Install for Codex

Codex support builds on top of Stride's shared .claude/ installation. Stride first installs its skills, commands and hooks there, then creates the extra .codex/ files that let Codex use them.

This means installing Stride for Codex also leaves the project ready for Claude Code. It does not install either CLI itself — install Claude Code or Codex separately before using it.

Follow the Install page for the shared prerequisites, Linear connection and Vision step. This page covers only what Codex adds.

Install with the Codex target

npx github:webventurer/stride --agent codex

Without --agent, Stride sets up Claude Code. With --agent codex, it keeps those shared, Claude-ready files and also writes the .codex/ files. They are built from .claude/, ignored by git and replaced on each install — never edit them by hand.

Trust the hooks, or they will not run

Codex quarantines every hook until you approve it. Open Codex in the project and run /hooks to review and trust stride's two hooks.

Codex records that approval against each hook's exact content. A stride update changes the content, which changes the hash, which quarantines the hooks again — so re-approve after every install. Skip this and the bare-commit block silently stops firing, with nothing in the output to say so.

Check the board before you start

The Linear steps on the Install page still apply. Codex reaches stride's commands as skills rather than slash commands, so they carry a $ and a hyphen — Running the commands covers the form in full:

$linear-check
$linear-setup

$linear-check confirms each API key authenticates. $linear-setup provisions the board with the exact states stride drives work through, then run $linear-check once more to confirm the columns landed. Do this first: without the right columns, the state transitions silently no-op and the card never moves.

Configure .stride.json

After $linear-setup, read .stride.json settings. Stride creates this shared project settings file with safe defaults, but you need to check and configure its Linear project, API key variable, output style, and attended or unattended behaviour.

Running the commands

Codex has no namespaced slash commands, so /linear:start is reachable as a skill instead:

$linear-start WB-123

Each command becomes a skill named linear-<command>, and the four shipped skills — commit, craft, vision, clear-speak — keep their names. Type $ in Codex to see them, or /skills to browse.

The generated skill points at the command file under .claude/commands/linear/ rather than copying it, so there is one place a command is authored and the two tools cannot drift.

Skip the approval prompts

Codex asks before it runs commands and edits files. To turn those prompts off, edit ~/.codex/config.toml:

default_permissions = ":danger-full-access"
approval_policy = "never"

This is Codex's equivalent of --dangerously-skip-permissions, and it is machine-wide. ~/.codex/config.toml is your user config, not the project's, so the setting applies to every repository you open in Codex — including ones you did not install stride into. The agent will run commands and write files without asking, so only set it on a machine where you are happy with that.

Stride never writes this file. Nothing in the install footprint lives outside your repo, and a setting that widens what an agent may do is yours to make deliberately.

What gets installed

.codex/
├── skills/              # stride's skills + one skill per Linear command
└── hooks.json           # registers the same hook scripts with Codex

Both hooks are the same scripts under .claude/hooks/ — hooks.json only wires them up.

Known limitations

An administrator can switch the guardrails off

Codex honours allow_managed_hooks_only = true in a managed requirements.toml. When that is set, Codex ignores user, project and session hook config entirely — including stride's .codex/hooks.json.

What this means: in a managed environment the bare-commit block may never run, with no error and nothing in the output to say so. If commits are landing without the multi-pass workflow, check that setting before assuming the hook is broken.

The Linear MCP block is Claude-only

Stride blocks the mcp__claude_ai_Linear__* tools so Linear writes go through linear_cli.py, which is the path every command is written against. That block is a Claude Code permission, and Codex has no matching primitive.

What this means: under Codex nothing stops an agent reaching Linear through an MCP server instead of the CLI. The commands still call linear_cli.py, so the common path is unaffected — but the guardrail is documentation rather than enforcement there.

Uninstall

The uninstall command removes the Codex footprint along with everything else — .codex/skills/, .codex/hooks.json, and the gitignore entries that cover them.