Module agent 

agent

Agent loop domain types.

These types define the core abstractions for the backend agentic loop. They are pure domain primitives: no LLM backend references, no MCP types, no infrastructure concerns.

Modules

Module	Contents
config	AgentConfig — loop control parameters
tool_types	ToolDefinition, ToolCall, ToolResult
messages	AgentMessage — closed conversation-turn enum
messages_serde	Custom Serialize/Deserialize impls for AssistantContent
events	AgentEvent (SSE units), LlmStreamEvent (stream protocol)
loop_detection	LoopDetector — repeated tool-call-batch guard (FNV-1a batch signatures)
stagnation	StagnationDetector — repeated assistant-text guard
fnv1a	fnv1a::fnv1a_64 — the hash backing both detectors

Design Principles

• AgentMessage is a closed enum so the type system prevents invalid states (e.g. a User message carrying tool_calls).
• ToolDefinition is a dedicated type — adapter layers convert McpTool → ToolDefinition; the agent domain must not depend on MCP domain types.
• ToolResult with success: false is context for the LLM, not an error; tool failures are fed back into the conversation so the model can reason about them and retry or adjust its approach.
• AgentEvent is the unit of SSE emission; every observable state change in the loop corresponds to exactly one variant.

Modules

Module	LOC	Complexity	Coverage
config.rs
events.rs
fnv1a.rs
messages.rs
messages_serde.rs
tool_display.rs
tool_types.rs
loop_detection/
stagnation/

Re-exports

pub use config::AgentConfig;

pub use config::AgentConfigError;

pub use config::DEFAULT_MAX_ITERATIONS;

pub use config::DEFAULT_MAX_PARALLEL_TOOLS;

pub use config::DEFAULT_MAX_STAGNATION_STEPS;

pub use config::MAX_ITERATIONS_CEILING;

pub use config::MAX_PARALLEL_TOOLS_CEILING;

pub use config::MAX_STAGNATION_STEPS_CEILING;

pub use config::MAX_TOOL_TIMEOUT_MS_CEILING;

pub use config::MIN_CONTEXT_BUDGET_CHARS;

pub use config::MIN_TOOL_TIMEOUT_MS;

Modules

config

AgentConfig — configuration for a single agentic loop run.

events 🔒

AgentEvent and LlmStreamEvent — observable events in the agentic loop.

fnv1a 🔒

FNV-1a 64-bit hash utility backing the loop and stagnation detectors.

loop_detection 🔒

Loop Detection

messages 🔒

AgentMessage — A single message in the agent conversation.

messages_serde 🔒

Custom Serialize / Deserialize implementations for agent message types.

stagnation 🔒

stagnation

tool_display

Shared tool display formatting — single source of truth for all surfaces.

tool_types 🔒

Tool-schema and tool-call/result types.

Structs

AssistantContent

Content carried by an AgentMessage::Assistant turn.

BatchRecord

Names the batch LoopDetector::check just counted.

LoopDetector

Stateful guard that detects when the same tool-call batch repeats back to back and gets the same answer back.

StagnationDetector

Stateful guard that detects when the assistant repeats the same text.

ToolCall

A tool invocation requested by the LLM.

ToolDefinition

A tool that the LLM may invoke, expressed as a name + JSON Schema.

ToolResult

The outcome of executing a ToolCall.

Enums

AgentEvent

An observable event emitted by the agentic loop.

AgentMessage

A single message in the agent conversation.

LlmStreamEvent

A single event produced by a streaming LLM response.

RepeatOutcome

What one occurrence turned out to be, once its answers were known.

Constants

AGENT_EVENT_CHANNEL_CAPACITY

[tokio::sync::mpsc] channel capacity for streaming AgentEvents produced by a single crate::ports::AgentLoopPort::run call.

Functions

batch_results_hash

Hash the answers to one batch of tool calls.

batch_signature

Compute the batch signature for a slice of ToolCalls.

hash_result_content

Hash one answer’s content, whatever shape it arrived in.

hash_result_text

Hash one answer’s text.

is_observation_batch

Return true if every call in calls is an observation-only tool.