diff --git a/apps/api/tests/fakes_projects.py b/apps/api/tests/fakes_projects.py
new file mode 100644
index 0000000..80b2662
--- /dev/null
+++ b/apps/api/tests/fakes_projects.py
@@ -0,0 +1,227 @@
+"""内存替身：项目/章节 repo + 写章网关（端点测试用，无 DB/无网络）。
+
+绝对导入 `from fakes_projects import ...`（包目录无 __init__.py，见 memory/gotchas）。
+全局唯一名（避免跨包同名碰撞）。
+"""
+
+from __future__ import annotations
+
+import uuid
+from collections.abc import AsyncIterator
+from typing import Any
+
+from pydantic import BaseModel
+from ww_core.domain.chapter_repo import ChapterDraftView, ChapterView
+from ww_core.domain.project_repo import ProjectCreate, ProjectView
+from ww_core.domain.repositories import DigestView
+from ww_core.domain.review_repo import ReviewView
+from ww_llm_gateway.types import Delta, LlmRequest, LlmResponse, ServedBy, Usage
+
+
+class FakeProjectRepo:
+    """实现 `ProjectRepo` Protocol 的内存版（按 owner_id 隔离）。"""
+
+    def __init__(self) -> None:
+        self.rows: dict[uuid.UUID, tuple[uuid.UUID, ProjectView]] = {}
+
+    async def create(self, owner_id: uuid.UUID, data: ProjectCreate) -> ProjectView:
+        pid = uuid.uuid4()
+        view = ProjectView(
+            id=pid,
+            title=data.title,
+            genre=data.genre,
+            logline=data.logline,
+            premise=data.premise,
+            theme=data.theme,
+            selling_points=list(data.selling_points),
+            structure=data.structure,
+        )
+        self.rows[pid] = (owner_id, view)
+        return view
+
+    async def list_for_owner(self, owner_id: uuid.UUID) -> list[ProjectView]:
+        return [v for (o, v) in self.rows.values() if o == owner_id]
+
+    async def get(self, owner_id: uuid.UUID, project_id: uuid.UUID) -> ProjectView | None:
+        entry = self.rows.get(project_id)
+        if entry is None or entry[0] != owner_id:
+            return None
+        return entry[1]
+
+
+class FakeChapterRepo:
+    """实现 `ChapterRepo` Protocol：幂等覆盖同 (project_id, chapter_no) 草稿。"""
+
+    def __init__(self) -> None:
+        self.drafts: dict[tuple[uuid.UUID, int], ChapterDraftView] = {}
+        self.accepted_versions: dict[tuple[uuid.UUID, int], int] = {}
+        self.accepted_content: dict[tuple[uuid.UUID, int, int], str] = {}
+
+    async def save_draft(
+        self, project_id: uuid.UUID, chapter_no: int, *, text: str, volume: int = 1
+    ) -> ChapterDraftView:
+        view = ChapterDraftView(
+            project_id=project_id,
+            chapter_no=chapter_no,
+            volume=volume,
+            content=text,
+            status="draft",
+            version=1,
+        )
+        self.drafts[(project_id, chapter_no)] = view
+        return view
+
+    async def get_draft(self, project_id: uuid.UUID, chapter_no: int) -> ChapterDraftView | None:
+        return self.drafts.get((project_id, chapter_no))
+
+    async def max_version(self, project_id: uuid.UUID, chapter_no: int) -> int:
+        versions = [
+            v for (p, c), v in self.accepted_versions.items() if p == project_id and c == chapter_no
+        ]
+        draft = 1 if (project_id, chapter_no) in self.drafts else 0
+        return max([*versions, draft], default=0)
+
+    async def promote_to_accepted(
+        self, project_id: uuid.UUID, chapter_no: int, *, content: str, volume: int = 1
+    ) -> ChapterView:
+        next_version = (await self.max_version(project_id, chapter_no)) + 1
+        self.accepted_versions[(project_id, chapter_no)] = next_version
+        self.accepted_content[(project_id, chapter_no, next_version)] = content
+        return ChapterView(
+            project_id=project_id,
+            chapter_no=chapter_no,
+            volume=volume,
+            content=content,
+            status="accepted",
+            version=next_version,
+        )
+
+    async def latest_accepted(self, project_id: uuid.UUID, chapter_no: int) -> ChapterView | None:
+        ver = self.accepted_versions.get((project_id, chapter_no))
+        if ver is None:
+            return None
+        return ChapterView(
+            project_id=project_id,
+            chapter_no=chapter_no,
+            volume=1,
+            content=self.accepted_content[(project_id, chapter_no, ver)],
+            status="accepted",
+            version=ver,
+        )
+
+
+class FakeReviewRepo:
+    """实现 `ReviewRepo` Protocol（内存留痕 + 历史 + 裁决；只 flush 语义无 DB）。"""
+
+    def __init__(self, *, fail_set_decisions: bool = False) -> None:
+        self.rows: list[ReviewView] = []
+        self._fail_set_decisions = fail_set_decisions
+
+    async def record(
+        self,
+        project_id: uuid.UUID,
+        chapter_no: int,
+        *,
+        chapter_version: int | None = None,
+        conflicts: list[dict[str, Any]],
+        foreshadow_sug: list[dict[str, Any]] | None = None,
+        style: dict[str, Any] | None = None,
+        pace: dict[str, Any] | None = None,
+        health_score: int | None = None,
+    ) -> ReviewView:
+        view = ReviewView(
+            id=uuid.uuid4(),
+            project_id=project_id,
+            chapter_no=chapter_no,
+            chapter_version=chapter_version,
+            conflicts=[dict(c) for c in conflicts],
+            foreshadow_sug=[dict(s) for s in (foreshadow_sug or [])],
+            style=style,
+            pace=pace,
+            health_score=health_score,
+        )
+        # 新→旧：插到列表头部。
+        self.rows.insert(0, view)
+        return view
+
+    async def list_for_chapter(self, project_id: uuid.UUID, chapter_no: int) -> list[ReviewView]:
+        return [r for r in self.rows if r.project_id == project_id and r.chapter_no == chapter_no]
+
+    async def set_decisions(self, review_id: uuid.UUID, *, decisions: dict[str, Any]) -> ReviewView:
+        if self._fail_set_decisions:
+            raise RuntimeError("boom: set_decisions failed (transaction rollback test)")
+        for i, r in enumerate(self.rows):
+            if r.id == review_id:
+                updated = r.model_copy(update={"decisions": dict(decisions)})
+                self.rows[i] = updated
+                return updated
+        raise LookupError(f"review not found: {review_id}")
+
+
+class FakeDigestAppendRepo:
+    """实现 `DigestAppendRepo` Protocol（内存 append-only）。"""
+
+    def __init__(self) -> None:
+        self.rows: list[tuple[uuid.UUID, int, dict[str, Any]]] = []
+
+    async def append(
+        self, project_id: uuid.UUID, chapter_no: int, *, facts: dict[str, Any]
+    ) -> DigestView:
+        self.rows.append((project_id, chapter_no, dict(facts)))
+        return DigestView(chapter_no=chapter_no, facts=dict(facts))
+
+
+class FakeSession:
+    """最小 fake：记录 commit 次数（验收事务边界断言）。"""
+
+    def __init__(self) -> None:
+        self.commits = 0
+
+    async def commit(self) -> None:
+        self.commits += 1
+
+
+class FakeReviewGateway:
+    """实现审稿/digest 节点最小依赖（`run`）：吐固定结构化 `parsed`，绝不联网。"""
+
+    def __init__(self, parsed: BaseModel | None = None, error: Exception | None = None) -> None:
+        self._parsed = parsed
+        self._error = error
+        self.requests: list[LlmRequest] = []
+
+    async def run(self, req: LlmRequest) -> LlmResponse:
+        self.requests.append(req)
+        if self._error is not None:
+            raise self._error
+        return LlmResponse(
+            text=self._parsed.model_dump_json() if self._parsed else "{}",
+            parsed=self._parsed,
+            usage=Usage(
+                provider="fake",
+                model="fake",
+                input_tokens=1,
+                output_tokens=1,
+                cost_minor=0,
+                currency="USD",
+            ),
+            served_by=ServedBy(provider="fake", model="fake"),
+        )
+
+
+class FakeWriterGateway:
+    """实现 write 节点最小依赖（`stream`）：吐固定 `Delta` 序列，绝不联网。
+
+    可注入 `error` 以验证 SSE 归一层把异常归一为 error 事件。
+    """
+
+    def __init__(self, chunks: list[str] | None = None, error: Exception | None = None) -> None:
+        self._chunks = chunks if chunks is not None else ["第一段。", "第二段。"]
+        self._error = error
+        self.requests: list[LlmRequest] = []
+
+    async def stream(self, req: LlmRequest) -> AsyncIterator[Delta]:
+        self.requests.append(req)
+        for c in self._chunks:
+            yield Delta(text=c)
+        if self._error is not None:
+            raise self._error
diff --git a/apps/api/tests/fakes_providers.py b/apps/api/tests/fakes_providers.py
new file mode 100644
index 0000000..a31d8b8
--- /dev/null
+++ b/apps/api/tests/fakes_providers.py
@@ -0,0 +1,56 @@
+"""内存替身：凭据存储 + 提供商探测（端点测试用，无 DB/无网络）。
+
+绝对导入 `from fakes import ...`（包目录无 __init__.py，见 memory/gotchas）。
+"""
+
+from __future__ import annotations
+
+import uuid
+
+from ww_api.services.credentials import StoredCredential, StoredRouting
+from ww_llm_gateway.adapters.base import Capabilities
+
+
+class FakeCredentialStore:
+    """实现 `CredentialStore` Protocol 的内存版。"""
+
+    def __init__(self) -> None:
+        # key: (owner_id, provider) → 密文
+        self.creds: dict[tuple[uuid.UUID, str], bytes] = {}
+        self.routing: dict[str, StoredRouting] = {}
+
+    async def list_credentials(self, owner_id: uuid.UUID) -> list[StoredCredential]:
+        return [
+            StoredCredential(provider=p, api_key_enc=blob)
+            for (o, p), blob in self.creds.items()
+            if o == owner_id
+        ]
+
+    async def list_routing(self) -> list[StoredRouting]:
+        return list(self.routing.values())
+
+    async def get_credential(self, owner_id: uuid.UUID, provider: str) -> StoredCredential | None:
+        blob = self.creds.get((owner_id, provider))
+        if blob is None:
+            return None
+        return StoredCredential(provider=provider, api_key_enc=blob)
+
+    async def upsert_credential(
+        self, owner_id: uuid.UUID, provider: str, api_key_enc: bytes
+    ) -> None:
+        self.creds[(owner_id, provider)] = api_key_enc
+
+    async def upsert_routing(self, routing: StoredRouting) -> None:
+        self.routing[routing.tier] = routing
+
+
+class FakeProviderProbe:
+    """实现 `ProviderProbe` Protocol：返回固定能力矩阵，绝不联网。"""
+
+    def __init__(self, caps: Capabilities | None = None) -> None:
+        self.caps = caps or Capabilities(structured_output=True, prefix_cache=True, thinking=False)
+        self.calls: list[tuple[uuid.UUID, str]] = []
+
+    async def probe(self, owner_id: uuid.UUID, provider: str) -> Capabilities:
+        self.calls.append((owner_id, provider))
+        return self.caps
diff --git a/apps/api/tests/test_credentials_crypto.py b/apps/api/tests/test_credentials_crypto.py
new file mode 100644
index 0000000..df1190f
--- /dev/null
+++ b/apps/api/tests/test_credentials_crypto.py
@@ -0,0 +1,59 @@
+"""T1.7 凭据加密工具单测——Fernet 往返 + 掩码（ARCH §4.7）。
+
+绝不在响应/日志回显明文 Key。
+"""
+
+from __future__ import annotations
+
+import pytest
+from cryptography.fernet import Fernet
+from ww_api.security.credentials import (
+    CredentialKeyError,
+    decrypt_api_key,
+    encrypt_api_key,
+    mask_api_key,
+)
+
+KEY = Fernet.generate_key().decode()
+
+
+def test_encrypt_decrypt_round_trip() -> None:
+    # Arrange
+    plaintext = "sk-abc123def456"
+    # Act
+    blob = encrypt_api_key(plaintext, key=KEY)
+    # Assert
+    assert isinstance(blob, bytes)
+    assert plaintext.encode() not in blob  # 密文不含明文
+    assert decrypt_api_key(blob, key=KEY) == plaintext
+
+
+def test_encrypt_is_non_deterministic() -> None:
+    # Fernet 含随机 IV——两次加密产物不同，但都能解回
+    a = encrypt_api_key("sk-secret", key=KEY)
+    b = encrypt_api_key("sk-secret", key=KEY)
+    assert a != b
+    assert decrypt_api_key(a, key=KEY) == decrypt_api_key(b, key=KEY) == "sk-secret"
+
+
+def test_missing_key_fails_fast() -> None:
+    with pytest.raises(CredentialKeyError):
+        encrypt_api_key("sk-x", key="")
+
+
+def test_invalid_key_fails_fast() -> None:
+    with pytest.raises(CredentialKeyError):
+        encrypt_api_key("sk-x", key="not-a-valid-fernet-key")
+
+
+def test_mask_shows_only_last_four() -> None:
+    assert mask_api_key("sk-abcdefgh1234") == "sk-…1234"
+
+
+def test_mask_short_key_fully_hidden() -> None:
+    # 极短 key 不泄露任何字符
+    assert mask_api_key("ab") == "sk-…••••"
+
+
+def test_mask_empty_key() -> None:
+    assert mask_api_key("") == "sk-…••••"
diff --git a/apps/api/tests/test_projects.py b/apps/api/tests/test_projects.py
new file mode 100644
index 0000000..089be46
--- /dev/null
+++ b/apps/api/tests/test_projects.py
@@ -0,0 +1,210 @@
+"""T1.4 端点测试：立项 CRUD + 写章 SSE + 自动保存（内存替身，无 DB/无网络）。
+
+DoD：端点入 OpenAPI；POST draft 回 SSE 流（mock 网关验证）；PUT draft 幂等。
+"""
+
+from __future__ import annotations
+
+import uuid
+
+import httpx
+import pytest
+from fakes_projects import FakeChapterRepo, FakeProjectRepo, FakeWriterGateway
+from ww_core.domain.repositories import (
+    CharacterView,
+    DigestView,
+    ForeshadowView,
+    MemoryRepos,
+    OutlineView,
+    RuleView,
+    StyleView,
+    WorldEntityView,
+)
+from ww_shared import ErrorCode
+
+
+class _EmptyOutlineRepo:
+    async def get(self, project_id: uuid.UUID, chapter_no: int) -> OutlineView | None:
+        return None
+
+
+class _EmptyCharacterRepo:
+    async def list_for_project(self, project_id: uuid.UUID) -> list[CharacterView]:
+        return []
+
+
+class _EmptyWorldEntityRepo:
+    async def list_for_project(self, project_id: uuid.UUID) -> list[WorldEntityView]:
+        return []
+
+
+class _EmptyDigestRepo:
+    async def recent(self, project_id: uuid.UUID, k: int) -> list[DigestView]:
+        return []
+
+
+class _EmptyForeshadowRepo:
+    async def list_for_codes(self, project_id: uuid.UUID, codes: list[str]) -> list[ForeshadowView]:
+        return []
+
+
+class _EmptyStyleRepo:
+    async def latest(self, project_id: uuid.UUID) -> StyleView | None:
+        return None
+
+
+class _EmptyRulesRepo:
+    async def all_for_project(self, project_id: uuid.UUID) -> list[RuleView]:
+        return []
+
+
+def _empty_memory_repos() -> MemoryRepos:
+    return MemoryRepos(
+        outline=_EmptyOutlineRepo(),
+        character=_EmptyCharacterRepo(),
+        world_entity=_EmptyWorldEntityRepo(),
+        digest=_EmptyDigestRepo(),
+        foreshadow=_EmptyForeshadowRepo(),
+        style=_EmptyStyleRepo(),
+        rules=_EmptyRulesRepo(),
+    )
+
+
+def _make_client(
+    *,
+    project_repo: FakeProjectRepo | None = None,
+    chapter_repo: FakeChapterRepo | None = None,
+    gateway: FakeWriterGateway | None = None,
+) -> tuple[httpx.AsyncClient, FakeProjectRepo, FakeChapterRepo, FakeWriterGateway]:
+    import os
+
+    os.environ.setdefault("CREDENTIAL_ENC_KEY", "x" * 44)
+    from ww_api.main import create_app
+    from ww_api.services.project_deps import (
+        get_chapter_repo,
+        get_memory_repos,
+        get_project_repo,
+        get_writer_gateway,
+    )
+
+    project_repo = project_repo or FakeProjectRepo()
+    chapter_repo = chapter_repo or FakeChapterRepo()
+    gateway = gateway or FakeWriterGateway()
+
+    app = create_app()
+    app.dependency_overrides[get_project_repo] = lambda: project_repo
+    app.dependency_overrides[get_chapter_repo] = lambda: chapter_repo
+    app.dependency_overrides[get_memory_repos] = _empty_memory_repos
+    app.dependency_overrides[get_writer_gateway] = lambda: gateway
+    transport = httpx.ASGITransport(app=app)
+    client = httpx.AsyncClient(transport=transport, base_url="http://test")
+    return client, project_repo, chapter_repo, gateway
+
+
+@pytest.mark.asyncio
+async def test_create_project_returns_201() -> None:
+    client, _, _, _ = _make_client()
+    async with client:
+        resp = await client.post(
+            "/projects",
+            json={"title": "我的网文", "genre": "玄幻", "selling_points": ["爽点密集"]},
+        )
+    assert resp.status_code == 201
+    body = resp.json()
+    assert body["title"] == "我的网文"
+    assert body["genre"] == "玄幻"
+    assert body["selling_points"] == ["爽点密集"]
+    assert uuid.UUID(body["id"])
+
+
+@pytest.mark.asyncio
+async def test_list_projects() -> None:
+    client, repo, _, _ = _make_client()
+    async with client:
+        await client.post("/projects", json={"title": "甲"})
+        await client.post("/projects", json={"title": "乙"})
+        resp = await client.get("/projects")
+    assert resp.status_code == 200
+    titles = {p["title"] for p in resp.json()["projects"]}
+    assert titles == {"甲", "乙"}
+
+
+@pytest.mark.asyncio
+async def test_get_project_detail() -> None:
+    client, _, _, _ = _make_client()
+    async with client:
+        created = (await client.post("/projects", json={"title": "详情"})).json()
+        resp = await client.get(f"/projects/{created['id']}")
+    assert resp.status_code == 200
+    assert resp.json()["title"] == "详情"
+
+
+@pytest.mark.asyncio
+async def test_get_unknown_project_404() -> None:
+    client, _, _, _ = _make_client()
+    async with client:
+        resp = await client.get(f"/projects/{uuid.uuid4()}")
+    assert resp.status_code == 404
+    assert resp.json()["error"]["code"] == ErrorCode.NOT_FOUND
+
+
+@pytest.mark.asyncio
+async def test_create_project_rejects_blank_title() -> None:
+    client, _, _, _ = _make_client()
+    async with client:
+        resp = await client.post("/projects", json={"title": ""})
+    assert resp.status_code == 422
+
+
+@pytest.mark.asyncio
+async def test_draft_stream_yields_sse_tokens_and_done() -> None:
+    gateway = FakeWriterGateway(chunks=["阿福", "走进门。"])
+    client, _, _, _ = _make_client(gateway=gateway)
+    pid = uuid.uuid4()
+    async with client:
+        resp = await client.post(f"/projects/{pid}/chapters/1/draft")
+    assert resp.status_code == 200
+    assert resp.headers["content-type"].startswith("text/event-stream")
+    text = resp.text
+    assert "event: token" in text
+    assert '"text": "阿福"' in text
+    assert "event: done" in text
+    # done 带累计长度（不回灌全文）
+    assert '"length": 6' in text
+    # 网关确实被调用，且只传 tier（不传具体 model）
+    assert len(gateway.requests) == 1
+    assert gateway.requests[0].tier == "writer"
+
+
+@pytest.mark.asyncio
+async def test_draft_stream_maps_error_to_sse_error_event() -> None:
+    from ww_shared import AppError
+
+    gateway = FakeWriterGateway(chunks=["半段"], error=AppError(ErrorCode.LLM_UNAVAILABLE, "boom"))
+    client, _, _, _ = _make_client(gateway=gateway)
+    pid = uuid.uuid4()
+    async with client:
+        resp = await client.post(f"/projects/{pid}/chapters/1/draft")
+    assert resp.status_code == 200
+    text = resp.text
+    assert "event: token" in text
+    assert "event: error" in text
+    assert ErrorCode.LLM_UNAVAILABLE in text
+
+
+@pytest.mark.asyncio
+async def test_put_draft_is_idempotent() -> None:
+    chapter_repo = FakeChapterRepo()
+    client, _, _, _ = _make_client(chapter_repo=chapter_repo)
+    pid = uuid.uuid4()
+    async with client:
+        r1 = await client.put(f"/projects/{pid}/chapters/3/draft", json={"text": "初稿"})
+        r2 = await client.put(f"/projects/{pid}/chapters/3/draft", json={"text": "改稿更长"})
+    assert r1.status_code == 200
+    assert r2.status_code == 200
+    # 同章节只一条草稿（版次不爆炸）
+    assert len(chapter_repo.drafts) == 1
+    saved = chapter_repo.drafts[(pid, 3)]
+    assert saved.content == "改稿更长"
+    assert saved.version == 1
+    assert r2.json()["length"] == len("改稿更长")
diff --git a/apps/api/tests/test_settings_providers.py b/apps/api/tests/test_settings_providers.py
new file mode 100644
index 0000000..1713416
--- /dev/null
+++ b/apps/api/tests/test_settings_providers.py
@@ -0,0 +1,93 @@
+"""T1.7 端点测试：凭据 upsert/列表/探测（内存替身，无 DB/无网络）。
+
+DoD：掩码往返、test-connection 经可注入探测回能力矩阵、明文绝不出响应。
+"""
+
+from __future__ import annotations
+
+from fakes_providers import FakeCredentialStore, FakeProviderProbe
+from fastapi.testclient import TestClient
+from ww_api.security.credentials import decrypt_api_key
+from ww_api.services.credentials import STUB_OWNER_ID
+
+
+def test_get_empty_providers(client: TestClient) -> None:
+    resp = client.get("/settings/providers")
+    assert resp.status_code == 200
+    body = resp.json()
+    assert body == {"providers": [], "tier_routing": []}
+
+
+def test_put_credential_encrypts_and_masks(
+    client: TestClient, store: FakeCredentialStore, enc_key: str
+) -> None:
+    # Act
+    resp = client.put(
+        "/settings/providers",
+        json={"credentials": [{"provider": "deepseek", "api_key": "sk-secret-9999"}]},
+    )
+    # Assert: 响应含掩码，绝不含明文
+    assert resp.status_code == 200
+    body = resp.json()
+    assert "sk-secret-9999" not in resp.text
+    assert body["providers"][0]["provider"] == "deepseek"
+    # 存的是密文且能解回明文（往返）
+    blob = store.creds[(STUB_OWNER_ID, "deepseek")]
+    assert b"sk-secret-9999" not in blob
+    assert decrypt_api_key(blob, key=enc_key) == "sk-secret-9999"
+
+
+def test_put_is_idempotent(client: TestClient, store: FakeCredentialStore) -> None:
+    payload = {"credentials": [{"provider": "deepseek", "api_key": "sk-a"}]}
+    client.put("/settings/providers", json=payload)
+    client.put(
+        "/settings/providers", json={"credentials": [{"provider": "deepseek", "api_key": "sk-b"}]}
+    )
+    # 同 (owner, provider) 只一条
+    assert len([k for k in store.creds if k[1] == "deepseek"]) == 1
+
+
+def test_put_tier_routing(client: TestClient, store: FakeCredentialStore) -> None:
+    resp = client.put(
+        "/settings/providers",
+        json={
+            "tier_routing": [
+                {
+                    "tier": "writer",
+                    "provider": "deepseek",
+                    "model": "deepseek-chat",
+                    "fallback": ["kimi"],
+                }
+            ]
+        },
+    )
+    assert resp.status_code == 200
+    routing = resp.json()["tier_routing"]
+    assert routing[0] == {
+        "tier": "writer",
+        "provider": "deepseek",
+        "model": "deepseek-chat",
+        "fallback": ["kimi"],
+    }
+
+
+def test_test_connection_returns_capabilities(client: TestClient, probe: FakeProviderProbe) -> None:
+    resp = client.post("/settings/providers/test", json={"provider": "deepseek"})
+    assert resp.status_code == 200
+    body = resp.json()
+    assert body["provider"] == "deepseek"
+    assert body["ok"] is True
+    assert body["capabilities"] == {
+        "structured_output": True,
+        "prefix_cache": True,
+        "thinking": False,
+    }
+    assert probe.calls == [(STUB_OWNER_ID, "deepseek")]
+
+
+def test_put_validation_rejects_blank_provider(client: TestClient) -> None:
+    resp = client.put(
+        "/settings/providers",
+        json={"credentials": [{"provider": "", "api_key": "sk-x"}]},
+    )
+    assert resp.status_code == 422
diff --git a/apps/api/ww_api/routers/projects.py b/apps/api/ww_api/routers/projects.py
new file mode 100644
index 0000000..9452a43
--- /dev/null
+++ b/apps/api/ww_api/routers/projects.py
@@ -0,0 +1,360 @@
+"""项目（立项）+ 章节草稿端点（C3 / ARCH §7.2, §7.3；不变量 #7）。
+
+- POST /projects                              立项向导 → projects 行（owner_id=stub）。
+- GET  /projects                              列出项目。
+- GET  /projects/:id                          项目详情（404 → NOT_FOUND）。
+- POST /projects/:id/chapters/:no/draft       流式写章草稿（SSE，text/event-stream）。
+- PUT  /projects/:id/chapters/:no/draft       自动保存草稿（幂等 upsert）。
+
+不变量：writer 只产草稿（不提升 accepted 版本/不抽 digest/不写状态——属 M2 验收）；
+agent 只传 tier；DB 按 project_id/owner_id 过滤；日志脱敏（只记长度，不记正文）。
+"""
+
+from __future__ import annotations
+
+import json
+import uuid
+from collections.abc import AsyncIterator
+from typing import Annotated
+
+from fastapi import APIRouter, Depends, Request
+from fastapi.responses import StreamingResponse
+from sqlalchemy.ext.asyncio import AsyncSession
+from ww_core.domain.chapter_repo import ChapterRepo
+from ww_core.domain.digest_repo import DigestAppendRepo
+from ww_core.domain.project_repo import ProjectCreate, ProjectRepo
+from ww_core.domain.repositories import MemoryRepos
+from ww_core.domain.review_repo import ReviewRepo
+from ww_core.memory import assemble
+from ww_core.orchestrator import (
+    ChapterState,
+    SseEvent,
+    build_review_context,
+    build_review_graph,
+    normalize_deltas,
+    normalize_review,
+    stream_chapter_draft,
+)
+from ww_db import get_session
