Critical multi-agent consensus for copy that survives its own reviewers
Forge exposes 2 tools (image-generation, graph-generation) with minimal schema rigor. Both tools have descriptions (50-60 chars), but input schemas lack critical structure: parameters are present but lack type definitions in the visible schema. No output schema is documented. Parameter descriptions are generic ('Text prompt describing the image', 'Data to visualize') without format constraints, ranges, or examples. No error handling guidance. No idempotency markers. Tools accept free-form strings with no validation hints. The ToolRunner.ts file shows permission gating logic (classifyToolRisk), but this is not reflected in tool annotations or schema metadata. Overall, definitions are skeletal, sufficient to invoke but insufficient for reliable LLM reasoning.
Generate a chart/graph visualization from data using Gemini
Generate an image from a text prompt using Gemini's multimodal output
Input schemas lack type definitions. Parameters 'prompt' and 'description' are listed but no JSON Schema types (string, number, object) are visible in the tool registration.
No output schema documented. LLMs cannot predict what fields to expect from tool results (e.g., does image-generation return a file path, base64 data, or a URL?). Agents cannot chain results to downstream tools.
Parameter descriptions are generic and lack format/constraint hints. 'Text prompt describing the image' does not specify length limits, character restrictions, or when to use 'prompt' vs 'description'. LLMs cannot validate inputs before calling.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | 2026-07-28+ | v2 |
No error handling guidance. GeminiTool.ts throws generic errors ('Gemini did not return an image') with no recovery hints. LLMs cannot determine if the error is retryable, user-fixable, or fatal.
Permission gating logic exists in ToolRunner (classifyToolRisk) but is not exposed via tool annotations (destructiveHint). LLMs cannot see that these tools are WRITE-risk and may require confirmation.