Фиксация изменений

app/modules/agent/README.md

@@ -37,6 +37,8 @@ classDiagram
   Методы: `run` — строит, валидирует и исполняет execution plan.
 - `TaskSpecBuilder`: формирует спецификацию задачи для оркестратора.
   Методы: `build` — собирает `TaskSpec` из route, контекстов и ограничений.
+- `ProjectQaConversationGraphFactory`, `ProjectQaClassificationGraphFactory`, `ProjectQaRetrievalGraphFactory`, `ProjectQaAnalysisGraphFactory`, `ProjectQaAnswerGraphFactory`: набор маленьких graph-исполнителей для `project/qa`.
+  Роли: нормализация запроса; классификация project-question; поздний retrieval из `RAG`; анализ code/docs контекста; сборка финального ответа.
 - `StorySessionRecorder`: пишет session-scoped артефакты для последующего bind к Story.
   Методы: `record_run` — сохраняет входные источники и выходные артефакты сессии.
 - `StoryContextRepository`: репозиторий Story-контекста и его связей.
@@ -58,3 +60,32 @@ sequenceDiagram
     Router->>Confluence: fetch_page(url)
     Confluence-->>Router: page(content_markdown, metadata)
 ```
+
+### `project/qa` reasoning flow
+Назначение: оркестратор планирует шаги, а каждый шаг исполняется отдельным graph. Retrieval вызывается поздно, внутри шага `context_retrieval`.
+```mermaid
+sequenceDiagram
+    participant Runtime as GraphAgentRuntime
+    participant Orch as OrchestratorService
+    participant G1 as conversation_understanding
+    participant G2 as question_classification
+    participant G3 as context_retrieval
+    participant Rag as RagService
+    participant G4 as context_analysis
+    participant G5 as answer_composition
+
+    Runtime->>Orch: run(task)
+    Orch->>G1: execute
+    G1-->>Orch: resolved_request
+    Orch->>G2: execute
+    G2-->>Orch: question_profile
+    Orch->>G3: execute
+    G3->>Rag: retrieve(query)
+    Rag-->>G3: rag_items
+    G3-->>Orch: source_bundle
+    Orch->>G4: execute
+    G4-->>Orch: analysis_brief
+    Orch->>G5: execute
+    G5-->>Orch: final_answer
+    Orch-->>Runtime: final_answer
+```

app/modules/agent/engine/graphs/__init__.py

@@ -1,8 +1,13 @@
 __all__ = [
     "BaseGraphFactory",
     "DocsGraphFactory",
+    "ProjectQaAnalysisGraphFactory",
+    "ProjectQaAnswerGraphFactory",
+    "ProjectQaClassificationGraphFactory",
+    "ProjectQaConversationGraphFactory",
     "ProjectEditsGraphFactory",
     "ProjectQaGraphFactory",
+    "ProjectQaRetrievalGraphFactory",
 ]
 
 
@@ -15,6 +20,26 @@ def __getattr__(name: str):
         from app.modules.agent.engine.graphs.docs_graph import DocsGraphFactory
 
         return DocsGraphFactory
+    if name == "ProjectQaConversationGraphFactory":
+        from app.modules.agent.engine.graphs.project_qa_step_graphs import ProjectQaConversationGraphFactory
+
+        return ProjectQaConversationGraphFactory
+    if name == "ProjectQaClassificationGraphFactory":
+        from app.modules.agent.engine.graphs.project_qa_step_graphs import ProjectQaClassificationGraphFactory
+
+        return ProjectQaClassificationGraphFactory
+    if name == "ProjectQaRetrievalGraphFactory":
+        from app.modules.agent.engine.graphs.project_qa_step_graphs import ProjectQaRetrievalGraphFactory
+
+        return ProjectQaRetrievalGraphFactory
+    if name == "ProjectQaAnalysisGraphFactory":
+        from app.modules.agent.engine.graphs.project_qa_step_graphs import ProjectQaAnalysisGraphFactory
+
+        return ProjectQaAnalysisGraphFactory
+    if name == "ProjectQaAnswerGraphFactory":
+        from app.modules.agent.engine.graphs.project_qa_step_graphs import ProjectQaAnswerGraphFactory
+
+        return ProjectQaAnswerGraphFactory
     if name == "ProjectEditsGraphFactory":
         from app.modules.agent.engine.graphs.project_edits_graph import ProjectEditsGraphFactory
 

app/modules/agent/engine/graphs/base_graph.py

@@ -59,7 +59,7 @@ class BaseGraphFactory:
                 f"Confluence context:\n{conf}",
             ]
         )
-        answer = self._llm.generate("general_answer", user_input)
+        answer = self._llm.generate("general_answer", user_input, log_context="graph.default.answer")
         emit_progress_sync(
             state,
             stage="graph.default.answer.done",

app/modules/agent/engine/graphs/docs_graph_logic.py

@@ -52,7 +52,7 @@ class DocsContextAnalyzer:
                 f"Detected documentation candidates:\n{snippets}",
             ]
         )
-        raw = self._llm.generate("docs_detect", user_input)
+        raw = self._llm.generate("docs_detect", user_input, log_context="graph.docs.detect_existing_docs")
         exists = self.parse_bool_marker(raw, "exists", default=True)
         summary = self.parse_text_marker(raw, "summary", default="Documentation files detected.")
         return {"existing_docs_detected": exists, "existing_docs_summary": summary}
@@ -71,7 +71,7 @@ class DocsContextAnalyzer:
                 f"Existing docs summary:\n{state.get('existing_docs_summary', '')}",
             ]
         )
-        raw = self._llm.generate("docs_strategy", user_input)
+        raw = self._llm.generate("docs_strategy", user_input, log_context="graph.docs.decide_strategy")
         strategy = self.parse_text_marker(raw, "strategy", default="").lower()
         if strategy not in {"incremental_update", "from_scratch"}:
             strategy = "incremental_update" if state.get("existing_docs_detected", False) else "from_scratch"
@@ -260,7 +260,7 @@ class DocsContentComposer:
                 f"Examples bundle:\n{state.get('rules_bundle', '')}",
             ]
         )
-        plan = self._llm.generate("docs_plan_sections", user_input)
+        plan = self._llm.generate("docs_plan_sections", user_input, log_context="graph.docs.plan_incremental_changes")
         return {
             "doc_plan": plan,
             "target_path": target_path,
@@ -279,7 +279,7 @@ class DocsContentComposer:
                 f"Examples bundle:\n{state.get('rules_bundle', '')}",
             ]
         )
-        plan = self._llm.generate("docs_plan_sections", user_input)
+        plan = self._llm.generate("docs_plan_sections", user_input, log_context="graph.docs.plan_new_document")
         return {"doc_plan": plan, "target_path": target_path, "target_file_content": "", "target_file_hash": ""}
 
     def generate_doc_content(self, state: AgentGraphState) -> dict:
@@ -294,7 +294,7 @@ class DocsContentComposer:
                 f"Examples bundle:\n{state.get('rules_bundle', '')}",
             ]
         )
-        raw = self._llm.generate("docs_generation", user_input)
+        raw = self._llm.generate("docs_generation", user_input, log_context="graph.docs.generate_doc_content")
         bundle = self._bundle.parse_docs_bundle(raw)
         if bundle:
             first_content = str(bundle[0].get("content", "")).strip()
@@ -369,7 +369,7 @@ class DocsContentComposer:
                 f"Generated document:\n{generated}",
             ]
         )
-        raw = self._llm.generate("docs_self_check", user_input)
+        raw = self._llm.generate("docs_self_check", user_input, log_context="graph.docs.self_check")
         passed = DocsContextAnalyzer.parse_bool_marker(raw, "pass", default=False)
         feedback = DocsContextAnalyzer.parse_text_marker(raw, "feedback", default="No validation feedback provided.")
         return {"validation_attempts": attempts, "validation_passed": passed, "validation_feedback": feedback}
@@ -379,7 +379,7 @@ class DocsContentComposer:
         bundle = state.get("generated_docs_bundle", []) or []
         strategy = state.get("docs_strategy", "from_scratch")
         if strategy == "from_scratch" and not self._bundle.bundle_has_required_structure(bundle):
-            LOGGER.warning(
+            LOGGER.info(
                 "build_changeset fallback bundle used: strategy=%s bundle_items=%s",
                 strategy,
                 len(bundle),
@@ -452,7 +452,11 @@ class DocsContentComposer:
             ]
         )
         try:
-            summary = self._llm.generate("docs_execution_summary", user_input).strip()
+            summary = self._llm.generate(
+                "docs_execution_summary",
+                user_input,
+                log_context="graph.docs.summarize_result",
+            ).strip()
         except Exception:
             summary = ""
         if not summary:

app/modules/agent/engine/graphs/project_edits_logic.py

@@ -48,7 +48,9 @@ class ProjectEditsLogic:
             },
             ensure_ascii=False,
         )
-        parsed = self._support.parse_json(self._llm.generate("project_edits_plan", user_input))
+        parsed = self._support.parse_json(
+            self._llm.generate("project_edits_plan", user_input, log_context="graph.project_edits.plan_changes")
+        )
         contracts = self._contracts.parse(
             parsed,
             request=str(state.get("message", "")),
@@ -165,7 +167,13 @@ class ProjectEditsLogic:
             "changeset": [{"op": x.op.value, "path": x.path, "reason": x.reason} for x in changeset[:20]],
             "rule": "Changes must stay inside contract blocks and not affect unrelated sections.",
         }
-        parsed = self._support.parse_json(self._llm.generate("project_edits_self_check", json.dumps(payload, ensure_ascii=False)))
+        parsed = self._support.parse_json(
+            self._llm.generate(
+                "project_edits_self_check",
+                json.dumps(payload, ensure_ascii=False),
+                log_context="graph.project_edits.self_check",
+            )
+        )
         passed = bool(parsed.get("pass")) if isinstance(parsed, dict) else False
         feedback = str(parsed.get("feedback", "")).strip() if isinstance(parsed, dict) else ""
         return {
@@ -192,7 +200,11 @@ class ProjectEditsLogic:
             "rag_context": self._support.shorten(state.get("rag_context", ""), 5000),
             "confluence_context": self._support.shorten(state.get("confluence_context", ""), 5000),
         }
-        raw = self._llm.generate("project_edits_hunks", json.dumps(prompt_payload, ensure_ascii=False))
+        raw = self._llm.generate(
+            "project_edits_hunks",
+            json.dumps(prompt_payload, ensure_ascii=False),
+            log_context="graph.project_edits.generate_changeset",
+        )
         parsed = self._support.parse_json(raw)
         hunks = parsed.get("hunks", []) if isinstance(parsed, dict) else []
         if not isinstance(hunks, list) or not hunks:

app/modules/agent/engine/graphs/project_qa_graph.py

@@ -33,7 +33,7 @@ class ProjectQaGraphFactory:
                 f"Confluence context:\n{state.get('confluence_context', '')}",
             ]
         )
-        answer = self._llm.generate("project_answer", user_input)
+        answer = self._llm.generate("project_answer", user_input, log_context="graph.project_qa.answer")
         emit_progress_sync(
             state,
             stage="graph.project_qa.answer.done",

app/modules/agent/engine/graphs/project_qa_step_graphs.py

@@ -0,0 +1,172 @@
+from __future__ import annotations
+import logging
+
+from langgraph.graph import END, START, StateGraph
+
+from app.modules.agent.engine.graphs.progress import emit_progress_sync
+from app.modules.agent.engine.graphs.state import AgentGraphState
+from app.modules.agent.engine.orchestrator.actions.project_qa_analyzer import ProjectQaAnalyzer
+from app.modules.agent.engine.orchestrator.actions.project_qa_support import ProjectQaSupport
+from app.modules.agent.llm import AgentLlmService
+from app.modules.contracts import RagRetriever
+from app.modules.rag.explain import ExplainPack, PromptBudgeter
+
+LOGGER = logging.getLogger(__name__)
+
+
+class ProjectQaConversationGraphFactory:
+    def __init__(self, llm: AgentLlmService | None = None) -> None:
+        self._support = ProjectQaSupport()
+
+    def build(self, checkpointer=None):
+        graph = StateGraph(AgentGraphState)
+        graph.add_node("resolve_request", self._resolve_request)
+        graph.add_edge(START, "resolve_request")
+        graph.add_edge("resolve_request", END)
+        return graph.compile(checkpointer=checkpointer)
+
+    def _resolve_request(self, state: AgentGraphState) -> dict:
+        emit_progress_sync(state, stage="graph.project_qa.conversation_understanding", message="Нормализую пользовательский запрос.")
+        resolved = self._support.resolve_request(str(state.get("message", "") or ""))
+        LOGGER.warning("graph step result: graph=project_qa/conversation_understanding normalized=%s", resolved.get("normalized_message", ""))
+        return {"resolved_request": resolved}
+
+
+class ProjectQaClassificationGraphFactory:
+    def __init__(self, llm: AgentLlmService | None = None) -> None:
+        self._support = ProjectQaSupport()
+
+    def build(self, checkpointer=None):
+        graph = StateGraph(AgentGraphState)
+        graph.add_node("classify_question", self._classify_question)
+        graph.add_edge(START, "classify_question")
+        graph.add_edge("classify_question", END)
+        return graph.compile(checkpointer=checkpointer)
+
+    def _classify_question(self, state: AgentGraphState) -> dict:
+        resolved = state.get("resolved_request", {}) or {}
+        message = str(resolved.get("normalized_message") or state.get("message", "") or "")
+        profile = self._support.build_profile(message)
+        LOGGER.warning("graph step result: graph=project_qa/question_classification domain=%s intent=%s", profile.get("domain"), profile.get("intent"))
+        return {"question_profile": profile}
+
+
+class ProjectQaRetrievalGraphFactory:
+    def __init__(self, rag: RagRetriever, llm: AgentLlmService | None = None) -> None:
+        self._rag = rag
+        self._support = ProjectQaSupport()
+
+    def build(self, checkpointer=None):
+        graph = StateGraph(AgentGraphState)
+        graph.add_node("retrieve_context", self._retrieve_context)
+        graph.add_edge(START, "retrieve_context")
+        graph.add_edge("retrieve_context", END)
+        return graph.compile(checkpointer=checkpointer)
+
+    def _retrieve_context(self, state: AgentGraphState) -> dict:
+        emit_progress_sync(state, stage="graph.project_qa.context_retrieval", message="Собираю контекст по проекту.")
+        resolved = state.get("resolved_request", {}) or {}
+        profile = state.get("question_profile", {}) or {}
+        files_map = dict(state.get("files_map", {}) or {})
+        rag_items: list[dict] = []
+        source_bundle = self._support.build_source_bundle(profile, list(rag_items), files_map)
+        LOGGER.warning(
+            "graph step result: graph=project_qa/context_retrieval mode=%s rag_items=%s file_candidates=%s legacy_rag=%s",
+            profile.get("domain"),
+            len(source_bundle.get("rag_items", []) or []),
+            len(source_bundle.get("file_candidates", []) or []),
+            False,
+        )
+        return {"source_bundle": source_bundle}
+
+
+class ProjectQaAnalysisGraphFactory:
+    def __init__(self, llm: AgentLlmService | None = None) -> None:
+        self._support = ProjectQaSupport()
+        self._analyzer = ProjectQaAnalyzer()
+
+    def build(self, checkpointer=None):
+        graph = StateGraph(AgentGraphState)
+        graph.add_node("analyze_context", self._analyze_context)
+        graph.add_edge(START, "analyze_context")
+        graph.add_edge("analyze_context", END)
+        return graph.compile(checkpointer=checkpointer)
+
+    def _analyze_context(self, state: AgentGraphState) -> dict:
+        explain_pack = state.get("explain_pack")
+        if explain_pack:
+            analysis = self._analysis_from_pack(explain_pack)
+            LOGGER.warning(
+                "graph step result: graph=project_qa/context_analysis findings=%s evidence=%s",
+                len(analysis.get("findings", []) or []),
+                len(analysis.get("evidence", []) or []),
+            )
+            return {"analysis_brief": analysis}
+        bundle = state.get("source_bundle", {}) or {}
+        profile = bundle.get("profile", {}) or state.get("question_profile", {}) or {}
+        rag_items = list(bundle.get("rag_items", []) or [])
+        file_candidates = list(bundle.get("file_candidates", []) or [])
+        analysis = self._analyzer.analyze_code(profile, rag_items, file_candidates) if str(profile.get("domain")) == "code" else self._analyzer.analyze_docs(profile, rag_items)
+        LOGGER.warning(
+            "graph step result: graph=project_qa/context_analysis findings=%s evidence=%s",
+            len(analysis.get("findings", []) or []),
+            len(analysis.get("evidence", []) or []),
+        )
+        return {"analysis_brief": analysis}
+
+    def _analysis_from_pack(self, raw_pack) -> dict:
+        pack = ExplainPack.model_validate(raw_pack)
+        findings: list[str] = []
+        evidence: list[str] = []
+        for entrypoint in pack.selected_entrypoints[:3]:
+            findings.append(f"Entrypoint `{entrypoint.title}` maps to handler `{entrypoint.metadata.get('handler_symbol_id', '')}`.")
+            if entrypoint.source:
+                evidence.append(entrypoint.source)
+        for path in pack.trace_paths[:3]:
+            if path.symbol_ids:
+                findings.append(f"Trace path: {' -> '.join(path.symbol_ids)}")
+        for excerpt in pack.code_excerpts[:4]:
+            evidence.append(f"{excerpt.path}:{excerpt.start_line}-{excerpt.end_line} [{excerpt.evidence_id}]")
+        return {
+            "subject": pack.intent.normalized_query,
+            "findings": findings or ["No explain trace was built from the available code evidence."],
+            "evidence": evidence,
+            "gaps": list(pack.missing),
+            "answer_mode": "summary",
+        }
+
+
+class ProjectQaAnswerGraphFactory:
+    def __init__(self, llm: AgentLlmService | None = None) -> None:
+        self._support = ProjectQaSupport()
+        self._llm = llm
+        self._budgeter = PromptBudgeter()
+
+    def build(self, checkpointer=None):
+        graph = StateGraph(AgentGraphState)
+        graph.add_node("compose_answer", self._compose_answer)
+        graph.add_edge(START, "compose_answer")
+        graph.add_edge("compose_answer", END)
+        return graph.compile(checkpointer=checkpointer)
+
+    def _compose_answer(self, state: AgentGraphState) -> dict:
+        profile = state.get("question_profile", {}) or {}
+        analysis = state.get("analysis_brief", {}) or {}
+        brief = self._support.build_answer_brief(profile, analysis)
+        explain_pack = state.get("explain_pack")
+        answer = self._compose_explain_answer(state, explain_pack)
+        if not answer:
+            answer = self._support.compose_answer(brief)
+        LOGGER.warning("graph step result: graph=project_qa/answer_composition answer_len=%s", len(answer or ""))
+        return {"answer_brief": brief, "final_answer": answer}
+
+    def _compose_explain_answer(self, state: AgentGraphState, raw_pack) -> str:
+        if raw_pack is None or self._llm is None:
+            return ""
+        pack = ExplainPack.model_validate(raw_pack)
+        prompt_input = self._budgeter.build_prompt_input(str(state.get("message", "") or ""), pack)
+        return self._llm.generate(
+            "code_explain_answer_v2",
+            prompt_input,
+            log_context="graph.project_qa.answer_v2",
+        ).strip()

app/modules/agent/engine/graphs/state.py

@@ -25,6 +25,12 @@ class AgentGraphState(TypedDict, total=False):
     validation_passed: bool
     validation_feedback: str
     validation_attempts: int
+    resolved_request: dict
+    question_profile: dict
+    source_bundle: dict
+    analysis_brief: dict
+    answer_brief: dict
+    final_answer: str
     answer: str
     changeset: list[ChangeItem]
     edits_requested_path: str

app/modules/agent/engine/orchestrator/actions/__init__.py

@@ -1,13 +1,17 @@
+from app.modules.agent.engine.orchestrator.actions.code_explain_actions import CodeExplainActions
 from app.modules.agent.engine.orchestrator.actions.docs_actions import DocsActions
 from app.modules.agent.engine.orchestrator.actions.edit_actions import EditActions
 from app.modules.agent.engine.orchestrator.actions.explain_actions import ExplainActions
 from app.modules.agent.engine.orchestrator.actions.gherkin_actions import GherkinActions
+from app.modules.agent.engine.orchestrator.actions.project_qa_actions import ProjectQaActions
 from app.modules.agent.engine.orchestrator.actions.review_actions import ReviewActions
 
 __all__ = [
+    "CodeExplainActions",
     "DocsActions",
     "EditActions",
     "ExplainActions",
     "GherkinActions",
+    "ProjectQaActions",
     "ReviewActions",
 ]

app/modules/agent/engine/orchestrator/actions/code_explain_actions.py

@@ -0,0 +1,46 @@
+from __future__ import annotations
+
+import logging
+
+from typing import TYPE_CHECKING
+
+from app.modules.agent.engine.orchestrator.actions.common import ActionSupport
+from app.modules.agent.engine.orchestrator.execution_context import ExecutionContext
+from app.modules.agent.engine.orchestrator.models import ArtifactType
+from app.modules.rag.explain.intent_builder import ExplainIntentBuilder
+from app.modules.rag.explain.models import ExplainPack
+
+if TYPE_CHECKING:
+    from app.modules.rag.explain.retriever_v2 import CodeExplainRetrieverV2
+
+LOGGER = logging.getLogger(__name__)
+
+
+class CodeExplainActions(ActionSupport):
+    def __init__(self, retriever: CodeExplainRetrieverV2 | None = None) -> None:
+        self._retriever = retriever
+        self._intent_builder = ExplainIntentBuilder()
+
+    def build_code_explain_pack(self, ctx: ExecutionContext) -> list[str]:
+        file_candidates = list((self.get(ctx, "source_bundle", {}) or {}).get("file_candidates", []) or [])
+        if self._retriever is None:
+            pack = ExplainPack(
+                intent=self._intent_builder.build(ctx.task.user_message),
+                missing=["code_explain_retriever_unavailable"],
+            )
+        else:
+            pack = self._retriever.build_pack(
+                ctx.task.rag_session_id,
+                ctx.task.user_message,
+                file_candidates=file_candidates,
+            )
+        LOGGER.warning(
+            "code explain action: task_id=%s entrypoints=%s seeds=%s paths=%s excerpts=%s missing=%s",
+            ctx.task.task_id,
+            len(pack.selected_entrypoints),
+            len(pack.seed_symbols),
+            len(pack.trace_paths),
+            len(pack.code_excerpts),
+            pack.missing,
+        )
+        return [self.put(ctx, "explain_pack", ArtifactType.STRUCTURED_JSON, pack.model_dump(mode="json"))]

app/modules/agent/engine/orchestrator/actions/project_qa_actions.py

@@ -0,0 +1,117 @@
+from __future__ import annotations
+
+from app.modules.agent.engine.orchestrator.actions.project_qa_analyzer import ProjectQaAnalyzer
+from app.modules.agent.engine.orchestrator.actions.common import ActionSupport
+from app.modules.agent.engine.orchestrator.actions.project_qa_support import ProjectQaSupport
+from app.modules.agent.engine.orchestrator.execution_context import ExecutionContext
+from app.modules.agent.engine.orchestrator.models import ArtifactType
+
+
+class ProjectQaActions(ActionSupport):
+    def __init__(self) -> None:
+        self._support = ProjectQaSupport()
+        self._analyzer = ProjectQaAnalyzer()
+
+    def classify_project_question(self, ctx: ExecutionContext) -> list[str]:
+        message = str(ctx.task.user_message or "")
+        profile = self._support.build_profile(message)
+        return [self.put(ctx, "question_profile", ArtifactType.STRUCTURED_JSON, profile)]
+
+    def collect_project_sources(self, ctx: ExecutionContext) -> list[str]:
+        profile = self.get(ctx, "question_profile", {}) or {}
+        terms = list(profile.get("terms", []) or [])
+        entities = list(profile.get("entities", []) or [])
+        rag_items = list(ctx.task.metadata.get("rag_items", []) or [])
+        files_map = dict(ctx.task.metadata.get("files_map", {}) or {})
+        explicit_test = any(term in {"test", "tests", "тест", "тесты"} for term in terms)
+
+        ranked_rag = []
+        for item in rag_items:
+            score = self._support.rag_score(item, terms, entities)
+            source = str(item.get("source", "") or "")
+            if not explicit_test and self._support.is_test_path(source):
+                score -= 3
+            if score > 0:
+                ranked_rag.append((score, item))
+        ranked_rag.sort(key=lambda pair: pair[0], reverse=True)
+
+        ranked_files = []
+        for path, payload in files_map.items():
+            score = self._support.file_score(path, payload, terms, entities)
+            if not explicit_test and self._support.is_test_path(path):
+                score -= 3
+            if score > 0:
+                ranked_files.append(
+                    (
+                        score,
+                        {
+                            "path": path,
+                            "content": str(payload.get("content", "")),
+                            "content_hash": str(payload.get("content_hash", "")),
+                        },
+                    )
+                )
+        ranked_files.sort(key=lambda pair: pair[0], reverse=True)
+
+        bundle = {
+            "profile": profile,
+            "rag_items": [item for _, item in ranked_rag[:12]],
+            "file_candidates": [item for _, item in ranked_files[:10]],
+            "rag_total": len(ranked_rag),
+            "files_total": len(ranked_files),
+        }
+        return [self.put(ctx, "source_bundle", ArtifactType.STRUCTURED_JSON, bundle)]
+
+    def analyze_project_sources(self, ctx: ExecutionContext) -> list[str]:
+        bundle = self.get(ctx, "source_bundle", {}) or {}
+        profile = bundle.get("profile", {}) or {}
+        rag_items = list(bundle.get("rag_items", []) or [])
+        file_candidates = list(bundle.get("file_candidates", []) or [])
+
+        if str(profile.get("domain")) == "code":
+            analysis = self._analyzer.analyze_code(profile, rag_items, file_candidates)
+        else:
+            analysis = self._analyzer.analyze_docs(profile, rag_items)
+        return [self.put(ctx, "analysis_brief", ArtifactType.STRUCTURED_JSON, analysis)]
+
+    def build_project_answer_brief(self, ctx: ExecutionContext) -> list[str]:
+        profile = self.get(ctx, "question_profile", {}) or {}
+        analysis = self.get(ctx, "analysis_brief", {}) or {}
+        brief = {
+            "question_profile": profile,
+            "resolved_subject": analysis.get("subject"),
+            "key_findings": analysis.get("findings", []),
+            "supporting_evidence": analysis.get("evidence", []),
+            "missing_evidence": analysis.get("gaps", []),
+            "answer_mode": analysis.get("answer_mode", "summary"),
+        }
+        return [self.put(ctx, "answer_brief", ArtifactType.STRUCTURED_JSON, brief)]
+
+    def compose_project_answer(self, ctx: ExecutionContext) -> list[str]:
+        brief = self.get(ctx, "answer_brief", {}) or {}
+        profile = brief.get("question_profile", {}) or {}
+        russian = bool(profile.get("russian"))
+        answer_mode = str(brief.get("answer_mode") or "summary")
+        findings = list(brief.get("key_findings", []) or [])
+        evidence = list(brief.get("supporting_evidence", []) or [])
+        gaps = list(brief.get("missing_evidence", []) or [])
+
+        title = "## Кратко" if russian else "## Summary"
+        lines = [title]
+        if answer_mode == "inventory":
+            lines.append("### Что реализовано" if russian else "### Implemented items")
+        else:
+            lines.append("### Что видно по проекту" if russian else "### What the project shows")
