Skip to content

armgit / kong

Public Imported from GitHub
0

The world's first agentic reverse engineer.

rm inline imports

Amruth3ef303bb5d85
diff --git a/kong/__main__.py b/kong/__main__.py
index db12a48..7731c1a 100644
--- a/kong/__main__.py
+++ b/kong/__main__.py
@@ -19,6 +19,11 @@ from kong.config import GhidraConfig, KongConfig, LLMConfig, LLMProvider, Output
from kong.evals.harness import score as eval_score
from kong.ghidra.client import GhidraClient, GhidraClientError
from kong.llm.usage import TokenUsage
+from kong.banner import _ENV_VARS, _KEY_EXAMPLES, _KEY_URLS
+from kong.llm.openai_client import OpenAIClient
+from kong.llm.client import AnthropicClient
+from kong.tui.app import KongApp
+
if TYPE_CHECKING:
from kong.agent.analyzer import LLMClient
@@ -37,9 +42,7 @@ def create_llm_client(config: LLMConfig) -> LLMClient:
# will generalize further in the future
model = config.model or _DEFAULT_MODELS[config.provider]
if config.provider is LLMProvider.OPENAI:
- from kong.llm.openai_client import OpenAIClient
return OpenAIClient(model=model, api_key=config.api_key)
- from kong.llm.client import AnthropicClient
return AnthropicClient(model=model, api_key=config.api_key)
@@ -204,7 +207,6 @@ def analyze(
_print_final_stats(supervisor, llm_client)
client.close()
else:
- from kong.tui.app import KongApp
app = KongApp(supervisor)
try:
app.run()
@@ -262,8 +264,6 @@ def setup() -> None:
print_banner(console)
console.print()
- from kong.banner import _ENV_VARS, _KEY_EXAMPLES, _KEY_URLS
-
step = 1
for provider in LLMProvider:
env_var = _ENV_VARS[provider]
diff --git a/kong/agent/analyzer.py b/kong/agent/analyzer.py
index 977d70c..ad4a3c8 100644
--- a/kong/agent/analyzer.py
+++ b/kong/agent/analyzer.py
@@ -19,6 +19,7 @@ from kong.agent.queue import WorkItem
from kong.ghidra.client import GhidraClient
from kong.ghidra.types import BinaryInfo, FunctionInfo, StringEntry, StructDefinition
from kong.normalizer.syntactic import normalize
+from kong.agent.deobfuscator import classify_obfuscation
if TYPE_CHECKING:
from kong.agent.deobfuscator import Deobfuscator
@@ -171,7 +172,6 @@ class Analyzer:
model: str | None = None,
) -> FunctionResult:
"""Full analysis pipeline for one function."""
- from kong.agent.deobfuscator import classify_obfuscation
func = item.function
context = self._build_context(item, binary_info, known_results, strings, known_types)
diff --git a/kong/evals/metrics.py b/kong/evals/metrics.py
index 6ddc726..70c02d1 100644
--- a/kong/evals/metrics.py
+++ b/kong/evals/metrics.py
@@ -4,6 +4,8 @@ from __future__ import annotations
import re
+from kong.evals.harness import match_functions
+
_SYNONYM_GROUPS: list[set[str]] = [
{"search", "find", "lookup"},
@@ -189,7 +191,6 @@ def overall_score(
truth: list[dict[str, str]],
) -> dict[str, float]:
"""Compute aggregate symbol and type accuracy across matched functions."""
- from kong.evals.harness import match_functions
matches = match_functions(predicted, truth)
diff --git a/kong/llm/openai_client.py b/kong/llm/openai_client.py
index f8b9d6b..83b1f51 100644
--- a/kong/llm/openai_client.py
+++ b/kong/llm/openai_client.py
@@ -144,9 +144,9 @@ class OpenAIClient:
self._record_usage(response, self.model)
choice = response.choices[0]
+ last_text = choice.message.content or ""
if choice.finish_reason != "tool_calls":
- last_text = choice.message.content or ""
result = Analyzer.parse_llm_json(last_text)
result.input_tokens = total_input
result.output_tokens = total_output
diff --git a/tests/test_deobfuscator.py b/tests/test_deobfuscator.py
index 50797ce..60da23d 100644
--- a/tests/test_deobfuscator.py
+++ b/tests/test_deobfuscator.py
@@ -19,6 +19,11 @@ from kong.agent.deobfuscator import (
)
from kong.agent.models import FunctionResult
from kong.ghidra.types import BinaryInfo, FunctionInfo
+from kong.agent.prompts import DEOBFUSCATION_SYSTEM_PROMPT
