benchflow 0.7.6.dev2250__tar.gz → 0.7.6.dev2253__tar.gz

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: benchflow
-Version: 0.7.6.dev2250
+Version: 0.7.6.dev2253
 Summary: Multi-turn agent benchmarking with ACP — run any agent, any model, any provider.
 Project-URL: Homepage, https://github.com/benchflow-ai/benchflow
 Project-URL: Repository, https://github.com/benchflow-ai/benchflow

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "benchflow"
-version = "0.7.6.dev2250"
+version = "0.7.6.dev2253"
 description = "Multi-turn agent benchmarking with ACP — run any agent, any model, any provider."
 readme = "README.md"
 requires-python = ">=3.12"

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/src/benchflow/acp/runtime.py

@@ -26,6 +26,10 @@ from pathlib import Path
 from benchflow.acp.client import ACPClient
 from benchflow.acp.container_transport import ContainerTransport
 from benchflow.acp.selection import selected_acp_transport
+from benchflow.acp.timeout_cleanup import (
+    cancel_and_drain_prompt_task,
+    cancel_prompt_after_timeout,
+)
 from benchflow.acp.types import McpServerSpec
 from benchflow.acp.watchdog import IdleWatchdog
 from benchflow.agents.codex_config import apply_codex_launch_config
@@ -70,7 +74,6 @@ logger = logging.getLogger(__name__)
 
 _ACP_CONNECT_MAX_RETRIES = 3
 _ACP_CONNECT_BASE_DELAY = 2.0
-_PROMPT_CANCEL_DRAIN_TIMEOUT_SEC = 0.25
 _ACP_HANDSHAKE_TIMEOUT_ENV = "BENCHFLOW_ACP_HANDSHAKE_TIMEOUT"
 _ACP_HANDSHAKE_TIMEOUT_DEFAULT_SEC = 60.0
 _OPENHANDS_DISABLE_SUBAGENTS_ENV = "BENCHFLOW_OPENHANDS_DISABLE_SUBAGENTS"
@@ -773,31 +776,6 @@ async def execute_prompts(
     return trajectory, len(session.tool_calls)
 
 
-async def _cancel_and_drain_prompt_task(prompt_task: asyncio.Task) -> bool:
-    if prompt_task.done():
-        return True
-    prompt_task.cancel()
-    done, _pending = await asyncio.wait(
-        {prompt_task}, timeout=_PROMPT_CANCEL_DRAIN_TIMEOUT_SEC
-    )
-    if done:
-        with contextlib.suppress(BaseException):
-            prompt_task.result()
-        return True
-
-    logger.warning(
-        "ACP prompt task did not finish within %.2fs after cancellation",
-        _PROMPT_CANCEL_DRAIN_TIMEOUT_SEC,
-    )
-
-    def _consume_prompt_result(task: asyncio.Task) -> None:
-        with contextlib.suppress(BaseException):
-            task.result()
-
-    prompt_task.add_done_callback(_consume_prompt_result)
-    return False
-
-
 def _agent_prompt_timeout_error(session, timeout: int) -> AgentPromptTimeoutError:
     session.mark_prompt_end()
     pending_tool_call_ids = session.pending_tool_call_ids()
@@ -829,16 +807,19 @@ async def _prompt_with_wall_clock_budget(
 ):
     """Run a prompt until either it finishes or BenchFlow's budget expires."""
     prompt_task = asyncio.create_task(acp_client.prompt(prompt))
+    cleanup_attempted = False
     try:
         done, _pending = await asyncio.wait({prompt_task}, timeout=timeout)
         if done:
             return prompt_task.result()
-        if await _cancel_and_drain_prompt_task(prompt_task):
+        cleanup_attempted = True
+        if await cancel_prompt_after_timeout(acp_client, prompt_task):
             raise _agent_prompt_timeout_error(session, timeout)
         raise TimeoutError(f"Agent prompt exceeded wall-clock budget {timeout}s")
     finally:
-        if not prompt_task.done():
-            await _cancel_and_drain_prompt_task(prompt_task)
+        if not prompt_task.done() and not cleanup_attempted:
+            cleanup_attempted = True
+            await cancel_and_drain_prompt_task(prompt_task)
 
