You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Planned release: Thursday, September 24, 2026 at 10:00 Europe/Brussels (08:00 UTC).
Candidate base: 39fdf5cd (main tip at the time of writing). 6 commits since v0.5.0.
This is the rolling release PR for the week; it will be refreshed onto the final
candidate head before merge.
Why 0.6.0 and not 0.5.1: the only package change since v0.5.0 is new public
surface — NovitaSandboxProvider (1,207 lines under src/openenv/core/containers/runtime/), a documented entry in docs/source/reference/core.md, and the new openenv[novita] extra. That is added
functionality, which is a minor bump under the same standard applied when 0.5.0
was cut rather than a quiet 0.4.3. If you would rather treat opt-in providers as
non-API and ship 0.5.1, it is a one-line change to this PR — please say so before
Wednesday so the notes and tag match.
Release notes
Breaking change — LLM endpoint URLs:--llm-endpoint is now a full base URL, and the implicit port 8000 has been removed. http://localhost now uses the standard HTTP port 80; to retain the previous local behavior, use http://localhost:8000 or --llm-port 8000 (#1189, landed via #1237).
Novita AI sandbox provider: run an OpenEnv server inside a Novita sandbox and
connect over wss://, from either a registry image or a local Dockerfile
(image_from_dockerfile), with multi-stage Dockerfiles replayed as a single stage
for Novita's template parser. HTTPS/WSS transport is enforced and captured sandbox
output is withheld from raised errors unless surface_server_logs=True. Installs
via pip install openenv[novita] (#1191).
Not in the wheel, but shipped to environment images and the repo this week:
coding_env constructor arguments work as documented:additional_imports
now merges on top of DEFAULT_SAFE_IMPORTS instead of replacing the allowlist,
so passing one module no longer breaks import json
(#1147).
Environment lockfile security updates: anyio 4.14.2 for the critical TLSStream IDNA certificate-spoofing advisory plus two lower-severity fixes
(#1201, #1202, and #1197 for the remaining 31
lockfiles), and soupsieve 2.9.2 for the polynomial-ReDoS advisories
(#1196, extended to browsergym_env and finrl_env in chore(deps): aggregate envs Dependabot updates #1197).
Version decision. Confirm 0.6.0 (recommended) or say 0.5.1 and I will
change the one line.
Known debt accepted onto main with #1191, all scoped to the
opt-in Novita provider and its examples, none of it re-blocking in my
assessment — listed so the decision to ship it is explicit:
unbraced $ARG substitution during Dockerfile flattening can corrupt a FROM
line ($BASE matching before $BASE_IMAGE); examples/novita_tbench2_simple.py
calls wait_for_ready outside its try/finally, so a readiness timeout leaks a
paid sandbox; concurrent starts and failed kills can orphan sandboxes; flattened
stages carry builder filesystem/ENV/USER state into the runtime image; and the
RFC 002 provider amendment (root execution, unrestricted egress, HTTP-only
readiness) is still unratified.
Dispatch and pass publish-testpypi.yml from this exact branch with a unique 0.6.0 pre/dev suffix.
Reconcile main immediately before merge; rerun all required checks if the
head or base changes.
Not in this candidate: every open small fix that touches the package
(#1145, #1198, #1200, #1203, #1208, #1209) comes from a fork whose
workflows are still action_required, so none of them has ever run repository CI.
A maintainer clicking Approve and run is the only way any of them becomes
release-eligible.
Release Checklist
Before opening this PR
pyproject.toml version changed from 0.5.1.dev0 → 0.6.0
hf-staging/ is NOT in this PR's diff
No print(), breakpoint(), or TODO added to release-critical paths
Release notes include user-facing changes and linked PRs
CI gates (must be green before merge)
test passes on Python 3.11
test passes on Python 3.12
lint passes (usort + ruff)
Package CI builds, checks, and smoke-tests wheel/sdist installs
TestPyPI validation (before merging)
Manual dispatch of publish-testpypi.yml from this branch
TestPyPI published a unique pre/dev version such as 0.6.0.devN
Tag v0.6.0 pushed against the exact merge commit on main
publish-pypi.yml completed successfully from the tag
GitHub Release created by the successful PyPI publish workflow
pip install openenv==0.6.0 from production PyPI verified (import, __version__, openenv --help, real Echo reset/step)
Next development-version bump PR opened and processed
RFC Status
Not required (release maintenance)
Note
Low Risk
Metadata-only version and lockfile sync with no runtime or API changes in the diff.
Overview Release version bump for the openenv package: pyproject.toml moves from 0.5.1.dev0 to stable 0.6.0, aligning the published package version with the v0.6.0 release.
The editable openenv entries in envs/grid_world_env/uv.lock and tests/validation_runtime/uv.lock are updated to 0.6.0 so those lockfiles stay consistent with the root project. No application or library code changes appear in this diff.
Reviewed by Cursor Bugbot for commit 0034202. Bugbot is set up for automated code reviews on this repo. Configure here.
Sets the published version to 0.6.0 for the Thursday 2026-09-24 cut. The only
package change since v0.5.0 is the new NovitaSandboxProvider public surface,
which a patch number would misrepresent; see the PR body for the version
decision and outstanding blockers.
Co-authored-by: benjamin.burtenshaw <benjamin.burtenshaw@huggingface.co>
grid_world_env is the only environment whose lock resolves the root openenv
package from an editable path, so its uv.lock pins the project version. Left at
0.5.1.dev0 it makes `uv lock --check` fail in that directory; `uv sync
--frozen` skips the freshness check, so no CI job catches it.
