verifiers 0.3.2.dev50__py3-none-any.whl → 0.3.2.dev52__py3-none-any.whl

verifiers/v1/acp/__init__.py

@@ -78,12 +78,11 @@ class ACPHarness(Harness[ConfigT]):
         self, trace: Trace, response_metadata: dict[str, Any]
     ) -> None:
         """Attach optional protocol extensions understood by every ACP harness."""
-        if ACP_SEMANTIC_EDGES_METADATA_KEY not in response_metadata:
-            return
-        edge_set = SemanticEdgeSet.model_validate(
-            response_metadata[ACP_SEMANTIC_EDGES_METADATA_KEY]
-        )
-        trace.add_semantic_edges(edge_set)
+        if ACP_SEMANTIC_EDGES_METADATA_KEY in response_metadata:
+            edge_set = SemanticEdgeSet.model_validate(
+                response_metadata[ACP_SEMANTIC_EDGES_METADATA_KEY]
+            )
+            trace.add_semantic_edges(edge_set)
 
     @abstractmethod
     async def prepare_acp(

verifiers/v1/harnesses/rlm/harness.py

@@ -51,10 +51,13 @@ class CompactionConfig(BaseConfig):
     max_compactions: PositiveInt | None = None
     """Compactions per session before the engine stops compacting; `None` =
     nano-rlm's default."""
+    max_attempts: PositiveInt | None = None
+    """Summary-generation attempts within one compaction cycle; `None` =
+    nano-rlm's default (5)."""
 
 
 class RLMHarnessConfig(HarnessConfig):
-    version: str = Field(default="dd2c04f", min_length=1)
+    version: str = Field(default="e384006", 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
@@ -170,6 +173,7 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
             policy_knobs["compaction"] = True
             policy_knobs["summarize_at_tokens"] = compaction.summarize_at_tokens
             policy_knobs["max_compactions"] = compaction.max_compactions
+            policy_knobs["max_compaction_attempts"] = compaction.max_attempts
         appends = [
             text
             for text in (system_prompt, self.config.append_to_system_prompt)

verifiers/v1/runtimes/modal.py

@@ -27,6 +27,7 @@ from verifiers.v1.runtimes.base import (
     RuntimeProcess,
 )
 from verifiers.v1.runtimes.limiters import creation_limiter
+from verifiers.v1.utils.aio import run_shielded
 
 logger = logging.getLogger(__name__)
 
@@ -128,25 +129,7 @@ class ModalRuntime(Runtime):
                 creation_limiter(self.config.creates_per_sec, "modal-sandbox")
                 or contextlib.nullcontext()
             ):
-                self._sandbox = await modal.Sandbox.create.aio(
-                    "sleep",
-                    "infinity",  # keep-alive entrypoint; the harness runs via `exec`
-                    app=app,
-                    name=self.name,
-                    # Clear the image's ENTRYPOINT so `sleep infinity` runs as the command
-                    # rather than as args to it — otherwise an image with its own entrypoint
-                    # (e.g. SWE task images) never starts the keep-alive and the sandbox dies.
-                    image=modal.Image.from_registry(self.config.image).entrypoint([]),
-                    workdir=self.config.workdir,
-                    env=self.env,
-                    cpu=self.config.cpu,
-                    memory=int(self.config.memory * 1024),  # Modal memory is MB
-                    gpu=self.config.gpu,
-                    region=self.config.region,
-                    block_network=not self.config.network_access,
-                    timeout=24 * 60 * 60,  # Maximum lifetime of any sandbox.
-                    encrypted_ports=[SERVICE_PORT],
-                )
+                await run_shielded(self._create_sandbox(app))
             self.info.id = self._sandbox.object_id
             logger.info(
                 "modal: sandbox %s up (image=%s)", self.info.id, self.config.image
@@ -157,6 +140,37 @@ class ModalRuntime(Runtime):
         ) as e:  # provisioning failure is one rollout's problem, not the eval's
             raise SandboxError(f"modal sandbox provisioning failed: {e}") from e
 
