clodds 1.6.14 → 1.6.15

package/dist/agents/index.js

@@ -16410,12 +16410,21 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                 }
                 return createMessageStreamed(params);
             };
+            // Detect tool/skill hints early — reused for skill budget + tool preloading
+            const hints = processedMessage.text ? (0, tool_registry_js_1.detectToolHints)(processedMessage.text) : { platforms: [], categories: [], hasIntent: false };
             // Build final system prompt (Clawdbot-style)
             // Priority: routed agent prompt > default system prompt
-            const skillContext = skills.getSkillContextForMessage(processedMessage.text || '');
-            const baseSystemPrompt = SYSTEM_PROMPT.replace('{{SKILLS}}', skillContext ? `\n## Skills Reference\n${skillContext}` : '');
-            let finalSystemPrompt = session.context.routedAgentPrompt || baseSystemPrompt;
-            // Add memory context if available
+            const skillContext = skills.getSkillContextForMessage(processedMessage.text || '', hints, messages.length);
+            // Split system prompt into cacheable blocks for prompt caching.
+            // Block 1 (cached): Base system prompt without skills — stable across messages.
+            // Block 2 (uncached): Skills + memory — changes per query.
+            const coreSystemPrompt = session.context.routedAgentPrompt
+                || SYSTEM_PROMPT.replace('{{SKILLS}}', '');
+            // Build dynamic context (skills + memory) — changes every query, not cached
+            let dynamicContext = '';
+            if (skillContext) {
+                dynamicContext += `\n## Skills Reference\n${skillContext}`;
+            }
             if (memory) {
                 const memoryAuto = config.memory?.auto || {};
                 const channelKey = processedMessage.chatId || processedMessage.platform;
@@ -16423,7 +16432,7 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                 if (memoryAuto.includeMemoryContext !== false) {
                     const memoryContext = memory.buildContextString(session.userId, scope);
                     if (memoryContext) {
-                        finalSystemPrompt += `\n\n## User Memory\n${memoryContext}`;
+                        dynamicContext += `\n\n## User Memory\n${memoryContext}`;
                     }
                 }
                 const semanticTopK = memoryAuto.semanticSearchTopK ?? (process.env.CLODDS_MEMORY_SEARCH === '1'
@@ -16434,7 +16443,7 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                         const results = await memory.semanticSearch(session.userId, scope, processedMessage.text, semanticTopK);
                         if (results.length > 0) {
                             const lines = results.map((r) => `- ${r.entry.key}: ${r.entry.value} (score ${r.score.toFixed(2)})`);
-                            finalSystemPrompt += `\n\n## Relevant Memory (semantic search)\n${lines.join('\n')}`;
+                            dynamicContext += `\n\n## Relevant Memory (semantic search)\n${lines.join('\n')}`;
                         }
                     }
                     catch (error) {
@@ -16442,6 +16451,8 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                     }
                 }
             }
+            // Backward-compatible string for hooks
+            let finalSystemPrompt = coreSystemPrompt + dynamicContext;
             // =========================================================================
             // HOOKS: agent:before_start - Can modify system prompt
             // =========================================================================
@@ -16461,6 +16472,22 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
             if (agentStartResult?.prependContext) {
                 finalSystemPrompt = `${agentStartResult.prependContext}\n\n${finalSystemPrompt}`;
             }
+            const hookModified = !!(agentStartResult?.systemPrompt || agentStartResult?.prependContext);
+            let systemBlocks;
+            if (hookModified) {
+                // Hooks changed the prompt — use as single block, still cache it
+                systemBlocks = [{ type: 'text', text: finalSystemPrompt, cache_control: { type: 'ephemeral' } }];
+            }
+            else {
+                systemBlocks = [
+                    // Block 1: Core system prompt (stable — cached)
+                    { type: 'text', text: coreSystemPrompt, cache_control: { type: 'ephemeral' } },
+                ];
+                // Block 2: Dynamic context (changes per query — not cached)
+                if (dynamicContext) {
+                    systemBlocks.push({ type: 'text', text: dynamicContext });
+                }
+            }
             // =========================================================================
             // CONTEXT MANAGEMENT - Check token usage and compact if needed
             // =========================================================================
@@ -16507,7 +16534,6 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
             // to avoid loading all tools from multiple platforms (~150+).
             // Also checks conversation context for platform hints in multi-turn chats.
             if (TOOL_SEARCH_ENABLED && processedMessage.text) {
-                const hints = (0, tool_registry_js_1.detectToolHints)(processedMessage.text);
                 // CONVERSATION CONTEXT: If no platform in current message, borrow from recent history.
                 // "buy YES at 40 cents" after discussing polymarket → still loads polymarket tools.
                 if (hints.platforms.length === 0 && messages.length > 1) {
@@ -16667,10 +16693,14 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
             let response;
             try {
                 const activeTools = getActiveTools();
+                // Add cache_control to last tool for tool definition caching
+                if (activeTools.length > 0) {
+                    activeTools[activeTools.length - 1].cache_control = { type: 'ephemeral' };
+                }
                 response = await createMessage({
                     model: modelId,
                     max_tokens: 1024,
-                    system: finalSystemPrompt,
+                    system: systemBlocks,
                     tools: activeTools,
                     messages,
                 });
@@ -16687,7 +16717,20 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
             // Use actual API token count for accurate context tracking
             if (response.usage) {
                 lastKnownInputTokens = response.usage.input_tokens;
-                logger_1.logger.info({ inputTokens: lastKnownInputTokens, max: modelContextWindow }, 'Actual API token usage');
+                // Track prompt cache performance
+                const usage = response.usage;
+                const cacheCreation = usage.cache_creation_input_tokens ?? 0;
+                const cacheRead = usage.cache_read_input_tokens ?? 0;
+                const cacheHitRate = cacheRead > 0
+                    ? (cacheRead / (cacheRead + lastKnownInputTokens)) * 100
+                    : 0;
+                logger_1.logger.info({
+                    inputTokens: lastKnownInputTokens,
+                    max: modelContextWindow,
+                    cacheCreation,
+                    cacheRead,
+                    cacheHitRate: `${cacheHitRate.toFixed(1)}%`,
+                }, 'API token usage (with cache stats)');
             }
             // Tool use loop — capped to prevent runaway token costs
             const MAX_TOOL_TURNS = 10;
@@ -16863,10 +16906,14 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                 }
                 try {
                     const activeTools = getActiveTools();
+                    // Add cache_control to last tool for tool definition caching
+                    if (activeTools.length > 0) {
+                        activeTools[activeTools.length - 1].cache_control = { type: 'ephemeral' };
+                    }
                     response = await createMessage({
                         model: modelId,
                         max_tokens: 1024,
-                        system: finalSystemPrompt,
+                        system: systemBlocks,
                         tools: activeTools,
                         messages,
                     });
@@ -16882,6 +16929,12 @@ async function createAgentManager(config, feeds, db, sessionManager, sendMessage
                 // Update actual token usage after each API call
                 if (response.usage) {
                     lastKnownInputTokens = response.usage.input_tokens;
+                    const loopUsage = response.usage;
+                    logger_1.logger.debug({
+                        inputTokens: lastKnownInputTokens,
+                        cacheRead: loopUsage.cache_read_input_tokens ?? 0,
+                        turn: toolTurnCount,
+                    }, 'Tool loop token usage');
                     // If actual usage is approaching limit, force compaction next iteration
                     if (lastKnownInputTokens > modelContextWindow * 0.85) {
                         logger_1.logger.info({ inputTokens: lastKnownInputTokens }, 'API reports high token usage, will compact');