fix: use first entry per message ID for correct output_tokens

Claude Code JSONL writes multiple entries per message (streaming chunks).
The FIRST entry contains the correct final usage from the API response.
Later entries are intermediate states with inflated cumulative output_tokens.

Previously we kept the entry with the highest output_tokens, causing output
to be ~9x higher than actual. Now we keep the first entry, matching ccusage
CLI's behavior (which also deduplicates by messageId:requestId, keeping first).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: jx453331958

agent/agent.js

@@ -329,9 +329,9 @@ function parseJsonlFile(filePath) {
           _msgId: msgId,
         };
 
-        // Keep the entry with the largest output_tokens per message ID
-        const existing = msgMap.get(msgId);
-        if (!existing || outputTokens >= existing.output_tokens) {
+        // Keep the FIRST entry per message ID (it has the correct final usage)
+        // Later entries are streaming chunks with cumulative intermediate values
+        if (!msgMap.has(msgId)) {
           msgMap.set(msgId, recordData);
         }
       } catch (err) {

agent/agent.py

@@ -261,8 +261,9 @@ def parse_jsonl_file(file_path: Path, state: State) -> List[Dict]:
                         '_msg_id': msg_id,
                     }
 
-                    # Keep the entry with the largest output_tokens per message ID
-                    if msg_id not in msg_map or output_tokens >= msg_map[msg_id]['output_tokens']:
+                    # Keep the FIRST entry per message ID (it has the correct final usage)
+                    # Later entries are streaming chunks with cumulative intermediate values
+                    if msg_id not in msg_map:
                         msg_map[msg_id] = record_data
 
                 except json.JSONDecodeError: