Permission behavior matrix
A single lookup table for what each agent's permission modes actually mean, since the labels don't line up 1:1 across providers.
| Concept | Claude Code | Codex | Antigravity |
|---|---|---|---|
| Standard mode | default | default | default |
| Auto-accept edits | acceptEdits | workspace-write | accept-edits |
| Plan/read-only | plan | read-only | plan |
| No sandbox, no prompts | bypassPermissions | danger-full-access / bypass | bypass (--dangerously-skip-permissions) |
| Live mid-session switch | Yes | No — fixed at thread creation | No — fixed at launch |
| Mechanism | SDK canUseTool callback | Sandbox tier set at thread creation | --mode launch flag + terminal-output classification |
Warning
This is a behavioral mapping to help you reason across providers, not a promise that these exact flag/mode names never change on the provider side. Verify against Permissions and each provider's own current documentation before relying on an exact name.
Compare Mode and AIgency overrides
Compare Mode forces every pane into the "plan/read-only" row above, per agent — see Workspace behaviour. AIgency layers its own run-level policy (ask / allow-approved / bypass) plus per-agent overrides (inherit / trusted / restricted / bypass) on top of these same underlying modes — see Approvals.