You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Docker sandbox: container cleanup on timeout -- timed-out containers are now killed via docker rm -f instead of leaking. Each container gets a unique name (initrunner-{uuid}) and initrunner.managed=true label for identification
Docker sandbox: image pre-pull at startup -- ensure_image_available() checks and pulls the configured image before the first tool call, with logging and clear error messages for private/offline images. Also reported in initrunner doctor --deep
Docker sandbox: --user flag -- new user field ("auto" | "<uid>:<gid>" | null, default "auto") runs containers as the current user when writable mounts exist, preventing root-owned output files
Docker sandbox: --init flag -- tini is now always used to reap zombie processes
Docker sandbox: OOM detection -- exit code 137 now appends a clear message suggesting to increase memory_limit
Docker sandbox: bind mount validation -- missing source paths raise an error at container start instead of Docker silently creating empty root-owned directories
Docker CLI in published image -- the container image now includes the Docker CLI for socket-passthrough sandbox support. Docker Compose file includes a commented socket mount with security warning
Multi-platform Docker image -- CI now builds linux/amd64 and linux/arm64 images via QEMU
Changed
Docker sandbox: script tool no longer leaks host environment -- when Docker is enabled, only script parameter values are passed into the container. env_passthrough in the Docker config handles explicit host vars
Docker sandbox: Python tool always mounts a working directory -- when working_dir is not set, a temp dir is created, mounted at /work, and cleaned up after execution
Example role docker-sandbox.yaml uses model auto-detection -- removed hardcoded provider: openai / name: gpt-4.1-mini
Fixed
16 ty type errors in test_doctor.py and test_a2a.py -- added missing None guards and type-ignore annotations for external SDK types
CI: pnpm version 9 -> 10 -- matches local dev toolchain, fixes ERR_PNPM_LOCKFILE_CONFIG_MISMATCH in release/testpypi/security workflows