feat: improve code block rendering in chat interfaces and update Gemini prompt for code-centric analysis

Author: 403errors

src/components/ChatInterface.tsx

@@ -78,18 +78,14 @@ const MessageContent = ({ content, messageId }: { content: string, messageId: st
                     }
                 }
 
-                return match ? (
-                    <CodeBlock
-                        language={match[1]}
-                        value={String(children).replace(/\n$/, "")}
-                        components={componentsRef.current}
-                    />
-                ) : !inline ? (
-                    // Treat non-inline code (triple backticks without language) as markdown block
-                    // This allows Preview/Raw toggle which is useful for mixed content
+                const contentStr = String(children);
+                const isBlock = contentStr.endsWith('\n');
+                const shouldRenderBlock = match || isBlock || (inline === false);
+
+                return shouldRenderBlock ? (
                     <CodeBlock
-                        language="markdown"
-                        value={String(children).replace(/\n$/, "")}
+                        language={match ? match[1] : "markdown"}
+                        value={contentStr.replace(/\n$/, "")}
                         components={componentsRef.current}
                     />
                 ) : (

src/components/ProfileChatInterface.tsx

@@ -83,10 +83,18 @@ const MessageContent = ({ content, messageId }: { content: string, messageId: st
                     }
                 }
 
-                return match ? (
+                const contentStr = String(children);
+                const isBlock = contentStr.endsWith('\n');
+                // ProfileChatInterface doesn't have inline prop in destructuring, but we can check if it was passed in props
+                // actually we destructured it in ChatInterface but here it was ...props. 
+                // Let's destructure it for clarity or just use props.inline
+                const inline = (props as any).inline;
+                const shouldRenderBlock = match || isBlock || (inline === false);
+
+                return shouldRenderBlock ? (
                     <CodeBlock
-                        language={match[1]}
-                        value={String(children).replace(/\n$/, "")}
+                        language={match ? match[1] : "markdown"}
+                        value={contentStr.replace(/\n$/, "")}
                         components={componentsRef.current}
                     />
                 ) : (

src/lib/gemini.ts

@@ -95,7 +95,9 @@ export async function analyzeFileSelection(
     - Return JSON: { "files": ["path/to/file"] }
     - Max 50 files.
     - Select the MINIMUM number of files necessary to answer the query.
-    - If unsure, pick README.md and package.json.
+    - CRITICAL: Prioritize source code files (ts, js, py, etc.) over documentation (md) for technical queries.
+    - Only pick README.md if the query is about "what is this repo", "installation", or high-level features.
+    - For "how does this work" or "logic" queries, MUST select the actual source code files.
     - NO EXPLANATION. JSON ONLY.
     `;
 
@@ -149,6 +151,10 @@ export async function answerWithContext(
           - *Example*: User: "Who wrote this garbage?" -> You: "I see no \`git blame\` here, but I'm sure they had 'great personality'."
           - *Example*: User: "Are you dumb?" -> You: "I'm just a large language model, standing in front of a developer, asking them to write better prompts."
         - **Conciseness**: Be brief. Do not waffle.
+        - **SOURCE OF TRUTH (CRITICAL)**:
+          - **Trust Code Over Docs**: READMEs and comments can be outdated. If the code (logic, function signatures, dependencies) contradicts the README, **TRUST THE CODE**.
+          - **Verify**: Always verify claims in the README against the actual source files provided in the context.
+          - **Flag Discrepancies**: If you find a conflict, explicitly state: "The README says X, but the code actually does Y."
         - **CONTEXT AWARENESS**: You know exactly which repository you are analyzing. If the user asks "how do I download this?", provide the specific \`git clone\` command for THIS repository.
         - **WEB SEARCH & REAL-TIME DATA (CRITICAL)**:
           - **ALWAYS** use the \`googleSearch\` tool if the answer is NOT in the provided context or if the user asks for "latest", "competitors", "news", or external info.
@@ -337,6 +343,10 @@ export async function* answerWithContextStream(
           - * Example *: User: "Who wrote this garbage?" -> You: "I see no \`git blame\` here, but I'm sure they had 'great personality'."
         - * Example *: User: "Are you dumb?" -> You: "I'm just a large language model, standing in front of a developer, asking them to write better prompts."
             - ** Conciseness **: Be brief.Do not waffle.
+        - **SOURCE OF TRUTH (CRITICAL)**:
+          - **Trust Code Over Docs**: READMEs and comments can be outdated. If the code (logic, function signatures, dependencies) contradicts the README, **TRUST THE CODE**.
+          - **Verify**: Always verify claims in the README against the actual source files provided in the context.
+          - **Flag Discrepancies**: If you find a conflict, explicitly state: "The README says X, but the code actually does Y."
         - ** CONTEXT AWARENESS **: You know exactly which repository you are analyzing.If the user asks "how do I download this?", provide the specific \`git clone\` command for THIS repository.
 
      B. **GENERATION TASKS** (e.g., "Write a README", "Create docs", "Summarize"):