Agent actions and gral function. Need agent classes to separate advanced agent functionality.

backend/gral.py

@@ -0,0 +1,39 @@
+from inference import LlamaMessage, replicate_llama_completion
+
+message_format_section = """
+When suggesting code to the user, keep the code as small snippets.
+If the update is meant to be applied to a file, include a line linking the file immediately before the code block like this: "[filename.txt]
+```
+updated segment
+```"
+Omit the file link when the block is not related to a known file
+Code blocks must be under 30 lines. Assistant responses MUST NOT contain horizontal rules.
+When suggesting bash code meant to be directly executed, use the same syntax as above but marking the code block with "[eval]" instead."""
+
+async def gral_message(app, system_prompt, message_history: list, file_map: dict):
+    messages = [
+        LlamaMessage("system", "\n".join([
+            system_prompt,
+            message_format_section,
+        ])),
+    ]
+
+    file_parts = []
+    for filename, content in file_map.items():
+        file_parts.append(f"\n---\n[{filename}]\n\n{content}\n")
+    
+    if file_parts:
+        file_lines = [
+            "This message contains all files relevant to the prompt: each file section in this message starts with a horizontal rule, followed by the file name in square brackets, followed by the file content.",
+            "Only this message contains horizontal rules, assistant responses must not include them.",
+            "The prompt and conversation start in the next message after this."
+        ]
+        file_lines.extend(file_parts)
+        messages.append(
+            LlamaMessage("user", "\n".join(file_lines))
+        )
+
+    messages.extend([LlamaMessage(msg['role'], msg['content']) for msg in message_history])
+
+    response = await replicate_llama_completion(app, messages)
+    return response