Checked against official documentation on May 24, 2026, this post focuses on the setup and failure points behind Codex config.toml approval_policy and sandbox_mode Setup Guide. The practical baseline is: Start with interactive approvals and workspace-limited writes, then create narrower profiles only after tests and review rules are stable.
Quick Answer
Start with interactive approvals and workspace-limited writes, then create narrower profiles only after tests and review rules are stable.
The practical rule is simple: keep the agentโs authority narrower than the task. Installation, settings, MCP tools, and repository instructions should be treated as part of the engineering system, not as one-time setup trivia.
When This Setup Matters
~/.codex/config.toml approval and sandbox settings decide when Codex pauses and what commands or files it can touch. This matters when a developer wants repeatable AI-agent help instead of a long ad hoc prompt. A good setup makes three things visible: what the agent may read, what it may change, and how the human will verify the result.
If the tool is being introduced to a team, write the decision down before broad use. Name the account or authentication path, the directory where the agent should start, the files it must not touch, and the command that proves a change is acceptable.
Baseline Commands
mkdir -p ~/.codex
codex --profile safe-edit "Explain the current repository structure."
Run commands from the same shell and project root where the agent will work. If a command succeeds in one terminal but fails in another, fix the shell, PATH, account, or working-directory issue before asking the agent to edit code.
Configuration Pattern
model = "gpt-5.5-codex"
approval_policy = "on-request"
sandbox_mode = "workspace-write"
[profiles.safe-edit]
approval_policy = "on-request"
sandbox_mode = "workspace-write"
Treat this block as a starting pattern, not a universal default. A personal laptop, a locked-down company workstation, and a CI job should not have the same permission model. Prefer read-only or planning modes until the repositoryโs tests and rollback path are clear.
For recurring use, keep a short setup note beside the repository. Include the CLI version, the selected permission mode, the instruction file that loaded, and the exact command used for the final verification. That note becomes the baseline when a teammate reports different behavior.
Verification Checklist
- Confirm writes stay inside the repo.
- Require approval for network or privileged commands.
- Record the profile used in pr notes.
After the setup works, ask the agent for a read-only summary first. Then ask for a narrow plan. Only after those two responses match the repository reality should you allow edits or tool calls that can change files.
Common Mistakes
- Using full access for exploratory prompts.
- Forgetting which profile is active.
- Treating sandboxing as a substitute for review.
The costly mistake is usually not a bad model answer; it is an unclear operating boundary. If authentication, MCP scope, settings precedence, or instruction files are ambiguous, the session can appear productive while quietly moving risk into code review.
FAQ
Should this be configured globally or per project?
Put personal preferences globally, but put repository rules in project files so every teammate and future session sees the same constraints. Secrets, local paths, and experiments should stay out of committed project files.
When should I allow the agent to edit files?
Allow edits only after the agent can restate the task, name the files it expects to touch, and identify the verification command. For unfamiliar repositories, start in planning or read-only mode.
What should I record after the setup works?
Record the install method, version check, account or API-key policy, permission mode, instruction file location, MCP scope, and the first verification command. This gives the next session a reproducible baseline.
Leave a comment