v2.1.91
Claude CodeSummary
This release introduces significant new capabilities, including support for larger MCP tool results (up to 500K chars) and the ability for plugins to ship and execute their own binaries. Several critical bugs were resolved, such as fixing conversation history loss during resume operations and correcting `cmd+delete` behavior across major terminals.
New Features
- Added MCP tool result persistence override via `_meta["anthropic/maxResultSizeChars"]` annotation, allowing up to 500K characters for results like DB schemas.
- Added `disableSkillShellExecution` setting to completely disable inline shell execution within skills, custom slash commands, and plugin commands.
- Added support for multi-line prompts when using `claude-cli://open?q=` deep links, correctly handling encoded newlines.
- Enabled plugins to ship executables under the `bin/` directory and invoke them directly as bare commands from the Bash tool.
Bug Fixes
- Fixed transcript chain breaks occurring with `--resume` that previously caused silent failures in async transcript writes, leading to lost conversation history.
- Fixed `cmd+delete` functionality not deleting text to the start of the line across iTerm2, kitty, WezTerm, Ghostty, and Windows Terminal.
- Fixed plan mode in remote sessions losing track of the plan file after a container restart, which previously caused permission prompts on plan edits and an empty plan-approval modal.
- Fixed JSON schema validation errors when setting `permissions.defaultMode: "auto"` in settings.json.
- Fixed an issue where Windows version cleanup was incorrectly protecting the rollback copy of the active version.
Improvements
- Improved guidance for the `/claude-api` skill, specifically detailing agent design patterns like tool surface decisions, context management, and caching strategy.
- Improved performance by routing `stripAnsi` through `Bun.stripANSI` for faster processing when using Bun.
- The Edit tool now uses shorter `old_string` anchors, resulting in reduced output tokens.
- The `/feedback` command now displays an explanation when it is unavailable instead of simply disappearing from the slash menu.