@miller-tech/uap 1.172.13 → 1.172.14

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.172.13",
+  "version": "1.172.14",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",

package/tools/agents/scripts/anthropic_proxy.py

@@ -2970,9 +2970,30 @@ _current_request_session: contextvars.ContextVar[str | None] = contextvars.Conte
 #
 # Same ContextVar trick as the session above: request-local without threading a
 # Request through every signature. Holds Starlette's `request.is_disconnected`.
-_current_client_gone: contextvars.ContextVar[
-    Callable[[], Awaitable[bool]] | None
-] = contextvars.ContextVar("uap_current_client_gone", default=None)
+# Is the DOWNSTREAM client still there?
+#
+# Nothing used to ask. Once a turn was accepted, the guardrail loops kept
+# retrying and re-POSTing upstream on their own — so when the client died the
+# work carried on regardless. Observed live: two `uap deliver` runs were killed,
+# no socket remained on :4000, and the proxy still drove llama for minutes,
+# generating 32k tokens per orphaned turn that nobody would ever read.
+#
+# WHY NOT request.is_disconnected(). That was tried and MEASURED not to work
+# here: it returned False 16 times across 30s while the caller was demonstrably
+# gone. The cause is `@app.middleware("http")` — Starlette's BaseHTTPMiddleware
+# runs the endpoint in a separate anyio task behind its own receive channel, so
+# the endpoint polls a channel that never delivers http.disconnect. It works
+# fine in an app with no middleware, which is exactly why three rounds of
+# isolated testing said it was sound.
+#
+# So the signal is captured by a pure-ASGI middleware OUTSIDE all of that (see
+# DisconnectWatcherMiddleware) and published here as a mutable holder. A dict
+# rather than a bool because the middleware's poller task and the endpoint must
+# see the same object: the ContextVar is set before the app is invoked, so every
+# downstream task inherits the reference, and mutations are visible across them.
+_disconnect_holder: contextvars.ContextVar[dict | None] = contextvars.ContextVar(
+    "uap_disconnect_holder", default=None
+)
 
 
 class ClientGoneError(Exception):
@@ -2986,19 +3007,17 @@ class ClientGoneError(Exception):
 
 
 async def _client_gone() -> bool:
-    """True when the caller has hung up. Never raises, never blocks meaningfully.
+    """True when the caller has hung up.
 
-    A disconnect check that can itself fail or stall would be worse than no
-    check: it sits in front of every upstream call. Any error means "assume the
-    client is still there" — the pre-existing behaviour.
+    Now a flag read rather than a call into Starlette, so there is nothing to
+    raise and nothing for a fail-safe except to swallow. The previous version
+    wrapped request.is_disconnected() in `except Exception: return False`, which
+    would have hidden a failing probe as "client still present" — it turned out
+    the probe was not raising, but the shape was wrong regardless.
     """
-    probe = _current_client_gone.get()
-    if probe is None:
-        return False
-    try:
-        return bool(await probe())
-    except Exception:  # noqa: BLE001
-        return False
+    holder = _disconnect_holder.get()
+    return bool(holder and holder.get("gone"))
+
 
 # Session admission state. _admitted_sessions maps session_id -> last-seen
 # monotonic ts; OrderedDict insertion order is the LRU (oldest = front).
@@ -3385,8 +3404,8 @@ class DisconnectAwareClient(httpx.AsyncClient):
     """
 
     async def send(self, request, **kwargs):  # type: ignore[override]
-        probe = _current_client_gone.get()
-        if probe is None:
+        holder = _disconnect_holder.get()
+        if holder is None:
             return await super().send(request, **kwargs)
 
         task = asyncio.ensure_future(super().send(request, **kwargs))
@@ -3635,6 +3654,73 @@ async def lifespan(app: FastAPI):
     logger.info("Proxy shut down")
 
