Port presence feature forward from orphaned agent/dd2a126b57a1 branch

Recovers a real August feature that was never merged (3 WIP checkpoints +
fix commit 47ab492, 225 commits behind main): a Redis-backed heartbeat
presence store plus /api/presence/{heartbeat,presence,status} endpoints.
Ported onto current main's conventions rather than rebuilt:

- app/models/user.py: last_seen_at/status_message columns (unchanged from
  the orphaned branch; still compatible with the current model).
- alembic/versions/0006_add_presence_columns.py: same column migration,
  rebased as the new head after 0005_add_world_content_tables (was
  2a9b9873244a off 0001_initial).
- app/routers/presence.py: same three endpoints, updated to the DbDep
  pattern other routers (avatar.py) now use; retains the 47ab492 fix
  (is_online_batch via MGET, correct StatusUpdate Field annotation).
- app/services/presence.py: real adaptation beyond a mechanical port --
  the original PresenceStore required an explicit connect() that nothing
  ever called (main.py's lifespan never wired it up), so every real
  request would have raised "not connected". Made the Redis client connect
  lazily on first use instead, and cache it per running event loop rather
  than globally: this repo's TestClient(app) pattern (no `with` block) can
  dispatch different requests onto different event loops, which broke a
  single cached connection the same way NullPool's comment in conftest.py
  describes for asyncpg. Dropped the unused get_presence()/
  update_status_message() methods and the never-wired init/shutdown
  singleton helpers -- dead code even in the original.
- tests/conftest.py: extended the existing ephemeral-Postgres bootstrap to
  also bring up ephemeral Redis (docker-compose.yml already defined the
  service) when REDIS_URL isn't already reachable, so `pytest` stays
  hermetic.
- tests/test_presence.py: heartbeat sets online, batch presence over
  mixed online/offline/unknown ids (the exact 47ab492 regression), status
  message update/strip/length-limit, and TTL expiry via a short-TTL
  PresenceStore instance (real Redis SET/EXISTS/MGET, no 60s sleep needed).

Verified: full backend pytest suite (144 tests) passes against a fresh
Postgres+Redis. Also drove the real endpoints end-to-end against a running
uvicorn + docker-compose Postgres/Redis: heartbeat -> online, batch query,
forced TTL expiry -> offline with last_seen_at retained, status message
trim + 422 over 140 chars.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y2FESCaMXeJpC9gSctiEeD

backend/alembic/versions/0006_add_presence_columns.py

@@ -0,0 +1,33 @@
+"""add presence columns to users
+
+Revision ID: 0006_add_presence_columns
+Revises: 0005_add_world_content_tables
+Create Date: 2026-09-14 00:00:00.000000
+
+Ported forward from the orphaned agent/dd2a126b57a1 branch (originally
+2a9b9873244a_add_presence_columns.py, down_revision 0001_initial) and
+rebased onto the current migration head. See app/services/presence.py /
+app/routers/presence.py.
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+# revision identifiers, used by Alembic.
+revision: str = "0006_add_presence_columns"
+down_revision: Union[str, None] = "0005_add_world_content_tables"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    """Add last_seen_at and status_message columns to users table."""
+    op.add_column("users", sa.Column("last_seen_at", sa.DateTime(timezone=True), nullable=True))
+    op.add_column("users", sa.Column("status_message", sa.Text(), nullable=True))
+
+
+def downgrade() -> None:
+    """Remove presence columns from users table."""
+    op.drop_column("users", "status_message")
+    op.drop_column("users", "last_seen_at")

backend/app/main.py

@@ -6,7 +6,7 @@ from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
 
 from app.config import settings
-from app.routers import auth, users, worlds, assets, avatar, world_content, internal
+from app.routers import auth, users, worlds, assets, avatar, world_content, internal, presence
 
 
 from app.services.game_servers import supervisor
@@ -44,6 +44,7 @@ app.include_router(world_content.router, prefix="/api/worlds", tags=["world-cont
 app.include_router(assets.router, prefix="/api/assets", tags=["assets"])
 app.include_router(avatar.router, prefix="/api/avatar", tags=["avatar"])
 app.include_router(internal.router, prefix="/api/internal", tags=["internal"])
+app.include_router(presence.router, prefix="/api/presence", tags=["presence"])
 
 
 @app.get("/api/health")

backend/app/models/user.py

@@ -1,7 +1,7 @@
 from datetime import date
 from typing import Optional
 
-from sqlalchemy import Column, Integer, String, Boolean, DateTime, Date
+from sqlalchemy import Column, Integer, String, Boolean, DateTime, Date, Text
 from sqlalchemy.orm import relationship
 from sqlalchemy.sql import func
 
@@ -20,6 +20,10 @@ class User(Base):
     created_at = Column(DateTime(timezone=True), server_default=func.now())
     updated_at = Column(DateTime(timezone=True), onupdate=func.now())
 
+    # Presence / status (backend/app/services/presence.py, app/routers/presence.py)
+    last_seen_at = Column(DateTime(timezone=True), nullable=True)
+    status_message = Column(Text, nullable=True)
+
     sessions = relationship("Session", back_populates="user", cascade="all, delete-orphan")
     refresh_tokens = relationship("RefreshToken", back_populates="user", cascade="all, delete-orphan")
     worlds = relationship("World", back_populates="owner")

backend/app/routers/presence.py

@@ -0,0 +1,146 @@
+"""Presence & status router.
+
+Endpoints:
+- POST  /api/presence/heartbeat  -- record a heartbeat (sets Redis TTL key)
+- GET   /api/presence/presence   -- query presence for a list of user ids
+- PATCH /api/presence/status     -- update the authenticated user's status message
+"""
+
+from datetime import datetime, timezone
+from typing import Annotated
+
+from fastapi import APIRouter, HTTPException, status, Depends
+from pydantic import BaseModel, Field, field_validator
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.database import get_db
+from app.dependencies import CurrentUser
+from app.models.user import User
+from app.services.presence import presence_store
+
+router = APIRouter()
+
+DbDep = Annotated[AsyncSession, Depends(get_db)]
+
+# ──────────────────────────── Pydantic schemas ────────────────────────────────
+
+_STATUS_MAX_LEN = 140
+
+
+class HeartbeatRequest(BaseModel):
+    """Empty body -- the caller's identity comes from the JWT."""
+
+
+class PresenceUser(BaseModel):
+    user_id: int
+    is_online: bool
+    last_seen_at: datetime | None
+    status_message: str | None
+
+
+class PresenceResponse(BaseModel):
+    users: list[PresenceUser]
+
+
+class StatusUpdate(BaseModel):
+    status_message: str | None = Field(default=None, max_length=_STATUS_MAX_LEN)
+
+    @field_validator("status_message")
+    @classmethod
+    def strip_message(cls, v: str | None) -> str | None:
+        if v is not None:
+            v = v.strip()
+        return v
+
+
+# ──────────────────────────────── Endpoints ───────────────────────────────────
+
+
+@router.post("/heartbeat", response_model=dict)
+async def heartbeat(
+    _req: HeartbeatRequest,
+    current_user: CurrentUser,
+    db: DbDep,
+):
+    """Record a heartbeat for the authenticated user.
+
+    Sets a Redis key with a TTL so the server can derive online/offline
+    status without trusting client claims. Also persists last_seen_at to
+    the database.
+    """
+    # Update Redis TTL (presence heartbeat)
+    await presence_store.heartbeat(current_user.id)
+
+    # Persist last_seen_at to the database
+    now = datetime.now(timezone.utc)
+    result = await db.execute(select(User).where(User.id == current_user.id))
+    user = result.scalar_one_or_none()
+    if user:
+        user.last_seen_at = now
+        await db.commit()
+
+    return {"status": "ok", "user_id": current_user.id}
+
+
+@router.get("/presence", response_model=PresenceResponse)
+async def get_presence(
+    user_ids: str,
+    current_user: CurrentUser,
+    db: DbDep,
+):
+    """Return presence info for a comma-separated list of user ids.
+
+    Online/offline is derived from the Redis heartbeat TTL.
+    last_seen_at and status_message come from the database.
+    """
+    ids = [int(uid.strip()) for uid in user_ids.split(",") if uid.strip()]
+    if not ids:
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail="user_ids must contain at least one integer",
+        )
+
+    # Check Redis for online status (single round trip via MGET).
+    online_ids = await presence_store.is_online_batch(ids)
+
+    # Fetch DB records for last_seen_at and status_message
+    result = await db.execute(select(User).where(User.id.in_(ids)))
+    users = {u.id: u for u in result.scalars().all()}
+
+    presence_users = []
+    for uid in ids:
+        user = users.get(uid)
+        if user is None:
+            continue
+        presence_users.append(
+            PresenceUser(
+                user_id=uid,
+                is_online=uid in online_ids,
+                last_seen_at=user.last_seen_at,
+                status_message=user.status_message,
+            )
+        )
+
+    return PresenceResponse(users=presence_users)
+
+
+@router.patch("/status", response_model=dict)
+async def update_status(
+    body: StatusUpdate,
+    current_user: CurrentUser,
+    db: DbDep,
+):
+    """Update the authenticated user's status message (length-limited)."""
+    result = await db.execute(select(User).where(User.id == current_user.id))
+    user = result.scalar_one_or_none()
+    if not user:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail="User not found",
+        )
+
+    user.status_message = body.status_message
+    await db.commit()
+
+    return {"status": "ok", "status_message": body.status_message}

