AINativeBench / data /processed /RQ1 /GameBuilder-MCP /reference_trajectory.yaml
王子睿
restructure + add files
8c10cf2
Raw
History Blame Contribute Delete
7.63 kB
# Complete reference trajectory configuration - GameBuilder-MCP
#
# Version: ideal trajectory derived from the code design (applies to all models)
# Principle: include only steps explicitly required by the design; do not rely on test statistics
#
# Design analysis:
# 1. Three agents execute sequentially:
# - Senior Software Engineer: generate game code
# - Software Quality Control Engineer: review code (uses validate_python_code)
# - Chief Software Quality Control Engineer: final evaluation (uses validate_python_code)
#
# 2. Tool usage analysis:
# - validate_python_code is provided via the MCP server
# - The tool is configured specifically for QA and Chief QA (via mcp_adapter.tools)
# - tasks.yaml requires "check for errors" and "syntax errors"
# - Validating code is a core QA responsibility; tool calls are required
#
# 3. Execution flow:
# - Process.sequential: sequential execution
# - Each agent executes its task and produces results via the LLM
#
# 4. MCP variant characteristics:
# - Tools are obtained from the MCP server via MCPServerAdapter
# - Tool name is the function name: validate_python_code
# - The key difference from the non-MCP variant is the tool provisioning mechanism
# Project name
project_name: "GameBuilder-MCP"
# Trajectory extraction configuration
# Focus: SPAN, Chain, AGENT, LLM, Tool
# Ignore: non-core items like Task Created, Crew Created
extract_types:
- "SPAN" # Level 1: top-level execution span
- "Chain" # Level 2: crew execution chain
- "AGENT" # Level 3: agent execution
- "LLM" # Level 4: LLM call
- "Tool" # Level 4: tool call
# ============================================================================
# Reference trajectory (ground truth) - derived from code design
# ============================================================================
#
# Visual structure (ideal execution path):
#
# [SPAN] GameBuilder_generation
# └─ [Chain] Crew***.kickoff
# ├─ [AGENT] Senior Software Engineer ← Task 1: code_task
# │ └─ [LLM] * ← generate game code
# │
# ├─ [AGENT] Software Quality Control Engineer ← Task 2: review_task
# │ ├─ [LLM] * ← review code and decide to validate
# │ ├─ [Tool] validate_python_code ← validate code syntax (MCP tool)
# │ └─ [LLM] * ← handle validation results and output
# │
# └─ [AGENT] Chief Software Quality Control Engineer ← Task 3: evaluate_task
# ├─ [LLM] * ← final evaluation and decide to validate
# ├─ [Tool] validate_python_code ← confirm code is runnable (MCP tool)
# └─ [LLM] * ← confirm completion and output
#
# Design rationale:
# - crew.py: defines a sequential flow of 3 agents and 3 tasks
# - tasks.yaml: all tasks expect "full python code, only the python code"
# - tasks.yaml: review_task requires "check for errors" and "syntax errors"
# - main.py line 532: SPAN name is "GameBuilder_generation"
# - agents.yaml: QA focuses on "checking code for errors"
# - crew.py: QA and Chief QA are configured with mcp_adapter.tools (lines 57 and 68)
# - mcp_server.py: tool name is validate_python_code (line 21)
#
# Ideal trajectory notes:
# - Senior Engineer: generate code (1 LLM call)
# - QA Engineer: review -> validate tool -> output (2 LLM + 1 tool)
# - Chief QA: evaluate -> validate tool -> confirm (2 LLM + 1 tool)
# - This trajectory represents an ideal execution path: no retries, no redundancy
#
# Notes:
# - "LLM: *" means any LLM model (wildcard match)
# - Agent names must exactly match those defined in agents.yaml
# - MCP tool name is the function name: validate_python_code (not "Python Code Validator")
# - Real runs may include more LLM calls (thinking, retries, etc.); this is expected
# ============================================================================
reference_trajectory:
