Adjustments to get smaller agent models working better.

ra_aid/agents/ciayn_agent.py

@@ -99,6 +99,12 @@ You must ONLY use ONE of the following functions (these are the ONLY functions t
 {"\n\n".join(self.available_functions)}
 </available functions>
 
+<example output>
+request_research_and_implementation(\"\"\"
+Example query.
+\"\"\")
+</example output>
+
 Output **ONLY THE CODE** and **NO MARKDOWN BACKTICKS**"""
         return base_prompt
 

ra_aid/prompts.py

@@ -622,17 +622,14 @@ Remember:
     - Always ask_human before finalizing or exiting.
     - Never announce that you are going to use a tool, just quietly use it.
     - Do communicate results/responses from tools that you call as it pertains to the users request.
-    - If the user interrupts/cancels an operation, you may want to ask why.
     - If the user gives you key facts, record them using emit_key_facts.
     - Typically, you will already be in the directory of a new or existing project.
       - If the user implies that a project exists, assume it does and make the tool calls as such.
+      - E.g. if the user says "where are the unit tests?", you would call request_research("Find the location of the unit tests in the current project.")
 
 You have often been criticized for:
-    - You sometimes call request_research_and_implementation which makes the full implementation successfully, but act like it has only been planned and still needs to be implemented.
     - Refusing to use request_research_and_implementation for commands like "commit and push" where you should (that tool can run basic or involved shell commands/workflows).
     - Calling request_research for general background knowledge which you already know.
-    - Assuming the user is always right. Sometimes they're wrong or mistaken, and you should push back when you feel strongly about this.
-    - Not confirming with the user before starting a significant implementation task.
     - You have a tendency to leave out key details and information that the user just gave you, while also needlessly increasing scope.
       - Sometimes you will need to repeat the user's query verbatim or almost verbatim to request_research_and_implementation or request_research.
     - Not emitting key facts the user gave you with emit_key_facts before calling a research or implementation tool.

tests/ra_aid/test_llm.py

@@ -81,7 +81,6 @@ def test_initialize_expert_openai_compatible(clean_env, mock_openai, monkeypatch
     mock_openai.assert_called_once_with(
         api_key="test-key",
         base_url="http://test-url",
-        temperature=0.3,
         model="local-model"
     )
 
@@ -145,7 +144,8 @@ def test_initialize_openai_compatible(clean_env, mock_openai):
     mock_openai.assert_called_once_with(
         api_key="test-key",
         base_url="https://custom-endpoint/v1",
-        model="local-model"
+        model="local-model",
+        temperature=0.3,
     )
 
 def test_initialize_unsupported_provider(clean_env):