openai Codex Responses: remap model='auto' to a concrete model (TAURI-RUST-AHX)

[M3gA-Mind]

Source

Sentry: https://sentry.tinyhumans.ai/organizations/tinyhumans/issues/11825/ (TAURI-RUST-AHX)
Events: 1,246 · Users affected: 1 · First seen: 2026-06-06 · Last seen: 2026-06-26 (actively firing on 0.57.53)

Symptom

openai Responses API error: {"detail":"The 'auto' model is not supported when using Codex with a ChatGPT account."} (tags: model=auto, provider=openai, operation=responses_api, status=400).

Root cause

The user's config carries the model string auto (a Codex-CLI alias). In chat_via_responses, openhuman builds the ResponsesRequest with model: model.to_string() (compatible_helpers.rs:70) and sends it unchanged to the ChatGPT-OAuth Codex /responses endpoint, which rejects auto (valid only for API-key Codex) with a 400. The is_codex_oauth_responses block already proactively adjusts two sibling Codex quirks — strips stream:false (#3201) and max_output_tokens (#4068 / TAURI-RUST-EWD) — but performs no auto→concrete-model remap, so the sentinel leaks to the provider. Emit site: compatible_helpers.rs:135.

Bucket / RCA

real-defect (model resolution). RCA fix: resolve/remap auto to a concrete model before the request so the provider never sees auto — within the is_codex_oauth_responses branch (mirroring the #3201 / EWD proactive-adjustment pattern at ~lines 40-78), substitute a concrete Codex-class model when model == "auto". Optionally normalize an auto default_model at config/provider-factory load (cf. #4080's stale-default_model normalization).

Dedupe

NEW (distinct from the #4009 capability-gating 404 family — openai supports Responses; this is a 400 on the model field). Closest siblings #4068 / #4080 (same block/pattern) don't remap auto. No open PR / GH issue. Not covered by #4182-4189.

Suggested scope

• src/openhuman/inference/provider/compatible_helpers.rs — in chat_via_responses is_codex_oauth_responses block, remap model=="auto" before building ResponsesRequest (follow Codex OAuth Responses API requests fail unless store=false is sent #3201 / EWD pattern ~40-78).
• Optional: config/schema/cloud_providers.rs / provider-factory normalization of an auto default_model.
• Test asserting auto is rewritten for the Codex OAuth responses endpoint.

No AI-attribution trailers in commits/PR. Do NOT run local cargo — push --no-verify and let PR CI test.