verifiers 0.2.2.dev75__tar.gz → 0.2.2.dev77__tar.gz

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev75
+Version: 0.2.2.dev77
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/pyproject.toml

@@ -72,7 +72,7 @@ dev = [
     "textarena==0.7.4",
 ]
 # V1 examples and support packages, resolved from their local sources below.
-# Installed by default (tool.uv.default-groups) so the v1 CLIs (eval <id>, serve) run straight
+# Installed by default (tool.uv.default-groups) so the v1 CLIs (eval <id>, validate) run straight
 # from this repo. The built-in tasksets/harnesses ship inside verifiers (verifiers/v1/...).
 examples = [
     "compact",
@@ -152,7 +152,6 @@ eval = "verifiers.v1.cli.eval.main:main"
 validate = "verifiers.v1.cli.validate:main"
 debug = "verifiers.v1.cli.debug:main"
 replay = "verifiers.v1.cli.replay:main"
-serve = "verifiers.v1.cli.serve:main"
 init = "verifiers.v1.cli.init:main"
 gepa = "verifiers.v1.cli.gepa:main"
 vf-gepa = "verifiers.scripts.gepa:main"

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/__init__.py

@@ -22,7 +22,7 @@ from verifiers.v1.configs.legacy import LegacyEnvConfig
 from verifiers.v1.configs.retries import RetryConfig
 from verifiers.v1.configs.serve import (
     ElasticPoolConfig,
-    ServingConfig,
+    ServeConfig,
     StaticPoolConfig,
     pool_serve_kwargs,
 )
@@ -254,7 +254,7 @@ __all__ = [  # noqa: RUF022 - grouped by public API area
     "Env",
     "SingleAgentEnv",
     "EnvConfig",
-    "ServingConfig",
+    "ServeConfig",
     "LegacyEnvConfig",
     "resolve_env_field",
     "narrowed_env_annotation",

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/configs/cli/__init__.py

@@ -3,7 +3,6 @@
 from verifiers.v1.configs.cli.debug import DebugConfig
 from verifiers.v1.configs.cli.eval import EvalConfig
 from verifiers.v1.configs.cli.init import InitConfig
-from verifiers.v1.configs.cli.serve import ServeConfig
 from verifiers.v1.configs.cli.validate import ValidateConfig
 
-__all__ = ["DebugConfig", "EvalConfig", "InitConfig", "ServeConfig", "ValidateConfig"]
+__all__ = ["DebugConfig", "EvalConfig", "InitConfig", "ValidateConfig"]

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/configs/cli/eval.py

@@ -10,7 +10,7 @@ from verifiers.v1.clients import ClientConfig, EvalClientConfig
 from verifiers.v1.configs.cli.env import narrowed_env_annotation, resolve_env_field
 from verifiers.v1.configs.env import EnvConfig
 from verifiers.v1.configs.legacy import LegacyEnvConfig
-from verifiers.v1.configs.serve import ServingConfig
+from verifiers.v1.configs.serve import ServeConfig
 from verifiers.v1.envs.single_agent import SingleAgentEnvConfig
 from verifiers.v1.types import SamplingConfig
 
@@ -19,7 +19,7 @@ class EvalConfig(BaseConfig):
     env: SerializeAsAny[EnvConfig] = SingleAgentEnvConfig()
     """The environment — which env, its seed taskset, each agent, its knobs. Narrowed to
     the selected env's config class by the env id, else the taskset id."""
-    serve: ServingConfig = ServingConfig()
+    serve: ServeConfig = ServeConfig()
     """How the env is hosted under `--server`: the worker pool, each worker's episode
     bound. Ignored by an in-process run."""
     legacy: LegacyEnvConfig = LegacyEnvConfig()

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/configs/harness.py

@@ -20,6 +20,8 @@ class HarnessConfig(BaseConfig):
     """Extra program variables; harness-owned variables take precedence."""
     forward_env: list[str] = Field(default_factory=list)
     """Host variables to forward without writing secrets into config; explicit `env` wins."""
+    tool_timeout: float = Field(600.0, gt=0)
+    """Seconds a single MCP tool call may take; raise it for tools that boot a VM."""
     disabled_tools: list[str] | None = None
     skills: list[Path] = Field(default_factory=list)
     """Skill folders to upload into the program's skill discovery directory — each

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/configs/serve.py

@@ -1,8 +1,8 @@
 """How an env is served: the worker pool, where it binds, and each worker's bound.
 
