memleaf 0.2.13__tar.gz → 0.2.14__tar.gz

{memleaf-0.2.13 → memleaf-0.2.14}/CHANGELOG.md

@@ -2,6 +2,17 @@
 
 All notable changes to memleaf are documented here.
 
+## 0.2.14 — 2026-09-02
+
+- Reject a permanent-memory candidate that combines a durable project rule or
+  implementation-plan change with an independent dated todo, while preserving
+  multiple dated milestones that belong to one project plan.
+- Normalize explicit implementation plans and plan adjustments to `project`
+  without promoting adjacent mail, attachment, meeting, or archive records.
+- Keep the gate-selected type immutable through summarization for both CREATE
+  and UPDATE, preserve multiple legal update candidates, and expose the new
+  `mixed_future_use` diagnostic through the Hermes provider.
+
 ## 0.2.13 — 2026-09-02
 
 - Preserve the full active project-plan body, stable title, and retrieval

{memleaf-0.2.13/src/memleaf.egg-info → memleaf-0.2.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memleaf
-Version: 0.2.13
+Version: 0.2.14
 Summary: A local-first Markdown memory core for AI agents
 Author: memleaf contributors
 License-Expression: MIT
@@ -23,8 +23,8 @@ Dynamic: license-file
 
 [English](README.en.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **当前版本：0.2.13。**
-> 核心库、Vault、stdio MCP Server、初始化 CLI、模型路由、提炼流程、受控检索协议和宿主适配器已经实现。memleaf 0.2.13 通过 PyPI 分发。
+> **当前版本：0.2.14。**
+> 核心库、Vault、stdio MCP Server、初始化 CLI、模型路由、提炼流程、受控检索协议和宿主适配器已经实现。memleaf 0.2.14 通过 PyPI 分发。
 > **当前版本支持 Hermes 和 Codex。** Antigravity（反重力）不检测、不安装、不配置。
 
 ## 项目定位

{memleaf-0.2.13 → memleaf-0.2.14}/README.en.md

@@ -4,8 +4,8 @@
 
 [中文](README.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **Version: 0.2.13.**
-> The core library, Vault, stdio MCP server, initialization CLI, model routing, memory extraction, controlled retrieval protocol, and host adapters are implemented. memleaf 0.2.13 is distributed through PyPI.
+> **Version: 0.2.14.**
+> The core library, Vault, stdio MCP server, initialization CLI, model routing, memory extraction, controlled retrieval protocol, and host adapters are implemented. memleaf 0.2.14 is distributed through PyPI.
 > **The current release supports Hermes and Codex.** Antigravity is not detected, installed, or configured.
 
 ## Project scope

{memleaf-0.2.13 → memleaf-0.2.14}/README.md

@@ -4,8 +4,8 @@
 
 [English](README.en.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **当前版本：0.2.13。**
-> 核心库、Vault、stdio MCP Server、初始化 CLI、模型路由、提炼流程、受控检索协议和宿主适配器已经实现。memleaf 0.2.13 通过 PyPI 分发。
+> **当前版本：0.2.14。**
+> 核心库、Vault、stdio MCP Server、初始化 CLI、模型路由、提炼流程、受控检索协议和宿主适配器已经实现。memleaf 0.2.14 通过 PyPI 分发。
 > **当前版本支持 Hermes 和 Codex。** Antigravity（反重力）不检测、不安装、不配置。
 
 ## 项目定位

{memleaf-0.2.13 → memleaf-0.2.14}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "memleaf"
-version = "0.2.13"
+version = "0.2.14"
 description = "A local-first Markdown memory core for AI agents"
 readme = "README.md"
 requires-python = ">=3.11"

{memleaf-0.2.13 → memleaf-0.2.14}/src/memleaf/__init__.py

@@ -1,6 +1,6 @@
 """Local-first Markdown memory core for AI agents."""
 
