@hupan56/wlkj 3.1.32 → 3.3.0

package/templates/qoder/scripts/domain/kg/build/kg_signatures.py

@@ -0,0 +1,169 @@
+# -*- coding: utf-8 -*-
+# v3.0 路径自举: 引导到 common/bootstrap, 统一 sys.path 逻辑
+import os as _o, sys as _s
+_f = _o.path.abspath(__file__)
+for _ in range(10):
+    _f = _o.path.dirname(_f)
+    _cp = _o.path.join(_f, 'foundation')
+    if _o.path.isfile(_o.path.join(_cp, 'bootstrap.py')):
+        break
+if _cp not in _s.path: _s.path.insert(0, _cp)
+from bootstrap import setup; setup()
+
+"""
+File-signature based incremental change detection (P0: fingerprint 三级增量).
+
+Goal: skip re-extraction for files whose changes don't affect the graph.
+The knowledge graph only records *topology* (who clicks/calls/handles whom),
+not implementations — so a file that only changed comments/whitespace has
+zero impact on the graph and can be safely skipped.
+
+Three levels (mirrors Understand-Anything's fingerprint approach, adapted to
+our regex/topology-only graph):
+  - NONE       : content_hash unchanged → file is byte-identical, skip entirely.
+  - COSMETIC   : structural_hash unchanged but content_hash changed
+                 → only comments/whitespace differ, skip re-extraction.
+  - STRUCTURAL : structural_hash changed → real code changed, re-extract.
+
+The signature is intentionally coarse (strip comments + all whitespace, hash
+the rest) — language-agnostic and robust. Worst case (e.g. pure variable
+rename) degrades to STRUCTURAL = re-extract = current behaviour, never wrong.
+"""
+
+import hashlib
+import os
+import re
+
+
+# ── Comment stripping (language-aware, single regex each) ────────────────────
+# Block comments /* ... */ and HTML comments <!-- ... --> span both JS-family
+# (.ts/.vue/.tsx/.jsx) and SQL (-- / /* */). Line comments vary by language.
+
+# /* block */ and <!-- block --> — shared by JS-family + SQL + Vue template
+_RE_BLOCK = re.compile(r'/\*.*?\*/', re.DOTALL)
+_RE_HTML_BLOCK = re.compile(r'<!--.*?-->', re.DOTALL)
+
+# // line comment — TS/JS/Vue script + Java. Robust against URL-ish tokens
+# (only strips when // is at start of a token, not inside a string — see below).
+_RE_LINE_SLASH = re.compile(r'(^|[\s;{}()\[\],])//[^\n]*')
+
+# # line comment — Python / shell / YAML / .properties
+_RE_LINE_HASH = re.compile(r'(^|[\s;{}()\[\],])#[^\n]*')
+
+# -- line comment — SQL / XML DTD
+_RE_LINE_DASH = re.compile(r'(^|[\s;{}()\[\],])--[^\n]*')
+
+
+def strip_comments(content, suffix):
+    """Remove comments from source text by file suffix.
+
+    Args:
+        content: source text (str)
+        suffix: file suffix lowercased with dot, e.g. '.vue', '.ts', '.java'
+
+    Returns:
+        Text with comments removed. Block comments always stripped for
+        JS-family/SQL; line-comment style picked by language.
+    """
+    s = content
+    # Block comments: JS-family + SQL share /* */; Vue/HTML share <!-- -->.
+    if suffix in ('.ts', '.tsx', '.js', '.jsx', '.vue', '.java', '.sql'):
+        s = _RE_BLOCK.sub(' ', s)
+    if suffix in ('.vue', '.html', '.xml'):
+        s = _RE_HTML_BLOCK.sub(' ', s)
+    # Line comments by language family.
+    if suffix in ('.ts', '.tsx', '.js', '.jsx', '.vue', '.java'):
+        s = _RE_LINE_SLASH.sub(r'\1', s)
+    elif suffix in ('.py', '.sh', '.yml', '.yaml', '.properties', '.cfg', '.ini'):
