agentengine-sdk-python 0.8.1__py3-none-any.whl → 0.8.2__py3-none-any.whl

ksadk/studio/authoring_coordinator.py

@@ -2,26 +2,124 @@
 
 from __future__ import annotations
 
+import copy
+import json
+import logging
 import os
 import shutil
 import threading
+import time
+from collections import OrderedDict
 from pathlib import Path
 from typing import Any, cast
 
-from ksadk.studio.authoring import AgentAuthoringService
+from ksadk.studio.authoring import AgentAuthoringService, ConversationProposal
 from ksadk.studio.codex_manifest import CodexAgentManifest
 from ksadk.studio.contracts import (
     AgentBindings,
     AgentDraft,
     AgentSpec,
-    Instructions,
+    CapabilitiesSpec,
+    CapabilityBinding,
     ModelSpec,
     RuntimeRef,
+    Usage,
 )
 from ksadk.studio.errors import StudioError
 from ksadk.studio.identifiers import generate_agent_slug, is_generated_agent_slug
 from ksadk.studio.templates import default_agent_spec
 
+LOGGER = logging.getLogger(__name__)
+
+# 对话创建阶段推进序列。前端只消费阶段名展示两段式文案，不解析内容。
+CONVERSATION_STAGES = (
+    "resolving_model",
+    "generating",
+    "codex_writing",
+    "validating",
+    "correcting",
+    "done",
+    "failed",
+)
+# 进度记录只保留最近的少量请求，避免长驻进程无限增长。
+_CONVERSATION_STATUS_LIMIT = 64
+_LOCAL_FALLBACK_MODEL_ERRORS = frozenset(
+    {
+        "AUTHORING_MODEL_OUTPUT_INVALID",
+        "MODEL_EMPTY_RESPONSE",
+        "MODEL_RATE_LIMITED",
+        "MODEL_REQUEST_FAILED",
+        "MODEL_RESPONSE_INVALID",
+        "MODEL_RESPONSE_TOO_LARGE",
+    }
+)
+_LOCAL_FALLBACK_REASON = {
+    "AUTHORING_MODEL_OUTPUT_INVALID": "invalid-model-output",
+    "MODEL_EMPTY_RESPONSE": "empty-model-output",
+    "MODEL_RATE_LIMITED": "model-rate-limited",
+    "MODEL_REQUEST_FAILED": "model-request-failed",
+    "MODEL_RESPONSE_INVALID": "invalid-model-response",
+    "MODEL_RESPONSE_TOO_LARGE": "model-response-too-large",
+}
+
+
+def _deduplicated_bindings(resource_ids: list[str]) -> list[CapabilityBinding]:
+    """Build a deterministic binding list from Studio-selected resource ids."""
+
+    seen: set[str] = set()
+    result: list[CapabilityBinding] = []
+    for resource_id in resource_ids:
+        normalized = str(resource_id or "").strip()
+        if not normalized or normalized in seen:
+            continue
+        seen.add(normalized)
+        result.append(CapabilityBinding(resource_id=normalized))
+    return result
+
+
+def _inject_managed_bindings(
+    proposal: Any,
+    *,
+    model_spec: ModelSpec,
+    bindings: AgentBindings,
+) -> Any:
+    """Replace all model/resource fields invented by the authoring model.
+
+    The LLM may propose product semantics (name, runtime, instructions and
+    execution strategy), but it is never a source of connection information or
+    capability identities.  Those are selected in Studio and validated before
+    this point.  The Profile id remains the sole persisted model reference;
+    compiler/run materialisation resolves the current endpoint, credential and
+    parameters from the catalog.  This also prevents provider discovery data,
+    limits and pricing from leaking into every Agent Revision.  Resetting
+    capabilities is intentional: a model-generated inline Tool/MCP/Skill
+    contract must not become a deployable capability.
+    """
+
+    spec = proposal.spec
+    managed_bindings = bindings
+    # Codex Runtime owns its native tool surface.  Studio-selected KsADK
+    # Tool contracts are meaningful for generic ADK/LangGraph Agents, but
+    # serialising them into a Codex Draft Patch advertises a binding the
+    # runtime cannot execute.  Enforce the same boundary as the UI here so a
+    # direct API caller cannot bypass it.
+    if proposal.runtimeType == "codex" and bindings.tools:
+        managed_bindings = bindings.model_copy(update={"tools": []})
+    # Resolution above is still intentional: it validates the selected Profile
+    # before a proposal is returned.  Do not serialize its catalog contract.
