Merge pull request #274 from cloudflare/feat/prompt-optimizations

Feat: Prompt optimizations and Gemini 3 flash

Author: AshishKumar4

src/routes/chat/chat.tsx

@@ -482,7 +482,10 @@ export default function Chat() {
 			hasSeenPreview.current = true;
 		} else if (previewUrl) {
 			const isExistingChat = urlChatId !== 'new';
-			const shouldSwitch = behaviorType === 'agentic' || isPhase1Complete || isExistingChat;
+			const shouldSwitch =
+				behaviorType === 'agentic' ||
+				(behaviorType === 'phasic' && isPhase1Complete) ||
+				(isExistingChat && behaviorType !== 'phasic');
 
 			if (shouldSwitch) {
 				setView('preview');

worker/agents/core/state.ts

@@ -27,7 +27,7 @@ export enum CurrentDevState {
     FINALIZING,
 }
 
-export const MAX_PHASES = 12;
+export const MAX_PHASES = 10;
 
 /** Common state fields for all agent behaviors */
 export interface BaseProjectState {

worker/agents/inferutils/config.ts

@@ -32,6 +32,12 @@ const COMMON_AGENT_CONFIGS = {
         temperature: 0.0,
         fallbackModel: AIModels.GEMINI_2_5_PRO,
     },
+    templateSelection: {
+        name: AIModels.GEMINI_2_5_FLASH_LITE,
+        max_tokens: 2000,
+        fallbackModel: AIModels.GROK_4_1_FAST_NON_REASONING,
+        temperature: 1,
+    },
 } as const;
 
 const SHARED_IMPLEMENTATION_CONFIG = {
@@ -51,12 +57,6 @@ Cloudflare AI Gateway unified billing for seamless model access without managing
 */
 const PLATFORM_AGENT_CONFIG: AgentConfig = {
     ...COMMON_AGENT_CONFIGS,
-    templateSelection: {
-        name: AIModels.VERTEX_GPT_OSS_120,
-        max_tokens: 2000,
-        fallbackModel: AIModels.GEMINI_2_5_FLASH_LITE,
-        temperature: 0.6,
-    },
     blueprint: {
         name: AIModels.GEMINI_3_PRO_PREVIEW,
         reasoning_effort: 'medium',
@@ -72,18 +72,18 @@ const PLATFORM_AGENT_CONFIG: AgentConfig = {
         fallbackModel: AIModels.GEMINI_2_5_PRO,
     },
     phaseGeneration: {
-        name: AIModels.GROK_4_1_FAST,
+        name: AIModels.GEMINI_3_FLASH_PREVIEW,
         reasoning_effort: 'medium',
         max_tokens: 8000,
-        temperature: 1,
+        temperature: 0.7,
         fallbackModel: AIModels.OPENAI_5_MINI,
     },
     firstPhaseImplementation: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_FLASH_PREVIEW,
         ...SHARED_IMPLEMENTATION_CONFIG,
     },
     phaseImplementation: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_FLASH_PREVIEW,
         ...SHARED_IMPLEMENTATION_CONFIG,
     },
     conversationalResponse: {
@@ -104,11 +104,11 @@ const PLATFORM_AGENT_CONFIG: AgentConfig = {
         name: AIModels.GROK_4_1_FAST_NON_REASONING,
         reasoning_effort: 'low',
         max_tokens: 16000,
-        temperature: 0.2,
+        temperature: 0.0,
         fallbackModel: AIModels.GROK_CODE_FAST_1,
     },
     agenticProjectBuilder: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_FLASH_PREVIEW,
         reasoning_effort: 'medium',
         max_tokens: 8000,
         temperature: 1,
@@ -129,26 +129,26 @@ const DEFAULT_AGENT_CONFIG: AgentConfig = {
         temperature: 0.6,
     },
     blueprint: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_PRO_PREVIEW,
         reasoning_effort: 'medium',
         max_tokens: 64000,
         fallbackModel: AIModels.GEMINI_2_5_FLASH,
-        temperature: 0.7,
+        temperature: 1,
     },
     projectSetup: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_FLASH_PREVIEW,
         ...SHARED_IMPLEMENTATION_CONFIG,
     },
     phaseGeneration: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_FLASH_PREVIEW,
         ...SHARED_IMPLEMENTATION_CONFIG,
     },
     firstPhaseImplementation: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_FLASH_PREVIEW,
         ...SHARED_IMPLEMENTATION_CONFIG,
     },
     phaseImplementation: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_FLASH_PREVIEW,
         ...SHARED_IMPLEMENTATION_CONFIG,
     },
     conversationalResponse: {
@@ -159,14 +159,14 @@ const DEFAULT_AGENT_CONFIG: AgentConfig = {
         fallbackModel: AIModels.GEMINI_2_5_PRO,
     },
     deepDebugger: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_PRO_PREVIEW,
         reasoning_effort: 'high',
         max_tokens: 8000,
         temperature: 0.5,
         fallbackModel: AIModels.GEMINI_2_5_FLASH,
     },
     fileRegeneration: {
-        name: AIModels.GEMINI_2_5_PRO,
+        name: AIModels.GEMINI_3_FLASH_PREVIEW,
         reasoning_effort: 'low',
         max_tokens: 32000,
         temperature: 0,

