verifiers 0.3.2.dev36__py3-none-any.whl → 0.3.2.dev38__py3-none-any.whl

verifiers/v1/harnesses/rlm/harness.py

@@ -5,7 +5,14 @@ import logging
 import shlex
 from typing import Any, Literal
 
-from pydantic import BaseModel, ConfigDict, Field, PositiveInt, model_validator
+from pydantic import (
+    BaseModel,
+    ConfigDict,
+    Field,
+    NonNegativeInt,
+    PositiveInt,
+    model_validator,
+)
 from pydantic_config import BaseConfig
 
 from verifiers.v1.acp import ACPConfig, ACPHarness, ACPTurn, JsonObject
@@ -17,7 +24,7 @@ from verifiers.v1.trace import Trace
 
 logger = logging.getLogger(__name__)
 
-BuiltinSkill = Literal["edit", "search"]
+BuiltinSkill = Literal["bash", "edit", "fetch", "search"]
 
 RLM_REPO = "github.com/PrimeIntellect-ai/nano-rlm.git"
 RLM_CACHE_DIR = "/tmp/vf-rlm"
@@ -40,18 +47,45 @@ class CompactionConfig(BaseConfig):
     summarize_at_tokens: PositiveInt | None = None
     """Compact at this token count. When unset, compact when 16k tokens remain below the
     model context window when the provider advertises it."""
+    max_compactions: PositiveInt | None = None
+    """Compactions per session before the engine stops compacting; `None` =
+    nano-rlm's default."""
 
 
 class RLMHarnessConfig(HarnessConfig):
-    version: str = Field(default="4ef3438", min_length=1)
-    """Git ref (branch, tag, or commit) of nano-rlm to install."""
-    max_depth: int = 0
-    """Recursion depth RLM may spawn sub-harnesses to."""
+    version: str = Field(default="dd2c04f", 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."""
-    compaction: CompactionConfig | None = None
-    """Context compaction policy. Set an empty config to use automatic thresholds."""
+    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
+    unset defers to nano-rlm's default (on, automatic thresholds, bounded by its
+    default 1M tree-token budget)."""
+    max_concurrent_subagents: PositiveInt | None = None
+    """Sub-agents running at once per session tree; `None` = nano-rlm's default (4).
+    nano-rlm requires it to be at least `max_depth`."""
+    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."""
+    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."""
+    append_to_system_prompt: str | None = None
+    """Appended to the root engine's system prompt, after the taskset's system prompt."""
+    subagent_append_to_system_prompt: str | None = None
+    """Append for sub-agent engines that can still delegate (depth >= 1, below
+    `max_depth`); unset falls back to `append_to_system_prompt`."""
+    leaf_append_to_system_prompt: str | None = None
+    """Append for depth-capped leaf engines; unset falls back to the sub-agent append."""
 
     @model_validator(mode="after")
     def reject_disabled_tools(self) -> "RLMHarnessConfig":
@@ -60,6 +94,15 @@ class RLMHarnessConfig(HarnessConfig):
                 "the rlm harness has a fixed tool set (ipython) and does not support "
                 "`disabled_tools`; use `builtin_skills` to enable built-in skills instead."
             )
+        if (
+            self.max_depth is not None
+            and self.max_concurrent_subagents is not None
+            and self.max_concurrent_subagents < self.max_depth
+        ):
+            raise ValueError(
+                "`max_concurrent_subagents` must be at least `max_depth` "
+                "(nano-rlm rejects the policy otherwise)."
+            )
         return self
 
 
@@ -109,6 +152,24 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
         system_prompt: str | None,
     ) -> JsonObject:
         compaction = self.config.compaction
+        policy_knobs: dict[str, Any] = {
+            "max_depth": self.config.max_depth,
+            "max_concurrent_subagents": self.config.max_concurrent_subagents,
+            "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,
+        }
+        if isinstance(compaction, bool):
+            policy_knobs["compaction"] = compaction
+        elif compaction is not None:
+            policy_knobs["compaction"] = True
+            policy_knobs["summarize_at_tokens"] = compaction.summarize_at_tokens
+            policy_knobs["max_compactions"] = compaction.max_compactions
+        appends = [
+            text
+            for text in (system_prompt, self.config.append_to_system_prompt)
+            if text
+        ]
         payload = {
             "session_id": trace.id,
             "model": ctx.model,
@@ -116,16 +177,17 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
                 "base_url": endpoint,
                 "api_key": secret,
             },
+            # None = passthrough: the key stays off the wire and nano-rlm's own
+            # default applies.
             "policy": {
-                "max_depth": self.config.max_depth,
-                "compaction": compaction is not None,
-                "summarize_at_tokens": (
-                    compaction.summarize_at_tokens if compaction else None
-                ),
-                "max_concurrent_subagents": max(4, self.config.max_depth),
+                key: value for key, value in policy_knobs.items() if value is not None
             },
             "system_prompt_path": None,
-            "append_to_system_prompt": system_prompt,
+            "append_to_system_prompt": "\n\n".join(appends) or None,
+            "subagent_append_to_system_prompt": (
+                self.config.subagent_append_to_system_prompt
+            ),
+            "leaf_append_to_system_prompt": self.config.leaf_append_to_system_prompt,
             "skills": list(self.config.builtin_skills),
             "kernel_env": runtime.env,
             "search_api_key": self.config.resolved_env.get("SERPER_API_KEY"),

