File size: 37,640 Bytes
b5b9c2e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 | """Dangerous command approval -- detection, prompting, and per-session state.
This module is the single source of truth for the dangerous command system:
- Pattern detection (DANGEROUS_PATTERNS, detect_dangerous_command)
- Per-session approval state (thread-safe, keyed by session_key)
- Approval prompting (CLI interactive + gateway async)
- Smart approval via auxiliary LLM (auto-approve low-risk commands)
- Permanent allowlist persistence (config.yaml)
"""
import contextvars
import logging
import os
import re
import sys
import threading
import unicodedata
from typing import Optional
logger = logging.getLogger(__name__)
# Per-thread/per-task gateway session identity.
# Gateway runs agent turns concurrently in executor threads, so reading a
# process-global env var for session identity is racy. Keep env fallback for
# legacy single-threaded callers, but prefer the context-local value when set.
_approval_session_key: contextvars.ContextVar[str] = contextvars.ContextVar(
"approval_session_key",
default="",
)
def set_current_session_key(session_key: str) -> contextvars.Token[str]:
"""Bind the active approval session key to the current context."""
return _approval_session_key.set(session_key or "")
def reset_current_session_key(token: contextvars.Token[str]) -> None:
"""Restore the prior approval session key context."""
_approval_session_key.reset(token)
def get_current_session_key(default: str = "default") -> str:
"""Return the active session key, preferring context-local state.
Resolution order:
1. approval-specific contextvars (set by gateway before agent.run)
2. session_context contextvars (set by _set_session_env)
3. os.environ fallback (CLI, cron, tests)
"""
session_key = _approval_session_key.get()
if session_key:
return session_key
from gateway.session_context import get_session_env
return get_session_env("HERMES_SESSION_KEY", default)
# Sensitive write targets that should trigger approval even when referenced
# via shell expansions like $HOME or $HERMES_HOME.
_SSH_SENSITIVE_PATH = r'(?:~|\$home|\$\{home\})/\.ssh(?:/|$)'
_HERMES_ENV_PATH = (
r'(?:~\/\.hermes/|'
r'(?:\$home|\$\{home\})/\.hermes/|'
r'(?:\$hermes_home|\$\{hermes_home\})/)'
r'\.env\b'
)
_SENSITIVE_WRITE_TARGET = (
r'(?:/etc/|/dev/sd|'
rf'{_SSH_SENSITIVE_PATH}|'
rf'{_HERMES_ENV_PATH})'
)
# =========================================================================
# Dangerous command patterns
# =========================================================================
DANGEROUS_PATTERNS = [
(r'\brm\s+(-[^\s]*\s+)*/', "delete in root path"),
(r'\brm\s+-[^\s]*r', "recursive delete"),
(r'\brm\s+--recursive\b', "recursive delete (long flag)"),
(r'\bchmod\s+(-[^\s]*\s+)*(777|666|o\+[rwx]*w|a\+[rwx]*w)\b', "world/other-writable permissions"),
(r'\bchmod\s+--recursive\b.*(777|666|o\+[rwx]*w|a\+[rwx]*w)', "recursive world/other-writable (long flag)"),
