GPTDiscord v13.2.2 — E2B code chat

Restores E2B code-chat execution, file handling, and cleanup with the current SDK.

• Shell and package commands return exit code, stdout, and stderr when they fail, allowing the agent to diagnose and correct them.
• All message attachments are uploaded with their exact sandbox paths. Python state and installed packages persist throughout the session.
• Generated CSVs and rich plot results can be downloaded through Download Artifacts. Downloads stream within the server's file-size limit, and ended-session controls give clear guidance.
• Code chats announce readiness after sandbox/agent registration; their opening message explains when to wait.
• Failed or canceled startup and graceful shutdown close active and late-created sandboxes. Close is idempotent, and SDK requests have explicit deadlines.
• OpenAI Responses use bounded output and disable provider storage while retaining local conversation history.

Install pip install '.[code]', configure E2B_API_KEY alongside OpenAI access, and restart the bot. Use /code chat, wait for its ready message, and type end to close it. Operators should allow a 120-second graceful service stop window.

Validation: 922 full-install tests and 638 base-install tests passed, including 29 focused code tests. Ruff, dependency checks, package builds, and Twine checks pass. A real live workflow verified numeric execution, persistent Python state, aggregation of two uploaded CSVs, CSV and rich PNG downloads through the actual button callback, Python package installation/import, nonzero command diagnostics, system package installation, duplicate suppression, End and cleanup. It used nine bounded Responses requests, one sandbox capped at 300 seconds, and eight executions/commands. All owned resources were removed. Human authorship and interaction acknowledgements are adapted by the diagnostic runner; execution tools and results are real.

The same 11 live checks also passed from the installed 13.2.2 package on llm.kaveenk.com, using 9 bounded Responses requests, one 300-second sandbox, and 8 executions/commands. Installed modules matched the merged source; the service stayed healthy, all owned resources were cleaned up, and canonical databases, saved indexes, conversations, permissions, and configuration remained unchanged during testing.

Cursor Bugbot reached its usage limit; independent reviews were completed and their findings were addressed.