+        s = _RE_LINE_HASH.sub(r'\1', s)
+    elif suffix == '.sql':
+        s = _RE_LINE_DASH.sub(r'\1', s)
+    return s
+
+
+def _strip_all_whitespace(s):
+    """Remove every whitespace char so formatting/indentation never affects hash."""
+    return re.sub(r'\s+', '', s)
+
+
+def compute_hashes(filepath):
+    """Compute (content_hash, structural_hash) for a file.
+
+    Args:
+        filepath: absolute or relative path to the source file.
+
+    Returns:
+        (content_hash, structural_hash) — both hex sha256, or (None, None) if
+        the file can't be read.
+    """
+    try:
+        with open(filepath, 'rb') as f:
+            raw = f.read()
+    except Exception:
+        return (None, None)
+    content_hash = hashlib.sha256(raw).hexdigest()
+
+    suffix = os.path.splitext(filepath)[1].lower()
+    try:
+        text = raw.decode('utf-8', errors='replace')
+    except Exception:
+        text = ''
+    stripped = strip_comments(text, suffix)
+    structural = _strip_all_whitespace(stripped)
+    structural_hash = hashlib.sha256(structural.encode('utf-8')).hexdigest()
+    return (content_hash, structural_hash)
+
+
+def classify(old_sig, new_sig):
+    """Compare old vs new (content_hash, structural_hash) → change level.
+
+    Args:
+        old_sig: dict {content_hash, structural_hash}, or tuple of the two,
+                 or None (never seen → STRUCTURAL).
+        new_sig: (content_hash, structural_hash) tuple from compute_hashes.
+
+    Returns:
+        'NONE' | 'COSMETIC' | 'STRUCTURAL'.
+        Never-seen files (old_sig is None) → 'STRUCTURAL' (safe default: extract).
+    """
+    if old_sig is None:
+        return 'STRUCTURAL'
+    # Accept both dict (from get_file_signature) and tuple (from compute_hashes)
+    # so callers don't have to convert.
+    if isinstance(old_sig, dict):
+        old_content = old_sig.get('content_hash')
+        old_struct = old_sig.get('structural_hash')
+    else:
+        old_content, old_struct = old_sig[0], old_sig[1]
+    new_content, new_struct = new_sig
+    if old_content == new_content:
+        return 'NONE'
+    if old_struct == new_struct:
+        return 'COSMETIC'
+    return 'STRUCTURAL'
+
+
+# ── Suffixes the incremental builder cares about ─────────────────────────────
+# Keep in sync with kg_incremental.py's filter: ('.vue', '.ts', '.java').
+TRACKED_SUFFIXES = ('.vue', '.ts', '.java')
+
+
+if __name__ == '__main__':
+    # Quick self-test: same code + extra comment → COSMETIC; changed code → STRUCTURAL.
+    import tempfile
+    tmp = tempfile.NamedTemporaryFile(suffix='.ts', delete=False, mode='w', encoding='utf-8')
+    tmp.write('export function add(a, b) {\n  return a + b\n}\n')
+    tmp.close()
+    h1 = compute_hashes(tmp.name)
+    print('original:', h1)
+
+    with open(tmp.name, 'w', encoding='utf-8') as f:
+        f.write('// leading comment\nexport function add(a, b) {\n  return a + b\n}\n')
+    h2 = compute_hashes(tmp.name)
+    print('comment-only change:', h2, '→', classify(h1, h2))
+
+    with open(tmp.name, 'w', encoding='utf-8') as f:
+        f.write('export function add(a, b, c) {\n  return a + b + c\n}\n')
+    h3 = compute_hashes(tmp.name)
+    print('structural change:', h3, '→', classify(h1, h3))
+    os.unlink(tmp.name)

