compresh-mcp 0.2.2__tar.gz → 0.2.4__tar.gz

{compresh_mcp-0.2.2/src/compresh_mcp.egg-info → compresh_mcp-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compresh-mcp
-Version: 0.2.2
+Version: 0.2.4
 Summary: MCP server for Compresh — production-grade context compression with Q-protective ranking, epistemic markers, and depth-aware adaptation
 Author-email: Compresh Ltd <hello@compre.sh>
 License: BUSL-1.1

{compresh_mcp-0.2.2 → compresh_mcp-0.2.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "compresh-mcp"
-version = "0.2.2"
+version = "0.2.4"
 description = "MCP server for Compresh — production-grade context compression with Q-protective ranking, epistemic markers, and depth-aware adaptation"
 readme = "README.md"
 license = { text = "BUSL-1.1" }

{compresh_mcp-0.2.2 → compresh_mcp-0.2.4}/src/compresh_mcp/__init__.py

@@ -39,7 +39,7 @@ Architecture history:
 
 from . import tulbase
 
-__version__ = "0.2.2"
+__version__ = "0.2.4"
 __author__ = "Compresh Ltd"
 __license__ = "BUSL-1.1"
 

{compresh_mcp-0.2.2 → compresh_mcp-0.2.4}/src/compresh_mcp/auth.py

@@ -136,6 +136,8 @@ def report_usage(
     session_id: str,
     saved_input_tokens: int,
     saved_chars: int,
+    original_chars: int = 0,
+    tulbase_chars: int = 0,
     n_turns: int,
     n_compressed_entries: int,
     provider_hint: Optional[str] = None,
@@ -158,6 +160,12 @@ def report_usage(
         "session_id": session_id,
         "saved_input_tokens": saved_input_tokens,
         "saved_chars": saved_chars,
+        # Absolute sizes (#9a): server only sees the compressed context, so it
+        # needs the client-reported original size to compute total savings; it
+        # can measure tulbase size itself but we send it for cross-check + the
+        # free-user portal saving display (#14).
+        "original_chars": original_chars,
+        "tulbase_chars": tulbase_chars,
         "n_turns": n_turns,
         "n_compressed_entries": n_compressed_entries,
         "provider_hint": provider_hint,

{compresh_mcp-0.2.2 → compresh_mcp-0.2.4}/src/compresh_mcp/server.py

@@ -127,10 +127,14 @@ class SessionState:
         self.log.ensure_schema()
         self.cold = ColdStorage(str(self.workdir / "cold"))
 
-        # Local pipeline runs the open-source tulbase core only — no Q-protective
-        # ranking, no epistemic markers, no semantic store. Those layers run on
-        # the Compresh server (see ``tul1_client.call_v1_tul1``). Local pipeline
-        # is still needed for cold storage + fetch_compressed retrieval support.
+        # Local pipeline runs the open-source tulbase core only — no Q matrix
+        # classifier, no Q-protective ranking, no epistemic markers, no semantic
+        # store. Those are the PAID layer and run server-side (see
+        # ``tul1_client.call_v1_tul1`` → /v1/tul1). [21 May 2026: a client-side
+        # classifier was explored (Decision B) then REVERSED — no bench basis;
+        # the classifier is deterministic + cheap and the server already
+        # receives the full messages, so client-side classification offered no
+        # measurable gain.] Local pipeline still backs cold storage + fetch.
         self.pipeline = Pipeline(
             log=self.log,
             cold=self.cold,
@@ -317,6 +321,43 @@ async def tool_compress(
         len(b.compressed_refs) for b in turn_boxes[: local_composed.n_compressed]
     )
 
