memleaf 0.2.95__tar.gz → 1.0.0__tar.gz

{memleaf-0.2.95 → memleaf-1.0.0}/CHANGELOG.md

@@ -5,6 +5,53 @@ All notable changes to memleaf are documented here.
 ## Unreleased
 
 
+## 1.0.0 — 2026-10-06
+
+### Stable public interfaces and production-only distributions
+
+- Define the 1.x compatibility scope for documented Python exports, CLI entry
+  points, MCP tools, configuration and persisted memory formats. Preserve the
+  0.2.96 production runtime, prompts, existing storage/recovery contracts and
+  five-request budget; no model request or Vault migration is added.
+- Move the isolated acceptance harness, synthetic acceptance suite and query
+  benchmark into the ignored local development directory. Remove the former
+  `python -m memleaf.acceptance` developer entry point from public artifacts;
+  normal Python/CLI/MCP interfaces and usage/discovery examples remain.
+- Remove obsolete packaged-tool instructions and prevent developer/private files
+  from re-entering wheel or source distributions through the release CI. Keep
+  the existing five-cell native installation matrix and publish verified bytes.
+- The 0.2.96 installed Hermes acceptance remains evidence for the unchanged
+  production runtime: four successful turns, eight Memleaf requests and 42,518
+  observed tokens. The 1.0.0 release verifies interface compatibility, isolated
+  installed regressions and clean artifact payloads separately; a stable version
+  is not a guarantee of perfect future model responses.
+
+
+## 0.2.96 — 2026-10-06
+
+### Preserve existing information when replacing memory bodies
+
+- Expose omitted old prose to semantic review for UPDATE and every MERGE
+  member. Require dispositions bound to the final body or new user evidence;
+  boolean body approval alone cannot erase existing information. Unconfirmed
+  replacements remain DEFERRED with the previous memory preserved.
+- Allow the existing review to repair the full body while retaining candidate
+  identity, evidence and other fields. Check the final replacement, accept valid
+  redundant coverage, and isolate invalid optional maintenance corrections from
+  independently valid decisions. Protect manual partial repairs as well.
+- Reuse the existing review request for ongoing actions; add a second review
+  only where a previously unreviewed fact update could omit existing prose.
+  Ordinary fact additions and repeated no-change turns remain single-request.
+  No third review, automatic retry or request-budget increase is added.
+- Validation: 122 affected checks and isolated frozen-candidate replays passed.
+  The final three live Hermes turns preserved unchanged requirements, avoided
+  duplicate writes and applied only an explicitly withdrawn requirement:
+  five Memleaf requests, 30,821 observed tokens, no issues or unresolved items.
+  Intermediate development failures were preserved separately; historical raw
+  review responses were not reconstructed. Equivalence still relies on semantic
+  review, while Core verifies complete coverage and source/quote bindings.
+
+
 ## 0.2.95 — 2026-10-05
 
 ### Preserve turn identity around internal Hermes continuations

{memleaf-0.2.95 → memleaf-1.0.0}/MANIFEST.in

@@ -5,9 +5,8 @@ include install.ps1
 include README.md
 include README.en.md
 recursive-include examples *.md *.ndjson *.py
-recursive-exclude examples *_acceptance.py
+recursive-exclude examples *_acceptance.py query_benchmark.py incremental_acceptance.json
 prune tests
+prune tests_public
 
 include docs/*.md
-
-include examples/incremental_acceptance.json

{memleaf-0.2.95/src/memleaf.egg-info → memleaf-1.0.0}/PKG-INFO

@@ -1,10 +1,11 @@
 Metadata-Version: 2.4
 Name: memleaf
-Version: 0.2.95
+Version: 1.0.0
 Summary: A local-first Markdown memory core for AI agents
 Author: memleaf contributors
 License-Expression: MIT
 Keywords: ai-agents,local-first,markdown,memory
+Classifier: Development Status :: 5 - Production/Stable
 Classifier: Intended Audience :: Developers
 Classifier: Programming Language :: Python :: 3
 Classifier: Programming Language :: Python :: 3 :: Only
@@ -23,7 +24,7 @@ Dynamic: license-file
 