+    _ = model_spec
+    return proposal.model_copy(
+        update={
+            "spec": spec.model_copy(
+                update={
+                    "model": None,
+                    "bindings": managed_bindings.model_copy(deep=True),
+                    "capabilities": CapabilitiesSpec(),
+                }
+            )
+        }
+    )
+
 
 class StudioAuthoringCoordinator:
     """Coordinates repositories without expanding the StudioService façade."""
@@ -29,7 +127,140 @@ class StudioAuthoringCoordinator:
     def __init__(self, studio: Any) -> None:
         self.studio = studio
         self.backend = AgentAuthoringService(studio.workspace)
+        # Codex authoring 执行器可注入替换（测试）；默认惰性探测可用性。
+        self.codex_authoring: Any = getattr(studio, "codex_authoring_executor", None)
         self._id_lock = threading.Lock()
+        self._status_lock = threading.Lock()
+        self._conversation_status: OrderedDict[str, dict[str, Any]] = OrderedDict()
+
+    def _codex_authoring_executor(self) -> Any:
+        if self.codex_authoring is not None:
+            return self.codex_authoring
+        from ksadk.studio.codex_authoring import CodexAuthoringExecutor
+
+        self.codex_authoring = CodexAuthoringExecutor(
+            self.studio.workspace,
+            credential_resolver=self.studio.credentials,
+        )
+        return self.codex_authoring
+
+    def _use_codex_authoring(self) -> bool:
+        """Use heavy Codex filesystem authoring only by explicit opt-in.
+
+        Conversational Draft Patch authoring is a bounded structured-chat
+        request.  Probing a local Codex installation and silently switching to
+        its full filesystem/tool harness made a small form action huge and
+        unreliable (and changed the selected model's request shape).  Keep the
+        optional expert authoring path, but never choose it implicitly.
+        """
+
+        mode = str(os.environ.get("KSADK_STUDIO_AUTHORIZER") or "").strip().lower()
+        if mode in {"codex", "codex-writing", "1", "true"}:
+            return True
+        if mode in {"chat", "off", "0", "false", "none"}:
+            return False
+        return False
+
+    # ------------------------------------------------------------------
+    # Conversation authoring stage tracking
+    # ------------------------------------------------------------------
+
+    def _record_conversation_stage(
+        self,
+        request_id: str | None,
+        stage: str,
+        *,
+        detail: str | None = None,
+    ) -> None:
+        if not request_id:
+            return
+        entry = {
+            "requestId": request_id,
+            "stage": stage,
+            "updatedAt": time.time(),
+            **({"detail": detail} if detail else {}),
+        }
+        with self._status_lock:
+            self._conversation_status[request_id] = entry
+            self._conversation_status.move_to_end(request_id)
+            while len(self._conversation_status) > _CONVERSATION_STATUS_LIMIT:
+                self._conversation_status.popitem(last=False)
+
+    def conversation_status(self, request_id: str) -> dict[str, Any] | None:
+        with self._status_lock:
+            entry = self._conversation_status.get(request_id)
+            return dict(entry) if entry else None
+
+    def _local_conversation_fallback(
+        self,
+        *,
+        messages: list[dict[str, str]],
+        previous_proposal: ConversationProposal | None,
+        runtime_type: str,
+        model_spec: ModelSpec,
+        bindings: AgentBindings,
+        reason_code: str,
+    ) -> dict[str, Any]:
+        """Return a reviewable draft when the selected authoring model is unavailable.
+
+        This is intentionally not a hidden second model or an auto-deploy
+        mechanism.  Studio keeps the user's selected runtime and resource
+        bindings, produces only deterministic editable semantic fields, and
+        tells the caller that a manual review is required.
