olo-dot-io/uni-cli/unicli-claude-code

Uni-CLI × Claude Code — Reliable Invocation

TL;DR — If the payload contains ANY of: quotes, emoji, newlines, JSON,

non-ASCII, or is longer than ~60 chars, PIPE IT. Never inline.

Why this skill exists

Transformers live in TC0. Matching balanced " and ' pairs with arbitrary nesting is a mod-2 state-tracking problem that TC0 cannot solve in constant decode depth. Empirically, top LLMs land <43% on 0→1 CLI generation (CLI-Tool-Bench arXiv:2604.06742) and 62% of agent-tool bugs cluster at the tool-invocation / execution stages (arXiv:2603.20847).

Uni-CLI gives you THREE invocation channels. One of them (shell-args) inherits the TC0 bottleneck. The other two do not.

The decision rule

When unsure → stdin-JSON. It never makes things worse.

The three channels

1 — stdin-JSON (preferred for complex payloads)

echo '{"query": "she said \"hi\" 🎉", "limit": 10}' | unicli twitter search

• Uni-CLI auto-detects non-TTY stdin whose body starts with {
• Precedence: stdin-JSON > --args-file > shell flags > defaults
• Works for EVERY command — no per-adapter opt-in needed

2 — --args-file (preferred when payload is reused)

cat > /tmp/q.json <<'EOF'
{ "query": "machine learning", "limit": 20, "locale": "en_US" }
EOF
unicli arxiv search --args-file /tmp/q.json

• Absolute paths recommended
• JSON only (YAML and TOML deliberately not supported — minimize format ambiguity)

3 — shell args (only for trivial payloads)

unicli hackernews top --limit 5

Fine when every value is ASCII, no quotes, no emoji.

Runtime introspection — unicli describe

Before invoking any unfamiliar command, run:

unicli describe <site> <command>

You get back a JSON object containing:

• args_schema — JSON Schema draft-2020-12 for the argument bag
• example_stdin — a realistic JSON body you can modify and pipe
• channels — the three invocation templates
• next_actions — hints for what to run next

This replaces reading docs. The CLI is the documentation.

The response envelope (v2)

Every successful call returns:

{
  "ok": true,
  "schema_version": "2",
  "command": "twitter.search",
  "meta": { "duration_ms": 847, "count": 10, "surface": "web" },
  "data": [ ... ],
  "error": null,
  "next_actions": [
    { "command": "unicli describe twitter search", "description": "…" },
    { "command": "unicli twitter search --args-file <path.json>", "description": "…" }
  ]
}

Errors return ok: false with a structured error object (adapter_path, step, suggestion) and failure-classified next_actions. unicli repair appears only for established adapter-drift classes and verifies an edit; it never edits automatically. If you see invalid_input from the hardening layer, the arg bag tripped a safety check — read error.suggestion for the exact fix.

--dry-run — preview before committing

echo '{"query": "🎉", "limit": 5}' | unicli twitter search --dry-run

Prints the resolved ArgBag, the source channel, and the pipeline step count without touching the network / filesystem. Use this whenever you want to confirm the agent built the payload correctly.

Common failure modes → fixes

Recipes

Search with a complex query

printf '%s\n' '{"query": "large language models \"emergent\" behaviour", "limit": 25}' \
  | unicli arxiv search

Multi-arg desktop command

cat > /tmp/render.json <<'EOF'
{ "scene": "/Users/me/project/main.blend", "output_dir": "~/renders", "frame_start": 1, "frame_end": 240 }
EOF
unicli blender render --args-file /tmp/render.json

Debug-first workflow

unicli describe youtube transcript         # 1. understand the contract
echo '{"url": "https://…"}' | unicli youtube transcript --dry-run   # 2. preview
echo '{"url": "https://…"}' | unicli youtube transcript             # 3. run

Version contract

• v0.214.0+ required for stdin-JSON auto-detect, --describe, and consistent MCP tool discovery metadata
• v0.213.1 or earlier agents should fall back to unicli schema <site> <cmd>

and explicit -- flag args