+    # Net-negative saving guard: on short conversations the Compresh memory
+    # header can outweigh the text it elides, so the "compressed" view is
+    # actually larger than the raw turns. Reporting that as a negative
+    # saving is wrong (it drags down dashboard totals and bills nothing).
+    # Treat it as a pass-through: applied=False, saving=0, raw messages
+    # returned unchanged. The calling hook then skips injection.
+    if saving_chars <= 0:
+        state.n_turns = max(state.n_turns, len(messages))
+        asyncio.create_task(
+            _report_usage_background(
+                session_id=session_id,
+                saved_input_tokens=0,
+                saved_chars=0,
+                original_chars=raw_chars,
+                tulbase_chars=optimized_chars,
+                n_turns=state.n_turns,
+                n_compressed_entries=0,
+                provider_hint=provider_hint,
+                model_hint=model_hint,
+            )
+        )
+        return {
+            "ok": True,
+            "applied": False,
+            "reason": "net_negative_saving",
+            "optimized_messages": messages,
+            "compresh_md": "",
+            "raw_tail": messages,
+            "n_compressed_turns": 0,
+            "n_compressed_entries": 0,
+            "n_total": len(messages),
+            "saving_chars": 0,
+            "session_id": session_id,
+            "protection_mode": protection_mode,
+            "tier": (get_auth().tier or "unknown"),
+        }
+
     state.saved_chars += saving_chars
     state.n_compressed_entries += n_compressed_entries
     state.n_turns = max(state.n_turns, len(messages))
@@ -342,6 +383,8 @@ async def tool_compress(
             session_id=session_id,
             saved_input_tokens=saving_chars // 4,
             saved_chars=saving_chars,
+            original_chars=raw_chars,
+            tulbase_chars=optimized_chars,
             n_turns=state.n_turns,
             n_compressed_entries=state.n_compressed_entries,
             provider_hint=provider_hint,
@@ -450,9 +493,12 @@ async def tool_stats(session_id: str) -> dict[str, Any]:
         "n_compressed_entries": state.n_compressed_entries,
         "saved_chars": state.saved_chars,
         "storage_path": str(state.workdir),
-        "protect_mode_active": DEFAULT_TIER1_PROTECT_MODE,
-        "q_classifier_enabled": True,
-        "epi_classifier_enabled": True,
+        # Local pipeline = open-source tulbase core only (no Q-protective).
+        # Q-protective ranking + epistemic markers run server-side (/v1/tul1).
+        # Report the LOCAL pipeline's real state, not a hardcoded flag.
+        "protect_mode_active": getattr(state.pipeline.summarizer, "protect_mode", "off"),
+        "q_classifier_enabled": state.pipeline.q_classifier is not None,
+        "epi_classifier_enabled": getattr(state.pipeline.summarizer, "epi_classifier", None) is not None,
     }
 
 
@@ -680,7 +726,7 @@ async def serve() -> None:
             read_stream, write_stream,
             InitializationOptions(
                 server_name="compresh-mcp",
-                server_version="0.2.2",
+                server_version="0.2.4",
                 capabilities=app.get_capabilities(
                     notification_options=NotificationOptions(),
                     experimental_capabilities={},

{compresh_mcp-0.2.2 → compresh_mcp-0.2.4/src/compresh_mcp.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compresh-mcp
-Version: 0.2.2
+Version: 0.2.4
 Summary: MCP server for Compresh — production-grade context compression with Q-protective ranking, epistemic markers, and depth-aware adaptation
 Author-email: Compresh Ltd <hello@compre.sh>
 License: BUSL-1.1

{compresh_mcp-0.2.2 → compresh_mcp-0.2.4}/tests/test_smoke.py

@@ -11,9 +11,15 @@ from __future__ import annotations
 
 
 def test_package_importable() -> None:
+    import re
+
     import compresh_mcp
 
-    assert compresh_mcp.__version__ == "0.2.0"
+    # Assert a valid semver, not a hardcoded version: the old "0.2.0" check
+    # went stale and failed on every release bump (2026-05-21 fix).
+    assert re.fullmatch(r"\d+\.\d+\.\d+", compresh_mcp.__version__), (
+        f"unexpected version: {compresh_mcp.__version__!r}"
+    )
     assert compresh_mcp.__license__ == "BUSL-1.1"