Use duplicate_repo for sandbox Spaces

[lewtun]

Summary

• Replace deprecated HfApi.duplicate_space with HfApi.duplicate_repo(..., repo_type="space") for sandbox Space creation.
• Map sandbox hardware and sleep-time options to the new space_hardware and space_sleep_time arguments.
• Update the sandbox unit-test fakes and assertions to cover the new Hub API call shape.

Tests

• uv run pytest tests/unit/test_sandbox_private_spaces.py
• uv run ruff check agent/tools/sandbox_client.py tests/unit/test_sandbox_private_spaces.py
• uv run ruff format --check agent/tools/sandbox_client.py tests/unit/test_sandbox_private_spaces.py
• uv run ruff check --exclude train_smollm2.py .
• uv run ruff format --check --exclude train_smollm2.py .

Note: the exact repo-wide uv run ruff check . and uv run ruff format --check . commands are currently blocked locally by unrelated untracked train_smollm2.py formatting/lint issues.

[github-actions]

Claude finished @lewtun's task in 2m 25s —— View job

---

Claude PR Review

No blocking issues — 1 P1, 1 P2.

This is a tightly scoped migration from the deprecated HfApi.duplicate_space to the unified HfApi.duplicate_repo(..., repo_type="space"), with the hardware/sleep_time kwargs renamed to their space-namespaced variants. Production code, unit-test fakes, and the PEP-723 inline-script dep pin (huggingface_hub>=1.12.0) are all updated consistently. pyproject.toml already requires huggingface-hub>=1.12.0 and uv.lock is on 1.12.0, so the runtime version is in sync.

I confirmed there are no remaining duplicate_space references and no other call sites use the old hardware=/sleep_time= kwargs.

---

P1 — Hub kwarg shape isn't actually exercised by the tests

agent/tools/sandbox_client.py:615-625 calls api.duplicate_repo(repo_type="space", space_hardware=..., space_sleep_time=...), but the four updated unit-test fakes at tests/unit/test_sandbox_private_spaces.py:25, 71, 111, 158 all use def duplicate_repo(self, **kwargs) — which silently accepts any kwarg name. If the real huggingface_hub 1.12.0 method names them differently (e.g. without the space_ prefix, or expects them inside a nested dict), the unit tests will still pass while the live call 400s.

This is the same mock-shape risk that existed pre-PR; I'd just suggest one of the following before merging, given it was the central reason for the change:

• Run uv run python -c "import inspect, huggingface_hub; print(inspect.signature(huggingface_hub.HfApi.duplicate_repo))" and paste it in the PR body, or
• Do one manual sandbox creation against the real Hub, or
• Tighten the fake's signature to def duplicate_repo(self, *, from_id, to_id, repo_type, private, space_hardware, space_sleep_time=None) so a typo in the production kwarg name fails tests.

Cite: agent/tools/sandbox_client.py:615-625, tests/unit/test_sandbox_private_spaces.py:25,71,111,158.

P2 — Inline-script dep header jumped from 0.20.0 → 1.12.0

agent/tools/sandbox_client.py:4 bumps the PEP-723 dep range across a major version. This is fine — it matches the project pin — but anyone who has been running this file standalone via uv run agent/tools/sandbox_client.py will now resolve a major-version-newer huggingface_hub. Worth a one-line note in the release notes if there is one.

---

· branch codex/fix-hf-duplicate-repo

[lewtun]

Addressed the P1 in c6e6942 by tightening the sandbox FakeApi.duplicate_repo methods to keyword-only signatures matching the Hub API shape, so old or misspelled kwargs fail tests instead of being swallowed by **kwargs.

Also verified the installed Hub signature locally:

(self, from_id: 'str', to_id: 'str | None' = None, *, repo_type: 'str | None' = None, private: 'bool | None' = None, visibility: 'RepoVisibility_T | None' = None, token: 'bool | str | None' = None, exist_ok: 'bool' = False, space_hardware: 'SpaceHardware | None' = None, space_storage: 'SpaceStorage | None' = None, space_sleep_time: 'int | None' = None, space_secrets: 'list[dict[str, str]] | None' = None, space_variables: 'list[dict[str, str]] | None' = None, space_volumes: 'list[Volume] | None' = None) -> 'RepoUrl'

P2 is noted; I do not see a release-notes file in this repo, and the PR body already calls out the standalone PEP-723 header bump.

[github-actions]

Claude finished @lewtun's task in 1m 7s —— View job

---

Claude PR Review (re-review of c6e6942)

LGTM.

• Read updated diff vs origin/main
• Check agent/tools/sandbox_client.py call site
• Check tightened test fakes
• Verify P1 fix from prior review
• Post review

---

P1 from the prior review is addressed. The four test fakes at tests/unit/test_sandbox_private_spaces.py:47, 110, 159, 223 are now strict keyword-only signatures (*, from_id, to_id, repo_type, private, space_hardware, space_sleep_time=None), so a renamed/typo'd kwarg in agent/tools/sandbox_client.py:615-625 would raise TypeError in tests instead of being silently swallowed. Three of the four tests also assert repo_type == "space" on the captured call. The kwarg names (from_id, to_id, repo_type, private, space_hardware, space_sleep_time) match the huggingface_hub 1.12.0 signature you pasted in the thread.

P2 (PEP-723 dep header 0.20.0 → 1.12.0) — accepted as noted; no release notes file exists in this repo.

Confirmed there are no remaining duplicate_space references in the tree.

--- · branch codex/fix-hf-duplicate-repo