Second release candidate on the 1.3.0 feature line, cut from release-1.3.0-rc.1 (2026-11-16). Baseline for every section is release-1.3.0-rc.1; rc users on that tag should upgrade. rc releases publish images but hold packages — SDK/CLI artifacts ship at the 1.3.0 stable bump.
Highlights
- Rootfs-only snapshots now cover native (S3) sandboxes:
create_snapshot(mode="rootfsOnly")works end to end for sandboxes on either artifact path — native S3 templates and OCI/OverlayBD templates alike; capture qualification derives from the deployment itself, and an unqualified node fails the snapshot instead of silently producing afullartifact (#2900 #2893 fast-sandbox#181 fast-sandbox#223 fast-sandbox#225) - BatchSandbox clean exits finally terminate: a sticky
Completedterminal phase replaces theSucceed → Pendinglimbo that kept an exit-0 sandbox reported as running until expiry deletion (#2715) - Two silent workspace-loss races fixed: on Docker, a successful renewal can no longer commit while a concurrently firing expiration deletes the live sandbox and its metadata (#2892); on Kubernetes, expiry deletion carries UID/resourceVersion preconditions so a stale informer copy cannot delete a just-renewed sandbox (#2905)
Server
- Native (S3) sandboxes accept
mode="rootfsOnly"on the public lifecycle API: the server drops its OCI-only preflight mirror, leaving mode compatibility to the FastPath gate (no intent is persisted on rejection) and the Fastlet's fail-closed node re-check; Docker and Kubernetes runtimes still return 409 (#2900) - Pool and allocation logic moves behind a
WorkloadProviderSPI: deployments with non-BatchSandbox workload providers no longer fail on hardcodedpools.sandbox.opensandbox.iolookups, and/v1/poolson a provider without OSB-managed Pool CRDs returns a clean501 KUBERNETES::POOL_NOT_SUPPORTED(#2884) - Docker renewal is serialized with the per-sandbox lifecycle lock: previously a concurrently firing expiration could delete a running sandbox and its metadata after
renew-expirationhad already returned 200 (#2892) - BatchSandbox
Completedmaps to the publicTerminatedstate with reasonSANDBOX_COMPLETED, carrying the controller's exit detail and completion time; pause/resume on a completed sandbox is rejected explicitly (#2715) - A snapshot informer watching a CRD that is not installed (LIST 404) logs a single warning that names the informer instead of a full traceback every 30 seconds (#2875)
- Smaller fixes: a malformed tenant-provider
ttl(null, negative,NaN/Infinity) falls back to the 30-second default instead of failing the lookup or caching forever (#2704); onePoolService/K8sClientis shared across/poolsrequests instead of rewriting the process-global kube config per request (#2852); list pagination for sandboxes, templates, and snapshots is consolidated with byte-identical responses (#2860); metadata patch validation is shared between the Kubernetes, Docker, and fast-sandbox backends (#2863)
SDKs
- Code-interpreter readiness is bounded by the ready timeout in Python, JavaScript, and C#: a stalled health probe no longer keeps
CodeInterpreter.create()waiting pastready_timeout/readyTimeoutSeconds/ReadyTimeoutSeconds(#2856 #2857 #2858) exit_code/exitCodeis populated aftercodes.run()in Python and JavaScript, matching Kotlin, C#, and Go:0on completion, otherwise the error's numeric exit value (#2854)
Python
- Streamed file downloads keep execd's structured error: a 404 carries
error.codeand the request id, and a 429 raisesSandboxRateLimitExceptionwithretry_afterinstead of a generic exception without the response body (#2696)
JavaScript
- File download and upload failures surface the server's error code and message instead of a hand-built
UNEXPECTED_RESPONSE, matching every other filesystem call (#2700)
C#
- SSE streams (commands, code, isolated sessions, and the code interpreter) are bounded by a per-read timeout of
max(RequestTimeoutSeconds, 30s): a stream that goes silent — for example after the sandbox was paused — now fails instead of blocking forever (#2848)
Controller
- BatchSandbox gains a sticky
Completedterminal phase for a clean exit-0 main container: previouslystatus.phaseregressed fromSucceedback toPendinguntilexpireTimedeleted the CRD, so every consumer kept reporting the sandbox as running for the whole TTL; a completed sandbox skips pod replacement and task scheduling, and pause/resume intents are refused with aSandboxCompletedcondition (#2715) - Status writes carry a
resourceVersionprecondition, so a stale runtime view can no longer overwrite a freshly writtenPauseFailed/ResumeFailedcondition — losing that condition also silently broke the server's pause/resume retry path (#2877) - All per-object in-memory controller state is forgotten when a BatchSandbox is deleted or replaced under the same name, so a successor no longer inherits stale scale expectations, pod-recovery budgets, or task schedulers (#2861)