 
 async def _prompt_with_idle_watchdog(
@@ -850,6 +831,7 @@ async def _prompt_with_idle_watchdog(
 ):
     """Run one ACP prompt with wall-clock and idle-watchdog budgets."""
     prompt_task = asyncio.create_task(acp_client.prompt(prompt))
+    cleanup_attempted = False
     loop = asyncio.get_running_loop()
     watchdog = IdleWatchdog.start(
         session,
@@ -869,9 +851,13 @@ async def _prompt_with_idle_watchdog(
             now = loop.time()
             watchdog.observe(session, now=now)
             if watchdog.idle_expired(now):
-                raise watchdog.timeout_error(session, now=now)
+                error = watchdog.timeout_error(session, now=now)
+                cleanup_attempted = True
+                await cancel_prompt_after_timeout(acp_client, prompt_task)
+                raise error
             if watchdog.wall_clock_expired(now):
-                if await _cancel_and_drain_prompt_task(prompt_task):
+                cleanup_attempted = True
+                if await cancel_prompt_after_timeout(acp_client, prompt_task):
                     raise _agent_prompt_timeout_error(session, timeout)
                 raise TimeoutError(
                     f"Agent prompt exceeded wall-clock budget {timeout}s"
@@ -883,5 +869,6 @@ async def _prompt_with_idle_watchdog(
         # external-cancellation path (CancelledError from sleep). Bound the
         # drain so a non-cooperative Daytona/ACP read cannot hide the watchdog
         # timeout forever; cleanup will tear down the live process.
-        if not prompt_task.done():
-            await _cancel_and_drain_prompt_task(prompt_task)
+        if not prompt_task.done() and not cleanup_attempted:
+            cleanup_attempted = True
+            await cancel_and_drain_prompt_task(prompt_task)

benchflow-0.7.6.dev2253/src/benchflow/acp/timeout_cleanup.py

@@ -0,0 +1,90 @@
+"""Bounded cooperative cleanup for timed-out ACP prompts.
+
+The original prompt task remains the sole ACP response reader. A best-effort
+``session/cancel`` request runs independently while the prompt gets a short
+grace period to receive peer-authored terminal usage and tool updates. At the
+shared deadline, every surviving task is cancelled and boundedly drained.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import contextlib
+import logging
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+PROMPT_TIMEOUT_CLEANUP_TOTAL_SEC = 5.0
+PROMPT_CANCEL_DRAIN_TIMEOUT_SEC = 0.25
+
+
+def _consume_task_result(task: asyncio.Task) -> None:
+    with contextlib.suppress(BaseException):
+        task.result()
+
+
+async def _cancel_and_drain_tasks(tasks: set[asyncio.Task], timeout: float) -> None:
+    pending = {task for task in tasks if not task.done()}
+    for task in pending:
+        task.cancel()
+    _done, pending = (
+        await asyncio.wait(pending, timeout=timeout) if pending else (set(), set())
+    )
+    for task in tasks - pending:
+        _consume_task_result(task)
+    for task in pending:
+        task.add_done_callback(_consume_task_result)
+
+
+async def cancel_and_drain_prompt_task(prompt_task: asyncio.Task) -> bool:
+    """Hard-cancel a prompt and return whether it completed within the drain."""
+    if prompt_task.done():
+        _consume_task_result(prompt_task)
+        return True
+    await _cancel_and_drain_tasks({prompt_task}, PROMPT_CANCEL_DRAIN_TIMEOUT_SEC)
+    if prompt_task.done():
+        return True
+
+    logger.warning(
+        "ACP prompt task did not finish within %.2fs after cancellation",
+        PROMPT_CANCEL_DRAIN_TIMEOUT_SEC,
+    )
+    return False
+
+
+async def cancel_prompt_after_timeout(
+    acp_client: Any, prompt_task: asyncio.Task
+) -> bool:
+    """Request peer cancellation, then bound cleanup without adding a reader."""
+    cancel = getattr(acp_client, "cancel", None)
+    if not callable(cancel):
+        return await cancel_and_drain_prompt_task(prompt_task)
+    if prompt_task.done():
+        _consume_task_result(prompt_task)
+        return True
+
+    async def _request_cancel() -> None:
+        try:
+            await cancel()
+        except asyncio.CancelledError:
+            raise
+        except Exception:
+            logger.warning("Failed to request ACP prompt cancellation", exc_info=True)
+
+    loop = asyncio.get_running_loop()
+    deadline = loop.time() + PROMPT_TIMEOUT_CLEANUP_TOTAL_SEC
+    cancel_task = asyncio.create_task(_request_cancel())
+    try:
+        cooperative_timeout = max(
+            0.0,
+            deadline - loop.time() - PROMPT_CANCEL_DRAIN_TIMEOUT_SEC,
+        )
+        await asyncio.wait({prompt_task}, timeout=cooperative_timeout)
+    finally:
+        remaining = max(0.0, deadline - loop.time())
+        await _cancel_and_drain_tasks(
+            {prompt_task, cancel_task},
+            min(PROMPT_CANCEL_DRAIN_TIMEOUT_SEC, remaining),
+        )
+    return prompt_task.done()

