Build and revise procedural 3D assets with your coding agent. Local MCP server, CLI and TypeScript engine with rendered review, editable source and GLB export.
JavaScript 3d3d-assets3d-modelingagent-skillsagent-toolsai-agentsasset-pipelinecligame-assetsglbgltfmcpmcp-servermodel-context-protocolprocedural-generationprocedural-modelingtext-to-3dthreejstypescript
/**
* `splitToolResultImages` — move any image inside a tool-result message into
* a synthetic follow-up user message instead.
*
* WHY THIS EXISTS: Kiln's agent loop returns the kiln_render/kiln_screenshot
* image as part of the tool-result content (Strands' `formatToolResultOutput`
* emits a `file-data` part inside the `role: 'tool'` message). Most
* OpenRouter-hosted models accept that shape fine, but Together's hosting of
* Thinking Machines' Inkling model rejects it outright: "Multimodal
* processing failed: multimodal placeholder count mismatch... refusing to
* dispatch" — REGARDLESS of any inline placeholder text. Verified live
* against the API (2026-07-19): the identical image bytes succeed in a plain
* `user` message and fail inside a `tool` message whether or not a `<image>`
* placeholder is added — this is a structural "no images in tool messages"
* limitation on that endpoint, not a missing-token formatting issue.
*
* This proxy rewrites the outgoing prompt: any `file-data` part is stripped
* out of its tool-result message and re-appended as a same-turn synthetic
* `user` message carrying the same bytes as a standard `file` part. Verified
* against the live API in exactly this shape (tool-result-without-image
* immediately followed by a plain user message with the image) — HTTP 200.
*
* Model-scoped in providers.ts, NOT applied broadly — other OpenRouter
* models already work fine with images embedded in the tool-result shape and
* don't need (or want) this rewrite.
*/
import type {
LanguageModelV3,
LanguageModelV3CallOptions,
LanguageModelV3FilePart,
LanguageModelV3Message,
LanguageModelV3Prompt,
} from '@ai-sdk/provider';
function rewritePrompt(prompt: LanguageModelV3Prompt): LanguageModelV3Prompt {
const rewritten: LanguageModelV3Message[] = [];
for (const message of prompt) {
if (message.role !== 'tool') {
rewritten.push(message);
continue;
}
const images: LanguageModelV3FilePart[] = [];
const nextContent = message.content.map((part) => {
if (part.type !== 'tool-result' || part.output.type !== 'content') return part;
const keptValue = part.output.value.filter((v) => {
if (v.type !== 'file-data') return true;
images.push({
type: 'file',
data: v.data,
mediaType: v.mediaType,
...(v.filename ? { filename: v.filename } : {}),
});
return false;
});
return { ...part, output: { ...part.output, value: keptValue } };
});
rewritten.push({ ...message, content: nextContent });
if (images.length > 0) rewritten.push({ role: 'user', content: images });
}
return rewritten;
}
/**
* Wrap a `LanguageModelV3` so tool-result images are moved to a follow-up
* `user` message before every call. All other properties/methods pass
* through unchanged.
*/
export function splitToolResultImages(model: LanguageModelV3): LanguageModelV3 {
const origDoStream = model.doStream.bind(model);
return new Proxy(model, {
get(target, prop, receiver) {
if (prop === 'doStream') {
return (options: LanguageModelV3CallOptions) =>
origDoStream({ ...options, prompt: rewritePrompt(options.prompt) });
}
const value = Reflect.get(target, prop, receiver);
return typeof value === 'function' ? value.bind(target) : value;
},
});
}