backend/app/services/presence.py

@@ -0,0 +1,145 @@
+"""Redis-backed presence store.
+
+Presence is determined by a heartbeat TTL key in Redis:
+- heartbeat(user_id) sets a TTL key per user (default 60 s).
+- is_online(user_id) / is_online_batch(user_ids) check whether the key(s)
+  exist.
+
+status_message and last_seen_at are persisted in the ``users`` table by the
+router (app/routers/presence.py); this store only owns the Redis side.
+
+The Redis client is created lazily, on first actual use, rather than
+requiring an explicit ``connect()`` call during app startup. This matters
+because FastAPI's lifespan handler does not run under the plain
+``TestClient(app)`` instantiation this codebase's tests use (no
+``with`` block), so anything that only connected during lifespan startup
+would never connect in tests. Lazy connection means the store works
+identically whether the app was started for real or is being driven by
+pytest.
+
+The client is also cached per-*event loop* rather than globally: the same
+plain ``TestClient(app)`` pattern can dispatch different requests onto
+different event loops (see the NullPool comment on ``_override_db`` in
+tests/conftest.py for the analogous asyncpg problem), and an
+``aioredis.Redis`` connection, like an asyncpg connection, breaks with
+"Event loop is closed" if reused across loops. In the real running server
+there is exactly one event loop for the process lifetime, so this
+degrades to a single cached, pooled connection there.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+
+import redis.asyncio as aioredis
+
+from app.config import settings
+
+logger = logging.getLogger(__name__)
+
+# Key prefix for heartbeat entries.
+_PRESENCE_KEY_PREFIX = "presence:heartbeat:"
+
+# Default heartbeat TTL in seconds (1 minute).
+# Clients should heartbeat at least once per interval.
+_DEFAULT_HEARTBEAT_TTL = 60
+
+
+class PresenceStore:
+    """Redis-backed presence store with async, lazily-connected operations."""
+
+    def __init__(self, redis_url: str | None = None, heartbeat_ttl: int = _DEFAULT_HEARTBEAT_TTL):
+        # redis_url is resolved from settings at first use (not at
+        # construction time) so tests that override settings.REDIS_URL
+        # before the first call still take effect -- this module is
+        # imported (and the module-level singleton below constructed)
+        # well before any such override runs.
+        self._redis_url = redis_url
+        self._heartbeat_ttl = heartbeat_ttl
+        # Keyed by id(running loop) -- see the module docstring.
+        self._clients: dict[int, aioredis.Redis] = {}
+        self._connect_lock = asyncio.Lock()
+
+    # ------------------------------------------------------------------
+    # Lifecycle
+    # ------------------------------------------------------------------
+
+    async def _ensure_client(self) -> aioredis.Redis:
+        """Return a connected client for the current event loop, connecting
+        on first use per loop."""
+        loop_key = id(asyncio.get_running_loop())
+        client = self._clients.get(loop_key)
+        if client is None:
+            async with self._connect_lock:
+                client = self._clients.get(loop_key)
+                if client is None:
+                    url = self._redis_url or settings.REDIS_URL
+                    client = aioredis.from_url(
+                        url,
+                        encoding="utf-8",
+                        decode_responses=True,
+                    )
+                    try:
+                        await client.ping()
+                    except Exception:
+                        logger.warning("Redis ping failed for %s", url)
+                        raise
+                    self._clients[loop_key] = client
+        return client
+
+    async def close(self) -> None:
+        """Close the Redis connection for the current event loop, if one was
+        opened."""
+        loop_key = id(asyncio.get_running_loop())
+        client = self._clients.pop(loop_key, None)
+        if client is not None:
+            await client.aclose()
+
+    # ------------------------------------------------------------------
+    # Heartbeat
+    # ------------------------------------------------------------------
+
+    async def heartbeat(self, user_id: int) -> None:
+        """Record a heartbeat for *user_id*.
+
+        Sets a TTL key so that the user is considered online as long as
+        heartbeats arrive at least once per heartbeat_ttl seconds.
+        """
+        client = await self._ensure_client()
+        await client.set(self._heartbeat_key(user_id), "1", ex=self._heartbeat_ttl)
+
+    # ------------------------------------------------------------------
+    # Online checks
+    # ------------------------------------------------------------------
+
+    async def is_online(self, user_id: int) -> bool:
+        """Return True if the user has a valid heartbeat key."""
+        client = await self._ensure_client()
+        exists = await client.exists(self._heartbeat_key(user_id))
+        return bool(exists)
+
+    async def is_online_batch(self, user_ids: list[int]) -> set[int]:
+        """Return the subset of user_ids that currently have a valid
+        heartbeat key. Uses MGET (one round trip) rather than one EXISTS
+        per user.
+        """
+        if not user_ids:
+            return set()
+        client = await self._ensure_client()
+        keys = [self._heartbeat_key(uid) for uid in user_ids]
+        values = await client.mget(keys)
+        return {uid for uid, v in zip(user_ids, values) if v is not None}
+
+    # ------------------------------------------------------------------
+    # Helpers
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def _heartbeat_key(user_id: int) -> str:
+        return f"{_PRESENCE_KEY_PREFIX}{user_id}"
+
+
+# Singleton instance shared by the router, so all requests share the same
+# connection pool.
+presence_store = PresenceStore()

backend/tests/conftest.py

@@ -1,15 +1,18 @@
-"""Shared pytest fixtures + hermetic-DB bootstrap.
+"""Shared pytest fixtures + hermetic-DB/Redis bootstrap.
 
 The integration tests hit a real Postgres (Postgres-specific types: JSONB,
-native ENUMs). To make ``cd backend && pytest`` work from a clean checkout
-with no manual setup, ``pytest_configure`` below brings up an ephemeral
-Postgres via docker compose and runs the migrations IF ``DATABASE_URL`` is
-not already pointing at a reachable database. When it is (dev loop, or the
-checkpoint runner which stands the DB up itself), that path is skipped and
-the existing database is reused.
+native ENUMs) and, for presence tests, a real Redis (TTL semantics aren't
+worth faking). To make ``cd backend && pytest`` work from a clean checkout
+with no manual setup, ``pytest_configure`` below brings up ephemeral
+Postgres/Redis containers via docker compose and runs the migrations IF
+``DATABASE_URL``/``REDIS_URL`` aren't already pointing at something
+reachable. When they are (dev loop, or the checkpoint runner which stands
+these up itself), that path is skipped and the existing services are
+reused.
 
-If Docker isn't available and no DB is reachable, the whole DB-backed suite
-is skipped with a clear message rather than erroring out 60+ times.
+If Docker isn't available and nothing is reachable, the whole
+DB/Redis-backed suite is skipped with a clear message rather than erroring
+out dozens of times.
 """
 
 import asyncio
@@ -44,8 +47,9 @@ TEST_TABLES = (
 BACKEND_DIR = Path(__file__).resolve().parents[1]
 _COMPOSE_PROJECT = "agp-pytest"
 _EPHEMERAL_PG_PORT = 25432
+_EPHEMERAL_REDIS_PORT = 26379
 
-# Set by _ensure_database() when it starts a container it owns; torn down in
+# Set by pytest_configure() when it starts container(s) it owns; torn down in
 # pytest_unconfigure.
 _started_container = False
 _db_unavailable_reason: str | None = None
@@ -64,6 +68,11 @@ def _db_reachable(url: str) -> bool:
     return _port_open(parts.hostname or "localhost", parts.port or 5432)
 
 
+def _redis_reachable(url: str) -> bool:
+    parts = urlsplit(url)
+    return _port_open(parts.hostname or "localhost", parts.port or 6379)
+
+
 def _run(cmd: list[str], **kw) -> subprocess.CompletedProcess:
     return subprocess.run(cmd, cwd=BACKEND_DIR, capture_output=True, text=True, **kw)
 
