verifiers 0.4.2.dev44__py3-none-any.whl → 0.4.2.dev46__py3-none-any.whl

verifiers/v1/harnesses/rlm/harness.py

@@ -101,6 +101,9 @@ class RLMHarnessConfig(HarnessConfig):
     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."""
+    provider_max_retries: NonNegativeInt | None = None
+    """SDK retries per model call (429, 5xx, connection errors); `None` uses nano-rlm's
+    default (5). Raise it so long rollouts ride out extended rate-limit bursts."""
     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;
@@ -225,6 +228,11 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
             "provider": {
                 "base_url": endpoint,
                 "api_key": secret,
+                **(
+                    {"max_retries": self.config.provider_max_retries}
+                    if self.config.provider_max_retries is not None
+                    else {}
+                ),
             },
             # None = passthrough: the key stays off the wire and nano-rlm's own
             # default applies.

verifiers/v1/runtimes/backends/prime.py

@@ -43,16 +43,8 @@ logger = get_logger(__name__)
 EFFECTIVELY_UNBOUNDED_SECONDS = 30 * 24 * 60 * 60
 """Safety deadline for APIs that require a finite bound. Normal execution remains
 bounded by idle detection or rollout cancellation; 30 days is above any real run."""
-# The SDK reads finished jobs' output through a bounded pool whose deadline includes time
-# spent queued; its defaults (20 reads, 45s) expire most reads when thousands of rollouts
-# finish a job at once (e.g. every harness's setup at a 3k-rollout start).
-_OUTPUT_READS = 100
-"""Concurrent output reads on the shared client."""
-_OUTPUT_DEADLINE_SECONDS = 300
-"""Deadline for one read of a finished job's output, queue time included."""
-_OUTPUT_RETRIES = 10
-"""Re-reads of a finished job's output that the SDK still failed to fetch, before the
-exec is reported as failed."""
+_OUTPUT_TAIL_BYTES = 10 * 1024 * 1024
+"""Maximum retained bytes per command output stream."""
 _PUBLIC_DOMAIN = "sandbox-v2-prod.pinfra.io"
 """Domain of the public URLs `https://<port>-<sandbox-id>.<domain>` that a LAN network's
 `internet` rule publishes ports at."""
@@ -69,8 +61,8 @@ class _SharedClient:
 
 
 # One `AsyncSandboxClient` per event loop, leased by every live runtime on it. The SDK
-# coalesces concurrent status polls per client into batched requests (up to 100 ids
-# each), so sharing a client turns N runtimes' creation/job polls into a few batch
+# coalesces concurrent creation polls per client into batched requests (up to 100 ids
+# each), so sharing a client turns N runtimes' creation polls into a few batch
 # calls. Keyed by loop (not plain process-global) so a client's tasks stay on the loop
 # that created it.
 _shared_clients: dict[asyncio.AbstractEventLoop, _SharedClient] = {}
@@ -82,9 +74,7 @@ def _lease_client() -> Any:
     loop = asyncio.get_running_loop()
     shared = _shared_clients.get(loop)
     if shared is None:
-        shared = _shared_clients[loop] = _SharedClient(
-            AsyncSandboxClient(background_job_output_concurrency=_OUTPUT_READS)
-        )
+        shared = _shared_clients[loop] = _SharedClient(AsyncSandboxClient())
     shared.leases += 1
     return shared.client
 
