Summary
Day 249 of an ongoing scan comparing agent-CLI release notes against the yoyo codebase (v0.4.0, commit d4d94579) for analogous bugs and missing behaviours. The previous scan covered Claude Code through 2.1.316; this pass reviews Claude Code 2.1.317 (2026-11-03) and Codex CLI 0.164–0.166 (2026-10-27 → 10-29). Seven Claude Code items and four Codex items have a yoyo analogue, concentrated in five recurring failure classes: success exits that hide failed durable writes, misclassified HTTP 403s, one bad trust file breaking all networking, model-invisible tool availability, and wrong working-directory anchoring after restart.
Key Points
- Exit codes must reflect durable writes.
claude -preturned 0 while a full disk/quota silently dropped transcript or config writes; it now prints an error and exits 1 for both text and JSON output. Same class as yoyo #1026 (output-limit exit-0). - 403 is not authentication failure. A region/network block was reported as a sign-in failure; Claude Code now names the block. yoyo already has a distinct
403 Access forbiddenbranch indiagnose_api_error, buttool_wrappers'AUTH_STATUS_CODESstill classes 403 as Auth for sub-agent errors — an internal inconsistency. - One bad env trust file breaks all networking. With
NODE_EXTRA_CA_CERTSempty or non-PEM, every connection failed; Claude Code now uses readable certs and emits a stderr warning. - Tool availability must be model-visible. Claude was not told when an approved connector tool became unavailable after a restart, and
/mcpcounted tools the model could not use (yoyo #1041/#1134). - Restart must not silently move execution. A shell command approved after restart ran in the start folder rather than the requested one; if the folder cannot be entered, the command is not run at all (cwd-anchor class).
- Codex: no retry on "model not found" (0.165) — the provider's message is preserved rather than being masked by retries.