From 78e7b5c3dbc94e4034820cff17df6a24767d2c1e Mon Sep 17 00:00:00 2001
From: Paul Gauthier <aider@paulg.org>
Date: Sat, 22 Jul 2023 08:45:40 -0300
Subject: [PATCH] wip

---
 aider/coders/base_coder.py |  2 ++
 aider/history.py           | 70 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 72 insertions(+)
 create mode 100644 aider/history.py

diff --git a/aider/coders/base_coder.py b/aider/coders/base_coder.py
index e54fb1087..9e245bbed 100755
--- a/aider/coders/base_coder.py
+++ b/aider/coders/base_coder.py
@@ -353,6 +353,8 @@ class Coder:
 
     def move_back_cur_messages(self, message):
         self.done_messages += self.cur_messages
+        self.done_messages = summarize_chat_history(self.done_messages)
+
         if message:
             self.done_messages += [
                 dict(role="user", content=message),
diff --git a/aider/history.py b/aider/history.py
new file mode 100644
index 000000000..f38e98e86
--- /dev/null
+++ b/aider/history.py
@@ -0,0 +1,70 @@
+import json
+
+import tiktoken
+from prompt_toolkit.completion import Completion
+
+from aider import prompts
+
+from .dump import dump  # noqa: F401
+
+
+class ChatSummary:
+    def __init__(self, model, max_tokens=1024):
+        self.tokenizer = tiktoken.encoding_for_model(model)
+        self.max_tokens = max_tokens
+
+    def summarize(self, messages):
+        num = len(messages)
+        if num < 2:
+            return messages
+
+        total = 0
+        sized = []
+        for msg in messages:
+            tokens = len(self.tokenizer.encode(json.dumps(msg)))
+            sized.append((tokens, msg))
+            total += tokens
+
+        if total <= self.max_tokens:
+            return messages
+
+        num = num // 2
+
+        # we want the head to end with an assistant msg
+        if messages[num]["role"] == "assistant":
+            num += 1
+
+        head = messages[:num]
+        tail = messages[num:]
+
+        summary = self.summarize_all(head)
+
+        tail_tokens = sum(tokens for tokens, msg in sized[num:])
+        summary_tokens = len(self.tokenizer.encode(json.dumps(summary)))
+
+        result = summary + tail
+        if summary_tokens + tail_tokens < self.max_tokens:
+            return result
+
+        return self.summarize(result)
+
+    def summarize_all(self, messages):
+        content = ""
+        for msg in messages:
+            role = msg["role"].upper()
+            if role not in ("USER", "ASSISTANT"):
+                continue
+            content += f"# {role}\n"
+            content += msg["content"]
+            if not content.endswith("\n"):
+                content += "\n"
+
+        dump(content)
+
+        messages = [
+            dict(role="system", content=prompts.summarize),
+            dict(role="user", content=content),
+        ]
+
+        summary = simple_send_with_retries(model=models.GPT35.name, messages=messages)
+        dump(summary)