[jayzuccarelli]

HarborSessionFactory (where #1276 points opencode_env users) can't pass purpose, provider or eval_sampling, although HarborEnv.run_rollout and the server's run_rollout tool accept all three. So a TRL loop-owning run can't ask for purpose="train", which the provider qualification guide recommends.

That matters most for training: a vLLM started without --return-tokens-as-token-ids still runs every rollout, as eval, and fetch_proxy_trace() hands the trainer an empty list per rollout with only a warning. With purpose="train" the server refuses the rollout before creating a sandbox.

The factory now takes the three kwargs and passes them through. It validates them at construction, like sampling, so a bad value fails once rather than on every rollout. Defaults are unchanged.

Tests: new cases in tests/envs/test_harbor_session_factory.py fail on main (TypeError on the new kwargs) and pass with this change; the harbor and capture suites pass (606 passed, 3 skipped). ruff and usort clean on the touched files.

Prepared with AI assistance (Claude Code); I reviewed the change and ran the tests.

---

Note

Low Risk
Harness plumbing and constructor validation only; rollout semantics align with existing server APIs without touching auth or data paths.

Overview
HarborSessionFactory and HarborSession now accept provider, purpose, and eval_sampling and pass them through to HarborEnv.run_rollout, matching what the harbor server already supports.

Training loops (e.g. TRL loop-owning mode) can set purpose="train" so rollouts fail fast when the engine cannot return token ids, instead of running as eval and yielding empty fetch_proxy_trace() results. provider selects the LLM wire protocol (e.g. Anthropic), and eval_sampling configures eval-only temperature/top-p/top-k when purpose="eval".

The factory validates purpose and the sampling / eval_sampling combination at construction time (same pattern as training sampling), so misconfiguration fails once rather than on every rollout. Defaults remain openai, auto, and no eval sampling.

Tests cover forwarding to run_rollout and the new ValueError cases for invalid purpose or mixed train/eval sampling.

^{Reviewed by Cursor Bugbot for commit 25a7a13. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jayzuccarelli]

@sergiopaniego @adithya-s-k re-checked this against current main: it merges cleanly, and tests/envs/test_harbor_.py plus test_capture_.py pass on the merged tree (566 passed vs 559 on main; the 7 extra are the new tests in test_harbor_session_factory.py). With opencode_env removed in #1346, this is how a TRL loop-owning run gets purpose="train" through HarborSessionFactory. Happy to reshape it if you'd rather expose it another way.

[burtenshaw]

Reviewed factory-to-session-to-rollout forwarding and validation against the existing server contract. Focused factory tests pass (34 passed, 1 skipped), with clean changed-file lint. No blocking findings.

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.