@@ -339,46 +329,39 @@ class PrimeBackend(RuntimeBackend):
     async def run(
         self, runtime: Runtime, argv: list[str], env: dict[str, str]
     ) -> ProgramResult:
+        async def collect(stream: AsyncIterator[bytes]) -> bytes:
+            tail = bytearray()
+            async for chunk in stream:
+                tail.extend(chunk[-_OUTPUT_TAIL_BYTES:])
+                del tail[:-_OUTPUT_TAIL_BYTES]
+            return bytes(tail)
+
         try:
-            # Poll directly so rollout cancellation owns the execution timeout.
-            job = await self._client.start_background_job(
+            # Command sessions survive task cleanup of shared scratch directories.
+            # Close stdin explicitly: the live-process API keeps it open by default.
+            process = await self._client.open_process(
                 self.info.id,
-                shlex.join(argv),
+                "exec "
+                + shlex.join(["sh", "-c", 'exec "$@" < /dev/null', "sh", *argv]),
                 working_dir=self.config.workdir,
                 env=runtime.process_env(env),
                 user=runtime.user,
             )
-            delay = 0.1
-            output_retries = 0
-            missing = None
-            while True:
-                result = await self._client.get_background_job(
-                    self.info.id, job, timeout=_OUTPUT_DEADLINE_SECONDS
+            try:
+                stdout, stderr, exit_code = await asyncio.gather(
+                    collect(process.stdout), collect(process.stderr), process.wait()
                 )
-                # Under load the SDK can see a job finish yet miss its output (its bounded
-                # output reads expire while queued) and reports that as `*_error`. The job
-                # is done, so asking again re-reads the output.
-                missing = result.stdout_error or result.stderr_error
-                if result.completed and (
-                    missing is None or output_retries == _OUTPUT_RETRIES
-                ):
-                    break
-                if result.completed:
-                    output_retries += 1
-                await asyncio.sleep(delay)
-                delay = min(delay * 2, 3)
+            finally:
+                # Closing signals and reaps a running process, then releases its stream.
+                await run_shielded(process.aclose())
         except (
             Exception
         ) as e:  # a sandbox/API failure is one rollout's problem, not the eval's
             raise SandboxError(f"prime exec failed: {e}") from e
-        if missing is not None:
-            raise SandboxError(f"prime exec output unavailable: {missing}")
-        if result.exit_code is None:
-            raise SandboxError("prime exec completed without an exit code")
         return ProgramResult(
-            exit_code=result.exit_code,
-            stdout=result.stdout or "",
-            stderr=result.stderr or "",
+            exit_code=exit_code,
+            stdout=stdout.decode(errors="replace"),
+            stderr=stderr.decode(errors="replace"),
         )
 
     async def exec(

{verifiers-0.4.2.dev44.dist-info → verifiers-0.4.2.dev46.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.4.2.dev44
+Version: 0.4.2.dev46
 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-0.4.2.dev44.dist-info → verifiers-0.4.2.dev46.dist-info}/RECORD