(r'\bchown\s+(-[^\s]*)?R\s+root', "recursive chown to root"),
(r'\bchown\s+--recursive\b.*root', "recursive chown to root (long flag)"),
(r'\bmkfs\b', "format filesystem"),
(r'\bdd\s+.*if=', "disk copy"),
(r'>\s*/dev/sd', "write to block device"),
(r'\bDROP\s+(TABLE|DATABASE)\b', "SQL DROP"),
(r'\bDELETE\s+FROM\b(?!.*\bWHERE\b)', "SQL DELETE without WHERE"),
(r'\bTRUNCATE\s+(TABLE)?\s*\w', "SQL TRUNCATE"),
(r'>\s*/etc/', "overwrite system config"),
(r'\bsystemctl\s+(stop|disable|mask)\b', "stop/disable system service"),
(r'\bkill\s+-9\s+-1\b', "kill all processes"),
(r'\bpkill\s+-9\b', "force kill processes"),
(r':\(\)\s*\{\s*:\s*\|\s*:\s*&\s*\}\s*;\s*:', "fork bomb"),
# Any shell invocation via -c or combined flags like -lc, -ic, etc.
(r'\b(bash|sh|zsh|ksh)\s+-[^\s]*c(\s+|$)', "shell command via -c/-lc flag"),
(r'\b(python[23]?|perl|ruby|node)\s+-[ec]\s+', "script execution via -e/-c flag"),
(r'\b(curl|wget)\b.*\|\s*(ba)?sh\b', "pipe remote content to shell"),
(r'\b(bash|sh|zsh|ksh)\s+<\s*<?\s*\(\s*(curl|wget)\b', "execute remote script via process substitution"),
(rf'\btee\b.*["\']?{_SENSITIVE_WRITE_TARGET}', "overwrite system file via tee"),
(rf'>>?\s*["\']?{_SENSITIVE_WRITE_TARGET}', "overwrite system file via redirection"),
(r'\bxargs\s+.*\brm\b', "xargs with rm"),
(r'\bfind\b.*-exec\s+(/\S*/)?rm\b', "find -exec rm"),
(r'\bfind\b.*-delete\b', "find -delete"),
# Gateway protection: never start gateway outside systemd management
(r'gateway\s+run\b.*(&\s*$|&\s*;|\bdisown\b|\bsetsid\b)', "start gateway outside systemd (use 'systemctl --user restart hermes-gateway')"),
(r'\bnohup\b.*gateway\s+run\b', "start gateway outside systemd (use 'systemctl --user restart hermes-gateway')"),
# Self-termination protection: prevent agent from killing its own process
(r'\b(pkill|killall)\b.*\b(hermes|gateway|cli\.py)\b', "kill hermes/gateway process (self-termination)"),
# Self-termination via kill + command substitution (pgrep/pidof).
# The name-based pattern above catches `pkill hermes` but not
# `kill -9 $(pgrep -f hermes)` because the substitution is opaque
# to regex at detection time. Catch the structural pattern instead.
(r'\bkill\b.*\$\(\s*pgrep\b', "kill process via pgrep expansion (self-termination)"),
(r'\bkill\b.*`\s*pgrep\b', "kill process via backtick pgrep expansion (self-termination)"),
# File copy/move/edit into sensitive system paths
(r'\b(cp|mv|install)\b.*\s/etc/', "copy/move file into /etc/"),
(r'\bsed\s+-[^\s]*i.*\s/etc/', "in-place edit of system config"),
(r'\bsed\s+--in-place\b.*\s/etc/', "in-place edit of system config (long flag)"),
# Script execution via heredoc — bypasses the -e/-c flag patterns above.
# `python3 << 'EOF'` feeds arbitrary code via stdin without -c/-e flags.
(r'\b(python[23]?|perl|ruby|node)\s+<<', "script execution via heredoc"),
# Git destructive operations that can lose uncommitted work or rewrite
# shared history. Not captured by rm/chmod/etc patterns.
(r'\bgit\s+reset\s+--hard\b', "git reset --hard (destroys uncommitted changes)"),
(r'\bgit\s+push\b.*--force\b', "git force push (rewrites remote history)"),
(r'\bgit\s+push\b.*-f\b', "git force push short flag (rewrites remote history)"),
(r'\bgit\s+clean\s+-[^\s]*f', "git clean with force (deletes untracked files)"),
(r'\bgit\s+branch\s+-D\b', "git branch force delete"),
# Script execution after chmod +x — catches the two-step pattern where
# a script is first made executable then immediately run. The script
# content may contain dangerous commands that individual patterns miss.
(r'\bchmod\s+\+x\b.*[;&|]+\s*\./', "chmod +x followed by immediate execution"),
]
def _legacy_pattern_key(pattern: str) -> str:
"""Reproduce the old regex-derived approval key for backwards compatibility."""
return pattern.split(r'\b')[1] if r'\b' in pattern else pattern[:20]
_PATTERN_KEY_ALIASES: dict[str, set[str]] = {}
for _pattern, _description in DANGEROUS_PATTERNS:
_legacy_key = _legacy_pattern_key(_pattern)
_canonical_key = _description
_PATTERN_KEY_ALIASES.setdefault(_canonical_key, set()).update({_canonical_key, _legacy_key})
_PATTERN_KEY_ALIASES.setdefault(_legacy_key, set()).update({_legacy_key, _canonical_key})
def _approval_key_aliases(pattern_key: str) -> set[str]:
"""Return all approval keys that should match this pattern.
