verifiers 0.3.2.dev74__tar.gz → 0.3.2.dev76__tar.gz

{verifiers-0.3.2.dev74 → verifiers-0.3.2.dev76}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev74
+Version: 0.3.2.dev76
 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.dev74 → verifiers-0.3.2.dev76}/verifiers/v1/harnesses/rlm/harness.py

@@ -26,6 +26,7 @@ from verifiers.v1.trace import Trace
 logger = logging.getLogger(__name__)
 
 BuiltinSkill = Literal["bash", "edit", "fetch", "search"]
+BuiltinTool = Literal["bash", "edit", "fetch", "ipython"]
 
 RLM_REPO = "github.com/PrimeIntellect-ai/nano-rlm.git"
 RLM_CACHE_DIR = "/tmp/vf-rlm"
@@ -57,15 +58,20 @@ class CompactionConfig(BaseConfig):
 
 
 class RLMHarnessConfig(HarnessConfig):
-    version: str = Field(default="e384006", min_length=1)
+    version: str = Field(
+        default="ad081dbcf5e8c1d4e5b431b4b7d4dd5f30b7367c", 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 (RLM_SKILLS), e.g. `["edit"]`; empty enables none.
-    The tool set is fixed (ipython); the base `skills` field takes SKILL.md paths."""
+    """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_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."""
     compaction: CompactionConfig | bool | None = None
     """Context compaction: a `[compaction]` section (or `true`) enables it with the
     given settings, `false` disables it (an overflowing session then fails), and
@@ -74,12 +80,22 @@ class RLMHarnessConfig(HarnessConfig):
     max_concurrent_subagents: PositiveInt | None = None
     """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)."""
+    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."""
+    allow_git: bool | None = None
+    """Allow unrestricted Git history queries (e.g. `git log --all`); `None` uses
+    nano-rlm's default (false). Ordinary Git commands remain allowed either way;
+    this is not filesystem or process isolation."""
     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
     """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` = unbounded."""
+    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."""
     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."""
@@ -95,8 +111,8 @@ class RLMHarnessConfig(HarnessConfig):
     def reject_disabled_tools(self) -> "RLMHarnessConfig":
         if self.disabled_tools:
             raise ValueError(
-                "the rlm harness has a fixed tool set (ipython) and does not support "
-                "`disabled_tools`; use `builtin_skills` to enable built-in skills instead."
+                "the rlm harness does not support `disabled_tools`; select the tool "
+                "set with `builtin_tools` and skills with `builtin_skills` instead."
             )
         if (
             self.max_depth is not None
@@ -163,6 +179,9 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
         policy_knobs: dict[str, Any] = {
             "max_depth": self.config.max_depth,
             "max_concurrent_subagents": max_concurrent_subagents,
+            "max_subagent_calls": self.config.max_subagent_calls,
+            "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_tool_output_bytes": self.config.max_tool_output_bytes,
@@ -201,6 +220,8 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
             "kernel_env": runtime.env,
             "search_api_key": self.config.resolved_env.get("SERPER_API_KEY"),
         }
+        if self.config.builtin_tools is not None:
+            payload["builtin_tools"] = list(self.config.builtin_tools)
         return {RLM_RUNTIME_METADATA_KEY: payload}
 
     async def prepare_acp(

{verifiers-0.3.2.dev74 → verifiers-0.3.2.dev76}/verifiers/v1/utils/platform.py

@@ -166,13 +166,13 @@ def _close(
 
 
 def _losses(run: pr.Run) -> str | None:
-    """What the uploader could not store, or `None`. A sink that switched itself off
-    quietly (Prime Traces outside the beta) is not a loss and is not counted here."""
-    parts = [
+    """What the run could not finish or store, or `None`."""
+    parts = list(run.errors)
+    parts.extend(
         f"{count} record(s) not stored by the {sink} sink"
         for sink, count in sorted(run.failed_records.items())
         if count
-    ]
+    )
     if run.dropped_records:
         parts.append(f"{run.dropped_records} record(s) never queued (uploader overrun)")
     return "; ".join(parts) or None