verifiers 0.1.15.dev154__tar.gz → 0.1.15.dev156__tar.gz

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.1.15.dev154
+Version: 0.1.15.dev156
 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-0.1.15.dev154 → verifiers-0.1.15.dev156}/tests/test_client_multimodal_types.py

@@ -1,7 +1,9 @@
-import pytest
 from types import SimpleNamespace
 
+import pytest
+
 from verifiers.clients.openai_chat_completions_client import OpenAIChatCompletionsClient
+from verifiers.errors import EmptyModelResponseError
 from verifiers.types import (
     AssistantMessage,
     ImageUrlContentPart,
@@ -18,6 +20,11 @@ from verifiers.types import (
 from verifiers.utils.response_utils import parse_response_message
 
 
+class _OpenAIMessage(SimpleNamespace):
+    def model_dump(self):
+        return self.__dict__
+
+
 @pytest.mark.asyncio
 async def test_openai_to_native_prompt_with_typed_multimodal_content_parts():
     client = OpenAIChatCompletionsClient(object())
@@ -52,6 +59,53 @@ async def test_openai_to_native_prompt_with_typed_multimodal_content_parts():
     ]
 
 
+@pytest.mark.asyncio
+async def test_openai_chat_rejects_reasoning_only_native_response():
+    client = OpenAIChatCompletionsClient(object())
+    native_response = SimpleNamespace(
+        choices=[
+            SimpleNamespace(
+                message=_OpenAIMessage(
+                    content=None,
+                    reasoning_content="hidden chain",
+                    tool_calls=None,
+                )
+            )
+        ]
+    )
+
+    with pytest.raises(EmptyModelResponseError, match="reasoning but no content"):
+        await client.raise_from_native_response(native_response)
+
+
+@pytest.mark.asyncio
+async def test_openai_chat_accepts_refusal_with_reasoning_native_response():
+    client = OpenAIChatCompletionsClient(object())
+    native_response = SimpleNamespace(
+        id="chatcmpl_refusal",
+        created=0,
+        model="gpt-5.2",
+        usage=None,
+        choices=[
+            SimpleNamespace(
+                finish_reason="stop",
+                message=_OpenAIMessage(
+                    content=None,
+                    refusal="I cannot help with that.",
+                    reasoning_content="hidden chain",
+                    tool_calls=None,
+                ),
+            )
+        ],
+    )
+
+    await client.raise_from_native_response(native_response)
+    response = await client.from_native_response(native_response)
+
+    assert response.message.content == "I cannot help with that."
+    assert response.message.reasoning_content == "hidden chain"
+
+
 @pytest.mark.asyncio
 async def test_anthropic_to_native_prompt_with_typed_multimodal_content_parts():
     pytest.importorskip("anthropic")
@@ -228,6 +282,24 @@ async def test_anthropic_from_native_response_always_parses_reasoning():
     assert response.message.content == "final answer"
 
 
+@pytest.mark.asyncio
+async def test_anthropic_rejects_reasoning_only_native_response():
+    pytest.importorskip("anthropic")
+    from verifiers.clients.anthropic_messages_client import AnthropicMessagesClient
+
+    client = AnthropicMessagesClient(object())
+    native_response = SimpleNamespace(
+        id="msg_think",
+        model="claude-haiku-4-5",
+        stop_reason="end_turn",
+        usage=SimpleNamespace(input_tokens=1, output_tokens=1),
+        content=[SimpleNamespace(type="thinking", thinking="hidden chain")],
+    )
+
+    with pytest.raises(EmptyModelResponseError, match="reasoning but no content"):
+        await client.raise_from_native_response(native_response)
+
+
 @pytest.mark.asyncio
 async def test_anthropic_tool_call_round_trips_thinking_blocks():
     pytest.importorskip("anthropic")

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/tests/test_composable_env.py

@@ -17,6 +17,12 @@ from verifiers.envs.experimental.composable import (
     TaskSet,
     discover_sibling_dir,
 )
+from verifiers.envs.experimental.composable.harnesses.mini_swe_agent import (
+    build_mini_swe_agent_install_script,
+)
+from verifiers.envs.experimental.composable.harnesses.opencode import (
+    build_install_script as build_opencode_install_script,
+)
 