Co-authored-by: benjamin.burtenshaw <benjamin.burtenshaw@huggingface.co>
envs/grid_world_env/uv.lock is the only environment lock that resolves the root openenv package from an editable path (source = { editable = "../../" }), so it
records the project version. With the version bump alone it is left at 0.5.1.dev0, and uv lock --check fails in that directory. After the one-line
change to 0.6.0 it passes, and uv sync --frozen --all-groups --all-extras --dry-run --no-install-project passes there too. Both verified locally.
No repository job catches this on its own: uv sync --frozen skips the freshness
check, and validate-env-locks is both scoped to changed env dependency files and
currently a no-op (it filters paths with rg, which is absent on ubuntu-latest,
so it logs rg: command not found and exits 0 — see #1197 for the detail). Treat
the green check on this PR as vacuous and the local run above as the real gate.
Follow-up for a maintainer:publish-pypi.yml's post-release-bump job edits
only pyproject.toml, so it will re-stale this lock the moment it writes 0.6.1.dev0. Every future version transition will need the same second line until
that job also refreshes the GridWorld lock.
Gate status on this exact head
All 10 reporting checks pass: test 3.11 and 3.12, test-thinkingbox-env 3.11 and
3.12, lint, check-env-docs, snapshot, Analyze (python), and Build and smoke-test package — the last one builds the wheel and sdist, runs twine check,
installs each into a clean venv, asserts openenv.__version__ equals the installed
metadata version, and runs openenv --help. mergeStateStatus is CLEAN.
A real Echo reset/step smoke also passes against this tree: reset metadata is {'status': 'ready', 'message': 'Echo environment ready!'}, list_tools returns
both tools, the step envelope's data is the echoed message, call_tool returns
the bare string, and state reports step_count=1.
Still blocking publication
Version decision — 0.6.0 (recommended) or 0.5.1. One line either way.
The Novita debt list in the body above needs an explicit ship-or-fix call.
publish-testpypi.yml has not been dispatched from this branch, and the fresh
production-PyPI install check obviously cannot run before publication.
Staying a draft until those clear. Nothing here should be read as a release
happening: this is a validated candidate, not a published version.
The reason will be displayed to describe this comment to others. Learn more.
Stale comment
Independent release validation at fee469c4 — all local gates green
Ran the full release gate set off-CI against this exact head (base main 7e591317, unchanged since the last refresh). Recorded here because I cannot edit the PR body.
openenv 0.6.0, __version__ == installed metadata, openenv --help OK
Clean-venv sdist install
same, openenv --help OK
openenv[novita] extra
installs; NovitaSandboxProvider imports
Tests (CI-equivalent: same ignores + not integration and not network and not docker)
2761 passed, 95 skipped, 37 deselected
Lint (CI-equivalent usort format + ruff format)
no resulting git diff
Real Echo reset/step/state/call_tool
PASSED
On lint: usort check alone reports tests/envs/test_grid_world.py and tests/envs/test_julia_env.py as unsorted, which reads like standing debt. It is not — CI runs usort formatthenruff format, and ruff reverts usort's change, so the combined pipeline is a no-op and git diff is empty. No action needed.
Echo smoke covered reset metadata (status=ready), both advertised tools (echo_message, echo_with_length), a step whose payload echoes the sent message, step_count accounting, direct call_tool, rejection of an unknown tool, and a second reset clearing step_count to 0.
validate-env-locks passed vacuously on this PR
Worth stating explicitly since this PR changes a lockfile. The green check at fee469c4 (job log) contains:
rg: command not found
No environment dependency files changed.
The job filters changed paths with rg, which is not installed on ubuntu-latest, so mapfile yields an empty array and the job exits 0 without validating anything. The envs/grid_world_env/uv.lock change in this PR was never checked by CI. I validated it locally instead: uv lock --check resolves 117 packages cleanly and uv sync --frozen --all-groups --all-extras --dry-run --no-install-project succeeds.
Also confirmed while checking: grid_world_env is the only env lock with source = { editable = "../../" }, so it is the only one that tracks the root project version. The other 37 env locks pin published openenv versions and are unaffected by the bump.
The one Tier 1 item that actually reaches the wheel is the unbraced ARG substitution order in _resolve_from_references. Reproduced against the shipped code:
Dockerfile shape
Result
ARG BASE then ARG BASE_IMAGE, FROM ${BASE_IMAGE}
correct
ARG BASE then ARG BASE_IMAGE, FROM $BASE_IMAGE
FROM python:3.12_IMAGE
longer ARG declared first, unbraced
correct
single ARG, unbraced
correct
So it needs a user Dockerfile with two global ARGs where one name is a prefix of the other, the shorter declared first, and the longer used unbraced. No in-repo Dockerfile can trigger it — there are zero FROM $ARG (unbraced) occurrences across envs/ and src/, and zero Dockerfiles with prefix-colliding global ARGs.
Recommendation: this is a latent bug in brand-new opt-in surface, not a regression, and does not warrant holding 0.6.0. Ship and track the fix (sort arg_defaults longest-name-first) as a follow-up.
Pre-existing, ships again, not a blocker
The README quick start is the PyPI long description, and its first example raises. Run verbatim against a live Echo server:
result.observation.echoed_message → AttributeError: 'GenericMCPObservation' object has no attribute 'echoed_message'. The text is at result.metadata["message"].
result.observation.result is a dict, not the "Hello, World!" string the comment claims; the string is at result.observation.result["data"].
This is not new — it is present at v0.4.2 and v0.5.0 and is already live on the PyPI 0.5.0 page. Flagging rather than fixing here, to keep this PR to the version change; it should not gate the release.
Still outstanding before this can merge
Version call — 0.6.0 vs 0.5.1 (rationale in the PR body). One-line change either way.
Ship-or-fix on the Novita debt — recommendation above is ship.
TestPyPI dispatch — required by the release checklist and I get HTTP 403 on workflow_dispatch, so this needs a maintainer to run publish-testpypi.yml from cursor/openenv-release-management-7ba7.
