feat(cad-agent-plugin): add Google Gemini provider support

packages/cad-agent-plugin/package.json

@@ -49,6 +49,7 @@
   },
   "dependencies": {
     "@ai-sdk/anthropic": "^2.0.0",
+    "@ai-sdk/google": "^2.0.0",
     "@ai-sdk/openai": "^2.0.0",
     "@ai-sdk/vue": "^2.0.0",
     "ai": "^5.0.0",

packages/cad-agent-plugin/src/agent/createCadAgent.ts

@@ -54,7 +54,7 @@ export function createCadAgent(settings: LlmSettings) {
     model: createModelFromSettings(settings),
     system: CAD_AGENT_SYSTEM_PROMPT,
     tools: createCadTools(),
-    stopWhen: stepCountIs(10)
+    stopWhen: stepCountIs(20)
   })
 }
 
@@ -136,6 +136,13 @@ export function createAgentChatTransport(
             const { userRequest, referenceImages } =
               extractConversationContext(validatedMessages)
             let verificationAttempts = 0
+            // Verifier memory: feedback from every prior round and the screenshot
+            // taken before the latest edits, so review is progress-aware.
+            const feedbackHistory: string[] = []
+            let previousPreviewDataUrl: string | undefined
+            // Consecutive rounds whose feedback is unchanged from the round
+            // before — a stalled loop we stop early instead of burning attempts.
+            let noProgressRounds = 0
 
             while (!abortSignal?.aborted) {
               workingMessages = await streamAgentRound({
@@ -181,6 +188,10 @@ export function createAgentChatTransport(
                   userRequest,
                   referenceImages,
                   preview.dataUrl,
+                  {
+                    previousFeedback: feedbackHistory,
+                    previousDrawingImage: previousPreviewDataUrl
+                  },
                   abortSignal
                 )
               } catch (error) {
@@ -201,6 +212,21 @@ export function createAgentChatTransport(
                 break
               }
 
+              // Detect a stalled loop: feedback unchanged from the prior round.
+              const normalize = (text: string) =>
+                text.trim().toLowerCase().replace(/\s+/g, ' ')
+              const lastFeedback = feedbackHistory[feedbackHistory.length - 1]
+              if (
+                lastFeedback &&
+                normalize(verification.feedback) === normalize(lastFeedback)
+              ) {
+                noProgressRounds += 1
+              } else {
+                noProgressRounds = 0
+              }
+              feedbackHistory.push(verification.feedback)
+              previousPreviewDataUrl = preview.dataUrl
+
               if (verificationAttempts >= MAX_VERIFICATION_ATTEMPTS) {
                 appendAssistantText(
                   write,
@@ -209,6 +235,14 @@ export function createAgentChatTransport(
                 break
               }
 
+              if (noProgressRounds >= 2) {
+                appendAssistantText(
+                  write,
+                  `\n${agentT('verificationNoProgress')}\n\n${verification.feedback.trim()}`
+                )
+                break
+              }
+
               appendAssistantText(
                 write,
                 `\n${agentT('verificationFailed')}\n${verification.feedback.trim()}\n\n${agentT('verificationContinuing')}`

packages/cad-agent-plugin/src/agent/createModel.ts

@@ -1,4 +1,5 @@
 import { createAnthropic } from '@ai-sdk/anthropic'
+import { createGoogleGenerativeAI } from '@ai-sdk/google'
 import { createOpenAI } from '@ai-sdk/openai'
 import type { LanguageModel } from 'ai'
 
@@ -28,6 +29,14 @@ export function createModelFromSettings(settings: LlmSettings): LanguageModel {
     return anthropic(settings.model)
   }
 
+  if (settings.provider === 'google') {
+    const google = createGoogleGenerativeAI({
+      apiKey: settings.apiKey,
+      baseURL: settings.baseUrl || undefined
+    })
+    return google(settings.model)
+  }
+
   if (settings.provider === 'openai') {
     const openai = createOpenAI({
       apiKey: settings.apiKey,

packages/cad-agent-plugin/src/agent/drawingVerifier.ts

@@ -5,7 +5,7 @@ import type { LlmSettings } from '../storage/LlmSettingsStore'
 import { createModelFromSettings } from './createModel'
 
 /** Maximum number of screenshot verification rounds per user message. */
-export const MAX_VERIFICATION_ATTEMPTS = 5
+export const MAX_VERIFICATION_ATTEMPTS = 8
 
 const verificationSchema = z.object({
   passed: z
@@ -25,27 +25,49 @@ const VERIFICATION_SYSTEM_PROMPT = `You verify CAD drawings produced by an autom
 Compare the current drawing screenshot against the user's request and any reference images.
 Focus on overall geometry, topology, proportions, and layout — not exact pixel alignment.
 
