dsh-secret-guard

A fail-closed secret-like payload guard for DeepSeek Harness (DSH) tool calls.

The plugin inspects model-supplied tool arguments at tools/pre-execute and returns allow, ask, or deny before the tool body runs. v0.1 deliberately does not rewrite/redact arguments: DSH freezes execution identity and arguments before policy, so mutating them inside a guard would violate the runtime contract.

v0.1

• detects common private-key/token/key patterns;
• detects obvious sensitive field names;
• bounds serialized payload size and fails closed above the configured limit;
• optional tool-name scope (protectedTools); empty means inspect every tool;
• actionOnFinding: block|ask;
• no raw secret values in the denial reason.

Non-claims

• heuristic secret detection, not complete DLP;
• no guarantee every credential format is detected;
• no argument redaction in v0.1;
• not a sandbox or malware detector;
• should be composed with a policy boundary such as AgentFuse when broader action authorization is required.

Development

npm test

License

MIT