Module coordinator 

Multi-turn tool coordinator — data types and runtime.

Provides a provider-agnostic orchestration layer for agent tool-calling loops: send a prompt with tool definitions, handle tool call requests, execute tools, feed results back, repeat until the model produces a final response or hits an iteration cap.

Types reused from crate::agent:

• ToolCallRequest — what the model asks for
• ToolCallRecord — what actually happened
• TokenUsage — accumulated counts

Types defined here:

• ToolCallingConfig — iteration / parallelism / timeout knobs
• FinishReason — why the session ended
• Role — re-exported from crate::agent (system/user/assistant/tool)
• ConversationMessage — a single turn in the history
• CoordinatorResult — everything the caller gets back
• ToolCoordinator — the runtime that drives the LLM+tool loop

Design notes

• ToolCallRecord is reused from crate::agent rather than duplicated. Failed tool calls land in result as a {"error": "..."} JSON object with success: false, matching pawan’s existing agent loop — there’s no separate error field on the record.
• ConversationMessage::tool_call_id is only populated on Role::Tool turns and links the result back to the assistant message that requested it.

Re-exports

pub use types::*;

Modules

types

Coordinator wire types — ToolCallingConfig, FinishReason, ConversationMessage, CoordinatorResult

Structs

ToolCoordinator

Runtime that drives the LLM + tool-calling loop.