Crate swarm_engine_llm 

LLM Integration Module

Action selection and batch inference using lightweight LLMs

Module Structure

• decider: LLM Decider trait and related types
• ollama: Ollama backend (HTTP API)
• invoker: BatchInvoker implementation (LLM batch calls)
• prompt_builder: ResolvedContext → prompt generation
• registry: Model registry
• response_parser: LLM response parsing and repair (shared module)

Design Approach

BatchInvoker System (for Manager)

ManagerAgent batch processing is implemented in the following layers:

Core Layer
├── ManagerAgent trait (prepare / finalize)
├── DefaultBatchManagerAgent   ← Core layer default implementation
├── ContextStore / ContextView ← Normalized context
└── ContextResolver            ← Scope resolution

LLM Layer
├── PromptBuilder              ← ResolvedContext → prompt
└── BatchInvoker implementations ← LLM batch calls
    └── LlmBatchInvoker

ManagerAgent implementation uses Core layer’s DefaultBatchManagerAgent. LLM layer provides PromptBuilder (prompt generation) and BatchInvoker (LLM calls).

Re-exports

pub use llama_cpp_server::ChatTemplate;

pub use llama_cpp_server::LlamaCppServerConfig;

pub use llama_cpp_server::LlamaCppServerDecider;

pub use batch_processor::BatchProcessError;

pub use batch_processor::BatchProcessResult;

pub use batch_processor::BatchProcessor;

pub use batch_processor::LlmBatchProcessor;

pub use batch_processor::LlmBatchProcessorConfig;

pub use debug_channel::LlmDebugChannel;

pub use debug_channel::LlmDebugEvent;

pub use debug_channel::StderrLlmSubscriber;

pub use decider::LlmDecider;

pub use decider::LlmDeciderConfig;

pub use decider::LlmError;

pub use invoker::create_llm_invoker;

pub use invoker::LlmBatchInvoker;

pub use ollama::OllamaConfig;

pub use ollama::OllamaDecider;

pub use prompt_builder::PromptBuilder;

pub use registry::ModelInfo;

pub use registry::ModelRegistry;

pub use registry::RegistryError;

pub use strategy_advisor::parse_selection_kind_fuzzy;

pub use strategy_advisor::LlmStrategyAdvisor;

pub use strategy_advisor::StrategyPromptBuilder;

pub use strategy_advisor::StrategyResponseParser;

Modules

batch_processor

Batch Processor - ManagerAgent 向け Batch LLM 処理

debug_channel

LLM Debug Channel - LLM呼び出しのデバッグ出力

decider

LLM Decider - Action 選択のための LLM 抽象

invoker

BatchInvoker 実装 - Core の BatchInvoker trait を LLM で実装

json_prompt

JSON出力専用プロンプトテンプレート

llama_cpp_server

llama-server Decider - HTTP API 連携

ollama

Ollama Decider - Ollama HTTP API 連携

prompt_builder

PromptBuilder - ResolvedContext からプロンプトを生成

registry

Model Registry - Ollamaモデルの動的検出と管理

response_parser

LLM レスポンスのパース処理（共通モジュール）

strategy_advisor

Strategy Advisor - LLM による探索戦略アドバイス

Structs

DecisionResponse

LLM からの判断レスポンス

LoraConfig

LoRA アダプター設定

StrategyAdvice

戦略アドバイスの結果

StrategyContext

戦略判断に必要なコンテキスト

WorkerDecisionRequest

個別 Worker への判断リクエスト

Enums

SelectionKind

Selection アルゴリズムの種類

StrategyAdviceError

戦略アドバイスエラー

Traits

StrategyAdvisor

戦略アドバイザー trait