@@ -71,14 +80,17 @@ def _run(cmd: list[str], **kw) -> subprocess.CompletedProcess:
 def pytest_configure(config):
     global _started_container, _db_unavailable_reason
 
-    if _db_reachable(settings.DATABASE_URL):
-        return  # dev loop / checkpoint runner already provided one
+    need_db = not _db_reachable(settings.DATABASE_URL)
+    need_redis = not _redis_reachable(settings.REDIS_URL)
+
+    if not need_db and not need_redis:
+        return  # dev loop / checkpoint runner already provided both
 
     if not _run(["docker", "--version"]).returncode == 0:
         _db_unavailable_reason = (
-            "no reachable DATABASE_URL and docker is unavailable -- run the "
-            "checkpoint suite (scripts/run_checkpoint_tests.py) or start Postgres "
-            "manually"
+            "no reachable DATABASE_URL/REDIS_URL and docker is unavailable -- run "
+            "the checkpoint suite (scripts/run_checkpoint_tests.py) or start "
+            "Postgres/Redis manually"
         )
         return
 
@@ -86,12 +98,16 @@ def pytest_configure(config):
     env["COMPOSE_PROJECT_NAME"] = _COMPOSE_PROJECT
     env["POSTGRES_HOST_PORT"] = str(_EPHEMERAL_PG_PORT)
     env["POSTGRES_PORT"] = str(_EPHEMERAL_PG_PORT)
-    up = _run(["docker", "compose", "up", "-d", "postgres"], env=env)
+    env["REDIS_HOST_PORT"] = str(_EPHEMERAL_REDIS_PORT)
+
+    services = [s for s, need in (("postgres", need_db), ("redis", need_redis)) if need]
+    up = _run(["docker", "compose", "up", "-d", *services], env=env)
     if up.returncode != 0:
-        _db_unavailable_reason = f"docker compose up postgres failed:\n{up.stderr[-800:]}"
+        _db_unavailable_reason = f"docker compose up {' '.join(services)} failed:\n{up.stderr[-800:]}"
         return
     _started_container = True
 
+    if need_db:
         for _ in range(30):
             r = _run(
                 ["docker", "compose", "exec", "-T", "postgres",
@@ -113,6 +129,24 @@ def pytest_configure(config):
                    env={**os.environ, "DATABASE_URL": url})
         if mig.returncode != 0:
             _db_unavailable_reason = f"alembic upgrade head failed:\n{mig.stderr[-800:]}"
+            return
+
+    if need_redis:
+        for _ in range(30):
+            r = _run(
+                ["docker", "compose", "exec", "-T", "redis", "redis-cli", "ping"],
+                env=env,
+            )
+            if r.returncode == 0 and "PONG" in r.stdout:
+                break
+            time.sleep(1)
+        else:
+            _db_unavailable_reason = "ephemeral Redis never became ready"
+            return
+
+        redis_url = f"redis://localhost:{_EPHEMERAL_REDIS_PORT}/0"
+        os.environ["REDIS_URL"] = redis_url
+        settings.REDIS_URL = redis_url
 
 
 def pytest_unconfigure(config):
@@ -122,6 +156,7 @@ def pytest_unconfigure(config):
     env["COMPOSE_PROJECT_NAME"] = _COMPOSE_PROJECT
     env["POSTGRES_HOST_PORT"] = str(_EPHEMERAL_PG_PORT)
     env["POSTGRES_PORT"] = str(_EPHEMERAL_PG_PORT)
+    env["REDIS_HOST_PORT"] = str(_EPHEMERAL_REDIS_PORT)
     _run(["docker", "compose", "down", "-v"], env=env)
 
 

backend/tests/test_presence.py

@@ -0,0 +1,221 @@
+"""Integration tests for the presence feature:
+
+- POST  /api/presence/heartbeat
+- GET   /api/presence/presence
+- PATCH /api/presence/status
+
+Hits the real app against the configured Postgres *and* a real Redis
+(conftest.py brings up an ephemeral instance of each when neither is
+already reachable) -- presence is fundamentally "does a TTL key exist in
+Redis", which isn't worth faking with a mock.
+
+Also covers the regression fixed in 47ab492 ("Fix StatusUpdate field
+annotation and is_online batch-call mismatch"): the router used to call
+``PresenceStore.is_online(ids)`` with a *list* of ids and then evaluate
+``uid in online_ids`` against the returned bool, which raised a TypeError
+on the first real multi-id batch call, and ``StatusUpdate.status_message``
+used the invalid annotation ``str | Field(...)``. Both are exercised below.
+"""
+
+import asyncio
+
+import pytest
+from fastapi.testclient import TestClient
+
+from app.main import app
+from app.services.presence import PresenceStore
+
+
+def _register(client: TestClient, username: str) -> dict:
+    resp = client.post(
+        "/api/auth/register",
+        json={
+            "username": username,
+            "email": f"{username}@example.com",
+            "password": "secret123",
+            "date_of_birth": "1990-01-01",
+        },
+    )
+    assert resp.status_code == 201, resp.text
+    token = resp.json()["access_token"]
+    headers = {"Authorization": f"Bearer {token}"}
+    me = client.get("/api/users/me", headers=headers)
+    assert me.status_code == 200, me.text
+    return {"headers": headers, "id": me.json()["id"]}
+
+
+class TestHeartbeat:
+    def test_heartbeat_sets_online(self):
+        client = TestClient(app)
+        user = _register(client, "presence_hb1")
+
+        resp = client.post("/api/presence/heartbeat", json={}, headers=user["headers"])
+        assert resp.status_code == 200, resp.text
+        assert resp.json() == {"status": "ok", "user_id": user["id"]}
+
+        presence = client.get(
+            "/api/presence/presence",
+            params={"user_ids": str(user["id"])},
+            headers=user["headers"],
+        )
+        assert presence.status_code == 200, presence.text
+        body = presence.json()["users"]
+        assert len(body) == 1
+        assert body[0]["user_id"] == user["id"]
+        assert body[0]["is_online"] is True
+        assert body[0]["last_seen_at"] is not None
+
+    def test_requires_auth(self):
+        client = TestClient(app)
+        resp = client.post("/api/presence/heartbeat", json={})
+        assert resp.status_code == 401
+
+    def test_never_seen_user_is_offline_with_null_last_seen(self):
+        client = TestClient(app)
+        user = _register(client, "presence_never_hb")
+
+        presence = client.get(
+            "/api/presence/presence",
+            params={"user_ids": str(user["id"])},
+            headers=user["headers"],
+        )
+        assert presence.status_code == 200
+        body = presence.json()["users"][0]
+        assert body["is_online"] is False
+        assert body["last_seen_at"] is None
+
+
+class TestBatchPresence:
+    """Covers the batch-call path fixed in 47ab492 (is_online_batch via MGET)."""
+
+    def test_batch_query_mixes_online_offline_and_unknown_ids(self):
+        client = TestClient(app)
+        online_user = _register(client, "presence_batch_online")
+        offline_user = _register(client, "presence_batch_offline")
+
+        client.post("/api/presence/heartbeat", json={}, headers=online_user["headers"])
+
+        unknown_id = max(online_user["id"], offline_user["id"]) + 999_999
+        ids = f"{online_user['id']},{offline_user['id']},{unknown_id}"
+
+        resp = client.get(
+            "/api/presence/presence",
+            params={"user_ids": ids},
+            headers=online_user["headers"],
+        )
+        assert resp.status_code == 200, resp.text
+        by_id = {u["user_id"]: u for u in resp.json()["users"]}
+
+        # Unknown id is silently dropped (not in the DB), not a 500 or 404.
+        assert unknown_id not in by_id
+        assert by_id[online_user["id"]]["is_online"] is True
+        assert by_id[offline_user["id"]]["is_online"] is False
+
+    def test_empty_user_ids_is_a_400(self):
+        client = TestClient(app)
+        user = _register(client, "presence_empty_ids")
+        resp = client.get(
+            "/api/presence/presence",
+            params={"user_ids": ""},
+            headers=user["headers"],
+        )
+        assert resp.status_code == 400
+
+
+class TestStatusMessage:
+    def test_update_and_read_back_status_message(self):
+        client = TestClient(app)
+        user = _register(client, "presence_status1")
+
+        resp = client.patch(
+            "/api/presence/status",
+            json={"status_message": "  brb, making coffee  "},
+            headers=user["headers"],
+        )
+        assert resp.status_code == 200, resp.text
+        # The router strips whitespace (StatusUpdate.strip_message).
+        assert resp.json()["status_message"] == "brb, making coffee"
+
+        presence = client.get(
+            "/api/presence/presence",
+            params={"user_ids": str(user["id"])},
+            headers=user["headers"],
+        )
+        assert presence.json()["users"][0]["status_message"] == "brb, making coffee"
+
+    def test_status_message_over_max_length_is_rejected(self):
+        client = TestClient(app)
+        user = _register(client, "presence_status_toolong")
+        resp = client.patch(
+            "/api/presence/status",
+            json={"status_message": "x" * 141},
+            headers=user["headers"],
+        )
+        # StatusUpdate.status_message: str | None = Field(..., max_length=140)
+        # This is exactly the annotation 47ab492 fixed (it used to be the
+        # invalid `str | Field(...)`, which breaks model construction
+        # entirely rather than validating length).
+        assert resp.status_code == 422
+
+    def test_status_message_at_max_length_is_accepted(self):
+        client = TestClient(app)
+        user = _register(client, "presence_status_maxlen")
+        resp = client.patch(
+            "/api/presence/status",
+            json={"status_message": "x" * 140},
+            headers=user["headers"],
+        )
+        assert resp.status_code == 200, resp.text
+
+    def test_null_status_message_clears_it(self):
+        client = TestClient(app)
+        user = _register(client, "presence_status_clear")
+        client.patch("/api/presence/status", json={"status_message": "hi"}, headers=user["headers"])
+        resp = client.patch("/api/presence/status", json={"status_message": None}, headers=user["headers"])
+        assert resp.status_code == 200
+        assert resp.json()["status_message"] is None
+
+
+class TestPresenceStoreTTL:
+    """Exercises the real Redis TTL semantics directly against the service,
+    with a short heartbeat_ttl so the test doesn't need to sleep for the
+    real 60s default used in production. This is the cheap, real proxy for
+    "TTL expiry means offline" the task asked for -- same code path
+    (real Redis SET ... EX / EXISTS / MGET), just a much shorter window.
+    """
+
+    async def test_ttl_expiry_flips_online_to_offline(self):
+        store = PresenceStore(heartbeat_ttl=1)
+        try:
+            uid = 987654321  # doesn't need to exist in Postgres; Redis-only check
+            assert await store.is_online(uid) is False
+
+            await store.heartbeat(uid)
+            assert await store.is_online(uid) is True
+
+            await asyncio.sleep(1.5)
+            assert await store.is_online(uid) is False
+        finally:
+            await store.close()
+
+    async def test_is_online_batch_matches_is_online_semantics(self):
+        """Regression coverage for 47ab492: is_online_batch must return a
+        set of ids usable with `in`, not a bool."""
+        store = PresenceStore(heartbeat_ttl=30)
+        try:
+            online_uid, offline_uid = 111222333, 444555666
+            await store.heartbeat(online_uid)
+
+            result = await store.is_online_batch([online_uid, offline_uid])
+            assert isinstance(result, set)
+            assert online_uid in result
+            assert offline_uid not in result
+        finally:
+            await store.close()
+
+    async def test_is_online_batch_empty_list(self):
+        store = PresenceStore()
+        try:
+            assert await store.is_online_batch([]) == set()
+        finally:
+            await store.close()

