verifiers 0.3.2.dev98__tar.gz → 0.3.2.dev100__tar.gz

{verifiers-0.3.2.dev98 → verifiers-0.3.2.dev100}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev98
+Version: 0.3.2.dev100
 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-0.3.2.dev98 → verifiers-0.3.2.dev100}/tests/v1/test_e2e.py

@@ -352,6 +352,10 @@ async def test_acp_resume_with_tool(run_v1, harness, harness_runtime, tmp_path):
     # populates trace.tools; the ACP transcript is the source of truth for use.
     assert "tool" in segments[1]["roles"]
     assert segments[1]["tool_outputs"]
+    if harness.id == "pi":
+        recall = next(tool for tool in trace.tools if tool.name == "resume_recall")
+        assert recall.parameters["properties"]["codeword"]["type"] == "string"
+        assert "codeword" in recall.parameters["required"]
     if harness.id == "rlm":
         assert "turns_since_last_compaction" in trace.metrics
         assert all(call.acp is not None for call in trace.calls)

{verifiers-0.3.2.dev98 → verifiers-0.3.2.dev100}/verifiers/v1/agent.py

@@ -54,11 +54,19 @@ __all__ = ["Agent", "AgentConfig", "Agents", "TimeoutConfig", "make_agent"]
 logger = logging.getLogger(__name__)
 
 
+DEFAULT_ROLLOUT_TIMEOUT = 4 * 3600.0
+"""Agent solve-attempt budget when neither the eval config nor the task sets one."""
+
+
 def resolve_rollout_timeouts(timeout: TimeoutConfig, task: Task) -> RolloutTimeouts:
     """Apply an agent's stage-timeout precedence to one task."""
     agent_timeout = (
         timeout.rollout if timeout.rollout is not None else task.data.timeout.agent
     )
+    if agent_timeout is None:
+        agent_timeout = DEFAULT_ROLLOUT_TIMEOUT
+    elif agent_timeout == 0:
+        agent_timeout = None  # explicit: unbounded
     return RolloutTimeouts(
         setup=timeout.setup if timeout.setup is not None else task.data.timeout.setup,
         agent=agent_timeout,

{verifiers-0.3.2.dev98 → verifiers-0.3.2.dev100}/verifiers/v1/configs/agent.py

@@ -16,7 +16,8 @@ class TimeoutConfig(BaseConfig):
     setup: float | None = None  # one shared budget: task setup + provisioning
     """Timeout (in seconds) for the task + harness setup hooks."""
     rollout: float | None = None
-    """Timeout (in seconds) for the agent's solve attempt."""
+    """Timeout (in seconds) for the agent's solve attempt. Unset: the task's own
+    timeout, else 4 hours. `0` disables the timeout."""
     finalize: float | None = None
     """Timeout (in seconds) for the task + harness finalize hooks."""
     scoring: float | None = None

{verifiers-0.3.2.dev98 → verifiers-0.3.2.dev100}/verifiers/v1/harnesses/pi/harness.py

@@ -137,7 +137,7 @@ class PiHarness(ACPHarness[PiHarnessConfig]):
             extension_path = f"{agent_dir}/mcp.js"
             mcp = {
                 "mcpServers": {
-                    name: {"url": url, "lifecycle": "eager"}
+                    name: {"url": url, "lifecycle": "eager", "directTools": True}
                     for name, url in mcp_urls.items()
                 }
             }

{verifiers-0.3.2.dev98 → verifiers-0.3.2.dev100}/verifiers/v1/harnesses/rlm/harness.py

@@ -68,16 +68,16 @@ class CompactionConfig(BaseConfig):
 