+from ww_llm_gateway import Gateway
+from ww_shared import AppError, ErrorCode
+
+from ww_api.logging_config import get_logger
+from ww_api.schemas.projects import (
+    AcceptRequest,
+    AcceptResponse,
+    DraftResponse,
+    DraftSaveRequest,
+    ProjectCreateRequest,
+    ProjectListResponse,
+    ProjectResponse,
+    ReviewHistoryItem,
+    ReviewHistoryResponse,
+    ReviewRequest,
+)
+from ww_api.services.accept_service import (
+    AcceptOutcome,
+    assert_conflicts_resolved,
+    run_accept_transaction,
+)
+from ww_api.services.credentials import STUB_OWNER_ID
+from ww_api.services.digest_extraction import extract_digest_facts
+from ww_api.services.project_deps import (
+    get_chapter_repo,
+    get_digest_append_repo,
+    get_digest_gateway,
+    get_memory_repos,
+    get_project_repo,
+    get_review_gateway,
+    get_review_repo,
+    get_writer_gateway,
+)
+
+log = get_logger("ww.api.projects")
+
+router = APIRouter(prefix="/projects", tags=["projects"])
+
+ProjectRepoDep = Annotated[ProjectRepo, Depends(get_project_repo)]
+ChapterRepoDep = Annotated[ChapterRepo, Depends(get_chapter_repo)]
+GatewayDep = Annotated[Gateway, Depends(get_writer_gateway)]
+ReviewGatewayDep = Annotated[Gateway, Depends(get_review_gateway)]
+DigestGatewayDep = Annotated[Gateway, Depends(get_digest_gateway)]
+MemoryReposDep = Annotated[MemoryRepos, Depends(get_memory_repos)]
+ReviewRepoDep = Annotated[ReviewRepo, Depends(get_review_repo)]
+DigestRepoDep = Annotated[DigestAppendRepo, Depends(get_digest_append_repo)]
+
+
+def _to_response(view: object) -> ProjectResponse:
+    # ProjectView 与 ProjectResponse 字段同名，逐字段映射（snake_case 契约）。
+    return ProjectResponse.model_validate(view, from_attributes=True)
+
+
+@router.post("", status_code=201)
+async def create_project(body: ProjectCreateRequest, repo: ProjectRepoDep) -> ProjectResponse:
+    view = await repo.create(
+        STUB_OWNER_ID,
+        ProjectCreate(
+            title=body.title,
+            genre=body.genre,
+            logline=body.logline,
+            premise=body.premise,
+            theme=body.theme,
+            selling_points=body.selling_points,
+            structure=body.structure,
+        ),
+    )
+    log.info("project_created", project_id=str(view.id), title_len=len(view.title))
+    return _to_response(view)
+
+
+@router.get("")
+async def list_projects(repo: ProjectRepoDep) -> ProjectListResponse:
+    views = await repo.list_for_owner(STUB_OWNER_ID)
+    return ProjectListResponse(projects=[_to_response(v) for v in views])
+
+
+@router.get("/{project_id}")
+async def get_project(project_id: uuid.UUID, repo: ProjectRepoDep) -> ProjectResponse:
+    view = await repo.get(STUB_OWNER_ID, project_id)
+    if view is None:
+        raise AppError(ErrorCode.NOT_FOUND, f"project {project_id} not found")
+    return _to_response(view)
+
+
+def _encode_sse(event: SseEvent) -> str:
+    """把归一事件编码为 text/event-stream 帧：`event: <name>\\ndata: <json>\\n\\n`。"""
+    payload = json.dumps(event.data, ensure_ascii=False)
+    return f"event: {event.event}\ndata: {payload}\n\n"
+
+
+@router.post("/{project_id}/chapters/{chapter_no}/draft")
+async def stream_draft(
+    project_id: uuid.UUID,
+    chapter_no: int,
+    request: Request,
+    repos: MemoryReposDep,
+    gateway: GatewayDep,
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> StreamingResponse:
+    """流式写章草稿：组装记忆 → 网关流 → 归一为 SSE 事件 → text/event-stream。"""
+    request_id = getattr(request.state, "request_id", None)
+    context = await assemble(repos, project_id, chapter_no)
+    log.info(
+        "draft_stream_start",
+        project_id=str(project_id),
+        chapter_no=chapter_no,
+        request_id=request_id,
+        stable_core_len=len(context.stable_core),
+        volatile_len=len(context.volatile),
+    )
+
+    deltas = stream_chapter_draft(
+        gateway,
+        stable_core=context.stable_core,
+        volatile=context.volatile,
+        user_id=STUB_OWNER_ID,
+        project_id=project_id,
+    )
+
+    async def _frames() -> AsyncIterator[str]:
+        async for event in normalize_deltas(deltas, request_id=request_id):
+            yield _encode_sse(event)
+        # 网关在流末经 SqlAlchemyLedgerSink.record 把 usage_ledger 行 flush 进本请求 session；
+        # sink 按设计不提交（写库事务由编排层控制，见不变量）。draft 端点无其他写副作用，
+        # 故流耗尽后在此提交，确保「每次调用一条 usage_ledger」真正落库（T1.9 暴露）。
+        await session.commit()
+
+    return StreamingResponse(
+        _frames(),
+        media_type="text/event-stream",
+        headers={"cache-control": "no-cache", "x-accel-buffering": "no"},
+    )
+
+
+@router.put("/{project_id}/chapters/{chapter_no}/draft")
+async def save_draft(
+    project_id: uuid.UUID,
+    chapter_no: int,
+    body: DraftSaveRequest,
+    repo: ChapterRepoDep,
+) -> DraftResponse:
+    """自动保存：幂等 upsert 草稿（同章节覆盖同一行，版次不爆炸）。"""
+    view = await repo.save_draft(project_id, chapter_no, text=body.text)
+    log.info(
+        "draft_saved",
+        project_id=str(project_id),
+        chapter_no=chapter_no,
+        length=len(view.content),
+    )
+    return DraftResponse(
+        project_id=view.project_id,
+        chapter_no=view.chapter_no,
+        volume=view.volume,
+        status=view.status,
+        version=view.version,
+        length=len(view.content),
+    )
+
+
+async def _resolve_review_draft(
+    body: ReviewRequest,
+    chapter_repo: ChapterRepo,
+    project_id: uuid.UUID,
+    chapter_no: int,
+) -> str:
+    """取待审正文：请求体 `draft` 优先；否则回退到已保存草稿；都无 → NOT_FOUND。"""
+    if body.draft is not None and body.draft.strip():
+        return body.draft
+    saved = await chapter_repo.get_draft(project_id, chapter_no)
+    if saved is None or not saved.content.strip():
+        raise AppError(
+            ErrorCode.NOT_FOUND,
+            f"chapter {chapter_no} has no draft to review; provide `draft` in body",
+        )
+    return saved.content
+
+
+@router.post("/{project_id}/chapters/{chapter_no}/review")
+async def review_chapter(
+    project_id: uuid.UUID,
+    chapter_no: int,
+    body: ReviewRequest,
+    request: Request,
+    repos: MemoryReposDep,
+    chapter_repo: ChapterRepoDep,
+    review_repo: ReviewRepoDep,
+    gateway: ReviewGatewayDep,
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> StreamingResponse:
+    """续审（SSE）：组审稿上下文 → 跑审稿子图 → 归一为 section/conflict/done 事件。
+
+    提交边界：网关 ledger + collect 经 review_repo.record 均只 flush；端点在**流耗尽后**
+    `await session.commit()`（镜像 draft 端点，否则记账/留痕静默丢失）。
+    """
+    request_id = getattr(request.state, "request_id", None)
+    draft = await _resolve_review_draft(body, chapter_repo, project_id, chapter_no)
+    context = await assemble(repos, project_id, chapter_no)
+    review_context = build_review_context(
+        draft=draft, stable_core=context.stable_core, volatile=context.volatile
+    )
+    log.info(
+        "review_stream_start",
+        project_id=str(project_id),
+        chapter_no=chapter_no,
+        request_id=request_id,
+        draft_len=len(draft),
+        review_context_len=len(review_context),
+    )
+
+    graph = build_review_graph(gateway, review_repo)
+    initial: ChapterState = {
+        "project_id": project_id,
+        "chapter_no": chapter_no,
+        "user_id": STUB_OWNER_ID,
+        "review_context": review_context,
+    }
+
+    async def _frames() -> AsyncIterator[str]:
+        final = await graph.ainvoke(initial)
+        reviews = final.get("reviews") or {}
+        async for event in normalize_review(reviews, request_id=request_id):
+            yield _encode_sse(event)
+        # collect 经 review_repo.record 落 chapter_reviews（只 flush）+ 网关 ledger 只 flush
+        # → 流耗尽后在此提交，确保审稿留痕 + usage_ledger 真正落库（同 draft 端点）。
+        await session.commit()
+
+    return StreamingResponse(
+        _frames(),
+        media_type="text/event-stream",
+        headers={"cache-control": "no-cache", "x-accel-buffering": "no"},
+    )
+
+
+@router.get("/{project_id}/chapters/{chapter_no}/reviews")
+async def list_reviews(
+    project_id: uuid.UUID,
+    chapter_no: int,
+    review_repo: ReviewRepoDep,
+) -> ReviewHistoryResponse:
+    """审稿历史（新→旧）：供前端审稿页加载既往审稿留痕 + 裁决。"""
+    views = await review_repo.list_for_chapter(project_id, chapter_no)
+    items = [
+        ReviewHistoryItem(
+            id=v.id,
+            project_id=v.project_id,
+            chapter_no=v.chapter_no,
+            chapter_version=v.chapter_version,
+            conflicts=v.conflicts,
+            foreshadow_sug=v.foreshadow_sug,
+            style=v.style,
+            pace=v.pace,
+            health_score=v.health_score,
+            decisions=v.decisions,
+        )
+        for v in views
+    ]
+    return ReviewHistoryResponse(reviews=items)
+
+
+@router.post("/{project_id}/chapters/{chapter_no}/accept")
+async def accept_chapter(
+    project_id: uuid.UUID,
+    chapter_no: int,
+    body: AcceptRequest,
+    request: Request,
+    chapter_repo: ChapterRepoDep,
+    digest_repo: DigestRepoDep,
+    review_repo: ReviewRepoDep,
+    gateway: DigestGatewayDep,
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> AcceptResponse:
+    """验收事务 + 冲突 gate（§5.5）：gate（事务前）→ 终稿提炼 digest（事务外，R2）→
+    单原子事务（晋升 + digest + 裁决留痕）→ 一次 commit。
+    """
+    request_id = getattr(request.state, "request_id", None)
+    # R3：审稿真相从领域表重读（最近一条 chapter_reviews），不依赖 checkpoint。
+    history = await review_repo.list_for_chapter(project_id, chapter_no)
+    latest_review = history[0] if history else None
+
+    # 冲突 gate（R5，事务前拦截，不写库）。
+    assert_conflicts_resolved(latest_review, body.decisions)
+
+    log.info(
+        "accept_start",
+        project_id=str(project_id),
+        chapter_no=chapter_no,
+        request_id=request_id,
+        final_text_len=len(body.final_text),
+        decision_count=len(body.decisions),
+        has_review=latest_review is not None,
+    )
+
+    # R2：终稿 digest 提炼在**事务外**做（别在持开事务里跨网络调 LLM）。
+    digest_facts = await extract_digest_facts(
+        gateway,
+        final_text=body.final_text,
+        user_id=STUB_OWNER_ID,
+        project_id=project_id,
+        chapter_no=chapter_no,
+    )
+
+    outcome: AcceptOutcome = await run_accept_transaction(
+        session=session,
+        chapter_repo=chapter_repo,
+        digest_repo=digest_repo,
+        review_repo=review_repo,
+        project_id=project_id,
+        chapter_no=chapter_no,
+        final_text=body.final_text,
+        digest_facts=digest_facts,
+        latest_review=latest_review,
+        decisions=body.decisions,
+    )
+    return AcceptResponse(
+        project_id=project_id,
+        chapter_no=chapter_no,
+        accepted_version=outcome.accepted_version,
+        digest_added=outcome.digest_added,
+        decisions_recorded=outcome.decisions_recorded,
+        review_id=outcome.review_id,
+    )
diff --git a/apps/api/ww_api/routers/settings_providers.py b/apps/api/ww_api/routers/settings_providers.py
new file mode 100644
index 0000000..13f3831
--- /dev/null
+++ b/apps/api/ww_api/routers/settings_providers.py
@@ -0,0 +1,127 @@
+"""提供商凭据与档位路由端点（C3 / ARCH §4.7, §7.2；UX §6.10）。
+
+- GET  /settings/providers       列出已配置提供商（掩码）+ 档位路由。
+- PUT  /settings/providers       幂等 upsert 凭据/档位路由，回掩码视图。
+- POST /settings/providers/test  最小探测验 Key + 拉能力矩阵。
+
+不变量：响应/日志绝不含明文 Key（加密入库、仅掩码出站）。
+"""
+
+from __future__ import annotations
+
+from typing import Annotated
+
+from fastapi import APIRouter, Depends
+from ww_config import get_settings
+from ww_llm_gateway.adapters.base import Capabilities
+
+from ww_api.logging_config import get_logger
+from ww_api.schemas.providers import (
+    CapabilitiesView,
+    ProvidersResponse,
+    ProvidersUpsertRequest,
+    ProviderView,
+    TestConnectionRequest,
+    TestConnectionResponse,
+    TierRoutingView,
+)
+from ww_api.security.credentials import encrypt_api_key, mask_api_key
+from ww_api.services.credentials import (
+    STUB_OWNER_ID,
+    CredentialStore,
+    ProviderProbe,
+    StoredCredential,
+    StoredRouting,
+)
+from ww_api.services.provider_deps import (
+    get_credential_store,
+    get_provider_probe,
+)
+
+log = get_logger("ww.api.providers")
+
+router = APIRouter(prefix="/settings/providers", tags=["settings"])
+
+StoreDep = Annotated[CredentialStore, Depends(get_credential_store)]
+ProbeDep = Annotated[ProviderProbe, Depends(get_provider_probe)]
+
+
+def _mask_credential(cred: StoredCredential, plaintext: str | None) -> ProviderView:
+    # 存储层只有密文；掩码需明文末四位——若没有则全隐占位。
+    masked = mask_api_key(plaintext) if plaintext is not None else mask_api_key("")
+    return ProviderView(provider=cred.provider, masked_key=masked)
+
+
+def _routing_view(r: StoredRouting) -> TierRoutingView:
+    return TierRoutingView(tier=r.tier, provider=r.provider, model=r.model, fallback=r.fallback)
+
+
+async def _build_response(store: CredentialStore) -> ProvidersResponse:
+    creds = await store.list_credentials(STUB_OWNER_ID)
+    routing = await store.list_routing()
+    # 列表视图无明文，掩码占位（不解密历史密文以免无谓 IO/泄露面）。
+    providers = [ProviderView(provider=c.provider, masked_key=mask_api_key("")) for c in creds]
+    return ProvidersResponse(
+        providers=providers,
+        tier_routing=[_routing_view(r) for r in routing],
+    )
+
+
+@router.get("")
+async def list_providers(store: StoreDep) -> ProvidersResponse:
+    return await _build_response(store)
+
+
+@router.put("")
+async def upsert_providers(
+    body: ProvidersUpsertRequest,
+    store: StoreDep,
+) -> ProvidersResponse:
+    enc_key = get_settings().credential_enc_key
+    for cred in body.credentials:
+        api_key_enc = encrypt_api_key(cred.api_key, key=enc_key)
+        await store.upsert_credential(STUB_OWNER_ID, cred.provider, api_key_enc)
+        # 仅记 provider + 末四位掩码，绝不记明文
+        log.info(
+            "provider_credential_upserted",
+            provider=cred.provider,
+            masked_key=mask_api_key(cred.api_key),
+        )
+    for routing in body.tier_routing:
+        await store.upsert_routing(
+            StoredRouting(
+                tier=routing.tier,
+                provider=routing.provider,
+                model=routing.model,
+                fallback=routing.fallback,
+            )
+        )
+        log.info(
+            "tier_routing_upserted",
+            tier=routing.tier,
+            provider=routing.provider,
+            model=routing.model,
+        )
+    return await _build_response(store)
+
+
+def _capabilities_view(caps: Capabilities) -> CapabilitiesView:
+    return CapabilitiesView(
+        structured_output=caps.structured_output,
+        prefix_cache=caps.prefix_cache,
+        thinking=caps.thinking,
+    )
+
+
+@router.post("/test")
+async def test_connection(
+    body: TestConnectionRequest,
+    probe: ProbeDep,
+) -> TestConnectionResponse:
+    caps = await probe.probe(STUB_OWNER_ID, body.provider)
+    log.info("provider_probe_ok", provider=body.provider)
+    return TestConnectionResponse(
+        provider=body.provider,
+        ok=True,
+        capabilities=_capabilities_view(caps),
+    )
diff --git a/apps/api/ww_api/schemas/projects.py b/apps/api/ww_api/schemas/projects.py
new file mode 100644
index 0000000..8307c2e
--- /dev/null
+++ b/apps/api/ww_api/schemas/projects.py
@@ -0,0 +1,126 @@
+"""项目（立项）与章节草稿的请求/响应 schema（C3 / ARCH §7.2）。
+
+snake_case；前端经 OpenAPI 生成 TS 类型消费。改字段 → 前端必须 `pnpm gen:api`。
+"""
+
+from __future__ import annotations
+
+import uuid
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+
+class ProjectCreateRequest(BaseModel):
+    """POST /projects：立项向导字段（owner_id 由后端补 stub，不入参）。"""
+
+    title: str = Field(min_length=1)
+    genre: str | None = None
+    logline: str | None = None
+    premise: str | None = None
+    theme: str | None = None
+    selling_points: list[Any] = Field(default_factory=list)
+    structure: str | None = None
+
+
+class ProjectResponse(BaseModel):
+    """项目视图（创建/列表/详情共用）。"""
+
+    id: uuid.UUID
+    title: str
+    genre: str | None = None
+    logline: str | None = None
+    premise: str | None = None
+    theme: str | None = None
+    selling_points: list[Any] = Field(default_factory=list)
+    structure: str | None = None
+
+
+class ProjectListResponse(BaseModel):
+    """GET /projects：项目列表。"""
+
+    projects: list[ProjectResponse] = Field(default_factory=list)
+
+
+class DraftSaveRequest(BaseModel):
+    """PUT /projects/:id/chapters/:no/draft：自动保存草稿正文。"""
+
+    text: str
+
+
+class DraftResponse(BaseModel):
+    """草稿保存结果（脱敏：只回元信息 + 长度，不回灌正文以外的衍生）。"""
+
+    project_id: uuid.UUID
+    chapter_no: int
+    volume: int
+    status: str
+    version: int
+    length: int
+
+
+# ---- 审稿（T2.5）----
+
+
+class ReviewRequest(BaseModel):
+    """POST /projects/:id/chapters/:no/review：可选携带待审草稿正文。
+
+    不传 `draft` 时端点回退到已保存的草稿（chapter_repo）。
+    """
+
+    draft: str | None = None
+
+
+class ReviewHistoryItem(BaseModel):
+    """单条审稿留痕（GET .../reviews 历史项；snake_case）。"""
+
+    id: uuid.UUID
+    project_id: uuid.UUID
+    chapter_no: int
+    chapter_version: int | None = None
+    conflicts: list[dict[str, Any]] = Field(default_factory=list)
+    foreshadow_sug: list[dict[str, Any]] = Field(default_factory=list)
+    style: dict[str, Any] | None = None
+    pace: dict[str, Any] | None = None
+    health_score: int | None = None
+    decisions: dict[str, Any] | None = None
+
+
+class ReviewHistoryResponse(BaseModel):
+    """GET /projects/:id/chapters/:no/reviews：审稿历史（新→旧）。"""
+
+    reviews: list[ReviewHistoryItem] = Field(default_factory=list)
+
+
+# ---- 验收（T2.4）----
+
+# 每个冲突的裁决：采纳改法 / 忽略 / 手改（R5）。
+Verdict = Literal["accept", "ignore", "manual"]
+
+
+class ConflictDecision(BaseModel):
+    """对最近一次审稿留痕里**某个冲突**（按其在 conflicts 列表的下标定位）的裁决。"""
+
+    conflict_index: int = Field(ge=0, description="冲突在最近审稿 conflicts 列表中的下标")
+    verdict: Verdict = Field(description="采纳改法 / 忽略 / 手改")
+    note: str | None = Field(default=None, description="可选裁决备注（如手改说明）")
+
+
+class AcceptRequest(BaseModel):
+    """POST /projects/:id/chapters/:no/accept：裁决清单 + 可能改过的终稿。"""
+
+    final_text: str = Field(min_length=1, description="作者裁决/改稿后的最终验收文本")
+    decisions: list[ConflictDecision] = Field(
+        default_factory=list, description="对最近审稿每个冲突的裁决（每冲突必有其一，R5）"
+    )
+
+
+class AcceptResponse(BaseModel):
+    """验收「本次将更新」清单（ARCH §7.2 写回结果）。"""
+
+    project_id: uuid.UUID
+    chapter_no: int
+    accepted_version: int = Field(description="晋升到的 accepted 版次（max+1）")
+    digest_added: bool = Field(description="是否新增了一行 chapter_digests")
+    decisions_recorded: int = Field(description="本次写回的裁决条数")
+    review_id: uuid.UUID | None = Field(default=None, description="写回裁决的审稿留痕行 id")
diff --git a/apps/api/ww_api/schemas/providers.py b/apps/api/ww_api/schemas/providers.py
new file mode 100644
index 0000000..3ea8366
--- /dev/null
+++ b/apps/api/ww_api/schemas/providers.py
@@ -0,0 +1,76 @@
+"""提供商凭据/档位路由的请求/响应 schema（C3 / ARCH §4.7, §7.2）。
+
+snake_case；响应一律 **掩码 Key**，绝不含明文。前端经 OpenAPI 生成 TS 类型消费。
+"""
+
+from __future__ import annotations
+
+from pydantic import BaseModel, Field
+
+
+class ProviderView(BaseModel):
+    """已配置提供商（掩码视图）。"""
+
+    provider: str
+    masked_key: str  # 形如 sk-…1234；明文永不出现
+
+
+class TierRoutingView(BaseModel):
+    """档位 → provider:model 路由（含回退链）。"""
+
+    tier: str
+    provider: str
+    model: str
+    fallback: list[str] = Field(default_factory=list)
+
+
+class ProvidersResponse(BaseModel):
+    """GET/PUT 响应：已配置提供商（掩码）+ 当前档位路由。"""
+
+    providers: list[ProviderView] = Field(default_factory=list)
+    tier_routing: list[TierRoutingView] = Field(default_factory=list)
+
+
+class ProviderCredentialInput(BaseModel):
+    """单条提供商凭据写入。"""
+
+    provider: str = Field(min_length=1)
+    api_key: str = Field(min_length=1)  # 明文入站，加密入库，绝不回显
+
+
+class TierRoutingInput(BaseModel):
+    """单条档位路由写入。"""
+
+    tier: str = Field(min_length=1)
+    provider: str = Field(min_length=1)
+    model: str = Field(min_length=1)
+    fallback: list[str] = Field(default_factory=list)
+
+
+class ProvidersUpsertRequest(BaseModel):
+    """PUT 请求：可同时 upsert 若干凭据与档位路由（幂等）。"""
+
+    credentials: list[ProviderCredentialInput] = Field(default_factory=list)
+    tier_routing: list[TierRoutingInput] = Field(default_factory=list)
+
+
+class TestConnectionRequest(BaseModel):
+    """POST /test：最小探测请求。"""
+
+    provider: str = Field(min_length=1)
+
+
+class CapabilitiesView(BaseModel):
+    """探测得到的能力矩阵（镜像网关 `Capabilities`）。"""
+
+    structured_output: bool = False
+    prefix_cache: bool = False
+    thinking: bool = False
+
+
+class TestConnectionResponse(BaseModel):
+    """POST /test 响应：连通性 + 能力矩阵。"""
+
+    provider: str
+    ok: bool
+    capabilities: CapabilitiesView
diff --git a/apps/api/ww_api/security/credentials.py b/apps/api/ww_api/security/credentials.py
new file mode 100644
index 0000000..df06a05
--- /dev/null
+++ b/apps/api/ww_api/security/credentials.py
@@ -0,0 +1,53 @@
+"""提供商 API Key 的对称加解密与掩码（ARCH §4.7）。
+
+- 加密用 `cryptography` 的 Fernet，密钥取 `settings.credential_enc_key`。
+- 缺失/非法密钥 → 立即失败（`CredentialKeyError`），绝不静默。
+- `mask_api_key` 产出可安全展示/落库回显的掩码；明文永不出响应、永不进日志。
+"""
+
+from __future__ import annotations
+
+from cryptography.fernet import Fernet, InvalidToken
+
+# 掩码常量：前缀 + 省略号 + 末四位。短 key 用 bullet 占位，绝不泄露字符。
+_MASK_PREFIX = "sk-…"
+_MASK_PLACEHOLDER = "••••"
+_LAST_N = 4
+
+
+class CredentialKeyError(RuntimeError):
+    """加密密钥缺失或非法——快速失败，调用方负责映射为 5xx/配置错误。"""
+
+
+def _fernet(key: str) -> Fernet:
+    if not key:
+        raise CredentialKeyError(
+            "credential_enc_key 未配置；无法加解密提供商凭据。请设置环境变量。"
+        )
+    try:
+        return Fernet(key.encode())
+    except (ValueError, TypeError) as exc:
+        # 不回显 key 内容
+        raise CredentialKeyError(
+            "credential_enc_key 非法（需为 32 字节 url-safe base64 Fernet key）。"
+        ) from exc
+
+
+def encrypt_api_key(plaintext: str, *, key: str) -> bytes:
+    """加密明文 API Key，返回可入库的密文字节（`api_key_enc`）。"""
+    return _fernet(key).encrypt(plaintext.encode())
+
+
+def decrypt_api_key(blob: bytes, *, key: str) -> str:
+    """解密密文字节回明文。仅供探测/调用使用，绝不进响应或日志。"""
+    try:
+        return _fernet(key).decrypt(blob).decode()
+    except InvalidToken as exc:
+        raise CredentialKeyError("凭据密文无法解密（密钥不匹配或数据损坏）。") from exc
+
+
+def mask_api_key(plaintext: str) -> str:
+    """掩码：仅露末四位，如 `sk-…1234`；过短/空则全隐。"""
+    if len(plaintext) < _LAST_N:
+        return f"{_MASK_PREFIX}{_MASK_PLACEHOLDER}"
+    return f"{_MASK_PREFIX}{plaintext[-_LAST_N:]}"
diff --git a/apps/api/ww_api/services/credentials.py b/apps/api/ww_api/services/credentials.py
new file mode 100644
index 0000000..dff9fda
--- /dev/null
+++ b/apps/api/ww_api/services/credentials.py
@@ -0,0 +1,149 @@
+"""凭据存储与提供商探测的接口 + SQLAlchemy/网关实现（ARCH §4.7）。
+
+路由依赖这里的 **接口**（Protocol），测试注入内存替身；运行时用 SQLAlchemy/网关实现。
+单用户原型：`owner_id` 用固定 stub（见 `STUB_OWNER_ID`），多租户化时改为按认证主体取。
+"""
+
+from __future__ import annotations
+
+import uuid
+from dataclasses import dataclass
+from typing import Protocol
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+from ww_db.models import ProviderCredential, TierRouting
+from ww_llm_gateway.adapters.base import Capabilities
+
+# 单用户 stub owner（与网关 Scope.user_id 的 stub 约定一致：UUID(int=1)）。
+# 多租户化时此常量由认证主体替换（见 ARCH §4.7 隔离）。
+STUB_OWNER_ID = uuid.UUID(int=1)
+
+
+@dataclass(frozen=True)
+class StoredCredential:
+    """存储层视图：含密文，绝不出 API 边界（路由仅取 provider 并掩码）。"""
+
+    provider: str
+    api_key_enc: bytes
+
+
+@dataclass(frozen=True)
+class StoredRouting:
+    tier: str
+    provider: str
+    model: str
+    fallback: list[str]
+
+
+class CredentialStore(Protocol):
+    """凭据 + 档位路由的读写接口（按 owner_id 隔离）。"""
+
+    async def list_credentials(self, owner_id: uuid.UUID) -> list[StoredCredential]: ...
+
+    async def list_routing(self) -> list[StoredRouting]: ...
+
+    async def get_credential(
+        self, owner_id: uuid.UUID, provider: str
+    ) -> StoredCredential | None: ...
+
+    async def upsert_credential(
+        self, owner_id: uuid.UUID, provider: str, api_key_enc: bytes
+    ) -> None: ...
+
+    async def upsert_routing(self, routing: StoredRouting) -> None: ...
+
+
+class ProviderProbe(Protocol):
+    """最小连通探测：验证 Key + 返回能力矩阵。测试注入假探测，绝不联网。"""
+
+    async def probe(self, owner_id: uuid.UUID, provider: str) -> Capabilities: ...
