Reason Harbor adapter

Run Reason as an agent in Harbor. Harbor starts the task container and grades the result. The adapter installs the published Reason CLI in that container, checks its checksum, and starts it as a worker; the agent runs as a normal Reason Session. Your API key stays on your machine; the container only receives a one-use enrollment ticket.

Install

uv tool install "harbor==0.25.0" --with "reason-machines-harbor>=0.3.44"

The adapter works with Harbor 0.22 to 0.25 and Python 3.12 or later. For cloud environments, add the matching Harbor extra, for example "harbor[e2b,modal]==0.25.0". To reproduce our published results, use the Harbor version the benchmark scripts pin.

Run

Get a Reason API key with run, sessions:read, deployment:read, devices:read, devices:write, devices:use and org:write, and enable the model under Models.

export REASON_API_KEY="YOUR_REASON_API_KEY"
DOCKER_DEFAULT_PLATFORM=linux/amd64 harbor run \
  --dataset datacurve/deep-swe-1-1 --n-tasks 1 \
  --agent reasonmachines_harbor:ReasonAgent \
  --model openai/gpt-6-luna \
  --ak reasoning_effort=xhigh \
  --allow-agent-host api.reasonmachines.com \
  --allow-agent-host relay.reasonmachines.com \
  --env reasonmachines_harbor.environment:ReasonDockerEnvironment

The worker connects out to the Reason API and relay, so both --allow-agent-host flags are required, even for offline tasks. Drop --n-tasks 1 to run the whole dataset. On e2b, use --env reasonmachines_harbor.environment:ReasonE2BEnvironment: it keeps the task image's environment variables, PATH included, which e2b otherwise drops.

Each trial's agent/ folder holds the Session as the model saw it (trajectory.json, ATIF), its link, status and cost (session.json) and what it ran on (receipt.json, deployment.json). API errors are reported by code, for example missing_scope or ephemeral_workspace_parent_payment_required.

Ephemeral mode needs an owner or admin key of a workspace with paid model access. To bill the model to your own provider account, connect its API key under Models and pass the model as byok:<provider>:<model>, for example byok:openai:gpt-6-luna.

To run the full benchmarks and reproduce our results, see reason-machines/benchmark.

Source

reason-machines/reason-machines-harbor is a read-only copy of the adapter in the Reason monorepo, where it is developed and tested. Releases are tagged vX.Y.Z.

License

Apache-2.0.