-__version__ = "0.2.13"
+__version__ = "0.2.14"
 
 from .config import DEFAULT_CONFIG, default_config, load_config, save_config
 from .frontmatter import FrontmatterError, dump_frontmatter, dump_yaml, load_yaml, parse_frontmatter

{memleaf-0.2.13 → memleaf-0.2.14}/src/memleaf/hermes_provider/__init__.py

@@ -94,6 +94,7 @@ _MODEL_VALIDATION_DETAILS = frozenset(
         "source_shape",
         "todo_fields",
         "relative_time",
+        "mixed_future_use",
         "other_schema_violation",
     }
 )

{memleaf-0.2.13 → memleaf-0.2.14}/src/memleaf/hermes_provider/plugin.yaml

@@ -1,5 +1,5 @@
 name: memleaf
-version: 0.2.13
+version: 0.2.14
 description: "Hermes-native external MemoryProvider for local-first Markdown memory shared with the memleaf MCP server."
 hooks:
   - prefetch

{memleaf-0.2.13 → memleaf-0.2.14}/src/memleaf/processing.py

@@ -34,6 +34,7 @@ from .prompts import (
     DUPLICATE_TARGET_CORRECTION,
     GATE_TYPE_CORRECTION,
     JSON_CORRECTION,
+    MIXED_FUTURE_USE_CORRECTION,
     MIXED_PROJECT_SCOPES_CORRECTION,
     RELATIVE_TIME_CORRECTION,
     GATE_SYSTEM,
@@ -66,6 +67,7 @@ from .validation import (
     normalize_relative_calendar_text,
     is_aggregate_operational_text,
     is_attachment_followup_only_text,
+    is_project_plan_text,
 )
 from .vault import safe_component
 
@@ -599,6 +601,8 @@ class Processor:
             return DUPLICATE_TARGET_CORRECTION
         if hint == "mixed_project_scopes":
             return MIXED_PROJECT_SCOPES_CORRECTION
+        if hint == "mixed_future_use":
+            return MIXED_FUTURE_USE_CORRECTION
         if stage == "gate" and hint == "update_target_type_mismatch":
             return UPDATE_TARGET_TYPE_CORRECTION
         if stage == "gate" and hint == "invalid_type":
@@ -1696,22 +1700,18 @@ class Processor:
 
     @staticmethod
     def _is_project_plan_title(value: Any) -> bool:
-        text = normalize_term(value) if isinstance(value, str) else ""
-        return any(
-            marker in text
-            for marker in ("实施计划", "项目计划", "implementation plan", "project plan")
-        )
+        return is_project_plan_text(value)
 
     @staticmethod
     def _is_adjacent_plan_record(value: Any) -> bool:
         text = normalize_term(value) if isinstance(value, str) else ""
-        return any(
+        return bool(text) and any(
             marker in text
             for marker in (
                 "已发送", "发送", "邮件", "附件", "存档", "会议", "启动会", "纪要",
                 "sent", "email", "mail", "attachment", "archive", "meeting", "minutes",
             )
-        )
+        ) and not is_project_plan_text(value)
 
     @classmethod
     def _merge_additive_project_plan_update(
@@ -2335,20 +2335,13 @@ class Processor:
                         expected_scopes=candidate["scopes"],
                         expected_scope_source=candidate["scope_source"],
                         allow_no_change=True,
-                        expected_type=(
-                            candidate.get("type")
-                            if (
-                                gate_update_target is not None
-                                and gate_target_type == candidate.get("type")
-                            )
-                            else None
-                        ),
+                        # The summarize stage may not reinterpret a gate
+                        # candidate, including CREATE candidates.  Updates
+                        # additionally retain the active target's immutable
+                        # type below.
+                        expected_type=candidate.get("type"),
                         expected_update_memory_id=gate_update_target,
-                        expected_target_type=(
-                            gate_target_type
-                            if gate_target_type == candidate.get("type")
-                            else None
-                        ),
+                        expected_target_type=gate_target_type,
                     ),
                     diagnostic_context={
                         "source": turn.source,

