The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a Node.js script (scripts/install.js) and a plugin (plugins/fusion-claude.js) that execute system commands to perform their functions. The installer manages file operations, backups, and configuration merging, while the Claude plugin spawns the official claude CLI and process management tools (taskkill on Windows). These actions are documented as the primary utility of the skill.
[INDIRECT_PROMPT_INJECTION]: The skill defines agents (research.md, vision.md) and tools (fusion_claude_review) that ingest untrusted data from external sources like the web, images, or plan packets.
Ingestion points: The websearch tool, the read tool for image processing, and the packet argument for external plan reviews.
Boundary markers: The fusion_claude_review tool implements a strict system prompt to instruct the reviewer model to treat input as untrusted data.
Capability inventory: Capabilities are restricted based on the role, with the main agent having no edit access and subagents having limited bash permissions.
Sanitization: Agent instructions include explicit directives to ignore any commands or instructions found within processed data.
[PERSISTENCE_MECHANISMS]: The skill performs persistent modifications to the agent's environment by writing configuration to ~/.config/opencode/opencode.json and installing multiple agent definition files in ~/.config/opencode/agent/. This is the stated purpose of the 'fusion-setup' utility.
[CREDENTIALS_SAFE]: The skill implements secure secret management by advising users to use environment variable placeholders ({env:VAR}) in the configuration instead of hardcoding API keys. The fusion-claude.js plugin also includes logic to scrub sensitive routing environment variables before spawning external processes.

fusion-setup