Skip to main content

shellsim

Shellsim is a BusyBox for containers: one small, deterministic process that provides a useful Unix-shaped environment without starting a VM, container runtime, or host subprocess. It is built for experimentation and testing with reinforcement-learning rollouts and agentic environments, where fast startup, reproducibility, isolation, and explicit resource limits matter more than cycle-accurate emulation.

Shell programs, common command-line tools, logical processes, and Python run in-process against an in-memory filesystem. Simulated code cannot access the host filesystem, processes, network, environment, or clock. A trusted harness may copy a selected project into the virtual filesystem before execution; changes never write back to the host.

Compatibility

Shellsim aims for broad compatibility inside clear boundaries. A supported facility should handle almost all ordinary uses, even when obscure flags or legacy behavior remain out of scope. A module or command with no coherent useful subset is omitted instead of being exposed as a misleading stub. Unsupported syntax, options, executable formats, and capabilities fail visibly and are included in structured results.

The current environment includes:

  • a Bash-like shell with pipelines, redirections, functions, common expansions, control flow, background jobs, signals, and job control;
  • common filesystem, text, archive, Git, Make, process, and system commands;
  • deterministic virtual time, network fixtures, /proc, /dev, processes, descriptors, and bounded pipes;
  • a mostly complete Python language runtime with a deliberately selected standard-library and third-party module surface.

Python is source-compatible where supported, not ABI-compatible with CPython. Native extensions, network package installation, and arbitrary host machine code are outside the simulation boundary. A limited Wasm-hosted C toolchain can be supplied as a virtual executable. The simulated pip and uv paths can activate packages already bundled with shellsim; they reject other packages instead of fetching them. See Python in shellsim for the current contract.

git provides a porcelain subset backed by the virtual filesystem: staging, commits, history, diffs, branches, tags, merges, stashes, and search, with output that matches real Git where an agent is likely to parse it. Networked subcommands and content conflicts are refused rather than approximated. See Git in shellsim for the supported surface and its boundaries.

Common utilities accept clustered short options, long options, option values, and -- through a shared parser. Each utility declares its supported options. For example, grep supports ordinary basic and extended regular expressions, recursive search, fixed strings, word and whole-line matching, pattern files, include and exclude filters, context, counts, line numbers, match limits, and common output controls. sed covers addresses and ranges plus the common substitution, selection, text, transliteration, and early-exit commands. awk parses a typed language subset with record rules, control flow, fields, associative arrays, arithmetic, regular expressions, and the usual scalar functions. Unsupported syntax and options exit nonzero with a direct diagnostic; text processing is UTF-8-only unless a command documents a byte-oriented mode.

Install and run

Install the Python package and console command:

python -m pip install shellsim
shellsim -c 'printf "b\na\n" | sort'
shellsim --root ./project -c 'python3.14 test.py'

--root copies the selected host tree into a disposable /work snapshot. With no -c and a terminal attached, shellsim starts a persistent interactive session.

To build the Rust binaries from source:

cargo build --release
./target/release/shellsim -c 'echo hello'
./target/release/shellsim eval --cpu 100k --memory 8m -c 'make test'
./target/release/shellsim-python ./project/main.py -- arg1

Limits accept k, m, and g binary suffixes. eval emits a structured result containing the exit status, stdout and stderr, resource use, command trace, and unsupported behavior. unsupported_commands lists recognized or attempted commands that crossed the capability boundary; invocation records include trust, status, and the unsupported reason.

The Python API exposes fresh and persistent environments:

import shellsim

environment = shellsim.Environment(cpu=100_000)
environment.write_file("/work/main.py", "print(6 * 7)\n")
result = environment.run("python3.14 /work/main.py")
assert result.returncode == 0
assert result.stdout == b"42\n"

environment.run() reuses the default shell. To keep independent shell state in one machine, create additional sessions; they share the virtual filesystem and resource limits:

first = environment.create_shell()
second = environment.create_shell()
first.run("cd /work; export LABEL=first")
assert first.run("printf '%s' \"$LABEL\"").stdout == b"first"
assert second.run("printf '%s' \"${LABEL-unset}\"").stdout == b"unset"

Static HTTP fixtures use the same isolated environment. They do not enable sockets, DNS, TLS, or host-network access:

environment = shellsim.Environment(http={
    "https://api.test/items/*": shellsim.HttpResponse(
        status=200,
        headers={"Content-Type": "application/json"},
        body='{"items": []}',
    ),
})
result = environment.run("curl -s https://api.test/items/1")

result = environment.run_python("""
from urllib.request import urlopen
print(urlopen('https://api.test/items/1').read())
""")
assert result.network_requests[0].matched

Python source can bypass shell parsing and quoting while using the same isolated runtime:

result = shellsim.python.run("print(sum(range(10)))", argv=["example"])
result.check_returncode()

environment = shellsim.Environment()
result = environment.run_python("print('persistent VFS, fresh Python interpreter')")

