2. The Deep Agent

The flagship of v1.0.0 is hawki deep, an autonomous exploit-hunting agent. It runs an asynchronous loop, gated by a budget manager:

Planner.next_attack  ->  Executor.execute  ->  memory.record  ->  budget.consume

Planners escalate in creativity: RulePlanner drains the known attack scripts, HybridPlanner blends known patterns with contract context, and LLMPlanner invents novel attacks the rules never encoded. For a novel attack, the NovelExecutor uses an LLM code generator to synthesize a Hardhat or Foundry proof-of-concept and runs it in the Docker sandbox, so the result is a reproducible PoC rather than a bare claim.

Agent memory is pluggable: a SQLite store (default, at ~/.hawki/deep_memory.db) or a JSON store. The budget is a dual limit on max_attempts and max_tokens, both defaulting to unlimited. Continuous mode and a target-contract focus round out the agent.

The agent has been exercised on real incident code: against DeFiVulnLabs' reproductions, it invented and landed a live reentrancy drain in the sandbox, growing the attacker balance from 1 ETH to 2 ETH.

$ hawki deep ./contracts --sandbox --max-attempts 25