fix(knowledge): 模型判定 should_ingest=false 时直接跳过，不再落草稿

日志显示寒暄/无效会话并非 LLM 调用失败，而是模型正确返回 should_ingest=false。
上一版把这种情况也当失败回退、堆成草稿，反而把无价值对话存进了库。

- llm_extractor: 模型显式 should_ingest=false 且无 notes → 抛 LLMRejectedIngest
- capture_thread: 捕获后抛 EmptyThreadError → 整条跳过、不入库（寒暄/总结此页这类不再进库）
- 仅当真正失败（API 报错 / 非 JSON）才回退规则抽取并落草稿
- 兼容包装 extract_thread_llm 吞掉该异常返回 None
- 更新测试（should_ingest=false 跳过）+ CLAUDE.md

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

offline-backend-20260512/backend/CLAUDE.md

@@ -286,7 +286,8 @@ Self-contained product knowledge base. **All knowledge capability lives in this
 
 **Wikilinks & multi-note capture** (`knowledge.{related_notes_enabled,related_notes_limit,multi_note_enabled}`, all default-on; `project_name` defaults to `zncm`):
 - **Navigable `[[wikilinks]]`** — `service.resolve_wikilink(s)` maps a `[[target]]` to a knowledge note (by id → `vault_path` → title) or a vault meta page (`read_vault_page`, path-traversal-proof), else `missing`. `markdown_vault.normalize_wikilink_target` canonicalizes targets (drops `|alias`, slashes, `.md`). The frontend `MarkdownRenderer` renders non-numeric `[[...]]` as clickable links from the resolver map (numeric `[[123]]` stay citation badges); note links navigate, vault links open a `VaultPageDialog`, missing links are greyed.
-- **Real related notes only (no structural meta-links)** — `extractor.related_links()` now returns `[]`: notes no longer auto-link `index` / `_meta/taxonomy` / a project hub (they were noise in every `## Related`). On capture, `service._find_related_notes` runs keyword/hybrid search over the draft's title + summary + *distinctive* tags (project/`knowledge-base` defaults excluded) and `_inject_related_notes` writes real `[[Title]]` cross-links into `## Related` (plus note→note `related` graph edges); the renderers omit the `## Related` section entirely when there are no related notes and no sources. The vault's own `index.md`/`_meta/taxonomy.md` still exist (obsidian scaffold) and remain resolvable for manually-written links. The router enriches `GET /notes/{id}` with `created_by_name`/`updated_by_name` (auth-provider email lookup) so the UI shows a person, not a UUID.
+- **Real related notes only (no structural meta-links)** — `extractor.related_links()` now returns `[]`: notes no longer auto-link `index` / `_meta/taxonomy` / a project hub (they were noise in every `## Related`). On capture, `service._find_related_notes` cross-links **only by shared *distinctive* tags** (project/`knowledge-base` defaults excluded; free-text title/summary matching was removed because CJK tokenization linked unrelated notes via generic words like 助手/知识图谱) and `_inject_related_notes` writes real `[[Title]]` links into `## Related` (plus note→note `related` graph edges); no distinctive tags → no related notes. The renderers omit the `## Related` section entirely when there are no related notes and no sources. The vault's own `index.md`/`_meta/taxonomy.md` still exist (obsidian scaffold) and remain resolvable for manually-written links. The router enriches `GET /notes/{id}` with `created_by_name`/`updated_by_name` (auth-provider email lookup) so the UI shows a person, not a UUID. The frontend `MarkdownRenderer` renders wiki-capture provenance markers `^[inferred]`/`^[ambiguous]` as small "推测"/"存疑" badges.
+- **LLM-extraction outcomes** — capture distinguishes three LLM outcomes: (1) **success** → declarative note(s); (2) **deliberate rejection** — the model returns `should_ingest:false` with no notes (greeting / meta chatter) → `extract_thread_llm_multi` raises `LLMRejectedIngest` and `capture_thread` raises `EmptyThreadError` (skip — nothing persisted, so worthless chats never enter the KB); (3) **genuine failure** — API error / non-JSON → returns `[]`, capture falls back to the rule-based verbatim extractor and lands the note as a **`draft`** (not `approved`), logs a WARNING, and returns `llm_fallback:true` (surfaced by the router). `llm_extractor` logs an actionable WARNING/INFO for each non-success path so extraction issues are diagnosable.
 - **Multi-note split** — `llm_extractor.extract_thread_llm_multi` lets one conversation distill into 1–4 cross-linked notes (`notes:[…]` JSON array; legacy single-object output still accepted). `capture_thread` persists each, sibling-cross-links them by title, and returns `{note, created, notes:[…]}`. Gated by `multi_note_enabled` (false → first note only). Tests: `tests/test_knowledge_wikilinks.py`.
 
