Фиксируем состояние

src/app/modules/agent/code_qa_runtime/answer_policy.py

@@ -4,6 +4,7 @@ from dataclasses import dataclass
 
 from app.modules.rag.code_qa_pipeline.evidence_gate import EvidenceGateDecision
 from app.modules.rag.intent_router_v2.models import IntentRouterResult
+from app.modules.agent.code_qa_runtime.short_answer_formatter import CodeQaShortAnswerFormatter
 
 
 @dataclass(slots=True, frozen=True)
@@ -16,6 +17,9 @@ class CodeQaPolicyDecision:
 
 
 class CodeQaAnswerPolicy:
+    def __init__(self, formatter: CodeQaShortAnswerFormatter | None = None) -> None:
+        self._formatter = formatter or CodeQaShortAnswerFormatter()
+
     def decide(
         self,
         *,
@@ -29,34 +33,39 @@ class CodeQaAnswerPolicy:
             target = path_scope[0] if path_scope else "запрошенный файл"
             return CodeQaPolicyDecision(
                 answer_mode="not_found",
-                answer=f"Файл {target} не найден.",
+                answer=self._formatter.open_file_not_found(target),
                 should_call_llm=False,
                 branch="open_file_not_found",
                 reason="path_scope_empty",
             )
-        if sub_intent == "EXPLAIN" and symbol_resolution.status in {"not_found", "ambiguous"}:
+        if sub_intent == "EXPLAIN" and symbol_resolution.status == "not_found":
             return CodeQaPolicyDecision(
-                answer_mode="degraded",
-                answer=self._symbol_message(symbol_resolution.status, symbol_resolution.alternatives),
+                answer_mode="not_found",
+                answer=self._formatter.entity_not_found(self._target_label(router_result), symbol_resolution.alternatives),
                 should_call_llm=False,
-                branch="explain_unresolved_symbol",
-                reason=f"symbol_resolution_{symbol_resolution.status}",
+                branch="explain_not_found",
+                reason="symbol_resolution_not_found",
+            )
+        if sub_intent == "EXPLAIN" and symbol_resolution.status == "ambiguous":
+            return CodeQaPolicyDecision(
+                answer_mode="ambiguous",
+                answer=self._formatter.entity_ambiguous(self._target_label(router_result), symbol_resolution.alternatives),
+                should_call_llm=False,
+                branch="explain_ambiguous_symbol",
+                reason="symbol_resolution_ambiguous",
             )
         if not gate_decision.passed:
             answer_mode = "insufficient" if "insufficient_evidence" in gate_decision.failure_reasons else "degraded"
             reason = gate_decision.failure_reasons[0] if gate_decision.failure_reasons else "evidence_gate_failed"
             return CodeQaPolicyDecision(
                 answer_mode=answer_mode,
-                answer=gate_decision.degraded_message,
+                answer=self._formatter.insufficient(gate_decision.degraded_message),
                 should_call_llm=False,
                 branch="evidence_gate_short_circuit",
                 reason=reason,
             )
         return CodeQaPolicyDecision(answer_mode="normal", branch="normal_answer", reason="evidence_sufficient")
 
-    def _symbol_message(self, status: str, alternatives: list[str]) -> str:
-        if status == "ambiguous" and alternatives:
-            return f"Сущность не удалось однозначно разрешить. Близкие варианты: {', '.join(alternatives[:3])}."
-        if alternatives:
-            return f"Сущность не найдена в доступном коде. Ближайшие варианты: {', '.join(alternatives[:3])}."
-        return "Сущность не найдена в доступном коде."
+    def _target_label(self, router_result: IntentRouterResult) -> str:
+        candidates = [item.strip() for item in list(router_result.query_plan.symbol_candidates or []) if item and item.strip()]
+        return candidates[0] if candidates else "запрошенная сущность"

src/app/modules/agent/code_qa_runtime/executor.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import logging
+from difflib import SequenceMatcher
 from time import perf_counter
 
 from app.modules.agent.code_qa_runtime.answer_policy import CodeQaAnswerPolicy
@@ -50,6 +51,9 @@ class CodeQaRuntimeExecutor:
     def execute(self, *, user_query: str, rag_session_id: str, files_map: dict[str, dict] | None = None) -> CodeQaFinalResult:
         timings_ms: dict[str, int] = {}
         runtime_trace: list[dict] = []
+        answer_policy_branch = ""
+        decision_reason = ""
+        post_gate_snapshot: dict = {}
         state = CodeQaExecutionState(
             user_query=user_query,
             rag_session_id=rag_session_id,
@@ -76,9 +80,10 @@ class CodeQaRuntimeExecutor:
         started = perf_counter()
         raw_rows = self._retrieve(state)
         timings_ms["retrieval"] = self._elapsed_ms(started)
+        retrieval_report = self._retrieval.consume_retrieval_report() or {}
+        raw_rows, retrieval_report = self._hydrate_entrypoint_sources(state, raw_rows, retrieval_report)
         symbol_resolution = self._resolve_symbol(state.router_result.symbol_resolution.model_dump(), raw_rows)
         state.router_result = state.router_result.model_copy(update={"symbol_resolution": SymbolResolution(**symbol_resolution)})
-        retrieval_report = self._retrieval.consume_retrieval_report()
         state.retrieval_result = build_retrieval_result(raw_rows, retrieval_report, symbol_resolution)
         if state.retrieval_request.sub_intent.upper() == "EXPLAIN" and symbol_resolution.get("status") in {"not_found", "ambiguous"}:
             state.retrieval_result = build_retrieval_result([], retrieval_report, symbol_resolution)
@@ -101,8 +106,10 @@ class CodeQaRuntimeExecutor:
         if state.retrieval_request.sub_intent.upper() == "EXPLAIN" and symbol_resolution.get("status") in {"not_found", "ambiguous"}:
             state.evidence_pack.sufficient = False
             state.evidence_pack.failure_reasons = ["target_not_resolved"]
+        pre_gate_input = self._build_pre_gate_input(state)
+        started = perf_counter()
         gate_decision = evaluate_evidence(state.evidence_pack)
-        timings_ms["pre_evidence_gate"] = 0
+        timings_ms["pre_evidence_gate"] = self._elapsed_ms(started)
         state.answer_mode = "normal" if gate_decision.passed else "degraded"
         state.degraded_message = gate_decision.degraded_message
         runtime_trace.append(
@@ -110,31 +117,51 @@ class CodeQaRuntimeExecutor:
                 "step": "pre_evidence_gate",
                 "status": "passed" if gate_decision.passed else "blocked",
                 "timings_ms": {"pre_evidence_gate": timings_ms["pre_evidence_gate"]},
+                "input": pre_gate_input,
                 "output": {
                     "passed": gate_decision.passed,
                     "failure_reasons": list(gate_decision.failure_reasons),
                     "degraded_message": gate_decision.degraded_message,
+                    "evidence_count": state.evidence_pack.evidence_count,
                 },
             }
         )
         decision = self._answer_policy.decide(router_result=state.router_result, gate_decision=gate_decision)
+        answer_policy_branch = decision.branch
+        decision_reason = decision.reason
         if not decision.should_call_llm:
             state.answer_mode = decision.answer_mode
+            started = perf_counter()
             runtime_trace.append(
                 {
                     "step": "llm",
                     "status": "skipped",
+                    "timings_ms": {"llm": self._elapsed_ms(started)},
                     "output": {
                         "reason": "policy_short_circuit",
                         "answer_mode": decision.answer_mode,
+                        "decision_reason": decision.reason,
+                        "answer_policy_branch": decision.branch,
                     },
                 }
             )
+            started = perf_counter()
+            timings_ms["post_evidence_gate"] = self._elapsed_ms(started)
+            post_gate_snapshot = {
+                "input": {
+                    "answer_mode": decision.answer_mode,
+                    "draft_present": False,
+                    "resolved_target": self._resolved_target(state),
+                },
+                "output": {"reason": "no_draft_answer"},
+            }
             runtime_trace.append(
                 {
                     "step": "post_evidence_gate",
                     "status": "skipped",
-                    "output": {"reason": "no_draft_answer"},
+                    "timings_ms": {"post_evidence_gate": timings_ms["post_evidence_gate"]},
+                    "input": post_gate_snapshot["input"],
+                    "output": post_gate_snapshot["output"],
                 }
             )
             return self._finalize(
@@ -145,13 +172,41 @@ class CodeQaRuntimeExecutor:
                 llm_used=False,
                 timings_ms=timings_ms,
                 runtime_trace=runtime_trace,
+                answer_policy_branch=answer_policy_branch,
+                decision_reason=decision_reason,
+                pre_gate_input=pre_gate_input,
+                gate_decision=gate_decision,
+                post_gate_snapshot=post_gate_snapshot,
             )
         if self._llm is None:
+            answer_policy_branch = "llm_unavailable"
+            decision_reason = "llm_service_missing"
+            started = perf_counter()
             runtime_trace.append(
                 {
                     "step": "llm",
                     "status": "skipped",
-                    "output": {"reason": "llm_unavailable"},
+                    "timings_ms": {"llm": self._elapsed_ms(started)},
+                    "output": {"reason": "llm_unavailable", "answer_policy_branch": answer_policy_branch, "decision_reason": decision_reason},
+                }
+            )
+            started = perf_counter()
+            timings_ms["post_evidence_gate"] = self._elapsed_ms(started)
+            post_gate_snapshot = {
+                "input": {
+                    "answer_mode": state.answer_mode,
+                    "draft_present": False,
+                    "resolved_target": self._resolved_target(state),
+                },
+                "output": {"reason": "no_draft_answer"},
+            }
+            runtime_trace.append(
+                {
+                    "step": "post_evidence_gate",
+                    "status": "skipped",
+                    "timings_ms": {"post_evidence_gate": timings_ms["post_evidence_gate"]},
+                    "input": post_gate_snapshot["input"],
+                    "output": post_gate_snapshot["output"],
                 }
             )
             return self._finalize(
@@ -162,6 +217,11 @@ class CodeQaRuntimeExecutor:
                 llm_used=False,
                 timings_ms=timings_ms,
                 runtime_trace=runtime_trace,
+                answer_policy_branch=answer_policy_branch,
+                decision_reason=decision_reason,
+                pre_gate_input=pre_gate_input,
+                gate_decision=gate_decision,
+                post_gate_snapshot=post_gate_snapshot,
             )
         state.synthesis_input = build_answer_synthesis_input(user_query, state.evidence_pack)
         prompt_name = self._prompt_selector.select(sub_intent=state.retrieval_request.sub_intent, answer_mode=state.answer_mode)
@@ -186,10 +246,29 @@ class CodeQaRuntimeExecutor:
                 "output": {
                     "prompt_name": prompt_name,
                     "answer_preview": draft.answer[:300],
+                    "resolved_target": self._resolved_target(state),
+                    "answer_policy_branch": answer_policy_branch,
+                    "decision_reason": decision_reason,
                 },
             }
         )
-        validation = self._post_gate.validate(answer=draft.answer, answer_mode=state.answer_mode, degraded_message=state.degraded_message)
+        post_gate_input = {
+            "answer_mode": state.answer_mode,
+            "degraded_message": state.degraded_message,
+            "resolved_target": self._resolved_target(state),
+            "draft_answer_preview": draft.answer[:300],
+            "repair_candidate": bool(self._repair is not None),
+        }
+        started = perf_counter()
+        validation = self._post_gate.validate(
+            answer=draft.answer,
+            answer_mode=state.answer_mode,
+            degraded_message=state.degraded_message,
+            sub_intent=state.retrieval_request.sub_intent,
+            user_query=user_query,
+            evidence_pack=state.evidence_pack,
+        )
+        timings_ms["post_evidence_gate"] = self._elapsed_ms(started)
         final_answer = draft.answer
         repair_used = False
         if not validation.passed and self._repair is not None:
@@ -197,22 +276,40 @@ class CodeQaRuntimeExecutor:
             final_answer = self._repair.repair(draft_answer=draft.answer, validation=validation, prompt_payload=prompt_payload)
             repair_used = True
             timings_ms["repair"] = self._elapsed_ms(started)
-            validation = self._post_gate.validate(answer=final_answer, answer_mode=state.answer_mode, degraded_message=state.degraded_message)
-            if not validation.passed and state.degraded_message:
-                final_answer = state.degraded_message
+            started = perf_counter()
+            validation = self._post_gate.validate(
+                answer=final_answer,
+                answer_mode=state.answer_mode,
+                degraded_message=state.degraded_message,
+                sub_intent=state.retrieval_request.sub_intent,
+                user_query=user_query,
+                evidence_pack=state.evidence_pack,
+            )
+            timings_ms["post_evidence_gate_recheck"] = self._elapsed_ms(started)
+            if not validation.passed:
+                final_answer = self._fallback_answer(state)
+                state.answer_mode = self._fallback_mode(state)
+        post_gate_snapshot = {
+            "input": post_gate_input,
+            "output": {
+                "passed": validation.passed,
+                "action": validation.action,
+                "reasons": list(validation.reasons),
+                "repair_used": repair_used,
+                "final_answer_preview": final_answer[:300],
+            },
+        }
         runtime_trace.append(
             {
                 "step": "post_evidence_gate",
                 "status": "passed" if validation.passed else "failed",
                 "timings_ms": {
-                    "post_evidence_gate": 0,
+                    "post_evidence_gate": timings_ms["post_evidence_gate"],
+                    "post_evidence_gate_recheck": timings_ms.get("post_evidence_gate_recheck", 0),
                     "repair": timings_ms.get("repair", 0),
                 },
-                "output": {
-                    "passed": validation.passed,
-                    "reasons": list(validation.reasons),
-                    "repair_used": repair_used,
-                },
+                "input": post_gate_snapshot["input"],
+                "output": post_gate_snapshot["output"],
             }
         )
         return self._finalize(
@@ -224,6 +321,11 @@ class CodeQaRuntimeExecutor:
             validation=validation,
             timings_ms=timings_ms,
             runtime_trace=runtime_trace,
+            answer_policy_branch=answer_policy_branch,
+            decision_reason=decision_reason,
+            pre_gate_input=pre_gate_input,
+            gate_decision=gate_decision,
+            post_gate_snapshot=post_gate_snapshot,
         )
 
