Claude Code 2.1.88 fixes StructuredOutput cache bugs and adds PermissionDenied hook

• Anthropic shipped Claude Code 2.1.88 with 41 CLI changes and 3 system-prompt changes, led by a fix for a StructuredOutput schema cache bug that had caused "~50% failure rate" in multi-schema workflows, according to the release thread and the changelog.
• The release adds a new PermissionDenied hook that fires after auto-mode classifier denials and can return {retry: true}, which matters more now that Anthropic's auto mode rollout has expanded to Enterprise and API users.
• Claude Code's changelog also targets long-session stability with fixes for prompt cache misses, repeated nested CLAUDE.md injection, an LRU-key memory leak, and a potential out-of-memory crash on files larger than 1 GiB.
• The update lands while Anthropic is investigating reports of unexpectedly fast Claude Code quota burn; in a widely shared thread, one user said community findings around cache invalidation might explain the spike, while Anthropic's response said the team was "actively investigating."

Claude Code 2.1.88 is a reliability-heavy release, but it also changes how agented sessions behave. The headline addition is the PermissionDenied hook: after auto mode blocks a command, the hook can fire and a {retry: true} response tells the model it may try again, as described in the CLI changelog. That arrives a day after Anthropic's auto mode rollout said auto mode is now available on Enterprise plans and for API users via claude --enable-auto-mode.

The same build adds CLAUDE_CODE_NO_FLICKER=1 for flicker-free alt-screen rendering with virtualized scrollback, plus named subagents in @ mention typeahead changelog. On the prompt side, the prompt diff summary says Anthropic moved workflow rules into a new "Session-specific guidance" block and added a stricter agent rule: "Never delegate understanding." The updated prompt also tightens Read tool pagination so offset and limit must be bounded integers, which should reduce invalid reads in long file inspections prompt updates.

The most concrete fix is the StructuredOutput cache bug. Claude Code's release notes say it had been causing about half of requests to fail in workflows that switch across multiple schemas. The same changelog also fixes prompt cache misses when tool schema bytes changed mid-session, stops nested CLAUDE.md files from being re-injected "dozens of times," and patches a memory leak where large JSON inputs were retained as LRU cache keys changelog. Other recovery fixes include restarting the LSP server after a crash instead of leaving it in a zombie state, correcting SDK error results to set is_error: true, and avoiding an Edit-tool out-of-memory crash on files over 1 GiB changelog.

Those fixes overlap with, but do not fully explain, the quota complaints that flared up before the release. In altryne's thread, the reported theory was cache invalidation: he wrote that uncached tokens are "10x-20x more expensive" and said community findings pointed to a Bun-related string replacement issue and --resume breaking cache, while also stressing he had "not independently confirmed this." The linked GitHub issue describes conversation history being invalidated on subsequent turns, and Anthropic's response said it was aware users were hitting usage limits "way faster than expected" and was actively investigating. So 2.1.88 clearly fixes several cache and session bugs, but Anthropic has not said this release closes the broader quota-burn incident.