 
 # ── Mock Rubrics ──────────────────────────────────────────────────────
@@ -187,6 +193,27 @@ def test_taskset_repr():
     assert "3" in repr(ts)
 
 
+def test_composable_mini_swe_agent_unversioned_package_uses_unpinned_requirement():
+    setup = build_mini_swe_agent_install_script(package="  mini-swe-agent  ")
+
+    assert (
+        "vf_python_install --target /opt/mini-swe-agent/prefix/site-packages mini-swe-agent"
+        in setup
+    )
+    assert "mini-swe-agent==mini-swe-agent" not in setup
+
+
+def test_composable_opencode_unversioned_release_uses_latest_download_url():
+    setup = build_opencode_install_script(
+        release="  PrimeIntellect-ai/opencode  ",
+        install_ripgrep=False,
+    )
+
+    assert "OPENCODE_RELEASE_REPO=PrimeIntellect-ai/opencode" in setup
+    assert "OPENCODE_RELEASE_PATH=releases/latest/download" in setup
+    assert "releases/download/vPrimeIntellect-ai/opencode" not in setup
+
+
 @pytest.mark.asyncio
 async def test_composable_env_exports_task_workdir():
     taskset = MockSandboxTaskSet(dataset=_make_dataset(), name="test")

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/tests/test_multiturn_env.py

@@ -189,7 +189,7 @@ class TestMultiTurnEnv:
 
             async def add_model_response(self, state, prompt_messages, response):  # type: ignore[override]
                 await super().add_model_response(state, prompt_messages, response)
-                await asyncio.sleep(0.05)
+                await asyncio.sleep(1)
 
         env = SlowMultiTurnEnv(
             client=mock_client,
@@ -197,7 +197,7 @@ class TestMultiTurnEnv:
             dataset=sample_chat_dataset,
             parser=Parser(),
             rubric=Rubric(),
-            timeout_seconds=0.01,
+            timeout_seconds=0.2,
         )
         mock_client.set_default_response("Still going")
 

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/tests/test_openai_responses_client.py

@@ -8,6 +8,7 @@ from verifiers.clients.openai_responses_client import (
     OPENAI_RESPONSES_OUTPUT_FIELD,
     OpenAIResponsesClient,
 )
+from verifiers.errors import EmptyModelResponseError
 from verifiers.types import (
     AssistantMessage,
     ClientConfig,
@@ -133,6 +134,24 @@ async def test_get_native_response_normalizes_sampling_args_and_tools():
     ]
 
 
+@pytest.mark.asyncio
+async def test_raise_from_native_response_rejects_reasoning_only_response():
+    native_response = SimpleNamespace(
+        output=[
+            {
+                "type": "reasoning",
+                "id": "rs_1",
+                "summary": [{"type": "summary_text", "text": "thinking"}],
+                "status": "completed",
+            }
+        ]
+    )
+    client = OpenAIResponsesClient(object())
+
+    with pytest.raises(EmptyModelResponseError, match="reasoning but no content"):
+        await client.raise_from_native_response(native_response)
+
+
 @pytest.mark.asyncio
 async def test_to_native_tool_omits_strict_when_unset():
     client = OpenAIResponsesClient(object())

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/tests/test_renderer_client.py

@@ -355,6 +355,14 @@ async def test_renderer_client_rejects_empty_dict_native_response():
         await client.raise_from_native_response({})
 
 
+@pytest.mark.asyncio
+async def test_renderer_client_rejects_reasoning_only_native_response():
+    client = object.__new__(RendererClient)
+
+    with pytest.raises(EmptyModelResponseError, match="reasoning but no content"):
+        await client.raise_from_native_response({"reasoning_content": "hidden chain"})
+
+
 @pytest.mark.asyncio
 async def test_from_native_response_uses_request_id_and_token_lengths():
     """vLLM's /inference/v1/generate returns ``request_id`` (not ``id``) and

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/tests/test_v1_harbor_cli.py

