The world's first agentic reverse engineer.
refactor: fix inline imports, move AnalysisStats to models, clean up init files
Amruth15c0050ab8fc
diff --git a/kong/__main__.py b/kong/__main__.pyindex fe0e7d9..0dcb652 100644--- a/kong/__main__.py+++ b/kong/__main__.py@@ -151,6 +151,7 @@ def analyze(_print_final_stats(supervisor, llm_client)client.close()else:+ # Deferred to avoid loading Textual when running --headlessfrom kong.tui.app import KongAppapp = KongApp(supervisor)try:diff --git a/kong/agent/__init__.py b/kong/agent/__init__.pyindex 1adc5c9..f48c55c 100644--- a/kong/agent/__init__.py+++ b/kong/agent/__init__.py@@ -5,7 +5,8 @@ from kong.agent.events import Event, EventType, Phasefrom kong.agent.models import FunctionResultfrom kong.agent.queue import WorkItem, WorkQueuefrom kong.agent.signatures import SignatureDB, SignatureMatch-from kong.agent.supervisor import AnalysisStats, Supervisor+from kong.agent.models import AnalysisStats+from kong.agent.supervisor import Supervisorfrom kong.agent.triage import CallGraph, LanguageHints, TriageAgent, TriageResult__all__ = [diff --git a/kong/agent/analyzer.py b/kong/agent/analyzer.pyindex b1fed84..ed130da 100644--- a/kong/agent/analyzer.py+++ b/kong/agent/analyzer.py@@ -10,13 +10,17 @@ from __future__ import annotationsimport jsonimport loggingfrom dataclasses import dataclass, field-from typing import Protocol+from typing import TYPE_CHECKING, Protocolfrom kong.agent.models import FunctionResultfrom kong.agent.queue import WorkItemfrom kong.ghidra.client import GhidraClientfrom kong.ghidra.types import BinaryInfo, FunctionInfo, StringEntry, StructDefinition+if TYPE_CHECKING:+ from kong.agent.deobfuscator import Deobfuscator+ from kong.llm.tools import ToolExecutor, ToolSchema+logger = logging.getLogger(__name__)@@ -29,8 +33,8 @@ class LLMClient(Protocol):self,prompt: str,system: str,- tools: list[dict[str, object]],- tool_executor: object,+ tools: list[ToolSchema],+ tool_executor: ToolExecutor,max_rounds: int = 10,) -> LLMResponse: ...@@ -115,7 +119,7 @@ class Analyzer:self,client: GhidraClient,llm_client: LLMClient,- deobfuscator: object | None = None,+ deobfuscator: Deobfuscator | None = None,) -> None:self.client = clientself.llm = llm_client@@ -131,14 +135,15 @@ class Analyzer:model: str | None = None,) -> FunctionResult:"""Full analysis pipeline for one function."""- from kong.agent.deobfuscator import Deobfuscator, classify_obfuscation+ # Inline import: deobfuscator imports from analyzer at module level (circular).+ from kong.agent.deobfuscator import classify_obfuscationfunc = item.functioncontext = self._build_context(item, binary_info, known_results, strings, known_types)techniques = classify_obfuscation(context.decompilation) if self._deobfuscator else []- if techniques and isinstance(self._deobfuscator, Deobfuscator):+ if techniques and self._deobfuscator:response, tool_calls = self._deobfuscator.deobfuscate(context, techniques)sig_applied = self._write_back(func.address, response)return FunctionResult(diff --git a/kong/agent/models.py b/kong/agent/models.pyindex 9f5d3df..e5981df 100644--- a/kong/agent/models.py+++ b/kong/agent/models.py@@ -2,6 +2,7 @@from __future__ import annotations+import timefrom dataclasses import dataclass, fieldfrom typing import TYPE_CHECKING@@ -9,6 +10,60 @@ if TYPE_CHECKING:from kong.agent.analyzer import StructProposal+@dataclass+class AnalysisStats:+ """Aggregate statistics for the full run."""