;
+    PRESERVED_RESULT_SOFT_TAIL = 1500;
+    HOT_TAIL_TOOL_PAIRS = 4;
     ContextCompactor = class {
       config;
       constructor(config) {
@@ -7899,7 +7927,9 @@ var init_compactor = __esm({
           }
         }
         const messagesToSummarize = conversationMessages.slice(0, preserveStart);
-        const messagesToPreserve = conversationMessages.slice(preserveStart);
+        const messagesToPreserve = this.trimPreservedToolResults(
+          conversationMessages.slice(preserveStart)
+        );
         if (messagesToSummarize.length === 0) {
           return {
             messages,
@@ -7997,6 +8027,60 @@ ${summary}
           return `[Summary generation failed: ${errorMessage}. Previous conversation had ${conversationText.length} characters.]`;
         }
       }
+      /**
+       * Hot-tail policy: apply a soft cap to tool results in the preserved window.
+       *
+       * The last HOT_TAIL_TOOL_PAIRS tool-result pairs are kept verbatim (hot tail).
+       * Older pairs in the preserved window that contain results larger than
+       * PRESERVED_RESULT_SOFT_CAP are trimmed to head+tail with a marker.
+       *
+       * This handles legacy results that were stored before the inline cap was in
+       * place, ensuring that a single large stale tree/grep/web result cannot fill
+       * the context even after compaction.
+       */
+      trimPreservedToolResults(messages) {
+        let hotTailStart = messages.length;
+        let pairsFound = 0;
+        for (let i = messages.length - 1; i >= 0; i--) {
+          const msg = messages[i];
+          if (!msg) continue;
+          const isToolResultMsg = Array.isArray(msg.content) && msg.content.length > 0 && msg.content[0]?.type === "tool_result";
+          if (isToolResultMsg) {
+            pairsFound++;
+            if (pairsFound >= HOT_TAIL_TOOL_PAIRS) {
+              hotTailStart = i > 0 ? i - 1 : i;
+              break;
+            }
+          }
+        }
+        return messages.map((msg, idx) => {
+          if (idx >= hotTailStart) return msg;
+          if (!Array.isArray(msg.content)) return msg;
+          const hasOversized = msg.content.some((block) => {
+            const b = block;
+            return b.type === "tool_result" && typeof b.content === "string" && b.content.length > PRESERVED_RESULT_SOFT_CAP;
+          });
+          if (!hasOversized) return msg;
+          const blocks = msg.content;
+          const trimmedContent = blocks.map((block) => {
+            if (block.type === "tool_result" && block.content.length > PRESERVED_RESULT_SOFT_CAP) {
+              const full = block.content;
+              const head = full.slice(0, PRESERVED_RESULT_SOFT_HEAD);
+              const tail = full.slice(-PRESERVED_RESULT_SOFT_TAIL);
+              const omitted = full.length - PRESERVED_RESULT_SOFT_HEAD - PRESERVED_RESULT_SOFT_TAIL;
+              const trimmedResult = {
+                ...block,
+                content: `${head}
+[... ${omitted.toLocaleString()} chars trimmed (compaction soft-cap) ...]
+${tail}`
+              };
+              return trimmedResult;
+            }
+            return block;
+          });
+          return { ...msg, content: trimmedContent };
+        });
+      }
       /**
        * Estimate token count for messages
        */
@@ -8526,7 +8610,7 @@ async function checkAndCompactContext(session, provider, signal, toolRegistry) {
     return null;
   }
   const compactor = createContextCompactor({
-    preserveLastN: 4,
+    preserveLastN: 8,
     summaryMaxTokens: 1e3
   });
   const result = await compactor.compact(session.messages, provider, signal);
@@ -9401,7 +9485,7 @@ function humanizeError(message, toolName) {
   )) {
     return msg;
   }
-  if (/run git_init\b/.test(msg)) {
+  if (/run git_init\b/i.test(msg)) {
     return msg;
   }
   if (/ECONNREFUSED/i.test(msg)) {
@@ -39183,10 +39267,37 @@ Examples:
     }
   }
 });
+var TREE_IGNORED_DIRS = /* @__PURE__ */ new Set([
+  "node_modules",
+  "dist",
+  "build",
+  "out",
+  ".next",
+  ".nuxt",
+  ".cache",
+  ".turbo",
+  ".parcel-cache",
+  "coverage",
+  ".nyc_output",
+  "vendor",
+  "__pycache__",
+  ".venv",
+  "venv",
+  "env",
+  "target",
+  ".gradle",
+  ".mvn",
+  "bin",
+  "obj"
+]);
+var MAX_TREE_LINES = 500;
 var treeTool = defineTool({
   name: "tree",
   description: `Display directory structure as a tree.
 
+Large dependency directories (node_modules, dist, .next, etc.) are excluded
+automatically. Output is capped at ${MAX_TREE_LINES} lines to keep context lean.
