Patch release: artifact pinning on claude.ai, new permission-prompt CLI flags, clearer skill-description rules, and tighter permission-denial handling.
Changelog
[new] Pin and unpin artifacts to the user’s private claude.ai sidebar (reversible; does not change who can see the artifact). Cloud sessions that do not support pinning say so instead of retrying.
[new] CLI flags --permission-prompt-tool and --permission-prompts for controlling how permission prompts are presented.
[changed] Skill descriptions must be one short sentence (aim under 200 characters, hard cap 1024). On skill improvements, reuse the existing description unless when-to-use actually changes.
[changed] When a tool use is denied, the model is told what the user said and should explain what was blocked and why, then continue with work that does not need approval.
[changed] Artifact / publish responses can surface remaining publish quota and reset time (publishesRemaining, publishesResetAt).
[changed] SDK compatibility matrix no longer lists wrapper 0.3.217 among tested versions (0.3.218+ remain).
Models & prompts
Guidance for pinning artifacts to the claude.ai sidebar, including unpin and “not available in this cloud session” paths.
Permission-denied copy now carries the user’s denial reason and steers the agent to disclose the block and keep going on unblocked work.
Skill create/improve flows emphasize short, stable “when to use” descriptions rather than restating them on every edit.
Autonomous / focus-mode / subagent and related system guidance refreshed in place (behavioural polish; no model ID or default model change in this release).
Under the hood
Session and skill plumbing: CLAUDE_CODE_SESSION_ORIGIN, CLAUDE_CODE_SKILL_ATTRIBUTION, and CLAUDE_CODE_ARTIFACT_PIN for origin tracking, skill attribution, and pin support.
Optional CLAUDE_CODE_ENABLE_FUNCTION_HOOKS and CLAUDE_CODE_PLUGIN_DIR_WATCH for function-hook experiments and watching plugin directories.
CCR path: early hydrate prefetch (CLAUDE_CODE_CCR_EARLY_HYDRATE_PREFETCH), less redundant heartbeating, and remote-tool restart adoption handling when routes drop or become unroutable.
Clearer handling of orphaned permissions (applied vs still unapplied at turn end or on resume) and of config fallback when the file is unreadable or under contention.