-Rules:
-- Ignore dimensions, dimension lines, and dimension text in reference images. The drawing tools cannot create dimensions, so do not require them in the output drawing.
-- Ignore missing annotations or labels unless the user explicitly asked for text entities.
-- Pass when the drawing reasonably captures the requested shape and layout.
-- Fail only when important geometry is missing, wrong, or clearly misaligned.`
+Tool limitations — NEVER fail a drawing for any of these, they are impossible with the available tools:
+- The tools can only draw solid, continuous lines. They cannot produce dashed, dotted, or "hidden" line styles. Treat every edge as solid; do not ask for hidden or dashed lines.
+- The tools cannot create dimensions, dimension lines, extension lines, arrows, or tolerance/GD&T symbols. Ignore these entirely.
+- Ignore missing annotations, labels, hatching styles, or line weights unless the user explicitly asked for text entities.
+
+Memory and progress:
+- You may be shown the issues you reported in previous rounds and the previous screenshot (before the latest edits). Use them to judge PROGRESS.
+- Confirm whether each previously reported issue is now resolved. Do NOT re-raise an issue that has already been fixed.
+- Do NOT introduce new nitpicks each round to keep failing the drawing. Only report issues that genuinely remain and are fixable with the tools above.
+- If the only remaining differences are unfixable (line style, dimensions, cosmetic) or the drawing already captures the requested geometry and layout, set passed = true.
+
+Verdict:
+- Pass when the drawing reasonably captures the requested shape, features, and layout.
+- Fail only when important geometry is missing, wrong, or clearly misaligned AND it can be fixed with solid-line drawing tools.
+- In feedback, list only the concrete remaining issues and the specific fix for each. Keep it short.`
 
 export type DrawingVerificationResult = z.infer<typeof verificationSchema>
 
+/**
+ * Prior-round context that gives the verifier memory across verification rounds.
+ */
+export interface VerificationHistory {
+  /** Feedback strings reported in previous rounds, oldest first. */
+  previousFeedback: string[]
+  /** Screenshot data URL from before the latest round of edits, if available. */
+  previousDrawingImage?: string
+}
+
 /**
  * Sends the current drawing screenshot to the LLM for visual verification.
  *
  * @param settings - LLM provider configuration (must support vision).
  * @param userRequest - Original user drawing request text.
  * @param referenceImages - User-attached reference image data URLs, if any.
  * @param drawingImage - PNG data URL of the current drawing.
+ * @param history - Previous rounds' feedback and screenshot for progress-aware review.
  */
 export async function verifyDrawing(
   settings: LlmSettings,
   userRequest: string,
   referenceImages: string[],
   drawingImage: string,
+  history?: VerificationHistory,
   abortSignal?: AbortSignal
 ): Promise<DrawingVerificationResult> {
   const model = createModelFromSettings(settings)
@@ -68,6 +90,25 @@ export async function verifyDrawing(
     }
   }
 
+  if (history && history.previousFeedback.length > 0) {
+    userContent.push({
+      type: 'text',
+      text:
+        'Issues you reported in previous rounds (oldest first). Check whether each is now resolved and do not re-raise fixed or unfixable ones:\n' +
+        history.previousFeedback
+          .map((f, i) => `Round ${i + 1}: ${f.trim()}`)
+          .join('\n')
+    })
+  }
+
+  if (history?.previousDrawingImage) {
+    userContent.push({
+      type: 'text',
+      text: 'Previous drawing screenshot (before the latest edits), for comparing progress:'
+    })
+    userContent.push({ type: 'image', image: history.previousDrawingImage })
+  }
+
   userContent.push({
     type: 'text',
     text: 'Current drawing screenshot to verify:'
@@ -97,9 +138,9 @@ export function buildVerificationFeedbackMessage(
   feedback: string
 ): string {
   return [
-    `Drawing verification failed (attempt ${attempt}/${maxAttempts}).`,
+    `Drawing verification (attempt ${attempt}/${maxAttempts}) found remaining issues:`,
     feedback.trim(),
-    'Please update the drawing to address these issues.'
+    'Fix ONLY these issues. Do not delete or redraw geometry that is already correct — modify or add just what is needed, using the entityIds you recorded. The tools draw solid continuous lines only, so do not attempt hidden, dashed, or dimension lines.'
   ]
     .filter(Boolean)
     .join('\n\n')