 
+class DisconnectWatcherMiddleware:
+    """Pure-ASGI middleware that records when the caller hangs up.
+
+    Sits OUTERMOST, above the auth middleware, and is deliberately pure ASGI:
+    Starlette's BaseHTTPMiddleware (`@app.middleware("http")`) re-tasks the
+    endpoint behind its own receive channel, which is precisely why
+    request.is_disconnected() reported False 16 times across 30s while the
+    caller was gone. Being outside that, this sees the real channel.
+
+    THE HARD PART is that wrapping receive() is not enough on its own. After the
+    request body is fully read nothing calls receive() again, so an
+    http.disconnect just sits in the channel unobserved. So once the body is
+    complete this starts polling the channel itself — safe precisely because the
+    app has no more body to ask for, and the only remaining message type is the
+    disconnect.
+
+    The flag lives in a mutable dict published on a ContextVar before the app is
+    invoked: downstream tasks inherit the reference when their context is copied,
+    so a mutation made by the poller is visible to the endpoint and to every
+    upstream call it makes.
+    """
+
+    def __init__(self, app):
+        self.app = app
+
+    async def __call__(self, scope, receive, send):
+        if scope.get("type") != "http":
+            return await self.app(scope, receive, send)
+
+        holder: dict = {"gone": False}
+        _disconnect_holder.set(holder)
+        poller: asyncio.Task | None = None
+
+        async def poll_for_disconnect() -> None:
+            # Only reached once the body is complete, so this cannot steal a
+            # message the app still needs.
+            try:
+                while not holder["gone"]:
+                    msg = await receive()
+                    if msg.get("type") == "http.disconnect":
+                        holder["gone"] = True
+                        return
+            except asyncio.CancelledError:
+                raise
+            except Exception:  # noqa: BLE001
+                return  # channel closed/erroring: stop watching, never crash
+
+        async def watched_receive():
+            nonlocal poller
+            msg = await receive()
+            if msg.get("type") == "http.disconnect":
+                holder["gone"] = True
+            elif msg.get("type") == "http.request" and not msg.get("more_body", False):
+                # Body complete — nobody will call receive() again, so take over.
+                if poller is None:
+                    poller = asyncio.ensure_future(poll_for_disconnect())
+            return msg
+
+        try:
+            await self.app(scope, watched_receive, send)
+        finally:
+            if poller is not None and not poller.done():
+                poller.cancel()
+
+
+
+
 app = FastAPI(
     title="UAP Anthropic Proxy",
     description="Translates Anthropic Messages API to OpenAI Chat Completions API",
@@ -3642,6 +3728,7 @@ app = FastAPI(
     lifespan=lifespan,
 )
 
+
 @app.exception_handler(ClientGoneError)
 async def _client_gone_handler(request: Request, exc: ClientGoneError):
     """The caller hung up mid-turn: stop, and say so at INFO rather than ERROR.
@@ -3696,6 +3783,18 @@ async def _pool_timeout_handler(request: Request, exc: httpx.PoolTimeout):
 _PROXY_AUTH_OPEN_PATHS = frozenset({"/health", "/", "/v1/models"})
 
 
+# OUTERMOST — and it must be registered AFTER the auth middleware below to BE
+# outermost: Starlette's add_middleware inserts at position 0, so the last one
+# registered is the outermost wrapper. (Registered before auth first time round,
+# which would have put it INSIDE the very BaseHTTPMiddleware that hides
+# http.disconnect — the exact bug it exists to work around.)
+#
+# The registration therefore trails the auth definition on purpose. Auth itself
+# is deliberately untouched.
+def _install_disconnect_watcher() -> None:
+    app.add_middleware(DisconnectWatcherMiddleware)
+
+
 @app.middleware("http")
 async def _shared_secret_auth(request: Request, call_next):
     """Gate every request behind PROXY_AUTH_TOKEN when it is set.
@@ -3754,6 +3853,10 @@ async def _shared_secret_auth(request: Request, call_next):
 # ===========================================================================
 
 
+
+# Auth is now defined, so this lands OUTSIDE it (add_middleware inserts at 0).
+_install_disconnect_watcher()
+
 def _image_block_to_openai(block: dict) -> dict | None:
     """Anthropic image block → OpenAI image_url part (data URI or URL)."""
     src = block.get("source") or {}
@@ -10908,10 +11011,6 @@ async def messages(request: Request):
     """
     global last_session_id
 
-    # Publish this request's disconnect probe for the whole call tree. Set before
-    # ANY upstream work so an already-dead client is caught on the first call.
-    _current_client_gone.set(request.is_disconnected)
-
     body = await request.json()
     is_stream = body.get("stream", False)
     model = body.get("model", "default")

package/tools/agents/tests/test_client_disconnect.py

@@ -28,33 +28,32 @@ proxy = _load_proxy()
 
 
 class ClientGoneProbeTest(unittest.TestCase):
-    def test_absent_probe_means_client_is_present(self):
-        """No probe set (background task, test harness) must not read as gone —
-        that would abandon perfectly live turns."""
-        proxy._current_client_gone.set(None)
+    """The probe is now a FLAG READ, not a call into Starlette.
+
+    It used to call request.is_disconnected() wrapped in `except Exception:
+    return False`. That was measured not to work at all behind
+    BaseHTTPMiddleware (False 16x across 30s while the caller was gone), and the
+    swallow-everything shape would have hidden a failing probe as "client still
+    present". Both are gone: there is nothing left to raise.
+    """
+
+    def test_no_holder_means_client_is_present(self):
+        """Background tasks and health checks have no request context. They must
+        not read as disconnected."""
+        proxy._disconnect_holder.set(None)
         self.assertFalse(asyncio.run(proxy._client_gone()))
 