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/src/benchflow/agents/env.py

@@ -469,6 +469,7 @@ def resolve_provider_env(
 ) -> None:
     """Detect provider for model, inject BENCHFLOW_PROVIDER_* and env_mapping."""
     from benchflow.agents.providers import (
+        ZAI_CODING_REGISTRY_BASE_ENV,
         find_provider,
         find_provider_for_bare_model,
         resolve_base_url,
@@ -515,6 +516,8 @@ def resolve_provider_env(
                 "BENCHFLOW_PROVIDER_BASE_URL",
                 base_url,
             )
+            if _prov_name == "zai-coding" and base_url:
+                agent_env[ZAI_CODING_REGISTRY_BASE_ENV] = "1"
         agent_env.setdefault(
             "BENCHFLOW_PROVIDER_PROTOCOL",
             agent_protocol or _prov_cfg.api_protocol,

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/src/benchflow/agents/providers.py

@@ -61,6 +61,8 @@ credential_files, url_params), ``vllm`` (user-supplied base_url).
 
 from dataclasses import dataclass, field
 
+ZAI_CODING_REGISTRY_BASE_ENV = "_BENCHFLOW_ZAI_CODING_REGISTRY_BASE"
+
 
 @dataclass
 class ProviderConfig:
@@ -275,6 +277,30 @@ PROVIDERS: dict[str, ProviderConfig] = {
             },
         ],
     ),
+    "zai-coding": ProviderConfig(
+        name="zai-coding",
+        base_url="https://api.z.ai/api/coding/paas/v4",
+        api_protocol="openai-completions",
+        auth_type="api_key",
+        auth_env="ZAI_API_KEY",
+        endpoints={
+            "openai-completions": "https://api.z.ai/api/coding/paas/v4",
+            "openai-responses": "https://api.z.ai/api/coding/paas/v4",
+            "anthropic-messages": "https://api.z.ai/api/anthropic",
+        },
+        models=[
+            {
+                "id": model,
+                "name": model.upper(),
+                "reasoning": True,
+                "input": ["text"],
+                "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0},
+                "contextWindow": 200000,
+                "maxTokens": 131072,
+            }
+            for model in ("glm-5.3", "glm-5.3-flash")
+        ],
+    ),
     "kimi": ProviderConfig(
         name="kimi",
         base_url="{base_url}",

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/src/benchflow/providers/litellm_config.py

