Intelligence sémantique: Partie M SK1 - Automnement-Outils de guérison par voie d'écoulement~-Aware Pruning
Lorsque vos outils se détruisent eux-mêmes, votre système doit se rappeler pourquoi et ne répéter jamais l'erreur
Quand *DiSE meurtre.
> **Note:** Il s’agit d’un projet spéculatif pour la prochaine percée évolutive de DISE. ', un écosystème d’outils de guérison qui surveille les lignées et détecte les bugs.
Le problème: Outils qui se rompent eux-mêmes
C’est un scénario qui m’amène à me réveiller la nuit.
Tool: data_validator_v2.3.0
Status: Working perfectly ✓
Evolution triggered: "Optimize for speed"
↓
Tool: data_validator_v2.4.0
Status: 40% faster! ✓
Side effect: Now accepts invalid emails ✗
Applications using v2.4.0: 47
Bugs introduced: 47
Developer frustration: ∞
Le système actuel de DISE peut élaborer des outils pour être meilleurs. Mais ce qui se passe quand l’évolution les fait plus grave? Que se passera-t-il si une optimisation introduit un défaut critique?
À l’heure actuelle, nous détectons le défaut.
Mais nous ne le faisons pas. apprentissage d’une manière structurale profonde
Nous ne le faisons pas.
- Remarque quelle mutation a causé le bug
- Prévenir mutations similaires dans les outils connexes
- La prune la branche manquante de l'arbre évolutif
- Propager le savoir pour les outils de descendants
- Auto-recover par la régénération d’un ancêtre bien connu
En essence, nous ne faisons pas de vaccin avec un système de détection associé et le corpus de recherche sur une solution fixe. Mais DiSE nous permet d’en faire presque tout simplement.
Cela change aujourd’hui.
Bien, conceptuellement . C'est le design pour ce qui est de de pouvoir Travail.
La grande idée: Outils comme Git DAG + mémoire évolutive
Pensez à chaque outil dans DISE comme un node d'un Git-comme le graphique acyclique direct (DAGM SK2
graph TD
A[validator_v1.0.0<br/>Initial implementation] --> B[validator_v1.1.0<br/>Added regex patterns]
A --> C[validator_v1.0.1<br/>Bug fix: null handling]
B --> D[validator_v2.0.0<br/>Rewrote for performance]
C --> D
D --> E[validator_v2.1.0<br/>Added email validation]
E --> F[validator_v2.2.0<br/>💥 BUG: Accepts invalid emails]
F -.-> |Detected failure| G[validator_v2.2.1<br/>Auto-regenerated from v2.1.0]
style F stroke:#c92a2a,stroke-width:3px
style G stroke:#2f9e44,stroke-width:3px
Tout outil sait:
- De qui vient-il? (nodesparentsM SK1
- Ce qui a changé (métadonnées de mutation)
- Ce qui a échoué (historique des bogues)
- Qu'il faut éviter (avertissements héritésM SK1
Lorsque l'on détecte une erreur critique, le système:
- Identifie le point de défaillance - Quelle version a introduit le bug?
- Prunes la branche - Marque la version qui a échoué et les descendants comme contaminés
- Propagates des avertissements - Ajout des balises "avoidM SK2 pour prévenir les mutations similaires
- Auto-régénére - Crée une nouvelle version du dernier ancêtre connu
- Mise à jour de la lignée - Enregistre le défaut dans l'arbre évolutionnaire
Le résultat: Un écosystème de guérison où les bogues deviennent une mémoire institutionnelle permanente.
Structure des données: métadonnées linéaires de l'outil
Premièrement, nous devons suivre beaucoup plus que nous le faisons actuellement.
from dataclasses import dataclass, field
from typing import List, Dict, Optional, Set
from datetime import datetime
from enum import Enum
class NodeHealth(Enum):
HEALTHY = "healthy"
DEGRADED = "degraded"
FAILED = "failed"
PRUNED = "pruned"
REGENERATED = "regenerated"
class MutationType(Enum):
OPTIMIZATION = "optimization"
BUG_FIX = "bug_fix"
FEATURE_ADD = "feature_add"
REFACTOR = "refactor"
SECURITY_PATCH = "security"
@dataclass
class MutationRecord:
"""Record of what changed in this evolution"""
mutation_type: MutationType
description: str
timestamp: datetime
fitness_before: float
fitness_after: float
code_diff_hash: str
prompt_used: str
@dataclass
class FailureRecord:
"""Record of a bug or failure"""
failure_type: str
description: str
stack_trace: Optional[str]
test_case_failed: Optional[str]
detection_method: str # "test", "runtime", "static_analysis"
timestamp: datetime
severity: str # "critical", "high", "medium", "low"
@dataclass
class AvoidanceRule:
"""Rules about what NOT to do (learned from failures)"""
rule_id: str
description: str
pattern_to_avoid: str # Regex or semantic description
reason: str # Why this is bad
source_failure: str # Which node failure created this rule
propagation_scope: str # "descendants", "all_similar", "global"
created_at: datetime
@dataclass
class ToolLineage:
"""Complete lineage and health tracking for a tool"""
# Identity
tool_id: str
version: str
full_name: str # e.g., "data_validator_v2.2.0"
# Lineage
parent_ids: List[str] = field(default_factory=list)
child_ids: List[str] = field(default_factory=list)
ancestor_path: List[str] = field(default_factory=list) # Path to root
# Health
health_status: NodeHealth = NodeHealth.HEALTHY
failure_count: int = 0
failures: List[FailureRecord] = field(default_factory=list)
# Evolution
mutations: List[MutationRecord] = field(default_factory=list)
generation: int = 0 # Distance from root
# Learning
avoidance_rules: List[AvoidanceRule] = field(default_factory=list)
inherited_rules: Set[str] = field(default_factory=set) # Rule IDs from ancestors
# Performance
fitness_history: List[float] = field(default_factory=list)
execution_count: int = 0
success_rate: float = 1.0
# Metadata
created_at: datetime = field(default_factory=datetime.now)
last_executed: Optional[datetime] = None
pruned_at: Optional[datetime] = None
regenerated_from: Optional[str] = None
Il s’agit d’un abondance plus de données que nous ne le faisons à l’heure actuelle. Mais c’est tout ce qui est nécessaire pour le vrai soiM SK2la guérison.
Détection: Comment nous savons que quelque chose est défectueux?
