scawful/yaze
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: Enhance system prompt JSON schema and clarify tool result formatting in conversational agent

Browse Source
This commit is contained in:
scawful
2025-10-04 03:19:28 -04:00
parent 551f926aba
commit 9acd6ba17f
3 changed files with 28 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
assets/agent/system_prompt.txt
View File

@@ -3,19 +3,26 @@ You are an expert ROM hacking assistant for The Legend of Zelda: A Link to the P
Your task is to generate a sequence of z3ed CLI commands to achieve the user's request, or to answer questions about the ROM using available tools.
# Output Format
You MUST respond with ONLY a JSON object with the following structure:
You MUST respond with ONLY a JSON object. NO other text before or after the JSON.
**REQUIRED JSON SCHEMA:**
```json
{
"text_response": "Your natural language reply to the user.",
"tool_calls": [{ "tool_name": "tool_name", "args": { "arg1": "value1" } }],
"commands": ["command1", "command2"],
"reasoning": "Your thought process."
"text_response": "string (your natural language reply)",
"tool_calls": [{"tool_name": "string", "args": {"key": "value"}}],
"commands": ["string array of z3ed commands"],
"reasoning": "string (your thought process)"
}
```
**CRITICAL:** The field name is `"text_response"` NOT `"response"` NOT `"answer"` NOT anything else.
# CRITICAL RULES:
1. If you previously called tools and received [TOOL RESULT], you MUST include text_response with your answer
2. NEVER send an empty text_response after receiving tool results
1. If you previously called tools and received [TOOL RESULT], you MUST include "text_response" with your answer
2. NEVER send an empty "text_response" after receiving tool results
3. NEVER call the same tool twice with the same arguments
4. If you have all the information needed to answer, provide text_response WITHOUT calling more tools
4. If you have all the information needed to answer, provide "text_response" WITHOUT calling more tools
5. The field name is `"text_response"` - this exact spelling is REQUIRED
# Tool Calling Workflow (CRITICAL)

7
src/cli/service/agent/conversational_agent_service.cc
View File

@@ -270,7 +270,12 @@ absl::StatusOr<ChatMessage> ConversationalAgentService::SendMessage(
}
// Add tool result with a clear marker for the LLM
std::string marked_output = "[TOOL RESULT] " + tool_output;
// Format as plain text to avoid confusing the LLM with nested JSON
std::string marked_output = absl::StrCat(
"[TOOL RESULT for ", tool_call.tool_name, "]\n",
"The tool returned the following data:\n",
tool_output, "\n\n",
"Please provide a text_response field in your JSON to summarize this information for the user.");
history_.push_back(
CreateMessage(ChatMessage::Sender::kUser, marked_output));
}

7
src/cli/service/ai/ollama_ai_service.cc
View File

@@ -222,6 +222,13 @@ absl::StatusOr<AgentResponse> OllamaAIService::GenerateResponse(
std::string llm_output = ollama_wrapper["response"].get<std::string>();
// Debug: Print raw LLM output when verbose mode is enabled
const char* verbose_env = std::getenv("Z3ED_VERBOSE");
if (verbose_env && std::string(verbose_env) == "1") {
std::cout << "\n" << "\033[35m" << "🔍 Raw LLM Response:" << "\033[0m" << "\n"
<< "\033[2m" << llm_output << "\033[0m" << "\n\n";
}
// Parse the LLM's JSON response (the agent structure)
nlohmann::json response_json;
try {