feat(cubesandbox): exec-only provider (Phase 1, stacked on RFC #136)

[JackWeiw]

What

Phase 1 of the cubesandbox provider (Cloud Hypervisor microVM, lifecycle-capable), stacked on RFC #136 (feat/cubesandbox-provider-rfc). This PR ships the exec-only baseline — the same surface cloud e2b has today — over the native cubesandbox PyPI SDK.

Scope (Phase 1 — exec-only)

New opt-in leaf submodule src/env_provider/cubesandbox/:

• config.py — Config (CUBE_API_URL/KEY/TEMPLATE_ID/SANDBOX_DOMAIN env vars, template, idle TTL, ids_file, snapshot_dir) + setup_cube_env() (placeholder-credential skip; cube auth is optional, no CLI fallback)
• schemas.py — CubeSandboxState (handle cube_sandbox + lifecycle flags + creation metrics)
• manager.py — CubesandboxManager(BaseSandboxManager) SDK seams + class attrs + detect_from_file + check_alive (control-plane get_info)
• __init__.py — CubesandboxProvider(EnvironmentProvider) adapter: create_all/detect_existing/detect_from_ids/save_ids/check_alive/cleanup_all/cleanup_existing/prepare_env/exec + _to_instance + build_provider

Registration + config:

• bench_core.bench: --provider cubesandbox (CLI choices + lazy _build_provider import; kernel stays SDK-free)
• pyproject.toml: cubesandbox optional-dependency (cubesandbox>=0.7.0)
• config/common/replay{,-exec-only,-trajectory}.yaml: cubesandbox: blocks (CUBE_* env, idle TTL)

default_replay_mode="exec_only", vmm_type=None (vm_monitor deferred).

SDK surface (verified against source)

Read directly from CubeSandbox/sdk/python/cubesandbox/ (v0.7.0) — the manager seams bind to the real API, not assumptions:

• Sandbox.list() returns a flat list[dict] (key sandboxID) — not a paginator like e2b
• commands.run(cmd, user, timeout, cwd, envs) → CommandResult(stdout, stderr, exit_code); raises only on transport/timeout (no partial output) → exec maps timeout → exit 124 so the replay runner records a timed_out step and continues
• native pause(wait=True); resume via Sandbox.connect(id) (deprecated resume()); get_info() 404 → dead (clean check_alive, no envd exec round-trip)
• SnapshotInfo has no size fields → snapshot_sizes deferred to Phase 2 (returns None)

Tests

36 new tests, no live SDK/cluster needed (fakes injected via monkeypatch, mirroring test_e2b_manager.py / test_aenv_provider.py):

• test_cubesandbox_manager.py — SDK seams (create/envs=None, flat-list _list_existing, dict-key _external_id, connect _attach, exec probe, get_info check_alive) + base lifecycle (create_all / detect_existing / detect_from_file / cleanup_all / cleanup_existing)
• test_cubesandbox_provider.py — name/replay-mode/protocols, build_provider reads cubesandbox block, setup_cube_env placeholder skip, exec success/timeout→124/transport-raise/cwd+envs forwarding, save_ids, _to_instance template+id

Verification

python -m pytest src/bench_core/tests src/env_provider/tests   # full suite green
python -m ruff check src/env_provider/cubesandbox src/bench_core/bench.py
python -m ruff format --check src/env_provider/cubesandbox
python -m pre_commit run --files <phase-1 files>              # all hooks pass

Stacking

• Base: RFC 0002: Add cubesandbox provider (lifecycle-capable, Cloud Hypervisor microVM) #136 (RFC feat/cubesandbox-provider-rfc) — this PR targets that branch.
• Next: Phase 2 (separate PR off this branch) adds LifecycleCapable (pause/resume), EphemeralCapable (create_one/kill_one), SnapshotSizeCapable (returns None for v1), and flips default_replay_mode to lifecycle. vm_monitor integration (cube-hypervisor VMM type) is a follow-on.

When #136 merges to main, this PR retargets to main.