telegram: add conversation context to group pre-filter

Pre-filter now sees last 5 messages so it can detect when user
is continuing a conversation with Ava, even without explicit mention.

- Fetch recent messages before shouldEngageInGroup
- Update classifier prompt to understand Ava context
- Handle follow-up messages to bot's previous responses

1 files changed, 25 insertions, 8 deletions

diff --git a/Omni/Agent/Telegram.hs b/Omni/Agent/Telegram.hs
index 27d7413..ee6784b 100644
--- a/Omni/Agent/Telegram.hs
+++ b/Omni/Agent/Telegram.hs
@@ -437,7 +437,21 @@ handleAuthorizedMessage tgConfig provider engineCfg msg uid userName chatId = do
     if Types.isGroupChat msg
       then do
         putText "Checking if should engage (group chat)..."
-        shouldEngageInGroup (Types.tgOpenRouterApiKey tgConfig) userMessage
+        recentMsgs <- Memory.getRecentMessages uid chatId 5
+        let recentContext =
+              if null recentMsgs
+                then ""
+                else
+                  Text.unlines
+                    [ "[Recent conversation for context]",
+                      Text.unlines
+                        [ (if Memory.cmRole m == Memory.UserRole then "User: " else "Ava: ") <> Memory.cmContent m
+                          | m <- reverse recentMsgs
+                        ],
+                      "",
+                      "[New message to classify]"
+                    ]
+        shouldEngageInGroup (Types.tgOpenRouterApiKey tgConfig) (recentContext <> userMessage)
       else pure True
 
   if not shouldEngage
@@ -597,20 +611,23 @@ shouldEngageInGroup openRouterKey messageText = do
       [ Provider.Message
           Provider.System
           ( Text.unlines
-              [ "You are a classifier that decides if an AI assistant should respond to a message in a group chat.",
+              [ "You are a classifier that decides if an AI assistant named 'Ava' should respond to a message in a group chat.",
+                "You may be given recent conversation context to help decide.",
                 "Respond with ONLY 'yes' or 'no' (lowercase, nothing else).",
                 "",
                 "Say 'yes' if:",
-                "- The message is a direct question the assistant could answer",
+                "- The message is a direct question Ava could answer",
                 "- The message contains a factual error worth correcting",
-                "- The message mentions the bot or asks for help",
+                "- The message mentions Ava or asks for help",
                 "- The message shares a link or document to analyze",
+                "- The message is a follow-up to a conversation Ava was just participating in",
+                "- The user is clearly talking to Ava based on context (e.g. Ava just responded)",
                 "",
                 "Say 'no' if:",
-                "- It's casual banter or chit-chat between people",
-                "- It's a greeting or farewell",
-                "- It's an inside joke or personal conversation",
-                "- It doesn't require or benefit from bot input"
+                "- It's casual banter or chit-chat between people (not involving Ava)",
+                "- It's a greeting or farewell not directed at Ava",
+                "- It's an inside joke or personal conversation between humans",
+                "- It doesn't require or benefit from Ava's input"
               ]
           )
           Nothing