inspect-swe 0.2.3__tar.gz → 0.2.6__tar.gz

{inspect_swe-0.2.3 → inspect_swe-0.2.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: inspect_swe
-Version: 0.2.3
+Version: 0.2.6
 Summary: Software engineering agents for Inspect AI.
 Project-URL: Documentation, https://meridianlabs-ai.github.io/inspect_swe/
 Project-URL: Source Code, https://github.com/meridianlabs-ai/inspect_swe
@@ -10,10 +10,11 @@ License: MIT License
 License-File: LICENSE
 Requires-Python: >=3.10
 Requires-Dist: httpx
-Requires-Dist: inspect-ai>=0.3.125
+Requires-Dist: inspect-ai>=0.3.126
 Requires-Dist: nest-asyncio
 Requires-Dist: platformdirs
 Requires-Dist: pydantic>=2.11.4
+Requires-Dist: pyyaml
 Requires-Dist: sniffio
 Requires-Dist: typing-extensions>=4.9.0
 Provides-Extra: dev
@@ -23,6 +24,7 @@ Requires-Dist: openai; extra == 'dev'
 Requires-Dist: pytest; extra == 'dev'
 Requires-Dist: pytest-dotenv; extra == 'dev'
 Requires-Dist: ruff; extra == 'dev'
+Requires-Dist: types-pyyaml; extra == 'dev'
 Provides-Extra: doc
 Requires-Dist: quarto-cli==1.7.31; extra == 'doc'
 Description-Content-Type: text/markdown

{inspect_swe-0.2.3 → inspect_swe-0.2.6}/pyproject.toml

@@ -12,10 +12,11 @@ requires-python = ">=3.10"
 license = { text = "MIT License" }
 dependencies = [
     "httpx",
-    "inspect_ai>=0.3.125",
+    "inspect_ai>=0.3.126",
     "nest_asyncio",
     "platformdirs",
     "pydantic>=2.11.4",
+    "pyyaml",
     "sniffio",
     "typing_extensions>=4.9.0",
 ]
@@ -30,7 +31,15 @@ inspect_swe = "inspect_swe._registry"
 
 
 [project.optional-dependencies]
-dev = ["ruff", "mypy", "pytest", "anthropic", "openai", "pytest-dotenv"]
+dev = [
+    "ruff",
+    "mypy",
+    "pytest",
+    "anthropic",
+    "openai",
+    "pytest-dotenv",
+    "types-PyYAML",
+]
 doc = ["quarto-cli==1.7.31"]
 
 [tool.hatch.build]

{inspect_swe-0.2.3 → inspect_swe-0.2.6}/src/inspect_swe/__init__.py

@@ -1,5 +1,5 @@
 from ._claude_code.claude_code import claude_code
-from ._claude_code.install.download import download_claude_code
+from ._tools.download import download_agent_binary
 from ._util.sandbox import SandboxPlatform
 
 try:
@@ -8,4 +8,9 @@ except ImportError:
     __version__ = "unknown"
 
 
-__all__ = ["claude_code", "download_claude_code", "SandboxPlatform", "__version__"]
+__all__ = [
+    "claude_code",
+    "download_agent_binary",
+    "SandboxPlatform",
+    "__version__",
+]

