The Agent Skills Directory

[COMMAND_EXECUTION]: The collab-canvas terminal write command enables the execution of arbitrary shell input within terminal tiles, which can be misused for unauthorized system commands. Evidence: collab-canvas terminal write <id> <input> in SKILL.md.
[COMMAND_EXECUTION]: The skill provides instructions to modify user shell profiles (~/.zshrc or ~/.bashrc) to update the PATH, which constitutes a persistence mechanism. Evidence: "Setup" section instructions in SKILL.md.
[DATA_EXFILTRATION]: The collab-canvas terminal read command allows an agent to read raw terminal buffers, creating a risk for the exposure of sensitive credentials or environment data. Evidence: collab-canvas terminal read <id> [--lines N] in SKILL.md.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted data from terminal outputs and files. 1. Ingestion points: collab-canvas terminal read and collab-canvas tile create --file in SKILL.md. 2. Boundary markers: Absent. 3. Capability inventory: Arbitrary shell execution via terminal write. 4. Sanitization: Absent; the skill does not filter or validate external content before processing.

collab-canvas