fix(models): preserve FIFO across request-admission handoff (#5459)

* fix(models): make request admission enqueue atomic

* test(models): cover atomic request admission FIFO entry

* docs(models): record atomic admission FIFO invariant

* fix(models): address request-admission review follow-ups

backend/packages/harness/deerflow/models/AGENTS.md

@@ -3,7 +3,10 @@
 Request-admission waits follow the next scheduled admission and configured
 interval, capped at 50 ms; the cap must not become a minimum poll interval that
 limits high-RPM throughput. Local `AdmissionError` is structurally non-retriable
-in LLM error handling regardless of its message text.
+in LLM error handling regardless of its message text. Immediate admission and
+joining the blocking FIFO are one lock-protected decision: do not split the
+fast-path permit check from queue insertion, or an older caller can be overtaken
+while handing off to the wait queue.
 
 - `create_chat_model(name, thinking_enabled)` instantiates LLM from config via reflection
 - Supports `thinking_enabled` flag with per-model `when_thinking_enabled` overrides
@@ -12,7 +15,7 @@ in LLM error handling regardless of its message text.
 - Supports `supports_vision` flag for image understanding models
 - Config values starting with `$` resolved as environment variables
 - Missing provider modules surface actionable install hints from reflection resolvers (for example `uv add langchain-google-genai`)
-- Optional `models[].request_admission` attaches a process-shared `BaseRateLimiter` at the model factory. Identical explicit groups share one FIFO across model instances, threads and event loops; implicit groups use the configured model name. Policies are immutable once registered and conflicting settings fail construction. A monotonic minimum interval spaces requests without idle-time burst credit; bounded waiters poll without occupying executor threads and unregister in `finally`. The factory strips the policy from provider kwargs and sets exposed SDK `max_retries=0` so middleware retries re-enter admission. This limits model invocations, not tokens or a distributed provider account; custom providers bypassing BaseChatModel hooks are outside the contract. Tests: `test_model_request_admission.py`.
+- Optional `models[].request_admission` attaches a process-shared `BaseRateLimiter` at the model factory. Identical explicit groups share one FIFO across model instances, threads and event loops; implicit groups use the configured model name. Policies are immutable once registered and conflicting settings fail construction. A monotonic minimum interval spaces requests without idle-time burst credit; bounded waiters poll without occupying executor threads and unregister in `finally`. The factory strips the policy from provider kwargs and sets exposed SDK `max_retries=0` so middleware retries re-enter admission. This limits model invocations, not tokens or a distributed provider account; custom providers bypassing BaseChatModel hooks are outside the contract. Tests: `test_model_request_admission.py` and `test_model_request_admission_fifo_atomic.py`.
 
 ### Claude Code Credentials (`packages/harness/deerflow/models/credential_loader.py`)
 

backend/packages/harness/deerflow/models/request_admission.py

@@ -28,7 +28,7 @@ class RequestAdmission(BaseRateLimiter):
         self._lock = threading.Lock()
         self._waiters: deque[object] = deque()
 
-    def _try(self, ticket: object | None) -> bool:
+    def _try(self, ticket: object) -> bool:
         with self._lock:
             now = monotonic()
             if self._waiters and self._waiters[0] is not ticket:
@@ -38,13 +38,20 @@ class RequestAdmission(BaseRateLimiter):
             self._next = now + self._interval
             return True
 
-    def _enqueue(self) -> object:
-        ticket = object()
+    def _try_or_enqueue(self, *, blocking: bool) -> tuple[bool, object | None]:
+        """Atomically admit immediately or join the FIFO before newcomers can pass."""
         with self._lock:
+            now = monotonic()
+            if not self._waiters and now >= self._next:
+                self._next = now + self._interval
+                return True, None
+            if not blocking:
+                return False, None
             if len(self._waiters) >= self.config.max_queue_size:
                 raise AdmissionError("LLM admission queue is full; reduce workload or increase queue capacity.")
+            ticket = object()
             self._waiters.append(ticket)
-        return ticket
+            return False, ticket
 
     def _remove(self, ticket: object) -> None:
         with self._lock:
@@ -62,12 +69,12 @@ class RequestAdmission(BaseRateLimiter):
         return min(0.05, self._interval, until_next if until_next > 0 else self._interval, remaining)
 
     def acquire(self, *, blocking: bool = True) -> bool:
-        if self._try(None):
+        acquired, ticket = self._try_or_enqueue(blocking=blocking)
+        if acquired:
             return True
-        if not blocking:
+        if ticket is None:
             return False
         deadline = monotonic() + self.config.max_wait_seconds
-        ticket = self._enqueue()
         try:
             while True:
                 delay = self._delay(deadline)
@@ -78,12 +85,12 @@ class RequestAdmission(BaseRateLimiter):
             self._remove(ticket)
 
     async def aacquire(self, *, blocking: bool = True) -> bool:
-        if self._try(None):
+        acquired, ticket = self._try_or_enqueue(blocking=blocking)
+        if acquired:
             return True
-        if not blocking:
+        if ticket is None:
             return False
         deadline = monotonic() + self.config.max_wait_seconds
-        ticket = self._enqueue()
         try:
             while True:
                 delay = self._delay(deadline)

backend/tests/test_model_request_admission_fifo_atomic.py

@@ -0,0 +1,39 @@
+"""Regression coverage for atomic request-admission FIFO entry."""
+
+import pytest
+
+from deerflow.config.model_config import RequestAdmissionConfig
+from deerflow.models import request_admission as admission
+
+
+@pytest.fixture
+def clock(monkeypatch):
+    now = [0.0]
+    monkeypatch.setattr(admission, "monotonic", lambda: now[0])
+    return now
+
+
+def test_blocking_waiter_reserves_fifo_position_before_newcomer_can_take_due_slot(clock):
+    limiter = admission.RequestAdmission(RequestAdmissionConfig(requests_per_minute=60))
+    assert limiter.acquire(blocking=False)
+
+    acquired, ticket = limiter._try_or_enqueue(blocking=True)
+    assert acquired is False
+    assert ticket is limiter._waiters[0]
+
+    clock[0] = 1
+    assert limiter.acquire(blocking=False) is False
+
+    limiter._remove(ticket)
+    assert limiter.acquire(blocking=False) is True
+
+
+def test_nonblocking_probe_never_joins_the_fifo(clock):
+    limiter = admission.RequestAdmission(RequestAdmissionConfig(requests_per_minute=60))
+    assert limiter.acquire(blocking=False)
+
+    acquired, ticket = limiter._try_or_enqueue(blocking=False)
+
+    assert acquired is False
+    assert ticket is None
+    assert not limiter._waiters