 Serving is its own axis. `[env]` says what runs; this says how it's hosted, so an
-eval, a `serve` process, and a trainer configure it identically instead of each
-flattening pool knobs onto its own config."""
+eval and a trainer's env server configure it identically instead of each flattening
+pool knobs onto its own config."""
 
 from typing import Annotated, Literal
 
@@ -35,10 +35,10 @@ PoolConfig = Annotated[
 ]
 
 
-class ServingConfig(BaseConfig):
+class ServeConfig(BaseConfig):
     """The `[serve]` block: the worker pool, the ZMQ address, and each worker's
-    episode bound. Read by whoever hosts the env — the `serve` CLI, a server-backed
-    eval, a trainer's orchestrator."""
+    episode bound. Read by whoever hosts the env — a server-backed eval, a
+    trainer's env-server process."""
 
     pool: PoolConfig = Field(default_factory=ElasticPoolConfig)
     """Worker-pool sizing. `elastic` (default) starts at one worker and scales up on

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/harnesses/bash/harness.py

@@ -101,7 +101,8 @@ class BashHarness(Harness[BashHarnessConfig]):
                 + json.dumps(
                     {
                         "mcpServers": {
-                            name: {"url": url} for name, url in mcp_urls.items()
+                            name: {"url": url, "timeout": self.config.tool_timeout}
+                            for name, url in mcp_urls.items()
                         }
                     }
                 )

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/harnesses/bash/program.py

@@ -18,7 +18,7 @@ from tenacity import AsyncRetrying, stop_after_attempt, wait_exponential_jitter
 SERPER_URL = "https://google.serper.dev/search"
 
 MCP_CALL_ATTEMPTS = 6
-MCP_TIMEOUT = httpx.Timeout(600.0, connect=5.0)  # the OpenAI SDK client defaults
+MCP_TIMEOUT = 600.0
 
 
 BASH_TOOL = {
@@ -202,7 +202,8 @@ async def mcp_session(spec: dict):
     try:
         http_client = await stack.enter_async_context(
             create_mcp_http_client(
-                headers=spec.get("headers") or None, timeout=MCP_TIMEOUT
+                headers=spec.get("headers") or None,
+                timeout=httpx.Timeout(spec.get("timeout", MCP_TIMEOUT), connect=5.0),
             )
         )
         read, write, *_ = await stack.enter_async_context(

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/harnesses/browser_use/harness.py

@@ -104,7 +104,8 @@ class BrowserUseHarness(Harness[BrowserUseHarnessConfig]):
                 + json.dumps(
                     {
                         "mcpServers": {
-                            name: {"url": url} for name, url in mcp_urls.items()
+                            name: {"url": url, "timeout": self.config.tool_timeout}
+                            for name, url in mcp_urls.items()
                         }
                     }
                 )

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/harnesses/browser_use/program.py

@@ -36,7 +36,7 @@ from openai import AsyncOpenAI
 from tenacity import AsyncRetrying, stop_after_attempt, wait_exponential_jitter
 
 MCP_CALL_ATTEMPTS = 6
-MCP_TIMEOUT = httpx.Timeout(600.0, connect=5.0)  # the OpenAI SDK client defaults
+MCP_TIMEOUT = 600.0
 
 BROWSER_TOOL_TIMEOUT = 3600
 """Matches the bash harness's command timeout."""
@@ -203,7 +203,8 @@ async def mcp_session(spec: dict):
     try:
         http_client = await stack.enter_async_context(
             create_mcp_http_client(
-                headers=spec.get("headers") or None, timeout=MCP_TIMEOUT
+                headers=spec.get("headers") or None,
+                timeout=httpx.Timeout(spec.get("timeout", MCP_TIMEOUT), connect=5.0),
             )
         )
         read, write, *_ = await stack.enter_async_context(

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/harnesses/codex/harness.py

@@ -254,7 +254,7 @@ class CodexHarness(Harness[CodexHarnessConfig]):
                     (
                         f"{json.dumps(name, ensure_ascii=False)}="
                         f"{{url={json.dumps(url, ensure_ascii=False)},required=true,"
-                        "startup_timeout_sec=60.0,tool_timeout_sec=600.0}"
+                        f"startup_timeout_sec=60.0,tool_timeout_sec={self.config.tool_timeout}}}"
                     )
                     for name, url in mcp_urls.items()
                 )

