Skill detail

call-agy

Delegates a bounded task to the locally authenticated Google Antigravity CLI (agy) and returns a host-verified handoff.

MatchDirectRecommended for Antigravity
Sourcef1rstdan/​call-agyExternal source
Reported installs60Popularity signal only

Inspect before use

Read the source instructions and check the requirements before you install this skill.

Saved source preview

SKILL.md

The saved excerpt is a snapshot from when this skill was listed. The external source remains the complete and most current version.

---
name: call-agy
description: "Delegate a bounded task to the locally authenticated Google Antigravity CLI (`agy`) and return a host-verified handoff. Use when the user explicitly asks to call Antigravity, `agy`, or `$call-agy` for coding, review, analysis, testing, or verification. Exclude informational questions and generic subagent requests."
---

# call-agy

Use the official local `agy` CLI as a bounded sub-agent. The calling Agent remains the task owner: it scopes the request, reviews the handoff and workspace, verifies material claims, and reports the combined result.

## Router Rules

- Trigger when the user explicitly asks to call `$call-agy`, `agy`, or Antigravity. Explicit invocation wins even when the task is trivial.
- Also trigger when an authorized host policy calls for a specifically Antigravity-backed implementation, debugging, review, test, analysis, or verification pass.
- Do not trigger for installation/docs/pricing questions, unsolicited trivial work, or generic “use a subagent” requests that do not identify Antigravity.
- Do not delegate when the host policy or user forbids cross-agent work.
- One invocation owns one focused goal with observable completion criteria. Split unrelated work before delegating.
- Keep the CLI default model, effort, and agent unless the user explicitly pins one. Never guess or silently substitute a slug.

## Compact Workflow

1. For a smoke test or direct text reply, skip repository discovery and file hints. Otherwise read enough local context to state the goal, completion criteria, constraints, expected verification, workspace, and up to six priority paths. For broad inspection, tell AGY to keep exploration proportional, avoid unbounded recursion, and deliver gathered evidence before time expires; do not impose fixed exclusions that could hide an explicit target.
2. Separate execution intent from authorization. Use `plan` only for planning/read-only work, `accept-edits` for authorized changes, and scoped permission rules for commands. Announce the trusted-workspace preset before using it.
3. Invoke the OS-native launcher once. Use `--task` for a short request, stdin for a multiline request, and `--file` for priority paths.
4. On `status=SUCCESS`, read the reported `output_path`, inspect any actual diff, and run host-side verification proportional to risk.
5. On failure, follow [Permissions and recovery](references/permissions-and-recovery.md): diagnose, make one safe local correction, and retry the same task at most once.
6. Return the substantive result, verification, limitations, and any known process-artifact paths.

## Fast Path

The canonical implementation is `scripts/call_agy.py`; use the native launcher from this Skill directory.

On Windows, stay in the current PowerShell process and invoke the launcher directly with `&`. Do not start a nested `powershell -Command`. Short tasks use `--task`; the wrapper stages large-prompt fallbacks under system temp and groups them by conversation after the run.

Explicit connectivity smoke test (no repository discovery):

```powershell
& "<skill-directory>\scripts\call_agy.ps1" `
  --task "Reply exactly: Antigravity is connected. Do not inspect files or run tools." `
  --workspace "<absolute-workspace>" --timeout 90s
```

Read-only or planning task that needs repository tools:

```powershell
& ".\scripts\call_agy.ps1" --task "Review the request flow and cite relevant files" `
  --workspace . --mode plan --sandbox
```

`plan` is an execution intent, not a host permission grant. In headless mode, shell commands
still need scoped `permissions.allow` rules. The wrapper prefers native read-only file tools
for plan tasks and reports `HEADLESS_PERMISSION_BLOCKED` rather than silently escalating.

```bash
sh ./scripts/call_agy.sh "Review the request flow and cite relevant files" --workspace .
```

Authorized edits in a trusted workspace use the success-first preset selected by this Skill:

```powershell
.\scripts\call_agy.ps1 "Fix the bug a
Read the full source on GitHub (opens external page)
Context

Related work