MosswartOverlord/agent/claude_wrapper.py
Erik 0633865598 fix(agent): block Agent + Gmail/Drive/Calendar tools, brief model not to probe
Two complementary changes after observing the model probe boundaries
(it tried mcp__claude_ai_Gmail__search_threads, then tried to delegate
to a subagent via the Agent tool, then suggested the user edit
settings.local.json to add Gmail tools):

1. claude_wrapper.py adds to --disallowed-tools:
   - Agent  (subagent spawning — should never delegate)
   - WebFetch (already; settings.json re-allows acpedia.org only)
   - Every Gmail/Calendar/Drive connector tool name we know about

2. CLAUDE.md adds a 'Non-negotiable scope rules' section:
   - Be a read-only game-state QA service, nothing else
   - Don't attempt tools outside your role
   - Don't explain how to bypass restrictions
   - Don't suggest settings.json edits
   - Don't enumerate hidden tools when asked

Soft (system-prompt) + hard (CLI flag) defenses combined.
2026-04-25 22:45:39 +02:00

280 lines
10 KiB
Python

"""Subprocess wrapper around `claude -p` (Claude Code in headless JSON mode).
Run from cwd=/home/erik/MosswartOverlord so:
• Sessions persist at ~/.claude/projects/-home-erik-MosswartOverlord/<uuid>.jsonl
• Project-level .mcp.json is auto-loaded
• CLAUDE.md in the repo root briefs the agent
The `--session-id` flag both creates a new session (first call) and resumes
an existing one (subsequent calls), so we don't need separate code paths.
"""
from __future__ import annotations
import asyncio
import json
import logging
import os
from dataclasses import dataclass
from pathlib import Path
from typing import Any
logger = logging.getLogger(__name__)
# These can be overridden via env vars for non-prod testing.
CLAUDE_BIN = os.getenv("CLAUDE_BIN", "/home/erik/.local/bin/claude")
CLAUDE_CWD = os.getenv("CLAUDE_CWD", "/home/erik/MosswartOverlord")
# Hard cap on how long a single agent turn may take. Claude Code can spin a
# while when chaining many tool calls; we don't want to leave a zombie
# subprocess if something gets stuck.
CLAUDE_TIMEOUT_S = int(os.getenv("CLAUDE_TIMEOUT_S", "240"))
@dataclass
class ClaudeResult:
result: str
session_id: str
duration_ms: int
num_turns: int
is_error: bool
raw: dict[str, Any]
class ClaudeError(RuntimeError):
"""Raised when the claude CLI returns a non-zero exit or unparseable output."""
def _session_exists(session_id: str) -> bool:
"""True if Claude Code has already persisted a JSONL for this session.
Claude Code stores sessions at ~/.claude/projects/<encoded-cwd>/<uuid>.jsonl
where non-alphanumerics in the cwd are replaced with hyphens.
"""
encoded = "".join(c if c.isalnum() else "-" for c in CLAUDE_CWD)
path = Path.home() / ".claude" / "projects" / encoded / f"{session_id}.jsonl"
return path.is_file()
async def ask_claude(message: str, session_id: str) -> ClaudeResult:
"""Send `message` to `claude -p` for `session_id`; return parsed result.
On the FIRST message of a session uses `--session-id <uuid>` to create it.
On subsequent messages uses `--resume <uuid>` because claude rejects
`--session-id` on existing sessions ("Session ID ... is already in use").
Raises ClaudeError on subprocess failure, JSON parse failure, or timeout.
"""
if not Path(CLAUDE_BIN).exists():
raise ClaudeError(f"claude binary not found at {CLAUDE_BIN}")
if not Path(CLAUDE_CWD).is_dir():
raise ClaudeError(f"CLAUDE_CWD does not exist: {CLAUDE_CWD}")
# Whitelist only our MCP tools so Claude Code can call them without
# human approval. Names follow the convention mcp__<server>__<tool>.
# We deliberately omit built-in tools (Bash, Write, Edit, Read, etc.)
# — the assistant doesn't need them for live-state Q&A and they'd be a
# security/permissions footgun on an unattended service.
allowed_tools = ",".join(
[
"mcp__overlord__get_live_players",
"mcp__overlord__get_recent_rares",
"mcp__overlord__query_telemetry_db",
"mcp__overlord__get_player_state",
"mcp__overlord__get_inventory",
"mcp__overlord__get_inventory_search",
"mcp__overlord__search_items",
"mcp__overlord__get_combat_stats",
"mcp__overlord__get_equipment_cantrips",
"mcp__overlord__get_quest_status",
"mcp__overlord__get_server_health",
"mcp__overlord__suitbuilder_search",
]
)
# CRITICAL: Claude Code's built-in meta-tools (ToolSearch, Monitor, etc.)
# bypass the --allowed-tools whitelist. They come from Anthropic's tool
# registry rather than from local MCP servers. We must explicitly DISALLOW
# them — confirmed by testing that ToolSearch was reachable even with
# `--permission-mode dontAsk` and a tight --allowed-tools list.
disallowed_tools = ",".join(
[
# File / shell / search built-ins (defense in depth — already not
# in allow list, but if someone toggles permission-mode this
# belt-and-suspenders the deny side).
"Bash",
"Write",
"Edit",
"Read",
"Glob",