     def test_reports_disconnect(self):
-        async def gone() -> bool:
-            return True
-
-        proxy._current_client_gone.set(gone)
+        proxy._disconnect_holder.set({"gone": True})
         self.assertTrue(asyncio.run(proxy._client_gone()))
 
     def test_reports_connected(self):
-        async def here() -> bool:
-            return False
-
-        proxy._current_client_gone.set(here)
+        proxy._disconnect_holder.set({"gone": False})
         self.assertFalse(asyncio.run(proxy._client_gone()))
 
-    def test_a_failing_probe_never_breaks_the_request(self):
-        """This check sits in front of every upstream call. A probe that can
-        raise would be worse than no probe, so failure means 'still there'."""
-        async def broken() -> bool:
-            raise RuntimeError("receive channel exploded")
-
-        proxy._current_client_gone.set(broken)
+    def test_a_malformed_holder_reads_as_present(self):
+        """Defensive: an empty dict must not be mistaken for a disconnect."""
+        proxy._disconnect_holder.set({})
         self.assertFalse(asyncio.run(proxy._client_gone()))
 
 

package/tools/agents/tests/test_disconnect_watcher.py

@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+"""Pure-ASGI disconnect watcher — the thing that finally made the abort work.
+
+Four attempts failed before this. The measured cause: `@app.middleware("http")`
+is Starlette's BaseHTTPMiddleware, which runs the endpoint in a separate anyio
+task behind its own receive channel, so request.is_disconnected() in the endpoint
+polls a channel that never delivers http.disconnect. Instrumented in production
+it returned False 16 times across 30s while the caller was demonstrably gone —
+and it worked perfectly in an app with no middleware, which is exactly why three
+rounds of isolated testing said the approach was sound.
+
+Two properties are load-bearing and each has a test that fails if it breaks:
+
+  1. ORDER. The watcher must be OUTERMOST. Starlette's add_middleware inserts at
+     position 0, so it must be registered LAST. The first implementation
+     registered it before auth, putting it inside the very middleware whose
+     behaviour it exists to work around.
+  2. POLLING. Wrapping receive() is not enough. Once the body is fully read
+     nothing calls receive() again, so an http.disconnect sits in the channel
+     unobserved. The watcher must take over polling after the final body message.
+
+Verified end to end after these were in place: the llama slot freed within 5s of
+the client hanging up, against 30s+ of continued generation before.
+"""
+
+import ast
+import asyncio
+import importlib.util
+import unittest
+from pathlib import Path
+
+PROXY_PATH = Path(__file__).resolve().parents[1] / "scripts" / "anthropic_proxy.py"
+
+
+def _load_proxy():
+    spec = importlib.util.spec_from_file_location("anthropic_proxy_watcher", PROXY_PATH)
+    assert spec is not None and spec.loader is not None
+    m = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(m)
+    return m
+
+
+proxy = _load_proxy()
+
+
+class MiddlewareOrderTest(unittest.TestCase):
+    def test_the_watcher_is_outermost(self):
+        """If it ends up inside BaseHTTPMiddleware it cannot see http.disconnect,
+        which is the whole failure this replaced."""
+        names = [getattr(mw.cls, "__name__", str(mw.cls)) for mw in proxy.app.user_middleware]
+        self.assertTrue(names, "no middleware registered at all")
+        self.assertEqual(
+            names[0],
+            "DisconnectWatcherMiddleware",
+            f"watcher is not outermost; stack is {names}",
+        )
+
+    def test_auth_middleware_is_still_installed(self):
+        """The brief was to leave auth alone. Losing it would silently open the
+        LAN token gate — a far worse outcome than the bug being fixed."""
+        names = [getattr(mw.cls, "__name__", str(mw.cls)) for mw in proxy.app.user_middleware]
+        self.assertIn("BaseHTTPMiddleware", names)
+
+    def test_probe_is_a_flag_read_not_a_starlette_call(self):
+        """_client_gone must not call back into the request object: that is the
+        path that measured False-forever behind BaseHTTPMiddleware.
+
+        Checked with ast rather than substring matching — the first version of
+        this test failed on the word appearing in the function's own docstring,
+        which is a false positive, not a finding.
+        """
+        tree = ast.parse(PROXY_PATH.read_text())
+        fn = next(
+            n for n in ast.walk(tree)
+            if isinstance(n, ast.AsyncFunctionDef) and n.name == "_client_gone"
+        )
+        attrs = {
