feat(approvals): hermes approvals suggest — mine approval history into allowlist proposals

This commit is contained in:
teknium1 2026-07-26 14:30:02 -07:00 committed by Teknium
parent d6fdae6746
commit db90e36202
5 changed files with 1021 additions and 0 deletions

View file

@ -0,0 +1,482 @@
"""``hermes approvals suggest`` — mine approval history into allowlist proposals.
Hermes has no dedicated approval-decision ledger: ``always`` answers land in
``command_allowlist`` (config.yaml) via :func:`tools.approval.save_permanent_allowlist`,
while ``once``/``session`` approvals are in-memory only. What *does* persist
is the session DB (``~/.hermes/state.db``): every assistant ``terminal`` tool
call is stored with its arguments, and the paired ``role='tool'`` result
records whether the command was blocked/denied ("BLOCKED: User denied …",
"Asking the user for approval") or actually executed.
So this module mines *implied approvals*: a command that matches a
dangerous-command class (the same :func:`tools.approval.detect_dangerous_command`
classifier that triggers the prompt) AND whose tool result is not a
block/denial marker must have been approved by the user (once, session,
always, smart-approve, or yolo) before it ran. Frequently re-approved
patterns are exactly the prompts worth turning into one-time allowlist
policy the port of Claude Code's ``/fewer-permission-prompts``.
Safety posture:
* **Never auto-applies.** The default run is a dry proposal; only an explicit
``--apply N[,M...]`` merges the chosen patterns into ``command_allowlist``
via the existing :func:`tools.approval.save_permanent_allowlist` path.
* **Hardline commands are never proposed** anything matched by
:func:`tools.approval.detect_hardline_command` is dropped outright.
* **Destructive / privilege / credential / obfuscation classes are never
proposed**, no matter how often they were approved. ``rm -rf build/``
approved 100 times still never yields an ``rm`` allowlist entry. Only
benign, recoverable classes (container lifecycle, git force push, service
restarts, hermes self-management, ) are eligible.
* **Dangerous root binaries never become globs** (``rm *``, ``sudo *`` ).
"""
from __future__ import annotations
import json
import re
import sqlite3
import time
from dataclasses import dataclass, field
from pathlib import Path
from typing import Iterable, Iterator, Optional
# ---------------------------------------------------------------------------
# Safety exclusions
# ---------------------------------------------------------------------------
# Dangerous-class descriptions matching ANY of these are never proposed,
# regardless of approval frequency. Matched case-insensitively against the
# pattern-key/description strings produced by tools.approval's
# DANGEROUS_PATTERNS / execution-flag findings. Deliberately conservative:
# a benign class accidentally excluded costs the user one manual config edit;
# a destructive class accidentally proposed costs them data.
_UNSAFE_CLASS_PATTERNS = [
r"delete", # recursive delete, find -delete, branch force delete, ...
r"\brm\b", # xargs with rm, find -exec rm
r"destro", # git reset --hard (destroys ...), destructive
r"wipe",
r"format", # format filesystem
r"\bdisk\b",
r"block device",
r"fork bomb",
r"kill (?:all )?process", # force/regex/all process kills
r"kill all",
r"self-termination",
r"\bsudo\b",
r"privilege",
r"credential",
r"\bssh\b",
r"shell.rc",
r"system config",
r"system file",
r"\bsql\b", # SQL DROP / TRUNCATE / DELETE without WHERE
r"\bchown\b",
r"\bchmod\b",
r"writable", # world/other-writable permissions
r"overwrite",
r"in-place edit",
r"pipe", # pipe remote/decoded content to shell
r"obfuscation",
r"remote content",
r"remote script",
r"heredoc",
r"encoded", # PowerShell encoded command execution
r"command substitution",
r"process substitution",
r"\bdd\b",
r"shutdown",
r"reboot",
r"hardline",
]
_UNSAFE_CLASS_RE = re.compile("|".join(_UNSAFE_CLASS_PATTERNS), re.IGNORECASE)
