OpenSandbox live-test helpers

This directory is tier T4 of the provider end-to-end plan (中文). T0–T3 do not need this server. T5 is the Alibaba Cloud suite in the parent e2e/ directory.

Secret-free helpers for running the Agent Sandbox opensandbox provider against a real, local OpenSandbox server (Docker runtime). Design and rationale: docs/en/design/provider-opensandbox-mapping.md.

File	Purpose
`local-up.sh`	`up` (default) / `down` / `status` / `env`. Installs Docker (`docker.io`) and `opensandbox-server` into `$OSB_WORKDIR` if missing, starts both, waits for `/health`, prints `export` lines. Needs Linux, Docker-capable kernel, passwordless sudo (or root), internet.
`live_smoke.py`	Stdlib-only, 35 checks of every REST call the provider needs (auth, create/get/list/endpoint, SSE exec, background commands, files, renew, pause/resume, kill).
`sdk_smoke.py`	Same flow through the official `opensandbox` Python SDK (`pip install opensandbox`); oracle only.

cd python-sdk/e2e/opensandbox
./local-up.sh up
eval "$(./local-up.sh env)"
python3 live_smoke.py          # prints ALL OK on success
python3 sdk_smoke.py           # optional
./local-up.sh down

Tunables (environment): OSB_SERVER_VERSION (1.1.1), OSB_EXECD_IMAGE (opensandbox/execd:latest; v1.1.0 lacks argv), OSB_PORT (8080), OSB_API_KEY (local-dev-key, not a secret), OSB_WORKDIR (~/.cache/agent-sandbox/opensandbox-local), OSB_WITH_EGRESS, OSB_WITH_CODE_INTERPRETER. Test overrides read by the smokes: OPENSANDBOX_BASE_URL, OPENSANDBOX_API_KEY, OPENSANDBOX_TEST_IMAGE, OPENSANDBOX_DOMAIN.

These files are not pytest tests (no test_* names) and are never collected by make e2e.

Provider plugin E2E

Two further scripts exercise the exec-provider plugin against the same local server. The plugin source is alibaba-agent-sandbox/awesome-providers (opensandbox/). The scripts clone that catalog (or use ABX_OPENSANDBOX_PLUGIN / ABX_AWESOME_PROVIDERS); they are not a copy of the plugin. They use a throwaway HOME (your real ~/.abx is never touched) and read the server from the environment: ABX_OPENSANDBOX_BASE_URL (default http://127.0.0.1:8080) and the API key from ABX_OPENSANDBOX_API_KEY, OPEN_SANDBOX_API_KEY or OPENSANDBOX_API_KEY. Needs Node >= 20, git, pytest and an importable agent_sandbox (the scripts add the sibling client-sdk/src to PYTHONPATH).

File	Purpose
`provider_e2e.py`	Stages `node-unit`, `node-live`, `install`, `kit` (shared `ProviderContractTests --provider-under-test opensandbox`), `sdk` (`Sandbox.create(provider="opensandbox")` scenarios) and `env` (canonical `ABX_OPENSANDBOX_` beats native `OPEN_SANDBOX_`). `--stages a,b` selects a subset.
`provider_install_e2e.py`	Install flows with `python -m agent_sandbox.providers`: local path, git `file://` bare repo pinned to a commit, tree-digest receipt, tamper detection, update (check / re-consent / rollback), uninstall, `--purge`, `--link`. Needs only `git` and Node; uses the server for one sandbox round trip when it is reachable.

./local-up.sh up && eval "$(./local-up.sh env)"
python3 provider_e2e.py                # the OPENSANDBOX_* names printed by local-up.sh env are accepted too
python3 provider_install_e2e.py

Both exit non-zero when a check fails. provider_install_e2e.py used to print non-fatal [GAP ] / [BUG ] lines for installer defects found by the authoring evaluation (docs/en/design/provider-authoring-evaluation.md); all of them are fixed and are now ordinary checks. The kit stage runs python -m agent_sandbox.providers test opensandbox. Like the files above, these are not pytest tests and make e2e never collects them.