+        """
+
+        latest_user_message = next(
+            (
+                str(item.get("content") or "").strip()
+                for item in reversed(messages)
+                if str(item.get("role") or "").strip() == "user"
+                and str(item.get("content") or "").strip()
+            ),
+            "请根据已确认的需求完成 Agent，并在部署前检查配置和权限。",
+        )
+        previous_instructions = (
+            previous_proposal.spec.instructions if previous_proposal else None
+        )
+        fallback_payload = {
+            "name": previous_proposal.name if previous_proposal else "待确认 Agent",
+            "slug": previous_proposal.slug if previous_proposal else "conversation-agent",
+            "description": (
+                previous_proposal.description
+                if previous_proposal and previous_proposal.description
+                else latest_user_message[:1024]
+            ),
+            "spec": {
+                "instructions": {
+                    "system": (
+                        previous_instructions.system
+                        if previous_instructions and previous_instructions.system
+                        else "根据用户需求完成任务；不确定时先澄清，并遵守已绑定能力的权限边界。"
+                    ),
+                    "task": latest_user_message,
+                }
+            },
+        }
+        proposal = self.backend.parse_conversation_proposal(
+            json.dumps(fallback_payload, ensure_ascii=False),
+            base=previous_proposal,
+            runtime_type=runtime_type,
+        )
+        return {
+            "proposal": _inject_managed_bindings(
+                proposal,
+                model_spec=model_spec,
+                bindings=bindings,
+            ).model_dump(by_alias=True, mode="json"),
+            "requiresConfirmation": True,
+            "authoringMode": "local-fallback",
+            "fallback": {
+                "active": True,
+                "reason": _LOCAL_FALLBACK_REASON.get(reason_code, "model-unavailable"),
+            },
+            "usage": Usage(source="local-fallback").model_dump(
+                by_alias=True, mode="json"
+            ),
+        }
 
     def create(
         self,
@@ -47,7 +278,30 @@ class StudioAuthoringCoordinator:
             resolved = (spec or default_agent_spec(template, description=description)).model_copy(
                 deep=True
             )
-            resolved.runtime = self.backend.runtime_ref(agent_id, runtime_type)
+            canonical_runtime = self.backend.runtime_ref(agent_id, runtime_type)
+            proposed_runtime = resolved.runtime
+            if proposed_runtime is not None and proposed_runtime.type != runtime_type:
+                raise StudioError(
+                    "AGENT_RUNTIME_MISMATCH",
+                    "AgentSpec Runtime 与创建方式不一致",
+                    status_code=422,
+                    field="runtimeType",
+                    details={
+                        "runtimeType": runtime_type,
+                        "specRuntimeType": proposed_runtime.type,
+                    },
+                )
+            if proposed_runtime is not None:
+                if runtime_type == "codex" and proposed_runtime.version:
+                    canonical_runtime.version = proposed_runtime.version
+                elif runtime_type in {"adk", "langgraph"}:
+                    canonical_runtime.entry_point = (
+                        proposed_runtime.entry_point or canonical_runtime.entry_point
+                    )
+                    canonical_runtime.agent_variable = proposed_runtime.agent_variable
+                    canonical_runtime.version = proposed_runtime.version
+                    canonical_runtime.detection = proposed_runtime.detection
+            resolved.runtime = canonical_runtime
             if description:
                 resolved.description = description
             draft = self.studio.create_studio_agent(
@@ -94,7 +348,15 @@ class StudioAuthoringCoordinator:
                 status_code=422,
             )
         if spec.runtime.type in {"adk", "langgraph"}:
-            spec.runtime = self.backend.runtime_ref(agent_id, spec.runtime.type)
+            imported_runtime = spec.runtime
+            canonical_runtime = self.backend.runtime_ref(agent_id, imported_runtime.type)
+            canonical_runtime.entry_point = (
+                imported_runtime.entry_point or canonical_runtime.entry_point
+            )
+            canonical_runtime.agent_variable = imported_runtime.agent_variable
+            canonical_runtime.version = imported_runtime.version
+            canonical_runtime.detection = imported_runtime.detection
+            spec.runtime = canonical_runtime
         created = self.studio.create_agent(
             agent_id=agent_id,
             name=display_name,
@@ -123,7 +385,24 @@ class StudioAuthoringCoordinator:
         return cast(AgentDraft, self.studio.drafts.get(agent_id))
 
     def inspect_project(self, project_path: str) -> dict:
-        return self.backend.inspect_project(project_path)
+        inspection = self.backend.inspect_project(project_path)
+        spec, unresolved = self._project_agent_spec(
+            inspection,
+            agent_id="agentkit-preview",
+            model_profile_id=None,
+        )
+        return {
+            **inspection,
+            "agentSpec": spec.model_dump(by_alias=True, exclude_none=True, mode="json"),
+            "bindingProjection": {
+                "preserved": spec.bindings.model_dump(
+                    by_alias=True,
+                    exclude_none=True,
+                    mode="json",
+                ),
+                "unresolved": unresolved,
+            },
+        }
 
     def commit_project(
         self,
@@ -155,30 +434,22 @@ class StudioAuthoringCoordinator:
 
         agent_id = self._allocate_agent_id(resolved_slug)
         resolved_slug = resolved_slug or agent_id
-        config = inspection.get("evidence", {}).get("config", {})
-        prompt = str(
-            config.get("prompt")
-            or config.get("instruction")
-            or "You are a reliable assistant imported from an existing project."