New approvals use the human-readable description string, but older
command_allowlist entries and session approvals may still contain the
historical regex-derived key.
"""
return _PATTERN_KEY_ALIASES.get(pattern_key, {pattern_key})
# =========================================================================
# Detection
# =========================================================================
def _normalize_command_for_detection(command: str) -> str:
"""Normalize a command string before dangerous-pattern matching.
Strips ANSI escape sequences (full ECMA-48 via tools.ansi_strip),
null bytes, and normalizes Unicode fullwidth characters so that
obfuscation techniques cannot bypass the pattern-based detection.
"""
from tools.ansi_strip import strip_ansi
# Strip all ANSI escape sequences (CSI, OSC, DCS, 8-bit C1, etc.)
command = strip_ansi(command)
# Strip null bytes
command = command.replace('\x00', '')
# Normalize Unicode (fullwidth Latin, halfwidth Katakana, etc.)
command = unicodedata.normalize('NFKC', command)
return command
def detect_dangerous_command(command: str) -> tuple:
"""Check if a command matches any dangerous patterns.
Returns:
(is_dangerous, pattern_key, description) or (False, None, None)
"""
command_lower = _normalize_command_for_detection(command).lower()
for pattern, description in DANGEROUS_PATTERNS:
if re.search(pattern, command_lower, re.IGNORECASE | re.DOTALL):
pattern_key = description
return (True, pattern_key, description)
return (False, None, None)
# =========================================================================
# Per-session approval state (thread-safe)
# =========================================================================
_lock = threading.Lock()
_pending: dict[str, dict] = {}
_session_approved: dict[str, set] = {}
_session_yolo: set[str] = set()
_permanent_approved: set = set()
# =========================================================================
# Blocking gateway approval (mirrors CLI's synchronous input() flow)
# =========================================================================
# Per-session QUEUE of pending approvals. Multiple threads (parallel
# subagents, execute_code RPC handlers) can block concurrently — each gets
# its own threading.Event. /approve resolves the oldest, /approve all
# resolves every pending approval in the session.
class _ApprovalEntry:
"""One pending dangerous-command approval inside a gateway session."""
__slots__ = ("event", "data", "result")
def __init__(self, data: dict):
self.event = threading.Event()
self.data = data # command, description, pattern_keys, …
self.result: Optional[str] = None # "once"|"session"|"always"|"deny"
_gateway_queues: dict[str, list] = {} # session_key → [_ApprovalEntry, …]
_gateway_notify_cbs: dict[str, object] = {} # session_key → callable(approval_data)
def register_gateway_notify(session_key: str, cb) -> None:
"""Register a per-session callback for sending approval requests to the user.
The callback signature is ``cb(approval_data: dict) -> None`` where
*approval_data* contains ``command``, ``description``, and
``pattern_keys``. The callback bridges sync→async (runs in the agent
thread, must schedule the actual send on the event loop).
"""
with _lock:
_gateway_notify_cbs[session_key] = cb
def unregister_gateway_notify(session_key: str) -> None:
"""Unregister the per-session gateway approval callback.
Signals ALL blocked threads for this session so they don't hang forever
(e.g. when the agent run finishes or is interrupted).
"""
with _lock:
_gateway_notify_cbs.pop(session_key, None)
entries = _gateway_queues.pop(session_key, [])
for entry in entries:
entry.event.set()
def resolve_gateway_approval(session_key: str, choice: str,
resolve_all: bool = False) -> int:
"""Called by the gateway's /approve or /deny handler to unblock
waiting agent thread(s).
When *resolve_all* is True every pending approval in the session is
resolved at once (``/approve all``). Otherwise only the oldest one
is resolved (FIFO).
