## Summary
Adds SDK support for the new `POST /sandboxes/{sandboxID}/fork` endpoint
(e2b-dev/infra#3202): checkpoint a running sandbox in place (briefly
paused, snapshotted with full memory state, and resumed — its ID and
expiration stay untouched) and boot `count` new sandboxes from that
snapshot.
- **spec**: adds `SandboxForkRequest` / `SandboxForkResult` schemas and
the `/sandboxes/{sandboxID}/fork` path (mirroring the infra spec); JS
and Python API clients regenerated via `make codegen`.
- **js-sdk**: `sandbox.fork(opts)` instance method and
`Sandbox.fork(sandboxId, opts)` static method. Returns
`Promise<Array<Sandbox | Error>>` — one entry per requested fork, each
either a connected `Sandbox` instance or an `Error` describing why that
fork failed to start (`Promise.allSettled`-style, matching the per-fork
results of the API). Per-fork error codes go through the same code→class
mapping as other API errors (extracted from `handleApiError` into
`apiErrorFromCode`), so e.g. a per-fork 429 (sandbox limit) surfaces as
`RateLimitError`. `SandboxForkOpts` extends the full `ConnectionOpts`
(like `SandboxConnectOpts`), so `proxy`, `logger`, `apiUrl`, etc. work
with fork-by-ID. `timeoutMs` defaults to 5 minutes like
`create`/`connect`; `count` defaults to 1 and is validated client-side
(`InvalidArgumentError` for `count < 1`); a whole-request 404 maps to
`SandboxNotFoundError` (the source sandbox is the missing resource —
same semantics as `pause`/`connect`/`setTimeout`), carrying the API
error message when present; per-fork 404 error codes map to generic
`NotFoundError` (the missing resource is fork-internal, e.g. the
snapshot).
- **python-sdk**: `sandbox.fork(timeout=..., count=...)` /
`Sandbox.fork(sandbox_id, ...)` and the `AsyncSandbox` equivalents (same
`@class_method_variant` instance/static pattern as `connect`/`pause`),
returning `List[Union[Sandbox, Exception]]`. Per-fork errors map through
the shared `api_exception_from_code` (extracted from
`handle_api_exception`). `timeout` is in seconds per Python SDK
convention; an explicit `timeout=0` is preserved. Whole-request 404
raises `SandboxNotFoundException`; per-fork 404 codes map to generic
`NotFoundException`.
- **changesets**: minor bumps for `e2b` and `@e2b/python-sdk`.
## Usage
JS:
```ts
const sandbox = await Sandbox.create()
const [fork1, fork2] = await sandbox.fork({ count: 2, timeoutMs: 60_000 })
if (fork1 instanceof Sandbox) {
await fork1.commands.run('echo "hello from fork"')
}
// or by ID
const forks = await Sandbox.fork(sandbox.sandboxId, { count: 2 })
```
Python (sync / async):
```python
sandbox = Sandbox.create()
fork1, fork2 = sandbox.fork(count=2, timeout=60)
if isinstance(fork1, Sandbox):
fork1.commands.run('echo "hello from fork"')
# or by ID
forks = Sandbox.fork(sandbox.sandbox_id, count=2)
```
```python
sandbox = await AsyncSandbox.create()
fork1, fork2 = await sandbox.fork(count=2)
```
## Notes
- The JS option is named `timeoutMs` (milliseconds) to match
`SandboxOpts.timeoutMs` / `SandboxConnectOpts.timeoutMs`; the API
receives seconds via `timeoutToSeconds` as elsewhere.
- Failed forks are returned as error **values** in the array rather than
rejected promises, so a partial failure doesn't throw away the
successful forks and there are no unhandled-rejection hazards. A
per-fork error message includes the API error code only when the API
returned one.
## Test plan
- [x] `pnpm run format`, `pnpm run lint`, `pnpm run typecheck` pass at
the repo root (`ty` diagnostics identical to baseline)
- [x] Offline tests pass: `count < 1` → `InvalidArgumentError` /
`InvalidArgumentException` in JS, Python sync, and Python async;
`handleApiError` suite passes after the `apiErrorFromCode` extraction
(plus a behavior-parity check of the Python `handle_api_exception`
refactor)
- [ ] Integration tests (single fork with FS state inheritance +
independence, multi-fork with unique IDs, fork-by-ID, fork of killed
sandbox → `SandboxNotFoundError`) are written but currently fail against
prod with 404 because the fork endpoint (e2b-dev/infra#3202) is not
deployed yet — they should pass once it lands.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Adds an optional `name` filter to `Sandbox.listSnapshots()` /
`Sandbox.list_snapshots()`, mirroring the infra snapshots list endpoint
([e2b-dev/infra#3184](https://github.com/e2b-dev/infra/pull/3184)). The
filter accepts a snapshot name or ID, optionally tag-qualified (e.g.
`"my-snapshot"`, `"my-team/my-snapshot"` or `"my-snapshot:v1"`); unknown
names return an empty list. It's a flat top-level option alongside the
existing `sandboxId` filter (non-breaking) and can be combined with it —
the backend applies both with AND, matching the `metadata`+`state`
behavior of `Sandbox.list()`. Applied equivalently across the OpenAPI
spec, generated clients, and the JS + Python sync/async SDKs, with tests
and a changeset.
## Usage
```ts
// JS/TS
const paginator = Sandbox.listSnapshots({ name: 'my-snapshot' })
const snapshots = await paginator.nextItems()
// combine filters (snapshots from a sandbox matching a name)
Sandbox.listSnapshots({ sandboxId: 'sandbox-id', name: 'my-snapshot' })
```
```python
# Python (sync)
paginator = Sandbox.list_snapshots(name="my-snapshot")
snapshots = paginator.next_items()
# Python (async)
paginator = AsyncSandbox.list_snapshots(name="my-snapshot")
snapshots = await paginator.next_items()
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Promotes the merged #1532 (by @anxkhn) from the staging branch
`fix/no-install-recommends-docstring` into `main`.
`TemplateBuilder.apt_install()` documents its `no_install_recommends`
parameter as
"Whether to install recommended packages", but the generated command
does the
opposite. In `packages/python-sdk/e2b/template/main.py` the command adds
apt-get's
`--no-install-recommends` flag when the argument is `True`:
```python
f"... apt-get install -y {'--no-install-recommends ' if no_install_recommends else ''}..."
```
`--no-install-recommends` tells apt to *skip* recommended packages, so
`no_install_recommends=True` skips them rather than installing them. A
user who
follows the docstring gets the inverse of the documented behavior. The
parameter
name and apt-get's own semantics confirm the code is correct and the
docstring was
wrong; this rewords the docstring line to match the real behavior.
