v0.2.0: Add purpose_agent/__init__.py

purpose_agent/__init__.py

@@ -1,50 +1,104 @@
 """
-purpose_agent — A Self-Improving Agentic Framework via State-Value Evaluation
-
-Architecture based on:
-  - MUSE (arxiv:2510.08002): 3-tier hierarchical memory (strategic/procedural/tool)
-  - LATS (arxiv:2310.04406): LLM-as-value-function V(s) = λ·LM_score + (1-λ)·SC_score
-  - REMEMBERER (arxiv:2306.07929): Q-value experience replay with Bellman updates
-  - Reflexion (arxiv:2303.11366): Verbal reinforcement via episodic self-reflection
-  - SPC (arxiv:2504.19162): Anti-reward-hacking via adversarial critic patterns
-
-Core philosophy: The agent improves via a "Purpose Function" Φ(s) that evaluates
-intermediate state improvements (distance to goal) rather than binary outcome success.
-No real-time backprop — improvement comes from expanding external memory with
-learned heuristics extracted from high-reward trajectories.
+purpose_agent — The World's First SLM-Native Self-Improving Agentic Framework
+
+Works with both Small Language Models (SLMs, 0.6B-3B params, local) and
+Large Language Models (LLMs, cloud APIs) with equal efficiency.
+
+Architecture based on 8 published papers:
+  - MUSE (arxiv:2510.08002): 3-tier hierarchical memory
+  - LATS (arxiv:2310.04406): LLM-as-value-function
+  - REMEMBERER (arxiv:2306.07929): Q-value experience replay
+  - Reflexion (arxiv:2303.11366): Verbal reinforcement
+  - SPC (arxiv:2504.19162): Anti-reward-hacking
+  - CER (arxiv:2506.06698): Contextual experience distillation
+  - MemRL (arxiv:2601.03192): Two-phase retrieval
+  - TinyAgent (arxiv:2409.00608): SLM-native agent patterns
+
+Modules:
+  Core:       types, llm_backend, actor, purpose_function, experience_replay, optimizer, orchestrator
+  SLM:        slm_backends (Ollama, llama-cpp, prompt compression)
+  Streaming:  streaming (async generators, event streaming)
+  Tools:      tools (Tool base class, built-in tools, Tool RAG)
+  Observe:    observability (cost tracking, callbacks, metrics)
+  Multi:      multi_agent (shared memory, agent delegation, teams)
+  HITL:       hitl (checkpoint, interrupt, resume, Φ overrides)
+  Eval:       evaluation (benchmark runner, improvement curves)
 """
 
-__version__ = "0.1.0"
+__version__ = "0.2.0"
 
+# Core
 from purpose_agent.types import (
-    State,
-    Action,
-    Trajectory,
-    TrajectoryStep,
-    Heuristic,
-    PurposeScore,
-    MemoryRecord,
+    State, Action, Trajectory, TrajectoryStep,
+    Heuristic, PurposeScore, MemoryRecord, MemoryTier,
+)
+from purpose_agent.llm_backend import (
+    LLMBackend, MockLLMBackend, HFInferenceBackend,
+    OpenAICompatibleBackend, ChatMessage,
 )
-from purpose_agent.llm_backend import LLMBackend, MockLLMBackend
 from purpose_agent.actor import Actor
 from purpose_agent.purpose_function import PurposeFunction
 from purpose_agent.experience_replay import ExperienceReplay
 from purpose_agent.optimizer import HeuristicOptimizer
-from purpose_agent.orchestrator import Orchestrator
+from purpose_agent.orchestrator import Orchestrator, Environment, SimpleEnvironment, TaskResult
+
+# SLM-Native Backends
+from purpose_agent.slm_backends import (
+    OllamaBackend, LlamaCppBackend, SLMPromptCompressor,
+    create_slm_backend, SLM_REGISTRY,
+)
+
+# Streaming & Async
+from purpose_agent.streaming import StreamingMixin, StreamEvent, AsyncOrchestrator
+
+# Tools
+from purpose_agent.tools import (
+    Tool, FunctionTool, ToolResult, ToolRegistry,
+    CalculatorTool, PythonExecTool, ReadFileTool, WriteFileTool,
+)
+
+# Observability
+from purpose_agent.observability import (
+    CostTracker, TokenUsage, CallbackManager,
+    AgentEvent, EventType, LoggingCallback, MetricsCollector,
+)
+
+# Multi-Agent
+from purpose_agent.multi_agent import AgentSpec, AgentTeam
+
+# Human-in-the-Loop
+from purpose_agent.hitl import (
+    HITLOrchestrator, Checkpoint, HumanInputHandler,
+    CLIInputHandler, AutoApproveHandler, InterruptType,
+)
+
+# Evaluation
+from purpose_agent.evaluation import BenchmarkTask, BenchmarkRunner, BenchmarkResult
 
 __all__ = [
-    "State",
-    "Action",
-    "Trajectory",
-    "TrajectoryStep",
-    "Heuristic",
-    "PurposeScore",
-    "MemoryRecord",
-    "LLMBackend",
-    "MockLLMBackend",
-    "Actor",
-    "PurposeFunction",
-    "ExperienceReplay",
-    "HeuristicOptimizer",
-    "Orchestrator",
+    # Core
+    "State", "Action", "Trajectory", "TrajectoryStep", "Heuristic",
+    "PurposeScore", "MemoryRecord", "MemoryTier",
+    "LLMBackend", "MockLLMBackend", "HFInferenceBackend",
+    "OpenAICompatibleBackend", "ChatMessage",
+    "Actor", "PurposeFunction", "ExperienceReplay", "HeuristicOptimizer",
+    "Orchestrator", "Environment", "SimpleEnvironment", "TaskResult",
+    # SLM
+    "OllamaBackend", "LlamaCppBackend", "SLMPromptCompressor",
+    "create_slm_backend", "SLM_REGISTRY",
+    # Streaming
+    "StreamingMixin", "StreamEvent", "AsyncOrchestrator",
+    # Tools
+    "Tool", "FunctionTool", "ToolResult", "ToolRegistry",
+    "CalculatorTool", "PythonExecTool", "ReadFileTool", "WriteFileTool",
+    # Observability
+    "CostTracker", "TokenUsage", "CallbackManager",
+    "AgentEvent", "EventType", "LoggingCallback", "MetricsCollector",
+    # Multi-Agent
+    "AgentSpec", "AgentTeam",
+    # HITL
+    "HITLOrchestrator", "Checkpoint", "HumanInputHandler",
+    "CLIInputHandler", "AutoApproveHandler", "InterruptType",
+    # Evaluation
+    "BenchmarkTask", "BenchmarkRunner", "BenchmarkResult",
 ]