dsh-egress-guard

Most existing security plugins in the DeepSeek Harness ecosystem scan configuration files before an agent starts. The real risk, however, lives in the execution moment itself: a curl to an unfamiliar paste site, a git push to an unexpected remote, an API key in a tool result that the model reads outright. dsh-egress-guard embeds security checks directly into the tool-execution pipeline, denying offending network destinations before they fire and redacting credentials after results return, so every security decision happens in real time with the call.

Three rule pipelines do the heavy lifting: an egress allowlist that intercepts requests to hosts outside your whitelist at the pre-execute stage; secret redaction that rewrites tokens and keys from tool results at the post-execute stage, before the model, the durable session log, or any Code Mode program can read them; and a decision audit log that appends every verdict, including the ones monitor mode would have made, to a JSONL file. The plugin attaches via three listeners on documented extension points and disposes cleanly on unload without forking or patching the main loop.

It ships in monitor mode by default, meaning every rule is evaluated and logged but nothing is blocked or rewritten, guaranteeing installation cannot break a working setup. You can read the audit log for a day, build an allowlist from real traffic, then flip to enforce mode. It is aimed at ops and security teams who need egress control and credential protection for Harness agents, as well as individual developers who want to establish an audit baseline before introducing a full sandbox.