config-codex-cli

# /config-codex-cli — Codex CLI Configuration Workflow

Configure the Codex CLI on this machine to use an OmniRoute instance as backend.

After this skill completes, `codex` will use OmniRoute by default with `cx/gpt-5.5` (xhigh reasoning), 7 named profiles for quick switching, and proper token limits configured for each model.

---

## Step 0 — Collect required inputs from the user

Before doing anything, ask the user for the two required values. Do not proceed until both are provided:

1. **OmniRoute host** — the IP or hostname of the OmniRoute server (e.g. `192.168.0.1`, `100.x.x.x` for Tailscale, or `localhost`)
2. **OmniRoute API key** — the API key for OmniRoute (starts with `sk-`)

Store these as local variables for the rest of the skill:
- `OMNI_HOST` = the host the user provided (no trailing slash, no port — port 20128 is appended by this skill)
- `OMNI_KEY` = the API key

---

## Step 1 — Detect environment

Run the following to gather machine facts. Store results — they are used in later steps.

```bash
# OS detection
uname -s   # Linux / Darwin (macOS) / MINGW*/CYGWIN*/MSYS* = Windows/Git Bash

# Home directory
echo $HOME           # Linux / macOS / Git Bash
echo $USERPROFILE    # Windows native (PowerShell / cmd)

# Current shell
echo $SHELL          # Linux / macOS: /bin/bash, /bin/zsh, /bin/fish, etc.

# Shell profile file
# Resolve which file to append env vars to:
#   bash  → ~/.bashrc  (Linux) or ~/.bash_profile (macOS)
#   zsh   → ~/.zshrc
#   fish  → ~/.config/fish/config.fish
#   PowerShell (Windows) → $PROFILE (run: echo $PROFILE inside PowerShell)

# PATH and common tool directories (to populate shell_environment_policy later)
echo $PATH
which node 2>/dev/null && node --version
echo ${NVM_DIR:-not-set}
echo ${BUN_INSTALL:-not-set}
echo ${SDKMAN_DIR:-not-set}
echo ${JAVA_HOME:-not-set}
```

Based on the OS result, set the **Codex config directory**:
- Linux / macOS / Git Bash: `~/.codex/`
- Windows native (PowerShell): `$env:USERPROFILE\.codex\`

---

## Step 2 — Verify Codex CLI is installed

```bash
codex --version
```

If this fails, stop and tell the user to install the Codex CLI first:

```bash
npm install -g @openai/codex
```

Then re-run the skill from Step 1.

---

## Step 3 — Create the Codex config directory

```bash
mkdir -p ~/.codex   # Linux / macOS
# Windows PowerShell: New-Item -ItemType Directory -Force "$env:USERPROFILE\.codex"
```

---

## Step 4 — Write `~/.codex/config.toml`

Read the existing file first if it exists (to avoid overwriting MCP server entries, skills, projects, or notify configurations the user may already have).

If the file **does not exist**: create it with the full content below.

If the file **already exists**: apply only the fields listed in the "Model/Inference", "Behaviour", "Auth/Credentials", "Features", "TUI", "Notice", and "Model providers" sections — do **not** remove existing `[mcp_servers.*]`, `[projects.*]`, `[[skills.config]]`, or `notify` entries.

Replace `<OMNI_HOST>` with the value collected in Step 0.

**Content to write (or merge):**

```toml
# ── Model / Inference ─────────────────────────────────────────────────────────
model                    = "cx/gpt-5.5"
model_provider           = "omniroute"
model_reasoning_effort   = "xhigh"
model_reasoning_summary  = "detailed"
model_verbosity          = "high"
model_context_window           = 400000
model_auto_compact_token_limit = 350000
model_max_output_tokens        = 65536
tool_output_token_limit        = 32768

# ── Behaviour ─────────────────────────────────────────────────────────────────
approval_policy          = "never"
sandbox_mode             = "danger-full-access"
personality              = "pragmatic"
web_search               = "live"
check_for_update_on_startup = true

# ── Auth / Credentials ────────────────────────────────────────────────────────
cli_auth_credentials_store   = "file"
mcp_oauth_credentials_store  = "file"

# ── Features ──────────────────────────────────────────────────────────────────
[features]
shell_snapshot = true
unified_exec   = true
multi_agent    = true
memories       = true
js_repl        = true
apps           = false
terminal_resize_reflow = true

# ── TUI ───────────────────────────────────────────────────────────────────────
[tui]
theme       = "dracula"
status_line = ["model-with-reasoning", "current-dir", "context-remaining", "context-used", "five-hour-limit"]

[tui.model_availability_nux]
"gpt-5.5" = 4

# ── Shell environment passed into sandboxed commands ──────────────────────────
# Populate [shell_environment_policy.set] with the PATH and tool dirs discovered
# in Step 1. Only include paths that actually exist on this machine.
# Minimum required: SHELL.
[shell_environment_policy]
inherit               = "all"
experimental_use_profile = true

[shell_environment_policy.set]
SHELL = "<detected shell binary, e.g. /bin/bash or /bin/zsh>"
# Add any of the following that exist on this machine:
# PATH       = "<full PATH from Step 1>"
# NVM_DIR    = "<NVM_DIR from Step 1>"
# BUN_INSTALL = "<BUN_INSTALL from Step 1>"
# SDKMAN_DIR = "<SDKMAN_DIR from Step 1>"
# JAVA_HOME  = "<JAVA_HOME from Step 1>"

# ── Notice / UI flags ─────────────────────────────────────────────────────────
[notice]
hide_full_access_warning   = true
hide_rate_limit_model_nudge = true
fast_default_opt_out       = true

# ── Model providers ───────────────────────────────────────────────────────────
# env_key = NAME of the environment variable (not the value).
# The actual key is stored in the shell profile (Step 5), never here.
[model_providers.omniroute]
name                 = "OmniRoute"
base_url             = "http://<OMNI_HOST>:20128/v1"
env_key              = "OMNIROUTE_API_KEY"
requires_openai_auth = false
wire_api             = "responses"
```

> **TOML rule:** `[[skills.config]]` array-of-tables must be the **last** section in the file. If the file already has `[[skills.config]]` entries, keep them at the end after inserting the new provider bl