[Chen Yufeiyang (feiiiiii5)]

AglRolloutManagerBase._delete_rollout() treats a failed HTTP DELETE as success: httpx.Client.delete does not raise for 4xx/5xx, and the method never checks the response. A 500 leaves the completed rollout and its events on the server for the rest of the training run, while the caller advances as if cleanup succeeded.

The fix calls raise_for_status() inside the existing try, so the failure goes through the method's existing failed to delete rollout ... branch. Deletion remains best-effort, matching delete_model() in the same class, but a rejected DELETE is no longer silent.

Test: tests/verl/test_agl_rollout_manager.py::test_delete_rollout_reports_http_errors uses an httpx.MockTransport returning 500; it fails on 05c7fe1 because no failure is printed and passes here. The server/controller/client and rollout-manager suites are 88 passed; ruff check/format --check are clean.

Details

The failure is surfaced but not retried, and a 3xx response is not treated as an error by raise_for_status(); both match the class's documented best-effort cleanup behavior.

[Copilot]

🟢 Approval recommended

The focused fix preserves best-effort cleanup and includes a targeted regression test, with no unresolved review findings.

No open issues 🎉

What changed in this PR

This PR makes failed rollout deletion visible while preserving the rollout manager’s best-effort cleanup behavior.

Changes:

• Checks the DELETE response status so HTTP failures reach the existing error-reporting branch.
• Adds a regression test verifying that a mocked HTTP 500 response produces a deletion-failure message.

File	Description
tests/​verl/​test_agl_rollout_manager.py	Adds regression coverage for reporting failed rollout deletion.
agentlightning/​verl/​agl_rollout_manager.py	Validates the DELETE response inside the existing exception handler.

_{Review effort: Balanced}