Skip to content

armgit / kong

Public Imported from GitHub
0

The world's first agentic reverse engineer.

feat: add _get_effective_limits for custom model limit overrides

Amruth226bff7e9aa3
diff --git a/kong/agent/supervisor.py b/kong/agent/supervisor.py
index 25d3c50..6bd7663 100644
--- a/kong/agent/supervisor.py
+++ b/kong/agent/supervisor.py
@@ -18,12 +18,12 @@ from kong.agent.queue import WorkItem, WorkQueue
from kong.agent.signatures import SignatureDB
from kong.agent.triage import TriageAgent, TriageResult
from kong.agent.type_recovery import StructAccumulator, apply_unified_structs
-from kong.config import KongConfig
+from kong.config import KongConfig, LLMProvider
from kong.export.source import ExportData, export_source
from kong.export.structured import export_json
from kong.ghidra.client import GhidraClient
from kong.ghidra.types import BinaryInfo, FunctionInfo, StringEntry
-from kong.llm.limits import get_model_limits
+from kong.llm.limits import ModelLimits, get_model_limits
from kong.llm.usage import TokenUsage
from kong.normalizer.syntactic import normalize
from kong.synthesis.semantic import SemanticSynthesizer, SynthesisResult
@@ -295,6 +295,17 @@ class Supervisor:
),
))
+ def _get_effective_limits(self) -> ModelLimits:
+ cfg = self.config.llm
+ base = get_model_limits(getattr(self.llm_client, "model", "") or "")
+ if cfg.provider is not LLMProvider.CUSTOM:
+ return base
+ return ModelLimits(
+ max_prompt_chars=cfg.max_prompt_chars or base.max_prompt_chars,
+ max_chunk_functions=cfg.max_chunk_functions or base.max_chunk_functions,
+ max_output_tokens=cfg.max_output_tokens or base.max_output_tokens,
+ )
+
def _split_into_chunks(
self,
items: list[tuple[WorkItem, str]],
@@ -302,8 +313,7 @@ class Supervisor:
"""Split items into chunks by building the actual prompt and measuring size."""
assert self.binary_info is not None
- model_name = getattr(self.llm_client, "model", "") or ""
- limits = get_model_limits(model_name)
+ limits = self._get_effective_limits()
overhead = self._build_chunk_prompt([])
overhead_len = len(overhead)
diff --git a/tests/test_agent_supervisor.py b/tests/test_agent_supervisor.py
index 1a2d998..04ecce1 100644
--- a/tests/test_agent_supervisor.py
+++ b/tests/test_agent_supervisor.py
@@ -8,7 +8,7 @@ from kong.agent.analyzer import LLMResponse
from kong.agent.events import EventType, Phase
from kong.agent.models import AnalysisStats, FunctionResult
from kong.agent.supervisor import Supervisor
-from kong.config import KongConfig, OutputConfig
+from kong.config import KongConfig, LLMConfig, LLMProvider, OutputConfig
from kong.ghidra.types import BinaryInfo, FunctionClassification, FunctionInfo
@@ -474,3 +474,50 @@ class TestChunkedPipelineIntegration:
sup.run()
assert mock_llm.analyze_function_batch.call_count == 1
+
+
+class TestGetEffectiveLimits:
+ def test_non_custom_returns_base_limits(self, tmp_path):
+ from kong.llm.limits import _DEFAULT_LIMITS
+
+ config = KongConfig(
+ llm=LLMConfig(provider=LLMProvider.ANTHROPIC),
+ output=OutputConfig(directory=tmp_path / "out"),
+ )
+ sup = Supervisor(_make_client(), config)
+ limits = sup._get_effective_limits()
+ assert limits == _DEFAULT_LIMITS
+
+ def test_custom_with_overrides(self, tmp_path):
+ config = KongConfig(
+ llm=LLMConfig(
+ provider=LLMProvider.CUSTOM,
+ model="llama3:8b",
+ max_prompt_chars=32000,
+ max_chunk_functions=20,
+ max_output_tokens=4096,
+ ),
+ output=OutputConfig(directory=tmp_path / "out"),
+ )
+ sup = Supervisor(_make_client(), config)
+ limits = sup._get_effective_limits()
+ assert limits.max_prompt_chars == 32000
+ assert limits.max_chunk_functions == 20
+ assert limits.max_output_tokens == 4096
+
+ def test_custom_partial_overrides_fall_back(self, tmp_path):
+ from kong.llm.limits import _DEFAULT_LIMITS
+
+ config = KongConfig(
+ llm=LLMConfig(
+ provider=LLMProvider.CUSTOM,
+ model="llama3:8b",
+ max_prompt_chars=32000,
+ ),
+ output=OutputConfig(directory=tmp_path / "out"),
+ )
+ sup = Supervisor(_make_client(), config)
+ limits = sup._get_effective_limits()
+ assert limits.max_prompt_chars == 32000
+ assert limits.max_chunk_functions == _DEFAULT_LIMITS.max_chunk_functions
+ assert limits.max_output_tokens == _DEFAULT_LIMITS.max_output_tokens