Skip to main content

Sandbox container tool code for inspect_ai

Project description

Design

diagram

Inspect calls into the sandboxed image are done statelessly via docker exec inspect-tool-support.

Some tools can be implemented without the need for any in-process state. For those tools, the tool code will be executed within the inspect-tool-support process.

For tools that require the maintenance of state over the lifetime of and sandbox, this image marshals tool calls into a long running process via JSON RPC to an http server process. That server then dispatches tool calls to tool specific @method handlers.

Stateful Tool Design Pattern

Each stateful tool should have its own subdirectory that contains the following files:

  • json_rpc_methods.py

    This module contains all of the JSON RPC @method functions — one for each tool (e.g. the web browser tool is actually a set of distinct tools). It is responsible for unpacking the JSON RPC request and forwarding the call to a transport-agnostic, strongly typed, stateful controller.

  • tool_types.py

    This module includes the pydantic models representing the types for tool call parameters and results.

  • controller.py

    This is transport-agnostic, strongly typed code that manages the tool specific in-process state and performs requested commands.

Release Process

Overview

The release process for this project separates code commits from package publishing. While developers can make frequent commits to the repository, package releases are performed on a different tempo, typically when enough meaningful changes have accumulated. This approach allows for continuous development while providing stable, well-documented releases.

Pending changelog items in the unreleased_changes/ directory serve two key purposes:

  1. They document all changes that will be included in the next release
  2. They determine what type of semantic version bump (major, minor, or patch) will be needed when publishing these pending changes

This system ensures that version numbers accurately reflect the nature of changes between releases according to semantic versioning principles.

Documenting Changes

All changes should be documented using towncrier. When making changes to the codebase, developers should create pending changelog items that will be used to update the changelog at release time:

  1. Use towncrier create to create a new pending changelog item:

    This will interactively prompt you for:

    • The (optional) related issue
    • The type of semantic version change (major, minor, or patch)
    • A description of your change (supporting markdown)

    Alternatively, all options can be provided directly on the command line:

    towncrier create <issue-number>.[major|minor|patch].md
    

    For more details on towncrier's command line options, refer to the towncrier documentation.

  2. Pending changelog items are stored in the unreleased_changes/ directory and accumulate until the next release.

Creating a Release

When it's time to make a release:

  1. Ensure all changes are committed.

  2. Run the make-release-commit script:

  3. The script automatically:

    • Determines the version bump type (major, minor, or patch) based on the pending changelog items
    • Runs towncrier build to incorporate all pending changelog items into CHANGELOG.md
    • Updates the version in pyproject.toml using bump2version
    • Commits the changes into a commit with a message like Bump inspect-tool-support version: 1.0.2 → 1.1.0
    • Tags the commit with the new version number inspect-tool-support-1.1.0

All changelog items are consumed during the release process and converted into entries in the CHANGELOG.md file. After the release, the unreleased_changes/ directory will be empty, ready to collect changes for the next release cycle.

Testing

When running pytest with inspect to test interactions with this package, you may wish to test your local version of the inspect_tool_support code instead of the latest published package. Passing the flag --local-inspect-tools to pytest when running tests from test_inspect_tool_support.py will build and install the package from source, for example:

pytest tests/tools/test_inspect_tool_support.py --runslow --local-inspect-tools

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

inspect_tool_support-1.2.0.tar.gz (94.6 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

inspect_tool_support-1.2.0-py3-none-any.whl (85.8 kB view details)

Uploaded Python 3

File details

Details for the file inspect_tool_support-1.2.0.tar.gz.

File metadata

  • Download URL: inspect_tool_support-1.2.0.tar.gz
  • Upload date:
  • Size: 94.6 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.13.1

File hashes

Hashes for inspect_tool_support-1.2.0.tar.gz
Algorithm Hash digest
SHA256 0173e4d7fd8d5b7042819dc0e017765a48de989177953f7784048cf18a099877
MD5 379a99266f323ae7aefd6db43460958f
BLAKE2b-256 2c1301a69d5e84e80a869fd4b8f9a7ec5f3f7fdf99e086e56e6ee94b75d8d2af

See more details on using hashes here.

File details

Details for the file inspect_tool_support-1.2.0-py3-none-any.whl.

File metadata

File hashes

Hashes for inspect_tool_support-1.2.0-py3-none-any.whl
Algorithm Hash digest
SHA256 be4ed407a38fb5b8f321f3c201d8b239fb6b518f46463b3b86cb1d47056ce717
MD5 7f44feac98c6e9a9a161ec9077ddb0dc
BLAKE2b-256 4bcfce3fc4650d58368317061a0aee717cf74d29a16bb556ed21c8ffdccf59e0

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page