 class RLMHarnessConfig(HarnessConfig):
     version: str = Field(
-        default="ad081dbcf5e8c1d4e5b431b4b7d4dd5f30b7367c", min_length=1
+        default="00d8aa713c820023033bce6aa2eacea6c37e675d", min_length=1
     )
     """Git ref (branch, tag, or commit) of nano-rlm to install. Must know every
     field this harness puts on the wire, i.e. be at least the default ref."""
     max_depth: NonNegativeInt | None = None
     """Recursion depth RLM may spawn sub-agents to; `None` = nano-rlm's default (1).
     Set 0 to disable recursion."""
-    builtin_skills: list[BuiltinSkill] = Field(default_factory=list)
-    """Built-in rlm skills to enable (the contract's `skills`), e.g. `["edit"]`;
-    empty enables none. The base `skills` field takes SKILL.md paths."""
+    builtin_skills: list[BuiltinSkill] = Field(default_factory=lambda: ["edit"])
+    """Built-in rlm skills to enable (the contract's `skills`); `["edit"]` by default,
+    `[]` enables none. The base `skills` field takes SKILL.md paths."""
     builtin_tools: list[BuiltinTool] | None = None
     """Native tool selection; None uses nano-rlm's default and is omitted from the
     runtime contract. Explicit selection requires a nano-rlm ref supporting it."""
@@ -90,7 +90,7 @@ class RLMHarnessConfig(HarnessConfig):
     """Sub-agents running at once per session tree; `None` = nano-rlm's default (4),
     raised to an explicit `max_depth` when needed to keep the policy valid."""
     max_subagent_calls: PositiveInt | None = None
-    """Tree-total recursive call cap; `None` uses nano-rlm's default (64)."""
+    """Tree-total recursive call cap; `None` uses nano-rlm's default (uncapped)."""
     exec_timeout: PositiveInt | None = None
     """IPython/native tool execution timeout in seconds; `None` uses nano-rlm's
     default (300). Separate from `tool_timeout`, which controls MCP calls."""
@@ -101,10 +101,11 @@ class RLMHarnessConfig(HarnessConfig):
     max_total_turns: PositiveInt | None = None
     """Tree-total turn budget (one turn = one work-loop model call, any engine); every
     engine stops before its next call once spent. `None` = uncapped."""
-    max_total_tokens: PositiveInt | None = None
+    max_total_tokens: NonNegativeInt | None = 10_000_000
     """Tree-total budget of NEW tokens (completion + uncached prompt) across the session
-    tree; once spent every engine stops and no further sub-agents spawn. `None` uses
-    nano-rlm's default (1,000,000); it does not disable the budget."""
+    tree; once spent every engine stops and no further sub-agents spawn. 10M by default;
+    `0` removes the budget (the rollout timeout is then the only terminator); `None`
+    defers to nano-rlm's own default (1,000,000)."""
     max_tool_output_bytes: PositiveInt | None = None
     """Byte budget for a single tool result entering the conversation (middle truncation);
     overrides rlm's built-in 20KB default in either direction."""
@@ -192,7 +193,7 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
             "exec_timeout": self.config.exec_timeout,
             "allow_git": self.config.allow_git,
             "max_total_turns": self.config.max_total_turns,
-            "max_total_tokens": self.config.max_total_tokens,
+            "max_total_tokens": self.config.max_total_tokens or None,
             "max_tool_output_bytes": self.config.max_tool_output_bytes,
         }
         if isinstance(compaction, bool):
@@ -217,7 +218,13 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
             # None = passthrough: the key stays off the wire and nano-rlm's own
             # default applies.
             "policy": {
-                key: value for key, value in policy_knobs.items() if value is not None
+                **{k: v for k, v in policy_knobs.items() if v is not None},
+                # explicit null: nano-rlm reads it as unbounded (its own default is 1M)
+                **(
+                    {"max_total_tokens": None}
+                    if self.config.max_total_tokens == 0
+                    else {}
+                ),
             },
             "system_prompt_path": None,
             "append_to_system_prompt": "\n\n".join(appends) or None,

{verifiers-0.3.2.dev98 → verifiers-0.3.2.dev100}/verifiers/v1/tasksets/harbor/taskset.py

@@ -75,10 +75,10 @@ class HarborConfig(TasksetConfig):
     tasks: list[str] | None = None
     """Optional subset of task names to load (None = all)."""
     ignore_timeouts: bool = True
-    """Drop each task's declared agent and verifier timeouts so rollouts run
-    unbounded (unless run-level `--timeout.*` limits are set). Task timeouts are
-    authored against Harbor's runtime and confound model capability with inference
-    speed; set False to apply them anyway."""
+    """Drop each task's declared agent and verifier timeouts so rollouts use the
+    run-level `--timeout.*` limits or their defaults (4 h for the agent; `--timeout.rollout 0`
+    runs unbounded). Task timeouts are authored against Harbor's runtime and confound
+    model capability with inference speed; set False to apply them anyway."""
     timeout_multiplier: float = Field(1.0, gt=0)
     """Scale each task's agent and verifier timeouts. Only applies with
     `ignore_timeouts=False`."""