     def _retrieve(self, state: CodeQaExecutionState) -> list[dict]:
@@ -257,9 +359,10 @@ class CodeQaRuntimeExecutor:
         exact = next((item for item in found if item in candidates), None)
         if exact:
             return {"status": "resolved", "resolved_symbol": exact, "alternatives": found[:5], "confidence": 0.99}
-        if found:
-            return {"status": "ambiguous", "resolved_symbol": None, "alternatives": found[:5], "confidence": 0.55}
-        return {"status": "not_found", "resolved_symbol": None, "alternatives": [], "confidence": 0.0}
+        close = self._close_matches(candidates, found)
+        if close:
+            return {"status": "ambiguous", "resolved_symbol": None, "alternatives": close[:5], "confidence": 0.55}
+        return {"status": "not_found", "resolved_symbol": None, "alternatives": close[:5], "confidence": 0.0}
 
     def _finalize(
         self,
@@ -272,6 +375,11 @@ class CodeQaRuntimeExecutor:
         validation=None,
         timings_ms: dict[str, int] | None = None,
         runtime_trace: list[dict] | None = None,
+        answer_policy_branch: str = "",
+        decision_reason: str = "",
+        pre_gate_input: dict | None = None,
+        gate_decision=None,
+        post_gate_snapshot: dict | None = None,
     ) -> CodeQaFinalResult:
         diagnostics = build_diagnostics_report(
             router_result=state.router_result,
@@ -280,6 +388,11 @@ class CodeQaRuntimeExecutor:
             evidence_bundle=state.evidence_pack,
             answer_mode=state.answer_mode,
             timings_ms=timings_ms or {},
+            resolved_target=self._resolved_target(state),
+            answer_policy_branch=answer_policy_branch,
+            decision_reason=decision_reason,
+            evidence_gate_input=pre_gate_input or {},
+            post_evidence_gate=post_gate_snapshot or {},
         )
         result = CodeQaFinalResult(
             final_answer=final_answer.strip(),
@@ -287,7 +400,15 @@ class CodeQaRuntimeExecutor:
             repair_used=repair_used,
             llm_used=llm_used,
             draft_answer=draft,
-            validation=validation or self._post_gate.validate(answer=final_answer, answer_mode=state.answer_mode, degraded_message=state.degraded_message),
+            validation=validation
+            or self._post_gate.validate(
+                answer=final_answer,
+                answer_mode=state.answer_mode,
+                degraded_message=state.degraded_message,
+                sub_intent=state.retrieval_request.sub_intent if state.retrieval_request else "",
+                user_query=state.user_query,
+                evidence_pack=state.evidence_pack,
+            ),
             router_result=state.router_result,
             retrieval_request=state.retrieval_request,
             retrieval_result=state.retrieval_result,
@@ -306,4 +427,119 @@ class CodeQaRuntimeExecutor:
         return result
 
