dsh-native-codex-oauth

Running OpenAI Codex models through DeepSeek Harness often trips up at the login step: you either paste an API key into a config file or shuffle a third-party OAuth shim in the browser. dsh-native-codex-oauth takes the direct route and wires the native ChatGPT OAuth flow straight into Harness's built-in Web Models page. Click "Sign in with ChatGPT", finish authorization, save the provider — no tokens cached in browser state, no auth.json to hunt down.

The plugin is deliberately thin. Harness already owns the heavy lifting: llm-pi-ai manages the model catalog and request lifecycle, credentials anchors grants on the host, and authorization defines the one-attempt-per-key login cycle. What the plugin contributes is the connective tissue between those native seams and the Settings → Models provider-card slot. It does not spawn a second Codex adapter, does not duplicate credential files, and keeps the browser's view limited to flow labels, prompts, and value-free status. Sign-out deletes the credential record through the host service. Prebuilt artifacts ship in the repo and a Git install triggers no build script, so deployment is friction-free.

Best suited to developers working on the DeepSeek Harness 0.1.2-alpha.3 source line who want Codex models without bypassing the native OAuth path. The plugin depends on the models.provider-card slot and the llm-pi-ai authorization flow introduced in that line; published 0.1.0-rc.x releases lacking either capability are not supported.