R6A-R4-R8 front matter JSON output and bounded schema retry

core/pipeline_agent.py

@@ -587,6 +587,7 @@ class SSciPipelineAgent:
         temp,
         stream,
         reasoning_effort,
+        response_format=None,
     ):
         payload = {
             "model": model_name,
@@ -609,6 +610,9 @@ class SSciPipelineAgent:
                 min(float(temp), 2.0),
             )
 
+        if response_format is not None:
+            payload["response_format"] = dict(response_format)
+
         return payload
 
     @staticmethod
@@ -729,6 +733,7 @@ class SSciPipelineAgent:
         thinking=None,
         reasoning_effort=None,
         allow_fallback=None,
+        response_format=None,
     ):
         (
             requested_model,
@@ -779,6 +784,7 @@ class SSciPipelineAgent:
                 temp=temp,
                 stream=False,
                 reasoning_effort=resolved_effort,
+                response_format=response_format,
             )
             start_time = time.time()
 
@@ -1384,6 +1390,58 @@ Return a concise PROVISIONAL RESEARCH LOGIC CONTRACT in plain English prose. It
             )
         return text
 
+    def _generate_structured_front_matter(
+        self,
+        system_prompt,
+        user_prompt,
+        *,
+        phase,
+        max_tokens,
+        validator,
+    ):
+        """Generate one front-matter object with one bounded content-only retry."""
+        base_prompt = str(user_prompt)
+        current_prompt = base_prompt
+        last_error = None
+        for local_attempt in range(2):
+            # API/auth/balance/cancellation/transport behavior remains owned by
+            # call_deepseek(). Those failures are deliberately not caught here.
+            response = self.call_deepseek(
+                system_prompt,
+                current_prompt,
+                max_tokens=max_tokens,
+                temp=0.0,
+                model_role='writing',
+                thinking=False,
+                allow_fallback=False,
+                response_format={"type": "json_object"},
+            )
+            try:
+                parsed = self._extract_json_from_text(response)
+                if not isinstance(parsed, dict):
+                    raise DeepSeekRequestError(f'{phase} front matter was not valid JSON.')
+                if not isinstance(parsed.get('abstract'), str) or not str(parsed.get('abstract') or '').strip():
+                    raise DeepSeekRequestError(f'{phase} front matter abstract must be a non-empty string.')
+                raw_keywords = parsed.get('keywords')
+                if not isinstance(raw_keywords, list) or not all(isinstance(item, str) for item in raw_keywords):
+                    raise DeepSeekRequestError(f'{phase} front matter keywords must be a JSON array of strings.')
+                return validator(parsed)
+            except DeepSeekRequestError as exc:
+                last_error = exc
+                if local_attempt >= 1:
+                    raise
+                logging.warning(
+                    'Front matter structured-output local retry | phase=%s | attempt=%s/2 | reason=%s',
+                    phase, local_attempt + 1, self._safe_source_error(exc),
+                )
+                current_prompt = base_prompt + (
+                    '\n\nLOCAL VALIDATION RETRY: The previous response failed local JSON/schema/content validation. '
+                    'Return a fresh complete JSON object only, with exactly a non-empty string field "abstract" '
+                    'and a JSON array field "keywords" containing 4-6 distinct non-empty academic terms. '
+                    'Do not describe the correction and do not wrap the JSON in Markdown.'
+                )
+        raise DeepSeekRequestError(f'{phase} front matter validation failed: {last_error}')
+
     def generate_working_front_matter(
         self, topic, keywords, research_logic_contract, synthesis_result, data_evidence='', mode='amj'
     ):
@@ -1406,16 +1464,15 @@ USER KEYWORD SEEDS: {keywords}
 <FROZEN_RESEARCH_EVIDENCE>
 {evidence}
 </FROZEN_RESEARCH_EVIDENCE>"""
-        response = self.call_deepseek(
+        def validate_working(parsed):
+            abstract = self._validate_working_abstract(parsed.get('abstract'))
+            working_keywords = self._front_matter_keywords(parsed.get('keywords'))
+            return {'abstract': abstract, 'keywords': working_keywords}
+
+        return self._generate_structured_front_matter(
             'You are a senior management-journal editor creating provisional front matter without anticipating results.',
-            prompt, max_tokens=1100, temp=0.0, model_role='writing', thinking=False, allow_fallback=False,
+            prompt, phase='Working', max_tokens=1100, validator=validate_working,
         )
-        parsed = self._extract_json_from_text(response)
-        if not isinstance(parsed, dict):
-            raise DeepSeekRequestError('Working front matter was not valid JSON.')
-        abstract = self._validate_working_abstract(parsed.get('abstract'))
-        working_keywords = self._front_matter_keywords(parsed.get('keywords'))
-        return {'abstract': abstract, 'keywords': working_keywords}
 
