skillfed
REPO

HKUDS/CLI-Anything

The core bet is that CLI is the right abstraction layer between AI agents and existing software — not APIs, not GUI automation, but a structured command-line harness wrapping whatever the software already does. Point the tool at a codebase or repo, run /cli-anything, and a 7-phase generation process produces a Click-based CLI with REPL mode, --json output on every command, undo/redo state, and a full test suite. The generated harness calls the real software backend — LibreOffice headless for PDF export, bpy for Blender rendering, sox for Audacity audio processing — rather than reimplementing anything.

The HARNESS.md methodology document is the honest center of the project. It names failure modes explicitly: the "Rendering Gap" where GUI apps apply effects at render time that naive file manipulation silently drops; timecode rounding errors at non-integer frame rates like 29.97fps; the need to verify export success by checking magic bytes and ZIP structure rather than trusting a zero exit code. Encoding these as a reusable SOP is the project's most transferable contribution.

Test numbers are specific and traceable. The harness table totals 2,461 passing tests — 1,732 unit tests, 579 end-to-end tests, and 19 Node.js tests. The Blender harness carries 208 tests; s&box has 244. LibreOffice E2E tests verify actual %PDF- magic bytes in output. The project is explicit that tests fail rather than skip when backends are missing, preventing harnesses from appearing healthy when real software isn't installed.

The CLI-Hub package manager (pip install cli-anything-hub) adds a registry layer so agents can autonomously discover and install harnesses. Each harness ships a SKILL.md with YAML frontmatter, command groups, and agent-specific guidance for JSON output and error handling.

The stated limitations are worth taking seriously. Generation depends on frontier-class models — the README names Claude Opus 4.6, Claude Sonnet 4.6, and GPT-5.4 — and warns that smaller models produce incomplete harnesses requiring manual correction. Software shipping only compiled binaries degrades harness quality substantially. A single generation run often needs one or more /refine passes to reach production coverage. None of this is hidden.

A source-grounded methodology for wrapping real software in agent-callable CLIs, with explicit failure modes and 2,461 tests backing the claims.

Sources & links