Smart tools: cut Claude Code’s needless reads and output

If a file already read in the conversation, and unchanged since, is read again with cat, head, tail or sed -n, the request is refused and the agent is told the content is already in its context. On the first full read of a file over 800 lines and 24 KB, the agent gets the file’s outline from the project map instead and reads only the line ranges it needs. Repeating the same request goes through; /compact or /clear resets this memory.

A command that runs the whole test suite, such as “npx vitest run”, “jest” or an “npm test” whose script is vitest or jest, is narrowed to the tests of the uncommitted changes and of files importing them, up to two levels. If more than 40 test files are affected, the whole suite runs. The agent is told what was narrowed and can run everything with AGENTVERA_ALL_TESTS=1.

Before a TS/JS file is edited, the current errors of that file and of up to 8 files importing it are recorded. After the edit they are checked again with the project’s own TypeScript and ESLint, and only new errors go back to Claude in the same turn. Warnings are left out and the check is capped at 8 seconds.

When the same lint, type check, test or build command runs again in the same folder on exactly the same code, whether this agent or a parallel one ran it, the earlier output comes back and the agent is told who ran it and how many minutes ago.

• Covered: eslint, tsc, vitest run, jest, biome, ruff, mypy, pytest, cargo, go vet, test and build, and npm, pnpm and yarn lint, typecheck, check, test and build scripts.
• Chained commands, --watch, --fix and --write aren’t cached; results are kept for 30 minutes.
• Running the command again right after a cached answer forces a fresh run.