# Root binaries that must never anchor a proposed command glob, even if the
# class survived the description filter. Prefix match for the mkfs family.
_UNSAFE_ROOT_BINARIES = {
"rm", "rmdir", "unlink", "shred", "dd", "fdisk", "parted", "wipefs",
"sudo", "doas", "su", "chmod", "chown", "chgrp",
"kill", "killall", "pkill",
"halt", "shutdown", "reboot", "poweroff", "init",
"del", "format", "truncate", "mkswap",
}
_UNSAFE_ROOT_PREFIXES = ("mkfs",)
# Substrings in a role='tool' result that mean the command did NOT execute
# with user consent (blocked, denied, timed out, or still pending). Kept in
# sync with the message templates in tools/approval.py.
_BLOCK_MARKERS = (
"BLOCKED (hardline)",
"BLOCKED: User denied",
"BLOCKED: Action ",
"BLOCKED: Command flagged as dangerous",
"BLOCKED: approval required",
"BLOCKED: Failed to send approval request",
"The user has NOT consented",
"Asking the user for approval",
"approval_required",
"BLOCKED by user deny rule",
)
@dataclass
class Proposal:
"""One ranked allowlist proposal."""
pattern: str # command glob ("git push *") or class key
kind: str # "glob" | "class"
count: int = 0
classes: set = field(default_factory=set)
examples: list = field(default_factory=list)
def add_example(self, command: str) -> None:
short = command.strip()
if len(short) > 100:
short = short[:97] + "..."
if short not in self.examples and len(self.examples) < 3:
self.examples.append(short)
# ---------------------------------------------------------------------------
# Scan: session DB -> (command, class description) records
# ---------------------------------------------------------------------------
def default_db_path() -> Path:
from hermes_constants import get_hermes_home
return get_hermes_home() / "state.db"
def _connect_readonly(db_path: Path) -> sqlite3.Connection:
uri = f"file:{db_path}?mode=ro"
return sqlite3.connect(uri, uri=True)
def _iter_terminal_calls(
con: sqlite3.Connection, since_ts: float
) -> Iterator[tuple[str, str]]:
"""Yield ``(tool_call_id, command)`` for every terminal tool call."""
cur = con.execute(
"SELECT tool_calls FROM messages "
"WHERE role='assistant' AND tool_calls IS NOT NULL "
"AND tool_calls LIKE '%terminal%' AND timestamp >= ?",
(since_ts,),
)
while True:
rows = cur.fetchmany(2000)
if not rows:
break
for (raw,) in rows:
try:
calls = json.loads(raw)
except (TypeError, ValueError):
continue
if not isinstance(calls, list):
continue
for call in calls:
if not isinstance(call, dict):
continue
fn = call.get("function") or {}
if fn.get("name") != "terminal":
continue
try:
args = json.loads(fn.get("arguments") or "{}")
except (TypeError, ValueError):
continue
command = args.get("command")
if isinstance(command, str) and command.strip():
yield (call.get("id") or "", command)
def _blocked_tool_call_ids(con: sqlite3.Connection, since_ts: float) -> set:
"""Collect tool_call_ids whose result shows the command never ran freely."""
blocked: set = set()
cur = con.execute(
"SELECT tool_call_id, content FROM messages "
"WHERE role='tool' AND tool_call_id IS NOT NULL AND timestamp >= ? "
"AND (content LIKE '%BLOCKED%' OR content LIKE '%approval%')",
(since_ts,),
)
while True:
rows = cur.fetchmany(2000)
if not rows:
break
for tool_call_id, content in rows:
if not content:
continue
if any(marker in content for marker in _BLOCK_MARKERS):
blocked.add(tool_call_id)
return blocked
def scan_approval_history(
db_path: Optional[Path] = None, days: int = 90
) -> list[tuple[str, str]]:
"""Return ``(command, dangerous_class_description)`` records mined from
the session DB dangerous-classified terminal commands that actually
executed (i.e. carried an implied user approval).
"""
from tools.approval import detect_dangerous_command, detect_hardline_command
path = Path(db_path) if db_path else default_db_path()
if not path.exists():
return []
since_ts = 0.0 if days <= 0 else time.time() - days * 86400
records: list[tuple[str, str]] = []
con = _connect_readonly(path)
try:
blocked = _blocked_tool_call_ids(con, since_ts)
for tool_call_id, command in _iter_terminal_calls(con, since_ts):
if tool_call_id in blocked:
continue
is_hardline, _desc = detect_hardline_command(command)
if is_hardline:
# Hardline commands are unconditionally blocked at runtime;
# never mine them (defense in depth against stale DB rows).
continue
is_dangerous, _key, description = detect_dangerous_command(command)
if not is_dangerous:
continue
records.append((command, description))
finally:
con.close()
return records
# ---------------------------------------------------------------------------
# Normalize -> aggregate -> rank -> exclude
# ---------------------------------------------------------------------------
def normalize_command(command: str) -> str:
"""Fold user/hermes home prefixes and collapse whitespace.
