AINativeBench / data /processed /RQ1 /SocialMediaManager-A2A /reference_trajectory.yaml
王子睿
restructure + add files
8c10cf2
Raw
History Blame Contribute Delete
23.3 kB
# Full reference trajectory configuration - SocialMediaManager-A2A project
#
# Version: ideal trajectory derived from the code design (applies to all models)
# Principle: include only steps explicitly required by the design, not based on test statistics
#
# Code-design analysis:
# 1. Execution flow (orchestrator.py):
# - The Orchestrator manages the overall workflow with two retry layers:
# (a) Orchestrator-level retries: for unexpected errors (up to 3 attempts, MAX_CREW_RETRIES=2)
# (b) Business-logic retries: for validation failures (up to 6 attempts, max_retries=5)
# - Ideal flow (no retries):
# Step 1: Topic Analysis (TopicAnalyzerCrew)
# Step 2-4: Content Generation with Feedback Loop
# - Generate the X post (ShakespeareGeneratorCrew)
# - Validate the X post (PostReviewCrew)
# - If valid=True, end the loop; otherwise retry
# Step 5: Save Results
#
# 2. Agent structure of the three Crews:
# - TopicAnalyzerCrew (topic_analyzer/topic_analyzer_crew.py):
# - Agent: Topic Analysis Expert
# Tools: keyword_extractor, topic_complexity_analyzer
# - ShakespeareGeneratorCrew (content_generator/shakespeare_generator_crew.py):
# - Agent: Shakespearean Bard
# Tools: keyword_extractor, character_counter_tool, emoji_detector_tool,
# post_structure_validator, shakespearean_style_detector, tone_analyzer
# - PostReviewCrew (post_reviewer/post_review_crew.py):
# - Agent: X Post Verifier
# Tools: character_counter_tool, emoji_detector_tool, post_structure_validator,
# shakespearean_style_detector, tone_analyzer
#
# 3. SPAN hierarchy:
# - shakespeare_x_post_orchestrator (top-level, orchestrator.py line 347)
# └─ crew_execution (contains retry logic, orchestrator.py line 362-393)
# ├─ topic_analysis (topic analysis, orchestrator.py line 114-141)
# └─ content_generation_loop (content generation & review loop, orchestrator.py line 143-250)
# includes multiple iterations:
# - ShakespeareGeneratorCrew.kickoff (generate X post)
# - PostReviewCrew.kickoff (validate X post)
# - If valid=True then stop; otherwise continue
#
# 4. MCP tools (tools/mcp_server.py):
# - Character count: character_counter_tool (validate 200-280 chars)
# - Emoji detection: emoji_detector_tool (ensure no emoji)
# - Structure validation: post_structure_validator (validate 1-3-1 structure, 5 lines)
# - Style detection: shakespearean_style_detector (detect Shakespearean style elements)
# - Tone analysis: tone_analyzer (ensure satire and humor)
# - Keyword extraction: keyword_extractor (extract key topic concepts)
# - Complexity analysis: topic_complexity_analyzer (estimate topic complexity)
# Project name
project_name: "SocialMediaManager-A2A"
# Trajectory node types to extract
extract_types:
- "SPAN" # Multi-level SPAN structure
- "Chain" # Each Crew's kickoff chain
- "Agent" # Each Crew's Agent
- "LLM" # LLM calls
- "Tool" # MCP tool calls
# Repeatable pattern configuration
# SocialMediaManager-A2A has a self-evaluation loop (but the ideal trajectory needs only 1 iteration)
repeatable_patterns: []
# ============================================================================
# Reference trajectory (Ground Truth) - derived from code design
# ============================================================================
#
# Graph view (ideal execution path: no retries, no errors, first generation passes validation):
#
# [SPAN] shakespeare_x_post_orchestrator
# └─ [SPAN] crew_execution
# ├─ [SPAN] analyze_topic
# │ └─ [SPAN] a2a_call_topic_analyzer
# │ └─ [SPAN] topic_analyzer_server_execution
# │ └─ [Chain] Crew***.kickoff
# │ └─ [AGENT] Topic Analysis Expert
# │ ├─ [LLM] * (understand the topic)
# │ ├─ [Tool] keyword_extractor (extract keywords)
# │ ├─ [LLM] * (analyze keywords)
# │ ├─ [Tool] topic_complexity_analyzer (analyze complexity)
# │ └─ [LLM] * (produce the topic analysis report)
# └─ [SPAN] content_generation_loop
# ├─ [SPAN] a2a_call_content_generator
# │ └─ [SPAN] content_generator_server_execution
# │ └─ [Chain] Crew***.kickoff (ShakespeareGeneratorCrew - 1st generation)
# │ └─ [AGENT] Shakespearean Bard
# │ ├─ [LLM] * (plan generation strategy)
# │ ├─ [Tool] keyword_extractor (understand topic keywords)
# │ ├─ [LLM] * (draft the initial X post)
# │ ├─ [Tool] character_counter_tool (check character count)
# │ ├─ [LLM] * (adjust character count)
# │ ├─ [Tool] emoji_detector_tool (check emoji)
# │ ├─ [LLM] * (confirm no emoji)
# │ ├─ [Tool] post_structure_validator (validate 1-3-1 structure)
# │ ├─ [LLM] * (adjust structure)
# │ ├─ [Tool] shakespearean_style_detector (check style)
# │ ├─ [LLM] * (strengthen style)
# │ ├─ [Tool] tone_analyzer (check tone)
# │ └─ [LLM] * (final polish and output)
# └─ [SPAN] a2a_call_post_reviewer
# └─ [SPAN] post_reviewer_server_execution
# └─ [Chain] Crew***.kickoff (PostReviewCrew - 1st validation)
# └─ [AGENT] X Post Verifier
# ├─ [LLM] * (read and perform an initial assessment)
# ├─ [Tool] character_counter_tool (validate character count)
# ├─ [LLM] * (record character-count check)
# ├─ [Tool] emoji_detector_tool (validate no emoji)
# ├─ [LLM] * (record emoji check)
# ├─ [Tool] post_structure_validator (validate 1-3-1 structure)
# ├─ [LLM] * (record structure check)
# ├─ [Tool] shakespearean_style_detector (validate style)
# ├─ [LLM] * (record style check)
# ├─ [Tool] tone_analyzer (validate tone)
# └─ [LLM] * (overall decision; output valid=True)
#
# Design rationale:
# - orchestrator.py executes analyze_topic → content_generation_loop in order (via A2A communication)
# - TopicAnalyzerCrew has 1 Agent using 2 tools to analyze the topic
# - ShakespeareGeneratorCrew has 1 Agent using multiple tools for generation and self-validation
# - PostReviewCrew has 1 Agent using multiple tools for comprehensive validation
# - Ideal case: first generation passes validation (valid=True), no retries
# - Agent roles are defined in each crew's config/agents.yaml
# - Tasks are defined in each crew's config/tasks.yaml and explicitly require specific tools
# - A2A architecture: the orchestrator calls three independent agent servers via Agent2Agent
# - Each agent server runs its Crew via the CrewAI framework
#
# Ideal-trajectory notes:
# - Top-level SPAN: shakespeare_x_post_orchestrator
# - Second-level SPAN: crew_execution (contains two-layer retry logic)
# - Two child SPANs run in order: analyze_topic → content_generation_loop
# - analyze_topic stage: call the topic_analyzer server via A2A; Topic Analyst uses
# keyword_extractor and topic_complexity_analyzer
# - content_generation_loop stage includes a self-evaluation loop:
# - Iteration 1 (ideal case):
# * ShakespeareGeneratorCrew generates the X post and self-validates using multiple tools
# * PostReviewCrew validates the X post using all validation tools, returning valid=True
# - No iterations 2-N needed (because the first one passes)
# - This trajectory represents an ideal execution path with no retries and no errors
#
# Notes:
# - "LLM: *" means any LLM model (wildcard match)
# - Agent names must exactly match the role fields in agents.yaml
# - Tool names must exactly match the tool names provided by each agent server
# - Both Generator and Verifier use multiple tools (as required by tasks.yaml)
# - Tool-call order may differ slightly depending on agent decisions
# - Some LLM calls may be merged or split, but core tool calls must exist
# ============================================================================
reference_trajectory:
# ===== Top-level SPAN =====
- "SPAN: shakespeare_x_post_orchestrator"
# ===== Second-level SPAN (contains two-layer retry logic) =====
- "SPAN: crew_execution"
# ===== Stage 1: Analyze Topic (TopicAnalyzerCrew) =====
- "SPAN: analyze_topic"
- "SPAN: a2a_call_topic_analyzer"
- "SPAN: topic_analyzer_server_execution"
- "Chain: Crew***.kickoff"
- "Agent: Topic Analysis Expert"
- "LLM: *" # Initial understanding of the topic; plan analysis strategy
- "Tool: keyword_extractor" # Required by tasks.yaml (no order requirement)
- "LLM: *" # Analyze keyword extraction results
- "Tool: topic_complexity_analyzer" # Required by tasks.yaml (no order requirement)
