provider: drop --gpu-mode from CAPE run face (flag removed upstream; pool-level sharing)

[Meirtz]

Upstream change

CAPE removed the cape run --gpu-mode flag entirely: requests now inherit exclusive/shared GPU behavior from the target pool, with no per-request override. In the same release, cape status gained three cotenancy fields: gpu_mode (derived, null | "exclusive" | "shared"), gpu_cotenant_count, and gpu_cotenant_counts.

Failure mode without this fix

The provider unconditionally emitted --gpu-mode whole on every cape run. Against the current CLI that is an unknown-argument error — argparse exits 2 with a usage message before anything is submitted — so every create() fails and no sandbox can be created at all.

Changes

• agentix/provider/cape.py: drop the --gpu-mode whole pair from the run argv and from the verified-contract docstring.
• tests/test_cape_provider.py:
  • Fake CLI parser no longer accepts --gpu-mode, matching the real parser by construction (unknown flag → rc=2, usage on stderr).
  • New test_fake_cape_rejects_dropped_gpu_mode_flag documents the exit-2 behavior of the removed flag.
  • New test_run_face_never_emits_dropped_gpu_mode_flag regression test asserts the run argv never contains --gpu-mode (gpu unset and gpu=4).
  • Run-face assertion updated from --gpu-mode whole present to --gpu-mode absent.
  • Fake cape status now emits the three new cotenancy fields, proving the provider's status parsing tolerates (ignores) them — it only reads state and the request_id echo, so no parsing change was needed.
• README.md: flag list updated; Contract status section notes the upstream flag removal (pool-level exclusive/shared) and the new status cotenancy fields tolerated by our parser.

Gates

• uv run ruff check . — clean
• uv run pyright — 0 errors
• uv run pytest plugins/providers/cape/tests -q — 31 passed
• uv run pytest tests/ plugins/ -q — green
• uv lock --check — up to date

🤖 Generated with Claude Code