Skip to content

v1.4.0

Choose a tag to compare

@nrzz nrzz released this 15 Oct 07:55
· 34 commits to main since this release
c3073e3
  • A twelfth tool: claude-permission-doctor tells you why Claude Code asked for a permission, finds the rules that do nothing (dead, repeated, hidden behind a deny or ask rule, or skipped by Claude Code) and proposes rules from the commands that ran without one. Like md-doctor, session replay and chat ferry it only runs on demand, so there is nothing to install or switch on: the setup page lists it with three buttons (show this project's rules, check them, suggest rules from history), none of which changes a file, and status says it runs on demand. The marketplace carries its permission-doctor plugin. The end-to-end test gives it a config of its own with a prefix rule, a :* rule, a literal rule and a deny, and a synthetic transcript in which one command ran four times without a rule: show --json lists the four rules with their file, why "git status && npm test" --json allows the first part by the :* rule and finds no rule for the second, why "rm -rf build" --json is denied by the deny rule, lint --json calls the literal rule nobody used dead (and no prefix rule), suggest --json proposes Bash(docker compose up *), and none of the five runs changes a file. With a Claude Code CLI at hand it also installs the plugin and runs its skill's two commands from the installed copy.
  • Always-on tokens: all ten plugins together add about 165 by Claude Code's own estimate (the nine before added 148; permission doctor's skill description adds 17), so the limit in the end-to-end test is now 200, not 150. The estimate counts the description of every user-only skill, and Claude Code never sends those to the model: session tidy and permission doctor each captured the request the model receives, with a fake API, and found no trace of their skill. A new tool's skill adds 15 to 17 to the estimate, so a limit of 150 would have meant shortening descriptions until they stop being readable. The test keeps the estimate as a check and adds what it stands for: with all ten plugins installed it resumes a session through Claude Code against a fake API and fails if any request mentions any of the 11 user-only skills (found by their disable-model-invocation flag and named in the check), after first checking that a skill the model may invoke, planted in the project, is listed in the first request, so that an absence means something. Claude Code lists skills in the request's messages, not in its system prompt or tool list, so the whole request is searched.
  • Website and README: twelve tools, ten plugins, the new tool's card and rows, over 5,700 tests, and the preview images say twelve. Two more tests: the version in package.json must be the newest one in this changelog and in the website's What's new, and the limit the README and CONTRIBUTING.md state must be the one the end-to-end test enforces.