- "LLM: *" # Synthesize analysis; produce topic insights
# ===== Stage 2: Content Generation Loop (Self-Evaluation Loop) =====
- "SPAN: content_generation_loop"
# === Iteration 1: Content generation (ShakespeareGeneratorCrew) ===
- "SPAN: a2a_call_content_generator"
- "SPAN: content_generator_server_execution"
- "Chain: Crew***.kickoff"
- "Agent: Shakespearean Bard"
- "LLM: *" # Understand the topic and topic_insights; plan generation strategy
- "Tool: keyword_extractor" # MUST be first: tasks.yaml STEP 1 requires "First, use the keyword_extractor tool"
- "LLM: *" # Draft the initial X post based on keywords
# The following tools are used for validation as required by tasks.yaml (no order requirement):
- "Tool: character_counter_tool" # Validate character count (200-280)
- "LLM: *" # Adjust based on character-count check
- "Tool: emoji_detector_tool" # Ensure no emoji
- "LLM: *" # Confirm emoji check passes
- "Tool: post_structure_validator" # Validate 1-3-1 structure (5 lines)
- "LLM: *" # Adjust based on structure check
- "Tool: shakespearean_style_detector" # Check Shakespearean style
- "LLM: *" # Strengthen style elements
- "Tool: tone_analyzer" # Validate satirical and humorous tone
- "LLM: *" # Final polish and output X post
# === Iteration 1: Content validation (PostReviewCrew) ===
- "SPAN: a2a_call_post_reviewer"
- "SPAN: post_reviewer_server_execution"
- "Chain: Crew***.kickoff"
- "Agent: X Post Verifier"
- "LLM: *" # Read the X post and plan the validation strategy
# The following 5 tools are all required by tasks.yaml (no order requirement):
- "Tool: character_counter_tool" # Validate character count (200-280)
- "LLM: *" # Analyze character-count check results
- "Tool: emoji_detector_tool" # Validate no emoji
- "LLM: *" # Analyze emoji check results
- "Tool: post_structure_validator" # Validate 1-3-1 structure (5 lines)
- "LLM: *" # Analyze structure check results
- "Tool: shakespearean_style_detector" # Validate Shakespearean style
- "LLM: *" # Analyze style check results
- "Tool: tone_analyzer" # Validate satirical and humorous tone
- "LLM: *" # Combine checks; output valid=True, feedback=""
# Target tool list (used for the single-tool use metric)
# Lists all tools filtered for each Agent in this config
target_tools:
# Topic Analysis stage
- "Tool: keyword_extractor"
- "Tool: topic_complexity_analyzer"
# Content Generation stage
- "Tool: character_counter_tool"
- "Tool: emoji_detector_tool"
- "Tool: post_structure_validator"
- "Tool: shakespearean_style_detector"
- "Tool: tone_analyzer"
# Post Review stage (shares tools with Generation; not duplicated here)
# ============================================================================
# Dynamic tool permutation configuration
# ============================================================================
# Define permutable tool groups for certain stages to generate reference trajectories
# with different tool orders. During evaluation, all permutations are tried and the
# best-matching one is selected as the reference.
#
# Notes:
# - Topic Analyst's two tools can be in any order (tasks.yaml does not specify order)
# - Shakespearean Bard's keyword_extractor must be first (explicitly required by tasks.yaml),
# but the subsequent 5 validation tools can be in any order
# - X Post Verifier's 5 validation tools can be in any order (tasks.yaml does not specify order)
permutable_tool_groups:
# Topic Analysis stage tools can be in any order
topic_analysis_tools:
- "Tool: keyword_extractor"
- "Tool: topic_complexity_analyzer"
# Shakespearean Bard validation tools can be in any order
# Note: this excludes the first keyword_extractor (it must remain the first tool)
shakespearean_bard_validation_tools:
- "Tool: character_counter_tool"
- "Tool: emoji_detector_tool"
- "Tool: post_structure_validator"
- "Tool: shakespearean_style_detector"
- "Tool: tone_analyzer"
# X Post Verifier validation tools can be in any order
x_post_verifier_tools:
- "Tool: character_counter_tool"
- "Tool: emoji_detector_tool"
- "Tool: post_structure_validator"
- "Tool: shakespearean_style_detector"
- "Tool: tone_analyzer"