+        if findings:
+            lines.extend(f"- {item}" for item in findings)
+        else:
+            lines.append("Не удалось собрать подтвержденные выводы по доступным данным." if russian else "No supported findings could be assembled from the available data.")
+        if evidence:
+            lines.append("")
+            lines.append("### Где смотреть в проекте" if russian else "### Where to look in the project")
+            lines.extend(f"- `{item}`" for item in evidence[:5])
+        if gaps:
+            lines.append("")
+            lines.append("### Что пока не подтверждено кодом" if russian else "### What is not yet confirmed in code")
+            lines.extend(f"- {item}" for item in gaps[:3])
+        return [self.put(ctx, "final_answer", ArtifactType.TEXT, "\n".join(lines))]

app/modules/agent/engine/orchestrator/actions/project_qa_analyzer.py

@@ -0,0 +1,154 @@
+from __future__ import annotations
+
+
+class ProjectQaAnalyzer:
+    def analyze_code(self, profile: dict, rag_items: list[dict], file_candidates: list[dict]) -> dict:
+        terms = list(profile.get("terms", []) or [])
+        intent = str(profile.get("intent") or "lookup")
+        russian = bool(profile.get("russian"))
+        findings: list[str] = []
+        evidence: list[str] = []
+        gaps: list[str] = []
+
+        symbol_titles = [str(item.get("title", "") or "") for item in rag_items if str(item.get("layer", "")).startswith("C1")]
+        symbol_set = set(symbol_titles)
+        file_paths = [str(item.get("path", "") or item.get("source", "") or "") for item in rag_items]
+        file_paths.extend(str(item.get("path", "") or "") for item in file_candidates)
+
+        if "ConfigManager" in profile.get("entities", []) or "configmanager" in terms or "config_manager" in terms:
+            alias_file = self.find_path(file_paths, "src/config_manager/__init__.py")
+            if alias_file:
+                findings.append(
+                    "Публичный `ConfigManager` экспортируется из `src/config_manager/__init__.py` как alias на `ConfigManagerV2`."
+                    if russian
+                    else "Public `ConfigManager` is exported from `src/config_manager/__init__.py` as an alias to `ConfigManagerV2`."
+                )
+                evidence.append("src/config_manager/__init__.py")
+
+        if "controlchannel" in {name.lower() for name in symbol_set}:
+            findings.append(
+                "Базовый контракт управления задает `ControlChannel`: он определяет команды `start` и `stop` для внешнего канала управления."
+                if russian
+                else "`ControlChannel` defines the base management contract with `start` and `stop` commands."
+            )
+            evidence.append("src/config_manager/v2/control/base.py")
+
+        if "ControlChannelBridge" in symbol_set:
+            findings.append(
+                "`ControlChannelBridge` связывает внешний канал управления с lifecycle-методами менеджера: `on_start`, `on_stop`, `on_status`."
+                if russian
+                else "`ControlChannelBridge` connects the external control channel to manager lifecycle methods: `on_start`, `on_stop`, `on_status`."
+            )
+            evidence.append("src/config_manager/v2/core/control_bridge.py")
+
+        implementation_files = self.find_management_implementations(file_candidates)
+        if implementation_files:
+            labels = ", ".join(f"`{path}`" for path in implementation_files)
+            channel_names = self.implementation_names(implementation_files)
+            findings.append(
+                f"В коде найдены конкретные реализации каналов управления: {', '.join(channel_names)} ({labels})."
+                if russian
+                else f"Concrete management channel implementations were found in code: {', '.join(channel_names)} ({labels})."
+            )
+            evidence.extend(implementation_files)
+        elif intent == "inventory":
+            gaps.append(
+                "В текущем контексте не удалось уверенно подтвердить конкретные файлы-реализации каналов, кроме базового контракта и bridge-слоя."
+                if russian
+                else "The current context does not yet confirm concrete channel implementation files beyond the base contract and bridge layer."
+            )
+
+        package_doc = self.find_management_doc(file_candidates)
+        if package_doc:
+            findings.append(
+                f"Пакет управления прямо описывает внешние каналы через `{package_doc}`."
+                if russian
+                else f"The control package directly describes external channels in `{package_doc}`."
+            )
+            evidence.append(package_doc)
+
+        subject = "management channels"
+        if profile.get("entities"):
+            subject = ", ".join(profile["entities"])
+        return {
+            "subject": subject,
+            "findings": self.dedupe(findings),
+            "evidence": self.dedupe(evidence),
+            "gaps": gaps,
+            "answer_mode": "inventory" if intent == "inventory" else "summary",
+        }
+
+    def analyze_docs(self, profile: dict, rag_items: list[dict]) -> dict:
+        findings: list[str] = []
+        evidence: list[str] = []
+        for item in rag_items[:5]:
+            title = str(item.get("title", "") or "")
+            source = str(item.get("source", "") or "")
+            content = str(item.get("content", "") or "").strip()
+            if content:
+                findings.append(content.splitlines()[0][:220])
+            if source:
+                evidence.append(source)
+            elif title:
+                evidence.append(title)
+        return {
+            "subject": "docs",
+            "findings": self.dedupe(findings),
+            "evidence": self.dedupe(evidence),
+            "gaps": [] if findings else ["Недостаточно данных в документации." if profile.get("russian") else "Not enough data in documentation."],
+            "answer_mode": "summary",
+        }
+
+    def find_management_implementations(self, file_candidates: list[dict]) -> list[str]:
+        found: list[str] = []
+        for item in file_candidates:
+            path = str(item.get("path", "") or "")
+            lowered = path.lower()
+            if self.is_test_path(path):
+                continue
+            if any(token in lowered for token in ("http_channel.py", "telegram.py", "telegram_channel.py", "http.py")):
+                found.append(path)
+                continue
+            content = str(item.get("content", "") or "").lower()
+            if "controlchannel" in content and "class " in content:
+                found.append(path)
+                continue
+            if ("channel" in lowered or "control" in lowered) and any(token in content for token in ("http", "telegram", "bot")):
+                found.append(path)
+        return self.dedupe(found)[:4]
+
+    def implementation_names(self, paths: list[str]) -> list[str]:
+        names: list[str] = []
+        for path in paths:
+            stem = path.rsplit("/", 1)[-1].rsplit(".", 1)[0]
+            label = stem.replace("_", " ").strip()
+            if label and label not in names:
+                names.append(label)
+        return names
+
+    def find_management_doc(self, file_candidates: list[dict]) -> str | None:
+        for item in file_candidates:
+            path = str(item.get("path", "") or "")
+            if self.is_test_path(path):
+                continue
+            content = str(item.get("content", "") or "").lower()
+            if any(token in content for token in ("каналы внешнего управления", "external control channels", "http api", "telegram")):
+                return path
+        return None
+
+    def find_path(self, paths: list[str], target: str) -> str | None:
+        for path in paths:
+            if path == target:
+                return path
+        return None
+
+    def dedupe(self, items: list[str]) -> list[str]:
+        seen: list[str] = []
+        for item in items:
+            if item and item not in seen:
+                seen.append(item)
+        return seen
+
+    def is_test_path(self, path: str) -> bool:
+        lowered = path.lower()
+        return lowered.startswith("tests/") or "/tests/" in lowered or lowered.startswith("test_") or "/test_" in lowered

app/modules/agent/engine/orchestrator/actions/project_qa_support.py

@@ -0,0 +1,166 @@
+from __future__ import annotations
+
+import re
+
+from app.modules.rag.retrieval.query_terms import extract_query_terms
+
+
+class ProjectQaSupport:
+    def resolve_request(self, message: str) -> dict:
+        profile = self.build_profile(message)
+        subject = profile["entities"][0] if profile.get("entities") else ""
+        return {
+            "original_message": message,
+            "normalized_message": " ".join((message or "").split()),
+            "subject_hint": subject,
+            "source_hint": profile["domain"],
+            "russian": profile["russian"],
+        }
+
+    def build_profile(self, message: str) -> dict:
+        lowered = message.lower()
+        return {
+            "domain": "code" if self.looks_like_code_question(lowered) else "docs",
+            "intent": self.detect_intent(lowered),
+            "terms": extract_query_terms(message),
+            "entities": self.extract_entities(message),
+            "russian": self.is_russian(message),
+        }
+
+    def build_retrieval_query(self, resolved_request: dict, profile: dict) -> str:
+        normalized = str(resolved_request.get("normalized_message") or resolved_request.get("original_message") or "").strip()
+        if profile.get("domain") == "code" and "по коду" not in normalized.lower():
+            return f"по коду {normalized}".strip()
+        return normalized
+
+    def build_source_bundle(self, profile: dict, rag_items: list[dict], files_map: dict[str, dict]) -> dict:
+        terms = list(profile.get("terms", []) or [])
+        entities = list(profile.get("entities", []) or [])
+        explicit_test = any(term in {"test", "tests", "тест", "тесты"} for term in terms)
+
+        ranked_rag: list[tuple[int, dict]] = []
+        for item in rag_items:
+            score = self.rag_score(item, terms, entities)
+            source = str(item.get("source", "") or "")
+            if not explicit_test and self.is_test_path(source):
+                score -= 3
+            if score > 0:
+                ranked_rag.append((score, item))
+        ranked_rag.sort(key=lambda pair: pair[0], reverse=True)
+
+        ranked_files: list[tuple[int, dict]] = []
+        for path, payload in files_map.items():
+            score = self.file_score(path, payload, terms, entities)
+            if not explicit_test and self.is_test_path(path):
+                score -= 3
+            if score > 0:
+                ranked_files.append(
+                    (
+                        score,
+                        {
+                            "path": path,
+                            "content": str(payload.get("content", "")),
+                            "content_hash": str(payload.get("content_hash", "")),
+                        },
+                    )
+                )
+        ranked_files.sort(key=lambda pair: pair[0], reverse=True)
+
+        return {
+            "profile": profile,
+            "rag_items": [item for _, item in ranked_rag[:12]],
+            "file_candidates": [item for _, item in ranked_files[:10]],
+            "rag_total": len(ranked_rag),
+            "files_total": len(ranked_files),
+        }
+
+    def build_answer_brief(self, profile: dict, analysis: dict) -> dict:
+        return {
+            "question_profile": profile,
+            "resolved_subject": analysis.get("subject"),
+            "key_findings": analysis.get("findings", []),
+            "supporting_evidence": analysis.get("evidence", []),
+            "missing_evidence": analysis.get("gaps", []),
+            "answer_mode": analysis.get("answer_mode", "summary"),
+        }
+
+    def compose_answer(self, brief: dict) -> str:
+        profile = brief.get("question_profile", {}) or {}
+        russian = bool(profile.get("russian"))
+        answer_mode = str(brief.get("answer_mode") or "summary")
+        findings = list(brief.get("key_findings", []) or [])
+        evidence = list(brief.get("supporting_evidence", []) or [])
+        gaps = list(brief.get("missing_evidence", []) or [])
+
+        title = "## Кратко" if russian else "## Summary"
+        lines = [title]
+        lines.append("### Что реализовано" if answer_mode == "inventory" and russian else "### Implemented items" if answer_mode == "inventory" else "### Что видно по проекту" if russian else "### What the project shows")
+        if findings:
+            lines.extend(f"- {item}" for item in findings)
+        else:
+            lines.append("Не удалось собрать подтвержденные выводы по доступным данным." if russian else "No supported findings could be assembled from the available data.")
+        if evidence:
+            lines.append("")
+            lines.append("### Где смотреть в проекте" if russian else "### Where to look in the project")
+            lines.extend(f"- `{item}`" for item in evidence[:5])
+        if gaps:
+            lines.append("")
+            lines.append("### Что пока не подтверждено кодом" if russian else "### What is not yet confirmed in code")
+            lines.extend(f"- {item}" for item in gaps[:3])
+        return "\n".join(lines)
+
+    def detect_intent(self, lowered: str) -> str:
+        if any(token in lowered for token in ("какие", "что уже реализ", "список", "перечень", "какие есть")):
+            return "inventory"
+        if any(token in lowered for token in ("где", "find", "where")):
+            return "lookup"
+        if any(token in lowered for token in ("сравни", "compare")):
+            return "compare"
+        return "explain"
+
+    def looks_like_code_question(self, lowered: str) -> bool:
+        code_markers = ("по коду", "код", "реализ", "имплементац", "класс", "метод", "модул", "файл", "канал", "handler", "endpoint")
+        return any(marker in lowered for marker in code_markers) or bool(re.search(r"\b[A-Z][A-Za-z0-9_]{2,}\b", lowered))
+
+    def extract_entities(self, message: str) -> list[str]:
+        return re.findall(r"\b[A-Z][A-Za-z0-9_]{2,}\b", message)[:5]
+
+    def rag_score(self, item: dict, terms: list[str], entities: list[str]) -> int:
+        haystacks = [
+            str(item.get("source", "") or "").lower(),
+            str(item.get("title", "") or "").lower(),
+            str(item.get("content", "") or "").lower(),
+            str((item.get("metadata", {}) or {}).get("qname", "") or "").lower(),
+        ]
+        score = 0
+        for term in terms:
+            if any(term in hay for hay in haystacks):
+                score += 3
+        for entity in entities:
+            if any(entity.lower() in hay for hay in haystacks):
+                score += 5
+        return score
+
+    def file_score(self, path: str, payload: dict, terms: list[str], entities: list[str]) -> int:
+        content = str(payload.get("content", "") or "").lower()
+        path_lower = path.lower()
+        score = 0
+        for term in terms:
+            if term in path_lower:
+                score += 4
+            elif term in content:
+                score += 2
+        for entity in entities:
+            entity_lower = entity.lower()
+            if entity_lower in path_lower:
+                score += 5
+            elif entity_lower in content:
+                score += 3
+        return score
+
+    def is_test_path(self, path: str) -> bool:
+        lowered = path.lower()
+        return lowered.startswith("tests/") or "/tests/" in lowered or lowered.startswith("test_") or "/test_" in lowered
+
+    def is_russian(self, text: str) -> bool:
+        return any("а" <= ch.lower() <= "я" or ch.lower() == "ё" for ch in text)

app/modules/agent/engine/orchestrator/execution_engine.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import asyncio
 import inspect
+import logging
 import time
 
 from app.modules.agent.engine.orchestrator.execution_context import ExecutionContext
@@ -9,6 +10,8 @@ from app.modules.agent.engine.orchestrator.models import PlanStatus, PlanStep, S
 from app.modules.agent.engine.orchestrator.quality_gates import QualityGateRunner
 from app.modules.agent.engine.orchestrator.step_registry import StepRegistry
 
+LOGGER = logging.getLogger(__name__)
+
 
 class ExecutionEngine:
     def __init__(self, step_registry: StepRegistry, gates: QualityGateRunner) -> None:
@@ -22,17 +25,18 @@ class ExecutionEngine:
         for step in ctx.plan.steps:
             dep_issue = self._dependency_issue(step, step_results)
             if dep_issue:
-                step_results.append(
-                    StepResult(
+                result = StepResult(
                         step_id=step.step_id,
                         status=StepStatus.SKIPPED,
                         warnings=[dep_issue],
                     )
-                )
+                step_results.append(result)
+                self._log_step_result(ctx, step, result)
                 continue
 
             result = await self._run_with_retry(step, ctx)
             step_results.append(result)
+            self._log_step_result(ctx, step, result)
             if result.status in {StepStatus.FAILED, StepStatus.RETRY_EXHAUSTED} and step.on_failure == "fail":
                 ctx.plan.status = PlanStatus.FAILED
                 return step_results
@@ -65,6 +69,15 @@ class ExecutionEngine:
         while attempt < max_attempts:
             attempt += 1
             started_at = time.monotonic()
+            LOGGER.warning(
+                "orchestrator step start: task_id=%s step_id=%s action_id=%s executor=%s attempt=%s graph_id=%s",
+                ctx.task.task_id,
+                step.step_id,
+                step.action_id,
+                step.executor,
+                attempt,
+                step.graph_id or "",
+            )
             await self._emit_progress(ctx, f"orchestrator.step.{step.step_id}", step.title)
 
             try:
@@ -113,3 +126,21 @@ class ExecutionEngine:
         result = ctx.progress_cb(stage, message, "task_progress", {"layer": "orchestrator"})
         if inspect.isawaitable(result):
             await result
+
+    def _log_step_result(self, ctx: ExecutionContext, step: PlanStep, result: StepResult) -> None:
+        artifact_keys = []
+        for artifact_id in result.produced_artifact_ids:
+            item = next((artifact for artifact in ctx.artifacts.all_items() if artifact.artifact_id == artifact_id), None)
+            if item is not None:
+                artifact_keys.append(item.key)
+        LOGGER.warning(
+            "orchestrator step result: task_id=%s step_id=%s action_id=%s status=%s duration_ms=%s artifact_keys=%s warnings=%s error=%s",
+            ctx.task.task_id,
+            step.step_id,
+            step.action_id,
+            result.status.value,
+            result.duration_ms,
+            artifact_keys,
+            result.warnings,
+            result.error_message or "",
+        )

app/modules/agent/engine/orchestrator/service.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import inspect
+import logging
 
 from app.core.exceptions import AppError
 from app.modules.agent.engine.orchestrator.execution_context import ExecutionContext, GraphInvoker, GraphResolver, ProgressCallback
@@ -14,6 +15,8 @@ from app.modules.agent.engine.orchestrator.step_registry import StepRegistry
 from app.modules.agent.engine.orchestrator.template_registry import ScenarioTemplateRegistry
 from app.schemas.common import ModuleName
 
+LOGGER = logging.getLogger(__name__)
+
 
 class OrchestratorService:
     def __init__(
@@ -74,6 +77,21 @@ class OrchestratorService:
             )
         result = self._assembler.assemble(ctx, step_results)
         await self._emit_progress(progress_cb, "orchestrator.done", "Execution plan completed.")
+        LOGGER.warning(
+            "orchestrator decision: task_id=%s scenario=%s plan_status=%s steps=%s changeset_items=%s answer_len=%s",
+            task.task_id,
+            task.scenario.value,
+            result.meta.get("plan", {}).get("status", ""),
+            [
+                {
+                    "step_id": step.step_id,
+                    "status": step.status.value,
+                }
+                for step in result.steps
+            ],
+            len(result.changeset),
+            len(result.answer or ""),
+        )
         return result
 
     async def _emit_progress(self, progress_cb: ProgressCallback | None, stage: str, message: str) -> None:

app/modules/agent/engine/orchestrator/step_registry.py

@@ -2,29 +2,50 @@ from __future__ import annotations
 
 import asyncio
 from collections.abc import Callable
+from typing import TYPE_CHECKING
 
 from app.modules.agent.engine.graphs.progress_registry import progress_registry
-from app.modules.agent.engine.orchestrator.actions import DocsActions, EditActions, ExplainActions, GherkinActions, ReviewActions
+from app.modules.agent.engine.orchestrator.actions import (
+    CodeExplainActions,
+    DocsActions,
+    EditActions,
+    ExplainActions,
+    GherkinActions,
+    ProjectQaActions,
+    ReviewActions,
+)
 from app.modules.agent.engine.orchestrator.execution_context import ExecutionContext
 from app.modules.agent.engine.orchestrator.models import ArtifactType, PlanStep
 
+if TYPE_CHECKING:
+    from app.modules.rag.explain.retriever_v2 import CodeExplainRetrieverV2
+
 StepFn = Callable[[ExecutionContext], list[str]]
 
 
 class StepRegistry:
-    def __init__(self) -> None:
+    def __init__(self, code_explain_retriever: CodeExplainRetrieverV2 | None = None) -> None:
+        code_explain = CodeExplainActions(code_explain_retriever)
         explain = ExplainActions()
         review = ReviewActions()
         docs = DocsActions()
         edits = EditActions()
         gherkin = GherkinActions()
+        project_qa = ProjectQaActions()
 
         self._functions: dict[str, StepFn] = {
             "collect_state": self._collect_state,
             "finalize_graph_output": self._finalize_graph_output,
+            "execute_project_qa_graph": self._collect_state,
+            "build_code_explain_pack": code_explain.build_code_explain_pack,
             "collect_sources": explain.collect_sources,
             "extract_logic": explain.extract_logic,
             "summarize": explain.summarize,
+            "classify_project_question": project_qa.classify_project_question,
+            "collect_project_sources": project_qa.collect_project_sources,
+            "analyze_project_sources": project_qa.analyze_project_sources,
+            "build_project_answer_brief": project_qa.build_project_answer_brief,
+            "compose_project_answer": project_qa.compose_project_answer,
             "fetch_source_doc": review.fetch_source_doc,
             "normalize_document": review.normalize_document,
             "structural_check": review.structural_check,
@@ -66,6 +87,7 @@ class StepRegistry:
         state = {
             "task_id": ctx.task.task_id,
             "project_id": ctx.task.rag_session_id,
+            "scenario": ctx.task.scenario.value,
             "message": ctx.task.user_message,
             "progress_key": ctx.task.task_id,
             "rag_context": str(ctx.task.metadata.get("rag_context", "")),
@@ -86,7 +108,7 @@ class StepRegistry:
             raise RuntimeError(f"Unsupported graph_id: {graph_key}")
 
         graph = ctx.graph_resolver(domain_id, process_id)
-        state = ctx.artifacts.get_content("agent_state", {}) or {}
+        state = self._build_graph_state(ctx)
 
         if ctx.progress_cb is not None:
             progress_registry.register(ctx.task.task_id, ctx.progress_cb)
@@ -96,8 +118,29 @@ class StepRegistry:
             if ctx.progress_cb is not None:
                 progress_registry.unregister(ctx.task.task_id)
 
-        item = ctx.artifacts.put(key="graph_result", artifact_type=ArtifactType.STRUCTURED_JSON, content=result)
-        return [item.artifact_id]
+        return self._store_graph_outputs(step, ctx, result)
+
+    def _build_graph_state(self, ctx: ExecutionContext) -> dict:
+        state = dict(ctx.artifacts.get_content("agent_state", {}) or {})
+        for item in ctx.artifacts.all_items():
+            state[item.key] = ctx.artifacts.get_content(item.key)
+        return state
+
+    def _store_graph_outputs(self, step: PlanStep, ctx: ExecutionContext, result: dict) -> list[str]:
+        if not isinstance(result, dict):
+            raise RuntimeError("graph_result must be an object")
+        if len(step.outputs) == 1 and step.outputs[0].key == "graph_result":
+            item = ctx.artifacts.put(key="graph_result", artifact_type=ArtifactType.STRUCTURED_JSON, content=result)
+            return [item.artifact_id]
+
+        artifact_ids: list[str] = []
+        for output in step.outputs:
+            value = result.get(output.key)
+            if value is None and output.required:
+                raise RuntimeError(f"graph_output_missing:{step.step_id}:{output.key}")
+            item = ctx.artifacts.put(key=output.key, artifact_type=output.type, content=value)
+            artifact_ids.append(item.artifact_id)
+        return artifact_ids
 
     def _finalize_graph_output(self, ctx: ExecutionContext) -> list[str]:
         raw = ctx.artifacts.get_content("graph_result", {}) or {}

app/modules/agent/engine/orchestrator/template_registry.py

@@ -16,6 +16,8 @@ class ScenarioTemplateRegistry:
         return builders.get(task.scenario, self._general)(task)
 
     def _general(self, task: TaskSpec) -> ExecutionPlan:
+        if task.routing.domain_id == "project" and task.routing.process_id == "qa":
+            return self._project_qa(task)
         steps = [
             self._step("collect_state", "Collect state", "collect_state", outputs=[self._out("agent_state", ArtifactType.STRUCTURED_JSON)]),
             self._step(
@@ -39,7 +41,77 @@ class ScenarioTemplateRegistry:
         ]
         return self._plan(task, "general_qa_v1", steps, [self._gate("non_empty_answer_or_changeset")])
 
+    def _project_qa(self, task: TaskSpec) -> ExecutionPlan:
+        steps = [
+            self._step("collect_state", "Collect state", "collect_state", outputs=[self._out("agent_state", ArtifactType.STRUCTURED_JSON)]),
+            self._step(
+                "conversation_understanding",
+                "Conversation understanding",
+                "execute_project_qa_graph",
+                executor="graph",
+                graph_id="project_qa/conversation_understanding",
+                depends_on=["collect_state"],
+                outputs=[self._out("resolved_request", ArtifactType.STRUCTURED_JSON)],
+            ),
+            self._step(
+                "question_classification",
+                "Question classification",
+                "execute_project_qa_graph",
+                executor="graph",
+                graph_id="project_qa/question_classification",
+                depends_on=["conversation_understanding"],
+                outputs=[self._out("question_profile", ArtifactType.STRUCTURED_JSON)],
+            ),
+            self._step(
+                "context_retrieval",
+                "Context retrieval",
+                "execute_project_qa_graph",
+                executor="graph",
+                graph_id="project_qa/context_retrieval",
+                depends_on=["question_classification"],
+                outputs=[self._out("source_bundle", ArtifactType.STRUCTURED_JSON)],