+    async def _create_sandbox(self, app) -> None:
+        """Create the sandbox and take ownership of the handle, as one shielded step.
+
+        Modal schedules the sandbox before `Sandbox.create` returns, so a cancellation
+        landing between those two points leaves a sandbox that boots and bills with
+        `_sandbox` still unset — invisible to both `teardown` and the atexit backstop, and
+        alive until its 24h maximum lifetime. Assigning inside the coroutine that
+        `run_shielded` owns keeps the cancellation pending until the handle is ours.
+        """
+        import modal
+
+        self._sandbox = await modal.Sandbox.create.aio(
+            "sleep",
+            "infinity",  # keep-alive entrypoint; the harness runs via `exec`
+            app=app,
+            name=self.name,
+            # Clear the image's ENTRYPOINT so `sleep infinity` runs as the command rather
+            # than as args to it — otherwise an image with its own entrypoint (e.g. SWE
+            # task images) never starts the keep-alive and the sandbox dies.
+            image=modal.Image.from_registry(self.config.image).entrypoint([]),
+            workdir=self.config.workdir,
+            env=self.env,
+            cpu=self.config.cpu,
+            memory=int(self.config.memory * 1024),  # Modal memory is MB
+            gpu=self.config.gpu,
+            region=self.config.region,
+            block_network=not self.config.network_access,
+            timeout=24 * 60 * 60,  # Maximum lifetime of any sandbox.
+            encrypted_ports=[SERVICE_PORT],
+        )
+
     async def expose(self, port: int) -> str | None:
         # Publish a server hosted IN the sandbox: Modal forwards `port` (named via
         # `encrypted_ports` at creation) to a public URL, read back from `sandbox.tunnels()`.

verifiers/v1/semantic.py

@@ -63,8 +63,8 @@ class SemanticEdgeSet(BaseModel):
     """A harness-published set of semantic edges over logical request IDs.
 
     Edge labels are intentionally extensible. Initial harnesses use ``continuation``,
-    ``compaction``, ``subagent_call``, and ``subagent_return``; consumers must preserve
-    unknown labels.
+    ``compaction_attempt``, ``compaction``, ``subagent_call``, and ``subagent_return``;
+    consumers must preserve unknown labels.
     """
 
     model_config = ConfigDict(extra="forbid", strict=True)

verifiers/v1/trace.py

@@ -20,11 +20,7 @@ from verifiers.v1.configs.agent import AgentConfig, WireAgentConfig
 from verifiers.v1.errors import ProviderError
 from verifiers.v1.graph import RECORD_FLOAT_DECIMALS, MessageNode
 from verifiers.v1.runtimes import RuntimeInfo
-from verifiers.v1.semantic import (
-    ACPInfo,
-    ParentLink,
-    SemanticEdgeSet,
-)
+from verifiers.v1.semantic import ACPInfo, ParentLink, SemanticEdgeSet
 from verifiers.v1.state import State, StateT
 from verifiers.v1.task import DataT, WireTaskData
 from verifiers.v1.types import (
@@ -201,6 +197,8 @@ class Branch(BaseModel):
     index: int
     nodes: list[MessageNode]
     calls: list[ModelCall] = Field(default_factory=list)
+    trainable: bool = True
+    """Whether this physical path contributes a training sample."""
     mm_token_type_id_map: dict[int, int] = Field(default_factory=dict)
     """The trace's `mm_token_type_id_map`, carried so `mm_token_type_ids` is self-contained."""
 
