Module worker 

The turn worker: a child process (Role::Turn) that runs ONE turn over the context slice the supervisor handed it — a root / conversation turn, a bounded agent step, or a structured think — and reports the transcript delta, the usage and the outcome (TurnDone).

It calls the model and MCP tools itself; internal tools (memory, plan, subagents, sleep, finish, …) are round-tripped to the supervisor (ToolRequest → ToolResult) so state changes are made by the state owner. Before each model call it may ask for budget admission (BudgetRequest → BudgetGrant).

The turn keeps tool results structured (structuredContent first, text-JSON second, text last), validates a schema’d final answer and re-asks on a miss, keeps a serializable transcript, estimates tokens for admission, and detects call loops. The loop body is run_turn, generic over a Bridge so it is unit-testable in-process; run_turn_child is the process entry.

Structs

BudgetReply

A budget grant as the worker sees it.

ChildBridge

The process bridge: frames over the control channel.

McpClients

The connected MCP clients as an McpCaller.

TurnLimits

The knobs of one turn (from the payload).

Constants

DEFAULT_MAX_TOKENS_PER_CALL

Default per-response completion cap.

LOOP_REPEATS

The same tool call (name + args) this many times in one turn is a loop.

SCHEMA_REASKS

Re-asks when a schema’d answer misses.

TOOL_CALL_CAP

Longest a single MCP tool call may take inside a turn.

Traits

Bridge

The supervisor link: internal-tool and budget round-trips + cancellation.

McpCaller

How MCP tools are reached from the turn.

Functions

parse_json_answer

Parse a model’s JSON answer tolerantly (fences, prose around the object).

run_turn

Run ONE turn. Never panics; every failure is a TurnResult status.

run_turn_child

The process entry for a Role::Turn child: run the turn, report Usage + TurnDone, return the exit code.

tool_result_value

The value of an MCP tool result: structuredContent, else the text parsed as JSON, else the text.