packages/cad-agent-plugin/src/i18n/en.ts

@@ -17,6 +17,7 @@ export const agentEn = {
     'Uses an OpenAI-compatible vision endpoint (default: SiliconFlow). For self-hosted vLLM, set Base URL to your server, e.g. http://127.0.0.1:8000/v1.',
   providerOpenai: 'OpenAI',
   providerAnthropic: 'Anthropic',
+  providerGoogle: 'Google Gemini',
   providerOpenaiCompatible: 'OpenAI Compatible',
   baseUrl: 'Base URL',
   model: 'Model',

packages/cad-agent-plugin/src/i18n/tr.ts

@@ -17,6 +17,7 @@ export const agentTr = {
     'OpenAI uyumlu bir görsel uç nokta kullanır (varsayılan: SiliconFlow). Kendi barındırdığınız vLLM için Taban URL alanını sunucunuza ayarlayın, ör. http://127.0.0.1:8000/v1.',
   providerOpenai: 'OpenAI',
   providerAnthropic: 'Anthropic',
+  providerGoogle: 'Google Gemini',
   providerOpenaiCompatible: 'OpenAI Uyumlu',
   baseUrl: 'Taban URL',
   model: 'Model',

packages/cad-agent-plugin/src/i18n/zh.ts

@@ -17,6 +17,7 @@ export const agentZh = {
     '使用 OpenAI 兼容的视觉接口（默认：SiliconFlow）。自托管 vLLM 时请将接口地址改为本地服务，例如 http://127.0.0.1:8000/v1。',
   providerOpenai: 'OpenAI',
   providerAnthropic: 'Anthropic',
+  providerGoogle: 'Google Gemini',
   providerOpenaiCompatible: 'OpenAI 兼容',
   baseUrl: '接口地址',
   model: '模型',

packages/cad-agent-plugin/src/storage/LlmSettingsStore.ts

@@ -4,6 +4,7 @@ import { decryptApiKey, encryptApiKey } from './apiKeyCrypto'
 export type LlmProviderId =
   | 'openai'
   | 'anthropic'
+  | 'google'
   | 'openai-compatible'
   | 'deepseek'
   | 'deepseek-vl'
@@ -51,6 +52,10 @@ export const PROVIDER_DEFAULTS: Record<
     baseUrl: 'https://api.anthropic.com/v1',
     model: 'claude-3-5-haiku-latest'
   },
+  google: {
+    baseUrl: 'https://generativelanguage.googleapis.com/v1beta',
+    model: 'gemini-2.5-flash'
+  },
   'openai-compatible': {
     baseUrl: 'https://api.openai.com/v1',
     model: 'gpt-4o-mini'

packages/cad-agent-plugin/src/storage/modelCatalog.ts

@@ -70,6 +70,38 @@ export const PROVIDER_MODEL_OPTIONS: Record<
       supportsVision: true
     }
   ],
+  google: [
+    {
+      value: 'gemini-2.5-pro',
+      labelEn: 'Gemini 2.5 Pro',
+      labelZh: 'Gemini 2.5 Pro',
+      supportsVision: true
+    },
+    {
+      value: 'gemini-2.5-flash',
+      labelEn: 'Gemini 2.5 Flash',
+      labelZh: 'Gemini 2.5 Flash',
+      supportsVision: true
+    },
+    {
+      value: 'gemini-2.0-flash',
+      labelEn: 'Gemini 2.0 Flash',
+      labelZh: 'Gemini 2.0 Flash',
+      supportsVision: true
+    },
+    {
+      value: 'gemini-1.5-pro',
+      labelEn: 'Gemini 1.5 Pro',
+      labelZh: 'Gemini 1.5 Pro',
+      supportsVision: true
+    },
+    {
+      value: 'gemini-1.5-flash',
+      labelEn: 'Gemini 1.5 Flash',
+      labelZh: 'Gemini 1.5 Flash',
+      supportsVision: true
+    }
+  ],
   deepseek: [
     {
       value: 'deepseek-v4-flash',

packages/cad-agent-plugin/src/ui/AgentChatPanel.vue

@@ -461,6 +461,7 @@ function isVerificationMessage(message: UIMessage): boolean {
           <option value="deepseek-vl">{{ labels.providerDeepseekVl }}</option>
           <option value="openai">{{ labels.providerOpenai }}</option>
           <option value="anthropic">{{ labels.providerAnthropic }}</option>
+          <option value="google">{{ labels.providerGoogle }}</option>
           <option value="openai-compatible">
             {{ labels.providerOpenaiCompatible }}
           </option>