verifiers 0.3.2.dev49__tar.gz → 0.3.2.dev51__tar.gz

{verifiers-0.3.2.dev49 → verifiers-0.3.2.dev51}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev49
+Version: 0.3.2.dev51
 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.dev49 → verifiers-0.3.2.dev51}/verifiers/v1/harnesses/prime_agent/harness.py

@@ -37,9 +37,9 @@ INSTALL = r"""
 set -e
 export PATH="/var/tmp/vf-node/bin:$PATH"
 prefix="$VF_PRIME_AGENT_DIR/$PRIME_AGENT_COMMIT"
-[ -x "$prefix/bin/prime-agent" ] && exit 0
+[ -x "$prefix/bin/prime-agent" ] && [ -f "$HOME/.prime/agent/kernel-venv/.bootstrap-version" ] && exit 0
 export NPM_CONFIG_PREFIX="$prefix"
-export PRIME_AGENT_BOOTSTRAP_KERNEL_ON_INSTALL=0
+export PRIME_AGENT_BOOTSTRAP_KERNEL_ON_INSTALL=1
 release_url="$VF_PRIME_AGENT_GITHUB_RELEASE_URL/v$PRIME_AGENT_RELEASE_VERSION"
 agent_tarball="prime-agent-$PRIME_AGENT_RELEASE_VERSION.tgz"
 ai_tarball="prime-agent-ai-$PRIME_AGENT_RELEASE_VERSION.tgz"
@@ -84,6 +84,7 @@ repacked="$(npm pack "$download_dir/package-root/package" \
 PRIME_AGENT_BOOTSTRAP_TOOLS_ON_INSTALL=1 npm install -g \
     --no-fund --no-audit --loglevel=error --progress=false \
     "$download_dir/repacked/$repacked"
+[ -f "$HOME/.prime/agent/kernel-venv/.bootstrap-version" ]
 """
 
 

{verifiers-0.3.2.dev49 → verifiers-0.3.2.dev51}/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()`.