+
+
+class SqlCredentialStore:
+    """SQLAlchemy 实现：写 `provider_credentials` / `tier_routing`，幂等 upsert。"""
+
+    def __init__(self, session: AsyncSession) -> None:
+        self._session = session
+
+    async def list_credentials(self, owner_id: uuid.UUID) -> list[StoredCredential]:
+        rows = (
+            await self._session.execute(
+                select(ProviderCredential).where(ProviderCredential.owner_id == owner_id)
+            )
+        ).scalars()
+        return [StoredCredential(provider=r.provider, api_key_enc=r.api_key_enc) for r in rows]
+
+    async def list_routing(self) -> list[StoredRouting]:
+        rows = (await self._session.execute(select(TierRouting))).scalars()
+        return [
+            StoredRouting(
+                tier=r.tier, provider=r.provider, model=r.model, fallback=list(r.fallback)
+            )
+            for r in rows
+        ]
+
+    async def get_credential(self, owner_id: uuid.UUID, provider: str) -> StoredCredential | None:
+        row = (
+            await self._session.execute(
+                select(ProviderCredential).where(
+                    ProviderCredential.owner_id == owner_id,
+                    ProviderCredential.provider == provider,
+                )
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            return None
+        return StoredCredential(provider=row.provider, api_key_enc=row.api_key_enc)
+
+    async def upsert_credential(
+        self, owner_id: uuid.UUID, provider: str, api_key_enc: bytes
+    ) -> None:
+        # 显式 read-modify-write：唯一约束含可空 project_id，PG ON CONFLICT
+        # 在 NULL 上不去重（NULLS DISTINCT），故不用 on_conflict。
+        existing = (
+            await self._session.execute(
+                select(ProviderCredential).where(
+                    ProviderCredential.owner_id == owner_id,
+                    ProviderCredential.project_id.is_(None),
+                    ProviderCredential.provider == provider,
+                )
+            )
+        ).scalar_one_or_none()
+        if existing is None:
+            self._session.add(
+                ProviderCredential(
+                    owner_id=owner_id,
+                    project_id=None,
+                    provider=provider,
+                    api_key_enc=api_key_enc,
+                )
+            )
+        else:
+            existing.api_key_enc = api_key_enc
+        await self._session.commit()
+
+    async def upsert_routing(self, routing: StoredRouting) -> None:
+        existing = (
+            await self._session.execute(
+                select(TierRouting).where(
+                    TierRouting.project_id.is_(None),
+                    TierRouting.tier == routing.tier,
+                )
+            )
+        ).scalar_one_or_none()
+        if existing is None:
+            self._session.add(
+                TierRouting(
+                    project_id=None,
+                    tier=routing.tier,
+                    provider=routing.provider,
+                    model=routing.model,
+                    fallback=routing.fallback,
+                )
+            )
+        else:
+            existing.provider = routing.provider
+            existing.model = routing.model
+            existing.fallback = routing.fallback
+        await self._session.commit()
diff --git a/apps/api/ww_api/services/project_deps.py b/apps/api/ww_api/services/project_deps.py
new file mode 100644
index 0000000..e4bc8be
--- /dev/null
+++ b/apps/api/ww_api/services/project_deps.py
@@ -0,0 +1,162 @@
+"""项目/章节端点的依赖装配（运行时实现）。
+
+- `get_project_repo` / `get_chapter_repo`：把请求 session 装配成 SQLAlchemy repo。
+- `get_writer_gateway`：据 writer 档位路由从已存凭据解密 → 建 OpenAI 兼容适配器
+  → `Gateway`（注入 `SqlAlchemyLedgerSink` + `resolve_route`）。这是 **draft SSE 的可注入缝**——
+  测试经 `app.dependency_overrides[get_writer_gateway]` 注入 mock 网关（产 `Delta`，绝不联网）。
+- `seed_stub_user`：幂等 seed 单用户 stub（owner_id FK 依赖它，见 memory/gotchas）。
+"""
+
+from __future__ import annotations
+
+from typing import Annotated
+
+from fastapi import Depends
+from openai import AsyncOpenAI
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+from ww_config import get_settings
+from ww_core.domain.chapter_repo import ChapterRepo, SqlChapterRepo
+from ww_core.domain.digest_repo import DigestAppendRepo, SqlDigestAppendRepo
+from ww_core.domain.project_repo import ProjectRepo, SqlProjectRepo
+from ww_core.domain.repositories import MemoryRepos
+from ww_core.domain.review_repo import ReviewRepo, SqlReviewRepo
+from ww_core.memory.sql_repositories import sql_memory_repos
+from ww_db import get_session
+from ww_db.models import User
+from ww_llm_gateway import (
+    Gateway,
+    OpenAICompatAdapter,
+    SqlAlchemyLedgerSink,
+    resolve_route,
+)
+from ww_llm_gateway.types import Tier
+from ww_shared import AppError, ErrorCode
+
+from ww_api.security.credentials import (
+    CredentialKeyError,
+    decrypt_api_key,
+)
+from ww_api.services.credentials import (
+    STUB_OWNER_ID,
+    CredentialStore,
+    SqlCredentialStore,
+)
+from ww_api.services.provider_deps import _PROVIDER_BASE_URLS
+
+# 单用户 stub 的占位邮箱（多租户化时由真实主体替换）。
+_STUB_USER_EMAIL = "stub@local"
+
+
+async def seed_stub_user(session: AsyncSession) -> None:
+    """幂等 seed 单用户 stub 行——所有 owner_id FK（projects/usage_ledger/...）依赖它。
+
+    无该行时插入；已存在则跳过。在 app lifespan 启动时调用一次。
+    """
+    existing = (
+        await session.execute(select(User).where(User.id == STUB_OWNER_ID))
+    ).scalar_one_or_none()
+    if existing is not None:
+        return
+    session.add(User(id=STUB_OWNER_ID, email=_STUB_USER_EMAIL, display_name="stub"))
+    await session.commit()
+
+
+def get_project_repo(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> ProjectRepo:
+    return SqlProjectRepo(session)
+
+
+def get_chapter_repo(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> ChapterRepo:
+    return SqlChapterRepo(session)
+
+
+def get_memory_repos(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> MemoryRepos:
+    """记忆组装的 7-repo 捆绑（draft SSE 用）。测试覆盖此依赖注入内存 fake。"""
+    return sql_memory_repos(session)
+
+
+def get_review_repo(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> ReviewRepo:
+    """审稿留痕 repo（review SSE collect / 历史 / accept 裁决）。"""
+    return SqlReviewRepo(session)
+
+
+def get_digest_append_repo(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> DigestAppendRepo:
+    """章节摘要写侧 repo（验收事务追加终稿 digest）。"""
+    return SqlDigestAppendRepo(session)
+
+
+async def build_gateway_for_tier(
+    session: AsyncSession, store: CredentialStore, tier: Tier
+) -> Gateway:
+    """据指定档位路由解密对应 provider 凭据 → 建网关（解析器仍为全局 `resolve_route`）。
+
+    无凭据/未知 provider → `LLM_UNAVAILABLE`（友好提示，前端引导去配置）。
+    解析器用 `resolve_route`（按 tier 路由）；这里只决定**要预备哪个 provider 的适配器**。
+    """
+    settings = get_settings()
+    route = resolve_route(tier)
+    base_url = _PROVIDER_BASE_URLS.get(route.provider)
+    if base_url is None:
+        raise AppError(
+            ErrorCode.LLM_UNAVAILABLE,
+            f"{tier} 档位 provider {route.provider} 暂不支持",
+            {"provider": route.provider, "tier": tier},
+        )
+    cred = await store.get_credential(STUB_OWNER_ID, route.provider)
+    if cred is None:
+        raise AppError(
+            ErrorCode.LLM_UNAVAILABLE,
+            f"{tier} 档位 provider {route.provider} 未配置凭据，请先在设置中配置",
+            {"provider": route.provider, "tier": tier},
+        )
+    try:
+        api_key = decrypt_api_key(cred.api_key_enc, key=settings.credential_enc_key)
+    except CredentialKeyError as exc:
+        raise AppError(ErrorCode.INTERNAL, str(exc)) from exc
+
+    client = AsyncOpenAI(api_key=api_key, base_url=base_url)
+    adapter = OpenAICompatAdapter(provider=route.provider, client=client)
+    return Gateway(
+        adapters={route.provider: adapter},
+        ledger=SqlAlchemyLedgerSink(session),
+        resolver=resolve_route,
+    )
+
+
+async def build_writer_gateway(session: AsyncSession, store: CredentialStore) -> Gateway:
+    """据 writer 档位路由解密对应 provider 凭据 → 建网关。"""
+    return await build_gateway_for_tier(session, store, "writer")
+
+
+async def get_writer_gateway(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> Gateway:
+    """draft SSE 的可注入网关缝。测试覆盖此依赖注入 mock（产 `Delta`，绝不联网）。"""
+    store = SqlCredentialStore(session)
+    return await build_writer_gateway(session, store)
+
+
+async def get_review_gateway(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> Gateway:
+    """续审（analyst 档位）的可注入网关缝。测试经 override 注 mock（产 `parsed`，绝不联网）。"""
+    store = SqlCredentialStore(session)
+    return await build_gateway_for_tier(session, store, "analyst")
+
+
+async def get_digest_gateway(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> Gateway:
+    """验收终稿 digest 提炼（light 档位）的可注入网关缝。测试经 override 注 mock。"""
+    store = SqlCredentialStore(session)
+    return await build_gateway_for_tier(session, store, "light")
diff --git a/apps/api/ww_api/services/provider_deps.py b/apps/api/ww_api/services/provider_deps.py
new file mode 100644
index 0000000..a82c946
--- /dev/null
+++ b/apps/api/ww_api/services/provider_deps.py
@@ -0,0 +1,93 @@
+"""提供商凭据端点的依赖装配（运行时实现）。
+
+把 `get_session` 装配成 `SqlCredentialStore`；把网关适配器装配成探测器。
+测试经 `app.dependency_overrides` 注入内存替身——不联网。
+"""
+
+from __future__ import annotations
+
+import uuid
+from typing import Annotated
+
+from fastapi import Depends
+from openai import AsyncOpenAI
+from sqlalchemy.ext.asyncio import AsyncSession
+from ww_config import get_settings
+from ww_db import get_session
+from ww_llm_gateway.adapters.base import Capabilities
+from ww_llm_gateway.adapters.openai_compat import OpenAICompatAdapter
+from ww_shared import AppError, ErrorCode
+
+from ww_api.security.credentials import (
+    CredentialKeyError,
+    decrypt_api_key,
+)
+from ww_api.services.credentials import (
+    CredentialStore,
+    SqlCredentialStore,
+)
+
+# 已知 OpenAI 兼容提供商 → base_url（ARCH §4.2）。
+_PROVIDER_BASE_URLS: dict[str, str] = {
+    "deepseek": "https://api.deepseek.com",
+    "kimi": "https://api.moonshot.cn/v1",
+    "qwen": "https://dashscope.aliyuncs.com/compatible-mode/v1",
+    "glm": "https://open.bigmodel.cn/api/paas/v4",
+    "openai": "https://api.openai.com/v1",
+}
+
+
+def get_credential_store(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> CredentialStore:
+    return SqlCredentialStore(session)
+
+
+class GatewayProviderProbe:
+    """运行时探测：解密 Key→建 OpenAI 兼容适配器→最小请求验 Key→回能力矩阵。
+
+    依赖 store 取密文 + settings 取加密 key。绝不在日志/响应回显明文。
+    """
+
+    def __init__(self, store: CredentialStore, enc_key: str) -> None:
+        self._store = store
+        self._enc_key = enc_key
+
+    async def probe(self, owner_id: uuid.UUID, provider: str) -> Capabilities:
+        cred = await self._store.get_credential(owner_id, provider)
+        if cred is None:
+            raise AppError(
+                ErrorCode.NOT_FOUND,
+                f"provider {provider} 未配置凭据",
+                {"provider": provider},
+            )
+        base_url = _PROVIDER_BASE_URLS.get(provider)
+        if base_url is None:
+            raise AppError(
+                ErrorCode.VALIDATION,
+                f"未知提供商 {provider}",
+                {"provider": provider},
+            )
+        try:
+            api_key = decrypt_api_key(cred.api_key_enc, key=self._enc_key)
+        except CredentialKeyError as exc:
+            raise AppError(ErrorCode.INTERNAL, str(exc)) from exc
+
+        client = AsyncOpenAI(api_key=api_key, base_url=base_url)
+        adapter = OpenAICompatAdapter(provider=provider, client=client)
+        try:
+            # 最小探测：列模型即可验证 Key 有效（不消耗生成额度）。
+            await client.models.list()
+        except Exception as exc:  # noqa: BLE001 — 任一失败都映射为 LLM 不可用
+            raise AppError(
+                ErrorCode.LLM_UNAVAILABLE,
+                f"provider {provider} 连接探测失败",
+                {"provider": provider},
+            ) from exc
+        return adapter.capabilities()
+
+
+def get_provider_probe(
+    store: Annotated[CredentialStore, Depends(get_credential_store)],
+) -> GatewayProviderProbe:
+    return GatewayProviderProbe(store, get_settings().credential_enc_key)
diff --git a/apps/web/app/page.tsx b/apps/web/app/page.tsx
new file mode 100644
index 0000000..14bb366
--- /dev/null
+++ b/apps/web/app/page.tsx
@@ -0,0 +1,80 @@
+import Link from "next/link";
+
+import { AppShell } from "@/components/AppShell";
+import { ProjectCard } from "@/components/ProjectCard";
+import { fetchProjects } from "@/lib/api/server";
+import type { ProjectResponse } from "@/lib/api/types";
+
+// 作品库（Dashboard 入口，UX §6.1）。Server Component 读取。
+export default async function DashboardPage() {
+  let projects: ProjectResponse[] = [];
+  let loadError = false;
+  try {
+    const data = await fetchProjects();
+    projects = data.projects ?? [];
+  } catch {
+    loadError = true;
+  }
+
+  return (
+    <AppShell>
+      <div className="mx-auto max-w-5xl px-8 py-10">
+        <div className="mb-8 flex items-center justify-between">
+          <h1 className="font-serif text-3xl text-ink">我的作品</h1>
+          <Link
+            href="/projects/new"
+            className="rounded bg-cinnabar px-4 py-2 text-sm text-panel hover:opacity-90"
+          >
+            ＋ 新建作品
+          </Link>
+        </div>
+
+        {loadError ? (
+          <p className="rounded border border-conflict bg-panel p-6 text-conflict">
+            无法连接后端服务，请确认 API 已启动后刷新。
+          </p>
+        ) : projects.length === 0 ? (
+          <EmptyState />
+        ) : (
+          <ul className="grid grid-cols-1 gap-5 sm:grid-cols-2 lg:grid-cols-3">
+            {projects.map((p) => (
+              <li key={p.id}>
+                <ProjectCard project={p} />
+              </li>
+            ))}
+            <li>
+              <NewProjectCard />
+            </li>
+          </ul>
+        )}
+      </div>
+    </AppShell>
+  );
+}
+
+function EmptyState() {
+  return (
+    <div className="flex flex-col items-center gap-4 rounded border border-line bg-panel py-20 text-center">
+      <p className="font-serif text-2xl text-ink">还没有作品</p>
+      <p className="text-ink-soft">从一句灵感开始你的第一本书。</p>
+      <Link
+        href="/projects/new"
+        className="rounded bg-cinnabar px-5 py-2 text-sm text-panel hover:opacity-90"
+      >
+        ＋ 新建作品
+      </Link>
+    </div>
+  );
+}
+
+function NewProjectCard() {
+  return (
+    <Link
+      href="/projects/new"
+      className="flex h-full min-h-[160px] flex-col items-center justify-center rounded border border-dashed border-line bg-panel text-center hover:border-cinnabar"
+    >
+      <span className="font-serif text-xl text-cinnabar">＋ 新建</span>
+      <span className="mt-2 text-sm text-ink-soft">从一句灵感开始一本书</span>
+    </Link>
+  );
+}
diff --git a/apps/web/app/projects/[id]/write/page.tsx b/apps/web/app/projects/[id]/write/page.tsx
new file mode 100644
index 0000000..696a8a1
--- /dev/null
+++ b/apps/web/app/projects/[id]/write/page.tsx
@@ -0,0 +1,19 @@
+import { notFound } from "next/navigation";
+
+import { Workbench } from "@/components/workbench/Workbench";
+import { fetchProject } from "@/lib/api/server";
+
+interface PageProps {
+  params: Promise<{ id: string }>;
+}
+
+// 写作工作台页（UX §6.3）。Server Component 取项目，Workbench 负责编辑/流式/保存。
+export default async function WritePage({ params }: PageProps) {
+  const { id } = await params;
+  try {
+    const project = await fetchProject(id);
+    return <Workbench project={project} />;
+  } catch {
+    notFound();
+  }
+}
diff --git a/apps/web/app/projects/new/page.tsx b/apps/web/app/projects/new/page.tsx
new file mode 100644
index 0000000..a206611
--- /dev/null
+++ b/apps/web/app/projects/new/page.tsx
@@ -0,0 +1,12 @@
+import { AppShell } from "@/components/AppShell";
+import { ProjectWizard } from "@/components/ProjectWizard";
+
+export default function NewProjectPage() {
+  return (
+    <AppShell>
+      <div className="px-8 py-10">
+        <ProjectWizard />
+      </div>
+    </AppShell>
+  );
+}
diff --git a/apps/web/app/settings/providers/page.tsx b/apps/web/app/settings/providers/page.tsx
new file mode 100644
index 0000000..1737780
--- /dev/null
+++ b/apps/web/app/settings/providers/page.tsx
@@ -0,0 +1,29 @@
+import { AppShell } from "@/components/AppShell";
+import { ProvidersSettings } from "@/components/settings/ProvidersSettings";
+import { fetchProviders } from "@/lib/api/server";
+import type { ProvidersResponse } from "@/lib/api/types";
+
+// 模型与提供商设置（UX §6.10）。Server Component 取初始数据。
+export default async function ProvidersSettingsPage() {
+  let initial: ProvidersResponse = { providers: [], tier_routing: [] };
+  let loadError = false;
+  try {
+    initial = await fetchProviders();
+  } catch {
+    loadError = true;
+  }
+
+  return (
+    <AppShell title="设置 › 模型与提供商">
+      <div className="mx-auto max-w-3xl px-8 py-10">
+        {loadError ? (
+          <p className="rounded border border-conflict bg-panel p-6 text-conflict">
+            无法连接后端服务，请确认 API 已启动后刷新。
+          </p>
+        ) : (
+          <ProvidersSettings initial={initial} />
+        )}
+      </div>
+    </AppShell>
+  );
+}
diff --git a/apps/web/components/ProjectCard.tsx b/apps/web/components/ProjectCard.tsx
new file mode 100644
index 0000000..e1b165b
--- /dev/null
+++ b/apps/web/components/ProjectCard.tsx
@@ -0,0 +1,26 @@
+import Link from "next/link";
+
+import type { ProjectResponse } from "@/lib/api/types";
+
+interface ProjectCardProps {
+  project: ProjectResponse;
+}
+
+// 作品卡（UX §6.1）：书名衬线大字、题材、一句话故事。
+// M1 无字数/章数/待办统计端点 → 暂不展示该徽标（避免编造 API）。
+export function ProjectCard({ project }: ProjectCardProps) {
+  return (
+    <Link
+      href={`/projects/${project.id}/write`}
+      className="flex h-full min-h-[160px] flex-col rounded border border-line bg-panel p-6 shadow-paper hover:border-cinnabar"
+    >
+      <h2 className="font-serif text-2xl text-ink">〈{project.title}〉</h2>
+      {project.genre ? (
+        <p className="mt-2 text-sm text-ink-soft">{project.genre}</p>
+      ) : null}
+      {project.logline ? (
+        <p className="mt-3 line-clamp-3 text-sm text-ink">{project.logline}</p>
+      ) : null}
+    </Link>
+  );
+}
diff --git a/apps/web/components/ProjectWizard.tsx b/apps/web/components/ProjectWizard.tsx
new file mode 100644
index 0000000..ed3ee62
--- /dev/null
+++ b/apps/web/components/ProjectWizard.tsx
@@ -0,0 +1,301 @@
+"use client";
+
+import { useRouter } from "next/navigation";
+import { useState } from "react";
+
+import { api } from "@/lib/api/client";
+import { useToast } from "@/components/Toast";
+import {
+  GENRES,
+  SELLING_POINT_PRESETS,
+  STRUCTURES,
+  WIZARD_STEPS,
+  canAdvance,
+  canSubmit,
+  clampStep,
+  emptyWizardForm,
+  toCreateRequest,
+  type WizardForm,
+} from "@/lib/wizard/wizard";
+
+const STEP_TITLES = [
+  "书名 + 题材",
+  "一句话故事 + 卖点 + 结构",
+  "立意 / 总纲",
+  "主角 / 金手指",
+  "基础世界观",
+];
+
+// 立项向导（UX §6.2）。Client Component：收集字段 → POST /projects → 进工作台。
+export function ProjectWizard() {
+  const router = useRouter();
+  const toast = useToast();
+  const [step, setStep] = useState(1);
+  const [form, setForm] = useState<WizardForm>(emptyWizardForm);
+  const [submitting, setSubmitting] = useState(false);
+
+  const update = (patch: Partial<WizardForm>): void =>
+    setForm((prev) => ({ ...prev, ...patch }));
+
+  const toggleSellingPoint = (point: string): void =>
+    setForm((prev) => ({
+      ...prev,
+      sellingPoints: prev.sellingPoints.includes(point)
+        ? prev.sellingPoints.filter((p) => p !== point)
+        : [...prev.sellingPoints, point],
+    }));
+
+  const goBack = (): void => setStep((s) => clampStep(s - 1));
+  const goNext = (): void => setStep((s) => clampStep(s + 1));
+
+  const submit = async (): Promise<void> => {
+    if (!canSubmit(form)) {
+      toast("请先填写书名", "error");
+      setStep(1);
+      return;
+    }
+    setSubmitting(true);
+    const { data, error } = await api.POST("/projects", {
+      body: toCreateRequest(form),
+    });
+    if (error || !data) {
+      setSubmitting(false);
+      toast("创建作品失败，请重试", "error");
+      return;
+    }
+    router.push(`/projects/${data.id}/write`);
+  };
+
+  const isLast = step === WIZARD_STEPS;
+
+  return (
+    <div className="mx-auto max-w-2xl rounded border border-line bg-panel p-8 shadow-paper">
+      <div className="mb-6 flex items-center justify-between">
+        <h1 className="font-serif text-2xl text-ink">新建作品</h1>
+        <StepDots current={step} />
+      </div>
+      <p className="mb-4 text-sm text-ink-soft">
+        步骤 {step}/{WIZARD_STEPS} · {STEP_TITLES[step - 1]}
+      </p>
+
+      <div className="min-h-[220px]">
+        {step === 1 && <StepBasics form={form} update={update} />}
+        {step === 2 && (
+          <StepStory
+            form={form}
+            update={update}
+            toggleSellingPoint={toggleSellingPoint}
+          />
+        )}
+        {step === 3 && <StepPremise form={form} update={update} />}
+        {step === 4 && <StepProtagonist form={form} update={update} />}
+        {step === 5 && <StepWorld />}
+      </div>
+
+      <div className="mt-8 flex items-center justify-between">
+        <button
+          type="button"
+          onClick={goBack}
+          disabled={step === 1}
+          className="rounded border border-line px-4 py-2 text-sm text-ink disabled:opacity-40"
+        >
+          ← 上一步
+        </button>
+        {isLast ? (
+          <button
+            type="button"
+            onClick={submit}
+            disabled={submitting || !canSubmit(form)}
+            className="rounded bg-cinnabar px-5 py-2 text-sm text-panel disabled:opacity-40"
+          >
+            {submitting ? "创建中…" : "完成立项 →"}
+          </button>
+        ) : (
+          <button
+            type="button"
+            onClick={goNext}
+            disabled={!canAdvance(step, form)}
+            className="rounded bg-cinnabar px-5 py-2 text-sm text-panel disabled:opacity-40"
+          >
+            下一步 →
+          </button>
+        )}
+      </div>
+    </div>
+  );
+}
+
+function StepDots({ current }: { current: number }) {
+  return (
+    <div className="flex gap-1.5" aria-hidden="true">
+      {Array.from({ length: WIZARD_STEPS }, (_, i) => (
+        <span
+          key={i}
+          className={`h-2 w-2 rounded-full ${
+            i + 1 <= current ? "bg-cinnabar" : "bg-line"
+          }`}
+        />
+      ))}
+    </div>
+  );
+}
+
+interface StepProps {
+  form: WizardForm;
+  update: (patch: Partial<WizardForm>) => void;
+}
+
+function Field({
+  label,
+  children,
+}: {
+  label: string;
+  children: React.ReactNode;
+}) {
+  return (
+    <label className="mb-4 block">
+      <span className="mb-1 block text-sm text-ink-soft">{label}</span>
+      {children}
+    </label>
+  );
+}
+
+const inputCls =
+  "w-full rounded border border-line bg-bg px-3 py-2 text-ink focus:border-cinnabar focus:outline-none";
+
+function StepBasics({ form, update }: StepProps) {
+  return (
+    <div>
+      <Field label="书名（必填）">
+        <input
+          className={inputCls}
+          value={form.title}
+          onChange={(e) => update({ title: e.target.value })}
+          placeholder="例：逐光而行"
+          autoFocus
+        />
+      </Field>
+      <Field label="题材">
+        <select
+          className={inputCls}
+          value={form.genre}
+          onChange={(e) => update({ genre: e.target.value })}
+        >
+          <option value="">未选择</option>
+          {GENRES.map((g) => (
+            <option key={g} value={g}>
+              {g}
+            </option>
+          ))}
+        </select>
+      </Field>
+    </div>
+  );
+}
+
+function StepStory({
+  form,
+  update,
+  toggleSellingPoint,
+}: StepProps & { toggleSellingPoint: (p: string) => void }) {
+  return (
+    <div>
+      <Field label="一句话故事（logline）">
+        <textarea
+          className={`${inputCls} h-20 resize-none`}
+          value={form.logline}
+          onChange={(e) => update({ logline: e.target.value })}
+          placeholder="废柴少年觉醒禁忌血脉，在仙门倾轧中逆势封神。"
+        />
+      </Field>
+      <Field label="故事结构">
+        <div className="flex gap-2">
+          {STRUCTURES.map((s) => (
+            <button
+              type="button"
+              key={s}
+              onClick={() => update({ structure: s })}
+              className={`rounded border px-3 py-1.5 text-sm ${
+                form.structure === s
+                  ? "border-cinnabar bg-[var(--color-cinnabar-wash)] text-cinnabar"
+                  : "border-line text-ink"
+              }`}
+            >
+              {s}
+            </button>
+          ))}
+        </div>
+      </Field>
+      <fieldset>
+        <legend className="mb-1 block text-sm text-ink-soft">核心卖点</legend>
+        <div className="flex flex-wrap gap-2">
+          {SELLING_POINT_PRESETS.map((p) => (
+            <button
+              type="button"
+              key={p}
+              aria-pressed={form.sellingPoints.includes(p)}
+              onClick={() => toggleSellingPoint(p)}
+              className={`rounded border px-3 py-1.5 text-sm ${
+                form.sellingPoints.includes(p)
+                  ? "border-cinnabar bg-[var(--color-cinnabar-wash)] text-cinnabar"
+                  : "border-line text-ink"
+              }`}
+            >
+              {p}
+            </button>
+          ))}
+        </div>
+      </fieldset>
+    </div>
+  );
+}
+
+function StepPremise({ form, update }: StepProps) {
+  return (
+    <div>
+      <Field label="立意（premise）">
+        <textarea
+          className={`${inputCls} h-20 resize-none`}
+          value={form.premise}
+          onChange={(e) => update({ premise: e.target.value })}
+          placeholder="故事的核心命题与前提。"
+        />
+      </Field>
+      <Field label="主题（theme）">
+        <input
+          className={inputCls}
+          value={form.theme}
+          onChange={(e) => update({ theme: e.target.value })}
+          placeholder="例：抗争与代价"
+        />
+      </Field>
+    </div>
+  );
+}
+
+function StepProtagonist({ form, update }: StepProps) {
+  // M1 projects 表无独立主角/金手指字段；先并入立意/总纲文本，避免编造 API。
+  return (
+    <div>
+      <p className="mb-3 text-sm text-ink-soft">
+        主角与金手指（M1 暂并入总纲，后续设定库独立建模）。
+      </p>
+      <Field label="主角 / 金手指概要">
+        <textarea
+          className={`${inputCls} h-28 resize-none`}
+          value={form.premise}
+          onChange={(e) => update({ premise: e.target.value })}
+          placeholder="主角设定、金手指来源与限制……"
+        />
+      </Field>
+    </div>
+  );
+}
+
+function StepWorld() {
+  return (
+    <div className="rounded border border-dashed border-line bg-bg p-6 text-sm text-ink-soft">
+      基础世界观将在「设定库」里建模（后续里程碑）。现在可直接完成立项，进入工作台开始写第一章。
+    </div>
+  );
+}
diff --git a/apps/web/components/settings/ProvidersSettings.tsx b/apps/web/components/settings/ProvidersSettings.tsx
new file mode 100644
index 0000000..837c0d4
--- /dev/null
+++ b/apps/web/components/settings/ProvidersSettings.tsx
@@ -0,0 +1,209 @@
+"use client";
+
+import { useState } from "react";
+
+import { api } from "@/lib/api/client";
+import { useToast } from "@/components/Toast";
+import { KNOWN_PROVIDERS, TIER_LABELS } from "@/lib/settings/providers";
+import type {
+  CapabilitiesView,
+  ProvidersResponse,
+  TierRoutingView,
+} from "@/lib/api/types";
+
+interface ProvidersSettingsProps {
+  initial: ProvidersResponse;
+}
+
+interface TestResult {
+  ok: boolean;
+  capabilities: CapabilitiesView;
+}
+
+// 设置页主体（UX §6.10）：档位路由（只读展示）+ 凭据行（脱敏 / 添加更新 / 测试连接）。
+export function ProvidersSettings({ initial }: ProvidersSettingsProps) {
+  const toast = useToast();
+  const [providers, setProviders] = useState(initial.providers ?? []);
+  const tierRouting: TierRoutingView[] = initial.tier_routing ?? [];
+  const [drafts, setDrafts] = useState<Record<string, string>>({});
+  const [savingId, setSavingId] = useState<string | null>(null);
+  const [testingId, setTestingId] = useState<string | null>(null);
+  const [results, setResults] = useState<Record<string, TestResult>>({});
+
+  const maskedFor = (id: string): string | null =>
+    providers.find((p) => p.provider === id)?.masked_key ?? null;
+
+  const saveCredential = async (providerId: string): Promise<void> => {
+    const apiKey = (drafts[providerId] ?? "").trim();
+    if (!apiKey) {
+      toast("请先输入 API Key", "error");
+      return;
+    }
+    setSavingId(providerId);
+    const { data, error } = await api.PUT("/settings/providers", {
+      body: { credentials: [{ provider: providerId, api_key: apiKey }] },
+    });
+    setSavingId(null);
+    if (error || !data) {
+      toast("保存凭据失败，请重试", "error");
+      return;
+    }
+    setProviders(data.providers ?? []);
+    setDrafts((prev) => ({ ...prev, [providerId]: "" }));
+    toast("凭据已保存", "success");
+  };
+
+  const testConnection = async (providerId: string): Promise<void> => {
+    setTestingId(providerId);
+    const { data, error } = await api.POST("/settings/providers/test", {
+      body: { provider: providerId },
+    });
+    setTestingId(null);
+    if (error || !data) {
+      toast("测试连接失败", "error");
+      return;
+    }
+    setResults((prev) => ({
+      ...prev,
+      [providerId]: { ok: data.ok, capabilities: data.capabilities },
+    }));
+    toast(data.ok ? "连接成功" : "连接未通过", data.ok ? "success" : "error");
+  };
+
+  return (
+    <div>
+      <section className="mb-10">
+        <h2 className="mb-3 font-serif text-xl text-ink">能力档位路由</h2>
+        {tierRouting.length === 0 ? (
+          <p className="rounded border border-dashed border-line bg-panel p-4 text-sm text-ink-soft">
+            尚未配置档位路由。连接提供商后将使用默认路由。
+          </p>
+        ) : (
+          <ul className="divide-y divide-line rounded border border-line bg-panel">
+            {tierRouting.map((t) => (
+              <li
+                key={t.tier}
+                className="flex items-center gap-4 px-4 py-3 text-sm"
+              >
+                <span className="w-20 text-ink">
+                  {TIER_LABELS[t.tier] ?? t.tier}
+                </span>
+                <span className="font-mono text-ink-soft">
+                  {t.provider} · {t.model}
+                </span>
+                {t.fallback && t.fallback.length > 0 ? (
+                  <span className="ml-auto font-mono text-xs text-ink-soft">
+                    回退 {t.fallback.join(", ")}
+                  </span>
+                ) : null}
+              </li>
+            ))}
+          </ul>
+        )}
+      </section>
+
+      <section>
+        <h2 className="mb-3 font-serif text-xl text-ink">提供商凭据</h2>
+        {providers.length === 0 ? (
+          <p className="mb-4 rounded border border-dashed border-line bg-panel p-4 text-sm text-ink-soft">
+            至少连接一个提供商即可开始写作（求质量选 Anthropic / 求性价比选
+            DeepSeek）。
+          </p>
+        ) : null}
+        <ul className="divide-y divide-line rounded border border-line bg-panel">
+          {KNOWN_PROVIDERS.map((prov) => {
+            const masked = maskedFor(prov.id);
+            const result = results[prov.id];
+            return (
+              <li key={prov.id} className="px-4 py-4">
+                <div className="flex flex-wrap items-center gap-3">
+                  <span
+                    className={`h-2 w-2 rounded-full ${masked ? "bg-pass" : "bg-line"}`}
+                    aria-hidden="true"
+                  />
+                  <span className="w-28 text-sm text-ink">{prov.label}</span>
+                  {masked ? (
+                    <span className="font-mono text-xs text-ink-soft">
+                      {masked}
+                    </span>
+                  ) : null}
+                  <label className="sr-only" htmlFor={`key-${prov.id}`}>
+                    {prov.label} API Key
+                  </label>
+                  <input
+                    id={`key-${prov.id}`}
+                    type="password"
+                    autoComplete="off"
+                    value={drafts[prov.id] ?? ""}
+                    onChange={(e) =>
+                      setDrafts((prev) => ({
+                        ...prev,
+                        [prov.id]: e.target.value,
+                      }))
+                    }
+                    placeholder={masked ? "输入新 Key 以更新" : "输入 API Key"}
+                    className="min-w-[12rem] flex-1 rounded border border-line bg-bg px-3 py-1.5 text-sm text-ink focus:border-cinnabar focus:outline-none"
+                  />
+                  <button
+                    type="button"
+                    onClick={() => saveCredential(prov.id)}
+                    disabled={savingId === prov.id}
+                    className="rounded bg-cinnabar px-3 py-1.5 text-sm text-panel disabled:opacity-40"
+                  >
+                    {savingId === prov.id
+                      ? "保存中…"
+                      : masked
+                        ? "更新凭据"
+                        : "添加凭据"}
+                  </button>
+                  <button
+                    type="button"
+                    onClick={() => testConnection(prov.id)}
+                    disabled={testingId === prov.id}
+                    className="rounded border border-line px-3 py-1.5 text-sm text-ink disabled:opacity-40"
+                  >
+                    {testingId === prov.id ? "测试中…" : "测试连接"}
+                  </button>
+                </div>
+                {result ? (
+                  <div className="mt-2 flex items-center gap-2 pl-5">
+                    <span
+                      className={`text-xs ${result.ok ? "text-pass" : "text-conflict"}`}
+                    >
+                      {result.ok ? "✓ 已连接" : "✗ 未连接"}
+                    </span>
+                    <CapabilityBadges caps={result.capabilities} />
+                  </div>
+                ) : null}
+              </li>
+            );
+          })}
+        </ul>
+      </section>
+    </div>
+  );
+}
+
+function CapabilityBadges({ caps }: { caps: CapabilitiesView }) {
+  const badges: { on: boolean; label: string }[] = [
+    { on: caps.structured_output, label: "结构化" },
+    { on: caps.prefix_cache, label: "前缀缓存" },
+    { on: caps.thinking, label: "思考" },
+  ];
+  return (
+    <div className="flex gap-1.5">
+      {badges.map((b) => (
+        <span
+          key={b.label}
+          className={`rounded px-2 py-0.5 text-[11px] ${
+            b.on
+              ? "bg-[var(--color-cinnabar-wash)] text-cinnabar"
+              : "bg-bg text-ink-soft/50"
+          }`}
+        >
+          {b.label}
+        </span>
+      ))}
+    </div>
+  );
+}
diff --git a/apps/web/components/workbench/ChapterAssistant.tsx b/apps/web/components/workbench/ChapterAssistant.tsx
new file mode 100644
index 0000000..7db6ee3
--- /dev/null
+++ b/apps/web/components/workbench/ChapterAssistant.tsx
@@ -0,0 +1,27 @@
+// 右栏「本章助手」（UX §6.3）。
+// M1 无端点暴露注入选择轨迹/四审结果 → 占位说明，不编造 API。
+export function ChapterAssistant() {
+  return (
+    <aside className="hidden border-l border-line bg-panel p-4 lg:block">
+      <h2 className="text-sm font-semibold text-ink">本章助手</h2>
+
+      <section className="mt-4">
+        <h3 className="text-xs font-semibold uppercase tracking-wide text-ink-soft">
+          本章注入（透明）
+        </h3>
+        <p className="mt-2 rounded border border-dashed border-line bg-bg p-3 text-xs text-ink-soft">
+          M1 暂未接：注入的设定/伏笔选择轨迹尚无对应端点，将在记忆服务接入后展示。
+        </p>
+      </section>
+
+      <section className="mt-4">
+        <h3 className="text-xs font-semibold uppercase tracking-wide text-ink-soft">
+          四审
+        </h3>
+        <p className="mt-2 rounded border border-dashed border-line bg-bg p-3 text-xs text-ink-soft">
+          M2 开放：一致性 / 伏笔 / 文风 / 节奏。
+        </p>
+      </section>
+    </aside>
+  );
+}
diff --git a/apps/web/components/workbench/ChapterList.tsx b/apps/web/components/workbench/ChapterList.tsx
new file mode 100644
index 0000000..f4ef596
--- /dev/null
+++ b/apps/web/components/workbench/ChapterList.tsx
@@ -0,0 +1,24 @@
+interface ChapterListProps {
+  currentChapterNo: number;
+}
+
+// 左栏目录（UX §6.3）。M1 无章节列表端点 → 仅展示当前章占位。
+export function ChapterList({ currentChapterNo }: ChapterListProps) {
+  return (
+    <aside className="hidden border-r border-line bg-panel py-4 lg:block">
+      <h2 className="px-4 text-xs font-semibold uppercase tracking-wide text-ink-soft">
+        目录
+      </h2>
+      <ul className="mt-2">
+        <li>
+          <span className="flex items-center gap-2 border-l-2 border-cinnabar bg-[var(--color-cinnabar-wash)] px-4 py-2 text-sm text-cinnabar">
+            <span aria-hidden="true">●</span>第 {currentChapterNo} 章
+          </span>
+        </li>
+      </ul>
+      <p className="mt-4 px-4 text-xs text-ink-soft/70">
+        多卷多章目录将在大纲模块开放（后续里程碑）。
+      </p>
+    </aside>
+  );
+}
diff --git a/apps/web/components/workbench/Editor.tsx b/apps/web/components/workbench/Editor.tsx
new file mode 100644
index 0000000..fedcd19
--- /dev/null
+++ b/apps/web/components/workbench/Editor.tsx
@@ -0,0 +1,34 @@
+"use client";
+
+interface EditorProps {
+  value: string;
+  onChange: (value: string) => void;
+  streaming: boolean;
+}
+
+// 中栏正文编辑器：宋体、720px 居中、行高 1.9（UX §2.3 / §6.3）。
+// 流式中显示朱砂光标提示（prefers-reduced-motion 下不闪烁，见 globals.css）。
+export function Editor({ value, onChange, streaming }: EditorProps) {
+  return (
+    <div className="mx-auto max-w-prose">
+      <label htmlFor="chapter-editor" className="sr-only">
+        正文编辑器
+      </label>
+      <textarea
+        id="chapter-editor"
+        value={value}
+        onChange={(e) => onChange(e.target.value)}
+        readOnly={streaming}
+        aria-busy={streaming}
+        placeholder="夜色如墨……点击「写本章」让 AI 起草，或直接在此书写。"
+        className="min-h-[60vh] w-full resize-none bg-transparent font-serif text-[18px] leading-[1.9] text-ink placeholder:text-ink-soft/50 focus:outline-none"
+      />
+      {streaming ? (
+        <span
+          className="typewriter-cursor ml-0.5 inline-block h-[1.2em] w-[2px] translate-y-1 bg-cinnabar align-middle"
+          aria-hidden="true"
+        />
+      ) : null}
+    </div>
+  );
+}
diff --git a/apps/web/components/workbench/Workbench.tsx b/apps/web/components/workbench/Workbench.tsx
new file mode 100644
index 0000000..7657b8a
--- /dev/null
+++ b/apps/web/components/workbench/Workbench.tsx
@@ -0,0 +1,156 @@
+"use client";
+
+import Link from "next/link";
+import { useEffect, useRef, useState } from "react";
+
+import { AppShell } from "@/components/AppShell";
+import type { ProjectResponse } from "@/lib/api/types";
+import { useAutosave } from "@/lib/autosave/useAutosave";
+import { useDraftStream } from "@/lib/stream/useDraftStream";
+import { ChapterList } from "./ChapterList";
+import { ChapterAssistant } from "./ChapterAssistant";
+import { Editor } from "./Editor";
+
+interface WorkbenchProps {
+  project: ProjectResponse;
+}
+
+// M1：单章工作台。章节列表为占位（无章节列表端点）；默认第 1 章。
+const M1_CHAPTER_NO = 1;
+
+export function Workbench({ project }: WorkbenchProps) {
+  const chapterNo = M1_CHAPTER_NO;
+  const [text, setText] = useState("");
+  const autosave = useAutosave(project.id, chapterNo);
+  const stream = useDraftStream();
+  const lastStreamText = useRef("");
+
+  // 流式 token 累积进编辑器（打字机）；停止/结束后已生成部分留在草稿。
+  useEffect(() => {
+    const streamed = stream.state.text;
+    if (
+      (stream.state.phase === "streaming" ||
+        stream.state.phase === "done" ||
+        stream.state.phase === "aborted") &&
+      streamed !== lastStreamText.current
+    ) {
+      lastStreamText.current = streamed;
+      setText(streamed);
+      autosave.onChange(streamed);
+    }
+  }, [stream.state.phase, stream.state.text, autosave]);
+
+  const onWrite = (): void => {
+    void stream.start(project.id, chapterNo);
+  };
+
+  const onEditorChange = (value: string): void => {
+    setText(value);
+    autosave.onChange(value);
+  };
+
+  const wordCount = text.length;
+
+  return (
+    <AppShell
+      title={`〈${project.title}〉`}
+      subtitle={`第 ${chapterNo} 章`}
+    >
+      <div className="grid h-[calc(100vh-4rem)] grid-cols-1 lg:grid-cols-[12rem_1fr_18rem]">
+        <ChapterList currentChapterNo={chapterNo} />
+
+        <section className="flex min-w-0 flex-col bg-bg">
+          <div className="flex-1 overflow-auto px-6 py-8">
+            <Editor
+              value={text}
+              onChange={onEditorChange}
+              streaming={stream.isStreaming}
+            />
+          </div>
+          <Toolbar
+            projectId={project.id}
+            chapterNo={chapterNo}
+            wordCount={wordCount}
+            savedLabel={autosave.savedLabel}
+            saveStatus={autosave.status}
+            streaming={stream.isStreaming}
+            streamError={stream.state.error}
+            onWrite={onWrite}
+            onStop={stream.stop}
+          />
+        </section>
+
+        <ChapterAssistant />
+      </div>
+    </AppShell>
+  );
+}
+
+interface ToolbarProps {
+  projectId: string;
+  chapterNo: number;
+  wordCount: number;
+  savedLabel: string | null;
+  saveStatus: ReturnType<typeof useAutosave>["status"];
+  streaming: boolean;
+  streamError: { code: string; message: string } | null;
+  onWrite: () => void;
+  onStop: () => void;
+}
+
+function Toolbar({
+  projectId,
+  chapterNo,
+  wordCount,
+  savedLabel,
+  saveStatus,
+  streaming,
+  streamError,
+  onWrite,
+  onStop,
+}: ToolbarProps) {
+  return (
+    <div className="border-t border-line bg-panel px-6 py-3">
+      {streamError ? (
+        <p className="mb-2 text-sm text-conflict">
+          写章失败（{streamError.code}）：{streamError.message}
+        </p>
+      ) : null}
+      <div className="flex items-center gap-4">
+        {streaming ? (
+          <button
+            type="button"
+            onClick={onStop}
+            className="rounded border border-conflict px-4 py-2 text-sm text-conflict"
+          >
+            停
+          </button>
+        ) : (
+          <button
+            type="button"
+            onClick={onWrite}
+            className="rounded bg-cinnabar px-4 py-2 text-sm text-panel hover:opacity-90"
+          >
+            ✍ 写本章
+          </button>
+        )}
+        <span className="font-mono text-xs text-ink-soft">
+          字数 {wordCount.toLocaleString()}
+        </span>
+        <Link
+          href={`/projects/${projectId}/review?chapter=${chapterNo}`}
+          className="rounded border border-cinnabar px-4 py-2 text-sm text-cinnabar hover:bg-[var(--color-cinnabar-wash)]"
+        >
+          审稿 →
+        </Link>
+        <span className="ml-auto font-mono text-xs text-ink-soft">
+          {saveStatus === "saving"
+            ? "保存中…"
+            : saveStatus === "error"
+              ? "保存失败"
+              : (savedLabel ?? "尚未保存")}
+        </span>
+      </div>
+    </div>
+  );
+}
diff --git a/apps/web/lib/autosave/autosave.test.ts b/apps/web/lib/autosave/autosave.test.ts
new file mode 100644
index 0000000..5455300
--- /dev/null
+++ b/apps/web/lib/autosave/autosave.test.ts
@@ -0,0 +1,61 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+
+import { createDebouncedSaver, formatSavedAt } from "./autosave";
+
+describe("createDebouncedSaver", () => {
+  beforeEach(() => vi.useFakeTimers());
+  afterEach(() => vi.useRealTimers());
+
+  it("saves only once after the debounce window for rapid edits", () => {
+    const save = vi.fn().mockResolvedValue(undefined);
+    const saver = createDebouncedSaver(save, 1000);
+    saver.schedule("a");
+    saver.schedule("ab");
+    saver.schedule("abc");
+    expect(save).not.toHaveBeenCalled();
+    vi.advanceTimersByTime(1000);
+    expect(save).toHaveBeenCalledTimes(1);
+    expect(save).toHaveBeenCalledWith("abc");
+  });
+
+  it("flush triggers the pending save immediately", () => {
+    const save = vi.fn().mockResolvedValue(undefined);
+    const saver = createDebouncedSaver(save, 1000);
+    saver.schedule("hello");
+    saver.flush();
+    expect(save).toHaveBeenCalledTimes(1);
+    expect(save).toHaveBeenCalledWith("hello");
+    // No double-fire when the timer would have elapsed.