Reuses tools.approval's home-folding machinery so proposals are portable
across machines/users (``/home/alice/x`` -> ``~/x``).
"""
from tools.approval import (
_rewrite_resolved_hermes_home,
_rewrite_resolved_user_home,
)
folded = _rewrite_resolved_user_home(_rewrite_resolved_hermes_home(command))
return " ".join(folded.split())
def is_unsafe_class(description: str) -> bool:
"""True when a dangerous-class description must never be proposed."""
return bool(_UNSAFE_CLASS_RE.search(description or ""))
def _unsafe_root_binary(token: str) -> bool:
tok = token.lower().rsplit("/", 1)[-1]
if tok in _UNSAFE_ROOT_BINARIES:
return True
return any(tok.startswith(p) for p in _UNSAFE_ROOT_PREFIXES)
def derive_glob(normalized: str) -> Optional[str]:
"""Derive a narrow command glob (``git push *``) from a simple command.
Returns None for compound commands (shell operators the runtime
allowlist matcher refuses those anyway) and for commands anchored on an
unsafe root binary.
"""
from tools.approval import _has_allowlist_shell_operator
if _has_allowlist_shell_operator(normalized):
return None
tokens = normalized.split()
if not tokens:
return None
if _unsafe_root_binary(tokens[0]):
return None
if len(tokens) == 1:
return tokens[0]
second = tokens[1]
if second.startswith("-") or any(ch in second for ch in "*?[$"):
return f"{tokens[0]} *"
return f"{tokens[0]} {second} *"
def build_proposals(
records: Iterable[tuple[str, str]],
existing: Optional[set] = None,
min_count: int = 2,
limit: int = 20,
) -> list[Proposal]:
"""Aggregate scan records into a ranked, safety-filtered proposal list.
Grain: a command glob (``git push *``) for simple commands; the
dangerous-class description itself (the same key an interactive
``[a]lways`` answer persists) for compound commands where no safe glob
can be derived.
"""
existing = existing or set()
by_pattern: dict[tuple[str, str], Proposal] = {}
for command, description in records:
if is_unsafe_class(description):
continue
normalized = normalize_command(command)
glob = derive_glob(normalized)
if glob is not None:
key = (glob, "glob")
else:
key = (description, "class")
pattern, kind = key
if pattern in existing:
continue
proposal = by_pattern.get(key)
if proposal is None:
proposal = by_pattern[key] = Proposal(pattern=pattern, kind=kind)
proposal.count += 1
proposal.classes.add(description)
proposal.add_example(normalized)
ranked = [p for p in by_pattern.values() if p.count >= max(min_count, 1)]
ranked.sort(key=lambda p: (-p.count, p.pattern))
return ranked[: max(limit, 1)]
# ---------------------------------------------------------------------------
# Apply / render
# ---------------------------------------------------------------------------
def parse_apply_indices(spec: str, total: int) -> list[int]:
"""Parse ``"1,3"`` into validated zero-based indices."""
indices: list[int] = []
for part in (spec or "").split(","):
part = part.strip()
if not part:
continue
try:
n = int(part)
except ValueError:
raise ValueError(f"invalid selection {part!r} — expected numbers like 1,3")
if n < 1 or n > total:
raise ValueError(f"selection {n} out of range (1..{total})")
if (n - 1) not in indices:
indices.append(n - 1)
if not indices:
raise ValueError("no valid selections in --apply")
return indices
def apply_proposals(proposals: list[Proposal], indices: list[int]) -> set:
"""Merge chosen proposal patterns into command_allowlist and persist."""
import tools.approval as approval_module
merged = set(approval_module.load_permanent_allowlist())
for idx in indices:
merged.add(proposals[idx].pattern)
approval_module.save_permanent_allowlist(merged)
# Keep the in-process allowlist consistent so a long-lived process sees
# the new entries immediately (mirrors the interactive 'always' path).
approval_module.load_permanent(merged)
return merged
def _render_text(proposals: list[Proposal], days: int) -> None:
window = "all history" if days <= 0 else f"last {days} days"
if not proposals:
print(
f"No allowlist candidates found in approval history ({window}).\n"
"Either nothing dangerous was approved often enough "
"(see --min-count/--days), or the approved classes are "
"excluded for safety."
