deepseek-harness-skillx

External Skills have become the standard way to extend the reach of an agent beyond its built-in capabilities, but that convenience carries real danger. Malicious repositories masquerade as legitimate Skills, prompt-injection payloads hide inside documentation, and candidate Skills routinely ask an agent to execute scripts, install dependencies, or read sensitive files. skillx tackles the root of the problem: it gives an agent a structured, auditable path to discover, compare, and vet external Skills before any action is taken, rather than defaulting to whatever the first README suggests.

At its core, skillx enforces a read-only, version-pinned discovery pipeline. The agent first confirms whether local capabilities already satisfy the task, then generates a structured capability request, searches candidate repositories in priority order, fetches documentation in read-only mode with version identifiers, benchmarks candidates side by side, flags injection signals and over-privileged requests, pauses for explicit user consent whenever risk is detected, and produces a human-readable adoption report. By default the agent references external content temporarily and never auto-installs third-party Skills; permanent adoption requires a full-file review and an explicit user decision.

skillx is aimed at developers who work in DeepSeek Harness, Codex, or any SKILL.md-compatible environment and want their agent to self-extend responsibly. If your workflow involves pulling in document generators, security analysis scripts, or niche utilities on a per-task basis, skillx provides a review-first, use-and-release discipline that keeps every external dependency transparent, traceable, and under your control.