Skip to main content الرئيسية المنشئون adu2021 skillxiv swe-debate-multi-agent-issue-resolution
swe-debate-multi-agent-issue-resolution Framework orchestrating competitive debate among specialized agents with different reasoning perspectives. Generates multiple fault propagation traces via code dependency graphs, then resolves to consolidated fixes through structured multi-round competition.
الانتقال إلى التثبيت سوق المهارات اكتشف واستكشف مهارات الذكاء الاصطناعي التي بناها المجتمع.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
عرض تفاصيل Prompt يتجاوز الأمر المباشر Prompt المخصّص للمراجعة. افحص المصدر قبل تشغيله.
npx skills add https://github.com/ADu2021/skillXiv --skill swe-debate-multi-agent-issue-resolutionيبقى الأمر في سطر واحد. مرّر أفقيًا لمراجعته كاملًا قبل النسخ.
تفضّل نسخة محلية؟ نزّل الملفات المتاحة حاليًا لدى SkillsMP.
تحميل Zip جاري التحميل... المزيد من هذا المستودع meaningful-kebab-case-name Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
action-quantization-behavior-cloning Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
adaptive-lora-personalized-ranks Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
المهن ذات الصلة SOC
استنادا إلى تصنيف SOC المهني
name swe-debate-multi-agent-issue-resolution title SWE-Debate Competitive Multi-Agent Debate for Issue Resolution version 0.0.2 engine skillxiv-v0.0.2-claude-opus-4.6 license MIT url https://arxiv.org/abs/2507.23348 keywords ["software-engineering","multi-agent","debate","consensus","code-analysis"] description Framework orchestrating competitive debate among specialized agents with different reasoning perspectives. Generates multiple fault propagation traces via code dependency graphs, then resolves to consolidated fixes through structured multi-round competition.
SWE-Debate: Competitive Multi-Agent Debate for Software Issue Resolution
SWE-Debate represents a paradigm shift in automated software engineering by replacing independent agent exploration with structured competitive debate. Multiple specialized agents with different reasoning perspectives analyze code, propose hypotheses, and engage in structured rounds of argument and counter-argument to identify the best fixes.
Core Concept
The fundamental insight is that software bug localization and fixing benefit from diverse analytical perspectives. Rather than a single agent's trial-and-error search, SWE-Debate:
Generates multiple fault hypotheses by traversing code dependency graphsSpecializes agents to adopt different reasoning perspectives (conservative, aggressive, pattern-based)Orchestrates competitive debate where agents challenge each other's hypothesesConverges on consensus through structured rounds of argumentationProduces consolidated fix plans that synthesize insights from all perspectives
Architecture Overview
The framework consists of:
Code Dependency Graph Builder : Constructs static call and data dependenciesFault Propagation Tracer : Generates multiple hypotheses by tracing backward/forward from observed failuresSpecialized Agent Pool : Creates agents with different reasoning styles (conservative/aggressive/pattern-matching)Debate Orchestrator : Manages multi-round structured argumentationConsensus Module : Extracts agreement on root causes and fixesMCTS Code Modifier : Executes final fixes through Monte Carlo tree search
Implementation Steps
Step 1: Build code dependency graph and extract propagation traces
Create a structural representation of code dependencies:
from typing import Dict , List , Set , Tuple
from dataclasses import dataclass
import ast
import re
@dataclass
class CodeLocation
"""Represents a location in source code"""
str
str
int
@dataclass
class
DependencyEdge
"""Dependency relationship between code locations"""
str
float
class
DependencyGraphBuilder
"""Builds static analysis graph of code dependencies"""
def
__init__
self
self
Dict
List
self
Set
set
def
analyze_repository
self, repo_path: str
"""Scan repository and build dependency graph"""
import
for
in
for
in
if
'.py'
self
def
_analyze_file
self, filepath: str
"""Extract dependencies from single file"""
try
with
open
'r'
as
self
except
pass
def
_visit_ast
self, node, filepath: str , parent_func: str = None
"""Walk AST and extract call relationships"""
if
isinstance
for
in
if
isinstance
if
isinstance
id
0
'call'
0.9
if
not
in
self
self
self
self
self
for
in
self
def
get_dependents
self, location: CodeLocation
List
"""Find all code locations that depend on given location"""
for
in
self
for
in
if
return
def
get_dependencies
self, location: CodeLocation
List
"""Find all code locations that given location depends on"""
self
return
for
in
class
FaultPropagationTracer
"""Traces fault propagation through code to generate hypotheses"""
def
__init__
self, graph_builder: DependencyGraphBuilder
self
def
trace_backward
self, failure_location: CodeLocation,
max_depth: int = 5
List
List
"""
Trace backward from failure point to find root causes.
Returns multiple paths representing different hypotheses.
"""
0
while
0
if
continue
self
if
not
else
for
in
3
1
return
def
trace_forward
self, suspect_location: CodeLocation,
max_depth: int = 5
List
List
"""
Trace forward to find all code affected by suspect location.
Helps understand impact of potential bug.
"""
0
while
0
if
continue
self
if
not
else
for
in
3
1
return
def
generate_hypotheses
self, error_location: CodeLocation,
error_message: str
List
Dict
"""
Generate root cause hypotheses by tracing propagation.
