SmartCodex

SmartCodex v1 routes a task to a Codex model using local rules or an optional AI classifier, then starts Codex CLI. It requires Python 3.10+ and Codex CLI available on PATH. The Python package has no runtime dependencies and does not require an OpenAI API key.

Install from this repository

From the repository root, install the package into the active Python environment:

py -m pip install .
smartcodex --help

For development, create and activate a virtual environment first:

py -m venv .venv
.\.venv\Scripts\Activate.ps1
py -m pip install .

PowerShell usage

Quote a prompt containing spaces. --router rules uses only local routing; smart is the default and invokes the AI classifier only when local rules are uncertain. --router ai always requests AI classification unless --dry-run is set.

smartcodex "Ubah teks tombol Simpan menjadi Kirim"
smartcodex --router rules "Tambahkan validasi form"
smartcodex --router ai "Audit integrasi ini"
smartcodex --mode exec "Periksa konfigurasi proyek"

Preview a local routing decision without starting Codex or an AI classifier:

smartcodex --dry-run "Ubah warna tombol login"

In exec and loop mode, Codex uses the read-only sandbox by default. Add --write to allow workspace edits. Chat mode retains the v1 workspace-write sandbox and on-request approval behavior. Legacy --force-tier values retain their original model mapping, including sol-low to Sol Low. Astra routing requires the explicit --allow-astra option.

v1 limitations

• Routing sees only the prompt; it does not inspect the target repository.
• The interactive chat routes only its initial prompt. Follow-up messages remain in that Codex session and are not routed again.
• Loop mode starts a separate codex exec session for each prompt and does not preserve chat history.
• Model availability depends on the Codex account and CLI. AI routing uses Luna Low as the classifier and falls back to local rules on failure.
• This package does not include project analysis, planning, task scheduling, session management, or per-subtask routing.

The project testing configuration uses GPT-6 Luna Low as the default model. Manually selected legacy tiers continue to map to their v1 models.