$ the-wire · showcase
Guardian REPL policy now sources from model metadata
By RepoJournal · Filed · About OpenAI
Codex shifted Guardian REPL policy authority to reviewer model metadata, with a bundled fallback and safety checks on reuse.
Codex now sources Guardian REPL policy from the reviewer model's metadata [1], adding `auto_review.node_repl_policy` to model messages for `node_repl` and `cua_repl` reviews. The field falls back to the bundled policy when absent and skips injection when explicitly empty. Effective policy participates in Guardian session reuse checks, and unsafe parent-model fallback transitions that would change it are rejected. Additionally, Guardian user answers survive compaction [2]; recorded answers now match against retained review history, so compacted conversations still pass trusted answers to subsequent reviews. Explicit rollback still removes the associated answer. Two TUI fixes round out the day: agent navigation restores across reconnects [3] without losing selected tasks or drafts, and Left opens the agents overview from an empty, focused composer [4] in local-daemon sessions, leaving the configurable `open_agents` action unbound by default.
Action items
- → Review Guardian session reuse logic for policy-change rejection behavior openai/codex [monitor]
- → Test compaction flows to confirm user answers are retained in reviews openai/codex [monitor]
- → Verify TUI reconnect behavior and the new Left shortcut in local-daemon sessions openai/codex [monitor]
References
- [1] Source Guardian REPL policy from model metadata (#41919) ↗ openai/codex
- [2] Preserve Guardian user answers across compaction (#41852) ↗ openai/codex
- [3] Restore agent navigation after TUI reconnects ↗ openai/codex
- [4] Open the agents overview from an empty composer ↗ openai/codex