inspect_swe-0.2.6/src/inspect_swe/_claude_code/claude_code.py

@@ -0,0 +1,207 @@
+import uuid
+from textwrap import dedent
+from typing import Any, Literal, Sequence
+
+from inspect_ai.agent import (
+    Agent,
+    AgentAttempts,
+    AgentState,
+    agent,
+    agent_with,
+    sandbox_agent_bridge,
+)
+from inspect_ai.model import ChatMessageSystem, ChatMessageUser
+from inspect_ai.scorer import score
+from inspect_ai.tool import MCPServerConfig
+from inspect_ai.util import sandbox as sandbox_env
+from pydantic_core import to_json
+
+from .._util._async import is_callable_coroutine
+from .install.install import ensure_claude_code_installed
+
+
+@agent
+def claude_code(
+    name: str = "Claude Code",
+    description: str = dedent("""
+       Autonomous coding agent capable of writing, testing, debugging,
+       and iterating on code across multiple languages.
+    """),
+    system_prompt: str | None = None,
+    mcp_servers: Sequence[MCPServerConfig] | None = None,
+    attempts: int | AgentAttempts = 1,
+    model: str | None = None,
+    small_model: str | None = None,
+    env: dict[str, str] | None = None,
+    version: Literal["auto", "sandbox", "stable", "latest"] | str = "auto",
+    user: str | None = None,
+    sandbox: str | None = None,
+) -> Agent:
+    """Claude Code agent.
+
+    Agent that uses [Claude Code](https://docs.anthropic.com/en/docs/claude-code/overview) running in a sandbox.
+
+    The agent can either use a version of Claude Code installed in the sandbox, or can download a version and install it in the sandbox (see docs on `version` option below for details).
+
+    Use the `attempts` option to enable additional submissions if the initial
+    submission(s) are incorrect (by default, no additional attempts are permitted).
+
+    Args:
+        name: Agent name (used in multi-agent systems with `as_tool()` and `handoff()`)
+        description: Agent description (used in multi-agent systems with `as_tool()` and `handoff()`)
+        system_prompt: Additional system prompt to append to default system prompt.
+        mcp_servers: MCP servers to make available to the agent.
+        attempts: Configure agent to make multiple attempts.
+        model: Model name to use for Opus and Sonnet calls (defaults to main model for task).
+        small_model: Model to use for Haiku calls (defaults to main model for task).
+        env: Environment variables to set for claude code.
+        version: Version of claude code to use. One of:
+            - "auto": Use any available version of claude code in the sandbox, otherwise download the current stable version.
+            - "sandbox": Use the version of claude code in the sandbox (raises `RuntimeError` if claude is not available in the sandbox)
+            - "stable": Download and use the current stable version of claude code.
+            - "latest": Download and use the very latest version of claude code.
+            - "x.x.x": Download and use a specific version of claude code.
+        user: User to execute claude code with.
+        sandbox: Optional sandbox environment name.
+    """
+    # resolve models
+    model = f"inspect/{model}" if model is not None else "inspect"
+    small_model = f"inspect/{small_model}" if small_model is not None else "inspect"
+
+    # resolve attempts
+    attempts = AgentAttempts(attempts) if isinstance(attempts, int) else attempts
+
+    async def execute(state: AgentState) -> AgentState:
+        async with sandbox_agent_bridge(state) as bridge:
+            # ensure claude is installed and get binary location
+            claude_binary = await ensure_claude_code_installed(
+                version, user, sandbox_env(sandbox)
+            )
+
+            # allocate session_id
+            session_id = str(uuid.uuid4())
+
+            # base options
+            cmd = [
+                "--print",  # run without interactions
+                "--dangerously-skip-permissions",
+                "--model",
+                model,
+            ]
+
+            # system prompt
+            system_messages = [
+                m.text for m in state.messages if isinstance(m, ChatMessageSystem)
+            ]
+            if system_prompt is not None:
+                system_messages.append(system_prompt)
+            if system_messages:
+                cmd.extend(["--append-system-prompt", "\n\n".join(system_messages)])
+
+            # mcp servers
+            if mcp_servers:
+                cmd.extend(mcp_server_args(mcp_servers))
+
+            # user prompt
+            prompt = "\n\n".join(
+                [m.text for m in state.messages if isinstance(m, ChatMessageUser)]
+            )
+
+            # resolve sandbox
+            sbox = sandbox_env(sandbox)
+
+            # execute the agent
+            agent_prompt = prompt
+            attempt_count = 0
+            while True:
+                # either starting a new session or resuming one
+                id_param = "--session-id" if attempt_count == 0 else "--resume"
+                agent_cmd = (
+                    [claude_binary, id_param, session_id] + cmd + ["--", agent_prompt]
+                )
+
+                # run agent
+                result = await sbox.exec(
+                    cmd=agent_cmd,
+                    env={
+                        "ANTHROPIC_BASE_URL": f"http://localhost:{bridge.port}",
+                        "ANTHROPIC_API_KEY": "sk-ant-api03-DOq5tyLPrk9M4hPE",
+                        "ANTHROPIC_MODEL": model,
+                        "ANTHROPIC_DEFAULT_OPUS_MODEL": model,
+                        "ANTHROPIC_DEFAULT_SONNET_MODEL": model,
+                        "CLAUDE_CODE_SUBAGENT_MODEL": model,
+                        "ANTHROPIC_DEFAULT_HAIKU_MODEL": small_model,
+                        "ANTHROPIC_SMALL_FAST_MODEL": small_model,
+                        "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
+                        "IS_SANDBOX": "1",
+                    }
+                    | (env or {}),
+                    user=user,
+                )
+
+                # raise for error
+                if not result.success:
+                    f"Error executing claude code agent: {result.stdout}\n{result.stderr}"
+
+                # exit if we are at max_attempts
+                attempt_count += 1
+                if attempt_count >= attempts.attempts:
+                    break
+
+                # score this attempt
+                answer_scores = await score(state)
+
+                # break if we score 'correct'
+                if attempts.score_value(answer_scores[0].value) == 1.0:
+                    break
+
+                # otherwise update prompt with incorrect message and continue
+                else:
+                    if callable(attempts.incorrect_message):
+                        if not is_callable_coroutine(attempts.incorrect_message):
+                            raise ValueError(
+                                "The incorrect_message function must be async."
+                            )
+                        agent_prompt = await attempts.incorrect_message(
+                            state, answer_scores
+                        )
+                    else:
+                        agent_prompt = attempts.incorrect_message
+
+        return bridge.state
+
+    # return agent with specified name and descritpion
+    return agent_with(execute, name=name, description=description)
+
+
+def mcp_server_args(mcp_servers: Sequence[MCPServerConfig]) -> list[str]:
+    # build servers and allowed tools
+    mcp_servers_json: dict[str, dict[str, Any]] = {}
+    allowed_tools: list[str] = []
+    for mcp_server in mcp_servers:
+        mcp_servers_json[mcp_server.name] = mcp_server.model_dump(
+            exclude={"name", "tools"}, exclude_none=True
+        )
+        if mcp_server.tools == "all":
+            allowed_tools.append(f"mcp__{mcp_server.name}_*")
+        elif isinstance(mcp_server.tools, list):
+            allowed_tools.extend(
+                [f"mcp__{mcp_server.name}__{tool}" for tool in mcp_server.tools]
+            )
+        else:
+            raise ValueError(
+                f"Unexpected value for mcp server tools: {mcp_server.tools}"
+            )
+
+    # map to cli args
+    cmds: list[str] = []
+    if len(mcp_servers_json) > 0:
+        cmds.append("--mcp-config")
+        cmds.append(
+            to_json({"mcpServers": mcp_servers_json}, exclude_none=True).decode()
+        )
+    if len(allowed_tools):
+        cmds.append("--allowed-tools")
+        cmds.append(",".join(allowed_tools))
+
+    return cmds

