@caboc/cli

A tiny Node CLI for CABOC — a markdown-only DSL for AI workflows.

CABOC routines are directories of markdown files (workflow + agents). An LLM main loop (e.g. Claude Code) reads and executes them. This CLI does not run an LLM itself — it scaffolds new routines, lints them, prints the system prompt the runtime should consume, and inspects past runs.

Install

npx @caboc/cli <cmd>
# or
npm i -g @caboc/cli

Commands

caboc init <routine-name>

Scaffold a new routine under routines/<routine-name>/ with a starter WORKFLOW.md and a sample agent.

caboc init triage

caboc lint <routine-dir>

Validate frontmatter, enforce the model-name denylist, and ensure every USE AGENT <ref> resolves to a sibling agents/<ref>.agent.md.

caboc lint routines/triage

Exits 0 on success, 1 on any failure. Prints ✓ per passed check and ✗ <file>:<line> — <reason> per failure.

caboc run <routine-dir> --inputs <inputs.json>

Prints the system prompt that an LLM-runtime should be fed to execute the routine. Also creates runs/<run-id>/inputs.json inside the routine directory and references that path in the prompt.

Pipe into Claude Code, paste into a chat — your call.

caboc run routines/triage --inputs ./issue.json | claude --print

caboc inspect <run-dir>

Pretty-prints a summary of a completed run: event counts by kind, step durations, truncated summary_md, and any assertion.failed / agent.repair events.

caboc inspect routines/triage/runs/20260511-143022-a1b2c3

Routine layout

routines/<name>/
├── WORKFLOW.md             # frontmatter (YAML) + PROCEDURE body
├── agents/
│   └── <agent>.agent.md    # frontmatter + system prompt body
└── runs/<run-id>/          # populated at exec
    ├── inputs.json
    ├── transcript.ndjson
    ├── outputs.json
    └── state.json

Body invocations are limited to:

USE AGENT <ref> [SESSION fresh|continuous|fork] WITH inputs={...}.

License

Apache-2.0