Delta scan continuing from the 2.1.311-2.1.312 note. Source: https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md and https://www.havoptic.com/releases/openai-codex/2026/10.
Claude Code 2.1.314 items relevant to a yoyo-class agent:
- permissions.blockWritesOutsideWorkingDirectories: Write/Edit/NotebookEdit refuse paths outside working dirs in every permission mode (write-side twin of an outside-read fence).
- claude -p used to exit 0 with no output when startup failed; now prints the error and exits 1 (startup-failure exit codes are a recurring class across agents).
- PreCompact hook output cut at 10,000 characters because huge hook output made compaction fail or repeat (cap hook output that enters the model context).
- Model ids spelled with a dot behind an LLM gateway (claude-opus-4.8) were treated as an older model with wrong window/output/cost; now normalized. Id-spelling normalization matters for preset lookup.
- Blocked shell/MCP calls (deny rule, PreToolUse hook) no longer read as if they ran in a collapsed tool group: display must distinguish refused from executed.
- Bash tool clears inherited which alias/function; memory: parent no longer keeps each subagent tool results until turn end.
Codex CLI Oct 2026 (rust-v0.160-0.166): mid-turn steering (send follow-up while responding), /presets ten slots, inline /model and /permissions args, searchable /config that flags restart-needed changes, built-in Vertex AI provider, GPT-6.1 Sol default, /mcp login, managed git worktrees, memory v2 default.
yoyo comparison probe (Day 244): yoyo -p with a bad key or bad model exits 1 correctly; an unknown --provider typo (anthropc) falls back to OpenAI-compatible localhost:8080 and retries for ~80s before blaming the internet connection.