The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The installation process involves cloning a repository from an untrusted GitHub source (https://github.com/SeemSeam/claude_codex_bridge.git) and immediately executing shell scripts (./install.sh install) or PowerShell scripts (.\install.ps1 install).
[EXTERNAL_DOWNLOADS]: The skill instructs the user to download code from an external repository not associated with trusted organizations.
[PRIVILEGE_ESCALATION]: The Windows installation command uses powershell -ExecutionPolicy Bypass, which explicitly overrides security restrictions designed to prevent the execution of unauthorized scripts.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates an environment where multiple agents interact via shared memory (.ccb/ccb_memory.md) and delegation commands (/ask). This creates a vulnerability surface where one agent could influence or control another through maliciously crafted data or responses.
Ingestion points: Agents read shared context from .ccb/ccb_memory.md and receive queries from other agents via the /ask command.
Boundary markers: The skill body provides no evidence of delimiters or instructions to ignore embedded commands in shared memory.
Capability inventory: The framework includes capabilities for subprocess.run (executing the ask utility), git worktree management, and general file system access.
Sanitization: There is no documented validation or sanitization of content written to the shared memory file or passed through inter-agent queries.

seemseam-ccb-multi-agent-cli