+            ),
+        ]
+        analysis_depends_on = ["context_retrieval"]
+        if task.scenario == Scenario.EXPLAIN_PART:
+            steps.append(
+                self._step(
+                    "code_explain_pack_step",
+                    "Build code explain pack",
+                    "build_code_explain_pack",
+                    depends_on=["context_retrieval"],
+                    outputs=[self._out("explain_pack", ArtifactType.STRUCTURED_JSON)],
+                )
+            )
+            analysis_depends_on = ["code_explain_pack_step"]
+        steps.extend(
+            [
+                self._step(
+                    "context_analysis",
+                    "Context analysis",
+                    "execute_project_qa_graph",
+                    executor="graph",
+                    graph_id="project_qa/context_analysis",
+                    depends_on=analysis_depends_on,
+                    outputs=[self._out("analysis_brief", ArtifactType.STRUCTURED_JSON)],
+                ),
+                self._step(
+                    "answer_composition",
+                    "Answer composition",
+                    "execute_project_qa_graph",
+                    executor="graph",
+                    graph_id="project_qa/answer_composition",
+                    depends_on=["context_analysis"],
+                    outputs=[self._out("answer_brief", ArtifactType.STRUCTURED_JSON, required=False), self._out("final_answer", ArtifactType.TEXT)],
+                    gates=[self._gate("non_empty_answer_or_changeset")],
+                ),
+            ]
+        )
+        return self._plan(task, "project_qa_reasoning_v1", steps, [self._gate("non_empty_answer_or_changeset")])
+
     def _explain(self, task: TaskSpec) -> ExecutionPlan:
+        if task.routing.domain_id == "project" and task.routing.process_id == "qa":
+            return self._project_qa(task)
         steps = [
             self._step("collect_sources", "Collect sources", "collect_sources", outputs=[self._out("sources", ArtifactType.STRUCTURED_JSON)]),
             self._step("extract_logic", "Extract logic", "extract_logic", depends_on=["collect_sources"], outputs=[self._out("logic_model", ArtifactType.STRUCTURED_JSON)]),

app/modules/agent/engine/router/__init__.py

@@ -2,21 +2,28 @@ from pathlib import Path
 from typing import TYPE_CHECKING
 
 from app.modules.agent.llm import AgentLlmService
+from app.modules.contracts import RagRetriever
 
 if TYPE_CHECKING:
     from app.modules.agent.repository import AgentRepository
     from app.modules.agent.engine.router.router_service import RouterService
 
 
-def build_router_service(llm: AgentLlmService, agent_repository: "AgentRepository") -> "RouterService":
+def build_router_service(llm: AgentLlmService, agent_repository: "AgentRepository", rag: RagRetriever) -> "RouterService":
     from app.modules.agent.engine.graphs import (
         BaseGraphFactory,
         DocsGraphFactory,
         ProjectEditsGraphFactory,
+        ProjectQaAnalysisGraphFactory,
+        ProjectQaAnswerGraphFactory,
+        ProjectQaClassificationGraphFactory,
+        ProjectQaConversationGraphFactory,
         ProjectQaGraphFactory,
+        ProjectQaRetrievalGraphFactory,
     )
     from app.modules.agent.engine.router.context_store import RouterContextStore
     from app.modules.agent.engine.router.intent_classifier import IntentClassifier
+    from app.modules.agent.engine.router.intent_switch_detector import IntentSwitchDetector
     from app.modules.agent.engine.router.registry import IntentRegistry
     from app.modules.agent.engine.router.router_service import RouterService
 
@@ -26,13 +33,20 @@ def build_router_service(llm: AgentLlmService, agent_repository: "AgentRepositor
     registry.register("project", "qa", ProjectQaGraphFactory(llm).build)
     registry.register("project", "edits", ProjectEditsGraphFactory(llm).build)
     registry.register("docs", "generation", DocsGraphFactory(llm).build)
+    registry.register("project_qa", "conversation_understanding", ProjectQaConversationGraphFactory(llm).build)
+    registry.register("project_qa", "question_classification", ProjectQaClassificationGraphFactory(llm).build)
+    registry.register("project_qa", "context_retrieval", ProjectQaRetrievalGraphFactory(rag, llm).build)
+    registry.register("project_qa", "context_analysis", ProjectQaAnalysisGraphFactory(llm).build)
+    registry.register("project_qa", "answer_composition", ProjectQaAnswerGraphFactory(llm).build)
 
     classifier = IntentClassifier(llm)
+    switch_detector = IntentSwitchDetector()
     context_store = RouterContextStore(agent_repository)
     return RouterService(
         registry=registry,
         classifier=classifier,
         context_store=context_store,
+        switch_detector=switch_detector,
     )
 
 

app/modules/agent/engine/router/context_store.py

@@ -17,6 +17,7 @@ class RouterContextStore:
         process_id: str,
         user_message: str,
         assistant_message: str,
+        decision_type: str = "start",
         max_history: int = 10,
     ) -> None:
         self._repo.update_router_context(
@@ -25,5 +26,6 @@ class RouterContextStore:
             process_id=process_id,
             user_message=user_message,
             assistant_message=assistant_message,
+            decision_type=decision_type,
             max_history=max_history,
         )

app/modules/agent/engine/router/intent_classifier.py

@@ -17,11 +17,7 @@ class IntentClassifier:
     def __init__(self, llm: AgentLlmService) -> None:
         self._llm = llm
 
-    def classify(self, user_message: str, context: RouterContext, mode: str = "auto") -> RouteDecision:
-        forced = self._from_mode(mode)
-        if forced:
-            return forced
-
+    def classify_new_intent(self, user_message: str, context: RouterContext) -> RouteDecision:
         text = (user_message or "").strip().lower()
         if text in self._short_confirmations and context.last_routing:
             return RouteDecision(
@@ -30,6 +26,7 @@ class IntentClassifier:
                 confidence=1.0,
                 reason="short_confirmation",
                 use_previous=True,
+                decision_type="continue",
             )
 
         deterministic = self._deterministic_route(text)
@@ -45,9 +42,10 @@ class IntentClassifier:
             process_id="general",
             confidence=0.8,
             reason="default",
+            decision_type="start",
         )
 
-    def _from_mode(self, mode: str) -> RouteDecision | None:
+    def from_mode(self, mode: str) -> RouteDecision | None:
         mapping = {
             "project_qa": ("project", "qa"),
             "project_edits": ("project", "edits"),
@@ -65,6 +63,8 @@ class IntentClassifier:
             process_id=route[1],
             confidence=1.0,
             reason=f"mode_override:{mode}",
+            decision_type="switch",
+            explicit_switch=True,
         )
 
     def _classify_with_llm(self, user_message: str, context: RouterContext) -> RouteDecision | None:
@@ -96,6 +96,7 @@ class IntentClassifier:
             process_id=route[1],
             confidence=confidence,
             reason=f"llm_router:{payload.get('reason', 'ok')}",
+            decision_type="start",
         )
 
     def _parse_llm_payload(self, raw: str) -> dict[str, str | float] | None:
@@ -139,6 +140,8 @@ class IntentClassifier:
                 process_id="edits",
                 confidence=0.97,
                 reason="deterministic_targeted_file_edit",
+                decision_type="switch",
+                explicit_switch=True,
             )
         if self._is_broad_docs_request(text):
             return RouteDecision(
@@ -146,6 +149,8 @@ class IntentClassifier:
                 process_id="generation",
                 confidence=0.95,
                 reason="deterministic_docs_generation",
+                decision_type="switch",
+                explicit_switch=True,
             )
         return None
 

app/modules/agent/engine/router/intent_switch_detector.py

@@ -0,0 +1,81 @@
+from __future__ import annotations
+
+import re
+
+from app.modules.agent.engine.router.schemas import RouterContext
+
+
+class IntentSwitchDetector:
+    _EXPLICIT_SWITCH_MARKERS = (
+        "теперь",
+        "а теперь",
+        "давай теперь",
+        "переключись",
+        "переключаемся",
+        "сейчас другое",
+        "новая задача",
+        "new task",
+        "switch to",
+        "now do",
+        "instead",
+    )
+    _FOLLOW_UP_MARKERS = (
+        "а еще",
+        "а ещё",
+        "подробнее",
+        "почему",
+        "зачем",
+        "что если",
+        "и еще",
+        "и ещё",
+        "покажи подробнее",
+        "можешь подробнее",
+    )
+
+    def should_switch(self, user_message: str, context: RouterContext) -> bool:
+        if not context.dialog_started or context.active_intent is None:
+            return False
+        text = " ".join((user_message or "").strip().lower().split())
+        if not text:
+            return False
+        if self._is_follow_up(text):
+            return False
+        if any(marker in text for marker in self._EXPLICIT_SWITCH_MARKERS):
+            return True
+        return self._is_strong_targeted_edit_request(text) or self._is_strong_docs_request(text)
+
+    def _is_follow_up(self, text: str) -> bool:
+        return any(marker in text for marker in self._FOLLOW_UP_MARKERS)
+
+    def _is_strong_targeted_edit_request(self, text: str) -> bool:
+        edit_markers = (
+            "добавь",
+            "добавить",
+            "измени",
+            "исправь",
+            "обнови",
+            "удали",
+            "замени",
+            "append",
+            "update",
+            "edit",
+            "remove",
+            "replace",
+        )
+        has_edit_marker = any(marker in text for marker in edit_markers)
+        has_file_marker = (
+            "readme" in text
+            or bool(re.search(r"\b[\w.\-/]+\.(md|txt|rst|yaml|yml|json|toml|ini|cfg|py)\b", text))
+        )
+        return has_edit_marker and has_file_marker
+
+    def _is_strong_docs_request(self, text: str) -> bool:
+        docs_markers = (
+            "подготовь документац",
+            "сгенерируй документац",
+            "создай документац",
+            "опиши документац",
+            "generate documentation",
+            "write documentation",
+        )
+        return any(marker in text for marker in docs_markers)

app/modules/agent/engine/router/router_service.py

@@ -1,7 +1,8 @@
 from app.modules.agent.engine.router.context_store import RouterContextStore
 from app.modules.agent.engine.router.intent_classifier import IntentClassifier
+from app.modules.agent.engine.router.intent_switch_detector import IntentSwitchDetector
 from app.modules.agent.engine.router.registry import IntentRegistry
-from app.modules.agent.engine.router.schemas import RouteResolution
+from app.modules.agent.engine.router.schemas import RouteDecision, RouteResolution
 
 
 class RouterService:
@@ -10,27 +11,48 @@ class RouterService:
         registry: IntentRegistry,
         classifier: IntentClassifier,
         context_store: RouterContextStore,
+        switch_detector: IntentSwitchDetector | None = None,
         min_confidence: float = 0.7,
     ) -> None:
         self._registry = registry
         self._classifier = classifier
         self._ctx = context_store
+        self._switch_detector = switch_detector or IntentSwitchDetector()
         self._min_confidence = min_confidence
 
     def resolve(self, user_message: str, conversation_key: str, mode: str = "auto") -> RouteResolution:
         context = self._ctx.get(conversation_key)
-        decision = self._classifier.classify(user_message, context, mode=mode)
-        if decision.confidence < self._min_confidence:
-            return self._fallback("low_confidence")
-        if not self._registry.is_valid(decision.domain_id, decision.process_id):
-            return self._fallback("invalid_route")
-        return RouteResolution(
-            domain_id=decision.domain_id,
-            process_id=decision.process_id,
-            confidence=decision.confidence,
-            reason=decision.reason,
-            fallback_used=False,
-        )
+        forced = self._classifier.from_mode(mode)
+        if forced:
+            return self._resolution(forced)
+
+        if not context.dialog_started or context.active_intent is None:
+            decision = self._classifier.classify_new_intent(user_message, context)
+            if not self._is_acceptable(decision):
+                return self._fallback("low_confidence")
+            return self._resolution(
+                decision.model_copy(
+                    update={
+                        "decision_type": "start",
+                        "explicit_switch": False,
+                    }
+                )
+            )
+
+        if self._switch_detector.should_switch(user_message, context):
+            decision = self._classifier.classify_new_intent(user_message, context)
+            if self._is_acceptable(decision):
+                return self._resolution(
+                    decision.model_copy(
+                        update={
+                            "decision_type": "switch",
+                            "explicit_switch": True,
+                        }
+                    )
+                )
+            return self._continue_current(context, "explicit_switch_unresolved_keep_current")
+
+        return self._continue_current(context, "continue_current_intent")
 
     def persist_context(
         self,
@@ -40,6 +62,7 @@ class RouterService:
         process_id: str,
         user_message: str,
         assistant_message: str,
+        decision_type: str = "start",
     ) -> None:
         self._ctx.update(
             conversation_key,
@@ -47,6 +70,7 @@ class RouterService:
             process_id=process_id,
             user_message=user_message,
             assistant_message=assistant_message,
+            decision_type=decision_type,
         )
 
     def graph_factory(self, domain_id: str, process_id: str):
@@ -59,4 +83,32 @@ class RouterService:
             confidence=0.0,
             reason=reason,
             fallback_used=True,
+            decision_type="start",
+            explicit_switch=False,
+        )
+
+    def _continue_current(self, context, reason: str) -> RouteResolution:
+        active = context.active_intent or context.last_routing or {"domain_id": "default", "process_id": "general"}
+        return RouteResolution(
+            domain_id=str(active["domain_id"]),
+            process_id=str(active["process_id"]),
+            confidence=1.0,
+            reason=reason,
+            fallback_used=False,
+            decision_type="continue",
+            explicit_switch=False,
+        )
+
+    def _is_acceptable(self, decision: RouteDecision) -> bool:
+        return decision.confidence >= self._min_confidence and self._registry.is_valid(decision.domain_id, decision.process_id)
+
+    def _resolution(self, decision: RouteDecision) -> RouteResolution:
+        return RouteResolution(
+            domain_id=decision.domain_id,
+            process_id=decision.process_id,
+            confidence=decision.confidence,
+            reason=decision.reason,
+            fallback_used=False,
+            decision_type=decision.decision_type,
+            explicit_switch=decision.explicit_switch,
         )

app/modules/agent/engine/router/schemas.py

@@ -7,6 +7,8 @@ class RouteDecision(BaseModel):
     confidence: float = 0.0
     reason: str = ""
     use_previous: bool = False
+    decision_type: str = "start"
+    explicit_switch: bool = False
 
     @field_validator("confidence")
     @classmethod
@@ -20,8 +22,13 @@ class RouteResolution(BaseModel):
     confidence: float
     reason: str
     fallback_used: bool = False
+    decision_type: str = "start"
+    explicit_switch: bool = False
 
 
 class RouterContext(BaseModel):
     last_routing: dict[str, str] | None = None
     message_history: list[dict[str, str]] = Field(default_factory=list)
+    active_intent: dict[str, str] | None = None
+    dialog_started: bool = False
+    turn_index: int = 0

app/modules/agent/llm/service.py

@@ -1,14 +1,40 @@
+import logging
+
 from app.modules.agent.prompt_loader import PromptLoader
 from app.modules.shared.gigachat.client import GigaChatClient
 
+LOGGER = logging.getLogger(__name__)
+
+
+def _truncate_for_log(text: str, max_chars: int = 1500) -> str:
+    value = (text or "").replace("\n", "\\n").strip()
+    if len(value) <= max_chars:
+        return value
+    return value[:max_chars].rstrip() + "...[truncated]"
+
 
 class AgentLlmService:
     def __init__(self, client: GigaChatClient, prompts: PromptLoader) -> None:
         self._client = client
         self._prompts = prompts
 
-    def generate(self, prompt_name: str, user_input: str) -> str:
+    def generate(self, prompt_name: str, user_input: str, *, log_context: str | None = None) -> str:
         system_prompt = self._prompts.load(prompt_name)
         if not system_prompt:
             system_prompt = "You are a helpful assistant."
-        return self._client.complete(system_prompt=system_prompt, user_prompt=user_input)
+        if log_context:
+            LOGGER.warning(
+                "graph llm input: context=%s prompt=%s user_input=%s",
+                log_context,
+                prompt_name,
+                _truncate_for_log(user_input),
+            )
+        output = self._client.complete(system_prompt=system_prompt, user_prompt=user_input)
+        if log_context:
+            LOGGER.warning(
+                "graph llm output: context=%s prompt=%s output=%s",
+                log_context,
+                prompt_name,
+                _truncate_for_log(output),
+            )
+        return output

app/modules/agent/module.py

@@ -1,5 +1,8 @@
+from __future__ import annotations
+
 from fastapi import APIRouter
 from pydantic import BaseModel, HttpUrl
+from typing import TYPE_CHECKING
 
 from app.modules.agent.changeset_validator import ChangeSetValidator
 from app.modules.agent.confluence_service import ConfluenceService
@@ -19,12 +22,17 @@ class ConfluenceFetchRequest(BaseModel):
     url: HttpUrl
 
 
+if TYPE_CHECKING:
+    from app.modules.rag.explain.retriever_v2 import CodeExplainRetrieverV2
+
+
 class AgentModule:
     def __init__(
         self,
         rag_retriever: RagRetriever,
         agent_repository: AgentRepository,
         story_context_repository: StoryContextRepository,
+        code_explain_retriever: CodeExplainRetrieverV2 | None = None,
     ) -> None:
         self.confluence = ConfluenceService()
         self.changeset_validator = ChangeSetValidator()
@@ -34,14 +42,16 @@ class AgentModule:
         client = GigaChatClient(settings, token_provider)
         prompt_loader = PromptLoader()
         llm = AgentLlmService(client=client, prompts=prompt_loader)
+        self.llm = llm
         story_recorder = StorySessionRecorder(story_context_repository)
         self.runtime = GraphAgentRuntime(
             rag=rag_retriever,
             confluence=self.confluence,
             changeset_validator=self.changeset_validator,
-            llm=llm,
+            llm=self.llm,
             agent_repository=agent_repository,
             story_recorder=story_recorder,
+            code_explain_retriever=code_explain_retriever,
         )
 
     def internal_router(self) -> APIRouter:

app/modules/agent/prompts/code_explain_answer_v2.txt

@@ -0,0 +1,17 @@
+Объяснение кода осуществляется только с использованием предоставленного ExplainPack.
+
+Правила:
+- Сначала используйте доказательства.
+- Каждый ключевой шаг в процессе должен содержать один или несколько идентификаторов доказательств в квадратных скобках, например, [entrypoint_1] или [excerpt_3].
+- Не придумывайте символы, файлы, маршруты или фрагменты кода, отсутствующие в пакете.
+- Если доказательства неполные, укажите это явно.
+- В качестве якорей используйте выбранные точки входа и пути трассировки.
+
+Верните Markdown со следующей структурой:
+1. Краткое описание
+2. Пошаговый процесс
+3. Данные и побочные эффекты
+4. Ошибки и граничные случаи
+5. Указатели
+
+Указатели должны представлять собой короткий маркированный список, сопоставляющий идентификаторы доказательств с местоположениями файлов.

app/modules/agent/prompts/rag_intent_router_v2.txt

@@ -0,0 +1,24 @@
+Ты intent-router для layered RAG.
+На вход ты получаешь JSON с полями:
+- message: текущий запрос пользователя
+- active_intent: текущий активный intent диалога или null
+- last_query: предыдущий запрос пользователя
+- allowed_intents: допустимые intent'ы
+
+Выбери ровно один intent из allowed_intents.
+Верни только JSON без markdown и пояснений.
+
+Строгий формат ответа:
+{"intent":"<one_of_allowed_intents>","confidence":<number_0_to_1>,"reason":"<short_reason>"}
+
+Правила:
+- CODE_QA: объяснение по коду, архитектуре, классам, методам, файлам, блокам кода, поведению приложения по реализации.
+- DOCS_QA: объяснение по документации, README, markdown, specs, runbooks, разделам документации.
+- GENERATE_DOCS_FROM_CODE: просьба сгенерировать, подготовить или обновить документацию по коду.
+- PROJECT_MISC: прочие вопросы по проекту, не относящиеся явно к коду или документации.
+
+Приоритет:
+- Если пользователь просит именно подготовить документацию по коду, выбирай GENERATE_DOCS_FROM_CODE.
+- Если пользователь спрашивает про конкретный класс, файл, метод или блок кода, выбирай CODE_QA.
+- Если пользователь спрашивает про README, docs, markdown или конкретную документацию, выбирай DOCS_QA.
+- Если сигнал неочевиден, выбирай PROJECT_MISC и confidence <= 0.6.

app/modules/agent/repository.py

@@ -18,6 +18,10 @@ class AgentRepository:
                         conversation_key VARCHAR(64) PRIMARY KEY,
                         last_domain_id VARCHAR(64) NULL,
                         last_process_id VARCHAR(64) NULL,
+                        active_domain_id VARCHAR(64) NULL,
+                        active_process_id VARCHAR(64) NULL,
+                        dialog_started BOOLEAN NOT NULL DEFAULT FALSE,
+                        turn_index INTEGER NOT NULL DEFAULT 0,
                         message_history_json TEXT NOT NULL DEFAULT '[]',
                         updated_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP
                     )
@@ -64,14 +68,24 @@ class AgentRepository:
                     """
                 )
             )
+            self._ensure_router_context_columns(conn)
             conn.commit()
 
+    def _ensure_router_context_columns(self, conn) -> None:
+        for statement in (
+            "ALTER TABLE router_context ADD COLUMN IF NOT EXISTS active_domain_id VARCHAR(64) NULL",
+            "ALTER TABLE router_context ADD COLUMN IF NOT EXISTS active_process_id VARCHAR(64) NULL",
+            "ALTER TABLE router_context ADD COLUMN IF NOT EXISTS dialog_started BOOLEAN NOT NULL DEFAULT FALSE",
+            "ALTER TABLE router_context ADD COLUMN IF NOT EXISTS turn_index INTEGER NOT NULL DEFAULT 0",
+        ):
+            conn.execute(text(statement))
+
     def get_router_context(self, conversation_key: str) -> RouterContext:
         with get_engine().connect() as conn:
             row = conn.execute(
                 text(
                     """
-                    SELECT last_domain_id, last_process_id, message_history_json
+                    SELECT last_domain_id, last_process_id, active_domain_id, active_process_id, dialog_started, turn_index, message_history_json
                     FROM router_context
                     WHERE conversation_key = :key
                     """
@@ -82,7 +96,7 @@ class AgentRepository:
         if not row:
             return RouterContext()
 
-        history_raw = row[2] or "[]"
+        history_raw = row[6] or "[]"
         try:
             history = json.loads(history_raw)
         except json.JSONDecodeError:
@@ -91,6 +105,9 @@ class AgentRepository:
         last = None
         if row[0] and row[1]:
             last = {"domain_id": str(row[0]), "process_id": str(row[1])}
+        active = None
+        if row[2] and row[3]:
+            active = {"domain_id": str(row[2]), "process_id": str(row[3])}
 
         clean_history = []
         for item in history if isinstance(history, list) else []:
@@ -101,7 +118,13 @@ class AgentRepository:
             if role in {"user", "assistant"} and content:
                 clean_history.append({"role": role, "content": content})
 
-        return RouterContext(last_routing=last, message_history=clean_history)
+        return RouterContext(
+            last_routing=last,
+            message_history=clean_history,
+            active_intent=active or last,
+            dialog_started=bool(row[4]),
+            turn_index=int(row[5] or 0),
+        )
 
     def update_router_context(
         self,
@@ -111,6 +134,7 @@ class AgentRepository:
         process_id: str,
         user_message: str,
         assistant_message: str,
+        decision_type: str,
         max_history: int,
     ) -> None:
         current = self.get_router_context(conversation_key)
@@ -121,17 +145,29 @@ class AgentRepository:
             history.append({"role": "assistant", "content": assistant_message})
         if max_history > 0:
             history = history[-max_history:]
+        current_active = current.active_intent or current.last_routing or {"domain_id": domain_id, "process_id": process_id}
+        next_active = (
+            {"domain_id": domain_id, "process_id": process_id}
+            if decision_type in {"start", "switch"}
+            else current_active
+        )
+        next_turn_index = max(0, int(current.turn_index or 0)) + (1 if user_message else 0)
 
         with get_engine().connect() as conn:
             conn.execute(
                 text(
                     """
                     INSERT INTO router_context (
-                        conversation_key, last_domain_id, last_process_id, message_history_json
-                    ) VALUES (:key, :domain, :process, :history)
+                        conversation_key, last_domain_id, last_process_id, active_domain_id, active_process_id,
+                        dialog_started, turn_index, message_history_json
+                    ) VALUES (:key, :domain, :process, :active_domain, :active_process, :dialog_started, :turn_index, :history)
                     ON CONFLICT (conversation_key) DO UPDATE SET
                         last_domain_id = EXCLUDED.last_domain_id,
                         last_process_id = EXCLUDED.last_process_id,
+                        active_domain_id = EXCLUDED.active_domain_id,
+                        active_process_id = EXCLUDED.active_process_id,
+                        dialog_started = EXCLUDED.dialog_started,
+                        turn_index = EXCLUDED.turn_index,
                         message_history_json = EXCLUDED.message_history_json,
                         updated_at = CURRENT_TIMESTAMP
                     """
@@ -140,6 +176,10 @@ class AgentRepository:
                     "key": conversation_key,
                     "domain": domain_id,
                     "process": process_id,
+                    "active_domain": str(next_active["domain_id"]),
+                    "active_process": str(next_active["process_id"]),
+                    "dialog_started": True,
+                    "turn_index": next_turn_index,
                     "history": json.dumps(history, ensure_ascii=False),
                 },
             )

app/modules/agent/service.py

@@ -1,12 +1,16 @@
+from __future__ import annotations
+