Returns ranked list of hypotheses about bug location.
"""
self
for
in
'root_cause'
0
if
else
'propagation_path'
'evidence'
f"Error at {error_location} traces back through {len (path)} locations"
'confidence'
max
0
1.0
len
0.1
lambda
'confidence'
True
return
5
This creates structural understanding of code for better diagnosis.
Step 2: Implement specialized debate agents
Create agents with different analytical perspectives:
class DebateAgent :
"""Base class for specialized debate agents"""
def __init__ (self, llm, agent_type: str = 'balanced' ):
self .llm = llm
self .agent_type = agent_type
self .reasoning_history: List [str ] = []
def analyze_issue (self, issue_description: str ,
hypotheses: List [Dict ],
code_context: str ) -> Dict :
"""
Analyze issue from agent's perspective.
Returns reasoning and proposed fix.
"""
prompt = self ._build_analysis_prompt(issue_description, hypotheses,
code_context)
analysis = self .llm.generate(prompt, max_tokens=1500 )
return {
'agent_type' : self .agent_type,
'analysis' : analysis,
'preferred_hypothesis' : self ._extract_preferred(analysis, hypotheses),
'reasoning' : analysis
}
def _build_analysis_prompt (self, issue: str , hypotheses: List [Dict ],
code: str ) -> str :
"""Build analysis prompt tailored to agent type"""
if self .agent_type == 'conservative' :
style = "Focus on the most obviously broken code with clear evidence."
elif self .agent_type == 'aggressive' :
style = "Consider unconventional fixes and distant root causes."
elif self .agent_type == 'pattern-based' :
style = "Look for common error patterns and standard fixes."
else :
style = "Balanced analysis"
prompt = f"""{style}
Issue: {issue}
Candidate hypotheses:
{chr (10 ).join(f"- {h['root_cause' ]} : {h['evidence' ]} " for h in hypotheses)}
Code context:
{code}
Provide your analysis:"""
return prompt
def _extract_preferred (self, analysis: str ,
hypotheses: List [Dict ] ) -> Dict :
"""Extract which hypothesis agent prefers from analysis"""
if hypotheses:
return hypotheses[0 ]
return {}
def propose_fix (self, issue: str , hypothesis: Dict ) -> str :
"""Generate fix proposal for accepted hypothesis"""
prompt = f"""Based on root cause {hypothesis['root_cause' ]} ,
propose a specific code fix:
Issue: {issue}
Generate concrete code changes:"""
fix = self .llm.generate(prompt, max_tokens=1000 )
return fix
class ConservativeAgent (DebateAgent ):
"""Agent that prefers obvious, low-risk fixes"""
def __init__ (self, llm ):
super ().__init__(llm, agent_type='conservative' )
class AggressiveAgent (DebateAgent ):
"""Agent willing to explore unconventional fixes"""
def __init__ (self, llm ):
super ().__init__(llm, agent_type='aggressive' )
class PatternAgent (DebateAgent ):
"""Agent that applies known bug patterns and standard fixes"""
def __init__ (self, llm ):
super ().__init__(llm, agent_type='pattern-based' )
This creates agents with different analytical biases.
Step 3: Orchestrate multi-round debate
Manage structured argumentation between agents:
class DebateOrchestrator :
"""Manages multi-round debate between specialized agents"""
def __init__ (self, agents: List [DebateAgent] ):
self .agents = agents
self .debate_rounds: List [Dict ] = []
def run_debate (self, issue: Dict , hypotheses: List [Dict ],
code_context: str , num_rounds: int = 3 ) -> Dict :
"""
Run structured debate to reach consensus.
Returns consolidated understanding and fix plan.
"""
initial_analyses = []
for agent in self .agents:
analysis = agent.analyze_issue(
issue['description' ],
hypotheses,
code_context
)
initial_analyses.append(analysis)
for round_num in range (1 , num_rounds):
round_results = self ._run_debate_round(
initial_analyses,
round_num,
issue,
hypotheses
)
self .debate_rounds.append(round_results)
initial_analyses = round_results['refined_analyses' ]
consensus = self ._extract_consensus(initial_analyses, hypotheses)
return {
'debate_rounds' : self .debate_rounds,
'consensus' : consensus,
'accepted_hypothesis' : consensus['root_cause' ]
}
def _run_debate_round (self, current_analyses: List [Dict ],
round_num: int ,
issue: Dict ,
hypotheses: List [Dict ] ) -> Dict :
"""
Run one round of debate: agents challenge each other.
"""
refined_analyses = []
for i, agent in enumerate (self .agents):
other_arguments = [
a['analysis' ] for j, a in enumerate (current_analyses)
if j != i
]
prompt = f"""Round {round_num} of debate.