+from kong.llm.tools import DEOBFUSCATION_TOOLS
+from kong.agent.analyzer import Analyzer
+from kong.agent.queue import WorkItem
+from kong.agent.deobfuscator import Deobfuscator
_CFF_CODE = """\
@@ -218,8 +223,6 @@ class TestDeobfuscator:
assert "bogus_cf" in prompt
def test_passes_deobfuscation_system_prompt(self, mock_ghidra, mock_llm):
- from kong.agent.prompts import DEOBFUSCATION_SYSTEM_PROMPT
-
deob = Deobfuscator(mock_ghidra, mock_llm)
context = _make_context()
deob.deobfuscate(context, [ObfuscationType.BOGUS_CONTROL_FLOW])
@@ -228,8 +231,6 @@ class TestDeobfuscator:
assert system == DEOBFUSCATION_SYSTEM_PROMPT
def test_passes_tools(self, mock_ghidra, mock_llm):
- from kong.llm.tools import DEOBFUSCATION_TOOLS
-
deob = Deobfuscator(mock_ghidra, mock_llm)
context = _make_context()
deob.deobfuscate(context, [ObfuscationType.INSTRUCTION_SUBSTITUTION])
@@ -240,9 +241,6 @@ class TestDeobfuscator:
class TestAnalyzerDeobfuscationIntegration:
def test_obfuscated_function_delegates(self):
- from kong.agent.analyzer import Analyzer
- from kong.agent.queue import WorkItem
-
mock_ghidra = MagicMock()
mock_ghidra.get_decompilation.return_value = _CFF_CODE
mock_ghidra.get_xrefs_from.return_value = []
@@ -254,7 +252,6 @@ class TestAnalyzerDeobfuscationIntegration:
classification="crypto",
)
- from kong.agent.deobfuscator import Deobfuscator
deob = Deobfuscator(mock_ghidra, mock_llm)
analyzer = Analyzer(mock_ghidra, mock_llm, deobfuscator=deob)
@@ -277,8 +274,6 @@ class TestAnalyzerDeobfuscationIntegration:
mock_llm.analyze_with_tools.assert_called_once()
def test_clean_function_uses_normal_path(self):
- from kong.agent.analyzer import Analyzer
- from kong.agent.queue import WorkItem
mock_ghidra = MagicMock()
mock_ghidra.get_decompilation.return_value = _CLEAN_CODE
@@ -291,7 +286,6 @@ class TestAnalyzerDeobfuscationIntegration:
classification="math",
)
- from kong.agent.deobfuscator import Deobfuscator
deob = Deobfuscator(mock_ghidra, mock_llm)
analyzer = Analyzer(mock_ghidra, mock_llm, deobfuscator=deob)
diff --git a/tests/test_llm_tools.py b/tests/test_llm_tools.py
index fd4bc8f..d64bd0a 100644
--- a/tests/test_llm_tools.py
+++ b/tests/test_llm_tools.py
@@ -12,6 +12,7 @@ from kong.llm.tools import (
ToolCallRecord,
ToolExecutor,
)
+from kong.llm.client import AnthropicClient
class TestToolSchemas:
@@ -135,7 +136,6 @@ class TestToolUseLoop:
return msg
def test_no_tool_calls(self):
- from kong.llm.client import AnthropicClient
json_response = json.dumps({
"name": "simple_func",
@@ -163,7 +163,6 @@ class TestToolUseLoop:
assert response.confidence == 85
def test_single_tool_call(self):
- from kong.llm.client import AnthropicClient
json_response = json.dumps({
"name": "deobfuscated_func",
@@ -198,7 +197,6 @@ class TestToolUseLoop:
assert mock_client.messages.create.call_count == 2
def test_max_rounds_limit(self):
- from kong.llm.client import AnthropicClient
tool_msg = self._mock_tool_use_message(
"simplify_expression",
diff --git a/tests/test_openai_client.py b/tests/test_openai_client.py
index 36fcaa5..99e96d9 100644
--- a/tests/test_openai_client.py
+++ b/tests/test_openai_client.py
@@ -253,6 +253,45 @@ class TestOpenAIToolUse:
assert result.output_tokens == 60
executor.execute.assert_called_once_with("simplify_expression", {"expression": "x + 0"})
+ @patch("kong.llm.openai_client.openai.OpenAI")
+ def test_max_rounds_exhausted_uses_last_content(self, mock_openai_cls):
+ mock_client = MagicMock()
+ mock_openai_cls.return_value = mock_client
+
+ tool_call = MagicMock()
+ tool_call.id = "call_abc"
+ tool_call.function.name = "lookup"
+ tool_call.function.arguments = '{"key": "val"}'
+
+ def make_tool_response(content: str | None):
+ resp = _mock_response("", prompt_tokens=50, completion_tokens=20)
+ resp.choices[0].finish_reason = "tool_calls"
+ resp.choices[0].message.content = content
+ resp.choices[0].message.tool_calls = [tool_call]
