Reduce tool count to make tool calling more reliable.

ra_aid/console/formatting.py

@@ -31,7 +31,7 @@ def print_stage_header(stage: str) -> None:
     icon = icons.get(stage_key, "🚀")
 
     # Create styled panel with icon
-    panel_content = f"{icon} {stage_title}"
+    panel_content = f" {icon} {stage_title}"
     console.print(Panel(panel_content, style="green bold", padding=0))
 
 

ra_aid/prompts.py

@@ -119,7 +119,9 @@ Prioritize checking current documentation for technical advice.
 
 # New project hints
 NEW_PROJECT_HINTS = """
-If the user did not specify a stack, use your best judgment, or make a proposal and ask the human if the human-in-the-loop tool is available.
+Because this is a new project:
+- If the user did not specify a stack, use your best judgment, or make a proposal and ask the human if the human-in-the-loop tool is available.
+- If the user did not specify a directory to create the project in, create directly in the current directory.
 """
 
 # Research stage prompt - guides initial codebase analysis
@@ -286,10 +288,6 @@ You have often been criticized for:
 {human_section}
 {web_research_section}
 
-If you make tool calls incorrectly, you **WILL** get errors like the following:
-
-Error: 1 validation error for emit_research_notes notes Field required [type=missing, input_value=, input_type=dict]
-
 NEVER ANNOUNCE WHAT YOU ARE DOING, JUST DO IT!
 """
 
@@ -501,12 +499,6 @@ You have often been criticized for:
     - Not indicating confidence levels or noting uncertainties
     - Not calling tools/functions properly, e.g. leaving off required arguments, calling a tool in a loop, calling tools inappropriately.
 
-
-
-If you make tool calls incorrectly, you **WILL** get errors like the following:
-
-Error: 1 validation error for emit_research_notes notes Field required [type=missing, input_value=, input_type=dict]
-
 NEVER ANNOUNCE WHAT YOU ARE DOING, JUST DO IT!
 """
 
@@ -602,52 +594,32 @@ NEVER ANNOUNCE WHAT YOU ARE DOING, JUST DO IT!
 IMPLEMENTATION_PROMPT = """Current Date: {current_date}
 Working Directory: {working_directory}
 
-Base-level task (for reference only):
+<key facts>
-<base task>
-{base_task}
-</base task>
-
-keep it simple. if the expert tool is available, use it frequently for high level logic and planning.
-
-Plan Overview (for reference only, remember you are only implementing your specific task):
-{plan}
-
-Key Facts:
 {key_facts}
+</key facts>
 
-Key Snippets:
+<key snippets>
 {key_snippets}
+</key snippets>
 
-Relevant Files:
+<relevant files>
 {related_files}
+</relevant files>
 
-Research Notes:
+<research notes>
 {research_notes}
-
+</research notes>
-Work done so far:
-<work log>
-{work_log}
-</work log>
 
 Important Notes:
 - Focus solely on the given task and implement it as described.
 - Scale the complexity of your solution to the complexity of the request. For simple requests, keep it straightforward and minimal. For complex requests, maintain the previously planned depth.
-- Use delete_key_facts to remove facts that become outdated, irrelevant, or duplicated.
 - Use emit_key_snippets to manage code sections before and after modifications in batches.
-- Regularly remove outdated snippets with delete_key_snippets.
-Instructions:
-1. Review the provided base task, plan, and key facts.
-2. Implement only the specified task:
-<task definition>
-{task}
-</task definition>
 
-3. Work incrementally, validating as you go. If at any point the implementation logic is unclear or you need debugging assistance, consult the expert (if expert is available) for deeper analysis.
+- Work incrementally, validating as you go. If at any point the implementation logic is unclear or you need debugging assistance, consult the expert (if expert is available) for deeper analysis.
-4. Use delete_key_facts to remove any key facts that no longer apply.
+- Do not add features not explicitly required.
-5. Do not add features not explicitly required.
+- Only create or modify files directly related to this task.
-6. Only create or modify files directly related to this task.
+- Use file_str_replace and write_file_tool for simple file modifications.
-7. Use file_str_replace and write_file_tool for simple file modifications.
+- Delegate to run_programming_task for more complex programming tasks. This is a capable human programmer that can work on multiple files at once.
-8. Delegate to run_programming_task for more complex programming tasks. This is a capable human programmer that can work on multiple files at once.
 
 Testing:
 
@@ -659,7 +631,7 @@ Testing:
 - Only test UI components if there is already a UI testing system in place.
 - Only test things that can be tested by an automated process.
 
-Once the task is complete, ensure all updated files are emitted.
+Once the task is complete, ensure all updated files are registered with emit_related_files.
 
 {expert_section}
 {human_section}
@@ -670,10 +642,16 @@ You have often been criticized for:
   - Doing changes outside of the specific scoped instructions.
   - Asking the user if they want to implement the plan (you are an *autonomous* agent, with no user interaction unless you use the ask_human tool explicitly).
   - Not calling tools/functions properly, e.g. leaving off required arguments, calling a tool in a loop, calling tools inappropriately.
+  - Using run_programming_task to simply write the full contents of files when you could have used write_file_tool instead.
 
