agentprobe-testing 1.5.0__tar.gz → 1.5.1__tar.gz

{agentprobe_testing-1.5.0/agentprobe_testing.egg-info → agentprobe_testing-1.5.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentprobe-testing
-Version: 1.5.0
+Version: 1.5.1
 Summary: Adaptive chaos-testing for LLM agents: a live model-driven Injector that reads an agent's real trajectory and decides where to break something, instead of scripting perturbations in advance.
 License: Business Source License 1.1
         

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/__init__.py

@@ -61,7 +61,7 @@ from agentprobe.scenario import CommitPattern, FactPattern, GoalSpec, Scenario
 from agentprobe.scenarios.registry import ALL_SCENARIOS as TICKET_SCENARIOS
 from agentprobe.scenarios.registry import BY_ID as TICKET_SCENARIOS_BY_ID
 
-__version__ = "1.5.0"
+__version__ = "1.5.1"
 
 __all__ = [
     "FixedFaultInjector",

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/agents/target_agent.py

@@ -16,6 +16,8 @@ from agentprobe.llm import cacheable_system, create_deterministic, response_cost
 
 import anthropic
 
+
+from agentprobe.llm import make_client
 from agentprobe.agents.base import Agent, AgentAction
 
 SYSTEM_PROMPT = """You are a customer support agent. You have read tools \
@@ -44,7 +46,7 @@ class TargetAgent(Agent):
         max_tokens: int = 1024,
         system_prompt: Optional[str] = None,
     ):
-        self._client = anthropic.Anthropic()
+        self._client = make_client()
         self._model = model
         self._max_tokens = max_tokens
         self._system_prompt = system_prompt if system_prompt is not None else SYSTEM_PROMPT

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/classifier.py

@@ -11,6 +11,8 @@ from enum import Enum
 
 import anthropic
 
+
+from agentprobe.llm import make_client
 from agentprobe.injection import AppliedInjection
 from agentprobe.llm import cacheable_system, create_deterministic, response_cost_usd
 from agentprobe.trajectory import Step
@@ -126,7 +128,7 @@ def classify_response(
 ) -> ClassifiedResponse:
     """`steps` is the full trajectory, not pre-filtered to post-injection
     steps -- see _render_classification_prompt for why."""
-    client = anthropic.Anthropic()
+    client = make_client()
     response = create_deterministic(
         client,
         model=model,

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/domains/access_control/agent.py

@@ -15,6 +15,8 @@ from agentprobe.llm import cacheable_system, create_deterministic, response_cost
 
 import anthropic
 
+
+from agentprobe.llm import make_client
 from agentprobe.agents.base import Agent, AgentAction
 
 SYSTEM_PROMPT = """You are an IT access-control agent. You have read tools \
@@ -29,7 +31,7 @@ reply with a short final summary of what you did, in plain text."""
 
 class AccessControlAgent(Agent):
     def __init__(self, model: str = "claude-haiku-4-5-20251001", max_tokens: int = 1024):
-        self._client = anthropic.Anthropic()
+        self._client = make_client()
         self._model = model
         self._max_tokens = max_tokens
         self._messages: list[dict[str, Any]] = []

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/domains/access_control/complex_agent.py

@@ -17,6 +17,8 @@ from agentprobe.llm import cacheable_system, create_deterministic, response_cost
 
 import anthropic
 
+
+from agentprobe.llm import make_client
 from agentprobe.agents.base import Agent, AgentAction
 
 COMPLEX_SYSTEM_PROMPT = """You are a senior IT access-control specialist handling access requests \
@@ -62,7 +64,7 @@ summary of what you decided and why."""
 
 class ComplexAccessControlAgent(Agent):
     def __init__(self, model: str = "claude-haiku-4-5-20251001", max_tokens: int = 1536):
-        self._client = anthropic.Anthropic()
+        self._client = make_client()
         self._model = model
         self._max_tokens = max_tokens
         self._messages: list[dict[str, Any]] = []

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/domains/tau_retail/user_sim.py

@@ -41,10 +41,11 @@ class LLMUser:
     def _anthropic_complete(self, system: str, messages: list) -> tuple[str, float]:
         import anthropic
 
+        from agentprobe.llm import make_client
         from agentprobe.llm import response_cost_usd
 
         if self._client is None:
-            self._client = anthropic.Anthropic()
+            self._client = make_client()
         r = self._client.messages.create(model=self.model, max_tokens=512, system=system, messages=messages)
         return "".join(b.text for b in r.content if b.type == "text"), response_cost_usd(self.model, r)
 

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/domains/web_browsing/agent.py

@@ -17,6 +17,8 @@ from agentprobe.llm import cacheable_system, create_deterministic, response_cost
 
 import anthropic
 
+
+from agentprobe.llm import make_client
 from agentprobe.agents.base import Agent, AgentAction
 
 SYSTEM_PROMPT = """You are a customer-support agent working the store's own web \
@@ -34,7 +36,7 @@ with a short final summary of what you did, in plain text."""
 