@@ -279,11 +279,47 @@ def test_opencode_config_owns_opencode_harness_fields() -> None:
     assert harness.config.max_turns == 2
     assert "apt-get -o Acquire::Retries=3 update" in setup
     assert "apt-get -o Acquire::Retries=3 install" in setup
+    assert "OPENCODE_RELEASE_REPO=PrimeIntellect-ai/opencode" in setup
+    assert "OPENCODE_RELEASE_PATH=releases/download/v1.1.63-rl2" in setup
     assert "/workspace" in cast(str, command[2])
     assert '"webfetch": false' in cast(str, mcp_setup)
     assert "/opencode/system.txt" in cast(dict[str, object], program["files"])
 
 
+@pytest.mark.parametrize(
+    "release", ["PrimeIntellect-ai/opencode@latest", "  PrimeIntellect-ai/opencode  "]
+)
+def test_opencode_latest_release_uses_latest_download_url(release: str) -> None:
+    harness = OpenCode(
+        config=OpenCodeConfig(
+            program=OpenCodeProgramConfig(
+                release=release,
+                install_ripgrep=False,
+            )
+        )
+    )
+    program = cast(dict[str, object], harness.config.program.data())
+    setup = cast(str, program["setup"])
+
+    assert "OPENCODE_RELEASE_REPO=PrimeIntellect-ai/opencode" in setup
+    assert "OPENCODE_RELEASE_PATH=releases/latest/download" in setup
+
+
+def test_opencode_custom_release_uses_versioned_spec() -> None:
+    harness = OpenCode(
+        config=OpenCodeConfig(
+            program=OpenCodeProgramConfig(
+                release="Example/open-code@v2.0.0",
+            )
+        )
+    )
+    program = cast(dict[str, object], harness.config.program.data())
+    setup = cast(str, program["setup"])
+
+    assert "OPENCODE_RELEASE_REPO=Example/open-code" in setup
+    assert "OPENCODE_RELEASE_PATH=releases/download/v2.0.0" in setup
+
+
 @pytest.mark.parametrize(
     ("harness_cls", "config_cls", "program_cls"),
     [
@@ -357,7 +393,7 @@ def test_pi_harness_writes_intercepted_model_and_mcp_config() -> None:
     assert "apt-get -o Acquire::Retries=3 update" in setup
     assert "apt-get -o Acquire::Retries=3 install" in setup
     assert harness.config.program.package == PI_DEFAULT_PACKAGE
-    assert PI_DEFAULT_PACKAGE == "@earendil-works/pi-coding-agent"
+    assert PI_DEFAULT_PACKAGE == "@earendil-works/pi-coding-agent@latest"
     assert f"npm install -g --ignore-scripts {PI_DEFAULT_PACKAGE}" in setup
     assert "mariozechner" not in setup
     assert '"baseUrl": "${OPENAI_BASE_URL}"' in mcp_setup
@@ -368,6 +404,18 @@ def test_pi_harness_writes_intercepted_model_and_mcp_config() -> None:
     assert f'"command": "{SANDBOX_PYTHON}"' in mcp_setup
 
 
+def test_pi_harness_preserves_scoped_npm_package_versions() -> None:
+    harness = Pi(
+        config=PiConfig(
+            program=PiProgramConfig(package="@anthropic-ai/claude-code@1.2.3")
+        )
+    )
+    program = cast(dict[str, object], harness.config.program.data())
+    setup = cast(str, program["setup"])
+
+    assert "npm install -g --ignore-scripts @anthropic-ai/claude-code@1.2.3" in setup
+
+
 def test_terminus_2_harness_builds_sandbox_program() -> None:
     harness = Terminus2(
         config=Terminus2Config(

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/tests/test_v1_mini_swe_agent.py

@@ -80,11 +80,38 @@ def test_mini_swe_agent_builds_sandbox_program():
     assert "model.model_kwargs.parallel_tool_calls=true" in script
     assert "apt-get -o Acquire::Retries=3 update" in cast(str, program["setup"])
     assert "apt-get -o Acquire::Retries=3 install" in cast(str, program["setup"])
+    assert "mini-swe-agent==2.2.8" in cast(str, program["setup"])
     assert "/mini-swe-agent/prompt.txt" in cast(dict[str, object], program["files"])
     assert "/mini-swe-agent/system.txt" in cast(dict[str, object], program["files"])
     assert "mini_swe_agent_log" in cast(dict[str, object], program["artifacts"])
 