Returns the number of approvals resolved (0 means nothing was pending).
"""
with _lock:
queue = _gateway_queues.get(session_key)
if not queue:
return 0
if resolve_all:
targets = list(queue)
queue.clear()
else:
targets = [queue.pop(0)]
if not queue:
_gateway_queues.pop(session_key, None)
for entry in targets:
entry.result = choice
entry.event.set()
return len(targets)
def has_blocking_approval(session_key: str) -> bool:
"""Check if a session has one or more blocking gateway approvals waiting."""
with _lock:
return bool(_gateway_queues.get(session_key))
def submit_pending(session_key: str, approval: dict):
"""Store a pending approval request for a session."""
with _lock:
_pending[session_key] = approval
def approve_session(session_key: str, pattern_key: str):
"""Approve a pattern for this session only."""
with _lock:
_session_approved.setdefault(session_key, set()).add(pattern_key)
def enable_session_yolo(session_key: str) -> None:
"""Enable YOLO bypass for a single session key."""
if not session_key:
return
with _lock:
_session_yolo.add(session_key)
def disable_session_yolo(session_key: str) -> None:
"""Disable YOLO bypass for a single session key."""
if not session_key:
return
with _lock:
_session_yolo.discard(session_key)
def is_session_yolo_enabled(session_key: str) -> bool:
"""Return True when YOLO bypass is enabled for a specific session."""
if not session_key:
return False
with _lock:
return session_key in _session_yolo
def is_current_session_yolo_enabled() -> bool:
"""Return True when the active approval session has YOLO bypass enabled."""
return is_session_yolo_enabled(get_current_session_key(default=""))
def is_approved(session_key: str, pattern_key: str) -> bool:
"""Check if a pattern is approved (session-scoped or permanent).
Accept both the current canonical key and the legacy regex-derived key so
existing command_allowlist entries continue to work after key migrations.
"""
aliases = _approval_key_aliases(pattern_key)
with _lock:
if any(alias in _permanent_approved for alias in aliases):
return True
session_approvals = _session_approved.get(session_key, set())
return any(alias in session_approvals for alias in aliases)
def approve_permanent(pattern_key: str):
"""Add a pattern to the permanent allowlist."""
with _lock:
_permanent_approved.add(pattern_key)
def load_permanent(patterns: set):
"""Bulk-load permanent allowlist entries from config."""
with _lock:
_permanent_approved.update(patterns)
def clear_session(session_key: str):
"""Clear all approvals and pending requests for a session."""
with _lock:
_session_approved.pop(session_key, None)
_session_yolo.discard(session_key)
_pending.pop(session_key, None)
_gateway_notify_cbs.pop(session_key, None)
# Signal ALL blocked threads so they don't hang forever
entries = _gateway_queues.pop(session_key, [])
for entry in entries:
entry.event.set()
# =========================================================================
# Config persistence for permanent allowlist
# =========================================================================
def load_permanent_allowlist() -> set:
"""Load permanently allowed command patterns from config.
Also syncs them into the approval module so is_approved() works for
patterns added via 'always' in a previous session.
"""
try:
from hermes_cli.config import load_config
config = load_config()
patterns = set(config.get("command_allowlist", []) or [])
if patterns:
load_permanent(patterns)
return patterns
except Exception as e:
logger.warning("Failed to load permanent allowlist: %s", e)
return set()
def save_permanent_allowlist(patterns: set):
"""Save permanently allowed command patterns to config."""
try:
from hermes_cli.config import load_config, save_config
config = load_config()
config["command_allowlist"] = list(patterns)
save_config(config)
except Exception as e:
logger.warning("Could not save allowlist: %s", e)
# =========================================================================
# Approval prompting + orchestration
# =========================================================================
def prompt_dangerous_approval(command: str, description: str,
timeout_seconds: int | None = None,
allow_permanent: bool = True,
approval_callback=None) -> str:
"""Prompt the user to approve a dangerous command (CLI only).
