Python interceptor reentrancy guard is task-local, not thread-local

[AndresL230]

Python interceptor reentrancy guard is task-local, not thread-local

Severity: High
Affected repos: middleware-python
Component boundary: middleware-python interceptor

Symptom

_interceptor.py uses contextvars.ContextVar (_in_interceptor) as the reentrancy guard. ContextVars are scoped to the current asyncio task. They do not propagate when work moves to a different OS thread (e.g., ThreadPoolExecutor, sync HTTP libraries called from inside async patches, libraries that internally chain HTTP across threads).

Concrete failure mode: an async httpx call is being intercepted (the contextvar is set). The httpx implementation, for some operations, dispatches sync urllib3 work to a thread pool. The thread pool worker has a fresh context — the contextvar is unset — so the urllib3 patch fires too, double-counting the call. Conversely, a requests library that internally retries via a separate thread can recurse through the patched urllib3 and not be guarded.

Evidence

• middleware-python/recost/_interceptor.py — _in_interceptor: contextvars.ContextVar[bool] = ContextVar(...); _in_interceptor.set(True) is called inside each patch wrapper.
• Both sync (urllib3) and async (httpx) patches use the same guard.

Impact

• Double-counted events in mixed sync/async stacks. Telemetry overstates request volume.
• Possible infinite loop if the patched library does any internal thread hop that re-enters the patch.

Fix recommendation

Use both — a contextvar for async tasks and a threading.local for thread-bound calls:

import threading
import contextvars

_in_interceptor_task: contextvars.ContextVar[bool] = ContextVar("_in_interceptor_task", default=False)
_in_interceptor_thread = threading.local()

def _is_in_interceptor() -> bool:
    return _in_interceptor_task.get() or getattr(_in_interceptor_thread, "flag", False)

def _enter_interceptor():
    token = _in_interceptor_task.set(True)
    _in_interceptor_thread.flag = True
    return token

def _exit_interceptor(token):
    _in_interceptor_task.reset(token)
    _in_interceptor_thread.flag = False

Apply to every patch wrapper. Same pattern is unnecessary in the Node SDK because Node is single-threaded for user code.

Verification

• Add a test: asyncio.run(asyncio.gather(*[asyncio.to_thread(requests.get, url) for _ in range(10)])) and assert event count equals 10, not >10.
• Add a test that mixes httpx async + urllib3 sync in the same task, asserts exactly one event per outbound call.