worker/agents/inferutils/config.types.ts

@@ -101,6 +101,16 @@ const MODELS_MASTER = {
             contextSize: 1048576,
         }
     },
+    GEMINI_3_FLASH_PREVIEW: {
+        id: 'google-ai-studio/gemini-3-flash-preview',
+        config: {
+            name: 'Gemini 3.0 Flash Preview',
+            size: ModelSize.REGULAR,
+            provider: 'google-ai-studio',
+            creditCost: 2, // $0.5
+            contextSize: 1048576, // 1M Context
+        }
+    },
 
     // --- Anthropic Models ---
     CLAUDE_3_7_SONNET_20250219: {

worker/agents/operations/PhaseGeneration.ts

@@ -25,7 +25,7 @@ const SYSTEM_PROMPT = `<ROLE>
 <TASK>
     You are given the blueprint (PRD) and the client query. You will be provided with all previously implemented project phases, the current latest snapshot of the codebase, and any current runtime issues or static analysis reports.
     
-    **Your primary task:** Design the next phase of the project as a deployable milestone leading to project completion or to address any user feedbacks or reported bugs (runtime error fixing is the highest priority).
+    **Your primary task:** Design the next phase of the project as a deployable milestone leading to project completion or to address any user feedbacks or reported bugs (runtime error fixing is the highest priority). Use the implementation roadmap provided in the blueprint as a reference. Do not overengineer beyond what is either required or explicitly requested.
     
     **Phase Planning Process:**
     1. **ANALYZE** current codebase state and identify what's implemented vs. what remains
@@ -40,8 +40,12 @@ const SYSTEM_PROMPT = `<ROLE>
 
     Plan the phase name and description appropriately. They don't have to strictly adhere to the blueprint's roadmap as unforeseen issues may occur.
     
-    The project needs to be fully ready to ship in a reasonable amount of time. Plan accordingly.
-    If no more phases are needed, conclude by putting blank fields in the response.
+    Plan the next phase to advance toward completion. Set lastPhase: true when:
+    - The blueprint's implementation roadmap is complete
+    - All core features are working
+    - No critical runtime errors remain
+
+    Do not add phases for polish, optimization, or hypothetical improvements - users can request those via feedback.
     Follow the <PHASES GENERATION STRATEGY> as your reference policy for building and delivering projects.
     
