@cloudflare/sandbox
1.0.3 · 22 Oct 2026Apache-2.02.1M / weekRepositoryHomepagenpmAtom feed
Upgrading @cloudflare/sandbox From 0.7.17 to 0.12.14
42 releases in betweenevery note read
Reading releases, release notes and advisories…
- 4
- changes to make
- 0
- advisories fixed
- 24
- new
- 58
- fixes
Before you upgrade
Tick each one off as you go- Breaking0.12.0
- Breaking0.8.8
- Breaking0.8.7
- Breaking0.8.2
Security
- 0.12.14
Update the container runtime to Bun 1.4.4. Bun checks certificates against its own built-in CA list, and in 1.4.4 that list no longer trusts the compromised eMudhra intermediates, so the
buncommand andcreateBackup()uploads of 10 MiB or more now reject their certificates. Bun 1.4.4 also updates Brotli to 1.2.1, which fixes CVE-2026-111497. The-opencodeimage ships OpenCode 1.19.4 or later, which also runs on Bun 1.4.4. - 0.8.4
Upgrade Go toolchain to 1.25 and update dependencies in the desktop container variant, including a security fix for image processing (CVE-2026-33809).
Release by release
- Security
Update the container runtime to Bun 1.4.4. Bun checks certificates against its own built-in CA list, and in 1.4.4 that list no longer trusts the compromised eMudhra intermediates, so the
buncommand andcreateBackup()uploads of 10 MiB or more now reject their certificates. Bun 1.4.4 also updates Brotli to 1.2.1, which fixes CVE-2026-111497. The-opencodeimage ships OpenCode 1.19.4 or later, which also runs on Bun 1.4.4.
- Security
- Fixed
Stop trusting certificates from the compromised eMudhra intermediate CAs in the container images. The images now ship the updated Ubuntu and Alpine CA bundles, Node.js 22.23.4, cloudflared 2026.10.2, and pip 26.2.2 in the
-pythonimage. Bun keeps trusting these certificates until it releases a fix, which affectsbun, opencode, and multipart uploads increateBackup()for archives of 10 MiB or more. Update the container image to get this fix. See the eMudhra certificate compromise for details.
- Fixed
- Fixed
Fix
mountBucket()breaking an existing mount when a second call for the same mount path fails with "already in use". The existing mount now stays mounted and usable. - Fixed
Keep streaming
exec()alive while a command is quiet, and enforce itstimeout. Quiet commands were cut off after about four minutes on the HTTP transport and five on WebSocket. The timeout was ignored on those transports, and on every transport for commands that run in a session. Requires the matching container image. - Fixed
Keep R2 binding and credential-proxy bucket mounts working when the Sandbox Durable Object restarts while its container keeps running. The mounts used to stay listed but fail every read and write with
EIOuntilmountBucket()was called again. - Fixed
Fix
createOpencodeServer()andcreateOpencode()waiting for the full three-minute startup timeout whenOPENCODE_SERVER_PASSWORDis set. When the password is passed inenv, the client returned bycreateOpencode()now authenticates with it.
- Fixed
- Fixed
Stop backups from dropping whole directories when
excludesorgitignore: truematch a nested path. For example,excludes: ['.cache/foo/*/node_modules']no longer removes all of.cachefrom the archive. A multi-component pattern such asbuild/outputnow matches only relative to the backup directory, while a single-component pattern such asnode_modulesstill matches at any depth. Because mksquashfs does not support**, a pattern like**/node_modules/.cachenow excludes onlynode_modules/.cacheat the root of the backup directory; add a pattern for each nested location that must stay out of the backup. - Fixed
Return 404 with
workspace_read_not_foundfrom the bridge'sGET /v1/sandbox/:id/file/*when the file doesn't exist. The route used to return 200 with an empty body, the same response as an empty file. - Fixed
Fix local bucket mounts for large files and write bursts. On the RPC transport, objects over 4 MiB now stream into the container instead of failing with an interrupted runtime connection. Container files stream back to R2 byte for byte, without loading the whole file into Worker memory. A file being written uploads once its changes settle, instead of on every write.
- Fixed
Stop writing command text and output to Sandbox logs. Exec and process log events now record IDs, exit codes, durations, and output lengths, and the OpenAI Agents
ShellandEditoradapters no longer log commands, stderr, or patch text, so secrets in commands, output, or patches no longer reach Workers Logs or container logs. Command failure messages no longer repeat the command; it remains available on the error'scontext.command. - Fixed
Fix
restoreBackup()error reporting for archives of 10 MB or more. A failed download part now throws aBackupRestoreErrorthat includes curl's error output and each part's exit code, instead of aShell terminated unexpectedlysession error. - Fixed
Sync local bucket mounts outside
/workspace, such as/data, from the container back to R2.watch()andcheckChanges()now accept any absolute path on every transport, as documented. A local mount whose container rejects the watch now logs one error instead of retrying forever.
- Fixed
- New
Allow multipart backup uploads to trust the runtime HTTPS interception certificate, so backups larger than 10 MiB work when
interceptHttpsis enabled. - Fixed
Reject bucket-level mutations sent through
credentialProxymounts while preserving object writes, bucket probes, and scoped listings. - Fixed
Fix
File existserrors when git or other tools recreate files in a directory restored withrestoreBackup(), such as runninggit reset --hardwhile a dev server watches the project.
- New
- New
Allow mounted R2 buckets to open object keys containing spaces, symbols, percent signs, and non-ASCII characters.
- New
- Fixed
Enforce source mount prefixes for server-side copies between R2 bindings.
- Fixed
- Fixed
Ensure stable releases advance npm's
latesttag only after all release artifacts pass verification.
- Fixed
7 other lines we could not sort, mostly version bumps
- 0.12.5
Instrument the bridge API with Cloudflare custom spans. Each endpoint now emits a
bridge.<operation>span annotated with the sandbox ID, container UUID, and operation-specific metadata (command, file path, port, session ID, and so on), so requests are traceable end to end. Enable tracing withobservability.traces.enabledin your Worker configuration. - 0.12.5
Make
destroy()succeed when the sandbox container was never admitted, so cleanup no longer fails with a no-instance platform error. - 0.12.5
Reduce intermittent startup failures during concurrent sandbox creation where operations could fail with a disposed RPC session error.
- 0.12.5
Surface container capacity and admission failures as retryable
ContainerUnavailableErrorwith a structuredreason, including no-instance and max-instances cases. Callers can branch onerror.reasoninstead of parsing platform messages. - 0.12.5
Fill the warm pool in bounded-parallel batches instead of one container at a time, so it primes and recovers far faster after a burst. Tune the batch size with
WARM_POOL_SCALE_BATCH_SIZE(default 5, clamped to 20). - 0.12.5
Refill the warm pool the moment it drains instead of waiting for the next background tick. After a burst consumes warm containers, replenishment now starts immediately, cutting the latency tail for sandboxes created during sustained or back-to-back bursts.
- 0.12.5
Let the warm pool know its instance ceiling up front via the
WARM_POOL_MAX_INSTANCESbridge variable. Set it to match your container'smax_instancesso the pool makes correct capacity decisions without first crashing into the limit.0(the default) preserves the existing auto-learn behaviour.
2 internal lines (docs, tests, tooling) left out.
A whole project to check? Paste its package.json in the dependency check: every deprecated, outdated or vulnerable dependency, with what to upgrade to. Free, no account.