Mining session logs to build sandbox allowlists for coding agents

On 2026-10-31, DEV Community's AI section published a piece titled "Sandboxing Coding Agents: Generating Allowlists from Session Logs." The article notes that Claude Code's /sandbox and Codex's workspace-write both follow a deny-by-default design: refuse everything first, then allow whatever you list. The same mechanism appears in Anthropic's open-source sandbox-runtime and in nono, among others. Because the allowlist starts out empty, users cannot know up front which directories, websites, and commands their work actually requires. The author argues this is precisely why most people switch the sandbox off within an hour of turning it on: a deny-by-default sandbox poses a question no one can answer before the task begins, so the practical route is to turn it on, observe what breaks, and adjust from there. As an alternative, the author suggests reading session records stored under ~/.claude/projects or ~/.codex/sessions, extracting the directories, domains, and commands actually needed, and writing those into the allowlist as rules. For now this remains a single published proposal; whether any agent vendor has adopted the approach is not addressed in the available coverage.