 [English](README.en.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **版本：0.2.95。**
+> **稳定版：1.0.0。** 公开接口进入稳定维护，版本与兼容范围见[稳定性约定](docs/stability.md)。
 > 自动提炼和复核共用同一条“未来记忆价值”标准：模型综合未来复用、信息增量、再次读取时的直接可用性和忘记成本，只保留对未来理解、判断或行动有实质影响的最小核心；没有明确价值的信息不提炼，不按具体业务场景硬编码排除。
 > 提炼先比较旧事项中的编号、目标、进展、责任和期限；补全未知信息也应维护原事项，另建独立待办不能替代项目事实更新。确认重复的记忆在现有处理请求中合并，保留历史与来源；项目和独立任务保持各自生命周期，Core 不按标题强行合并。
 > 新请求兼容等价的字段表示；未知执行人保持 `null`，新增或变更负责人和期限需要真实来源；等待条件在正文维护，不再提炼 `waiting_on`。行动及相关字段经过有界语义复核，提炼与复核共用每份工作的最多五次模型请求额度，成功即停止；复核失败不写入。来源时间用于解析相对日期，不补造事实日期。

{memleaf-0.2.95 → memleaf-1.0.0}/README.en.md

@@ -4,7 +4,7 @@
 
 [中文](README.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **Version: 0.2.95.**
+> **Stable release: 1.0.0.** Public interfaces enter stable maintenance; see the [stability policy](docs/stability.md) for versioning and compatibility scope.
 > Automatic extraction and review share one future-value standard: the model weighs likely reuse, information gain, direct usability when read again, and the cost of forgetting. It keeps only the smallest core that can materially help future understanding, decisions, or actions; information without clear value is not extracted, and no business-specific exclusion rule is hard-coded.
 > Extraction compares existing identifiers, objectives, progress, responsibility and deadlines first. Filling an unknown also maintains the original item; a separate task does not replace updating shared project facts. Confirmed duplicates can be merged in the existing processing requests with their sources and history preserved. Independent project and task lifecycles remain separate; Core does not merge matching titles.
 > New requests accept equivalent field representations. Unknown executors remain `null`; new or changed responsibility and deadlines require actual source evidence. Dependencies belong in the body; `waiting_on` is no longer extracted. Proposed actions and related fields undergo bounded semantic verification within a shared five-request limit per work item, stopping on success; failed verification blocks writes. Source time resolves relative dates without inventing factual dates.

{memleaf-0.2.95 → memleaf-1.0.0}/README.md

@@ -4,7 +4,7 @@
 
 [English](README.en.md) · [PyPI](https://pypi.org/project/memleaf/) · [GitHub](https://github.com/miffyblueboo/memleaf)
 
-> **版本：0.2.95。**
+> **稳定版：1.0.0。** 公开接口进入稳定维护，版本与兼容范围见[稳定性约定](docs/stability.md)。
 > 自动提炼和复核共用同一条“未来记忆价值”标准：模型综合未来复用、信息增量、再次读取时的直接可用性和忘记成本，只保留对未来理解、判断或行动有实质影响的最小核心；没有明确价值的信息不提炼，不按具体业务场景硬编码排除。
 > 提炼先比较旧事项中的编号、目标、进展、责任和期限；补全未知信息也应维护原事项，另建独立待办不能替代项目事实更新。确认重复的记忆在现有处理请求中合并，保留历史与来源；项目和独立任务保持各自生命周期，Core 不按标题强行合并。
 > 新请求兼容等价的字段表示；未知执行人保持 `null`，新增或变更负责人和期限需要真实来源；等待条件在正文维护，不再提炼 `waiting_on`。行动及相关字段经过有界语义复核，提炼与复核共用每份工作的最多五次模型请求额度，成功即停止；复核失败不写入。来源时间用于解析相对日期，不补造事实日期。

{memleaf-0.2.95 → memleaf-1.0.0}/docs/acceptance-evidence-guide.md

@@ -24,7 +24,7 @@ the suite count or from module-name references.
 
 ## Current route semantics
 
-The v0.2.67 candidate has one processing engine: incremental. New Vaults default
+The supported runtime has one processing engine: incremental. New Vaults default
 both `process.automatic_pipeline` and `process.remember_pipeline` to
 `incremental`. A retained `legacy` value from an older Vault is not executable;
 migration preflight reports it as `legacy_pipeline_configuration`, and runtime
@@ -38,23 +38,17 @@ preflight result.
 