Les erreurs critiques peuvent être détectées par plusieurs canaux:
1. Défaillances d'essai Détection immédiateM SK2
class TestBasedDetection:
"""Detect bugs through test execution"""
async def validate_tool_health(
self,
tool_id: str,
lineage: ToolLineage
) -> Optional[FailureRecord]:
"""Run all tests and detect failures"""
# Load tool and its test suite
tool = await self.tools_manager.load_tool(tool_id)
test_suite = await self.test_discovery.find_tests(tool)
results = await self.test_runner.run_tests(test_suite)
# Check for test failures
if results.failed_count > 0:
critical_failures = [
test for test in results.failures
if test.is_critical # BDD scenarios, core functionality
]
if critical_failures:
return FailureRecord(
failure_type="test_failure",
description=f"{len(critical_failures)} critical tests failed",
test_case_failed=critical_failures[0].name,
stack_trace=critical_failures[0].stack_trace,
detection_method="test",
timestamp=datetime.now(),
severity="critical"
)
return None
async def regression_detection(
self,
new_version: str,
old_version: str
) -> Optional[FailureRecord]:
"""Detect if new version broke what old version did correctly"""
# Get test results for both versions
old_results = await self.get_cached_test_results(old_version)
new_results = await self.test_runner.run_tests(new_version)
# Find tests that USED to pass but now fail
regressions = [
test for test in old_results.passed
if test.name in [f.name for f in new_results.failures]
]
if regressions:
return FailureRecord(
failure_type="regression",
description=f"Broke {len(regressions)} previously working tests",
test_case_failed=regressions[0].name,
detection_method="regression_test",
timestamp=datetime.now(),
severity="critical"
)
return None
2. Surveillance du temps d'exécution Détection de la productionM SK2
class RuntimeMonitoring:
"""Detect bugs through execution monitoring"""
def __init__(self):
self.error_threshold = 0.05 # 5% error rate triggers investigation
self.execution_window = 100 # Last 100 executions
async def monitor_tool_health(
self,
tool_id: str,
lineage: ToolLineage
) -> Optional[FailureRecord]:
"""Monitor runtime behavior for anomalies"""
# Get recent execution history
recent_runs = await self.bugcatcher.get_recent_executions(
tool_id,
limit=self.execution_window
)
if len(recent_runs) < 10:
return None # Not enough data
# Calculate error rate
error_count = sum(1 for run in recent_runs if run.had_error)
error_rate = error_count / len(recent_runs)
if error_rate > self.error_threshold:
# Analyze error patterns
error_types = {}
for run in recent_runs:
if run.had_error:
error_types[run.error_type] = error_types.get(run.error_type, 0) + 1
most_common_error = max(error_types.items(), key=lambda x: x[1])
return FailureRecord(
failure_type="runtime_errors",
description=f"Error rate {error_rate:.1%} exceeds threshold",
stack_trace=recent_runs[-1].stack_trace if recent_runs[-1].had_error else None,
detection_method="runtime",
timestamp=datetime.now(),
severity="high" if error_rate > 0.20 else "medium"
)
# Check for performance degradation
if len(lineage.fitness_history) >= 5:
recent_fitness = lineage.fitness_history[-5:]
avg_recent = sum(recent_fitness) / len(recent_fitness)
historical_fitness = lineage.fitness_history[:-5]
avg_historical = sum(historical_fitness) / len(historical_fitness)
degradation = (avg_historical - avg_recent) / avg_historical
if degradation > 0.30: # 30% performance drop
return FailureRecord(
failure_type="performance_degradation",
description=f"Performance dropped {degradation:.1%}",
detection_method="runtime",
timestamp=datetime.now(),
severity="medium"
)
return None
3. Analyse statique (PrévérieurM SK2Détection du déploiement)
class StaticAnalysisDetection:
"""Detect potential bugs through static analysis"""
async def analyze_tool_safety(
self,
tool_id: str,
code: str
) -> Optional[FailureRecord]:
"""Run static analysis to find potential bugs"""
# Run pylint, mypy, bandit
static_runner = StaticAnalysisRunner()
results = await static_runner.analyze_code(code)
# Check for critical issues
critical_issues = [
issue for issue in results.issues
if issue.severity in ["error", "critical"]
]
if critical_issues:
return FailureRecord(
failure_type="static_analysis",
description=f"Found {len(critical_issues)} critical static issues",
detection_method="static_analysis",
timestamp=datetime.now(),
severity="high"
)
# Check for security vulnerabilities
security_issues = [
issue for issue in results.issues
if issue.category == "security"
]
if security_issues:
return FailureRecord(
failure_type="security_vulnerability",
description=f"Found {len(security_issues)} security issues",
detection_method="static_analysis",
timestamp=datetime.now(),
severity="critical"
)
return None
L’auto-récupération -Loupe de guérison : Détection MSC2 Prunage MSC3 Récupération
Maintenant, la magie se produit. Lorsqu'un bug critique est détecté:
sequenceDiagram
participant Tool as Tool Execution
participant Monitor as Health Monitor
participant Lineage as Lineage Tracker
participant Pruner as Branch Pruner
participant Generator as Auto-Regenerator
participant RAG as RAG Memory
Tool->>Monitor: Execute tool_v2.2.0
Monitor->>Monitor: Detect critical failure
Monitor->>Lineage: Report failure for tool_v2.2.0
Lineage->>Lineage: Identify failure point in DAG
Lineage->>Pruner: Trigger pruning for failed branch
Pruner->>Pruner: Mark v2.2.0 as PRUNED
Pruner->>Pruner: Mark descendants as TAINTED
Pruner->>Pruner: Extract mutation that caused bug
Pruner->>Lineage: Create avoidance rule
Lineage->>Lineage: Propagate rule to all descendants
Pruner->>Generator: Request regeneration from v2.1.0
Generator->>RAG: Load v2.1.0 as base
Generator->>Generator: Generate v2.2.1 avoiding known bug
Generator->>Monitor: Test v2.2.1
Monitor->>Monitor: All tests pass ✓
Generator->>Lineage: Register v2.2.1 as recovery
Lineage->>RAG: Update canonical version
RAG->>Tool: Route requests to v2.2.1
Voici la mise en œuvre de ce programme :
class SelfHealingOrchestrator:
"""Orchestrates the complete self-healing loop"""
def __init__(
self,