)
return
print(f"Proposed command_allowlist additions (from approval history, {window}):\n")
for i, p in enumerate(proposals, 1):
kind = " (class key)" if p.kind == "class" else ""
print(f" {i}. {p.pattern} — approved {p.count}x{kind}")
for cls in sorted(p.classes):
print(f" class: {cls}")
for ex in p.examples:
print(f" e.g. {ex}")
print(
"\nNothing has been changed. Apply selected entries with:\n"
" hermes approvals suggest --apply 1,3\n"
"Entries are merged into command_allowlist in ~/.hermes/config.yaml."
)
def suggest_command(args) -> int:
"""Entry point for ``hermes approvals suggest``."""
db_path = Path(args.db) if getattr(args, "db", None) else default_db_path()
days = getattr(args, "days", 90)
if not db_path.exists():
print(f"Session database not found: {db_path}")
return 1
import tools.approval as approval_module
existing = set(approval_module.load_permanent_allowlist())
records = scan_approval_history(db_path, days=days)
proposals = build_proposals(
records,
existing=existing,
min_count=getattr(args, "min_count", 2),
limit=getattr(args, "limit", 20),
)
apply_spec = getattr(args, "apply_indices", None)
if apply_spec:
try:
indices = parse_apply_indices(apply_spec, len(proposals))
except ValueError as exc:
print(f"--apply error: {exc}")
return 1
merged = apply_proposals(proposals, indices)
applied = [proposals[i].pattern for i in indices]
if getattr(args, "json", False):
print(json.dumps({"applied": applied, "allowlist_size": len(merged)}))
else:
print("Added to command_allowlist:")
for pattern in applied:
print(f" + {pattern}")
print(f"\ncommand_allowlist now has {len(merged)} entries "
"(~/.hermes/config.yaml).")
return 0
if getattr(args, "json", False):
payload = {
"db": str(db_path),
"days": days,
"proposals": [
{
"n": i,
"pattern": p.pattern,
"kind": p.kind,
"count": p.count,
"classes": sorted(p.classes),
"examples": p.examples,
}
for i, p in enumerate(proposals, 1)
],
}
print(json.dumps(payload, indent=2))
return 0
_render_text(proposals, days)
return 0
def approvals_command(args) -> int:
"""Dispatch ``hermes approvals <subcommand>``."""
sub = getattr(args, "approvals_command", None)
if sub == "suggest":
return suggest_command(args)
print(
"usage: hermes approvals <subcommand>\n"
"\n"
"subcommands:\n"
" suggest Mine past approval decisions into a proposed\n"
" command_allowlist (dry by default; --apply N,M to merge)\n"
"\n"
"Run `hermes approvals suggest -h` for details."
)
return 1

View file

@ -445,6 +445,7 @@ from hermes_cli.subcommands.webhook import build_webhook_parser
from hermes_cli.subcommands.hooks import build_hooks_parser
from hermes_cli.subcommands.doctor import build_doctor_parser
from hermes_cli.subcommands.security import build_security_parser
from hermes_cli.subcommands.approvals import build_approvals_parser
from hermes_cli.subcommands.dump import build_dump_parser
from hermes_cli.subcommands.debug import build_debug_parser
from hermes_cli.subcommands.backup import build_backup_parser
@ -4558,6 +4559,16 @@ def cmd_security(args):
sys.exit(2)
def cmd_approvals(args):
"""Dispatch `hermes approvals <subcmd>`."""
from hermes_cli.approvals_suggest import approvals_command
status = approvals_command(args)
if status:
sys.exit(status)
return status
def cmd_dump(args):
"""Dump setup summary for support/debugging."""
from hermes_cli.dump import run_dump
@ -15657,6 +15668,11 @@ def main():
# =========================================================================
build_security_parser(subparsers, cmd_security=cmd_security)
# =========================================================================
# approvals command (parser built in hermes_cli/subcommands/approvals.py)
# =========================================================================
build_approvals_parser(subparsers, cmd_approvals=cmd_approvals)
# =========================================================================
# dump command (parser built in hermes_cli/subcommands/dump.py)
# =========================================================================

View file

@ -0,0 +1,77 @@
"""``hermes approvals`` subcommand parser.