{memleaf-0.2.13 → memleaf-0.2.14}/src/memleaf/prompts.py

@@ -108,6 +108,13 @@ be independently retrieved and updated. Prefer one coherent current-state memory
 over adjacent overlapping snapshots; do not use a score or fixed category rule.
 Most ordinary turns should produce zero or one candidate; multiple require
 genuinely independent future questions or actions.
+Never combine a durable project rule or implementation-plan change with an
+independent dated todo in one candidate. For example, a project plan/rule and
+"reply or investigate by YYYY-MM-DD" are two future uses: emit separate
+candidates, preserving each candidate's own evidence and project scope. A
+date that is an intrinsic milestone of the same implementation plan may remain
+inside that one project candidate. A candidate that combines these independent
+uses is invalid and must be split before summarization.
 Never persist an aggregate mailbox sweep, daily report, or "items to watch"
 digest as one cross-project memory. Inspect each item and split genuinely
 independent project risks or actionable todo items; if an item has no
@@ -155,6 +162,16 @@ same-use project targets remain, do not guess an update target. A candidate
 that still combines a mailbox or daily digest shell with counts/statuses is not
 atomic: set worth=false for the aggregate, and emit any independent concrete
 action as a separate candidate.
+If the same confirmed change applies to multiple same-project active memories,
+emit one candidate per independently changed memory and set each exact
+update_memory_id; never select one target and silently omit the others. Do not
+target adjacent sent-mail, attachment, meeting, or archive records unless the
+current evidence changes their own future-use topic.
+An explicit implementation plan, project implementation plan, or plan
+adjustment is type project. Do not label it fact merely because it came from
+an email or because no existing project memory was retrieved. Keep an actual
+todo as todo only when its future use is the action itself, not the durable
+plan it mentions.
 Return no prose, markdown fences, comments, or trailing text."""
 
 
@@ -265,6 +282,16 @@ plan target over same-topic sent-mail, attachment/archive, or meeting records;
 the target type remains immutable. Never summarize a combined mailbox/daily
 digest shell with counts or statuses; return NO_CHANGE for that aggregate,
 while a separate atomic future-use action may still be summarized.
+If the current confirmed change applies to multiple same-project active
+memories, the gate will provide separate candidates and exact update targets;
+preserve each candidate's target and do not collapse them into one summary.
+Never combine a durable project rule or implementation-plan change with an
+independent dated todo in one summary. A plan/rule and a dated reply,
+investigation, submission, or follow-up are separate future uses; such output
+is invalid and must be retried as separate gate candidates. A date that is an
+intrinsic milestone of the same project plan may remain in its project summary.
+An explicit implementation plan or plan adjustment must use type project;
+preserve an existing update target's type when it is the same project plan.
 Return no prose, markdown fences, comments, or trailing text."""
 
 
@@ -325,6 +352,18 @@ MIXED_PROJECT_SCOPES_CORRECTION = (
 )
 
 
+MIXED_FUTURE_USE_CORRECTION = (
+    "Previous output violated: mixed_future_use. Keep every candidate and "
+    "summary atomic. Split a durable project rule or implementation-plan "
+    "change from an independent dated todo such as a reply, investigation, "
+    "submission, or follow-up; preserve each part's own evidence and project "
+    "scope. A date intrinsic to the same implementation plan may remain in "
+    "that project candidate. Return separate gate candidates or return "
+    "NO_CHANGE for a part with no independent future use. Return only the "
+    "strict JSON object."
+)
+
+
 DUPLICATE_TARGET_CORRECTION = (
     "Previous output violated: duplicate_update_target. Merge every candidate "
     "that refers to the same active memory target into one candidate for one "

{memleaf-0.2.13 → memleaf-0.2.14}/src/memleaf/validation.py

@@ -75,6 +75,7 @@ MODEL_VALIDATION_DETAILS = frozenset(
         "source_shape",
         "todo_fields",
         "relative_time",
+        "mixed_future_use",
         "other_schema_violation",
     )
 )
@@ -377,6 +378,229 @@ _AGGREGATE_COUNT = re.compile(
 )
 
