clementine-agent 1.18.37 → 1.18.38

package/dist/agent/assistant.js

@@ -417,7 +417,17 @@ const SESSIONS_FILE = path.join(BASE_DIR, '.sessions.json');
 const MAX_SESSION_EXCHANGES = 40;
 const SESSION_EXPIRY_MS = 24 * 60 * 60 * 1000;
 const AUTO_MEMORY_MIN_LENGTH = 80;
-const AUTO_MEMORY_MODEL = MODELS.sonnet;
+// Model used by the post-exchange memory extractor + the conversation
+// summarizer. Both are routine "read this exchange, extract facts, call
+// memory_write with structured JSON" tasks — Haiku handles them fine and
+// they fire on EVERY substantive exchange, so the multiplier matters.
+// Override with CLEMENTINE_AUTO_MEMORY_MODEL=sonnet if you observe
+// extraction quality drop.
+const AUTO_MEMORY_MODEL = process.env.CLEMENTINE_AUTO_MEMORY_MODEL?.includes('sonnet')
+    ? MODELS.sonnet
+    : process.env.CLEMENTINE_AUTO_MEMORY_MODEL?.includes('opus')
+        ? MODELS.opus
+        : MODELS.haiku;
 const OWNER = OWNER_NAME || 'the user';
 const MCP_SERVER_SCRIPT = path.join(PKG_DIR, 'dist', 'tools', 'mcp-server.js');
 const TOOLS_SERVER = `${ASSISTANT_NAME.toLowerCase()}-tools`;
@@ -4623,6 +4633,12 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                     const blocks = getContentBlocks(message);
                     summaryText += extractText(blocks);
                 }
+                else if (message.type === 'result') {
+                    // Make session-summarization cost visible in usage_log. Without
+                    // this, every session rotation spawned a Sonnet summarize call
+                    // that didn't appear in any metric.
+                    this.logQueryResult(message, 'summarize', `summarize:${sessionKey}`);
+                }
             }
             if (summaryText.trim()) {
                 if (this.memoryStore) {
@@ -4993,6 +5009,13 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             });
             const collectedText = [];
             for await (const message of stream) {
+                if (message.type === 'result') {
+                    // Auto-memory extraction fires after every substantive
+                    // exchange. Before this log call, its cost was invisible in
+                    // usage_log — a per-user-message Sonnet pass running silently.
+                    this.logQueryResult(message, 'auto_memory', `auto-memory:${sessionKey ?? 'unknown'}`, undefined, profile?.slug);
+                    continue;
+                }
                 if (message.type === 'assistant') {
                     const blocks = getContentBlocks(message);
                     for (const block of blocks) {
@@ -5651,6 +5674,11 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
                     const blocks = getContentBlocks(message);
                     responseText += extractText(blocks);
                 }
+                else if (message.type === 'result') {
+                    // Cron reflection (post-task quality check) fires after every
+                    // cron run. Cheap (Haiku, 1 turn, ~1KB) but should be visible.
+                    this.logQueryResult(message, 'cron_reflection', `reflection:${jobName}`, jobName);
+                }
             }
             if (responseText.trim()) {
                 const reflection = JSON.parse(responseText.trim());

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.18.37",
+  "version": "1.18.38",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",