mini_task_runner

Runs one Harbor task on its own EC2 box with full observability: everything about the trial lands in S3 for post-mortem analysis. The box deletes itself.

You need

• uv, the AWS CLI v2 and ssh.
• AWS credentials that can launch EC2 boxes (one-time setup in ec2/) and either call sts:GetFederationToken or take on a role that can write your bucket.
• An OpenAI-compatible endpoint the box can reach. To reach a model on a private machine, put it on your tailnet with tailscale_relay.

Run

./run examples/tasks/hello examples/run_configs/oracle.toml   # prints <trial-id> <ip> <record-url>

Copy a file from examples/run_configs/ to write your own; its comments explain every field.

• RECORD_ROLE=<role-arn>: if only a role can write your bucket.
• TS_AUTHKEY=<key>: if the endpoint is only on your tailnet.

Watch, stop

source ec2/env.sh && ssh "${SSH_OPTS[@]}" root@<ip> journalctl -fu trial   # follow the trial
bash ec2/terminate_one.sh <ip>                                            # stop it early; its record is lost

The record

<record>/<trial-id>/
├── spec.json      the run config, plus the trial id, task name and launch time
├── task/          the task as it ran
├── runner/        the box scripts that ran it
├── harbor/        Harbor's trial folder: result.json (the reward), agent/, verifier/, trial.log
├── model_calls/   every model call, one file per call, written as it happens
├── services/      each container's output and settings, the images, each image build's log
├── box/           the machine: facts, usage every second, kernel log; steps/ has each step's log
└── upload.json    every file uploaded and its size; written last, so it marks the record complete