 
+# These are intentionally narrow, high-signal markers.  The gate is still
+# responsible for deciding whether a candidate is worth retaining; these
+# helpers only protect the write boundary from a known class of semantic
+# contracts that cannot safely be represented by one memory.
+_PROJECT_PLAN_MARKERS = (
+    "实施计划",
+    "项目计划",
+    "实施方案",
+    "项目方案",
+    "计划调整",
+    "调整计划",
+    "调整实施计划",
+    "重新压实计划",
+    "重新压实实施计划",
+    "重新制定计划",
+    "部署计划",
+    "上线计划",
+    "implementation plan",
+    "project plan",
+    "implementation schedule",
+    "plan adjustment",
+    "adjust the plan",
+    "revise the plan",
+    "rebaseline the plan",
+)
+_ADJACENT_PLAN_RECORD_MARKERS = (
+    "邮件",
+    "附件",
+    "会议纪要",
+    "会议记录",
+    "启动会",
+    "存档",
+    "归档",
+    "email",
+    "mail",
+    "attachment",
+    "meeting minutes",
+    "meeting record",
+    "archive",
+)
+_PROJECT_RULE_MARKERS = (
+    "项目要求",
+    "项目约束",
+    "交付要求",
+    "技术路线",
+    "长期要求",
+    "长期约束",
+    "统一规范",
+    "固定规则",
+    "以后",
+    "project requirement",
+    "project constraint",
+    "delivery requirement",
+    "technical approach",
+)
+_DATED_TODO_MARKERS = (
+    "待办",
+    "排查",
+    "答复",
+    "回复",
+    "提交",
+    "跟进",
+    "提醒",
+    "联系",
+    "发给",
+    "发送",
+    "处理",
+    "确认",
+    "整改",
+    "修复",
+    "回访",
+    "催办",
+    "反馈",
+    "完成",
+    "补充材料",
+    "提供",
+    "todo",
+    "deadline",
+    "due",
+    "follow up",
+    "follow-up",
+)
+_INDEPENDENT_TODO_MARKERS = tuple(
+    marker for marker in _DATED_TODO_MARKERS if marker != "完成"
+)
+_PLAN_MILESTONE_MARKERS = (
+    "计划",
+    "上线日期",
+    "交付日期",
+    "完成日期",
+    "工期",
+    "工作日",
+    "里程碑",
+    "上线",
+    "验收",
+    "交付",
+    "迁移",
+    "milestone",
+)
+_PLAN_STAGE_TEXT = re.compile(
+    r"(?:第[一二三四五六七八九十百\d]+阶段|阶段[一二三四五六七八九十百\d]+|"
+    r"phase\s*(?:one|two|three|four|five|\d+))",
+    re.IGNORECASE,
+)
+_CALENDAR_TEXT = re.compile(
+    rf"(?:{_RELATIVE_DATE_TOKEN}|"
+    r"(?<![A-Za-z\d./-])(?:\d{4}[/\-](?:0?[1-9]|1[0-2])[/\-](?:0?[1-9]|[12]\d|3[01])|"
+    r"(?:\d{4}\s*年\s*)?(?:0?[1-9]|1[0-2])\s*月\s*(?:0?[1-9]|[12]\d|3[01])\s*日?)"
+    r")",
+    re.IGNORECASE,
+)
+
+
+def is_project_plan_text(value: Any) -> bool:
+    """Return whether text explicitly describes a project plan or adjustment."""
+
+    if not isinstance(value, str):
+        return False
+    folded = value.casefold()
+    # Only a record whose own label/phrase is a plan mail, attachment, meeting
+    # record, or archive is adjacent.  A durable plan may still mention that
+    # an attachment must be migrated, an email sent, or a meeting held.
+    compact = re.sub(r"[\s:：，,；;。.!！?？()（）\[\]【】_-]+", "", folded)
+    plan_markers = tuple(
+        re.sub(r"\s+", "", marker.casefold()) for marker in _PROJECT_PLAN_MARKERS
+    )
+    adjacent_markers = tuple(
+        re.sub(r"\s+", "", marker.casefold())
+        for marker in _ADJACENT_PLAN_RECORD_MARKERS
+    )
+    direct_adjacent = any(
+        f"{plan}{adjacent}" in compact or f"{adjacent}{plan}" in compact
+        for plan in plan_markers
+        for adjacent in adjacent_markers
+    )
+    archived_plan = any(
+        f"{plan}已{archive}" in compact or f"已{archive}{plan}" in compact
+        for plan in plan_markers
+        for archive in ("存档", "归档", "archive")
+    )
+    sent_plan_mail = bool(
+        re.search(
+            r"(?:已发送|发送|sent|sentto).{0,12}"
+            r"(?:实施计划|项目计划|实施方案|项目方案|部署计划|上线计划|"
+            r"implementationplan|projectplan|implementationschedule)"
+            r"(?:邮件|email|mail)",
+            compact,
+            re.IGNORECASE,
+        )
+    )
+    if direct_adjacent or archived_plan or sent_plan_mail:
+        return False
+    return any(marker in folded for marker in _PROJECT_PLAN_MARKERS)
+
+
+def _split_future_use_clauses(value: str) -> list[str]:
+    # Semicolons, sentence boundaries, and explicit contrast/conjunction
+    # markers are the only boundaries used here.  Commas are deliberately not
+    # boundaries because implementation plans commonly enumerate milestones.
+    return [
+        part.strip()
+        for part in re.split(
+            r"[;；。！？!?\n]+|(?:同时|另外|此外|其次|并且|并要求|还需|另需|in addition|separately)",
+            value,
+            flags=re.IGNORECASE,
+        )
+        if part.strip()
+    ]
+
+
+def is_mixed_future_use_text(value: Any) -> bool:
+    """Detect a plan/rule combined with an independent dated todo.
+
+    This is a safety boundary, not a general-purpose classifier.  It only
+    returns true when the durable plan/rule and a strong dated action occur in
+    separate clauses.  A plan's own milestone or delivery date therefore
+    remains representable as one project memory.
+    """
+
+    if not isinstance(value, str):
+        return False
+    clauses = _split_future_use_clauses(value)
+    if len(clauses) < 2:
+        return False
+    durable_indexes = {
+        index
+        for index, clause in enumerate(clauses)
+        if is_project_plan_text(clause)
+        or any(marker in clause.casefold() for marker in _PROJECT_RULE_MARKERS)
+    }
+    if not durable_indexes:
+        return False
+    plan_milestone_context = any(
+        any(marker in clauses[index].casefold() for marker in ("里程碑", "milestone"))
+        for index in durable_indexes
+    )
+    plan_stage_context = sum(bool(_PLAN_STAGE_TEXT.search(clause)) for clause in clauses) >= 2
+    todo_indexes = {
+        index
+        for index, clause in enumerate(clauses)
+        if _CALENDAR_TEXT.search(clause)
+        and any(marker in clause.casefold() for marker in _DATED_TODO_MARKERS)
+        and not (
+            "完成" in clause
+            and any(marker in clause.casefold() for marker in _PLAN_MILESTONE_MARKERS)
+            and not any(
+                marker in clause.casefold() for marker in _INDEPENDENT_TODO_MARKERS
+            )
+        )
+        and not (
+            (plan_milestone_context or plan_stage_context)
+            and "完成" in clause
+            and not any(
+                marker in clause.casefold() for marker in _INDEPENDENT_TODO_MARKERS
+            )
+        )
+        # A second clause that is itself the project plan is a plan milestone,
+        # not evidence of a second future-use object.
+        and not is_project_plan_text(clause)
+    }
+    return bool(todo_indexes - durable_indexes)
+
+
 def is_aggregate_operational_text(value: Any) -> bool:
     """Recognize a combined operational digest, not one concrete work item.
 
