$ the-wire · showcase
OpenEnv 0.5.0 ships new public surface, capture proxy gains an admin key
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4)
![hf-security-analysis[bot]](https://avatars.githubusercontent.com/u/25720743?v=4)
13 people shipped this
github-actions[bot]
@github-actions[bot]
1 cited
burtenshaw
@burtenshaw
1 cited
Quentin Gallouédec
@qgallouedec
1 cited
Albert Villanova del Moral
@albertvillanova
1 cited
hf-security-analysis[bot]
@hf-security-analysis[bot]
1 cited
Tarek Ziade
@tarekziade
cursor
@cursor
mugenkyou
@mugenkyou
kashif
@kashif
dependabot
@dependabot
hf-security-analysis
@hf-security-analysis
hf-dependantbot-rollout
@hf-dependantbot-rollout
Arthur
@ArthurZucker
OpenEnv cuts 0.5.0 with the openenv.harbor package and new catalog, discover, and validate commands, while its capture proxy control plane stops accepting unauthenticated callers and TRL deletes four trainers it once marked experimental.
v0.5.0 huggingface/OpenEnv
The release notes list the restored RFC 008 manifest and validation contracts and the openenv validate skeleton, but the release PR is blunt about why this is 0.5.0 and not a patch: it "adds new public surface," and "a patch number would misrepresent it." 47 commits have landed since v0.4.2.
fix(harbor): require an admin key on the capture proxy control plane huggingface/OpenEnv
The capture proxy's session routes were gated by _admin_ok(), which admitted every caller whenever app.state.admin_key was unset, leaving POST /sessions (an open relay to the upstream engine), GET /sessions, and the per-session rollout and trace endpoints reachable. The fix requires an admin key on that control plane, so anyone running the proxy without one should set it before exposing it.
Remove the experimental BCO trainer (#7136) huggingface/trl
Quentin Gallouédec removed the experimental BCO trainer, the second of several deletions. Anything importing it from trl no longer has it, so pinning to the last version that shipped it is your only route if you still depend on the class.
Remove the experimental Nash-MD trainer (#7255) huggingface/trl
Albert Villanova del Moral removed the experimental Nash-MD trainer, following PRM and XPO. Four trainers that were labeled experimental have now been dropped across this batch of changes.
fix(ci): harden workflow files flagged on #7 (#8) huggingface/tokbench
The security analysis bot's hardening PR declared job permissions in .github/workflows/bench.yml, and it carries its own warning: "This narrows what the workflow can reach," because each job now gets only the scopes its steps were read to need and will fail on the next run if a step does something the analysis could not see. The same bot separately opened a Dependabot bump for actions/checkout ...
Action items
- → Set app.state.admin_key on the OpenEnv capture proxy before exposing it, or session-management routes stay open to an... huggingface/OpenEnv [immediate]
- → Replace imports of the BCO, PRM, XPO, and Nash-MD trainers removed from trl, or pin to a version that still ships them huggingface/trl [plan]
- → Verify the bench.yml job scopes still cover every step after the workflow hardening, since jobs will fail on the next... huggingface/tokbench [plan]
