hy-memory 2.0.3__tar.gz → 2.0.4__tar.gz

{hy_memory-2.0.3/hy_memory.egg-info → hy_memory-2.0.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hy-memory
-Version: 2.0.3
+Version: 2.0.4
 Summary: HY Memory - Industrial-grade dual-system cognitive memory framework
 Author: alvinfei
 License-Expression: Apache-2.0

{hy_memory-2.0.3 → hy_memory-2.0.4}/hy_memory/agent/extractor.py

@@ -1123,6 +1123,7 @@ class Extractor:
                 max_tokens=self._llm_config.agent_max_tokens,
                 tools=None,
                 tool_choice=None,
+                audit_stage="extract",
             )
             self._call_count += 1
             total_tokens += response.tokens_used
@@ -1233,6 +1234,7 @@ class Extractor:
                         max_tokens=self._llm_config.agent_max_tokens,
                         tools=None,
                         tool_choice=None,
+                        audit_stage="extract",
                     )
                     self._call_count += 1
                     total_tokens += response.tokens_used

{hy_memory-2.0.3 → hy_memory-2.0.4}/hy_memory/agent/llm_provider.py

@@ -34,6 +34,7 @@ import os
 import weakref
 
 from ..config import MemoryConfig
+from ..utils.model_call_context import model_call_stage
 from ..core.workload import llm_slot
 
 logger = logging.getLogger(__name__)
@@ -348,6 +349,7 @@ class LLMProvider:
         stop: List[str] = None,
         tools: Optional[List[Dict[str, Any]]] = None,
         tool_choice: Optional[Any] = None,
+        audit_stage: Optional[str] = None,
         **kwargs
     ) -> LLMResponse:
         """
@@ -386,15 +388,16 @@ class LLMProvider:
                         len(prompt), _content_fingerprint(prompt),
                     )
 
-                response = await self._call_llm(
-                    prompt=prompt,
-                    max_tokens=max_tokens,
-                    temperature=temperature,
-                    stop=stop,
-                    tools=tools,
-                    tool_choice=tool_choice,
-                    **kwargs
-                )
+                with model_call_stage(audit_stage):
+                    response = await self._call_llm(
+                        prompt=prompt,
+                        max_tokens=max_tokens,
+                        temperature=temperature,
+                        stop=stop,
+                        tools=tools,
+                        tool_choice=tool_choice,
+                        **kwargs
+                    )
 
                 self._total_calls += 1
                 self._total_tokens += response.tokens_used
@@ -595,6 +598,7 @@ class LLMProvider:
         temperature: Optional[float] = None,
         tools: Optional[List[Dict[str, Any]]] = None,
         tool_choice: Optional[Any] = None,
+        audit_stage: Optional[str] = None,
     ) -> LLMResponse:
         """
         多轮 messages 调用（支持 tool results 回传）。
@@ -622,13 +626,14 @@ class LLMProvider:
                         f"msgs={len(messages)} tools={len(tools) if tools else 0}"
                     )
 
-                response = await self._chat_completions(
-                    messages,
-                    max_tokens=max_tokens,
-                    temperature=temperature,
-                    tools=tools,
-                    tool_choice=tool_choice,
-                )
+                with model_call_stage(audit_stage):
+                    response = await self._chat_completions(
+                        messages,
+                        max_tokens=max_tokens,
+                        temperature=temperature,
+                        tools=tools,
+                        tool_choice=tool_choice,
+                    )
 
                 self._total_calls += 1
                 self._total_tokens += response.tokens_used

{hy_memory-2.0.3 → hy_memory-2.0.4}/hy_memory/agent/reconciler.py

@@ -669,6 +669,7 @@ class MemoryReconciler:
             response = await llm.complete(
                 prompt=prompt,
                 max_tokens=500,
+                audit_stage="reconcile",
             )
             text = response.content.strip()
             if "```json" in text:
@@ -1383,6 +1384,7 @@ class MemoryReconciler:
                     response = await llm.complete(
                         prompt=prompt,
                         max_tokens=dyn_max_tokens,
+                        audit_stage="reconcile",
                     )
                 except Exception as e:
                     # LLM 彻底不可用（provider 已按 max_retries 重试过，网关 500 /

{hy_memory-2.0.3 → hy_memory-2.0.4}/hy_memory/coding/extractor.py

@@ -919,6 +919,7 @@ class CodingMemoryExtractor:
             resp = await self.llm_provider.complete(
                 prompt=prompt,
                 max_tokens=self.max_tokens,
+                audit_stage="coding_extract",
             )
         except Exception as e:
             logger.warning(f"[coding-extract] LLM call failed: {e}")

{hy_memory-2.0.3 → hy_memory-2.0.4}/hy_memory/coding/fact_reconciler.py

@@ -571,6 +571,7 @@ class WorkingFactReconciler:
         resp = await self.llm_provider.complete(
             prompt=prompt,
             max_tokens=self.max_tokens,
+            audit_stage="coding_fact_reconcile",
         )
         return self._parse_ops(resp.content)
 
