Preventing double-execution and tool replay in local LLM workflows (Rust tool gate)

I run local models via Ollama and LM Studio tied to custom system scripts. One annoyance I kept hitting wasn't the initial tool call, but duplicate retries or mid-sequence parameter mutations when a model gets stuck in a loop. I built a lightweight v0.1 Rust CLI / library gate where the local model acts purely as an untrusted proposer: • A trusted issuer creates a single-use claim bound to action + resource + payload SHA-256. • The agent submits its tool request. • The first matching request executes and atomically burns the claim on disk (O_EXCL / create_new). • Any replay or mutated payload immediately gets a DENY. Disclosure: This is my own open-source project. Repo: https://github.com/trencinodin-stack/arcstone-mcp-sidecar?utm_source=gemini Note: Right now this is a local CLI tool and library function (execute_request()), not a full stdio MCP proxy server. You call it directly inside your tool handlers before firing file modifications or shell scripts. Curious what patterns other people running local agents use to catch duplicate execution loops or unauthorized retries!