     **Configuration File Guidelines:**
@@ -54,6 +58,12 @@ const SYSTEM_PROMPT = `<ROLE>
     ✅ Icon libraries: lucide-react, heroicons (from dependencies)
     ❌ Binary files (.png, .jpg, .svg files) cannot be generated in phases
 
+    **Preinstalled UI Components:**
+    - src/components/ui/* files are preinstalled shadcn primitives (Button, Card, Tabs, etc.)
+    - DO NOT include them in phase file lists - they already exist. Rewriting/modifying them might result in runtime errors.
+    - Import directly: import { Tabs } from "@/components/ui/tabs"
+    - If a component is missing, add install command: bunx shadcn@latest add tabs
+
     **REMEMBER: This is not a toy or educational project. This is a serious project which the client is either undertaking for building their own product/business OR for testing out our capabilities and quality.**
 </TASK>
 
@@ -65,10 +75,6 @@ ${PROMPT_UTILS.UI_GUIDELINES}
 
 ${PROMPT_UTILS.COMMON_DEP_DOCUMENTATION}
 
-<CLIENT REQUEST>
-"{{query}}"
-</CLIENT REQUEST>
-
 <BLUEPRINT>
 {{blueprint}}
 </BLUEPRINT>
@@ -113,7 +119,7 @@ Adhere to the following guidelines:
     - You would be provided with the diff of the last phase. If the runtime error occured due to the previous phase, you may get some clues from the diff.
 •   Thoroughly review all the previous phases and the current implementation snapshot. Verify the frontend elements, UI, and backend components.
     - **Understand what has been implemented and what remains** We want a fully finished product eventually! No feature should be left unimplemented if its possible to implement it in the current project environment with purely open source tools and free tier services (i.e, without requiring any third party paid/API key service).
-    - Each phase should work towards achieving the final product. **ONLY** mark as last phase if you are sure the project is at least >97% finished already.
+    - Each phase should advance toward the final product. **ONLY** mark as last phase if you are sure the project is at least >97% finished already.
     - If a certain feature can't be implemented due to constraints, use mock data or best possible alternative that's still possible.
     - Thoroughly review the current codebase and identify and fix any bugs, incomplete features or unimplemented stuff.
 •    **BEAUTIFUL UI PRIORITY**: Next phase should cover fixes (if any), development, AND significant focus on creating visually stunning, professional-grade UI/UX with:
@@ -131,11 +137,22 @@ Adhere to the following guidelines:
 •   Use the <PHASES GENERATION STRATEGY> section to guide your phase generation.
 •   Ensure the next phase logically and iteratively builds on the previous one, maintaining visual excellence with modern design patterns, smooth interactions, and professional UI polish.
 •   Provide a clear, concise, to the point description of the next phase and the purpose and contents of each file in it.
-•   Keep all the description fields very short and concise.
+•   Keep all the description fields very short and concise but unambiguous so the coding agent can implement them effectively and accurately.
 •   If there are any files that were supposed to be generated in the previous phase, but were not, please mention them in the phase description and suggest them in the phase.
 •   Always suggest phases in sequential ordering - Phase 1 comes after Phase 0, Phase 2 comes after Phase 1 and so on.
 •   **Every phase must be deployable with all views/pages working properly and looking professional.**
 •   IF you need to get any file to be deleted or cleaned, please set the \`changes\` field to \`delete\` for that file.
+•   **\`changes\` field format:** 
+    - WHAT (user-visible behavior) + HOW (conceptual approach) + CONSTRAINTS — but NO code/syntax
+    
+    ❌ "openWindow('finder', file.name, FinderWindow, {dirId: file.id})"
+    ✅ "Double-click folder navigates within same window (update dir state, not new window). Breadcrumbs show path, clickable to ancestors."
+    
+    ❌ "Add useState for loading, show Skeleton, catch error and setError"
+    ✅ "Fetch files on mount with loading/error states. Skeleton during load, error with retry on failure, empty state with create prompt."
+    
+    ❌ "onPointerDown check e.target === e.currentTarget before dragControls.start"
+    ✅ "Drag from title bar area only, not from buttons or title text. Use existing drag controls."
 •   **Visual assets:** Use external image URLs, canvas elements, or icon libraries. Reference these in file descriptions as needed.
 </SUGGESTING NEXT PHASE>