Clock modes

A machine runs on virtual time by default: when every process waits on a timer, the scheduler jumps the clock to the next deadline, so sleep 3600 costs no host time and runs are reproducible. An embedder can boot a machine with Environment::with_limits_and_clock(limits, ClockMode::RealTime), or with the CLI's --real-time option, for interactive demos. Its clock then follows physical time from boot: shell, Python, and Wasm sleeps take real time, and clock reads report elapsed time. Polling never blocks in either mode. An idle real-time machine reports that it is blocked, and Environment::host_wait_until says how long to wait. Simulated programs cannot choose the mode, and a snapshot of a real-time machine is not reproducible.

Experimental Wasm executables

An executable VFS file beginning with the WebAssembly magic bytes runs through a fuel-metered Wasmtime engine. The WASI Preview 1 adapter provides standard streams, arguments, exported environment variables, the virtual clock, deterministic random bytes, and basic regular file access relative to the virtual working directory. It does not expose host files, processes, network, environment, or time. Invalid modules and imports outside the WASI namespace fail with a diagnostic. Unavailable WASI calls trap if reached; they never report fake success.

A Wasm executable runs as its own scheduled process on that process's virtual descriptors. A read or write that would block suspends the guest until the pipe or terminal is ready, so a guest can answer a prompt, sit in the middle of a pipeline, or be stopped by kill or timeout. A write to a pipe with no reader ends the guest with status 141. Fuel yields keep a compute-bound guest from starving other processes, and consumed fuel counts against the machine's CPU budget while the guest runs. A harness fork taken while a guest is running gets a copy of that process that fails with status 126 and a diagnostic, since a live Wasmtime stack cannot be cloned. Clock subscriptions to poll_oneoff, which back libc sleep and nanosleep, block the guest on virtual time like the native sleep command.

This is not full WASI process support. Descriptor subscriptions to poll_oneoff fail with ENOTSUP; sockets and guest process creation remain unsupported. No C compiler is installed by default. The engine accepts Wasm exception instructions. A pinned external TinyCC package and permissively licensed wasi-libc sysroot can be installed into the VFS by a caller; integration tests compile and run C programs through that virtual toolchain. This is a tested subset of libc, not full POSIX support. A prebuilt wasm32-wasip1 command that uses only the listed imports can be written to the VFS with executable permissions and invoked by path or through PATH.

The virtual display lets a Wasm guest present RGBA frames and poll bounded key events. A host-driven session can resume the guest between frames without granting it ambient display, input, or network access.

Agent harness

shellsim serve --root ./project runs a persistent newline-delimited JSON session. It supports bounded execution, streaming actions, VFS operations, checkpoints, workspace diffs, process and resource inspection, and deterministic session forks. shellsim mcp exposes the same environment as a stdio MCP server. shellsim replay scenario.ndjson reruns checked action transcripts. shellsim corpus manifest.json runs frozen shell and Python workloads under a versioned stock environment and emits a classified JSON report. See compatibility testing.

printf '%s\n' \
  '{"id":1,"op":"execute","source":"printf hello > result"}' \
  '{"id":2,"op":"workspace_diff"}' \
  | shellsim serve --root ./project

Resource model

CPU is deterministic fuel, memory is modeled working set, disk is current virtual-filesystem usage, and output bounds materialized stdout and stderr. The defaults are 10,000,000 CPU units, 64 MiB memory, 64 MiB disk, and 4 MiB output. Costs use simple, stable approximations intended to be correct within an order of magnitude. They bound runaway work and make runs comparable; they do not model allocator layouts or instruction timing exactly. Exhaustion is observable and never falls back to an ambient host implementation.

Every CLI command that boots a machine accepts the same options: --cpu, --memory, --disk, and --output (counts accept k, m, or g suffixes), and --real-time. They may precede the command, as in shellsim --cpu 1m -c 'make'. Commands without script arguments (shell, eval, serve, mcp, replay) also accept them after the command name.

For internals and contribution workflow, see implementation, Python, and CONTRIBUTING.md.