The `--no-install-recommends` flag was introduced in #983; the docstring
has been
inverted since then.
This is Python-only. The JS twin `aptInstall` applies the same flag but
has no
per-parameter JSDoc for `noInstallRecommends` (it appears only inside an
`@example`), so there is nothing contradictory to fix on the JS side.
There is a
single Python definition (no sync/async mirror for the template
builder).
No behavior change; documentation-only, plus a `@e2b/python-sdk: patch`
changeset.
### Usage
```python
from e2b import Template
template = Template().from_image("ubuntu:22.04")
# Install recommended packages as well (apt-get default):
template.apt_install("vim")
# Skip recommended packages (adds apt-get's --no-install-recommends):
template.apt_install("vim", no_install_recommends=True)
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Signed-off-by: Anas Khan <83116240+anxkhn@users.noreply.github.com>
Co-authored-by: Anas Khan <anxkhn28@gmail.com>
The Python volume content client factories passed both `proxy` and the
shared cached `transport` to httpx, so with a proxy configured (e.g.
`Volume.connect(volume_id, proxy="http://user:pass@127.0.0.1:8080")`),
every volume operation mounted a fresh, never-closed proxy transport
that bypassed the cached connection pool. The client-level `proxy`
argument is now dropped — the proxy is already baked into the cached
transport, so proxied requests keep working but reuse one pooled
transport per thread/event loop.
The volume transports also gained connect-level retries
(`E2B_CONNECTION_RETRIES`, default 3), matching the core API and envd
transports. Includes a changeset for a `@e2b/python-sdk` patch release.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Replaces the per-call `integration` connection option with a set-once,
process-wide setter — `ConnectionConfig.setIntegration()` in JS and
`ConnectionConfig.set_integration()` in Python — so integrations
wrapping the SDK tag themselves once at startup and every request
carries the identifier in the `User-Agent` header, with no threading
through individual SDK calls. The setter is internal and hidden from
generated docs; the `integration` option is removed from
`ConnectionConfigOpts` (kept as a deprecated alias of `ConnectionOpts`)
and from the Python constructor, and the round-trip machinery from #1459
is no longer needed since rebuilt configs read the process-wide value.
User-Agent handling now follows a single rule in both SDKs via one
shared helper per SDK: an explicitly provided `User-Agent` always wins,
otherwise the SDK sends its own tagged with the current integration —
and SDK-built values are recomputed whenever a config is rebuilt, so
clearing or changing the integration propagates. Tests cover
attribution, clearing, config rebuilds, and custom User-Agent precedence
in both SDKs, with changesets for `e2b` and `@e2b/python-sdk` (minor).
CLI attribution using this setter will follow in a separate PR.
Usage (internal integrations only):
```ts
import { ConnectionConfig } from 'e2b'
ConnectionConfig.setIntegration('e2b-code-interpreter/0.1.0') // once at startup
```
```python
from e2b import ConnectionConfig
ConnectionConfig.set_integration("e2b-code-interpreter/0.1.0") # once at startup
```
A caller-supplied `User-Agent` (via `headers`/`apiHeaders`) is preserved
in both SDKs:
```ts
const sbx = await Sandbox.create({ apiHeaders: { 'User-Agent': 'my-app/1.0' } })
// requests carry: my-app/1.0
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Live volume tests create real volumes against the API; this gates them
behind an `ENABLE_VOLUME_TESTS` env var so they skip by default. In the
JS SDK, the `volumeTest` fixture is chained with
`.skipIf(process.env.ENABLE_VOLUME_TESTS === undefined)`, skipping all
of `tests/volume/file.test.ts`. In the Python SDK, the `volume` and
`async_volume` fixtures call `pytest.skip` when the env var is unset,
gating `tests/{sync/volume_sync,async/volume_async}/test_file.py`.
Mocked and unit volume tests (msw-based `volume.test.ts`,
`test_volume.py`, `test_volume_content.py`, `test_volume_client.py`,
`test_volume_connection_config.py`) still run unconditionally. To run
the live tests: `ENABLE_VOLUME_TESTS=1 pnpm run test` or
`ENABLE_VOLUME_TESTS=1 poetry run pytest`.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Migrates the Python SDK's packaging and CI from Poetry to
[uv](https://docs.astral.sh/uv/): `pyproject.toml` is converted to PEP
621 metadata using uv's native `uv_build` backend (verified to produce a
byte-equivalent wheel containing both `e2b` and `e2b_connect`),
`poetry.lock` is replaced with `uv.lock`, and the `Makefile`,
`package.json` scripts, `.tool-versions`, `CLAUDE.md`, and all six
GitHub workflows now use `uv` (`astral-sh/setup-uv` + `uv
sync`/`build`/`version`/`publish`). It also drops the now-redundant
explicit sync steps (since `uv run` auto-syncs) and removes the orphaned
`pydoc-markdown` dev dependency, whose only consumer was deleted long
ago — trimming 58 packages from the dev lockfile.
## Usage
```sh
cd packages/python-sdk
uv sync # install deps (replaces `poetry install`)
uv run pytest # run tests
uv build # build the wheel/sdist
make lint # ruff (run via `uv run`)
```
No user-facing SDK change — packaging/tooling only — so no changeset is
included; the published package contents are unchanged.
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
### What
Cherry-picks the fix from #1519.
`strip_ansi_escape_codes` in the Python SDK only matched
semicolon-separated CSI
parameters, so colon-separated SGR sequences leaked literal escape
garbage into
template build-log messages. This widens the parameter class from `;` to
`[;:]`
so colon-separated sequences are stripped too, matching the JS SDK's
`stripAnsi`.
Modern terminals emit colon-separated SGR sequences:
- 256-color: `\x1b[38:5:82m`
- truecolor: `\x1b[38:2::255:0:0m`
- curly underline: `\x1b[4:3m`
The two SDKs share one source (chalk/ansi-regex) and the JS twin was
already
updated to support colons (`packages/js-sdk/src/utils.ts:95`, comment:
"supports
; and :"); the Python port lagged behind. `strip_ansi_escape_codes` is
consumed
by `LogEntry.__post_init__`
(`packages/python-sdk/e2b/template/logger.py`), so
the leftover escape bytes showed up in Python build logs only.