package/templates/qoder/scripts/domain/kg/extract/asset/__init__.py

@@ -0,0 +1,10 @@
+# -*- coding: utf-8 -*-
+"""资产中心 25 表吸收 (GOAL 4)。
+
+子模块:
+  prd_importer         prds + prd_versions → 业务知识 (调 LLM)
+  prototype_importer   prototypes + annotations → 意图知识
+  discussion_importer  discussions → 决策知识 (LLM)
+  asset_tree           assets → 模块结构
+  returns_importer     asset_returns → 实践知识
+"""

package/templates/qoder/scripts/domain/kg/extract/asset/asset_tree.py

@@ -0,0 +1,57 @@
+# -*- coding: utf-8 -*-
+"""资产树 → 模块结构知识。
+
+assets 表 (飞书式文件树, 330节点) → MODULE 实体。
+"01_PRD/02_设计/03_交付" 这种结构本身就是项目知识。
+parent_id 递归 → 模块层级。
+"""
+from __future__ import annotations
+import os, sys
+
+_THIS = os.path.dirname(os.path.abspath(__file__))
+for _i in range(8):
+    _p = os.path.dirname(_THIS)
+    if os.path.isfile(os.path.join(_p, 'foundation', 'bootstrap.py')):
+        sys.path.insert(0, _p); break
+    _THIS = _p
+
+
+def import_asset_tree(project_id: str = None, logger=print):
+    """assets 表 → MODULE 实体 + 层级边。"""
+    try:
+        from dotenv import load_dotenv
+        for p in ['wlinkj-workspace/backend/.env', '.env']:
+            if os.path.isfile(p): load_dotenv(p); break
+    except: pass
+    sys.path.insert(0, os.path.join(os.getcwd(), 'wlinkj-workspace', 'backend'))
+    from app.db import SessionLocal
+    from app.models import Asset
+    from domain.kg.extract.java import pg_upsert
+
+    db = SessionLocal()
+    q = db.query(Asset)
+    if project_id: q = q.filter(Asset.project_id == project_id)
+    assets = q.all()
+    logger('  [asset-tree] %d 个资产节点' % len(assets))
+
+    entities, edges = [], []
+    for a in assets:
+        repo_id = 'asset-%s' % (a.project_id or 'global')[:8]
+        entities.append({
+            'id': 'asset:%s' % a.id, 'repo_id': repo_id,
+            'type': 'MODULE' if a.is_folder else 'ASSET',
+            'canonical': a.name, 'cn': a.type or '',
+            'props': {'project_id': a.project_id, 'kind': 'folder' if a.is_folder else 'file',
+                      'mime': a.mime_type, 'source': 'asset_center'},
+        })
+        # parent_id → 层级边
+        if a.parent_id:
+            edges.append({
+                'from_id': 'asset:%s' % a.id, 'to_id': 'asset:%s' % a.parent_id,
+                'edge_type': 'child_of', 'from_repo': repo_id, 'to_repo': repo_id,
+                'confidence': 1.0, 'source': 'explicit',
+            })
+    db.close()
+
+    upsert_r = pg_upsert.upsert_batch(entities, edges) if (entities or edges) else {'ok': False}
+    return {'assets': len(assets), 'entities': len(entities), 'edges': len(edges), 'upsert': upsert_r}

