dsh-agent-doctor

A small DeepSeek Harness (DSH) health and safety diagnostics plugin.

/doctor inspects the effective model-visible tool surface for the receiving agent and reports bounded configuration signals without claiming to be a security scanner, sandbox, or danger classifier.

v0.1 checks

• model-visible tool count and names;
• whether reserved run_code is visible;
• heuristic warning for obviously side-effect-shaped tool names;
• optional AgentFuse presence requirement;
• optional approval-service availability signal.

The name-based side-effect check is a diagnostic hint only. It does not authorize or block execution. Use a real runtime policy boundary such as AgentFuse for enforcement.

DSH integration

This repository is intended to be mounted as a DSH bundle. DSH is currently Developer Preview; the plugin targets the pinned public architecture where human commands are registered through ctx.commands and effective tool schemas are exposed by ctx.tools.

- id: dsh-agent-doctor
  name: '@mkaliezz/dsh-agent-doctor'
  config:
    requireAgentFuse: false
    forbidRunCode: true

Run /doctor from an interactive DSH surface after the bundle is composed.

Development

npm test

The core checker has no runtime dependencies and is tested independently of the DSH monorepo. A monorepo compatibility proof should be run against each pinned DSH revision before claiming compatibility.

Non-claims

• not an antivirus or malware scanner;
• not a process sandbox;
• not a policy enforcement boundary;
• does not prove that a tool is safe from its name;
• does not prove effective profile immutability.

License

MIT