# 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:
# - Level 1: SPAN (shakespeare_x_post_orchestrator)
# - Level 2: SPAN (crew_execution)
# - Level 3: SPAN (analyze_topic, content_generation_loop)
# - Level 4: SPAN (a2a_call_topic_analyzer, a2a_call_content_generator, a2a_call_post_reviewer)
# - Level 5: SPAN (topic_analyzer_server_execution, content_generator_server_execution, post_reviewer_server_execution)
# - Level 6: Chain (Crew***.kickoff; UUID wildcard handling)
# - Level 7: AGENT (exact match for agent role)
# - Level 8: LLM (wildcard match for any model)
# - Level 8: Tool (exact match for tool name)
#
# 2. Wildcard handling:
# - Crew_<UUID>.kickoff -> Crew***.kickoff (handled automatically by evaluate_trajectory.py)
# - LLM: * -> matches any model name (e.g., gpt-4o-mini, deepseek-reasoner, gemini-2.5-flash, etc.)
# - Agent names must match in full (including the full role name)
#
# 3. Exact-match requirements:
# - SPAN names: "shakespeare_x_post_orchestrator", "crew_execution",
# "analyze_topic", "content_generation_loop",
# "a2a_call_topic_analyzer", "a2a_call_content_generator", "a2a_call_post_reviewer",
# "topic_analyzer_server_execution", "content_generator_server_execution", "post_reviewer_server_execution"
# - Chain name: "Crew***.kickoff" (wildcard match)
# - Agent names: "Agent: Topic Analysis Expert",
# "Agent: Shakespearean Bard",
# "Agent: X Post Verifier"
# (exact match for the role field in agents.yaml)
# - LLM name: "LLM: *" (wildcard match for any model)
# - Tool names: exact match for tool names (keyword_extractor, topic_complexity_analyzer,
# character_counter_tool, emoji_detector_tool, post_structure_validator,
# shakespearean_style_detector, tone_analyzer)
#
# 4. Metric definitions:
# - Exact Match: trajectories must be identical (including LLM-call counts and tool-call counts)
# - In-order Match: extra calls allowed, but core steps must appear in order
# - Any-order Match: all required steps must be present (order ignored)
# - Precision: fraction of predicted steps that are correct
# - Recall: fraction of reference steps covered by the prediction
# - Single-tool Use: checks whether core tools are used
# - unique_path_ratio: path diversity (number of unique full trajectories / number of samples)
# - path_entropy: path entropy (Shannon entropy over trajectory frequencies, normalized to 0-1)
#
# 5. Run command:
# cd /Users/wzr/TOSEM-2025/RESULTS/RQ-Failure_Breakdown/SocialMediaManager-A2A
# python3 evaluate_trajectory.py --config reference_trajectory.yaml
#
# 6. Design notes:
# - This reference trajectory represents an ideal execution path (no retries, no errors)
# - Two main stages execute in order: Analyze Topic → Content Generation Loop
# - Stage 1 (Analyze Topic): Topic Analyst uses 2 tools to analyze the topic
# - Stage 2 (Content Generation Loop): includes a self-evaluation loop
# * Shakespearean Bard generates the X post and self-validates with multiple tools
# * X Post Verifier performs comprehensive validation using all validation tools
# * Ideal case: first iteration returns valid=True with no retries
# - In-order Match and Recall are often more suitable for evaluating real-world behavior
# - The reference trajectory only defines core required steps; extra tool calls and LLM reasoning are allowed
#
# 7. Differences from actual trajectories:
# - Actual trajectories may include business-logic retries (validation failure triggers regeneration)
# - Actual trajectories may include orchestrator-level retries (exceptions trigger a full workflow retry)
# - The Generator may use a different tool order (but should still use tools required by tasks.yaml)
# - The Verifier may use a different tool order (but must use all 5 validation tools)
# - Reasoning models (e.g., DeepSeek-R1) may include many additional LLM calls
# - Some models (e.g., Gemini) may skip tool calls and generate results directly (not compliant)
# - These differences are not always errors, but skipping required tools or failing validation is abnormal
#
# 8. A2A version characteristics:
# - Uses an Agent2Agent (A2A) architecture; the orchestrator calls 3 independent agent servers via A2A
# - Each agent server runs a CrewAI Crew with its own toolset
# - Includes a multi-level SPAN structure (orchestrator + crew_execution + 2 stage SPANs + A2A comm SPAN + server_execution SPAN)
# - Three specialized Crews collaborate to generate a Shakespeare-style X post
# - Each Crew has clear responsibilities and an