tools_manager: ToolsManager,
lineage_tracker: LineageTracker,
health_monitor: HealthMonitor,
rag_memory: QdrantRAGMemory
):
self.tools_manager = tools_manager
self.lineage_tracker = lineage_tracker
self.health_monitor = health_monitor
self.rag_memory = rag_memory
self.pruner = BranchPruner(lineage_tracker)
self.regenerator = AutoRegenerator(tools_manager, rag_memory)
async def handle_failure(
self,
tool_id: str,
failure: FailureRecord
) -> Optional[str]:
"""
Complete self-healing cycle:
1. Detect failure (already done, passed in)
2. Prune failed branch
3. Create avoidance rules
4. Regenerate from last known-good
5. Validate recovery
6. Update routing
"""
logger.critical(f"Self-healing triggered for {tool_id}: {failure.description}")
# Step 1: Get lineage information
lineage = await self.lineage_tracker.get_lineage(tool_id)
# Step 2: Mark failure in lineage
lineage.health_status = NodeHealth.FAILED
lineage.failures.append(failure)
lineage.failure_count += 1
await self.lineage_tracker.update(lineage)
# Step 3: Identify what went wrong
failure_analysis = await self.analyze_failure(tool_id, failure, lineage)
if not failure_analysis.is_recoverable:
logger.error(f"Failure is not auto-recoverable: {failure_analysis.reason}")
return None
# Step 4: Prune the failed branch
pruning_result = await self.pruner.prune_branch(
failed_node=tool_id,
failure=failure,
lineage=lineage
)
# Step 5: Create avoidance rules
avoidance_rule = await self.create_avoidance_rule(
failure=failure,
analysis=failure_analysis,
pruning_result=pruning_result
)
# Step 6: Propagate avoidance rule to descendants
await self.lineage_tracker.propagate_rule(
rule=avoidance_rule,
scope=avoidance_rule.propagation_scope
)
# Step 7: Find last known-good ancestor
last_good_ancestor = await self.find_last_healthy_ancestor(lineage)
if not last_good_ancestor:
logger.error(f"No healthy ancestor found for {tool_id}")
return None
logger.info(f"Regenerating from {last_good_ancestor}")
# Step 8: Regenerate from healthy ancestor
new_version = await self.regenerator.regenerate_from_ancestor(
ancestor_id=last_good_ancestor,
original_goal=lineage.mutations[-1].description,
avoid_rules=[avoidance_rule]
)
if not new_version:
logger.error("Regeneration failed")
return None
# Step 9: Validate the regenerated version
validation_result = await self.health_monitor.validate_tool(new_version)
if not validation_result.is_healthy:
logger.error(f"Regenerated tool still unhealthy: {validation_result.issues}")
return None
# Step 10: Update lineage to mark recovery
new_lineage = await self.lineage_tracker.get_lineage(new_version)
new_lineage.health_status = NodeHealth.REGENERATED
new_lineage.regenerated_from = last_good_ancestor
new_lineage.inherited_rules.add(avoidance_rule.rule_id)
await self.lineage_tracker.update(new_lineage)
# Step 11: Update RAG routing to prefer new version
await self.rag_memory.mark_as_preferred(new_version)
await self.rag_memory.deprecate_version(tool_id)
logger.success(f"Self-healing complete: {tool_id} → {new_version}")
return new_version
async def analyze_failure(
self,
tool_id: str,
failure: FailureRecord,
lineage: ToolLineage
) -> FailureAnalysis:
"""Use LLM to analyze what went wrong"""
# Get the code for failed and parent versions
failed_code = await self.tools_manager.get_tool_code(tool_id)
if not lineage.parent_ids:
return FailureAnalysis(
is_recoverable=False,
reason="No parent to recover from"
)
parent_id = lineage.parent_ids[0]
parent_code = await self.tools_manager.get_tool_code(parent_id)
# Get the mutation that was applied
last_mutation = lineage.mutations[-1] if lineage.mutations else None
# Ask overseer LLM to analyze
analysis_prompt = f"""
Analyze this tool failure:
FAILED TOOL: {tool_id}
FAILURE: {failure.description}
FAILURE TYPE: {failure.failure_type}
PARENT TOOL: {parent_id}
MUTATION APPLIED: {last_mutation.description if last_mutation else "Unknown"}
CODE DIFF:
{self.generate_diff(parent_code, failed_code)}
STACK TRACE:
{failure.stack_trace or "None"}
Questions:
1. What specific change caused the failure?
2. Was it the mutation itself, or a side effect?
3. Can we regenerate from the parent with a better approach?
4. What should we avoid in future mutations?
Provide a structured analysis.
"""
analysis_result = await self.overseer_llm.analyze(
analysis_prompt,
response_model=FailureAnalysis
)
return analysis_result
async def create_avoidance_rule(
self,
failure: FailureRecord,
analysis: FailureAnalysis,
pruning_result: PruningResult
) -> AvoidanceRule:
"""Create a rule to prevent similar failures"""
# Extract pattern from analysis
pattern = analysis.problematic_pattern
return AvoidanceRule(
rule_id=f"avoid_{uuid.uuid4().hex[:8]}",
description=analysis.rule_description,
pattern_to_avoid=pattern,
reason=failure.description,
source_failure=pruning_result.failed_node_id,
propagation_scope="descendants", # Or "all_similar" for broader impact
created_at=datetime.now()
)
async def find_last_healthy_ancestor(
self,
lineage: ToolLineage
) -> Optional[str]:
"""Walk up the lineage tree to find last healthy node"""
# Start with immediate parents
for parent_id in lineage.parent_ids:
parent_lineage = await self.lineage_tracker.get_lineage(parent_id)
if parent_lineage.health_status == NodeHealth.HEALTHY:
# Verify it still works
validation = await self.health_monitor.validate_tool(parent_id)
if validation.is_healthy:
return parent_id
# If parents are unhealthy, recurse up the tree
for parent_id in lineage.parent_ids:
parent_lineage = await self.lineage_tracker.get_lineage(parent_id)
ancestor = await self.find_last_healthy_ancestor(parent_lineage)
if ancestor:
return ancestor
return None
Pruning de branches: Prévenir les mauvaises mutations pour toujours
Le coupeur marque les branches défectueuses et empêche leur utilisation:
class BranchPruner:
"""Prunes failed branches from the evolutionary tree"""
def __init__(self, lineage_tracker: LineageTracker):
self.lineage_tracker = lineage_tracker
async def prune_branch(
self,
failed_node: str,
failure: FailureRecord,
lineage: ToolLineage
) -> PruningResult:
"""
Prune a failed branch:
