Only the latest release on master is supported. This is a young project; please update before reporting.

Open a private security advisory at https://github.com/activeing123/dsh-forge/security/advisories/new or email the maintainer through GitHub. Do not open a public issue for a vulnerability.

dsh-forge is a DeepSeek Harness plugin that runs in your DSH process with the same permissions as your harness instance. Installing it runs its code with your own credentials. Review the source before installing, and run it somewhere that does not hold keys you cannot afford to lose — the standard warning for any community plugin.

• No network telemetry. dsh-forge never phones home. The only outbound model calls are the harness's own ctx.llm stream used to distill a session trace into a skill, using the harness's configured provider.
• Event stream only. It reads session/event and tool/session services through the plugin context; it stores only in-memory traces (intents, tool names + truncated args, turn counts) and drops nothing to third parties.
• Skill output is your data. Forged SKILL.md files are written under .dsh-forge/skills/ (configurable via skillsDir) and registered with ctx.skills — inside your workspace, no exfiltration.

Session traces are treated as untrusted input. The LLM's forged JSON is parsed defensively (bounded braces search, kebab-case name validation, truncated description/whenToUse/content); a parse failure aborts the forge rather than writing a partial skill.

• dsh version and profile (e.g. dsh --version, profile name)
• dsh-forge version / commit
• Steps to reproduce, expected vs actual behaviour
• Whether it involves network access, credentials, or files outside the workspace