@@ -492,8 +490,18 @@ class Trace(BaseModel, Generic[DataT, StateT, AgentConfigT]):
 
     @property
     def branches(self) -> list[Branch]:
-        """One root-to-leaf path per graph leaf, its calls attached in path order."""
+        """One root-to-leaf view per graph leaf, with no duplicated node storage.
+
+        A compaction attempt is a dangling physical leaf. It is trainable only when
+        the final semantic graph shows that the harness resumed from it.
+        """
         by_node = {c.node: c for c in self.calls if c.node is not None}
+        accepted_compaction_attempts = {
+            link.node
+            for node in self.nodes
+            for link in node.semantic_parents
+            if link.type == "compaction"
+        }
         branches: list[Branch] = []
         for i, leaf in enumerate(graph.leaves(self)):
             path: list[int] = []
@@ -502,11 +510,19 @@ class Trace(BaseModel, Generic[DataT, StateT, AgentConfigT]):
                 path.append(nid)
                 nid = self.nodes[nid].parent
             path.reverse()
+            is_compaction_attempt = any(
+                link.type == "compaction_attempt"
+                for link in self.nodes[leaf].semantic_parents
+            )
             branches.append(
                 Branch(
                     index=i,
                     nodes=[self.nodes[n] for n in path],
                     calls=[by_node[n] for n in path if n in by_node],
+                    trainable=(
+                        not is_compaction_attempt
+                        or leaf in accepted_compaction_attempts
+                    ),
                     mm_token_type_id_map=self.mm_token_type_id_map,
                 )
             )

{verifiers-0.3.2.dev50.dist-info → verifiers-0.3.2.dev52.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev50
+Version: 0.3.2.dev52
 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.dev50.dist-info → verifiers-0.3.2.dev52.dist-info}/RECORD