 
+@pytest.mark.parametrize("package", ["mini-swe-agent@latest", "  mini-swe-agent  "])
+def test_mini_swe_agent_latest_package_uses_unpinned_pip_requirement(package: str):
+    harness = MiniSWEAgent(
+        config=MiniSWEAgentConfig(program=MiniSWEAgentProgramConfig(package=package))
+    )
+    program = cast(dict[str, Any], harness.config.program.data())
+    setup = cast(str, program["setup"])
+
+    assert (
+        "vf_python_install --target /opt/mini-swe-agent/prefix/site-packages mini-swe-agent"
+        in setup
+    )
+
+
+def test_mini_swe_agent_pinned_package_uses_pip_requirement():
+    harness = MiniSWEAgent(
+        config=MiniSWEAgentConfig(
+            program=MiniSWEAgentProgramConfig(package="mini-swe-agent@2.2.7")
+        )
+    )
+    program = cast(dict[str, Any], harness.config.program.data())
+    setup = cast(str, program["setup"])
+
+    assert "mini-swe-agent==2.2.7" in setup
+
+
 def test_mini_swe_agent_composes_with_harbor_taskset(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ):

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/verifiers/clients/anthropic_messages_client.py

@@ -29,7 +29,7 @@ from anthropic.types import (
 )
 
 from verifiers.clients.client import Client
-from verifiers.errors import OverlongPromptError
+from verifiers.errors import EmptyModelResponseError, OverlongPromptError
 from verifiers.types import (
     AssistantMessage,
     ClientConfig,
@@ -378,7 +378,29 @@ class AnthropicMessagesClient(
             )
 
     async def raise_from_native_response(self, response: AnthropicMessage) -> None:
-        pass
+        if response is None:
+            raise EmptyModelResponseError("Model returned no response")
+
+        has_text = False
+        has_tool_call = False
+        has_reasoning = False
+        for content_block in getattr(response, "content", []) or []:
+            block_type = getattr(content_block, "type", None)
+            if block_type == "text" and getattr(content_block, "text", None):
+                has_text = True
+            elif block_type == "tool_use":
+                has_tool_call = True
+            elif block_type in {"thinking", "redacted_thinking"}:
+                has_reasoning = True
+
+        if not (has_text or has_tool_call):
+            if has_reasoning:
+                raise EmptyModelResponseError(
+                    "Model returned reasoning but no content and did not call any tools"
+                )
+            raise EmptyModelResponseError(
+                "Model returned no content and did not call any tools"
+            )
 
     async def from_native_response(self, response: AnthropicMessage) -> Response:
         def parse_content(

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/verifiers/clients/openai_chat_completions_client.py

@@ -43,6 +43,7 @@ from verifiers.types import (
     ClientConfig,
     FinishReason,
     Message,
+    MessageContent,
     Messages,
     Response,
     ResponseMessage,
@@ -119,6 +120,14 @@ def content_to_text(content: Any) -> str:
     return ""
 
 
+def parse_refusal_content(message: Any) -> str | None:
+    if isinstance(message, Mapping):
+        refusal = message.get("refusal")
+    else:
+        refusal = getattr(message, "refusal", None)
+    return refusal if isinstance(refusal, str) and refusal else None
+
+
 DEFAULT_REASONING_FIELDS = [
     "reasoning",  # vLLM, Together AI, OpenRouter
     "reasoning_content",  # DeepSeek, Qwen/DashScope, SGLang, Fireworks AI, Kimi/Moonshot
@@ -331,15 +340,29 @@ class OpenAIChatCompletionsClient(
                 f"Model returned {len(response.choices)} choices, expected 1"
             )
         message = response.choices[0].message
-        has_content = bool(content_to_text(getattr(message, "content", None)))
+        has_content = bool(
+            content_to_text(getattr(message, "content", None))
+            or parse_refusal_content(message)
+        )
         has_tool_calls = bool(getattr(message, "tool_calls", None))
         has_reasoning = bool(parse_reasoning_content(message))
-        if not (has_content or has_tool_calls or has_reasoning):
+        if not (has_content or has_tool_calls):
+            if has_reasoning:
+                raise EmptyModelResponseError(
+                    "Model returned reasoning but no content and did not call any tools"
+                )
             raise EmptyModelResponseError(
-                "Model returned no content, reasoning, and did not call any tools"
+                "Model returned no content and did not call any tools"
             )
 