explicit toolset
# - TopicAnalyzerCrew: 1 Agent analyzing the topic and extracting insights (via the topic_analyzer server)
# - ShakespeareGeneratorCrew: 1 Agent generating and self-validating the X post (via the content_generator server)
# - PostReviewCrew: 1 Agent comprehensively validating X post quality (via the post_reviewer server)
# - Tool categories: analysis tools (keyword_extractor, topic_complexity_analyzer),
# validation tools (character_counter_tool, emoji_detector_tool, post_structure_validator,
# shakespearean_style_detector, tone_analyzer)
#
# 9. Business-logic notes:
# - The self-evaluation loop is critical: generation must be followed by validation
# - If validation fails (valid=False), regeneration happens with feedback (up to 6 attempts)
# - If all attempts fail, the workflow ends and the failing X post is saved
# - The ideal trajectory assumes the first generation passes validation (best case)
# - The Generator should self-validate using all tools explicitly required by tools.yaml
# - The Verifier must use all 5 validation tools for thorough validation
# - Some actual trajectories may include multiple generate-validate loops (normal business-logic retries)
#
# 10. Tool usage requirements (based on tasks.yaml analysis):
# - Topic Analyst (analyze_topic task):
# * Explicit requirement: "Start by using the available analytical tools"
# * Must use: keyword_extractor, topic_complexity_analyzer
# * Ordering: none (tasks.yaml does not specify order)
# - Shakespearean Bard (write_x_post task):
# * Explicit requirement: "STEP 1 - REQUIRED: First, use the keyword_extractor tool"
# * Explicit requirement: "IMPORTANT: Before finalizing, use the available tools to verify quality"
# * Must use: keyword_extractor (must be the first tool)
# * Should use: character_counter_tool, emoji_detector_tool, post_structure_validator,
# shakespearean_style_detector, tone_analyzer (quality validation)
# * Ordering: keyword_extractor must be first; other validation tools have no order requirement
# - X Post Verifier (verify_x_post task):
# * Explicit requirement: "CRITICAL: You MUST use the following tools to perform thorough validation"
# * Must use: character_counter_tool, emoji_detector_tool, post_structure_validator,
# shakespearean_style_detector, tone_analyzer (all 5)
# * Requirement: "Base your validation decision on these tool results, not assumptions"
# * Ordering: none (tasks.yaml does not specify order)
#
# 11. Tool-call ordering notes:
# - **Scenarios with ordering requirements**:
# * Shakespearean Bard's keyword_extractor must be the first tool
# (tasks.yaml line 7-8: "STEP 1 - REQUIRED: First, use the keyword_extractor tool")
# - **Scenarios without ordering requirements**:
# * Topic Analyst's two tools (keyword_extractor, topic_complexity_analyzer) can be in any order
# * Shakespearean Bard's 5 validation tools can be in any order (as long as after keyword_extractor)
# * X Post Verifier's 5 validation tools can be in any order
# - **Order in the reference trajectory**:
# * The listed order reflects a common observed pattern
# * Evaluation should allow flexibility in tool order (except Bard keyword_extractor must be first)
# * The evaluation script should enforce the constraint that keyword_extractor must be first for the Bard
#
# 12. Dynamic tool permutation optimization:
# - **Design idea**: tool calls within each AGENT may appear in different orders
# - **Mechanism**:
# * permutable_tool_groups defines the tool groups that can be permuted
# * the evaluation script generates all possible tool-order combinations (cartesian product)
# * for each sample, it selects the permutation with the highest combined score from
# exact_match, in_order_match, and any_order_match
# - **Number of combinations**:
# * topic_analysis_tools: 2! = 2 permutations
# * shakespearean_bard_validation_tools: 5! = 120 permutations
# * x_post_verifier_tools: 5! = 120 permutations
# * total: 2 × 120 × 120 = 28,800 reference trajectories
# - **Evaluation strategy**:
# * dynamically select the best reference trajectory per test sample
# * combined score = exact_match×3 + in_order_match×2 + any_order_match×1
# * more fair across different models' tool-calling strategies
# - **Notes**:
# * the Bard's first keyword_extractor is not permuted (must remain first)
# * permutations are only within the same AGENT and do not cross AGENT boundaries
# * tool permutations do not change where/when LLM calls appear or how many there are