 ## Preparation with no paid calls
 
-Use the existing entry point on a known candidate installation:
-
-```sh
-python -m memleaf.acceptance --suite examples/incremental_acceptance.json --repeat 5
-```
-
-This is plan-only: it validates the suite and reports hashes and request bounds;
-it does not load a backend configuration or create a Vault. `process --dry-run`
-is different and may call the configured model. Do not use it as a free preview.
-
-The shipped suite is public synthetic regression material. Take case count, turn
-count, repetitions, normalized suite hash, prompt hash and request bounds from
-this command on the selected candidate; do not inherit numbers from an older
-report. Also record the original suite file SHA-256: it is not the normalized
-semantic suite hash. The reported bound is not a spending authorization. A private
-held-out suite requires its own provenance, permission and count. Merely adding
-a label `holdout` does not prove the examples were unseen in tuning.
+Acceptance runners and synthetic fixtures are local development tools, not
+entry points shipped with an installed package. Use a controlled development
+checkout to plan the approved suite without constructing a model, loading
+credentials or opening a production Vault. Record the suite file and normalized
+hashes, prompt hash, repetitions and request bound from that exact checkout.
+
+`process --dry-run` is a product processing preview and may call the configured
+model. It is not a free substitute for local acceptance planning. A request bound
+is not a spending authorization; a private held-out suite also needs provenance
+and explicit data/route permission. A `holdout` label alone proves nothing about
+whether those examples were used in development.
 
 Do not execute until the actual route/model, local credential reference, suite
 hash, repeats and total allowance are approved. Keep credentials local. Do not

{memleaf-0.2.95 → memleaf-1.0.0}/docs/host-acceptance-runbook.md

@@ -46,14 +46,11 @@ Do not return API keys, tokens, authentication headers, credential-bearing URLs,
 full configuration, chat bodies, native-memory text or backups. Keep raw paths
 and necessary private fingerprints in local evidence rather than public CI.
 
-Prepare the existing public suite with the exact candidate Python:
+Acceptance planning uses the exact candidate Python, locally retained tooling and approved
+fixtures. Neither the runner nor its suite is included in an installed package;
+do not construct a production runtime to replace a missing development tool.
 
-```sh
-python -m memleaf.acceptance --suite examples/incremental_acceptance.json --repeat 5
-```
-
-The suite path must point to the selected checkout; an installed wheel alone need
-not place `examples/` in the current directory. Planning accepts no backend,
+The suite path must point to the controlled local development checkout. Planning accepts no backend,
 output or execution flags. Save stdout by the caller into the evidence directory,
 not with the tool's `--output` option, which belongs to execution. Record both
 file SHA-256 and the normalized suite digest in the result. This planning command

{memleaf-0.2.95 → memleaf-1.0.0}/docs/installed-artifact-verification.md

@@ -11,12 +11,15 @@ The build job:
 
 1. compiles `src/` with Python's standard-library compiler;
 2. builds one wheel and one source distribution;
-3. uploads only those two distributions as the workflow artifact.
+3. rejects development harnesses, tests, private runtime directories and local
+   environment files in either artifact;
+4. uploads only those two distributions as the workflow artifact.
 
 The `verify-artifacts` job downloads those exact bytes and installs each artifact
 in its own fresh virtual environment. It imports `memleaf` and checks the
 reported package version. It does not read a source-tree checkout, run a model,
-access a production Vault or create a release.
+access a production Vault or create a release. The former developer-only
+`memleaf.acceptance` module must not be importable from either installation.
 
 The matrix is:
 

memleaf-1.0.0/docs/processing-quality-acceptance.md

@@ -0,0 +1,33 @@
+# Processing quality and release checks
+
+Memleaf uses the incremental capture, planner, compiler, review and commit path.
+Extraction, review and eligible maintenance share each work item's durable
+five-request budget. A successful result stops further extraction requests;
+missing support remains unknown or unresolved rather than being invented.
+
+## Separate kinds of evidence
+
+- Deterministic checks verify source binding, identity, permissions, state,
+  recovery and request accounting without external model calls.
+- Source-based review of actual retained memories verifies whether useful facts
+  were kept, unchanged requirements survived updates and independent matters
+  stayed independent. A JSON or schema pass alone is insufficient.
+- Real-host acceptance verifies the loaded Core/Provider, actual session model,
+  capture and background processing through the installed host.
+- Artifact CI verifies source syntax, clean distribution payloads and native
+  wheel/sdist installation. It does not establish model quality.
+
+Development acceptance runners, synthetic fixtures, benchmarks and private
+reports are retained locally and are not shipped as package modules, examples
+or user-facing diagnostic tools. The normal runtime APIs and configuration are
+used for supported product operations.
+
+Real model acceptance records the actual route, all attempted runs, request
+counts and obtainable usage metrics. Scripted replay is reported separately.
+Unknown usage remains unknown; failed attempts are not relabelled as passing
+checks. Normal `process --dry-run` may call the configured model and must not be
+used as a free development preview.
+
+See the [evidence guide](acceptance-evidence-guide.md),
+[installed artifact checks](installed-artifact-verification.md) and
+[stability policy](stability.md).