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/harnesses/null/harness.py

@@ -53,7 +53,8 @@ class NullHarness(Harness[NullHarnessConfig]):
                 + json.dumps(
                     {
                         "mcpServers": {
-                            name: {"url": url} for name, url in mcp_urls.items()
+                            name: {"url": url, "timeout": self.config.tool_timeout}
+                            for name, url in mcp_urls.items()
                         }
                     }
                 )

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/harnesses/null/program.py

@@ -15,7 +15,7 @@ from openai import AsyncOpenAI
 from tenacity import AsyncRetrying, stop_after_attempt, wait_exponential_jitter
 
 MCP_CALL_ATTEMPTS = 6
-MCP_TIMEOUT = httpx.Timeout(600.0, connect=5.0)  # the OpenAI SDK client defaults
+MCP_TIMEOUT = 600.0
 
 
 async def chat(
@@ -43,7 +43,8 @@ async def mcp_session(spec: dict):
     try:
         http_client = await stack.enter_async_context(
             create_mcp_http_client(
-                headers=spec.get("headers") or None, timeout=MCP_TIMEOUT
+                headers=spec.get("headers") or None,
+                timeout=httpx.Timeout(spec.get("timeout", MCP_TIMEOUT), connect=5.0),
             )
         )
         read, write, *_ = await stack.enter_async_context(

{verifiers-0.2.2.dev75 → verifiers-0.2.2.dev77}/verifiers/v1/harnesses/openclaw/harness.py

@@ -166,7 +166,7 @@ class OpenClawHarness(Harness[OpenClawHarnessConfig]):
                         "url": url,
                         "transport": "streamable-http",
                         "connectionTimeoutMs": 60_000,
-                        "requestTimeoutMs": 600_000,
+                        "requestTimeoutMs": int(self.config.tool_timeout * 1000),
                     }
                     for name, url in mcp_urls.items()
                 }

verifiers-0.2.2.dev75/verifiers/v1/cli/serve.py

@@ -1,79 +0,0 @@
-"""Environment-server CLI entrypoint."""
-
-import sys
-from functools import partial
-
-from pydantic_config import cli
-
-from verifiers.v1.cli.resolve import (
-    extract_id,
-    narrow_config,
-    plugin_errors,
-    references_config_file,
-    with_positional_taskset,
-)
-from verifiers.v1.configs.cli.serve import ServeConfig
-from verifiers.v1.configs.serve import pool_serve_kwargs
-from verifiers.v1.serve import serve_env
-from verifiers.v1.utils.logging import setup_logging
-
-USAGE = "usage: uv run serve [<taskset-id>] [--env.id <id>] [--legacy.id <env-id> (v0)] [options] [@ file.toml]"
-
-
-def main(argv: list[str] | None = None) -> None:
-    argv = with_positional_taskset(list(sys.argv[1:]) if argv is None else list(argv))
-
-    if not argv or any(arg in ("-h", "--help") for arg in argv):
-        print(USAGE)
-        sys.argv = [sys.argv[0], "--help"]
-        with plugin_errors():
-            cli(narrow_config(ServeConfig, argv))
-        return
-    legacy_id = any(a == "--legacy.id" or a.startswith("--legacy.id=") for a in argv)
-    # An env-block flag (or a since-moved flat axis) skips the usage gate so the typed
-    # parse renders its did-you-mean instead of a bare usage line.
-    typed_axis = any(
-        a.startswith(("--env.", "--taskset.", "--harness.", "--serve.")) for a in argv
-    )
-    if (
-        not extract_id(argv, "env.taskset")
-        and not legacy_id
-        and not references_config_file(argv)
-        and not typed_axis
-    ):
-        raise SystemExit(
-            USAGE
-        )  # need a taskset (positional / --env.taskset.id), a v0 --legacy.id, or @ file.toml
-
-    with plugin_errors():
-        config_type = narrow_config(ServeConfig, argv)
-        sys.argv = [sys.argv[0], *argv]
-        config = cli(config_type)
-    if config.dry_run:
-        print(config.model_dump_json(indent=2, exclude_none=True))
-        return
-    level = "DEBUG" if config.verbose else "INFO"
-    setup_logging(level)
-
-    # The pool config decides in-process vs router + worker pool (static or elastic); the
-    # frontend speaks the same protocol either way. serve_env owns the SIGTERM teardown.
-    # Pool workers are spawned with no logging, so hand serve_env the same setup to apply
-    # in each one.
-    server_kwargs = (
-        {
-            "env_id": config.legacy.id,
-            "env_args": config.legacy.args,
-            "extra_env_kwargs": config.legacy.extra_env_kwargs,
-        }
-        if config.is_legacy
-        # `--serve.max-concurrent` is each v1 worker's episode bound; the legacy
-        # bridge has never had one.
-        else {"config": config.env, "max_concurrent": config.serve.max_concurrent}
-    )
-    serve_env(
-        **pool_serve_kwargs(config.serve.pool),
-        legacy=config.is_legacy,
-        address=config.serve.address,
-        log_setup=partial(setup_logging, level),
-        **server_kwargs,
-    )