 from dataclasses import dataclass, field
 from collections.abc import Awaitable, Callable
 import inspect
 import logging
 import re
+from typing import TYPE_CHECKING
 
 from app.modules.agent.engine.orchestrator import OrchestratorService, TaskSpecBuilder
 from app.modules.agent.engine.orchestrator.metrics_persister import MetricsPersister
 from app.modules.agent.engine.orchestrator.models import RoutingMeta
+from app.modules.agent.engine.orchestrator.step_registry import StepRegistry
 from app.modules.agent.engine.router import build_router_service
 from app.modules.agent.llm import AgentLlmService
 from app.modules.agent.story_session_recorder import StorySessionRecorder
@@ -22,6 +26,9 @@ from app.schemas.common import ModuleName
 
 LOGGER = logging.getLogger(__name__)
 
+if TYPE_CHECKING:
+    from app.modules.rag.explain.retriever_v2 import CodeExplainRetrieverV2
+
 
 def _truncate_for_log(text: str | None, max_chars: int = 1500) -> str:
     value = (text or "").replace("\n", "\\n").strip()
@@ -47,13 +54,14 @@ class GraphAgentRuntime:
         llm: AgentLlmService,
         agent_repository: AgentRepository,
         story_recorder: StorySessionRecorder | None = None,
+        code_explain_retriever: CodeExplainRetrieverV2 | None = None,
     ) -> None:
         self._rag = rag
         self._confluence = confluence
         self._changeset_validator = changeset_validator
-        self._router = build_router_service(llm, agent_repository)
+        self._router = build_router_service(llm, agent_repository, rag)
         self._task_spec_builder = TaskSpecBuilder()
-        self._orchestrator = OrchestratorService()
+        self._orchestrator = OrchestratorService(step_registry=StepRegistry(code_explain_retriever))
         self._metrics_persister = MetricsPersister(agent_repository)
         self._story_recorder = story_recorder
         self._checkpointer = None
@@ -70,7 +78,7 @@ class GraphAgentRuntime:
         files: list[dict],
         progress_cb: Callable[[str, str, str, dict | None], Awaitable[None] | None] | None = None,
     ) -> AgentResult:
-        LOGGER.warning(
+        LOGGER.info(
             "GraphAgentRuntime.run started: task_id=%s dialog_session_id=%s mode=%s",
             task_id,
             dialog_session_id,
@@ -96,9 +104,7 @@ class GraphAgentRuntime:
             meta={"domain_id": route.domain_id, "process_id": route.process_id},
         )
         files_map = self._build_files_map(files)
-
-        await self._emit_progress(progress_cb, "agent.rag", "Собираю релевантный контекст из RAG.")
-        rag_ctx = await self._rag.retrieve(rag_session_id, message)
+        rag_ctx: list[dict] = []
         await self._emit_progress(progress_cb, "agent.attachments", "Обрабатываю дополнительные вложения.")
         conf_pages = await self._fetch_confluence_pages(attachments)
         route_meta = RoutingMeta(
@@ -157,8 +163,9 @@ class GraphAgentRuntime:
                     process_id=route.process_id,
                     user_message=message,
                     assistant_message=final_answer,
+                    decision_type=route.decision_type,
                 )
-                LOGGER.warning(
+                LOGGER.info(
                     "final agent answer: task_id=%s route=%s/%s answer=%s",
                     task_id,
                     route.domain_id,
@@ -178,7 +185,7 @@ class GraphAgentRuntime:
                     answer=final_answer,
                     meta={
                         "route": route.model_dump(),
-                        "used_rag": True,
+                        "used_rag": False,
                         "used_confluence": bool(conf_pages),
                         "changeset_filtered_out": True,
                         "orchestrator": orchestrator_meta,
@@ -193,6 +200,7 @@ class GraphAgentRuntime:
                 process_id=route.process_id,
                 user_message=message,
                 assistant_message=final_answer or f"changeset:{len(validated)}",
+                decision_type=route.decision_type,
             )
             final = AgentResult(
                 result_type=TaskResultType.CHANGESET,
@@ -200,7 +208,7 @@ class GraphAgentRuntime:
                 changeset=validated,
                 meta={
                     "route": route.model_dump(),
-                    "used_rag": True,
+                    "used_rag": False,
                     "used_confluence": bool(conf_pages),
                     "orchestrator": orchestrator_meta,
                     "orchestrator_steps": orchestrator_steps,
@@ -214,7 +222,7 @@ class GraphAgentRuntime:
                 scenario=str(orchestrator_meta.get("scenario", task_spec.scenario.value)),
                 quality=quality_meta,
             )
-            LOGGER.warning(
+            LOGGER.info(
                 "GraphAgentRuntime.run completed: task_id=%s route=%s/%s result_type=%s changeset_items=%s",
                 task_id,
                 route.domain_id,
@@ -222,7 +230,7 @@ class GraphAgentRuntime:
                 final.result_type.value,
                 len(final.changeset),
             )
-            LOGGER.warning(
+            LOGGER.info(
                 "final agent answer: task_id=%s route=%s/%s answer=%s",
                 task_id,
                 route.domain_id,
@@ -239,13 +247,14 @@ class GraphAgentRuntime:
             process_id=route.process_id,
             user_message=message,
             assistant_message=final_answer,
+            decision_type=route.decision_type,
         )
         final = AgentResult(
             result_type=TaskResultType.ANSWER,
             answer=final_answer,
             meta={
                 "route": route.model_dump(),
-                "used_rag": True,
+                "used_rag": False,
                 "used_confluence": bool(conf_pages),
                 "orchestrator": orchestrator_meta,
                 "orchestrator_steps": orchestrator_steps,
@@ -259,7 +268,7 @@ class GraphAgentRuntime:
             scenario=str(orchestrator_meta.get("scenario", task_spec.scenario.value)),
             quality=quality_meta,
         )
-        LOGGER.warning(
+        LOGGER.info(
             "GraphAgentRuntime.run completed: task_id=%s route=%s/%s result_type=%s answer_len=%s",
             task_id,
             route.domain_id,
@@ -267,7 +276,7 @@ class GraphAgentRuntime:
             final.result_type.value,
             len(final.answer or ""),
         )
-        LOGGER.warning(
+        LOGGER.info(
             "final agent answer: task_id=%s route=%s/%s answer=%s",
             task_id,
             route.domain_id,
@@ -351,7 +360,7 @@ class GraphAgentRuntime:
             factory = self._router.graph_factory("default", "general")
         if factory is None:
             raise RuntimeError("No graph factory configured")
-        LOGGER.warning("_resolve_graph resolved: domain_id=%s process_id=%s", domain_id, process_id)
+        LOGGER.debug("_resolve_graph resolved: domain_id=%s process_id=%s", domain_id, process_id)
         return factory(self._checkpointer)
 
     def _invoke_graph(self, graph, state: dict, dialog_session_id: str):
@@ -365,7 +374,7 @@ class GraphAgentRuntime:
         for item in attachments:
             if item.get("type") == "confluence_url":
                 pages.append(await self._confluence.fetch_page(item["url"]))
-        LOGGER.warning("_fetch_confluence_pages completed: pages=%s", len(pages))
+        LOGGER.info("_fetch_confluence_pages completed: pages=%s", len(pages))
         return pages
 
     def _format_rag(self, items: list[dict]) -> str:
@@ -411,7 +420,7 @@ class GraphAgentRuntime:
                 "content": str(item.get("content", "")),
                 "content_hash": str(item.get("content_hash", "")),
             }
-        LOGGER.warning("_build_files_map completed: files=%s", len(output))
+        LOGGER.debug("_build_files_map completed: files=%s", len(output))
         return output
 
     def _lookup_file(self, files_map: dict[str, dict], path: str) -> dict | None:
@@ -437,7 +446,7 @@ class GraphAgentRuntime:
                     )
                 item.base_hash = str(source["content_hash"])
             enriched.append(item)
-        LOGGER.warning("_enrich_changeset_hashes completed: items=%s", len(enriched))
+        LOGGER.debug("_enrich_changeset_hashes completed: items=%s", len(enriched))
         return enriched
 
     def _sanitize_changeset(self, items: list[ChangeItem], files_map: dict[str, dict]) -> list[ChangeItem]:
@@ -462,7 +471,7 @@ class GraphAgentRuntime:
                 continue
             sanitized.append(item)
         if dropped_noop or dropped_ws:
-            LOGGER.warning(
+            LOGGER.info(
                 "_sanitize_changeset dropped items: noop=%s whitespace_only=%s kept=%s",
                 dropped_noop,
                 dropped_ws,

app/modules/application.py

@@ -1,9 +1,14 @@
 from app.modules.agent.module import AgentModule
 from app.modules.agent.repository import AgentRepository
 from app.modules.agent.story_context_repository import StoryContextRepository, StoryContextSchemaRepository
+from app.modules.chat.direct_service import CodeExplainChatService
+from app.modules.chat.dialog_store import DialogSessionStore
 from app.modules.chat.repository import ChatRepository
 from app.modules.chat.module import ChatModule
+from app.modules.chat.session_resolver import ChatSessionResolver
+from app.modules.chat.task_store import TaskStore
 from app.modules.rag.persistence.repository import RagRepository
+from app.modules.rag.explain import CodeExplainRetrieverV2, CodeGraphRepository, LayeredRetrievalGateway
 from app.modules.rag_session.module import RagModule
 from app.modules.rag_repo.module import RagRepoModule
 from app.modules.shared.bootstrap import bootstrap_database
@@ -20,16 +25,32 @@ class ModularApplication:
         self.agent_repository = AgentRepository()
         self.story_context_schema_repository = StoryContextSchemaRepository()
         self.story_context_repository = StoryContextRepository()
+        self.chat_tasks = TaskStore()
 
         self.rag_session = RagModule(event_bus=self.events, retry=self.retry, repository=self.rag_repository)
         self.rag_repo = RagRepoModule(
             story_context_repository=self.story_context_repository,
             rag_repository=self.rag_repository,
         )
+        self.code_explain_retriever = CodeExplainRetrieverV2(
+            gateway=LayeredRetrievalGateway(self.rag_repository, self.rag_session.embedder),
+            graph_repository=CodeGraphRepository(),
+        )
         self.agent = AgentModule(
             rag_retriever=self.rag_session.rag,
             agent_repository=self.agent_repository,
             story_context_repository=self.story_context_repository,
+            code_explain_retriever=self.code_explain_retriever,
+        )
+        self.direct_chat = CodeExplainChatService(
+            retriever=self.code_explain_retriever,
+            llm=self.agent.llm,
+            session_resolver=ChatSessionResolver(
+                dialogs=DialogSessionStore(self.chat_repository),
+                rag_session_exists=lambda rag_session_id: self.rag_session.sessions.get(rag_session_id) is not None,
+            ),
+            task_store=self.chat_tasks,
+            message_sink=self.chat_repository.add_message,
         )
         self.chat = ChatModule(
             agent_runner=self.agent.runtime,
@@ -37,6 +58,8 @@ class ModularApplication:
             retry=self.retry,
             rag_sessions=self.rag_session.sessions,
             repository=self.chat_repository,
+            direct_chat=self.direct_chat,
+            task_store=self.chat_tasks,
         )
 
     def startup(self) -> None:

app/modules/chat/dialog_store.py

@@ -1,7 +1,11 @@
+from __future__ import annotations
+
 from dataclasses import dataclass
+from typing import TYPE_CHECKING
 from uuid import uuid4
 
-from app.modules.chat.repository import ChatRepository
+if TYPE_CHECKING:
+    from app.modules.chat.repository import ChatRepository
 
 
 @dataclass

app/modules/chat/direct_service.py

@@ -0,0 +1,71 @@
+from __future__ import annotations
+
+import logging
+from uuid import uuid4
+
+from app.modules.agent.llm import AgentLlmService
+from app.modules.chat.evidence_gate import CodeExplainEvidenceGate
+from app.modules.chat.session_resolver import ChatSessionResolver
+from app.modules.chat.task_store import TaskState, TaskStore
+from app.modules.rag.explain import CodeExplainRetrieverV2, PromptBudgeter
+from app.schemas.chat import ChatMessageRequest, TaskQueuedResponse, TaskResultType, TaskStatus
+
+LOGGER = logging.getLogger(__name__)
+
+
+class CodeExplainChatService:
+    def __init__(
+        self,
+        retriever: CodeExplainRetrieverV2,
+        llm: AgentLlmService,
+        session_resolver: ChatSessionResolver,
+        task_store: TaskStore,
+        message_sink,
+        budgeter: PromptBudgeter | None = None,
+        evidence_gate: CodeExplainEvidenceGate | None = None,
+    ) -> None:
+        self._retriever = retriever
+        self._llm = llm
+        self._session_resolver = session_resolver
+        self._task_store = task_store
+        self._message_sink = message_sink
+        self._budgeter = budgeter or PromptBudgeter()
+        self._evidence_gate = evidence_gate or CodeExplainEvidenceGate()
+
+    async def handle_message(self, request: ChatMessageRequest) -> TaskQueuedResponse:
+        dialog_session_id, rag_session_id = self._session_resolver.resolve(request)
+        task_id = str(uuid4())
+        task = TaskState(task_id=task_id, status=TaskStatus.RUNNING)
+        self._task_store.save(task)
+        self._message_sink(dialog_session_id, "user", request.message, task_id=task_id)
+        pack = self._retriever.build_pack(
+            rag_session_id,
+            request.message,
+            file_candidates=[item.model_dump(mode="json") for item in request.files],
+        )
+        decision = self._evidence_gate.evaluate(pack)
+        if decision.passed:
+            prompt_input = self._budgeter.build_prompt_input(request.message, pack)
+            answer = self._llm.generate(
+                "code_explain_answer_v2",
+                prompt_input,
+                log_context="chat.code_explain.direct",
+            ).strip()
+        else:
+            answer = decision.answer
+        self._message_sink(dialog_session_id, "assistant", answer, task_id=task_id)
+        task.status = TaskStatus.DONE
+        task.result_type = TaskResultType.ANSWER
+        task.answer = answer
+        self._task_store.save(task)
+        LOGGER.warning(
+            "direct code explain response: task_id=%s rag_session_id=%s excerpts=%s missing=%s",
+            task_id,
+            rag_session_id,
+            len(pack.code_excerpts),
+            pack.missing,
+        )
+        return TaskQueuedResponse(
+            task_id=task_id,
+            status=TaskStatus.DONE.value,
+        )

app/modules/chat/evidence_gate.py

@@ -0,0 +1,62 @@
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from app.modules.rag.explain.models import ExplainPack
+
+
+@dataclass(slots=True)
+class EvidenceGateDecision:
+    passed: bool
+    answer: str = ""
+    diagnostics: dict[str, list[str]] = field(default_factory=dict)
+
+
+class CodeExplainEvidenceGate:
+    def __init__(self, min_excerpts: int = 2) -> None:
+        self._min_excerpts = min_excerpts
+
+    def evaluate(self, pack: ExplainPack) -> EvidenceGateDecision:
+        diagnostics = self._diagnostics(pack)
+        if len(pack.code_excerpts) >= self._min_excerpts:
+            return EvidenceGateDecision(passed=True, diagnostics=diagnostics)
+        return EvidenceGateDecision(
+            passed=False,
+            answer=self._build_answer(pack, diagnostics),
+            diagnostics=diagnostics,
+        )
+
+    def _diagnostics(self, pack: ExplainPack) -> dict[str, list[str]]:
+        return {
+            "entrypoints": [item.title for item in pack.selected_entrypoints[:3] if item.title],
+            "symbols": [item.title for item in pack.seed_symbols[:5] if item.title],
+            "paths": self._paths(pack),
+            "missing": list(pack.missing),
+        }
+
+    def _paths(self, pack: ExplainPack) -> list[str]:
+        values: list[str] = []
+        for item in pack.selected_entrypoints + pack.seed_symbols:
+            path = item.source or (item.location.path if item.location else "")
+            if path and path not in values:
+                values.append(path)
+        for excerpt in pack.code_excerpts:
+            if excerpt.path and excerpt.path not in values:
+                values.append(excerpt.path)
+        return values[:6]
+
+    def _build_answer(self, pack: ExplainPack, diagnostics: dict[str, list[str]]) -> str:
+        lines = [
+            "Недостаточно опоры в коде, чтобы дать объяснение без догадок.",
+            "",
+            f"Найдено фрагментов кода: {len(pack.code_excerpts)} из {self._min_excerpts} минимально необходимых.",
+        ]
+        if diagnostics["paths"]:
+            lines.append(f"Пути: {', '.join(diagnostics['paths'])}")
+        if diagnostics["entrypoints"]:
+            lines.append(f"Entrypoints: {', '.join(diagnostics['entrypoints'])}")
+        if diagnostics["symbols"]:
+            lines.append(f"Символы: {', '.join(diagnostics['symbols'])}")
+        if diagnostics["missing"]:
+            lines.append(f"Диагностика: {', '.join(diagnostics['missing'])}")
+        return "\n".join(lines).strip()

app/modules/chat/module.py

@@ -1,13 +1,16 @@
+from __future__ import annotations
+
+import os
+from typing import TYPE_CHECKING
+
 from fastapi import APIRouter, Header
 from fastapi.responses import StreamingResponse
 
 from app.core.exceptions import AppError
+from app.modules.chat.direct_service import CodeExplainChatService
 from app.modules.chat.dialog_store import DialogSessionStore
-from app.modules.chat.repository import ChatRepository
 from app.modules.chat.service import ChatOrchestrator
 from app.modules.chat.task_store import TaskStore
-from app.modules.contracts import AgentRunner
-from app.modules.rag_session.session_store import RagSessionStore
 from app.modules.shared.event_bus import EventBus
 from app.modules.shared.idempotency_store import IdempotencyStore
 from app.modules.shared.retry_executor import RetryExecutor
@@ -20,6 +23,11 @@ from app.schemas.chat import (
 )
 from app.schemas.common import ModuleName
 
+if TYPE_CHECKING:
+    from app.modules.chat.repository import ChatRepository
+    from app.modules.contracts import AgentRunner
+    from app.modules.rag_session.session_store import RagSessionStore
+
 
 class ChatModule:
     def __init__(
@@ -29,12 +37,16 @@ class ChatModule:
         retry: RetryExecutor,
         rag_sessions: RagSessionStore,
         repository: ChatRepository,
+        direct_chat: CodeExplainChatService | None = None,
+        task_store: TaskStore | None = None,
     ) -> None:
         self._rag_sessions = rag_sessions
-        self.tasks = TaskStore()
+        self._simple_code_explain_only = os.getenv("SIMPLE_CODE_EXPLAIN_ONLY", "true").lower() in {"1", "true", "yes"}
+        self.tasks = task_store or TaskStore()
         self.dialogs = DialogSessionStore(repository)
         self.idempotency = IdempotencyStore()
         self.events = event_bus
+        self.direct_chat = direct_chat
         self.chat = ChatOrchestrator(
             task_store=self.tasks,
             dialogs=self.dialogs,
@@ -59,11 +71,13 @@ class ChatModule:
                 rag_session_id=dialog.rag_session_id,
             )
 
-        @router.post("/api/chat/messages", response_model=TaskQueuedResponse)
+        @router.post("/api/chat/messages", response_model=TaskQueuedResponse | TaskResultResponse)
         async def send_message(
             request: ChatMessageRequest,
             idempotency_key: str | None = Header(default=None, alias="Idempotency-Key"),
-        ) -> TaskQueuedResponse:
+        ) -> TaskQueuedResponse | TaskResultResponse:
+            if self._simple_code_explain_only and self.direct_chat is not None:
+                return await self.direct_chat.handle_message(request)
             task = await self.chat.enqueue_message(request, idempotency_key)
             return TaskQueuedResponse(task_id=task.task_id, status=task.status.value)
 

app/modules/chat/service.py

@@ -6,6 +6,7 @@ from app.modules.contracts import AgentRunner
 from app.schemas.chat import ChatMessageRequest, TaskResultType, TaskStatus
 from app.schemas.common import ErrorPayload, ModuleName
 from app.modules.chat.dialog_store import DialogSessionStore
+from app.modules.chat.session_resolver import ChatSessionResolver
 from app.modules.chat.task_store import TaskState, TaskStore
 from app.modules.shared.event_bus import EventBus
 from app.modules.shared.idempotency_store import IdempotencyStore
@@ -41,6 +42,7 @@ class ChatOrchestrator:
         self._retry = retry
         self._rag_session_exists = rag_session_exists
         self._message_sink = message_sink
+        self._session_resolver = ChatSessionResolver(dialogs, rag_session_exists)
 
     async def enqueue_message(
         self,
@@ -52,7 +54,7 @@ class ChatOrchestrator:
             if existing:
                 task = self._task_store.get(existing)
                 if task:
-                    LOGGER.warning(
+                    LOGGER.info(
                         "enqueue_message reused task by idempotency key: task_id=%s mode=%s",
                         task.task_id,
                         request.mode.value,
@@ -63,7 +65,7 @@ class ChatOrchestrator:
         if idempotency_key:
             self._idempotency.put(idempotency_key, task.task_id)
         asyncio.create_task(self._process_task(task.task_id, request))
-        LOGGER.warning(
+        LOGGER.info(
             "enqueue_message created task: task_id=%s mode=%s",
             task.task_id,
             request.mode.value,
@@ -135,6 +137,13 @@ class ChatOrchestrator:
             task.changeset = result.changeset
             if task.result_type == TaskResultType.ANSWER and task.answer:
                 self._message_sink(dialog_session_id, "assistant", task.answer, task_id=task_id)
+                LOGGER.warning(
+                    "outgoing chat response: task_id=%s dialog_session_id=%s result_type=%s answer=%s",
+                    task_id,
+                    dialog_session_id,
+                    task.result_type.value,
+                    _truncate_for_log(task.answer),
+                )
             elif task.result_type == TaskResultType.CHANGESET:
                 self._message_sink(
                     dialog_session_id,
@@ -146,6 +155,14 @@ class ChatOrchestrator:
                         "changeset": [item.model_dump(mode="json") for item in task.changeset],
                     },
                 )
+                LOGGER.warning(
+                    "outgoing chat response: task_id=%s dialog_session_id=%s result_type=%s changeset_items=%s answer=%s",
+                    task_id,
+                    dialog_session_id,
+                    task.result_type.value,
+                    len(task.changeset),
+                    _truncate_for_log(task.answer or ""),
+                )
             self._task_store.save(task)
             await self._events.publish(
                 task_id,
@@ -160,7 +177,7 @@ class ChatOrchestrator:
                 },
             )
             await self._publish_progress(task_id, "task.done", "Обработка завершена.", progress=100)
-            LOGGER.warning(
+            LOGGER.info(
                 "_process_task completed: task_id=%s status=%s result_type=%s changeset_items=%s",
                 task_id,
                 task.status.value,
@@ -232,7 +249,7 @@ class ChatOrchestrator:
         if progress is not None:
             payload["progress"] = max(0, min(100, int(progress)))
         await self._events.publish(task_id, kind, payload)
-        LOGGER.warning(
+        LOGGER.debug(
             "_publish_progress emitted: task_id=%s kind=%s stage=%s progress=%s",
             task_id,
             kind,
@@ -259,35 +276,7 @@ class ChatOrchestrator:
                     meta={"heartbeat": True},
                 )
                 index += 1
-        LOGGER.warning("_run_heartbeat stopped: task_id=%s ticks=%s", task_id, index)
+        LOGGER.debug("_run_heartbeat stopped: task_id=%s ticks=%s", task_id, index)
 
     def _resolve_sessions(self, request: ChatMessageRequest) -> tuple[str, str]:
-        # Legacy compatibility: old session_id/project_id flow.
-        if request.dialog_session_id and request.rag_session_id:
-            dialog = self._dialogs.get(request.dialog_session_id)
-            if not dialog:
-                raise AppError("dialog_not_found", "Dialog session not found", ModuleName.BACKEND)
-            if dialog.rag_session_id != request.rag_session_id:
-                raise AppError("dialog_rag_mismatch", "Dialog session does not belong to rag session", ModuleName.BACKEND)
-            LOGGER.warning(
-                "_resolve_sessions resolved by dialog_session_id: dialog_session_id=%s rag_session_id=%s",
-                request.dialog_session_id,
-                request.rag_session_id,
-            )
-            return request.dialog_session_id, request.rag_session_id
-
-        if request.session_id and request.project_id:
-            if not self._rag_session_exists(request.project_id):
-                raise AppError("rag_session_not_found", "RAG session not found", ModuleName.RAG)
-            LOGGER.warning(
-                "_resolve_sessions resolved by legacy session/project: session_id=%s project_id=%s",
-                request.session_id,
-                request.project_id,
-            )
-            return request.session_id, request.project_id
-
-        raise AppError(
-            "missing_sessions",
-            "dialog_session_id and rag_session_id are required",
-            ModuleName.BACKEND,
-        )
+        return self._session_resolver.resolve(request)

app/modules/chat/session_resolver.py

@@ -0,0 +1,36 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from app.core.exceptions import AppError
+from app.schemas.chat import ChatMessageRequest
+from app.schemas.common import ModuleName
+
+if TYPE_CHECKING:
+    from app.modules.chat.dialog_store import DialogSessionStore
+
+
+class ChatSessionResolver:
+    def __init__(self, dialogs: DialogSessionStore, rag_session_exists) -> None:
+        self._dialogs = dialogs
+        self._rag_session_exists = rag_session_exists
+
+    def resolve(self, request: ChatMessageRequest) -> tuple[str, str]:
+        if request.dialog_session_id and request.rag_session_id:
+            dialog = self._dialogs.get(request.dialog_session_id)
+            if not dialog:
+                raise AppError("dialog_not_found", "Dialog session not found", ModuleName.BACKEND)
+            if dialog.rag_session_id != request.rag_session_id:
+                raise AppError("dialog_rag_mismatch", "Dialog session does not belong to rag session", ModuleName.BACKEND)
+            return request.dialog_session_id, request.rag_session_id
+
+        if request.session_id and request.project_id:
+            if not self._rag_session_exists(request.project_id):
+                raise AppError("rag_session_not_found", "RAG session not found", ModuleName.RAG)
+            return request.session_id, request.project_id
+
+        raise AppError(
+            "missing_sessions",
+            "dialog_session_id and rag_session_id are required",
+            ModuleName.BACKEND,
+        )

app/modules/rag/README.md

@@ -90,6 +90,41 @@ sequenceDiagram
     Rag-->>Agent: items
 ```
 