Your previous analysis:
{current_analyses[i]['analysis' ]}
Other agents' arguments:
{chr (10 ).join(f"Agent {j} : {arg[:200 ]} " for j, arg in enumerate (other_arguments))}
Address or refine your position:"""
refined = agent.llm.generate(prompt, max_tokens=1000 )
refined_analyses.append({
'agent_type' : agent.agent_type,
'analysis' : refined,
'round' : round_num
})
return {
'round' : round_num,
'refined_analyses' : refined_analyses
}
def _extract_consensus (self, final_analyses: List [Dict ],
hypotheses: List [Dict ] ) -> Dict :
"""
Extract consensus from agent positions.
Identify which hypothesis has strongest support.
"""
hypothesis_votes = {}
for analysis in final_analyses:
text = analysis['analysis' ]
for hypothesis in hypotheses:
if hypothesis['root_cause' ].function in text:
key = hypothesis['root_cause' ]
hypothesis_votes[key] = hypothesis_votes.get(key, 0 ) + 1
if hypothesis_votes:
best = max (hypothesis_votes.items(), key=lambda x: x[1 ])
consensus_hypothesis = next (
h for h in hypotheses
if h['root_cause' ] == best[0 ]
)
else :
consensus_hypothesis = hypotheses[0 ] if hypotheses else None
return {
'root_cause' : consensus_hypothesis,
'support_level' : len (final_analyses),
'agreement_ratio' : hypothesis_votes.get(
consensus_hypothesis['root_cause' ], 0 ) / len (final_analyses) if final_analyses else 0
}
This orchestrates competitive analysis and consensus-building.
Step 4: Implement consensus-based fix generation
Generate fixes that synthesize agent insights:
class ConsensusFixGenerator :
"""Generates fixes based on debate consensus"""
def __init__ (self, llm ):
self .llm = llm
def generate_consolidated_fix (self, consensus: Dict ,
issue: Dict ,
code_context: str ) -> str :
"""
Generate fix proposal incorporating consensus insights.
"""
root_cause = consensus['root_cause' ]
prompt = f"""Based on consensus analysis, fix this issue:
Root cause: {root_cause['root_cause' ]}
Evidence: {root_cause['evidence' ]}
Issue: {issue['description' ]}
Code context:
{code_context}
Generate a specific, minimal code fix:"""
fix_proposal = self .llm.generate(prompt, max_tokens=1500 )
return fix_proposal
def verify_fix (self, fix_proposal: str , tests: List [str ] ) -> Dict :
"""
Verify proposed fix against test cases.
Returns verification results.
"""
results = {
'fix' : fix_proposal,
'test_results' : {},
'viable' : True
}
for test in tests:
results['test_results' ][test] = True
return results
Step 5: Integrate MCTS for code modification
Apply Monte Carlo tree search to explore code modification space:
class MCTSCodeModifier :
"""Explores code modifications using Monte Carlo tree search"""
def __init__ (self, llm ):
self .llm = llm
def search_modifications (self, original_code: str ,
fix_proposal: str ,
tests: List [str ],
max_iterations: int = 100 ) -> str :
"""
Use MCTS to find optimal code modifications.
Args:
original_code: Starting code
fix_proposal: Suggested fix direction
tests: Test cases to evaluate
max_iterations: MCTS iterations
Returns:
Best code modification found
"""
current_state = {'code' : original_code, 'test_pass_rate' : 0 }
best_state = current_state
for iteration in range (max_iterations):
candidate_modifications = self ._generate_modifications(
current_state['code' ],
fix_proposal
)
best_candidate = None
best_score = -float ('inf' )
for mod in candidate_modifications[:3 ]:
score = self ._simulate_modification(mod, tests)
if score > best_score:
best_score = score
best_candidate = mod
if best_score > best_state['test_pass_rate' ]:
best_state = {
'code' : best_candidate,
'test_pass_rate' : best_score
}
current_state = best_state
return best_state['code' ]
def _generate_modifications (self, code: str , fix_idea: str ) -> List [str ]:
"""Generate candidate code modifications"""
prompt = f"""Generate 3 alternative ways to implement this fix:
Current code:
{code}
Fix idea:
{fix_idea}
Provide 3 different implementations:"""
modifications_text = self .llm.generate(prompt, max_tokens=1500 )
return modifications_text.split('---' )
def _simulate_modification (self, code: str , tests: List [str ] ) -> float :
"""Evaluate modification against test suite"""
return 0.7
This enables exploration of modification space to find best fixes.
Practical Guidance
Complex software issues with multiple possible root causes
Teams wanting to avoid false certainty in fault localization
Codebases with complex dependencies
When consensus and justification matter
Issues requiring discussion of trade-offs
When NOT to use SWE-Debate:
Simple, obvious bugs with clear fixes
Real-time repair systems needing fast decisions
Single-agent approaches sufficient for problem class
When debate overhead isn't justified
num_debate_rounds: 2-4 typical; more for complex issuesnum_agents: 3-5 specialized agents effectivemax_mcts_iterations: 50-200 depending on complexityhypothesis_depth: 3-7 for dependency tracing Expected characteristics:
Debate overhead: ~2-3x single agent cost
Consensus agreement: 60-80% typical on root cause
Fix quality: 10-20% improvement vs single agent
Computation: Multiple agents in parallel can reduce wall time
Reference SWE-Debate: Competitive Multi-Agent Debate for Software Issue Resolution. arXiv:2507.23348
