The Wire · Showcase
CODEX HOTFIX CORRECTS GPT-5.6 CONTEXT WINDOWS AFTER BUNDLED INSTRUCTION REFRESH
By RepoJournal · Filed · About OpenAI
Codex 0.144.6 shipped overnight to fix context window mismatches across GPT-5.6 Sol, Terra, and Luna after bundled instruction metadata got out of sync with the models themselves.
The hotfix [1] is a critical catch for anyone running inference against the latest GPT-5.6 variants in production. Those three models now correctly report 272,000 token context windows instead of the inflated figures that shipped in 0.144.5, which means your prompt caching and token budgeting will no longer silently exceed actual limits. In parallel, the team landed four substantial feature additions across realtime and tool integrations. You can now seed realtime V3 sessions with initial text items [2], which eliminates the cold-start delay for resuming conversations and gives you better control over session bootstrap. Audio output support is live for dynamic tools and code mode [3], converting MCP audio blocks into model input and providing helpers for inline data URLs. The legacy thread history pagination support [4] ensures clients using full-history resume work consistently across both paginated and non-paginated thread implementations. Finally, a performance win landed: TUI redraws during streaming now only trigger when the visible tail actually changes [5], cutting unnecessary screen flicker and reducing CPU overhead during long-running operations.
Action items
- → Upgrade to Codex 0.144.6 immediately if you deploy GPT-5.6 Sol, Terra, or Luna to production openai/codex [immediate]
- → Test audio output in dynamic tools and code mode; adjust prompt engineering for new audio() helper syntax openai/codex [plan]
- → Migrate to initialItems for realtime V3 sessions to eliminate cold-start latency openai/codex [monitor]
References
- [1] 0.144.6 ↗ openai/codex
- [2] Seed realtime V3 sessions with initial text items ↗ openai/codex
- [3] Add audio output support to dynamic tools and code mode ↗ openai/codex
- [4] Support legacy views for paginated thread history ↗ openai/codex
- [5] Avoid redundant TUI redraws while streaming ↗ openai/codex