@@ -637,6 +638,7 @@ class WorkingFactReconciler:
             resp = await self.llm_provider.complete(
                 prompt=prompt,
                 max_tokens=_CHAIN_SUMMARY_MAX_TOKENS,
+                audit_stage="coding_fact_summary",
             )
         except Exception as e:
             logger.warning(f"[fact-reconcile] chain summary failed: {e}")

{hy_memory-2.0.3 → hy_memory-2.0.4}/hy_memory/coding/judge.py

@@ -29,6 +29,9 @@ logger = logging.getLogger(__name__)
 
 _JUDGE_MAX_TURNS = 20
 _JUDGE_HEAD_TURNS = 5
+_JUDGE_USER_CONTENT_LIMIT = 1_000
+_JUDGE_USER_CONTENT_HEAD = 200
+_JUDGE_USER_CONTENT_TAIL = 200
 _JUDGE_MAX_QUERY_CHARS = 300
 _JUDGE_MAX_QUERIES = 20
 
@@ -186,6 +189,7 @@ async def classify_messages_scene(
             prompt=prompt,
             max_tokens=max_tokens,
             temperature=temperature,
+            audit_stage="coding_judge",
         )
         _elapsed = round((time.perf_counter() - _t0) * 1000, 2)
         result = _parse_scene_json(resp.content)
@@ -631,9 +635,13 @@ def _format_turns(summary: List[Dict[str, Any]]) -> str:
     rows = []
     for t in summary:
         user = t.get("user", "") or ""
-        # 单 turn 用户文本截到 200 字符
-        if len(user) > 200:
-            user = user[:200] + "..."
+        if len(user) > _JUDGE_USER_CONTENT_LIMIT:
+            omitted = len(user) - _JUDGE_USER_CONTENT_HEAD - _JUDGE_USER_CONTENT_TAIL
+            user = (
+                user[:_JUDGE_USER_CONTENT_HEAD]
+                + f"\n[...{omitted} chars omitted...]\n"
+                + user[-_JUDGE_USER_CONTENT_TAIL:]
+            )
         tools = t.get("tools", []) or []
         rows.append(
             f"- turn {t.get('turn', '?')}: user={user!r}, tools={tools}"

{hy_memory-2.0.3 → hy_memory-2.0.4}/hy_memory/coding/reconciler.py

@@ -310,6 +310,7 @@ class CodingMemoryReconciler:
         resp = await self.llm_provider.complete(
             prompt=prompt,
             max_tokens=self.max_tokens,
+            audit_stage="coding_reconcile",
         )
         return self._parse_ops(resp.content), prompt, resp.content, resp
 

hy_memory-2.0.4/hy_memory/utils/model_call_context.py

@@ -0,0 +1,24 @@
+"""Context propagated to host-provided HTTP transports for cost accounting."""
+
+from collections.abc import Iterator
+from contextlib import contextmanager
+from contextvars import ContextVar, Token
+
+_model_call_stage: ContextVar[str] = ContextVar(
+    "hy_memory_model_call_stage",
+    default="other",
+)
+
+
+def current_model_call_stage() -> str:
+    return _model_call_stage.get()
+
+
+@contextmanager
+def model_call_stage(stage: str | None) -> Iterator[None]:
+    normalized = (stage or "other").strip().lower() or "other"
+    token: Token[str] = _model_call_stage.set(normalized)
+    try:
+        yield
+    finally:
+        _model_call_stage.reset(token)

{hy_memory-2.0.3 → hy_memory-2.0.4}/hy_memory/utils/pipeline_log_writer.py

@@ -80,7 +80,11 @@ class PipelineLogWriter:
                 split_namespace_user,
             )
 
-            tokens = (prompt_tokens or 0) + (completion_tokens or 0) + (total_tokens or 0)
+            # Providers define total_tokens as prompt+completion. Adding all
+            # three fields doubled every cost estimate in this JSONL.
+            tokens = (total_tokens or 0) or (
+                (prompt_tokens or 0) + (completion_tokens or 0)
+            )
             namespace, uid = split_namespace_user(user_id or "")
             if not is_llm_content_logging_enabled():
                 prompt = redact_content(prompt)

{hy_memory-2.0.3 → hy_memory-2.0.4/hy_memory.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hy-memory
-Version: 2.0.3
+Version: 2.0.4
 Summary: HY Memory - Industrial-grade dual-system cognitive memory framework
 Author: alvinfei
 License-Expression: Apache-2.0

{hy_memory-2.0.3 → hy_memory-2.0.4}/hy_memory.egg-info/SOURCES.txt

@@ -128,6 +128,7 @@ hy_memory/utils/keyword_ties.py
 hy_memory/utils/lang_detect.py
 hy_memory/utils/log_setup.py
 hy_memory/utils/metrics.py
+hy_memory/utils/model_call_context.py
 hy_memory/utils/pipeline_log_writer.py
 hy_memory/utils/pipeline_observability.py
 hy_memory/utils/source_filter.py

{hy_memory-2.0.3 → hy_memory-2.0.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "hy-memory"
-version = "2.0.3"
+version = "2.0.4"
 description = "HY Memory - Industrial-grade dual-system cognitive memory framework"
 readme = "README.md"
 requires-python = ">=3.9"