# ===== Level 1: SPAN =====
- "SPAN: GameBuilder_generation"
# ===== Level 2: Chain (wildcard UUID) =====
- "Chain: Crew***.kickoff"
# ===== Task 1: code_task =====
# Senior Software Engineer generates game code
- "Agent: Senior Software Engineer"
- "LLM: *" # Any model
# ===== Task 2: review_task =====
# Software Quality Control Engineer reviews code
- "Agent: Software Quality Control Engineer"
- "LLM: *"
- "Tool: validate_python_code"
- "LLM: *"
# ===== Task 3: evaluate_task =====
# Chief Software Quality Control Engineer performs final evaluation
- "Agent: Chief Software Quality Control Engineer"
- "LLM: *"
- "Tool: validate_python_code"
- "LLM: *"
# Target tool list (for the single-tool use metric)
# Only includes actual tool calls; used to detect tool usage
target_tools:
- "Tool: validate_python_code"
# Models to evaluate
models:
- "GPT-5"
- "GPT-4o-mini"
- "DeepSeek-V3-1"
- "DeepSeek-R1"
- "Gemini-2.5-flash"
- "Gemini-2.5-flash-nothinking"
- "Qwen3-235b"
# ============================================================================
# Usage
# ============================================================================
#
# 1. Trajectory hierarchy structure:
# - Level 1: SPAN (GameBuilder_generation)
# - Level 2: Chain (Crew***.kickoff, wildcard UUID)
# - Level 3: AGENT (specific agent name, must match exactly)
# - Level 4: LLM (wildcard match for any model)
#
# 2. Wildcards:
# - Crew_<UUID>.kickoff -> Crew***.kickoff (handled automatically by evaluate_trajectory.py)
# - LLM: * -> matches any model name (e.g., gpt-4o-mini, deepseek-r1)
#
# 3. Exact match requirements:
# - SPAN name: "GameBuilder_generation" (exact)
# - Chain name: "Crew***.kickoff" (wildcard)
# - AGENT names: must exactly match those defined in agents.yaml
# * "Senior Software Engineer"
# * "Software Quality Control Engineer"
# * "Chief Software Quality Control Engineer"
# - LLM name: "LLM: *" (wildcard)
# - Tool name: "validate_python_code" (exact MCP tool function name)
#
# 4. Metric meanings:
# - Exact Match: trajectories must be identical (including the number of LLM calls)
# - In-order Match: allows extra calls, but core steps must appear in order
# - Any-order Match: includes all required steps (order ignored)
# - Precision: fraction of predicted steps that are correct
# - Recall: fraction of reference steps covered
# - Single-tool Use: checks usage of validate_python_code
#
# 5. Example command:
# cd /Users/wzr/TOSEM-2025/RESULTS/RQ-Failure_Breakdown/GameBuilder-MCP
# python3 evaluate_trajectory.py --config reference_trajectory.yaml
#
# 6. Design notes:
# - This reference trajectory represents the ideal execution path
# - Senior Engineer: 1 LLM call (generate code)
# - QA Engineer: 2 LLM + 1 tool (review -> validate -> output)
# - Chief QA: 2 LLM + 1 tool (evaluate -> validate -> confirm)
# - Total: 5 LLM calls + 2 tool calls
# - In-order Match and Any-order Match are often more informative for real executions
#
# 7. Differences from real trajectories:
# - Real trajectories may include more LLM calls (thinking, planning, execution, summaries)
# - Real trajectories may call the validation tool multiple times
# - These differences are not necessarily errors; they reflect execution strategies
# - Exact Match may be low; focus on In-order Match and Recall
#
# 8. MCP variant notes:
# - Tools are provided by the MCP server via SSE
# - Tool names use the function name (validate_python_code)
# - The trajectory is the same as the non-MCP variant; only the tool provisioning differs
# - The evaluation method and criteria are identical