$ the-wire · showcase
CODEX FIXES RACE CONDITIONS AND STABILIZES STREAMING PIPELINE
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
Codex shipped five architectural fixes overnight that eliminate liveness races in side conversations, stabilize the Markdown streaming stack, and cut TUI layout time significantly.
The most critical change addresses a timing bug where side conversation notifications could arrive after the fork response, causing incorrect unavailability reports [1]. This is fixed by seeding agent navigation from the fork response itself before thread selection. In parallel, the streaming pipeline got a structural upgrade: the Markdown collector is now the source of truth for rendered output, with the streaming controller reading directly from committed byte ranges instead of maintaining duplicate buffers [2]. This eliminates a class of synchronization bugs between renderer and controller. On the performance side, TUI Markdown layout got a 3x speedup through bulk table width allocation and a single forward scan for hyperlink remapping instead of per-cell processing [3]. Two smaller but important fixes landed: legacy exec policy rules now migrate cleanly on startup without breaking user-created rules [4], and command completion is now tracked independently from output streaming, so interrupted turns won't be marked inactive prematurely [5].
Action items
- → Pull the latest Codex build to get the side conversation race fix before production openai/codex [immediate]
- → Review your TUI Markdown rendering if you've hit layout performance walls openai/codex [plan]
- → Monitor streaming output fidelity in your next deploy cycle openai/codex [monitor]
References
- [1] Avoid liveness races when starting side conversations ↗ openai/codex
- [2] Use the Markdown collector as the streaming source of truth ↗ openai/codex
- [3] Speed up TUI Markdown layout ↗ openai/codex
- [4] Migrate legacy exec policy allow rules (#34271) openai/codex ↗
- [5] Track TUI command completion separately from output ↗ openai/codex