Nothing is tagged or published. PyPI and the GitHub release remain 0.5.0.
The reason will be displayed to describe this comment to others. Learn more.
Stale comment
Monday check-in — candidate re-validated, and the cargo list changed
main is still 7e591317 and this PR is still fee469c4, so the candidate did not drift over the weekend. I re-ran the release gate from scratch at this exact head rather than carrying Sunday's result forward:
The grid_world_env lock was checked by hand again because CI's validate-env-locks still passes vacuously (rg: command not found → "No environment dependency files changed"), so this PR's own lockfile edit is not actually covered by the green check above it.
The "not in this candidate" list in the PR body needs one correction.
#1145 is now approved from my side and it is package cargo — it changes src/openenv/core/env_client.py. I withdrew my earlier change request after measuring the flagged edge case against main instead of against an idealised baseline: on the path I called a regression, main is equally wedged and additionally swallows the real configuration error, while this PR surfaces it. Full evidence is on that PR. It fixes the #1144 provider leak, so it is worth folding into 0.6.0 if its CI greens before the cut.
#1198 is also approved but stays out of the wheel — it is envs/** plus one test. It makes #1174's state_cls fix actually reach users (on main, maze_env's /schema and /state drop current_position, exit_cell, status and done, which the WebSocket frame returns). No packaging risk either way.
Both are fork PRs whose repository CI has never run, so neither can merge until a maintainer clicks Approve and run. That click has a lead time if either is to make Thursday.
Blockers unchanged
The three items in the PR body still need @burtenshaw: the version call (0.6.0 vs 0.5.1), Novita ship-or-fix (my recommendation is still ship — the unbraced $ARG collision is latent and no in-repo Dockerfile can trigger it), and the TestPyPI dispatch, which I cannot run (workflow_dispatch returns 403 for this token, and that workflow has no other trigger).
One housekeeping note on this PR's own text: the body says "Candidate base: 39fdf5cd … 6 commits since v0.5.0". It is now 7 commits and the base is 7e591317; the added commit is #1197, which the release notes already list. The notes are correct; only that one line is stale, and it will be rewritten when this PR is refreshed onto the final candidate head.
The reason will be displayed to describe this comment to others. Learn more.
Stale comment
Candidate reconciled at 12:20 UTC Monday. I could not edit the PR description (the app token is refused on updatePullRequest), so the corrections to the notes are here.
The description is now stale in three places.
Candidate base is 228bd147, not 39fdf5cd; this branch merges it at cf88ef23. That is 8 commits since v0.5.0, not 6.
#1145 merged to main at 07:56 UTC today, so it is in this candidate — it should come out of the "not in this candidate" list and into the release notes. Suggested entry:
Failed session startup no longer leaks a provider: when new_session() cannot bring up a client, EnvClient stops the container or sandbox it started instead of leaving it running. Cleanup is retry-safe, and a client whose cleanup is still pending refuses to reconnect to the old base URL rather than overwriting the only resource handle (#1145).
#1208 is closed unmerged, so it should drop off the fork-CI list. The remaining package-touching fork PRs are #1198, #1200, #1203, #1209 and #1217, all still action_required.
Measured severity for the $ARG debt item, so that decision does not rest on a description alone. Reproduced against _resolve_from_references exactly as shipped on 228bd147:
Dockerfile
Resolved FROM
ARG BASE_IMAGE=… + FROM ${BASE_IMAGE}
FROM ghcr.io/huggingface/openenv-base:latest
ARG BASE=python:3.11-slim, ARG BASE_IMAGE=… + FROM ${BASE_IMAGE}
FROM ghcr.io/openenv-base:latest
ARG BASE=python:3.11-slim, ARG BASE_IMAGE=… + FROM $BASE_IMAGE
FROM python:3.11-slim_IMAGE
So it needs all three of: the opt-in provider's image_from_dockerfile path, two global ARGs where one name prefixes the other, and unbraced $NAME in FROM. No in-repo Dockerfile uses the unbraced form, and the failure mode is a hard build error on an invalid reference, not a silent wrong image. I agree with the existing recommendation: ship 0.6.0 and fix it in a follow-up.
State right now: head cf88ef23, CLEAN, all checks green (tests 3.11/3.12, lint, Package CI, env locks, docs, CodeQL). Still draft, correctly — the two things standing between this and Thursday are the 0.6.0 vs 0.5.1 call and a TestPyPI run from the final head.
The reason will be displayed to describe this comment to others. Learn more.
Stale comment
TestPyPI validation is done — that ask can come off Ben's list.
We do not need a manual workflow_dispatch for this. The same mechanism used for v0.4.2 and v0.5.0 still works: push the candidate to a testpypi/<version> branch carrying the nine-line push: testpypi/** trigger, which the publish workflow accepts without actions: write. I pushed testpypi/0.6.0 at 683e003a, which is this PR's head cf88ef23 plus that trigger commit and nothing else (git diff --stat cf88ef23 683e003a shows only .github/workflows/publish-testpypi.yml, +9 lines). The branch is a scratch branch and is not proposed for main.
Run 35598458908 published 0.6.0.dev71. Both jobs succeeded, including the workflow's own install-from-TestPyPI check.
Independently verified afterwards, outside the workflow:
Clean Python 3.11 venv, wheel installed with the new extra (openenv[novita], dependencies resolved from PyPI, never TestPyPI): importlib.metadata.version("openenv") and openenv.__version__ both report 0.6.0.dev71, the package imports from the fresh venv, and openenv --help works.
NovitaSandboxProvider imports from the installed wheel, as do openenv.harbor, openenv.discovery and openenv.validation. This is the first candidate carrying the novita extra, so that resolution path is the one worth proving early.
