fix: provide local import allowlist before code execution

[EDDIWARD]

Summary

Expose the local Python import allowlist in the agent context before code execution. This lets the model choose an available approach or report the restriction instead of probing unavailable packages.

Root cause

In local execution, the interpreter enforces an import allowlist. The original prompt did not describe that boundary, so some models learned it only from failed executions and retried with different unavailable packages.

Before

The reproduction task triggered attempts to import requests and urllib.request, both rejected by the local interpreter.

After

The agent receives the permitted imports before planning. The same task no longer runs a restricted import and reports the execution boundary directly.

Validation

• Focused backend tests: 20 passed.
• Manual reproduction with the same network-request prompt: no restricted import attempt.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!