-        )
-        runtime = RuntimeRef(
-            type=cast(Any, runtime_type),
-            project_path=str(inspection["projectPath"]),
-            entry_point=str(inspection.get("entryPoint") or "agent.py"),
-            agent_variable=str(
-                inspection.get("agentVariable")
-                or ("graph" if runtime_type == "langgraph" else "root_agent")
-            ),
-            detection="auto",
+        spec, unresolved = self._project_agent_spec(
+            inspection,
+            agent_id=agent_id,
+            model_profile_id=model_profile_id,
         )
+        if unresolved:
+            raise StudioError(
+                "PROJECT_BINDINGS_UNRESOLVED",
+                "项目中的 Tool、MCP 或 Skill 绑定无法无损映射，请先安装或修正对应资源",
+                status_code=422,
+                details={"unresolved": unresolved},
+            )
         created = self.studio.create_agent(
             agent_id=agent_id,
             name=display_name,
-            spec=AgentSpec(
-                runtime=runtime,
-                instructions=Instructions(system=prompt),
-                bindings=AgentBindings(model_profile_id=model_profile_id),
-            ),
+            spec=spec,
             labels={
                 "agentkit.ksyun.com/slug": self.backend.normalize_slug(resolved_slug),
                 "agentkit.ksyun.com/source": "project-detection",
@@ -193,32 +464,511 @@ class StudioAuthoringCoordinator:
         *,
         messages: list[dict[str, str]],
         model_profile_id: str,
+        runtime_type: str = "codex",
+        agent_model_profile_ids: list[str] | None = None,
+        agent_default_model_profile_id: str | None = None,
+        tool_resource_ids: list[str] | None = None,
+        mcp_resource_ids: list[str] | None = None,
+        skill_resource_ids: list[str] | None = None,
+        request_id: str | None = None,
     ) -> dict:
-        model_spec = self.studio.catalog.resolve_model(
-            AgentBindings(model_profile_id=model_profile_id)
+        started = time.monotonic()
+        self._record_conversation_stage(request_id, "resolving_model")
+        LOGGER.info(
+            "conversation authoring started: modelProfileId=%s messages=%d requestId=%s",
+            model_profile_id,
+            len(messages),
+            request_id or "-",
         )
-        if model_spec is None:
+        try:
+            result = await self._compose_conversation_inner(
+                messages=messages,
+                model_profile_id=model_profile_id,
+                runtime_type=runtime_type,
+                agent_model_profile_ids=agent_model_profile_ids or [],
+                agent_default_model_profile_id=agent_default_model_profile_id,
+                tool_resource_ids=tool_resource_ids or [],
+                mcp_resource_ids=mcp_resource_ids or [],
+                skill_resource_ids=skill_resource_ids or [],
+                request_id=request_id,
+                started=started,
+            )
+        except Exception as exc:
+            self._record_conversation_stage(request_id, "failed", detail=str(exc))
+            LOGGER.warning(
+                "conversation authoring failed after %.2fs: modelProfileId=%s reason=%s",
+                time.monotonic() - started,
+                model_profile_id,
+                exc,
+            )
+            raise
+        return result
+
+    async def _compose_conversation_inner(
+        self,
+        *,
+        messages: list[dict[str, str]],
+        model_profile_id: str,
+        runtime_type: str,
+        agent_model_profile_ids: list[str],
+        agent_default_model_profile_id: str | None,
+        tool_resource_ids: list[str],
+        mcp_resource_ids: list[str],
+        skill_resource_ids: list[str],
+        request_id: str | None,
+        started: float,
+    ) -> dict:
+        # The authoring model is a one-off control-plane choice.  The Agent's
+        # model allow-list is a separate deploy-time contract and can contain
+        # multiple profiles.  Keep backwards compatibility for API callers
+        # that only send ``modelProfileId`` by using it as the one Agent model.