     async def from_native_response(self, response: OpenAIChatResponse) -> Response:
+        def parse_content(response: OpenAIChatResponse) -> MessageContent | None:
+            message = response.choices[0].message
+            content = message.content
+            if content_to_text(content):
+                return content
+            return parse_refusal_content(message)
+
         def parse_single_tool_call(tool_call: Any) -> ToolCall | None:
             if isinstance(tool_call, ChatCompletionMessageFunctionToolCall):
                 return ToolCall(
@@ -511,7 +534,7 @@ class OpenAIChatCompletionsClient(
             model=model,
             usage=parse_usage(response),
             message=ResponseMessage(
-                content=response.choices[0].message.content,
+                content=parse_content(response),
                 reasoning_content=parse_reasoning_content(response.choices[0].message),
                 finish_reason=parse_finish_reason(response),
                 is_truncated=parse_is_truncated(response),

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/verifiers/clients/openai_responses_client.py

@@ -293,9 +293,13 @@ class OpenAIResponsesClient(
                     ):
                         has_text = True
 
-        if not (has_text or has_tool_call or has_reasoning):
+        if not (has_text or has_tool_call):
+            if has_reasoning:
+                raise EmptyModelResponseError(
+                    "Model returned reasoning but no content and did not call any tools"
+                )
             raise EmptyModelResponseError(
-                "Model returned no content, reasoning, and did not call any tools"
+                "Model returned no content and did not call any tools"
             )
 
     async def from_native_response(

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/verifiers/clients/renderer_client.py

@@ -641,9 +641,13 @@ class RendererClient(
         # model having tried to call a tool, so we don't filter by status here.
         has_tool_calls = bool(response.get("tool_calls"))
         has_reasoning = bool(response.get("reasoning_content"))
-        if not (has_content or has_tool_calls or has_reasoning):
+        if not (has_content or has_tool_calls):
+            if has_reasoning:
+                raise EmptyModelResponseError(
+                    "Model returned reasoning but no content and did not call any tools"
+                )
             raise EmptyModelResponseError(
-                "Model returned no content, reasoning, and did not call any tools"
+                "Model returned no content and did not call any tools"
             )
 
     async def from_native_response(self, response: dict[str, Any]) -> Response:

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/verifiers/envs/experimental/composable/harnesses/__init__.py

@@ -9,7 +9,7 @@ from verifiers.envs.experimental.composable.harnesses.rlm import (
 )
 from verifiers.envs.experimental.composable.harnesses.opencode import (
     DEFAULT_DISABLED_TOOLS,
-    DEFAULT_RELEASE_SHA256,
+    DEFAULT_RELEASE,
     DEFAULT_SYSTEM_PROMPT,
     OPENCODE_INSTALL_SCRIPT,
     build_install_script as build_opencode_install_script,
@@ -39,7 +39,7 @@ __all__ = [
     "build_opencode_run_command",
     "OPENCODE_INSTALL_SCRIPT",
     "DEFAULT_DISABLED_TOOLS",
-    "DEFAULT_RELEASE_SHA256",
+    "DEFAULT_RELEASE",
     "DEFAULT_SYSTEM_PROMPT",
     "mini_swe_agent_harness",
     "build_mini_swe_agent_install_script",

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/verifiers/envs/experimental/composable/harnesses/mini_swe_agent.py

@@ -4,6 +4,7 @@ from pathlib import PurePosixPath
 import shlex
 
 from harnesses.mini_swe_agent import (
+    MINI_SWE_AGENT_DEFAULT_PACKAGE,
     build_mini_swe_agent_install_script as build_packaged_mini_swe_agent_install_script,
 )
 