+ total_functions: int = 0+ analyzed: int = 0+ renamed: int = 0+ confirmed: int = 0+ high_confidence: int = 0+ medium_confidence: int = 0+ low_confidence: int = 0+ skipped: int = 0+ errors: int = 0+ llm_calls: int = 0+ start_time: float = 0.0+ end_time: float = 0.0+ signature_matches: int = 0++ @property+ def named(self) -> int:+ return self.renamed + self.confirmed++ @property+ def duration_seconds(self) -> float:+ end = self.end_time or time.time()+ return end - self.start_time if self.start_time else 0.0++ @property+ def name_rate(self) -> float:+ return self.named / self.total_functions if self.total_functions else 0.0++ def record_result(self, result: FunctionResult) -> None:+ if result.skipped:+ self.skipped += 1+ return+ if result.error:+ self.errors += 1+ return+ self.analyzed += 1+ self.llm_calls += result.llm_calls+ if result.name:+ if result.name != result.original_name:+ self.renamed += 1+ else:+ self.confirmed += 1+ # TODO: calibrate these eventually. these buckets are arbitrary, need eval data to+ # determine meaningful confidence tiers for the LLM's self-reported scores.+ if result.confidence >= 80:+ self.high_confidence += 1+ elif result.confidence >= 50:+ self.medium_confidence += 1+ else:+ self.low_confidence += 1++@dataclassclass FunctionResult:"""Result of analyzing a single function."""diff --git a/kong/agent/supervisor.py b/kong/agent/supervisor.pyindex 8c5b93e..4872ce1 100644--- a/kong/agent/supervisor.py+++ b/kong/agent/supervisor.py@@ -9,19 +9,21 @@ from __future__ import annotationsimport loggingimport timefrom concurrent.futures import ThreadPoolExecutor, as_completed-from dataclasses import dataclassfrom kong.agent.analyzer import Analyzer, LLMClient, LLMResponsefrom kong.agent.deobfuscator import Deobfuscator, classify_obfuscationfrom kong.agent.events import Event, EventCallback, EventType, Phase-from kong.agent.models import FunctionResult+from kong.agent.models import AnalysisStats, FunctionResultfrom kong.agent.queue import WorkItem, WorkQueuefrom kong.agent.signatures import SignatureDBfrom kong.agent.triage import TriageAgent, TriageResultfrom kong.agent.type_recovery import StructAccumulator, apply_unified_structsfrom kong.config import KongConfig+from kong.export.source import ExportData, export_source+from kong.export.structured import export_jsonfrom kong.ghidra.client import GhidraClientfrom kong.ghidra.types import BinaryInfo, FunctionInfo, StringEntry+from kong.llm.client import TokenUsagelogger = logging.getLogger(__name__)@@ -32,60 +34,6 @@ MAX_PARALLEL_LLM = 4REQUEST_STAGGER_SECONDS = 0.5-@dataclass-class AnalysisStats:- """Aggregate statistics for the full run."""- total_functions: int = 0- analyzed: int = 0- renamed: int = 0- confirmed: int = 0- high_confidence: int = 0 # >= 80- medium_confidence: int = 0 # 50-79- low_confidence: int = 0 # < 50- skipped: int = 0- errors: int = 0- llm_calls: int = 0- start_time: float = 0.0- end_time: float = 0.0- signature_matches: int = 0-- @property- def named(self) -> int:- return self.renamed + self.confirmed-- @property- def duration_seconds(self) -> float:- end = self.end_time or time.time()- return end - self.start_time if self.start_time else 0.0-- @property- def name_rate(self) -> float:- return self.named / self.total_functions if self.total_functions else 0.0-- def record_result(self, result: FunctionResult) -> None:- if result.skipped:- self.skipped += 1- return- if result.error:- self.errors += 1- return- self.analyzed += 1- self.llm_calls += result.llm_calls- if result.name:- if result.name != result.original_name:- self.renamed += 1- else:- self.confirmed += 1- # TODO: calibrate these eventually. these buckets are arbitrary, need eval data to- # determine meaningful confidence tiers for the LLM's self-reported scores.- if result.confidence >= 80:- self.high_confidence += 1- elif result.confidence >= 50:- self.medium_confidence += 1- else:- self.low_confidence += 1--class Supervisor:"""Main agent loop that orchestrates the full analysis pipeline.