Codex

Codex talks to devin-2api over the OpenAI Responses surface (POST /v1/responses).

Setup

~/.codex/config.toml:

model_provider = "OpenAI"
model = "swe-2-max"
model_context_window = 262000
model_auto_compact_token_limit = 230000

[model_providers.OpenAI]
base_url = "http://127.0.0.1:8080/v1"
stream_max_retries = 100

The downstream token goes through the OPENAI_API_KEY environment variable.

Two details matter:

• model_context_window / model_auto_compact_token_limit must reflect the real window (262000). A default or mismatched larger value puts the auto-compact threshold beyond the upstream limit, so oversized requests fail instead of compacting first.
• stream_max_retries = 100. codex-rs terminates on HTTP 429 (retry_429 is hardcoded false), so devin-2api converts pre-stream rate limits into response.failed stream events — Codex reads try again in Ns and sleeps until the latch lifts. The default 5 retries covers under a minute of limiting; upstream minute-bucket episodes run longer, so raise it toward the 100 cap.

apply_patch arrives as a type: "custom" tool call and round-trips through the proxy untouched.

WebSocket transport (optional)

devin-2api also serves the Responses API over WebSocket — GET /v1/responses negotiates the upgrade, and multi-turn sessions live on the connection (response.create / response.append with incremental input, expanded into full transcripts server-side). Codex opts in with supports_websockets = true in [model_providers.OpenAI] plus prefer_websockets on the model entry. Plain HTTP POST works fine without this.