Block truncated/lossy file regenerations from shipping

agent-core/src/agents/base.ts

@@ -201,7 +201,9 @@ export class BackendAgent extends BaseAgent {
     // Single-shot with key files pre-loaded. Avoids tool-call overhead while
     // still giving the model the real source code it needs to match patterns.
     // Llama 4 Scout (30K TPM) keeps this in a separate rate-limit bucket from PM/Architect.
-    return callLLM(BACKEND_SYSTEM, input, SCOUT_MODEL, 4096);
+    // 8192 tokens: a full-file rewrite easily exceeds 4096, and a cut-off response
+    // produces invalid JSON (dropped silently) or a truncated file (shipped as-is).
+    return callLLM(BACKEND_SYSTEM, input, SCOUT_MODEL, 8192);
   }
 }
 
@@ -243,7 +245,8 @@ export class FrontendAgent extends BaseAgent {
 
   async run(input: string, context: ExecutionContext): Promise<string> {
     this.log(`Generating frontend code for ${context.repository}`);
-    return callLLM(FRONTEND_SYSTEM, input, SCOUT_MODEL, 4096);
+    // 8192 tokens — see BackendAgent: a truncated full-file response ships broken.
+    return callLLM(FRONTEND_SYSTEM, input, SCOUT_MODEL, 8192);
   }
 }
 

agent-core/src/index.ts

@@ -358,7 +358,7 @@ export class AgentCoordinator {
         ],
         context.repoContext
       );
-      const staticValidation = validateChanges(codeChanges);
+      const staticValidation = validateChanges(codeChanges, context.repoContext?.keyFiles);
       const staticSection = formatStaticResults(staticValidation);
 
       // DevOps only needs the task, the plan, and the generated files — not the

agent-core/src/llm/client.ts

@@ -37,6 +37,8 @@ export interface LLMToolCall {
 export interface ChatResultMessage {
   content: string | null;
   tool_calls?: LLMToolCall[];
+  /** Why generation stopped. 'length' means the response was cut off at max_tokens. */
+  finish_reason?: string | null;
 }
 
 export interface ChatRequest {
@@ -72,7 +74,7 @@ const chatCompletion: ChatTransport = async (req) => {
   while (attempt <= MAX_RETRIES) {
     try {
       const response = await axios.post<{
-        choices: Array<{ message: ChatResultMessage }>;
+        choices: Array<{ message: ChatResultMessage; finish_reason?: string | null }>;
       }>(`${GROQ_BASE_URL}/chat/completions`, req, {
         headers: {
           Authorization: `Bearer ${getApiKey()}`,
@@ -81,11 +83,14 @@ const chatCompletion: ChatTransport = async (req) => {
         timeout: 120_000,
       });
 
-      const message = response.data.choices[0]?.message;
+      const choice = response.data.choices[0];
+      const message = choice?.message;
       if (!message) {
         throw new Error('Empty response from Groq');
       }
-      return message;
+      // Surface the choice-level finish_reason on the message so callers can tell
+      // a complete response from one cut off at max_tokens.
+      return { ...message, finish_reason: choice.finish_reason ?? message.finish_reason };
     } catch (err) {
       const axiosErr = err as AxiosError;
       const status = axiosErr.response?.status;
@@ -131,6 +136,12 @@ export async function callLLM(
       { role: 'user', content: userMessage },
     ],
   });
+  if (message.finish_reason === 'length') {
+    console.warn(
+      `[LLM] Response cut off at max_tokens=${maxTokens} (finish_reason=length). ` +
+        `Output is incomplete — raise max_tokens or reduce scope. Model: ${model}.`
+    );
+  }
   return message.content ?? '';
 }
 

agent-core/src/validation/static.ts

@@ -28,7 +28,10 @@ function extname(basename: string): string {
  * failure modes LLMs actually produce: malformed JSON, syntax errors, empty or
  * truncated content, and unsafe paths. Real ground truth to back the DevOps verdict.
  */
-export function validateChanges(changes: CodeChange[]): StaticValidationResult {
+export function validateChanges(
+  changes: CodeChange[],
+  originals: Record<string, string> = {}
+): StaticValidationResult {
   const issues: ValidationIssue[] = [];
 
   for (const change of changes) {
@@ -61,6 +64,33 @@ export function validateChanges(changes: CodeChange[]): StaticValidationResult {
       issues.push({ path, severity: 'error', message: 'File content appears truncated (contains a truncation marker).' });
     }
 
+    // Lossy-regeneration check — when an agent modifies an EXISTING file, its
+    // output should not be a shrunken copy of the original. The failure mode this
+    // catches (and how it shipped a broken README before): the model echoes the
+    // file it was shown and trails off partway through, emitting syntactically
+    // valid output that is really the original truncated mid-document. Markdown
+    // and other non-code files get no syntax check, so this is their only guard.
+    const original = originals[path] ?? originals[normalized];
+    if (original != null) {
+      const o = original.trimEnd();
+      const n = content.trimEnd();
+      // Exact prefix of the original but shorter → an echo that stopped early.
+      if (n.length < o.length && o.startsWith(n)) {
+        issues.push({
+          path,
+          severity: 'error',
+          message: `Modified file is a truncated copy of the original (${n.length}/${o.length} chars, exact prefix) — the agent echoed the existing file and stopped early.`,
+        });
+      } else if (o.length >= 400 && n.length < o.length * 0.5) {
+        // Not a prefix, but lost more than half its content — likely lossy.
+        issues.push({
+          path,
+          severity: 'error',
+          message: `Modified file is ${Math.round((1 - n.length / o.length) * 100)}% shorter than the original (${n.length}/${o.length} chars) — likely truncated or lossy regeneration.`,
+        });
+      }
+    }
+
     // Type-specific syntax checks.
     const ext = extname(basename);
     if (ext === '.json') {