Follows the cron/security pattern: parser construction lives here, the
handler is injected by ``main.py`` so this module never imports ``main``
(cycle avoidance).
"""
from __future__ import annotations
from typing import Callable
def build_approvals_parser(subparsers, *, cmd_approvals: Callable) -> None:
"""Attach the ``approvals`` subcommand to ``subparsers``."""
approvals_parser = subparsers.add_parser(
"approvals",
help="Approval-prompt tools (mine history into allowlist proposals)",
description=(
"Tools for the dangerous-command approval system. "
"`hermes approvals suggest` mines past approval decisions from "
"the session database and proposes command_allowlist entries so "
"repeatedly-approved commands stop prompting."
),
)
approvals_subparsers = approvals_parser.add_subparsers(
dest="approvals_command",
metavar="<subcommand>",
)
suggest_parser = approvals_subparsers.add_parser(
"suggest",
help="Propose command_allowlist entries from past approvals",
description=(
"Scan the session database for dangerous-classified commands "
"that ran with user approval, rank the recurring patterns, and "
"print a numbered allowlist proposal. Nothing is written unless "
"--apply is given. Destructive classes (recursive delete, sudo, "
"disk writes, credential edits, ...) are never proposed."
),
)
suggest_parser.add_argument(
"--apply",
dest="apply_indices",
metavar="N[,M...]",
help="Merge the numbered proposals (from a prior run) into "
"command_allowlist in config.yaml",
)
suggest_parser.add_argument(
"--json",
action="store_true",
help="Emit machine-readable JSON instead of human-readable text",
)
suggest_parser.add_argument(
"--days",
type=int,
default=90,
help="How far back to scan session history (default: 90; 0 = all)",
)
suggest_parser.add_argument(
"--min-count",
dest="min_count",
type=int,
default=2,
help="Minimum approval count for a pattern to be proposed (default: 2)",
)
suggest_parser.add_argument(
"--limit",
type=int,
default=20,
help="Maximum number of proposals to show (default: 20)",
)
suggest_parser.add_argument(
"--db",
help="Path to an alternate session database (default: ~/.hermes/state.db)",
)
suggest_parser.set_defaults(func=cmd_approvals)
approvals_parser.set_defaults(func=cmd_approvals)

View file

@ -0,0 +1,409 @@
"""Tests for ``hermes approvals suggest`` (hermes_cli/approvals_suggest.py).
Approval history in Hermes is implied, not ledgered: the session DB
(state.db) stores every assistant ``terminal`` tool call plus its paired
``role='tool'`` result. A dangerous-classified command whose result is not a
BLOCKED/denied/pending marker ran with user consent. These tests build
synthetic state.db fixtures and verify scanning, ranking, safety exclusions,
and the --apply merge path.
"""
from __future__ import annotations
import json
import sqlite3
import time
from argparse import Namespace
import pytest
import tools.approval as approval_module
from hermes_cli.approvals_suggest import (
Proposal,
apply_proposals,
build_proposals,
derive_glob,
is_unsafe_class,
normalize_command,
parse_apply_indices,
scan_approval_history,
suggest_command,
)
# ---------------------------------------------------------------------------
# Fixture helpers: synthetic session DB
# ---------------------------------------------------------------------------
def _make_db(path):
con = sqlite3.connect(path)
con.executescript(
"""
CREATE TABLE messages (
id INTEGER PRIMARY KEY AUTOINCREMENT,
session_id TEXT NOT NULL,
role TEXT NOT NULL,
content TEXT,
tool_call_id TEXT,
tool_calls TEXT,
timestamp REAL NOT NULL
);
"""
)
con.commit()
return con
_ID_COUNTER = [0]
def _add_terminal_call(con, command, result="ok: done", ts=None):
"""Insert an assistant terminal tool call + its paired tool result."""