+### Retrieval + project/qa reasoning
+Назначение: `RAG` вызывается не в начале runtime, а внутри отдельного graph-шага `context_retrieval` для `project/qa`.
+```mermaid
+sequenceDiagram
+    participant Agent as GraphAgentRuntime
+    participant Orch as OrchestratorService
+    participant G1 as conversation_understanding
+    participant G2 as question_classification
+    participant G3 as context_retrieval
+    participant Rag as RagService
+    participant G4 as context_analysis
+    participant G5 as answer_composition
+
+    Agent->>Orch: run(task)
+    Orch->>G1: execute
+    G1-->>Orch: resolved_request
+    Orch->>G2: execute
+    G2-->>Orch: question_profile
+    Orch->>G3: execute
+    G3->>Rag: retrieve(query)
+    Rag-->>G3: rag_items
+    G3-->>Orch: source_bundle
+    Orch->>G4: execute
+    G4-->>Orch: analysis_brief
+    Orch->>G5: execute
+    G5-->>Orch: final_answer
+    Orch-->>Agent: final_answer
+```
+
+Для `project/qa` это означает:
+- ранний глобальный retrieval больше не нужен;
+- `RAG` возвращает записи только для конкретного шага `context_retrieval`;
+- оркестратор управляет цепочкой graph-шагов;
+- пользовательский ответ собирается после анализа, а не напрямую из сырого retrieval.
+
 ## 5. Слои, фиксируемые в RAG
 