+    vi.advanceTimersByTime(1000);
+    expect(save).toHaveBeenCalledTimes(1);
+  });
+
+  it("cancel prevents a scheduled save", () => {
+    const save = vi.fn().mockResolvedValue(undefined);
+    const saver = createDebouncedSaver(save, 1000);
+    saver.schedule("x");
+    saver.cancel();
+    vi.advanceTimersByTime(2000);
+    expect(save).not.toHaveBeenCalled();
+  });
+
+  it("each new schedule resets the timer (only last value persists)", () => {
+    const save = vi.fn().mockResolvedValue(undefined);
+    const saver = createDebouncedSaver(save, 1000);
+    saver.schedule("one");
+    vi.advanceTimersByTime(800);
+    saver.schedule("two");
+    vi.advanceTimersByTime(800);
+    expect(save).not.toHaveBeenCalled();
+    vi.advanceTimersByTime(200);
+    expect(save).toHaveBeenCalledTimes(1);
+    expect(save).toHaveBeenCalledWith("two");
+  });
+});
+
+describe("formatSavedAt", () => {
+  it("formats as 保存于 hh:mm zero-padded", () => {
+    const d = new Date(2026, 5, 18, 9, 4);
+    expect(formatSavedAt(d)).toBe("保存于 09:04");
+  });
+});
diff --git a/apps/web/lib/autosave/autosave.ts b/apps/web/lib/autosave/autosave.ts
new file mode 100644
index 0000000..1a16f0a
--- /dev/null
+++ b/apps/web/lib/autosave/autosave.ts
@@ -0,0 +1,52 @@
+// 自动保存的纯逻辑：防抖调度器 + 时间格式化。与 React 解耦，便于单测。
+
+export const AUTOSAVE_DELAY_MS = 1200;
+
+export type SaveFn = (text: string) => Promise<void>;
+
+export interface DebouncedSaver {
+  schedule: (text: string) => void;
+  flush: () => void;
+  cancel: () => void;
+}
+
+// 防抖：最后一次输入静默 delayMs 后才真正保存。flush 立即触发待保存项。
+export function createDebouncedSaver(
+  save: SaveFn,
+  delayMs: number = AUTOSAVE_DELAY_MS,
+): DebouncedSaver {
+  let timer: ReturnType<typeof setTimeout> | null = null;
+  let pending: string | null = null;
+
+  const run = (): void => {
+    if (pending === null) return;
+    const text = pending;
+    pending = null;
+    timer = null;
+    void save(text);
+  };
+
+  return {
+    schedule(text: string): void {
+      pending = text;
+      if (timer !== null) clearTimeout(timer);
+      timer = setTimeout(run, delayMs);
+    },
+    flush(): void {
+      if (timer !== null) clearTimeout(timer);
+      run();
+    },
+    cancel(): void {
+      if (timer !== null) clearTimeout(timer);
+      timer = null;
+      pending = null;
+    },
+  };
+}
+
+// 「保存于 hh:mm」标签。
+export function formatSavedAt(date: Date): string {
+  const hh = String(date.getHours()).padStart(2, "0");
+  const mm = String(date.getMinutes()).padStart(2, "0");
+  return `保存于 ${hh}:${mm}`;
+}
diff --git a/apps/web/lib/autosave/useAutosave.ts b/apps/web/lib/autosave/useAutosave.ts
new file mode 100644
index 0000000..c58e2da
--- /dev/null
+++ b/apps/web/lib/autosave/useAutosave.ts
@@ -0,0 +1,74 @@
+"use client";
+
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
+
+import { api } from "@/lib/api/client";
+import { useToast } from "@/components/Toast";
+import {
+  AUTOSAVE_DELAY_MS,
+  createDebouncedSaver,
+  formatSavedAt,
+  type DebouncedSaver,
+} from "./autosave";
+
+export type SaveStatus = "idle" | "saving" | "saved" | "error";
+
+export interface UseAutosave {
+  status: SaveStatus;
+  savedLabel: string | null;
+  onChange: (text: string) => void;
+  flush: () => void;
+}
+
+// 防抖自动保存草稿：乐观置 saving → PUT 成功置 saved（落「保存于 hh:mm」）；
+// 失败回滚状态 + toast，不丢正文（已在编辑器里）。
+export function useAutosave(
+  projectId: string,
+  chapterNo: number,
+): UseAutosave {
+  const [status, setStatus] = useState<SaveStatus>("idle");
+  const [savedLabel, setSavedLabel] = useState<string | null>(null);
+  const toast = useToast();
+  const lastSavedRef = useRef<string | null>(null);
+
+  const save = useCallback(
+    async (text: string): Promise<void> => {
+      if (text === lastSavedRef.current) return;
+      setStatus("saving");
+      const { error } = await api.PUT(
+        "/projects/{project_id}/chapters/{chapter_no}/draft",
+        {
+          params: { path: { project_id: projectId, chapter_no: chapterNo } },
+          body: { text },
+        },
+      );
+      if (error) {
+        setStatus("error");
+        toast("自动保存失败，稍后重试（正文未丢失）", "error");
+        return;
+      }
+      lastSavedRef.current = text;
+      setSavedLabel(formatSavedAt(new Date()));
+      setStatus("saved");
+    },
+    [projectId, chapterNo, toast],
+  );
+
+  const saver: DebouncedSaver = useMemo(
+    () => createDebouncedSaver(save, AUTOSAVE_DELAY_MS),
+    [save],
+  );
+
+  useEffect(() => () => saver.cancel(), [saver]);
+
+  const onChange = useCallback(
+    (text: string) => {
+      saver.schedule(text);
+    },
+    [saver],
+  );
+
+  const flush = useCallback(() => saver.flush(), [saver]);
+
+  return { status, savedLabel, onChange, flush };
+}
diff --git a/apps/web/lib/settings/providers.ts b/apps/web/lib/settings/providers.ts
new file mode 100644
index 0000000..3496567
--- /dev/null
+++ b/apps/web/lib/settings/providers.ts
@@ -0,0 +1,16 @@
+// 已知提供商（UX §6.10）。后端按 provider 字符串识别；这里只是 UI 候选列表。
+export const KNOWN_PROVIDERS: { id: string; label: string }[] = [
+  { id: "anthropic", label: "Anthropic" },
+  { id: "deepseek", label: "DeepSeek" },
+  { id: "kimi", label: "Kimi" },
+  { id: "openai", label: "OpenAI" },
+  { id: "qwen", label: "通义千问" },
+  { id: "glm", label: "智谱 GLM" },
+  { id: "gemini", label: "Gemini" },
+];
+
+export const TIER_LABELS: Record<string, string> = {
+  writer: "写手档",
+  analyst: "分析档",
+  light: "轻量档",
+};
diff --git a/apps/web/lib/stream/sse.test.ts b/apps/web/lib/stream/sse.test.ts
new file mode 100644
index 0000000..8923300
--- /dev/null
+++ b/apps/web/lib/stream/sse.test.ts
@@ -0,0 +1,96 @@
+import { describe, expect, it } from "vitest";
+
+import {
+  SseFrameBuffer,
+  initialStreamState,
+  parseSseBlock,
+  reduceStream,
+  type SseEvent,
+} from "./sse";
+
+describe("parseSseBlock", () => {
+  it("parses a token frame", () => {
+    const evt = parseSseBlock('event:token\ndata:{"text":"夜"}');
+    expect(evt).toEqual({ event: "token", data: { text: "夜" } });
+  });
+
+  it("parses a done frame", () => {
+    const evt = parseSseBlock('event:done\ndata:{"length":12}');
+    expect(evt).toEqual({ event: "done", data: { length: 12 } });
+  });
+
+  it("parses an error frame with request_id", () => {
+    const evt = parseSseBlock(
+      'event:error\ndata:{"code":"INTERNAL","message":"boom","request_id":"r1"}',
+    );
+    expect(evt).toEqual({
+      event: "error",
+      data: { code: "INTERNAL", message: "boom", request_id: "r1" },
+    });
+  });
+
+  it("ignores comment/heartbeat lines and tolerates leading space after colon", () => {
+    const evt = parseSseBlock(': keep-alive\nevent: token\ndata: {"text":"x"}');
+    expect(evt).toEqual({ event: "token", data: { text: "x" } });
+  });
+
+  it("returns null for unknown event types", () => {
+    expect(parseSseBlock('event:section\ndata:{"x":1}')).toBeNull();
+  });
+
+  it("returns null for malformed json", () => {
+    expect(parseSseBlock("event:token\ndata:{not json")).toBeNull();
+  });
+});
+
+describe("SseFrameBuffer", () => {
+  it("emits complete blocks split by blank line and buffers the tail", () => {
+    const buf = new SseFrameBuffer();
+    const first = buf.push('event:token\ndata:{"text":"a"}\n\nevent:tok');
+    expect(first).toEqual([{ event: "token", data: { text: "a" } }]);
+    const second = buf.push('en\ndata:{"text":"b"}\n\n');
+    expect(second).toEqual([{ event: "token", data: { text: "b" } }]);
+  });
+
+  it("handles CRLF boundaries", () => {
+    const buf = new SseFrameBuffer();
+    const out = buf.push('event:token\r\ndata:{"text":"z"}\r\n\r\n');
+    expect(out).toEqual([{ event: "token", data: { text: "z" } }]);
+  });
+});
+
+describe("reduceStream", () => {
+  it("accumulates token text and marks streaming", () => {
+    const tokens: SseEvent[] = [
+      { event: "token", data: { text: "夜色" } },
+      { event: "token", data: { text: "如墨" } },
+    ];
+    const state = tokens.reduce(reduceStream, initialStreamState);
+    expect(state.text).toBe("夜色如墨");
+    expect(state.phase).toBe("streaming");
+  });
+
+  it("marks done on done event preserving text", () => {
+    let s = reduceStream(initialStreamState, {
+      event: "token",
+      data: { text: "abc" },
+    });
+    s = reduceStream(s, { event: "done", data: { length: 3 } });
+    expect(s.phase).toBe("done");
+    expect(s.text).toBe("abc");
+  });
+
+  it("captures error and retains partial text", () => {
+    let s = reduceStream(initialStreamState, {
+      event: "token",
+      data: { text: "partial" },
+    });
+    s = reduceStream(s, {
+      event: "error",
+      data: { code: "LLM_UNAVAILABLE", message: "no key" },
+    });
+    expect(s.phase).toBe("error");
+    expect(s.text).toBe("partial");
+    expect(s.error?.code).toBe("LLM_UNAVAILABLE");
+  });
+});
diff --git a/apps/web/lib/stream/sse.ts b/apps/web/lib/stream/sse.ts
new file mode 100644
index 0000000..7a92e12
--- /dev/null
+++ b/apps/web/lib/stream/sse.ts
@@ -0,0 +1,106 @@
+// SSE 帧解析 + 草稿流归一（对齐 C3 / ARCH §7.3）。
+// 帧格式：`event:<token|done|error>\ndata:<json>\n\n`。
+// 纯逻辑，便于单测（不依赖浏览器/DOM）。
+
+export interface TokenEvent {
+  event: "token";
+  data: { text: string };
+}
+export interface DoneEvent {
+  event: "done";
+  data: { length: number };
+}
+export interface ErrorEvent {
+  event: "error";
+  data: { code: string; message: string; request_id?: string | null };
+}
+export type SseEvent = TokenEvent | DoneEvent | ErrorEvent;
+
+const KNOWN_EVENTS = new Set(["token", "done", "error"]);
+
+// 把一个完整 SSE 块（多行）解析成事件；无法解析则返回 null（跳过）。
+export function parseSseBlock(block: string): SseEvent | null {
+  let event = "";
+  const dataLines: string[] = [];
+  for (const rawLine of block.split("\n")) {
+    const line = rawLine.replace(/\r$/, "");
+    if (line.startsWith(":")) continue; // 注释/心跳
+    const sep = line.indexOf(":");
+    if (sep === -1) continue;
+    const field = line.slice(0, sep);
+    const value = line.slice(sep + 1).replace(/^ /, "");
+    if (field === "event") event = value;
+    else if (field === "data") dataLines.push(value);
+  }
+  if (!KNOWN_EVENTS.has(event) || dataLines.length === 0) return null;
+  let data: unknown;
+  try {
+    data = JSON.parse(dataLines.join("\n"));
+  } catch {
+    return null;
+  }
+  return { event, data } as SseEvent;
+}
+
+// 增量缓冲：吃进一段文本，吐出已完成的事件块（以空行分隔），保留未完成尾部。
+export class SseFrameBuffer {
+  private buf = "";
+
+  push(chunk: string): SseEvent[] {
+    this.buf += chunk;
+    const events: SseEvent[] = [];
+    let idx: number;
+    // 块之间以空行（\n\n，兼容 \r\n\r\n）分隔。
+    while ((idx = this.findBoundary(this.buf)) !== -1) {
+      const block = this.buf.slice(0, idx.valueOf());
+      this.buf = this.buf.slice(this.boundaryEnd(this.buf, idx));
+      const evt = parseSseBlock(block);
+      if (evt) events.push(evt);
+    }
+    return events;
+  }
+
+  private findBoundary(s: string): number {
+    const a = s.indexOf("\n\n");
+    const b = s.indexOf("\r\n\r\n");
+    if (a === -1) return b;
+    if (b === -1) return a;
+    return Math.min(a, b);
+  }
+
+  private boundaryEnd(s: string, idx: number): number {
+    return s.startsWith("\r\n\r\n", idx) ? idx + 4 : idx + 2;
+  }
+}
+
+export type StreamPhase = "idle" | "streaming" | "done" | "error" | "aborted";
+
+export interface StreamState {
+  phase: StreamPhase;
+  text: string;
+  error: { code: string; message: string; request_id?: string | null } | null;
+}
+
+export const initialStreamState: StreamState = {
+  phase: "idle",
+  text: "",
+  error: null,
+};
+
+// 纯 reducer：把单个事件折叠进状态（打字机文本累积）。
+export function reduceStream(state: StreamState, event: SseEvent): StreamState {
+  switch (event.event) {
+    case "token":
+      return {
+        ...state,
+        phase: "streaming",
+        text: state.text + event.data.text,
+      };
+    case "done":
+      return { ...state, phase: "done" };
+    case "error":
+      return { ...state, phase: "error", error: event.data };
+    default:
+      return state;
+  }
+}
diff --git a/apps/web/lib/stream/useDraftStream.ts b/apps/web/lib/stream/useDraftStream.ts
new file mode 100644
index 0000000..d232ed9
--- /dev/null
+++ b/apps/web/lib/stream/useDraftStream.ts
@@ -0,0 +1,122 @@
+"use client";
+
+import { useCallback, useReducer, useRef } from "react";
+
+import { API_BASE_PUBLIC } from "@/lib/api/config";
+import {
+  SseFrameBuffer,
+  initialStreamState,
+  reduceStream,
+  type StreamState,
+} from "./sse";
+
+type Action =
+  | { type: "start" }
+  | { type: "events"; events: ReturnType<SseFrameBuffer["push"]> }
+  | { type: "abort" }
+  | { type: "fail"; code: string; message: string }
+  | { type: "reset"; text: string };
+
+function reducer(state: StreamState, action: Action): StreamState {
+  switch (action.type) {
+    case "start":
+      return { phase: "streaming", text: "", error: null };
+    case "events":
+      return action.events.reduce(reduceStream, state);
+    case "abort":
+      return { ...state, phase: "aborted" };
+    case "fail":
+      return {
+        ...state,
+        phase: "error",
+        error: { code: action.code, message: action.message },
+      };
+    case "reset":
+      return { ...initialStreamState, text: action.text };
+    default:
+      return state;
+  }
+}
+
+export interface UseDraftStream {
+  state: StreamState;
+  isStreaming: boolean;
+  start: (projectId: string, chapterNo: number) => Promise<void>;
+  stop: () => void;
+  reset: (text: string) => void;
+}
+
+// 消费 POST .../draft 的 SSE 流。用 fetch+ReadableStream（EventSource 不支持 POST）。
+// "停" = abort 连接（已生成部分保留在 state.text，由调用方落入草稿）。
+export function useDraftStream(): UseDraftStream {
+  const [state, dispatch] = useReducer(reducer, initialStreamState);
+  const controllerRef = useRef<AbortController | null>(null);
+
+  const stop = useCallback(() => {
+    controllerRef.current?.abort();
+    controllerRef.current = null;
+    dispatch({ type: "abort" });
+  }, []);
+
+  const reset = useCallback((text: string) => {
+    dispatch({ type: "reset", text });
+  }, []);
+
+  const start = useCallback(async (projectId: string, chapterNo: number) => {
+    const controller = new AbortController();
+    controllerRef.current = controller;
+    dispatch({ type: "start" });
+    try {
+      const res = await fetch(
+        `${API_BASE_PUBLIC}/projects/${projectId}/chapters/${chapterNo}/draft`,
+        {
+          method: "POST",
+          headers: { Accept: "text/event-stream" },
+          signal: controller.signal,
+        },
+      );
+      if (!res.ok || !res.body) {
+        // 流前错误（如无凭据 → 503 LLM_UNAVAILABLE，JSON 信封而非帧）。
+        let code = "STREAM_FAILED";
+        let message = `写章请求失败（${res.status}）`;
+        try {
+          const body = (await res.json()) as {
+            error?: { code?: string; message?: string };
+          };
+          if (body.error?.code) code = body.error.code;
+          if (body.error?.message) message = body.error.message;
+        } catch {
+          // 非 JSON 信封，沿用默认文案。
+        }
+        dispatch({ type: "fail", code, message });
+        return;
+      }
+      const reader = res.body.getReader();
+      const decoder = new TextDecoder();
+      const buffer = new SseFrameBuffer();
+      for (;;) {
+        const { value, done } = await reader.read();
+        if (done) break;
+        const events = buffer.push(decoder.decode(value, { stream: true }));
+        if (events.length > 0) dispatch({ type: "events", events });
+      }
+    } catch (err: unknown) {
+      if (err instanceof DOMException && err.name === "AbortError") {
+        // 用户主动停止：state 已置 aborted。
+        return;
+      }
+      const message = err instanceof Error ? err.message : "未知网络错误";
+      dispatch({ type: "fail", code: "NETWORK", message });
+    } finally {
+      controllerRef.current = null;
+    }
+  }, []);
+
+  return {
+    state,
+    isStreaming: state.phase === "streaming",
+    start,
+    stop,
+    reset,
+  };
+}
diff --git a/apps/web/lib/wizard/wizard.test.ts b/apps/web/lib/wizard/wizard.test.ts
new file mode 100644
index 0000000..3894e7d
--- /dev/null
+++ b/apps/web/lib/wizard/wizard.test.ts
@@ -0,0 +1,81 @@
+import { describe, expect, it } from "vitest";
+
+import {
+  canAdvance,
+  canSubmit,
+  clampStep,
+  emptyWizardForm,
+  toCreateRequest,
+  WIZARD_STEPS,
+  type WizardForm,
+} from "./wizard";
+
+describe("wizard step gating", () => {
+  it("blocks advancing past step 1 without a title", () => {
+    expect(canAdvance(1, emptyWizardForm)).toBe(false);
+  });
+
+  it("allows advancing step 1 once title is set", () => {
+    expect(canAdvance(1, { ...emptyWizardForm, title: "逐光而行" })).toBe(true);
+  });
+
+  it("allows advancing later steps even when fields are empty", () => {
+    expect(canAdvance(3, emptyWizardForm)).toBe(true);
+  });
+
+  it("clamps step within bounds", () => {
+    expect(clampStep(0)).toBe(1);
+    expect(clampStep(WIZARD_STEPS + 3)).toBe(WIZARD_STEPS);
+    expect(clampStep(3)).toBe(3);
+  });
+
+  it("requires a title to submit", () => {
+    expect(canSubmit(emptyWizardForm)).toBe(false);
+    expect(canSubmit({ ...emptyWizardForm, title: "x" })).toBe(true);
+  });
+});
+
+describe("toCreateRequest", () => {
+  it("maps form to snake_case request, nulling empty optionals", () => {
+    const form: WizardForm = {
+      title: "  逐光而行  ",
+      genre: "玄幻",
+      logline: "",
+      sellingPoints: ["逆袭", "系统流"],
+      structure: "三幕",
+      premise: "  ",
+      theme: "抗争",
+    };
+    expect(toCreateRequest(form)).toEqual({
+      title: "逐光而行",
+      genre: "玄幻",
+      logline: null,
+      premise: null,
+      theme: "抗争",
+      selling_points: ["逆袭", "系统流"],
+      structure: "三幕",
+    });
+  });
+});
+
+// wizard→create 流程：归一后的请求体经 mock 客户端提交，返回新建项目 id。
+describe("wizard create flow (mocked client)", () => {
+  it("submits the normalized body and resolves a project id", async () => {
+    const post = async (
+      _path: string,
+      opts: { body: ReturnType<typeof toCreateRequest> },
+    ): Promise<{ data: { id: string }; error: null }> => {
+      expect(opts.body.title).toBe("青冥录");
+      expect(opts.body.selling_points).toEqual(["群像"]);
+      return { data: { id: "proj-123" }, error: null };
+    };
+
+    const form: WizardForm = {
+      ...emptyWizardForm,
+      title: "青冥录",
+      sellingPoints: ["群像"],
+    };
+    const { data } = await post("/projects", { body: toCreateRequest(form) });
+    expect(data.id).toBe("proj-123");
+  });
+});
diff --git a/apps/web/lib/wizard/wizard.ts b/apps/web/lib/wizard/wizard.ts
new file mode 100644
index 0000000..c1ab6cb
--- /dev/null
+++ b/apps/web/lib/wizard/wizard.ts
@@ -0,0 +1,62 @@
+import type { ProjectCreateRequest } from "@/lib/api/types";
+
+// 立项向导 5 步（UX §6.2）。表单状态用字符串/数组，提交时归一为 ProjectCreateRequest。
+export interface WizardForm {
+  title: string;
+  genre: string;
+  logline: string;
+  sellingPoints: string[];
+  structure: string;
+  premise: string;
+  theme: string;
+}
+
+export const WIZARD_STEPS = 5;
+
+export const emptyWizardForm: WizardForm = {
+  title: "",
+  genre: "",
+  logline: "",
+  sellingPoints: [],
+  structure: "",
+  premise: "",
+  theme: "",
+};
+
+export const GENRES = ["玄幻", "仙侠", "都市", "科幻", "历史", "悬疑"];
+export const STRUCTURES = ["三幕", "故事圈", "雪花"];
+export const SELLING_POINT_PRESETS = ["逆袭", "打脸", "系统流", "双男主", "群像"];
+
+// 每一步可否前进：仅第 1 步（书名）必填，其余可空着继续（草稿式立项）。
+export function canAdvance(step: number, form: WizardForm): boolean {
+  if (step === 1) return form.title.trim().length > 0;
+  return true;
+}
+
+// 整个向导可否提交：书名必填。
+export function canSubmit(form: WizardForm): boolean {
+  return form.title.trim().length > 0;
+}
+
+export function clampStep(step: number): number {
+  if (step < 1) return 1;
+  if (step > WIZARD_STEPS) return WIZARD_STEPS;
+  return step;
+}
+
+// 归一为后端请求体（snake_case，空值落 null/省略）。
+export function toCreateRequest(form: WizardForm): ProjectCreateRequest {
+  const trim = (s: string): string | null => {
+    const v = s.trim();
+    return v.length > 0 ? v : null;
+  };
+  return {
+    title: form.title.trim(),
+    genre: trim(form.genre),
+    logline: trim(form.logline),
+    premise: trim(form.premise),
+    theme: trim(form.theme),
+    selling_points: form.sellingPoints,
+    structure: trim(form.structure),
+  };
+}
diff --git a/packages/core/README.md b/packages/core/README.md
new file mode 100644
index 0000000..b346088
--- /dev/null
+++ b/packages/core/README.md
@@ -0,0 +1,2 @@
+# core (@llm/@backend) — 占位骨架
+领域核心(domain/状态机)、编排器(orchestrator/LangGraph 图)、记忆服务(memory)。Phase 1+ 由对应 owner 填充。
diff --git a/packages/core/pyproject.toml b/packages/core/pyproject.toml
new file mode 100644
index 0000000..579c724
--- /dev/null
+++ b/packages/core/pyproject.toml
@@ -0,0 +1,26 @@
+[project]
+name = "ww-core"
+version = "0.0.0"
+requires-python = ">=3.12"
+dependencies = [
+  "langgraph>=0.2.40",
+  "langgraph-checkpoint-postgres>=2.0",
+  "pydantic>=2.7",
+  "ww-shared",
+  "ww-config",
+  "ww-db",
+  "ww-llm-gateway",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["ww_core"]
+
+[tool.uv.sources]
+ww-shared = { workspace = true }
+ww-config = { workspace = true }
+ww-db = { workspace = true }
+ww-llm-gateway = { workspace = true }
diff --git a/packages/core/tests/fakes_orchestrator.py b/packages/core/tests/fakes_orchestrator.py
new file mode 100644
index 0000000..ea8ee51
--- /dev/null
+++ b/packages/core/tests/fakes_orchestrator.py
@@ -0,0 +1,122 @@
+"""编排器测试替身——mock 网关（`stream` / `run`）+ 内存审稿留痕 repo。
+
+绝对导入（无包目录下不能相对导入，见 gotchas 2026-06-18）：`from fakes_orchestrator import ...`。
+"""
+
+from __future__ import annotations
+
+import uuid
+from collections.abc import AsyncIterator
+from typing import Any
+
+from pydantic import BaseModel
+from ww_llm_gateway.types import (
+    Delta,
+    LlmRequest,
+    LlmResponse,
+    ServedBy,
+    Usage,
+)
+
+
+class FakeStreamGateway:
+    """按预设 token 列表吐 `Delta`，并记录收到的 `LlmRequest`（断言请求构造）。"""
+
+    def __init__(self, tokens: list[str]) -> None:
+        self._tokens = tokens
+        self.last_request: LlmRequest | None = None
+        self.call_count = 0
+
+    async def stream(self, req: LlmRequest) -> AsyncIterator[Delta]:
+        self.last_request = req
+        self.call_count += 1
+        for tok in self._tokens:
+            yield Delta(text=tok)
+
+
+class RaisingGateway:
+    """先吐几个 token，再抛指定异常——模拟流式中途失败（验 SSE error 归一）。"""
+
+    def __init__(self, tokens: list[str], exc: Exception) -> None:
+        self._tokens = tokens
+        self._exc = exc
+
+    async def stream(self, req: LlmRequest) -> AsyncIterator[Delta]:
+        for tok in self._tokens:
+            yield Delta(text=tok)
+        raise self._exc
+
+
+def _stub_usage() -> Usage:
+    return Usage(
+        provider="fake",
+        model="fake-analyst",
+        input_tokens=1,
+        output_tokens=1,
+        cost_minor=0,
+        currency="USD",
+    )
+
+
+class FakeRunGateway:
+    """按预设 `parsed` 实例回 `LlmResponse`，记录收到的请求（断言审稿请求构造）。"""
+
+    def __init__(self, parsed: BaseModel) -> None:
+        self._parsed = parsed
+        self.last_request: LlmRequest | None = None
+        self.call_count = 0
+
+    async def run(self, req: LlmRequest) -> LlmResponse:
+        self.last_request = req
+        self.call_count += 1
+        return LlmResponse(
+            text=self._parsed.model_dump_json(),
+            parsed=self._parsed,