+
 Examples:
 - Current dir: { }
 - Specific dir: { "path": "src" }
@@ -39206,9 +39317,12 @@ Examples:
       let totalFiles = 0;
       let totalDirs = 0;
       const lines = [path36__default.basename(absolutePath) + "/"];
+      let truncated = false;
       async function buildTree(dir, prefix, currentDepth) {
         if (currentDepth > (depth ?? 4)) return;
+        if (lines.length >= MAX_TREE_LINES) return;
         let items = await fs34__default.readdir(dir, { withFileTypes: true });
+        items = items.filter((item) => !TREE_IGNORED_DIRS.has(item.name));
         if (!showHidden) {
           items = items.filter((item) => !item.name.startsWith("."));
         }
@@ -39221,6 +39335,10 @@ Examples:
           return a.name.localeCompare(b.name);
         });
         for (let i = 0; i < items.length; i++) {
+          if (lines.length >= MAX_TREE_LINES) {
+            truncated = true;
+            return;
+          }
           const item = items[i];
           const isLast = i === items.length - 1;
           const connector = isLast ? "\u2514\u2500\u2500 " : "\u251C\u2500\u2500 ";
@@ -39236,10 +39354,17 @@ Examples:
         }
       }
       await buildTree(absolutePath, "", 1);
+      if (truncated) {
+        lines.push(
+          `
+[... output truncated at ${MAX_TREE_LINES} lines. Use a deeper path or lower depth to see more.]`
+        );
+      }
       return {
         tree: lines.join("\n"),
         totalFiles,
-        totalDirs
+        totalDirs,
+        truncated
       };
     } catch (error) {
       if (isENOENT(error)) {
@@ -41790,7 +41915,7 @@ init_registry4();
 init_errors();
 init_version();
 var DEFAULT_TIMEOUT_MS5 = 3e4;
-var DEFAULT_MAX_LENGTH = 5e4;
+var DEFAULT_MAX_LENGTH = 8e3;
 var MAX_DOWNLOAD_SIZE = 10 * 1024 * 1024;
 var BLOCKED_SCHEMES = ["file:", "ftp:", "data:", "javascript:"];
 var PRIVATE_IP_PATTERNS = [
@@ -48443,6 +48568,18 @@ async function executeAgentTurn(session, userMessage, provider, toolRegistry, op
   const maxIterations = session.config.agent.maxToolIterations;
   const toolErrorCounts = /* @__PURE__ */ new Map();
   const MAX_CONSECUTIVE_TOOL_ERRORS = 3;
+  const INLINE_RESULT_MAX_CHARS = 8e3;
+  const INLINE_RESULT_HEAD_CHARS = 6500;
+  const INLINE_RESULT_TAIL_CHARS = 1e3;
+  function truncateInlineResult(content, toolName) {
+    if (content.length <= INLINE_RESULT_MAX_CHARS) return content;
+    const head = content.slice(0, INLINE_RESULT_HEAD_CHARS);
+    const tail = content.slice(-INLINE_RESULT_TAIL_CHARS);
+    const omitted = content.length - INLINE_RESULT_HEAD_CHARS - INLINE_RESULT_TAIL_CHARS;
+    return `${head}
+[... ${omitted.toLocaleString()} characters omitted \u2014 use read_file with offset/limit to retrieve more of '${toolName}' output ...]
+${tail}`;
+  }
   while (iteration < maxIterations) {
     iteration++;
     if (options.signal?.aborted) {
@@ -48695,7 +48832,7 @@ async function executeAgentTurn(session, userMessage, provider, toolRegistry, op
         toolResults.push({
           type: "tool_result",
           tool_use_id: toolCall.id,
-          content: executedCall.result.output,
+          content: truncateInlineResult(executedCall.result.output, toolCall.name),
           is_error: !executedCall.result.success
         });
       } else {
@@ -50073,6 +50210,25 @@ async function startRepl(options = {}) {
         continue;
       }
       const errorMsg = error instanceof Error ? error.message : String(error);
+      if (errorMsg.includes("prompt token count") && errorMsg.includes("exceeds the limit")) {
+        renderError("Context window full \u2014 compacting conversation history...");
+        try {
+          const compactionResult = await checkAndCompactContext(
+            session,
+            provider,
+            void 0,
+            toolRegistry
+          );
+          if (compactionResult?.wasCompacted) {
+            console.log(chalk2.green("   \u2713 Context compacted. Please retry your message."));
+          } else {
+            console.log(chalk2.yellow("   \u26A0 Could not compact context. Use /clear to start fresh."));
+          }
+        } catch {
+          console.log(chalk2.yellow("   \u26A0 Context compaction failed. Use /clear to start fresh."));
+        }
+        continue;
+      }
       if (errorMsg.includes("context length") || errorMsg.includes("tokens to keep")) {
         renderError(errorMsg);
         console.log();