scripts/editor/chunk_save_service.gd

@@ -0,0 +1,280 @@
+class_name ChunkSaveService
+extends Node
+## HTTP save path for the in-game editor
+## (docs/architecture/world_persistence_impl.md §8.1 chunk write / §8.2 D5).
+##
+## Owns the write half of editor persistence: single-chunk PUT with
+## `X-Base-Revision`/`base_revision`, batch PUT for a multi-chunk edit, and
+## the §5.4 optimistic-concurrency retry loop on `409`.
+##
+## Conflict handling ("re-fetch current chunk, re-apply the edit, retry"):
+## per §8.2 D5 this backend stores and accepts **whole chunks**, not
+## JSON-Patch deltas -- "lost-update protection comes from `base_revision`,
+## not from the patch". That means there is no delta to replay against a
+## freshly-fetched chunk; "re-apply the edit" is exactly "keep the locally
+## edited payload, learn the chunk's current revision, and retry at that
+## revision". `fetch_chunk()` is still called on a conflict (rather than
+## just trusting the 409 body's `current_revision`) so the editor's own
+## last-known state stays in sync with the store, matching what a real
+## client does before retrying a write it doesn't fully trust yet.
+##
+## Deliberately does **not** use `ApiClient.request()`: that helper discards
+## the parsed JSON body on any non-2xx response (see its `_make_result`
+## calls), and the `409` body's `current_revision`/`current_checksum` are
+## exactly what this retry loop needs. Uses a raw per-request `HTTPRequest`
+## child instead, the same pattern `scripts/http_chunk_source.gd` uses for
+## the read side.
+##
+## Must be added to the scene tree before use.
+
+## The world this service saves chunks for (string form of the backend's
+## numeric world id).
+var world_id: String = ""
+
+## The room slug within that world. world.md's default room is "main".
+var room_slug: String = "main"
+
+## Bounded retries on a 409 conflict -- refuses to loop forever against a
+## chunk under sustained contention.
+const MAX_RETRIES := 3
+
+
+## Fetch the current chunk at [param coord].
+##
+## Returns:
+##     {"ok": bool, "status": int, "model": ChunkEditorModel|null,
+##      "revision": int, "checksum": String, "detail": String}
+##
+## A `404` (unwritten coordinate) is reported as `ok: true, model: null` --
+## world_persistence_impl.md §8.1: "an unwritten coordinate is not an error".
+## Callers wanting a fresh chunk to edit should fall back to
+## `ChunkEditorModel.create_empty()` at `base_revision: 0` in that case.
+func fetch_chunk(coord: Vector3i) -> Dictionary:
+	var http := HTTPRequest.new()
+	http.timeout = 15.0
+	add_child(http)
+
+	var api := _api_client()
+	var headers := PackedStringArray(["Accept: application/octet-stream"])
+	if api.has_token():
+		headers.append("Authorization: " + str(api.bearer_token))
+
+	var url: String = str(api.base_url) + _chunk_path(coord)
+	var err := http.request(url, headers, HTTPClient.METHOD_GET)
+	if err != OK:
+		http.queue_free()
+		return _fetch_result(false, 0, null, 0, "", "request failed to start (%d)" % err)
+
+	var args: Array = await http.request_completed
+	var result_code: int = args[0]
+	var response_code: int = args[1]
+	var response_headers: PackedStringArray = args[2]
+	var body: PackedByteArray = args[3]
+	http.queue_free()
+
+	if result_code != HTTPRequest.RESULT_SUCCESS:
+		return _fetch_result(false, 0, null, 0, "", "transport error (%d)" % result_code)
+
+	if response_code == 404:
+		return _fetch_result(true, 404, null, 0, "", "not found (unwritten chunk)")
+
+	if response_code != 200:
+		return _fetch_result(false, response_code, null, 0, "", _error_detail_from_body(body))
+
+	var model := ChunkEditorModel.from_binary(body)
+	if model == null:
+		return _fetch_result(false, response_code, null, 0, "", "failed to decode chunk payload")
+
+	return _fetch_result(
+		true, 200, model,
+		_int_header(response_headers, "x-chunk-revision"),
+		_etag_from_headers(response_headers),
+		"OK",
+	)
+
+
+## Save one chunk. `model.encode_binary()` is sent as the new payload;
+## `base_revision` is the revision the editor last knew about (0 = create).
+##
+## Returns:
+##     {"ok": bool, "status": int, "revision": int, "checksum": String,
+##      "live_session": bool, "retried": bool, "detail": String}
+func save_chunk(model: ChunkEditorModel, coord: Vector3i, base_revision: int) -> Dictionary:
+	var attempt_revision := base_revision
+	var retried := false
+	for attempt in range(MAX_RETRIES + 1):
+		var body := {
+			"base_revision": attempt_revision,
+			"payload_b64": Marshalls.raw_to_base64(model.encode_binary()),
+		}
+		var resp := await _request_json(HTTPClient.METHOD_PUT, _chunk_path(coord), body)
+
+		if resp["status"] == 200:
+			var d: Dictionary = resp["data"]
+			return {
+				"ok": true, "status": 200,
+				"revision": int(d.get("revision", attempt_revision)),
+				"checksum": str(d.get("checksum", "")),
+				"live_session": bool(d.get("live_session", false)),
+				"retried": retried, "detail": "OK",
+			}
+
+		if resp["status"] == 409 and attempt < MAX_RETRIES:
+			var conflict: Dictionary = resp["data"] if resp["data"] is Dictionary else {}
+			var current_rev := int(conflict.get("current_revision", attempt_revision))
+			# Doc-mandated "re-fetch current chunk" step before retrying --
+			# keeps this service's view of the chunk current even though the
+			# retry itself only needs the revision number from the 409 body.
+			await fetch_chunk(coord)
+			attempt_revision = current_rev
+			retried = true
+			continue
+
+		return {"ok": false, "status": resp["status"], "revision": 0, "checksum": "",
+			"live_session": false, "retried": retried, "detail": _error_detail(resp)}
+
+	return {"ok": false, "status": 409, "revision": 0, "checksum": "", "live_session": false,
+		"retried": true, "detail": "conflict retry limit exceeded"}
+
+
+## Save multiple chunks in one atomic batch PUT (world_persistence_impl.md
+## §8.1 batch write: "an editor brush stroke or an object drag routinely
+## straddles a chunk boundary"). `entries` is an Array of
+## `{"coord": Vector3i, "model": ChunkEditorModel, "base_revision": int}`.
+## All-or-nothing: a `409` names the offending coordinate and nothing is
+## written, so only that entry's `base_revision` is refreshed before the
+## whole batch is retried.
+##
+## Returns:
+##     {"ok": bool, "status": int, "live_session": bool,
+##      "chunks": Array, "retried": bool, "detail": String}
+func save_batch(entries: Array) -> Dictionary:
+	var working: Array = []
+	for e in entries:
+		working.append((e as Dictionary).duplicate())
+	var retried := false
+
+	for attempt in range(MAX_RETRIES + 1):
+		var chunks_json: Array = []
+		for e in working:
+			var c: Vector3i = e["coord"]
+			var m: ChunkEditorModel = e["model"]
+			chunks_json.append({
+				"cx": c.x, "cy": c.y, "cz": c.z,
+				"base_revision": int(e["base_revision"]),
+				"payload_b64": Marshalls.raw_to_base64(m.encode_binary()),
+			})
+		var resp := await _request_json(HTTPClient.METHOD_PUT, _batch_path(), {"chunks": chunks_json})
+
+		if resp["status"] == 200:
+			var d: Dictionary = resp["data"]
+			return {"ok": true, "status": 200, "live_session": bool(d.get("live_session", false)),
+				"chunks": d.get("chunks", []), "retried": retried, "detail": "OK"}
+
+		if resp["status"] == 409 and attempt < MAX_RETRIES:
+			var conflict: Dictionary = resp["data"] if resp["data"] is Dictionary else {}
+			var current_rev := int(conflict.get("current_revision", 0))
+			var offending: Variant = conflict.get("coord")
+			if offending is Array and offending.size() == 3:
+				var oc := Vector3i(int(offending[0]), int(offending[1]), int(offending[2]))
+				for e in working:
+					if (e["coord"] as Vector3i) == oc:
+						await fetch_chunk(oc)
+						e["base_revision"] = current_rev
+						break
+			retried = true
+			continue
+
+		return {"ok": false, "status": resp["status"], "live_session": false, "chunks": [],
+			"retried": retried, "detail": _error_detail(resp)}
+
+	return {"ok": false, "status": 409, "live_session": false, "chunks": [],
+		"retried": true, "detail": "conflict retry limit exceeded"}
+
+
+# ─────────────────────────── internals ───────────────────────────
+
+## Resolved via an absolute node path, not the bare `ApiClient` autoload
+## identifier -- see the identical note on `HttpChunkSource._api_client()`
+## (bare autoload identifiers don't resolve when a script is compiled for a
+## headless `godot --script ...` run, which is how this service's smoke
+## test drives it).
+func _api_client() -> Node:
+	return get_node("/root/ApiClient")
+
+
+func _chunk_path(coord: Vector3i) -> String:
+	return "/api/worlds/%s/rooms/%s/chunks/%d/%d/%d" % [world_id, room_slug, coord.x, coord.y, coord.z]
+
+
+func _batch_path() -> String:
+	return "/api/worlds/%s/rooms/%s/chunks" % [world_id, room_slug]
+
+
+func _request_json(method: HTTPClient.Method, path: String, body_dict: Dictionary) -> Dictionary:
+	var http := HTTPRequest.new()
+	http.timeout = 15.0
+	add_child(http)
+
+	var api := _api_client()
+	var headers := PackedStringArray(["Content-Type: application/json", "Accept: application/json"])
+	if api.has_token():
+		headers.append("Authorization: " + str(api.bearer_token))
+
+	var url: String = str(api.base_url) + path
+	var err := http.request(url, headers, method, JSON.stringify(body_dict))
+	if err != OK:
+		http.queue_free()
+		return {"status": 0, "data": null, "raw": "request failed to start (%d)" % err}
+
+	var args: Array = await http.request_completed
+	var result_code: int = args[0]
+	var response_code: int = args[1]
+	var body_bytes: PackedByteArray = args[3]
+	http.queue_free()
+
+	if result_code != HTTPRequest.RESULT_SUCCESS:
+		return {"status": 0, "data": null, "raw": "transport error (%d)" % result_code}
+
+	var text := body_bytes.get_string_from_utf8()
+	var parsed: Variant = JSON.parse_string(text) if text != "" else null
+	return {"status": response_code, "data": parsed, "raw": text}
+
+
+func _error_detail(resp: Dictionary) -> String:
+	var data: Variant = resp.get("data")
+	if data is Dictionary and (data as Dictionary).has("detail"):
+		return str((data as Dictionary)["detail"])
+	if int(resp.get("status", 0)) == 0:
+		return str(resp.get("raw", "network error"))
+	return "HTTP error %d" % int(resp.get("status", 0))
+
+
+func _error_detail_from_body(body: PackedByteArray) -> String:
+	var text := body.get_string_from_utf8()
+	var parsed: Variant = JSON.parse_string(text) if text != "" else null
+	if parsed is Dictionary and (parsed as Dictionary).has("detail"):
+		return str((parsed as Dictionary)["detail"])
+	return text
+
+
+func _fetch_result(ok: bool, status: int, model: ChunkEditorModel, revision: int,
+		checksum: String, detail: String) -> Dictionary:
+	return {"ok": ok, "status": status, "model": model, "revision": revision,
+		"checksum": checksum, "detail": detail}
+
+
+func _int_header(headers: PackedStringArray, key: String) -> int:
+	var prefix := key + ":"
+	for h in headers:
+		if h.to_lower().begins_with(prefix):
+			return int(h.substr(h.find(":") + 1).strip_edges())
+	return 0
+
+
+func _etag_from_headers(headers: PackedStringArray) -> String:
+	for h in headers:
+		if h.to_lower().begins_with("etag:"):
+			var v := h.substr(h.find(":") + 1).strip_edges()
+			return v.trim_prefix('"').trim_suffix('"')
+	return ""

scripts/editor/ui/in_game_editor.gd

@@ -16,10 +16,12 @@ extends Control
 ##   InspectorPanel  -- property editor for the selected object
 ##   EditorCore      -- shared core (passed in, not owned/created here)
 ##
-## This UI (and `EditorCore`) never talk to the network -- persisting edits
-## is a sibling task's job (world_persistence_impl.md's batch-PUT save
-## path). Closing this editor just discards the UI; the in-memory
-## `ChunkEditorModel` it was pointed at is unaffected either way.
+## Persisting edits (world_persistence_impl.md's single/batch-PUT save path,
+## §8.1-8.2) is wired through `ChunkSaveService`, set via
+## `setup_persistence()`. Without it (e.g. the F2 scratch-chunk path, or
+## `smoke_test_editor_core.gd`'s bare `InGameEditor.new(core)`), the Save
+## button is present but disabled -- closing the editor just discards the
+## UI and the in-memory `ChunkEditorModel` is unaffected either way.
 
 signal closed
 
@@ -30,6 +32,18 @@ var _palette: PalettePanel
 var _inspector: InspectorPanel
 var _viewport: EditorViewport
 var _close_button: Button
+var _save_button: Button
+var _save_status_label: Label
+
+## Save-path state (world_persistence_impl.md §8.1/§8.2), set via
+## `setup_persistence()`. `_save_service == null` means there is nothing to
+## save to (see the class doc above) -- the Save button stays disabled.
+var _save_service: ChunkSaveService = null
+var _world_id: String = ""
+var _room_slug: String = "main"
+var _coord: Vector3i = Vector3i.ZERO
+var _base_revision: int = 0
+var _saving: bool = false
 
 
 func _init(p_core: EditorCore) -> void:
@@ -39,6 +53,21 @@ func _init(p_core: EditorCore) -> void:
 	mouse_filter = Control.MOUSE_FILTER_STOP
 
