The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill executes a remote script piped directly to bash from the HKUDS/Upskill repository.
Evidence: curl -sSL https://raw.githubusercontent.com/HKUDS/Upskill/main/cc-integration/install.sh | bash -s -- --remote is used in both SKILL.md and scripts/install.sh.
[COMMAND_EXECUTION]: The skill frequently invokes shell commands to manage its environment, including installing external dependencies and modifying system-wide configuration.
Evidence: scripts/install.sh uses npx to globally install skills and executes local shell scripts like install.sh and upskill-build.sh to perform core logic.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests untrusted session data to influence future agent instructions.
Ingestion points: The upskill-build command reads session.log and metadata.txt from previous sessions (found in references/commands.md).
Boundary markers: None mentioned for the teacher model's analysis of failure trajectories.
Capability inventory: The skill has access to Bash, Read, Write, Edit, and WebFetch tools.
Sanitization: No evidence of sanitization or filtering of the session logs before they are processed by the "Teacher" model to generate new SKILL.md instructions.

upskill