Module loop_runner 

Agent loop implementation with Phase-based plugin execution.

The agent loop orchestrates the conversation between user, LLM, and tools:

User Input → LLM → Tool Calls? → Execute Tools → LLM → ... → Final Response

Phase Execution

Each phase dispatches to its typed plugin hook:

RunStart (once)
    │
    ▼
┌─────────────────────────┐
│      StepStart          │ ← plugins can apply state patches
├─────────────────────────┤
│    BeforeInference      │ ← plugins can inject prompt context, filter tools
├─────────────────────────┤
│      [LLM CALL]         │
├─────────────────────────┤
│    AfterInference       │
├─────────────────────────┤
│  ┌───────────────────┐  │
│  │ BeforeToolExecute │  │ ← plugins can block/pending
│  ├───────────────────┤  │
│  │   [TOOL EXEC]     │  │
│  ├───────────────────┤  │
│  │ AfterToolExecute  │  │ ← plugins can add reminders
│  └───────────────────┘  │
├─────────────────────────┤
│       StepEnd           │
└─────────────────────────┘
    │
    ▼
RunEnd (once)

Re-exports

pub use crate::runtime::run_context::await_or_cancel;

pub use crate::runtime::run_context::is_cancelled;

pub use crate::runtime::run_context::CancelAware;

pub use crate::runtime::run_context::RunCancellationToken;

pub use crate::runtime::run_context::StateCommitError;

pub use crate::runtime::run_context::StateCommitter;

pub use crate::runtime::run_context::TOOL_SCOPE_CALLER_AGENT_ID_KEY;

pub use crate::runtime::run_context::TOOL_SCOPE_CALLER_MESSAGES_KEY;

pub use crate::runtime::run_context::TOOL_SCOPE_CALLER_STATE_KEY;

pub use crate::runtime::run_context::TOOL_SCOPE_CALLER_THREAD_ID_KEY;

Structs

AgentConfig

Runtime configuration for the agent loop.

ChannelStateCommitter

GenaiLlmExecutor

Default LLM executor backed by genai::Client.

LlmRetryPolicy

Retry strategy for LLM inference calls.

LoopOutcome

Unified terminal state for loop execution.

LoopStats

Aggregated runtime metrics for one loop run.

LoopUsage

Aggregated token usage for one loop run.

ParallelToolExecutor

Executes all tool calls concurrently.

ResolvedRun

Fully resolved agent wiring ready for execution.

SequentialToolExecutor

Executes tool calls one-by-one in call order.

StepToolInput

Input context passed to per-step tool providers.

StepToolSnapshot

Tool snapshot resolved for one step.

Enums

AgentLoopError

Error type for agent loop operations.

Traits

LlmExecutor

Provider-neutral LLM execution contract consumed by the loop runtime.

StepToolProvider

Provider that resolves the tool snapshot for each step.

ToolExecutor

Strategy abstraction for tool execution.

Functions

execute_tools

Execute tool calls (simplified version without plugins).

execute_tools_with_config

Execute tool calls with phase-based plugin hooks.

execute_tools_with_plugins

Execute tool calls with plugin hooks.

execute_tools_with_plugins_and_executor

run_loop

Run the full agent loop until completion or a stop condition is met.

run_loop_stream

Run the agent loop with streaming output.

tool_map

Helper to create a tool map from an iterator of tools.

tool_map_from_arc

Helper to create a tool map from Arc.