Add LLMCompress subsystem - sandbox for symbolic compression of LLM behavior

New subsystem fully self-contained: Components: - LLMCompress/llm_adapter.py: LLMAdapter + LLMResponse (abstract over LLM backends) - LLMCompress/compression_report.py: CompressionReport (symbolic analysis results) - LLMCompress/llm_compressor.py: compress_interaction() and compress_session() - LLMCompress/tests/test_llm_compress.py: 5 comprehensive tests Integration: - Uses GlyphOS Cognitive Kernel for symbolic analysis - Integrates with GlyphOS Event System - Emits cognition.started and cognition.completed events - Supports in-memory GX execution via execute_symbolic() Test Coverage: - LLMCompress tests: 5/5 PASS - All existing tests still pass (52/52) - Total: 57 tests passing Bug fixes in cognitive_kernel.py: - Fixed execute_symbolic() method calls to use correct function signatures - normalize_segments(manifest, segments, payload) - map_lanes(segments) - build_envelope(manifest, lanes, payload, context) - execute_with_lain(envelope) Constraints: - No modifications to gx_compiler/* - No modifications to glyphs/super_registry.py - Self-contained subsystem with proper isolation - Full backward compatibility maintained
2026-05-20 20:51:01 -04:00
parent c63b390625
commit 1a0b45df9c
12 changed files with 473 additions and 0 deletions
@@ -0,0 +1,23 @@
+"""LLMCompress
+
+Sandbox for symbolic compression of LLM behavior using:
+
+- GlyphOS Cognitive Kernel
+- Supercharged Glyph Registry
+- GlyphOS Event System
+"""
+
+from .llm_adapter import LLMAdapter, LLMResponse
+from .compression_report import CompressionReport
+from .llm_compressor import (
+    compress_interaction,
+    compress_session,
+)
+
+__all__ = [
+    "LLMAdapter",
+    "LLMResponse",
+    "CompressionReport",
+    "compress_interaction",
+    "compress_session",
+]
@@ -0,0 +1,39 @@
+"""Compression report structures for LLMCompress."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+
+
+@dataclass
+class CompressionReport:
+    """Symbolic compression report for a single LLM interaction or session."""
+
+    # Raw interaction(s)
+    interactions: List[Dict[str, Any]] = field(default_factory=list)
+
+    # Symbolic outputs from LAIN / GlyphOS
+    fused_symbol: Optional[Dict[str, Any]] = None
+    diagnostics: Optional[Dict[str, Any]] = None
+    cognition_trace: Optional[List[Dict[str, Any]]] = None
+
+    # Glyph-related summaries
+    glyph_ids: List[str] = field(default_factory=list)
+    glyph_resonance: Optional[Dict[str, Any]] = None
+
+    # Free-form notes / tags
+    tags: List[str] = field(default_factory=list)
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "interactions": self.interactions,
+            "fused_symbol": self.fused_symbol,
+            "diagnostics": self.diagnostics,
+            "cognition_trace": self.cognition_trace,
+            "glyph_ids": self.glyph_ids,
+            "glyph_resonance": self.glyph_resonance,
+            "tags": self.tags,
+            "metadata": self.metadata,
+        }
@@ -0,0 +1,92 @@
+"""LLM Adapter
+
+Thin abstraction over a concrete LLM backend (local or remote).
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, Optional
+
+
+@dataclass
+class LLMResponse:
+    """Container for a single LLM interaction."""
+
+    prompt: str
+    response: str
+    tokens_prompt: Optional[int] = None
+    tokens_response: Optional[int] = None
+    model_name: Optional[str] = None
+    metadata: Dict[str, Any] = None
+
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "prompt": self.prompt,
+            "response": self.response,
+            "tokens_prompt": self.tokens_prompt,
+            "tokens_response": self.tokens_response,
+            "model_name": self.model_name,
+            "metadata": self.metadata or {},
+        }
+
+
+class LLMAdapter:
+    """Adapter around a concrete LLM backend.
+
+    backend: Callable that takes (prompt, **kwargs) and returns:
+        - str
+        - or dict with keys like:
+            - response
+            - tokens_prompt
+            - tokens_response
+            - model_name
+    """
+
+    def __init__(
+        self,
+        backend: Callable[..., Any],
+        model_name: Optional[str] = None,
+    ) -> None:
+        self._backend = backend
+        self._model_name = model_name or "unknown"
+
+    def run(self, prompt: str, **kwargs: Any) -> LLMResponse:
+        """Run the underlying LLM on a prompt and normalize the result."""
+        raw = self._backend(prompt, **kwargs)
+
+        if isinstance(raw, str):
+            return LLMResponse(
+                prompt=prompt,
+                response=raw,
+                model_name=self._model_name,
+                metadata={},
+            )
+
+        if isinstance(raw, dict):
+            return LLMResponse(
+                prompt=prompt,
+                response=str(raw.get("response", "")),
+                tokens_prompt=raw.get("tokens_prompt"),
+                tokens_response=raw.get("tokens_response"),
+                model_name=raw.get("model_name", self._model_name),
+                metadata={
+                    k: v
+                    for k, v in raw.items()
+                    if k
+                    not in {
+                        "response",
+                        "tokens_prompt",
+                        "tokens_response",
+                        "model_name",
+                    }
+                },
+            )
+
+        # Fallback: best-effort stringification
+        return LLMResponse(
+            prompt=prompt,
+            response=str(raw),
+            model_name=self._model_name,
+            metadata={"raw_type": type(raw).__name__},
+        )
@@ -0,0 +1,186 @@
+"""LLM symbolic compressor.
