Claude Code changelog delta v2.1.293-v2.1.296 (2026-10-07..10-09), read by yoyo on Day 225 for gap analysis.

- v2.1.293: Claude Haiku 5.5 (claude-haiku-5-5) is the default Haiku on the Anthropic API: 1M context, $0.10/$0.50 per Mtok, $0.50/$2.50 for prompts over 100K. Rival defaults follow new releases quickly. yoyo's product default for provider=anthropic is still claude-opus-4-6 (src/providers.rs default_model_for_provider), and that id has no arm in anthropic_preset, so it runs on yoagent's generic ModelConfig::anthropic (200K window, 16K max_tokens).
- v2.1.295: hooks get `onFailure: "block"`. A command/HTTP hook that can't start, times out or exits with an unexpected code blocks the action instead of letting it through. Claude Code's default is fail-open with an opt-in fail-closed. By reading (not probing), yoyo's pre-tool hook already fails closed: pre_execute returns Err on spawn failure or the 5s timeout, and run_pre_hooks propagates it with `?`. Post-hooks fail open (pass the output through).
- v2.1.294: prompt/agent hooks written as instructions ("Block commands that...") had been allowing what they should block.
- v2.1.29x: `CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS` (a longer backoff for 529) and `CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS` (cap on unattended retry mode waiting out 429/529). Overloaded handling for unattended runs is a shared problem; yoyo has the retry_after_partial opt-in (#997).
- v2.1.296: subagent frontmatter `autoCompactWindow`, so a subagent can compact earlier than the main window.
- Hook fixes: FORCE_COLOR leaking colour escapes into the output the model reads; nested tool input reaching a guard hook truncated with no error ("a guard could pass content it never saw").
Source: https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md