diff --git a/.chainlit/config.toml b/.chainlit/config.toml
index 4e2f3e7..9ff01b2 100644
--- a/.chainlit/config.toml
+++ b/.chainlit/config.toml
@@ -29,7 +29,10 @@ latex = false
auto_tag_thread = true
# Allow users to edit their own messages
-edit_message = true
+# Off: an edit removed later turns from the screen but not from the model's
+# history, so text a reader edited out kept being sent with every turn
+# (review, area 1b).
+edit_message = false
# Authorize users to spontaneously upload files with messages
[features.spontaneous_file_upload]
diff --git a/bin/chat-chainlit.py b/bin/chat-chainlit.py
index 37ea68b..b55cf6b 100644
--- a/bin/chat-chainlit.py
+++ b/bin/chat-chainlit.py
@@ -3,6 +3,7 @@
# or get_data_layer. Not fixable here; it needs stubs upstream. The file is
# named with a hyphen, so it cannot be listed in [[tool.mypy.overrides]].
import asyncio
+import contextlib
import os
import time
from collections.abc import Awaitable, Callable
@@ -476,7 +477,21 @@ async def on_window_message(message: object) -> None:
# the summary's data loads. Otherwise a question asked meanwhile ran on
# the same thread, and the seed landed beside it and was lost -- while
# the chat still said it was continuing from the summary.
- run_as_task(lambda: continue_from_handoff(handoff_id))
+ # Messages wait for this, below. Running the claim as a task alone did
+ # not hold them: Chainlit ends its own startup task with task_end,
+ # which unlocks the box mid-seed whenever the claim arrives first --
+ # 6 of 6 page loads at 200ms latency (review, area 1b).
+ seeding = asyncio.Event()
+ _seeding[session_id()] = seeding
+
+ async def claim() -> None:
+ try:
+ await continue_from_handoff(handoff_id)
+ finally:
+ seeding.set()
+ _seeding.pop(session_id(), None)
+
+ run_as_task(claim)
await cl.send_window_message(acknowledgement(handoff_id))
@@ -526,6 +541,20 @@ async def answer_with_model(content: str, message_id: str) -> None:
save_openai_metrics(message_id, openai_cb)
+#: Handoffs being seeded, per session. A message waits for its session's seed
+#: before touching the thread, or the seed lands beside a running turn and is
+#: lost while the chat says it is continuing from the summary.
+_seeding: dict[str, asyncio.Event] = {}
+SEED_WAIT_SECONDS = 45.0
+
+
+async def wait_for_seed() -> None:
+ event = _seeding.get(session_id())
+ if event is not None:
+ with contextlib.suppress(TimeoutError):
+ await asyncio.wait_for(event.wait(), SEED_WAIT_SECONDS)
+
+
#: Guests' messages, limited per human check rather than per session. The
#: per-session quota is keyed on the session id, which the client chooses, so
#: a reconnect -- or a script -- started a fresh quota every time (review, 1b).
@@ -554,6 +583,8 @@ def solve_key() -> str:
@cl.on_message
async def main(message: cl.Message) -> None:
+ await wait_for_seed()
+
# First, before any early return: a "yes" means the offer just made, so
# any other message -- rate limited, an attachment -- ends that meaning.
latest = proposals.take_latest(session_id())
diff --git a/bin/chat-fastapi.py b/bin/chat-fastapi.py
index 627aeed..78df461 100644
--- a/bin/chat-fastapi.py
+++ b/bin/chat-fastapi.py
@@ -6,9 +6,21 @@
from typing import Any
from urllib.parse import urlsplit
+from dotenv import load_dotenv
+
+from util.secrets import SECRET_NAMES, get_secret, load_secrets_to_environ
+
+# Before anything imports chainlit. `chainlit.utils` loads
+# `chainlit.oauth_providers`, which reads OAUTH_*_CLIENT_SECRET once, at import:
+# loaded after it, an OAuth secret supplied as a Docker secret was never seen,
+# and logged-in chat failed at the OAuth callback (review, area 1b).
+load_dotenv()
+# The same list chat-chainlit uses. This was a second, hand-maintained copy
+# that had already drifted from it in both directions.
+load_secrets_to_environ(SECRET_NAMES)
+
import httpx
from chainlit.utils import mount_chainlit
-from dotenv import load_dotenv
from fastapi import FastAPI, Request, Response
from fastapi.responses import HTMLResponse, RedirectResponse
@@ -21,12 +33,6 @@
from util.captcha_scope import is_captcha_exempt
from util.embedding_environment import EmbeddingEnvironment
from util.logging import logging
-from util.secrets import SECRET_NAMES, get_secret, load_secrets_to_environ
-
-load_dotenv()
-# The same list chat-chainlit uses. This was a second, hand-maintained copy
-# that had already drifted from it in both directions.