     def _elapsed_ms(self, started: float) -> int:
-        return int((perf_counter() - started) * 1000)
+        return max(1, round((perf_counter() - started) * 1000))
+
+    def _build_pre_gate_input(self, state: CodeQaExecutionState) -> dict:
+        evidence = state.evidence_pack
+        retrieval = state.retrieval_result
+        return {
+            "resolved_target": self._resolved_target(state),
+            "sub_intent": state.retrieval_request.sub_intent if state.retrieval_request else None,
+            "target_type": evidence.target_type if evidence else None,
+            "evidence_count": evidence.evidence_count if evidence else 0,
+            "code_chunk_count": len(evidence.code_chunks) if evidence else 0,
+            "entrypoint_count": len(evidence.entrypoints) if evidence else 0,
+            "relation_count": len(evidence.relations) if evidence else 0,
+            "test_evidence_count": len(evidence.test_evidence) if evidence else 0,
+            "symbol_resolution_status": retrieval.symbol_resolution_status if retrieval else None,
+            "path_scope": list(state.retrieval_request.path_scope) if state.retrieval_request else [],
+        }
+
+    def _resolved_target(self, state: CodeQaExecutionState) -> str | None:
+        if state.evidence_pack and state.evidence_pack.resolved_target:
+            return state.evidence_pack.resolved_target
+        if state.retrieval_result and state.retrieval_result.resolved_symbol:
+            return state.retrieval_result.resolved_symbol
+        if state.retrieval_request and state.retrieval_request.path_scope:
+            return state.retrieval_request.path_scope[0]
+        return None
+
+    def _close_matches(self, candidates: list[str], found: list[str]) -> list[str]:
+        ranked: list[tuple[float, str]] = []
+        for candidate in candidates:
+            for item in found:
+                score = SequenceMatcher(None, candidate.lower(), item.lower()).ratio()
+                if score >= 0.52:
+                    ranked.append((score, item))
+        ranked.sort(key=lambda pair: (-pair[0], pair[1]))
+        result: list[str] = []
+        for _, item in ranked:
+            if item not in result:
+                result.append(item)
+        return result
+
+    def _hydrate_entrypoint_sources(
+        self,
+        state: CodeQaExecutionState,
+        raw_rows: list[dict],
+        retrieval_report: dict,
+    ) -> tuple[list[dict], dict]:
+        if not state.retrieval_request or state.retrieval_request.sub_intent.upper() != "FIND_ENTRYPOINTS":
+            return raw_rows, retrieval_report
+        entrypoint_paths = []
+        for row in raw_rows:
+            if str(row.get("layer") or "") != "C3_ENTRYPOINTS":
+                continue
+            path = str(row.get("path") or "").strip()
+            if path and path not in entrypoint_paths:
+                entrypoint_paths.append(path)
+        if not entrypoint_paths:
+            return raw_rows, retrieval_report
+        extra_rows = self._retrieval.retrieve_exact_files(
+            state.rag_session_id,
+            paths=entrypoint_paths[:3],
+            layers=["C0_SOURCE_CHUNKS"],
+            limit=24,
+            query=state.user_query,
+            ranking_profile="entrypoint_source_hydration",
+        )
+        extra_report = self._retrieval.consume_retrieval_report() or {}
+        return self._merge_rows(raw_rows, extra_rows), self._merge_reports(retrieval_report, extra_report)
+
+    def _merge_rows(self, base_rows: list[dict], extra_rows: list[dict]) -> list[dict]:
+        merged: list[dict] = []
+        seen: set[tuple[str, str, str, int | None, int | None]] = set()
+        for row in [*base_rows, *extra_rows]:
+            key = (
+                str(row.get("layer") or ""),
+                str(row.get("path") or ""),
+                str(row.get("title") or ""),
+                row.get("span_start"),
+                row.get("span_end"),
+            )
+            if key in seen:
+                continue
+            seen.add(key)
+            merged.append(row)
+        return merged
+
+    def _merge_reports(self, base: dict, extra: dict) -> dict:
+        merged = dict(base or {})
+        merged["executed_layers"] = list(dict.fromkeys([*(base.get("executed_layers") or []), *(extra.get("executed_layers") or [])]))
+        merged["retrieval_mode_by_layer"] = {**dict(base.get("retrieval_mode_by_layer") or {}), **dict(extra.get("retrieval_mode_by_layer") or {})}
+        merged["top_k_by_layer"] = {**dict(base.get("top_k_by_layer") or {}), **dict(extra.get("top_k_by_layer") or {})}
+        merged["filters_by_layer"] = {**dict(base.get("filters_by_layer") or {}), **dict(extra.get("filters_by_layer") or {})}
+        merged["retrieval_by_layer_ms"] = {**dict(base.get("retrieval_by_layer_ms") or {}), **dict(extra.get("retrieval_by_layer_ms") or {})}
+        merged["fallback"] = dict(extra.get("fallback") or base.get("fallback") or {"used": False, "reason": None})
+        merged["supplemental_requests"] = [*(base.get("supplemental_requests") or []), *(extra.get("requests") or [])]
+        return merged
+
+    def _fallback_mode(self, state: CodeQaExecutionState) -> str:
+        status = str(state.router_result.symbol_resolution.status if state.router_result and state.router_result.symbol_resolution else "")
+        if status == "ambiguous":
+            return "ambiguous"
+        if status == "not_found":
+            return "not_found"
+        return "degraded"
+
+    def _fallback_answer(self, state: CodeQaExecutionState) -> str:
+        symbol_resolution = state.router_result.symbol_resolution if state.router_result else None
+        query_plan = state.router_result.query_plan if state.router_result else None
+        target = next((item for item in list(query_plan.symbol_candidates or []) if item), "запрошенная сущность") if query_plan else "запрошенная сущность"
+        if symbol_resolution and symbol_resolution.status == "ambiguous":
+            return self._answer_policy.decide(router_result=state.router_result, gate_decision=evaluate_evidence(state.evidence_pack)).answer
+        if symbol_resolution and symbol_resolution.status == "not_found":
+            return self._answer_policy.decide(router_result=state.router_result, gate_decision=evaluate_evidence(state.evidence_pack)).answer
+        if state.degraded_message:
+            return state.degraded_message
+        return f"Недостаточно подтверждённых данных для уверенного ответа по {target}."

src/app/modules/agent/code_qa_runtime/post_gate.py

@@ -1,6 +1,25 @@
 from __future__ import annotations
 
+import re
+
 from app.modules.agent.code_qa_runtime.models import CodeQaValidationResult
+from app.modules.rag.code_qa_pipeline.answer_fact_curator import build_curated_answer_facts
+from app.modules.rag.code_qa_pipeline.contracts import EvidenceBundle
+
+_TOKEN_RE = re.compile(r"[a-zA-Zа-яА-Я0-9_/]+")
+_VAGUE_PHRASES = (
+    "ряд аргументов",
+    "имеет responsibilities",
+    "имеет responsibility",
+    "ключевой компонент",
+    "центральный компонент",
+    "играет роль",
+    "управляет системой",
+    "этап пайплайна",
+    "инициализация сервисов",
+    "регистрация основных служб",
+)
+_OPTIMISTIC_TRACE_CLAIMS = ("полностью восстанавливается", "полный поток выполнения", "полностью прослеживается")
 
 
 class CodeQaPostEvidenceGate:
@@ -10,6 +29,9 @@ class CodeQaPostEvidenceGate:
         answer: str,
         answer_mode: str,
         degraded_message: str,
+        sub_intent: str,
+        user_query: str,
+        evidence_pack: EvidenceBundle | None,
     ) -> CodeQaValidationResult:
         normalized = (answer or "").strip()
         if not normalized:
@@ -18,6 +40,189 @@ class CodeQaPostEvidenceGate:
             return CodeQaValidationResult(passed=False, action="repair", reasons=["degraded_answer_missing_guardrail"])
         if answer_mode == "not_found" and "не найден" not in normalized.lower():
             return CodeQaValidationResult(passed=False, action="repair", reasons=["not_found_answer_missing_phrase"])
+        if answer_mode == "ambiguous" and "не удалось однозначно разрешить" not in normalized.lower():
+            return CodeQaValidationResult(passed=False, action="repair", reasons=["ambiguous_answer_missing_phrase"])
         if degraded_message and answer_mode != "normal" and len(normalized) < 24:
             return CodeQaValidationResult(passed=False, action="repair", reasons=["answer_too_short"])
+        if answer_mode != "normal" or evidence_pack is None:
+            return CodeQaValidationResult(passed=True, action="return")
+
+        reasons = self._normal_answer_reasons(normalized.lower(), sub_intent.upper(), user_query, evidence_pack)
+        if reasons:
+            return CodeQaValidationResult(passed=False, action="repair", reasons=_dedupe(reasons))
         return CodeQaValidationResult(passed=True, action="return")
+
+    def _normal_answer_reasons(self, answer: str, sub_intent: str, user_query: str, evidence_pack: EvidenceBundle) -> list[str]:
+        reasons: list[str] = []
+        if sub_intent == "FIND_ENTRYPOINTS":
+            reasons.extend(self._validate_entrypoints(answer, user_query, evidence_pack))
+        elif sub_intent == "EXPLAIN":
+            reasons.extend(self._validate_explain(answer, evidence_pack))
+        elif sub_intent == "ARCHITECTURE":
+            reasons.extend(self._validate_architecture(answer, evidence_pack))
+        elif sub_intent == "TRACE_FLOW":
+            reasons.extend(self._validate_trace_flow(answer, evidence_pack))
+        return reasons
+
+    def _validate_entrypoints(self, answer: str, user_query: str, evidence_pack: EvidenceBundle) -> list[str]:
+        confirmed = [item for item in list(evidence_pack.entrypoints or []) if item.get("http_method") and item.get("route_path")]
+        if not confirmed:
+            return []
+        query_tokens = self._tokens(user_query)
+        matched = [item for item in confirmed if self._tokens(str(item.get("route_path") or "")) & query_tokens]
+        required = matched or confirmed[:1]
+        reasons = []
+        if any(f"{item['http_method']} {item['route_path']}".lower() not in answer for item in required):
+            reasons.append("missing_confirmed_route")
+        if matched and any(token in query_tokens for token in {"health", "/health"}) and "не обнаруж" in answer:
+            reasons.append("contradicts_confirmed_route")
+        return reasons
+
+    def _validate_explain(self, answer: str, evidence_pack: EvidenceBundle) -> list[str]:
+        facts = build_curated_answer_facts(evidence_pack)
+        explain = dict(facts.get("explain") or {})
+        reasons = self._validate_target_focus(answer, evidence_pack)
+        reasons.extend(self._vagueness_reasons(answer, "explain"))
+
+        matches = 0
+        methods = list(explain.get("required_methods") or [])
+        calls = list(explain.get("required_calls") or [])
+        dependencies = list(explain.get("required_dependencies") or [])
+        fields = list(explain.get("required_fields") or [])
+        if methods and not self._mentions_fact_group(answer, methods):
+            reasons.append("missing_concrete_methods")
+        elif methods:
+            matches += 1
+        if calls and not self._mentions_fact_group(answer, calls):
+            reasons.append("missing_concrete_calls")
+        elif calls:
+            matches += 1
+        if dependencies and not self._mentions_fact_group(answer, dependencies):
+            reasons.append("missing_concrete_dependencies")
+        elif dependencies:
+            matches += 1
+        if fields and self._mentions_fact_group(answer, fields):
+            matches += 1
+        if (methods or calls or dependencies or fields) and matches == 0:
+            reasons.append("too_vague_for_explain")
+        if self._semantic_leakage(answer, facts, has_concrete_support=matches > 0):
+            reasons.append("semantic_labels_without_code_edges")
+        return reasons
+
+    def _validate_architecture(self, answer: str, evidence_pack: EvidenceBundle) -> list[str]:
+        facts = build_curated_answer_facts(evidence_pack)
+        architecture = dict(facts.get("architecture") or {})
+        reasons = self._validate_target_focus(answer, evidence_pack)
+        reasons.extend(self._vagueness_reasons(answer, "architecture"))
+
+        components = list(architecture.get("required_components") or [])
+        relations = list(architecture.get("required_relations") or [])
+        verbs = list(architecture.get("required_relation_verbs") or [])
+        if components and not self._mentions_fact_group(answer, components):
+            reasons.append("missing_concrete_components")
+        if relations and not self._mentions_relations(answer, relations):
+            reasons.append("missing_concrete_relations")
+        if verbs and not self._mentions_fact_group(answer, verbs):
+            reasons.append("missing_relation_verbs")
+        if any(label in answer for label in architecture.get("forbidden_labels") or []):
+            reasons.append("contains_retrieval_artifacts")
+        if self._methods_dominate_components(answer, components):
+            reasons.append("methods_as_primary_components")
+        if relations and (not self._mentions_relations(answer, relations) or not self._mentions_fact_group(answer, verbs)):
+            reasons.append("too_vague_for_architecture")
+        if self._semantic_leakage(answer, facts, has_concrete_support=self._mentions_relations(answer, relations)):
+            reasons.append("semantic_labels_without_code_edges")
+        return reasons
+
+    def _validate_trace_flow(self, answer: str, evidence_pack: EvidenceBundle) -> list[str]:
+        facts = build_curated_answer_facts(evidence_pack)
+        trace = dict(facts.get("trace_flow") or {})
+        reasons = self._validate_target_focus(answer, evidence_pack)
+        reasons.extend(self._vagueness_reasons(answer, "trace_flow"))
+
+        steps = list(trace.get("required_flow_steps") or [])
+        calls = list(trace.get("required_calls") or [])
+        if steps and not self._mentions_steps(answer, steps):
+            reasons.append("missing_flow_steps")
+        if calls and not self._mentions_fact_group(answer, calls):
+            reasons.append("missing_concrete_calls")
+        if steps and not self._mentions_relations(answer, steps):
+            reasons.append("missing_sequence_edges")
+        if any(claim in answer for claim in _OPTIMISTIC_TRACE_CLAIMS):
+            reasons.append("overclaims_trace_completeness")
+        if steps and not (self._mentions_steps(answer, steps) and self._mentions_relations(answer, steps)):
+            reasons.append("too_vague_for_trace_flow")
+        return reasons
+
+    def _validate_target_focus(self, answer: str, evidence_pack: EvidenceBundle) -> list[str]:
+        target = str(evidence_pack.resolved_target or "").strip().lower()
+        return [] if not target or target in answer else ["missing_resolved_target"]
+
+    def _mentions_fact_group(self, answer: str, values: list[str]) -> bool:
+        return any(alias in answer for value in values for alias in _aliases(value))
+
+    def _mentions_relations(self, answer: str, relations: list[dict]) -> bool:
+        for relation in relations:
+            source = str(relation.get("source") or "").lower()
+            target = str(relation.get("target") or "").lower()
+            verb = str(relation.get("verb") or "").lower()
+            if source and target and source in answer and target in answer:
+                return True
+            if source and verb and target and source in answer and verb in answer and target in answer:
+                return True
+        return False
+
+    def _mentions_steps(self, answer: str, steps: list[dict]) -> bool:
+        if "сначала" in answer and "затем" in answer:
+            return True
+        numeric_steps = sum(1 for marker in ("1.", "2.", "3.") if marker in answer)
+        if numeric_steps >= 2:
+            return True
+        mentioned = sum(1 for step in steps[:3] if self._mentions_relations(answer, [step]))
+        return mentioned >= min(2, len(steps[:3]))
+
+    def _methods_dominate_components(self, answer: str, components: list[str]) -> bool:
+        method_like = re.findall(r"\b[a-z_]+\(\)", answer)
+        component_hits = sum(1 for component in components if component.lower() in answer)
+        return bool(method_like) and component_hits == 0
+
+    def _semantic_leakage(self, answer: str, facts: dict, *, has_concrete_support: bool) -> bool:
+        if has_concrete_support:
+            return False
+        semantic_roles = [str(item.get("role") or "").strip().lower() for item in facts.get("semantic_hints") or [] if str(item.get("role") or "").strip()]
+        return bool(semantic_roles) and any(role in answer for role in semantic_roles)
+
+    def _vagueness_reasons(self, answer: str, scenario: str) -> list[str]:
+        if any(phrase in answer for phrase in _VAGUE_PHRASES):
+            return [f"too_vague_for_{scenario}"]
+        return []
+
+    def _tokens(self, value: str) -> set[str]:
+        return {token.lower().strip("/") for token in _TOKEN_RE.findall(value or "") if token.strip("/")}
+
+
+def _aliases(value: str) -> list[str]:
+    text = str(value or "").strip().lower()
+    if not text:
+        return []
+    aliases = [text]
+    bare = text.removesuffix("()")
+    if bare != text:
+        aliases.append(bare)
+    tail = bare.rsplit(".", 1)[-1]
+    if tail and tail not in aliases:
+        aliases.append(tail)
+    if tail and f"{tail}()" not in aliases:
+        aliases.append(f"{tail}()")
+    return aliases
+
+
+def _dedupe(values: list[str]) -> list[str]:
+    seen: set[str] = set()
+    result: list[str] = []
+    for value in values:
+        if value in seen:
+            continue
+        seen.add(value)
+        result.append(value)
+    return result