 **Phases 2–4 (all config-gated, defaults safe):**

offline-backend-20260512/backend/packages/harness/deerflow/knowledge/llm_extractor.py

@@ -27,6 +27,23 @@ logger = logging.getLogger(__name__)
 
 _VALID_CATEGORIES = {"synthesis", "concepts", "references", "decision", "journal"}
 
+
+class LLMRejectedIngest(Exception):
+    """The LLM judged the conversation as having no knowledge value.
+
+    Raised (instead of returning ``[]``) when the model explicitly sets
+    ``should_ingest: false`` with no notes — so callers can *skip* ingestion
+    entirely rather than falling back to the rule-based verbatim dump.
+    """
+
+
+def _is_falsey(value: Any) -> bool:
+    if value is False:
+        return True
+    if isinstance(value, str):
+        return value.strip().lower() in ("false", "0", "no")
+    return False
+
 _SYSTEM_PROMPT = """\
 你是知识库整理专家，遵循 obsidian-wiki 的 wiki-capture 规范，把一次对话提炼成\
 **陈述式知识（declarative knowledge）**，而不是聊天流水账。
@@ -257,11 +274,18 @@ async def extract_thread_llm_multi(
 
     note_dicts = _notes_from_data(data)
     if not note_dicts:
+        reason = str(data.get("reason") or "")[:200]
+        if _is_falsey(data.get("should_ingest")):
+            # The model deliberately judged this conversation worthless (greeting,
+            # meta chatter, no findings). Skip ingestion entirely — do NOT fall
+            # back to the rule-based verbatim dump.
+            logger.info("Knowledge LLM judged thread %s not worth ingesting: %s", thread_id, reason)
+            raise LLMRejectedIngest(reason or "no knowledge value")
         logger.warning(
             "Knowledge LLM extraction for thread %s produced no notes (should_ingest=%s, reason=%r); using rule-based fallback",
             thread_id,
             data.get("should_ingest"),
-            str(data.get("reason") or "")[:200],
+            reason,
         )
         return []
     if not allow_multi:
@@ -310,18 +334,21 @@ async def extract_thread_llm(
     Thin wrapper over :func:`extract_thread_llm_multi` that returns the first
     note, or ``None`` when nothing was produced.
     """
-    drafts = await extract_thread_llm_multi(
+    try:
-        messages,
+        drafts = await extract_thread_llm_multi(
-        thread_id=thread_id,
+            messages,
-        thread_title=thread_title,
+            thread_id=thread_id,
-        mode=mode,
+            thread_title=thread_title,
-        title=title,
+            mode=mode,
-        tags=tags,
+            title=title,
-        project_name=project_name,
+            tags=tags,
-        sources=sources,
+            project_name=project_name,
-        max_messages=max_messages,
+            sources=sources,
-        model_name=model_name,
+            max_messages=max_messages,
-        max_input_chars=max_input_chars,
+            model_name=model_name,
-        allow_multi=False,
+            max_input_chars=max_input_chars,
-    )
+            allow_multi=False,
+        )
+    except LLMRejectedIngest:
+        return None
     return drafts[0] if drafts else None

offline-backend-20260512/backend/packages/harness/deerflow/knowledge/service.py

@@ -247,25 +247,30 @@ class KnowledgeService:
         llm_attempted = self.config.llm_extract_enabled
         llm_fallback = False
         if llm_attempted:
-            drafts = await llm_extractor.extract_thread_llm_multi(
+            try:
-                messages,
+                drafts = await llm_extractor.extract_thread_llm_multi(
-                thread_id=thread_id,
+                    messages,
-                thread_title=thread_title,
+                    thread_id=thread_id,
-                mode=mode,
+                    thread_title=thread_title,
-                title=title,
+                    mode=mode,
-                tags=tags,
+                    title=title,
-                project_name=self.project_name,
+                    tags=tags,
-                sources=sources,
+                    project_name=self.project_name,
-                max_messages=self.config.max_thread_messages,
+                    sources=sources,
-                model_name=self.config.extract_model_name,
+                    max_messages=self.config.max_thread_messages,
-                max_input_chars=self.config.extract_max_input_chars,
+                    model_name=self.config.extract_model_name,
-                allow_multi=self.config.multi_note_enabled,
+                    max_input_chars=self.config.extract_max_input_chars,
-            )
+                    allow_multi=self.config.multi_note_enabled,
+                )
+            except llm_extractor.LLMRejectedIngest as exc:
+                # The model judged this conversation worthless (greeting / meta
+                # chatter). Skip ingestion entirely — do NOT dump the raw chat.
+                raise EmptyThreadError(f"Thread {thread_id} judged not worth ingesting: {exc}") from exc
             if not drafts:
                 # The LLM path was supposed to run but produced nothing usable
-                # (API error / non-JSON / empty). We will fall back to the
+                # (API error / non-JSON). We will fall back to the rule-based
-                # rule-based extractor, which yields a low-quality verbatim dump
+                # extractor, which yields a low-quality verbatim dump with no
-                # with no entities/tags — flag it so it lands as a draft.
+                # entities/tags — flag it so it lands as a draft.
                 llm_fallback = True
         if not drafts:
             single = extractor.extract_thread(