Args:
allow_permanent: When False, hide the [a]lways option (used when
tirith warnings are present, since broad permanent allowlisting
is inappropriate for content-level security findings).
approval_callback: Optional callback registered by the CLI for
prompt_toolkit integration. Signature:
(command, description, *, allow_permanent=True) -> str.
Returns: 'once', 'session', 'always', or 'deny'
"""
if timeout_seconds is None:
timeout_seconds = _get_approval_timeout()
if approval_callback is not None:
try:
return approval_callback(command, description,
allow_permanent=allow_permanent)
except Exception as e:
logger.error("Approval callback failed: %s", e, exc_info=True)
return "deny"
os.environ["HERMES_SPINNER_PAUSE"] = "1"
try:
while True:
print()
print(f" ⚠️ DANGEROUS COMMAND: {description}")
print(f" {command}")
print()
if allow_permanent:
print(" [o]nce | [s]ession | [a]lways | [d]eny")
else:
print(" [o]nce | [s]ession | [d]eny")
print()
sys.stdout.flush()
result = {"choice": ""}
def get_input():
try:
prompt = " Choice [o/s/a/D]: " if allow_permanent else " Choice [o/s/D]: "
result["choice"] = input(prompt).strip().lower()
except (EOFError, OSError):
result["choice"] = ""
thread = threading.Thread(target=get_input, daemon=True)
thread.start()
thread.join(timeout=timeout_seconds)
if thread.is_alive():
print("\n ⏱ Timeout - denying command")
return "deny"
choice = result["choice"]
if choice in ('o', 'once'):
print(" ✓ Allowed once")
return "once"
elif choice in ('s', 'session'):
print(" ✓ Allowed for this session")
return "session"
elif choice in ('a', 'always'):
if not allow_permanent:
print(" ✓ Allowed for this session")
return "session"
print(" ✓ Added to permanent allowlist")
return "always"
else:
print(" ✗ Denied")
return "deny"
except (EOFError, KeyboardInterrupt):
print("\n ✗ Cancelled")
return "deny"
finally:
if "HERMES_SPINNER_PAUSE" in os.environ:
del os.environ["HERMES_SPINNER_PAUSE"]
print()
sys.stdout.flush()
def _normalize_approval_mode(mode) -> str:
"""Normalize approval mode values loaded from YAML/config.
YAML 1.1 treats bare words like `off` as booleans, so a config entry like
`approvals:\n mode: off` is parsed as False unless quoted. Treat that as the
intended string mode instead of falling back to manual approvals.
"""
if isinstance(mode, bool):
return "off" if mode is False else "manual"
if isinstance(mode, str):
normalized = mode.strip().lower()
return normalized or "manual"
return "manual"
def _get_approval_config() -> dict:
"""Read the approvals config block. Returns a dict with 'mode', 'timeout', etc."""
try:
from hermes_cli.config import load_config
config = load_config()
return config.get("approvals", {}) or {}
except Exception as e:
logger.warning("Failed to load approval config: %s", e)
return {}
def _get_approval_mode() -> str:
"""Read the approval mode from config. Returns 'manual', 'smart', or 'off'."""
mode = _get_approval_config().get("mode", "manual")
return _normalize_approval_mode(mode)
def _get_approval_timeout() -> int:
"""Read the approval timeout from config. Defaults to 60 seconds."""
try:
return int(_get_approval_config().get("timeout", 60))
except (ValueError, TypeError):
return 60
def _smart_approve(command: str, description: str) -> str:
"""Use the auxiliary LLM to assess risk and decide approval.
Returns 'approve' if the LLM determines the command is safe,
'deny' if genuinely dangerous, or 'escalate' if uncertain.
Inspired by OpenAI Codex's Smart Approvals guardian subagent
(openai/codex#13860).
"""
try:
from agent.auxiliary_client import get_text_auxiliary_client, auxiliary_max_tokens_param
client, model = get_text_auxiliary_client(task="approval")
if not client or not model:
logger.debug("Smart approvals: no aux client available, escalating")
return "escalate"
prompt = f"""You are a security reviewer for an AI coding agent. A terminal command was flagged by pattern matching as potentially dangerous.