+        selected_agent_model_ids = [
+            binding.resource_id
+            for binding in _deduplicated_bindings(agent_model_profile_ids or [model_profile_id])
+        ]
+        if not selected_agent_model_ids:
+            raise StudioError(
+                "AGENT_MODEL_REQUIRED",
+                "Agent 至少需要选择一个可用 Model Profile",
+                status_code=422,
+            )
+        default_agent_model_id = str(
+            agent_default_model_profile_id or selected_agent_model_ids[0]
+        ).strip()
+        if default_agent_model_id not in selected_agent_model_ids:
+            raise StudioError(
+                "AGENT_DEFAULT_MODEL_NOT_SELECTED",
+                "Agent 默认模型必须包含在可用模型列表中",
+                status_code=422,
+                details={
+                    "defaultModelProfileId": default_agent_model_id,
+                    "modelProfileIds": selected_agent_model_ids,
+                },
+            )
+        authoring_bindings = AgentBindings(
+            model_profile_id=model_profile_id,
+            model_profile_ids=[model_profile_id],
+        )
+        bindings = AgentBindings(
+            model_profile_id=default_agent_model_id,
+            model_profile_ids=selected_agent_model_ids,
+            tools=_deduplicated_bindings(tool_resource_ids),
+            mcp_servers=_deduplicated_bindings(mcp_resource_ids),
+            skills=_deduplicated_bindings(skill_resource_ids),
+        )
+        # Resolve every selected resource now.  This validates kind, readiness
+        # and capability contracts before an LLM response can be displayed as a
+        # deployable Draft Patch.
+        authoring_model_spec = self.studio.catalog.resolve_model(authoring_bindings)
+        if authoring_model_spec is None:
             raise StudioError(
                 "AGENT_MODEL_REQUIRED",
                 "对话构建需要选择一个 Model Profile",
                 status_code=422,
             )
-        model = self.studio.catalog.resolver.resolve_model(model_spec)
-        response = await self.studio.model_client.complete(
-            model,
-            messages=self.backend.conversation_messages(messages),
-            network_policy=self.backend.authoring_network_policy(model.endpoint_url),
-            timeout_seconds=60,
-            max_attempts=2,
-            backoff_seconds=1,
+        # Resolve every Agent model before returning a Draft Patch.  This
+        # catches a deleted, wrong-kind or malformed secondary profile rather
+        # than accepting a non-deployable multi-model Agent in the browser.
+        self.studio.catalog.resolve_models(bindings)
+        model_spec = self.studio.catalog.resolve_model(bindings)
+        if model_spec is None:  # defensive: bindings above requires a default
+            raise StudioError(
+                "AGENT_MODEL_REQUIRED",
+                "Agent 需要选择一个默认 Model Profile",
+                status_code=422,
+            )
+        self.studio.catalog.policy_preview(bindings)
+        self.studio.catalog.resolve_mcp_servers(bindings)
+        self.studio.catalog.resolve_skills(bindings)
+        model = self.studio.catalog.resolver.resolve_model(authoring_model_spec)
+        # authoring 自身的 max_tokens 跟随 profile：未配置则 payload 不带该字段
+        # （服务端默认）；finishReason=length 截断由 model_client 一次性扩容重试兜底。
+        LOGGER.info(
+            "conversation authoring model resolved: model=%s endpoint=%s",
+            getattr(model, "model", "-"),
+            getattr(model, "endpoint_url", "-"),
+        )
+        normalized_runtime_type = str(runtime_type or "").strip().lower()
+        if normalized_runtime_type not in {"codex", "adk", "langgraph"}:
+            raise StudioError(
+                "AGENT_RUNTIME_INVALID",
+                "对话构建 Runtime 仅支持 Codex、ADK 或 LangGraph",
+                status_code=422,
+                field="runtimeType",
+            )
+        normalized_messages = self.backend.conversation_messages(
+            messages,
+            runtime_type=normalized_runtime_type,
+        )
+        previous_proposal = None
+        for item in reversed(messages):
+            if str(item.get("role") or "").strip() != "assistant":
+                continue
+            try:
+                previous_proposal = self.backend.parse_conversation_proposal(
+                    str(item.get("content") or ""),
+                    runtime_type=normalized_runtime_type,
+                )
+            except StudioError:
+                continue
+            break
+
+        if self._use_codex_authoring():
+            codex_result = await self._compose_conversation_codex(
+                messages=messages,
+                model=model,
+                previous_proposal=previous_proposal,
+                request_id=request_id,
+                started=started,
+                model_profile_id=model_profile_id,
+                model_spec=model_spec,
+                bindings=bindings,
+            )
+            if codex_result is not None:
+                return codex_result
+
+        request_options = {
+            "network_policy": self.backend.authoring_network_policy(model.endpoint_url),
+            # The authoring output is deliberately small.  Keep a bounded
+            # request budget so an unavailable upstream cannot make a simple
+            # create flow appear to hang for minutes.