package/templates/qoder/scripts/domain/kg/extract/asset/discussion_importer.py

@@ -0,0 +1,62 @@
+# -*- coding: utf-8 -*-
+"""讨论 → 决策知识。
+
+discussion_messages → DECISION 实体 (LLM 抽取决策点)。
+当前 0 数据, 管线先建好, 等团队用起来立刻生效。
+"""
+from __future__ import annotations
+import os, sys
+
+_THIS = os.path.dirname(os.path.abspath(__file__))
+for _i in range(8):
+    _p = os.path.dirname(_THIS)
+    if os.path.isfile(os.path.join(_p, 'foundation', 'bootstrap.py')):
+        sys.path.insert(0, _p); break
+    _THIS = _p
+
+
+def import_discussions(project_id: str = None, logger=print):
+    """讨论 → KG。"""
+    try:
+        from dotenv import load_dotenv
+        for p in ['wlinkj-workspace/backend/.env', '.env']:
+            if os.path.isfile(p): load_dotenv(p); break
+    except: pass
+    sys.path.insert(0, os.path.join(os.getcwd(), 'wlinkj-workspace', 'backend'))
+    from app.db import SessionLocal
+    from app.models import Discussion, DiscussionMessage
+    from domain.kg.extract.java import pg_upsert
+
+    db = SessionLocal()
+    q = db.query(Discussion)
+    if project_id: q = q.filter(Discussion.project_id == project_id)
+    discussions = q.all()
+    logger('  [asset-disc] %d 个讨论' % len(discussions))
+
+    entities, edges = [], []
+    for d in discussions:
+        repo_id = 'asset-%s' % (d.project_id or 'global')[:8]
+        # 讨论 → DECISION 实体
+        entities.append({
+            'id': 'disc:%s' % d.id, 'repo_id': repo_id, 'type': 'DECISION',
+            'canonical': d.title[:80], 'cn': '',
+            'props': {'kind': 'discussion', 'project_id': d.project_id, 'source': 'asset_center'},
+        })
+        # 消息 → 关联到讨论
+        msgs = db.query(DiscussionMessage).filter(DiscussionMessage.discussion_id == d.id).all()
+        for m in msgs:
+            if not m.content: continue
+            entities.append({
+                'id': 'discmsg:%s' % m.id, 'repo_id': repo_id, 'type': 'DECISION',
+                'canonical': m.content[:60], 'cn': m.author_name or '',
+                'props': {'discussion_id': d.id, 'author': m.author_name, 'source': 'discussion_msg'},
+            })
+            edges.append({
+                'from_id': 'discmsg:%s' % m.id, 'to_id': 'disc:%s' % d.id,
+                'edge_type': 'decided_in', 'from_repo': repo_id, 'to_repo': repo_id,
+                'confidence': 1.0, 'source': 'explicit',
+            })
+    db.close()
+
+    upsert_r = pg_upsert.upsert_batch(entities, edges) if (entities or edges) else {'ok': False}
+    return {'discussions': len(discussions), 'entities': len(entities), 'edges': len(edges), 'upsert': upsert_r}

