Add ModalProvider(ContainerProvider)

[joyliu-q]

Summary

Sample PR of adding Modal as a container provider to OpenEnv.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation
• New environment
• Refactoring

Alignment Checklist

Before submitting, verify:

• I have read .claude/docs/PRINCIPLES.md and this PR aligns with our principles
• I have checked .claude/docs/INVARIANTS.md and no invariants are violated
• I have run /pre-submit-pr (or bash .claude/hooks/lint.sh and tests) and addressed all issues
  • :flag: I ran this and actually a bunch of unrelated PRs failed lint formatting. We might want to fix this.

RFC Status

• Not required (bug fix, docs, minor refactoring)
• RFC exists: #___
• RFC needed (will create before merge)

Test Plan

Claude Code Review

Usage

  """E2E check for ModalProvider against real Modal infra (echo_env)."""

  import functools

  import modal
  import requests

  # --- WORKAROUND: Modal's encrypted tunnel doesn't forward WS ping/pong frames,
  # so disable the websockets client keepalive for tunnelled connections. ---
  import openenv.core.env_client as _ec
  _ec.ws_connect = functools.partial(_ec.ws_connect, ping_interval=None)

  from openenv.core.containers.runtime.modal_provider import ModalProvider

  DOCKERFILE = "envs/echo_env/server/Dockerfile"   # path relative to repo root


  def run(use_v2: bool) -> None:
      print(f"\n=== ModalProvider e2e (use_sandbox_v2={use_v2}) ===")

      # ---- provider config ----
      provider = ModalProvider(
          app_name="openenv-e2e",     # Modal app (created if missing)
          use_sandbox_v2=use_v2,      # False -> Sandbox.create, True -> _experimental_create
          timeout=600,                # max sandbox lifetime (s)
          cpu=1.0,                    # optional
          memory=1024,                # optional; ignored when use_sandbox_v2=True
          # cmd="cd /app/env && uvicorn server.app:app --host 0.0.0.0 --port 8000",  # optional override
      )

      image = ModalProvider.image_from_dockerfile(DOCKERFILE)  # context defaults to envs/echo_env/
      print("image:", image)

      with modal.enable_output():                 # stream the build logs
          base_url = provider.start_container(image)
      print("base_url:", base_url)

      try:
          provider.wait_for_ready(base_url, timeout_s=180)
          print("health ready ✅")

          r = requests.get(f"{base_url}/health", timeout=10)
          assert r.status_code == 200, r.status_code
          print(f"GET /health -> {r.status_code} ✅")

          from echo_env.client import EchoEnv
          with EchoEnv(base_url=base_url).sync() as env:   # .sync() — client is async by default
              result = env.reset()
              print("reset ->", type(result).__name__, "✅")
          print(f"MCP round-trip OK (v2={use_v2}) ✅")
      finally:
          provider.stop_container()
          print("sandbox terminated")


  if __name__ == "__main__":
      run(use_v2=False)   # stable Sandbox API
      run(use_v2=True)    # Sandbox v2 (beta)
      print("\nALL E2E CHECKS PASSED ✅")

Run it (from the repo root, src + envs on the path, with modal available):

  PYTHONPATH=src:envs uv run --extra core --with modal python modal_e2e.py

[burtenshaw]

Thanks for the PR. I'm down with the direction but I think this should first align with the provider-neutral cloud sandbox shape from #793.

Core asks:

• Make Modal work through stock EnvClient. The E2E currently relies on monkeypatching WebSocket ping_interval which imo we should avoid. lmk if we shoul make changes in core to facilitate sandbox services without the ws ping route.
• Mirror ACA lifecycle semantics: reject double-starts, implement close()/context cleanup, and avoid masking startup errors during cleanup. See feat: add Azure Container Apps cloud sandbox provider #793
• Treat sandbox logs and tunnel URLs as sensitive by default; redact or require explicit opt-in before surfacing logs.
• Consider a small Modal SDK adapter/feature check, especially because Sandbox._experimental_create is private.
• fix that PR example passes cpu/memory to ModalProvider(...), but the constructor rejects those; they currently must be passed to start_container()

Also, it would be nice to bring a simple example of inference with the modal sandbox, we can then build a training example from it.