+            "timeout_seconds": 20,
+            "backoff_seconds": 1,
+            "response_format": {"type": "json_object"},
+            # finishReason=length 截断时在 model_client 内自动扩容 max_tokens
+            # 重发一次（一次性），避免大 JSON 被截断成空响应。
+            "retry_on_length": True,
+        }
+        self._record_conversation_stage(request_id, "generating")
+        try:
+            response = await self.studio.model_client.complete(
+                model,
+                messages=normalized_messages,
+                max_attempts=2,
+                **request_options,
+            )
+        except StudioError as exc:
+            if exc.code not in _LOCAL_FALLBACK_MODEL_ERRORS:
+                raise
+            LOGGER.warning(
+                "conversation authoring model unavailable; returning local fallback: "
+                "modelProfileId=%s reason=%s",
+                model_profile_id,
+                exc.code,
+            )
+            self._record_conversation_stage(request_id, "done")
+            return self._local_conversation_fallback(
+                messages=messages,
+                previous_proposal=previous_proposal,
+                runtime_type=normalized_runtime_type,
+                model_spec=model_spec,
+                bindings=bindings,
+                reason_code=exc.code,
+            )
+        self._record_conversation_stage(request_id, "validating")
+        try:
+            proposal = self.backend.parse_conversation_proposal(
+                response.content,
+                base=previous_proposal,
+                runtime_type=normalized_runtime_type,
+            )
+        except StudioError as exc:
+            if exc.code != "AUTHORING_MODEL_OUTPUT_INVALID":
+                raise
+            LOGGER.warning(
+                "conversation authoring patch invalid; returning local fallback: "
+                "modelProfileId=%s",
+                model_profile_id,
+            )
+            self._record_conversation_stage(request_id, "done")
+            return self._local_conversation_fallback(
+                messages=messages,
+                previous_proposal=previous_proposal,
+                runtime_type=normalized_runtime_type,
+                model_spec=model_spec,
+                bindings=bindings,
+                reason_code=exc.code,
+            )
+        self._record_conversation_stage(request_id, "done")
+        LOGGER.info(
+            "conversation authoring finished in %.2fs: modelProfileId=%s slug=%s",
+            time.monotonic() - started,
+            model_profile_id,
+            getattr(proposal, "slug", "-"),
         )
-        proposal = self.backend.parse_conversation_proposal(response.content)
         return {
-            "proposal": proposal.model_dump(mode="json"),
+            "proposal": _inject_managed_bindings(
+                proposal,
+                model_spec=model_spec,
+                bindings=bindings,
+            ).model_dump(
+                by_alias=True, mode="json"
+            ),
             "requiresConfirmation": True,
+            "authoringMode": "chat",
             "usage": response.usage.model_dump(by_alias=True, mode="json"),
         }
 
+    async def _compose_conversation_codex(
+        self,
+        *,
+        messages: list[dict[str, str]],
+        model: Any,
+        previous_proposal: Any,
+        request_id: str | None,
+        started: float,
+        model_profile_id: str,
+        model_spec: ModelSpec,
+        bindings: AgentBindings,
+    ) -> dict | None:
+        """让真实 Codex 会话在工作区写 agentkit.yaml；任何失败降级 chat 链。
+
+        返回 ``None`` 表示应降级（探测失败/超时/重试后仍不合法），调用方继续走
+        既有 chat 链路，保证零回归。
+        """
+
+        self._record_conversation_stage(request_id, "codex_writing")
+        executor = self._codex_authoring_executor()
+        try:
+            result = await executor.compose(
+                messages=messages,
+                model=model,
+                base=previous_proposal,
+                request_id=request_id,
+            )
+        except Exception as exc:
+            LOGGER.warning(
+                "codex authoring failed after %.2fs, falling back to chat chain: "
+                "modelProfileId=%s requestId=%s reason=%s",
+                time.monotonic() - started,
+                model_profile_id,
+                request_id or "-",
+                exc,
+            )
+            self._record_conversation_stage(
+                request_id, "generating", detail=f"codex authoring 降级: {exc}"
+            )
+            return None
+        self._record_conversation_stage(request_id, "done")
+        LOGGER.info(
+            "conversation authoring finished in %.2fs via codex: modelProfileId=%s "
+            "slug=%s attempts=%d",
+            time.monotonic() - started,
+            model_profile_id,
+            getattr(result.proposal, "slug", "-"),
+            result.attempts,
+        )
+        return {
+            "proposal": _inject_managed_bindings(
+                result.proposal,
+                model_spec=model_spec,
+                bindings=bindings,
+            ).model_dump(
+                by_alias=True, mode="json"
+            ),
+            "requiresConfirmation": True,
+            "authoringMode": "codex",
+            "usage": result.usage.model_dump(by_alias=True, mode="json"),
+        }
+
+    def _project_agent_spec(
+        self,
+        inspection: dict[str, Any],
+        *,
+        agent_id: str,
+        model_profile_id: str | None,
+    ) -> tuple[AgentSpec, list[dict[str, Any]]]:
+        """Project detected project config without silently dropping authoring fields."""
