Custom code agents

A code agent is a single Python function:

def run(context: dict) -> dict:
    # context holds what you bound to this agent (run, repo, pr_diff, plan_output, …)
    logger.log("info", "reviewing the diff")
    # …your logic…
    return {"status": "done", "output": "looks good"}

You write run() (plus any imports/helpers). Premia wraps it in a managed function — the image, secrets, and a logger are provided — and deploys it to our cloud for your org. You manage no infrastructure.

• context contains the bindings you selected for the step (run metadata, repo, the PR diff, the plan output, org secrets, …).
• logger.log(level, message) writes to the run's logs (visible in the dashboard / premia runs logs).
• Return a dict. If the step is a gate, return a verdict (e.g. {"verdict": "approved"}) so pipeline edges can branch on it.

The dev loop (CLI)

The agent's code is the source of truth in Premia; the CLI syncs it to a local file you can keep in your repo:

premia agents pull my-agent          # → ./my-agent.py (scaffolds run(context) if new)
#   …edit ./my-agent.py…
premia agents push my-agent          # upload your changes
premia agents deploy my-agent --wait # build + deploy to our cloud

Then add the agent as a step in your pipeline and it runs like any built-in.

Requirements & dependencies

Extra pip packages an agent needs are installed into its container image at deploy time. Set org secrets (like API keys) via secrets and read them from context.