+            n.attr for n in ast.walk(fn) if isinstance(n, ast.Attribute)
+        }
+        self.assertNotIn("is_disconnected", attrs)
+        # And it must read the holder, so the test fails if the body is gutted.
+        self.assertIn("get", attrs)
+
+
+class WatcherBehaviourTest(unittest.TestCase):
+    """Drive the middleware directly with an ASGI receive/send pair."""
+
+    def _run(self, messages, app_body):
+        """messages: what receive() yields, in order. app_body: coroutine fn(scope, receive, send)."""
+        sent = []
+        queue = list(messages)
+
+        async def receive():
+            if queue:
+                return queue.pop(0)
+            await asyncio.sleep(3600)  # channel idle, like a live connection
+
+        async def send(msg):
+            sent.append(msg)
+
+        mw = proxy.DisconnectWatcherMiddleware(app_body)
+        asyncio.run(mw({"type": "http"}, receive, send))
+        return sent
+
+    def test_records_a_disconnect_the_app_itself_reads(self):
+        seen = {}
+
+        async def app(scope, receive, send):
+            await receive()  # body
+            msg = await receive()  # disconnect
+            seen["type"] = msg["type"]
+            seen["gone"] = await proxy._client_gone()
+
+        self._run(
+            [{"type": "http.request", "body": b"{}", "more_body": False},
+             {"type": "http.disconnect"}],
+            app,
+        )
+        self.assertEqual(seen["type"], "http.disconnect")
+        self.assertTrue(seen["gone"], "watcher did not record the disconnect")
+
+    def test_records_a_disconnect_the_app_NEVER_reads(self):
+        """The real case. After the body is consumed the app never calls receive()
+        again, so the disconnect is only ever seen by the watcher's own poller.
+        Wrapping receive() alone would leave `gone` False here forever."""
+        seen = {}
+
+        async def app(scope, receive, send):
+            await receive()  # body only — then just "work", like a generation
+            for _ in range(40):
+                await asyncio.sleep(0.02)
+                if await proxy._client_gone():
+                    seen["gone"] = True
+                    return
+            seen["gone"] = False
+
+        self._run(
+            [{"type": "http.request", "body": b"{}", "more_body": False},
+             {"type": "http.disconnect"}],
+            app,
+        )
+        self.assertTrue(seen.get("gone"), "poller did not observe the unread disconnect")
+
+    def test_a_live_request_is_never_marked_gone(self):
+        """A false positive would abort healthy turns — worse than the bug."""
+        seen = {}
+
+        async def app(scope, receive, send):
+            await receive()
+            for _ in range(10):
+                await asyncio.sleep(0.02)
+            seen["gone"] = await proxy._client_gone()
+
+        self._run([{"type": "http.request", "body": b"{}", "more_body": False}], app)
+        self.assertFalse(seen["gone"])
+
+    def test_non_http_scopes_pass_straight_through(self):
+        """Lifespan and websocket scopes have no disconnect semantics here and
+        must not be wrapped."""
+        called = {}
+
+        async def app(scope, receive, send):
+            called["type"] = scope["type"]
+
+        async def go():
+            mw = proxy.DisconnectWatcherMiddleware(app)
+            await mw({"type": "lifespan"}, None, None)
+
+        asyncio.run(go())
+        self.assertEqual(called["type"], "lifespan")
+
+
+if __name__ == "__main__":
+    unittest.main()

package/tools/agents/tests/test_upstream_chokepoint.py

@@ -115,9 +115,9 @@ class ChokePointBehaviourTest(unittest.TestCase):
         return client, state, (lambda: setattr(httpx.AsyncClient, "send", original))
 
     def _set_probe(self, gone: bool):
-        async def probe() -> bool:
-            return gone
-        proxy._current_client_gone.set(probe)
+        # The guard now reads a shared mutable holder published by the ASGI
+        # watcher, rather than calling back into the request object.
+        proxy._disconnect_holder.set({"gone": gone})
 
     def test_cancels_upstream_when_the_caller_leaves(self):
         client, state, restore = self._client_with_send("never")
@@ -147,7 +147,7 @@ class ChokePointBehaviourTest(unittest.TestCase):
         keep working rather than acquiring surprise cancellation."""
         client, state, restore = self._client_with_send("fast")
         try:
-            proxy._current_client_gone.set(None)
+            proxy._disconnect_holder.set(None)
             self.assertEqual(asyncio.run(client.send(self._Req())), "RESPONSE")
         finally:
             restore()