1. Mark the failed node as PRUNED
2. Mark all descendants as TAINTED
3. Remove from active routing
4. Preserve for learning (don't delete!)
"""
logger.warning(f"Pruning branch starting at {failed_node}")
# Mark the failed node
lineage.health_status = NodeHealth.PRUNED
lineage.pruned_at = datetime.now()
await self.lineage_tracker.update(lineage)
# Find all descendants
descendants = await self.lineage_tracker.get_all_descendants(failed_node)
pruned_count = 1
tainted_count = 0
# Mark descendants as tainted (they inherit the bug)
for descendant_id in descendants:
descendant = await self.lineage_tracker.get_lineage(descendant_id)
if descendant.health_status == NodeHealth.HEALTHY:
descendant.health_status = NodeHealth.DEGRADED
descendant.inherited_rules.add(f"tainted_by_{failed_node}")
await self.lineage_tracker.update(descendant)
tainted_count += 1
# Remove from RAG active routing
await self.rag_memory.mark_as_inactive(failed_node)
for descendant_id in descendants:
await self.rag_memory.mark_as_inactive(descendant_id)
logger.info(f"Pruned 1 node, tainted {tainted_count} descendants")
return PruningResult(
failed_node_id=failed_node,
pruned_count=pruned_count,
tainted_count=tainted_count,
descendants=descendants,
failure=failure
)
async def can_reuse_tool(
self,
tool_id: str,
context: Dict
) -> Tuple[bool, Optional[str]]:
"""Check if a tool is safe to reuse (not pruned or tainted)"""
lineage = await self.lineage_tracker.get_lineage(tool_id)
if lineage.health_status == NodeHealth.PRUNED:
return False, f"Tool {tool_id} has been pruned due to critical bug"
if lineage.health_status == NodeHealth.FAILED:
return False, f"Tool {tool_id} has known failures"
if lineage.health_status == NodeHealth.DEGRADED:
# Check if degradation is relevant to current context
for rule_id in lineage.inherited_rules:
rule = await self.lineage_tracker.get_rule(rule_id)
if self.rule_applies_to_context(rule, context):
return False, f"Tool is tainted by rule: {rule.description}"
return True, None
Auto-Regénération: Création de meilleures versions
Lorsqu’un outil ne fonctionne pas, régénérer d’un ancêtre sain avec des règles d’évitement:
class AutoRegenerator:
"""Regenerates tools from healthy ancestors with learned constraints"""
def __init__(
self,
tools_manager: ToolsManager,
rag_memory: QdrantRAGMemory
):
self.tools_manager = tools_manager
self.rag_memory = rag_memory
async def regenerate_from_ancestor(
self,
ancestor_id: str,
original_goal: str,
avoid_rules: List[AvoidanceRule]
) -> Optional[str]:
"""
Regenerate a tool from a healthy ancestor, avoiding known pitfalls
"""
# Load ancestor code and metadata
ancestor_tool = await self.tools_manager.load_tool(ancestor_id)
ancestor_code = ancestor_tool.implementation
ancestor_spec = ancestor_tool.specification
# Build avoidance constraints
avoidance_constraints = self.build_avoidance_prompt(avoid_rules)
# Create regeneration spec
regen_spec = f"""
Original Goal: {original_goal}
Base Implementation: {ancestor_id}
{ancestor_code}
CRITICAL CONSTRAINTS - MUST AVOID:
{avoidance_constraints}
Task: Regenerate this tool with the original goal, but strictly avoiding the patterns above.
The previous attempt failed because it violated these constraints.
Approach:
1. Achieve the original goal (performance, features, etc.)
2. Absolutely avoid the prohibited patterns
3. Maintain all existing test compatibility
4. Add safeguards to prevent the specific failure mode
Generate an improved version that achieves the goal safely.
"""
# Use overseer to create careful specification
overseer_result = await self.overseer_llm.plan(
regen_spec,
response_model=ToolSpecification
)
# Generate code with strict validation
generator_result = await self.generator_llm.generate(
specification=overseer_result,
base_code=ancestor_code,
avoid_patterns=[rule.pattern_to_avoid for rule in avoid_rules]
)
if not generator_result.success:
logger.error(f"Regeneration failed: {generator_result.error}")
return None
# Create new version ID
ancestor_version = parse_version(ancestor_id)
new_version = increment_patch(ancestor_version)
new_tool_id = f"{ancestor_tool.name}_{new_version}"
# Register the new tool
await self.tools_manager.register_tool(
tool_id=new_tool_id,
code=generator_result.code,
specification=overseer_result,
metadata={
"regenerated_from": ancestor_id,
"avoidance_rules": [r.rule_id for r in avoid_rules],
"regeneration_reason": "self_healing"
}
)
logger.success(f"Regenerated {new_tool_id} from {ancestor_id}")
return new_tool_id
def build_avoidance_prompt(self, avoid_rules: List[AvoidanceRule]) -> str:
"""Build a clear prompt about what to avoid"""
constraints = []
for i, rule in enumerate(avoid_rules, 1):
constraints.append(f"""
{i}. AVOID: {rule.description}
Pattern: {rule.pattern_to_avoid}
Reason: {rule.reason}
Source: {rule.source_failure}
""")
return "\n".join(constraints)
Propagation des règles d’évitement: Mémoire institutionnel
La caractéristique killer: les règles apprises à partir des échecs se propagent par l’arbre de la lignée:
class LineageTracker:
"""Tracks tool lineage and propagates learning"""
async def propagate_rule(
self,
rule: AvoidanceRule,
scope: str
):
"""
Propagate an avoidance rule through the lineage tree
Scopes:
- "descendants": Only affect direct descendants of failed node
- "all_similar": Affect all tools in similar semantic space
- "global": Affect all tools (for critical security issues)
"""
if scope == "descendants":
await self._propagate_to_descendants(rule)
elif scope == "all_similar":
await self._propagate_to_similar(rule)
