CQLS Agentic Coding Policies¶
Updated 2026-11-19.
These policies are part of the CQLS HPC Agent Terms of Use (https://docs.hpc.oregonstate.edu/cqls/ai-gateway/hpc-agent-terms/). They hold operational rules for agent work on the CQLS HPC cluster. OSU updates them separately from the terms, and it dates each update. Follow the version in force.
An agent in an hpcagent session gets the same rules as commands, in the site rules that hpcagent gives to each session.
Changes¶
- 2026-11-19: first version.
Where agents run¶
- Run each agent that runs on the cluster through
hpcagent. This includes an IDE extension that connects with Remote-SSH. An agent in a session can change its own harness: its settings, hooks, skills, plugins and binaries. If you run that harness outsidehpcagent, those changes run with your whole account. OSU does not support agents that run on the cluster outsidehpcagent. - An agent on your own computer that runs commands on the cluster over SSH runs outside the sandbox. You are accountable for its commands (rule 5). Use
hpcagentwhen you can. - Do not run agents on login hosts or on file-transfer hosts.
hpcagentrefuses to start there. On a compute node, start agents in asallocshell. On a server that is not a compute node, you can start agents without an allocation. - Use the transfer host
files.hpc.oregonstate.edufor downloads from external sources (for example NCBI) and for copies to or from your own computer. Do these outside a session. Copies inside the cluster can run on the node.
Accountability¶
- You are accountable for every command that your agent runs. The agent runs under your account.
- CQLS staff can stop agent processes that affect other users, without notice.
Data¶
- Before you start an agent in a project, decide what kind of data the project holds.
- Claude Code and Codex send your prompts, and the contents of each file that the agent reads, to Anthropic or OpenAI. Use them only on data that you may share with those companies.
hpcagent piwith the CQLS AI Gateway keeps prompts and files at CQLS. Use it for scientific data that is not released yet, for example unpublished results or data under embargo, when no agreement for that data forbids processing at CQLS.- Do not use any agent on personal or regulated data: human-subjects data, protected health information (HIPAA), student education records (FERPA), personal information about other people, controlled-access data, export-controlled data or Controlled Unclassified Information, and data that OSU classifies as confidential.
- Keep credentials and sensitive data out of the directories of a project. An agent reads what the session can read, and it sends that text to the models.
Shared resources¶
- Do not write data, installs or caches to your home directory. Home directories have a small quota, and a full home directory stops your sessions. Keep caches in the project or in
$TMPDIR. - Do not let an agent scan shared filesystems. On
/nfs*and/local, an agent searches only the project directory and the paths that you give it. A scan of a shared filesystem makes it slow for all users. - Wait at least 30 seconds between two status checks of the same job.
- Test with a small input first. Use array jobs, not many small jobs.
Slurm jobs¶
- An agent cannot submit a job. It can check a job with
hqsub --check. - Review a job script that an agent wrote before you submit it. Submit it yourself, outside the session, through
hqsub.
Git and GitLab¶
- Grant a project only the GitLab projects that its agent needs (
hpcagent gitlab allow). Use thereadprofile when the agent does not need to push. - Agents open merge requests. An agent merges only when you ask, and only when the project allows it.
Keys¶
- Restart pi after a key rotation or a new enrollment. A running pi keeps the key that it read at start.
- If you need a key outside a session, for example for a script or another tool, get a separate API key from the key page (https://chat.ceoas.oregonstate.edu/keys). Do not copy the agent key out of a session.