Loading ToolBench
Delegate bounded tasks to the local Google Antigravity CLI.
Two tools with complete schemas and proper Zod validation. Tool naming is clear (verb_noun pattern: agy_check, agy_delegate). Descriptions are present and adequate (60-90 chars each). agy_delegate has comprehensive parameter documentation with enums, ranges, and defaults. agy_check is simpler but well-defined. Error handling returns actionable messages ('Antigravity CLI was not found. Run agy_check for installation guidance.'). Tool annotations (readOnlyHint, destructiveHint, idempotentHint) are correctly applied. Main gaps: output schemas are not formally documented in the tool registration (only returned as structuredContent at runtime), and parameter descriptions could be more explicit about constraints and use cases.
Verify that Google Antigravity CLI is installed and return its executable path.
Run one bounded prompt through Google Antigravity CLI in non-interactive print mode and return the result.
Output schemas not formally documented in tool registration. structuredContent is returned at runtime but not declared in the tool definition, forcing LLMs to infer response structure.
Parameter descriptions lack explicit format/constraint guidance. 'cwd' says 'Absolute existing workspace directory' but does not state validation happens at runtime or what error to expect if invalid.
No pagination or result-limiting guidance in agy_delegate output. If Antigravity returns large output (up to 2 MiB), the tool truncates silently but does not offer a way to retrieve the rest.
agy_check returns both text content and structuredContent with 'available' and 'executable' fields, but the structuredContent structure is not declared in the schema.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2025-06-18+ | v2 |
No confirmation or dry-run mode for agy_delegate despite destructiveHint=true. Agents cannot preview changes before committing them.