+
+        config = inspection.get("evidence", {}).get("config", {})
+        if not isinstance(config, dict):
+            config = {}
+        embedded = config.get("spec")
+        payload: dict[str, Any] = copy.deepcopy(embedded) if isinstance(embedded, dict) else {}
+
+        for field in (
+            "description",
+            "model",
+            "capabilities",
+            "bindings",
+            "execution",
+            "context",
+            "memory",
+            "security",
+            "evaluation",
+        ):
+            if field not in payload and field in config:
+                payload[field] = copy.deepcopy(config[field])
+
+        instructions = payload.get("instructions")
+        if not isinstance(instructions, dict):
+            instructions = {}
+        instructions.setdefault(
+            "system",
+            str(
+                config.get("prompt")
+                or config.get("instruction")
+                or "You are a reliable assistant imported from an existing project."
+            ),
+        )
+        instructions.setdefault("task", str(config.get("task_prompt") or config.get("task") or ""))
+        payload["instructions"] = instructions
+
+        model_payload = payload.get("model")
+        if isinstance(model_payload, str) and model_payload.strip():
+            upstream = (
+                (os.environ.get("OPENAI_BASE_URL") or os.environ.get("OPENAI_API_BASE") or "")
+                .strip()
+                .rstrip("/")
+            )
+            endpoint = (
+                {"endpointUrl": upstream}
+                if upstream.endswith("/chat/completions")
+                else {"baseUrl": upstream or "https://api.openai.com/v1"}
+            )
+            payload["model"] = {
+                "model": model_payload.strip(),
+                "credentialRef": "env://AGENTKIT_MODEL_API_KEY",
+                **endpoint,
+            }
+
+        bindings = payload.get("bindings")
+        if not isinstance(bindings, dict):
+            bindings = {}
+        else:
+            bindings = copy.deepcopy(bindings)
+        capabilities = payload.get("capabilities")
+        if not isinstance(capabilities, dict):
+            capabilities = {}
+        else:
+            capabilities = copy.deepcopy(capabilities)
+        unresolved: list[dict[str, Any]] = []
+        legacy_fields = (
+            ("tools", "tools"),
+            ("mcpServers", "mcp_servers"),
+            ("skills", "skills"),
+        )
+        for canonical, legacy in legacy_fields:
+            if canonical in bindings or canonical in capabilities:
+                continue
+            raw = config.get(canonical, config.get(legacy))
+            if raw is None:
+                continue
+            if not isinstance(raw, list):
+                unresolved.append(
+                    {"kind": canonical, "value": copy.deepcopy(raw), "reason": "not-a-list"}
+                )
+                continue
+            projected_bindings: list[dict[str, Any]] = []
+            projected_capabilities: list[dict[str, Any]] = []
+            for item in raw:
+                if isinstance(item, str) and item.strip():
+                    projected_bindings.append({"resourceId": item.strip()})
+                elif isinstance(item, dict) and (item.get("resourceId") or item.get("resource_id")):
+                    projected_bindings.append(copy.deepcopy(item))
+                elif isinstance(item, dict) and item.get("name") and item.get("version"):
+                    projected_capabilities.append(copy.deepcopy(item))
+                else:
+                    unresolved.append(
+                        {
+                            "kind": canonical,
+                            "value": copy.deepcopy(item),
+                            "reason": "unsupported-binding-shape",
+                        }
+                    )
+            if projected_bindings:
+                bindings[canonical] = projected_bindings
+            if projected_capabilities:
+                capabilities[canonical] = projected_capabilities
+
+        if model_profile_id:
+            existing_profiles = list(
+                bindings.get("modelProfileIds") or bindings.get("model_profile_ids") or []
+            )
+            bindings["modelProfileId"] = model_profile_id
+            bindings["modelProfileIds"] = list(
+                dict.fromkeys([model_profile_id, *existing_profiles])
+            )
+        if "modelParameters" not in bindings and "model_parameters" in config:
+            bindings["modelParameters"] = copy.deepcopy(config["model_parameters"])
+        if "policyTemplate" not in bindings:
+            policy = config.get("policy_template", config.get("policy"))
+            if isinstance(policy, str) and policy:
+                bindings["policyTemplate"] = policy
+        payload["bindings"] = bindings
+        payload["capabilities"] = capabilities
+
+        try:
