[sergiopaniego]

Three small fixes, one commit each.

1. Remove dead openenv_core import fallbacks (maze_env, websearch_env, openapp_env, qed_math_env)

openenv_core is a deprecation shim inside the openenv package that re-exports openenv.core, so falling back to it after openenv.core fails can never help. The envs now import from openenv.core directly, as #1379 did for snake_env. Checked that the models, clients and servers still import both in-repo (PYTHONPATH=src:envs) and standalone (from the env directory, where the env supports it).

2. GitServerClient.execute_git_command breaks quoted arguments

The command was split with str.split(), so commit -m "two words" reached git as -m '"two' 'words"' and failed with pathspec 'words"' did not match. It now uses shlex.split. The new test commits with a two-word quoted message in a tmp repo (HOME pointed at tmp_path) and checks the message.

3. build_model_step leaks a thread per call

Since #1367, every build_model_step call starts a daemon thread with an event loop that never stops. The CLI builds one per run, but a library user building one per episode gains a thread per episode. The loop is now created lazily once and shared by all model steps, so every call still runs on the same long-lived loop and #1367's "Event loop is closed" fix holds (its regression test still passes). The new test calls build_model_step five times and checks that threading.active_count() grows by at most one.

How it was checked

• New tests fail before the fix and pass after (git: pathspec 'words"', collect: 16 threads vs at most 12).
• tests/core/test_git_server_client.py, tests/core/test_harness_collect.py, tests/test_cli/test_collect.py, tests/envs/test_openspiel_collect.py, tests/envs/test_maze_environment.py (including the gunicorn server tests), tests/envs/test_websearch_environment.py, tests/envs/test_qed_math_environment.py (excluding TestQEDMathServerIntegration, which needs a running server) all pass.

This PR was prepared with AI assistance.

---

Note

Low Risk
Changes are import cleanup, safer shell parsing for git, and a shared event loop with existing regression tests; no auth or data-path changes.

Overview
This PR bundles three small fixes across env imports, git tooling, and rollout collection.

Env imports: Several environments (maze_env, openapp_env, websearch_env, qed_math_env) now import from openenv.core only and drop nested try/except fallbacks to openenv_core, which cannot help when openenv.core is missing.

Git client: GitServerClient.execute_git_command parses the command string with shlex.split so quoted arguments (e.g. commit messages with spaces) are passed to git correctly; a regression test covers empty commits with a two-word message.

Collect harness: build_model_step lazily creates one shared background asyncio loop (via @functools.cache) for all model steps instead of starting a new daemon thread on every call, preserving the pooled-async-client fix while avoiding thread leaks when callers build a step per episode.

^{Reviewed by Cursor Bugbot for commit 916d3a1. 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]

Alignment Review Report

Automated Checks

• Lint: PASS — uv run usort check / ruff format --check / ruff check are clean on every file this PR touches (src/openenv/core/harness/collect.py, src/openenv/core/tools/git_server_client.py, both new test files, and all 12 envs/* import-fallback files). The repo-wide lint.sh run does report failures, but I verified each one is pre-existing on the merge base (8d3455ab) and untouched by this diff (mostly unrelated README/docstring formatting and two long-standing F401/F821 issues in envs/openapp_env/server/openapp_environment.py from an untouched try: import playwright.sync_api block). Nothing in this PR regresses lint.
• Debug code: CLEAN — check-debug.sh only flags print()/TODO occurrences in files this PR doesn't touch (src/openenv/auto/*, src/openenv/harbor/*, src/openenv/cli/*). No debug code in the diff.

Open RFCs Context

• RFC 005 (Agentic Harnesses, In Review) and RFC 006 (Agentic RL Harness Interception, In Review) both treat ModelStep as the fixed generate-hook contract. This PR's collect.py change only caches the event loop build_model_step schedules onto — it does not touch ModelStep's signature or the wrapping/session/capture abstractions those RFCs own, so no conflict.
• No other Draft/In-Review RFC (000–003, 008, 010–012) touches git tooling, env import shims, or this harness internal, so no further RFC overlap.

Tier 1: Fixes Required

None — mechanical checks are clean for the diff.

Tier 2: Alignment Discussion

Principle Conflicts

None identified. Specifically checked and ruled out:

• The openenv_core import-fallback removal (maze_env, openapp_env, websearch_env, qed_math_env) is not a breaking change requiring the INVARIANTS.md deprecation process: src/openenv_core/__init__.py imports openenv.core unconditionally at module load, so the removed except ImportError: from openenv_core... branches could never actually execute — they were dead code. The openenv_core shim package itself is untouched and still works exactly as before for any user importing it directly. This mirrors the precedent already set for snake_env in #1379.
• functools.cache-sharing one event loop across build_model_step calls doesn't change the public ModelStep contract or introduce new concurrency between environment instances — it only fixes a resource leak if the helper is (re)built per episode instead of per run, consistent with the existing "one env = one trajectory" guidance rather than conflicting with it.
• shlex.split in GitServerClient.execute_git_command is a straightforward correctness fix (quoted args were previously broken by naive .split()) with no interface change.

RFC Conflicts

None identified (see Open RFCs Context above).

Verification performed

• Ran the full lint pipeline (usort, ruff format --check, ruff check) scoped to every file in the diff — clean.
• Ran tests/core/test_git_server_client.py and tests/core/test_harness_collect.py — 54/54 passed, including both new tests (test_execute_git_command_keeps_quoted_args, test_repeated_builds_share_one_loop_thread).
• Ran the full suite (PYTHONPATH=src:envs pytest tests/, excluding the smolagents-dependent module that fails to collect without that optional dep): 3493 passed, 166 skipped, 2 failed. Confirmed both failures (test_protocol_client_state, test_protocol_client_multiple_episodes in tests/envs/test_websockets.py) reproduce identically on the unmodified merge-base commit, so they are pre-existing flakiness/bugs unrelated to this PR.

Summary

• 0 mechanical issues to fix
• 0 alignment points for human review
• 0 RFC conflicts to discuss

This is a clean, well-scoped follow-up PR: a genuine quoting bug fix in GitServerClient, a real thread-leak fix in build_model_step, and dead-code removal of unreachable import fallbacks, each covered by tests. No principle or RFC concerns.

  

_{Sent by Cursor Automation: Pre-review}