slipstream/discordbot
slipstream/discordbot
1
1
Fork 0
Code Issues Pull Requests 3 Actions Packages Projects Releases Wiki Activity

aa

Browse Source
This commit is contained in:
Slipstream 2025-04-29 10:51:26 -06:00
parent de09605313
commit ac298213a8
Signed by: slipstream
GPG Key ID: 13E498CE010AC6FD
2 changed files with 42 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
gurt/api.py
View File

@ -151,11 +151,12 @@ except Exception as e:
# Use actual types if import succeeded, otherwise fallback to Any
_HarmCategory = getattr(generative_models, 'HarmCategory', Any)
_HarmBlockThreshold = getattr(generative_models, 'HarmBlockThreshold', Any)
# To disable blocking, set all thresholds to BLOCK_NONE (or equivalent).
STANDARD_SAFETY_SETTINGS = {
getattr(_HarmCategory, 'HARM_CATEGORY_HATE_SPEECH', 'HARM_CATEGORY_HATE_SPEECH'): getattr(_HarmBlockThreshold, 'BLOCK_MEDIUM_AND_ABOVE', 'BLOCK_MEDIUM_AND_ABOVE'),
getattr(_HarmCategory, 'HARM_CATEGORY_DANGEROUS_CONTENT', 'HARM_CATEGORY_DANGEROUS_CONTENT'): getattr(_HarmBlockThreshold, 'BLOCK_MEDIUM_AND_ABOVE', 'BLOCK_MEDIUM_AND_ABOVE'),
getattr(_HarmCategory, 'HARM_CATEGORY_SEXUALLY_EXPLICIT', 'HARM_CATEGORY_SEXUALLY_EXPLICIT'): getattr(_HarmBlockThreshold, 'BLOCK_MEDIUM_AND_ABOVE', 'BLOCK_MEDIUM_AND_ABOVE'),
getattr(_HarmCategory, 'HARM_CATEGORY_HARASSMENT', 'HARM_CATEGORY_HARASSMENT'): getattr(_HarmBlockThreshold, 'BLOCK_MEDIUM_AND_ABOVE', 'BLOCK_MEDIUM_AND_ABOVE'),
getattr(_HarmCategory, 'HARM_CATEGORY_HATE_SPEECH', 'HARM_CATEGORY_HATE_SPEECH'): getattr(_HarmBlockThreshold, 'BLOCK_NONE', 'BLOCK_NONE'),
getattr(_HarmCategory, 'HARM_CATEGORY_DANGEROUS_CONTENT', 'HARM_CATEGORY_DANGEROUS_CONTENT'): getattr(_HarmBlockThreshold, 'BLOCK_NONE', 'BLOCK_NONE'),
getattr(_HarmCategory, 'HARM_CATEGORY_SEXUALLY_EXPLICIT', 'HARM_CATEGORY_SEXUALLY_EXPLICIT'): getattr(_HarmBlockThreshold, 'BLOCK_NONE', 'BLOCK_NONE'),
getattr(_HarmCategory, 'HARM_CATEGORY_HARASSMENT', 'HARM_CATEGORY_HARASSMENT'): getattr(_HarmBlockThreshold, 'BLOCK_NONE', 'BLOCK_NONE'),
}
# --- API Call Helper ---
@ -203,7 +204,23 @@ async def call_vertex_api_with_retry(
tool_config=tool_config # Pass tool_config here
)
# --- Success Logging ---
# --- Check Finish Reason (Safety) ---
# This check is primarily for non-streaming responses where a single finish_reason is available.
if not stream and response and response.candidates:
candidate = response.candidates[0]
# Ensure FinishReason is accessible (it should be if vertexai imported correctly)
_FinishReason = globals().get('FinishReason')
if _FinishReason and candidate.finish_reason == _FinishReason.SAFETY:
safety_ratings_str = ", ".join([f"{rating.category}: {rating.probability.name}" for rating in candidate.safety_ratings]) if candidate.safety_ratings else "N/A"
print(f"⚠️ SAFETY BLOCK: API request for {request_desc} ({model_name}) was blocked by safety filters. Finish Reason: SAFETY. Ratings: [{safety_ratings_str}]")
# Optionally, raise a specific exception here if needed downstream
# raise SafetyBlockError(f"Blocked by safety filters. Ratings: {safety_ratings_str}")
elif _FinishReason and candidate.finish_reason != _FinishReason.STOP and candidate.finish_reason != _FinishReason.MAX_TOKENS and candidate.finish_reason != _FinishReason.FUNCTION_CALL:
# Log other unexpected finish reasons
print(f"⚠️ UNEXPECTED FINISH REASON: API request for {request_desc} ({model_name}) finished with reason: {candidate.finish_reason.name}")
# --- Success Logging (Proceed even if safety blocked, but log occurred) ---
elapsed_time = time.monotonic() - start_time
# Ensure model_name exists in stats before incrementing
if model_name not in cog.api_stats:

