Agy

Antigravity CLI (agy) Skill

Leverage the agy CLI to perform live online web searches, conduct deep technical research, and delegate small, focused tasks.

Scripts live alongside this file in scripts/ and work from any project directory.

Critical Execution Rules

1. Always gemini-3.8-flash-high at --effort high
   • Every invocation, in every mode (search, deep, task), uses --effort high --model gemini-3.8-flash-high. This is not mode-dependent — there is no lower-effort or different-model path. The wrapper scripts hardcode this; if you ever call agy directly instead of through the scripts, set both flags yourself, every time.
2. Non-Interactive Flag Order
   • Always supply --dangerously-skip-permissions BEFORE --print (-p).
   • If --print is placed first without an attached argument, it treats subsequent flags as prompt text and fails.
   • Correct format: agy --dangerously-skip-permissions --print "Your prompt here"
3. Prefer the Wrapper Scripts
   • Web search: ~/.claude/skills/agy/scripts/agy_search.sh "<query>"
   • General: ~/.claude/skills/agy/scripts/agy_runner.sh [search|deep|task] "<prompt>"
   • They handle argument escaping, timeouts, and flag sequencing automatically.
4. Citations & Sources
   • Live web searches return grounded source URLs. Always preserve and surface these links in the final answer to the user.
5. Quoting
   • Always wrap the query in double quotes. The scripts join all trailing args, so an unquoted query still works, but quoting avoids shell globbing on *, ?, and ().

1. Online Web Search (Primary Capability)

Use online search whenever information is needed beyond the training cutoff or in real time.

When to Trigger Online Search

• API Documentation — latest syntax, parameters, or functions for recently updated SDKs.
• Breaking Changes & Release Notes — new framework versions (Next.js, React, PyTorch, Go, Node.js).
• Error Diagnostics & GitHub Issues — exact stack traces or error strings to find active discussions.
• Package Versions — newest stable release on npm, PyPI, or crates.io.

Execution

# Dedicated search script (preferred)
~/.claude/skills/agy/scripts/agy_search.sh "<query>"

# Multi-mode runner
~/.claude/skills/agy/scripts/agy_runner.sh search "<query>"

# Direct CLI
agy --dangerously-skip-permissions --effort high --model gemini-3.8-flash-high --print \
    "Search the live web for: <query>. Provide a concise summary with official source links."

Query Optimization Tips

• Include current year or version: "Next.js 15 Server Actions best practices"
• Restrict to documentation domains: "Python 3.13 changelog site:docs.python.org"
• Search error strings verbatim: "TypeError: Cannot read properties of undefined (reading 'digest')"

Typical latency: ~20–30s.

2. Deep Searching & Technical Research (deep)

For architectural trade-off analysis, complex debugging, or in-depth technical comparisons requiring high-effort reasoning:

~/.claude/skills/agy/scripts/agy_runner.sh deep "<topic>"

# Direct CLI
agy --dangerously-skip-permissions --effort high --model gemini-3.8-flash-high --print \
    "Conduct a thorough research investigation on: <topic>. Detail pros, cons, architectural implications, and sample implementations."

Typical latency: 35s–5m (6m timeout).

3. Simple and Small Tasks (task)

For quick standalone subtasks, boilerplate generation, or secondary verification:

~/.claude/skills/agy/scripts/agy_runner.sh task "<task description>"

# Direct CLI
agy --dangerously-skip-permissions --effort high --model gemini-3.8-flash-high --print \
    "Perform this focused task: <task description>."

Typical latency: ~5–20s.

Output Handling

• Parse the text printed by agy on stdout.
• Surface source URLs to the user for reference.
• On error (timeout or non-zero exit), inspect stderr and retry or fall back gracefully.
• The scripts exit 1 with a usage message when the query/mode is missing, and exit 1 with an explanatory message if agy is not on PATH.

See references/cheatsheet.md for the full flag and model reference.