Native Codex

Use this route when the codex executable is installed and you want Codex—not Ollama—to own authentication, the conversation, and model selection. JSAT adds its MCP server and global $jsat skill; it does not create project-local Codex files.

Prerequisites and connect

pip install jsat
curl -fsSL https://chatgpt.com/codex/install.sh | sh
codex                         # complete sign-in, then exit
cd /path/to/project
jsat index .
jsat connect codex

The connection writes ~/.codex/config.toml and ~/.codex/skills/jsat/SKILL.md. Restart a Codex process that was already open.

Start and select a model

jsat codex --repo .
# equivalent managed route
jsat start codex --via native --repo .

Codex selects its model using its own UI/configuration. Do not pass an Ollama model to this route. If JSAT's internal LLM-backed tools should also call native Codex, configure that separately:

jsat ai use codex-cli
jsat ai test

Use and verify JSAT

Inside Codex:

$jsat status
$jsat query what is the request path for checkout?
$jsat blast-radius src/payments/service.py

Or ask: Use jsat__get_index_status and tell me the node and edge counts. A successful call proves MCP is active. On the host, jsat connect list confirms the saved connection.

Lifecycle

jsat ps
jsat restart codex
jsat stop codex
jsat resume codex
jsat resume codex --session SESSION_ID

jsat codex resume SESSION_ID is also forwarded directly to Codex. jsat session resume is different: it resumes an interrupted JSAT workflow such as crack or magic.

Troubleshooting and removal

• If $jsat is missing, rerun jsat connect codex, then restart Codex completely.
• If MCP is missing, inspect ~/.codex/config.toml and run jsat doctor.
• If an error shows an Ollama tag with provider: openai, the model/provider pair is misrouted. Select codex_cli for native Codex or ollama for an Ollama tag; do not combine an Ollama model name with the OpenAI provider.
• If Codex is not installed, use an Ollama-launched client or install Codex.

jsat disconnect codex