clodds 1.6.1 → 1.6.3

package/dist/agents/index.js

@@ -16336,9 +16336,14 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                 'claude-3-opus-20240229': 200000,
             };
             const modelContextWindow = MODEL_CONTEXT_WINDOWS[modelId] || 200000;
+            // Estimate fixed overhead: tool definitions + system prompt (these don't change during conversation)
+            const toolsTokenEstimate = (0, context_1.estimateTokens)(JSON.stringify(tools), modelId);
+            const systemTokenEstimate = (0, context_1.estimateTokens)(finalSystemPrompt, modelId);
+            // Reserve enough for tools + system prompt + response buffer
+            const reserveForFixed = toolsTokenEstimate + systemTokenEstimate + 4096;
             const contextConfig = {
                 maxTokens: modelContextWindow,
-                reserveTokens: 4096,
+                reserveTokens: reserveForFixed,
                 compactThreshold: 0.85,
                 minMessagesAfterCompact: 6,
                 summarizer,
@@ -16349,14 +16354,14 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                 similarity: memory?.cosineSimilarity,
             };
             const contextManager = (0, context_1.createContextManager)(contextConfig, memory);
-            const effectiveMaxTokens = (contextConfig.maxTokens ?? 128000) - (contextConfig.reserveTokens ?? 4096);
+            const effectiveMaxTokens = (contextConfig.maxTokens ?? 128000) - reserveForFixed;
             const estimateSubmitTokens = () => {
                 const system = (0, context_1.estimateTokens)(finalSystemPrompt, modelId);
                 const msgs = messages.reduce((sum, m) => {
                     const content = typeof m.content === 'string' ? m.content : JSON.stringify(m.content);
                     return sum + (0, context_1.estimateTokens)(content, modelId) + 4;
                 }, 0);
-                return system + msgs;
+                return system + msgs + toolsTokenEstimate;
             };
             // Add all messages to context manager for tracking
             for (const msg of messages) {
@@ -16366,10 +16371,9 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                     content,
                 });
             }
-            // Add system prompt tokens
-            const systemTokens = (0, context_1.estimateTokens)(finalSystemPrompt, modelId);
             // Check if we need to compact before first API call
-            const guard = contextManager.checkGuard(systemTokens);
+            // (tools + system prompt are already accounted for in reserveTokens)
+            const guard = contextManager.checkGuard();
             if (guard.shouldCompact) {
                 logger_1.logger.info({ percentUsed: guard.percentUsed }, 'Context approaching limit, compacting');
                 // Trigger compaction:before hook
@@ -16411,6 +16415,11 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
             }
             const initialEstimate = estimateSubmitTokens();
             logger_1.logger.info({ tokens: initialEstimate, max: effectiveMaxTokens }, 'Token estimate before submit');
+            // Safety: if still over limit after compaction, return a friendly error
+            if (initialEstimate > effectiveMaxTokens * 1.1) {
+                logger_1.logger.warn({ tokens: initialEstimate, max: effectiveMaxTokens }, 'Context exceeds limit even after compaction');
+                return 'This conversation has gotten too long for me to process. Please start a new conversation and I\'ll be happy to help!';
+            }
             let response = await createMessage({
                 model: modelId,
                 max_tokens: 1024,
@@ -16515,7 +16524,7 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                         content,
                     });
                 }
-                const loopGuard = contextManager.checkGuard(0);
+                const loopGuard = contextManager.checkGuard();
                 if (loopGuard.shouldCompact) {
                     logger_1.logger.info({ percentUsed: loopGuard.percentUsed }, 'Compacting context during tool loop');
                     const loopCompactResult = await contextManager.compact();
@@ -16533,6 +16542,11 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                 }
                 const loopEstimate = estimateSubmitTokens();
                 logger_1.logger.info({ tokens: loopEstimate, max: effectiveMaxTokens }, 'Token estimate before submit (tool loop)');
+                // Safety: bail if over limit during tool loop
+                if (loopEstimate > effectiveMaxTokens * 1.1) {
+                    logger_1.logger.warn({ tokens: loopEstimate, max: effectiveMaxTokens }, 'Context exceeds limit during tool loop');
+                    break;
+                }
                 response = await createMessage({
                     model: modelId,
                     max_tokens: 1024,