Skip to content

[VDR][v0.6.0] CLI-00845c9f · gym env test with no arguments silently starts a ~2-hour, ~145 GB all-servers run #2690

Description

@anwithk

Severity: 🟠 high | Category: CLI | Finding ID: CLI-00845c9f

Location: nemo_gym/cli/main.py:490

Problem

gym env test with no NAME and no --resources-server falls through to nemo_gym.cli.env:test_all, which begins building an isolated venv for every one of the 164 resources servers and running their test suites. There is no confirmation prompt, no ETA warning, and gym env test --help says only 'Test a manifest-backed workload or resources server' with an optional [NAME] -- nothing signals the fan-out. The sibling gym env publish, which takes the identical optional NAME positional, correctly refuses with 'requires a catalog name'. Each venv measured ~900 MB inside the repo tree, so a completed run is roughly 145 GB.

Suggested fix

Require an explicit opt-in for the all-servers path (gym env test --all, or a y/N confirmation showing '164 servers, est. ~1h50m, ~145 GB'), and mirror gym env publish's behaviour by erroring when no target is given. At minimum document the no-argument fan-out in --help and print the server count and estimate before the first venv is created.

Evidence

$ gym env test  ->  'Running tests:   1%|          | 1/164 [00:40<1:51:21, 40.99s/it]'; du -sh resources_servers/aalcr/.venv = 924M, abstention/.venv = 897M

Found by the cli-explorer critic in DX review Gym-1-20260821203248 — incremental from v0.5.0 to fa0d25158. Parent tracker #2214.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    Status
    Label Sync

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions