Prism

Use PRISM when: (1) reviewing an architecture decision, security-sensitive change, or major refactor (>500 lines), (2) making a decision you'll live with for...

This skill is internally consistent for orchestrating multi-agent code reviews, but it will read your repository files and store review archives under analysis/prism/archive/. Before installing or running: (1) confirm you are comfortable with an automated process reading and quoting files from your workspace (sensitive secrets in repo files could end up in findings), (2) inspect or create the referenced completion script (~/.openclaw/scripts/sub-agent-complete.sh) so you know what 'completion' does and whether it triggers any network callbacks, (3) review the README/git-clone source (https://github.com/jeremyknows/PRISM.git) if you want the canonical implementation, and (4) consider retention: PRISM archives reviews by default—ensure that archive retention policies meet your data-sensitivity requirements. If any of the above is unacceptable (exposing secrets, automatic callbacks), restrict the skill's access or run it in a controlled sandbox.

Type: OpenClaw Skill Name: prism Version: 2.1.1 The PRISM skill bundle implements a multi-agent review protocol that uses sub-agents to perform adversarial analysis on code and architecture. While the logic is aligned with its stated purpose, the skill contains a significant prompt-injection vulnerability explicitly acknowledged in the 'Known Limitations' section of SKILL.md: prior review findings are retrieved from a local archive and injected into reviewer prompts without sanitization. This allows a potentially compromised archive file to influence agent behavior. Furthermore, the orchestrator performs broad file system operations and executes shell commands, including a call to a local script (~/.openclaw/scripts/sub-agent-complete.sh), which increases the risk profile in the presence of the injection vulnerability.