 
+## Wire this editor's Save button/Ctrl+S to a real backend round-trip.
+## `save_service` must already be in the scene tree (game.gd adds it as a
+## sibling before constructing this editor). `base_revision` is whatever the
+## caller last fetched for `coord` (0 if the chunk didn't exist yet).
+func setup_persistence(save_service: ChunkSaveService, world_id: String, room_slug: String,
+		coord: Vector3i, base_revision: int) -> void:
+	_save_service = save_service
+	_world_id = world_id
+	_room_slug = room_slug
+	_coord = coord
+	_base_revision = base_revision
+	if _save_button != null:
+		_save_button.disabled = false
+
+
 func _ready() -> void:
 	_build_ui()
 	_refresh_history_buttons()
@@ -66,11 +95,22 @@ func _build_ui() -> void:
 	_toolbar.size_flags_horizontal = Control.SIZE_EXPAND_FILL
 	top_bar.add_child(_toolbar)
 
+	_save_button = Button.new()
+	_save_button.text = "Save (Ctrl+S)"
+	_save_button.disabled = (_save_service == null)
+	_save_button.pressed.connect(_on_save_pressed)
+	top_bar.add_child(_save_button)
+
 	_close_button = Button.new()
 	_close_button.text = "Close (F2)"
 	_close_button.pressed.connect(_on_close_pressed)
 	top_bar.add_child(_close_button)
 
+	_save_status_label = Label.new()
+	_save_status_label.autowrap_mode = TextServer.AUTOWRAP_WORD
+	_save_status_label.visible = false
+	root_vbox.add_child(_save_status_label)
+
 	var body := HBoxContainer.new()
 	body.size_flags_vertical = Control.SIZE_EXPAND_FILL
 	root_vbox.add_child(body)
@@ -119,6 +159,9 @@ func _unhandled_key_input(event: InputEvent) -> void:
 	elif key.ctrl_pressed and key.keycode == KEY_Y:
 		_on_redo_pressed()
 		get_viewport().set_input_as_handled()
+	elif key.ctrl_pressed and key.keycode == KEY_S:
+		_on_save_pressed()
+		get_viewport().set_input_as_handled()
 	elif key.keycode == KEY_DELETE and core.selection.has_selection():
 		core.delete_object(core.selection.get_selected())
 		_on_edit_applied()
@@ -167,3 +210,41 @@ func _refresh_history_buttons() -> void:
 func _on_close_pressed() -> void:
 	closed.emit()
 	queue_free()
+
+
+## Single-chunk save (world_persistence_impl.md §8.1). `ChunkSaveService`
+## owns the actual HTTP round trip, retrying on `409` internally
+## (§8.2/§5.4) -- this just reflects the outcome and, per §6 D4, surfaces
+## the `live_session` flag as a warning instead of ignoring it.
+func _on_save_pressed() -> void:
+	if _save_service == null or _saving:
+		return
+	_saving = true
+	_save_button.disabled = true
+	_set_save_status("Saving…", Color(0.85, 0.85, 0.85))
+
+	var result: Dictionary = await _save_service.save_chunk(core.model, _coord, _base_revision)
+
+	_saving = false
+	_save_button.disabled = false
+
+	if result.get("ok", false):
+		_base_revision = int(result.get("revision", _base_revision))
+		var msg := "Saved (revision %d)." % _base_revision
+		if bool(result.get("retried", false)):
+			msg += " (resolved a save conflict automatically)"
+		if bool(result.get("live_session", false)):
+			msg += " This world has players in it right now -- they'll see your changes when the session restarts."
+			_set_save_status(msg, Color(1.0, 0.8, 0.25))
+		else:
+			_set_save_status(msg, Color(0.55, 1.0, 0.55))
+	else:
+		_set_save_status("Save failed: %s" % str(result.get("detail", "unknown error")), Color(1.0, 0.45, 0.45))
+
+
+func _set_save_status(text: String, color: Color) -> void:
+	if _save_status_label == null:
+		return
+	_save_status_label.text = text
+	_save_status_label.add_theme_color_override("font_color", color)
+	_save_status_label.visible = true

scripts/game.gd

@@ -39,12 +39,18 @@ var _build_editor: InGameEditor = null
 
 func _ready() -> void:
 	_show_build_stamp()
-	if Session.active:
+	if Session.editing:
+		_open_editor_for_world()
+	elif Session.active:
 		_setup_networked()
 
 
 func _unhandled_input(event: InputEvent) -> void:
 	if event.is_action_pressed("toggle_build_mode"):
+		# world_persistence_impl.md §6 D4: the non-joined "Edit world" entry
+		# already has its editor open from _ready() -- F2 here is the
+		# scratch/in-session build mode's own toggle and doesn't apply.
+		if not Session.editing:
 			get_viewport().set_input_as_handled()
 			_toggle_build_mode()
 		return
@@ -56,7 +62,11 @@ func _unhandled_input(event: InputEvent) -> void:
 	# against the scene change and looked like a crash.
 	if event.is_action_pressed("ui_cancel"):
 		get_viewport().set_input_as_handled()
-		if _build_editor != null and is_instance_valid(_build_editor):
+		if Session.editing:
+			# Non-joined editing has no play session to fall back into --
+			# closing it means leaving, same as the Close button.
+			_exit_editor_to_browser()
+		elif _build_editor != null and is_instance_valid(_build_editor):
 			_close_build_mode()
 		else:
 			_leave_world()
@@ -104,6 +114,64 @@ func _close_build_mode() -> void:
 	_player._refresh_mouse_mode()
 
 
+## "Edit world" entry point from the world browser
+## (world_persistence_impl.md §6 D4 / §9 "Client/editor task"). Reached only
+## via Session.editing (set by world_list.gd's "Edit World" button, which
+## never calls /join) -- non-joined, no ENet connection, no dedicated
+## server. Loads the real chunk at (0,0,0) in the "main" room over HTTP so
+## there is something real to edit and save, falling back to a fresh empty
+## chunk at base_revision 0 for an unwritten coordinate (§8.1: "an unwritten
+## coordinate is not an error").
+func _open_editor_for_world() -> void:
+	Input.mouse_mode = Input.MOUSE_MODE_VISIBLE
+	_player.set_physics_process(false)
+	_player.set_process_unhandled_input(false)
+
+	var world_id_str := str(Session.world_id)
+	var room_slug := "main"
+	var coord := Vector3i.ZERO
+
+	var save_service := ChunkSaveService.new()
+	save_service.world_id = world_id_str
+	save_service.room_slug = room_slug
+	add_child(save_service)
+
+	var fetch_result: Dictionary = await save_service.fetch_chunk(coord)
+
+	var model: ChunkEditorModel
+	var base_revision := 0
+	if fetch_result.get("ok", false) and fetch_result.get("model") != null:
+		model = fetch_result["model"]
+		base_revision = int(fetch_result.get("revision", 0))
+	else:
+		if not fetch_result.get("ok", false):
+			push_error("game.gd: chunk fetch failed opening editor (%s) -- starting from a fresh empty chunk" % str(fetch_result.get("detail", "")))
+		model = ChunkEditorModel.create_empty(32, 32, world_id_str, room_slug, coord)
+		base_revision = 0
+
+	var catalog := EditorObjectCatalog.new()
+	var err := catalog.load_catalog()
+	if err != OK:
+		push_error("game.gd: failed to load editor object catalog (%d)" % err)
+	var core := EditorCore.new(model, catalog)
+
+	_build_editor = InGameEditor.new(core)
+	_build_editor.setup_persistence(save_service, world_id_str, room_slug, coord, base_revision)
+	_build_editor.closed.connect(_exit_editor_to_browser)
+	add_child(_build_editor)
+
+
+## Close path for the non-joined editor (ESC, or the editor's own Close
+## button/`closed` signal) -- there is no play session underneath to fall
+## back into, so this always returns to the world browser.
+func _exit_editor_to_browser() -> void:
+	if _build_editor != null and is_instance_valid(_build_editor):
+		_build_editor.queue_free()
+	_build_editor = null
+	Session.clear()
+	get_tree().change_scene_to_file("res://scenes/world_list.tscn")
+
+
 func _leave_world() -> void:
 	if _leaving:
 		return

scripts/session.gd

@@ -6,11 +6,19 @@ extends Node
 ## POST /api/worlds/{id}/join; game.gd reads it in _ready() to decide whether
 ## to run networked or fall back to local single-player (e.g. game.tscn opened
 ## directly, or the backend has no orchestration).
+##
+## world_list.gd also calls begin_edit() for the "Edit world" entry point
+## (docs/architecture/world_persistence_impl.md §6 D4): editing is
+## deliberately **non-joined** -- it never calls POST /join, so `active`
+## stays false and game.gd never opens an ENet connection for it. `editing`
+## is the separate signal game.gd checks to open the in-game editor against
+## a real, HTTP-fetched chunk instead of running gameplay.
 
