verifiers 0.2.2.dev7__tar.gz → 0.2.2.dev9__tar.gz

{verifiers-0.2.2.dev7 → verifiers-0.2.2.dev9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.2.2.dev7
+Version: 0.2.2.dev9
 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.dev7 → verifiers-0.2.2.dev9}/verifiers/v1/agent.py

@@ -221,19 +221,19 @@ class Agent:
         task: Task,
         *,
         runtime: Runtime | None = None,
-        shared_tools: Mapping[str, SharedToolServer] | None = None,
+        tools: Mapping[str, SharedToolServer] | None = None,
         on_trace: Callable[[Trace], None] | None = None,
     ) -> Trace:
         """Run this agent on `task` once and return the trace: `runtime` places it
-        into a live borrowed box instead of provisioning one; `shared_tools` are
-        live servers borrowed from their owner, counted in the pairing check;
+        into a live borrowed box instead of provisioning one; `tools` are live
+        servers borrowed from their owner, counted in the pairing check;
         `on_trace` observes the trace the moment it's minted, before any I/O.
         Retries whole while the trace ends with a retryable error (`config.retries`)
         — never into a borrowed box; the final trace keeps earlier attempts' errors."""
         retry = self.config.retries
         history: list = []
         for attempt in range(retry.max_retries + 1):
-            trace = await self._run_once(task, runtime, shared_tools, on_trace)
+            trace = await self._run_once(task, runtime, tools, on_trace)
             if attempt == retry.max_retries or not trace_should_retry(trace, retry):
                 break
             if runtime is not None:
@@ -356,7 +356,7 @@ class _EpisodeAgent(Agent):
     they're created, finished ones land in `completed` (the episode's traces),
     each run takes the eval's gate. The taskset's shared tool servers ride only
     its own tasks — on an env-minted task they'd wrongly put MCP in play
-    (`shared_tools=` overrides)."""
+    (`tools=` overrides)."""
 
     def __init__(
         self,
@@ -392,7 +392,7 @@ class _EpisodeAgent(Agent):
         task: Task,
         *,
         runtime: Runtime | None = None,
-        shared_tools: Mapping[str, SharedToolServer] | None = None,
+        tools: Mapping[str, SharedToolServer] | None = None,
         on_trace: Callable[[Trace], None] | None = None,
     ) -> Trace:
         last: Trace | None = None
@@ -416,9 +416,7 @@ class _EpisodeAgent(Agent):
             trace = await super().run(
                 task,
                 runtime=runtime,
-                shared_tools=shared_tools
-                if shared_tools is not None
-                else self._shared_for(task),
+                tools=tools if tools is not None else self._shared_for(task),
                 on_trace=watch,
             )
         self._completed.append(trace)

verifiers-0.2.2.dev7/tests/AGENTS.md

@@ -1,5 +0,0 @@
-# Testing
-
-Use `uv run pytest` to run any tests, as it will include other dev dependencies. Running `pytest` directly will NOT include these dependencies.
-
-See `pyproject.toml` for more details.

verifiers-0.2.2.dev7/verifiers/envs/AGENTS.md

@@ -1,72 +0,0 @@
-# Environments
-
-**WIP** -- subject to change.
-
-## Overview
-
-Environment classes are organized into three categories:
-
-| Category | Location | Description |
-|----------|----------|-------------|
-| **Core** | `verifiers/envs/` | Stable base classes and sandbox environments |
-| **Integrations** | `verifiers/envs/integrations/` | Third-party library wrappers (lazy-loaded, require extras) |
-| **Experimental** | `verifiers/envs/experimental/` | Newer environments with sharper edges |
-
-Avoid storing global state in the environment. Instead, use the state argument to the environment's methods.
-
-## Core Environments
-
-Base classes for building environments:
-
-- `SingleTurnEnv` - Single response Q&A tasks
-- `MultiTurnEnv` - Multi-turn interactions (games, simulations, agents)
-- `ToolEnv` - Stateless, idempotent tools passed as Python functions with all arguments exposed to the model. For managing additional state, use `StatefulToolEnv`.
-- `StatefulToolEnv` - Tools requiring per-rollout state (e.g. sandbox ID). Use `update_tool_args` to inject state into tool calls. See `sandbox_env.py` for an example.
-- `SandboxEnv` - Sandboxed container execution using `prime` sandboxes. All sandbox setup logic should be included in the start command and queued via `setup_state`, but not awaited—await resources only when first needed to overlap provisioning with rollout. See `python_env.py` for an example.
-- `PythonEnv` - Persistent Python REPL in sandbox
-
-## Integrations
-
-Third-party library wrappers that require additional dependencies:
-
-| Environment | Extra | Install Command |
-|-------------|-------|-----------------|
-| `TextArenaEnv` | `ta` | `uv add 'verifiers[ta]'` |
-| `ReasoningGymEnv` | `rg` | `uv add 'verifiers[rg]'` |
-| `BrowserEnv` | `browser` | `uv add 'verifiers[browser]'` |
-
-When developing in the `verifiers` repo:
-```bash
-uv sync --extra ta       # for TextArenaEnv
-uv sync --extra rg       # for ReasoningGymEnv
-uv sync --extra browser  # for BrowserEnv
-```
-
-### TextArenaEnv
-
-Wrapper for text-based TextArena environments (games, simulations). When adding new TextArena environments, investigate the `textarena` source code to determine the observation format. Often you'll want to re-render observations via `feedback_fn`, as `verifiers` doesn't allow overwriting past messages (only concatenation), and many TextArena games return full game state rather than turn-level diffs.
-
-### ReasoningGymEnv
-
-Wrapper for [reasoning-gym](https://github.com/reasoning-gym/reasoning-gym) procedural datasets. Supports single datasets or composite mixtures via `DatasetSpec`.
-
-### BrowserEnv
-
-Browser automation environment using [Browserbase](https://browserbase.com) with two modes:
-
-**DOM mode** (`mode="dom"`): Natural language browser control via Stagehand SDK.
-
-**CUA mode** (`mode="cua"`): Vision-based browser control using coordinate-based primitives.
-
-CUA mode supports automatic sandbox deployment (`use_sandbox=True`, the default), where the CUA server is automatically deployed to a sandbox container. This eliminates the need for manual server setup. The server files are located in `assets/templates/browserbase/cua/` and are uploaded to sandboxes during rollout initialization.
-
-For local development, set `use_sandbox=False` and start the server manually from `assets/templates/browserbase/cua/`.
-
-## Experimental
-
-See `verifiers/envs/experimental/README.md` for documentation on:
-- `GymEnv` - Universal Gym-compatible environment runner
-- `MCPEnv` - MCP server integration
-- `CliAgentEnv` - Agent code in sandboxes with API call interception
-- V1 `vf.Env` / `vf.Taskset` / `vf.Harness` - taskset/harness agent composition
-- `HarborEnv` - Harbor-format task loading