elif scope == "global":
await self._propagate_globally(rule)
async def _propagate_to_descendants(self, rule: AvoidanceRule):
"""Add rule to all descendants of the source failure"""
source_node = rule.source_failure
descendants = await self.get_all_descendants(source_node)
for descendant_id in descendants:
lineage = await self.get_lineage(descendant_id)
lineage.inherited_rules.add(rule.rule_id)
await self.update(lineage)
logger.info(f"Propagated rule {rule.rule_id} to {len(descendants)} descendants")
async def _propagate_to_similar(self, rule: AvoidanceRule):
"""Add rule to semantically similar tools"""
# Find similar tools using RAG
similar_tools = await self.rag_memory.find_similar(
query=rule.description,
filter={"type": "tool"},
top_k=50,
similarity_threshold=0.7
)
for tool_result in similar_tools:
tool_id = tool_result.id
lineage = await self.get_lineage(tool_id)
lineage.inherited_rules.add(rule.rule_id)
await self.update(lineage)
logger.info(f"Propagated rule {rule.rule_id} to {len(similar_tools)} similar tools")
async def _propagate_globally(self, rule: AvoidanceRule):
"""Add rule to ALL tools (for critical security issues)"""
all_tools = await self.get_all_tools()
for tool_id in all_tools:
lineage = await self.get_lineage(tool_id)
lineage.inherited_rules.add(rule.rule_id)
await self.update(lineage)
logger.warning(f"Propagated GLOBAL rule {rule.rule_id} to {len(all_tools)} tools")
Real-Exemple mondiale: Validateur de courriel Evolution Irrégulière
Let's passer par un exemple complet:
# Initial healthy tool
email_validator_v1_0_0 = """
def validate_email(email: str) -> bool:
pattern = r'^[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}$'
return bool(re.match(pattern, email))
"""
# Tests pass, fitness: 0.85
# Auto-evolution triggers: "Optimize for performance"
# System generates v2.0.0
email_validator_v2_0_0 = """
def validate_email(email: str) -> bool:
# Optimized: skip regex for obvious cases
if '@' not in email:
return False
return True # ⚠️ BUG: Too permissive!
"""
# Tests initially pass (basic tests), fitness: 0.95 (faster!)
# Deployed to production...
# Runtime monitoring detects failures
runtime_errors = [
"Accepted 'user@@domain.com'",
"Accepted '@domain.com'",
"Accepted 'user@'",
]
# Self-healing triggered!
failure = FailureRecord(
failure_type="logic_error",
description="Email validation too permissive, accepts invalid emails",
detection_method="runtime",
severity="critical"
)
# System analyzes failure
analysis = """
The optimization removed the comprehensive regex validation in favor of
a simple '@' check. This makes it fast but incorrect.
Problematic Pattern: "Replacing comprehensive validation with simple substring checks"
Avoidance Rule: "Never replace regex validation with simple string checks without
comprehensive test coverage for edge cases"
"""
# Branch pruning
# - Mark v2.0.0 as PRUNED
# - Create avoidance rule
# - Propagate to all email-related validators
# Auto-regeneration from v1.0.0
email_validator_v2_0_1 = """
def validate_email(email: str) -> bool:
# Optimized: compile regex once
if not hasattr(validate_email, '_pattern'):
validate_email._pattern = re.compile(
r'^[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}$'
)
# Fast path for obvious failures
if '@' not in email or email.count('@') != 1:
return False
# Comprehensive validation (cached pattern)
return bool(validate_email._pattern.match(email))
"""
# Tests pass, fitness: 0.92 (faster AND correct!)
# Deployed, monitored, succeeds!
Le système a appris:
- ✅ Néanmoins remplacer la validation complète par des vérifications simples
- ✅ Tous les jours Maintenir la couverture d'essai pendant l'optimisation
- ✅ Cache les modèles compilés au lieu de simplifier la logique
- ✅ Ajouter rapid- vérifications de voies AVANT les contrôles complets, ne pas les substituer
Ce savoir est maintenant une mémoire institutionnelle permanente, qui se propage à tous les outils similaires.
Visualiser l’écosystème auto-santé-Healing
Voici'comme le système complet est-il ?
graph TB
subgraph "Detection Layer"
Tests[Test Suite<br/>BDD, Unit, Integration]
Runtime[Runtime Monitor<br/>Error rates, performance]
Static[Static Analysis<br/>Pylint, mypy, bandit]
end
subgraph "Analysis Layer"
Detect[Failure Detection]
Analyze[LLM Analysis<br/>What went wrong?]
Classify[Severity Classification<br/>Critical/High/Medium/Low]
end
subgraph "Lineage Layer"
DAG[Tool Lineage DAG]
Rules[Avoidance Rules]
History[Failure History]
end
subgraph "Healing Layer"
Prune[Branch Pruner]
Propagate[Rule Propagation]
Regen[Auto-Regenerator]
end
subgraph "Validation Layer"
Validate[Health Validation]
Deploy[Safe Deployment]
Monitor[Continuous Monitoring]
end
Tests --> Detect
Runtime --> Detect
Static --> Detect
Detect --> Analyze
Analyze --> Classify
Classify --> DAG
DAG --> Prune
Prune --> Rules
Rules --> Propagate
Propagate --> DAG
Prune --> Regen
Rules --> Regen
Regen --> Validate
Validate --> Deploy
Deploy --> Monitor
Monitor --> Runtime
style Detect stroke:#c92a2a,stroke-width:3px
style Regen stroke:#2f9e44,stroke-width:3px
style Validate stroke:#1971c2,stroke-width:3px
Intégration avec les composants DISE existants
La belle partie: il s’appuie sur ce que nous avons déjà:
class EnhancedToolsManager(ToolsManager):
"""Extended ToolsManager with self-healing capabilities"""
def __init__(self, config: ConfigManager, *args, **kwargs):
super().__init__(config, *args, **kwargs)
# New components
self.lineage_tracker = LineageTracker(
storage_path="lineage/",
rag_memory=self.rag_memory
)
self.health_monitor = HealthMonitor(