### The one-line fix
```python
# packages/python-sdk/e2b/template/utils.py:319
- r"(?:(?:\d{1,4}(?:;\d{0,4})*)?[\dA-PR-TZcf-nq-uy=><~]))",
+ r"(?:(?:\d{1,4}(?:[;:]\d{0,4})*)?[\dA-PR-TZcf-nq-uy=><~]))",
```
### Usage example (before / after)
```python
from e2b.template.utils import strip_ansi_escape_codes
# 256-color, colon-separated
strip_ansi_escape_codes("\x1b[38:5:82mX\x1b[0m")
# before: ":5:82mX" after: "X"
# truecolor, colon-separated
strip_ansi_escape_codes("\x1b[38:2::255:0:0mRED\x1b[0m")
# before: ":2::255:0:0mRED" after: "RED"
# semicolon variants already worked and still do
strip_ansi_escape_codes("\x1b[38;5;82mX\x1b[0m") # "X" (unchanged)
```
### Tests
Unit tests at
`packages/python-sdk/tests/shared/template/utils/test_strip_ansi_escape_codes.py`
(no API key / sandbox): colon-256, colon-truecolor, curly-underline,
plus
basic/semicolon regressions. All 7 pass locally.
### Changeset
`.changeset/python-strip-ansi-colon.md` (patch on `@e2b/python-sdk`).
### Notes
Original PR: #1519 (by @anxkhn). Opened against a fresh branch off
`main` per
request, rather than merging #1519 directly.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Anas Khan <83116240+anxkhn@users.noreply.github.com>
Adds a `gzip` option to the template `.copy()` / `copyItems` layer that
controls whether copied files are gzipped before upload, threaded from
the copy call through the build-time tar stream in both the JS SDK and
the sync/async Python SDKs. It is enabled by default to preserve
existing behavior, so passing `gzip: false` (`gzip=False`) uploads an
uncompressed tar — useful for already-compressed payloads where gzip
adds CPU cost without shrinking the upload. The option name matches
node-tar's own `gzip` option and the existing sandbox filesystem `gzip`
kwarg. Gzip is deliberately excluded from the file cache hash, so
toggling it does not bust the build cache. Tests in both SDKs were
updated for the new argument and extended with `gzip: false` cases
asserting the archive is not gzipped yet still extracts, and a changeset
(`minor` for both packages) is included.
> [!NOTE]
> The server that extracts these uploaded archives lives in another repo
and must auto-detect compression (peek the gzip `0x1f 0x8b` magic)
rather than assuming gzip; confirm it handles plain tars before release.
## Usage
```ts
// JS/TS
template.copy('model.bin', '/app/', { gzip: false })
template.copyItems([{ src: 'a.bin', dest: '/app/', gzip: false }])
```
```python
# Python (sync & async)
template.copy('model.bin', '/app/', gzip=False)
template.copy_items([{ 'src': 'a.bin', 'dest': '/app/', 'gzip': False }])
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Integration branch PR for #1500. Merges the docstring fix into `main`.
Once #1500 is merged into `python-sdk-list-docstring-base`, this PR will
carry those changes into `main`.
---------
Co-authored-by: Leinux <tristone13th@outlook.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Extracts the cursor-based pagination state machine into a reusable base
class — `Paginator` in the JS SDK's `utils`, `PaginatorBase` in
`e2b/utils.py` — that owns `hasNext`/`nextToken` and the `x-next-token`
header handling, and migrates the sandbox and snapshot paginators onto
it. Each concrete paginator now just implements `nextItems`/`next_items`
to fetch its own page, so future list endpoints (templates, builds,
etc.) can add pagination by subclassing without reimplementing the
bookkeeping. Applied equivalently to the JS SDK and both Python sync and
async implementations, with unit tests covering the shared base. There
are no public API changes — `Sandbox.list()` / `listSnapshots()` and the
existing paginator types behave identically.
## Usage (unchanged)
```ts
const paginator = Sandbox.list()
while (paginator.hasNext) {
const sandboxes = await paginator.nextItems()
console.log(sandboxes)
}
```
```python
paginator = Sandbox.list()
while paginator.has_next:
sandboxes = paginator.next_items()
print(sandboxes)
```
## Summary
Template builds previously buffered the entire gzipped build-context tar
archive in memory before uploading it. This PR spools the archive to a
temporary file and streams it from disk during upload — in the JS SDK
and both sync and async Python SDKs — so memory usage no longer scales
with the size of the build context.
The upload keeps an explicit `Content-Length` header (taken from the
spooled file's size), which S3 presigned PUT URLs require — they reject
`Transfer-Encoding: chunked` with `501 NotImplemented` (#1243).
## Changes
- **JS** (`packages/js-sdk/src/template/`):
`tarFileStream`/`tarFileStreamUpload` are replaced by `tarFileToStream`,
which writes the archive to a temp file and returns a self-cleaning read
stream plus its `size`. The spooled temp file deletes itself once the
stream is closed (consumed, errored, or destroyed) via the stream's
`close` event — mirroring the Python SDK's `tar_file_stream`. `buildApi`
streams this body with `duplex: 'half'` and an explicit `Content-Length`
from `size`; if `fetch` throws before consuming the body, it destroys
the stream to trigger the same cleanup. There is no separate cleanup
callback, so a cleanup failure can no longer mask the upload result.
- **Python** (`packages/python-sdk/e2b/template/utils.py`,
`template_async/build_api.py`, `template_sync/build_api.py`):
`tar_file_stream` now writes to a `tempfile.TemporaryFile` instead of
`io.BytesIO` and returns the file object positioned at the start; the
upload streams from it with an explicit `Content-Length` and closes it
(deleting the temp file) when done.
- Tests updated for the new return shapes (JS `tarFileToStream.test.ts`,
`uploadFile.test.ts`; Python upload/tar tests), including assertions
that the spooled archive is removed on both the consume and destroy
paths.
## Usage
No API changes — `Template.build()` / template builds behave the same,
just without holding the build context in memory:
```ts
await Template.build(template, { alias: 'my-template' })
```
```python
Template.build(template, alias="my-template")
```
Split out of #1433, which covers streaming for sandbox/volume file
uploads and downloads.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Adds a `logger` option (a standard library `logging.Logger`) to
`Sandbox.create`/`AsyncSandbox.create` and the static
`Sandbox.connect(sandbox_id, ...)`, wired into the API client, the envd
client, the volume content client, and the RPC (ConnectRPC) path. The
logger is stored on the sandbox and propagates to all of its later
operations — including control-plane calls like
`kill`/`pause`/`set_timeout`/`get_info` (via `get_api_params`) — so
logging keeps working after construction; mirroring the JS SDK, `logger`
is a construction-time option and not a public per-request parameter
those methods accept from the caller, and nothing is logged unless a
logger is supplied. The stdlib `logging.Logger` is used directly as the
adapter (no ported JS `Logger` interface), and log levels match JS:
requests at `INFO`, successful API and unary RPC responses at `INFO`,
streamed RPC messages at `DEBUG`, failed API responses (status >= 400)
at `ERROR`. The always-on module-level (`e2b.*`) request logging at the
transport layer was removed in favor of this opt-in client-layer
logging, and volume content operations continue to accept `logger` per
call via `VolumeApiParams` to match the JS Volume API. Includes a
changeset and unit tests in `tests/test_logging_option.py`.
