Remove the experimental Harbor integration (superseded by OpenEnv)

[adithya-s-k]

Removes trl.experimental.harbor and everything that hung off it (added in #6018).

Harbor support now lives in OpenEnv — see huggingface/OpenEnv#1036 — and is reached through the environment/harness surface rather than a TRL-side integration. Keeping an in-tree copy means maintaining two paths to the same framework and having to explain which one a reader should use.

The in-tree path was also the more limited of the two. It supported Harbor's external agents only, because TRL had to drive generation turn-by-turn to capture the policy's tokens, logprobs and env mask — which an opaque in-container agent cannot expose. OpenEnv's capture proxy exposes exactly that, so the installed agents Harbor ships become trainable without being reimplemented. #6947 is the replacement example.

Removed

trl/experimental/harbor/	__init__.py, _env.py, _spec.py
examples/grpo_harbor/	script plus the bash, jupyter and terminal_notes harnesses
tests/experimental/test_harbor.py
trl/import_utils.py	is_harbor_available()
tests/testing_utils.py	require_harbor — defined but never referenced by any test
pyproject.toml	the public trl[harbor] extra
docs	harbor.md, its _toctree entry, and its rows in example_overview.md and grpo_trainer.md

docs/source/openenv.md said "Two further integrations" and now names only OpenReward.

SECURITY.md is deliberately untouched — its "Safe harbor" section is the legal term, not the framework.

Note

This drops a public extra (trl[harbor]), so it is a breaking change for anyone importing trl.experimental.harbor. Happy to land it as a deprecation first if you'd prefer a release of warning.

---

Note

Medium Risk
Breaking removal of a public optional extra and experimental API; core GRPO trainers are unchanged, but Harbor users must migrate to OpenEnv.

Overview
Removes the experimental trl.experimental.harbor integration and everything tied to it: HarborSpec / HarborEnv, the examples/grpo_harbor GRPO script and custom harnesses, tests/experimental/test_harbor.py, the trl[harbor] optional extra, and is_harbor_available / require_harbor. Harbor training is expected to go through OpenEnv instead of an in-process TRL path.

Documentation drops harbor.md from the docs toctree and strips Harbor from the examples index, GRPO environment-integration table, and openenv.md (OpenReward is now the only other named environment_factory integration in that guide).

Adds examples/async_grpo_harbor/PR_DESCRIPTION.md as scaffolding for a follow-up AsyncGRPO + OpenEnv Harbor example (capture proxy, loop-owning harnesses); the diff does not include the actual async_grpo_harbor.py script.

This is a breaking change for pip install trl[harbor] and from trl.experimental.harbor import ....

^{Reviewed by Cursor Bugbot for commit 0729015. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 0729015. Configure here.}

[cursor]

Scaffolding file accidentally committed

Low Severity

examples/async_grpo_harbor/PR_DESCRIPTION.md is scaffolding that the file itself says to delete or leave untracked. The directory has no async_grpo_harbor.py and this removal PR does not add that example, so the leftover draft ships as an empty examples folder.

 

^{Reviewed by Cursor Bugbot for commit 0729015. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0729015743

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Remove the tracked scaffold directory

Tracking this scaffold creates an examples/async_grpo_harbor directory even though no corresponding row is added to docs/source/example_overview.md. tests/test_examples_index.py::test_examples_index_matches_folders enumerates every example directory and therefore reports async_grpo_harbor in missing_rows, causing the test suite to fail; delete or keep this scaffold untracked, as its own comment instructs, or add the actual indexed example.

Useful? React with 👍 / 👎.