@@ -910,6 +1134,23 @@ def validate_gate_output(
             raise ModelOutputError("duplicate candidate cannot be worth remembering", validation_detail="invalid_flags")
         if item["worth"] and candidate_type is None:
             raise ModelOutputError("worth candidate must have a type", validation_detail="invalid_type")
+        if item["worth"] and is_mixed_future_use_text(item["memory"]):
+            raise ModelOutputError(
+                "one candidate combines independent future uses",
+                validation_detail="mixed_future_use",
+            )
+        # A clearly named implementation plan/plan adjustment is a project
+        # memory, even when the model labels it as a generic fact.  Keep todo
+        # candidates intact: a task list that merely mentions a plan is still
+        # an actionable todo, while fact/event/other are the common accidental
+        # labels observed for durable plan changes.
+        if (
+            item["worth"]
+            and candidate_type in {"fact", "event", "other"}
+            and is_project_plan_text(item["memory"])
+        ):
+            item["type"] = "project"
+            candidate_type = "project"
         if "update_memory_id" in item and "duplicate_memory_id" in item:
             raise ModelOutputError(
                 "a candidate cannot set both duplicate_memory_id and update_memory_id",
@@ -1117,10 +1358,32 @@ def validate_summarize_output(
     _string(item["title"], "title")
     _string(item["body"], "body", multiline=True)
     _reject_relative_calendar_expression(item)
+    if is_mixed_future_use_text(f"{item['title']}\n{item['body']}"):
+        raise ModelOutputError(
+            "one summary combines independent future uses",
+            validation_detail="mixed_future_use",
+        )
     item["tags"] = _string_list(item["tags"], "tags")
     candidate_type = item["type"]
     if not isinstance(candidate_type, str) or candidate_type not in MEMORY_TYPES:
         raise ModelOutputError("invalid memory type", validation_detail="invalid_type")
+    inferred_project_type = (
+        candidate_type in {"fact", "event", "other"}
+        and is_project_plan_text(f"{item['title']}\n{item['body']}")
+    )
+    if inferred_project_type:
+        if expected_type is not None and expected_type != "project":
+            raise ModelOutputError(
+                "project plan summary cannot use a non-project gate type",
+                validation_detail="invalid_type",
+            )
+        if expected_target_type is not None and expected_target_type != "project":
+            raise ModelOutputError(
+                "project plan summary cannot update a non-project target",
+                validation_detail="invalid_type",
+            )
+        candidate_type = "project"
+        item["type"] = candidate_type
     if expected_type is not None and candidate_type != expected_type:
         raise ModelOutputError(
             "summary type differs from gate candidate",
@@ -1375,6 +1638,8 @@ __all__ = [
     "parse_strict_json",
     "is_aggregate_operational_text",
     "is_attachment_followup_only_text",
+    "is_project_plan_text",
+    "is_mixed_future_use_text",
     "normalize_relative_calendar_text",
     "parse_summarize_output",
     "parse_summary_output",

{memleaf-0.2.13 → memleaf-0.2.14/src/memleaf.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memleaf
-Version: 0.2.13
+Version: 0.2.14
 Summary: A local-first Markdown memory core for AI agents
 Author: memleaf contributors
 License-Expression: MIT
@@ -23,8 +23,8 @@ Dynamic: license-file
 
 [English](README.en.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **当前版本：0.2.13。**
-> 核心库、Vault、stdio MCP Server、初始化 CLI、模型路由、提炼流程、受控检索协议和宿主适配器已经实现。memleaf 0.2.13 通过 PyPI 分发。
+> **当前版本：0.2.14。**
+> 核心库、Vault、stdio MCP Server、初始化 CLI、模型路由、提炼流程、受控检索协议和宿主适配器已经实现。memleaf 0.2.14 通过 PyPI 分发。
 > **当前版本支持 Hermes 和 Codex。** Antigravity（反重力）不检测、不安装、不配置。
 
 ## 项目定位

{memleaf-0.2.13 → memleaf-0.2.14}/src/memleaf.egg-info/SOURCES.txt

@@ -67,6 +67,7 @@ tests/test_codex_native_cli.py
 tests/test_context_budget.py
 tests/test_credential_safety.py
 tests/test_cross_host_acceptance.py
+tests/test_extraction_quality_regressions.py
 tests/test_hermes_provider.py
 tests/test_hermes_stdio_transport.py
 tests/test_host_events.py