The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill requires the user to provide a full SSH command for remote access, which often includes file paths to private identity keys (e.g., -i ~/.ssh/id_ed25519). This connection string is stored in a configuration file (emergence.json) and used for automated sessions.
[REMOTE_CODE_EXECUTION]: The core functionality involves writing Python experiment code locally, synchronizing it to a remote GPU pod, and executing it autonomously. It uses nohup and background processes to maintain long-running execution sessions on the remote host.
[COMMAND_EXECUTION]: The skill executes a wide range of shell commands on the remote pod via SSH to probe the environment (nvidia-smi, pip show), manage the workspace (mkdir, rsync), and control experiment lifecycles (kill, grep).
[EXTERNAL_DOWNLOADS]: During the setup phase, the skill clones and installs the arxivcli tool from the author's GitHub repository (github.com/tylergibbs1/arxivcli) using bun install to facilitate literature grounding.
[DATA_EXFILTRATION]: The skill uses rsync to transfer local project code, training scripts, and configurations to an external remote pod for processing.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted external data from arXiv abstracts and documentation lookups without explicit boundary markers or sanitization. This data could potentially influence the agent's autonomous research decisions.
Ingestion points: ArXiv search results and paper summaries, remote experiment logs (runs/<id>.log), and documentation retrieved via Context7.
Boundary markers: None identified in the prompt interpolation logic.
Capability inventory: SSH command execution, file system modification (local and remote), process management (kill), and network operations via git and rsync (found in SKILL.md and references/RUNPOD.md).
Sanitization: No validation or escaping is performed on external strings before processing.

discovering-emergence