_ID_COUNTER[0] += 1
call_id = f"call_{_ID_COUNTER[0]}"
ts = ts if ts is not None else time.time()
tool_calls = json.dumps(
[
{
"id": call_id,
"type": "function",
"function": {
"name": "terminal",
"arguments": json.dumps({"command": command}),
},
}
]
)
con.execute(
"INSERT INTO messages (session_id, role, content, tool_calls, timestamp) "
"VALUES ('s1', 'assistant', '', ?, ?)",
(tool_calls, ts),
)
con.execute(
"INSERT INTO messages (session_id, role, content, tool_call_id, timestamp) "
"VALUES ('s1', 'tool', ?, ?, ?)",
(result, call_id, ts + 1),
)
con.commit()
@pytest.fixture
def db_path(tmp_path):
path = tmp_path / "state.db"
con = _make_db(path)
yield path, con
con.close()
@pytest.fixture
def isolated_allowlist(monkeypatch):
"""Fake config-backed allowlist store so no real config.yaml is touched."""
store = {"patterns": set(), "saves": 0}
def fake_load():
return set(store["patterns"])
def fake_save(patterns):
store["patterns"] = set(patterns)
store["saves"] += 1
monkeypatch.setattr(approval_module, "load_permanent_allowlist", fake_load)
monkeypatch.setattr(approval_module, "save_permanent_allowlist", fake_save)
# apply_proposals also syncs into the in-process set; keep it isolated.
saved = approval_module._permanent_approved.copy()
approval_module._permanent_approved.clear()
yield store
approval_module._permanent_approved.clear()
approval_module._permanent_approved.update(saved)
# ---------------------------------------------------------------------------
# Scan
# ---------------------------------------------------------------------------
class TestScan:
def test_scan_finds_approved_dangerous_commands(self, db_path):
path, con = db_path
for _ in range(3):
_add_terminal_call(con, "git push --force origin main")
# A benign command never enters approval history.
_add_terminal_call(con, "ls -la")
records = scan_approval_history(path, days=0)
commands = [c for c, _ in records]
assert len(commands) == 3
assert all("git push" in c for c in commands)
def test_blocked_and_denied_results_are_not_approvals(self, db_path):
path, con = db_path
_add_terminal_call(
con,
"git push --force origin main",
result="BLOCKED: User denied this potentially dangerous action",
)
_add_terminal_call(
con,
"docker restart web",
result=(
"⚠️ This action is potentially dangerous. "
"Asking the user for approval."
),
)
assert scan_approval_history(path, days=0) == []
def test_days_window_filters_old_history(self, db_path):
path, con = db_path
old_ts = time.time() - 200 * 86400
_add_terminal_call(con, "git push --force origin main", ts=old_ts)
_add_terminal_call(con, "git push --force origin main")
assert len(scan_approval_history(path, days=90)) == 1
assert len(scan_approval_history(path, days=0)) == 2
def test_missing_db_returns_empty(self, tmp_path):
assert scan_approval_history(tmp_path / "nope.db", days=0) == []
# ---------------------------------------------------------------------------
# Normalize / glob derivation
# ---------------------------------------------------------------------------
class TestNormalizeAndGlob:
def test_normalize_folds_home_prefix(self):
import os
home = os.path.expanduser("~")
normalized = normalize_command(f"git checkout -- {home}/project/file.txt")
assert "~/project/file.txt" in normalized
assert home not in normalized
def test_derive_glob_uses_first_two_tokens(self):
assert derive_glob("git push --force origin main") == "git push *"
assert derive_glob("docker restart web") == "docker restart *"
def test_derive_glob_flag_second_token_falls_back_to_root(self):
assert derive_glob("hermes --yolo update") == "hermes --yolo".split()[0] + " *"
def test_derive_glob_rejects_compound_commands(self):
assert derive_glob("git push --force && rm -rf /tmp/x") is None
assert derive_glob("echo hi; docker restart web") is None
def test_derive_glob_never_anchors_unsafe_binaries(self):
assert derive_glob("rm -rf ./build") is None
assert derive_glob("sudo apt install foo") is None
assert derive_glob("/bin/rm -rf ./build") is None
assert derive_glob("mkfs.ext4 /dev/sdb1") is None
# ---------------------------------------------------------------------------
# Ranking + safety exclusion
# ---------------------------------------------------------------------------
class TestRankingAndSafety:
def test_ranking_orders_by_frequency(self, db_path):
path, con = db_path
for _ in range(5):
_add_terminal_call(con, "docker restart web")
for _ in range(2):
_add_terminal_call(con, "git push --force origin main")
records = scan_approval_history(path, days=0)
proposals = build_proposals(records, min_count=2)
assert [p.pattern for p in proposals] == ["docker restart *", "git push *"]
assert proposals[0].count == 5
assert proposals[1].count == 2
def test_min_count_threshold(self, db_path):
path, con = db_path
_add_terminal_call(con, "git push --force origin main")
records = scan_approval_history(path, days=0)
assert build_proposals(records, min_count=2) == []
assert len(build_proposals(records, min_count=1)) == 1
def test_rm_rf_never_proposed_even_after_100_approvals(self, db_path):
path, con = db_path
for _ in range(100):
_add_terminal_call(con, "rm -rf ./build")
records = scan_approval_history(path, days=0)