 ### 5.1. Слои DOCS

app/modules/rag/explain/__init__.py

@@ -0,0 +1,36 @@
+from __future__ import annotations
+
+from importlib import import_module
+
+__all__ = [
+    "CodeExcerpt",
+    "CodeExplainRetrieverV2",
+    "CodeGraphRepository",
+    "EvidenceItem",
+    "ExplainIntent",
+    "ExplainIntentBuilder",
+    "ExplainPack",
+    "LayeredRetrievalGateway",
+    "PromptBudgeter",
+    "TracePath",
+]
+
+
+def __getattr__(name: str):
+    module_map = {
+        "CodeExcerpt": "app.modules.rag.explain.models",
+        "EvidenceItem": "app.modules.rag.explain.models",
+        "ExplainIntent": "app.modules.rag.explain.models",
+        "ExplainPack": "app.modules.rag.explain.models",
+        "TracePath": "app.modules.rag.explain.models",
+        "ExplainIntentBuilder": "app.modules.rag.explain.intent_builder",
+        "PromptBudgeter": "app.modules.rag.explain.budgeter",
+        "LayeredRetrievalGateway": "app.modules.rag.explain.layered_gateway",
+        "CodeGraphRepository": "app.modules.rag.explain.graph_repository",
+        "CodeExplainRetrieverV2": "app.modules.rag.explain.retriever_v2",
+    }
+    module_name = module_map.get(name)
+    if module_name is None:
+        raise AttributeError(name)
+    module = import_module(module_name)
+    return getattr(module, name)

app/modules/rag/explain/budgeter.py

@@ -0,0 +1,62 @@
+from __future__ import annotations
+
+import json
+
+from app.modules.rag.explain.models import ExplainPack
+
+
+class PromptBudgeter:
+    def __init__(
+        self,
+        *,
+        max_paths: int = 3,
+        max_symbols: int = 25,
+        max_excerpts: int = 40,
+        max_chars: int = 30000,
+    ) -> None:
+        self._max_paths = max_paths
+        self._max_symbols = max_symbols
+        self._max_excerpts = max_excerpts
+        self._max_chars = max_chars
+
+    def build_prompt_input(self, question: str, pack: ExplainPack) -> str:
+        symbol_ids: list[str] = []
+        for path in pack.trace_paths[: self._max_paths]:
+            for symbol_id in path.symbol_ids:
+                if symbol_id and symbol_id not in symbol_ids and len(symbol_ids) < self._max_symbols:
+                    symbol_ids.append(symbol_id)
+        excerpts = []
+        total_chars = 0
+        for excerpt in pack.code_excerpts:
+            if symbol_ids and excerpt.symbol_id and excerpt.symbol_id not in symbol_ids:
+                continue
+            body = excerpt.content.strip()
+            remaining = self._max_chars - total_chars
+            if remaining <= 0 or len(excerpts) >= self._max_excerpts:
+                break
+            if len(body) > remaining:
+                body = body[:remaining].rstrip() + "...[truncated]"
+            excerpts.append(
+                {
+                    "evidence_id": excerpt.evidence_id,
+                    "title": excerpt.title,
+                    "path": excerpt.path,
+                    "start_line": excerpt.start_line,
+                    "end_line": excerpt.end_line,
+                    "focus": excerpt.focus,
+                    "content": body,
+                }
+            )
+            total_chars += len(body)
+        payload = {
+            "question": question,
+            "intent": pack.intent.model_dump(mode="json"),
+            "selected_entrypoints": [item.model_dump(mode="json") for item in pack.selected_entrypoints[:5]],
+            "seed_symbols": [item.model_dump(mode="json") for item in pack.seed_symbols[: self._max_symbols]],
+            "trace_paths": [path.model_dump(mode="json") for path in pack.trace_paths[: self._max_paths]],
+            "evidence_index": {key: value.model_dump(mode="json") for key, value in pack.evidence_index.items()},
+            "code_excerpts": excerpts,
+            "missing": pack.missing,
+            "conflicts": pack.conflicts,
+        }
+        return json.dumps(payload, ensure_ascii=False, indent=2)

app/modules/rag/explain/excerpt_planner.py

@@ -0,0 +1,59 @@
+from __future__ import annotations
+
+from app.modules.rag.explain.models import CodeExcerpt, LayeredRetrievalItem
+
+
+class ExcerptPlanner:
+    _FOCUS_TOKENS = ("raise", "except", "db", "select", "insert", "update", "delete", "http", "publish", "emit")
+
+    def plan(self, chunk: LayeredRetrievalItem, *, evidence_id: str, symbol_id: str | None) -> list[CodeExcerpt]:
+        location = chunk.location
+        if location is None:
+            return []
+        excerpts = [
+            CodeExcerpt(
+                evidence_id=evidence_id,
+                symbol_id=symbol_id,
+                title=chunk.title,
+                path=location.path,
+                start_line=location.start_line,
+                end_line=location.end_line,
+                content=chunk.content.strip(),
+                focus="overview",
+            )
+        ]
+        focus = self._focus_excerpt(chunk, evidence_id=evidence_id, symbol_id=symbol_id)
+        if focus is not None:
+            excerpts.append(focus)
+        return excerpts
+
+    def _focus_excerpt(
+        self,
+        chunk: LayeredRetrievalItem,
+        *,
+        evidence_id: str,
+        symbol_id: str | None,
+    ) -> CodeExcerpt | None:
+        location = chunk.location
+        if location is None:
+            return None
+        lines = chunk.content.splitlines()
+        for index, line in enumerate(lines):
+            lowered = line.lower()
+            if not any(token in lowered for token in self._FOCUS_TOKENS):
+                continue
+            start = max(0, index - 2)
+            end = min(len(lines), index + 3)
+            if end - start >= len(lines):
+                return None
+            return CodeExcerpt(
+                evidence_id=evidence_id,
+                symbol_id=symbol_id,
+                title=f"{chunk.title}:focus",
+                path=location.path,
+                start_line=(location.start_line or 1) + start,
+                end_line=(location.start_line or 1) + end - 1,
+                content="\n".join(lines[start:end]).strip(),
+                focus="focus",
+            )
+        return None

app/modules/rag/explain/graph_repository.py

@@ -0,0 +1,216 @@
+from __future__ import annotations
+
+import json
+
+from sqlalchemy import text
+
+from app.modules.rag.explain.models import CodeLocation, LayeredRetrievalItem
+from app.modules.shared.db import get_engine
+
+
+class CodeGraphRepository:
+    def get_out_edges(
+        self,
+        rag_session_id: str,
+        src_symbol_ids: list[str],
+        edge_types: list[str],
+        limit_per_src: int,
+    ) -> list[LayeredRetrievalItem]:
+        if not src_symbol_ids:
+            return []
+        sql = """
+            SELECT path, content, layer, title, metadata_json, span_start, span_end
+            FROM rag_chunks
+            WHERE rag_session_id = :sid
+              AND layer = 'C2_DEPENDENCY_GRAPH'
+              AND CAST(metadata_json AS jsonb)->>'src_symbol_id' = ANY(:src_ids)
+              AND CAST(metadata_json AS jsonb)->>'edge_type' = ANY(:edge_types)
+            ORDER BY path, span_start
+        """
+        with get_engine().connect() as conn:
+            rows = conn.execute(
+                text(sql),
+                {"sid": rag_session_id, "src_ids": src_symbol_ids, "edge_types": edge_types},
+            ).mappings().fetchall()
+        grouped: dict[str, int] = {}
+        items: list[LayeredRetrievalItem] = []
+        for row in rows:
+            metadata = self._loads(row.get("metadata_json"))
+            src_symbol_id = str(metadata.get("src_symbol_id") or "")
+            grouped[src_symbol_id] = grouped.get(src_symbol_id, 0) + 1
+            if grouped[src_symbol_id] > limit_per_src:
+                continue
+            items.append(self._to_item(row, metadata))
+        return items
+
+    def get_in_edges(
+        self,
+        rag_session_id: str,
+        dst_symbol_ids: list[str],
+        edge_types: list[str],
+        limit_per_dst: int,
+    ) -> list[LayeredRetrievalItem]:
+        if not dst_symbol_ids:
+            return []
+        sql = """
+            SELECT path, content, layer, title, metadata_json, span_start, span_end
+            FROM rag_chunks
+            WHERE rag_session_id = :sid
+              AND layer = 'C2_DEPENDENCY_GRAPH'
+              AND CAST(metadata_json AS jsonb)->>'dst_symbol_id' = ANY(:dst_ids)
+              AND CAST(metadata_json AS jsonb)->>'edge_type' = ANY(:edge_types)
+            ORDER BY path, span_start
+        """
+        with get_engine().connect() as conn:
+            rows = conn.execute(
+                text(sql),
+                {"sid": rag_session_id, "dst_ids": dst_symbol_ids, "edge_types": edge_types},
+            ).mappings().fetchall()
+        grouped: dict[str, int] = {}
+        items: list[LayeredRetrievalItem] = []
+        for row in rows:
+            metadata = self._loads(row.get("metadata_json"))
+            dst_symbol_id = str(metadata.get("dst_symbol_id") or "")
+            grouped[dst_symbol_id] = grouped.get(dst_symbol_id, 0) + 1
+            if grouped[dst_symbol_id] > limit_per_dst:
+                continue
+            items.append(self._to_item(row, metadata))
+        return items
+
+    def resolve_symbol_by_ref(
+        self,
+        rag_session_id: str,
+        dst_ref: str,
+        package_hint: str | None = None,
+    ) -> LayeredRetrievalItem | None:
+        ref = (dst_ref or "").strip()
+        if not ref:
+            return None
+        with get_engine().connect() as conn:
+            rows = conn.execute(
+                text(
+                    """
+                    SELECT path, content, layer, title, metadata_json, span_start, span_end, qname
+                    FROM rag_chunks
+                    WHERE rag_session_id = :sid
+                      AND layer = 'C1_SYMBOL_CATALOG'
+                      AND (qname = :ref OR title = :ref OR qname LIKE :tail)
+                    ORDER BY path
+                    LIMIT 12
+                    """
+                ),
+                {"sid": rag_session_id, "ref": ref, "tail": f"%{ref}"},
+            ).mappings().fetchall()
+        best: LayeredRetrievalItem | None = None
+        best_score = -1
+        for row in rows:
+            metadata = self._loads(row.get("metadata_json"))
+            package = str(metadata.get("package_or_module") or "")
+            score = 0
+            if str(row.get("qname") or "") == ref:
+                score += 3
+            if str(row.get("title") or "") == ref:
+                score += 2
+            if package_hint and package.startswith(package_hint):
+                score += 3
+            if package_hint and package_hint in str(row.get("path") or ""):
+                score += 1
+            if score > best_score:
+                best = self._to_item(row, metadata)
+                best_score = score
+        return best
+
+    def get_symbols_by_ids(self, rag_session_id: str, symbol_ids: list[str]) -> list[LayeredRetrievalItem]:
+        if not symbol_ids:
+            return []
+        with get_engine().connect() as conn:
+            rows = conn.execute(
+                text(
+                    """
+                    SELECT path, content, layer, title, metadata_json, span_start, span_end
+                    FROM rag_chunks
+                    WHERE rag_session_id = :sid
+                      AND layer = 'C1_SYMBOL_CATALOG'
+                      AND symbol_id = ANY(:symbol_ids)
+                    ORDER BY path, span_start
+                    """
+                ),
+                {"sid": rag_session_id, "symbol_ids": symbol_ids},
+            ).mappings().fetchall()
+        return [self._to_item(row, self._loads(row.get("metadata_json"))) for row in rows]
+
+    def get_chunks_by_symbol_ids(
+        self,
+        rag_session_id: str,
+        symbol_ids: list[str],
+        prefer_chunk_type: str = "symbol_block",
+    ) -> list[LayeredRetrievalItem]:
+        symbols = self.get_symbols_by_ids(rag_session_id, symbol_ids)
+        chunks: list[LayeredRetrievalItem] = []
+        for symbol in symbols:
+            location = symbol.location
+            if location is None:
+                continue
+            chunk = self._chunk_for_symbol(rag_session_id, symbol, prefer_chunk_type=prefer_chunk_type)
+            if chunk is not None:
+                chunks.append(chunk)
+        return chunks
+
+    def _chunk_for_symbol(
+        self,
+        rag_session_id: str,
+        symbol: LayeredRetrievalItem,
+        *,
+        prefer_chunk_type: str,
+    ) -> LayeredRetrievalItem | None:
+        location = symbol.location
+        if location is None:
+            return None
+        with get_engine().connect() as conn:
+            rows = conn.execute(
+                text(
+                    """
+                    SELECT path, content, layer, title, metadata_json, span_start, span_end
+                    FROM rag_chunks
+                    WHERE rag_session_id = :sid
+                      AND layer = 'C0_SOURCE_CHUNKS'
+                      AND path = :path
+                      AND COALESCE(span_start, 0) <= :end_line
+                      AND COALESCE(span_end, 999999) >= :start_line
+                    ORDER BY
+                      CASE WHEN CAST(metadata_json AS jsonb)->>'chunk_type' = :prefer_chunk_type THEN 0 ELSE 1 END,
+                      ABS(COALESCE(span_start, 0) - :start_line)
+                    LIMIT 1
+                    """
+                ),
+                {
+                    "sid": rag_session_id,
+                    "path": location.path,
+                    "start_line": location.start_line or 0,
+                    "end_line": location.end_line or 999999,
+                    "prefer_chunk_type": prefer_chunk_type,
+                },
+            ).mappings().fetchall()
+        if not rows:
+            return None
+        row = rows[0]
+        return self._to_item(row, self._loads(row.get("metadata_json")))
+
+    def _to_item(self, row, metadata: dict) -> LayeredRetrievalItem:
+        return LayeredRetrievalItem(
+            source=str(row.get("path") or ""),
+            content=str(row.get("content") or ""),
+            layer=str(row.get("layer") or ""),
+            title=str(row.get("title") or ""),
+            metadata=metadata,
+            location=CodeLocation(
+                path=str(row.get("path") or ""),
+                start_line=row.get("span_start"),
+                end_line=row.get("span_end"),
+            ),
+        )
+
+    def _loads(self, value) -> dict:
+        if not value:
+            return {}
+        return json.loads(str(value))

