Codex gpt-oss workflows remain brittle

The real issue here is ecosystem drift, not model quality: Codex, gpt-oss, and llama.cpp are each moving fast enough that the integration contract keeps breaking.

• –This reads like a “what works today?” post for people trying to run Codex-style workflows locally with open-weight models
• –The mention of incomplete Responses API support in llama.cpp points to protocol mismatches, not just generic inference bugs
• –The useful answer here is probably a pinned version matrix plus a known-good serving/config recipe, not a conceptual explanation
• –Community guidance in this space tends to age quickly, so current docs and recent issue threads matter more than older setup guides