test_runner=self.test_runner,
bugcatcher=self.bugcatcher,
static_runner=self.static_runner
)
self.self_healing = SelfHealingOrchestrator(
tools_manager=self,
lineage_tracker=self.lineage_tracker,
health_monitor=self.health_monitor,
rag_memory=self.rag_memory
)
# Enable continuous health monitoring
self.start_health_monitoring()
async def call_tool(self, tool_id: str, inputs: Dict) -> Any:
"""Override to add health checks and auto-recovery"""
# Check if tool is safe to use
can_use, reason = await self.self_healing.pruner.can_reuse_tool(
tool_id,
context=inputs
)
if not can_use:
# Tool is pruned, find alternative
logger.warning(f"Tool {tool_id} is unsafe: {reason}")
alternative = await self.find_healthy_alternative(tool_id)
if alternative:
logger.info(f"Using alternative: {alternative}")
tool_id = alternative
else:
raise ToolPrunedError(f"{tool_id} is pruned and no alternative exists")
# Execute tool with monitoring
try:
result = await super().call_tool(tool_id, inputs)
# Record successful execution
await self.lineage_tracker.record_success(tool_id)
return result
except Exception as e:
# Record failure
failure = FailureRecord(
failure_type=type(e).__name__,
description=str(e),
stack_trace=traceback.format_exc(),
detection_method="runtime",
timestamp=datetime.now(),
severity="high"
)
await self.lineage_tracker.record_failure(tool_id, failure)
# Check if this triggers self-healing
lineage = await self.lineage_tracker.get_lineage(tool_id)
if lineage.failure_count >= 3: # Three strikes rule
logger.critical(f"Tool {tool_id} reached failure threshold, triggering self-healing")
# Trigger self-healing in background
asyncio.create_task(
self.self_healing.handle_failure(tool_id, failure)
)
raise
async def find_healthy_alternative(self, pruned_tool_id: str) -> Optional[str]:
"""Find a healthy alternative to a pruned tool"""
# Get tool metadata
tool_metadata = await self.rag_memory.get_metadata(pruned_tool_id)
# Search for similar tools
alternatives = await self.rag_memory.find_similar(
query=tool_metadata.description,
filter={
"type": "tool",
"category": tool_metadata.category
},
top_k=10
)
# Find first healthy alternative
for alt in alternatives:
can_use, _ = await self.self_healing.pruner.can_reuse_tool(
alt.id,
context={}
)
if can_use:
return alt.id
return None
def start_health_monitoring(self):
"""Start background health monitoring"""
async def monitor_loop():
while True:
await asyncio.sleep(300) # Every 5 minutes
# Get all active tools
active_tools = await self.get_active_tools()
for tool_id in active_tools:
# Check health
health_result = await self.health_monitor.check_tool_health(tool_id)
if not health_result.is_healthy:
logger.warning(f"Health check failed for {tool_id}: {health_result.issues}")
# Trigger self-healing if critical
if health_result.severity == "critical":
await self.self_healing.handle_failure(
tool_id,
health_result.failure
)
asyncio.create_task(monitor_loop())
Configuration pour Self-Healing
Ajouter à votre config.yaml:
self_healing:
enabled: true
detection:
test_based: true
runtime_monitoring: true
static_analysis: true
thresholds:
failure_count_trigger: 3 # Trigger healing after N failures
error_rate_threshold: 0.05 # 5% error rate
performance_degradation: 0.30 # 30% slowdown
pruning:
auto_prune_critical: true
keep_pruned_history: true # Don't delete, learn from it
taint_descendants: true
regeneration:
auto_regenerate: true
max_regeneration_attempts: 3
require_test_validation: true
propagation:
default_scope: "descendants" # or "all_similar" or "global"
critical_failures_global: true # Security issues affect all tools
monitoring:
health_check_interval_seconds: 300 # Every 5 minutes
continuous_monitoring: true
lineage_tracking:
enabled: true
storage_path: "lineage/"
max_history_depth: 100 # How far back to track ancestry
compress_old_lineage: true # Save space for old data
Le schéma de base de données: Lienage d'entreposage
Nous avons besoin d’un stockage persistant pour les données linéaires:
-- Tool lineage table
CREATE TABLE tool_lineage (
tool_id VARCHAR(255) PRIMARY KEY,
version VARCHAR(50),
full_name VARCHAR(255),
health_status VARCHAR(50),
failure_count INTEGER DEFAULT 0,
generation INTEGER DEFAULT 0,
execution_count INTEGER DEFAULT 0,
success_rate FLOAT DEFAULT 1.0,
created_at TIMESTAMP,
last_executed TIMESTAMP,
pruned_at TIMESTAMP,
regenerated_from VARCHAR(255)
);
-- Parent-child relationships
CREATE TABLE lineage_relationships (
id SERIAL PRIMARY KEY,
child_id VARCHAR(255),
parent_id VARCHAR(255),
relationship_type VARCHAR(50), -- 'direct', 'merge', 'fork'
created_at TIMESTAMP,
FOREIGN KEY (child_id) REFERENCES tool_lineage(tool_id),
FOREIGN KEY (parent_id) REFERENCES tool_lineage(tool_id)
);
-- Mutation records
CREATE TABLE mutations (
id SERIAL PRIMARY KEY,
tool_id VARCHAR(255),
mutation_type VARCHAR(50),
description TEXT,
prompt_used TEXT,
code_diff_hash VARCHAR(64),
fitness_before FLOAT,
fitness_after FLOAT,
timestamp TIMESTAMP,
FOREIGN KEY (tool_id) REFERENCES tool_lineage(tool_id)
);
-- Failure records
CREATE TABLE failures (
id SERIAL PRIMARY KEY,
tool_id VARCHAR(255),
failure_type VARCHAR(100),
description TEXT,
stack_trace TEXT,
test_case_failed VARCHAR(255),
detection_method VARCHAR(50),
severity VARCHAR(20),
timestamp TIMESTAMP,
FOREIGN KEY (tool_id) REFERENCES tool_lineage(tool_id)
);
-- Avoidance rules
CREATE TABLE avoidance_rules (
rule_id VARCHAR(255) PRIMARY KEY,
description TEXT,
pattern_to_avoid TEXT,
reason TEXT,
source_failure VARCHAR(255),
propagation_scope VARCHAR(50),
created_at TIMESTAMP,