@@ -9,6 +9,7 @@ from dataclasses import dataclass
 from urllib.parse import urlparse
 
 from benchflow.agents.providers import (
+    ZAI_CODING_REGISTRY_BASE_ENV,
     ProviderConfig,
     find_provider,
     resolve_base_url,
@@ -291,7 +292,11 @@ def _route_registered_provider(
     )
     explicit_api_base = (env.get("BENCHFLOW_PROVIDER_BASE_URL") or "").strip()
     explicit_api_key = (env.get("BENCHFLOW_PROVIDER_API_KEY") or "").strip()
-    if explicit_api_base and explicit_api_key:
+    zai_registry_base = (
+        provider_name == "zai-coding" and env.get(ZAI_CODING_REGISTRY_BASE_ENV) == "1"
+    )
+    explicit_route = explicit_api_base and explicit_api_key and not zai_registry_base
+    if explicit_api_base and not zai_registry_base:
         api_base = explicit_api_base
     else:
         try:
@@ -333,9 +338,13 @@ def _route_registered_provider(
     params: dict[str, str | int | float | bool | list[str]] = {"model": upstream}
     if api_base:
         params["api_base"] = api_base
+    native_key = (env.get(provider_cfg.auth_env or "") or "").strip()
+    explicit_zai_key = bool(
+        zai_registry_base and explicit_api_key and explicit_api_key != native_key
+    )
     api_key_ref = (
         _env_ref("BENCHFLOW_PROVIDER_API_KEY")
-        if explicit_api_base and explicit_api_key
+        if explicit_route or explicit_zai_key
         else _registered_api_key_ref(provider_cfg)
     )
     if api_key_ref:
@@ -344,7 +353,6 @@ def _route_registered_provider(
             required_env.append("BENCHFLOW_PROVIDER_API_KEY")
         elif provider_cfg.auth_env:
             required_env.append(provider_cfg.auth_env)
-
     return LiteLLMRoute(
         requested_model=model,
         model_alias=safe_model_alias(model),

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/src/benchflow/providers/litellm_runtime.py

@@ -1622,6 +1622,19 @@ async def ensure_litellm_runtime(
     routed through it: ``oracle`` (no model) and native-subscription auth (no
     API key to proxy). Gemini uses LiteLLM's native GenerateContent endpoints.
     """
+    # Re-entrant connects pass back proxy-owned env, which cannot reconstruct
+    # upstream routing or credentials. Restore controller-held source config.
+    if (
+        runtime is not None
+        and getattr(runtime, "kind", None) == "litellm"
+        and getattr(runtime, "source_agent", None) == agent
+        and getattr(runtime, "source_model", None) == model
+        and agent_env.get(LITELLM_MASTER_KEY_ENV)
+        == getattr(runtime, "master_key", None)
+        and getattr(runtime, "source_env", None) is not None
+    ):
+        agent_env = dict(runtime.source_env)
+
     usage_cfg = UsageTrackingConfig.coerce(usage_tracking).with_env_defaults()
 
     if uses_native_subscription_auth(agent, model, agent_env):
@@ -1745,6 +1758,9 @@ async def ensure_litellm_runtime(
         server=server,
         config_key=config_key,
         master_key=master_key,
+        source_agent=agent,
+        source_model=model,
+        source_env=dict(agent_env),
     )
     if live_trajectory_path is not None:
         server.start_live_capture(live_trajectory_path)

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/src/benchflow/providers/runtime.py

@@ -7,7 +7,7 @@ from the concrete host/sandbox process launcher.
 
 from __future__ import annotations
 
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
@@ -31,6 +31,10 @@ class ProviderRuntime:
     server: LiteLLMProcess | None = None
     config_key: str | None = None
     master_key: str | None = None
+    source_agent: str | None = None
+    source_model: str | None = None
+    # Controller-only upstream config for re-entrant connects and proxy restarts.
+    source_env: dict[str, str] | None = field(default=None, repr=False)
 