## Usage
```python
import logging
from e2b import Sandbox
logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger("my-app.e2b")
sbx = Sandbox.create(logger=logger)
sbx.commands.run("echo hello") # RPC logged via `logger`
sbx.set_timeout(60) # control-plane call also logged via `logger`
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: Matt Brockman <matt.brockman@e2b.dev>
## Filesystem-only auto-pause (`onTimeout` object form)
Adds an object form to the sandbox **lifecycle** `onTimeout`
(`on_timeout` in Python) that controls the snapshot kind taken when a
sandbox auto-pauses on timeout, via `keepMemory` (`keep_memory`).
`onTimeout` now accepts either the existing bare action (`'pause'` /
`'kill'`) or the object form `{ action, keepMemory }`. When `keepMemory`
is `false` (with `action: 'pause'`), a timeout auto-pause takes a
**filesystem-only** snapshot (no memory) instead of a full memory one,
so the sandbox cold-boots (reboots) from disk on resume — losing running
processes and open connections. Defaults to `true` (full memory
snapshot), so existing callers are unaffected. **The bare string form is
unchanged.**
It's the create-time / auto-pause counterpart to the explicit
`pause(keepMemory=false)` from #1465: same `keepMemory` naming, mapped
onto the `autoPauseMemory` create field.
### Type safety
The object form is a **discriminated union** on `action`: `keepMemory`
is only valid with `action: 'pause'`. Pairing it with `action: 'kill'`
is a **compile-time type error** (TS) / static error (`ty`), and is
additionally rejected at runtime (`InvalidArgumentError` /
`InvalidArgumentException`) for untyped callers.
### Behavior & validation
- `keepMemory` only applies to a `pause` action.
- **Incompatible with auto-resume** — auto-resume wakes a paused sandbox
on inbound traffic by restoring its memory snapshot in place; a
filesystem-only snapshot has no memory to restore (resuming cold-boots
it), so it must be resumed explicitly via `connect()`. Combining
`keepMemory: false` with `autoResume` is rejected client-side.
### Usage
```ts
// JS/TS — filesystem-only auto-pause on timeout
const sbx = await Sandbox.create({
lifecycle: { onTimeout: { action: 'pause', keepMemory: false } },
})
// bare string form still works (full memory snapshot)
const sbx2 = await Sandbox.create({ lifecycle: { onTimeout: 'pause' } })
```
```python
# Python
sbx = Sandbox.create(
lifecycle={"on_timeout": {"action": "pause", "keep_memory": False}}
)
```
### Changes
- `spec/openapi.yml`: `autoPauseMemory` on the create body (+
regenerated JS/Python clients).
- JS `SandboxOnTimeout` discriminated union (`'pause' | 'kill' | {
action: 'pause'; keepMemory? } | { action: 'kill' }`) and the Python
`SandboxOnTimeoutPause` / `SandboxOnTimeoutKill` TypedDicts, wired
through `createSandbox` / `_create_sandbox` (sync + async) to
`autoPauseMemory`, with the client-side guards.
- Tests: payload serialization + validation (offline, incl. the `action:
'kill'` type/runtime guard) and live cold-boot e2e in both SDKs;
changeset (`e2b` + `@e2b/python-sdk`, minor).
### Backend dependency
The live e2e tests exercise the real auto-pause→cold-boot path and
require the infra-side `autoPauseMemory` support (e2b-dev/infra#3055),
now merged and deployed.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Signed-off-by: Babis Chalios <babis.chalios@e2b.dev>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
## Summary
Adds an optional **`memory`** flag to `pause` in both the JS and Python
SDKs. When `memory` is `false`, the pause captures **only the
filesystem** (no memory snapshot); resuming such a snapshot **cold-boots
(reboots)** the sandbox from disk — losing in-memory state, running
processes, and open connections. Defaults to `true` (full memory
snapshot), so existing callers are unaffected.
This is the SDK surface for the filesystem-only snapshot feature on the
infra side.
## Usage
```ts
// JS / TS
const sbx = await Sandbox.create()
await sbx.pause({ memory: false }) // filesystem-only snapshot
const resumed = await sbx.connect() // resumes by cold-booting from disk
```
```python
# Python (sync)
sbx = Sandbox()
sbx.pause(memory=False) # filesystem-only snapshot
resumed = sbx.connect() # resumes by cold-booting from disk
# Python (async)
sbx = await AsyncSandbox.create()
await sbx.pause(memory=False)
resumed = await sbx.connect()
```
`memory` defaults to `true` — `pause()` / `pause({})` behave exactly as
before.
## What changed
- **spec**: optional `memory: boolean` (default `true`) on `POST
/sandboxes/{sandboxID}/pause` (`SandboxPauseRequest`); both API clients
regenerated via `make codegen`.
- **JS**: `Sandbox.pause` / `betaPause` accept `{ memory }` →
`SandboxApi.pause` sends the request body.
- **Python**: `pause(memory=...)` / `beta_pause` → `_cls_pause` (sync +
async) sends `SandboxPauseRequest(memory=...)`.
- **Tests**: filesystem-only pause+resume reboots the guest while the
filesystem survives — JS (`tests/sandbox/snapshot.test.ts`) and Python
sync + async. All pass against a local stack; `format` / `lint` /
`typecheck` clean.
- **Changeset**: `minor` for `e2b` and `@e2b/python-sdk`.
## Note (related infra observation, not addressed here)
While testing, a filesystem-only **resume cold-boots into a different
default exec context** (`root` / `/root`) than a memory resume (`user` /
`/home/user`). The filesystem itself is fully intact; tests use absolute
paths to be robust to this. Worth confirming on the infra reboot path
whether the template's default user should be restored after a cold
boot.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Signed-off-by: Babis Chalios <babis.chalios@e2b.dev>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
## Description
This PR fixes two related issues in the command handle's event handling.
### 1. JS `CommandHandle.disconnect()` leaked the output subscription
`disconnect()` was fire-and-forget — it only triggered the transport
abort and relied entirely on HTTP/2 abort propagation to stop events,
which is unreliable under keepalive: `onStdout`/`onStderr`/`onPty` could
keep firing for output produced after `disconnect()` returned.