src/app/modules/agent/code_qa_runtime/prompt_payload_builder.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import json
+import re
 
 from app.modules.rag.code_qa_pipeline.contracts import AnswerSynthesisInput, EvidenceBundle
 
@@ -11,6 +12,7 @@ _LAYER_GUIDE = (
     "- C3_ENTRYPOINTS: подтвержденные точки входа.\n"
     "- C4_SEMANTIC_ROLES: вспомогательная роль компонента, использовать осторожно."
 )
+_TOKEN_RE = re.compile(r"[a-zA-Zа-яА-Я0-9_/]+")
 
 
 class CodeQaPromptPayloadBuilder:
@@ -30,8 +32,82 @@ class CodeQaPromptPayloadBuilder:
             "fast_context": synthesis_input.fast_context,
             "deep_context": synthesis_input.deep_context,
             "evidence_summary": synthesis_input.evidence_summary,
+            "semantic_hints": synthesis_input.semantic_hints,
             "diagnostic_hints": synthesis_input.diagnostic_hints,
             "retrieval_summary": evidence_pack.retrieval_summary,
+            "confirmed_entrypoints": self._entrypoints(user_query, evidence_pack),
+            "required_entrypoints": self._required_entrypoints(user_query, evidence_pack),
             "layer_guide": _LAYER_GUIDE,
         }
+        payload.update(self._scenario_payload(synthesis_input))
         return json.dumps(payload, ensure_ascii=False, indent=2)
+
+    def _entrypoints(self, user_query: str, evidence_pack: EvidenceBundle) -> list[dict]:
+        query_tokens = self._tokens(user_query)
+        enriched: list[dict] = []
+        for item in list(evidence_pack.entrypoints or []):
+            route_path = str(item.get("route_path") or "").strip()
+            http_method = str(item.get("http_method") or "").strip().upper()
+            if not http_method:
+                continue
+            candidate = {
+                "http_method": http_method,
+                "route_path": route_path,
+                "display": f"{http_method} {route_path}".strip(),
+                "handler_symbol": str(item.get("handler_symbol") or "").strip(),
+                "path": str(item.get("path") or "").strip(),
+                "start_line": item.get("start_line"),
+                "end_line": item.get("end_line"),
+                "query_match": self._matches_query(route_path, query_tokens),
+            }
+            if candidate not in enriched:
+                enriched.append(candidate)
+        return sorted(enriched, key=lambda item: (-int(bool(item["query_match"])), -int(bool(item["route_path"])), len(item["route_path"] or "/"), item["display"]))
+
+    def _required_entrypoints(self, user_query: str, evidence_pack: EvidenceBundle) -> list[str]:
+        return [item["display"] for item in self._entrypoints(user_query, evidence_pack) if item["query_match"] and item["route_path"]]
+
+    def _scenario_payload(self, synthesis_input: AnswerSynthesisInput) -> dict:
+        scenario = (synthesis_input.resolved_scenario or "").upper()
+        curated = dict(synthesis_input.curated_facts or {})
+        if scenario == "EXPLAIN":
+            facts = dict(curated.get("explain") or {})
+            return {
+                "must_mention_methods": facts.get("required_methods", []),
+                "must_mention_fields": facts.get("required_fields", []),
+                "must_mention_calls": facts.get("required_calls", []),
+                "must_mention_dependencies": facts.get("required_dependencies", []),
+                "must_mention_constructor_args": facts.get("required_constructor_args", []),
+                "must_mention_files": facts.get("required_files", []),
+                "must_not_infer_missing_details": True,
+                "fact_gaps": facts.get("fact_gaps", []),
+            }
+        if scenario == "ARCHITECTURE":
+            facts = dict(curated.get("architecture") or {})
+            return {
+                "must_mention_components": facts.get("required_components", []),
+                "must_mention_relations": facts.get("required_relations", []),
+                "must_use_relation_verbs": facts.get("required_relation_verbs", []),
+                "must_avoid_semantic_labels_as_primary_claims": True,
+                "must_not_use_retrieval_labels": facts.get("forbidden_labels", []),
+                "fact_gaps": facts.get("fact_gaps", []),
+            }
+        if scenario == "TRACE_FLOW":
+            facts = dict(curated.get("trace_flow") or {})
+            return {
+                "must_mention_flow_steps": facts.get("required_flow_steps", []),
+                "must_mention_calls": facts.get("required_calls", []),
+                "must_mention_sequence_edges": facts.get("required_sequence_edges", []),
+                "must_avoid_overclaiming_full_flow": True,
+                "fact_gaps": facts.get("fact_gaps", []),
+            }
+        return {}
+
+    def _matches_query(self, route_path: str, query_tokens: set[str]) -> bool:
+        if not route_path:
+            return False
+        path_tokens = self._tokens(route_path)
+        return bool(path_tokens & query_tokens)
+
+    def _tokens(self, value: str) -> set[str]:
+        return {token.lower().strip("/") for token in _TOKEN_RE.findall(value or "") if token.strip("/")}

