OpenAI Codex CLI

The OpenAI Codex CLI harness (codex) runs Codex non-interactively with codex exec --json. Jimothy parses its JSONL event stream, so you see commands, file changes and reasoning live, and token usage is recorded per step. The built-in Feature template uses Codex for its cross-model review step.

Install and sign in

npm i -g @openai/codex
codex login        # or set OPENAI_API_KEY
codex --version

Then click Re-detect on the Harnesses page.

Default configuration

Field	Default
Command	codex
Arguments	exec --json --full-auto --skip-git-repo-check {{options}} -
Model arguments	--model {{model}}
System prompt arguments	none (the system prompt is prepended to the prompt)
Subagent arguments	none (subagents are ignored, with a note in the log)
Prompt delivery	stdin
Output format	Codex exec –json (codex-json)
Version check args	--version
Model list source	Command: debug models
Models	gpt-6-astra, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5
Default model	gpt-6-sol

The {{options}} line marks where model arguments go, so with a model selected a step runs:

codex exec --json --full-auto --skip-git-repo-check --model gpt-6-sol - < prompt

The trailing - tells Codex to read the prompt from stdin. --skip-git-repo-check lets Codex run in workspaces that aren’t git repositories (the Empty folder and Existing folder workspace modes).

What you see in the logs

Codex event	Log line
thread.started	thread <id> (stored as the step’s session id)
item.started for a command	$ <command>
item.completed command	exit <code> plus the command output (clipped at 1,500 chars); non-zero exits are shown as errors
agent message	assistant text
reasoning	thinking
file change	edit add src/a.ts, update src/b.ts
MCP tool call	mcp <server>.<tool>
web search	web search: <query>
turn.completed	turn completed · N in / M out tokens
turn.failed / error	error, and the step fails

The step Output is the last agent message.

Tokens, not cost

Codex reports tokens (input, output, cached input) per turn, and Jimothy sums them for the step. It does not report a dollar cost, so Codex steps add $0 to the run’s cost figures.

Models

Click Models on the Codex card (or Fetch latest in the editor) to run codex debug models and replace the model list with what your installed Codex offers. Entries marked hidden in that catalog are skipped, and the rest are sorted by Codex’s priority.

Check the model on the Feature template's review step

The Feature: Plan → Build → Review → PR template sets its Cross-model review step to the model gpt-5-codex, which isn’t in Codex’s default model list. If your Codex doesn’t accept it, the review step fails. Open the step and pick a model from the list, or choose Default.

Using Codex as a reviewer

A typical review step prompt asks for a verdict line, and the step’s Pass if output matches (regex) is VERDICT:\s*APPROVE with On failure, loop back to the implement step. Codex’s final message becomes {{steps.review.output}}, which the implementer’s prompt can include on the next pass. See Loops & verdicts.