{inspect_swe-0.2.3 → inspect_swe-0.2.6}/src/inspect_swe/_claude_code/install/download.py

@@ -3,7 +3,6 @@ from typing import Literal
 
 from pydantic import BaseModel
 
-from ..._util._async import run_coroutine
 from ..._util.checksum import verify_checksum
 from ..._util.download import download_file, download_text_file
 from ..._util.sandbox import SandboxPlatform
@@ -14,22 +13,6 @@ from .cache import (
 )
 
 
-def download_claude_code(
-    version: Literal["stable", "latest"] | str, platform: SandboxPlatform
-) -> None:
-    """Download Claude Code.
-
-    Download a version of Claude Code. This version will be added to the cache of downloaded versions (which retains the 5 most recently downloaded versions).
-
-    Use this if you need to ensure that a specific version of Claude Code is downloaded in advance (e.g. if you are going to run your evaluations offline). After downloading, explicit requests for the downloaded version (e.g. `claude_code(version="1.0.98")`) will not require network access.
-
-    Args:
-        version: Version to download ("stable", "latest", or an explicit version number).
-        platform: Target platform ("linux-x64", "linux-arm64", "linux-x64-musl", or "linux-arm64-musl")
-    """
-    run_coroutine(download_claude_code_async(version, platform))
-
-
 async def download_claude_code_async(
     version: Literal["stable", "latest"] | str, platform: SandboxPlatform
 ) -> bytes:

