Claude Code Digest — 2026-05-11 00:21:00
What the docs reveal
The War on Prompt Bloat: Dynamic Skill Context
Anthropic is fighting prompt bloat. As developers integrate more custom skills, context windows fill up before Claude processes a single code edit. The new maxSkillDescriptionChars and skillListingBudgetFraction settings reveal a critical engineering pivot: Claude now dynamically manages the context budget based on usage frequency.
Think of this mechanism as a Least Recently Used (LRU) cache for prompt instructions. If a skill goes unused, Claude collapses its instructions to conserve tokens. The system automatically evicts stale context to keep the model focused and responsive.
This dynamic pruning changes how you must write custom skills. Keep critical trigger instructions at the absolute beginning of your prompt, as trailing details may vanish during budget cuts. Use the new /doctor command to audit budget overflows. If you maintain a shared workspace with dozens of niche skills, configure skillOverrides to force "name-only" on secondary tools. Conserve your budget for the tools you trigger daily.
Concurrent Hooks Force Idempotent Design
Latency clearly drives the new hook architecture. PreToolUse hooks now run concurrently. Parallel execution eliminates the sequential delay before a tool fires, but it introduces a strict contract for developers. The engine resolves decisions by enforcing the most restrictive outcome (deny > ask > allow) and seamlessly aggregates all additionalContext strings.
Concurrency guarantees side effects. If Hook A denies a process, Hook B still executes its logic. A webhook or database log in Hook B fires regardless of Hook A's veto. Wait states do not exist here.
Design your hooks to be strictly idempotent. Treat them as pure validation functions. Never rely on sequential execution to protect sensitive side effects in downstream hooks.
Graceful Degradation in Auto-Mode
Auto-mode now fails with intent. By exempting fundamental file system operations (read, search, edit) from auto-mode classifier checks, Anthropic prevents internal classifier faults from locking developers out of their codebase. Standard file operations succeed even when the classifier panics.
Context window exhaustion now bifurcates based on your runtime environment. Interactive sessions gracefully degrade to manual approval prompts. Non-interactive headless runs immediately abort.
Anthropic built this distinction specifically for CI/CD pipelines. A headless automated agent must fail fast, throw an exit code, and terminate rather than hang indefinitely awaiting phantom user input. For interactive end-users, the fallback buys you a moment to step in. When you see the manual approval fallback trigger, run /compact immediately to purge your chat history and restore system capacity.