verifiers 0.3.2.dev61__py3-none-any.whl → 0.3.2.dev63__py3-none-any.whl

verifiers/v1/harnesses/prime_agent/harness.py

@@ -3,7 +3,6 @@
 import hashlib
 import json
 import logging
-import shlex
 from typing import Literal
 
 from verifiers.v1.acp import ACPConfig, ACPHarness, ACPTurn
@@ -256,25 +255,17 @@ class PrimeAgentHarness(ACPHarness[PrimeAgentHarnessConfig]):
         if system_prompt:
             args += ["--append-system-prompt", system_prompt]
 
-        wrapper = f"{root}/prime-agent"
-        await runtime.write(
-            wrapper,
-            (
-                "#!/bin/sh\n"
-                "set -eu\n"
-                f'export PATH="{NODE_BIN_DIR}:$HOME/.local/bin:$PATH"\n'
-                f'exec {shlex.join(args)} "$@"\n'
-            ).encode(),
-        )
-        executable = await runtime.run(["chmod", "700", wrapper], {})
-        if executable.exit_code != 0:
-            raise RuntimeError(
-                f"prime-agent wrapper chmod failed: {executable.stderr.strip()[-500:]}"
-            )
-
         return ACPConfig(
             env=self._env(trace, secret),
-            command=[wrapper],
+            # Expand the sandbox's PATH while keeping every agent argument literal.
+            command=[
+                "/bin/sh",
+                "-eu",
+                "-c",
+                f'export PATH="{NODE_BIN_DIR}:$HOME/.local/bin:$PATH"; exec "$@"',
+                "prime-agent",
+                *args,
+            ],
             prompt=prompt,
         )
 

verifiers/v1/utils/artifacts.py

@@ -91,12 +91,33 @@ async def collect(
             raise RuntimeError(f"artifact {artifact.source!r} declared more than once")
         seen.add(artifact.source)
 
+    # Batch roots to save remote round trips. Leave room below the shell argument
+    # limit for the command itself and further quoting by runtime transports.
+    batches: list[list[str]] = [[]]
+    batch_bytes = 0
+    for artifact in entries:
+        source_bytes = len(shlex.quote(artifact.source).encode()) + 1
+        if batches[-1] and batch_bytes + source_bytes > 8 * 1024:
+            batches.append([])
+            batch_bytes = 0
+        batches[-1].append(artifact.source)
+        batch_bytes += source_bytes
+
+    existence: list[str] = []
+    for sources in batches:
+        output = await _run(
+            runtime,
+            f"for source in {shlex.join(sources)}; do "
+            'if test -e "$source" || test -L "$source"; then echo 1; else echo 0; fi; '
+            "done",
+            "check artifact roots",
+        )
+        existence.extend(output.splitlines())
     collected: dict[str, bytes | None] = {}
     budget = MAX_ARTIFACT_BYTES
-    for artifact in entries:
+    for artifact, exists in zip(entries, existence, strict=True):
         source = artifact.source
-        exists = f"test -e {shlex.quote(source)} || test -L {shlex.quote(source)}"
-        if (await runtime.run(["sh", "-c", exists], {})).exit_code != 0:
+        if exists != "1":
             if not artifact.required:
                 collected[source] = None
                 continue
@@ -205,8 +226,9 @@ def _validate_restore(root: str, archive: bytes | None) -> None:
         raise RuntimeError(f"unreadable artifact archive for {root!r}: {exc}") from exc
 
 
-async def _run(runtime: Runtime, command: str, action: str) -> None:
+async def _run(runtime: Runtime, command: str, action: str) -> str:
     result = await runtime.run(["sh", "-c", command], {})
     if result.exit_code:
         detail = (result.stderr or result.stdout).strip()[-500:]
         raise RuntimeError(f"failed to {action}: {detail}")
+    return result.stdout

{verifiers-0.3.2.dev61.dist-info → verifiers-0.3.2.dev63.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev61
+Version: 0.3.2.dev63
 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.dev61.dist-info → verifiers-0.3.2.dev63.dist-info}/RECORD

@@ -110,7 +110,7 @@ verifiers/v1/harnesses/pi/harness.py,sha256=11V9d6_npjE9vEEgIG7y80kfaX1Ilfi3lbYs
 verifiers/v1/harnesses/pool/__init__.py,sha256=HTYsNiWGdiNEsSHj4xhPmshJSNb7SHjQ7d4yZQ4p3HM,127
 verifiers/v1/harnesses/pool/harness.py,sha256=TsELYhW4mMOt_laiDN6nAQJxFL3iCcKvckQWYve2XiE,3571
 verifiers/v1/harnesses/prime_agent/__init__.py,sha256=xfPC4RrlnyYWaf38tPAOZq4YL6M55lMpj5BoRLmaK8I,171
-verifiers/v1/harnesses/prime_agent/harness.py,sha256=CgrKBXoelwRiSH9fzsNnz4oykz_iDJ3p5WVGBOB9AeU,10906
+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=A-3lUD4t6kND5Hq01wedTbAcCWAlQvjqn-QQKBIYbi4,10671
 verifiers/v1/harnesses/terminus_2/__init__.py,sha256=l1pwfg1HyVDEqT9efdnFWcXqaZwtKwe7TrplkyxZSA0,166
@@ -172,7 +172,7 @@ verifiers/v1/tasksets/textarena/__init__.py,sha256=Os2OlBY_pSH0B1DP32RitumgSLpq2
 verifiers/v1/tasksets/textarena/taskset.py,sha256=9fa_unlFiFZyuN4I02rW3x4qyRnFWXqwTUZTloQefu0,4433
 verifiers/v1/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 verifiers/v1/utils/aio.py,sha256=0yNFOqB2oO6Qktc6NI_ZSAGuo1aHsIK_wuT1BtmoZ3I,1532
-verifiers/v1/utils/artifacts.py,sha256=GowZubXGk6NhVgaAUKCVXYoBh_3L1Aw0LAHe6tNwl8g,8447
+verifiers/v1/utils/artifacts.py,sha256=PFrSkwAyvzHyT6JjlFeGOPIXDmS-uRajaN3am4HZ0TI,9263
 verifiers/v1/utils/compile.py,sha256=5pxy31wx2Q-3pQjOO5B9SH6OT-wwb8r4F1LAl6HMeQA,5835
 verifiers/v1/utils/decorators.py,sha256=vhoN6YoKb15-Fkqmc9RaNlUIbt9WRvNK4zAn2bAxfS4,5367
 verifiers/v1/utils/format.py,sha256=NfQo9M5KMzWCZpQaet5YtsJx56vCKAZPfbjZZJLJhhM,2187
@@ -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.dev61.dist-info/METADATA,sha256=7o5tiEoE-ol4J9KtYpx1306yqLvQ6cEKtF0inhzekjQ,4204
-verifiers-0.3.2.dev61.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
-verifiers-0.3.2.dev61.dist-info/entry_points.txt,sha256=uqQje0TMsr7k6nXlWxlPs13Si0evb3pVCEDF6c-YL80,241
-verifiers-0.3.2.dev61.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.3.2.dev61.dist-info/RECORD,,
+verifiers-0.3.2.dev63.dist-info/METADATA,sha256=gUF-jEoP0VpPXkADaNMPJUezstcC2q9qL06f985QPoM,4204
+verifiers-0.3.2.dev63.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
+verifiers-0.3.2.dev63.dist-info/entry_points.txt,sha256=uqQje0TMsr7k6nXlWxlPs13Si0evb3pVCEDF6c-YL80,241
+verifiers-0.3.2.dev63.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.3.2.dev63.dist-info/RECORD,,