{inspect_swe-0.2.3 → inspect_swe-0.2.6}/src/inspect_swe/_claude_code/install/install.py

@@ -45,6 +45,8 @@ async def ensure_claude_code_installed(
             )
             if claude_binary_bytes is not None:
                 trace(f"Used claude code binary from cache: {version} ({platform})")
+        else:
+            claude_binary_bytes = None
 
         # download the binary
         if claude_binary_bytes is None:

inspect_swe-0.2.6/src/inspect_swe/_tools/download.py

@@ -0,0 +1,27 @@
+from typing import Literal
+
+from .._claude_code.install.download import download_claude_code_async
+from .._util._async import run_coroutine
+from .._util.sandbox import SandboxPlatform
+
+
+def download_agent_binary(
+    binary: Literal["claude_code"],
+    version: Literal["stable", "latest"] | str,
+    platform: SandboxPlatform,
+) -> None:
+    """Download agent binary.
+
+    Download an agent binary. This version will be added to the cache of downloaded versions (which retains the 5 most recently downloaded versions).
+
+    Use this if you need to ensure that a specific version of an agent binary is downloaded in advance (e.g. if you are going to run your evaluations offline). After downloading, explicit requests for the downloaded version (e.g. `claude_code(version="1.0.98")`) will not require network access.
+
+    Args:
+        binary: Type of binary to download (currently only "claude_code")
+        version: Version to download ("stable", "latest", or an explicit version number).
+        platform: Target platform ("linux-x64", "linux-arm64", "linux-x64-musl", or "linux-arm64-musl")
+    """
+    if binary == "claude_code":
+        run_coroutine(download_claude_code_async(version, platform))
+    else:
+        raise ValueError(f"Unsuported agent binary type: {binary}")

{inspect_swe-0.2.3 → inspect_swe-0.2.6}/src/inspect_swe/_util/_async.py

@@ -1,5 +1,6 @@
 import asyncio
-from typing import Coroutine, Literal, TypeVar, cast
+import inspect
+from typing import Any, Coroutine, Literal, TypeVar, cast
 
 import nest_asyncio  # type: ignore
 import sniffio
@@ -9,6 +10,14 @@ from .platform import running_in_notebook
 T = TypeVar("T")
 
 
+def is_callable_coroutine(func_or_cls: Any) -> bool:
+    if inspect.iscoroutinefunction(func_or_cls):
+        return True
+    elif callable(func_or_cls):
+        return inspect.iscoroutinefunction(func_or_cls.__call__)
+    return False
+
+
 def run_coroutine(coroutine: Coroutine[None, None, T]) -> T:
     if current_async_backend() == "trio":
         raise RuntimeError("run_coroutine cannot be used with trio")

inspect_swe-0.2.6/src/inspect_swe/_util/_yaml.py

@@ -0,0 +1,21 @@
+import re
+
+import yaml
+
+
+def read_front_matter_name(content: str) -> str | None:
+    # front-matter
+    frontmatter_match = re.match(r"^\s*---\s*\n(.*?)\n---", content, re.DOTALL)
+    if not frontmatter_match:
+        return None
+    frontmatter = frontmatter_match.group(1)
+
+    try:
+        # Parse as YAML
+        data = yaml.safe_load(frontmatter)
+        if "name" in data:
+            return str(data.get("name"))
+        else:
+            return None
+    except yaml.YAMLError:
+        return None

