kaizen

name	kaizen
description	Kailash Kaizen (Rust) — MANDATORY for AI agents/signatures/RAG via kailash-kaizen crate. Custom LLM agents BLOCKED.

Kaizen Skills

AI agent framework for building intelligent agents with kailash-enterprise.

Quick Links

Skill	Description
kaizen-agent-patterns	Agent building blocks, tools, memory, known issues
kaizen-signatures	Input/output contracts (Signature, InputField)
kaizen-hooks-lifecycle	Lifecycle hooks (9 events via HookManager)
kaizen-checkpoint-resume	Save/restore agent state
kaizen-interrupt-mechanism	Timeouts, budgets, manual interrupts
kaizen-pipelines	Sequential, Parallel, Ensemble agent composition
kaizen-control-protocol	Human-in-the-loop approval workflows
kaizen-chain-of-thought	Chain-of-thought agent implementation
kaizen-react-pattern	ReAct (Reasoning + Acting) agent pattern
kaizen-rag-agent	RAG agent implementation
kaizen-cost-tracking	LLM cost tracking and budget management
kaizen-budget-tracking	Two-phase budget enforcement for tool agents
kaizen-trust-architecture	Three-layer trust stack map (EATP/Trust-Plane/Kaizen)
kaizen-streaming	TaodRunner Python binding, CallerEvent streaming
kaizen-a2a-protocol	Agent-to-Agent protocol (cards & registry)
kaizen-a2a	A2A messaging, discovery, and delegation
kaizen-checkpoint	Checkpoint/resume with known issues & interrupts
kaizen-quickstart	First agent in 5 minutes
kaizen-llm-providers	LlmClient, provider detection, mock testing
kaizen-memory	SessionMemory, SharedMemory, PersistentMemory
kaizen-tools	ToolDef, ToolRegistry, ToolParam, handler pattern
kaizen-orchestration	OrchestrationRuntime, strategies, AgentExecutor
create-agent	Scaffold a new agent (template skill)

Core API

from kailash.kaizen import BaseAgent, LlmClient
from kailash.kaizen import ToolRegistry, ToolDef, ToolParam
from kailash.kaizen import SessionMemory, SharedMemory
from kailash.kaizen import AgentCheckpoint, CostTracker
from kailash.kaizen import Signature, InputField, OutputField
from kailash.kaizen import HookManager, InterruptManager, ControlProtocol
from kailash.kaizen import AgentCard, AgentRegistry
from kailash.kaizen import TrustLevel, TrustPosture
from kailash.kaizen.pipelines import SequentialPipeline, ParallelPipeline, EnsemblePipeline
from kailash.kaizen.pipelines import RouterPipeline, SupervisorPipeline

Known Issue: BaseAgent.execute() raises NotImplementedError by default — override it in your subclass. Convenience methods run(), extract_str(), extract_dict(), and write_to_memory() are available on BaseAgent (added in P17-002).

Known Issue: Memory methods in .pyi are wrong -- use store()/recall()/remove(), not set()/get()/delete().

Known Issue: ToolDef uses handler= kwarg, not callback=.

Known Issue: Signature.validate_inputs() returns a dict of validated inputs with defaults filled in. Raises ValueError if a required field is missing.

Known Issue: Checkpoint storage load() raises RuntimeError when checkpoint not found (does NOT return None).

Known Issue: HookManager.EVENTS is a tuple, not a list.

Known Issue: AgentCheckpoint.memory_snapshot and tool_state are None by default, not empty dicts.

Known Issue: InputField default is None when omitted, not "".

Note: Streaming is available via StreamingAgent and StreamHandler (from kailash.kaizen import StreamingAgent, StreamHandler).

Phase 17 Additions

Skill	Description
kaizen-structured-output	StructuredOutput for typed agent responses
kaizen-multi-agent	SupervisorAgent and WorkerAgent coordination
kaizen-observability	ObservabilityManager and MetricsCollector

New Imports (Phase 17)

from kailash.kaizen import StructuredOutput
from kailash.kaizen import SupervisorAgent, WorkerAgent
from kailash.kaizen import ObservabilityManager, MetricsCollector

Ontology Module (feature: ontology)

Embedding-backed concept classification via OntologyRegistry. Feature-gated behind ontology in kailash-kaizen.

Type	Purpose
OntologyRegistry	Main registry: load YAML seeds, classify text
EmbeddingProvider	Trait for embedding models (OpenAI, local, mock)
MockEmbeddingProvider	Deterministic mock for testing
Namespace	Named concept collection with SHA-256 version hash
Concept	Classification target with optional exemplar texts
ClassificationResult	Top match with confidence score
ConceptMatch	Single concept match (label, similarity)
ConceptEvaluation	Batch evaluation result
EvaluationReport	Precision/recall/F1 evaluation harness output
OntologyConfig	Registry configuration

Key operations: classify(text), classify_batch(texts), cosine similarity (in-crate, no external deps), YAML seed parsing.

Source: crates/kailash-kaizen/src/ontology/ (6 files: mod.rs, registry.rs, provider.rs, similarity.rs, types.rs, error.rs)

Specialist

For complex queries beyond these skills, use the kaizen-specialist agent.