command-scout

The most frustrating moment when delegating build tasks to an Agent is not when it runs the wrong command -- it is when it does not know which command to run at all. Makefiles carry a dozen targets, package.json lists a wall of scripts, justfiles and deno tasks each speak their own dialect. The Agent is left guessing, or you end up feeding it one command at a time. command-scout scans the workspace root once at activation, extracts every declared command, and registers them as ready-to-run agent tools. How a project builds, tests, or lints is decided by the project itself, not by guesswork.

It supports multiple build systems: Makefile targets, package.json scripts with automatic runner detection (npm / yarn / pnpm / bun), justfile recipes, and deno.json tasks. Each command carries rich metadata -- descriptions, prerequisite lists, referenced variables, and argument hints -- and is registered under a deterministic name (make_build, pnpm_dev, just_lint), with numeric suffixes resolving any collisions. Every tool accepts a single optional args parameter that is appended verbatim to the command line, so passing --port 8080 or overriding a version number is just one string. Execution goes through the platform shell with a hard per-command timeout and full process-tree cleanup; stdout and stderr are each truncated at 1 MiB, and the result is returned as structured JSON. Zero runtime dependencies, plain Node.js ESM, and a standalone CLI (scan / docs) works without any harness.

Built for developers who run DeepSeek Harness or similar agent workflows and want to stop manually dictating every command. If you juggle projects that use different build systems, or you want the Agent to see the project-specified command list before it executes anything, command-scout is the lightweight plugin that removes the guessing step.