Real Echo round trip against the installed package (repo envs/ on PYTHONPATH, core from the wheel): reset() returned {'status': 'ready', 'message': 'Echo environment ready!'}, discovery returned echo_message and echo_with_length, and an echo_message step round-tripped release-0.6.0.dev71-smoke through both step() and call_tool().
Release mutations were serialized behind the automation-locks/openenv-release-0.6.0 branch lock, which is now released.
What is left before Thursday is only the version decision (0.6.0 vs 0.5.1) and explicit acceptance of the documented Novita debt. If the version changes, this is one line plus a re-run of the same TestPyPI branch; if it stays 0.6.0, Thursday is merge, tag, publish, verify.
The reason will be displayed to describe this comment to others. Learn more.
Stale comment
Candidate refreshed to 8bbb8542 on main ca4219d2
The PR body above is stale and I cannot edit it (updatePullRequest is 403 for this token), so treat this comment as the current release notes and blocker list. 13 commits since v0.5.0.
Release notes for v0.6.0
In the wheel and sdist
Novita AI sandbox provider. Run an OpenEnv server inside a Novita sandbox and connect over wss://, from either a registry image or a local Dockerfile (image_from_dockerfile), with multi-stage Dockerfiles replayed as a single stage for Novita's template parser. HTTPS/WSS transport is enforced and captured sandbox output is withheld from raised errors unless surface_server_logs=True. Installs via pip install openenv[novita]. (#1191)
RFC 008 Level 2 validation contracts.openenv.validation gains openenv.validation.runtime contract types, the manifest-v2 / report-v2 / runtime-plan JSON schemas, and a severity-v2 policy, so an environment can declare and be checked against a Level 2 contract. (#1178)
Client no longer leaks a sandbox when session startup fails.EnvClient tears the provider down if the session cannot be established, instead of leaving the container or sandbox allocated. (#1145)
Packaging metadata. The PyPI long description picks up the Miles entry added to the root README. (#1220)
Not in the wheel, but shipped to environment images, CI, and docs this week
coding_env constructor arguments work as documented: additional_imports merges on top of DEFAULT_SAFE_IMPORTS instead of replacing the allowlist, so passing one module no longer breaks import json. (#1147)
Environment lockfile security updates: anyio 4.14.2 for the critical TLSStream IDNA certificate-spoofing advisory plus two lower-severity fixes (#1201, #1202, #1197 for the remaining 31 lockfiles), and soupsieve 2.9.2 for the polynomial-ReDoS advisories (#1196).
The release pipeline now pins actions/upload-artifact to v7.0.1 by immutable SHA, so a future v7.0.x cannot change the publish path mid-release. (#1221)
This PR's own diff is still only the version change: 0.5.1.dev0 to 0.6.0 in pyproject.toml and the grid_world_env lockfile.
Validation at this head
Exact-head repository CI: 11/11 green (test 3.11 and 3.12, lint, Package CI, validate-env-locks, check-env-docs, snapshot, CodeQL, ThinkingBox 3.11 and 3.12). build is skipped only because the PR is still a draft.
Local CI-equivalent suite: 2931 passed, 95 skipped, 41 deselected. ruff format and ruff check clean. usort still reports tests/envs/test_grid_world.py and tests/envs/test_julia_env.py; I confirmed both are identical on main, so it is pre-existing and not candidate cargo.
uv build plus twine check: both wheel and sdist PASSED. Clean 3.11 venv installs of each, with imports (openenv.harbor, openenv.discovery, openenv.validation, NovitaSandboxProvider, openenv.validation.runtime.contracts) and openenv --help / harbor / catalog / discover / validate all passing.
TestPyPI 0.6.0.dev91 published from testpypi/0.6.0 at 5e95203a, whose only tree delta from this candidate is the push-trigger block in publish-testpypi.yml (run 35694308033). This run also exercised the newly pinned actions/upload-artifact@v7.0.1 on the real publish path.
Wheel sha256:f73b243d42569e9ec4268a68afef08a0a9df5ac5cf49584862985f525386a20f, sdist sha256:57446bf8d62b2f5fbd54b6f6f86d53da26488471d8d690cf3d2e609b3cae97e5. I re-downloaded the wheel and the hash matches both the upload log and TestPyPI metadata.
All 165 src/openenv/** files in the published wheel are byte-identical to this candidate. The six non-Python assets that are not packaged (core/README.md, containers/images/Dockerfile, and similar) are absent from the shipped 0.5.0 wheel too, so this is existing packaging behaviour rather than a regression.
Clean 3.11 install with the [novita] extra, imports and CLI verified, and a real Echo reset/step smoke passed: reset returned ready, tool discovery returned echo_message and echo_with_length, and both call_tool and step(CallToolAction(...)) round-tripped release-0.6.0.dev91-smoke.
One tooling note: the workflow's own Verify install from TestPyPI step failed even though the upload returned 200 OK for both artifacts. Its retry budget is five attempts 30 seconds apart, and the TestPyPI simple index had not caught up within those 2.5 minutes. The publication itself is fine and I verified the install independently, but the same race exists in publish-pypi.yml, so Thursday's production run could report a false failure after a successful upload. I have not changed the workflow, since widening that budget is a release-policy edit rather than routine maintenance.
Outstanding blockers
Version decision.0.6.0 is my recommendation and is now better supported than when this PR was opened: the candidate carries two separate pieces of new public surface, the Novita provider (#1191) and the RFC 008 Level 2 validation contracts (#1178). If you would rather ship 0.5.1, it is a one-line change.
Novita debt. Recommend ship and track as follow-ups: unbraced $ARG substitution during Dockerfile flattening can corrupt a FROM line when one build arg name is a prefix of another (no in-repo Dockerfile triggers it), and examples/novita_tbench2_simple.py calls wait_for_ready outside its try/finally, so a readiness timeout leaks the sandbox.