`disconnect()` now sets a cooperative `disconnected` flag and aborts the
transport. The flag is checked before every callback dispatch in the
event loop, so once `disconnect()` returns no callback fires for output
that arrives (or was buffered) after the call — even if the underlying
abort hasn't torn the stream down yet. It does **not** wait for the
event handler to drain, so it returns promptly even for an idle command
(e.g. `sleep`) whose stream produces no further output, never blocks on
an in-flight callback, and does not deadlock when awaited from inside a
callback.
The async Python SDK was already correct here (`disconnect()` cancels
the event-handling task), and the sync Python SDK has no background
subscription (events are consumed only while the caller iterates). The
added Python tests confirm both.
### 2. Exit code was lost when a disconnected consumer stopped on a
flushed `end`-event chunk
When the `end` event flushes trailing decoder bytes (an incomplete
multibyte sequence → replacement character) and the consumer stops
iterating on the first flushed chunk, the generator was aborted before
the result was assigned, so `wait()` failed as if the process never
produced a result. The `end` handler now records the result **before**
yielding the flushed chunks, across the JS, async Python, and sync
Python SDKs.
## Usage
```js
const handle = await sandbox.commands.run(daemon, { background: true, stdin: true, onStdout })
await sandbox.commands.sendStdin(handle.pid, 'turn1\n')
await handle.disconnect() // resolves promptly; onStdout will not fire again
await sandbox.commands.sendStdin(handle.pid, 'turn2\n') // turn2 output never reaches onStdout
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
merged from https://github.com/e2b-dev/E2B/pull/1457
## Problem
`AsyncCommandHandle` / `CommandHandle` accumulate streamed output with
`self._stdout += out` per chunk. Because `self._stdout` is an instance
attribute (`STORE_ATTR`), CPython's in-place string-concatenation
optimization — which only applies to local `STORE_FAST` targets —
doesn't apply, so each append re-copies the entire buffer. For commands
that emit large volumes of output this becomes O(n²) in total bytes, and
in async contexts it stalls the event loop for hundreds of ms per chunk
near the tail.
## Fix
Buffer decoded chunks in a `list[str]` and `"".join()` them on read.
This restores linear-time accumulation and keeps streaming responsive,
with no change to the resulting `stdout`/`stderr` values or the public
API.
## Notes
- Applies the same change to both the sync and async command handles.
- Pure internal change; the incremental UTF-8 decoding behavior is
preserved.
- Changeset included (`@e2b/python-sdk`, patch).
Co-authored-by: davidzeng-pplx <david.zeng@perplexity.ai>
Follow-up to #1433. Builds on the shared streaming infrastructure
introduced there (`FILE_TIMEOUT_MS`, request-controller/stream-cleanup
helpers in `connectionConfig`, `io_utils` chunk iterators, the `runtime`
guard) and applies the same streaming model to volumes.
> [!NOTE]
> Based on `mishushakov/stream-write-file-upload` (#1433). Merge that PR
first; this PR's diff will then retarget to `main` automatically.
## What changed
- **`Volume.writeFile()` / `Volume.write_file()`** — stream the request
body instead of buffering it in memory.
- JS: `ReadableStream` data is streamed outside the browser
(half-duplex); browsers still buffer since they can't stream request
bodies.
- Python: file-like objects are streamed in chunks (async wraps them in
an async iterator; sync passes them to httpx directly, text-mode IO is
encoded chunk-by-chunk).
- **`Volume.readFile(format="stream")` / `read_file(format="stream")`**
— the request timeout now bounds only the initial handshake, not the
body read, matching the sandbox `files.read` stream path. A dropped
connection during the handshake surfaces the same typed, health-checked
error; JS supports `signal` to cancel an in-flight stream and cancels
unconsumed bodies on error so the pooled connection is released.
## Usage
JS — stream a file straight to a volume without buffering:
```ts
import { createReadStream } from 'node:fs'
import { Readable } from 'node:stream'
const stream = Readable.toWeb(createReadStream('large-input.bin'))
await volume.writeFile('/data/large-input.bin', stream)
// read back as a stream; the body lives until consumed/cancelled
const out = await volume.readFile('/data/large-input.bin', { format: 'stream' })
for await (const chunk of out) {
// process chunk
}
```
Python — stream a file-like object:
```python
with open("large-input.bin", "rb") as f:
volume.write_file("/data/large-input.bin", f) # streamed, not read() into memory
for chunk in volume.read_file("/data/large-input.bin", format="stream"):
... # process chunk
```
## Testing
- `pnpm run format`, `pnpm run lint`, `pnpm run typecheck` pass.
- Added volume streaming tests (JS `tests/volume/file.test.ts`; Python
sync/async `test_file.py` text-stream cases).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
## Description
Removes full in-memory buffering from the SDK **sandbox** file-transfer
paths, in both JS and Python (sync + async).
**Streamed uploads** — `Sandbox.files.write` / `write_files` streams
`ReadableStream` (JS, outside the browser) and file-like (Python) input
to the sandbox with chunk-by-chunk gzip compression, instead of
buffering the whole body in memory. `useOctetStream`/`use_octet_stream`
now defaults to auto-detect — octet-stream when any entry is streamable
(so streamed uploads aren't silently buffered), `multipart/form-data`
otherwise; browsers always use `multipart/form-data` since streaming
request bodies aren't supported there. A streamed upload is bounded by a
per-chunk timeout on the wire (Python's per-write `httpx` timeout,
default the request timeout); a stalled upload the wire can't observe is
bounded server-side. On Python's `AsyncSandbox`, the blocking file reads
and gzip compression of a streamed upload now run in a worker thread so
a large upload doesn't stall the event loop.
**Streamed downloads** — `Sandbox.files.read(format="stream")` now
streams the response body from the sandbox instead of downloading it
into memory before iterating (Python sync + async), and the 60s request
timeout no longer kills the stream while it's being consumed:
- The request timeout now bounds only the initial handshake.
- The body is bounded by a per-chunk **idle-read timeout** on the wire —
a per-`read()` option (`streamIdleTimeoutMs` in JS,
`stream_idle_timeout` in Python; default the request timeout — 60s —
`0`/`None` to disable). It's armed only while waiting on a network read
and cleared the moment a chunk arrives, so it aborts only when the
server stops sending mid-stream; a slow or paused consumer never trips
it (a held-but-unread stream is reclaimed server-side, not by this
timer).