{memleaf-0.2.95 → memleaf-1.0.0}/docs/query-performance.md

@@ -39,12 +39,11 @@ snapshot nor claimed to be eliminated.
 
 ## Reproduce an isolated local measurement
 
-```sh
-# Source checkout; for an installed wheel omit PYTHONPATH.
-PYTHONPATH=src python examples/query_benchmark.py --sizes 100 1000 10000 --repeat 5
-```
+The isolated benchmark is retained only in the local development directory.
+It is not distributed as an example or installed package tool. A measurement
+must record the exact source or installed artifact being imported.
 
-The example accepts sizes/repetitions only, not a production Vault, model config
+The local runner accepts sizes/repetitions only, not a production Vault, model config
 or credentials. Each size uses a fresh TemporaryDirectory, 1..10,000 synthetic
 independent Todo heads, and a fixed source clock. Fixture creation is outside the
 timed section and does not rebuild an index for each inserted file. One warmup is
@@ -86,9 +85,9 @@ unchanged recheck parsing fell from N to zero. No business records were omitted
 to improve latency. At 10,000 heads the fixture contains 7,550,000 bytes; it is not
 a sample of private user conversations or an actual production Vault.
 
-The harness is included in sdist through the existing examples manifest. Running
-it against an installed wheel must use that wheel's imports, not an editable
-checkout. The measurements do not cover provider latency, tokens, recall/semantic
+The harness is excluded from both wheel and sdist. Local runs against an
+installed wheel must use that wheel's imports, not an editable checkout.
+The measurements do not cover provider latency, tokens, recall/semantic
 quality, native Hermes installation or Windows/macOS performance. Those remain
 separate G5 acceptance gates; successful timings never authorize switching.
 

memleaf-1.0.0/docs/stability.md

@@ -0,0 +1,56 @@
+# Memleaf 1.x stability policy
+
+Memleaf 1.0.0 begins stable maintenance of its documented public interfaces.
+Version numbers follow [Semantic Versioning](https://semver.org/): compatible
+fixes use a patch release, compatible additions use a minor release, and
+incompatible public API changes require a new major version. Stable maintenance
+does not imply that every future conversation or model response will be correct.
+
+## Public interfaces
+
+The compatibility scope consists of:
+
+- The names exported by `memleaf.__all__`, their documented constructors and
+  functions, and the documented public methods of `Memleaf`, `Core`,
+  `MemoryService`, `Memory` and `Vault`.
+- The `memleaf`, `memleaf-mcp` and `memleaf-mcpw` entry points, documented CLI
+  commands and options, and their documented JSON fields and error codes.
+- The published MCP tool names, accepted parameters and documented result
+  fields. Clients should tolerate additional optional fields.
+- Documented configuration keys and the Markdown/frontmatter memory format.
+  Existing committed memories and persistent recovery state must retain their
+  identity, provenance and interpretation across compatible upgrades. Any
+  required migration must be documented before it is applied.
+
+Module internals, names beginning with `_`, internal planner/reviewer protocols,
+development harnesses and exact natural-language model wording are outside this
+scope. An internal protocol change may fence pending model responses and require
+an explicit recovery action; it must not silently reinterpret saved work, erase
+unresolved outcomes or reset consumed request budgets.
+
+## Upgrade to 1.0.0
+
+The 1.0.0 runtime preserves the 0.2.96 production interfaces, configuration,
+storage formats, prompts and five-request limit. It requires no new model calls
+or Vault migration from 0.2.96. Earlier unsupported pipeline settings and pending
+work retain their existing migration/recovery requirements; the version number
+does not mark them complete.
+
+Update Core and the copied Hermes Provider together through `memleaf install`,
+then restart Hermes to load the installed files. The existing build/protocol
+compatibility checks still apply. Codex connections must reconnect or restart
+their MCP process after an upgrade.
+
+The former `memleaf.acceptance` module, synthetic acceptance suite and query
+benchmark are development tools. They are removed from the public package and
+repository in 1.0.0 and retained only in the local development directory. Normal
+usage examples, MCP discovery examples and artifact verification in CI remain.
+
+## Release evidence
+
+Each release builds one wheel and source distribution, verifies their payloads
+and installs those same artifacts on the existing native CI matrix before
+publishing. Local regressions and real-host/model acceptance are separate gates:
+green installation checks do not prove model semantics, and a successful host
+conversation does not replace package verification. Development tests and private
+acceptance reports are kept outside the public source and distributions.