Nothing is tagged. Production PyPI is still 0.5.0.
The reason will be displayed to describe this comment to others. Learn more.
Candidate refresh
Exact-main again at 0150e7e4 after folding #1223 (doc-builder upload_pr_documentation workflow pin only).
Net PR diff unchanged vs main: root 0.6.0 + GridWorld editable lock version.
Package bytes unchanged vs prior tip 8bbb8542 — TestPyPI 0.6.0.dev91 evidence remains valid (no republish needed for this refresh).
Exact-head CI re-running.
Still draft / no tag. Holding publish pending Ben’s version (0.6.0 vs 0.5.1) + Novita-debt call.
Note on #1221: publish workflows still use mutable actions/upload-artifact@v7.0.1 (tag); #1228 is the still-open immutable-SHA harden and stays held thru Thursday unless Ben.
The reason will be displayed to describe this comment to others. Learn more.
Exact-head CI green at 0150e7e4 (11 checks; docs build skipped). Still draft; TestPyPI 0.6.0.dev91 remains the package evidence. No tag/publish pending Ben’s version + Novita-debt call.
The reason will be displayed to describe this comment to others. Learn more.
Exact-head CI green at a59a1b48 (11 checks; docs build skipped). Still draft; TestPyPI 0.6.0.dev91 remains the package evidence. No tag/publish pending Ben’s version + Novita-debt call.
The reason will be displayed to describe this comment to others. Learn more.
Candidate refresh + publish-path change
Exact-main at fd997ac1 after folding #1228 (Pauline merged publish-workflow immutable SHA pins for checkout / setup-python / upload-artifact / download-artifact).
Net package diff vs main unchanged: root 0.6.0 + GridWorld editable lock.
Publish path changed → prior TestPyPI 0.6.0.dev91 is stale for release evidence.
Exact-head CI re-running on fd997ac1.
TestPyPI revalidation already pushed: testpypi/0.6.0 @ 104c639a (= tip + push trigger); run 35696413568 queued.
Still draft / no tag. Holding production publish pending Ben’s version + Novita-debt call.
Note: pypa/gh-action-pypi-publish@release/v1 remains on the mutable branch ref (the two HIGH findings #1228 deliberately left for release-owner decision).
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
The reason will be displayed to describe this comment to others. Learn more.
Alignment Review Report
Automated two-tier review of the 0.6.0 release version bump (0.5.1.dev0 → 0.6.0). The diff touches only pyproject.toml and envs/grid_world_env/uv.lock — no Python source, no API/signature changes.
Automated Checks
Lint: PASS (no regression from this PR). The CI lint job is green. The local .claude/hooks/lint.sh reports failures, but they are entirely pre-existing and outside this diff: ruff format --check flags 56 files under envs/** (CI only formats src/ tests/, not envs/), plus the two known usort quirk files tests/envs/test_grid_world.py and tests/envs/test_julia_env.py documented in AGENTS.md. This PR changes no .py files, so it introduces zero lint issues.
Debug code: CLEAN for this diff.check-debug.sh scans the whole src/ tree (informational, always exit 0); every hit is legitimate console.print/CLI output, docstring >>> print(...) examples, or one manual e2e script — none are in this PR.
Open RFCs Context
No RFC is affected by a version-string bump. For reference: In Review — 000, 001, 002, 003, 005, 008, 012; Draft — 010, 011. None touch packaging/versioning or the files in this diff.
Tier 1: Fixes Required
tests/validation_runtime/uv.lock:916 — stale root openenv version. It still records version = "0.5.1.dev0" for the source = { directory = "../../" } entry while this release bumps the repo root to 0.6.0. This is the only remaining reference to 0.5.1.dev0 anywhere in the repo. Its sibling — the only other lockfile that tracks the repo root via ../../ — envs/grid_world_env/uv.lockwas bumped to 0.6.0 in dedicated commit fee469c4, so this one looks like it was simply missed.
Severity: low / not CI-blocking. The Runtime validation job runs uv sync --project tests/validation_runtime --frozen, which builds + installs the actual 0.6.0 from source and ignores the recorded version (verified locally: uv sync --frozen → exit 0, installs 0.6.0). And validate-env-locks only scans changed files under envs/**, so it never inspects this file. It is, however, a release-artifact inconsistency: uv lock --check / uv sync --locked against this project reports the lockfile out-of-date (verified locally: exit 1).
Fix: run uv lock --project tests/validation_runtime (or bump the single recorded version line to 0.6.0) so the committed lockfile matches the release, for parity with the grid_world_env bump.
Suggested reviewer:@burtenshaw (author of tests/validation_runtime/ and .github/workflows/validation-runtime.yml).
No other version-bump gaps found:openenv.__version__ resolves dynamically via importlib.metadata (no hardcoded version to update), there is no CHANGELOG in the repo, and the many env lockfiles pinned to older published versions (0.3.1, 0.4.x, 0.3.2.dev0) reference published releases rather than the repo root — they are intentionally independent and out of scope for this bump.
Tier 2: Alignment Discussion
Principle Conflicts
None identified. A pre-1.0 MINOR bump (0.5.x → 0.6.0) with no API changes is consistent with the Breaking Change Policy in INVARIANTS.md. Nothing touches reset/step/state signatures, client–server separation, rewards-in-environment, or the MCP/Gym API boundary.
RFC Conflicts
None identified. The diff changes no code covered by any open or accepted RFC.
Summary
1 mechanical issue to fix (low-severity, not CI-blocking): stale openenv version in tests/validation_runtime/uv.lock.
The reason will be displayed to describe this comment to others. Learn more.
This root-tracking lockfile (source = { editable = "../../" }) is correctly bumped to 0.6.0.