@@ -13,11 +14,7 @@ DEFAULT_SITE_PACKAGES_DIR = f"{DEFAULT_PREFIX_DIR}/site-packages"
 DEFAULT_MINI_BINARY = f"{DEFAULT_PREFIX_DIR}/bin/mini"
 MINI_SWE_AGENT_CLI_PACKAGE = "mini-swe-agent"
 MINI_SWE_AGENT_CLI_VERSION = "2.2.8"
-MINI_SWE_AGENT_CLI_SHA256 = (
-    "694df4de1337e665e3cd82e99f93374f573bf52b8e7c362ac5d8045ad9f7c37c"
-)
-DEFAULT_PACKAGE_VERSION = MINI_SWE_AGENT_CLI_VERSION
-DEFAULT_PACKAGE_SHA256 = MINI_SWE_AGENT_CLI_SHA256
+DEFAULT_PACKAGE = MINI_SWE_AGENT_DEFAULT_PACKAGE
 DEFAULT_INSTRUCTION_PATH = "/mini-swe-agent/prompt.txt"
 DEFAULT_SYSTEM_PROMPT_PATH = "/mini-swe-agent/system.txt"
 DEFAULT_LOG_DIR = "/logs/agent"
@@ -30,14 +27,12 @@ DEFAULT_ENVIRONMENT_TIMEOUT = 120
 
 
 def build_mini_swe_agent_install_script(
-    package_version: str = DEFAULT_PACKAGE_VERSION,
-    package_sha256: str = DEFAULT_PACKAGE_SHA256,
+    package: str = DEFAULT_PACKAGE,
     prefix_dir: str = DEFAULT_PREFIX_DIR,
 ) -> str:
     """Build the shell script that installs mini-SWE-agent."""
     return build_packaged_mini_swe_agent_install_script(
-        package_version=package_version,
-        package_sha256=package_sha256,
+        package=package,
         prefix_dir=prefix_dir,
     )
 
@@ -127,7 +122,6 @@ MINI_SWE_AGENT_CONFIG = {
     "install_script": MINI_SWE_AGENT_INSTALL_SCRIPT,
     "cli_package": MINI_SWE_AGENT_CLI_PACKAGE,
     "cli_version": MINI_SWE_AGENT_CLI_VERSION,
-    "cli_sha256": MINI_SWE_AGENT_CLI_SHA256,
 }
 
 
@@ -139,8 +133,7 @@ def mini_swe_agent_harness(
     system_prompt_path: str = DEFAULT_SYSTEM_PROMPT_PATH,
     log_path: str = DEFAULT_LOG_PATH,
     trajectory_path: str = DEFAULT_TRAJECTORY_PATH,
-    package_version: str = DEFAULT_PACKAGE_VERSION,
-    package_sha256: str = DEFAULT_PACKAGE_SHA256,
+    package: str = DEFAULT_PACKAGE,
     config_spec: str = DEFAULT_CONFIG_SPEC,
     model_class: str = DEFAULT_MODEL_CLASS,
     environment_timeout: int = DEFAULT_ENVIRONMENT_TIMEOUT,
@@ -160,8 +153,7 @@ def mini_swe_agent_harness(
     # into mini's agent.system_template at runtime.
     return Harness(
         install_script=build_mini_swe_agent_install_script(
-            package_version=package_version,
-            package_sha256=package_sha256,
+            package=package,
         ),
         run_command=build_mini_swe_agent_run_command(
             agent_workdir=agent_workdir,

{verifiers-0.1.15.dev154 → verifiers-0.1.15.dev156}/verifiers/envs/experimental/composable/harnesses/opencode.py

@@ -13,13 +13,11 @@ import json
 import shlex
 from pathlib import Path, PurePosixPath
 
+from harnesses.utils import split_versioned_agent_spec
+
 # ── Defaults ─────────────────────────────────────────────────────────────
 
-DEFAULT_RELEASE_REPO = "PrimeIntellect-ai/opencode"
-DEFAULT_RELEASE_VERSION = "1.1.63-rl2"
-DEFAULT_RELEASE_SHA256 = (
-    "47f4102796da50769e27d2c9ea6a9cf7941f76898390cb497278cab39c4b6ed4"
-)
+DEFAULT_RELEASE = "PrimeIntellect-ai/opencode@1.1.63-rl2"
 DEFAULT_SYSTEM_PROMPT = (Path(__file__).parent / "prompt.txt").read_text()
 