{memleaf-0.2.95 → memleaf-1.0.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "memleaf"
-version = "0.2.95"
+version = "1.0.0"
 description = "A local-first Markdown memory core for AI agents"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -13,6 +13,7 @@ license-files = ["LICENSE"]
 authors = [{name = "memleaf contributors"}]
 keywords = ["ai-agents", "local-first", "markdown", "memory"]
 classifiers = [
+    "Development Status :: 5 - Production/Stable",
     "Intended Audience :: Developers",
     "Programming Language :: Python :: 3",
     "Programming Language :: Python :: 3 :: Only",

{memleaf-0.2.95 → memleaf-1.0.0}/src/memleaf/__init__.py

@@ -1,6 +1,6 @@
 """Local-first Markdown memory core for AI agents."""
 
-__version__ = "0.2.95"
+__version__ = "1.0.0"
 
 from .config import DEFAULT_CONFIG, default_config, load_config, save_config
 from .frontmatter import FrontmatterError, dump_frontmatter, dump_yaml, load_yaml, parse_frontmatter

memleaf-1.0.0/src/memleaf/body_preservation.py

@@ -0,0 +1,112 @@
+"""Account for omitted old prose without guessing which facts remain valid.
+
+Text fragments are comparison aids, not a parser for facts. The existing
+semantic review judges equivalence or an authorized change; Core checks that
+every omission has a real, source-bound disposition for the final candidate.
+"""
+from __future__ import annotations
+
+import re
+
+
+def _compact(text):
+    return re.sub(r"\s+", "", text)
+
+
+def omissions(row, original):
+    if not isinstance(row, dict):
+        return []
+    action = row.get("action")
+    action = action.strip().upper() if isinstance(action, str) else ""
+    if action not in {"UPDATE", "MERGE"}:
+        return []
+    patch = row.get("patch")
+    if not isinstance(patch, dict):
+        return []
+    # Retraction erases the stored body even when the model omits patch.body.
+    body = "" if patch.get("validity") == "retracted" else patch.get("body")
+    if not isinstance(body, str):
+        return []
+    refs = [row.get("target")]
+    if action == "MERGE" and isinstance(row.get("duplicates"), list):
+        refs.extend(row["duplicates"])
+    refs = [ref.strip() for ref in refs if isinstance(ref, str)]
+    new = _compact(body)
+    missing = []
+    seen = set()
+    for memory in original.get("memories", []):
+        if memory.get("ref") not in refs or memory.get("validity", "valid") != "valid":
+            continue
+        old = memory.get("body", "")
+        if not isinstance(old, str):
+            continue
+        # Language-neutral punctuation/line boundaries. No names, keywords or
+        # expected facts influence the comparison. Short fragments count too.
+        for text in re.split(r"[。！？!?；;\n，,]+|(?<=\.)\s+", old):
+            text = text.strip()
+            key = (memory["ref"], text)
+            if text and _compact(text) not in new and key not in seen:
+                missing.append({"target": key[0], "old": text})
+                seen.add(key)
+    return missing
+
+
+def covered(row, original, coverage=None):
+    """Missing/malformed proofs defer this proposal, never erase the head."""
+    missing = omissions(row, original)
+    if not missing:
+        return True
+    if not isinstance(coverage, list):
+        return False
+    expected = {(part["target"], part["old"]) for part in missing}
+    # A reviewer may restore a fragment and still include its proof. Check
+    # against the real old fragments, not an exact output-count convention.
+    all_old = omissions({**row, "patch": {"body": ""}}, original)
+    allowed = {(part["target"], part["old"]) for part in all_old}
+    verified = set()
+    body = "" if row["patch"].get("validity") == "retracted" else row["patch"].get("body", "")
+    evidence = {e["ref"]: e for e in original.get("evidence", [])}
+    for proof in coverage:
+        if not isinstance(proof, dict):
+            return False
+        key = (proof.get("target"), proof.get("old"))
+        if not all(isinstance(value, str) for value in key) or key not in allowed or key in verified:
+            return False
+        if set(proof) == {"target", "old", "body"}:
+            quote = proof["body"]
+            if not isinstance(quote, str) or not quote.strip() or quote not in body:
+                return False
+        elif set(proof) == {"target", "old", "source"}:
+            selection = proof["source"]
+            if not isinstance(selection, dict) or set(selection) != {"ref", "text"}:
+                return False
+            if not isinstance(selection["ref"], str):
+                return False
+            event = evidence.get(selection.get("ref"))
+            quote = selection.get("text")
+            # An assistant suggestion cannot authorize discarding user facts.
+            if (event is None or event.get("use") != "new" or event.get("role") != "user"
+                    or event["ref"] not in row.get("evidence", [])
+                    or not isinstance(quote, str) or not quote.strip() or quote not in event.get("text", "")):
+                return False
+        else:
+            return False
+        verified.add(key)
+    return expected <= verified
+
+
+def defer_unreviewed(response, snapshot):
+    """A manual partial repair has no model reviewer; preserve unsafe heads."""
+    from .incremental_semantics import envelope
+    from .validation import parse_strict_json
+    from .incremental_protocol import compile_incremental
+    import json
+    draft = envelope(parse_strict_json(response))
+    original = snapshot.model_input()
+    for i, row in enumerate(draft["items"]):
+        if (omissions(row, original)
+                and not compile_incremental(json.dumps({"items": [row]}), snapshot)["issues"]):
+            draft["items"][i] = {"action": "DEFERRED", "evidence": row.get("evidence", []),
+                                 "reason": "missing_context",
+                                 "need": "The body replacement has unverified omissions from the previous memory."}
+    return json.dumps(draft, ensure_ascii=False, separators=(",", ":"))