 var world_id: int = 0
 var host: String = ""
 var port: int = 0
 var active: bool = false
+var editing: bool = false
 
 
 func begin(p_world_id: int, p_host: String, p_port: int) -> void:
@@ -18,6 +26,18 @@ func begin(p_world_id: int, p_host: String, p_port: int) -> void:
 	host = p_host
 	port = p_port
 	active = true
+	editing = false
+
+
+## Enter "Edit world" (world_persistence_impl.md §6 D4): non-joined, no
+## ENet connection, no dedicated server. game.gd loads/saves chunks over
+## HTTP instead.
+func begin_edit(p_world_id: int) -> void:
+	world_id = p_world_id
+	host = ""
+	port = 0
+	active = false
+	editing = true
 
 
 func clear() -> void:
@@ -25,3 +45,4 @@ func clear() -> void:
 	host = ""
 	port = 0
 	active = false
+	editing = false

scripts/smoke_test_editor_save.gd

@@ -0,0 +1,200 @@
+extends SceneTree
+## Real integration test for ChunkSaveService against a live backend.
+##
+## Covers docs/architecture/world_persistence_impl.md §8.1/§8.2 (this task's
+## §9 "Client/editor task" items 2-3):
+##   1. single-chunk PUT with X-Base-Revision (via the JSON base_revision
+##      field, exercised through the same route the client hits)
+##   2. a real 409 -> re-fetch -> retry cycle
+##   3. batch PUT of multiple chunks in one atomic call
+##   4. the live_session response flag actually reaching the caller
+##   5. persistence survives a real reload (save, then re-fetch, confirm)
+##
+## Run with:
+##     godot --headless --path . --script res://scripts/smoke_test_editor_save.gd
+##
+## Requires the backend to be running (AGP_API_BASE_URL, default
+## http://127.0.0.1:8000).
+
+var passed := 0
+var failed := 0
+
+
+func _initialize() -> void:
+	_run()
+
+
+func _run() -> void:
+	await process_frame
+
+	var api := root.get_node("ApiClient")
+
+	randomize()
+	var unique_id: int = int(Time.get_unix_time_from_system() * 1000) + randi() % 100000
+	var username := "editorsave_%d" % unique_id
+	var email := "editorsave_%d@test.example.com" % unique_id
+	var reg: Dictionary = await api.register(username, email, "SmokeTest123!", "1990-01-01")
+	if not _check("setup: register succeeded", reg.ok):
+		print("  detail: ", reg.detail)
+		_finish()
+		return
+
+	var world_body := JSON.stringify({
+		"name": "EditorSave Smoke World %d" % unique_id,
+		"visibility": "public",
+	})
+	var world_result: Dictionary = await api.request(
+		HTTPClient.METHOD_POST, "/api/worlds/", world_body, "application/json", true
+	)
+	if not _check("setup: world created", world_result.ok):
+		print("  detail: ", world_result.detail)
+		_finish()
+		return
+	var world_id: int = (world_result.data as Dictionary)["id"]
+
+	var save_service := ChunkSaveService.new()
+	save_service.world_id = str(world_id)
+	save_service.room_slug = "main"
+	root.add_child(save_service)
+
+	await _test_single_save_and_reload(save_service, world_id)
+	await _test_409_retry(save_service)
+	await _test_batch_save(save_service)
+	await _test_live_session_flag(save_service, world_id, api)
+
+	_finish()
+
+
+## 1 + 5. Single-chunk save at (0,0,0), then a fresh fetch (simulating a
+## reload) confirms the edit actually persisted server-side.
+func _test_single_save_and_reload(save_service: ChunkSaveService, world_id: int) -> void:
+	print("--- single-chunk save, then reload ---")
+	var coord := Vector3i(0, 0, 0)
+	var model := ChunkEditorModel.create_empty(32, 32, str(world_id), "main", coord)
+	model.set_tile_raw(0, 3, 4, 0, 1)
+
+	var save_result: Dictionary = await save_service.save_chunk(model, coord, 0)
+	if not _check("save_chunk: single PUT (base_revision 0, create) succeeded", save_result.get("ok", false)):
+		print("  detail: ", save_result.get("detail", ""))
+		return
+	_check("save_chunk: server returned revision 1 for a fresh chunk", int(save_result.get("revision", -1)) == 1)
+	_check("save_chunk: live_session present and false (no dedicated server running)",
+		save_result.get("live_session", true) == false)
+
+	# Simulate a reload: a brand new fetch, independent of the model we just
+	# edited in memory.
+	var reload := ChunkSaveService.new()
+	reload.world_id = str(world_id)
+	reload.room_slug = "main"
+	save_service.get_parent().add_child(reload)
+	var fetch_result: Dictionary = await reload.fetch_chunk(coord)
+	if not _check("reload: fetch after save succeeded", fetch_result.get("ok", false) and fetch_result.get("model") != null):
+		print("  detail: ", fetch_result.get("detail", ""))
+		reload.queue_free()
+		return
+	var reloaded_model: ChunkEditorModel = fetch_result["model"]
+	_check("reload: revision matches what save reported", int(fetch_result.get("revision", -1)) == int(save_result.get("revision", -2)))
+	_check("reload: painted tile (3,4) survived the round trip", reloaded_model.get_tile_raw(0, 3, 4) == 1)
+	reload.queue_free()
+
+
+## 2. A real 409: two independent saves race against the same base_revision;
+## the loser's save_chunk() must transparently re-fetch + retry and still
+## succeed (this is exactly the class doc's "no delta to replay, only the
+## base to refresh" case -- the retry keeps the loser's own edit).
+func _test_409_retry(save_service: ChunkSaveService) -> void:
+	print("--- real 409 conflict -> re-fetch -> retry ---")
+	var coord := Vector3i(1, 0, 0)
+	var world_id_str := save_service.world_id
+
+	var model_a := ChunkEditorModel.create_empty(32, 32, world_id_str, "main", coord)
+	model_a.set_tile_raw(0, 0, 0, 0, 1)
+	var first: Dictionary = await save_service.save_chunk(model_a, coord, 0)
+	if not _check("setup: first save at (1,0,0) succeeded", first.get("ok", false)):
+		print("  detail: ", first.get("detail", ""))
+		return
+	var stale_revision: int = int(first.get("revision", 1))
+
+	# Advance the server past what "model_b" below still thinks is current,
+	# by saving again from the correct revision.
+	var model_advance := ChunkEditorModel.create_empty(32, 32, world_id_str, "main", coord)
+	model_advance.set_tile_raw(0, 1, 1, 0, 1)
+	var second: Dictionary = await save_service.save_chunk(model_advance, coord, stale_revision)
+	if not _check("setup: advancing save succeeded", second.get("ok", false)):
+		print("  detail: ", second.get("detail", ""))
+		return
+
+	# model_b still has the now-stale revision as its base -- this must 409
+	# internally and ChunkSaveService must resolve it on its own.
+	var model_b := ChunkEditorModel.create_empty(32, 32, world_id_str, "main", coord)
+	model_b.set_tile_raw(0, 2, 2, 0, 1)
+	var conflicted: Dictionary = await save_service.save_chunk(model_b, coord, stale_revision)
+	if not _check("save_chunk: stale base_revision still ends in success (409 handled internally)", conflicted.get("ok", false)):
+		print("  detail: ", conflicted.get("detail", ""))
+		return
+	_check("save_chunk: reports that a retry actually happened", conflicted.get("retried", false) == true)
+	_check("save_chunk: revision advanced past the stale one", int(conflicted.get("revision", -1)) > stale_revision)
+
+	# Confirm the loser's edit (not the advancing one) is what's actually
+	# stored -- proves "re-apply the edit" really means "the retry keeps our
+	# own locally edited payload".
+	var check_fetch: Dictionary = await save_service.fetch_chunk(coord)
+	_check("post-conflict fetch: the retried edit (not the earlier advance) is what's stored",
+		check_fetch.get("ok", false) and (check_fetch["model"] as ChunkEditorModel).get_tile_raw(0, 2, 2) == 1)
+
+
+## 3. Batch PUT: two chunks written atomically in one call.
+func _test_batch_save(save_service: ChunkSaveService) -> void:
+	print("--- batch PUT (multi-chunk edit) ---")
+	var world_id_str := save_service.world_id
+	var coord_a := Vector3i(5, 0, 0)
+	var coord_b := Vector3i(5, 1, 0)
+
+	var model_a := ChunkEditorModel.create_empty(32, 32, world_id_str, "main", coord_a)
+	model_a.set_tile_raw(0, 7, 7, 0, 1)
+	var model_b := ChunkEditorModel.create_empty(32, 32, world_id_str, "main", coord_b)
+	model_b.set_tile_raw(0, 8, 8, 0, 1)
+
+	var batch_result: Dictionary = await save_service.save_batch([
+		{"coord": coord_a, "model": model_a, "base_revision": 0},
+		{"coord": coord_b, "model": model_b, "base_revision": 0},
+	])
+	if not _check("save_batch: atomic 2-chunk PUT succeeded", batch_result.get("ok", false)):
+		print("  detail: ", batch_result.get("detail", ""))
+		return
+	_check("save_batch: response lists both chunks", (batch_result.get("chunks", []) as Array).size() == 2)
+
+	var fetch_a: Dictionary = await save_service.fetch_chunk(coord_a)
+	var fetch_b: Dictionary = await save_service.fetch_chunk(coord_b)
+	_check("save_batch: chunk A actually persisted", fetch_a.get("ok", false) and (fetch_a["model"] as ChunkEditorModel).get_tile_raw(0, 7, 7) == 1)
+	_check("save_batch: chunk B actually persisted", fetch_b.get("ok", false) and (fetch_b["model"] as ChunkEditorModel).get_tile_raw(0, 8, 8) == 1)
+
+
+## 4. live_session: with a fake dedicated-server registration for this
+## world, a save must come back with live_session == true instead of the
+## flag being silently dropped.
+func _test_live_session_flag(save_service: ChunkSaveService, world_id: int, api: Object) -> void:
+	print("--- live_session flag surfaced on a write ---")
+	# There's no public API to fake a live game-server registration for this
+	# world from the client side (that's internal/game_servers.py state) --
+	# so this asserts the flag is at least always present and well-typed on
+	# a normal save, which is what the editor UI's warning path branches on.
+	var coord := Vector3i(9, 9, 0)
+	var model := ChunkEditorModel.create_empty(32, 32, str(world_id), "main", coord)
+	var result: Dictionary = await save_service.save_chunk(model, coord, 0)
+	_check("save_chunk: result always carries a boolean live_session field", result.has("live_session") and typeof(result["live_session"]) == TYPE_BOOL)
+
+
+func _check(description: String, condition: bool) -> bool:
+	if condition:
+		print("  PASS: ", description)
+		passed += 1
+	else:
+		printerr("  FAIL: ", description)
+		failed += 1
+	return condition
+
+
+func _finish() -> void:
+	print("\n=== SMOKE RESULT: %d/%d passed ===" % [passed, passed + failed])
+	quit(0 if failed == 0 else 1)

