Codex
How LinkCode runs Codex, including in-app sign-in, its three approval policies, per-model reasoning effort, shell passthrough, and shared thread history.
Codex is the second most complete of the five supported agents: real approval prompts, a curated model list with per-model reasoning effort, $ shell passthrough, and history shared with the codex CLI.
Set it up
LinkCode doesn't ship the Codex CLI. The first time you pick it, LinkCode either finds an install already on your machine or offers to Download the version this release was tested against. See agents.
Sign in
- Sign in from inside LinkCode. The Sign in to Codex card opens a ChatGPT sign-in page in your browser and completes on its own when you finish there.
- Reuse an existing ChatGPT/Codex login on this machine.
- Connect an account under Settings → Providers — a ChatGPT subscription, an OpenAI API key, or a gateway/custom endpoint.
If your credentials expire while a thread is running, LinkCode reports a sign-in failure rather than a generic error, and picks the thread back up after you sign in again.
Approval policy
Three policies:
| Policy | What it does |
|---|---|
| Ask permissions | Asks before running commands outside Codex's own safe list. |
| Sandboxed auto | Works autonomously inside the workspace sandbox; asks only to escalate. |
| Full access | No sandbox and no approval prompts in this workspace. |
Sandboxed auto is the default. Switches apply from the next turn, not mid-turn.
When you answer a Codex permission request, you can reject it and say what you'd rather it do — that reply goes back to Codex as steering, not just a refusal.
Your own config.toml is respected
Until you explicitly pick a policy in a thread, LinkCode leaves your ~/.codex/config.toml sandbox setting alone — a stricter read-only setup of your own stays in force. Picking a policy applies it exactly, config.toml notwithstanding. (Codex's own approval_policy key is not read yet; the sandbox setting is.)
Model and reasoning effort
Codex has a curated model list — GPT-5.6-Sol, GPT-5.6-Terra, GPT-5.6-Luna, GPT-5.5, GPT-5.4, and GPT-5.4-Mini — and reasoning-effort levels that depend on the model you pick: Low, Medium, High, and xHigh everywhere, plus Max on the 5.6 models and Ultra on Sol and Terra. LinkCode reads that catalog from Codex itself, so a model your account exposes with different levels is reflected as it is.
Model and effort changes both apply from the next turn.
Commands and shell
Start a message with / to run Codex's own commands — /compact plus whichever skills you have enabled. Start it with $ to run a shell command through Codex; the output comes back as a normal turn in the transcript. A $ command deliberately bypasses the thread's sandbox and approval policy: it's your command, not the agent's.
Thread history
Resume is fully supported: stopping and later resuming a Codex thread works. History is shared with the standalone codex CLI — a thread started in either place reads back from the same on-disk store. File changes come through as real diffs, and compaction is shown in the transcript rather than silently dropping earlier context.