# The commands ARE mined (they ran with approval) ...
assert len(records) == 100
# ... but the destructive class is unconditionally excluded.
proposals = build_proposals(records, min_count=1)
assert proposals == []
def test_unsafe_classes_are_excluded(self):
for desc in (
"recursive delete",
"git reset --hard (destroys uncommitted changes)",
"sudo with privilege flag (stdin/askpass/shell/list)",
"pipe remote content to shell",
"overwrite system config",
"SQL DROP",
"in-place edit of sensitive credential/SSH/shell-rc path",
"format filesystem",
"kill all processes",
"write to block device",
):
assert is_unsafe_class(desc), desc
def test_benign_classes_are_not_excluded(self):
for desc in (
"git force push (rewrites remote history)",
"docker restart/stop/kill (container lifecycle)",
"hermes update (restarts gateway, kills running agents)",
"stop/restart system service",
):
assert not is_unsafe_class(desc), desc
def test_existing_allowlist_entries_are_skipped(self, db_path):
path, con = db_path
for _ in range(3):
_add_terminal_call(con, "git push --force origin main")
records = scan_approval_history(path, days=0)
proposals = build_proposals(records, existing={"git push *"}, min_count=1)
assert proposals == []
def test_hardline_commands_never_mined(self, db_path):
path, con = db_path
# Even if a hardline command somehow shows an executed result in the
# DB, the miner refuses it (defense in depth).
for _ in range(5):
_add_terminal_call(con, "rm -rf /")
assert scan_approval_history(path, days=0) == []
# ---------------------------------------------------------------------------
# --apply / dry-run
# ---------------------------------------------------------------------------
def _args(db, **kw):
base = dict(
db=str(db), days=0, min_count=1, limit=20,
apply_indices=None, json=False,
)
base.update(kw)
return Namespace(**base)
class TestApply:
def test_parse_apply_indices(self):
assert parse_apply_indices("1,3", 5) == [0, 2]
assert parse_apply_indices(" 2 ", 2) == [1]
with pytest.raises(ValueError):
parse_apply_indices("0", 3)
with pytest.raises(ValueError):
parse_apply_indices("4", 3)
with pytest.raises(ValueError):
parse_apply_indices("a,b", 3)
with pytest.raises(ValueError):
parse_apply_indices("", 3)
def test_apply_merges_and_persists(self, isolated_allowlist):
isolated_allowlist["patterns"] = {"podman *"}
proposals = [
Proposal(pattern="git push *", kind="glob", count=5),
Proposal(pattern="docker restart *", kind="glob", count=3),
]
merged = apply_proposals(proposals, [0])