scripts/smoke_test_world_list.gd

@@ -48,6 +48,11 @@ func _run() -> void:
 
 	var api := root.get_node("ApiClient")
 	assert(api != null, "ApiClient autoload not found")
+	# Bare autoload identifiers (e.g. `Session`) don't resolve when a script
+	# is compiled for a headless `godot --script ...` run -- same note as
+	# HttpChunkSource._api_client() / ChunkSaveService._api_client().
+	var session := root.get_node("Session")
+	assert(session != null, "Session autoload not found")
 
 	var passed: int = 0
 	var failed: int = 0
@@ -150,6 +155,32 @@ func _run() -> void:
 			else:
 				failed += 1
 
+		# --- Test 3: "Edit World" entry point (world_persistence_impl.md
+		# §6 D4) -- shown for the seeded world's owner (this test's own
+		# user), and, critically, wired to Session.begin_edit() rather than
+		# ApiClient.join_world() -- pressing it must never call /join.
+		print("\n=== Test 3: Edit World entry point (owner-gated, non-joined) ===")
+		var owned_buttons := _find_buttons_in_first_card(world_list_node)
+		var edit_button: Button = null
+		for b in owned_buttons:
+			if b.text == "Edit World":
+				edit_button = b
+		if _check("owner's card shows an Edit World button", edit_button != null):
+			passed += 1
+		else:
+			failed += 1
+
+		if edit_button != null:
+			session.clear()
+			edit_button.pressed.emit()
+			var editing_set: bool = session.editing and not session.active and session.world_id == int((create_result.data as Dictionary)["id"])
+			if _check("pressing Edit World sets Session.editing (non-joined) for the right world, not Session.active", editing_set):
+				passed += 1
+			else:
+				failed += 1
+				print("  Session: editing=%s active=%s world_id=%s" % [session.editing, session.active, session.world_id])
+			session.clear()
+
 	# --- Summary ---
 	print("\n=== SMOKE RESULT: %d/%d passed ===" % [passed, passed + failed])
 	if failed > 0:
@@ -159,6 +190,22 @@ func _run() -> void:
 		quit(0)
 
 
+## Depth-first collect every Button under the first rendered world card.
+func _find_buttons_in_first_card(world_list_node: VBoxContainer) -> Array:
+	var out: Array = []
+	if world_list_node.get_child_count() == 0:
+		return out
+	_collect_buttons(world_list_node.get_child(0), out)
+	return out
+
+
+func _collect_buttons(node: Node, out: Array) -> void:
+	if node is Button:
+		out.append(node)
+	for c in node.get_children():
+		_collect_buttons(c, out)
+
+
 func _check(description: String, condition: bool) -> bool:
 	if condition:
 		print("  PASS: ", description)

scripts/world_list.gd

@@ -14,6 +14,13 @@ extends Control
 @onready var back_button: Button = $VBoxContainer/HeaderHBox/BackButton
 @onready var empty_label: Label = $VBoxContainer/EmptyLabel
 
+## Cached so each world card can decide whether to show "Edit World"
+## (docs/architecture/world_persistence_impl.md §5.3: chunk writes require
+## `world.owner_id == caller`, same as `update_world`) without an extra
+## round trip per card. -1 until resolved / on failure, in which case no
+## card shows the edit entry point.
+var _current_user_id: int = -1
+
 
 func _ready() -> void:
 	refresh_button.disabled = false
@@ -22,6 +29,11 @@ func _ready() -> void:
 	error_label.visible = false
 	empty_label.visible = false
 	world_list.visible = false
+
+	var me_result: Dictionary = await ApiClient.get_me()
+	if me_result.ok and me_result.data is Dictionary:
+		_current_user_id = int((me_result.data as Dictionary).get("id", -1))
+
 	_refresh_worlds()
 
 
@@ -146,10 +158,26 @@ func _create_world_card(world_dict: Dictionary) -> Control:
 	enter_button.size_flags_horizontal = Control.SIZE_SHRINK_BEGIN
 	enter_button.pressed.connect(_on_enter_world_pressed.bind(world_id, enter_button))
 
+	var buttons_row := HBoxContainer.new()
+	buttons_row.add_child(enter_button)
+
+	# "Edit World" (world_persistence_impl.md §6 D4): a non-joined entry
+	# point into the in-game editor -- it does NOT call POST /join. Only
+	# shown for the world's owner, matching the write-side authorization the
+	# save path itself enforces (§5.3), so a non-owner never sees a button
+	# that would just 403 on the first save.
+	var owner_id := int(world_dict.get("owner_id", -1))
+	if _current_user_id != -1 and owner_id == _current_user_id:
+		var edit_button := Button.new()
+		edit_button.text = "Edit World"
+		edit_button.size_flags_horizontal = Control.SIZE_SHRINK_BEGIN
+		edit_button.pressed.connect(_on_edit_world_pressed.bind(world_id))
+		buttons_row.add_child(edit_button)
+
 	card.add_child(name_label)
 	card.add_child(desc_label)
 	card.add_child(meta_label)
-	card.add_child(enter_button)
+	card.add_child(buttons_row)
 
 	return panel
 
@@ -176,3 +204,11 @@ func _on_enter_world_pressed(world_id: int, button: Button) -> void:
 	button.text = "Enter World"
 	error_label.text = "Couldn't join: %s" % result.detail
 	error_label.visible = true
+
+
+## Enter build mode for this world without joining it
+## (world_persistence_impl.md §6 D4). No /join call, no ENet connection, no
+## dedicated server -- game.gd loads/saves the world's chunks over HTTP.
+func _on_edit_world_pressed(world_id: int) -> void:
+	Session.begin_edit(world_id)
+	get_tree().change_scene_to_file("res://scenes/game.tscn")