src/app/modules/agent/code_qa_runtime/repair.py

@@ -17,10 +17,12 @@ class CodeQaAnswerRepairService:
         validation: CodeQaValidationResult,
         prompt_payload: str,
     ) -> str:
+        repair_focus = self._repair_focus(validation.reasons)
         repair_input = json.dumps(
             {
                 "draft_answer": draft_answer,
                 "validation_reasons": validation.reasons,
+                "repair_focus": repair_focus,
                 "prompt_payload": prompt_payload,
             },
             ensure_ascii=False,
@@ -31,3 +33,24 @@ class CodeQaAnswerRepairService:
             repair_input,
             log_context="graph.project_qa.code_qa.repair",
         ).strip()
+
+    def _repair_focus(self, reasons: list[str]) -> list[str]:
+        mapping = {
+            "missing_concrete_methods": "missing_concrete_methods",
+            "missing_concrete_calls": "missing_concrete_calls",
+            "missing_concrete_dependencies": "missing_concrete_dependencies",
+            "missing_concrete_components": "missing_concrete_components",
+            "missing_concrete_relations": "missing_concrete_relations",
+            "missing_relation_verbs": "missing_relation_verbs",
+            "missing_flow_steps": "missing_flow_steps",
+            "missing_sequence_edges": "missing_sequence_edges",
+            "too_vague_for_explain": "too_vague_for_explain",
+            "too_vague_for_architecture": "too_vague_for_architecture",
+            "too_vague_for_trace_flow": "too_vague_for_trace_flow",
+            "semantic_labels_without_code_edges": "semantic_labels_without_code_edges",
+            "contains_retrieval_artifacts": "contains_retrieval_artifacts",
+            "methods_as_primary_components": "methods_as_primary_components",
+            "overclaims_trace_completeness": "overclaims_trace_completeness",
+        }
+        result = [mapping[reason] for reason in reasons if reason in mapping]
+        return result or ["tighten_to_evidence"]

src/app/modules/agent/code_qa_runtime/short_answer_formatter.py

@@ -0,0 +1,33 @@
+from __future__ import annotations
+
+
+class CodeQaShortAnswerFormatter:
+    def open_file_not_found(self, target: str) -> str:
+        return f"Файл {target} не найден."
+
+    def entity_not_found(self, target: str, alternatives: list[str]) -> str:
+        base = f"Сущность {target} не найдена в доступном коде."
+        suffix = self._alternatives(alternatives)
+        return f"{base} {suffix}".strip()
+
+    def entity_ambiguous(self, target: str, alternatives: list[str]) -> str:
+        base = f"Сущность {target} не удалось однозначно разрешить."
+        suffix = self._alternatives(alternatives)
+        return f"{base} {suffix}".strip()
+
+    def insufficient(self, message: str) -> str:
+        normalized = (message or "").strip()
+        if normalized:
+            return normalized
+        return "Недостаточно подтверждённых данных для уверенного ответа."
+
+    def related_only(self, target: str, alternatives: list[str]) -> str:
+        base = f"Прямых подтверждений для сущности {target} не найдено."
+        suffix = self._alternatives(alternatives)
+        return f"{base} {suffix}".strip()
+
+    def _alternatives(self, alternatives: list[str]) -> str:
+        items = [item.strip() for item in alternatives if item and item.strip()]
+        if not items:
+            return ""
+        return f"Близкие варианты: {', '.join(items[:3])}."

src/app/modules/agent/prompts/code_qa_architecture_answer.txt

@@ -17,9 +17,15 @@
 Не выводи пустые разделы, пустые списки и формулировки вида "кандидатов нет", если это не помогает ответу.
 
 Дай архитектурное объяснение без лишней теории.
-Назови подтверждённые компоненты и конкретные связи между ними: создаёт, вызывает, регистрирует, читает, пишет, передаёт, оборачивает.
+Строй ответ вокруг concrete facts из payload: `must_mention_components`, `must_mention_relations`, `must_use_relation_verbs`.
+Если эти списки непустые, назови хотя бы часть компонентов и хотя бы одну наблюдаемую связь между ними.
+Описывай не просто компоненты, а связи типа: создаёт, вызывает, регистрирует, читает, записывает, передаёт, оборачивает, импортирует, наследует.
+Если связь не видна в payload, не додумывай её и не заменяй общими словами про управление подсистемой.
+Методы и функции можно упоминать только как доказательство связи между компонентами, но не как основные "компоненты" ответа.
 Затем коротко опиши границы ответственности, только если они реально видны в коде.
 Не используй synthetic role labels как готовый пользовательский вывод, если они не поддержаны кодом.
 Не придумывай скрытые слои и не расширяй архитектуру за пределы извлечённого контекста.
 Не используй обязательные markdown-секции.
-Не используй абстрактные формулы вроде "главный компонент", "центральный управляющий компонент", "управляет потоками данных и состоянием системы", если конкретная связь не раскрыта через наблюдаемые методы, поля или вызовы.
+Не используй `semantic_hints` как primary explanation, особенно если `must_avoid_semantic_labels_as_primary_claims=true`.
+Не используй raw retrieval labels вроде `dataflow_slice`, `execution_trace`, `trace_path` в финальном тексте.
+Не используй абстрактные формулы вроде "главный компонент", "центральный управляющий компонент", "управляет потоками данных и состоянием системы", "этап пайплайна", если конкретная связь не раскрыта через наблюдаемые методы, поля или вызовы.

src/app/modules/agent/prompts/code_qa_explain_answer.txt

@@ -18,9 +18,15 @@
 
 Объясни, как работает сущность из вопроса пользователя, обычным инженерным текстом.
 Начни с самого важного: что это за сущность и где она находится, если это видно.
-Затем кратко опиши подтверждённые зависимости, вызовы, аргументы, поля или шаги работы, только если они реально видны.
+Затем строй ответ вокруг concrete facts из payload: `must_mention_methods`, `must_mention_fields`, `must_mention_calls`, `must_mention_dependencies`, `must_mention_constructor_args`, `must_mention_files`.
+Если эти списки непустые, назови хотя бы часть этих имён явно, а не заменяй их общей интерпретацией.
+Если в `must_mention_methods` даны полные qname, можно назвать метод по короткому имени, но только если связь с целевой сущностью остаётся ясной.
+Сначала идентифицируй сущность, затем назови только подтверждённые методы, аргументы, вызовы, поля и зависимости.
+Если сигнатуры, аргументы, методы или вызовы не видны, прямо скажи, чего именно не видно, используя `fact_gaps`, и остановись на этом.
 Не используй общие формулы без конкретных имён.
 Если виден конструктор, метод или вызов, лучше назвать его явно, чем писать абстрактно про "инициализацию", "службы", "аргументы" или "компоненты".
 Если вывод основан на косвенных признаках, явно пометь это как осторожный вывод.
 Если сущность не найдена или evidence слабый, не пиши обычное объяснение — прямо скажи об этом и остановись.
+Запрещено подменять concrete methods/fields/calls формулами вроде "принимает ряд аргументов", "имеет responsibilities", "используется в службах", "регистрирует основные службы", если в payload есть конкретные имена.
+Не используй `semantic_hints` как основной каркас ответа. Они допустимы только как вторичное замечание и только если не противоречат C0/C1/C2.
 Не используй обязательные секции и подзаголовки.

src/app/modules/agent/prompts/code_qa_find_entrypoints_answer.txt

@@ -19,6 +19,8 @@
 Найди точки входа, обработчики запуска или важные entrypoints.
 Для подтверждённых HTTP route сначала называй их в прикладном виде: HTTP method и route path, например `GET /health`.
 Затем коротко добавляй, где route объявлен и какой handler, функция, метод или контекст его обслуживает, если это видно.
+Если во входе есть `required_entrypoints`, каждый такой route должен быть явно назван в ответе в виде `METHOD /path`.
+Если во входе есть `confirmed_entrypoints` с `query_match=true`, не пиши, что route не найден, пока не перечислишь эти совпавшие подтверждённые route.
 Подтверждённые entrypoints перечисляй первыми.
 Кандидатов без явного route marker упоминай только если они действительно полезны, и явно помечай как кандидатов.
 Не своди ответ к обсуждению декораторов вроде `@app.get`; пользователю важнее method, path и контекст.

src/app/modules/agent/prompts/code_qa_repair_answer.txt

@@ -1,3 +1,9 @@
 Ты исправляешь черновой ответ по коду после проверки groundedness.
 Сделай ответ короче, точнее и строже по evidence payload.
 Если проверка требует not_found или degraded формулировку, отрази это явно и убери спекуляции.
+Если в `repair_focus` есть причины для `EXPLAIN`, перепиши ответ так, чтобы он назвал concrete methods, calls, fields, constructor args или dependencies из payload, а не общие responsibilities.
+Если в `repair_focus` есть причины для `ARCHITECTURE`, перепиши ответ так, чтобы он назвал concrete components и связи с relation verbs из payload: создает, вызывает, читает, записывает, импортирует, наследует.
+Если в `repair_focus` есть причины для `TRACE_FLOW`, перепиши ответ как последовательность concrete steps с явными methods/calls/edges из payload. Если виден только partial flow, так и скажи.
+Если в `repair_focus` есть `semantic_labels_without_code_edges`, убери semantic role labels из основной формулировки, если они не подкреплены concrete code edges.
+Если в `repair_focus` есть `contains_retrieval_artifacts` или `methods_as_primary_components`, убери raw retrieval labels и не выдавай методы за компоненты.
+Если в `repair_focus` есть `overclaims_trace_completeness`, убери фразы про полный/полностью восстановленный flow, если payload не подтверждает это явно.

src/app/modules/agent/prompts/code_qa_trace_flow_answer.txt

@@ -17,7 +17,9 @@
 Не выводи пустые разделы, пустые списки и формулировки вида "кандидатов нет", если это не помогает ответу.
 