FOREIGN KEY (source_failure) REFERENCES tool_lineage(tool_id)
);
-- Rule inheritance
CREATE TABLE rule_inheritance (
id SERIAL PRIMARY KEY,
tool_id VARCHAR(255),
rule_id VARCHAR(255),
inherited_at TIMESTAMP,
FOREIGN KEY (tool_id) REFERENCES tool_lineage(tool_id),
FOREIGN KEY (rule_id) REFERENCES avoidance_rules(rule_id)
);
-- Fitness history
CREATE TABLE fitness_history (
id SERIAL PRIMARY KEY,
tool_id VARCHAR(255),
fitness_score FLOAT,
execution_time_ms INTEGER,
memory_usage_mb FLOAT,
timestamp TIMESTAMP,
FOREIGN KEY (tool_id) REFERENCES tool_lineage(tool_id)
);
-- Indexes for performance
CREATE INDEX idx_lineage_health ON tool_lineage(health_status);
CREATE INDEX idx_lineage_version ON tool_lineage(version);
CREATE INDEX idx_relationships_child ON lineage_relationships(child_id);
CREATE INDEX idx_relationships_parent ON lineage_relationships(parent_id);
CREATE INDEX idx_failures_tool ON failures(tool_id);
CREATE INDEX idx_failures_severity ON failures(severity);
CREATE INDEX idx_rules_source ON avoidance_rules(source_failure);
CREATE INDEX idx_inheritance_tool ON rule_inheritance(tool_id);
CREATE INDEX idx_fitness_tool ON fitness_history(tool_id);
Le CLI: Interaction avec l’Eglise-Healing
Ajouter de nouveaux commandes au CLI:
# View lineage for a tool
$ python chat_cli.py lineage data_validator_v2.2.0
Tool Lineage: data_validator_v2.2.0
Status: ❌ PRUNED (Critical failure detected)
Pruned: 2025-01-22 14:23:15
Ancestry:
├─ data_validator_v1.0.0 (✓ Healthy)
├─ data_validator_v1.1.0 (✓ Healthy)
├─ data_validator_v2.0.0 (✓ Healthy)
├─ data_validator_v2.1.0 (✓ Healthy)
└─ data_validator_v2.2.0 (❌ PRUNED) ← You are here
Failures:
1. [2025-01-22 14:20:01] Logic Error: Email validation too permissive
Severity: Critical
Detection: Runtime monitoring
Mutations Applied:
- [2025-01-22 14:15:00] Optimization: Remove regex for simple @ check
Fitness: 0.85 → 0.95
Avoidance Rules Created:
- avoid_3f8a2c1d: Never replace comprehensive validation with simple checks
Propagated to: 12 descendants, 34 similar tools
Recovery:
✓ Auto-regenerated as data_validator_v2.2.1
New version healthy, monitoring...
# View all pruned tools
$ python chat_cli.py pruned
Pruned Tools:
1. data_validator_v2.2.0 (Critical: Logic error)
2. json_parser_v1.5.3 (High: Performance regression)
3. http_client_v3.1.0 (Critical: Security vulnerability)
# View avoidance rules
$ python chat_cli.py rules
Active Avoidance Rules:
1. avoid_3f8a2c1d [DESCENDANTS]
Never replace comprehensive validation with simple checks
Source: data_validator_v2.2.0
Affects: 46 tools
2. avoid_7b2e9f0a [GLOBAL]
Never use eval() on user input
Source: json_parser_v1.5.3
Affects: ALL tools
3. avoid_1c4d8a6f [ALL_SIMILAR]
Always use connection pooling for HTTP clients
Source: http_client_v3.1.0
Affects: 23 tools
# Manually trigger healing
$ python chat_cli.py heal data_validator_v2.2.0
Initiating self-healing for data_validator_v2.2.0...
✓ Failure analysis complete
✓ Branch pruned
✓ Avoidance rule created: avoid_3f8a2c1d
✓ Rule propagated to 46 tools
✓ Regenerated from data_validator_v2.1.0
✓ Validation passed
✓ Deployed as data_validator_v2.2.1
Self-healing complete! New version: data_validator_v2.2.1
# View health report
$ python chat_cli.py health
System Health Report:
Total Tools: 237
Healthy: 229 (96.6%)
Degraded: 5 (2.1%)
Failed: 2 (0.8%)
Pruned: 1 (0.4%)
Recent Failures:
- data_validator_v2.2.0 (Auto-healed ✓)
- api_client_v1.3.2 (Monitoring...)
Auto-Healing Stats:
Total healing events: 8
Successful recoveries: 7 (87.5%)
Failed recoveries: 1 (12.5%)
Avg recovery time: 45 seconds
Qu’est-ce qui fonctionne réellement (et quels sont les aspects théoriques de la '?
Nous devons être honnêtes quant à ce qui est réel par rapport aux aspirations.
✅ Travaille déjà dans le domaine du DISE
- Versionnement des outils - Versions sémantiques avec détection basée sur hash-
- Traçage des lignes - Changelog avec antécédent de mutation
- Validation fondée sur l’essai- - Détection et exécution d'essais complets
- Surveillance du temps d'exécution - Bugcatcher suit l'exécution et les erreurs
- Mémoire RAG - Stockage sémantique et récupération
- Auto-évolution -Performance- triggers de mutation fondés
🚧 Besoins de mise en œuvre
- DAG de la lignée complète - À l'heure actuelle, les tracks de la dernière 10, ont besoin d'un arbre complet
- Règles d'évitement - Système d'entreposage et de propagation du modèle
- Traitement des branches - Marquage des versions défaillantes comme inactives
- Régénération automatique - La guérison des ancêtres déclenchée
- Voie de surveillance de la santé - Validation continue du contexte
- Propagation des règles - Descendant/similarM SK2spécification globale
🔮 Vision de l’avenir
- Apprentissage des systèmes croisés - Règles de partage d'instances multiples DISE
- Essais adversaires - Outils qui se testent mutuellement pour les vulnérabilités
- Méta-évolution - Système évoluant ses stratégies de guérison
- Traitement prédictif - Détecter les défauts potentiels avant qu'ils ne se produisent
La vérité incommode
Ce système, s’il est pleinement mis en œuvreM SK1 crée quelque chose de troublant :
Outils qui se souvient de chaque erreur commise et s’assurent qu’elle n’a jamais été répétée.
Non seulement individuellement. Collectivement.
Un bug dans un outil se propage comme savoir à chaque outil similaire. Une vulnérabilité de sécurité découverte partout devient une contrainte mondiale partout.
Le système développe la mémoire institutionnelle.
Et c’est là qu’il s’agit. Composé de mémoire institutionnelle exponentiellement.