Command: {command}
Flagged reason: {description}
Assess the ACTUAL risk of this command. Many flagged commands are false positives — for example, `python -c "print('hello')"` is flagged as "script execution via -c flag" but is completely harmless.
Rules:
- APPROVE if the command is clearly safe (benign script execution, safe file operations, development tools, package installs, git operations, etc.)
- DENY if the command could genuinely damage the system (recursive delete of important paths, overwriting system files, fork bombs, wiping disks, dropping databases, etc.)
- ESCALATE if you're uncertain
Respond with exactly one word: APPROVE, DENY, or ESCALATE"""
response = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": prompt}],
**auxiliary_max_tokens_param(16),
temperature=0,
)
answer = (response.choices[0].message.content or "").strip().upper()
if "APPROVE" in answer:
return "approve"
elif "DENY" in answer:
return "deny"
else:
return "escalate"
except Exception as e:
logger.debug("Smart approvals: LLM call failed (%s), escalating", e)
return "escalate"
def check_dangerous_command(command: str, env_type: str,
approval_callback=None) -> dict:
"""Check if a command is dangerous and handle approval.
This is the main entry point called by terminal_tool before executing
any command. It orchestrates detection, session checks, and prompting.
Args:
command: The shell command to check.
env_type: Terminal backend type ('local', 'ssh', 'docker', etc.).
approval_callback: Optional CLI callback for interactive prompts.
Returns:
{"approved": True/False, "message": str or None, ...}
"""
if env_type in ("docker", "singularity", "modal", "daytona"):
return {"approved": True, "message": None}
# --yolo: bypass all approval prompts. Gateway /yolo is session-scoped;
# CLI --yolo remains process-scoped via the env var for local use.
if os.getenv("HERMES_YOLO_MODE") or is_current_session_yolo_enabled():
return {"approved": True, "message": None}
is_dangerous, pattern_key, description = detect_dangerous_command(command)
if not is_dangerous:
return {"approved": True, "message": None}
session_key = get_current_session_key()
if is_approved(session_key, pattern_key):
return {"approved": True, "message": None}
is_cli = os.getenv("HERMES_INTERACTIVE")
is_gateway = os.getenv("HERMES_GATEWAY_SESSION")
if not is_cli and not is_gateway:
return {"approved": True, "message": None}
if is_gateway or os.getenv("HERMES_EXEC_ASK"):
submit_pending(session_key, {
"command": command,
"pattern_key": pattern_key,
"description": description,
})
return {
"approved": False,
"pattern_key": pattern_key,
"status": "approval_required",
"command": command,
"description": description,
"message": (
f"⚠️ This command is potentially dangerous ({description}). "
f"Asking the user for approval.\n\n**Command:**\n```\n{command}\n```"
),
}
choice = prompt_dangerous_approval(command, description,
approval_callback=approval_callback)
if choice == "deny":
return {
"approved": False,
"message": f"BLOCKED: User denied this potentially dangerous command (matched '{description}' pattern). Do NOT retry this command - the user has explicitly rejected it.",
"pattern_key": pattern_key,
"description": description,
}
if choice == "session":
approve_session(session_key, pattern_key)
elif choice == "always":
approve_session(session_key, pattern_key)
approve_permanent(pattern_key)
save_permanent_allowlist(_permanent_approved)
return {"approved": True, "message": None}
# =========================================================================
# Combined pre-exec guard (tirith + dangerous command detection)
# =========================================================================
def _format_tirith_description(tirith_result: dict) -> str:
"""Build a human-readable description from tirith findings.
Includes severity, title, and description for each finding so users
can make an informed approval decision.
"""
findings = tirith_result.get("findings") or []
if not findings:
summary = tirith_result.get("summary") or "security issue detected"
return f"Security scan: {summary}"
parts = []
for f in findings:
severity = f.get("severity", "")
title = f.get("title", "")
desc = f.get("description", "")
if title and desc:
parts.append(f"[{severity}] {title}: {desc}" if severity else f"{title}: {desc}")
elif title:
parts.append(f"[{severity}] {title}" if severity else title)
if not parts:
summary = tirith_result.get("summary") or "security issue detected"
return f"Security scan: {summary}"
return "Security scan — " + "; ".join(parts)
def check_all_command_guards(command: str, env_type: str,
approval_callback=None) -> dict:
"""Run all pre-exec security checks and return a single approval decision.