@@ -9,14 +9,14 @@ verifiers/v1/graph.py,sha256=K6wKkoT_8TpNqPQISu2Xf0eHtAYW_KUGw0CteN5zr3c,36529
 verifiers/v1/harness.py,sha256=Yx-XCIs1HAmqv1F-nhCTDBUwAx51xQWNWgFcWL1WRTM,13777
 verifiers/v1/judge.py,sha256=uqn-djkK2_7Q3kOjkDo6ipOvV_HOvuLa9y6AoDjRux8,9618
 verifiers/v1/rollout.py,sha256=FT6ZYfSqkKOFa5tm52XDTLijKxqpeLr3x2MBdLIk4xk,24021
-verifiers/v1/semantic.py,sha256=wssvBSEpY_tgzpFUADG9-N4HkDXgCZvNQ1pdf6qAsS0,4680
+verifiers/v1/semantic.py,sha256=VQ_buw6Ya6QtRoh_fiYI5Avs5RmIQCfWH0H0bQc4aoc,4704
 verifiers/v1/session.py,sha256=xM7JJJVJACk0JfH9RAN1Ts8gJJKr2tGGRiIgQTrdieY,18692
 verifiers/v1/state.py,sha256=EckF2bWp-vV4b1jYJ9sLI5xrfGuI5spIgYYwW926toI,595
 verifiers/v1/task.py,sha256=9YhllEDgoHSz72-1fcYwBADl-rOzASnrSVimLheWmwI,12237
 verifiers/v1/taskset.py,sha256=fp2E0IEhL_Ybj9cegZwljfTmW27p_30zTWHFKw_kXE4,4374
-verifiers/v1/trace.py,sha256=8KCGMF2iaUNH_PypKNx--QwbSsl5dUCGeXp9Dmkm-24,29623
+verifiers/v1/trace.py,sha256=dBrStZXoEhVIKtcUzG6lZN0SEAjATOqBZCrgbaPdIXw,30398
 verifiers/v1/types.py,sha256=PH1SxRWXOjPCgQD9aZgM18_bECvFSNQ25gkCbulmTKU,9857
-verifiers/v1/acp/__init__.py,sha256=wUPZVYBFjJuOnIIoYONBFuA04A1DJf_dOyKY3QwK37Q,12722
+verifiers/v1/acp/__init__.py,sha256=sbJL7vTE_BEmw34wwgbNPuWuAmQ_gwELsK3wQbT08f4,12715
 verifiers/v1/acp/runner.py,sha256=gMXFmsVg2qD0Zhrx7kvC6Me2Z-wsfBamRIO0i45UW8g,11097
 verifiers/v1/cli/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 verifiers/v1/cli/debug.py,sha256=x3bqOwwbwwfwQQpxHrZeKdxsW3ZVtUT9cAn_r4aOro4,11631
@@ -112,7 +112,7 @@ verifiers/v1/harnesses/pool/harness.py,sha256=TsELYhW4mMOt_laiDN6nAQJxFL3iCcKvck
 verifiers/v1/harnesses/prime_agent/__init__.py,sha256=xfPC4RrlnyYWaf38tPAOZq4YL6M55lMpj5BoRLmaK8I,171
 verifiers/v1/harnesses/prime_agent/harness.py,sha256=CgrKBXoelwRiSH9fzsNnz4oykz_iDJ3p5WVGBOB9AeU,10906
 verifiers/v1/harnesses/rlm/__init__.py,sha256=hwx51xTdEWzzYgMj-p4ETNcptoub3IPMDbEdJ4uB8Ns,122
-verifiers/v1/harnesses/rlm/harness.py,sha256=DX6xPBdxdXhJz-LClThAj7f1ZJ5tgN03H_WBbYhKce4,10445
+verifiers/v1/harnesses/rlm/harness.py,sha256=A-3lUD4t6kND5Hq01wedTbAcCWAlQvjqn-QQKBIYbi4,10671
 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
@@ -143,7 +143,7 @@ verifiers/v1/mcp/toolset.py,sha256=G2wzNdEMGLXoQUcrwF0TartJkqokNPI3o7YUOAaaa6U,1
 verifiers/v1/runtimes/__init__.py,sha256=ppc1R_sx-aMcuJSx4t99rTSjloUXI0M6EUj676-1PeY,2862
 verifiers/v1/runtimes/base.py,sha256=SRTUNAcdW0xJIWjfS5RcijFXqdKyXmGvp2tvV58eu58,16826
 verifiers/v1/runtimes/limiters.py,sha256=C6cWDD4pwyFwUY4QFke3oiMXwydSNIfSbJSIglUC2kE,3062
-verifiers/v1/runtimes/modal.py,sha256=Lgc0OkQEWHVBHCqBE2ja-mxtRYGserixlO4B9b5Vwn4,12661
+verifiers/v1/runtimes/modal.py,sha256=yyrqI6agpzi0lp5jR_ExqKwAq8tRgLvHVa3Sd3oh6RU,13222
 verifiers/v1/runtimes/prime.py,sha256=1IWRLKdGgscs8JmvkBdaS3au4tV0KUqv4W1gHJvOE2Q,18781
 verifiers/v1/runtimes/subprocess.py,sha256=IO82fI1wFbeTtV7T1AHSF8iTUyT9yGJFIUmCKKw6CnQ,7763
 verifiers/v1/runtimes/docker/__init__.py,sha256=K6xbJAT1F3-PtgP5Gum2UJKQ8Re1Yz90BP8Uc7Npec4,20594
@@ -189,8 +189,8 @@ 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.dev50.dist-info/METADATA,sha256=mTpWsGyXsT89Id7rFwue1VwNmTo5zmMJDt_E-X-trlk,4153
-verifiers-0.3.2.dev50.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
-verifiers-0.3.2.dev50.dist-info/entry_points.txt,sha256=uqQje0TMsr7k6nXlWxlPs13Si0evb3pVCEDF6c-YL80,241
-verifiers-0.3.2.dev50.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.3.2.dev50.dist-info/RECORD,,
+verifiers-0.3.2.dev52.dist-info/METADATA,sha256=g2XajQ0FKtnxFleFE_R5mxzDMXVNOapplm1pEpfhJ0o,4153
+verifiers-0.3.2.dev52.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
+verifiers-0.3.2.dev52.dist-info/entry_points.txt,sha256=uqQje0TMsr7k6nXlWxlPs13Si0evb3pVCEDF6c-YL80,241
+verifiers-0.3.2.dev52.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.3.2.dev52.dist-info/RECORD,,