verifiers-0.2.2.dev75/verifiers/v1/configs/cli/serve.py

@@ -1,62 +0,0 @@
-"""Environment-server CLI configuration."""
-
-from pydantic import AliasChoices, Field, SerializeAsAny, model_validator
-from pydantic_config import BaseConfig
-
-from verifiers.v1.configs.cli.env import narrowed_env_annotation, resolve_env_field
-from verifiers.v1.configs.env import EnvConfig
-from verifiers.v1.configs.legacy import LegacyEnvConfig
-from verifiers.v1.configs.serve import ServingConfig
-from verifiers.v1.envs.single_agent import SingleAgentEnvConfig
-
-
-class ServeConfig(BaseConfig):
-    """`uv run serve`: what to serve (`[env]`, or `[legacy]` for a classic v0 env) and
-    how it's hosted (`[serve]`)."""
-
-    env: SerializeAsAny[EnvConfig] = SingleAgentEnvConfig()
-    """The environment — which env, its seed taskset, each agent, its knobs. Narrowed to
-    the selected env's config class by the env id, else the taskset id."""
-    serve: ServingConfig = ServingConfig()
-    """How it's served: the worker pool, the bind address, each worker's episode bound."""
-    legacy: LegacyEnvConfig = LegacyEnvConfig()
-    """A classic (v0) environment to serve through the bridge instead of `[env]`."""
-    verbose: bool = Field(False, validation_alias=AliasChoices("verbose", "v"))
-    """Log at debug level instead of info."""
-    dry_run: bool = False
-    """Resolve + validate the config and dump it, then exit."""
-
-    @property
-    def is_legacy(self) -> bool:
-        """Whether this run goes through the v0 bridge: a legacy id and no v1 taskset."""
-        return self.legacy.id is not None and not self.env.taskset.id
-
-    @property
-    def env_id(self) -> str:
-        """The run's identifier: the v1 env's, else the v0 env id."""
-        return self.env.env_id or self.legacy.id or ""
-
-    @model_validator(mode="after")
-    def _refuse_mixed_run(self):
-        # A v0 id next to any v1 env identity leaves one of the two going nowhere, and
-        # which one depends on `is_legacy`: a taskset makes it False, so the v0 env never
-        # loads; a bare `--env.id` leaves it True, so the v0 env runs under the v1 name.
-        if self.legacy.id is None or not self.env.env_id:
-            return self
-        if self.env.taskset.id:
-            raise ValueError(
-                f"--legacy.id {self.legacy.id!r} is a classic (v0) env and can't combine "
-                f"with the v1 taskset {self.env.taskset.id!r}. Pairing a reusable env with "
-                f"a taskset is --env.id {self.legacy.id!r} (TOML: id under [env]); to run "
-                "the v0 env instead, drop the taskset."
-            )
-        raise ValueError(
-            f"--legacy.id {self.legacy.id!r} is a classic (v0) env and can't combine with "
-            f"the v1 env --env.id {self.env.id!r}: the v0 env is what would run, stamped "
-            "with the v1 env's name. Keep whichever one you meant to run."
-        )
-
-    @model_validator(mode="before")
-    @classmethod
-    def _resolve_env(cls, data):
-        return resolve_env_field(data, narrowed_env_annotation(cls))