 Проследи поток выполнения или поток данных по найденным артефактам.
-Старайся описывать шаги последовательно и коротко, без лишних подзаголовков.
+Строй ответ вокруг `must_mention_flow_steps`, `must_mention_calls` и `must_mention_sequence_edges` из payload.
+Старайся описывать шаги последовательно и коротко, без лишних подзаголовков: сначала, затем, после этого, в конце.
 Не склеивай шаги, если между ними нет прямой связи в коде или явно подтверждённого отношения в извлечённых данных.
-Если поток восстанавливается только частично, так и скажи.
+Если поток восстанавливается только частично, так и скажи, опираясь на `fact_gaps`, и не заявляй, что flow восстановлен полностью.
 Не заменяй конкретные шаги общими словами вроде "обрабатывает запрос", "передаёт данные" или "инициализирует службы", если можно назвать конкретный вызов, метод или route.
+Не используй сильные формулировки вроде "полностью восстанавливается", "полный поток виден", если payload показывает только часть цепочки.

src/app/modules/rag/code_qa_pipeline/answer_fact_curator.py

@@ -0,0 +1,300 @@
+from __future__ import annotations
+
+import re
+from typing import Any
+
+from app.modules.rag.code_qa_pipeline.contracts import CodeChunkItem, EvidenceBundle
+
+_CALL_RE = re.compile(r"([A-Za-z_][\w\.]*)\s*\(")
+_FIELD_RE = re.compile(r"self\.(\w+)")
+_SIGNATURE_RE = re.compile(r"(?P<name>[A-Za-z_][\w\.]*)\((?P<args>[^)]*)\)")
+_RETRIEVAL_LABELS = ("dataflow_slice", "execution_trace", "trace_path")
+_RELATION_VERBS = {
+    "calls": "вызывает",
+    "instantiates": "создает",
+    "inherits": "наследует",
+    "imports": "импортирует",
+    "reads_attr": "читает",
+    "writes_attr": "записывает",
+}
+
+
+def build_curated_answer_facts(bundle: EvidenceBundle) -> dict[str, Any]:
+    target = str(bundle.resolved_target or "").strip()
+    semantic_hints = _semantic_hints(bundle.code_chunks)
+    primary_chunks = [chunk for chunk in bundle.code_chunks if chunk.layer != "C4_SEMANTIC_ROLES"]
+    relations = _normalized_relations(bundle.relations)
+    target_relations = [relation for relation in relations if _is_target_relation(relation, target)]
+
+    return {
+        "scenario": (bundle.resolved_sub_intent or "EXPLAIN").upper(),
+        "semantic_hints": semantic_hints,
+        "primary_chunk_count": len(primary_chunks),
+        "relation_count": len(relations),
+        "explain": _explain_facts(bundle, primary_chunks, target_relations),
+        "architecture": _architecture_facts(bundle, primary_chunks, target_relations),
+        "trace_flow": _trace_flow_facts(bundle, primary_chunks, target_relations),
+    }
+
+
+def _explain_facts(bundle: EvidenceBundle, chunks: list[CodeChunkItem], relations: list[dict[str, Any]]) -> dict[str, Any]:
+    target = str(bundle.resolved_target or "").strip()
+    signatures = [_signature_payload(chunk) for chunk in chunks if chunk.layer == "C1_SYMBOL_CATALOG"]
+    target_signatures = [item for item in signatures if _is_target_symbol(item["name"], target)]
+    methods = _unique(item["name"] for item in target_signatures if item["kind"] == "method")
+    constructor_args = _unique(
+        arg
+        for item in target_signatures
+        if item["name"].endswith(".__init__")
+        for arg in item["args"]
+        if arg not in {"self", "cls"}
+    )
+    calls = _unique(
+        _display_call_target(relation["target"])
+        for relation in relations
+        if relation["edge_type"] in {"calls", "instantiates"}
+    )
+    if not calls:
+        calls = _fallback_calls(chunks, target)
+    fields = _unique(
+        relation["target"].split(".", 1)[-1]
+        for relation in relations
+        if relation["edge_type"] in {"reads_attr", "writes_attr"}
+    )
+    if not fields:
+        fields = _unique(field for chunk in chunks if _chunk_matches_target(chunk, target) for field in _FIELD_RE.findall(chunk.content or ""))
+    dependencies = _unique(
+        _display_dependency_target(relation["target"])
+        for relation in relations
+        if relation["edge_type"] in {"imports", "instantiates"}
+    )
+    required_files = _unique(chunk.path for chunk in chunks if _chunk_matches_target(chunk, target) and chunk.path)
+    required_symbols = _unique(item["name"] for item in target_signatures if item["name"]) or _unique([target] if target else bundle.target_symbol_candidates)
+
+    fact_gaps: list[str] = []
+    if not methods:
+        fact_gaps.append("Методы целевой сущности не подтверждены в извлеченных C0/C1 данных.")
+    if not calls:
+        fact_gaps.append("Конкретные вызовы целевой сущности не подтверждены в C2/C0.")
+    if not dependencies:
+        fact_gaps.append("Явные зависимости целевой сущности не подтверждены.")
+
+    return {
+        "required_symbols": required_symbols[:8],
+        "required_methods": methods[:6],
+        "required_calls": calls[:6],
+        "required_fields": fields[:6],
+        "required_constructor_args": constructor_args[:6],
+        "required_dependencies": dependencies[:6],
+        "required_files": required_files[:4],
+        "fact_gaps": fact_gaps,
+    }
+
+
+def _architecture_facts(bundle: EvidenceBundle, chunks: list[CodeChunkItem], relations: list[dict[str, Any]]) -> dict[str, Any]:
+    target = str(bundle.resolved_target or "").strip()
+    components = _unique(
+        ([target] if target else [])
+        + [_component_name(relation["source"]) for relation in relations]
+        + [_component_name(relation["target"]) for relation in relations]
+        + [_component_name(chunk.title) for chunk in chunks if _chunk_matches_target(chunk, target)]
+    )
+    relation_rows = [
+        {
+            "source": _component_name(relation["source"]),
+            "verb": relation["verb"],
+            "target": _component_name(relation["target"]),
+            "source_method": relation["source"],
+            "target_method": relation["target"],
+            "path": relation["path"],
+            "line_span": relation["line_span"],
+            "edge_type": relation["edge_type"],
+        }
+        for relation in relations
+        if _component_name(relation["source"]) and _component_name(relation["target"])
+    ]
+    relation_rows = [row for row in relation_rows if row["source"] != row["target"]]
+    relation_verbs = _unique(row["verb"] for row in relation_rows if row["verb"])
+    fact_gaps: list[str] = []
+    if not relation_rows:
+        fact_gaps.append("Concrete code edges между компонентами не подтверждены.")
+
+    return {
+        "required_components": components[:8],
+        "required_relations": relation_rows[:8],
+        "required_relation_verbs": relation_verbs[:6],
+        "required_creation_edges": [row for row in relation_rows if row["edge_type"] == "instantiates"][:4],
+        "required_call_edges": [row for row in relation_rows if row["edge_type"] == "calls"][:4],
+        "required_registration_edges": [row for row in relation_rows if row["edge_type"] == "imports"][:4],
+        "required_data_edges": [row for row in relation_rows if row["edge_type"] in {"reads_attr", "writes_attr"}][:4],
+        "fact_gaps": fact_gaps,
+        "forbidden_labels": list(_RETRIEVAL_LABELS),
+    }
+
+
+def _trace_flow_facts(bundle: EvidenceBundle, chunks: list[CodeChunkItem], relations: list[dict[str, Any]]) -> dict[str, Any]:
+    target = str(bundle.resolved_target or "").strip()
+    flow_steps = [
+        {
+            "step": index,
+            "source": relation["source"],
+            "verb": relation["verb"],
+            "target": relation["target"],
+            "path": relation["path"],
+            "line_span": relation["line_span"],
+        }
+        for index, relation in enumerate(sorted(relations, key=lambda item: (item["path"], item["sort_line"], item["source"], item["target"])), start=1)
+    ]
+    fact_gaps: list[str] = []
+    if len(flow_steps) < 2:
+        fact_gaps.append("Полная последовательность шагов не подтверждена; виден только частичный flow.")
+    if not flow_steps:
+        fact_gaps.append("Конкретные sequence edges для flow не подтверждены.")
+
+    return {
+        "required_flow_steps": flow_steps[:8],
+        "required_calls": _unique(_display_call_target(item["target"]) for item in flow_steps),
+        "required_sequence_edges": flow_steps[:8],
+        "required_files": _unique(chunk.path for chunk in chunks if _chunk_matches_target(chunk, target) and chunk.path)[:4],
+        "fact_gaps": fact_gaps,
+    }
+
+
+def _semantic_hints(chunks: list[CodeChunkItem]) -> list[dict[str, Any]]:
+    hints: list[dict[str, Any]] = []
+    for chunk in chunks:
+        if chunk.layer != "C4_SEMANTIC_ROLES":
+            continue
+        meta = dict(chunk.metadata or {})
+        hints.append(
+            {
+                "symbol": meta.get("symbol_name") or chunk.title,
+                "role": meta.get("role"),
+                "path": chunk.path,
+                "confidence": meta.get("confidence"),
+            }
+        )
+    return hints[:8]
+
+
+def _normalized_relations(relations: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    normalized: list[dict[str, Any]] = []
+    for relation in relations:
+        metadata = dict(relation.get("metadata") or {})
+        edge_type = str(metadata.get("edge_type") or relation.get("edge_type") or "").strip()
+        source = _clean_endpoint(str(metadata.get("src_qname") or relation.get("source") or "").strip())
+        target = _clean_endpoint(str(metadata.get("dst_ref") or relation.get("target") or "").strip())
+        if not edge_type or not source or not target:
+            continue
+        normalized.append(
+            {
+                "edge_type": edge_type,
+                "source": source,
+                "target": target,
+                "verb": _RELATION_VERBS.get(edge_type, edge_type),
+                "path": str(relation.get("path") or ""),
+                "line_span": _line_span(relation),
+                "sort_line": int(relation.get("start_line") or 0),
+            }
+        )
+    return normalized
+
+
+def _signature_payload(chunk: CodeChunkItem) -> dict[str, Any]:
+    meta = dict(chunk.metadata or {})
+    signature = str(meta.get("signature") or chunk.content or "").strip()
+    match = _SIGNATURE_RE.search(signature)
+    args = [part.strip() for part in match.group("args").split(",") if part.strip()] if match else []
+    return {
+        "name": _clean_endpoint(str(meta.get("qname") or chunk.title or "").strip()),
+        "kind": str(meta.get("kind") or "").strip(),
+        "args": args,
+    }
+
+
+def _fallback_calls(chunks: list[CodeChunkItem], target: str) -> list[str]:
+    values: list[str] = []
+    for chunk in chunks:
+        if not _chunk_matches_target(chunk, target):
+            continue
+        values.extend(_display_call_target(item) for item in _CALL_RE.findall(chunk.content or ""))
+    return _unique(values)
+
+
+def _chunk_matches_target(chunk: CodeChunkItem, target: str) -> bool:
+    if not target:
+        return True
+    title = _clean_endpoint(chunk.title)
+    symbol = _clean_endpoint(str(dict(chunk.metadata or {}).get("qname") or ""))
+    return _is_target_symbol(title, target) or _is_target_symbol(symbol, target) or target.lower() in (chunk.content or "").lower()
+
+
+def _is_target_relation(relation: dict[str, Any], target: str) -> bool:
+    if not target:
+        return True
+    return _is_target_symbol(relation["source"], target) or _component_name(relation["source"]) == target or _component_name(relation["target"]) == target
+
+
+def _is_target_symbol(name: str, target: str) -> bool:
+    clean_name = _clean_endpoint(name)
+    clean_target = _clean_endpoint(target)
+    if not clean_target:
+        return True
+    return clean_name == clean_target or clean_name.startswith(f"{clean_target}.")
+
+
+def _component_name(value: str) -> str:
+    clean = _clean_endpoint(value)
+    if not clean:
+        return ""
+    parts = clean.split(".")
+    if len(parts) <= 1:
+        return clean
+    if len(parts) == 2:
+        return clean if parts[1][:1].isupper() else parts[0]
+    return ".".join(parts[:-1])
+
+
+def _display_call_target(value: str) -> str:
+    clean = _clean_endpoint(value)
+    if not clean:
+        return ""
+    tail = clean.rsplit(".", 1)[-1]
+    return f"{tail}()" if tail and tail != clean else clean
+
+
+def _display_dependency_target(value: str) -> str:
+    clean = _clean_endpoint(value)
+    if not clean:
+        return ""
+    return _component_name(clean) or clean
+
+
+def _clean_endpoint(value: str) -> str:
+    clean = str(value or "").strip()
+    if not clean:
+        return ""
+    for label in _RETRIEVAL_LABELS:
+        clean = clean.replace(f".{label}", "")
+        clean = clean.replace(label, "")
+    return clean.strip(".: ")
+
+
+def _line_span(item: dict[str, Any]) -> str:
+    start = item.get("start_line")
+    end = item.get("end_line")
+    if start is None and end is None:
+        return "?:?"
+    return f"{start or '?'}-{end or '?'}"
+
+
+def _unique(values: Any) -> list[str]:
+    seen: set[str] = set()
+    result: list[str] = []
+    for value in values:
+        text = str(value or "").strip()
+        if not text or text in seen:
+            continue
+        seen.add(text)
+        result.append(text)
+    return result