- A dropped connection during the handshake surfaces the same typed,
health-checked error as non-stream reads. In JS, `signal` can still
cancel an in-flight stream.
- The stream holds its pooled connection until it is consumed to the
end, cancelled/closed, errors, or the idle timeout fires — consume it
fully, use the context manager, or close it. (This replaces the earlier
GC-finalizer net.) Python returns a
`FileStreamReader`/`AsyncFileStreamReader` supporting deterministic
cleanup via `close()`/`aclose()` and (async) context-manager use; both
still satisfy `Iterator[bytes]`/`AsyncIterator[bytes]`, so existing
iteration is unchanged.
**Empty files** — JS `Sandbox.files.read()` with `blob` or `stream`
format now returns a format-correct empty value (empty `Blob` / empty
`ReadableStream`) for empty files instead of `""`.
> [!NOTE]
> The equivalent **volume** streaming changes
(`Volume.writeFile`/`write_file`, `Volume.readFile`/`read_file` streams)
live in a follow-up PR, #1453, which is based on this branch.
## Usage
```ts
// JS: upload a large file without holding it in memory
const file = createReadStream('large.bin')
await sandbox.files.write('large.bin', Readable.toWeb(file), { gzip: true })
// JS: consume a download for longer than 60s without it being killed
const stream = await sandbox.files.read('large.bin', { format: 'stream' })
for await (const chunk of stream) { /* ... */ }
// JS: tune (or disable) the per-chunk idle-read timeout for a read
const stream = await sandbox.files.read('large.bin', {
format: 'stream',
streamIdleTimeoutMs: 120_000, // 0 to disable
})
// JS: empty files now return format-correct empty values
const blob = await sandbox.files.read('empty.txt', { format: 'blob' }) // Blob (size 0), not ''
```
```python
# Python: streamed upload and download
with open("large.bin", "rb") as f:
sandbox.files.write("large.bin", f, gzip=True)
for chunk in sandbox.files.read("large.bin", format="stream"):
...
# Python: deterministic cleanup when not reading the stream to the end
with sandbox.files.read("large.bin", format="stream") as stream:
first_chunk = next(iter(stream)) # connection released on block exit
# Python: tune (or disable) the per-chunk idle-read timeout for a read
for chunk in sandbox.files.read(
"large.bin", format="stream", stream_idle_timeout=120.0 # None to disable
):
...
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
## Why
The Release run
[27844631141](https://github.com/e2b-dev/E2B/actions/runs/27844631141/job/82412481993)
**published all packages successfully** but then failed at the final
*"Commit new versions"* step — the version-bump commit-back to \`main\`
was rejected as non-fast-forward (another PR landed on \`main\` during
the release window).
As a result the registries are ahead of the repo:
| Package | Published | Repo (main) before this PR |
|---|---|---|
| \`e2b\` (JS) | 2.30.4 (npm) | 2.30.3 |
| \`@e2b/cli\` | 2.12.2 (npm) | 2.12.1 |
| \`e2b\` (Python) | 2.29.4 (PyPI) | 2.29.3 |
The changeset \`fix-logo-pypi-npm.md\` was also never consumed and is
still on \`main\`.
## What this PR does
Replays exactly what the failed *"Commit new versions"* step would have
committed — i.e. \`pnpm run version\` (changeset version +
\`postVersion\` poetry sync) + lockfile update:
- Bumps \`e2b\` → 2.30.4, \`@e2b/cli\` → 2.12.2, \`@e2b/python-sdk\` →
2.29.4 (matching what's already published)
- Deletes the consumed changeset \`fix-logo-pypi-npm.md\`
- Updates \`pnpm-lock.yaml\` (CLI's \`e2b\` dep → 2.30.4)
No new packages are published by merging this — it only syncs the repo
to the registries. **Do not re-run the Release workflow** for this
changeset; the versions already exist on npm/PyPI.
The Python SDK's `ApiClient` applied the deprecated `access_token`
*after* merging `config.headers` (which includes `api_headers`), so a
custom `Authorization` passed via `api_headers` was silently overwritten
— the opposite of the JS SDK, where a custom `Authorization` wins. This
moves the deprecated access token before `config.headers` so
`api_headers` now takes precedence, matching JS. No change when only
`access_token` is set.
```python
# Custom Authorization via api_headers now wins over the deprecated access_token
ConnectionConfig(api_key="e2b_...", access_token="old", api_headers={"Authorization": "Bearer custom"})
# -> Authorization: Bearer custom
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Summary
Fixes the duplicate logo issue on NPM and PyPI caused by #1462. The
`#gh-light-mode-only` / `#gh-dark-mode-only` URL fragments are
GitHub-specific — NPM and PyPI ignore them and render both `<img>` tags.
Switches all three package READMEs (CLI, JS SDK, Python SDK) to
`<picture>` elements:
```html
<picture>
<source media="(prefers-color-scheme: dark)" srcset=".../logo-white.png">
<source media="(prefers-color-scheme: light)" srcset=".../logo-black.png">
<img alt="E2B Logo" src=".../logo-black.png" width="200">
</picture>
```
- **GitHub**: `<picture>` + `prefers-color-scheme` handles theme
switching
- **NPM/PyPI**: `<picture>` not supported, falls back to the single
`<img>` (black logo)
Link to Devin session:
https://app.devin.ai/sessions/4983f23d23934d2c9a51733f5f9920f3
Requested by: @mlejva
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: vasek <vasek.mlejnsky@gmail.com>
## Summary
Replace the old `logo-circle.png` in the CLI, JS SDK, and Python SDK
READMEs with the new E2B wordmark logos that adapt to GitHub's theme
setting.
Each package README now uses a `<picture>` element:
```html
<picture>
<source media="(prefers-color-scheme: dark)" srcset=".../logo-white.png">
<source media="(prefers-color-scheme: light)" srcset=".../logo-black.png">
<img alt="E2B Logo" src=".../logo-black.png" width="200">
</picture>
```
- **Light theme** → black logo (`logo-black.png`)
- **Dark theme** → white logo (`logo-white.png`)
- **NPM/PyPI** (no `<picture>` support) → falls back to the black logo
via the `<img>` tag
New logo assets added to `readme-assets/`: `logo-black.png`,
`logo-white.png`.
Includes a patch changeset for `@e2b/cli`, `e2b` (JS SDK), and
`@e2b/python-sdk`.