package/templates/qoder/scripts/domain/kg/extract/asset/prd_importer.py

@@ -0,0 +1,146 @@
+# -*- coding: utf-8 -*-
+"""资产中心 prds + prd_versions → KG 业务知识吸收。
+
+把工作台 SQLite/PG 的 prds 表灌进 KG (当前0吸收)。
+对每个有正文的 PRD:
+  1. 调 prd_llm_extract 抽功能点/字段/规则/端点/权限
+  2. 调 prd_chunk_embed 做向量 (LightRAG vector 路)
+  3. 调 req_anchor 锚到 REQ-ID
+"""
+from __future__ import annotations
+import os
+import sys
+
+_THIS = os.path.dirname(os.path.abspath(__file__))
+for _i in range(8):
+    _p = os.path.dirname(_THIS)
+    if os.path.isfile(os.path.join(_p, 'foundation', 'bootstrap.py')):
+        sys.path.insert(0, _p); break
+    _THIS = _p
+
+
+def import_all_prds(project_id: str = None, do_embed: bool = False, llm_limit: int = 10, logger=print):
+    """扫资产中心所有 PRD → LLM 抽取 + 锚定。
+
+    Args:
+        project_id: 限定某项目 (None=全部)
+        do_embed: 是否做 embedding (慢, 默认False)
+        llm_limit: LLM 抽取上限 (避免154个PRD全调qwen超时; 0=不调LLM只锚定)
+    Returns: {stats}
+    """
+    from domain.kg.extract.prd import prd_llm_extract, prd_chunk_embed, req_anchor
+
+    # 连工作台 DB 读 PRD
+    try:
+        from dotenv import load_dotenv
+        for p in ['wlinkj-workspace/backend/.env', '.env']:
+            if os.path.isfile(p):
+                load_dotenv(p); break
+    except Exception:
+        pass
+    sys.path.insert(0, os.path.join(os.getcwd(), 'wlinkj-workspace', 'backend'))
+    from app.db import SessionLocal
+    from app.models import PRD, PRDVersion
+    from sqlalchemy import text
+
+    db = SessionLocal()
+    q = db.query(PRD)
+    if project_id:
+        q = q.filter(PRD.project_id == project_id)
+    prds = q.all()
+    logger('  [asset-prd] %d 个 PRD 待吸收' % len(prds))
+
+    all_entities, all_edges = [], []
+    embedded = 0
+    llm_done = 0
+    for prd in prds:
+        # 取最新版本正文
+        ver = db.query(PRDVersion).filter(PRDVersion.prd_id == prd.id).order_by(PRDVersion.version.desc()).first()
+        content_md = ver.content_md if ver else ''
+        repo_id = 'asset-%s' % (prd.project_id or 'global')[:8]
+
+        # ① PRD 实体本身 (写 entities, group=prd)
+        all_entities.append({
+            'id': 'prd:%s' % prd.id, 'repo_id': repo_id, 'type': 'PRD',
+            'canonical': prd.title[:80], 'cn': prd.category or '',
+            'props': {'project_id': prd.project_id, 'status': prd.status,
+                      'version': prd.current_version, 'source': 'asset_center'},
+        })
+
+        # ② REQ-ID 锚定 (PRD 标题 → anchors)
+        a_ent, a_edg = req_anchor.anchor_prd(prd.id, prd.title, prd.project_id, repo_id)
+        all_entities += a_ent
+        all_edges += a_edg
+
+        # ③ 有正文且未超 LLM 上限 → LLM 抽取
+        if content_md and len(content_md) > 100 and llm_done < llm_limit:
+            extracted = prd_llm_extract.extract_prd_knowledge(content_md)
+            e_ent, e_edg = prd_llm_extract.to_entities_edges(
+                extracted, prd.id, prd.project_id, repo_id)
+            all_entities += e_ent
+            all_edges += e_edg
+            llm_done += 1
+
+            # ④ embedding (可选, 慢)
+            if do_embed and len(content_md) > 500:
+                try:
+                    r = prd_chunk_embed.embed_prd(prd.id, content_md, repo_id)
+                    embedded += r.get('embedded', 0)
+                except Exception:
+                    pass
+
+    db.close()
+
+    # 写 PG
+    from domain.kg.extract.java import pg_upsert
+    upsert_r = pg_upsert.upsert_batch(all_entities, all_edges) if (all_entities or all_edges) else {'ok': False}
+
+    # ★ 同时把锚点写进 anchors 专用表 (GOAL 1 建的)
+    anchor_records = []
+    seen_anchor = set()
+    for e in all_entities:
+        if e.get('type') == 'ANCHOR':
+            aid = e['id']
+            if aid in seen_anchor:
+                continue
+            seen_anchor.add(aid)
+            anchor_records.append({
+                'id': aid,
+                'project_id': e['props'].get('project_id'),
+                'title': e.get('cn', ''),
+                'kind': e['props'].get('kind', 'feature'),
+                'prd_id': e['props'].get('prd_id'),
+            })
+    anchor_ok = _upsert_anchors(anchor_records) if anchor_records else 0
+
+    return {
+        'prds_total': len(prds),
+        'entities': len(all_entities),
+        'edges': len(all_edges),
+        'embedded': embedded,
+        'anchors': anchor_ok,
+        'upsert': upsert_r,
+    }
+
+
+def _upsert_anchors(records):
+    """写 anchors 专用表。"""
+    from domain.kg.extract.java import pg_upsert as _pg
+    engine, _ = _pg._get_pg_engine()
+    if engine is None or not records:
+        return 0
+    from sqlalchemy import text
+    ok = 0
+    with engine.begin() as conn:
+        for r in records:
+            try:
+                conn.execute(text("""
+                    INSERT INTO anchors (id, project_id, title, kind, prd_id)
+                    VALUES (:id, :pid, :title, :kind, :prd)
+                    ON CONFLICT (id) DO UPDATE SET title=EXCLUDED.title, updated_at=now()
+                """), {'id': r['id'], 'pid': r.get('project_id'), 'title': r.get('title', ''),
+                       'kind': r.get('kind', 'feature'), 'prd': r.get('prd_id')})
+                ok += 1
+            except Exception:
+                pass
+    return ok

