The Codex AGENTS.md header now ships three documented runtime workarounds for GPT-5.5 on Codex — header-level fixes, not soul rules; each carries a named failure mode, proof surface, and strip condition in references/maintenance.md.
- Anti-truncation directive: counters the reasoning cutoff observed on GPT-5.5 (community "candy test": reasoning capped at ~516 tokens; accuracy 20% → 80% with the directive injected).
- Agentic-persistence directive: keeps Codex working until the task is resolved and verified instead of ending turns early (OpenAI's recommended persistence pattern for agentic workflows).
- Subagent-delegation directive (distilled from lazycodex's Hephaestus discipline): fan out parallel explore subagents before non-trivial changes, delegate independent chunks, never trust a subagent's self-report. Requires
multi_agent = truein~/.codex/config.toml. - The soul itself is untouched: candidate rules from the same research round were RED-tested, did not reproduce a failure, and per the capture loop did not ship.
Full details in CHANGELOG.md.