@@ -125,13 +73,20 @@ class Supervisor:for cb in self._listeners:cb(event)+ @property+ def is_paused(self) -> bool:+ return self._paused+def pause(self) -> None:self._paused = Truedef resume(self) -> None:self._paused = False-+ def export(self) -> None:+ """Trigger export manually (e.g. from TUI keybind)."""+ self._run_export()+def run(self) -> dict[int, FunctionResult]:"""Run the full analysis pipeline. Returns addr -> FunctionResult."""self.stats.start_time = time.time()@@ -658,10 +613,6 @@ class Supervisor:def _run_export(self) -> None:"""Generate output files."""- from kong.export.source import ExportData, export_source- from kong.export.structured import export_json- from kong.llm.client import TokenUsage-self._emit(Event(type=EventType.PHASE_START,phase=Phase.EXPORT,diff --git a/kong/agent/type_recovery.py b/kong/agent/type_recovery.pyindex 9ae79d9..33a85ba 100644--- a/kong/agent/type_recovery.py+++ b/kong/agent/type_recovery.py@@ -90,33 +90,6 @@ class StructAccumulator:groups.append(same_name)return groups- @staticmethod- def _split_by_field_overlap(proposals: list[StructProposal]) -> list[list[StructProposal]]:- """Within same-size proposals, cluster by overlapping field offsets."""- if len(proposals) <= 1:- return [proposals]-- assigned: list[int] = [-1] * len(proposals)- group_id = 0-- for i, pi in enumerate(proposals):- if assigned[i] == -1:- assigned[i] = group_id- group_id += 1-- offsets_i = {f.offset for f in pi.fields}- for j in range(i + 1, len(proposals)):- if assigned[j] != -1:- continue- offsets_j = {f.offset for f in proposals[j].fields}- if offsets_i & offsets_j:- assigned[j] = assigned[i]-- clusters: dict[int, list[StructProposal]] = defaultdict(list)- for idx, gid in enumerate(assigned):- clusters[gid].append(proposals[idx])- return list(clusters.values())-@staticmethoddef _merge_group(group: list[StructProposal]) -> UnifiedStruct:name_counts: dict[str, int] = defaultdict(int)diff --git a/kong/evals/__init__.py b/kong/evals/__init__.pynew file mode 100644index 0000000..e69de29diff --git a/kong/export/__init__.py b/kong/export/__init__.pyindex c5ed44d..8b13789 100644--- a/kong/export/__init__.py+++ b/kong/export/__init__.py@@ -1 +1 @@-"""Export pipeline for Kong analysis results."""+diff --git a/kong/export/source.py b/kong/export/source.pyindex d5cb032..d49f4ad 100644--- a/kong/export/source.py+++ b/kong/export/source.py@@ -4,7 +4,7 @@ from dataclasses import dataclassfrom pathlib import Pathfrom kong.agent.models import FunctionResult-from kong.agent.supervisor import AnalysisStats+from kong.agent.models import AnalysisStatsfrom kong.ghidra.types import BinaryInfofrom kong.llm.client import TokenUsagediff --git a/kong/llm/tools.py b/kong/llm/tools.pyindex 1715287..2edd892 100644--- a/kong/llm/tools.py+++ b/kong/llm/tools.py@@ -13,6 +13,7 @@ from typing import Any, TYPE_CHECKINGfrom kong.symbolic.dead_code import ResolvedPredicate, Resolution, eliminate_dead_codefrom kong.symbolic.simplifier import simplify_expression+from kong.symbolic.state_machine import trace_state_machineif TYPE_CHECKING:from kong.ghidra.client import GhidraClient@@ -199,8 +200,7 @@ class ToolExecutor:def __init__(self, ghidra_client: GhidraClient) -> None:self.client = ghidra_client- self.call_log: list[ToolCallRecord] = field(default_factory=list)- self.call_log = []+ self.call_log: list[ToolCallRecord] = []@propertydef call_count(self) -> int:@@ -262,8 +262,6 @@ class ToolExecutor:return cleaneddef _trace_state_machine(self, inputs: dict[str, Any]) -> str:- from kong.symbolic.state_machine import trace_state_machine-func_addr = inputs["function_address"]state_var = inputs["state_variable"]result = trace_state_machine(self.client, func_addr, state_var)diff --git a/kong/state/__init__.py b/kong/state/__init__.pynew file mode 100644index 0000000..e69de29diff --git a/kong/tui/__init__.py b/kong/tui/__init__.pyindex ebb87db..8b13789 100644--- a/kong/tui/__init__.py+++ b/kong/tui/__init__.py@@ -1 +1 @@-"""Kong TUI package."""