clementine-agent 1.13.1 → 1.13.2

package/dist/integrations/tool-preferences.d.ts

@@ -57,13 +57,15 @@ export interface ServiceAvailability {
 export declare function computeAvailability(composioConnectedSlugs: Set<string>, claudeDesktopActiveNames: Set<string>, preferences: Record<string, ToolSource>): ServiceAvailability[];
 /**
  * Build the system-prompt instruction listing which tool source the agent
- * should use for each service. Only includes services where:
- *   - There IS a conflict (both sources connected), AND
- *   - The user has explicitly picked a non-default preference, OR
- *   - The user picked 'off' (so we tell the agent NOT to use it)
+ * should use for each service. Output ONLY for services with a real
+ * conflict (both sources connected). Includes the silent-default state
+ * too, so the agent knows the current rule without falling back to
+ * memory recall (which can hold stale preferences from past conversations).
  *
- * Returns empty string when no instruction is needed — that's the goal:
- * silent default, zero prompt overhead, until the user actually configures.
+ * Token shape:
+ *   - 0 conflicts → empty string (zero overhead)
+ *   - N conflicts → ~70 char header + ~50 chars per conflict
+ *   Compare to the previous always-on hardcoded block which was ~700 chars.
  */
 export declare function buildPromptInstruction(availability: ServiceAvailability[], preferences: Record<string, ToolSource>): string;
 //# sourceMappingURL=tool-preferences.d.ts.map

package/dist/integrations/tool-preferences.js

@@ -81,13 +81,15 @@ export function computeAvailability(composioConnectedSlugs, claudeDesktopActiveN
 }
 /**
  * Build the system-prompt instruction listing which tool source the agent
- * should use for each service. Only includes services where:
- *   - There IS a conflict (both sources connected), AND
- *   - The user has explicitly picked a non-default preference, OR
- *   - The user picked 'off' (so we tell the agent NOT to use it)
+ * should use for each service. Output ONLY for services with a real
+ * conflict (both sources connected). Includes the silent-default state
+ * too, so the agent knows the current rule without falling back to
+ * memory recall (which can hold stale preferences from past conversations).
  *
- * Returns empty string when no instruction is needed — that's the goal:
- * silent default, zero prompt overhead, until the user actually configures.
+ * Token shape:
+ *   - 0 conflicts → empty string (zero overhead)
+ *   - N conflicts → ~70 char header + ~50 chars per conflict
+ *   Compare to the previous always-on hardcoded block which was ~700 chars.
  */
 export function buildPromptInstruction(availability, preferences) {
     const lines = [];
@@ -95,20 +97,23 @@ export function buildPromptInstruction(availability, preferences) {
         if (!a.hasConflict)
             continue;
         const userPref = preferences[a.service.id];
-        if (!userPref)
-            continue; // no explicit pref → silent default, no prompt cost
-        if (userPref === 'off') {
+        // Effective rule when no explicit preference: default to Composio.
+        // We surface this in the prompt anyway so the agent has a deterministic
+        // rule and doesn't recall a contradicting preference from memory.
+        const effective = userPref ?? 'composio';
+        if (effective === 'off') {
             lines.push(`- ${a.service.label}: do NOT use any of its tools (user disabled)`);
         }
-        else if (userPref === 'composio' && a.service.composioSlug) {
-            lines.push(`- ${a.service.label}: use \`mcp__${a.service.composioSlug}__*\` (NOT \`mcp__claude_ai_${a.service.claudeDesktopName}__*\`)`);
+        else if (effective === 'composio' && a.service.composioSlug) {
+            const note = userPref ? '' : ' [default]';
+            lines.push(`- ${a.service.label}: \`mcp__${a.service.composioSlug}__*\` (NOT \`mcp__claude_ai_${a.service.claudeDesktopName}__*\`)${note}`);
         }
-        else if (userPref === 'claude-desktop' && a.service.claudeDesktopName) {
-            lines.push(`- ${a.service.label}: use \`mcp__claude_ai_${a.service.claudeDesktopName}__*\` (NOT \`mcp__${a.service.composioSlug}__*\`)`);
+        else if (effective === 'claude-desktop' && a.service.claudeDesktopName) {
+            lines.push(`- ${a.service.label}: \`mcp__claude_ai_${a.service.claudeDesktopName}__*\` (NOT \`mcp__${a.service.composioSlug}__*\`)`);
         }
     }
     if (lines.length === 0)
         return '';
-    return `## Tool Source Preferences\n\n${lines.join('\n')}`;
+    return `## Tool Source Preferences\n\nThese rules OVERRIDE any tool-source preference you may recall from memory. Do not consult memory for this — the canonical source is the dashboard's Tool Source Preferences (Settings → Integrations).\n\n${lines.join('\n')}`;
 }
 //# sourceMappingURL=tool-preferences.js.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.13.1",
+  "version": "1.13.2",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",