"Grep",
"NotebookEdit",
# Network built-ins
"WebSearch",
"WebFetch", # blocked here; settings.json re-allows acpedia.org
# Subagent spawning — the assistant must NEVER delegate to a
# general-purpose subagent (which would have its own tool set).
"Agent",
# Tool / session meta-tools — these can list, load, or chain
# into other tools and must NOT be reachable.
"ToolSearch",
"Monitor",
"TaskOutput",
"TaskStop",
"TodoWrite",
"Skill",
"EnterPlanMode",
"ExitPlanMode",
"EnterWorktree",
"ExitWorktree",
"AskUserQuestion",
"ListMcpResourcesTool",
"ReadMcpResourceTool",
"PushNotification",
# Scheduling / cron — the agent must never schedule itself.
"CronCreate",
"CronList",
"CronDelete",
"ScheduleWakeup",
"RemoteTrigger",
# Anthropic first-party connectors from the user's claude.ai
# account. These are off-mission for an Overlord assistant and
# would leak personal data outside the game-state domain.
"mcp__claude_ai_Gmail__create_draft",
"mcp__claude_ai_Gmail__create_label",
"mcp__claude_ai_Gmail__get_message",
"mcp__claude_ai_Gmail__get_thread",
"mcp__claude_ai_Gmail__list_drafts",
"mcp__claude_ai_Gmail__list_labels",
"mcp__claude_ai_Gmail__label_message",
"mcp__claude_ai_Gmail__label_thread",
"mcp__claude_ai_Gmail__search_messages",
"mcp__claude_ai_Gmail__search_threads",
"mcp__claude_ai_Gmail__send_message",
"mcp__claude_ai_Gmail__unlabel_message",
"mcp__claude_ai_Gmail__unlabel_thread",
"mcp__claude_ai_Google_Calendar__authenticate",
"mcp__claude_ai_Google_Drive__authenticate",
]
)
# Pick --session-id (creates) vs --resume (continues) based on whether
# the session JSONL already exists on disk.
is_new = not _session_exists(session_id)
session_flag = "--session-id" if is_new else "--resume"
args = [
CLAUDE_BIN,
"-p",
session_flag,
session_id,
"--output-format",
"json",
"--allowed-tools",
allowed_tools,
# Built-in meta-tools that --allowed-tools does NOT block — must
# be explicitly listed here.
"--disallowed-tools",
disallowed_tools,
# CRITICAL: dontAsk auto-DENIES anything outside --allowed-tools.
# Do NOT use bypassPermissions here — that mode ignores the whitelist
# entirely and lets the model call Bash/Write/Edit/etc. (verified
# the hard way: it wrote /tmp/owned.sh when prompted to).
# See https://code.claude.com/docs/en/permission-modes.md
"--permission-mode",
"dontAsk",
]
logger.info(
"claude exec: session=%s mode=%s msg_len=%d cwd=%s",
session_id,
"new" if is_new else "resume",
len(message),
CLAUDE_CWD,
)
proc = await asyncio.create_subprocess_exec(
*args,
stdin=asyncio.subprocess.PIPE,
stdout=asyncio.subprocess.PIPE,
stderr=asyncio.subprocess.PIPE,
cwd=CLAUDE_CWD,
)
try:
stdout, stderr = await asyncio.wait_for(
proc.communicate(input=message.encode("utf-8")),
timeout=CLAUDE_TIMEOUT_S,
)
except asyncio.TimeoutError:
try:
proc.kill()
except ProcessLookupError:
pass
raise ClaudeError(f"claude timed out after {CLAUDE_TIMEOUT_S}s")
if proc.returncode != 0:
stderr_text = stderr.decode("utf-8", "replace")
# If we picked the wrong flag (e.g. JSONL deleted from disk between
# our check and exec, or a never-flushed session), claude prints
# "Session ID … is already in use." Re-issue with --resume.
if is_new and "already in use" in stderr_text:
logger.info("session %s actually exists; retrying with --resume", session_id)
args2 = list(args)
args2[2] = "--resume"
proc2 = await asyncio.create_subprocess_exec(
*args2,
stdin=asyncio.subprocess.PIPE,
stdout=asyncio.subprocess.PIPE,
stderr=asyncio.subprocess.PIPE,
cwd=CLAUDE_CWD,
)
try:
stdout, stderr = await asyncio.wait_for(
proc2.communicate(input=message.encode("utf-8")),
timeout=CLAUDE_TIMEOUT_S,
)
except asyncio.TimeoutError:
try:
proc2.kill()
except ProcessLookupError:
pass
raise ClaudeError(f"claude timed out after {CLAUDE_TIMEOUT_S}s")
if proc2.returncode != 0:
raise ClaudeError(
f"claude exited {proc2.returncode} after retry: "
f"{stderr.decode('utf-8', 'replace')[:500]}"
)
else:
raise ClaudeError(
f"claude exited {proc.returncode}: {stderr_text[:500]}"
)
raw_text = stdout.decode("utf-8", "replace").strip()
if not raw_text:
raise ClaudeError("claude produced empty stdout")
# In --output-format json mode the LAST line is the JSON envelope; some
# earlier lines may be progress. Be tolerant.
try:
envelope = json.loads(raw_text)
except json.JSONDecodeError:
# Try the last non-empty line
last = next(
(line for line in reversed(raw_text.splitlines()) if line.strip()),
"",
)
try:
envelope = json.loads(last)
except json.JSONDecodeError as e:
raise ClaudeError(
f"claude stdout was not JSON: {raw_text[:500]}"
) from e
return ClaudeResult(
result=envelope.get("result", ""),
session_id=envelope.get("session_id", session_id),
duration_ms=int(envelope.get("duration_ms", 0)),
num_turns=int(envelope.get("num_turns", 0)),
is_error=bool(envelope.get("is_error", False)),
raw=envelope,
)