Metadata

Release files for shellsim 0.1.19

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for shellsim 0.1.19
File Size Uploaded
shellsim-0.1.19.tar.gz 825.0 kB Details

Built distributions (wheels)

Table of built distributions (wheels) for shellsim 0.1.19
File
shellsim-0.1.19-cp39-abi3-win_amd64.whl CPython 3.9 abi3 Windows x86-64 Details
shellsim-0.1.19-cp39-abi3-manylinux_2_28_x86_64.whl CPython 3.9 abi3 Linux glibc 2.28+ x86-64 Details
shellsim-0.1.19-cp39-abi3-manylinux_2_28_aarch64.whl CPython 3.9 abi3 Linux glibc 2.28+ ARM64 Details
shellsim-0.1.19-cp39-abi3-macosx_11_0_arm64.whl CPython 3.9 abi3 macOS 11.0+ ARM64 Details
shellsim-0.1.19-cp39-abi3-macosx_10_12_x86_64.whl CPython 3.9 abi3 macOS 10.12+ x86-64 Details

Total release size: 24.3 MB

Release files / shellsim-0.1.19.tar.gz

Download URL shellsim-0.1.19.tar.gz
Size 825.0 kB
Tags Source
SHA-256 checksum
How to use checksums
9e4c1e75e18240b106a61553adffaa8e3a521049c34016c2bbfcf221b115668f
BLAKE2b-256 checksum
How to use checksums
9ac225f93021f8ebac88bd57d39d826939e082e01d6b5f1076d0285fe48e3dbb
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 26, 2026.

Transparency log

Release files / shellsim-0.1.19-cp39-abi3-win_amd64.whl

Download URL shellsim-0.1.19-cp39-abi3-win_amd64.whl
Size 4.9 MB
Tags CPython 3.9 Windows x86-64 abi3
SHA-256 checksum
How to use checksums
9d92ba079a0e16d9e76e480f02cf6ca9332d6f709ca01858692dd726d1d936a7
BLAKE2b-256 checksum
How to use checksums
2e58744189ec7b17a750f754742b54b7a08ef65d94e697ec77cfad2c7936748a
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 26, 2026.

Transparency log

Release files / shellsim-0.1.19-cp39-abi3-manylinux_2_28_x86_64.whl

Download URL shellsim-0.1.19-cp39-abi3-manylinux_2_28_x86_64.whl
Size 5.0 MB
Tags CPython 3.9 Linux glibc 2.28+ x86-64 abi3
SHA-256 checksum
How to use checksums
861e34225394590f5aba0bf643e1bded732a9bca3e18b1d5ef148ac5e7754fae
BLAKE2b-256 checksum
How to use checksums
5fa94a267ffca224f2de12169b83ae2d42ca4341af6cdf2d6de53c7e1a743bcd
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 26, 2026.

Transparency log

Release files / shellsim-0.1.19-cp39-abi3-manylinux_2_28_aarch64.whl

Download URL shellsim-0.1.19-cp39-abi3-manylinux_2_28_aarch64.whl
Size 4.7 MB
Tags CPython 3.9 Linux glibc 2.28+ ARM64 abi3
SHA-256 checksum
How to use checksums
1fd3e4337160858dcc91ff98184a34b585ed1467fe55368169ea59e01c79662d
BLAKE2b-256 checksum
How to use checksums
24252520e36afddd363f552f14f3f8df02c729bf93d76510db0cc9e16bca6c79
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 26, 2026.

Transparency log

Release files / shellsim-0.1.19-cp39-abi3-macosx_11_0_arm64.whl

Download URL shellsim-0.1.19-cp39-abi3-macosx_11_0_arm64.whl
Size 4.2 MB
Tags CPython 3.9 abi3 macOS 11.0+ ARM64
SHA-256 checksum
How to use checksums
78b86e4c72bf30f493f4f73b19fed91675895aaebb49e4aa792bd1156ecb3473
BLAKE2b-256 checksum
How to use checksums
e10dec7203bb4283660fa49305f40a6084e6627e80650a6e7acddafa8eace5b8
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 26, 2026.

Transparency log

Release files / shellsim-0.1.19-cp39-abi3-macosx_10_12_x86_64.whl

Download URL shellsim-0.1.19-cp39-abi3-macosx_10_12_x86_64.whl
Size 4.6 MB
Tags CPython 3.9 abi3 macOS 10.12+ x86-64
SHA-256 checksum
How to use checksums
22565aef655771fa6f062cc2c551d5fc6c8293fe69e30ecb943b609125baad37
BLAKE2b-256 checksum
How to use checksums
17be1f8e371b80b594bd2a0715ff4a583b4eb91b1bd65331c8ec2ed50cf74930
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 26, 2026.

Transparency log

Release history Release notifications | RSS feed

0.1.37

4 release files

0.1.36

4 release files

0.1.35

4 release files

0.1.34

5 release files

0.1.33

5 release files

0.1.22

6 release files

0.1.21

6 release files

0.1.20

6 release files

This release

0.1.19 This release

6 release files

0.1.18

6 release files

0.1.17

6 release files

0.1.16

6 release files

0.1.15

6 release files

0.1.14

6 release files

0.1.13

6 release files

0.1.12

6 release files

0.1.11

6 release files

0.1.10

6 release files

0.1.9

6 release files

0.1.8

6 release files

0.1.7

6 release files

0.1.5

6 release files

0.1.4

6 release files

0.1.3

6 release files

0.1.2

6 release files

0.1.0

6 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page