fix(ci): finalize DeepSeek review retries

.github/scripts/deepseek-common.mjs

@@ -343,9 +343,11 @@ export async function callDeepSeekJson({
 
   const parsed = parseJsonObject(content);
   if (!parsed || typeof parsed !== 'object') {
-    throw new Error(
+    const error = new Error(
       `DeepSeek returned non-JSON content: ${truncate(content, 1000, 'model output')}`
     );
+    error.modelOutput = content;
+    throw error;
   }
 
   return {
@@ -391,32 +393,50 @@ export async function callDeepSeekJsonWithRetries(options) {
     ...requestOptions
   } = options
   let lastError = null
+  let previousModelOutput = null
 
   for (let attempt = 1; attempt <= maxAttempts; attempt += 1) {
     const attemptMaxTokens = attempt === 1 ? maxTokens : Math.max(maxTokens, 16384)
+    const retryInstructions = [
+      'AUTOMATION RETRY NOTE:',
+      '- Return ONLY valid JSON.',
+      `- The JSON MUST include a non-empty string field named "${fieldName}".`,
+      '- Do not wrap the JSON in code fences.',
+      `- Do not return an empty, null, or missing "${fieldName}" field.`,
+    ]
     const attemptPrompt =
       attempt === 1
         ? userPrompt
-        : [
-            userPrompt,
-            '',
-            'AUTOMATION RETRY NOTE:',
-            '- Your previous response was invalid for automation.',
-            '- Return ONLY valid JSON.',
-            `- The JSON MUST include a non-empty string field named "${fieldName}".`,
-            '- Do not wrap the JSON in code fences.',
-            `- Do not return an empty, null, or missing "${fieldName}" field.`,
-          ].join('\n')
-
+        : previousModelOutput
+          ? [
+              'Finalize the prior model analysis below into the requested JSON response.',
+              'Do not repeat the analysis and do not inspect the pull request again.',
+              ...retryInstructions,
+              '',
+              'PRIOR MODEL ANALYSIS:',
+              truncate(previousModelOutput, 60000, 'prior model analysis'),
+            ].join('\n')
+          : [userPrompt, '', ...retryInstructions].join('\n')
+
+    let result = null
     try {
-      const result = await callDeepSeekJson({
+      result = await callDeepSeekJson({
         ...requestOptions,
         maxTokens: attemptMaxTokens,
         userPrompt: attemptPrompt,
       })
       assertNonEmptyParsedString(result.parsed, fieldName)
       return result
     } catch (error) {
+      const modelOutput =
+        typeof result?.content === 'string'
+          ? result.content
+          : typeof error?.modelOutput === 'string'
+            ? error.modelOutput
+            : null
+      if (modelOutput?.trim()) {
+        previousModelOutput = modelOutput
+      }
       lastError = error
       if (attempt >= maxAttempts || !isRetryableDeepSeekOutputError(error)) {
         throw error

tests/deepseek-common.test.ts

@@ -266,7 +266,7 @@ describe('deepseek-common structured response parsing', () => {
               {
                 message: {
                   content: '',
-                  reasoning_content: 'Analysis was truncated before the final JSON object.',
+                  reasoning_content: 'Analysis was truncated before the final JSON object.\n{}',
                   role: 'assistant',
                 },
               },
@@ -302,13 +302,18 @@ describe('deepseek-common structured response parsing', () => {
         effort: 'high',
         model: 'deepseek-v4-flash',
         systemPrompt: 'Review the pull request.',
-        userPrompt: 'Return JSON.',
+        userPrompt: 'ORIGINAL LARGE PR PROMPT',
       })
     ).resolves.toMatchObject({
       parsed: { body: 'No findings.' },
     });
 
     const requestBodies = fetchMock.mock.calls.map(([, init]) => JSON.parse(String(init?.body)));
     expect(requestBodies.map((body) => body.max_tokens)).toEqual([8192, 16384]);
+    expect(requestBodies[1].messages[1].content).toContain('PRIOR MODEL ANALYSIS:');
+    expect(requestBodies[1].messages[1].content).toContain(
+      'Analysis was truncated before the final JSON object.'
+    );
+    expect(requestBodies[1].messages[1].content).not.toContain('ORIGINAL LARGE PR PROMPT');
   });
 });