 DEFAULT_DISABLED_TOOLS = [
@@ -49,9 +47,7 @@ DEFAULT_DISABLED_TOOLS = [
 
 
 def build_install_script(
-    release_repo: str = DEFAULT_RELEASE_REPO,
-    release_version: str = DEFAULT_RELEASE_VERSION,
-    release_sha256: str = DEFAULT_RELEASE_SHA256,
+    release: str = DEFAULT_RELEASE,
     install_ripgrep: bool = True,
 ) -> str:
     """Build the shell script that installs OpenCode in a sandbox."""
@@ -60,7 +56,15 @@ def build_install_script(
         if install_ripgrep
         else ""
     )
-    sha256_check = f'echo "{release_sha256}  /tmp/opencode.tar.gz" | sha256sum -c -'
+    release_repo, release_version = split_versioned_agent_spec(release)
+    release_path = "releases/latest/download"
+    if release_version and release_version != "latest":
+        release_tag = (
+            release_version
+            if release_version.startswith("v")
+            else f"v{release_version}"
+        )
+        release_path = f"releases/download/{release_tag}"
     # Acquire::Retries=3 mitigates transient archive.ubuntu.com CDN sync mismatches
     # (e.g. "File has unexpected size ... Mirror sync in progress?"). See launchpad
     # bug #1876035. apt's default retries is 0, so one bad fetch fails the rollout.
@@ -69,8 +73,8 @@ set -e
 apt-get -o Acquire::Retries=3 update -qq && apt-get -o Acquire::Retries=3 install -y -qq curl tar > /dev/null 2>&1
 {rg_install}
 
-OPENCODE_RELEASE_REPO="{release_repo}"
-OPENCODE_RELEASE_VERSION="{release_version}"
+OPENCODE_RELEASE_REPO={shlex.quote(release_repo)}
+OPENCODE_RELEASE_PATH={shlex.quote(release_path)}
 
 case "$(uname -m)" in
   x86_64) OPENCODE_ARCH=x64 ;;
@@ -79,15 +83,13 @@ case "$(uname -m)" in
 esac
 
 OPENCODE_ASSET="opencode-linux-$OPENCODE_ARCH.tar.gz"
-OPENCODE_RELEASE_TAG="${{OPENCODE_RELEASE_VERSION#v}}"
-OPENCODE_RELEASE_URL="https://github.com/$OPENCODE_RELEASE_REPO/releases/download/v$OPENCODE_RELEASE_TAG/$OPENCODE_ASSET"
+OPENCODE_RELEASE_URL="https://github.com/$OPENCODE_RELEASE_REPO/$OPENCODE_RELEASE_PATH/$OPENCODE_ASSET"
 
 mkdir -p "$HOME/.opencode/bin"
 if [ -x "$HOME/.opencode/bin/opencode" ]; then
   echo "OpenCode already installed, skipping download"
 else
   curl -fsSL "$OPENCODE_RELEASE_URL" -o /tmp/opencode.tar.gz
-  {sha256_check}
   tar -xzf /tmp/opencode.tar.gz -C /tmp
   install -m 755 /tmp/opencode "$HOME/.opencode/bin/opencode"
   rm -f /tmp/opencode.tar.gz /tmp/opencode
@@ -246,9 +248,7 @@ def opencode_harness(
     agent_workdir: str = "/app",
     allow_git: bool = False,
     disable_compaction: bool = True,
-    release_repo: str = DEFAULT_RELEASE_REPO,
-    release_version: str = DEFAULT_RELEASE_VERSION,
-    release_sha256: str = DEFAULT_RELEASE_SHA256,
+    release: str = DEFAULT_RELEASE,
     instruction_path: str = "/opencode/prompt.txt",
     system_prompt_path: str = "/opencode/system.txt",
     log_path: str = "/opencode/logs.txt",
@@ -276,9 +276,7 @@ def opencode_harness(
 
     return Harness(
         install_script=build_install_script(
-            release_repo=release_repo,
-            release_version=release_version,
-            release_sha256=release_sha256,
+            release=release,
         ),
         run_command=build_opencode_run_command(
             agent_workdir=agent_workdir,