Gathers findings from tirith and dangerous-command detection, then
presents them as a single combined approval request. This prevents
a gateway force=True replay from bypassing one check when only the
other was shown to the user.
"""
# Skip containers for both checks
if env_type in ("docker", "singularity", "modal", "daytona"):
return {"approved": True, "message": None}
# --yolo or approvals.mode=off: bypass all approval prompts.
# Gateway /yolo is session-scoped; CLI --yolo remains process-scoped.
approval_mode = _get_approval_mode()
if os.getenv("HERMES_YOLO_MODE") or is_current_session_yolo_enabled() or approval_mode == "off":
return {"approved": True, "message": None}
is_cli = os.getenv("HERMES_INTERACTIVE")
is_gateway = os.getenv("HERMES_GATEWAY_SESSION")
is_ask = os.getenv("HERMES_EXEC_ASK")
# Preserve the existing non-interactive behavior: outside CLI/gateway/ask
# flows, we do not block on approvals and we skip external guard work.
if not is_cli and not is_gateway and not is_ask:
return {"approved": True, "message": None}
# --- Phase 1: Gather findings from both checks ---
# Tirith check — wrapper guarantees no raise for expected failures.
# Only catch ImportError (module not installed).
tirith_result = {"action": "allow", "findings": [], "summary": ""}
try:
from tools.tirith_security import check_command_security
tirith_result = check_command_security(command)
except ImportError:
pass # tirith module not installed — allow
# Dangerous command check (detection only, no approval)
is_dangerous, pattern_key, description = detect_dangerous_command(command)
# --- Phase 2: Decide ---
# Collect warnings that need approval
warnings = [] # list of (pattern_key, description, is_tirith)
session_key = get_current_session_key()
# Tirith block/warn → approvable warning with rich findings.
# Previously, tirith "block" was a hard block with no approval prompt.
# Now both block and warn go through the approval flow so users can
# inspect the explanation and approve if they understand the risk.
if tirith_result["action"] in ("block", "warn"):
findings = tirith_result.get("findings") or []
rule_id = findings[0].get("rule_id", "unknown") if findings else "unknown"
tirith_key = f"tirith:{rule_id}"
tirith_desc = _format_tirith_description(tirith_result)
if not is_approved(session_key, tirith_key):
warnings.append((tirith_key, tirith_desc, True))
if is_dangerous:
if not is_approved(session_key, pattern_key):
warnings.append((pattern_key, description, False))
# Nothing to warn about
if not warnings:
return {"approved": True, "message": None}
# --- Phase 2.5: Smart approval (auxiliary LLM risk assessment) ---
# When approvals.mode=smart, ask the aux LLM before prompting the user.
# Inspired by OpenAI Codex's Smart Approvals guardian subagent
# (openai/codex#13860).
if approval_mode == "smart":
combined_desc_for_llm = "; ".join(desc for _, desc, _ in warnings)
verdict = _smart_approve(command, combined_desc_for_llm)
if verdict == "approve":
# Auto-approve and grant session-level approval for these patterns
for key, _, _ in warnings:
approve_session(session_key, key)
logger.debug("Smart approval: auto-approved '%s' (%s)",
command[:60], combined_desc_for_llm)
return {"approved": True, "message": None,
"smart_approved": True,
"description": combined_desc_for_llm}
elif verdict == "deny":
combined_desc_for_llm = "; ".join(desc for _, desc, _ in warnings)
return {
"approved": False,
"message": f"BLOCKED by smart approval: {combined_desc_for_llm}. "
"The command was assessed as genuinely dangerous. Do NOT retry.",
"smart_denied": True,
}
# verdict == "escalate" → fall through to manual prompt
# --- Phase 3: Approval ---
# Combine descriptions for a single approval prompt
combined_desc = "; ".join(desc for _, desc, _ in warnings)
primary_key = warnings[0][0]
all_keys = [key for key, _, _ in warnings]
has_tirith = any(is_t for _, _, is_t in warnings)