package/templates/qoder/scripts/domain/kg/extract/asset/prototype_importer.py

@@ -0,0 +1,64 @@
+# -*- coding: utf-8 -*-
+"""原型 + 标注 → 意图知识。
+
+把 prototypes 表灌进 KG (当前10个原型, 0标注 - 管线先建好)。
+prototype → PAGE 实体; annotation → CONSTRAINT 实体。
+"""
+from __future__ import annotations
+import os, sys
+
+_THIS = os.path.dirname(os.path.abspath(__file__))
+for _i in range(8):
+    _p = os.path.dirname(_THIS)
+    if os.path.isfile(os.path.join(_p, 'foundation', 'bootstrap.py')):
+        sys.path.insert(0, _p); break
+    _THIS = _p
+
+
+def import_prototypes(project_id: str = None, logger=print):
+    """原型 + 标注 → KG。"""
+    try:
+        from dotenv import load_dotenv
+        for p in ['wlinkj-workspace/backend/.env', '.env']:
+            if os.path.isfile(p): load_dotenv(p); break
+    except: pass
+    sys.path.insert(0, os.path.join(os.getcwd(), 'wlinkj-workspace', 'backend'))
+    from app.db import SessionLocal
+    from app.models import Prototype, PrototypeAnnotation
+    from domain.kg.extract.java import pg_upsert
+
+    db = SessionLocal()
+    q = db.query(Prototype)
+    if project_id: q = q.filter(Prototype.project_id == project_id)
+    protos = q.all()
+    logger('  [asset-proto] %d 个原型' % len(protos))
+
+    entities, edges = [], []
+    for p in protos:
+        repo_id = 'asset-%s' % (p.project_id or 'global')[:8]
+        # 原型 → PAGE 实体
+        entities.append({
+            'id': 'proto:%s' % p.id, 'repo_id': repo_id, 'type': 'PAGE',
+            'canonical': p.feature or p.id[:8], 'cn': p.platform or '',
+            'props': {'kind': 'prototype', 'project_id': p.project_id,
+                      'status': p.status, 'source': 'asset_center'},
+        })
+        # 该原型的标注 → CONSTRAINT
+        anns = db.query(PrototypeAnnotation).filter(PrototypeAnnotation.prototype_id == p.id).all()
+        for a in anns:
+            if not a.comment: continue
+            entities.append({
+                'id': 'ANNO:%s' % a.id, 'repo_id': repo_id, 'type': 'CONSTRAINT',
+                'canonical': a.comment[:60], 'cn': a.type or '',
+                'props': {'prototype_id': p.id, 'status': a.status,
+                          'coords': {'x': a.x, 'y': a.y}, 'source': 'annotation'},
+            })
+            edges.append({
+                'from_id': 'ANNO:%s' % a.id, 'to_id': 'proto:%s' % p.id,
+                'edge_type': 'annotates', 'from_repo': repo_id, 'to_repo': repo_id,
+                'confidence': 1.0, 'source': 'explicit',
+            })
+    db.close()
+
+    upsert_r = pg_upsert.upsert_batch(entities, edges) if (entities or edges) else {'ok': False}
+    return {'prototypes': len(protos), 'entities': len(entities), 'edges': len(edges), 'upsert': upsert_r}