     @property
     def base_url(self) -> str:

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/src/benchflow/rollout/__init__.py

@@ -1371,6 +1371,9 @@ class Rollout:
             self._capture_partial_session_factory_trajectory()
         else:
             self._capture_partial_acp_trajectory()
+        collect_native_usage = getattr(self, "_collect_native_acp_usage", None)
+        if callable(collect_native_usage):
+            collect_native_usage()
         if self._acp_client:
             try:
                 await self._acp_client.close()
@@ -2385,6 +2388,7 @@ class Rollout:
                     role.agent, getattr(self, "_task", None), agent_cfg
                 ),
             )
+        self._native_usage_checkpoint = None
         self._reapply_ask_user_handler()
         self._attach_trajectory_writer(rollout_dir)
         self._active_role = role

{benchflow-0.7.6.dev2250 → benchflow-0.7.6.dev2253}/tests/test_acp.py

@@ -830,74 +830,6 @@ class TestACPInterleaving:
             await client.close()
 
 
-class TestACPIdleWatchdog:
-    @pytest.mark.asyncio
-    async def test_idle_watchdog_returns_even_when_prompt_cancel_drain_stalls(
-        self,
-    ) -> None:
-        """Guards the 2026-05-22 Daytona/Gemini blocker fix against stuck cancel drain.
-
-        When the idle watchdog fires it cancels the prompt task and drains it.
-        A non-cooperative agent — here one that swallows the cancellation and
-        blocks on an event that never arrives — must not be able to wedge the
-        watchdog: it bounds the drain and raises ``IdleTimeoutError`` anyway.
-
-        Determinism: this asserts the *behaviour* (idle error raised; the stuck
-        prompt task abandoned while still pending) rather than a wall-clock upper
-        bound, so it cannot be squeezed into a spurious failure when the full
-        suite loads the event loop. The outer ``wait_for`` is only a hang guard;
-        its timeout is deliberately loose (far above the real ~1.25s runtime), so
-        load can never trip it, while a regression to an unbounded drain hangs
-        past it and fails.
-        """
-        from benchflow.acp.runtime import IdleTimeoutError, execute_prompts
-
-        # release is never set while the watchdog runs, so the prompt task stays
-        # wedged in its cancellation handler — exactly the stuck-drain scenario.
-        class StubbornPromptClient:
-            def __init__(self) -> None:
-                self.release = asyncio.Event()
-                self.task: asyncio.Task | None = None
-
-            async def prompt(self, _prompt: str):
-                self.task = asyncio.current_task()
-                try:
-                    await asyncio.Future()
-                except asyncio.CancelledError:
-                    await self.release.wait()
-                    raise
-
-        client = StubbornPromptClient()
-        session = ACPSession("idle-session")
-
-        # Loose hang guard: ~4x the real runtime (1s idle detect + 0.25s bounded
-        # drain). Not a tight assertion — it only catches a true hang/regression.
-        hang_guard_sec = 5.0
-
-        try:
-            with pytest.raises(IdleTimeoutError, match="Agent idle for 1s"):
-                await asyncio.wait_for(
-                    execute_prompts(
-                        client,  # type: ignore[arg-type]
-                        session,
-                        ["solve"],
-                        timeout=30,
-                        idle_timeout=1,
-                    ),
-                    timeout=hang_guard_sec,
-                )
-            # The watchdog returned while the stuck prompt task is still wedged on
-            # release.wait(): proves it bounded the drain instead of waiting for a
-            # cancellation that never completes. Load-independent — no clock math.
-            assert client.task is not None
-            assert not client.task.done()
-        finally:
-            client.release.set()
-            if client.task is not None:
-                with pytest.raises(asyncio.CancelledError):
-                    await client.task
-
-
 class TestPendingToolCallGrace:
     @pytest.mark.asyncio
     async def test_lost_tool_call_completion_trips_idle_after_grace(self):