-load_secrets_to_environ(SECRET_NAMES)
@asynccontextmanager
@@ -62,7 +68,9 @@ async def lifespan(_app: FastAPI) -> AsyncIterator[None]:
app = FastAPI(lifespan=lifespan)
-CHAINLIT_URI = os.getenv("CHAINLIT_URI")
+# Empty means unset. Compose turns an unset variable into "", and an empty
+# CHAINLIT_URI left the captcha page redirecting to itself (review, 1b).
+CHAINLIT_URI = os.getenv("CHAINLIT_URI") or None
# Defined after CHAINLIT_URI, which it is built from. The endpoint lives under
# the Chainlit mount point so one nginx location covers both.
@@ -99,6 +107,13 @@ async def lifespan(_app: FastAPI) -> AsyncIterator[None]:
"this deployment does not want one."
)
CLOUDFLARE_SITE_KEY = os.getenv("CLOUDFLARE_SITE_KEY")
+if CHAT_REQUIRES_HUMAN and not CLOUDFLARE_SITE_KEY:
+ # Without it the check page renders data-sitekey="None", which no one can
+ # pass, while startup reported healthy (review, area 1b).
+ raise RuntimeError(
+ "CHAT_REQUIRES_HUMAN is on and no CLOUDFLARE_SITE_KEY is configured, "
+ "so the human check could not be shown."
+ )
ERROR_PAGE_TEMPLATE = Template(
f"""
@@ -225,7 +240,7 @@ async def captcha_page() -> Response:
">',
+ citations=(),
+ created_at=time.time(),
+ )
+ )
+ assert re.search(r"(? None:
+ # No leeway refused 20% of fresh tokens at 200ms of skew (review, 1b).
+ private, public = keys
+ assert verify(_mint(private, iat=int(time.time()) + 10), public)
+
+
+def test_a_token_from_well_in_the_future_is_refused(keys: tuple[str, str]) -> None:
+ private, public = keys
+ with pytest.raises(TokenRejectedError, match="future"):
+ verify(_mint(private, iat=int(time.time()) + 600), public)
+
+
+@pytest.mark.parametrize("which", ["private", "garbage"])
+def test_a_key_that_cannot_verify_stops_startup(
+ keys: tuple[str, str], tmp_path: Path, which: str
+) -> None:
+ # It passed the emptiness check, then every token was refused as
+ # "unusable" in a log line that blamed the website (review, 1b).
+ private, _ = keys
+ key_file = tmp_path / "key.pem"
+ key_file.write_text(
+ private if which == "private" else "-----BEGIN PUBLIC KEY-----\nAAAA\n"
+ )
+ with pytest.raises(RuntimeError, match="not a PEM public key"):
+ load_verifying_key(str(key_file))
+
+
+def test_the_real_public_key_loads(keys: tuple[str, str], tmp_path: Path) -> None:
+ _, public = keys
+ key_file = tmp_path / "key.pem"
+ key_file.write_text(public)
+ assert load_verifying_key(str(key_file)).startswith("-----BEGIN PUBLIC KEY-----")
diff --git a/tests/util/test_logging_tokens.py b/tests/util/test_logging_tokens.py
index 208c8ac..6140b6c 100644
--- a/tests/util/test_logging_tokens.py
+++ b/tests/util/test_logging_tokens.py
@@ -20,3 +20,26 @@ def test_our_own_info_logs_still_are() -> None:
# The fix must lower only the request logger, not logging generally.
root = logging.getLogger()
assert root.isEnabledFor(logging.getLevelName(util.logging.DEFAULT_LOG_LEVEL))
+
+
+def test_session_ids_are_redacted_from_the_access_log() -> None:
+ # The access log line uvicorn writes for a Chainlit file download.
+ record = logging.LogRecord(
+ "uvicorn.access",
+ logging.INFO,
+ __file__,
+ 0,
+ '%s - "%s %s HTTP/%s" %d',
+ (
+ "1.2.3.4:5",
+ "GET",
+ "/chat/guest/project/file/abc?session_id=SECRET-SID",
+ "1.1",
+ 200,
+ ),
+ None,
+ )
+ assert logging.getLogger("uvicorn.access").filter(record)
+ line = record.getMessage()
+ assert "SECRET-SID" not in line
+ assert "session_id=[redacted]" in line