{memleaf-0.2.95 → memleaf-1.0.0}/src/memleaf/hermes_provider/plugin.yaml

@@ -1,5 +1,5 @@
 name: memleaf
-version: 0.2.95
+version: 1.0.0
 description: "Hermes-native external MemoryProvider for local-first Markdown memory shared with the memleaf MCP server."
 hooks:
   - prefetch

{memleaf-0.2.95 → memleaf-1.0.0}/src/memleaf/incremental_execution.py

@@ -232,11 +232,14 @@ def _drive(service, run_id, token, backend, calls, metric_ordinals):
                     return public_result(run, calls=calls[0], metric_ordinals=metric_ordinals)
                 if run["status"] in {"response_ready", "committing"}:
                     try:
+                        if (run.get("partial_used") and run.get("partial_recovery", {}).get("mode") == "repair"):
+                            from .body_preservation import defer_unreviewed
+                            run["response"] = defer_unreviewed(run["response"], snapshot)
                         compiled = compile_incremental(run["response"], snapshot)
                         from .incremental_semantics import needs_review, review_request
                         if (not (run.get("partial_used") and run.get("partial_recovery", {}).get("mode") == "repair")
                                 and run.get("semantic_stage") != "reviewed"
-                                and needs_review(compiled, run["response"])):
+                                and needs_review(compiled, run["response"], snapshot)):
                             if run["reserved_requests"] >= MAX_INCREMENTAL_REQUESTS:
                                 _finish(service, processed, run, "blocked", "semantic_review_budget_exhausted")
                                 return public_result(run, calls=calls[0], metric_ordinals=metric_ordinals)
@@ -391,7 +394,7 @@ def _drive(service, run_id, token, backend, calls, metric_ordinals):
             elif not run.get("partial_used"):
                 try:
                     from .incremental_semantics import needs_review
-                    run["semantic_review_required"] = needs_review(compile_incremental(response, snapshot), response)
+                    run["semantic_review_required"] = needs_review(compile_incremental(response, snapshot), response, snapshot)
                 except (ValueError, TypeError):
                     run["semantic_review_required"] = False
             save_run(service, processed, run)