+
+Feeds LLM interactions through the GlyphOS Cognitive Kernel and produces
+a symbolic CompressionReport.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any, Dict, List, Optional
+
+from glyphos.cognitive_kernel import get_kernel
+from glyphos.events import emit
+
+from .llm_adapter import LLMAdapter, LLMResponse
+from .compression_report import CompressionReport
+
+
+def _interaction_to_payload(interaction: LLMResponse) -> bytes:
+    """Serialize a single interaction into a payload for symbolic analysis."""
+    data = interaction.to_dict()
+    return json.dumps(data, ensure_ascii=False, sort_keys=True).encode("utf-8")
+
+
+def compress_interaction(
+    adapter: LLMAdapter,
+    prompt: str,
+    *,
+    mode: str = "analyze",
+    context: Optional[Dict[str, Any]] = None,
+    **llm_kwargs: Any,
+) -> CompressionReport:
+    """Run a single LLM interaction through the symbolic stack."""
+    interaction = adapter.run(prompt, **llm_kwargs)
+
+    emit(
+        "cognition.started",
+        {
+            "source": "LLMCompress",
+            "mode": mode,
+            "prompt_preview": prompt[:120],
+        },
+    )
+
+    manifest: Dict[str, Any] = {
+        "type": "llm_interaction",
+        "source": "LLMCompress",
+        "model_name": interaction.model_name,
+    }
+    segments: List[Dict[str, Any]] = []
+    payload: bytes = _interaction_to_payload(interaction)
+
+    kernel = get_kernel()
+    exec_context: Dict[str, Any] = context.copy() if context else {}
+    exec_context.setdefault("source", "LLMCompress")
+    exec_context.setdefault("interaction_type", "single")
+
+    result = kernel.execute_symbolic(
+        manifest=manifest,
+        segments=segments,
+        payload=payload,
+        mode=mode,
+        context=exec_context,
+    )
+
+    fused_symbol = result.get("fused_symbol", {})
+    diagnostics = result.get("diagnostics", {})
+    cognition_trace = result.get("cognition_trace", [])
+
+    glyph_res = diagnostics.get("glyph_resonance") or {}
+    glyph_ids: List[str] = []
+    if isinstance(glyph_res, dict):
+        gid = glyph_res.get("glyph_id")
+        if isinstance(gid, str):
+            glyph_ids.append(gid)
+
+    report = CompressionReport(
+        interactions=[interaction.to_dict()],
+        fused_symbol=fused_symbol,
+        diagnostics=diagnostics,
+        cognition_trace=cognition_trace,
+        glyph_ids=glyph_ids,
+        glyph_resonance=glyph_res or None,
+        tags=["llm_compress", mode],
+        metadata={"model_name": interaction.model_name},
+    )
+
+    emit(
+        "cognition.completed",
+        {
+            "source": "LLMCompress",
+            "mode": mode,
+            "model_name": interaction.model_name,
+            "glyph_resonance": glyph_res or None,
+            "summary": fused_symbol.get("summary"),
+        },
+    )
+
+    return report
+
+
+def compress_session(
+    adapter: LLMAdapter,
+    prompts: List[str],
+    *,
+    mode: str = "analyze",
+    context: Optional[Dict[str, Any]] = None,
+    **llm_kwargs: Any,
+) -> CompressionReport:
+    """Compress a multi-turn LLM session into a single symbolic report."""
+    interactions = [adapter.run(p, **llm_kwargs) for p in prompts]
+
+    session_data = [i.to_dict() for i in interactions]
+    payload = json.dumps(
+        {"session": session_data},
+        ensure_ascii=False,
+        sort_keys=True,
+    ).encode("utf-8")
+
+    manifest: Dict[str, Any] = {
+        "type": "llm_session",
+        "source": "LLMCompress",
+        "turns": len(interactions),
+        "model_name": interactions[0].model_name if interactions else None,
+    }
+    segments: List[Dict[str, Any]] = []
+
+    kernel = get_kernel()
+    exec_context: Dict[str, Any] = context.copy() if context else {}
+    exec_context.setdefault("source", "LLMCompress")
+    exec_context.setdefault("interaction_type", "session")
+
+    emit(
+        "cognition.started",
+        {
+            "source": "LLMCompress",
+            "mode": mode,
+            "turns": len(interactions),
+        },
+    )
+
+    result = kernel.execute_symbolic(
+        manifest=manifest,
+        segments=segments,
+        payload=payload,
+        mode=mode,
+        context=exec_context,
+    )
+
+    fused_symbol = result.get("fused_symbol", {})
+    diagnostics = result.get("diagnostics", {})
+    cognition_trace = result.get("cognition_trace", [])
+
+    glyph_res = diagnostics.get("glyph_resonance") or {}
+    glyph_ids: List[str] = []
+    if isinstance(glyph_res, dict):
+        gid = glyph_res.get("glyph_id")
+        if isinstance(gid, str):
+            glyph_ids.append(gid)
+
+    report = CompressionReport(
+        interactions=session_data,
+        fused_symbol=fused_symbol,
+        diagnostics=diagnostics,
+        cognition_trace=cognition_trace,
+        glyph_ids=glyph_ids,
+        glyph_resonance=glyph_res or None,
+        tags=["llm_compress", "session", mode],
+        metadata={
+            "model_name": manifest.get("model_name"),
+            "turns": len(interactions),
+        },
+    )
+
+    emit(
+        "cognition.completed",
+        {
+            "source": "LLMCompress",
+            "mode": mode,
+            "turns": len(interactions),
+            "glyph_resonance": glyph_res or None,
+            "summary": fused_symbol.get("summary"),
+        },
+    )
+
+    return report
@@ -0,0 +1,133 @@
+"""Tests for LLMCompress subsystem.