25
gurt/tools.py
View File

@ -620,12 +620,25 @@ async def _check_command_safety(cog: commands.Cog, command: str) -> Dict[str, An
"type": "object",
"properties": {
"is_safe": {"type": "boolean", "description": "True if safe for restricted container, False otherwise."},
"reason": {"type": "string", "description": "Brief explanation."}
"reason": {"type": "string", "description": "Brief explanation why the command is safe or unsafe."}
}, "required": ["is_safe", "reason"]
}
# Enhanced system prompt with more examples of safe commands
system_prompt_content = (
f"Analyze shell command safety for execution in an isolated, network-disabled Docker container ({DOCKER_EXEC_IMAGE}) "
f"with CPU ({DOCKER_CPU_LIMIT} core) and Memory ({DOCKER_MEM_LIMIT}) limits. "
"Focus on preventing: data destruction (outside container's ephemeral storage), resource exhaustion (fork bombs, crypto mining), "
"container escape vulnerabilities, network attacks (network is disabled), sensitive environment variable leakage (assume only safe vars are mounted). "
"Commands that only read system info, list files, print text, or manipulate text are generally SAFE. "
"Examples of SAFE commands: whoami, id, uname, hostname, pwd, ls, echo, cat, grep, sed, awk, date, time, env, df, du, ps, top, htop, find (read-only), file. "
"Examples of UNSAFE commands: rm, mkfs, shutdown, reboot, poweroff, wget, curl, apt, yum, apk, pip install, npm install, git clone (network disabled, but still potentially risky), "
"any command trying to modify system files outside /tmp or /home, fork bombs like ':(){ :|:& };:', commands enabling network access. "
"Be cautious with file writing/modification commands if not clearly limited to temporary directories. "
"Respond ONLY with the raw JSON object matching the provided schema."
)
prompt_messages = [
{"role": "system", "content": f"Analyze shell command safety for execution in isolated, network-disabled Docker ({DOCKER_EXEC_IMAGE}) with CPU/Mem limits. Focus on data destruction, resource exhaustion, container escape, network attacks (disabled), env var leaks. Simple echo/ls/pwd safe. rm/mkfs/shutdown/wget/curl/install/fork bombs unsafe. Respond ONLY with JSON matching the provided schema."},
{"role": "user", "content": f"Analyze safety: ```{command}```"}
{"role": "system", "content": system_prompt_content},
{"role": "user", "content": f"Analyze safety of this command: ```\n{command}\n```"}
]
safety_response = await get_internal_ai_json_response(
cog=cog,
@ -642,8 +655,10 @@ async def _check_command_safety(cog: commands.Cog, command: str) -> Dict[str, An
print(f"AI Safety Check Result: is_safe={is_safe}, reason='{reason}'")
return {"safe": is_safe, "reason": reason}
else:
error_msg = "AI safety check failed or returned invalid format."
print(f"AI Safety Check Error: Response was {safety_response}")
# Include part of the invalid response in the error for debugging
raw_response_excerpt = str(safety_response)[:200] # Get first 200 chars
error_msg = f"AI safety check failed or returned invalid format. Response: {raw_response_excerpt}"
print(f"AI Safety Check Error: {error_msg}")
return {"safe": False, "reason": error_msg}
async def run_terminal_command(cog: commands.Cog, command: str) -> Dict[str, Any]: