Codex & ChatGPT passthrough

Route the Codex CLI through llama-dash to log and inspect ChatGPT subscription requests. Codex keeps ownership of the subscription OAuth bearer; llama-dash preserves it while forwarding requests to the ChatGPT Codex backend.

Client config

Sign in to Codex with your ChatGPT subscription first (codex login). Then add a custom model provider in ~/.codex/config.toml:

model_provider = "llama-dash"

[model_providers.llama-dash]
name = "llama-dash"
base_url = "https://<llama-dash-host>/v1"
wire_api = "responses"
requires_openai_auth = true
# Only necessary for request attribution to a specific llama-dash API key
http_headers = { "x-llama-dash-api-key" = "sk-..." }

• wire_api = "responses" makes Codex call the OpenAI Responses API at /v1/responses.
• requires_openai_auth = true makes Codex attach its own ChatGPT credentials to requests sent to the custom provider.
• http_headers is optional. Set x-llama-dash-api-key to attribute requests to a specific llama-dash API key.

Do not set an OpenAI API key for this flow. Codex supplies its subscription credentials, and llama-dash forwards them unchanged.

Routing rule

In Policies, add a first-match routing rule for the OpenAI Responses API:

• Match endpoint /v1/responses.
• Match the exact model IDs you use in Codex.
• Set action to continue.
• Set auth mode to passthrough and preserve the client Authorization header.
• Set direct target to https://chatgpt.com/backend-api/codex.

Do not leave Requested model empty. Other OpenAI-compatible clients can use /v1/responses; exact model matching limits this rule to the selected Codex models.

The resulting requests remain visible in Requests, including routing metadata and the redacted authorization header.