- Smaller fixes: deletion can stop a task by name even when the stored shard patch is malformed, instead of sticking in
Terminatingforever (#2554); scale expectations that stay unsatisfied past the timeout are dropped so pod creation resumes, like the Pool controller (#2851); snapshot template tests no longer write the process-globaltime.Localunder-race(#2885)
Execd
- Disk telemetry arrives over OTLP: per-mount usage gauges (
execd.system.disk.usage_bytes/total_bytes) withmountpoint/fstype/sourceattributes, plus cumulative sandbox IO counters (execd.system.disk.io.bytes/io.operations) — each mount stat is bounded by a 2-second timeout so a stale network mount cannot stall the pipeline, and IO prefers sandbox cgroup accounting with a/proc/diskstatsfallback (#2912) - Foreground command timeouts are reported as
CommandTimeoutinstead ofCommandExecError, with the requested timeout recorded in the traceback, logs, and command status (#2684) POST /codeno longer hangs forever when the kernel's completion signal is lost: finalization is bounded by a reply-anchored deadline (new--jupyter-reply-finalize-timeoutflag, default 2s, restarted on output silence) so the stream closes with a terminal event and the session lock is released (#2804)- A
/commandoutput line longer than 16 MiB is delivered in UTF-8-boundary pieces instead of being buffered without limit — a 200 MB line previously OOM-killed execd (#2729) - Rename batches are validated before any item moves: a missing or empty path returns 400 up front instead of moving the daemon's working directory or partially applying the batch (#2818); an empty expanded path is rejected in every file API, closing the
{"": {"mode": 700}}chmod-on-working-directory hole (#2897) /metrics/watchactually sends an update every second as documented, instead of one every two seconds (#2853)- Smaller fixes: execution tracking is shared between the three streaming handlers with unchanged metrics (#2862); isolated-session non-zero exits use a typed error so
ExitErrorclassification survives future message changes (#2864); rename tests that assumed${}expands to empty are dropped (#2899)
Networking
egress
- Credential-vault resolvers now receive the outbound request headers (minus a structural/hop-by-hop denylist, capped at 64 headers and 8 KiB per value, failing closed with 407), completing OSEP-0025's header-bound signature support so exchange resolvers can re-sign schemes like SigV4 and ACS (#2917)
- [EXPERIMENTAL] The credential-bound Fast Sandbox campaign advances on the readiness and verification fronts: an Actions v1
data-plane-readyHook waits up to 50 ms for the subject snapshot acknowledgement, turning the guaranteed first-create rejection in this mode into millisecond readiness (#2922); denied TLS handshakes close the client transport instead of leaving it to time out (#2886); and the integration environment gains opt-in mode wiring with honest BLOCKED/NOT_RUN reporting (#2850), trusted TLS protocol fixtures (#2876), a subject-isolation and recovery verifier with a serial credential-bound CI leg and first real KVM acceptance evidence (#2881), restart-recovery tests (#2888 #2906), and operator/client documentation for bindings and persistent connections (#2910). The feature remains opt-in;alland unset behavior is unchanged. - Upstream TLS verification failures are reported in Egress logs with a connection ID and a fixed reason code (for example
self_signed), so a 502 no longer requires verbose proxy output to diagnose; no hostnames, certificates, or credentials are emitted (#2918) - Policy targets written with a trailing root dot (
pypi.org.,*.github.com.) now match like their dot-less spelling — previously they were silently inert and fell through todefaultAction(#2810) - Smaller fixes: the fast-sandbox nftables applier reuses the sidecar's helpers, so a failed batch is retried only when the target table was actually missing (#2865)
Fast Sandbox
- Rootfs-only snapshots extend to native (S3) sandboxes upstream: FastPath admits RootfsOnly at the gRPC gate for native artifacts (fast-sandbox#223), capture qualification self-derives from the standard managed VMM directory and StateRoot, dropping the
FAST_SANDBOX_ROOTFS_QUALIFICATIONprofile (fast-sandbox#225), and native delivery ships with template-state import and restore-manifest readiness fixes (fast-sandbox#181 fast-sandbox#212 fast-sandbox#214) — carriers (#2893 #2900) - FSEP-0003 Fastlet control authentication ships: fastlets authenticate the control API with ServiceAccount tokens verified via JWKS, the controller derives its default caller allow list from its own namespace, and the base chart grants the fenced token-minting Role plus anonymous issuer discovery (fast-sandbox#207 #2900 #2871)
- Guest wall clocks are calibrated after restore and crash-recovery resume: Firecracker microVMs have no RTC, so a restored guest previously believed the snapshot moment was still now — breaking TLS certificate validation, token expiry, and cron schedules inside the guest (#2900 fast-sandbox#218)
