SciDER / scider /tools /shell /run_bash_cmd /__init__.py
leonardklin's picture
Upload 355 files
1499363 verified
Raw
History Blame Contribute Delete
11.4 kB
"""BashTool — execute shell commands with safety classification.
Modeled after Claude Code's BashTool security model:
- Commands classified as read-only / write / dangerous per invocation
- Dangerous commands are denied
- Write commands constrained to workspace directory
"""
from __future__ import annotations
import os
import re
import shlex
import shutil
import subprocess
from pathlib import Path
from loguru import logger
from pydantic import BaseModel, Field
from ...base import BaseTool, ToolContext
def _detect_shell() -> str:
if shutil.which("bash"):
return "bash"
return "sh"
_SHELL = _detect_shell()
# Commands that only read — safe to run anywhere
READ_ONLY_COMMANDS = {
"ls",
"cat",
"head",
"tail",
"less",
"more",
"wc",
"file",
"stat",
"find",
"grep",
"egrep",
"fgrep",
"rg",
"ag",
"awk",
"sed", # sed without -i
"diff",
"cmp",
"sort",
"uniq",
"cut",
"tr",
"tee",
"echo",
"printf",
"date",
"whoami",
"pwd",
"env",
"printenv",
"which",
"type",
"whereis",
"man",
"help",
"du",
"df",
"free",
"top",
"ps",
"uname",
"id",
"git status",
"git log",
"git diff",
"git show",
"git branch",
"git remote",
"git tag",
"git stash list",
"python --version",
"python3 --version",
"pip list",
"pip show",
"node --version",
"npm list",
"uv --version",
"tree",
"realpath",
"readlink",
"basename",
"dirname",
"md5sum",
"sha256sum",
"sha1sum",
}
# Dangerous compound patterns — matched as substrings because the composition
# itself is the danger signal (e.g. "rm -rf /" never appears legitimately).
DANGEROUS_COMPOUND_PATTERNS = {
"rm -rf /",
"rm -rf /*",
"rm -rf ~",
"dd if=",
"chmod 777",
"chmod -R 777",
"> /dev/sda",
"curl | sh",
"curl | bash",
"wget | sh",
"wget | bash",
":(){:|:&};:", # fork bomb
}
# Dangerous command NAMES — matched only at shell-command boundaries so that
# `.format()` / `literal_eval` / `# Frame format:` etc. inside Python code
# don't trigger false positives. Previously these were substring-matched,
# which blocked harmless Python one-liners.
DANGEROUS_COMMAND_NAMES = {
"mkfs",
"shutdown",
"reboot",
"halt",
"poweroff",
"eval", # shell `eval "..."`, not Python eval()
}
# Regex matches any of the dangerous command names at a shell-command
# boundary: start-of-string OR after whitespace/; /& /| /`, AND followed by
# the same or end-of-string. `(?:^|...)` is non-capturing so only the name
# itself lands in group 1.
# Suffix allows `.` so `mkfs.ext4` / `mkfs.xfs` still match the `mkfs`
# family. Prefix deliberately does NOT include `.` — that's what keeps
# Python's `.format()` / `.literal_eval` from triggering.
_DANGEROUS_NAME_RE = re.compile(
r"(?:^|[\s;&|`])("
+ "|".join(re.escape(n) for n in DANGEROUS_COMMAND_NAMES)
+ r")(?=[\s;&|`.]|$)",
re.IGNORECASE,
)
# Dangerous environment variables that should never be set
DANGEROUS_ENV_VARS = {
"PATH",
"LD_PRELOAD",
"LD_LIBRARY_PATH",
"DYLD_INSERT_LIBRARIES",
"DYLD_LIBRARY_PATH",
"NODE_OPTIONS",
"PYTHONPATH",
"BASH_ENV",
}
def _get_first_command(command: str) -> str:
"""Extract the first command from a compound command."""
# Split on &&, ||, ;, | and take first
for sep in ["&&", "||", ";", "|"]:
if sep in command:
return command.split(sep)[0].strip()
return command.strip()
def _classify_command(command: str) -> str:
"""Classify a command as 'read-only', 'write', or 'dangerous'.
Returns one of: 'read-only', 'write', 'dangerous'
"""
cmd = command.strip()
cmd_lower = cmd.lower()
# Compound patterns (substring): the phrasing IS the signal.
for pattern in DANGEROUS_COMPOUND_PATTERNS:
if pattern in cmd_lower:
return "dangerous"