+            usage=_stub_usage(),
+            served_by=ServedBy(provider="fake", model="fake-analyst"),
+        )
+
+
+class FailingRunGateway:
+    """`run` 抛指定异常——验「某审失败不阻塞其余」+ 未完成占位归一。"""
+
+    def __init__(self, exc: Exception) -> None:
+        self._exc = exc
+        self.call_count = 0
+
+    async def run(self, req: LlmRequest) -> LlmResponse:
+        self.call_count += 1
+        raise self._exc
+
+
+class FakeReviewRepo:
+    """内存审稿留痕 repo（实现 `ReviewRepo` Protocol 的 `record`）。"""
+
+    def __init__(self) -> None:
+        self.records: list[dict[str, Any]] = []
+
+    async def record(
+        self,
+        project_id: uuid.UUID,
+        chapter_no: int,
+        *,
+        chapter_version: int | None = None,
+        conflicts: list[dict[str, Any]],
+        foreshadow_sug: list[dict[str, Any]] | None = None,
+        style: dict[str, Any] | None = None,
+        pace: dict[str, Any] | None = None,
+        health_score: int | None = None,
+    ) -> Any:
+        entry = {
+            "id": uuid.uuid4(),
+            "project_id": project_id,
+            "chapter_no": chapter_no,
+            "chapter_version": chapter_version,
+            "conflicts": list(conflicts),
+            "foreshadow_sug": list(foreshadow_sug or []),
+            "style": style,
+            "pace": pace,
+            "health_score": health_score,
+        }
+        self.records.append(entry)
+        return entry
diff --git a/packages/core/tests/test_memory.py b/packages/core/tests/test_memory.py
new file mode 100644
index 0000000..a693c61
--- /dev/null
+++ b/packages/core/tests/test_memory.py
@@ -0,0 +1,270 @@
+"""T1.2 记忆服务单测——确定性选择 + 渲染 + 组装（ARCH §3.4/§5.3）。
+
+全部用内存 fake repo，无 DB；断言确定性、稳定块排序、无时间戳/UUID。
+"""
+
+from __future__ import annotations
+
+import uuid
+from dataclasses import dataclass, field
+
+from ww_core.domain.repositories import (
+    CharacterView,
+    DigestView,
+    ForeshadowView,
+    MemoryRepos,
+    OutlineView,
+    RuleView,
+    StyleView,
+    WorldEntityView,
+)
+from ww_core.memory import (
+    AssembledContext,
+    EntityKind,
+    SelectedEntity,
+    SelectionTrace,
+    assemble,
+    render_cards,
+    select_relevant_entities,
+)
+
+PROJECT = uuid.UUID("00000000-0000-0000-0000-000000000001")
+
+
+# ---- 内存 fake repositories ----
+
+
+@dataclass
+class FakeOutlineRepo:
+    rows: dict[int, OutlineView] = field(default_factory=dict)
+
+    async def get(self, project_id: uuid.UUID, chapter_no: int) -> OutlineView | None:
+        return self.rows.get(chapter_no)
+
+
+@dataclass
+class FakeCharacterRepo:
+    rows: list[CharacterView] = field(default_factory=list)
+
+    async def list_for_project(self, project_id: uuid.UUID) -> list[CharacterView]:
+        return list(self.rows)
+
+
+@dataclass
+class FakeWorldEntityRepo:
+    rows: list[WorldEntityView] = field(default_factory=list)
+
+    async def list_for_project(self, project_id: uuid.UUID) -> list[WorldEntityView]:
+        return list(self.rows)
+
+
+@dataclass
+class FakeDigestRepo:
+    rows: list[DigestView] = field(default_factory=list)
+
+    async def recent(self, project_id: uuid.UUID, k: int) -> list[DigestView]:
+        return sorted(self.rows, key=lambda d: d.chapter_no, reverse=True)[:k]
+
+
+@dataclass
+class FakeForeshadowRepo:
+    rows: list[ForeshadowView] = field(default_factory=list)
+
+    async def list_for_codes(self, project_id: uuid.UUID, codes: list[str]) -> list[ForeshadowView]:
+        wanted = set(codes)
+        return [r for r in self.rows if r.code in wanted]
+
+
+@dataclass
+class FakeStyleRepo:
+    row: StyleView | None = None
+
+    async def latest(self, project_id: uuid.UUID) -> StyleView | None:
+        return self.row
+
+
+@dataclass
+class FakeRulesRepo:
+    rows: list[RuleView] = field(default_factory=list)
+
+    async def all_for_project(self, project_id: uuid.UUID) -> list[RuleView]:
+        return list(self.rows)
+
+
+def build_repos(
+    *,
+    outline: dict[int, OutlineView] | None = None,
+    characters: list[CharacterView] | None = None,
+    world: list[WorldEntityView] | None = None,
+    digests: list[DigestView] | None = None,
+    foreshadows: list[ForeshadowView] | None = None,
+    style: StyleView | None = None,
+    rules: list[RuleView] | None = None,
+) -> MemoryRepos:
+    return MemoryRepos(
+        outline=FakeOutlineRepo(outline or {}),
+        character=FakeCharacterRepo(characters or []),
+        world_entity=FakeWorldEntityRepo(world or []),
+        digest=FakeDigestRepo(digests or []),
+        foreshadow=FakeForeshadowRepo(foreshadows or []),
+        style=FakeStyleRepo(style),
+        rules=FakeRulesRepo(rules or []),
+    )
+
+
+# ---- select_relevant_entities ----
+
+
+def char(name: str, role: str | None = None, latest_state: str | None = None) -> CharacterView:
+    return CharacterView(name=name, role=role, latest_state=latest_state)
+
+
+def world(name: str, type_: str = "设定", latest_state: str | None = None) -> WorldEntityView:
+    return WorldEntityView(type=type_, name=name, latest_state=latest_state)
+
+
+def test_selects_main_character_even_when_not_in_beats() -> None:
+    outline = OutlineView(volume=1, chapter_no=1, beats={"summary": "无人提及"})
+    trace = select_relevant_entities(
+        outline=outline,
+        characters=[char("林动", role="主角"), char("路人甲", role="配角")],
+        world_entities=[],
+        recent_digests=[],
+    )
+    names = {e.name for e in trace.selected}
+    assert "林动" in names
+    assert "路人甲" not in names
+    main = next(e for e in trace.selected if e.name == "林动")
+    assert "main_character" in main.reasons
+
+
+def test_selects_entity_named_in_beats() -> None:
+    outline = OutlineView(volume=1, chapter_no=2, beats={"entities": ["青檀"], "text": "青檀出场"})
+    trace = select_relevant_entities(
+        outline=outline,
+        characters=[char("青檀", role="配角")],
+        world_entities=[],
+        recent_digests=[],
+    )
+    assert any(e.name == "青檀" and "explicit_beat" in e.reasons for e in trace.selected)
+
+
+def test_selects_recent_digest_entity() -> None:
+    outline = OutlineView(volume=1, chapter_no=10, beats={})
+    digests = [DigestView(chapter_no=9, facts={"entities": ["玄机阁"]})]
+    trace = select_relevant_entities(
+        outline=outline,
+        characters=[],
+        world_entities=[world("玄机阁")],
+        recent_digests=digests,
+    )
+    assert any(e.name == "玄机阁" and "recent_digest" in e.reasons for e in trace.selected)
+
+
+def test_union_dedups_reasons_sorted() -> None:
+    outline = OutlineView(
+        volume=1,
+        chapter_no=3,
+        beats={"entities": ["林动"]},
+        foreshadow_windows=[{"code": "F1", "entities": ["林动"]}],
+    )
+    digests = [DigestView(chapter_no=2, facts={"entities": ["林动"]})]
+    trace = select_relevant_entities(
+        outline=outline,
+        characters=[char("林动", role="主角")],
+        world_entities=[],
+        recent_digests=digests,
+    )
+    selected = [e for e in trace.selected if e.name == "林动"]
+    assert len(selected) == 1  # 去重为一条
+    # 四个理由全命中且按规范顺序排列
+    assert selected[0].reasons == [
+        "explicit_beat",
+        "main_character",
+        "recent_digest",
+        "foreshadow_window",
+    ]
+
+
+# ---- render_cards ----
+
+
+def test_render_cards_deterministic_and_sorted() -> None:
+    selection = SelectionTrace(
+        selected=[
+            _selected("character", "乙"),
+            _selected("character", "甲"),
+        ]
+    )
+    chars = [char("乙", role="配角", latest_state="受伤"), char("甲", role="主角")]
+    out1 = render_cards(selection, chars, [])
+    out2 = render_cards(selection, chars, [])
+    assert out1 == out2  # 确定性
+    # 按名排序（Python 默认 codepoint 序：乙 U+4E59 < 甲 U+7532）
+    assert out1.index("乙") < out1.index("甲")
+
+
+# ---- assemble ----
+
+
+async def test_assemble_deterministic_no_timestamps() -> None:
+    repos = _full_repos()
+    ctx1 = await assemble(repos, PROJECT, 5)
+    ctx2 = await assemble(repos, PROJECT, 5)
+    assert isinstance(ctx1, AssembledContext)
+    assert ctx1.stable_core == ctx2.stable_core
+    assert ctx1.volatile == ctx2.volatile
+    # 无 UUID / 时间戳泄漏进缓存前缀
+    assert str(PROJECT) not in ctx1.stable_core
+    assert "T00:00" not in ctx1.stable_core
+
+
+async def test_latest_state_in_volatile_not_stable() -> None:
+    repos = _full_repos()
+    ctx = await assemble(repos, PROJECT, 5)
+    # latest_state 是易变态——必须在断点之后，不污染缓存前缀
+    assert "断左臂" in ctx.volatile
+    assert "断左臂" not in ctx.stable_core
+
+
+async def test_rule_merge_precedence() -> None:
+    rules = [
+        RuleView(level="global", content="全局规则"),
+        RuleView(level="project", content="作品规则"),
+        RuleView(level="genre", content="题材规则"),
+        RuleView(level="style", content="文风规则"),
+    ]
+    repos = build_repos(rules=rules)
+    ctx = await assemble(repos, PROJECT, 1)
+    # 四级都出现，且 global→genre→style→project 顺序（越具体越靠后/越优先）
+    g = ctx.stable_core.index("全局规则")
+    ge = ctx.stable_core.index("题材规则")
+    s = ctx.stable_core.index("文风规则")
+    p = ctx.stable_core.index("作品规则")
+    assert g < ge < s < p
+
+
+# ---- helpers ----
+
+
+def _selected(kind: EntityKind, name: str) -> SelectedEntity:
+    return SelectedEntity(kind=kind, name=name, reasons=["main_character"])
+
+
+def _full_repos() -> MemoryRepos:
+    return build_repos(
+        outline={
+            5: OutlineView(
+                volume=1,
+                chapter_no=5,
+                beats={"entities": ["林动"], "text": "林动突破"},
+                foreshadow_windows=[{"code": "F1", "entities": []}],
+            )
+        },
+        characters=[char("林动", role="主角", latest_state="断左臂")],
+        world=[world("元力", type_="力量体系")],
+        digests=[DigestView(chapter_no=4, facts={"entities": ["元力"]})],
+        foreshadows=[ForeshadowView(code="F1", title="神秘石符", status="OPEN", planted_at=1)],
+        style=StyleView(dimensions={"句长": "短句为主"}),
+        rules=[RuleView(level="global", content="全局规则")],
+    )
diff --git a/packages/core/tests/test_orchestrator.py b/packages/core/tests/test_orchestrator.py
new file mode 100644
index 0000000..fe42035
--- /dev/null
+++ b/packages/core/tests/test_orchestrator.py
@@ -0,0 +1,165 @@
+"""T1.3 编排器单测——write 节点 + SSE 归一 + 图编译（ARCH §5.2/§7.3）。
+
+全部注入 mock 网关，无真 LLM、无真 Postgres（图用 MemorySaver）。asyncio_mode=auto。
+"""
+
+from __future__ import annotations
+
+import uuid
+
+from fakes_orchestrator import FakeStreamGateway, RaisingGateway
+from langchain_core.runnables import RunnableConfig
+from langgraph.checkpoint.memory import MemorySaver
+from ww_core.orchestrator import (
+    EVENT_DONE,
+    EVENT_ERROR,
+    EVENT_TOKEN,
+    WRITE_NODE,
+    ChapterState,
+    build_write_graph,
+    build_write_request,
+    normalize_deltas,
+    stream_chapter_draft,
+    write_node,
+)
+from ww_shared import AppError, ErrorCode
+
+PROJECT = uuid.UUID("00000000-0000-0000-0000-000000000001")
+USER = uuid.UUID("00000000-0000-0000-0000-0000000000aa")
+
+STABLE = "## 世界观硬规则\n灵气可凝丹"
+VOLATILE = "## 本章节拍\n主角破境"
+
+
+# ---- build_write_request：纯函数、缓存断点、tier 不变量 ----
+
+
+def test_build_write_request_puts_stable_core_in_cached_system_block() -> None:
+    req = build_write_request(
+        stable_core=STABLE, volatile=VOLATILE, user_id=USER, project_id=PROJECT
+    )
+
+    assert req.tier == "writer"  # 不变量②：只传档位，不传 model
+    assert req.stream is True
+    assert len(req.system) == 1
+    assert req.system[0].text == STABLE
+    assert req.system[0].cache is True  # 不变量#9：稳定内核进缓存断点前
+    assert req.input == VOLATILE
+    assert req.scope.user_id == USER
+    assert req.scope.project_id == PROJECT
+
+
+# ---- stream_chapter_draft：转发网关 Delta、构造正确请求 ----
+
+
+async def test_stream_chapter_draft_yields_deltas_and_builds_request() -> None:
+    gateway = FakeStreamGateway(["灵", "气", "破", "境"])
+
+    texts = [
+        d.text
+        async for d in stream_chapter_draft(
+            gateway, stable_core=STABLE, volatile=VOLATILE, user_id=USER, project_id=PROJECT
+        )
+    ]
+
+    assert texts == ["灵", "气", "破", "境"]
+    assert gateway.call_count == 1
+    assert gateway.last_request is not None
+    assert gateway.last_request.tier == "writer"
+    assert gateway.last_request.system[0].cache is True
+
+
+# ---- write_node：累积草稿、不可变增量返回 ----
+
+
+async def test_write_node_accumulates_draft() -> None:
+    gateway = FakeStreamGateway(["第一段。", "第二段。"])
+    state: ChapterState = {
+        "project_id": PROJECT,
+        "chapter_no": 7,
+        "user_id": USER,
+        "stable_core": STABLE,
+        "volatile": VOLATILE,
+        "draft": "",
+    }
+
+    result = await write_node(state, gateway=gateway)
+
+    assert result == {"draft": "第一段。第二段。"}
+    assert state["draft"] == ""  # 不可变：未原地改入参
+
+
+# ---- SSE 归一：token / done ----
+
+
+async def test_normalize_deltas_emits_tokens_then_done() -> None:
+    gateway = FakeStreamGateway(["a", "bc", ""])  # 空 Delta 不应产 token
+    deltas = stream_chapter_draft(
+        gateway, stable_core=STABLE, volatile=VOLATILE, user_id=USER, project_id=PROJECT
+    )
+
+    events = [e async for e in normalize_deltas(deltas)]
+
+    assert [e.event for e in events] == [EVENT_TOKEN, EVENT_TOKEN, EVENT_DONE]
+    assert events[0].data == {"text": "a"}
+    assert events[1].data == {"text": "bc"}
+    assert events[-1].data == {"length": 3}  # "a"+"bc"
+
+
+# ---- SSE 归一：中途失败 → error 事件（不向上抛） ----
+
+
+async def test_normalize_deltas_emits_error_on_app_error() -> None:
+    gateway = RaisingGateway(["半句"], AppError(ErrorCode.LLM_UNAVAILABLE, "fallbacks exhausted"))
+    deltas = stream_chapter_draft(
+        gateway, stable_core=STABLE, volatile=VOLATILE, user_id=USER, project_id=PROJECT
+    )
+
+    events = [e async for e in normalize_deltas(deltas, request_id="req-1")]
+
+    assert [e.event for e in events] == [EVENT_TOKEN, EVENT_ERROR]  # 无 done
+    err = events[-1].data
+    assert err["code"] == ErrorCode.LLM_UNAVAILABLE
+    assert err["request_id"] == "req-1"
+
+
+async def test_normalize_deltas_maps_unexpected_error_to_internal() -> None:
+    gateway = RaisingGateway([], RuntimeError("boom"))
+    deltas = stream_chapter_draft(
+        gateway, stable_core=STABLE, volatile=VOLATILE, user_id=USER, project_id=PROJECT
+    )
+
+    events = [e async for e in normalize_deltas(deltas)]
+
+    assert [e.event for e in events] == [EVENT_ERROR]
+    assert events[0].data["code"] == ErrorCode.INTERNAL
+
+
+# ---- 图编译：装节点 + checkpointer（MemorySaver，不连真 Postgres） ----
+
+
+def test_build_write_graph_compiles_with_checkpointer() -> None:
+    gateway = FakeStreamGateway(["x"])
+
+    graph = build_write_graph(gateway, checkpointer=MemorySaver())
+
+    assert WRITE_NODE in graph.get_graph().nodes
+    assert graph.checkpointer is not None
+
+
+async def test_compiled_graph_runs_write_node_end_to_end() -> None:
+    gateway = FakeStreamGateway(["云", "深"])
+    graph = build_write_graph(gateway, checkpointer=MemorySaver())
+    config: RunnableConfig = {"configurable": {"thread_id": "t-1"}}
+    initial: ChapterState = {
+        "project_id": PROJECT,
+        "chapter_no": 1,
+        "user_id": USER,
+        "stable_core": STABLE,
+        "volatile": VOLATILE,
+        "draft": "",
+    }
+
+    final = await graph.ainvoke(initial, config=config)
+
+    assert final["draft"] == "云深"
diff --git a/packages/core/ww_core/__init__.py b/packages/core/ww_core/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/packages/core/ww_core/domain/chapter_repo.py b/packages/core/ww_core/domain/chapter_repo.py
new file mode 100644
index 0000000..c36058a
--- /dev/null
+++ b/packages/core/ww_core/domain/chapter_repo.py
@@ -0,0 +1,189 @@
+"""章节草稿/验收 Repository 协议 + SQLAlchemy async 实现（ARCH §3.3 / §5.5 / §7.2）。
+
+- 草稿：自动保存把流式产出的文本落到 `chapters`（status='draft'）。幂等：同一
+  (project_id, chapter_no) 的草稿重复保存覆盖同一行、版次固定为草稿版次。
+- 验收晋升（M2/T2.3）：`promote_to_accepted` 计算该章 max(version)+1，插入 status='accepted'
+  的**新行**（草稿行保留，符合「多版本行」§3.3）。真正「何时晋升 + 从终稿提炼 digest」
+  在 T2.4 验收事务里组合调用——本 repo 只提供确定性的版次计算 + 插入能力。
+"""
+
+from __future__ import annotations
+
+import uuid
+from dataclasses import dataclass
+from typing import Protocol
+
+from sqlalchemy import func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+from ww_db.models import Chapter
+
+# 草稿固定版次（accepted 版本提升属 M2，不在此处递增）。
+DRAFT_STATUS = "draft"
+DRAFT_VERSION = 1
+ACCEPTED_STATUS = "accepted"
+
+
+@dataclass(frozen=True)
+class ChapterDraftView:
+    """只读章节草稿快照（snake_case）。"""
+
+    project_id: uuid.UUID
+    chapter_no: int
+    volume: int
+    content: str
+    status: str
+    version: int
+
+
+@dataclass(frozen=True)
+class ChapterView:
+    """只读章节版本快照（任意 status/version；验收晋升返回此形）。"""
+
+    project_id: uuid.UUID
+    chapter_no: int
+    volume: int
+    content: str
+    status: str
+    version: int
+
+
+class ChapterRepo(Protocol):
+    """章节草稿读写 + 验收晋升接口（按 project_id 隔离）。"""
+
+    async def save_draft(
+        self, project_id: uuid.UUID, chapter_no: int, *, text: str, volume: int = 1
+    ) -> ChapterDraftView: ...
+
+    async def get_draft(
+        self, project_id: uuid.UUID, chapter_no: int
+    ) -> ChapterDraftView | None: ...
+
+    async def max_version(self, project_id: uuid.UUID, chapter_no: int) -> int:
+        """该章现有最大 version（无任何行则 0）——验收取 max+1。"""
+        ...
+
+    async def promote_to_accepted(
+        self, project_id: uuid.UUID, chapter_no: int, *, content: str, volume: int = 1
+    ) -> ChapterView:
+        """从终稿晋升：插入 status='accepted'、version=max+1 的新行（草稿行保留）。"""
+        ...
+
+    async def latest_accepted(self, project_id: uuid.UUID, chapter_no: int) -> ChapterView | None:
+        """该章最新（最高 version）的 accepted 版本，无则 None。"""
+        ...
+
+
+def _to_view(row: Chapter) -> ChapterDraftView:
+    return ChapterDraftView(
+        project_id=row.project_id,
+        chapter_no=row.chapter_no,
+        volume=row.volume,
+        content=row.content or "",
+        status=row.status,
+        version=row.version,
+    )
+
+
+def _to_chapter_view(row: Chapter) -> ChapterView:
+    return ChapterView(
+        project_id=row.project_id,
+        chapter_no=row.chapter_no,
+        volume=row.volume,
+        content=row.content or "",
+        status=row.status,
+        version=row.version,
+    )
+
+
+class SqlChapterRepo:
+    """SQLAlchemy 实现：草稿幂等 upsert（显式 read-modify-write，版次固定）。"""
+
+    def __init__(self, session: AsyncSession) -> None:
+        self._s = session
+
+    async def _find_draft(self, project_id: uuid.UUID, chapter_no: int) -> Chapter | None:
+        return (
+            await self._s.execute(
+                select(Chapter).where(
+                    Chapter.project_id == project_id,
+                    Chapter.chapter_no == chapter_no,
+                    Chapter.version == DRAFT_VERSION,
+                )
+            )
+        ).scalar_one_or_none()
+
+    async def save_draft(
+        self, project_id: uuid.UUID, chapter_no: int, *, text: str, volume: int = 1
+    ) -> ChapterDraftView:
+        # 唯一约束 (project_id, chapter_no, version)；草稿版次固定 → 覆盖同一行，幂等。
+        existing = await self._find_draft(project_id, chapter_no)
+        if existing is None:
+            row = Chapter(
+                project_id=project_id,
+                volume=volume,
+                chapter_no=chapter_no,
+                content=text,
+                status=DRAFT_STATUS,
+                version=DRAFT_VERSION,
+            )
+            self._s.add(row)
+            await self._s.commit()
+            await self._s.refresh(row)
+            return _to_view(row)
+        existing.content = text
+        existing.status = DRAFT_STATUS
+        await self._s.commit()
+        await self._s.refresh(existing)
+        return _to_view(existing)
+
+    async def get_draft(self, project_id: uuid.UUID, chapter_no: int) -> ChapterDraftView | None:
+        row = await self._find_draft(project_id, chapter_no)
+        if row is None:
+            return None
+        return _to_view(row)
+
+    async def max_version(self, project_id: uuid.UUID, chapter_no: int) -> int:
+        result = await self._s.execute(
+            select(func.max(Chapter.version)).where(
+                Chapter.project_id == project_id,
+                Chapter.chapter_no == chapter_no,
+            )
+        )
+        return result.scalar_one_or_none() or 0
+
+    async def promote_to_accepted(
+        self, project_id: uuid.UUID, chapter_no: int, *, content: str, volume: int = 1
+    ) -> ChapterView:
+        # 取 max+1 晋升新 accepted 行；草稿行保留（多版本行，§3.3）。
+        # 唯一约束 (project_id, chapter_no, version) 由 DB 强制。
+        # **只 flush 不 commit**：验收事务（T2.4）把晋升+digest+裁决留痕作为单事务提交。
+        next_version = (await self.max_version(project_id, chapter_no)) + 1
+        row = Chapter(
+            project_id=project_id,
+            volume=volume,
+            chapter_no=chapter_no,
+            content=content,
+            status=ACCEPTED_STATUS,
+            version=next_version,
+        )
+        self._s.add(row)
+        await self._s.flush()
+        await self._s.refresh(row)
+        return _to_chapter_view(row)
+
+    async def latest_accepted(self, project_id: uuid.UUID, chapter_no: int) -> ChapterView | None:
+        row = (
+            await self._s.execute(
+                select(Chapter)
+                .where(
+                    Chapter.project_id == project_id,
+                    Chapter.chapter_no == chapter_no,
+                    Chapter.status == ACCEPTED_STATUS,
+                )
+                .order_by(Chapter.version.desc())
+                .limit(1)
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            return None
+        return _to_chapter_view(row)
diff --git a/packages/core/ww_core/domain/project_repo.py b/packages/core/ww_core/domain/project_repo.py
new file mode 100644
index 0000000..0b53961
--- /dev/null
+++ b/packages/core/ww_core/domain/project_repo.py
@@ -0,0 +1,107 @@
+"""项目（立项）Repository 协议 + SQLAlchemy async 实现（ARCH §3.5 / §7.2）。
+
+记忆/记账依赖这里的 **Protocol**，单测注入内存 fake，运行时注入 SQLAlchemy 实现。
+统一按 owner_id 过滤（单用户原型 stub，多租户化时由认证主体替换）。
+视图是与 ORM 解耦的只读 Pydantic 快照——路由不碰 SQLAlchemy 行。
+"""
+
+from __future__ import annotations
+
+import uuid
+from dataclasses import dataclass, field
+from typing import Any, Protocol
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+from ww_db.models import Project
+
+
+@dataclass(frozen=True)
+class ProjectView:
+    """只读项目快照（snake_case）。"""
+
+    id: uuid.UUID
+    title: str
+    genre: str | None = None
+    logline: str | None = None
+    premise: str | None = None
+    theme: str | None = None
+    selling_points: list[Any] = field(default_factory=list)
+    structure: str | None = None
+
+
+@dataclass(frozen=True)
+class ProjectCreate:
+    """立项向导写入字段（owner_id 由服务层补 stub）。"""
+
+    title: str
+    genre: str | None = None
+    logline: str | None = None
+    premise: str | None = None
+    theme: str | None = None
+    selling_points: list[Any] = field(default_factory=list)
+    structure: str | None = None
+
+
+class ProjectRepo(Protocol):
+    """项目读写接口（按 owner_id 隔离）。"""
+
+    async def create(self, owner_id: uuid.UUID, data: ProjectCreate) -> ProjectView: ...
+
+    async def list_for_owner(self, owner_id: uuid.UUID) -> list[ProjectView]: ...
+
+    async def get(self, owner_id: uuid.UUID, project_id: uuid.UUID) -> ProjectView | None: ...