Link to Devin session:
https://app.devin.ai/sessions/4983f23d23934d2c9a51733f5f9920f3
Requested by: @mlejva
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: vasek <vasek.mlejnsky@gmail.com>
## Summary
The access token was only ever used by the CLI, never by any SDK
operation — sandbox, template, and volume calls all authenticate with
the API key. This cleans up the auth plumbing and **deprecates** (rather
than removes) the access token on `ConnectionConfig`, so there's no
breaking change for direct SDK consumers.
## Changes
- **Deprecated** the `accessToken` (JS) / `access_token` (Python) option
on `ConnectionConfig`. It still works exactly as before — when set (or
via `E2B_ACCESS_TOKEN`) the `Authorization: Bearer` header is still sent
— but `apiHeaders` is now the recommended way to pass custom auth.
- **Clear error when the API key is missing**, pointing to the API Keys
tab (`https://e2b.dev/dashboard?tab=keys`). In JS this is gated by a
`requireApiKey` option (default `true`) so callers that authenticate
differently — like the CLI hitting `/teams` with an access token — can
opt out; in Python the API key is always required.
- Removed the unused access-token toggle from the API clients:
`requireAccessToken` (JS) / `require_access_token` (Python). No caller
ever set it to a non-default value, so behavior is unchanged.
- The CLI now passes the access token to the `/teams` endpoint via
`apiHeaders` instead of the deprecated option, and opts out of the
API-key requirement on its own clients.
- Decoupled the sandbox-scoped envd access token from
`ConnectionConfig`: `EnvdApiClient` now owns its own `envdAccessToken`
field and sets the `X-Access-Token` header itself, removing a redundant
manually-set header.
## Recommended usage
```ts
// Deprecated
new ConnectionConfig({ accessToken: 'my-token' })
// Preferred
new ConnectionConfig({ apiHeaders: { Authorization: 'Bearer my-token' } })
```
```python
# Deprecated
ConnectionConfig(access_token="my-token")
# Preferred
ConnectionConfig(api_headers={"Authorization": "Bearer my-token"})
```
## Verification
`pnpm run typecheck`, `pnpm run lint`, Python `make typecheck`, and the
unit tests all pass — including new tests for the API-key requirement
(and its opt-out) in both SDKs. Confirmed the `Authorization: Bearer`
header is still sent for both the deprecated option and
`E2B_ACCESS_TOKEN`.
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fixes seven template-build correctness bugs across the JS and Python
(sync + async) SDKs, plus a small shell-quoting cleanup. Each fix has
unit/regression coverage, and real end-to-end builds were run against
the API in all three SDK variants.
**Template fixes**
- `getAllFilesInPath` now sorts by full path so the files hash no longer
depends on filesystem traversal order (the JS `sort()` was a no-op on
glob `Path` objects).
- `waitForPort` anchors the port match so port 80 no longer matches
8080.
- The readycmd helpers (`waitForURL`/`waitForFile`/`waitForProcess`) and
the file-op helpers (`remove`/`rename`/`makeDir`/`makeSymlink`) now
shell-quote interpolated values/paths.
- `waitForBuildFinish` keeps fetching logs after a terminal status so
the tail of the build logs (beyond the API's 100-entries-per-call limit)
is no longer dropped.
- COPY instructions now collect one stack trace each, so failed-step
traces stay aligned after `copy()` with multiple sources or
`copyItems()`.
- JS `LogEntry` strips ANSI escape codes in its constructor, matching
the Python SDK.
**Cleanup:** consolidated three duplicate single-quote shell helpers
(`shellQuote`, the new `quoteShellArg`, and git's `shellEscape`) into
one faithful `shlex.quote` port in `utils.ts` — safe values stay
unquoted, keeping generated commands and layer-cache hashes stable.
**Behavior note:** templates with paths/URLs containing spaces or shell
metacharacters now build correctly; plain paths are unchanged, so
existing layer caches are preserved.
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
## Summary
Follow-up to #1444 (the `httpcore` → `TimeoutException` mapping, now
merged). That mapping made the flaky timeout deterministic, but the
underlying cause remained: the streaming HTTP `read` timeout was set to
the command `timeout`, so it raced the server's own `deadline_exceeded`
response. This PR removes the read timeout on streaming calls entirely
and relies on the server-side `connect-timeout-ms` header to enforce the
command timeout — matching the JS SDK, which has no per-chunk read
timeout. The race is now structurally impossible rather than mapped
over.
## Usage
```python
cmd = sandbox.commands.run("sleep 10", timeout=2, background=True)
try:
for _ in cmd:
pass
except TimeoutException:
print("command timed out") # server-side deadline_exceeded, no transport race
```
## Tradeoff
A silently dropped connection (no RST) on a command with `timeout=0`
(disabled) now has no client-side read backstop and relies on keepalive
pings — the same posture as the JS SDK.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Summary
The flaky test `test_run_with_too_short_timeout_iterating` failed
intermittently because, when iterating a background command's output,
the Python SDK sets the HTTP stream `read` timeout to the command
`timeout` — so it races the server's own `deadline_exceeded` response.
When the client read timeout won, a raw `httpcore.ReadTimeout` leaked
out instead of a `TimeoutException`. This PR maps
`httpcore.TimeoutException` to `TimeoutException` in
`handle_rpc_exception`, so callers get a consistent timeout error
regardless of which side fires first, plus unit tests for the mapping.
It also adds a JS parity test (JS was never affected — connect-es always
normalizes timeouts into an already-mapped `ConnectError`).
## Usage
```python
cmd = sandbox.commands.run("sleep 10", timeout=2, background=True)
try:
for _ in cmd:
pass
except TimeoutException:
print("command timed out") # now raised reliably, no raw httpcore.ReadTimeout
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Summary
Allow disabling client-side API key **format** validation. Previously
the SDKs hard-required keys to match the `e2b_<hex>` pattern, which
blocked deployments that issue API keys with a different format. Instead
of a custom-prefix override, this adds a simple on/off toggle.
The default behaviour is unchanged (validation stays **on**).
## Configuration
| Form | JS | Python |
| --- | --- | --- |
| Env var | `E2B_VALIDATE_API_KEY=false` | `E2B_VALIDATE_API_KEY=false`
|
| Connection option | `validateApiKey: false` | `validate_api_key=False`
|
The connection option takes priority over the environment variable.
## Usage
**JavaScript / TypeScript**
```ts
import { Sandbox } from 'e2b'
// Via connection option
const sandbox = await Sandbox.create({
apiKey: 'custom_key_format',
validateApiKey: false,
})
// Or via env var: E2B_VALIDATE_API_KEY=false
```
**Python**
```python
from e2b import Sandbox
# Via connection option
sandbox = Sandbox(
api_key="custom_key_format",
validate_api_key=False,
)
# Or via env var: E2B_VALIDATE_API_KEY=false
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Summary
- Python `upload_url`/`download_url` now raise
`InvalidArgumentException` when `use_signature_expiration` is passed for
an unsecured sandbox, matching the JS SDK (which now throws
`InvalidArgumentError` instead of a plain `Error`).