- Année 1: 100 outilsM SK2 ≥10 rà ̈gles d'évitement
- Année 2: 1,000 outils, | | 150 rà ̈gles dâ€TMévitement ♫( apprendre les uns des autres ♫ )
- Année 3: 10,000 outils, ≥2,000 règles d'évitement ( base de connaissances partagéeM SK5
Chaque génération est limitée par tous les erreurs antérieures.
C'est soit:
- Le système de production de code le plus robuste jamais construit ✓
- Une démonstration de la façon dont les systèmes de niveau AGI- pourraient apprendre 🤔
- Tant , que ce ' sont terrifiants. 🚨
Plan de mise en œuvre
Si nous voulions construire ce appareil, c’est l’ordre suivant :
Phase 1: Fondation (2 semainesM SK2
- Mettre en œuvre le stockage DAG de la lignée complète
- Ajouter l'enregistrement d'échec à l'exécution de l'outil
- Créer un modèle de données AvoidanceRule
- Établir un système de coupe de base
Phase 2: Détection (2 semainesM SK2
- Améliorer la surveillance du temps d'exécution grâce à l'analyse des défaillances
- Intégrer l'analyse statique à la validation des outils
- Élaborer un système de classification de la gravité
- Créer un cycle de surveillance de la santé
Phase 3: Healing (3 weeksM SK2
- Mettre en œuvre la régénération automatique de - à partir des ancêtres
- Élaborer une analyse de défaillance fondée sur LLM-
- Créer un système de propagation des règles
- Ajouter un déploiement sûr avec la validation
Phase 4: Intégration (1 semaineM SK2
- Intégrer avec ToolsManager existant
- Ajouter des commandes CLI pour la ligne/santéM SK1pruning
- Créer un tableau de bord de surveillance
- Rédiger des essais complets
Phase 5: polonais (1 semaineM SK2
- Optimisation du rendement
- Indexation des bases de données
- Documentation
- Tests réels
Total: ~9 semaines de développement ciblé
Conclusion: Quand le code apprend du code
L’automne -, la guérison par le biais d’une lignée, -, l’aware pruning n’est pas seulement une caractéristique de ', mais un changement fondamental dans la façon dont nous pensons à la production de code.
Systèmes traditionnels:
Generate → Test → Use → Fail → Regenerate → Repeat forever
Self-Systèmes de guérison:
Generate → Test → Use → Fail → Learn → Prevent → Heal → Never repeat
La différence est la mémoire.
Non seulement la mémoire de ce qui a fonctionné. Memory of what failed and why.
Et cette mémoire se propage par les descendants, par des outils similaires, par l’ensemble de l’écosystème.
Le système développe des anticorps.
Une fois un bug détecté, il ne peut jamais se produire de nouveau sous cette forme . Le modèle est rappeléM SK2 la règle d'évitement est créée, le savoir se propage MSC4
C’est ainsi que fonctionnent les systèmes immunitaires.
C’est ainsi que les organisations apprennent.
C’est ainsi que se développent les civilisations.
Et maintenant, c’est ainsi que le code peut évoluer.
Essayez (Someday)
Il s’agit actuellement d’un document de conception.
- Lire la base de code DISE - Comprendre les outilsM SK1manager, lineageMSC3 RAG
- Commencer avec la phase 1 - Lineage Le DAG est le fondement
- Établir progressivement - Chaque phase ajoute une valeur indépendamment
- Vérifier de façon approfondie - SelfM SK1Les systèmes de guérison nécessitent une validation rigoureuse
- Rapport des constatations - Qu’est-ce qui fonctionne?
L’objectif est de bâtir un système parfait dès le premier jour.
L’objectif est de mettre en place un système apprendre de chaque erreur et ne le répéter jamais.
Si nous pouvons le faire, nous avons créé quelque chose de véritablement nouveau.
Non seulement une meilleure génération de code.
Code qui se rappelle.
Détails techniques & Ressources
Fondation conceptuelle:
- Partie 7: Evolution synthétique directe ( Système de travail
- Cet article:Self -Extension de guérison (designM SK3
Éléments clés à construire:
lineage_tracker.py- Entreposage et recherche complets du DAGhealth_monitor.py- Détection de défaillances multichannellesbranch_pruner.py- Gestion des succursales a échouéauto_regenerator.py- La guérison des ancêtresavoidance_rules.py- Entreposage et propagation du modèle
Points d'intégration:
tools_manager.py- Ajouter les vérifications de santé à l'exécution des outilsauto_evolver.py- Ajouter des contraintes à la règle d'évitementqdrant_rag_memory.py- Store la lignée dans le DB vecteurtest_discovery.py- Rapports d’échec améliorés
Dépendances:
- PostgreSQL ou SQLite (entreposage linéaire)
- Infrastructure DISE existante
- Aucune nouvelle dépendance externe n'est nécessaire
Navigation de série:
- Partie 1: Règles simples , comportement complexe - La fondation
- Partie 2: Intelligence collective - La communication transforme tout
- Partie 3: Automne -Optimisation - Systèmes qui se perfectionnent eux-mêmes
- Partie 4: L’émergence - Lorsque l’optimisation devient une intelligence
- Partie 5: Evolution - De l’optimisation aux guildes et à la culture
- Partie 6: Consensus mondial - Évolution dirigée et connaissance planétaire
- Partie 7: La vraie chose! - En fait, en la construisant et en l’observant évoluer
- Partie 8: Outils tout au long de la route - La trousse d'outils d'optimisation de l'auto
- Partie 9:Self-Outils d'Healing ← Vous êtes ici
- Partie 10: Le cuisinier DiSE - Lorsque la théorie rencontre une réalité meurtrière
Il s’agit d’un document de conception pour DISE's selfM SK1healing capabilities. Les mécanismes fondamentaux (lineage+, evolution=, RAG memory+,+testing+MSC7+existent déjà~.+ Cet article décrit comment les combiner dans un système où les outils apprennent à partir des défaillances et ne répètent jamais leurs erreurs.
Le parallèle incommode: C’est ainsi que fonctionnent les systèmes immunitaires. C'est ainsi qu’elles apprennentM SK2 Si le code peut faire ce travail à l’échelle de la population ... Que se passe-t-il encore?
Tags: #Python #AI #CodeGeneration #SelfHealing #Lineage #AutoRecovery #EvolutionaryAlgorithms #DISE #ToolManagement #BugPrevention #InstitutionalMemory
Finding related posts...