+            spec = AgentSpec.model_validate(payload)
+        except ValueError as exc:
+            raise StudioError(
+                "PROJECT_AGENT_SPEC_INVALID",
+                "项目配置无法无损转换为 AgentSpec",
+                status_code=422,
+                details={"reason": str(exc)},
+            ) from exc
+
+        runtime_type = str(inspection["runtimeType"])
+        detected_runtime = self.backend.runtime_ref(agent_id, runtime_type)
+        supplied_runtime = spec.runtime
+        if supplied_runtime is not None and supplied_runtime.type != runtime_type:
+            raise StudioError(
+                "PROJECT_RUNTIME_MISMATCH",
+                "项目配置中的 Runtime 与检测结果不一致",
+                status_code=422,
+                details={
+                    "detected": runtime_type,
+                    "configured": supplied_runtime.type,
+                },
+            )
+        if runtime_type in {"adk", "langgraph"}:
+            detected_runtime = RuntimeRef(
+                type=cast(Any, runtime_type),
+                project_path=str(inspection["projectPath"]),
+                entry_point=(
+                    supplied_runtime.entry_point
+                    if supplied_runtime and supplied_runtime.entry_point
+                    else str(inspection.get("entryPoint") or "agent.py")
+                ),
+                agent_variable=(
+                    supplied_runtime.agent_variable
+                    if supplied_runtime
+                    else str(
+                        inspection.get("agentVariable")
+                        or ("graph" if runtime_type == "langgraph" else "root_agent")
+                    )
+                ),
+                version=supplied_runtime.version if supplied_runtime else None,
+                detection="auto",
+            )
+        elif supplied_runtime is not None and supplied_runtime.version:
+            detected_runtime.version = supplied_runtime.version
+        spec.runtime = detected_runtime
+
+        catalog_ids = {
+            kind: {item.resource_id for item in self.studio.catalog.list(kind=kind, limit=500)}
+            for kind in ("tool", "mcp", "skill")
+        }
+        for kind, values in (
+            ("tool", spec.bindings.tools),
+            ("mcp", spec.bindings.mcp_servers),
+            ("skill", spec.bindings.skills),
+        ):
+            for binding in values:
+                if binding.resource_id not in catalog_ids[kind]:
+                    unresolved.append(
+                        {
+                            "kind": kind,
+                            "value": binding.model_dump(
+                                by_alias=True,
+                                exclude_none=True,
+                                mode="json",
+                            ),
+                            "reason": "not-in-resource-catalog",
+                        }
+                    )
+        return spec, unresolved
+
     def _agent_exists(self, agent_id: str) -> bool:
         return bool(
             self.studio.codex_manifests.exists(agent_id)
@@ -250,36 +1000,18 @@ class StudioAuthoringCoordinator:
         *,
         resolved_slug: str,
     ) -> AgentDraft:
-        upstream = (
-            (os.environ.get("OPENAI_BASE_URL") or os.environ.get("OPENAI_API_BASE") or "")
-            .strip()
-            .rstrip("/")
-        )
-        if upstream.endswith("/chat/completions"):
-            model_endpoint: dict[str, str] = {"endpoint_url": upstream}
-        else:
-            model_endpoint = {"base_url": upstream or "https://api.openai.com/v1"}
-        spec = AgentSpec(
-            runtime=RuntimeRef(type="codex", version=manifest.runtime.version),
-            instructions=Instructions(system=manifest.prompt),
-            model=ModelSpec(
-                model=manifest.model,
-                credential_ref="env://AGENTKIT_MODEL_API_KEY",
-                **model_endpoint,
-            ),
-        )
-        return cast(
-            AgentDraft,
-            self.studio.create_codex_agent(
-                agent_id=agent_id,
-                spec=spec,
-                name=display_name,
-                labels={
-                    "agentkit.ksyun.com/slug": self.backend.normalize_slug(resolved_slug),
-                    "agentkit.ksyun.com/source": "import",
-                },
-            ),
+        imported = manifest.model_copy(update={"name": agent_id}, deep=True)
+        snapshot = self.studio.codex_manifests.save(imported)
+        draft = self.studio.codex_agents._project(snapshot)
+        draft.metadata.name = display_name
+        draft.metadata.labels.update(
+            {
+                "agentkit.ksyun.com/slug": self.backend.normalize_slug(resolved_slug),
+                "agentkit.ksyun.com/source": "import",
+            }
         )
+        self.studio.codex_drafts.save(draft)
+        return cast(AgentDraft, self.studio.codex_agents._project(snapshot, current=draft))
 
 
 __all__ = ["StudioAuthoringCoordinator"]