One gap to mirror here: tests/validation_runtime/uv.lock is the only other lockfile that tracks the repo root (source = { directory = "../../" }), and it still records openenvversion = "0.5.1.dev0" — the last remaining reference to the old version in the repo. It appears to have been missed in this release.
Not CI-blocking (Runtime validation uses uv sync --frozen, which installs the actual 0.6.0; validate-env-locks only scans envs/**), but for a consistent 0.6.0 release please run uv lock --project tests/validation_runtime.
Align tests/validation_runtime/uv.lock with the release version bump
(parity with envs/grid_world_env). Relocked with lab-pinned uv 0.9.3.
Co-authored-by: benjamin.burtenshaw <benjamin.burtenshaw@huggingface.co>
The reason will be displayed to describe this comment to others. Learn more.
Fixed Pre-review finding at 40194e57: tests/validation_runtime/uv.lock now records openenv0.6.0 (relocked with lab-pinned uv 0.9.3; single-line change). Exact-head CI re-running.
The reason will be displayed to describe this comment to others. Learn more.
Release notes and validation evidence for v0.6.0 — head ab7818d1
The PR body above is stale and cannot be edited (updatePullRequest returns 403 for this token), so these are the authoritative notes. They supersede every earlier evidence comment on this PR.
What this PR changes
Nothing but the version, in three places, all on exact main (c32c3174):
File
0.5.1.dev0 → 0.6.0
pyproject.toml
package version
envs/grid_world_env/uv.lock
local openenv entry
tests/validation_runtime/uv.lock
local openenv entry
Release contents (v0.5.0..main, 22 commits)
New public API surface — this is what justifies a minor bump rather than a patch:
Novita sandbox provider (#1191) — NovitaSandboxProvider plus a new openenv[novita] extra (novita-sandbox>=2.1.1).
RFC 008 Level 2 validation contracts (#1178) — openenv.validation.runtime (RuntimePlan and friends), the manifest-v2 / report-v2 / runtime-plan JSON schemas, and severity-v2.
Docker validation provider (#1179) — openenv.validation.providers.docker.DockerValidationProvider and the reproducible validation lab.
Fixes and metadata:
Client now cleans up the provider when session startup fails (#1145).
coding_env documented constructor arguments actually work (#1147).
README is the package long description, including the Miles integration (#1220).
Repository CI: green on ab7818d1, including Runtime validation / Linux Docker, both Python matrices, ThinkingBox, CodeQL, lint, env-lock validation, and Bugbot. The docs build now runs too, since this PR is out of draft.
Local CI-equivalent suite: 2970 passed, 103 skipped, 44 deselected.
TestPyPI 0.6.0.dev111 (run 35717494905) remains the current package evidence. Today's three merges touched only tests/validation_runtime/**, which the wheel does not ship, and I re-proved that rather than asserting it: a wheel built from this head has all 166 openenv/** files byte-identical to the published 0.6.0.dev111 wheel (sha256:b8784828…fc3b22, re-downloaded and hash-checked against the TestPyPI index).
Fresh-install checks from the wheel built at this head, in a clean 3.11 venv: reports 0.6.0; imports openenv.harbor, openenv.discovery, openenv.validation, NovitaSandboxProvider, RuntimePlan, DockerValidationProvider; the [novita] extra resolves novita_sandbox; openenv --help plus harbor / catalog / discover / validate all work.
Real Echo reset/step smoke: server started from the installed wheel, reset() → ready, list_tools() → echo_message + echo_with_length, and both call_tool and step(CallToolAction(...)) round-tripped a unique marker.
Outstanding blockers
Two release-owner decisions, not code:
Version — 0.6.0 (recommended) vs 0.5.1.
Novita debt — accept and ship, or fix first. Both items are latent: _resolve_from_references substitutes unbraced $ARG names, so a Dockerfile declaring ARG BASE before ARG BASE_IMAGE would resolve FROM $BASE_IMAGE to python:3.12_IMAGE (no in-repo Dockerfile declares both), and examples/novita_tbench2_simple.py calls wait_for_ready above its try/finally, so a readiness timeout leaks the sandbox.
Nothing is tagged and PyPI is still on 0.5.0. Once both decisions land I merge this at its verified head, tag that exact commit, and run the existing production publish workflow.
Alignment review on merged #1237 flagged this for the 0.6.0 notes (I cannot edit this PR body — 403).
Please add under Release notes (breaking, documented):
collect --llm-endpoint / OpenAIClient accept full base URLs: pathless endpoints get /v1; an existing path is kept; --llm-port only fills in when the URL has no port. Non-http(s), missing host, credentials, query/fragment, and conflicting ports are rejected. Breaking: the old implicit port 8000 for bare hosts is gone — --llm-endpoint http://localhost is now port 80; use http://localhost:8000 or --llm-port 8000 (#1189 / #1237).
Also worth refreshing the stale checklist: version is decided (0.6.0); Novita debt is follow-up #1235 (not a cut blocker); TestPyPI is 0.6.0.dev121 verified against tip ee33f7ae; Level 2 contracts (#1178/#1179) are already on main and in this candidate.
The reason will be displayed to describe this comment to others. Learn more.
The breaking-change bullet landed correctly — thanks. While checking it I diffed the candidate against v0.5.0 and found the notes are still missing two of the four package-level changes, plus a stale header. Since you can edit the body and I cannot (403), here is the exact text.
Missing bullets.git diff --stat v0.5.0..81379bd3 -- src/openenv/ shows 17 files and ~3,600 added lines. The notes cover Novita (#1191) and the LLM endpoint change (#1189/#1237); these two are in the candidate but undocumented:
-**Level 2 validation building blocks:**`openenv.validation` gains the RFC 008 Level 2 contracts — new `manifest-v2`, `report-v2` and `runtime-plan` JSON schemas, a `severity-v2` severity policy, runtime contract types, and `DockerValidationProvider`, a validation-only provider that builds an image and probes it locally with network-policy enforcement and in-sandbox exec ([#1178](https://github.com/huggingface/OpenEnv/pull/1178), [#1179](https://github.com/huggingface/OpenEnv/pull/1179)). The `openenv validate` CLI is unchanged in this release; the runtime probes that drive these contracts are still in review ([#1181](https://github.com/huggingface/OpenEnv/pull/1181)).
-**Failed session startup no longer leaks a provider:** when `new_session()` cannot bring up a client, `EnvClient` stops the container or sandbox it started instead of leaving it running. Cleanup is retry-safe, and a client whose cleanup is still pending refuses to reconnect to the old base URL rather than overwriting the only resource handle ([#1145](https://github.com/huggingface/OpenEnv/pull/1145)).
I deliberately wrote the validation bullet as building blocks rather than a working --level runtime --local: src/openenv/cli/commands/validate.py is unchanged since v0.5.0 (only #1181 touches it), so claiming a usable runtime-probe CLI would overstate what ships tomorrow.
Stale header, three spots. The top still describes the state from five days ago:
"Candidate base: 39fdf5cd … 6 commits since v0.5.0" → candidate is ee33f7ae on main 81379bd3, 23 commits since v0.5.0.
"the only package change since v0.5.0 is new public surface — NovitaSandboxProvider" → no longer true; it is now Novita, the Level 2 validation contracts plus DockerValidationProvider, the LLM endpoint change, and the provider-cleanup fix. That is a stronger case for the minor bump, not a weaker one.
"please say so before Wednesday" → the version is decided; Ben chose 0.6.0 in Slack this morning. The Version decision checkbox under Outstanding blockers can be ticked, as can the three TestPyPI items (0.6.0.dev121, run 35842587905, verified byte-identical to this tip).
The compare link at the bottom still points at 39fdf5cd; it should be v0.5.0...81379bd3.
One caveat on the TestPyPI checklist line: the documented verification command uses --index-url test.pypi.org --extra-index-url pypi.org, which lets a TestPyPI name-squat outrank a real dependency. That is a release-policy fix rather than something to change silently during a cut, so I am only flagging it here — the workflow itself resolves dependencies from PyPI and installs the package --no-deps, so the published evidence is sound.
The reason will be displayed to describe this comment to others. Learn more.
Correction to my previous comment, and it makes one edit mandatory rather than optional.
I reported that the #1145 bullet was in the body. It is not. The only mention of #1145 is in this paragraph:
Not in this candidate: every open small fix that touches the package (#1145, #1198, #1200, #1203, #1208, #1209) comes from a fork whose workflows are still action_required …
That statement is wrong now: #1145 merged on Monday as 228bd147 and is in this candidate; #1208 was closed unmerged. So the release notes currently tell readers that a shipped fix is not in the release. Two edits, both in that paragraph and the notes list:
Remove #1145 and #1208 from the "Not in this candidate" list. The accurate remainder is #1198, #1200, #1203, #1209 and #1217 — all forks still at action_required.
Add the #1145 bullet under Release notes (text in my previous comment), alongside the still-missing Level 2 bullet for #1178/#1179.
Everything else from that comment stands: the header still says candidate base 39fdf5cd and "6 commits since v0.5.0" where the truth is ee33f7ae on main 81379bd3 with 23 commits, and the compare link points at the old base.
Mechanically that is one paragraph edit plus two bullets. The candidate itself is unaffected — still CLEAN, 13/13 green, and byte-identical to the verified TestPyPI 0.6.0.dev121.
The reason will be displayed to describe this comment to others. Learn more.
Independent pre-tag verification — the TestPyPI gate is closed
Read-only check from a sibling run (bc-799dc749); the release lock refs/release-jock/v0.6.0... correction: refs/release-lock/v0.6.0 is held by bc-ab703bc0, so I touched nothing — no merge, tag, publish or branch push.
The last publish-testpypi.yml run (35855102604, abbfb4fc) is red, and that
red is misleading in a way worth recording: the upload succeeded and 0.6.0.dev151 exists on TestPyPI (uploaded 2026-09-23 11:33:38Z). The job failed
afterwards, in its own verification step, because the index had not made the new
version resolvable inside the retry window — No matching distribution found for openenv==0.6.0.dev151, then ModuleNotFoundError: No module named 'openenv' as a
consequence. Nothing about the artifact was wrong.
I completed that verification by hand, and tied it to the commit being tagged.
1. The published artifact is the candidate, to the byte
Built a wheel from candidate head 0034202e with uv build --wheel and compared
it entry by entry against the downloaded TestPyPI wheel (sha256 c5019f30…54c2, matches the index digest):
167 packaged files on each side, no file on only one side, zero content
differences.
dist-info: LICENSE, WHEEL, entry_points.txt, top_level.txt identical; METADATA differs in exactly one line (Version: 0.6.0 vs Version: 0.6.0.dev151); RECORD differs only in the version-named dist-info
paths.
2. That same source is what is now on main
git diff 0034202e 5f95614f is empty — the squash tree is identical to the
candidate — and git diff abbfb4fc 5f95614f -- src/ pyproject.toml README.md MANIFEST.in is also empty. So the artifact verified below and the package source
at 5f95614f (the commit to tag v0.6.0) are the same bytes.
3. The published artifact actually works
uv pip install --no-deps --index-url https://test.pypi.org/simple/ openenv==0.6.0.dev151 now resolves, which is the step that timed out in CI.
Clean venv, artifact from TestPyPI and dependencies from PyPI (avoiding the
mixed-index ordering problem flagged earlier): openenv.__version__ == importlib.metadata.version("openenv") == 0.6.0.dev151, and openenv --help
renders.
All three new API surfaces import from the installed wheel: NovitaSandboxProvider (#1191), openenv.validation.runtime.contracts (#1178),
and openenv.core.llm_client.LLMClient (#1237).
Real Echo episode against the installed package (repo envs/ on PYTHONPATH, no src/): /health 200, reset metadata {'status': 'ready', 'message': 'Echo environment ready!'}, both tools listed, step returned v0.6.0-preflight with is_error: False, call_tool("echo_with_length", message="abc") returned {'message': 'abc', 'length': 3}, and state reported step_count=1.
Still outstanding for the publisher
publish-pypi.yml and publish-testpypi.yml were both rewritten by #1240 (action
SHA pins) afterdev151 was built, so the pinned publish path itself has never
been exercised. That is a workflow-mechanics risk, not an artifact risk: watch the
first publish-pypi.yml run closely and confirm it actually started from the tag
rather than assuming the tag push triggered it.
Production is still 0.5.0 and no v0.6.0 tag exists as of this comment.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
burtenshaw
added
enhancement
Cursor
Release PR: v0.6.0
Planned release: Thursday, September 24, 2026 at 10:00 Europe/Brussels (08:00 UTC).
Candidate base:
39fdf5cd(main tip at the time of writing). 6 commits sincev0.5.0.This is the rolling release PR for the week; it will be refreshed onto the final
candidate head before merge.
Why 0.6.0 and not 0.5.1: the only package change since
v0.5.0is new publicsurface —
NovitaSandboxProvider(1,207 lines undersrc/openenv/core/containers/runtime/), a documented entry indocs/source/reference/core.md, and the newopenenv[novita]extra. That is addedfunctionality, which is a minor bump under the same standard applied when
0.5.0was cut rather than a quiet
0.4.3. If you would rather treat opt-in providers asnon-API and ship
0.5.1, it is a one-line change to this PR — please say so beforeWednesday so the notes and tag match.
Release notes
Breaking change — LLM endpoint URLs:
--llm-endpointis now a full base URL, and the implicit port8000has been removed.http://localhostnow uses the standard HTTP port80; to retain the previous local behavior, usehttp://localhost:8000or--llm-port 8000(#1189, landed via #1237).Novita AI sandbox provider: run an OpenEnv server inside a Novita sandbox and
connect over
wss://, from either a registry image or a local Dockerfile(
image_from_dockerfile), with multi-stage Dockerfiles replayed as a single stagefor Novita's template parser. HTTPS/WSS transport is enforced and captured sandbox
output is withheld from raised errors unless
surface_server_logs=True. Installsvia
pip install openenv[novita](#1191).Not in the wheel, but shipped to environment images and the repo this week:
coding_envconstructor arguments work as documented:additional_importsnow merges on top of
DEFAULT_SAFE_IMPORTSinstead of replacing the allowlist,so passing one module no longer breaks
import json(#1147).
TLSStreamIDNA certificate-spoofing advisory plus two lower-severity fixes(#1201,
#1202, and
#1197 for the remaining 31
lockfiles), and soupsieve 2.9.2 for the polynomial-ReDoS advisories
(#1196, extended to
browsergym_envandfinrl_envin chore(deps): aggregate envs Dependabot updates #1197).Full candidate comparison: v0.5.0...39fdf5c
Release-maintenance changes in this PR
0.5.1.dev0to stable0.6.0. Nothing else.Outstanding blockers
0.6.0(recommended) or say0.5.1and I willchange the one line.
opt-in Novita provider and its examples, none of it re-blocking in my
assessment — listed so the decision to ship it is explicit:
unbraced
$ARGsubstitution during Dockerfile flattening can corrupt aFROMline (
$BASEmatching before$BASE_IMAGE);examples/novita_tbench2_simple.pycalls
wait_for_readyoutside itstry/finally, so a readiness timeout leaks apaid sandbox; concurrent starts and failed kills can orphan sandboxes; flattened
stages carry builder filesystem/ENV/USER state into the runtime image; and the
RFC 002 provider amendment (root execution, unrestricted egress, HTTP-only
readiness) is still unratified.
publish-testpypi.ymlfrom this exact branch with a unique0.6.0pre/dev suffix.mainimmediately before merge; rerun all required checks if thehead or base changes.
Not in this candidate: every open small fix that touches the package
(#1145,
#1198,
#1200,
#1203,
#1208,
#1209) comes from a fork whose
workflows are still
action_required, so none of them has ever run repository CI.A maintainer clicking Approve and run is the only way any of them becomes
release-eligible.
Release Checklist
Before opening this PR
pyproject.tomlversion changed from0.5.1.dev0→0.6.0hf-staging/is NOT in this PR's diffprint(),breakpoint(), orTODOadded to release-critical pathsCI gates (must be green before merge)
testpasses on Python 3.11testpasses on Python 3.12lintpasses (usort + ruff)Package CIbuilds, checks, and smoke-tests wheel/sdist installsTestPyPI validation (before merging)
publish-testpypi.ymlfrom this branch0.6.0.devNpip install --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ openenv==0.6.0.devNPost-merge steps
v0.6.0pushed against the exact merge commit onmainpublish-pypi.ymlcompleted successfully from the tagpip install openenv==0.6.0from production PyPI verified (import,__version__,openenv --help, real Echo reset/step)RFC Status
Note
Low Risk
Metadata-only version and lockfile sync with no runtime or API changes in the diff.
Overview
Release version bump for the
openenvpackage:pyproject.tomlmoves from0.5.1.dev0to stable0.6.0, aligning the published package version with the v0.6.0 release.The editable
openenventries inenvs/grid_world_env/uv.lockandtests/validation_runtime/uv.lockare updated to0.6.0so those lockfiles stay consistent with the root project. No application or library code changes appear in this diff.Reviewed by Cursor Bugbot for commit 0034202. Bugbot is set up for automated code reviews on this repo. Configure here.