# Gateway/async approval — block the agent thread until the user
# responds with /approve or /deny, mirroring the CLI's synchronous
# input() flow. The agent never sees "approval_required"; it either
# gets the command output (approved) or a definitive "BLOCKED" message.
if is_gateway or is_ask:
notify_cb = None
with _lock:
notify_cb = _gateway_notify_cbs.get(session_key)
if notify_cb is not None:
# --- Blocking gateway approval (queue-based) ---
# Each call gets its own _ApprovalEntry so parallel subagents
# and execute_code threads can block concurrently.
approval_data = {
"command": command,
"pattern_key": primary_key,
"pattern_keys": all_keys,
"description": combined_desc,
}
entry = _ApprovalEntry(approval_data)
with _lock:
_gateway_queues.setdefault(session_key, []).append(entry)
# Notify the user (bridges sync agent thread → async gateway)
try:
notify_cb(approval_data)
except Exception as exc:
logger.warning("Gateway approval notify failed: %s", exc)
with _lock:
queue = _gateway_queues.get(session_key, [])
if entry in queue:
queue.remove(entry)
if not queue:
_gateway_queues.pop(session_key, None)
return {
"approved": False,
"message": "BLOCKED: Failed to send approval request to user. Do NOT retry.",
"pattern_key": primary_key,
"description": combined_desc,
}
# Block until the user responds or timeout (default 5 min)
timeout = _get_approval_config().get("gateway_timeout", 300)
try:
timeout = int(timeout)
except (ValueError, TypeError):
timeout = 300
resolved = entry.event.wait(timeout=timeout)
# Clean up this entry from the queue
with _lock:
queue = _gateway_queues.get(session_key, [])
if entry in queue:
queue.remove(entry)
if not queue:
_gateway_queues.pop(session_key, None)
choice = entry.result
if not resolved or choice is None or choice == "deny":
reason = "timed out" if not resolved else "denied by user"
return {
"approved": False,
"message": f"BLOCKED: Command {reason}. Do NOT retry this command.",
"pattern_key": primary_key,
"description": combined_desc,
}
# User approved — persist based on scope (same logic as CLI)
for key, _, is_tirith in warnings:
if choice == "session" or (choice == "always" and is_tirith):
approve_session(session_key, key)
elif choice == "always":
approve_session(session_key, key)
approve_permanent(key)
save_permanent_allowlist(_permanent_approved)
# choice == "once": no persistence — command allowed this
# single time only, matching the CLI's behavior.
return {"approved": True, "message": None,
"user_approved": True, "description": combined_desc}
# Fallback: no gateway callback registered (e.g. cron, batch).
# Return approval_required for backward compat.
submit_pending(session_key, {
"command": command,
"pattern_key": primary_key,
"pattern_keys": all_keys,
"description": combined_desc,
})
return {
"approved": False,
"pattern_key": primary_key,
"status": "approval_required",
"command": command,
"description": combined_desc,
"message": (
f"⚠️ {combined_desc}. Asking the user for approval.\n\n**Command:**\n```\n{command}\n```"
),
}
# CLI interactive: single combined prompt
# Hide [a]lways when any tirith warning is present
choice = prompt_dangerous_approval(command, combined_desc,
allow_permanent=not has_tirith,
approval_callback=approval_callback)
if choice == "deny":
return {
"approved": False,
"message": "BLOCKED: User denied. Do NOT retry.",
"pattern_key": primary_key,
"description": combined_desc,
}
# Persist approval for each warning individually
for key, _, is_tirith in warnings:
if choice == "session" or (choice == "always" and is_tirith):
# tirith: session only (no permanent broad allowlisting)
approve_session(session_key, key)
elif choice == "always":
# dangerous patterns: permanent allowed
approve_session(session_key, key)
approve_permanent(key)
save_permanent_allowlist(_permanent_approved)
return {"approved": True, "message": None,
"user_approved": True, "description": combined_desc}
# Load permanent allowlist from config on module import
load_permanent_allowlist()
|