Fuel-Metered Agent Runtime-Control Plane

# Fuel-Metered Agent Runtime-Control Plane

Resource governance and runaway prevention for autonomous coding agents: every action spends from a fuel budget, and exhaustion traps to a safe-stop with a human catch.

## Overview
Agent actions (generation steps, tool calls, file writes) consume a configured fuel/instruction budget. When the budget is exhausted, the runtime traps rather than continuing, routes the agent to an operator handoff node, and uses human-in-the-loop approval to either replenish fuel or safely terminate. For platform/AI engineers who need a hard boundary on cost, runaway loops, and compute.

## Architecture (data flow)
Initial state holds a fuel budget. Fuel-metered execution evaluates each step's cost and raises an OutOfFuelTrap when insufficient. A reducer merges decremented budget and event history into global state. A streaming trace renderer prints thoughts, tool calls, and remaining fuel to the terminal in real time. Conditional routing sends budget>0 back to the worker and budget<=0 (or trapped) to the handoff node, where HITL approval gates replenish-or-exit.

## Components
1. Fuel-Metered Execution — (VMState, FuelBudget) -> Either<(VMState, FuelBudget), OutOfFuelTrap>. Source: wasmi-labs/wasmi.
2. Reducer-Based State Merging — (State, StateUpdate) -> State. Source: langchain-ai/langgraph.
3. State-Based Conditional Routing — State -> NodeId. Source: langchain-ai/langgraph.
4. In-the-Loop Tool Approval — ToolCall -> DeferredToolCall. Source: pydantic/pydantic-ai.
5. Streaming Execution Trace Terminal Renderer — AsyncIterable<AgentEvent> -> None. Source: microsoft/autogen.

## Build
An async Python reference (RunawayControlRuntime) steps a simulated coding-agent task list, emitting color-coded trace events, decrementing fuel per step, trapping on exhaustion, and routing to an interactive operator handoff that replenishes (+fuel) or safe-exits.

## Acceptance Check
Initialize with 60 fuel; run steps costing 20, 25, 40. The third trips the trap, control routes to the operator console; entering "yes" replenishes and resumes, "no" safe-stops.

## Honest status
Architecture is sound. The "fuel" is a Python integer counter, not real instruction-weighted metering — the wasmi attribution is to a genuine fuel-metered VM that this build only simulates. A production version meters real work (token cost, wall-clock, or VM instruction weights). Star atom: fuel-metered execution (the budget-and-trap). Differs from trust-gating: this governs resource burn, not tool admission.