Summary
Delta analysis covering releases after the 2026-10 scans: Claude Code 2.1.317 (2026-11-02) and Codex CLI 0.165–0.166. The theme across both agents is hardening of failure semantics and per-model behavior: hook failure policies extended to the Agent SDK, misleading error presentations corrected, retries narrowed to genuinely retriable errors, and sub-agent settings resolved from the child model rather than the session. Secondary themes are privacy/credential hygiene (account-email disclosure prompt, credential masking in remote bash) and model catalog churn (GPT-6.1 Terra, built-in Vertex provider).
Key Points
- Claude Code 2.1.317 extends
onFailure:blockfrom command/HTTP hooks (2.1.295) to Agent SDK hook callbacks — hook failure policy is now uniform across hook types. - A privacy guard now asks before a web or shell tool call sends the account email over the network.
- Sub-agents on a different model resolve effort from that model's saved/default effort, not the session's.
claude -pexits 1 (text and json modes) when a full disk or exhausted quota drops transcript writes; previously it exited 0 — a silent-success bug.- A region/network 403 is no longer misreported as a sign-in failure.
- The Read tool's "file unchanged" stub was removed.
- Anthropic Message threads entered public beta: send only new messages plus the previous response id; ~35% median time-to-first-token reduction at 500K–1M token contexts, same token billing, 5-minute expiry.
- Codex CLI stops retrying when the provider reports model not found (and preserves the message); resume/fork now carry over approval/permission settings and reasoning-token accounting (avoiding premature compaction).
- Codex adds
codex doctor --network-only, GPT-6.1 Terra in catalogs (including Bedrock and Vertex), built-in Vertex provider, and on by default with external-agent memory import.