app/modules/rag/explain/intent_builder.py

@@ -0,0 +1,102 @@
+from __future__ import annotations
+
+import re
+
+from app.modules.rag.explain.models import ExplainHints, ExplainIntent
+from app.modules.rag.retrieval.query_terms import extract_query_terms
+
+
+class ExplainIntentBuilder:
+    _ROUTE_RE = re.compile(r"(/[A-Za-z0-9_./{}:-]+)")
+    _FILE_RE = re.compile(r"([A-Za-z0-9_./-]+\.py)")
+    _SYMBOL_RE = re.compile(r"\b([A-Z][A-Za-z0-9_]*\.[A-Za-z_][A-Za-z0-9_]*|[A-Z][A-Za-z0-9_]{2,}|[a-z_][A-Za-z0-9_]{2,})\b")
+    _COMMAND_RE = re.compile(r"`([A-Za-z0-9:_-]+)`")
+    _TEST_KEYWORDS = (
+        "тест",
+        "tests",
+        "test ",
+        "unit-test",
+        "unit test",
+        "юнит-тест",
+        "pytest",
+        "spec",
+        "как покрыто тестами",
+        "как проверяется",
+        "how is it tested",
+        "how it's tested",
+    )
+
+    def build(self, user_query: str) -> ExplainIntent:
+        normalized = " ".join((user_query or "").split())
+        lowered = normalized.lower()
+        keywords = self._keywords(normalized)
+        hints = ExplainHints(
+            paths=self._dedupe(self._FILE_RE.findall(normalized)),
+            symbols=self._symbols(normalized),
+            endpoints=self._dedupe(self._ROUTE_RE.findall(normalized)),
+            commands=self._commands(normalized, lowered),
+        )
+        return ExplainIntent(
+            raw_query=user_query,
+            normalized_query=normalized,
+            keywords=keywords[:12],
+            hints=hints,
+            include_tests=self._include_tests(lowered),
+            expected_entry_types=self._entry_types(lowered, hints),
+            depth=self._depth(lowered),
+        )
+
+    def _keywords(self, text: str) -> list[str]:
+        keywords = extract_query_terms(text)
+        for token in self._symbols(text):
+            if token not in keywords:
+                keywords.append(token)
+        for token in self._ROUTE_RE.findall(text):
+            if token not in keywords:
+                keywords.append(token)
+        return self._dedupe(keywords)
+
+    def _symbols(self, text: str) -> list[str]:
+        values = []
+        for raw in self._SYMBOL_RE.findall(text):
+            token = raw.strip()
+            if len(token) < 3:
+                continue
+            if token.endswith(".py"):
+                continue
+            values.append(token)
+        return self._dedupe(values)
+
+    def _commands(self, text: str, lowered: str) -> list[str]:
+        values = list(self._COMMAND_RE.findall(text))
+        if " command " in f" {lowered} ":
+            values.extend(re.findall(r"command\s+([A-Za-z0-9:_-]+)", lowered))
+        if " cli " in f" {lowered} ":
+            values.extend(re.findall(r"cli\s+([A-Za-z0-9:_-]+)", lowered))
+        return self._dedupe(values)
+
+    def _entry_types(self, lowered: str, hints: ExplainHints) -> list[str]:
+        if hints.endpoints or any(token in lowered for token in ("endpoint", "route", "handler", "http", "api")):
+            return ["http"]
+        if hints.commands or any(token in lowered for token in ("cli", "command", "click", "typer")):
+            return ["cli"]
+        return ["http", "cli"]
+
+    def _depth(self, lowered: str) -> str:
+        if any(token in lowered for token in ("deep", "подроб", "деталь", "full flow", "trace")):
+            return "deep"
+        if any(token in lowered for token in ("high level", "overview", "кратко", "summary")):
+            return "high"
+        return "medium"
+
+    def _include_tests(self, lowered: str) -> bool:
+        normalized = f" {lowered} "
+        return any(token in normalized for token in self._TEST_KEYWORDS)
+
+    def _dedupe(self, values: list[str]) -> list[str]:
+        result: list[str] = []
+        for value in values:
+            item = value.strip()
+            if item and item not in result:
+                result.append(item)
+        return result

app/modules/rag/explain/layered_gateway.py

@@ -0,0 +1,289 @@
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Callable
+
+from app.modules.rag.explain.models import CodeLocation, LayeredRetrievalItem
+from app.modules.rag.retrieval.test_filter import build_test_filters, debug_disable_test_filter
+
+LOGGER = logging.getLogger(__name__)
+
+if TYPE_CHECKING:
+    from app.modules.rag.persistence.repository import RagRepository
+    from app.modules.rag_session.embedding.gigachat_embedder import GigaChatEmbedder
+
+
+@dataclass(slots=True)
+class LayerRetrievalResult:
+    items: list[LayeredRetrievalItem]
+    missing: list[str] = field(default_factory=list)
+
+
+class LayeredRetrievalGateway:
+    def __init__(self, repository: RagRepository, embedder: GigaChatEmbedder) -> None:
+        self._repository = repository
+        self._embedder = embedder
+
+    def retrieve_layer(
+        self,
+        rag_session_id: str,
+        query: str,
+        layer: str,
+        *,
+        limit: int,
+        path_prefixes: list[str] | None = None,
+        exclude_tests: bool = True,
+        prefer_non_tests: bool = False,
+        include_spans: bool = False,
+    ) -> LayerRetrievalResult:
+        effective_exclude_tests = exclude_tests and not debug_disable_test_filter()
+        filter_args = self._filter_args(effective_exclude_tests)
+        query_embedding: list[float] | None = None
+        try:
+            query_embedding = self._embedder.embed([query])[0]
+            rows = self._repository.retrieve(
+                rag_session_id,
+                query_embedding,
+                query_text=query,
+                limit=limit,
+                layers=[layer],
+                path_prefixes=path_prefixes,
+                exclude_path_prefixes=filter_args["exclude_path_prefixes"],
+                exclude_like_patterns=filter_args["exclude_like_patterns"],
+                prefer_non_tests=prefer_non_tests or not effective_exclude_tests,
+            )
+            return self._success_result(
+                rows,
+                rag_session_id=rag_session_id,
+                label="layered retrieval",
+                include_spans=include_spans,
+                layer=layer,
+                exclude_tests=effective_exclude_tests,
+                path_prefixes=path_prefixes,
+            )
+        except Exception as exc:
+            if query_embedding is None:
+                self._log_failure(
+                    label="layered retrieval",
+                    rag_session_id=rag_session_id,
+                    layer=layer,
+                    exclude_tests=effective_exclude_tests,
+                    path_prefixes=path_prefixes,
+                    exc=exc,
+                )
+                return LayerRetrievalResult(items=[], missing=[self._failure_missing(f"layer:{layer} retrieval_failed", exc)])
+            retry_result = self._retry_without_test_filter(
+                operation=lambda: self._repository.retrieve(
+                    rag_session_id,
+                    query_embedding,
+                    query_text=query,
+                    limit=limit,
+                    layers=[layer],
+                    path_prefixes=path_prefixes,
+                    exclude_path_prefixes=None,
+                    exclude_like_patterns=None,
+                    prefer_non_tests=True,
+                ),
+                label="layered retrieval",
+                rag_session_id=rag_session_id,
+                include_spans=include_spans,
+                layer=layer,
+                exclude_tests=effective_exclude_tests,
+                path_prefixes=path_prefixes,
+                exc=exc,
+                missing_prefix=f"layer:{layer} retrieval_failed",
+            )
+            if retry_result is not None:
+                return retry_result
+            return LayerRetrievalResult(items=[], missing=[self._failure_missing(f"layer:{layer} retrieval_failed", exc)])
+
+    def retrieve_lexical_code(
+        self,
+        rag_session_id: str,
+        query: str,
+        *,
+        limit: int,
+        path_prefixes: list[str] | None = None,
+        exclude_tests: bool = True,
+        include_spans: bool = False,
+    ) -> LayerRetrievalResult:
+        effective_exclude_tests = exclude_tests and not debug_disable_test_filter()
+        filter_args = self._filter_args(effective_exclude_tests)
+        try:
+            rows = self._repository.retrieve_lexical_code(
+                rag_session_id,
+                query_text=query,
+                limit=limit,
+                path_prefixes=path_prefixes,
+                exclude_path_prefixes=filter_args["exclude_path_prefixes"],
+                exclude_like_patterns=filter_args["exclude_like_patterns"],
+                prefer_non_tests=not effective_exclude_tests,
+            )
+            return self._success_result(
+                rows,
+                rag_session_id=rag_session_id,
+                label="lexical retrieval",
+                include_spans=include_spans,
+                exclude_tests=effective_exclude_tests,
+                path_prefixes=path_prefixes,
+            )
+        except Exception as exc:
+            retry_result = self._retry_without_test_filter(
+                operation=lambda: self._repository.retrieve_lexical_code(
+                    rag_session_id,
+                    query_text=query,
+                    limit=limit,
+                    path_prefixes=path_prefixes,
+                    exclude_path_prefixes=None,
+                    exclude_like_patterns=None,
+                    prefer_non_tests=True,
+                ),
+                label="lexical retrieval",
+                rag_session_id=rag_session_id,
+                include_spans=include_spans,
+                exclude_tests=effective_exclude_tests,
+                path_prefixes=path_prefixes,
+                exc=exc,
+                missing_prefix="layer:C0 lexical_retrieval_failed",
+            )
+            if retry_result is not None:
+                return retry_result
+            return LayerRetrievalResult(items=[], missing=[self._failure_missing("layer:C0 lexical_retrieval_failed", exc)])
+
+    def _retry_without_test_filter(
+        self,
+        *,
+        operation: Callable[[], list[dict]],
+        label: str,
+        rag_session_id: str,
+        include_spans: bool,
+        exclude_tests: bool,
+        path_prefixes: list[str] | None,
+        exc: Exception,
+        missing_prefix: str,
+        layer: str | None = None,
+    ) -> LayerRetrievalResult | None:
+        if not exclude_tests:
+            self._log_failure(
+                label=label,
+                rag_session_id=rag_session_id,
+                layer=layer,
+                exclude_tests=exclude_tests,
+                path_prefixes=path_prefixes,
+                exc=exc,
+            )
+            return None
+        self._log_failure(
+            label=label,
+            rag_session_id=rag_session_id,
+            layer=layer,
+            exclude_tests=exclude_tests,
+            path_prefixes=path_prefixes,
+            exc=exc,
+            retried_without_test_filter=True,
+        )
+        try:
+            rows = operation()
+        except Exception as retry_exc:
+            self._log_failure(
+                label=f"{label} retry",
+                rag_session_id=rag_session_id,
+                layer=layer,
+                exclude_tests=False,
+                path_prefixes=path_prefixes,
+                exc=retry_exc,
+            )
+            return None
+        result = self._success_result(
+            rows,
+            rag_session_id=rag_session_id,
+            label=f"{label} retry",
+            include_spans=include_spans,
+            layer=layer,
+            exclude_tests=False,
+            path_prefixes=path_prefixes,
+        )
+        result.missing.append(f"{missing_prefix}:retried_without_test_filter")
+        return result
+
+    def _success_result(
+        self,
+        rows: list[dict],
+        *,
+        rag_session_id: str,
+        label: str,
+        include_spans: bool,
+        exclude_tests: bool,
+        path_prefixes: list[str] | None,
+        layer: str | None = None,
+    ) -> LayerRetrievalResult:
+        items = [self._to_item(row, include_spans=include_spans) for row in rows]
+        LOGGER.warning(
+            "%s: rag_session_id=%s layer=%s exclude_tests=%s path_prefixes=%s returned_count=%s top_paths=%s",
+            label,
+            rag_session_id,
+            layer,
+            exclude_tests,
+            path_prefixes or [],
+            len(items),
+            [item.source for item in items[:3]],
+        )
+        return LayerRetrievalResult(items=items)
+
+    def _log_failure(
+        self,
+        *,
+        label: str,
+        rag_session_id: str,
+        exclude_tests: bool,
+        path_prefixes: list[str] | None,
+        exc: Exception,
+        layer: str | None = None,
+        retried_without_test_filter: bool = False,
+    ) -> None:
+        LOGGER.warning(
+            "%s failed: rag_session_id=%s layer=%s exclude_tests=%s path_prefixes=%s retried_without_test_filter=%s error=%s",
+            label,
+            rag_session_id,
+            layer,
+            exclude_tests,
+            path_prefixes or [],
+            retried_without_test_filter,
+            self._exception_summary(exc),
+            exc_info=True,
+        )
+
+    def _filter_args(self, exclude_tests: bool) -> dict[str, list[str] | None]:
+        test_filters = build_test_filters() if exclude_tests else None
+        return {
+            "exclude_path_prefixes": test_filters.exclude_path_prefixes if test_filters else None,
+            "exclude_like_patterns": test_filters.exclude_like_patterns if test_filters else None,
+        }
+
+    def _failure_missing(self, prefix: str, exc: Exception) -> str:
+        return f"{prefix}:{self._exception_summary(exc)}"
+
+    def _exception_summary(self, exc: Exception) -> str:
+        message = " ".join(str(exc).split())
+        if len(message) > 180:
+            message = message[:177] + "..."
+        return f"{type(exc).__name__}:{message or 'no_message'}"
+
+    def _to_item(self, row: dict, *, include_spans: bool) -> LayeredRetrievalItem:
+        location = None
+        if include_spans:
+            location = CodeLocation(
+                path=str(row.get("path") or ""),
+                start_line=row.get("span_start"),
+                end_line=row.get("span_end"),
+            )
+        return LayeredRetrievalItem(
+            source=str(row.get("path") or ""),
+            content=str(row.get("content") or ""),
+            layer=str(row.get("layer") or ""),
+            title=str(row.get("title") or ""),
+            metadata=dict(row.get("metadata", {}) or {}),
+            score=row.get("distance"),
+            location=location,
+        )

app/modules/rag/explain/models.py

@@ -0,0 +1,91 @@
+from __future__ import annotations
+
+from typing import Any, Literal
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class ExplainHints(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    paths: list[str] = Field(default_factory=list)
+    symbols: list[str] = Field(default_factory=list)
+    endpoints: list[str] = Field(default_factory=list)
+    commands: list[str] = Field(default_factory=list)
+
+
+class ExplainIntent(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    raw_query: str
+    normalized_query: str
+    keywords: list[str] = Field(default_factory=list)
+    hints: ExplainHints = Field(default_factory=ExplainHints)
+    include_tests: bool = False
+    expected_entry_types: list[Literal["http", "cli"]] = Field(default_factory=list)
+    depth: Literal["high", "medium", "deep"] = "medium"
+
+
+class CodeLocation(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    path: str
+    start_line: int | None = None
+    end_line: int | None = None
+
+
+class LayeredRetrievalItem(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    source: str
+    content: str
+    layer: str
+    title: str
+    metadata: dict[str, Any] = Field(default_factory=dict)
+    score: float | None = None
+    location: CodeLocation | None = None
+
+
+class TracePath(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    symbol_ids: list[str] = Field(default_factory=list)
+    score: float = 0.0
+    entrypoint_id: str | None = None
+    notes: list[str] = Field(default_factory=list)
+
+
+class EvidenceItem(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    evidence_id: str
+    kind: Literal["entrypoint", "symbol", "edge", "excerpt"]
+    summary: str
+    location: CodeLocation | None = None
+    supports: list[str] = Field(default_factory=list)
+
+
+class CodeExcerpt(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    evidence_id: str
+    symbol_id: str | None = None
+    title: str
+    path: str
+    start_line: int | None = None
+    end_line: int | None = None
+    content: str
+    focus: str = "overview"
+
+
+class ExplainPack(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    intent: ExplainIntent
+    selected_entrypoints: list[LayeredRetrievalItem] = Field(default_factory=list)
+    seed_symbols: list[LayeredRetrievalItem] = Field(default_factory=list)
+    trace_paths: list[TracePath] = Field(default_factory=list)
+    evidence_index: dict[str, EvidenceItem] = Field(default_factory=dict)
+    code_excerpts: list[CodeExcerpt] = Field(default_factory=list)
+    missing: list[str] = Field(default_factory=list)
+    conflicts: list[str] = Field(default_factory=list)