- A signature expiration of `0` was treated as falsy and silently
produced a never-expiring signed URL; it now produces an immediately
expiring URL in both SDKs.
- Adds unit tests mirrored across both SDKs
(`tests/sandbox/urls.test.ts` ↔ `tests/test_sandbox_urls.py`) plus a
changeset.
## Usage
```python
sbx = Sandbox() # not secure=True
sbx.download_url("a.txt", use_signature_expiration=120) # now raises InvalidArgumentException instead of silently ignoring the expiration
```
```ts
const sbx = await Sandbox.create({ secure: true })
await sbx.downloadUrl('a.txt', { useSignatureExpiration: 0 }) // URL now expires immediately instead of never
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
## Summary
- Python `write()` / `write_files()` now return `WriteInfo.type` as the
`FileType` enum instead of the raw API string (sync and async; the JS
string union was already correct, and the volumes client already
converts to `VolumeEntryStatType`).
- `EntryInfo.modified_time` is now timezone-aware UTC (protobuf
`ToDatetime()` returns naive datetimes by default), and naive volume
`atime`/`mtime`/`ctime` timestamps are normalized to UTC.
- `gzip=true` uploads now imply the `application/octet-stream` path in
both JS and Python instead of being silently ignored on the default
`multipart/form-data` path; on envd < 0.5.7 the upload falls back to
uncompressed multipart, matching the existing `use_octet_stream`
fallback.
- Adds sandbox-free unit tests for the model conversions, strengthens
write/info integration test assertions, and includes changesets for
`@e2b/python-sdk` and `e2b`.
## Usage examples
```python
info = sandbox.files.write("hello.txt", "hi")
info.type == FileType.FILE # was the raw string "file"
entry = sandbox.files.get_info("hello.txt")
entry.modified_time.tzinfo # datetime.timezone.utc (was None)
sandbox.files.write("big.bin", data, gzip=True) # now actually gzip-compressed
```
## Test plan
- [x] `pytest tests/test_filesystem_models.py` (new unit tests, 5
passed)
- [x] Python sync + async integration tests for `write`, `info`,
`content_encoding` (16 each, passed against live sandboxes)
- [x] JS `write.test.ts` + `contentEncoding.test.ts` (14 passed)
- [x] `pnpm run format`, `pnpm run lint`, `pnpm run typecheck`
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
## Summary
Fixes three command/PTY streaming issues in the Python and JS SDKs:
- **Multibyte UTF-8 corruption (JS + Python sync/async):** stdout/stderr
were decoded per-chunk, so a UTF-8 character split across two stream
chunks turned into replacement characters. Each handle now keeps a
persistent incremental decoder per stream
(`codecs.getincrementaldecoder` in Python, a shared `TextDecoder` with
`{ stream: true }` in JS) and flushes any incomplete trailing bytes to
`�` on the end event, preserving the existing broken-UTF-8 behavior.
- **`commands.list()` optionals (Python):** now returns `None` instead
of `""` for unset proto3-optional `tag` and `cwd` fields, matching the
declared `Optional[str]` types and the JS SDK.
- **Leaked connections (Python):** command/PTY/watch streams are now
closed when stream setup fails, instead of abandoning the generator (and
its pooled HTTP connection) until GC.
## Usage example
```python
# Split multibyte output is now decoded correctly instead of returning "ð\x9f\x98\x80"-style garbage
result = sandbox.commands.run("printf '😀'")
assert result.stdout == "😀"
# Unset fields are None rather than ""
proc = sandbox.commands.list()[0]
assert proc.tag is None # previously ""
```
## Testing
- New unit tests for incremental/trailing UTF-8 decoding (Python sync +
async, JS).
- Live command/PTY/watch integration suites pass.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
## Description
`Sandbox.connect()` now short-circuits in debug mode instead of calling
the control plane, matching `Sandbox.create()` — fixed in the JS SDK and
both sync/async Python SDKs (static and instance variants). The Python
SDK's `connect()` also previously passed the generated client's `Unset`
sentinel through as the envd/traffic access tokens when they were absent
(non-secure sandboxes), which made `download_url()`/`upload_url()` emit
broken signed URLs; `_cls_connect` now normalizes the response into
`SandboxCreateResponse` with proper `None` values, the same pattern
`_create_sandbox` already uses. Dead `Unset` checks and the now-unused
generated `Sandbox` model import were cleaned up, and unit tests cover
both behaviors in all three implementations. Debug mode is resolved
through `ConnectionConfig`, so the `E2B_DEBUG` env var triggers the
short-circuit in both `connect()` and `create()`, not just an explicit
`debug=True`.
## Usage
```ts
// JS: works fully offline with E2B_DEBUG / debug: true (no control plane call)
const sbx = await Sandbox.connect(sandboxId, { debug: true })
```
```python
# Python: non-secure sandboxes get unsigned URLs again instead of broken signatures
sbx = Sandbox.connect(sandbox_id)
print(sbx.download_url("file.txt")) # no garbage signature when envd token is absent
# Debug mode skips the control plane, like Sandbox.create()
sbx = Sandbox.connect(sandbox_id, debug=True)
```
## Testing
New unit tests in `tests/sandbox/connect.test.ts`,
`tests/sync/sandbox_sync/test_connect.py`, and
`tests/async/sandbox_async/test_connect.py`; existing connect
integration suites pass against the live API (8/8 sync Python, 8/8 async
Python, 6/6 JS).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
## Summary
Python's `$` regex anchor also matches just before a trailing newline,
so file-metadata keys/values ending in `\n` passed client-side
validation (unlike the JS SDK, where `$` matches only the true end of
string) and then failed deep in the HTTP stack with an opaque "illegal
header value" error. This re-anchors both validation regexes with
`\A`/`\Z` so such inputs are rejected upfront with
`InvalidArgumentException`, matching JS behavior and the existing
convention used for the API-key pattern.
Also adds trailing-newline rejection test cases to the sync/async Python
suites and, for parity of coverage, to the JS SDK suite (JS already
rejected them — no behavior change there).
## Usage example
```python
sandbox.files.write("file.txt", "x", metadata={"author": "mish\n"})
# before: passed validation, then httpcore raised 'Illegal header value'
# after: raises InvalidArgumentException with a clear message
```
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
github-actions[bot]