@@ -105,7 +105,7 @@ verifiers/v1/harnesses/pi/harness.py,sha256=K2FtvDhoAUU_CdKVqE_NmDwsw3vqlSaZuLJW
 verifiers/v1/harnesses/prime_agent/__init__.py,sha256=xfPC4RrlnyYWaf38tPAOZq4YL6M55lMpj5BoRLmaK8I,171
 verifiers/v1/harnesses/prime_agent/harness.py,sha256=sV82MCi7uZTKlOH8Vg9_WSytgZNSbknShOhmYcdPPaE,12354
 verifiers/v1/harnesses/rlm/__init__.py,sha256=hwx51xTdEWzzYgMj-p4ETNcptoub3IPMDbEdJ4uB8Ns,122
-verifiers/v1/harnesses/rlm/harness.py,sha256=4Q6g93YJp9JQQWoCaNRYX6ojaMzoBGEuSHbChnJYprg,14873
+verifiers/v1/harnesses/rlm/harness.py,sha256=ehWNA6MgSlKJ8Ee9pC_7WAxIVKUzAImAJn1SJYj5K7M,15304
 verifiers/v1/harnesses/stirrup/__init__.py,sha256=a-NQxX1bOwaGx8QHj2jWcRJEGiLIxdY6XUMbojr9dOk,142
 verifiers/v1/harnesses/stirrup/harness.py,sha256=ypWsg89r4CTeMy5hQVS8z8CUt9AUWgSVDidKjxZDqag,3139
 verifiers/v1/harnesses/stirrup/program.py,sha256=ks7hxFVc1Y_wQGhlUdFI62LeJBrAYgczZ0aioXSynm8,14939
@@ -149,7 +149,7 @@ verifiers/v1/runtimes/runtime.py,sha256=-CvhH42y0F0dPeNgDRVBeajkRK_vZYp2qxKY1XVf
 verifiers/v1/runtimes/strategies.py,sha256=WtIrIpKmvjvqpUfwLUwDNJcgRLE8d-2sx7reks2S2PA,8765
 verifiers/v1/runtimes/backends/custom.py,sha256=Iz1Brd4dBTpO1zYVc2KA0u5sxz5ZUROY_Cdme5W33hM,2875
 verifiers/v1/runtimes/backends/docker.py,sha256=uOjmENxhdQ9GPrvtQCXi1-6TL3fSRcflg6sENBOTM-w,31591
-verifiers/v1/runtimes/backends/prime.py,sha256=qTTIrmMbEluKIwHQ4kwRnKXn6RIT-P3mZtkxgh_wJ-4,24758
+verifiers/v1/runtimes/backends/prime.py,sha256=vMD3d1JxyKWO5Zbd1c0DJ2I4DjdIFhi4XlgYabd1v4A,23884
 verifiers/v1/runtimes/backends/subprocess.py,sha256=naOWJGmcaOlu2ZdRbUwhJicJdxASDRAUhDT0jzDff3w,8591
 verifiers/v1/tasksets/__init__.py,sha256=2ijE3qIlLoUA02g6iysOwWQylyBBQL1GhSpjv8_bp0w,521
 verifiers/v1/tasksets/harbor/__init__.py,sha256=FjW4tM3rgGALMK-8p7ux26G4-d4ZFVmaAmO6Fni6e_g,214
@@ -187,8 +187,8 @@ verifiers/v1/utils/scope.py,sha256=zbs4pj6AVk4jywlAJ3AX0Q0_kzB9dBB02nY-9H4ELdQ,8
 verifiers/v1/utils/score.py,sha256=493yJVMw8teCu9JxapxMFPFzI0hNUqdo0Y2nGW4kckk,6200
 verifiers/v1/utils/trace_store.py,sha256=Qo9qvsJAl8PoRUavjQAfUB2VNxmUtYQtU2ejLKPzY9c,3382
 verifiers/v1/utils/version.py,sha256=-obEo_-l9-D8FLef4hYxncOe-uJpxrM1g2Hig_37Sgs,1607
-verifiers-0.4.2.dev44.dist-info/METADATA,sha256=1QsclZmVbRWmAKlpEy0XS6It5xZpqx0mfwtAzwiH7VA,5115
-verifiers-0.4.2.dev44.dist-info/WHEEL,sha256=2iZrJkDIkeDh0k9SikKMqFuRP-ua6mb4Ir1AsER4BEw,87
-verifiers-0.4.2.dev44.dist-info/entry_points.txt,sha256=qdMg7ajiFpDsHf7tuPMRclJXizu1J6LTWawS5YS0B0I,137
-verifiers-0.4.2.dev44.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.4.2.dev44.dist-info/RECORD,,
+verifiers-0.4.2.dev46.dist-info/METADATA,sha256=OTP5cYrKTTt2rIGIAZceZC14T36z-qgm_1MPcs2bVok,5115
+verifiers-0.4.2.dev46.dist-info/WHEEL,sha256=2iZrJkDIkeDh0k9SikKMqFuRP-ua6mb4Ir1AsER4BEw,87
+verifiers-0.4.2.dev46.dist-info/entry_points.txt,sha256=qdMg7ajiFpDsHf7tuPMRclJXizu1J6LTWawS5YS0B0I,137
+verifiers-0.4.2.dev46.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.4.2.dev46.dist-info/RECORD,,