{verifiers-0.3.2.dev36.dist-info → verifiers-0.3.2.dev38.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev36
+Version: 0.3.2.dev38
 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.dev36.dist-info → verifiers-0.3.2.dev38.dist-info}/RECORD

@@ -109,7 +109,7 @@ verifiers/v1/harnesses/pool/harness.py,sha256=HKAogPbyLK5QpvPQ1teHOh_IyRQX-or2JF
 verifiers/v1/harnesses/prime_agent/__init__.py,sha256=xfPC4RrlnyYWaf38tPAOZq4YL6M55lMpj5BoRLmaK8I,171
 verifiers/v1/harnesses/prime_agent/harness.py,sha256=byplbbIOFACCn2BvVZAxBSHdaPKBGERYz_hA1rgBj0E,11205
 verifiers/v1/harnesses/rlm/__init__.py,sha256=hwx51xTdEWzzYgMj-p4ETNcptoub3IPMDbEdJ4uB8Ns,122
-verifiers/v1/harnesses/rlm/harness.py,sha256=clbx7-HQtEyEiOYu3JPEH6LaH66VYgvG3k3qLn1Re_w,7062
+verifiers/v1/harnesses/rlm/harness.py,sha256=1a88tUM8l4BRf-In-dPyhtW9ilyH2GJdbh1ezL2BRcQ,10297
 verifiers/v1/harnesses/terminus_2/__init__.py,sha256=l1pwfg1HyVDEqT9efdnFWcXqaZwtKwe7TrplkyxZSA0,166
 verifiers/v1/harnesses/terminus_2/harness.py,sha256=kXKhi-aqqsoXGnvet0Tq_v7LYYpRYEzARyQL2MTEGWs,3101
 verifiers/v1/harnesses/terminus_2/program.py,sha256=X8MZdyItU3gbXGXnGOlJ9LWcUTnqIDaxlhkZvoRiqKY,2703
@@ -185,8 +185,8 @@ verifiers/v1/utils/prime.py,sha256=UTYRjp9cbjNb6CVmBHda-1wWZAIIfxNOmTyNuT7_wL4,9
 verifiers/v1/utils/retries.py,sha256=Y2ZgrAjn-qNkRKZP_RVNL_7EK0iaRcZeCa404lpGYi4,5417
 verifiers/v1/utils/score.py,sha256=493yJVMw8teCu9JxapxMFPFzI0hNUqdo0Y2nGW4kckk,6200
 verifiers/v1/utils/version.py,sha256=-obEo_-l9-D8FLef4hYxncOe-uJpxrM1g2Hig_37Sgs,1607
-verifiers-0.3.2.dev36.dist-info/METADATA,sha256=CPiuEseqw899eF8UoezmBH1wUJVCm4_DB26wx4jIXnA,4065
-verifiers-0.3.2.dev36.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
-verifiers-0.3.2.dev36.dist-info/entry_points.txt,sha256=uqQje0TMsr7k6nXlWxlPs13Si0evb3pVCEDF6c-YL80,241
-verifiers-0.3.2.dev36.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.3.2.dev36.dist-info/RECORD,,
+verifiers-0.3.2.dev38.dist-info/METADATA,sha256=XtmwcXdAJSovO0iDRdqt2zmGCOe8Mh7oaGIxQNa2Euo,4065
+verifiers-0.3.2.dev38.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
+verifiers-0.3.2.dev38.dist-info/entry_points.txt,sha256=uqQje0TMsr7k6nXlWxlPs13Si0evb3pVCEDF6c-YL80,241
+verifiers-0.3.2.dev38.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.3.2.dev38.dist-info/RECORD,,