release-sandbox-tools

name	release-sandbox-tools
description	Build, validate, and publish a new version of the inspect-sandbox-tools injectable executables. Use when code under src/inspect_sandbox_tools/ changed and a new injectable version must ship, or when asked to build/validate/upload sandbox tools binaries.

Releasing sandbox tools injectables

Builds the four injectable artifacts (amd64/arm64 × glibc/musl), validates them across Linux distros, and publishes them to S3. Background: src/inspect_sandbox_tools/design/RELEASING.md.

All commands run from the repo root using the repo venv (.venv/bin/python — do not use uv run or system python).

1. Preconditions

• Docker must be running with multi-arch (buildx) support: docker info succeeds.

• The version must be bumped. It's a plain integer in src/inspect_ai/tool/_sandbox_tools_utils/sandbox_tools_version.txt and must be greater than main's:

  git show origin/main:src/inspect_ai/tool/_sandbox_tools_utils/sandbox_tools_version.txt
  

  If not yet bumped, increment it by 1 (this is a committed source change — it rides in the same PR as the sandbox tools code change).

Call the bumped value {V} below.

2. Build

.venv/bin/python src/inspect_ai/tool/_sandbox_tools_utils/build_within_container.py --all --dev=false

Slow (Docker builds for 4 variants; can exceed 10 minutes) — run it in the background and monitor. On success, verify all four artifacts exist in src/inspect_ai/binaries/:

• inspect-sandbox-tools-amd64-v{V}
• inspect-sandbox-tools-arm64-v{V}
• inspect-sandbox-tools-amd64-musl-v{V}
• inspect-sandbox-tools-arm64-musl-v{V}

(These are gitignored — never commit them.)

3. Validate across distros

.venv/bin/python -m inspect_ai.tool._sandbox_tools_utils.validate_distros

Must be run with -m (module mode) for imports to work. It runs each artifact's healthcheck in Docker containers: glibc variants on Ubuntu 16.04–24.04, Debian 11/12, Kali; musl variants on Alpine 3.16/3.18/latest. Also slow — background it. Every distro must pass; any failure blocks the release.

4. Upload to S3 (user runs this — not the agent)

The upload needs the user's AWS credentials (SSO session). Do not run it yourself; instead ask the user to run it in-session so the output lands in the conversation:

Uploading needs your AWS credentials. If your SSO session is stale, run aws sso login in a terminal first. Then type this in the prompt:

! .venv/bin/python src/inspect_ai/tool/_sandbox_tools_utils/upload_to_s3.py {V}

The ! prefix runs the command inside the session. It uploads all four artifacts to s3://inspect-sandbox-tools/ (us-east-2) with --acl public-read.

5. Verify the upload (no credentials needed)

Objects are public-read, so confirm each returns HTTP 200:

for f in amd64 arm64 amd64-musl arm64-musl; do
  curl -sI -o /dev/null -w "%{http_code} inspect-sandbox-tools-$f-v{V}\n" \
    "https://inspect-sandbox-tools.s3.us-east-2.amazonaws.com/inspect-sandbox-tools-$f-v{V}"
done

6. Nothing else to do here

PyPI bundling is not part of this process — the inspect_ai release script (scripts/pypi-release.py) pulls the glibc binaries from S3 automatically at release time. Merging the PR with the version bump completes the release.