package/templates/qoder/scripts/domain/kg/extract/asset/returns_importer.py

@@ -0,0 +1,52 @@
+# -*- coding: utf-8 -*-
+"""回流历史 → 实践知识。
+
+asset_returns (引擎→平台产出) → embedding (能搜"之前怎么做的")。
+当前3条, 是引擎回流的PRD/原型/资产。
+"""
+from __future__ import annotations
+import os, sys
+
+_THIS = os.path.dirname(os.path.abspath(__file__))
+for _i in range(8):
+    _p = os.path.dirname(_THIS)
+    if os.path.isfile(os.path.join(_p, 'foundation', 'bootstrap.py')):
+        sys.path.insert(0, _p); break
+    _THIS = _p
+
+
+def import_returns(project_id: str = None, logger=print):
+    """asset_returns → KG 实践知识。"""
+    try:
+        from dotenv import load_dotenv
+        for p in ['wlinkj-workspace/backend/.env', '.env']:
+            if os.path.isfile(p): load_dotenv(p); break
+    except: pass
+    sys.path.insert(0, os.path.join(os.getcwd(), 'wlinkj-workspace', 'backend'))
+    from app.db import SessionLocal
+    from app.models import AssetReturn
+    from domain.kg.extract.java import pg_upsert
+
+    db = SessionLocal()
+    q = db.query(AssetReturn)
+    if project_id: q = q.filter(AssetReturn.project_id == project_id)
+    returns = q.all()
+    logger('  [asset-ret] %d 条回流' % len(returns))
+
+    entities, edges = [], []
+    for r in returns:
+        repo_id = 'asset-%s' % (r.project_id or 'global')[:8]
+        # 回流产出 → 实体 (按 return_type 分类)
+        etype = {'prd': 'PRD', 'prototype': 'PAGE', 'asset': 'ASSET'}.get(r.return_type, 'ASSET')
+        entities.append({
+            'id': 'return:%s' % r.id, 'repo_id': repo_id, 'type': etype,
+            'canonical': r.title[:80], 'cn': r.return_type or '',
+            'props': {'project_id': r.project_id, 'return_type': r.return_type,
+                      'source': r.source, 'status': r.status,
+                      'preview': (r.content_preview or '')[:200],
+                      'source': 'asset_return'},
+        })
+
+    db.close()
+    upsert_r = pg_upsert.upsert_batch(entities, edges) if (entities or edges) else {'ok': False}
+    return {'returns': len(returns), 'entities': len(entities), 'edges': len(edges), 'upsert': upsert_r}

