wrapper-scionos

wrapper-scionos

Extensible ScioNos CLI wrapper for RouterLab-backed coding assistants.

Current version: 3.2.0.

This release targets Claude Code, Claude Desktop, and Codex CLI without coupling every client integration into one large module.

Lire en français

Requirements

• Node.js 22 or later
• Claude Code installed if you want to launch Claude Code through the wrapper
• A RouterLab token
• Windows, macOS, or the claude-desktop-debian Linux port for Claude Desktop 3P profile configuration

Quick Start

From the project folder:

cd D:\Serveurs\Projet_ScioNos\Wrapper-ScioNos
npm install
node index.js

node index.js opens an interactive menu with:

• Claude Code
• Claude Desktop
• Codex CLI
• Auth
• Doctor

Commands

node index.js
node index.js --service llm
node index.js claude-code --strategy aws -- -p "Summarize this repo"
node index.js auth login
node index.js auth status --service llm
node index.js auth test --service llm
node index.js doctor
node index.js strategies --service routerlab
node index.js claude-desktop status
node index.js claude-desktop apply --service llm --strategy claude --dry-run
node index.js claude-desktop apply --service llm --strategy claude --yes
node index.js claude-desktop apply-proxy --service routerlab --strategy claude-gpt --yes
node index.js claude-desktop proxy --service routerlab
node index.js claude-desktop proxy --service llm
node index.js codex launch --service routerlab
node index.js codex launch --service llm
node index.js codex template --service routerlab
node index.js codex restore --yes

When installed globally, use wrapper-scionos or scionos instead of node index.js.

RouterLab services

• routerlab: https://api.routerlab.ch
• llm: https://llm-api.routerlab.ch

Tokens are service-scoped. New tokens are stored under wrapper-scionos; existing claude-scionos secure token files or keychain entries are also read as a migration fallback. On Linux, the wrapper uses Secret Service through secret-tool when available and falls back to a user-only 0600 token file when secret-tool is missing or unavailable.

Claude Code

Launch through the wrapper's local resilient proxy to RouterLab:

node index.js claude-code --service routerlab --strategy aws

With Claude Code arguments:

node index.js claude-code --strategy aws -- -p "Summarize this project"

RouterLab LLM-specific strategies include:

node index.js claude-code --service llm --strategy glm-5.2
node index.js claude-code --service llm --strategy claude-qwen3.7-max
node index.js claude-code --service llm --strategy claude-MiniMax-M3
node index.js claude-code --service llm --strategy deepseek-v4

In the --service llm menu, the order is Claude, OpenAI GPT, glm-5.2, qwen3.7-max, MiniMax-M3, then deepseek-v4.

claude-MiniMax-M3 is shown as MiniMax-M3 in the guided menu. claude-qwen3.7-max is shown as qwen3.7-max. The strategy uses claude-qwen3.7-max for Opus, Sonnet, and Haiku, with claude-qwen3.6-flash for subagents.

All LLM strategies accept --subagent-model. With Strategy default, claude-MiniMax-M3 keeps claude-MiniMax-M3, claude-qwen3.7-max keeps claude-qwen3.6-flash, glm-5.2 keeps claude-glm-5.2, and deepseek-v4 keeps claude-deepseek-v4-flash.

The wrapper starts a local long-running proxy for Claude Code launches and configures:

• ANTHROPIC_BASE_URL
• ANTHROPIC_AUTH_TOKEN

ANTHROPIC_BASE_URL points to the local proxy, while the proxy forwards to the selected RouterLab service with the stored RouterLab token. This avoids body timeout crashes during long thinking, tool, and subagent pauses.

Claude Code launches also receive temporary child-process environment:

• CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1

Strategy mappings can also configure:

• ANTHROPIC_DEFAULT_OPUS_MODEL
• ANTHROPIC_DEFAULT_SONNET_MODEL
• ANTHROPIC_DEFAULT_HAIKU_MODEL
• CLAUDE_CODE_SUBAGENT_MODEL

Unknown CLI arguments are forwarded to Claude Code.

