dsh-secret-guard

When a model generates tool-call arguments in DeepSeek Harness, it can inadvertently embed private keys, API tokens, or other sensitive payloads. If executed unchecked, those secrets may leak into logs, external services, or downstream systems. dsh-secret-guard intercepts at the tools/pre-execute hook, inspects the model-supplied arguments, and returns allow, ask, or deny before the tool body ever runs, keeping suspected secrets out of execution.

In v0.1 the guard detects common private-key, token, and key-naming patterns, flags obviously sensitive field names, and enforces a serialized payload size limit, failing closed when exceeded. An optional protectedTools scope narrows inspection to specific tools; an empty value means every tool is checked. A key design constraint: v0.1 does not rewrite or redact arguments, because DSH freezes execution identity and arguments before the policy layer, so mutating them inside a guard would violate the runtime contract.

The plugin suits teams and individual developers running DSH workflows that process user-supplied or externally sourced data and want a lightweight secret-leak checkpoint before tool execution. It is a heuristic detector, not a full DLP solution, and makes no guarantee that every credential format will be caught. For broader action-authorization requirements, compose it with a policy boundary such as AgentFuse.