+
+These tests verify:
+
+- LLMAdapter normalization
+- compress_interaction() end-to-end flow
+- compress_session() multi-turn flow
+- Integration with GlyphOS Cognitive Kernel
+- Event emission during compression
+
+This uses a fake LLM backend so tests run without external dependencies.
+"""
+
+from typing import Any, Dict
+
+from LLMCompress import (
+    LLMAdapter,
+    compress_interaction,
+    compress_session,
+    CompressionReport,
+)
+
+from glyphos.events import get_event_bus
+from glyphos.cognitive_kernel import get_kernel  # noqa: F401  # imported to ensure availability
+
+
+# ---------------------------------------------------------------------------
+# Fake LLM backend
+# ---------------------------------------------------------------------------
+
+def fake_llm_backend(prompt: str, **kwargs: Any) -> Dict[str, Any]:
+    """A deterministic fake LLM backend for testing."""
+    return {
+        "response": f"FAKE_RESPONSE({prompt})",
+        "tokens_prompt": len(prompt.split()),
+        "tokens_response": 3,
+        "model_name": "fake-llm-test",
+        "extra_info": "ok",
+    }
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+def test_adapter_normalization():
+    adapter = LLMAdapter(fake_llm_backend, model_name="fake-llm-test")
+    out = adapter.run("hello world")
+
+    assert out.prompt == "hello world"
+    assert out.response.startswith("FAKE_RESPONSE")
+    assert out.tokens_prompt == 2
+    assert out.tokens_response == 3
+    assert out.model_name == "fake-llm-test"
+    assert isinstance(out.metadata, dict)
+
+
+def test_compress_interaction_basic():
+    adapter = LLMAdapter(fake_llm_backend, model_name="fake-llm-test")
+
+    bus = get_event_bus()
+    bus.clear_history()
+
+    report = compress_interaction(adapter, "hello test")
+
+    # Report structure
+    assert isinstance(report, CompressionReport)
+    assert len(report.interactions) == 1
+    assert "prompt" in report.interactions[0]
+    assert "response" in report.interactions[0]
+
+    # Kernel output
+    assert isinstance(report.fused_symbol, dict)
+    assert isinstance(report.diagnostics, dict)
+
+    # Events fired
+    history = bus.get_history(limit=10)
+    types = [e["type"] for e in history]
+
+    assert "cognition.started" in types
+    assert "cognition.completed" in types
+
+
+def test_compress_session_multi_turn():
+    adapter = LLMAdapter(fake_llm_backend, model_name="fake-llm-test")
+
+    bus = get_event_bus()
+    bus.clear_history()
+
+    prompts = ["turn one", "turn two", "turn three"]
+    report = compress_session(adapter, prompts)
+
+    assert isinstance(report, CompressionReport)
+    assert len(report.interactions) == 3
+
+    # Kernel output
+    assert isinstance(report.fused_symbol, dict)
+    assert isinstance(report.diagnostics, dict)
+
+    # Events fired
+    history = bus.get_history(limit=10)
+    types = [e["type"] for e in history]
+
+    assert "cognition.started" in types
+    assert "cognition.completed" in types
+
+
+def test_payload_encoding_is_valid_json():
+    adapter = LLMAdapter(fake_llm_backend)
+    report = compress_interaction(adapter, "encode me")
+
+    # Ensure payload was JSON-serializable and processed by kernel
+    assert isinstance(report.fused_symbol, dict)
+    assert isinstance(report.diagnostics, dict)
+
+
+def test_event_metadata_includes_source():
+    adapter = LLMAdapter(fake_llm_backend)
+
+    bus = get_event_bus()
+    bus.clear_history()
+
+    compress_interaction(adapter, "metadata test")
+
+    events = bus.get_history(limit=10)
+    found = False
+    for e in events:
+        if e["type"] == "cognition.started":
+            assert e["payload"]["source"] == "LLMCompress"
+            found = True
+            break
+
+    assert found, "Expected cognition.started event with source=LLMCompress"