{memleaf-0.2.95 → memleaf-1.0.0}/src/memleaf/incremental_semantics.py

@@ -10,8 +10,10 @@ from copy import deepcopy
 import json
 
 from .validation import parse_strict_json, ModelOutputError
+from .body_preservation import omissions, covered
 
-REVIEW_SYSTEM = """核验拟写入的持续行动。original_input 是唯一事实和权限来源；draft 是待审候选，不能当作事实。返回 JSON 对象，必填 decisions，值为数组；每项有 item（整数）、keep（布尔值）、field_support（对象）。按输入 fields_to_review 的实际字段填 field_support，不能复制其他事项的字段名。
+REVIEW_SYSTEM = """核验拟写入的记忆及持续行动。original_input 是唯一事实和权限来源；draft 是待审候选，不能当作事实。返回 JSON 对象，必填 decisions，值为数组；每项有 item（整数）、keep（布尔值）、field_support（对象）。按输入 fields_to_review 的实际字段填 field_support，不能复制其他事项的字段名。
+存在 body_omissions 时，keep:true 的 UPDATE/MERGE 必须逐项说明省略的去向；仅 body:true 不足以通过。body_omissions 按 item 列出拟议正文中不再原样出现的旧片段（比较辅助，不表示已经失效）。逐项核对完整含义，不能因主题相同、进展重复或用户未再次提及就删除。需要补回或修正文句时，直接在 decision 增加 body（修正后的完整正文字符串），field_support.body=true；不必复制 action、target、evidence，也不要把该 target 放入 updates。Core 沿用该候选的身份、证据和其他字段；已有 replace 方式仍可用，但不能同时使用 body 和 replace。对最终 UPDATE/MERGE（包括 body、replace 和 updates 修正）的剩余省略片段，在对应 decision/updates 项增加 body_coverage 数组：每项 {"target":"旧目标ref","old":"旧片段原文","body":"最终正文中完整承接该片段含义的逐字引文"}，或 {"target":"旧目标ref","old":"旧片段原文","source":{"ref":"new用户证据ref","text":"明确纠正/撤回该内容的逐字引文"}}。body 引文必须完整等价且真实存在于最终正文；不能只引用旧目标中的文字却忘了将它写入修正正文。source 不能引用旧上下文或助手建议。修正后原样保留的片段无需列入。MERGE 对所有被合并目标逐项核对，撤回整个目标也需明确用户依据。无法确认则 keep:false；不要猜测或提供主题相符但实际无关的引文。没有剩余省略时省略 body_coverage。
 每个 draft.items 都必须有且只有一个 decision，item 是从 0 开始的位置。keep 判断该事项本身是否由原文确立且持续值得保留。临时操作及助手额外建议不能产生用户长期待办，原话只证明另一个操作时 keep:false。独立事实不因其他候选有误而拒绝。
 keep:true 时，field_support 必须逐个列出候选 memory/patch 中的业务字段，每个布尔值判断该字段的拟议值是否受来源支持；不列 responsibility_basis、effective、at、reopen 等引用/控制字段。没有 replace 的 NO_CHANGE/DEFERRED/NO_MEMORY 的 field_support 为 {}。keep:false 时 field_support 为 {}。unverified_fields 是系统发现未能验证引用的可选责任字段，必须判为 false；不能补造引用。这不影响其他有依据的事实或任务。归属/客户尚未确定不自动表示行动被阻塞。
 不能仅检查引文存在：必须判断原话是否确立该具体行动、责任、状态和期限。assistant 的建议/计划不能赋予 user 新义务。转发、协调、记录不等于本人执行；支持旧任务不等于接受新安排。
@@ -45,15 +47,23 @@ def fields(row):
     return {key for key in payload if key not in _METADATA} if isinstance(payload, dict) else set()
 