+diff --git a/kong/tui/app.py b/kong/tui/app.pyindex d324405..dd698c0 100644--- a/kong/tui/app.py+++ b/kong/tui/app.py@@ -123,11 +123,11 @@ class KongApp(App):elapsed=elapsed,llm_calls=llm_calls,cost=cost,- paused=self.supervisor._paused,+ paused=self.supervisor.is_paused,)def action_toggle_pause(self) -> None:- if self.supervisor._paused:+ if self.supervisor.is_paused:self.supervisor.resume()else:self.supervisor.pause()@@ -140,4 +140,4 @@ class KongApp(App):self.run_worker(self._run_export, thread=True)def _run_export(self) -> None:- self.supervisor._run_export()+ self.supervisor.export()diff --git a/tests/test_agent_supervisor.py b/tests/test_agent_supervisor.pyindex 5d91e4b..9948b5a 100644--- a/tests/test_agent_supervisor.py+++ b/tests/test_agent_supervisor.py@@ -6,7 +6,8 @@ from unittest.mock import MagicMockfrom kong.agent.events import EventType, Phasefrom kong.agent.models import FunctionResult-from kong.agent.supervisor import AnalysisStats, Supervisor+from kong.agent.models import AnalysisStats+from kong.agent.supervisor import Supervisorfrom kong.config import KongConfig, OutputConfigfrom kong.ghidra.types import BinaryInfo, FunctionClassification, FunctionInfodiff --git a/tests/test_export_source.py b/tests/test_export_source.pyindex 6c2ff4e..5ad3298 100644--- a/tests/test_export_source.py+++ b/tests/test_export_source.py@@ -5,7 +5,7 @@ from pathlib import Pathimport pytestfrom kong.agent.models import FunctionResult-from kong.agent.supervisor import AnalysisStats+from kong.agent.models import AnalysisStatsfrom kong.export.source import ExportData, export_sourcefrom kong.ghidra.types import BinaryInfofrom kong.llm.client import TokenUsagediff --git a/tests/test_export_structured.py b/tests/test_export_structured.pyindex 143e52d..dd00d9a 100644--- a/tests/test_export_structured.py+++ b/tests/test_export_structured.py@@ -6,7 +6,7 @@ from pathlib import Pathimport pytestfrom kong.agent.models import FunctionResult-from kong.agent.supervisor import AnalysisStats+from kong.agent.models import AnalysisStatsfrom kong.export.source import ExportDatafrom kong.export.structured import export_jsonfrom kong.ghidra.types import BinaryInfodiff --git a/tests/test_tui.py b/tests/test_tui.pyindex 8a525a9..878c497 100644--- a/tests/test_tui.py+++ b/tests/test_tui.py@@ -8,6 +8,7 @@ from textual.app import Appfrom textual.widget import Widgetfrom textual.widgets import RichLog, Static+from kong.agent.events import Event, EventTypefrom kong.agent.supervisor import Supervisorfrom kong.tui.app import EVENT_STYLES, AgentEvent, KongAppfrom kong.tui.widgets import AgentLog, BinaryHeader, ProgressWidget, StatusBar@@ -93,7 +94,7 @@ class TestStatusBar:class TestKongApp:def test_is_app_subclass(self) -> None:supervisor = create_autospec(Supervisor, instance=True)- supervisor._paused = False+ supervisor.is_paused = Falsesupervisor.stats = MagicMock()supervisor.stats.llm_calls = 0app = KongApp(supervisor)@@ -101,7 +102,7 @@ class TestKongApp:def test_stores_supervisor(self) -> None:supervisor = create_autospec(Supervisor, instance=True)- supervisor._paused = False+ supervisor.is_paused = Falsesupervisor.stats = MagicMock()supervisor.stats.llm_calls = 0app = KongApp(supervisor)@@ -113,7 +114,6 @@ class TestEventStyles:assert isinstance(EVENT_STYLES, dict)def test_event_styles_has_expected_keys(self) -> None:- from kong.agent.events import EventTypeexpected_keys = {EventType.PHASE_START,EventType.PHASE_COMPLETE,@@ -132,7 +132,6 @@ class TestEventStyles:class TestAgentEvent:def test_agent_event_wraps_event(self) -> None:- from kong.agent.events import Event, EventTypeevent = Event(type=EventType.RUN_START, message="test")msg = AgentEvent(event)assert msg.event is event