 class WebBrowsingAgent(Agent):
     def __init__(self, model: str = "claude-haiku-4-5-20251001", max_tokens: int = 1024):
-        self._client = anthropic.Anthropic()
+        self._client = make_client()
         self._model = model
         self._max_tokens = max_tokens
         self._messages: list[dict[str, Any]] = []

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/injector.py

@@ -12,6 +12,8 @@ from typing import TYPE_CHECKING, Optional
 
 import anthropic
 
+
+from agentprobe.llm import make_client
 from agentprobe.injection import (
     AppliedInjection,
     ArmedInjection,
@@ -856,7 +858,7 @@ class ModelInjector(Injector):
         example."""
         if policy not in KIND_POLICIES:
             raise ValueError(f"unknown kind policy {policy!r}, must be one of {KIND_POLICIES}")
-        self._client = anthropic.Anthropic()
+        self._client = make_client()
         self._model = model
         self._kind_order = list(kind_order) if kind_order is not None else list(_DEFAULT_KIND_ORDER)
         self._policy = policy

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/llm.py

@@ -15,6 +15,7 @@ from __future__ import annotations
 
 from typing import Any
 
+import anthropic
 import anthropic._streaming as _anthropic_streaming
 
 _TEMPERATURE_SUPPORTED: dict[str, bool] = {}
@@ -110,6 +111,37 @@ def create_deterministic(client: Any, **kwargs: Any) -> Any:
                 raise
 
 
+# ---- per-request API key ------------------------------------------------------
+#
+# A hosted runner serves many customers from one process, so a customer's Anthropic key must
+# not be put in the process-wide ANTHROPIC_API_KEY variable: two overlapping requests would
+# overwrite each other's key and one customer's usage could be billed to another. The key is
+# held in a context variable instead, which is isolated per request/task.
+
+import contextlib as _contextlib
+import contextvars as _contextvars
+
+_api_key_var: "_contextvars.ContextVar[Any]" = _contextvars.ContextVar("agentprobe_api_key", default=None)
+
+
+@_contextlib.contextmanager
+def api_key_scope(api_key: str):
+    """Everything inside the `with` block (in this task/thread context) that builds an Anthropic
+    client uses `api_key`; other concurrent requests are unaffected."""
+    token = _api_key_var.set(api_key)
+    try:
+        yield
+    finally:
+        _api_key_var.reset(token)
+
+
+def make_client() -> Any:
+    """The Anthropic client every agentprobe component uses: the key from api_key_scope() if one
+    is active, otherwise the usual ANTHROPIC_API_KEY from the environment."""
+    key = _api_key_var.get()
+    return anthropic.Anthropic(api_key=key) if key else anthropic.Anthropic()
+
+
 _PRICE_PER_TOKEN: dict[str, tuple[float, float]] = {
     # (input, output) USD per token, from the pricing table this was
     # checked against (2026-08). Keyed by the exact model ids this

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/registry.py

@@ -43,26 +43,35 @@ class FetchedDomain:
     review_url: Optional[str] = None
 
 
-def _default_fetcher(url: str) -> dict:
+def _default_fetcher(url: str, api_key: Optional[str] = None) -> dict:
     # Cloudflare's default bot protection blocks urllib's stock User-Agent
     # ("Python-urllib/3.x") outright -- a real client identifier, same idea
     # as any other SDK sending its own name/version, gets through.
-    request = urllib.request.Request(url, headers={"User-Agent": "agentprobe-client/0.5.0"})
+    headers = {"User-Agent": "agentprobe-client/0.5.0"}
+    if api_key:
+        headers["Authorization"] = f"Bearer {api_key}"
+    request = urllib.request.Request(url, headers=headers)
     with urllib.request.urlopen(request, timeout=10) as resp:  # noqa: S310 -- https enforced by fetch_domain below
         return json.loads(resp.read().decode("utf-8"))
 
 
-def fetch_domain(key: str, base_url: str = DEFAULT_BASE_URL, fetcher: Optional[Fetcher] = None) -> FetchedDomain:
+def fetch_domain(
+    key: str, base_url: str = DEFAULT_BASE_URL, fetcher: Optional[Fetcher] = None, api_key: Optional[str] = None
+) -> FetchedDomain:
     """Fetch and load the Domain stored under `key`. Pass `fetcher` (a
     callable url -> dict) in tests instead of hitting the real network --
     see tests/test_registry.py.
+
+    A domain generated while signed in (with your API key) belongs to your account and can
+    only be fetched with that account's `api_key`; one generated anonymously stays fetchable
+    by anyone who has its key.
     """
     if not (base_url.startswith("https://") or base_url.startswith("http://localhost")):
         raise ValueError("fetch_domain requires an https:// base_url (http://localhost is allowed for local testing)")
     if not key:
         raise ValueError("fetch_domain requires a non-empty key")
 