 
-def needs_review(compiled, response=None):
+def needs_review(compiled, response=None, snapshot=None):
     if response is not None:
         rows = envelope(parse_strict_json(response))["items"]
+        original = snapshot.model_input() if snapshot is not None else None
+        update_ids = {op.get("target") for op in compiled["operations"] if op["action"] == "UPDATE"}
+        writable_updates = ({ref for ref, target in snapshot.state()["targets"].items()
+                             if target["memory"]["memory_id"] in update_ids} if snapshot is not None else set())
         actions = {str(row.get("action", "")).strip().upper() for row in rows if isinstance(row, dict)}
         if "MERGE" in actions or {"UPDATE", "NO_CHANGE"} <= actions:
             return True
         for row in rows:
             if not isinstance(row, dict):
                 continue
+            target = row.get("target")
+            if (original is not None and isinstance(target, str) and target.strip() in writable_updates
+                    and omissions(row, original)):
+                return True
             action = row.get("action")
             action = action.strip().upper() if isinstance(action, str) else ""
             payload = row.get("memory") if action == "CREATE" else row.get("patch")
@@ -118,10 +128,12 @@ def review_request(original, response, snapshot=None):
             row["action"] = row["action"].strip().upper()
         rows.append(row)
     draft["items"] = rows
+    missing = [omissions(row, source) for row in rows]
     return {"system": REVIEW_SYSTEM,
             "user": json.dumps({"original_input": source, "draft": draft,
                                 "fields_to_review": [sorted(fields(row)) for row in draft["items"]],
                                 "unverified_fields": [unverified_fields(row, source) for row in draft["items"]],
+                                **({"body_omissions": missing} if any(missing) else {}),
                                 **({"invalid_representation_items": invalid} if invalid else {})},
                                ensure_ascii=False, separators=(",", ":"))}
 
@@ -164,7 +176,7 @@ def apply_review(request, response):
                 rejected.append(deepcopy(row))
         for decision in decisions:
             if (not isinstance(decision, dict) or not {"item", "keep", "field_support"} <= set(decision)
-                    or set(decision) - {"item", "keep", "field_support", "replace"}):
+                    or set(decision) - {"item", "keep", "field_support", "replace", "body", "body_coverage"}):
                 raise ValueError
             i = decision["item"]
             if type(i) is not int or not 0 <= i < len(rows) or i in positions or type(decision["keep"]) is not bool:
@@ -174,6 +186,14 @@ def apply_review(request, response):
             if not isinstance(support, dict) or any(type(v) is not bool for v in support.values()):
                 raise ValueError
             row = rows[i]
+            if "body" in decision:
+                if (not decision["keep"] or "replace" in decision
+                        or i in data.get("invalid_representation_items", [])
+                        or row.get("action") not in {"CREATE", "UPDATE", "MERGE"}
+                        or "body" not in fields(row) or support.get("body") is not True
+                        or not isinstance(decision["body"], str)):
+                    raise ValueError
+                row["memory" if row["action"] == "CREATE" else "patch"]["body"] = decision["body"]
             if "replace" in decision:
                 if i in data.get("invalid_representation_items", []) or not decision["keep"]:
                     raise ValueError
@@ -253,6 +273,9 @@ def apply_review(request, response):
             if row["action"].upper() == "UPDATE" and not payload:
                 discard(row)
                 rows[i] = None
+            elif not covered(row, data["original_input"], decision.get("body_coverage")):
+                discard(row)
+                rows[i] = None
         rows = [r for r in rows if r is not None]
         additions = review.get("updates", [])
         from .incremental_protocol import MAX_ITEMS
@@ -260,12 +283,24 @@ def apply_review(request, response):
             raise ValueError
         used = {row.get("target") for row in rows if isinstance(row, dict)}
         for entry in additions:
-            if not isinstance(entry, dict) or set(entry) != {"row", "field_support"}:
-                raise ValueError
-            row = _maintenance_update(entry["row"], entry["field_support"], data["original_input"])
-            if row["target"] in used:
-                raise ValueError
-            rows.append(row)
+            try:
+                if (not isinstance(entry, dict) or not {"row", "field_support"} <= set(entry)
+                        or set(entry) - {"row", "field_support", "body_coverage"}):
+                    raise ValueError
+                row = _maintenance_update(entry["row"], entry["field_support"], data["original_input"])
+                if row["target"] in used:
+                    raise ValueError
+            except (KeyError, TypeError, ValueError):
+                # Optional correction is not an authority grant. Reject it
+                # explicitly without erasing independently valid decisions.
+                rows.append({"action": "DEFERRED", "evidence": [e["ref"] for e in data["original_input"]["evidence"]
+                             if e["use"] == "new"], "reason": "missing_context",
+                             "need": "The additional maintenance correction is not source-verified."})
+                continue
+            if covered(row, data["original_input"], entry.get("body_coverage")):
+                rows.append(row)
+            else:
+                discard(row)
             used.add(row["target"])
         for rejected_row in rejected:
             # The same review can supply a complete, independently validated