Плагин agent-kit

Implements a written spec in the listed files; never runs tests unless the project allows it; returns a constraint ledger and the tests to run.

Answers ONE question about a file over the read threshold, a multi-file question or a large diff, in path:line bullets; I/O only, never judges.

Writes predictable code (scaffolds, fixtures, stubs) when over 80% derives from a reference file; design decisions are Builder work.

Hard debugging only: root-causes a failure others could not explain; returns the mechanism, evidence and a Builder-ready fix brief.

Independently reviews Builder output against the spec and re-runs static checks; reports merge-blocking defects with proof; reports, never fixes.

Reads docs and source and returns verified facts with a source for each; flags anything it could not confirm.

Finds files, symbols and call sites and returns paths with line numbers only, no contents; use it to locate code before reading or editing.

Check that agent-kit is set up correctly in this project and on this machine — hooks, agents, model aliases, the Agent contract and conflicts with a script install. Read-only.

Set up the current project for agent-kit — write a filled-in "## Agent contract" section into its CLAUDE.md (verify commands, test policy, excluded paths, generated files) and gitignore .claude/tmp/

Read|Bash

startup|resume|clear|compact|fork

В файле нет описания.