What does Codex --full-auto actually do? Approvals and sandbox explained
--full-auto is shorthand for on-request approvals plus a workspace-write sandbox, and it is now deprecated. What runs unattended, what stays blocked, and what to use instead.
Contents
codex --full-auto reads like "do everything automatically", but what it actually grants is freedom inside the workspace. Writing outside the workspace and reaching the network remain restricted by the sandbox.
One thing to know up front: the documentation now marks --full-auto as deprecated and points to codex exec --sandbox workspace-write instead. The flag still works, and knowing what it maps to is what lets you replace it.
KEY POINT
What you will learn
- The exact approval policy and sandbox mode
--full-autois shorthand for - What runs unattended and what stays restricted
- How to pin the same behavior in
config.toml, and what to use now
What it maps to
--full-auto is shorthand for two settings:
approval_policy = "on-request"
sandbox_mode = "workspace-write"
| Setting | Value | Meaning |
|---|---|---|
| Sandbox | workspace-write | Read files, edit within the workspace, run routine local commands inside that boundary. Network access disabled by default |
| Approval policy | on-request | The default: approval is required for sandbox escalations and network access |
The other sandbox modes are read-only, where Codex can read files and run commands inside a read-only sandbox, and danger-full-access, which removes all sandbox restrictions and approval requirements and is not recommended for untrusted code.
Approval policies are on-failure (approval only when operations fail), on-request (the default), and never (no approval prompts). The older untrusted value has been retired and now causes startup failures.
用語解説
Sandbox versus approvals: approvals decide whether Codex asks you before running something; the sandbox decides what an action can reach once it runs. --full-auto reduces how often you are asked, but the sandbox still sets the boundary. See Codex CLI approval modes versus sandbox modes.
What runs and what stops
| Operation | Under --full-auto |
|---|---|
| Reading and editing files in the workspace | Runs |
| Running commands in the workspace (tests, builds) | Runs |
Writing outside the workspace (~/.ssh, for example) | Sandbox refuses |
Network access (npm install, curl) | Disabled by default; escalation requires approval |
| Git commits (inside the workspace) | Runs |
git push | Needs the network, so it triggers an approval request |
When npm install fails for lack of network, approving the escalation lets it retry outside the sandbox. If answering that every time gets tedious, allow network access in config instead.
Pinning it in config.toml
approval_policy = "on-request"
sandbox_mode = "workspace-write"
[sandbox_workspace_write]
network_access = true
This gives the same behavior without the flag, with network access enabled. Enabling the network also makes outbound data transfer possible, so keep it to repositories you trust.
Where it fits
| Situation | Verdict |
|---|---|
| Refactoring a repository with a solid test suite | Good fit; it can iterate until tests pass |
| Bulk-fixing lint errors | Good fit |
| Exploring an unfamiliar codebase | --sandbox read-only is safer |
| An environment holding production credentials | Avoid; the sandbox does not restrict reads |
| Fully automated CI runs | Consider -a never in a disposable environment |
Reads are not restricted
workspace-write restricts writes, but reading files inside the workspace is unrestricted. Keep .env and other secrets outside the workspace. See Stop AI coding tools from reading your API keys and secrets.
Switching mid-session
Running /approvals during a session manages approval settings interactively, so you can adjust the policy without restarting Codex. It suits starting in a read-only investigation and moving to implementation partway through.
For the config file layout in general, see The Codex config.toml file.
Summary
--full-autois shorthand forapproval_policy = "on-request"andsandbox_mode = "workspace-write"- The documentation marks it deprecated and points to
codex exec --sandbox workspace-write - Network access is off by default under
workspace-write; enable it withnetwork_access = true - The sandbox does not restrict reads, so secrets belong outside the workspace
- Use
read-onlyfor investigation, and--full-autosemantics for repositories with tests you trust
FAQ
- Does --full-auto mean everything is automatic?
- No. It makes reads and edits inside the workspace automatic, while writes outside the workspace and network access stay restricted by the sandbox.
- Is --full-auto still the recommended flag?
- No. The documentation marks it as deprecated and points to codex exec --sandbox workspace-write instead. The flag still works and is equivalent to approval_policy on-request plus sandbox_mode workspace-write.
- How is it different from --dangerously-bypass-approvals-and-sandbox?
- That flag, aliased --yolo, removes all sandbox enforcement and approval requirements. --full-auto keeps the sandbox on, so the blast radius stays bounded.
Primary sources
This article was drafted by AI from official documentation and reviewed by the site operator before publishing. Found a mistake? Let us know via the contact page.