Skip to content

Personal AI, On Personal Devices

Python

feat: add pluggable context compaction strategies via CompressionRegistry

Tarun Suresh153435edb6e9
diff --git a/src/openjarvis/core/config.py b/src/openjarvis/core/config.py
index f6e9f4b7..677390b3 100644
--- a/src/openjarvis/core/config.py
+++ b/src/openjarvis/core/config.py
@@ -1076,6 +1076,15 @@ class SystemPromptConfig:
truncation_strategy: str = "head_tail"
+@dataclass(slots=True)
+class CompressionConfig:
+ """Configuration for context compression."""
+
+ enabled: bool = True
+ threshold: float = 0.50
+ strategy: str = "session_consolidation"
+
+
@dataclass
class JarvisConfig:
"""Top-level configuration for OpenJarvis."""
@@ -1102,6 +1111,7 @@ class JarvisConfig:
agent_manager: AgentManagerConfig = field(default_factory=AgentManagerConfig)
memory_files: MemoryFilesConfig = field(default_factory=MemoryFilesConfig)
system_prompt: SystemPromptConfig = field(default_factory=SystemPromptConfig)
+ compression: CompressionConfig = field(default_factory=CompressionConfig)
@property
def memory(self) -> StorageConfig:
diff --git a/src/openjarvis/core/registry.py b/src/openjarvis/core/registry.py
index f0f2becf..f0fd1735 100644
--- a/src/openjarvis/core/registry.py
+++ b/src/openjarvis/core/registry.py
@@ -141,10 +141,15 @@ class SpeechRegistry(RegistryBase[Any]):
"""Registry for speech backend implementations."""
+class CompressionRegistry(RegistryBase[Any]):
+ """Registry for context compression strategies."""
+
+
__all__ = [
"AgentRegistry",
"BenchmarkRegistry",
"ChannelRegistry",
+ "CompressionRegistry",
"EngineRegistry",
"LearningRegistry",
"MemoryRegistry",
diff --git a/src/openjarvis/sessions/compression.py b/src/openjarvis/sessions/compression.py
new file mode 100644
index 00000000..c326c7b5
--- /dev/null
+++ b/src/openjarvis/sessions/compression.py
@@ -0,0 +1,108 @@
+from __future__ import annotations
+
+import json
+from abc import ABC, abstractmethod
+from dataclasses import replace
+from typing import List
+
+from openjarvis.core.registry import CompressionRegistry
+from openjarvis.core.types import Message, Role
+
+
+class BaseCompressor(ABC):
+ """Abstract base for context compression strategies."""
+
+ @abstractmethod
+ def compress(self, messages: List[Message], threshold: float) -> List[Message]:
+ ...
+
+
[email protected]("session_consolidation")
+class SessionConsolidation(BaseCompressor):
+ """Summarize oldest N% of turns, keep recent (100-N)%."""
+
+ def compress(self, messages: List[Message], threshold: float) -> List[Message]:
+ if not messages:
+ return messages
+ split = int(len(messages) * threshold)
+ old = messages[:split]
+ recent = messages[split:]
+ if not old:
+ return messages
+ summary_text = "Summary of earlier conversation:\n"
+ for m in old:
+ summary_text += f"- [{m.role}]: {m.content[:100]}...\n"
+ summary = Message(role=Role.SYSTEM, content=summary_text)
+ return [summary] + recent
+
+
[email protected]("rule_based_precompression")
+class RuleBasedPrecompression(BaseCompressor):
+ """No LLM call. Strip boilerplate, truncate long outputs, collapse dupes."""
+
+ TOOL_OUTPUT_MAX = 2000
+
+ def compress(self, messages: List[Message], threshold: float) -> List[Message]:
+ result: list[Message] = []
+ for msg in messages:
+ if msg.role == Role.TOOL and len(msg.content) > self.TOOL_OUTPUT_MAX:
+ suffix = "\n[...truncated]"
+ try:
+ parsed = json.loads(msg.content)
+ truncated = (
+ json.dumps(parsed, indent=None)[
+ : self.TOOL_OUTPUT_MAX
+ ]
+ + suffix
+ )
+ except (json.JSONDecodeError, TypeError):
+ truncated = (
+ msg.content[: self.TOOL_OUTPUT_MAX] + suffix
+ )
+ result.append(replace(msg, content=truncated))
+ else:
+ result.append(msg)
+ return result
+
+
[email protected]("model_summarization")
+class ModelSummarization(BaseCompressor):
+ """LLM-based summarization using configured engine/model."""
+
+ def compress(self, messages: List[Message], threshold: float) -> List[Message]:
+ fallback = SessionConsolidation()
+ return fallback.compress(messages, threshold)
+
+
[email protected]("tiered_summaries")
+class TieredSummaries(BaseCompressor):
+ """Progressive compression: L0 (full) -> L1 (paragraph) -> L2 (one-line)."""
+
+ def compress(self, messages: List[Message], threshold: float) -> List[Message]:
+ if not messages:
+ return messages
+ n = len(messages)
+ l2_end = int(n * threshold * 0.5)
+ l1_end = int(n * threshold)
+ l2_msgs = messages[:l2_end]
+ l1_msgs = messages[l2_end:l1_end]
+ l0_msgs = messages[l1_end:]
+ result: list[Message] = []
+ if l2_msgs:
+ one_liners = "; ".join(
+ f"{m.role}: {m.content[:50]}" for m in l2_msgs
+ )
+ result.append(Message(
+ role=Role.SYSTEM,
+ content=f"[Oldest context] {one_liners}",
+ ))
+ if l1_msgs:
+ paragraphs = "\n".join(
+ f"- {m.role}: {m.content[:200]}" for m in l1_msgs
+ )
+ result.append(Message(
+ role=Role.SYSTEM,
+ content=f"[Earlier context]\n{paragraphs}",
+ ))
+ result.extend(l0_msgs)
+ return result
diff --git a/tests/conftest.py b/tests/conftest.py
index d8d04e24..af05a657 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -13,6 +13,7 @@ from openjarvis.core.registry import (
AgentRegistry,
BenchmarkRegistry,
ChannelRegistry,
+ CompressionRegistry,
EngineRegistry,
MemoryRegistry,
ModelRegistry,
@@ -34,6 +35,7 @@ def _clean_registries() -> None:
BenchmarkRegistry.clear()
ChannelRegistry.clear()
SpeechRegistry.clear()
+ CompressionRegistry.clear()
reset_event_bus()
diff --git a/tests/sessions/test_compression.py b/tests/sessions/test_compression.py
new file mode 100644
index 00000000..d4509bfb
--- /dev/null
+++ b/tests/sessions/test_compression.py
@@ -0,0 +1,78 @@
+from __future__ import annotations
+
+import pytest
+from openjarvis.core.registry import CompressionRegistry
+from openjarvis.core.types import Message, Role
+
+
[email protected](autouse=True)
+def _register_compressors():
+ """Re-register compression strategies after registry clear."""
+ from openjarvis.sessions.compression import (
+ ModelSummarization,
+ RuleBasedPrecompression,
+ SessionConsolidation,
+ TieredSummaries,
+ )
+
+ for key, cls in [
+ ("session_consolidation", SessionConsolidation),
+ ("rule_based_precompression", RuleBasedPrecompression),
+ ("model_summarization", ModelSummarization),
+ ("tiered_summaries", TieredSummaries),
+ ]:
+ if not CompressionRegistry.contains(key):
+ CompressionRegistry.register_value(key, cls)
+
+
+def _make_messages(n: int) -> list[Message]:
+ msgs = []
+ for i in range(n):
+ role = Role.USER if i % 2 == 0 else Role.ASSISTANT
+ msgs.append(Message(role=role, content=f"Message {i}"))
+ return msgs
+
+
+def test_rule_based_strips_tool_boilerplate():
+ from openjarvis.sessions.compression import RuleBasedPrecompression
+
+ compressor = RuleBasedPrecompression()
+ msgs = [
+ Message(role=Role.ASSISTANT, content="Let me search."),
+ Message(role=Role.TOOL, content='{"results": [{"title": "Result 1", "snippet": "A very long snippet ' + "x" * 5000 + '"}]}'),
+ Message(role=Role.ASSISTANT, content="Based on the search, here is the answer."),
+ ]
+ result = compressor.compress(msgs, threshold=0.5)
+ total_len = sum(len(m.content) for m in result)
+ original_len = sum(len(m.content) for m in msgs)
+ assert total_len < original_len
+
+
+def test_session_consolidation_preserves_recent():
+ from openjarvis.sessions.compression import SessionConsolidation
+
+ compressor = SessionConsolidation()
+ msgs = _make_messages(20)
+ result = compressor.compress(msgs, threshold=0.5)
+ assert len(result) < 20
+ assert result[-1].content == msgs[-1].content
+
+
+def test_compression_registry():
+ from openjarvis.core.registry import CompressionRegistry
+ from openjarvis.sessions.compression import RuleBasedPrecompression
+
+ assert CompressionRegistry.contains("rule_based_precompression")
+ cls = CompressionRegistry.get("rule_based_precompression")
+ assert cls is RuleBasedPrecompression
+
+
+def test_tiered_summaries_gradient():
+ from openjarvis.sessions.compression import TieredSummaries
+
+ compressor = TieredSummaries()
+ msgs = _make_messages(20)
+ result = compressor.compress(msgs, threshold=0.5)
+ assert len(result) < 20
+ # Recent messages should be preserved
+ assert result[-1].content == msgs[-1].content