# Command names at shell-command boundaries only.
if _DANGEROUS_NAME_RE.search(cmd):
return "dangerous"
# Check for dangerous env var manipulation
for var in DANGEROUS_ENV_VARS:
if f"{var}=" in cmd and not cmd.startswith("echo"):
return "dangerous"
# Check for command substitution / eval
if "$()" in cmd or "`" in cmd:
# Could be read-only (e.g., echo $(date)) but treat as write conservatively
pass
# Extract the base command
first_cmd = _get_first_command(cmd)
try:
parts = shlex.split(first_cmd)
except ValueError:
parts = first_cmd.split()
if not parts:
return "read-only"
base = os.path.basename(parts[0])
# Check read-only commands
if base in READ_ONLY_COMMANDS:
# sed with -i is a write, not read-only
if base == "sed" and "-i" in parts:
return "write"
return "read-only"
# Check multi-word read-only (e.g., "git status")
two_word = f"{base} {parts[1]}" if len(parts) > 1 else ""
if two_word in READ_ONLY_COMMANDS:
return "read-only"
# Known write commands
if base in {
"rm",
"mv",
"cp",
"mkdir",
"rmdir",
"touch",
"chmod",
"chown",
"ln",
"install",
"git commit",
"git push",
"git checkout",
"git reset",
"git merge",
"git rebase",
}:
return "write"
# Package managers, compilers, etc. — write
if base in {
"pip",
"pip3",
"npm",
"yarn",
"pnpm",
"uv",
"apt",
"apt-get",
"brew",
"cargo",
"go",
"gcc",
"g++",
"make",
"cmake",
"python",
"python3",
"node",
"docker",
"docker-compose",
}:
return "write"
# Default: write (conservative)
return "write"
class RunBashCmdInput(BaseModel):
command: str = Field(description=f"The {_SHELL} command to execute")
cwd: str | None = Field(
default=None,
description="Working directory for the command (optional)",
)
timeout: int = Field(
default=120,
description="Timeout in seconds (default 120, max 43200 i.e. 12 hours)",
)
run_in_background: bool = Field(
default=False,
description=(
"Set to true to run this command in the background. "
"Returns a task ID immediately. Use TaskOutput to read the output later."
),
)
class RunBashCmdTool(BaseTool):
name = "Bash"
description = (
f"Execute a {_SHELL} command and return its output (stdout and stderr combined). "
f"Use this for shell operations. Prefer dedicated tools (Read, Glob, Grep) "
f"for file reading and searching."
)
input_schema = RunBashCmdInput
prompt = (
"# Bash tool usage\n"
"- Do NOT use Bash to read files (use Read), search files (use Glob), "
"or search content (use Grep). Reserve Bash for commands that require shell execution.\n"
"- Always quote file paths with spaces using double quotes.\n"
"- When issuing multiple independent commands, make separate tool calls in parallel.\n"
"- For long-running commands (training, builds, large data processing), "
"use `run_in_background=true` and check results later with TaskOutput.\n"
"- Foreground commands are capped at 10 minutes. Background commands allow up to 12 hours.\n"
"- If a command fails, diagnose why before retrying. Do not blindly re-run.\n"
"- Avoid unnecessary `sleep` commands — use `run_in_background` instead.\n"
)
def is_read_only(self, **kwargs) -> bool:
"""Bash is read-only only if the specific command is classified as read-only.
When called without a command (e.g., plan mode static check), returns False
because Bash CAN write — it depends on the specific command.
"""
command = kwargs.get("command")
if not command:
return False # Bash can write, so not read-only by default
return _classify_command(command) == "read-only"
def check_permissions(self, kwargs: dict, context: ToolContext) -> "PermissionResult":
"""Check command safety and path constraints."""
from scider.core.permissions import allow, deny, is_dangerous_path
command = kwargs.get("command", "")
classification = _classify_command(command)
# Dangerous commands are always denied
if classification == "dangerous":
return deny(f"Dangerous command blocked: {command[:100]}")
# Check for dangerous path access in write commands
if classification == "write":
# Extract paths from command (simple heuristic)
parts = command.split()
for part in parts:
if part.startswith("/") or part.startswith("~") or part.startswith("."):
if is_dangerous_path(part):
return deny(f"Write to dangerous path blocked: {part}")
return allow()
def call(
self,
context: ToolContext,
*,
command: str,
cwd: str | None = None,
timeout: int = 120,
run_in_background: bool = False,
) -> str:
working_dir = None
if cwd:
working_dir_path = Path(os.path.expandvars(cwd)).expanduser()
if not working_dir_path.exists():
return f"Error: Working directory '{cwd}' does not exist"
working_dir = str(working_dir_path)
# Background execution: spawn task and return immediately
# Background allows up to 12 hours; foreground capped at 10 minutes
if run_in_background:
timeout = max(1, min(timeout, 43200)) # up to 12 hours
from scider.core.task import TaskManager
task_id = TaskManager.spawn_shell(
command=command,
cwd=working_dir,
timeout=timeout,
description=command[:100],
)
return (
f"Command is running in the background.\n"
f"Task ID: {task_id}\n"
f'Use TaskOutput(task_id="{task_id}") to check status and read output.'
)
# Foreground execution (capped at 10 min to avoid blocking the agent loop)
timeout = max(1, min(timeout, 600))
try:
result = subprocess.run(
[_SHELL, "-c", command],
capture_output=True,
text=True,
cwd=working_dir,
timeout=timeout,
)
output = []
if result.stdout:
output.append(f"STDOUT:\n{result.stdout}")
if result.stderr:
output.append(f"STDERR:\n{result.stderr}")
output.append(f"\nReturn code: {result.returncode}")
return "\n".join(output) if output else "Command executed with no output"
except subprocess.TimeoutExpired:
return f"Error: Command timed out after {timeout} seconds"
except Exception as e:
return f"Error executing command: {e}"