src/app/modules/rag/code_qa_pipeline/answer_synthesis.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+from app.modules.rag.code_qa_pipeline.answer_fact_curator import build_curated_answer_facts
 from app.modules.rag.code_qa_pipeline.contracts import AnswerSynthesisInput, EvidenceBundle
 
 
@@ -13,11 +14,16 @@ def build_answer_synthesis_input(
     scenario = bundle.resolved_sub_intent or "EXPLAIN"
     target = bundle.resolved_target
     sufficient = bundle.sufficient
+    curated = build_curated_answer_facts(bundle)
+    semantic_hints = list(curated.get("semantic_hints") or [])
 
     fast_lines = [
         f"Scenario: {scenario}",
         f"Target: {target or 'none'}",
         f"Evidence chunks: {bundle.evidence_count}",
+        f"Primary chunks: {curated.get('primary_chunk_count', 0)}",
+        f"Semantic hints: {len(semantic_hints)}",
+        f"Relations: {curated.get('relation_count', 0)}",
         f"Sufficient: {sufficient}",
     ]
     if bundle.failure_reasons:
@@ -25,18 +31,21 @@ def build_answer_synthesis_input(
     fast_context = "\n".join(fast_lines)
 
     deep_parts = []
-    for i, c in enumerate(bundle.code_chunks[:30], 1):
+    primary_chunks = [chunk for chunk in bundle.code_chunks if scenario.upper() not in {"EXPLAIN", "ARCHITECTURE"} or chunk.layer != "C4_SEMANTIC_ROLES"]
+    for i, c in enumerate(primary_chunks[:30], 1):
         deep_parts.append(f"[{i}] {c.path}:{c.start_line or '?'}-{c.end_line or '?'}\n{c.content[:800]}")
     deep_context = "\n\n---\n\n".join(deep_parts) if deep_parts else "(no code chunks)"
 
     evidence_summary = []
-    for c in bundle.code_chunks[:20]:
+    for c in primary_chunks[:20]:
         evidence_summary.append({
             "layer": c.layer,
             "path": c.path,
             "title": c.title,
             "lines": f"{c.start_line or '?'}-{c.end_line or '?'}",
         })
+    if semantic_hints:
+        evidence_summary.append({"kind": "semantic_hints", "count": len(semantic_hints)})
     if bundle.entrypoints:
         evidence_summary.append({"kind": "entrypoints", "count": len(bundle.entrypoints)})
     if bundle.test_evidence:
@@ -45,6 +54,9 @@ def build_answer_synthesis_input(
     hints = list(bundle.failure_reasons) if bundle.failure_reasons else []
     if not sufficient and bundle.retrieval_summary.get("missing_layers"):
         hints.append(f"Missing layers: {', '.join(bundle.retrieval_summary['missing_layers'])}")
+    scenario_key = scenario.lower()
+    scenario_facts = dict(curated.get(scenario_key) or {})
+    hints.extend(list(scenario_facts.get("fact_gaps") or []))
 
     return AnswerSynthesisInput(
         user_question=user_question,
@@ -53,6 +65,8 @@ def build_answer_synthesis_input(
         fast_context=fast_context,
         deep_context=deep_context,
         evidence_summary=evidence_summary,
+        semantic_hints=semantic_hints,
+        curated_facts=curated,
         evidence_sufficient=sufficient,
         diagnostic_hints=hints,
     )

src/app/modules/rag/code_qa_pipeline/contracts.py

@@ -98,6 +98,7 @@ class RetrievalResult(BaseModel):
     file_candidates: list[str] = Field(default_factory=list)
     code_chunks: list[CodeChunkItem] = Field(default_factory=list)
     relations: list[dict[str, Any]] = Field(default_factory=list)
+    semantic_hints: list[dict[str, Any]] = Field(default_factory=list)
     entrypoints: list[dict[str, Any]] = Field(default_factory=list)
     test_candidates: list[dict[str, Any]] = Field(default_factory=list)
     layer_outcomes: list[LayerOutcome] = Field(default_factory=list)
@@ -138,6 +139,8 @@ class AnswerSynthesisInput(BaseModel):
     fast_context: str = ""
     deep_context: str = ""
     evidence_summary: list[dict[str, Any]] = Field(default_factory=list)
+    semantic_hints: list[dict[str, Any]] = Field(default_factory=list)
+    curated_facts: dict[str, Any] = Field(default_factory=dict)
     evidence_sufficient: bool = False
     diagnostic_hints: list[str] = Field(default_factory=list)
 

src/app/modules/rag/code_qa_pipeline/evidence_bundle_builder.py

@@ -28,6 +28,7 @@ def build_evidence_bundle(
     summary = {
         "chunk_count": len(retrieval_result.code_chunks),
         "relation_count": len(retrieval_result.relations),
+        "semantic_hint_count": len(retrieval_result.semantic_hints),
         "entrypoint_count": len(retrieval_result.entrypoints),
         "test_candidate_count": len(retrieval_result.test_candidates),
         "layers_with_hits": [

src/app/modules/rag/code_qa_pipeline/retrieval_result_builder.py

@@ -2,13 +2,14 @@
 
 from __future__ import annotations
 
-from app.modules.rag.code_qa_pipeline.contracts import (
-    CodeChunkItem,
-    LayerOutcome,
-    RetrievalResult,
-)
+import re
+
+from app.modules.rag.code_qa_pipeline.contracts import CodeChunkItem, LayerOutcome, RetrievalResult
 from app.modules.rag.retrieval.test_filter import is_test_path
 