{inspect_swe-0.2.3 → inspect_swe-0.2.6}/src/inspect_swe/_util/sandbox.py

@@ -45,7 +45,7 @@ async def detect_sandbox_platform(sandbox: SandboxEnvironment) -> SandboxPlatfor
 
 
 def bash_command(cmd: str) -> list[str]:
-    return ["bash", "--login", "-c", cmd]
+    return ["bash", "-c", cmd]
 
 
 async def sandbox_exec(
@@ -53,7 +53,5 @@ async def sandbox_exec(
 ) -> str:
     result = await sandbox.exec(bash_command(cmd), user=user)
     if not result.success:
-        raise RuntimeError(
-            f"Error executing sandbox command {','.join(cmd)}: {result.stderr}"
-        )
+        raise RuntimeError(f"Error executing sandbox command {cmd}: {result.stderr}")
     return result.stdout.strip()

{inspect_swe-0.2.3 → inspect_swe-0.2.6}/src/inspect_swe/_version.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.2.3'
-__version_tuple__ = version_tuple = (0, 2, 3)
+__version__ = version = '0.2.6'
+__version_tuple__ = version_tuple = (0, 2, 6)
 
 __commit_id__ = commit_id = None

inspect_swe-0.2.3/src/inspect_swe/_claude_code/claude_code.py

@@ -1,87 +0,0 @@
-from typing import Literal
-
-from inspect_ai.agent import (
-    Agent,
-    AgentState,
-    agent,
-    sandbox_agent_bridge,
-)
-from inspect_ai.model import ChatMessageSystem, ChatMessageUser
-from inspect_ai.util import sandbox as sandbox_env
-
-from inspect_swe._claude_code.install.install import ensure_claude_code_installed
-
-
-@agent
-def claude_code(
-    version: Literal["auto", "sandbox", "stable", "latest"] | str = "auto",
-    user: str | None = None,
-    sandbox: str | None = None,
-) -> Agent:
-    """Claude Code agent.
-
-    Agent that uses [Claude Code](https://docs.anthropic.com/en/docs/claude-code/overview) running in a sandbox.
-
-    The agent can either use a version of Claude Code installed in the sandbox, or can download a version and install it in the sandbox (see docs on `version` option below for details).
-
-    Args:
-        version: Version of claude code to use. One of:
-            - "auto": Use any available version of claude code in the sandbox, otherwise download the current stable version.
-            - "sandbox": Use the version of claude code in the sandbox (raises `RuntimeError` if claude is not available in the sandbox)
-            - "stable": Download and use the current stable version of claude code.
-            - "latest": Download and use the very latest version of claude code.
-            - "x.x.x": Download and use a specific version of claude code.
-        user: User to execute claude code with.
-        sandbox: Optional sandbox environment name.
-    """
-
-    async def execute(state: AgentState) -> AgentState:
-        async with sandbox_agent_bridge(state) as bridge:
-            # ensure claude is installed and get binary location
-            claude_binary = await ensure_claude_code_installed(
-                version, user, sandbox_env(sandbox)
-            )
-
-            # base options
-            cmd = [
-                claude_binary,
-                "--print",  # run without interactions
-                "--dangerously-skip-permissions",
-                "--model",  # use current inspect model
-                "inspect",
-            ]
-
-            # system message
-            system_message = "\n\n".join(
-                [m.text for m in state.messages if isinstance(m, ChatMessageSystem)]
-            )
-            if system_message:
-                cmd.extend(["--append-system-prompt", system_message])
-
-            # user prompt
-            prompt = "\n\n".join(
-                [m.text for m in state.messages if isinstance(m, ChatMessageUser)]
-            )
-            cmd.append(prompt)
-
-            # execute the agent
-            result = await sandbox_env(sandbox).exec(
-                cmd=cmd,
-                env={
-                    "ANTHROPIC_BASE_URL": f"http://localhost:{bridge.port}",
-                    "ANTHROPIC_API_KEY": "sk-ant-api03-DOq5tyLPrk9M4hPE",
-                    "ANTHROPIC_SMALL_FAST_MODEL": "inspect",
-                    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
-                    "IS_SANDBOX": "1",
-                },
-                user=user,
-            )
-
-        if result.success:
-            return bridge.state
-        else:
-            raise RuntimeError(
-                f"Error executing claude code agent: {result.stdout}\n{result.stderr}"
-            )
-
-    return execute