verifiers 0.3.2.dev99__py3-none-any.whl → 0.3.2.dev100__py3-none-any.whl

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/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/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/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`."""

{verifiers-0.3.2.dev99.dist-info → verifiers-0.3.2.dev100.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev99
+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.dev99.dist-info → verifiers-0.3.2.dev100.dist-info}/RECORD

@@ -1,7 +1,7 @@
 verifiers/__init__.py,sha256=yDYsckVQbUkwoh9mU_hBeCLGvS_aWgAxuoNm5LxQpDI,517
 verifiers/cli/commands/init.py,sha256=DaT1NGrK2lnQWvHlrLkZ1D2Te49ll2Eh-DyOAhmCzfU,147
 verifiers/v1/__init__.py,sha256=qJKOsiUADKX4C1tMKsAPQXiJMo9-80FNeyPpAyRVshY,8352
-verifiers/v1/agent.py,sha256=h4zI924A-tLgezoGQaus9msogj1fSoC1Ugty01RFqU4,32206
+verifiers/v1/agent.py,sha256=rFjoAn9vpRXoyIZTzMgIuGb6HczTSuOenaB-U-x09to,32489
 verifiers/v1/env.py,sha256=CEhzjojhg5jqbOYruUnc5U7TtgDHDSx8NiiFA0qfbt8,18234
 verifiers/v1/episode.py,sha256=Fft0R2W6INAVeJhoE9b8h7L_KQSz7XuxP4FT_XhcsUY,5741
 verifiers/v1/errors.py,sha256=uQySmCUrjDYFkBIii1vBFPiEcPFPQZ9EGpas_6HvcQI,5279
@@ -42,7 +42,7 @@ verifiers/v1/clients/client.py,sha256=zqC_AkiD9pl0kxxIbupNiehbS-YnVdp3EMZ0LOHTd8
 verifiers/v1/clients/eval.py,sha256=cJucyUFsbF1JlG8nF0Tlw2EBMSoAF0YMklgYiOSoNpg,8322
 verifiers/v1/clients/train.py,sha256=w6bUY6GCBYlscOxz-TBWjbDbFoNSyEfQgDynbB7Ef80,18698
 verifiers/v1/configs/__init__.py,sha256=X7u6X7B3ieD1HZl79Tv1kg-yPHbT0QLvAebaGqxVtvA,317
-verifiers/v1/configs/agent.py,sha256=L5P6ombAi4U5wwsspMHAuM8111sB9AVdm8q6B5iV1lw,3251
+verifiers/v1/configs/agent.py,sha256=Aczd-QlNYin6yGjierybBu9wDocjSPZXtKyC0knE5QU,3326
 verifiers/v1/configs/client.py,sha256=E7ZxeWBFV_AP8MyfxXncTZs-Yn-QKdDS57HCAzqVTrQ,4437
 verifiers/v1/configs/env.py,sha256=-Bcdd4MEPfekVc_fnpiwsb9UDwCNdNCKrAHGoQuuXu8,7784
 verifiers/v1/configs/harness.py,sha256=OetAXz5ErJPvNxUd2_6iiP0JItCnJl5eS9aH6VoNywE,2569
@@ -110,7 +110,7 @@ verifiers/v1/harnesses/pi/harness.py,sha256=l3iZYXhSm4_dGHuxsfN20phCkTzNWYtTr--a
 verifiers/v1/harnesses/prime_agent/__init__.py,sha256=xfPC4RrlnyYWaf38tPAOZq4YL6M55lMpj5BoRLmaK8I,171
 verifiers/v1/harnesses/prime_agent/harness.py,sha256=cMvkNdqEutzCoX3lrmCQ5dAi_HSDV3Vhw6K1N-LiBtA,10630
 verifiers/v1/harnesses/rlm/__init__.py,sha256=hwx51xTdEWzzYgMj-p4ETNcptoub3IPMDbEdJ4uB8Ns,122
-verifiers/v1/harnesses/rlm/harness.py,sha256=70HXfw88G1dfDhSCjrqFUKRvY7a8ZIp0jifPc4EJNM8,12250
+verifiers/v1/harnesses/rlm/harness.py,sha256=xADdXtDey4w4zs8YR_CP-hR55beHDUXUMkMJD13vSsQ,12609
 verifiers/v1/harnesses/terminus_2/__init__.py,sha256=l1pwfg1HyVDEqT9efdnFWcXqaZwtKwe7TrplkyxZSA0,166
 verifiers/v1/harnesses/terminus_2/harness.py,sha256=DMTY1bzvydyjHqzmvBSBGKJsmUzYEgxqd215a7cG7tg,3052
 verifiers/v1/harnesses/terminus_2/program.py,sha256=X8MZdyItU3gbXGXnGOlJ9LWcUTnqIDaxlhkZvoRiqKY,2703
@@ -155,7 +155,7 @@ verifiers/v1/serve/types.py,sha256=CjZ6PWQlzSRkBG77gXfzKHnw_S_qKUm9dZSSxVgkv1w,1
 verifiers/v1/tasksets/__init__.py,sha256=5hTKSPLXSZTgKBbeKQI7FEp06c0ZsaQfuwvrvVFLgb8,712
 verifiers/v1/tasksets/harbor/__init__.py,sha256=JxThzHFYEkuSfQbVf3pEnjordgdfe4MU-QOsSSbyovU,326
 verifiers/v1/tasksets/harbor/env.py,sha256=rZEh-1IaHpfsrwSOn7qTq7uV0K7cZZ28bHB7BEWQdnc,3622
-verifiers/v1/tasksets/harbor/taskset.py,sha256=7riwbznDpxWj3ObJyxdviZ-JO7TjF7DJxON0BSKY-iw,33640
+verifiers/v1/tasksets/harbor/taskset.py,sha256=o7weK05AFjeyxQOFqkSvumx_1iGNe7wViF2zskJrR1I,33693
 verifiers/v1/tasksets/harbor/toolset.py,sha256=_7HXu-uuvqzhmJ1EX6c3KWLg0g0eRrLjCq1kBsZhIwo,3159
 verifiers/v1/tasksets/lean/__init__.py,sha256=oyv-GbJQBvCyGwo7Ysxm2NjZW-aX4j1uDoHU0f-E7xQ,767
 verifiers/v1/tasksets/lean/scoring.py,sha256=sfzsT6MUK0zdMgAQ57_P_0CBaPHy3M7go71QMr5hDrs,10599
@@ -188,8 +188,8 @@ 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.dev99.dist-info/METADATA,sha256=TkRluztpX99xbTGp_ofgFihjGJAEMBRCrAI6FAe487E,4237
-verifiers-0.3.2.dev99.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
-verifiers-0.3.2.dev99.dist-info/entry_points.txt,sha256=uqQje0TMsr7k6nXlWxlPs13Si0evb3pVCEDF6c-YL80,241
-verifiers-0.3.2.dev99.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.3.2.dev99.dist-info/RECORD,,
+verifiers-0.3.2.dev100.dist-info/METADATA,sha256=Bm_j6pVdQmjLPgQ2TAP7LS-ouM_sySd-EVJumEekPGY,4238
+verifiers-0.3.2.dev100.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
+verifiers-0.3.2.dev100.dist-info/entry_points.txt,sha256=uqQje0TMsr7k6nXlWxlPs13Si0evb3pVCEDF6c-YL80,241
+verifiers-0.3.2.dev100.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.3.2.dev100.dist-info/RECORD,,