Cache Keepalive: send ping as a user message (depth-0 injection) and update default message

6803ba16e2fd0e1a5c64a2d836107d430410fcda

permissionBRICK <permissionBRICK@gmail.com>

public/scripts/extensions/keepalive/index.js+24 -2 Viewed

     deactivateSendButtons,
     eventSource,
     event_types,
 +    extension_prompt_roles,
 +    extension_prompt_types,
     generateQuietPrompt,
     isGenerating,
     saveSettingsDebounced,
 +    setExtensionPrompt,
     setSendButtonState,
     streamingProcessor,
 } from '../../../script.js';

 // cache), but we never pay for a full completion.
 const KEEPALIVE_RESPONSE_LENGTH = 1;
 +// Extension-prompt key used to inject the keepalive message as a USER message at depth 0.
 +const KEEPALIVE_INJECT_ID = 'keepalive_ping';
++
 +// Original default message; used to migrate untouched installs to the current default wording.
 +const LEGACY_DEFAULT_MESSAGE = 'Keep the cache warm.';
++
 // Idle-tracking key used when no connection profile is active (manual API panel).
 const NONE_KEY = '__none__';

 const defaultSettings = {
     enabled: false,
     timeoutSeconds: 295,
 -    message: 'Keep the cache warm.',
 +    message: '[Note: just reply with an empty response to keep the session alive]',
     // Per-profile enable map: { [profileId]: boolean }. Unlisted profiles default to enabled.
     profiles: {},
 };

     deactivateSendButtons();
     try {
 -        await generateQuietPrompt({ quietPrompt: extension_settings[MODULE].message, responseLength: KEEPALIVE_RESPONSE_LENGTH });
 +        // Deliver the keepalive as a USER message at depth 0 (the quiet-prompt path forces a
 +        // system role — openai.js hardcodes role:'system' for it), using the same depth+role
 +        // injection Author's Note uses. The quiet prompt itself stays empty so the full chat
 +        // context is still assembled (which is what keeps the cached prefix warm).
 +        setExtensionPrompt(KEEPALIVE_INJECT_ID, extension_settings[MODULE].message, extension_prompt_types.IN_CHAT, 0, false, extension_prompt_roles.USER);
 +        await generateQuietPrompt({ quietPrompt: '', responseLength: KEEPALIVE_RESPONSE_LENGTH });
         // The ping kept this profile warm; treat it as fresh activity so the next
         // keepalive for it is a full timeout away.
         lastActivity[getActiveProfileKey()] = Date.now();

         // Best-effort: any failure here is harmless and intentionally swallowed.
         console.debug('[Keepalive] Background generation ended:', error);
     } finally {
 +        setExtensionPrompt(KEEPALIVE_INJECT_ID, '', extension_prompt_types.IN_CHAT, 0, false, extension_prompt_roles.USER);
         setSendButtonState(false);
         activateSendButtons();
         keepaliveActive = false;

         extension_settings[MODULE].profiles = {};
+    }
 +    // Migrate the original default message to the current default so installs that never
 +    // customized it pick up the improved wording (custom messages are left untouched).
 +    if (extension_settings[MODULE].message === LEGACY_DEFAULT_MESSAGE) {
 +        extension_settings[MODULE].message = defaultSettings.message;
 +        saveSettingsDebounced();
 +    }
++
     $('#keepalive_enabled').prop('checked', extension_settings[MODULE].enabled);
     $('#keepalive_timeout').val(extension_settings[MODULE].timeoutSeconds);
     $('#keepalive_message').val(extension_settings[MODULE].message);

public/scripts/extensions/keepalive/settings.html+1 -1 Viewed

             <label for="keepalive_message" data-i18n="ext_keepalive_message">Keepalive message</label>
             <input id="keepalive_message" class="text_pole" type="text" />
             <small data-i18n="ext_keepalive_help">After this many seconds without activity on the active connection profile, a hidden background prompt (the current chat plus this message (sent as a hidden user message) is sentsubmitted to that profile to keep its server-side prompt cache warm. The completion is capped to a single token and is never saved to the chat. Sending is briefly locked while it runs. Only the profile you're currently using is pinged; each profile has its own idle timer.</small>
             <hr>