Files
strix/strix/llm/dedupe.py
T
0xallam f4b32ba222 refactor: nuke gratuitous XML serialization + delete argument_parser
Argument parser:
- Delete ``strix/tools/argument_parser.py`` and its tests. The SDK
  validates and types tool arguments via Pydantic before they hit our
  wrappers, and the in-container tool server receives JSON-typed
  kwargs over the wire. The string-coercion belt-and-suspenders is no
  longer pulling its weight.

XML → JSON / typed structures:
- ``create_vulnerability_report``: ``cvss_breakdown`` is now a
  ``dict[str, str]`` of the 8 metrics; ``code_locations`` is a
  ``list[dict]``. No more XML parsing in the tool or the renderer.
- ``check_duplicate``: the dedup judge now emits a single JSON object
  instead of an ``<dedupe_result>`` block. Strict JSON parser handles
  optional code-fence wrappers.
- ``agent_finish``: completion report posted to the parent inbox is a
  JSON object (``kind``, ``from``, ``agent_id``, ``success``,
  ``summary``, ``findings``, ``recommendations``) rather than a
  hand-rolled ``<agent_completion_report>`` XML envelope.
- ``create_agent``: identity preamble + inherited-context markers are
  plain bracketed labels rather than ``<agent_delegation>`` /
  ``<inherited_context_from_parent>`` envelopes.
- ``inject_messages_filter``: peer messages get a
  ``[Message from agent <id> | type=... | priority=...]`` header line
  instead of an ``<inter_agent_message>`` envelope.
- Crash + system-warning messages: bracketed labels, no XML.
- System prompt: the inter-agent block now describes the new header
  format and drops the "never echo XML envelope" rule.
- ``strix/llm/utils.py``: deleted. ``clean_content`` collapsed into a
  one-line blank-line normalizer in the agent-message renderer (the
  XML envelope scrub had nothing left to scrub).

Tests updated to match the new shapes.
2026-04-25 12:31:07 -07:00

200 lines
6.3 KiB
Python

import json
import logging
from typing import Any
import litellm
from strix.config.config import resolve_llm_config
logger = logging.getLogger(__name__)
DEDUPE_SYSTEM_PROMPT = """You are an expert vulnerability report deduplication judge.
Your task is to determine if a candidate vulnerability report describes the SAME vulnerability
as any existing report.
CRITICAL DEDUPLICATION RULES:
1. SAME VULNERABILITY means:
- Same root cause (e.g., "missing input validation" not just "SQL injection")
- Same affected component/endpoint/file (exact match or clear overlap)
- Same exploitation method or attack vector
- Would be fixed by the same code change/patch
2. NOT DUPLICATES if:
- Different endpoints even with same vulnerability type (e.g., SQLi in /login vs /search)
- Different parameters in same endpoint (e.g., XSS in 'name' vs 'comment' field)
- Different root causes (e.g., stored XSS vs reflected XSS in same field)
- Different severity levels due to different impact
- One is authenticated, other is unauthenticated
3. ARE DUPLICATES even if:
- Titles are worded differently
- Descriptions have different level of detail
- PoC uses different payloads but exploits same issue
- One report is more thorough than another
- Minor variations in technical analysis
COMPARISON GUIDELINES:
- Focus on the technical root cause, not surface-level similarities
- Same vulnerability type (SQLi, XSS) doesn't mean duplicate - location matters
- Consider the fix: would fixing one also fix the other?
- When uncertain, lean towards NOT duplicate
FIELDS TO ANALYZE:
- title, description: General vulnerability info
- target, endpoint, method: Exact location of vulnerability
- technical_analysis: Root cause details
- poc_description: How it's exploited
- impact: What damage it can cause
Respond with a single JSON object and nothing else:
{
"is_duplicate": true,
"duplicate_id": "vuln-0001",
"confidence": 0.95,
"reason": "Both reports describe SQL injection in /api/login via the username parameter"
}
Or, if not a duplicate:
{
"is_duplicate": false,
"duplicate_id": "",
"confidence": 0.90,
"reason": "Different endpoints: candidate is /api/search, existing is /api/login"
}
Rules:
- ``is_duplicate`` is a boolean.
- ``duplicate_id`` is the exact id from existing reports, or "" if not a duplicate.
- ``confidence`` is a number between 0 and 1.
- ``reason`` is a specific explanation mentioning endpoint/parameter/root cause.
- Output ONLY the JSON object — no surrounding prose, no code fences."""
def _prepare_report_for_comparison(report: dict[str, Any]) -> dict[str, Any]:
relevant_fields = [
"id",
"title",
"description",
"impact",
"target",
"technical_analysis",
"poc_description",
"endpoint",
"method",
]
cleaned = {}
for field in relevant_fields:
if report.get(field):
value = report[field]
if isinstance(value, str) and len(value) > 8000:
value = value[:8000] + "...[truncated]"
cleaned[field] = value
return cleaned
def _parse_dedupe_response(content: str) -> dict[str, Any]:
text = content.strip()
if text.startswith("```"):
text = text.strip("`")
if text.lower().startswith("json"):
text = text[4:]
text = text.strip()
start = text.find("{")
end = text.rfind("}")
if start == -1 or end == -1 or end <= start:
raise ValueError(f"No JSON object found in dedupe response: {content[:500]}")
parsed = json.loads(text[start : end + 1])
duplicate_id = str(parsed.get("duplicate_id") or "")[:64]
reason = str(parsed.get("reason") or "")[:500]
try:
confidence = float(parsed.get("confidence", 0.0))
except (TypeError, ValueError):
confidence = 0.0
return {
"is_duplicate": bool(parsed.get("is_duplicate", False)),
"duplicate_id": duplicate_id,
"confidence": confidence,
"reason": reason,
}
def check_duplicate(
candidate: dict[str, Any], existing_reports: list[dict[str, Any]]
) -> dict[str, Any]:
if not existing_reports:
return {
"is_duplicate": False,
"duplicate_id": "",
"confidence": 1.0,
"reason": "No existing reports to compare against",
}
try:
candidate_cleaned = _prepare_report_for_comparison(candidate)
existing_cleaned = [_prepare_report_for_comparison(r) for r in existing_reports]
comparison_data = {"candidate": candidate_cleaned, "existing_reports": existing_cleaned}
model_name, api_key, api_base = resolve_llm_config()
litellm_model: str | None = model_name
messages = [
{"role": "system", "content": DEDUPE_SYSTEM_PROMPT},
{
"role": "user",
"content": (
f"Compare this candidate vulnerability against existing reports:\n\n"
f"{json.dumps(comparison_data, indent=2)}\n\n"
f"Respond with ONLY the JSON object described in the system prompt."
),
},
]
completion_kwargs: dict[str, Any] = {
"model": litellm_model,
"messages": messages,
"timeout": 120,
}
if api_key:
completion_kwargs["api_key"] = api_key
if api_base:
completion_kwargs["api_base"] = api_base
response = litellm.completion(**completion_kwargs)
content = response.choices[0].message.content
if not content:
return {
"is_duplicate": False,
"duplicate_id": "",
"confidence": 0.0,
"reason": "Empty response from LLM",
}
result = _parse_dedupe_response(content)
logger.info(
f"Deduplication check: is_duplicate={result['is_duplicate']}, "
f"confidence={result['confidence']}, reason={result['reason'][:100]}"
)
except Exception as e:
logger.exception("Error during vulnerability deduplication check")
return {
"is_duplicate": False,
"duplicate_id": "",
"confidence": 0.0,
"reason": f"Deduplication check failed: {e}",
"error": str(e),
}
else:
return result