+
+
+def _to_view(row: Project) -> ProjectView:
+    return ProjectView(
+        id=row.id,
+        title=row.title,
+        genre=row.genre,
+        logline=row.logline,
+        premise=row.premise,
+        theme=row.theme,
+        selling_points=list(row.selling_points or []),
+        structure=row.structure,
+    )
+
+
+class SqlProjectRepo:
+    """SQLAlchemy 实现：写 `projects`，按 owner_id 过滤读取。"""
+
+    def __init__(self, session: AsyncSession) -> None:
+        self._s = session
+
+    async def create(self, owner_id: uuid.UUID, data: ProjectCreate) -> ProjectView:
+        row = Project(
+            owner_id=owner_id,
+            title=data.title,
+            genre=data.genre,
+            logline=data.logline,
+            premise=data.premise,
+            theme=data.theme,
+            selling_points=list(data.selling_points),
+            structure=data.structure,
+        )
+        self._s.add(row)
+        await self._s.commit()
+        await self._s.refresh(row)
+        return _to_view(row)
+
+    async def list_for_owner(self, owner_id: uuid.UUID) -> list[ProjectView]:
+        rows = (
+            await self._s.execute(
+                select(Project).where(Project.owner_id == owner_id).order_by(Project.created_at)
+            )
+        ).scalars()
+        return [_to_view(r) for r in rows]
+
+    async def get(self, owner_id: uuid.UUID, project_id: uuid.UUID) -> ProjectView | None:
+        row = (
+            await self._s.execute(
+                select(Project).where(Project.owner_id == owner_id, Project.id == project_id)
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            return None
+        return _to_view(row)
diff --git a/packages/core/ww_core/domain/repositories.py b/packages/core/ww_core/domain/repositories.py
new file mode 100644
index 0000000..af0e150
--- /dev/null
+++ b/packages/core/ww_core/domain/repositories.py
@@ -0,0 +1,132 @@
+"""领域视图 + Repository 协议（ARCH §3.5）。
+
+记忆服务依赖这些 **Protocol** 而非具体实现（依赖倒置）：单测注入内存 fake，
+运行时注入 SQLAlchemy 实现（见 ww_core.memory.sql_repositories）。
+
+视图是与 ORM 解耦的只读 Pydantic 快照——记忆逻辑不碰 SQLAlchemy 行，
+保证确定性选择可纯函数化、可单测。所有读取统一按 project_id 过滤（数据隔离）。
+"""
+
+from __future__ import annotations
+
+import uuid
+from dataclasses import dataclass
+from typing import Any, Protocol
+
+from pydantic import BaseModel, Field
+
+# ---- 只读视图（snake_case，frozen 防止意外突变）----
+
+
+class OutlineView(BaseModel):
+    model_config = {"frozen": True}
+
+    volume: int
+    chapter_no: int
+    beats: dict[str, Any] = Field(default_factory=dict)
+    foreshadow_windows: list[dict[str, Any]] = Field(default_factory=list)
+
+
+class CharacterView(BaseModel):
+    model_config = {"frozen": True}
+
+    name: str
+    role: str | None = None
+    traits: dict[str, Any] | None = None
+    appearance: str | None = None
+    motive: str | None = None
+    backstory: str | None = None
+    arc: dict[str, Any] | None = None
+    speech_tics: dict[str, Any] | None = None
+    tags: list[Any] = Field(default_factory=list)
+    relations: list[Any] = Field(default_factory=list)
+    first_chapter: int | None = None
+    latest_state: str | None = None
+
+
+class WorldEntityView(BaseModel):
+    model_config = {"frozen": True}
+
+    type: str
+    name: str
+    rules: dict[str, Any] | None = None
+    first_chapter: int | None = None
+    latest_state: str | None = None
+
+
+class DigestView(BaseModel):
+    model_config = {"frozen": True}
+
+    chapter_no: int
+    facts: dict[str, Any] = Field(default_factory=dict)
+
+
+class ForeshadowView(BaseModel):
+    model_config = {"frozen": True}
+
+    code: str
+    title: str
+    status: str
+    planted_at: int | None = None
+    expected_close_from: int | None = None
+    expected_close_to: int | None = None
+    content: str | None = None
+
+
+class StyleView(BaseModel):
+    model_config = {"frozen": True}
+
+    dimensions: dict[str, Any] = Field(default_factory=dict)
+
+
+class RuleView(BaseModel):
+    model_config = {"frozen": True}
+
+    level: str
+    content: str
+
+
+# ---- Repository 协议（async；统一按 project_id 过滤）----
+
+
+class OutlineRepo(Protocol):
+    async def get(self, project_id: uuid.UUID, chapter_no: int) -> OutlineView | None: ...
+
+
+class CharacterRepo(Protocol):
+    async def list_for_project(self, project_id: uuid.UUID) -> list[CharacterView]: ...
+
+
+class WorldEntityRepo(Protocol):
+    async def list_for_project(self, project_id: uuid.UUID) -> list[WorldEntityView]: ...
+
+
+class DigestRepo(Protocol):
+    async def recent(self, project_id: uuid.UUID, k: int) -> list[DigestView]: ...
+
+
+class ForeshadowRepo(Protocol):
+    async def list_for_codes(
+        self, project_id: uuid.UUID, codes: list[str]
+    ) -> list[ForeshadowView]: ...
+
+
+class StyleRepo(Protocol):
+    async def latest(self, project_id: uuid.UUID) -> StyleView | None: ...
+
+
+class RulesRepo(Protocol):
+    async def all_for_project(self, project_id: uuid.UUID) -> list[RuleView]: ...
+
+
+@dataclass(frozen=True)
+class MemoryRepos:
+    """记忆服务所需的 7 个 repo 的依赖捆绑（注入点）。"""
+
+    outline: OutlineRepo
+    character: CharacterRepo
+    world_entity: WorldEntityRepo
+    digest: DigestRepo
+    foreshadow: ForeshadowRepo
+    style: StyleRepo
+    rules: RulesRepo
diff --git a/packages/core/ww_core/memory/__init__.py b/packages/core/ww_core/memory/__init__.py
new file mode 100644
index 0000000..3035041
--- /dev/null
+++ b/packages/core/ww_core/memory/__init__.py
@@ -0,0 +1,28 @@
+"""记忆服务（C5）——确定性选择 + 卡片渲染 + prompt 组装（ARCH §3.4/§5.3）。"""
+
+from __future__ import annotations
+
+from .assemble import assemble, merge_rules
+from .render import render_cards
+from .selection import MAIN_ROLES, RECENT_DIGEST_COUNT, select_relevant_entities
+from .types import (
+    AssembledContext,
+    EntityKind,
+    SelectedEntity,
+    SelectionReason,
+    SelectionTrace,
+)
+
+__all__ = [
+    "MAIN_ROLES",
+    "RECENT_DIGEST_COUNT",
+    "AssembledContext",
+    "EntityKind",
+    "SelectedEntity",
+    "SelectionReason",
+    "SelectionTrace",
+    "assemble",
+    "merge_rules",
+    "render_cards",
+    "select_relevant_entities",
+]
diff --git a/packages/core/ww_core/memory/assemble.py b/packages/core/ww_core/memory/assemble.py
new file mode 100644
index 0000000..e38abe3
--- /dev/null
+++ b/packages/core/ww_core/memory/assemble.py
@@ -0,0 +1,129 @@
+"""记忆注入与 prompt 组装（ARCH §5.3）。
+
+assemble = 确定性选择 → 渲染卡片 → 序列化为 stable_core(缓存前缀) + volatile(断点后)。
+稳定内核只放真正定型、无 latest_state 的内容；latest_state/本章大纲等易变内容入 volatile。
+全程排序、无时间戳/UUID，保证缓存前缀字节稳定（不变量 #9）。
+"""
+
+from __future__ import annotations
+
+import json
+import uuid
+from typing import Any
+
+from ww_core.domain.repositories import (
+    CharacterView,
+    DigestView,
+    ForeshadowView,
+    MemoryRepos,
+    OutlineView,
+    RuleView,
+    StyleView,
+    WorldEntityView,
+)
+
+from .render import render_cards
+from .selection import MAIN_ROLES, RECENT_DIGEST_COUNT, select_relevant_entities
+from .types import AssembledContext
+
+# 四级规则合并优先级：global → genre → style → project（越具体越靠后/越优先）
+_LEVEL_RANK: dict[str, int] = {"global": 0, "genre": 1, "style": 2, "project": 3}
+
+
+def _ser(value: Any) -> str:
+    return json.dumps(value, ensure_ascii=False, sort_keys=True)
+
+
+def merge_rules(rules: list[RuleView]) -> list[RuleView]:
+    """按 global→genre→style→project 排序（同级按内容排序，确定性）。"""
+    return sorted(rules, key=lambda r: (_LEVEL_RANK.get(r.level, 99), r.level, r.content))
+
+
+def _section(title: str, body: str) -> str | None:
+    return f"## {title}\n{body}" if body else None
+
+
+def _build_stable(
+    world_entities: list[WorldEntityView],
+    main_characters: list[CharacterView],
+    style: StyleView | None,
+    rules: list[RuleView],
+) -> str:
+    world_lines = [
+        f"【世界规则】{w.name}：{_ser(w.rules)}"
+        for w in sorted(world_entities, key=lambda w: w.name)
+        if w.rules
+    ]
+    # 定型主角：只放稳定字段，绝不含 latest_state（易变）
+    char_lines = [
+        f"【主角】{c.name}｜定位：{c.role or '未定'}"
+        + (f"｜动机：{c.motive}" if c.motive else "")
+        + (f"｜背景：{c.backstory}" if c.backstory else "")
+        for c in sorted(main_characters, key=lambda c: c.name)
+    ]
+    rule_lines = [f"[{r.level}] {r.content}" for r in merge_rules(rules)]
+
+    sections = [
+        _section("世界观硬规则", "\n".join(world_lines)),
+        _section("定型主角", "\n".join(char_lines)),
+        _section("文风指纹", _ser(style.dimensions) if style else ""),
+        _section("规则", "\n".join(rule_lines)),
+    ]
+    return "\n\n".join(s for s in sections if s)
+
+
+def _build_volatile(
+    cards: str,
+    foreshadows: list[ForeshadowView],
+    digests: list[DigestView],
+    beats: dict[str, Any],
+) -> str:
+    fore_lines = [
+        f"【伏笔】{f.code} {f.title} [{f.status}]"
+        for f in sorted(foreshadows, key=lambda f: f.code)
+    ]
+    digest_lines = [
+        f"第{d.chapter_no}章：{_ser(d.facts)}" for d in sorted(digests, key=lambda d: d.chapter_no)
+    ]
+    sections = [
+        _section("本章注入卡片", cards),
+        _section("相关伏笔窗口", "\n".join(fore_lines)),
+        _section("近况摘要", "\n".join(digest_lines)),
+        _section("本章节拍", _ser(beats) if beats else ""),
+    ]
+    return "\n\n".join(s for s in sections if s)
+
+
+async def assemble(
+    repos: MemoryRepos,
+    project_id: uuid.UUID,
+    chapter_no: int,
+    recent_k: int = RECENT_DIGEST_COUNT,
+) -> AssembledContext:
+    """组装本章 prompt 上下文（确定性、可缓存）。"""
+    outline = await repos.outline.get(project_id, chapter_no) or OutlineView(
+        volume=0, chapter_no=chapter_no
+    )
+    characters = await repos.character.list_for_project(project_id)
+    world_entities = await repos.world_entity.list_for_project(project_id)
+    recent_digests = await repos.digest.recent(project_id, recent_k)
+
+    selection = select_relevant_entities(
+        outline=outline,
+        characters=characters,
+        world_entities=world_entities,
+        recent_digests=recent_digests,
+    )
+
+    codes = [str(w["code"]) for w in outline.foreshadow_windows if w.get("code")]
+    foreshadows = await repos.foreshadow.list_for_codes(project_id, codes)
+    style = await repos.style.latest(project_id)
+    rules = await repos.rules.all_for_project(project_id)
+
+    main_characters = [c for c in characters if (c.role or "") in MAIN_ROLES]
+
+    stable_core = _build_stable(world_entities, main_characters, style, rules)
+    cards = render_cards(selection, characters, world_entities)
+    volatile = _build_volatile(cards, foreshadows, recent_digests, outline.beats)
+
+    return AssembledContext(stable_core=stable_core, volatile=volatile, selection=selection)
diff --git a/packages/core/ww_core/memory/render.py b/packages/core/ww_core/memory/render.py
new file mode 100644
index 0000000..c75351d
--- /dev/null
+++ b/packages/core/ww_core/memory/render.py
@@ -0,0 +1,62 @@
+"""把选中实体渲染成「角色/设定卡」文本块（ARCH §3.4 渲染为卡片）。
+
+卡片含 latest_state（易变态）——故渲染产物归入 volatile，不进缓存前缀。
+确定性：按 SelectionTrace 既定顺序渲染；字段缺省则略过该行。
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from ww_core.domain.repositories import CharacterView, WorldEntityView
+
+from .types import SelectionTrace
+
+
+def _serialize(value: Any) -> str:
+    return json.dumps(value, ensure_ascii=False, sort_keys=True)
+
+
+def _line(label: str, value: str | None) -> str | None:
+    if value is None or value == "":
+        return None
+    return f"- {label}：{value}"
+
+
+def _character_card(c: CharacterView) -> str:
+    head = f"【角色】{c.name}｜定位：{c.role or '未定'}"
+    lines = [
+        _line("外貌", c.appearance),
+        _line("动机", c.motive),
+        _line("说话习惯", _serialize(c.speech_tics) if c.speech_tics else None),
+        _line("当前状态", c.latest_state),
+    ]
+    return "\n".join([head, *[ln for ln in lines if ln]])
+
+
+def _world_card(w: WorldEntityView) -> str:
+    head = f"【设定·{w.type}】{w.name}"
+    lines = [
+        _line("规则", _serialize(w.rules) if w.rules else None),
+        _line("当前状态", w.latest_state),
+    ]
+    return "\n".join([head, *[ln for ln in lines if ln]])
+
+
+def render_cards(
+    selection: SelectionTrace,
+    characters: list[CharacterView],
+    world_entities: list[WorldEntityView],
+) -> str:
+    """按选择顺序渲染卡片；找不到底层行的选中项跳过。"""
+    char_by_name = {c.name: c for c in characters}
+    world_by_name = {w.name: w for w in world_entities}
+
+    cards: list[str] = []
+    for e in sorted(selection.selected, key=lambda e: (e.kind, e.name)):
+        if e.kind == "character" and e.name in char_by_name:
+            cards.append(_character_card(char_by_name[e.name]))
+        elif e.kind == "world_entity" and e.name in world_by_name:
+            cards.append(_world_card(world_by_name[e.name]))
+    return "\n\n".join(cards)
diff --git a/packages/core/ww_core/memory/selection.py b/packages/core/ww_core/memory/selection.py
new file mode 100644
index 0000000..4e6919e
--- /dev/null
+++ b/packages/core/ww_core/memory/selection.py
@@ -0,0 +1,112 @@
+"""确定性记忆选择（ARCH §3.4）——无向量，纯函数、可单测、可调试。
+
+选择来源（并集去重）：①显式引用(beats 点名) ②主角常驻 ③近况实体(近 N 章摘要)
+④命中本章伏笔窗口。pg_trgm 按名匹配兜底属 P-later，此处不做。
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from ww_core.domain.repositories import CharacterView, DigestView, OutlineView, WorldEntityView
+
+from .types import EntityKind, SelectedEntity, SelectionReason, SelectionTrace
+
+# 视为常驻主角的 role 取值
+MAIN_ROLES: frozenset[str] = frozenset({"主角", "核心", "protagonist", "lead"})
+# 近况实体回看的章数（ARCH §3.4 默认 3–5）
+RECENT_DIGEST_COUNT = 5
+# 选择结果排序时的 kind 次序
+_KIND_RANK: dict[EntityKind, int] = {"character": 0, "world_entity": 1}
+_REASON_RANK: dict[SelectionReason, int] = {
+    "explicit_beat": 0,
+    "main_character": 1,
+    "recent_digest": 2,
+    "foreshadow_window": 3,
+}
+
+
+def _flatten_strings(obj: Any) -> list[str]:
+    """递归收集任意 JSON 结构里的字符串叶子（确定性、用于按名匹配）。"""
+    if isinstance(obj, str):
+        return [obj]
+    if isinstance(obj, dict):
+        out: list[str] = []
+        for key in sorted(obj):
+            out.extend(_flatten_strings(obj[key]))
+        return out
+    if isinstance(obj, (list, tuple)):
+        out = []
+        for item in obj:
+            out.extend(_flatten_strings(item))
+        return out
+    return []
+
+
+def _explicit_names(beats: dict[str, Any]) -> tuple[set[str], str]:
+    """返回 (显式列出的实体名集合, beats 扁平文本) 供按名匹配。"""
+    listed = {str(n) for n in beats.get("entities", []) if isinstance(n, str)}
+    text = "\n".join(_flatten_strings({k: v for k, v in beats.items() if k != "entities"}))
+    return listed, text
+
+
+def _digest_names(digests: list[DigestView]) -> tuple[set[str], str]:
+    listed: set[str] = set()
+    texts: list[str] = []
+    for d in digests:
+        listed |= {str(n) for n in d.facts.get("entities", []) if isinstance(n, str)}
+        texts.extend(_flatten_strings({k: v for k, v in d.facts.items() if k != "entities"}))
+    return listed, "\n".join(texts)
+
+
+def _window_entity_names(outline: OutlineView) -> set[str]:
+    names: set[str] = set()
+    for win in outline.foreshadow_windows:
+        for n in win.get("entities", []):
+            if isinstance(n, str):
+                names.add(n)
+    return names
+
+
+def _matches(name: str, listed: set[str], text: str) -> bool:
+    return name in listed or (name in text)
+
+
+def select_relevant_entities(
+    *,
+    outline: OutlineView,
+    characters: list[CharacterView],
+    world_entities: list[WorldEntityView],
+    recent_digests: list[DigestView],
+) -> SelectionTrace:
+    """确定性并集选择 → SelectionTrace（每实体带去重、排序的入选理由）。"""
+    beat_listed, beat_text = _explicit_names(outline.beats)
+    digest_listed, digest_text = _digest_names(recent_digests)
+    window_names = _window_entity_names(outline)
+
+    selected: list[SelectedEntity] = []
+
+    def reasons_for(name: str, *, is_main: bool) -> list[SelectionReason]:
+        reasons: list[SelectionReason] = []
+        if _matches(name, beat_listed, beat_text):
+            reasons.append("explicit_beat")
+        if is_main:
+            reasons.append("main_character")
+        if _matches(name, digest_listed, digest_text):
+            reasons.append("recent_digest")
+        if name in window_names:
+            reasons.append("foreshadow_window")
+        return sorted(set(reasons), key=lambda r: _REASON_RANK[r])
+
+    for c in characters:
+        reasons = reasons_for(c.name, is_main=(c.role or "") in MAIN_ROLES)
+        if reasons:
+            selected.append(SelectedEntity(kind="character", name=c.name, reasons=reasons))
+
+    for w in world_entities:
+        reasons = reasons_for(w.name, is_main=False)
+        if reasons:
+            selected.append(SelectedEntity(kind="world_entity", name=w.name, reasons=reasons))
+
+    selected.sort(key=lambda e: (_KIND_RANK[e.kind], e.name))
+    return SelectionTrace(selected=selected)
diff --git a/packages/core/ww_core/memory/sql_repositories.py b/packages/core/ww_core/memory/sql_repositories.py
new file mode 100644
index 0000000..58a2f04
--- /dev/null
+++ b/packages/core/ww_core/memory/sql_repositories.py
@@ -0,0 +1,190 @@
+"""Repository 协议的 SQLAlchemy async 实现（ARCH §3.5）。
+
+ORM 行 → 只读视图的映射层；统一按 project_id 过滤（数据隔离，多租户化时此层加 owner_id）。
+单测用内存 fake（见 tests），此实现由 mypy 静态校验、运行时（T1.4）注入真实 session。
+"""
+
+from __future__ import annotations
+
+import uuid
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+from ww_db.models import (
+    ChapterDigest,
+    Character,
+    Foreshadow,
+    Outline,
+    Rule,
+    StyleFingerprint,
+    WorldEntity,
+)
+
+from ww_core.domain.repositories import (
+    CharacterView,
+    DigestView,
+    ForeshadowView,
+    MemoryRepos,
+    OutlineView,
+    RuleView,
+    StyleView,
+    WorldEntityView,
+)
+
+
+class SqlOutlineRepo:
+    def __init__(self, session: AsyncSession) -> None:
+        self._s = session
+
+    async def get(self, project_id: uuid.UUID, chapter_no: int) -> OutlineView | None:
+        row = (
+            await self._s.execute(
+                select(Outline).where(
+                    Outline.project_id == project_id, Outline.chapter_no == chapter_no
+                )
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            return None
+        return OutlineView(
+            volume=row.volume,
+            chapter_no=row.chapter_no,
+            beats=row.beats or {},
+            foreshadow_windows=list(row.foreshadow_windows or []),
+        )
+
+
+class SqlCharacterRepo:
+    def __init__(self, session: AsyncSession) -> None:
+        self._s = session
+
+    async def list_for_project(self, project_id: uuid.UUID) -> list[CharacterView]:
+        rows = (
+            await self._s.execute(select(Character).where(Character.project_id == project_id))
+        ).scalars()
+        return [
+            CharacterView(
+                name=r.name,
+                role=r.role,
+                traits=r.traits,
+                appearance=r.appearance,
+                motive=r.motive,
+                backstory=r.backstory,
+                arc=r.arc,
+                speech_tics=r.speech_tics,
+                tags=list(r.tags or []),
+                relations=list(r.relations or []),
+                first_chapter=r.first_chapter,
+                latest_state=r.latest_state,
+            )
+            for r in rows
+        ]
+
+
+class SqlWorldEntityRepo:
+    def __init__(self, session: AsyncSession) -> None:
+        self._s = session
+
+    async def list_for_project(self, project_id: uuid.UUID) -> list[WorldEntityView]:
+        rows = (
+            await self._s.execute(select(WorldEntity).where(WorldEntity.project_id == project_id))
+        ).scalars()
+        return [
+            WorldEntityView(
+                type=r.type,
+                name=r.name,
+                rules=r.rules,
+                first_chapter=r.first_chapter,
+                latest_state=r.latest_state,
+            )
+            for r in rows
+        ]
+
+
+class SqlDigestRepo:
+    def __init__(self, session: AsyncSession) -> None:
+        self._s = session
+
+    async def recent(self, project_id: uuid.UUID, k: int) -> list[DigestView]:
+        rows = (
+            await self._s.execute(
+                select(ChapterDigest)
+                .where(ChapterDigest.project_id == project_id)
+                .order_by(ChapterDigest.chapter_no.desc())
+                .limit(k)
+            )
+        ).scalars()
+        return [DigestView(chapter_no=r.chapter_no, facts=r.facts) for r in rows]
+
+
+class SqlForeshadowRepo:
+    def __init__(self, session: AsyncSession) -> None:
+        self._s = session
+
+    async def list_for_codes(self, project_id: uuid.UUID, codes: list[str]) -> list[ForeshadowView]:
+        if not codes:
+            return []
+        rows = (
+            await self._s.execute(
+                select(Foreshadow).where(
+                    Foreshadow.project_id == project_id, Foreshadow.code.in_(codes)
+                )
+            )
+        ).scalars()
+        return [
+            ForeshadowView(
+                code=r.code,
+                title=r.title,
+                status=r.status,
+                planted_at=r.planted_at,
+                expected_close_from=r.expected_close_from,
+                expected_close_to=r.expected_close_to,
+                content=r.content,
+            )
+            for r in rows
+        ]
+
+
+class SqlStyleRepo:
+    def __init__(self, session: AsyncSession) -> None:
+        self._s = session
+
+    async def latest(self, project_id: uuid.UUID) -> StyleView | None:
+        row = (
+            await self._s.execute(
+                select(StyleFingerprint)
+                .where(StyleFingerprint.project_id == project_id)
+                .order_by(StyleFingerprint.version.desc())
+                .limit(1)
+            )
+        ).scalar_one_or_none()
+        if row is None:
+            return None
+        return StyleView(dimensions=row.dimensions_json)
+
+
+class SqlRulesRepo:
+    def __init__(self, session: AsyncSession) -> None:
+        self._s = session
+
+    async def all_for_project(self, project_id: uuid.UUID) -> list[RuleView]:
+        # 全局规则(project_id 为空) + 本作品规则——合并优先级在 assemble.merge_rules 处理
+        rows = (
+            await self._s.execute(
+                select(Rule).where((Rule.project_id == project_id) | (Rule.project_id.is_(None)))
+            )
+        ).scalars()
+        return [RuleView(level=r.level, content=r.content) for r in rows]
+
+
+def sql_memory_repos(session: AsyncSession) -> MemoryRepos:
+    """用一个 AsyncSession 装配全部 7 个 SQLAlchemy repo（T1.4 注入点）。"""
+    return MemoryRepos(
+        outline=SqlOutlineRepo(session),
+        character=SqlCharacterRepo(session),
+        world_entity=SqlWorldEntityRepo(session),
+        digest=SqlDigestRepo(session),
+        foreshadow=SqlForeshadowRepo(session),
+        style=SqlStyleRepo(session),
+        rules=SqlRulesRepo(session),
+    )
diff --git a/packages/core/ww_core/memory/types.py b/packages/core/ww_core/memory/types.py
new file mode 100644
index 0000000..91cd8e2
--- /dev/null
+++ b/packages/core/ww_core/memory/types.py
@@ -0,0 +1,60 @@
+"""记忆服务输出契约（C5）——AssembledContext / SelectionTrace。
+
+决策（memory/decisions.md 2026-06-18）：assemble 返回中性序列化文本，
+不返回 gateway 的 LlmRequest——记忆服务只经 DB 通信，不依赖网关类型。
+write 节点(T1.3)再据此构造 LlmRequest(system=[Block(stable_core, cache=True)], input=volatile)。
+"""
+
+from __future__ import annotations
+
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+EntityKind = Literal["character", "world_entity"]
+SelectionReason = Literal[
+    "explicit_beat",  # 本章大纲 beats 点名
+    "main_character",  # 主角/核心常驻
+    "recent_digest",  # 近 N 章摘要出现
+    "foreshadow_window",  # 命中本章伏笔窗口
+]
+
+# 理由的规范排序（保证 SelectionTrace 确定性）
+REASON_ORDER: tuple[SelectionReason, ...] = (
+    "explicit_beat",
+    "main_character",
+    "recent_digest",
+    "foreshadow_window",
+)
+
+
+class SelectedEntity(BaseModel):
+    """一个被选中的实体 + 入选理由（喂给 UX 注入透明面板，T1.6）。"""
+
+    model_config = {"frozen": True}
+
+    kind: EntityKind
+    name: str
+    reasons: list[SelectionReason] = Field(default_factory=list)
+
+
+class SelectionTrace(BaseModel):
+    """确定性选择的可解释结果：选了谁、为什么。"""
+
+    model_config = {"frozen": True}
+
+    selected: list[SelectedEntity] = Field(default_factory=list)
+
+
+class AssembledContext(BaseModel):
+    """组装结果：缓存断点前的稳定内核 + 断点后的易变内容 + 选择留痕。
+
+    stable_core / volatile 均为**已确定性排序、无时间戳/无 UUID** 的字符串
+    （缓存前缀字节稳定，ARCH §4.6 / 不变量 #9）。
+    """
+
+    model_config = {"frozen": True}
+
+    stable_core: str
+    volatile: str
+    selection: SelectionTrace
diff --git a/packages/core/ww_core/orchestrator/__init__.py b/packages/core/ww_core/orchestrator/__init__.py
new file mode 100644
index 0000000..16c4914
--- /dev/null
+++ b/packages/core/ww_core/orchestrator/__init__.py
@@ -0,0 +1,97 @@
+"""编排器（C4 / ARCH §5.2）——写章图 + 审稿子图 + SSE 归一 + Postgres checkpointer。
+
+- M1：单 `write` 节点（`build_write_graph`）。
+- M2：并行审子图（`build_review_graph`：START→各审并行→collect 落 `chapter_reviews`）+
+  审稿 SSE（`section`/`conflict` 事件 + `normalize_review`）。
+- accept（验收事务/`interrupt_before`）不在本图——属 T2.4（确定性代码读领域表，R3）。
+"""
+
+from __future__ import annotations
+
+from .collect import (
+    CONTINUITY,
+    ReviewRecorder,
+    collect_reviews,
+    extract_conflicts,
+)
+from .graph import (
+    COLLECT_NODE,
+    REVIEW_SPECS,
+    WRITE_NODE,
+    build_review_graph,
+    build_write_graph,
+    setup_checkpointer,
+)
+from .review_node import (
+    REVIEW_INCOMPLETE,
+    REVIEW_OK,
+    BoundReviewNode,
+    GatewayRun,
+    build_review_context,
+    build_review_request,
+    make_review_node,
+    run_review,
+)
+from .sse import (
+    EVENT_CONFLICT,
+    EVENT_DONE,
+    EVENT_ERROR,
+    EVENT_SECTION,
+    EVENT_TOKEN,
+    SECTION_DONE,
+    SECTION_INCOMPLETE,
+    SECTION_STARTED,
+    SseEvent,
+    conflict_event,
+    done_event,
+    error_event,
+    normalize_deltas,
+    normalize_review,
+    section_event,
+    token_event,
+)
+from .state import ChapterState, merge_reviews
+from .write_node import GatewayStream, build_write_request, stream_chapter_draft, write_node
+
+__all__ = [
+    "COLLECT_NODE",
+    "CONTINUITY",
+    "EVENT_CONFLICT",
+    "EVENT_DONE",
+    "EVENT_ERROR",
+    "EVENT_SECTION",
+    "EVENT_TOKEN",
+    "REVIEW_INCOMPLETE",
+    "REVIEW_OK",
+    "REVIEW_SPECS",
+    "SECTION_DONE",
+    "SECTION_INCOMPLETE",
+    "SECTION_STARTED",
+    "WRITE_NODE",
+    "BoundReviewNode",
+    "ChapterState",
+    "GatewayRun",
+    "GatewayStream",
+    "ReviewRecorder",
+    "SseEvent",
+    "build_review_context",
+    "build_review_graph",
+    "build_review_request",
+    "build_write_graph",
+    "build_write_request",
+    "collect_reviews",
+    "conflict_event",
+    "done_event",
+    "error_event",
+    "extract_conflicts",
+    "make_review_node",
+    "merge_reviews",
+    "normalize_deltas",
+    "normalize_review",
+    "run_review",
+    "section_event",
+    "setup_checkpointer",
+    "stream_chapter_draft",
+    "token_event",
+    "write_node",
+]
diff --git a/packages/core/ww_core/orchestrator/graph.py b/packages/core/ww_core/orchestrator/graph.py
new file mode 100644
index 0000000..c5f3753
--- /dev/null
+++ b/packages/core/ww_core/orchestrator/graph.py
@@ -0,0 +1,100 @@
+"""写章图工厂 + Postgres checkpointer（C4 / ARCH §5.2）。
+
+M1：单 `write` 节点图（四审/collect/interrupt 验收属 M2）。
+
+checkpointer 纪律（CLAUDE.md「LangGraph」）：
+- 用 **Postgres checkpointer**（durable/resumable，跨 write→accept 的请求间隙）。
+- `setup()` **只在 migrations/CI 跑**，绝不在 import/app-runtime 跑——故本模块只
+  暴露 `setup_checkpointer` 入口，build 函数不触碰 DDL。
+- 单测用 `MemorySaver`（或直接测节点函数），永不连真 Postgres。
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import TYPE_CHECKING, Any
+
+from langgraph.graph import END, START, StateGraph
+from ww_agents import AgentSpec, continuity_spec
+
+from .collect import ReviewRecorder, collect_reviews
+from .review_node import GatewayRun, run_review
+from .state import ChapterState
+from .write_node import GatewayStream, write_node
+
+if TYPE_CHECKING:
+    from langgraph.checkpoint.base import BaseCheckpointSaver
+    from langgraph.graph.state import CompiledStateGraph
+
+WRITE_NODE = "write"
+COLLECT_NODE = "collect"
+
+
+def build_write_graph(
+    gateway: GatewayStream,
+    *,
+    checkpointer: BaseCheckpointSaver[Any] | None = None,
+) -> CompiledStateGraph[ChapterState, None, ChapterState, ChapterState]:
+    """构建并编译单 `write` 节点图。
+
+    `gateway` 经闭包绑进节点（节点对图只暴露 `(state)->dict`，确定性可单测）。
+    `checkpointer`：运行时传 Postgres saver；单测传 `MemorySaver` 或不传。
+    """
+
+    async def _write(state: ChapterState) -> dict[str, str]:
+        return await write_node(state, gateway=gateway)
+
+    builder: StateGraph[ChapterState, None, ChapterState, ChapterState] = StateGraph(ChapterState)
+    builder.add_node(WRITE_NODE, _write)
+    builder.add_edge(START, WRITE_NODE)
+    builder.add_edge(WRITE_NODE, END)
+    return builder.compile(checkpointer=checkpointer)
+
+
+REVIEW_SPECS: tuple[AgentSpec, ...] = (continuity_spec,)
+
+
+def build_review_graph(
+    gateway: GatewayRun,
+    review_repo: ReviewRecorder,
+    *,
+    review_specs: Sequence[AgentSpec] = REVIEW_SPECS,
+    checkpointer: BaseCheckpointSaver[Any] | None = None,
+) -> CompiledStateGraph[ChapterState, None, ChapterState, ChapterState]:
+    """构建并编译审稿子图：`START →`（各 review spec 并行节点）`→ collect → END`。
+
+    M2 默认仅 continuity；`review_specs` 可扩（M3/M4 加 foreshadow/style/pace 三审），
+    图按这组 spec 循环加并行分支，全部汇入 collect（任一审失败不阻塞，§5.2——失败隔离
+    在 `run_review` 内、标 incomplete 占位）。
+
+    `gateway` / `review_repo` 经闭包绑进节点（langgraph 不收 functools.partial，见 gotcha）。
+    **accept 不在本图**：本任务只交付到 collect（审稿留痕落 `chapter_reviews`）+ review SSE；
+    `interrupt_before=["accept"]` / accept 节点属 T2.4（验收事务，确定性代码读领域表，R3）。
+    """
+
+    async def _collect(state: ChapterState) -> dict[str, Any]:
+        return await collect_reviews(state, review_repo=review_repo)
+
+    builder: StateGraph[ChapterState, None, ChapterState, ChapterState] = StateGraph(ChapterState)
+    builder.add_node(COLLECT_NODE, _collect)
+    for spec in review_specs:
+        # spec 经默认参绑定（避免闭包晚绑定 + langgraph 不收 functools.partial，见 gotcha）。
+        async def _review(state: ChapterState, *, _spec: AgentSpec = spec) -> dict[str, Any]:
+            return await run_review(_spec, state, gateway=gateway)
+
+        builder.add_node(spec.name, _review)
+        builder.add_edge(START, spec.name)
+        builder.add_edge(spec.name, COLLECT_NODE)
+    builder.add_edge(COLLECT_NODE, END)
+    return builder.compile(checkpointer=checkpointer)
+
+
+async def setup_checkpointer(conn_string: str) -> None:
+    """在 **migrations/CI** 创建 checkpointer 所需表（绝不在 app-runtime 调用）。
+
+    用 async saver 的上下文管理器拿连接后跑一次性 `setup()` DDL。
+    """
+    from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver
+
+    async with AsyncPostgresSaver.from_conn_string(conn_string) as saver:
+        await saver.setup()
diff --git a/packages/core/ww_core/orchestrator/sse.py b/packages/core/ww_core/orchestrator/sse.py
new file mode 100644
index 0000000..374ccac
--- /dev/null
+++ b/packages/core/ww_core/orchestrator/sse.py
@@ -0,0 +1,167 @@
+"""SSE 归一层（C4 / ARCH §7.3）——把网关 `Delta` 流归一为 SSE 事件对象。
+
+§7.3 全量事件：`token`/`section`/`conflict`/`done`/`error`。**M1 写章只需 `token`/`done`/`error`**
+（`section`/`conflict` 属 M2 四审）。本模块产出**与传输无关的事件对象**——
+HTTP `StreamingResponse`/`text/event-stream` 编码归 T1.4（apps/api），不在此处。
+
+错误处理（CLAUDE.md「Logging」）：底层流抛错 → 发一条 `error` 事件（带 `code`/`message`/
+`request_id`，对齐错误信封 §7.1）后**正常收尾**，绝不把异常泄到调用方循环里。
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator, Mapping
+from typing import Any
+
+import structlog
+from pydantic import BaseModel
+from ww_llm_gateway.types import Delta
+from ww_shared import AppError, ErrorCode
+
+from .review_node import REVIEW_OK
+
+log = structlog.get_logger(__name__)
+
+# §7.3 事件名（全集）。完整清单以 ARCHITECTURE §7.3 为准。
+EVENT_TOKEN = "token"  # 正文增量（写章 draft）
+EVENT_SECTION = "section"  # 四审分项开始/完成（审稿 review）
+EVENT_CONFLICT = "conflict"  # 冲突命中（审稿 review）
+EVENT_DONE = "done"
+EVENT_ERROR = "error"
+
+# section 状态（对齐 review_node.REVIEW_OK / REVIEW_INCOMPLETE，外加分项「开始」）。
+SECTION_STARTED = "started"
+SECTION_DONE = "done"
+SECTION_INCOMPLETE = "incomplete"
+
+
+class SseEvent(BaseModel):
+    """归一后的 SSE 事件对象（snake_case）；T1.4 据此编码为 text/event-stream。"""
+
+    model_config = {"frozen": True}
+
+    event: str
+    data: dict[str, object]
+
+
+def token_event(text: str) -> SseEvent:
+    return SseEvent(event=EVENT_TOKEN, data={"text": text})
+
+
+def done_event(*, length: int) -> SseEvent:
+    """完成事件：只带长度（不回灌全文——脱敏 + 前端已逐 token 累积）。"""
+    return SseEvent(event=EVENT_DONE, data={"length": length})
+
+
+def section_event(*, name: str, status: str) -> SseEvent:
+    """四审分项事件（§7.3 `section`）：`{name, status}`——前端报告分区进度。"""
+    return SseEvent(event=EVENT_SECTION, data={"name": name, "status": status})
+
+
+def conflict_event(
+    *,
+    type: str,
+    where: str,
+    refs: list[str],
+    suggestion: str,
+) -> SseEvent:
+    """冲突命中事件（§7.3 `conflict`）——形对齐 C6 `Conflict`，供正文就地朱砂标注。"""
+    return SseEvent(
+        event=EVENT_CONFLICT,
+        data={"type": type, "where": where, "refs": refs, "suggestion": suggestion},
+    )
+
+
+def error_event(*, code: str, message: str, request_id: str | None = None) -> SseEvent:
+    """错误事件，形对齐错误信封 §7.1（`code`/`message`），附 `request_id` 便于贯通排查。"""
+    return SseEvent(
+        event=EVENT_ERROR,
+        data={"code": code, "message": message, "request_id": request_id},
+    )
+
+
+async def normalize_deltas(
+    deltas: AsyncIterator[Delta],
+    *,
+    request_id: str | None = None,
+) -> AsyncIterator[SseEvent]:
+    """把 `Delta` 流归一为 SSE 事件流：每个非空 `Delta` → `token`；正常收尾发 `done`；
+    底层异常 → `error`（不再向上抛）。
+
+    这是 **T1.4 消费的缝**：apps/api 拿 `stream_chapter_draft(...)` 传进来，
+    本函数产出 `SseEvent`，T1.4 包成 `StreamingResponse`。
+    """
+    total = 0
+    try:
+        async for delta in deltas:
+            if delta.text:
+                total += len(delta.text)
+                yield token_event(delta.text)
+    except AppError as exc:
+        log.warning(
+            "sse_stream_error",
+            code=str(exc.code),
+            request_id=request_id,
+            chars_before_error=total,
+        )
+        yield error_event(code=str(exc.code), message=exc.message, request_id=request_id)
+        return
+    except Exception as exc:  # noqa: BLE001 — 边界兜底：任何意外都归一为 error 事件，不泄异常
+        log.error(
+            "sse_stream_unexpected_error",
+            error=str(exc),
+            request_id=request_id,
+            chars_before_error=total,
+        )
+        yield error_event(
+            code=str(ErrorCode.INTERNAL),
+            message="internal error during streaming",
+            request_id=request_id,
+        )
+        return
+    yield done_event(length=total)
+
+
+async def normalize_review(
+    reviews: Mapping[str, Any],
+    *,
+    request_id: str | None = None,
+) -> AsyncIterator[SseEvent]:
+    """把审稿子图的结构化产出（`state["reviews"]`）归一为 SSE 事件序列。
+
+    供 **T2.5** 的 `POST .../review` 端点消费：跑 `build_review_graph` 后拿 `final["reviews"]`
+    （形 `{name: {status, result}}`），本函数产 `section`（每审 done/incomplete）+ `conflict`
+    （命中冲突，形对齐 C6 `Conflict`）+ `done`。HTTP event-stream 编码归 T2.5，不在此处。
+
+    错误处理纪律（同 `normalize_deltas`）：任何意外 → 一条 `error` 事件后正常收尾，不上抛。
+    某审 `incomplete`（网关失败被隔离，§5.2）→ 发 `section{status:incomplete}`，不发其冲突。
+    `reviews` 键按字典序遍历 → 确定性事件顺序（便于单测/前端稳定渲染）。
+    """
+    try:
+        section_count = 0
+        for name in sorted(reviews.keys()):
+            entry = reviews.get(name) or {}
+            status = entry.get("status")
+            if status == REVIEW_OK:
+                yield section_event(name=name, status=SECTION_DONE)
+                section_count += 1
+                result = entry.get("result") or {}
+                for conflict in result.get("conflicts") or []:
+                    yield conflict_event(
+                        type=conflict.get("type", ""),
+                        where=conflict.get("where", ""),
+                        refs=list(conflict.get("refs") or []),
+                        suggestion=conflict.get("suggestion", ""),
+                    )
+            else:
+                yield section_event(name=name, status=SECTION_INCOMPLETE)
+                section_count += 1
+    except Exception as exc:  # noqa: BLE001 — 边界兜底：任何意外都归一为 error 事件，不泄异常
+        log.error("sse_review_unexpected_error", error=str(exc), request_id=request_id)
+        yield error_event(
+            code=str(ErrorCode.INTERNAL),
+            message="internal error during review",
+            request_id=request_id,
+        )
+        return
+    yield done_event(length=section_count)
diff --git a/packages/core/ww_core/orchestrator/state.py b/packages/core/ww_core/orchestrator/state.py
new file mode 100644
index 0000000..57a9d79
--- /dev/null
+++ b/packages/core/ww_core/orchestrator/state.py
@@ -0,0 +1,48 @@
+"""编排器图状态（C4 / ARCH §5.2）——只存**控制流 + 组装上下文 + 累积草稿**。
+
+不变量 #5：checkpoint 只承载控制流位置 + 组装上下文 + 已生成草稿增量 + 过程态审稿产物，
+正文/审稿以领域表为权威；resume 时从 `chapters`/`chapter_reviews` 重读（M2）。
+M2 加并行审 + collect；验收/interrupt 的真相源重建归 T2.4（accept 不在本图）。
+
+`ChapterState` 用 TypedDict（LangGraph 原生状态形态），字段全 snake_case。
+"""
+
+from __future__ import annotations
+
+import uuid
+from typing import Annotated, Any, TypedDict
+
+
+def merge_reviews(left: dict[str, Any], right: dict[str, Any]) -> dict[str, Any]:
+    """并行审分支的 `reviews` 归并 reducer（不可变：返回新 dict）。
+
+    四审为并行分支，同一 superstep 各写 `{spec.name: ...}`；无 reducer 时 LangGraph
+    会因并发写同一键抛 `InvalidUpdateError`。本 reducer 把各分项浅合并入一个 dict。
+    """
+    return {**left, **right}
+
+
+class ChapterState(TypedDict, total=False):
+    """写章 + 审稿图的控制流状态。
+
+    - `project_id` / `chapter_no`：定位本章（resume 时据此重读领域表）。
+    - `user_id`：调用作用域（落账 owner_id；原型单用户 stub）。
+    - `stable_core` / `volatile`：来自 `memory.assemble` 的中性文本
+      （stable_core=缓存断点前；volatile=断点后）。write 节点据此构造 `LlmRequest`。
+    - `draft`：write 节点累积的草稿正文（流式增量拼接的最终结果）。
+    - `review_context`（M2）：审稿注入文本——草稿 + 近况摘要/人物卡/世界观硬规则的
+      确定性拼装（复用 assemble 的 stable/volatile，无时间戳）。
+    - `reviews`（M2）：collect 汇总的各审分项结果（**过程态**）。落 `chapter_reviews`
+      表后真相在表；不变量 #5：checkpoint 不作审稿真相源，resume 从领域表重读。
+
+    M2 起字段按节点逐步填充，故 `total=False`（write 阶段无 review 字段亦合法）。
+    """
+
+    project_id: uuid.UUID
+    chapter_no: int
+    user_id: uuid.UUID
+    stable_core: str
+    volatile: str
+    draft: str
+    review_context: str
+    reviews: Annotated[dict[str, Any], merge_reviews]
diff --git a/packages/core/ww_core/orchestrator/write_node.py b/packages/core/ww_core/orchestrator/write_node.py
new file mode 100644
index 0000000..60a6081
--- /dev/null
+++ b/packages/core/ww_core/orchestrator/write_node.py
@@ -0,0 +1,89 @@
+"""write 节点（C4 / ARCH §5.2 §5.4 writer 行）——把组装上下文流式写成草稿。
+
+确定性边界（CLAUDE.md「LangGraph」纪律）：
+- 节点逻辑里**无 LLM 非确定性**——所有不确定性藏在网关后；节点只做
+  `AssembledContext → LlmRequest` 的纯构造 + 转发 `Gateway.stream` 的 `Delta`。
+- 因此 `build_write_request` 是纯函数、`stream_chapter_draft` 注入 mock 网关即可单测，
+  不需要图运行时、不需要真 Postgres。
+- 瞬时失败重试在网关（§4.5），不在节点；节点遇错只干净抛出，SSE 层据此发 `error`（§7.3）。
+
+不变量 ②：只传 `tier="writer"`，绝不传具体 model。
+不变量 #9：`stable_core` 进 `system` 缓存断点前（cache=True），`volatile` 进 `input`（断点后）。
+"""
+
+from __future__ import annotations
+
+import uuid
+from collections.abc import AsyncIterator
+from typing import Protocol
+
+from ww_llm_gateway.types import Block, Delta, LlmRequest, Scope
+
+from .state import ChapterState
+
+
+class GatewayStream(Protocol):
+    """write 节点对网关的最小依赖——只需 `stream`（注入真网关或 mock）。"""
+
+    def stream(self, req: LlmRequest) -> AsyncIterator[Delta]: ...
+
+
+def build_write_request(
+    *,
+    stable_core: str,
+    volatile: str,
+    user_id: uuid.UUID,
+    project_id: uuid.UUID,
+) -> LlmRequest:
+    """据组装上下文构造写章请求（纯函数，决策 2026-06-18）。
+
+    `stable_core` → `system` 缓存断点前块（cache=True）；`volatile` → `input`（断点后）。
+    """
+    return LlmRequest(
+        tier="writer",
+        system=[Block(text=stable_core, cache=True)],
+        input=volatile,
+        stream=True,
+        scope=Scope(user_id=user_id, project_id=project_id),
+    )
+
+
+async def stream_chapter_draft(
+    gateway: GatewayStream,
+    *,
+    stable_core: str,
+    volatile: str,
+    user_id: uuid.UUID,
+    project_id: uuid.UUID,
+) -> AsyncIterator[Delta]:
+    """流式生成本章草稿增量（T1.4 的底层流）——构造请求 → 转发网关 `Delta`。
+
+    本身不累积/不落库；累积归 `write_node`（图状态），落库（自动保存）归 T1.4。
+    瞬时失败已在网关重试；此处的异常向上抛给 SSE 归一层（发 `error` 事件）。
+    """
+    req = build_write_request(
+        stable_core=stable_core,
+        volatile=volatile,
+        user_id=user_id,
+        project_id=project_id,
+    )
+    async for delta in gateway.stream(req):
+        yield delta
+
+
+async def write_node(state: ChapterState, *, gateway: GatewayStream) -> dict[str, str]:
+    """LangGraph write 节点：流式生成 → 累积草稿 → 不可变返回 `{"draft": ...}`。
+
+    直接调用本函数（注入 mock 网关）即可单测，无需图运行时。
+    返回**增量字典**（LangGraph 合并进状态），不原地改 `state`（不可变更新）。
+    """
+    parts: list[str] = []
+    async for delta in stream_chapter_draft(
+        gateway,
+        stable_core=state["stable_core"],
+        volatile=state["volatile"],
+        user_id=state["user_id"],
+        project_id=state["project_id"],
+    ):
+        parts.append(delta.text)
+    return {"draft": "".join(parts)}
diff --git a/packages/llm_gateway/README.md b/packages/llm_gateway/README.md
new file mode 100644
index 0000000..b346088
--- /dev/null
+++ b/packages/llm_gateway/README.md
@@ -0,0 +1,2 @@
+# core (@llm/@backend) — 占位骨架
+领域核心(domain/状态机)、编排器(orchestrator/LangGraph 图)、记忆服务(memory)。Phase 1+ 由对应 owner 填充。
diff --git a/packages/llm_gateway/pyproject.toml b/packages/llm_gateway/pyproject.toml
new file mode 100644
index 0000000..d206eef
--- /dev/null
+++ b/packages/llm_gateway/pyproject.toml
@@ -0,0 +1,24 @@
+[project]
+name = "ww-llm-gateway"
+version = "0.0.0"
+requires-python = ">=3.12"
+dependencies = [
+  "openai>=1.40",
+  "instructor>=1.5",
+  "pydantic>=2.7",
+  "ww-shared",
+  "ww-config",
+  "ww-db",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["ww_llm_gateway"]
+
+[tool.uv.sources]
+ww-shared = { workspace = true }
+ww-config = { workspace = true }
+ww-db = { workspace = true }
diff --git a/packages/llm_gateway/tests/conftest.py b/packages/llm_gateway/tests/conftest.py
new file mode 100644
index 0000000..5a393ac
--- /dev/null
+++ b/packages/llm_gateway/tests/conftest.py
@@ -0,0 +1,18 @@
+"""网关单测公用 fixtures。"""
+
+from __future__ import annotations
+
+import uuid
+
+import pytest
+from ww_llm_gateway.types import LlmRequest, Scope
+
+
+@pytest.fixture
+def scope() -> Scope:
+    return Scope(user_id=uuid.UUID(int=1), project_id=uuid.UUID(int=2))
+
+
+@pytest.fixture
+def req(scope: Scope) -> LlmRequest:
+    return LlmRequest(tier="writer", input="写第 1 章", scope=scope)
diff --git a/packages/llm_gateway/tests/fakes.py b/packages/llm_gateway/tests/fakes.py
new file mode 100644
index 0000000..dde29cc
--- /dev/null
+++ b/packages/llm_gateway/tests/fakes.py
@@ -0,0 +1,56 @@
+"""网关单测替身（mock provider + 内存 ledger）——不碰真实 API / DB。
+
+放在独立模块（非 conftest）以便测试用绝对导入 `from fakes import ...`；
+本目录无 __init__.py（避免与顶层 tests 包同名冲突），故走 pytest 路径注入。
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator
+
+from ww_llm_gateway.adapters.base import (
+    Capabilities,
+    ProviderResult,
+    ProviderUsage,
+    StreamChunk,
+)
+from ww_llm_gateway.routing import Route
+from ww_llm_gateway.types import LlmRequest, Scope, Tier, Usage
+
+
+class FakeAdapter:
+    provider = "deepseek"
+
+    def __init__(self, text: str = "hello world", deltas: list[str] | None = None) -> None:
+        self.text = text
+        self.deltas = deltas if deltas is not None else ["hel", "lo ", "world"]
+        self.complete_calls = 0
+        self.stream_calls = 0
+
+    def capabilities(self) -> Capabilities:
+        return Capabilities(structured_output=True, prefix_cache=True)
+
+    async def complete(self, req: LlmRequest, model: str) -> ProviderResult:
+        self.complete_calls += 1
+        return ProviderResult(
+            text=self.text,
+            usage=ProviderUsage(input_tokens=100, output_tokens=50),
+        )
+
+    async def stream(self, req: LlmRequest, model: str) -> AsyncIterator[StreamChunk]:
+        self.stream_calls += 1
+        for d in self.deltas:
+            yield StreamChunk(text=d)
+        yield StreamChunk(usage=ProviderUsage(input_tokens=100, output_tokens=50))
+
+
+class FakeLedger:
+    def __init__(self) -> None:
+        self.records: list[Usage] = []
+
+    async def record(self, scope: Scope, usage: Usage) -> None:
+        self.records.append(usage)
+
+
+def fake_route(tier: Tier) -> Route:
+    return Route(provider="deepseek", model="deepseek-chat")
diff --git a/packages/llm_gateway/tests/test_gateway.py b/packages/llm_gateway/tests/test_gateway.py
new file mode 100644
index 0000000..2045677
--- /dev/null
+++ b/packages/llm_gateway/tests/test_gateway.py
@@ -0,0 +1,82 @@
+"""T1.1 网关核心单测：run / stream / 记账 / 路由 / 错误。"""
+
+from __future__ import annotations
+
+import uuid
+
+import pytest
+from fakes import FakeAdapter, FakeLedger, fake_route
+from ww_llm_gateway.gateway import Gateway
+from ww_llm_gateway.routing import resolve_route
+from ww_llm_gateway.types import LlmRequest, Scope
+from ww_shared import AppError, ErrorCode
+
+
+async def test_run_returns_text_usage_and_served_by(req: LlmRequest) -> None:
+    adapter = FakeAdapter(text="第一章正文")
+    ledger = FakeLedger()
+    gw = Gateway({"deepseek": adapter}, ledger, resolver=fake_route)
+
+    resp = await gw.run(req)
+
+    assert resp.text == "第一章正文"
+    assert resp.served_by.provider == "deepseek"
+    assert resp.served_by.fell_back is False
+    assert resp.usage.input_tokens == 100
+    assert resp.usage.output_tokens == 50
+    assert adapter.complete_calls == 1
+
+
+async def test_run_writes_exactly_one_ledger_record(req: LlmRequest) -> None:
+    ledger = FakeLedger()
+    gw = Gateway({"deepseek": FakeAdapter()}, ledger, resolver=fake_route)
+
+    await gw.run(req)
+
+    assert len(ledger.records) == 1
+    assert ledger.records[0].provider == "deepseek"
+    assert ledger.records[0].model == "deepseek-chat"
+
+
+async def test_run_computes_cost_from_pricing(req: LlmRequest) -> None:
+    ledger = FakeLedger()
+    gw = Gateway({"deepseek": FakeAdapter()}, ledger, resolver=fake_route)
+
+    resp = await gw.run(req)
+
+    # ceil(100/1e6*27 + 50/1e6*110) == 1 cent
+    assert resp.usage.cost_minor == 1
+    assert resp.usage.currency == "USD"
+
+
+async def test_stream_yields_deltas_and_records_once(req: LlmRequest) -> None:
+    ledger = FakeLedger()
+    gw = Gateway({"deepseek": FakeAdapter(deltas=["a", "b", "c"])}, ledger, resolver=fake_route)
+
+    collected = [d.text async for d in gw.stream(req)]
+
+    assert "".join(collected) == "abc"
+    assert len(ledger.records) == 1
+    assert ledger.records[0].output_tokens == 50
+
+
+async def test_unknown_provider_raises_llm_unavailable(req: LlmRequest) -> None:
+    gw = Gateway({}, FakeLedger(), resolver=fake_route)
+
+    with pytest.raises(AppError) as exc:
+        await gw.run(req)
+
+    assert exc.value.code == ErrorCode.LLM_UNAVAILABLE
+
+
+def test_resolve_route_parses_tier_defaults() -> None:
+    route = resolve_route("writer")
+    assert route.provider == "deepseek"
+    assert route.model == "deepseek-chat"
+
+
+def test_agent_passes_only_tier_never_model() -> None:
+    # 不变量 ②：LlmRequest 无 model 字段，只有 tier
+    req = LlmRequest(tier="analyst", input="x", scope=Scope(user_id=uuid.UUID(int=9)))
+    assert "model" not in LlmRequest.model_fields
+    assert req.tier == "analyst"
diff --git a/packages/llm_gateway/tests/test_openai_adapter.py b/packages/llm_gateway/tests/test_openai_adapter.py
new file mode 100644
index 0000000..5b935ba
--- /dev/null
+++ b/packages/llm_gateway/tests/test_openai_adapter.py
@@ -0,0 +1,102 @@
+"""OpenAI 兼容适配器单测：用替身 client 验证消息翻译 + usage 映射（不联网）。"""
+
+from __future__ import annotations
+
+import uuid
+from collections.abc import AsyncIterator
+from types import SimpleNamespace
+from typing import Any, cast
+
+from openai import AsyncOpenAI
+from ww_llm_gateway.adapters.openai_compat import OpenAICompatAdapter, _messages
+from ww_llm_gateway.types import Block, LlmRequest, Scope
+
+
+def _req(**kw: Any) -> LlmRequest:
+    kw.setdefault("tier", "writer")
+    return LlmRequest(scope=Scope(user_id=uuid.UUID(int=1)), **kw)
+
+
+class _FakeStream:
+    def __init__(self, chunks: list[Any]) -> None:
+        self._chunks = chunks
+
+    async def __aiter__(self) -> AsyncIterator[Any]:
+        for c in self._chunks:
+            yield c
+
+
+class _FakeCompletions:
+    def __init__(self, response: Any, stream_chunks: list[Any]) -> None:
+        self._response = response
+        self._stream_chunks = stream_chunks
+        self.last_messages: Any = None
+
+    async def create(self, **kw: Any) -> Any:
+        self.last_messages = kw.get("messages")
+        if kw.get("stream"):
+            return _FakeStream(self._stream_chunks)
+        return self._response
+
+
+class _FakeClient:
+    def __init__(self, completions: _FakeCompletions) -> None:
+        self.chat = SimpleNamespace(completions=completions)
+
+
+def test_messages_put_system_first_then_user() -> None:
+    req = _req(input="正文", system=[Block(text="世界观硬规则", cache=True)])
+    msgs = _messages(req)
+    assert msgs[0]["role"] == "system"
+    assert msgs[0]["content"] == "世界观硬规则"
+    assert msgs[1]["role"] == "user"
+    assert msgs[1]["content"] == "正文"
+
+
+async def test_complete_maps_text_and_usage() -> None:
+    response = SimpleNamespace(
+        choices=[SimpleNamespace(message=SimpleNamespace(content="草稿"))],
+        usage=SimpleNamespace(
+            prompt_tokens=120,
+            completion_tokens=80,
+            prompt_tokens_details=SimpleNamespace(cached_tokens=30),
+        ),
+    )
+    completions = _FakeCompletions(response, [])
+    adapter = OpenAICompatAdapter("deepseek", cast(AsyncOpenAI, _FakeClient(completions)))
+
+    result = await adapter.complete(_req(input="x"), "deepseek-chat")
+
+    assert result.text == "草稿"
+    assert result.usage.input_tokens == 120
+    assert result.usage.output_tokens == 80
+    assert result.usage.cache_read_tokens == 30
+
+
+async def test_stream_yields_text_then_final_usage() -> None:
+    chunks = [
+        SimpleNamespace(choices=[SimpleNamespace(delta=SimpleNamespace(content="他"))], usage=None),
+        SimpleNamespace(choices=[SimpleNamespace(delta=SimpleNamespace(content="说"))], usage=None),
+        SimpleNamespace(
+            choices=[],
+            usage=SimpleNamespace(
+                prompt_tokens=10,
+                completion_tokens=2,
+                prompt_tokens_details=None,
+            ),
+        ),
+    ]
+    completions = _FakeCompletions(None, chunks)
+    adapter = OpenAICompatAdapter("deepseek", cast(AsyncOpenAI, _FakeClient(completions)))
+
+    texts: list[str] = []
+    usage_seen = None
+    async for ch in adapter.stream(_req(input="x"), "deepseek-chat"):
+        if ch.text:
+            texts.append(ch.text)
+        if ch.usage is not None:
+            usage_seen = ch.usage
+
+    assert "".join(texts) == "他说"
+    assert usage_seen is not None
+    assert usage_seen.output_tokens == 2
diff --git a/packages/llm_gateway/tests/test_structured_output.py b/packages/llm_gateway/tests/test_structured_output.py
new file mode 100644
index 0000000..661e36e
--- /dev/null
+++ b/packages/llm_gateway/tests/test_structured_output.py
@@ -0,0 +1,123 @@
+"""T2.1 结构化输出接线单测：output_schema → instructor → LlmResponse.parsed。
+
+注入 fake instructor client（返回 (parsed, raw_completion)），不联网。
+覆盖：① 带 schema → parsed 是实例且字段正确 + 仍记 1 条 ledger；
+② 无 schema → parsed is None 且纯文本路径不变。
+"""
+
+from __future__ import annotations
+
+import uuid
+from types import SimpleNamespace
+from typing import Any, cast
+
+from openai import AsyncOpenAI
+from pydantic import BaseModel
+from ww_llm_gateway.adapters.openai_compat import OpenAICompatAdapter
+from ww_llm_gateway.gateway import Gateway
+from ww_llm_gateway.routing import Route
+from ww_llm_gateway.types import LlmRequest, Scope, Tier
+
+
+class _Out(BaseModel):
+    label: str
+    score: int
+
+
+def _req(**kw: Any) -> LlmRequest:
+    kw.setdefault("tier", "analyst")
+    return LlmRequest(scope=Scope(user_id=uuid.UUID(int=1)), **kw)
+
+
+def _route(_tier: Tier) -> Route:
+    return Route(provider="deepseek", model="deepseek-chat")
+
+
+class _FakeStructured:
+    """模拟 instructor.AsyncInstructor：create_with_completion 返回 (parsed, raw)。"""
+
+    def __init__(self, parsed: BaseModel, raw: Any) -> None:
+        self._parsed = parsed
+        self._raw = raw
+        self.calls = 0
+        self.last_response_model: Any = None
+
+    async def create_with_completion(
+        self, *, messages: Any, response_model: Any, **kw: Any
+    ) -> tuple[BaseModel, Any]:
+        self.calls += 1
+        self.last_response_model = response_model
+        return self._parsed, self._raw
+
+
+def _raw_with_usage() -> Any:
+    return SimpleNamespace(
+        usage=SimpleNamespace(
+            prompt_tokens=42,
+            completion_tokens=7,
+            prompt_tokens_details=SimpleNamespace(cached_tokens=5),
+        )
+    )
+
+
+def _client() -> AsyncOpenAI:
+    return cast(AsyncOpenAI, SimpleNamespace(chat=SimpleNamespace(completions=None)))
+
+
+async def test_complete_with_schema_returns_parsed_instance() -> None:
+    structured = _FakeStructured(_Out(label="ok", score=9), _raw_with_usage())
+    adapter = OpenAICompatAdapter("deepseek", _client(), structured_client=structured)
+
+    result = await adapter.complete(_req(input="x", output_schema=_Out), "deepseek-chat")
+
+    assert isinstance(result.parsed, _Out)
+    assert result.parsed.label == "ok"
+    assert result.parsed.score == 9
+    assert structured.last_response_model is _Out
+    # usage 从 raw completion 提取
+    assert result.usage.input_tokens == 42
+    assert result.usage.output_tokens == 7
+    assert result.usage.cache_read_tokens == 5
+
+
+async def test_complete_without_schema_keeps_parsed_none() -> None:
+    response = SimpleNamespace(
+        choices=[SimpleNamespace(message=SimpleNamespace(content="纯文本"))],
+        usage=SimpleNamespace(prompt_tokens=3, completion_tokens=2, prompt_tokens_details=None),
+    )
+
+    class _Completions:
+        async def create(self, **kw: Any) -> Any:
+            return response
+
+    client = cast(
+        AsyncOpenAI,
+        SimpleNamespace(chat=SimpleNamespace(completions=_Completions())),
+    )
+    adapter = OpenAICompatAdapter("deepseek", client)
+
+    result = await adapter.complete(_req(input="x"), "deepseek-chat")
+
+    assert result.parsed is None
+    assert result.text == "纯文本"
+
+
+class _FakeLedger:
+    def __init__(self) -> None:
+        self.records: list[Any] = []
+
+    async def record(self, scope: Any, usage: Any) -> None:
+        self.records.append(usage)
+
+
+async def test_gateway_run_passes_parsed_through_and_records_once() -> None:
+    structured = _FakeStructured(_Out(label="hit", score=1), _raw_with_usage())
+    adapter = OpenAICompatAdapter("deepseek", _client(), structured_client=structured)
+    ledger = _FakeLedger()
+    gw = Gateway({"deepseek": adapter}, ledger, resolver=_route)
+
+    resp = await gw.run(_req(input="x", output_schema=_Out))
+
+    assert isinstance(resp.parsed, _Out)
+    assert resp.parsed.label == "hit"
+    assert len(ledger.records) == 1
diff --git a/packages/llm_gateway/ww_llm_gateway/__init__.py b/packages/llm_gateway/ww_llm_gateway/__init__.py
new file mode 100644
index 0000000..3e02718
--- /dev/null
+++ b/packages/llm_gateway/ww_llm_gateway/__init__.py
@@ -0,0 +1,47 @@
+"""LLM 网关（C1 / ARCH §4）：薄自建，tier→provider+model，屏蔽厂商差异。"""
+
+from __future__ import annotations
+
+from .adapters.base import (
+    Capabilities,
+    ProviderAdapter,
+    ProviderResult,
+    ProviderUsage,
+    StreamChunk,
+)
+from .adapters.openai_compat import OpenAICompatAdapter
+from .gateway import Gateway
+from .ledger import LedgerSink, SqlAlchemyLedgerSink
+from .routing import Route, resolve_route
+from .types import (
+    Block,
+    Delta,
+    LlmRequest,
+    LlmResponse,
+    Scope,
+    ServedBy,
+    Tier,
+    Usage,
+)
+
+__all__ = [
+    "Block",
+    "Capabilities",
+    "Delta",
+    "Gateway",
+    "LedgerSink",
+    "LlmRequest",
+    "LlmResponse",
+    "OpenAICompatAdapter",
+    "ProviderAdapter",
+    "ProviderResult",
+    "ProviderUsage",
+    "Route",
+    "Scope",
+    "ServedBy",
+    "SqlAlchemyLedgerSink",
+    "StreamChunk",
+    "Tier",
+    "Usage",
+    "resolve_route",
+]
diff --git a/packages/llm_gateway/ww_llm_gateway/adapters/__init__.py b/packages/llm_gateway/ww_llm_gateway/adapters/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/packages/llm_gateway/ww_llm_gateway/adapters/base.py b/packages/llm_gateway/ww_llm_gateway/adapters/base.py
new file mode 100644
index 0000000..2787a75
--- /dev/null
+++ b/packages/llm_gateway/ww_llm_gateway/adapters/base.py
@@ -0,0 +1,51 @@
+"""适配器接口与中间数据形（ARCH §4.2/§4.4）。
+
+适配器把 `LlmRequest` 翻译成目标厂商请求，并把响应/流/usage 翻译回统一中间形。
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator
+from typing import Protocol, runtime_checkable
+
+from pydantic import BaseModel, ConfigDict
+
+from ..types import LlmRequest
+
+
+class Capabilities(BaseModel):
+    structured_output: bool = False
+    prefix_cache: bool = False
+    thinking: bool = False
+
+
+class ProviderUsage(BaseModel):
+    input_tokens: int
+    output_tokens: int
+    cache_read_tokens: int = 0
+
+
+class ProviderResult(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    text: str
+    usage: ProviderUsage
+    parsed: BaseModel | None = None  # output_schema 命中时的结构化结果（§4.4）
+
+
+class StreamChunk(BaseModel):
+    """流式块：文本增量（usage=None），或末尾用量块（text=""）。"""
+
+    text: str = ""
+    usage: ProviderUsage | None = None
+
+
+@runtime_checkable
+class ProviderAdapter(Protocol):
+    provider: str
+
+    def capabilities(self) -> Capabilities: ...
+
+    async def complete(self, req: LlmRequest, model: str) -> ProviderResult: ...
+
+    def stream(self, req: LlmRequest, model: str) -> AsyncIterator[StreamChunk]: ...
diff --git a/packages/llm_gateway/ww_llm_gateway/adapters/openai_compat.py b/packages/llm_gateway/ww_llm_gateway/adapters/openai_compat.py
new file mode 100644
index 0000000..3575034
--- /dev/null
+++ b/packages/llm_gateway/ww_llm_gateway/adapters/openai_compat.py
@@ -0,0 +1,129 @@
+"""OpenAI 兼容适配器：一套覆盖 DeepSeek/Kimi/Qwen/GLM/OpenAI（ARCH §4.2）。
+
+仅 base_url + model + key 不同。注入 `AsyncOpenAI` 客户端以便测试用替身。
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator
+from typing import Any, Protocol
+
+import instructor
+from openai import AsyncOpenAI
+from openai.types.chat import ChatCompletionMessageParam
+from pydantic import BaseModel
+
+from ..types import LlmRequest
+from .base import Capabilities, ProviderResult, ProviderUsage, StreamChunk
+
+
+class StructuredClient(Protocol):
+    """instructor 风格的结构化客户端缝（`AsyncInstructor` 即满足此协议）。
+
+    抽成 Protocol 以便测试注入 fake，绝不联网（不变量：测试零真实 LLM）。
+    """
+
+    async def create_with_completion(
+        self, *, messages: Any, response_model: type[BaseModel], **kwargs: Any
+    ) -> tuple[BaseModel, Any]: ...
+
+
+def _system_text(req: LlmRequest) -> str:
+    return "\n\n".join(b.text for b in req.system)
+
+
+def _input_text(req: LlmRequest) -> str:
+    if isinstance(req.input, str):
+        return req.input
+    return "\n\n".join(b.text for b in req.input)
+
+
+def _messages(req: LlmRequest) -> list[ChatCompletionMessageParam]:
+    msgs: list[ChatCompletionMessageParam] = []
+    system = _system_text(req)
+    if system:
+        msgs.append({"role": "system", "content": system})
+    msgs.append({"role": "user", "content": _input_text(req)})
+    return msgs
+
+
+def _cache_read(usage: Any) -> int:
+    details = getattr(usage, "prompt_tokens_details", None)
+    if details is None:
+        return 0
+    return int(getattr(details, "cached_tokens", 0) or 0)
+
+
+def _usage_from(usage: Any) -> ProviderUsage:
+    if usage is None:
+        return ProviderUsage(input_tokens=0, output_tokens=0)
+    return ProviderUsage(
+        input_tokens=getattr(usage, "prompt_tokens", 0) or 0,
+        output_tokens=getattr(usage, "completion_tokens", 0) or 0,
+        cache_read_tokens=_cache_read(usage),
+    )
+
+
+class OpenAICompatAdapter:
+    def __init__(
+        self,
+        provider: str,
+        client: AsyncOpenAI,
+        *,
+        structured_client: StructuredClient | None = None,
+    ) -> None:
+        self.provider = provider
+        self._client = client
+        # 结构化输出走 instructor（Pydantic 校验 + 重试，锁定栈）；可注入便于测试。
+        self._structured_client = structured_client
+
+    def capabilities(self) -> Capabilities:
+        return Capabilities(structured_output=True, prefix_cache=True, thinking=False)
+
+    def _structured(self) -> StructuredClient:
+        if self._structured_client is None:
+            # 懒构建：从同一 AsyncOpenAI client patch 出 instructor 客户端。
+            self._structured_client = instructor.from_openai(self._client)
+        return self._structured_client
+
+    async def complete(self, req: LlmRequest, model: str) -> ProviderResult:
+        if req.output_schema is not None:
+            return await self._complete_structured(req, model)
+        return await self._complete_text(req, model)
+
+    async def _complete_text(self, req: LlmRequest, model: str) -> ProviderResult:
+        resp = await self._client.chat.completions.create(
+            model=model,
+            messages=_messages(req),
+            max_tokens=req.max_tokens,
+        )
+        text = resp.choices[0].message.content or ""
+        return ProviderResult(text=text, usage=_usage_from(resp.usage))
+
+    async def _complete_structured(self, req: LlmRequest, model: str) -> ProviderResult:
+        assert req.output_schema is not None
+        parsed, raw = await self._structured().create_with_completion(
+            messages=_messages(req),
+            response_model=req.output_schema,
+            model=model,
+            max_tokens=req.max_tokens,
+        )
+        usage = _usage_from(getattr(raw, "usage", None))
+        # 文本载体保留校验后的 JSON（便于日志/留痕）；程序消费走 parsed。
+        return ProviderResult(text=parsed.model_dump_json(), usage=usage, parsed=parsed)
+
+    async def stream(self, req: LlmRequest, model: str) -> AsyncIterator[StreamChunk]:
+        stream = await self._client.chat.completions.create(
+            model=model,
+            messages=_messages(req),
+            max_tokens=req.max_tokens,
+            stream=True,
+            stream_options={"include_usage": True},
+        )
+        async for chunk in stream:
+            if chunk.choices:
+                delta = chunk.choices[0].delta
+                if delta and delta.content:
+                    yield StreamChunk(text=delta.content)
+            if getattr(chunk, "usage", None):
+                yield StreamChunk(usage=_usage_from(chunk.usage))
diff --git a/packages/llm_gateway/ww_llm_gateway/gateway.py b/packages/llm_gateway/ww_llm_gateway/gateway.py
new file mode 100644
index 0000000..3d1f44b
--- /dev/null
+++ b/packages/llm_gateway/ww_llm_gateway/gateway.py
@@ -0,0 +1,99 @@
+"""网关核心：路由 → 调用适配器 → 记账 → 返回（ARCH §4.1–4.3/§4.8）。
+
+M1 单 provider，无回退/熔断（那是 M5/T5.4）。流式经 `stream()` 归一为 `Delta`。
+日志脱敏：只记长度，绝不记原文/api key（不变量、§9.3）。
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator, Callable
+
+import structlog
+from ww_shared import AppError, ErrorCode
+
+from .adapters.base import ProviderAdapter, ProviderUsage
+from .ledger import LedgerSink
+from .pricing import cost_minor
+from .routing import Route, resolve_route
+from .types import Delta, LlmRequest, LlmResponse, ServedBy, Tier, Usage
+
+log = structlog.get_logger(__name__)
+
+
+def _input_len(req: LlmRequest) -> int:
+    if isinstance(req.input, str):
+        return len(req.input)
+    return sum(len(b.text) for b in req.input)
+
+
+class Gateway:
+    def __init__(
+        self,
+        adapters: dict[str, ProviderAdapter],
+        ledger: LedgerSink,
+        resolver: Callable[[Tier], Route] = resolve_route,
+    ) -> None:
+        self._adapters = adapters
+        self._ledger = ledger
+        self._resolve = resolver
+
+    def _adapter_for(self, provider: str) -> ProviderAdapter:
+        adapter = self._adapters.get(provider)
+        if adapter is None:
+            raise AppError(ErrorCode.LLM_UNAVAILABLE, f"no adapter for provider {provider!r}")
+        return adapter
+
+    def _usage(self, route: Route, pu: ProviderUsage) -> Usage:
+        cost, currency = cost_minor(route.provider, route.model, pu.input_tokens, pu.output_tokens)
+        return Usage(
+            provider=route.provider,
+            model=route.model,
+            input_tokens=pu.input_tokens,
+            output_tokens=pu.output_tokens,
+            cache_read_tokens=pu.cache_read_tokens,
+            cost_minor=cost,
+            currency=currency,
+        )
+
+    def _log_call(self, req: LlmRequest, usage: Usage, *, stream: bool) -> None:
+        log.info(
+            "llm_call",
+            provider=usage.provider,
+            model=usage.model,
+            tier=req.tier,
+            input_tokens=usage.input_tokens,
+            output_tokens=usage.output_tokens,
+            cache_read_tokens=usage.cache_read_tokens,
+            cost_minor=usage.cost_minor,
+            currency=usage.currency,
+            stream=stream,
+            input_chars=_input_len(req),
+            project_id=str(req.scope.project_id) if req.scope.project_id else None,
+        )
+
+    async def run(self, req: LlmRequest) -> LlmResponse:
+        route = self._resolve(req.tier)
+        adapter = self._adapter_for(route.provider)
+        result = await adapter.complete(req, route.model)
+        usage = self._usage(route, result.usage)
+        await self._ledger.record(req.scope, usage)
+        self._log_call(req, usage, stream=False)
+        return LlmResponse(
+            text=result.text,
+            parsed=result.parsed,
+            usage=usage,
+            served_by=ServedBy(provider=route.provider, model=route.model),
+        )
+
+    async def stream(self, req: LlmRequest) -> AsyncIterator[Delta]:
+        route = self._resolve(req.tier)
+        adapter = self._adapter_for(route.provider)
+        final = ProviderUsage(input_tokens=0, output_tokens=0)
+        async for chunk in adapter.stream(req, route.model):
+            if chunk.text:
+                yield Delta(text=chunk.text)
+            if chunk.usage is not None:
+                final = chunk.usage
+        usage = self._usage(route, final)
+        await self._ledger.record(req.scope, usage)
+        self._log_call(req, usage, stream=True)
diff --git a/packages/llm_gateway/ww_llm_gateway/ledger.py b/packages/llm_gateway/ww_llm_gateway/ledger.py
new file mode 100644
index 0000000..1e03530
--- /dev/null
+++ b/packages/llm_gateway/ww_llm_gateway/ledger.py
@@ -0,0 +1,39 @@
+"""用量记账落库（ARCH §4.8）。
+
+`LedgerSink` 为接口，便于测试注入内存替身；生产用 SQLAlchemy 实现写 usage_ledger。
+"""
+
+from __future__ import annotations
+
+from typing import Protocol
+
+from sqlalchemy.ext.asyncio import AsyncSession
+from ww_db.models import UsageLedger
+
+from .types import Scope, Usage
+
+
+class LedgerSink(Protocol):
+    async def record(self, scope: Scope, usage: Usage) -> None: ...
+
+
+class SqlAlchemyLedgerSink:
+    """把每次调用写入 usage_ledger（owner_id 取 scope.user_id，单用户 stub）。"""
+
+    def __init__(self, session: AsyncSession) -> None:
+        self._session = session
+
+    async def record(self, scope: Scope, usage: Usage) -> None:
+        row = UsageLedger(
+            owner_id=scope.user_id,
+            project_id=scope.project_id,
+            provider=usage.provider,
+            model=usage.model,
+            input_tokens=usage.input_tokens,
+            output_tokens=usage.output_tokens,
+            cache_read=usage.cache_read_tokens,
+            cost_minor=usage.cost_minor,
+            currency=usage.currency,
+        )
+        self._session.add(row)
+        await self._session.flush()
diff --git a/packages/llm_gateway/ww_llm_gateway/pricing.py b/packages/llm_gateway/ww_llm_gateway/pricing.py
new file mode 100644
index 0000000..e866c8b
--- /dev/null
+++ b/packages/llm_gateway/ww_llm_gateway/pricing.py
@@ -0,0 +1,34 @@
+"""提供商价格表与成本换算（ARCH §4.8）。
+
+价格以「每百万 token 的最小货币单位（如分/cent）」表示，随 provider 配置维护；
+未知 (provider, model) 则成本计 0（仍记账，便于观测）。
+"""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class Price:
+    in_per_mtok: int
+    out_per_mtok: int
+    currency: str
+
+
+# 近似价（可后续移入 config / provider 配置维护）
+_PRICING: dict[tuple[str, str], Price] = {
+    ("deepseek", "deepseek-chat"): Price(in_per_mtok=27, out_per_mtok=110, currency="USD"),
+}
+
+
+def cost_minor(provider: str, model: str, input_tokens: int, output_tokens: int) -> tuple[int, str]:
+    price = _PRICING.get((provider, model))
+    if price is None:
+        return 0, "USD"
+    cost = math.ceil(
+        input_tokens / 1_000_000 * price.in_per_mtok
+        + output_tokens / 1_000_000 * price.out_per_mtok
+    )
+    return cost, price.currency
diff --git a/packages/llm_gateway/ww_llm_gateway/routing.py b/packages/llm_gateway/ww_llm_gateway/routing.py
new file mode 100644
index 0000000..9737365
--- /dev/null
+++ b/packages/llm_gateway/ww_llm_gateway/routing.py
@@ -0,0 +1,29 @@
+"""档位路由：tier -> (provider, model)（ARCH §4.3）。
+
+M1 只读全局默认（config.tier_defaults，形如 "deepseek:deepseek-chat"）；
+作品级 / Skill 级覆盖留待后续（§4.3 三级解析）。
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from ww_config import get_settings
+
+from .types import Tier
+
+
+@dataclass(frozen=True)
+class Route:
+    provider: str
+    model: str
+
+
+def resolve_route(tier: Tier) -> Route:
+    spec = get_settings().tier_defaults.get(tier)
+    if not spec:
+        raise ValueError(f"no tier_defaults entry for tier={tier!r}")
+    provider, sep, model = spec.partition(":")
+    if not sep or not provider or not model:
+        raise ValueError(f"invalid tier route {spec!r}; expected 'provider:model'")
+    return Route(provider=provider, model=model)
diff --git a/packages/llm_gateway/ww_llm_gateway/types.py b/packages/llm_gateway/ww_llm_gateway/types.py
new file mode 100644
index 0000000..9b09022
--- /dev/null
+++ b/packages/llm_gateway/ww_llm_gateway/types.py
@@ -0,0 +1,78 @@
+"""LLM 网关统一接口契约（C1 / ARCH §4.1）——snake_case，Pydantic v2。
+
+上层（编排器/Agent）只碰这些类型，永不接触具体厂商字段。Agent 只声明 `tier`，
+不传具体 model（不变量 ②）。
+"""
+
+from __future__ import annotations
+
+import uuid
+from typing import Literal
+
+from pydantic import BaseModel, ConfigDict, Field
+
+Tier = Literal["writer", "analyst", "light"]
+
+
+class Block(BaseModel):
+    """一个 prompt 文本块；`cache=True` 标记缓存断点前的稳定块（ARCH §4.6）。"""
+
+    text: str
+    cache: bool = False
+
+
+class Scope(BaseModel):
+    """调用作用域。原型单用户：`user_id` 可固定 stub，`project_id` 可空。"""
+
+    user_id: uuid.UUID
+    project_id: uuid.UUID | None = None
+
+
+class LlmRequest(BaseModel):
+    """统一请求。`system` 稳定块在前（断点前），`input` 易变内容在后。"""
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    tier: Tier
+    input: str | list[Block]
+    system: list[Block] = Field(default_factory=list)
+    stream: bool = False
+    output_schema: type[BaseModel] | None = None
+    thinking: bool = False
+    max_tokens: int | None = None
+    scope: Scope
+
+
+class Usage(BaseModel):
+    """一次调用的用量与成本（落 usage_ledger，ARCH §4.8）。"""
+
+    provider: str
+    model: str
+    input_tokens: int
+    output_tokens: int
+    cache_read_tokens: int = 0
+    cost_minor: int  # 最小货币单位（如分/cent）
+    currency: str
+
+
+class ServedBy(BaseModel):
+    """实际服务方；`fell_back` 标记是否走了回退链（M5 才有回退，M1 恒 False）。"""
+
+    provider: str
+    model: str
+    fell_back: bool = False
+
+
+class LlmResponse(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    text: str
+    parsed: BaseModel | None = None
+    usage: Usage
+    served_by: ServedBy
+
+
+class Delta(BaseModel):
+    """流式增量：归一各家 SSE 的统一 token 块。"""
+
+    text: str
diff --git a/tests/test_m1_e2e.py b/tests/test_m1_e2e.py
new file mode 100644
index 0000000..f74c7f9
--- /dev/null
+++ b/tests/test_m1_e2e.py
@@ -0,0 +1,256 @@
+"""M1 端到端：立项 → 写一章草稿(mock 网关) → 自动保存（真实 DB，零 token）。
+
+证明 M1 闭环：`POST /projects` → `GET /projects/{id}` → `POST .../draft`(SSE) →
+`PUT .../draft`(自动保存)，且 DB 为真源（projects/chapters/usage_ledger 落库）。
+
+确定性 & 零成本：网关用**真实 `Gateway`** 但注入一个吐固定 `Delta` 的假适配器
+（不联网、不花 token），其 `ledger` 是**真实 `SqlAlchemyLedgerSink`** —— 闭环含用量
+记账也被走通。无 DB 时跳过（对齐 `tests/test_jobs_integration.py`）。
+
+坑（见 memory/gotchas）：
+- `ASGITransport` 不跑 lifespan → 用 `asgi-lifespan` 的 `LifespanManager` 触发
+  `seed_stub_user`，否则 `projects.owner_id`/`usage_ledger.owner_id` FK 报错。
+- `get_sessionmaker` 缓存的 async engine 绑定首个事件循环；每个 DB 测试清缓存重建、
+  结束 dispose。
+- 网关 `SqlAlchemyLedgerSink` 只 `flush()` 不 `commit()`（写库事务由编排层控制，
+  见不变量）；draft 端点在 SSE 流耗尽后对**请求 session** `commit()`，usage_ledger
+  方落库。本测试用请求 session 做 ledger，故无需手动提交——即对该提交的回归校验。
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator
+from typing import Annotated
+
+import httpx
+import pytest
+from asgi_lifespan import LifespanManager
+from fastapi import Depends
+from sqlalchemy import delete, func, select
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker
+from ww_db import get_session, get_sessionmaker
+from ww_db.models import Chapter, Project, UsageLedger
+from ww_llm_gateway import (
+    Gateway,
+    SqlAlchemyLedgerSink,
+    resolve_route,
+)
+from ww_llm_gateway.adapters.base import (
+    Capabilities,
+    ProviderResult,
+    ProviderUsage,
+    StreamChunk,
+)
+from ww_llm_gateway.types import LlmRequest
+
+# writer 档位的全局默认 provider（config.tier_defaults["writer"] = "deepseek:..."）。
+_WRITER_PROVIDER = "deepseek"
+# 确定性流式 token（不联网、不花 token）。
+_TOKENS = ["第", "一", "章", "：", "开端。"]
+# 假适配器在末尾块回报的 token 数（喂记账，证明用量闭环走通）。
+_FAKE_INPUT_TOKENS = 7
+_FAKE_OUTPUT_TOKENS = 5
+
+
+class _FakeStreamingAdapter:
+    """实现 `ProviderAdapter` Protocol：吐固定 `StreamChunk`，绝不联网。
+
+    末尾块带 `ProviderUsage` → 网关据此记账（cost 经 pricing 表，未知 model→0）。
+    """
+
+    provider = _WRITER_PROVIDER
+
+    def capabilities(self) -> Capabilities:
+        return Capabilities()
+
+    async def complete(self, req: LlmRequest, model: str) -> ProviderResult:
+        # M1 草稿走 stream；complete 不参与本 E2E，仍给确定性实现。
+        return ProviderResult(
+            text="".join(_TOKENS),
+            usage=ProviderUsage(input_tokens=_FAKE_INPUT_TOKENS, output_tokens=_FAKE_OUTPUT_TOKENS),
+        )
+
+    async def stream(self, req: LlmRequest, model: str) -> AsyncIterator[StreamChunk]:
+        for token in _TOKENS:
+            yield StreamChunk(text=token)
+        # 末尾用量块（text=""）：网关收尾时落 1 条 usage_ledger。
+        yield StreamChunk(
+            usage=ProviderUsage(input_tokens=_FAKE_INPUT_TOKENS, output_tokens=_FAKE_OUTPUT_TOKENS)
+        )
+
+
+@pytest.fixture
+async def e2e_sm() -> AsyncIterator[async_sessionmaker[AsyncSession]]:
+    """真实 DB session 工厂；无 pg 时跳过（每测试清缓存重建 engine、结束 dispose）。"""
+    get_sessionmaker.cache_clear()
+    maker = get_sessionmaker()
+    try:
+        async with maker() as probe:
+            await probe.execute(select(1))
+    except Exception:
+        pytest.skip("postgres not reachable")
+    yield maker
+    await maker.kw["bind"].dispose()
+    get_sessionmaker.cache_clear()
+
+
+def _parse_sse(raw: str) -> list[tuple[str, str]]:
+    """把 text/event-stream 原文解析为 `(event, data)` 帧列表。"""
+    frames: list[tuple[str, str]] = []
+    event: str | None = None
+    data: str | None = None
+    for line in raw.splitlines():
+        if line.startswith("event:"):
+            event = line[len("event:") :].strip()
+        elif line.startswith("data:"):
+            data = line[len("data:") :].strip()
+        elif line == "":
+            if event is not None and data is not None:
+                frames.append((event, data))
+            event, data = None, None
+    if event is not None and data is not None:
+        frames.append((event, data))
+    return frames
+
+
+async def test_m1_closed_loop_project_to_draft_to_autosave(
+    e2e_sm: async_sessionmaker[AsyncSession],
+) -> None:
+    import json
+
+    from ww_api.main import create_app
+    from ww_api.services.project_deps import get_writer_gateway
+
+    # 注入：真实 Gateway + 假适配器 + 真实 ledger。ledger 用**请求 session**
+    # （`Depends(get_session)` 被 FastAPI 按请求缓存，与 draft 端点同一实例）——
+    # 故由端点流末的 `session.commit()` 落库，不在测试里手动提交。这正是对端点
+    # 提交修复的回归校验：若端点不提交，下面 usage_ledger 断言会失败。
+    def _override_gateway(
+        session: Annotated[AsyncSession, Depends(get_session)],
+    ) -> Gateway:
+        return Gateway(
+            adapters={_WRITER_PROVIDER: _FakeStreamingAdapter()},
+            ledger=SqlAlchemyLedgerSink(session),
+            resolver=resolve_route,
+        )
+
+    app = create_app()
+    app.dependency_overrides[get_writer_gateway] = _override_gateway
+
+    transport = httpx.ASGITransport(app=app)
+    # LifespanManager 触发 lifespan → seed_stub_user（owner_id FK 依赖它）。
+    async with LifespanManager(app):
+        async with httpx.AsyncClient(transport=transport, base_url="http://test") as client:
+            # 1) 立项 → 201。唯一标题避免跨次干扰。
+            title = "M1 E2E 闭环验证作品"
+            create_resp = await client.post(
+                "/projects",
+                json={
+                    "title": title,
+                    "genre": "玄幻",
+                    "logline": "少年逆袭",
+                    "selling_points": ["爽点密集"],
+                },
+            )
+            assert create_resp.status_code == 201
+            created = create_resp.json()
+            project_id = created["id"]
+            assert created["title"] == title
+
+            # 2) GET 详情 → 200，与立项一致。
+            get_resp = await client.get(f"/projects/{project_id}")
+            assert get_resp.status_code == 200
+            fetched = get_resp.json()
+            assert fetched["id"] == project_id
+            assert fetched["title"] == title
+            assert fetched["genre"] == "玄幻"
+            assert fetched["selling_points"] == ["爽点密集"]
+
+            # 3) 流式写章草稿 → 消费 SSE：≥1 token 帧 + 终结 done 帧，重组文本。
+            draft_resp = await client.post(f"/projects/{project_id}/chapters/1/draft")
+            assert draft_resp.status_code == 200
+            assert draft_resp.headers["content-type"].startswith("text/event-stream")
+            frames = _parse_sse(draft_resp.text)
+            token_frames = [d for (ev, d) in frames if ev == "token"]
+            done_frames = [d for (ev, d) in frames if ev == "done"]
+            error_frames = [d for (ev, d) in frames if ev == "error"]
+            assert len(token_frames) >= 1
+            assert len(done_frames) == 1
+            assert error_frames == []
+            streamed_text = "".join(json.loads(d)["text"] for d in token_frames)
+            assert streamed_text == "".join(_TOKENS)
+            # done 帧带累计长度。
+            assert json.loads(done_frames[0])["length"] == len(streamed_text)
+            # 不手动提交：draft 端点流末 `session.commit()` 已把 usage_ledger 落库。
+
+            # 4) 自动保存草稿 ← 重组文本 → 200 DraftResponse{status:'draft', version:1}。
+            save_resp = await client.put(
+                f"/projects/{project_id}/chapters/1/draft",
+                json={"text": streamed_text},
+            )
+            assert save_resp.status_code == 200
+            saved = save_resp.json()
+            assert saved["project_id"] == project_id
+            assert saved["chapter_no"] == 1
+            assert saved["status"] == "draft"
+            assert saved["version"] == 1
+            assert saved["length"] == len(streamed_text)
+
+            # PUT 再次 → 幂等：不新增章节版本（覆盖同一行）。
+            save_resp2 = await client.put(
+                f"/projects/{project_id}/chapters/1/draft",
+                json={"text": streamed_text},
+            )
+            assert save_resp2.status_code == 200
+            assert save_resp2.json()["version"] == 1
+
+    # 5) DB 断言（经 e2e session 查询）——DB 是唯一真源。
+    project_uuid = created["id"]
+    async with e2e_sm() as verify:
+        # projects 行存在且字段一致。
+        project_row = (
+            await verify.execute(select(Project).where(Project.id == project_uuid))
+        ).scalar_one()
+        assert project_row.title == title
+
+        # chapters 草稿行：status='draft', version=1, content=已保存文本。
+        chapter_rows = (
+            (await verify.execute(select(Chapter).where(Chapter.project_id == project_uuid)))
+            .scalars()
+            .all()
+        )
+        assert len(chapter_rows) == 1  # 幂等：两次 PUT 仍一行
+        chapter = chapter_rows[0]
+        assert chapter.chapter_no == 1
+        assert chapter.status == "draft"
+        assert chapter.version == 1
+        assert chapter.content == streamed_text
+
+        # usage_ledger：草稿流至少写 1 条（用量记账闭环走通）。
+        ledger_count = (
+            await verify.execute(
+                select(func.count())
+                .select_from(UsageLedger)
+                .where(UsageLedger.project_id == project_uuid)
+            )
+        ).scalar_one()
+        assert ledger_count >= 1
+        ledger_row = (
+            (
+                await verify.execute(
+                    select(UsageLedger).where(UsageLedger.project_id == project_uuid)
+                )
+            )
+            .scalars()
+            .first()
+        )
+        assert ledger_row is not None
+        assert ledger_row.provider == _WRITER_PROVIDER
+        assert ledger_row.input_tokens == _FAKE_INPUT_TOKENS
+        assert ledger_row.output_tokens == _FAKE_OUTPUT_TOKENS
+
+    # 清理：usage_ledger 的 project FK 无级联 → 先删它，再删 project（chapters 经 FK 级联）。
+    async with e2e_sm() as cleanup:
+        await cleanup.execute(delete(UsageLedger).where(UsageLedger.project_id == project_uuid))
+        await cleanup.execute(delete(Project).where(Project.id == project_uuid))
+        await cleanup.commit()