+ return resp
+
+ mock_client.chat.completions.create.side_effect = [
+ make_tool_response(None),
+ make_tool_response('{"name": "last_round", "confidence": 60}'),
+ ]
+
+ executor = MagicMock()
+ executor.execute.return_value = '{"result": "ok"}'
+
+ tools = [{
+ "name": "lookup",
+ "description": "Look up a value.",
+ "input_schema": {"type": "object", "properties": {"key": {"type": "string"}}},
+ }]
+
+ client = OpenAIClient(api_key="test-key")
+ result = client.analyze_with_tools("prompt", "system", tools, executor, max_rounds=2)
+
+ assert result.name == "last_round"
+ assert result.confidence == 60
+ assert result.input_tokens == 100
+ assert result.output_tokens == 40
+
class TestConvertTools:
def test_converts_anthropic_to_openai_format(self):
diff --git a/tests/test_synthesis.py b/tests/test_synthesis.py
index 641ef6e..a09b2b2 100644
--- a/tests/test_synthesis.py
+++ b/tests/test_synthesis.py
@@ -7,7 +7,7 @@ from dataclasses import dataclass
import pytest
from kong.agent.models import FunctionResult
-from kong.synthesis.semantic import SemanticSynthesizer, SynthesisResult
+from kong.synthesis.semantic import SemanticSynthesizer, SynthesisResult, SYNTHESIS_FUNCTION_CAP
@dataclass
@@ -183,7 +183,6 @@ Trailing text'''
class TestSynthesisCap:
def test_synthesis_caps_at_50_functions(self) -> None:
- from kong.synthesis.semantic import SYNTHESIS_FUNCTION_CAP
results = [_make_result(i, f"func_{i}") for i in range(100)]
decomps = {i: f"void func_{i}(void) {{ DAT_1000 = 1; }}" for i in range(100)}
diff --git a/tests/test_type_recovery.py b/tests/test_type_recovery.py
index 8dd916d..a31a36e 100644
--- a/tests/test_type_recovery.py
+++ b/tests/test_type_recovery.py
@@ -27,6 +27,10 @@ from kong.ghidra.types import (
StructDefinition,
StructField,
)
+from kong.agent.events import EventType
+from kong.agent.supervisor import Supervisor
+from kong.config import KongConfig, OutputConfig
+from kong.agent.queue import WorkItem
def _field(name: str, dtype: str, offset: int, size: int) -> StructFieldProposal:
@@ -331,7 +335,6 @@ class TestAnalyzerContextTypes:
)
client.get_xrefs_from.return_value = []
- from kong.agent.queue import WorkItem
item = WorkItem(
function=FunctionInfo(
address=0x1000, name="FUN_00001000", size=100,
@@ -374,7 +377,6 @@ class TestAnalyzerContextTypes:
)
client.get_xrefs_from.return_value = []
- from kong.agent.queue import WorkItem
item = WorkItem(
function=FunctionInfo(
address=0x1000, name="f", size=100,
@@ -418,7 +420,6 @@ class TestAnalyzerStructProposalPassthrough:
)
client.get_xrefs_from.return_value = []
- from kong.agent.queue import WorkItem
item = WorkItem(
function=FunctionInfo(
address=0x1000, name="f", size=100,
@@ -440,10 +441,6 @@ class TestAnalyzerStructProposalPassthrough:
class TestSupervisorCleanupIntegration:
def test_cleanup_with_struct_proposals(self, tmp_path):
- from kong.agent.events import EventType, Phase
- from kong.agent.supervisor import Supervisor
- from kong.config import KongConfig, OutputConfig
-
funcs = [
FunctionInfo(
address=0x1000, name="process", size=100,
@@ -495,10 +492,6 @@ class TestSupervisorCleanupIntegration:
assert EventType.CLEANUP_TYPE_CREATED in event_types
def test_cleanup_without_struct_proposals(self, tmp_path):
- from kong.agent.events import EventType
- from kong.agent.supervisor import Supervisor
- from kong.config import KongConfig, OutputConfig
-
funcs = [
FunctionInfo(
address=0x1000, name="f", size=100,
@@ -535,10 +528,7 @@ class TestSupervisorCleanupIntegration:
def test_cleanup_retries_failed_signatures(self, tmp_path):
"""Signatures that fail during Phase 2 (type not yet created) get
- retried after struct creation in cleanup."""
- from kong.agent.events import EventType
- from kong.agent.supervisor import Supervisor
- from kong.config import KongConfig, OutputConfig
+ retried after struct creation in cleanup."""
funcs = [
FunctionInfo(