- An optional FastPath gRPC-Web listener (HTTP/1.1 protobuf, same RPC methods) can be enabled with the new
controller.fastpathHTTPPortchart value (default0, off); ingress gateway and server keep native gRPC and are unaffected (#2938 fast-sandbox#234) - The all-in-one (
--role=all) controller is enforced to a single replica: chart values with more than one replica are rejected and upgrades use theRecreatestrategy so two controllers never reconcile the same SandboxPool concurrently (#2901) - Security: Go dependency and toolchain vulnerabilities in the companion image builds are remediated upstream (#2938 fast-sandbox#203)
- FSEP-0004 (chunked snapshot transfer) and FSEP-0005 (rootfs-only create CPU/memory overrides) land as design proposals with no behavior change (#2938 fast-sandbox#216 fast-sandbox#219)
Misc
- OSEPs: OSEP-0028 proposes CPU and memory overrides for RootfsOnly sandbox creation via the existing
resourceLimitsfield, reusingPOST /sandboxes; it is already in implementation (#2902 #2908) - E2B shim (OSEP-0027) [EXPERIMENTAL]: the new
components/e2b-shimGo module lands withadapter/routerskeletons and liveness endpoints plus component CI (#2920), then validated non-secret configuration loading (#2932); ownership is assigned via CODEOWNERS (#2931). No E2B routes yet. - Charts: the Kubernetes minimum is raised to 1.22 across all sub-charts and release metadata — Helm rejects installs on 1.21 clusters (#2874)
- Docs: the lifecycle API and spec state explicitly that a DELETE 204 means deletion accepted, not confirmed runtime termination, and that a later 404 must not be read as termination (#2738)
- CI: the host
GOPROXYis passed into Go-based integration image builds, unblocking dependency-remediation PRs on restricted hosts (#2925); fast-sandbox PR gates run a single egress leg with per-PR cancellation instead of always occupying the KVM runner with both legs (#2911)
Upgrade & Compatibility
- Kubernetes 1.21 no longer supported: all charts gate on
>=1.22.0-0; upgrade the cluster before installing (#2874) - BatchSandbox
Completedrollout order: the CRD must be applied before the new controller — an old CRD rejects the new phase with a validation error. Helm installs both together; operators upgrading separately apply the CRD first. Clients mapping CRD phases see the newCompletedvalue, and the public lifecycle state becomesTerminatedwith reasonSANDBOX_COMPLETED(#2715) - Native templates must be rebuilt for rootfs-only capture: native template builds now always inject guest control and validate freeze/thaw at build time, and capture from older native templates fails closed (#2900)
- execd foreground timeouts:
CommandTimeoutreplacesCommandExecError; callers matching error names must handle the new name, and timeout status text now includes the requested duration (#2684) - execd rename validation: empty-string
src/destare rejected with 400; callers relying on empty strings to select the current or session-view directory send an explicit path such as"."(#2818) - Egress trailing-dot rules take effect: rules written with a trailing root dot were silently inert before and now enforce like their dot-less spelling; review policies — especially deny rules under a default-allow policy — when adopting (#2810)
- Fast Sandbox upgrades: upgrade the chart and companion images together — control-auth fastlets fail closed without the base chart's token-minting RBAC (#2871 #2900); the all-in-one controller rejects
replicaCount > 1and upgrades carry a brief FastPath interruption under theRecreatestrategy (#2901); the gRPC-Web listener stays off unlesscontroller.fastpathHTTPPortis set (#2938) - Images:
opensandbox/<component>:release-1.3.0-rc.2(all three registries; fsb images for fast-sandbox) - Packages: unchanged — SDK/CLI artifacts publish at the
1.3.0stable release - Charts: render from this tag (
helm template ./manifests/charts/opensandbox) - Kubernetes: supported range v1.22 – v1.34
- Skew: server ↔ CLI/SDK same line supported; ±1 minor warns
👥 Contributors
Thanks to these contributors ❤️
- @AdityaxDeore
- @Dev-next-gen
- @Jackmeson1
- @JasonBuildAI
- @LyleLiu666
- @Pangjiping
- @Yeaury
- @acodercat
- @arun-elr
- @cwj2001
- @fengcone
- @fxia0018
- @gezerui888
- @hittyt
- @hpliStartAgain
- @jwx0925
- @kaustubh-765
- @mengdehong
- @rudycelekli
- @yjjnls
Artifacts
- Images:
opensandbox/<component>:release-1.3.0-rc.2(Docker Hub / GHCR / ACR)
Installation
# Platform (Kubernetes) — render the chart at this tag and apply
git clone https://github.com/opensandbox-group/OpenSandbox
cd OpenSandbox
git checkout release-1.3.0-rc.2
helm dependency build manifests/charts/opensandbox # package file:// sub-charts (not committed)
helm template ./manifests/charts/opensandbox | kubectl apply -f -
# or point your GitOps platform (Argo / Flux) at the repo path + tagVerify what you installed against the BOM: see Release Verification.