[abhishekgahlot2]

• deepgym.agents: ClaudeCodeAgent, CodexAgent (CLI inside the sandbox as the unprivileged agent), AnthropicComputerAgent, OpenAIComputerAgent (provider computer-use loops over Session.step).
• Shared turn, wall-time and USD budgets with usage/cost provenance; one cli action per CLI run with nested events, recorded in traces (JSONL, ATIF, Parquet).
• CLI egress through a root-owned CONNECT proxy limited to the provider host, enabled per environment with cli_provider_egress; the agent stays in its loopback-only namespace and local app traffic bypasses the proxy. PID-namespace containment so every agent process stops before grading.
• MultiTurnEnvironment.root_service_start for images whose services start as root (rejected for inline API specs, as is cli_provider_egress).
• Pinned CLI versions in one place (agents/versions.py, cli_install_command()), checked inside the sandbox before launch.

Testing: full suite 2829 passed; the 9 failures also fail on main. Live on Daytona: computer-use agent on a desktop snapshot (iTop) passed before and after the containment change; Claude Code CLI on the same app passed (score 1.0, 12 s). Codex CLI live run pending.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}