     def generate_final_front_matter(self, topic, keywords, manuscript, data_evidence='', mode='amj'):
         """Generate final Abstract + Keywords only from the frozen completed body/evidence."""
@@ -1433,16 +1490,18 @@ USER KEYWORD SEEDS: {keywords}
 <COMPLETED_FROZEN_MANUSCRIPT>
 {body}
 </COMPLETED_FROZEN_MANUSCRIPT>"""
-        response = self.call_deepseek(
+        def validate_final(parsed):
+            abstract = re.sub(r'(?im)^#+\s*abstract\s*', '', str(parsed.get('abstract') or '')).strip()
+            if self._front_matter_has_citation(abstract):
+                raise DeepSeekRequestError('Final Abstract must not contain citations.')
+            final_keywords = self._front_matter_keywords(parsed.get('keywords'))
+            return {'abstract': abstract, 'keywords': final_keywords}
+
+        structured = self._generate_structured_front_matter(
             'You are a senior journal editor finalizing front matter from frozen manuscript evidence.',
-            prompt, max_tokens=1200, temp=0.0, model_role='writing', thinking=False, allow_fallback=False,
+            prompt, phase='Final', max_tokens=1200, validator=validate_final,
         )
-        parsed = self._extract_json_from_text(response)
-        if not isinstance(parsed, dict):
-            raise DeepSeekRequestError('Final front matter was not valid JSON.')
-        abstract = re.sub(r'(?im)^#+\s*abstract\s*', '', str(parsed.get('abstract') or '')).strip()
-        if self._front_matter_has_citation(abstract):
-            raise DeepSeekRequestError('Final Abstract must not contain citations.')
+        abstract = structured['abstract']
         abstract = self.polish_academic_section(abstract, 'Abstract', abstract=True, mode=mode)
         if self._front_matter_has_citation(abstract):
             raise DeepSeekRequestError('Final Abstract revision introduced a citation.')
@@ -1455,8 +1514,7 @@ USER KEYWORD SEEDS: {keywords}
                 'Final Abstract length diagnostic | words=%s | preferred=80-%s | nonfatal=true',
                 final_word_count, max_words,
             )
-        final_keywords = self._front_matter_keywords(parsed.get('keywords'))
-        return {'abstract': abstract, 'keywords': final_keywords}
+        return {'abstract': abstract, 'keywords': structured['keywords']}
 
     # Backward-compatible name retained for old checkpoints/tests.
     def stream_generate_anchors(self, topic, keywords, synthesis_result, refs_text, data_evidence="", mode="amj"):

static/previews/AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA.json

@@ -0,0 +1 @@
+{"task_id": "task_r6a_r4_r4_progress", "request_hash": "e954898a5de4d01e361ec143cb36e8d46838844310afe6e74a5b9448335eb673", "execution_epoch": 1, "stage": "", "text": "", "complete": false, "target_stage": "translation_qa:Findings", "target_text": "# Кіріспе\n\nДайын.", "target_complete": false, "target_english_commit_text": "", "target_progress": {"qa_passed": 0, "repair_pending": 1, "queued": 2, "in_flight": 1, "total": 3}, "updated_at": "2026-08-17T05:08:49.171257+00:00"}