package/templates/qoder/scripts/domain/kg/extract/build_goal3.py

@@ -0,0 +1,104 @@
+# -*- coding: utf-8 -*-
+"""GOAL 3 统一构建入口: MyBatis XML + DB Schema + 写 PG。
+
+被 kg_build.py 调用, 或独立 CLI:
+    python -m domain.kg.extract.build_goal3 --repo-dir data/code/fywl-ics --repo-id fywl-ics
+"""
+from __future__ import annotations
+import os
+import sys
+
+_THIS = os.path.dirname(os.path.abspath(__file__))
+for _i in range(8):
+    _p = os.path.dirname(_THIS)
+    if os.path.isfile(os.path.join(_p, 'foundation', 'bootstrap.py')):
+        sys.path.insert(0, _p); break
+    _THIS = _p
+
+from domain.kg.extract.mybatis import all as mybatis_all
+from domain.kg.extract.db import schema_extractor, fk_extractor
+from domain.kg.extract.java import pg_upsert
+
+
+def build(repo_dir: str, repo_id: str, write_pg: bool = True, logger=print):
+    """GOAL 3 主入口: mybatis 提取 + db schema + 写 PG。"""
+    logger('  [goal3] === MyBatis XML 提取 ===')
+    mb_entities, mb_edges = mybatis_all.extract_repo(repo_dir, repo_id, logger=logger)
+
+    logger('  [goal3] === DB Schema 真源 ===')
+    db_tables, db_columns = schema_extractor.extract_all_schema(repo_id, logger=logger)
+    fk_edges = fk_extractor.extract_foreign_keys(repo_id, logger=logger)
+
+    # db_tables/db_columns 写专用表 (不是 entities)
+    db_pg_result = None
+    if write_pg and (db_tables or db_columns):
+        db_pg_result = _upsert_db_schema(db_tables, db_columns)
+
+    # mybatis 实体/边 写 entities/edges (复用 GOAL 2 的 pg_upsert)
+    mb_pg_result = None
+    if write_pg and (mb_entities or mb_edges):
+        # 加上 FK 边
+        all_edges = mb_edges + fk_edges
+        mb_pg_result = pg_upsert.upsert_batch(mb_entities, all_edges)
+
+    return {
+        'mybatis_entities': len(mb_entities), 'mybatis_edges': len(mb_edges),
+        'db_tables': len(db_tables), 'db_columns': len(db_columns),
+        'fk_edges': len(fk_edges),
+        'mb_pg': mb_pg_result, 'db_pg': db_pg_result,
+    }
+
+
+def _upsert_db_schema(tables: list, columns: list):
+    """把 db_tables / db_columns 写进 PG 专用表 (GOAL 1 建的)。"""
+    engine, _ = pg_upsert._get_pg_engine()
+    if engine is None:
+        return {'ok': False, 'error': 'PG 不可用'}
+    from sqlalchemy import text
+    t_ok = c_ok = 0
+    with engine.begin() as conn:
+        for t in tables:
+            try:
+                conn.execute(text("""
+                    INSERT INTO db_tables (repo_id, schema_name, table_name, table_comment, rows_count)
+                    VALUES (:r, :s, :t, :c, :n)
+                    ON CONFLICT (repo_id, schema_name, table_name) DO UPDATE SET
+                        table_comment=EXCLUDED.table_comment, rows_count=EXCLUDED.rows_count
+                """), {'r': t['repo_id'], 's': t['schema_name'], 't': t['table_name'],
+                       'c': t['table_comment'], 'n': t.get('rows_count')})
+                t_ok += 1
+            except Exception:
+                pass
+        for c in columns:
+            try:
+                conn.execute(text("""
+                    INSERT INTO db_columns (repo_id, schema_name, table_name, column_name,
+                                           data_type, is_nullable, column_comment, ordinal_position)
+                    VALUES (:r, :s, :t, :cn, :dt, :nn, :cc, :op)
+                    ON CONFLICT (repo_id, schema_name, table_name, column_name) DO UPDATE SET
+                        data_type=EXCLUDED.data_type, is_nullable=EXCLUDED.is_nullable,
+                        column_comment=EXCLUDED.column_comment
+                """), {'r': c['repo_id'], 's': c['schema_name'], 't': c['table_name'],
+                       'cn': c['column_name'], 'dt': c['data_type'], 'nn': c['is_nullable'],
+                       'cc': c['column_comment'], 'op': c.get('ordinal_position')})
+                c_ok += 1
+            except Exception:
+                pass
+    return {'ok': True, 'tables_ok': t_ok, 'columns_ok': c_ok}
+
+
+if __name__ == '__main__':
+    import argparse, json
+    ap = argparse.ArgumentParser()
+    ap.add_argument('--repo-dir', required=True)
+    ap.add_argument('--repo-id', default='')
+    ap.add_argument('--dry-run', action='store_true')
+    args = ap.parse_args()
+    try:
+        from dotenv import load_dotenv
+        for p in ['wlinkj-workspace/backend/.env', '.env']:
+            if os.path.isfile(p): load_dotenv(p); break
+    except: pass
+    rid = args.repo_id or os.path.basename(args.repo_dir.rstrip('/\\'))
+    st = build(args.repo_dir, rid, write_pg=not args.dry_run)
+    print(json.dumps(st, ensure_ascii=False, indent=2, default=str))