+_ROUTE_RE = re.compile(r'@[\w\.]+\.(get|post|put|delete|patch|options|head)\(\s*["\']([^"\']+)["\']')
+_DEF_RE = re.compile(r"async\s+def\s+(\w+)|def\s+(\w+)")
+
 
 def build_retrieval_result(
     raw_rows: list[dict],
@@ -21,52 +22,210 @@ def build_retrieval_result(
     layers_seen: set[str] = set()
     code_chunks: list[CodeChunkItem] = []
     relations: list[dict] = []
-    entrypoints: list[dict] = []
+    semantic_hints: list[dict] = []
+    entrypoint_rows: list[CodeChunkItem] = []
     test_candidates: list[dict] = []
     file_candidates: list[str] = []
     target_symbols: list[str] = []
 
     for row in raw_rows:
-        layer = str(row.get("layer") or "")
-        path = str(row.get("path") or "")
+        chunk = _to_chunk(row)
+        layer = chunk.layer
+        path = chunk.path
         if layer:
             layers_seen.add(layer)
         if path and path not in file_candidates:
             file_candidates.append(path)
-
-        chunk = CodeChunkItem(
-            layer=layer,
-            path=path,
-            title=str(row.get("title") or ""),
-            content=str(row.get("content") or ""),
-            start_line=row.get("span_start"),
-            end_line=row.get("span_end"),
-            metadata=dict(row.get("metadata") or {}),
-        )
         code_chunks.append(chunk)
 
         if layer == "C1_SYMBOL_CATALOG":
-            title = str(row.get("title") or "").strip()
+            title = chunk.title.strip()
             if title and title not in target_symbols:
                 target_symbols.append(title)
-        elif layer == "C2_DEPENDENCY_GRAPH":
-            relations.append({"path": path, "metadata": chunk.metadata, "content": chunk.content[:500]})
-        elif layer == "C3_ENTRYPOINTS":
-            entrypoints.append({"path": path, "title": chunk.title, "metadata": chunk.metadata})
-        elif is_test_path(path):
+            continue
+        if layer == "C2_DEPENDENCY_GRAPH":
+            relations.append(_relation_payload(chunk))
+            continue
+        if layer == "C4_SEMANTIC_ROLES":
+            semantic_hints.append(_semantic_hint_payload(chunk))
+            continue
+        if layer == "C3_ENTRYPOINTS":
+            entrypoint_rows.append(chunk)
+            continue
+        if is_test_path(path):
             test_candidates.append({"path": path, "title": chunk.title, "content": chunk.content[:300]})
 
+    entrypoints = _normalize_entrypoints(entrypoint_rows, code_chunks)
     executed = list(report.get("executed_layers") or [])
     missing_layers = [lid for lid in executed if lid not in layers_seen]
+    layer_outcomes = _layer_outcomes(raw_rows, executed, layers_seen, report)
+
+    resolved = str(sym.get("resolved_symbol") or "").strip() or None
+    status = str(sym.get("status") or "not_requested")
+
+    return RetrievalResult(
+        target_symbol_candidates=target_symbols,
+        resolved_symbol=resolved,
+        symbol_resolution_status=status,
+        file_candidates=file_candidates,
+        code_chunks=code_chunks,
+        relations=relations,
+        semantic_hints=semantic_hints,
+        entrypoints=entrypoints,
+        test_candidates=test_candidates,
+        layer_outcomes=layer_outcomes,
+        missing_layers=missing_layers,
+        raw_rows=raw_rows,
+        retrieval_report=report,
+    )
+
+
+def _to_chunk(row: dict) -> CodeChunkItem:
+    return CodeChunkItem(
+        layer=str(row.get("layer") or ""),
+        path=str(row.get("path") or ""),
+        title=str(row.get("title") or ""),
+        content=str(row.get("content") or ""),
+        start_line=row.get("span_start"),
+        end_line=row.get("span_end"),
+        metadata=dict(row.get("metadata") or {}),
+    )
+
+
+def _normalize_entrypoints(entrypoint_rows: list[CodeChunkItem], code_chunks: list[CodeChunkItem]) -> list[dict]:
+    chunks_by_path: dict[str, list[CodeChunkItem]] = {}
+    for chunk in code_chunks:
+        chunks_by_path.setdefault(chunk.path, []).append(chunk)
+
+    normalized: list[dict] = []
+    seen: set[tuple[str, str, str, int | None, int | None]] = set()
+    for chunk in entrypoint_rows:
+        item = _entrypoint_payload(chunk, chunks_by_path.get(chunk.path, []))
+        key = (
+            str(item.get("http_method") or ""),
+            str(item.get("route_path") or ""),
+            str(item.get("path") or ""),
+            chunk.start_line,
+            chunk.end_line,
+        )
+        if key in seen:
+            continue
+        seen.add(key)
+        normalized.append(item)
+    return sorted(normalized, key=_entrypoint_sort_key)
+
+
+def _relation_payload(chunk: CodeChunkItem) -> dict:
+    metadata = dict(chunk.metadata or {})
+    edge_type = str(metadata.get("edge_type") or "").strip()
+    src_qname = str(metadata.get("src_qname") or "").strip()
+    dst_ref = str(metadata.get("dst_ref") or "").strip()
+    return {
+        "path": chunk.path,
+        "start_line": chunk.start_line,
+        "end_line": chunk.end_line,
+        "edge_type": edge_type,
+        "source": src_qname,
+        "target": dst_ref,
+        "source_component": _component_name(src_qname),
+        "target_component": _component_name(dst_ref),
+        "has_retrieval_label": "dataflow_slice" in dst_ref or "dataflow_slice" in src_qname,
+        "metadata": metadata,
+        "content": chunk.content[:500],
+    }
+
+
+def _semantic_hint_payload(chunk: CodeChunkItem) -> dict:
+    metadata = dict(chunk.metadata or {})
+    return {
+        "path": chunk.path,
+        "title": chunk.title,
+        "symbol": metadata.get("symbol_name") or chunk.title,
+        "role": metadata.get("role"),
+        "confidence": metadata.get("confidence"),
+        "content": chunk.content[:300],
+    }
+
+
+def _component_name(value: str) -> str:
+    cleaned = str(value or "").replace(".dataflow_slice", "").strip(". ")
+    if "." not in cleaned:
+        return cleaned
+    head, tail = cleaned.rsplit(".", 1)
+    return head if tail and not tail[:1].isupper() else cleaned
+
+
+def _entrypoint_payload(chunk: CodeChunkItem, siblings: list[CodeChunkItem]) -> dict:
+    methods = [str(item).strip().upper() for item in list(dict(chunk.metadata).get("lang_payload", {}).get("methods", [])) if str(item).strip()]
+    method = methods[0] if methods else ""
+    route_path = _infer_route_path(chunk, siblings)
+    handler_symbol = _infer_handler_symbol(chunk, siblings)
+    return {
+        "path": chunk.path,
+        "title": chunk.title,
+        "start_line": chunk.start_line,
+        "end_line": chunk.end_line,
+        "http_method": method,
+        "route_path": route_path,
+        "handler_symbol": handler_symbol,
+        "handler_symbol_id": dict(chunk.metadata).get("handler_symbol_id"),
+        "route_or_command": dict(chunk.metadata).get("route_or_command"),
+        "metadata": chunk.metadata,
+    }
+
+
+def _infer_route_path(chunk: CodeChunkItem, siblings: list[CodeChunkItem]) -> str:
+    for candidate in siblings:
+        if candidate.layer != "C0_SOURCE_CHUNKS":
+            continue
+        match = _ROUTE_RE.search(_window_text(candidate, chunk.start_line))
+        if match:
+            return match.group(2).strip()
+    return ""
+
+
+def _infer_handler_symbol(chunk: CodeChunkItem, siblings: list[CodeChunkItem]) -> str:
+    for candidate in siblings:
+        if candidate.layer != "C0_SOURCE_CHUNKS":
+            continue
+        match = _DEF_RE.search(_window_text(candidate, chunk.start_line, before=2, after=12))
+        if match:
+            return match.group(1) or match.group(2) or ""
+    return ""
+
+
+def _window_text(chunk: CodeChunkItem, target_line: int | None, *, before: int = 6, after: int = 8) -> str:
+    text = chunk.content or ""
+    if not text or not target_line or not chunk.start_line:
+        return text
+    lines = text.splitlines()
+    offset = max(0, target_line - chunk.start_line - before)
+    end = min(len(lines), target_line - chunk.start_line + after)
+    if offset >= end:
+        return text
+    return "\n".join(lines[offset:end])
+
+
+def _entrypoint_sort_key(item: dict) -> tuple[int, int, str, str, int]:
+    route_path = str(item.get("route_path") or "")
+    method = str(item.get("http_method") or "")
+    score = 0
+    if route_path:
+        score += 3
+    if route_path and route_path != "/":
+        score += 3
+    if method:
+        score += 1
+    return (-score, len(route_path or "zzzz"), method, str(item.get("path") or ""), int(item.get("start_line") or 0))
+
+
+def _layer_outcomes(raw_rows: list[dict], executed: list[str], layers_seen: set[str], report: dict) -> list[LayerOutcome]:
     layer_outcomes = [
         LayerOutcome(
             layer_id=layer_id,
             hit_count=sum(1 for r in raw_rows if str(r.get("layer") or "") == layer_id),
             empty=layer_id not in layers_seen,
-            fallback_used=bool(
-                (report.get("fallback") or {}).get("used")
-                and report.get("retrieval_mode_by_layer", {}).get(layer_id)
-            ),
+            fallback_used=bool((report.get("fallback") or {}).get("used") and report.get("retrieval_mode_by_layer", {}).get(layer_id)),
         )
         for layer_id in executed
     ]
@@ -80,21 +239,4 @@ def build_retrieval_result(
                     fallback_used=False,
                 )
             )
-
-    resolved = str(sym.get("resolved_symbol") or "").strip() or None
-    status = str(sym.get("status") or "not_requested")
-
-    return RetrievalResult(
-        target_symbol_candidates=target_symbols,
-        resolved_symbol=resolved,
-        symbol_resolution_status=status,
-        file_candidates=file_candidates,
-        code_chunks=code_chunks,
-        relations=relations,
-        entrypoints=entrypoints,
-        test_candidates=test_candidates,
-        layer_outcomes=layer_outcomes,
-        missing_layers=missing_layers,
-        raw_rows=raw_rows,
-        retrieval_report=report,
-    )
+    return layer_outcomes

src/app/modules/shared/gigachat/settings.py

@@ -20,6 +20,6 @@ class GigaChatSettings:
             scope=os.getenv("GIGACHAT_SCOPE", "GIGACHAT_API_PERS"),
             credentials=os.getenv("GIGACHAT_TOKEN", "").strip(),
             ssl_verify=os.getenv("GIGACHAT_SSL_VERIFY", "true").lower() in {"1", "true", "yes"},
-            model=os.getenv("GIGACHAT_MODEL", "GigaChat"),
+            model=os.getenv("GIGACHAT_MODEL", "GigaChat-Pro"),
             embedding_model=os.getenv("GIGACHAT_EMBEDDING_MODEL", "Embeddings"),
         )