Claude Desktop

The wrapper supports:

• Windows: %LOCALAPPDATA%\Claude and %LOCALAPPDATA%\Claude-3p
• macOS: ~/Library/Application Support/Claude and ~/Library/Application Support/Claude-3p
• Linux with claude-desktop-debian: ${XDG_CONFIG_HOME:-~/.config}/Claude and ${XDG_CONFIG_HOME:-~/.config}/Claude-3p

Restore official Claude Desktop mode:

node index.js claude-desktop restore-official --yes

Direct profile configuration:

node index.js claude-desktop apply --service routerlab --yes

For strategies that Claude Desktop hides from the model menu, use local proxy mode:

node index.js claude-desktop apply-proxy --service routerlab --strategy claude-gpt --yes
node index.js claude-desktop proxy --service routerlab

The proxy terminal must stay open while Claude Desktop uses mapped models. This mode uses the same long-running proxy infrastructure that Claude Code and Codex use internally.

With claude-desktop apply and no strategy, Claude Desktop reads the model catalog from RouterLab directly. Some non-Claude-family model ids can be hidden by Claude Desktop even when RouterLab returns them.

Local proxy mode exposes Desktop-safe model ids, then forwards requests to the real RouterLab strategy models. The default RouterLab Desktop catalog is ordered as:

claude-opus-4-8
claude-sonnet-4-6
claude-haiku-4-5
aws-claude-opus-4-8
aws-claude-sonnet-4-6
aws-claude-haiku-4-5
gpt-5.5
gpt-5.4
gpt-5.4-mini
kimi-k2.7-code
glm-5.1

For --service llm, the Desktop local mapping mirrors the Claude Code LLM strategies: Claude, OpenAI GPT, GLM, Qwen, MiniMax, and DeepSeek. Display names remove the RouterLab claude- routing prefix where helpful, for example gpt-5.5, deepseek-v4-pro, qwen3.7-max, and glm-5.2.

claude-desktop apply is dry-run by default. Pass --yes to write files. claude-desktop apply-proxy writes a profile pointing to http://127.0.0.1:15721. claude-desktop proxy must stay running while Claude Desktop uses mapped routes.

The 1M context flag is applied per upstream model. Haiku, Kimi, GLM, and GPT mini routes do not get 1M variants, while GPT 5.4 and GPT 5.5 do.

Codex CLI

Launch the official Codex CLI through the wrapper-managed local proxy for the current session without rewriting ~/.codex/config.toml:

node index.js codex launch --service routerlab
node index.js codex launch --service llm

codex launch starts Codex with the RouterLab model catalog for the selected service. For scripted launches, pass --model <value> to choose the initial model; otherwise the service default is used.

For debugging only, --direct bypasses the local proxy and points Codex straight at RouterLab.

The wrapper also includes a Codex CLI config template generator:

node index.js codex template --service routerlab

If you previously used an older wrapper version to persistently rewrite Codex config, you can restore the saved backup:

node index.js codex restore --yes

RouterLab Codex CLI models are offered in this order:

gpt-5.5
gpt-5.4
gpt-5.4-mini
deepseek-v4-pro
deepseek-v4-flash
kimi-k2.7-code
glm-5.1

RouterLab LLM Codex CLI models are offered in this order:

gpt-5.5
gpt-5.4
gpt-5.4-mini
glm-5.2
qwen3.7-max
MiniMax-M3
deepseek-v4-pro

codex launch is non-destructive by default: it starts the official codex binary with runtime -c overrides for model_providers.custom.base_url and passes the selected RouterLab token through a local OPENAI_API_KEY to the child process; the proxy swaps it for the selected RouterLab token upstream. It does not rewrite config.toml and does not touch auth.json. The wrapper writes a temporary RouterLab model catalog under the system temp directory for the duration of the Codex process, then removes it.

The persistent codex apply flow was removed because replacing the user's Codex config.toml can overwrite unrelated Codex settings such as MCP, hooks, features, and sandbox preferences. codex restore --yes remains available only as a recovery command: it restores config.toml.wrapper-scionos-backup when present; if no backup exists, it only removes a config that clearly looks like a wrapper-generated RouterLab config.

Development

npm test
node index.js doctor

See docs/architecture-notes.md for the current architecture notes.