smartpipe for AI agents

Agents (Claude Code, Codex, and friends) drive smartpipe well - it is a CLI with machine-parseable output, disclosed costs, and deterministic free verbs.

The agent-facing skill lives at the repository root - SKILL.md - as a lean entry point whose quick-reference table routes into focused reference files under skills/smartpipe/ (ingestion, extraction, cost-and-reliability, output, recipes) that agents load on demand. Together they cover invocation patterns, the free-vs-paid cost model, structured extraction, reliability semantics (exit codes, skips, the circuit breaker), and the machine-output contract (parse stdout JSONL; never parse stderr).

Two rules worth repeating to any agent:

• belt exploratory runs with --max-calls N
• cut with free verbs (where, sample) before paid ones