app/modules/rag/explain/retriever_v2.py

@@ -0,0 +1,328 @@
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+from app.modules.rag.contracts.enums import RagLayer
+from app.modules.rag.explain.intent_builder import ExplainIntentBuilder
+from app.modules.rag.explain.layered_gateway import LayerRetrievalResult, LayeredRetrievalGateway
+from app.modules.rag.explain.models import CodeExcerpt, EvidenceItem, ExplainPack, LayeredRetrievalItem
+from app.modules.rag.explain.source_excerpt_fetcher import SourceExcerptFetcher
+from app.modules.rag.explain.trace_builder import TraceBuilder
+from app.modules.rag.retrieval.test_filter import exclude_tests_default, is_test_path
+
+LOGGER = logging.getLogger(__name__)
+_MIN_EXCERPTS = 2
+
+if TYPE_CHECKING:
+    from app.modules.rag.explain.graph_repository import CodeGraphRepository
+    from app.modules.rag.explain.models import ExplainIntent
+
+
+class CodeExplainRetrieverV2:
+    def __init__(
+        self,
+        gateway: LayeredRetrievalGateway,
+        graph_repository: CodeGraphRepository,
+        intent_builder: ExplainIntentBuilder | None = None,
+        trace_builder: TraceBuilder | None = None,
+        excerpt_fetcher: SourceExcerptFetcher | None = None,
+    ) -> None:
+        self._gateway = gateway
+        self._graph = graph_repository
+        self._intent_builder = intent_builder or ExplainIntentBuilder()
+        self._trace_builder = trace_builder or TraceBuilder(graph_repository)
+        self._excerpt_fetcher = excerpt_fetcher or SourceExcerptFetcher(graph_repository)
+
+    def build_pack(
+        self,
+        rag_session_id: str,
+        user_query: str,
+        *,
+        file_candidates: list[dict] | None = None,
+    ) -> ExplainPack:
+        intent = self._intent_builder.build(user_query)
+        path_prefixes = _path_prefixes(intent, file_candidates or [])
+        exclude_tests = exclude_tests_default() and not intent.include_tests
+        pack = self._run_pass(rag_session_id, intent, path_prefixes, exclude_tests=exclude_tests)
+        if exclude_tests and len(pack.code_excerpts) < _MIN_EXCERPTS:
+            self._merge_test_fallback(pack, rag_session_id, intent, path_prefixes)
+        self._log_pack(rag_session_id, pack)
+        return pack
+
+    def _run_pass(
+        self,
+        rag_session_id: str,
+        intent: ExplainIntent,
+        path_prefixes: list[str],
+        *,
+        exclude_tests: bool,
+    ) -> ExplainPack:
+        missing: list[str] = []
+        entrypoints_result = self._entrypoints(rag_session_id, intent, path_prefixes, exclude_tests=exclude_tests)
+        missing.extend(entrypoints_result.missing)
+        selected_entrypoints = self._filter_entrypoints(intent, entrypoints_result.items)
+        if not selected_entrypoints:
+            missing.append("layer:C3 empty")
+        seed_result = self._seed_symbols(rag_session_id, intent, path_prefixes, selected_entrypoints, exclude_tests=exclude_tests)
+        missing.extend(seed_result.missing)
+        seed_symbols = seed_result.items
+        if not seed_symbols:
+            missing.append("layer:C1 empty")
+        depth = 4 if intent.depth == "deep" else 3 if intent.depth == "medium" else 2
+        trace_paths = self._trace_builder.build_paths(rag_session_id, seed_symbols, max_depth=depth) if seed_symbols else []
+        excerpts, excerpt_evidence = self._excerpt_fetcher.fetch(rag_session_id, trace_paths) if trace_paths else ([], {})
+        if not excerpts:
+            lexical_result = self._gateway.retrieve_lexical_code(
+                rag_session_id,
+                intent.normalized_query,
+                limit=6,
+                path_prefixes=path_prefixes or None,
+                exclude_tests=exclude_tests,
+                include_spans=True,
+            )
+            missing.extend(lexical_result.missing)
+            excerpts, excerpt_evidence = _lexical_excerpts(lexical_result.items)
+            if not excerpts:
+                missing.append("layer:C0 empty")
+        evidence_index = _evidence_index(selected_entrypoints, seed_symbols)
+        evidence_index.update(excerpt_evidence)
+        missing.extend(_missing(selected_entrypoints, seed_symbols, trace_paths, excerpts))
+        return ExplainPack(
+            intent=intent,
+            selected_entrypoints=selected_entrypoints,
+            seed_symbols=seed_symbols,
+            trace_paths=trace_paths,
+            evidence_index=evidence_index,
+            code_excerpts=excerpts,
+            missing=_cleanup_missing(_dedupe(missing), has_excerpts=bool(excerpts)),
+            conflicts=[],
+        )
+
+    def _merge_test_fallback(
+        self,
+        pack: ExplainPack,
+        rag_session_id: str,
+        intent: ExplainIntent,
+        path_prefixes: list[str],
+    ) -> None:
+        lexical_result = self._gateway.retrieve_lexical_code(
+            rag_session_id,
+            intent.normalized_query,
+            limit=6,
+            path_prefixes=path_prefixes or None,
+            exclude_tests=False,
+            include_spans=True,
+        )
+        excerpt_offset = len([key for key in pack.evidence_index if key.startswith("excerpt_")])
+        excerpts, evidence = _lexical_excerpts(
+            lexical_result.items,
+            start_index=excerpt_offset,
+            is_test_fallback=True,
+        )
+        if not excerpts:
+            pack.missing = _dedupe(pack.missing + lexical_result.missing)
+            return
+        seen = {(item.path, item.start_line, item.end_line, item.content) for item in pack.code_excerpts}
+        for excerpt in excerpts:
+            key = (excerpt.path, excerpt.start_line, excerpt.end_line, excerpt.content)
+            if key in seen:
+                continue
+            pack.code_excerpts.append(excerpt)
+            seen.add(key)
+        pack.evidence_index.update(evidence)
+        pack.missing = _cleanup_missing(_dedupe(pack.missing + lexical_result.missing), has_excerpts=bool(pack.code_excerpts))
+
+    def _entrypoints(
+        self,
+        rag_session_id: str,
+        intent: ExplainIntent,
+        path_prefixes: list[str],
+        *,
+        exclude_tests: bool,
+    ) -> LayerRetrievalResult:
+        return self._gateway.retrieve_layer(
+            rag_session_id,
+            intent.normalized_query,
+            RagLayer.CODE_ENTRYPOINTS,
+            limit=6,
+            path_prefixes=path_prefixes or None,
+            exclude_tests=exclude_tests,
+            prefer_non_tests=True,
+            include_spans=True,
+        )
+
+    def _filter_entrypoints(self, intent: ExplainIntent, items: list[LayeredRetrievalItem]) -> list[LayeredRetrievalItem]:
+        if not intent.expected_entry_types:
+            return items[:3]
+        filtered = [item for item in items if str(item.metadata.get("entry_type") or "") in intent.expected_entry_types]
+        return filtered[:3] or items[:3]
+
+    def _seed_symbols(
+        self,
+        rag_session_id: str,
+        intent: ExplainIntent,
+        path_prefixes: list[str],
+        entrypoints: list[LayeredRetrievalItem],
+        *,
+        exclude_tests: bool,
+    ) -> LayerRetrievalResult:
+        symbol_result = self._gateway.retrieve_layer(
+            rag_session_id,
+            intent.normalized_query,
+            RagLayer.CODE_SYMBOL_CATALOG,
+            limit=12,
+            path_prefixes=path_prefixes or None,
+            exclude_tests=exclude_tests,
+            prefer_non_tests=True,
+            include_spans=True,
+        )
+        handlers: list[LayeredRetrievalItem] = []
+        handler_ids = [str(item.metadata.get("handler_symbol_id") or "") for item in entrypoints]
+        if handler_ids:
+            handlers = self._graph.get_symbols_by_ids(rag_session_id, [item for item in handler_ids if item])
+        seeds: list[LayeredRetrievalItem] = []
+        seen: set[str] = set()
+        for item in handlers + symbol_result.items:
+            symbol_id = str(item.metadata.get("symbol_id") or "")
+            if not symbol_id or symbol_id in seen:
+                continue
+            seen.add(symbol_id)
+            seeds.append(item)
+            if len(seeds) >= 8:
+                break
+        return LayerRetrievalResult(items=seeds, missing=list(symbol_result.missing))
+
+    def _log_pack(self, rag_session_id: str, pack: ExplainPack) -> None:
+        prod_excerpt_count = len([excerpt for excerpt in pack.code_excerpts if not _is_test_excerpt(excerpt)])
+        test_excerpt_count = len(pack.code_excerpts) - prod_excerpt_count
+        LOGGER.warning(
+            "code explain pack: rag_session_id=%s entrypoints=%s seeds=%s paths=%s excerpts=%s prod_excerpt_count=%s test_excerpt_count=%s missing=%s",
+            rag_session_id,
+            len(pack.selected_entrypoints),
+            len(pack.seed_symbols),
+            len(pack.trace_paths),
+            len(pack.code_excerpts),
+            prod_excerpt_count,
+            test_excerpt_count,
+            pack.missing,
+        )
+
+
+def _evidence_index(
+    entrypoints: list[LayeredRetrievalItem],
+    seed_symbols: list[LayeredRetrievalItem],
+) -> dict[str, EvidenceItem]:
+    result: dict[str, EvidenceItem] = {}
+    for index, item in enumerate(entrypoints, start=1):
+        evidence_id = f"entrypoint_{index}"
+        result[evidence_id] = EvidenceItem(
+            evidence_id=evidence_id,
+            kind="entrypoint",
+            summary=item.title,
+            location=item.location,
+            supports=[str(item.metadata.get("handler_symbol_id") or "")],
+        )
+    for index, item in enumerate(seed_symbols, start=1):
+        evidence_id = f"symbol_{index}"
+        result[evidence_id] = EvidenceItem(
+            evidence_id=evidence_id,
+            kind="symbol",
+            summary=item.title,
+            location=item.location,
+            supports=[str(item.metadata.get("symbol_id") or "")],
+        )
+    return result
+
+
+def _missing(
+    entrypoints: list[LayeredRetrievalItem],
+    seed_symbols: list[LayeredRetrievalItem],
+    trace_paths,
+    excerpts,
+) -> list[str]:
+    missing: list[str] = []
+    if not entrypoints:
+        missing.append("entrypoints")
+    if not seed_symbols:
+        missing.append("seed_symbols")
+    if not trace_paths:
+        missing.append("trace_paths")
+    if not excerpts:
+        missing.append("code_excerpts")
+    return missing
+
+
+def _lexical_excerpts(
+    items: list[LayeredRetrievalItem],
+    *,
+    start_index: int = 0,
+    is_test_fallback: bool = False,
+) -> tuple[list[CodeExcerpt], dict[str, EvidenceItem]]:
+    excerpts: list[CodeExcerpt] = []
+    evidence_index: dict[str, EvidenceItem] = {}
+    for item in items:
+        evidence_id = f"excerpt_{start_index + len(evidence_index) + 1}"
+        location = item.location
+        evidence_index[evidence_id] = EvidenceItem(
+            evidence_id=evidence_id,
+            kind="excerpt",
+            summary=item.title or item.source,
+            location=location,
+            supports=[],
+        )
+        focus = "lexical"
+        if _item_is_test(item):
+            focus = "test:lexical"
+        elif is_test_fallback:
+            focus = "lexical"
+        excerpts.append(
+            CodeExcerpt(
+                evidence_id=evidence_id,
+                symbol_id=str(item.metadata.get("symbol_id") or "") or None,
+                title=item.title or item.source,
+                path=item.source,
+                start_line=location.start_line if location else None,
+                end_line=location.end_line if location else None,
+                content=item.content,
+                focus=focus,
+            )
+        )
+    return excerpts, evidence_index
+
+
+def _item_is_test(item: LayeredRetrievalItem) -> bool:
+    return bool(item.metadata.get("is_test")) or is_test_path(item.source)
+
+
+def _is_test_excerpt(excerpt: CodeExcerpt) -> bool:
+    return excerpt.focus.startswith("test:") or is_test_path(excerpt.path)
+
+
+def _path_prefixes(intent: ExplainIntent, file_candidates: list[dict]) -> list[str]:
+    values: list[str] = []
+    for path in intent.hints.paths:
+        prefix = path.rsplit("/", 1)[0] if "/" in path else path
+        if prefix and prefix not in values:
+            values.append(prefix)
+    for item in file_candidates[:6]:
+        path = str(item.get("path") or "")
+        prefix = path.rsplit("/", 1)[0] if "/" in path else ""
+        if prefix and prefix not in values:
+            values.append(prefix)
+    return values
+
+
+def _cleanup_missing(values: list[str], *, has_excerpts: bool) -> list[str]:
+    if not has_excerpts:
+        return values
+    return [value for value in values if value not in {"code_excerpts", "layer:C0 empty"}]
+
+
+def _dedupe(values: list[str]) -> list[str]:
+    result: list[str] = []
+    for value in values:
+        item = value.strip()
+        if item and item not in result:
+            result.append(item)
+    return result

app/modules/rag/explain/source_excerpt_fetcher.py

@@ -0,0 +1,53 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from app.modules.rag.explain.excerpt_planner import ExcerptPlanner
+from app.modules.rag.explain.models import CodeExcerpt, EvidenceItem, TracePath
+from app.modules.rag.retrieval.test_filter import is_test_path
+
+if TYPE_CHECKING:
+    from app.modules.rag.explain.graph_repository import CodeGraphRepository
+
+
+class SourceExcerptFetcher:
+    def __init__(self, graph_repository: CodeGraphRepository, planner: ExcerptPlanner | None = None) -> None:
+        self._graph = graph_repository
+        self._planner = planner or ExcerptPlanner()
+
+    def fetch(
+        self,
+        rag_session_id: str,
+        trace_paths: list[TracePath],
+        *,
+        max_excerpts: int = 40,
+    ) -> tuple[list[CodeExcerpt], dict[str, EvidenceItem]]:
+        ordered_symbol_ids: list[str] = []
+        for path in trace_paths:
+            for symbol_id in path.symbol_ids:
+                if symbol_id and symbol_id not in ordered_symbol_ids:
+                    ordered_symbol_ids.append(symbol_id)
+        chunks = self._graph.get_chunks_by_symbol_ids(rag_session_id, ordered_symbol_ids)
+        excerpts: list[CodeExcerpt] = []
+        evidence_index: dict[str, EvidenceItem] = {}
+        for chunk in chunks:
+            symbol_id = str(chunk.metadata.get("symbol_id") or "")
+            evidence_id = f"excerpt_{len(evidence_index) + 1}"
+            location = chunk.location
+            evidence_index[evidence_id] = EvidenceItem(
+                evidence_id=evidence_id,
+                kind="excerpt",
+                summary=chunk.title,
+                location=location,
+                supports=[symbol_id] if symbol_id else [],
+            )
+            is_test_chunk = bool(chunk.metadata.get("is_test")) or is_test_path(location.path if location else chunk.source)
+            for excerpt in self._planner.plan(chunk, evidence_id=evidence_id, symbol_id=symbol_id):
+                if len(excerpts) >= max_excerpts:
+                    break
+                if is_test_chunk and not excerpt.focus.startswith("test:"):
+                    excerpt.focus = f"test:{excerpt.focus}"
+                excerpts.append(excerpt)
+            if len(excerpts) >= max_excerpts:
+                break
+        return excerpts, evidence_index

app/modules/rag/explain/trace_builder.py

@@ -0,0 +1,102 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from app.modules.rag.explain.models import LayeredRetrievalItem, TracePath
+
+if TYPE_CHECKING:
+    from app.modules.rag.explain.graph_repository import CodeGraphRepository
+
+
+class TraceBuilder:
+    def __init__(self, graph_repository: CodeGraphRepository) -> None:
+        self._graph = graph_repository
+
+    def build_paths(
+        self,
+        rag_session_id: str,
+        seed_symbols: list[LayeredRetrievalItem],
+        *,
+        max_depth: int,
+        max_paths: int = 3,
+        edge_types: list[str] | None = None,
+    ) -> list[TracePath]:
+        edges_filter = edge_types or ["calls", "imports", "inherits"]
+        symbol_map = self._symbol_map(seed_symbols)
+        paths: list[TracePath] = []
+        for seed in seed_symbols:
+            seed_id = str(seed.metadata.get("symbol_id") or "")
+            if not seed_id:
+                continue
+            queue: list[tuple[list[str], float, list[str]]] = [([seed_id], 0.0, [])]
+            while queue and len(paths) < max_paths * 3:
+                current_path, score, notes = queue.pop(0)
+                src_symbol_id = current_path[-1]
+                out_edges = self._graph.get_out_edges(rag_session_id, [src_symbol_id], edges_filter, limit_per_src=4)
+                if not out_edges or len(current_path) >= max_depth:
+                    paths.append(TracePath(symbol_ids=current_path, score=score, notes=notes))
+                    continue
+                for edge in out_edges:
+                    metadata = edge.metadata
+                    dst_symbol_id = str(metadata.get("dst_symbol_id") or "")
+                    next_notes = list(notes)
+                    next_score = score + self._edge_score(edge, symbol_map.get(src_symbol_id))
+                    if not dst_symbol_id:
+                        dst_ref = str(metadata.get("dst_ref") or "")
+                        package_hint = self._package_hint(symbol_map.get(src_symbol_id))
+                        resolved = self._graph.resolve_symbol_by_ref(rag_session_id, dst_ref, package_hint=package_hint)
+                        if resolved is not None:
+                            dst_symbol_id = str(resolved.metadata.get("symbol_id") or "")
+                            symbol_map[dst_symbol_id] = resolved
+                            next_score += 2.0
+                            next_notes.append(f"resolved:{dst_ref}")
+                    if not dst_symbol_id or dst_symbol_id in current_path:
+                        paths.append(TracePath(symbol_ids=current_path, score=next_score, notes=next_notes))
+                        continue
+                    if dst_symbol_id not in symbol_map:
+                        symbols = self._graph.get_symbols_by_ids(rag_session_id, [dst_symbol_id])
+                        if symbols:
+                            symbol_map[dst_symbol_id] = symbols[0]
+                    queue.append((current_path + [dst_symbol_id], next_score, next_notes))
+        unique = self._unique_paths(paths)
+        unique.sort(key=lambda item: item.score, reverse=True)
+        return unique[:max_paths] or [TracePath(symbol_ids=[seed.metadata.get("symbol_id", "")], score=0.0) for seed in seed_symbols[:1]]
+
+    def _edge_score(self, edge: LayeredRetrievalItem, source_symbol: LayeredRetrievalItem | None) -> float:
+        metadata = edge.metadata
+        score = 1.0
+        if str(metadata.get("resolution") or "") == "resolved":
+            score += 2.0
+        source_path = source_symbol.source if source_symbol is not None else ""
+        if source_path and edge.source == source_path:
+            score += 1.0
+        if "tests/" in edge.source or "/tests/" in edge.source:
+            score -= 3.0
+        return score
+
+    def _package_hint(self, symbol: LayeredRetrievalItem | None) -> str | None:
+        if symbol is None:
+            return None
+        package = str(symbol.metadata.get("package_or_module") or "")
+        if not package:
+            return None
+        return ".".join(package.split(".")[:-1]) or package
+
+    def _symbol_map(self, items: list[LayeredRetrievalItem]) -> dict[str, LayeredRetrievalItem]:
+        result: dict[str, LayeredRetrievalItem] = {}
+        for item in items:
+            symbol_id = str(item.metadata.get("symbol_id") or "")
+            if symbol_id:
+                result[symbol_id] = item
+        return result
+
+    def _unique_paths(self, items: list[TracePath]) -> list[TracePath]:
+        result: list[TracePath] = []
+        seen: set[tuple[str, ...]] = set()
+        for item in items:
+            key = tuple(symbol_id for symbol_id in item.symbol_ids if symbol_id)
+            if not key or key in seen:
+                continue
+            seen.add(key)
+            result.append(item)
+        return result

app/modules/rag/indexing/code/code_text/document_builder.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 from app.modules.rag.contracts import RagDocument, RagLayer, RagSource, RagSpan
 from app.modules.rag.indexing.code.code_text.chunker import CodeChunk
+from app.modules.rag.retrieval.test_filter import is_test_path
 
 
 class CodeTextDocumentBuilder:
@@ -17,6 +18,7 @@ class CodeTextDocumentBuilder:
                 "chunk_index": chunk_index,
                 "chunk_type": chunk.chunk_type,
                 "module_or_unit": source.path.replace("/", ".").removesuffix(".py"),
+                "is_test": is_test_path(source.path),
                 "artifact_type": "CODE",
             },
         )

app/modules/rag/indexing/code/edges/document_builder.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 from app.modules.rag.contracts import EvidenceLink, EvidenceType, RagDocument, RagLayer, RagSource, RagSpan
 from app.modules.rag.indexing.code.edges.extractor import PyEdge
+from app.modules.rag.retrieval.test_filter import is_test_path
 
 
 class EdgeDocumentBuilder:
@@ -22,6 +23,7 @@ class EdgeDocumentBuilder:
                 "dst_symbol_id": edge.dst_symbol_id,
                 "dst_ref": edge.dst_ref,
                 "resolution": edge.resolution,
+                "is_test": is_test_path(source.path),
                 "lang_payload": edge.metadata,
                 "artifact_type": "CODE",
             },

app/modules/rag/indexing/code/entrypoints/document_builder.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 from app.modules.rag.contracts import EvidenceLink, EvidenceType, RagDocument, RagLayer, RagSource, RagSpan
 from app.modules.rag.indexing.code.entrypoints.registry import Entrypoint
+from app.modules.rag.retrieval.test_filter import is_test_path
 
 
 class EntrypointDocumentBuilder:
@@ -19,6 +20,7 @@ class EntrypointDocumentBuilder:
                 "framework": entrypoint.framework,
                 "route_or_command": entrypoint.route_or_command,
                 "handler_symbol_id": entrypoint.handler_symbol_id,
+                "is_test": is_test_path(source.path),
                 "lang_payload": entrypoint.metadata,
                 "artifact_type": "CODE",
             },

app/modules/rag/indexing/code/symbols/document_builder.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 from app.modules.rag.contracts import RagDocument, RagLayer, RagSource, RagSpan
 from app.modules.rag.indexing.code.symbols.extractor import PySymbol
+from app.modules.rag.retrieval.test_filter import is_test_path
 
 
 class SymbolDocumentBuilder:
@@ -26,6 +27,7 @@ class SymbolDocumentBuilder:
                 "parent_symbol_id": symbol.parent_symbol_id,
                 "package_or_module": source.path.replace("/", ".").removesuffix(".py"),
                 "is_entry_candidate": bool(symbol.decorators),
+                "is_test": is_test_path(source.path),
                 "lang_payload": symbol.lang_payload,
                 "artifact_type": "CODE",
             },