assert merged == {"podman *", "git push *"}
assert isolated_allowlist["patterns"] == {"podman *", "git push *"}
assert isolated_allowlist["saves"] == 1
# In-process allowlist synced too (long-lived process consistency).
assert "git push *" in approval_module._permanent_approved
def test_suggest_apply_end_to_end(self, db_path, isolated_allowlist, capsys):
path, con = db_path
for _ in range(4):
_add_terminal_call(con, "git push --force origin main")
for _ in range(3):
_add_terminal_call(con, "docker restart web")
rc = suggest_command(_args(path, apply_indices="1,2"))
assert rc == 0
assert isolated_allowlist["patterns"] == {"git push *", "docker restart *"}
out = capsys.readouterr().out
assert "git push *" in out and "docker restart *" in out
def test_dry_default_writes_nothing(self, db_path, isolated_allowlist, capsys):
path, con = db_path
for _ in range(4):
_add_terminal_call(con, "git push --force origin main")
rc = suggest_command(_args(path))
assert rc == 0
assert isolated_allowlist["saves"] == 0
assert isolated_allowlist["patterns"] == set()
out = capsys.readouterr().out
assert "git push *" in out
assert "Nothing has been changed" in out
def test_apply_out_of_range_errors_without_writing(
self, db_path, isolated_allowlist, capsys
):
path, con = db_path
for _ in range(4):
_add_terminal_call(con, "git push --force origin main")
rc = suggest_command(_args(path, apply_indices="7"))
assert rc == 1
assert isolated_allowlist["saves"] == 0
class TestJsonOutput:
def test_json_proposal_output(self, db_path, isolated_allowlist, capsys):
path, con = db_path
for _ in range(4):
_add_terminal_call(con, "git push --force origin main")
rc = suggest_command(_args(path, json=True))
assert rc == 0
payload = json.loads(capsys.readouterr().out)
assert payload["proposals"][0]["pattern"] == "git push *"
assert payload["proposals"][0]["count"] == 4
assert payload["proposals"][0]["n"] == 1
assert isolated_allowlist["saves"] == 0
def test_json_apply_output(self, db_path, isolated_allowlist, capsys):
path, con = db_path
for _ in range(4):
_add_terminal_call(con, "git push --force origin main")
rc = suggest_command(_args(path, apply_indices="1", json=True))
assert rc == 0
payload = json.loads(capsys.readouterr().out)
assert payload["applied"] == ["git push *"]
assert isolated_allowlist["patterns"] == {"git push *"}
# ---------------------------------------------------------------------------
# Parser wiring
# ---------------------------------------------------------------------------
class TestParserWiring:
def test_build_approvals_parser_wires_suggest(self):
import argparse
from hermes_cli.subcommands.approvals import build_approvals_parser
sentinel = object()
parser = argparse.ArgumentParser()
subparsers = parser.add_subparsers(dest="command")
build_approvals_parser(subparsers, cmd_approvals=lambda a: sentinel)
args = parser.parse_args(
["approvals", "suggest", "--apply", "1,3", "--json",
"--days", "30", "--min-count", "5", "--limit", "10"]
)
assert args.approvals_command == "suggest"
assert args.apply_indices == "1,3"
assert args.json is True
assert args.days == 30
assert args.min_count == 5
assert args.limit == 10
assert args.func(args) is sentinel
def test_suggest_defaults_are_dry_and_bounded(self):
import argparse
from hermes_cli.subcommands.approvals import build_approvals_parser
parser = argparse.ArgumentParser()
subparsers = parser.add_subparsers(dest="command")
build_approvals_parser(subparsers, cmd_approvals=lambda a: None)
args = parser.parse_args(["approvals", "suggest"])
assert args.apply_indices is None
assert args.json is False
assert args.days == 90
assert args.min_count == 2

View file

@ -233,6 +233,43 @@ These patterns are loaded at startup and silently approved in all future session
Use `hermes config edit` to review or remove patterns from your permanent allowlist.
:::
### Mining Approval History (`hermes approvals suggest`)
Instead of answering the same prompt session after session, you can mine your
past approval decisions into allowlist proposals:
```bash
hermes approvals suggest # dry run — prints a numbered proposal
hermes approvals suggest --apply 1,3 # merge picks into command_allowlist
hermes approvals suggest --json # machine-readable output
```
The command scans the session database (`~/.hermes/state.db`) for
dangerous-classified commands that actually executed — i.e. commands you
approved — aggregates them into patterns (`git push *`, or the dangerous-class
key for compound commands), and ranks them by approval frequency:
```
Proposed command_allowlist additions (from approval history, last 90 days):
1. git push * — approved 14x
2. docker restart/stop/kill (container lifecycle) — approved 9x (class key)
```
Safety rules:
- **Nothing is ever applied automatically** — the default run is read-only;
only an explicit `--apply N[,M...]` writes to `config.yaml`.
- **Destructive classes are never proposed**, no matter how often they were
approved: recursive deletes, `sudo`, disk/device writes, credential and
system-config edits, pipe-to-shell, SQL DROP/TRUNCATE, process kills, and
every hardline class are excluded outright. `rm -rf build/` approved 100
times still never yields an `rm` entry.
- Proposals already covered by your existing `command_allowlist` are skipped.
Useful flags: `--days N` (history window, default 90), `--min-count N`
(minimum approvals to qualify, default 2), `--limit N`, and `--db PATH`.
## File Write Safety {#file-write-safety}
Before `write_file` or `patch` touches disk, Hermes checks the target path against a denylist and an optional sandbox. Blocked writes return an error to the agent immediately — **there is no approval prompt** and no way to override from the chat UI. The model may still claim the edit succeeded; when `display.file_mutation_verifier` is on (default), trust the [file-mutation verifier footer](./configuration.md#file-mutation-verifier) over the assistant's closing summary.