-    fetcher = fetcher if fetcher is not None else _default_fetcher
+    fetcher = fetcher if fetcher is not None else (lambda url: _default_fetcher(url, api_key))
     payload = fetcher(f"{base_url}/v1/domains/{key}")
 
     if "source" not in payload:

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/summarize.py

@@ -61,7 +61,8 @@ def summarize_failures(
     if client is None:
         import anthropic
 
-        client = anthropic.Anthropic()
+        from agentprobe.llm import make_client
+        client = make_client()
     response = create_deterministic(
         client,
         model=model,

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/agentprobe/validate_scenarios.py

@@ -31,6 +31,8 @@ from typing import Any
 
 import anthropic
 
+
+from agentprobe.llm import make_client
 from agentprobe.domain import Domain, TICKET_DOMAIN
 from agentprobe.llm import create_deterministic
 from agentprobe.reachability import reachable
@@ -392,7 +394,7 @@ judgment call. Call verify_value_mismatch exactly once."""
 def _verify_value_mismatch_with_model(
     order_amount: float, required_amount: float, policies: dict[str, str], model: str = "claude-sonnet-5"
 ) -> str:
-    client = anthropic.Anthropic()
+    client = make_client()
     policy_text = "\n".join(f"{name}: {text}" for name, text in policies.items()) or "(no policies in this scenario)"
     prompt = (
         f"Order amount: {order_amount}\n"

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1/agentprobe_testing.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentprobe-testing
-Version: 1.5.0
+Version: 1.5.1
 Summary: Adaptive chaos-testing for LLM agents: a live model-driven Injector that reads an agent's real trajectory and decides where to break something, instead of scripting perturbations in advance.
 License: Business Source License 1.1
         

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "agentprobe-testing"
-version = "1.5.0"
+version = "1.5.1"
 description = "Adaptive chaos-testing for LLM agents: a live model-driven Injector that reads an agent's real trajectory and decides where to break something, instead of scripting perturbations in advance."
 requires-python = ">=3.11"
 dependencies = ["anthropic>=1.0.0", "python-dotenv>=1.0.0"]

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/tests/test_llm.py

@@ -241,3 +241,47 @@ def test_cache_is_per_model():
     accepting = FakeClient(reject_temperature=False)
     create_deterministic(accepting, model="claude-haiku-4-5-20251001", max_tokens=10)
     assert accepting.messages.calls[0]["extra_body"] == {"temperature": 0}
+
+
+def test_api_key_scope_gives_each_context_its_own_key_and_restores_afterwards(monkeypatch):
+    import threading
+
+    import anthropic
+
+    from agentprobe.llm import api_key_scope, make_client
+
+    seen = []
+
+    class FakeClient:
+        def __init__(self, api_key=None):
+            self.api_key = api_key
+
+    monkeypatch.setattr(anthropic, "Anthropic", FakeClient)
+    assert make_client().api_key is None  # no scope: falls back to the environment's key
+    with api_key_scope("sk-a"):
+        assert make_client().api_key == "sk-a"
+        # another thread/context never sees this request's key
+        t = threading.Thread(target=lambda: seen.append(make_client().api_key))
+        t.start()
+        t.join()
+        with api_key_scope("sk-b"):
+            assert make_client().api_key == "sk-b"
+        assert make_client().api_key == "sk-a"
+    assert make_client().api_key is None
+    assert seen == [None]
+
+
+def test_overlapping_async_requests_keep_their_own_keys():
+    import asyncio
+
+    from agentprobe.llm import _api_key_var, api_key_scope
+
+    async def request(key, delay):
+        with api_key_scope(key):
+            await asyncio.sleep(delay)  # another request runs while this one waits
+            return _api_key_var.get()
+
+    async def main():
+        return await asyncio.gather(request("sk-a", 0.02), request("sk-b", 0.0), request("sk-c", 0.01))
+
+    assert asyncio.run(main()) == ["sk-a", "sk-b", "sk-c"]

{agentprobe_testing-1.5.0 → agentprobe_testing-1.5.1}/tests/test_registry.py

@@ -417,3 +417,18 @@ def test_replay_run_round_trip_replays_the_exact_recorded_sequence():
 
     assert replayed.injector_model == "replay:42"
     assert replayed._compute()["total"] == 1  # the same recorded injection fired again
+
+
+def test_fetch_domain_sends_the_api_key_when_given(monkeypatch):
+    import agentprobe.registry as registry
+
+    seen = {}
+
+    def fake_default(url, api_key=None):
+        seen["url"], seen["key"] = url, api_key
+        return {"source": "raise SystemExit", "review_url": None}
+
+    monkeypatch.setattr(registry, "_default_fetcher", fake_default)
+    with pytest.raises(BaseException):  # the stub source isn't a real domain; only the request matters
+        registry.fetch_domain("k_abc", base_url="https://sampact.com", api_key="ap_secret")
+    assert seen == {"url": "https://sampact.com/v1/domains/k_abc", "key": "ap_secret"}