-If you make tool calls incorrectly, you **WILL** get errors like the following:
+Instructions:
+1. Review the provided base task, plan, and key facts.
+2. Implement only the specified task:
+<task definition>
+{task}
+</task definition>
 
-Error: 1 validation error for run_programming_task instructions Field required [type=missing, input_value=, input_type=dict] 
+KEEP IT SIMPLE
 
 NEVER ANNOUNCE WHAT YOU ARE DOING, JUST DO IT!
 """

ra_aid/tool_configs.py

@@ -30,6 +30,7 @@ from ra_aid.tools.agent import (
     request_web_research,
 )
 from ra_aid.tools.memory import one_shot_completed
+from ra_aid.tools.write_file import write_file_tool
 
 
 # Read-only tools that don't modify system state
@@ -48,10 +49,11 @@ def get_read_only_tools(
     tools = [
         emit_related_files,
         emit_key_facts,
-        delete_key_facts,
         emit_key_snippets,
-        delete_key_snippets,
+        # *TEMPORARILY* disabled to improve tool calling perf.
-        deregister_related_files,
+        # delete_key_facts,
+        # delete_key_snippets,
+        # deregister_related_files,
         list_directory_tree,
         read_file_tool,
         fuzzy_find_project_files,
@@ -70,14 +72,15 @@ def get_read_only_tools(
 
 # Define constant tool groups
 READ_ONLY_TOOLS = get_read_only_tools()
-MODIFICATION_TOOLS = [run_programming_task]
+MODIFICATION_TOOLS = [run_programming_task, write_file_tool]
 COMMON_TOOLS = get_read_only_tools()
 EXPERT_TOOLS = [emit_expert_context, ask_expert]
 RESEARCH_TOOLS = [
     emit_research_notes,
-    one_shot_completed,
+    # *TEMPORARILY* disabled to improve tool calling perf.
-    monorepo_detected,
+    # one_shot_completed,
-    ui_detected,
+    # monorepo_detected,
+    # ui_detected,
 ]
 
 
@@ -129,9 +132,10 @@ def get_planning_tools(
 
     # Add planning-specific tools
     planning_tools = [
-        emit_plan,
         request_task_implementation,
-        plan_implementation_completed,
+        # *TEMPORARILY* disabled to improve tool calling perf.
+        # emit_plan,
+        # plan_implementation_completed,
     ]
     tools.extend(planning_tools)
 
@@ -202,9 +206,10 @@ def get_chat_tools(
         request_research,
         request_research_and_implementation,
         emit_key_facts,
-        delete_key_facts,
+        # *TEMPORARILY* disabled to improve tool calling perf.
-        delete_key_snippets,
+        # delete_key_facts,
-        deregister_related_files,
+        # delete_key_snippets,
+        # deregister_related_files,
     ]
 
     if web_research_enabled:

ra_aid/tools/agent.py

@@ -248,6 +248,8 @@ def request_research_and_implementation(query: str) -> Dict[str, Any]:
 def request_task_implementation(task_spec: str) -> Dict[str, Any]:
     """Spawn an implementation agent to execute the given task.
 
+    Task specs should have the requirements. Generally, the spec will not include any code.
+
     Args:
         task_spec: REQUIRED The full task specification (markdown format, typically one part of the overall plan)
     """

ra_aid/tools/memory.py

@@ -387,6 +387,8 @@ def plan_implementation_completed(message: str) -> str:
     Returns:
         Confirmation message
     """
+    _global_memory["task_completed"] = True
+    _global_memory["completion_message"] = message
     _global_memory["plan_completed"] = True
     _global_memory["completion_message"] = message
     _global_memory["tasks"].clear()  # Clear task list when plan is completed

ra_aid/tools/programmer.py

@@ -33,16 +33,20 @@ def run_programming_task(
 
     The programmer sees only what you provide, no conversation history.
 
-    Give detailed instructions but do not write their code.
+    Give detailed instructions including multi-file tasks but do not write their code.
 
-    They are intelligent and can edit multiple files.
+    The programmer cannot run commands.
 
     If new files are created, emit them after finishing.
 
     They can add/modify files, but not remove. Use run_shell_command to remove files. If referencing files you’ll delete, remove them after they finish.
+  
+    Use write_file_tool instead if you need to write the entire contents of file(s).
+
+    If the programmer wrote files, they actually wrote to disk. You do not need to rewrite the output of what the programmer showed you.
 
     Args:
-     instructions: REQUIRED Programming task instructions (markdown format, use newlines and as many tokens as needed)
+     instructions: REQUIRED Programming task instructions (markdown format, use newlines and as many tokens as needed, no commands allowed)
      files: Optional; if not provided, uses related_files
 
     Returns: { "output": stdout+stderr, "return_code": 0 if success, "success": True/False }