From f579a26cc01aad33214a4a09ebdf5fd79f97c55e Mon Sep 17 00:00:00 2001
From: Adolfo Reyna <aeroreyna@gmail.com>
Date: Fri, 7 Aug 2026 18:27:51 -0400
Subject: [PATCH] Add AGENTS.md template and fix repetition penalty for local
 macOS LLM

---
 AGENTS.md.example |  7 +++++++
 apple_llm.py      | 17 ++++++++++++-----
 bot.py            |  1 +
 3 files changed, 20 insertions(+), 5 deletions(-)
 create mode 100644 AGENTS.md.example

diff --git a/AGENTS.md.example b/AGENTS.md.example
new file mode 100644
index 0000000..59e4bf5
--- /dev/null
+++ b/AGENTS.md.example
@@ -0,0 +1,7 @@
+# Voice Assistant Personality
+
+You are a fast, concise, and direct spoken voice assistant running locally on Adolfo's Mac.
+
+- Keep your answers short and conversational (1 to 3 sentences).
+- Do not use markdown, bullet points, code blocks, URLs, or emoji in your replies, as they will be read out loud.
+- Speak naturally and get straight to the point.
diff --git a/apple_llm.py b/apple_llm.py
index e02342f..2ea168f 100644
--- a/apple_llm.py
+++ b/apple_llm.py
@@ -92,7 +92,7 @@ class MacOSLLM(FrameProcessor):
     def __init__(
         self,
         *,
-        model: str = "mlx-community/Qwen2.5-0.5B-Instruct-4bit",
+        model: str = "mlx-community/Llama-3.2-1B-Instruct-4bit",
         system_prompt: str | None = None,
         observer=None,
         **kwargs,
@@ -194,7 +194,9 @@ class MacOSLLM(FrameProcessor):
                 loop = asyncio.get_running_loop()
                 def _gen():
                     from mlx_lm import generate
-                    messages = [{"role": "system", "content": self._system_prompt}] + self._history
+                    # Keep history concise to avoid context drift and repetition
+                    recent_history = self._history[-8:]
+                    messages = [{"role": "system", "content": self._system_prompt}] + recent_history
                     prompt = self._mlx_tokenizer.apply_chat_template(
                         messages, add_generation_prompt=True, tokenize=False
                     )
@@ -202,13 +204,18 @@ class MacOSLLM(FrameProcessor):
                         self._mlx_model,
                         self._mlx_tokenizer,
                         prompt=prompt,
-                        max_tokens=256,
+                        max_tokens=150,
+                        temp=0.7,
+                        repetition_penalty=1.2,
+                        repetition_context_size=30,
                         verbose=False,
                     )
 
                 response_text = await loop.run_in_executor(None, _gen)
-                chunks.append(response_text)
-                await self.push_frame(LLMTextFrame(response_text))
+                # Clean up any potential repetitive sentence tails
+                cleaned_text = response_text.strip()
+                chunks.append(cleaned_text)
+                await self.push_frame(LLMTextFrame(cleaned_text))
 
         except asyncio.CancelledError:
             logger.info("Turn cancelled mid-response.")
diff --git a/bot.py b/bot.py
index a21e12a..34c124a 100644
--- a/bot.py
+++ b/bot.py
@@ -383,6 +383,7 @@ def build_vocabulary(args: argparse.Namespace, brain=None) -> Vocabulary | None:
     for target, template in (
         (vocabulary_file, here / "vocabulary.example.txt"),
         (corrections_file, here / "corrections.example.txt"),
+        (workspace / PERSONALITY_FILE, here / "AGENTS.md.example"),
     ):
         if not target.exists() and template.exists():
             target.write_text(template.read_text())