kebab/docs/superpowers/specs/2026-04-27-kb-final-form-design.md

title, date, status, purpose, source_report, related_tasks

title	date	status	purpose	source_report	related_tasks
KB v1 최종 결과물 형태 — Frozen Design	2026-04-27	frozen	작은 단위 분해 작업 시 spec 변경을 막기 위한 단단한 contract 동결	../../../kb_local_rust_report.md	../../../tasks/INDEX.md

KB v1 최종 결과물 형태 — Frozen Design

이 문서는 사용자가 만족할 최종 결과물의 매우 구체적 형태를 동결한다. 각 phase 의 task 분해는 이 contract 위에서 수행되며, 이 문서가 바뀌지 않는 한 task 들의 인터페이스는 변하지 않는다.

전제 보고서는 kb_local_rust_report.md. 그 보고서가 방향과 근거를 제공하며, 이 문서가 형태를 못박는다.

---

0. 동결된 결정 요약

#	결정	값	근거
Q1	scope 우선순위	UX → Data 역도출	사용자 만족이 spec 안정성 lever
Q2	headline UX	kb ask 답변 화면	검색/citation/RAG/refusal/모델메타 모두 노출
–	ask 기본 형식	inline numeric refs [1]…[n] + footer	일상 가독성
–	ask --explain	per-claim 분해 + verbose footer + retrieval trace	디버그 단일 플래그
Q3	citation 문자열	URI fragment (path#k=v…, W3C Media Fragments)	표준 정합 + Windows path 안전 + 브라우저 자동 스크롤
Q4	refusal 정책	양층: score gate + LLM self-judge + citation 후처리 검증	환각/false-negative 양쪽 차단
Q5	streaming	always (tty 토큰, pipe buffered)	체감 속도 + LLM trait 단일화
Q6	JSON 모드	별도 stable wire schema (*.v1), schema_version 명시	internal 자유 진화 + 외부 contract 동결
Q7	footer	toggle (default minimal / --explain verbose)	일상-디버그 분리
–	search 출력	dense 4줄 (rank+score / path#frag / heading / snippet)	line-oriented 파싱 + fzf 친화
Q8	ID 인코딩	hybrid: blake3(canonical_json(tuple))[..32] PK + path/heading human ref	짧은 PK + path-based citation
Q9	frontmatter	모두 optional + auto-derive + 미지 키 metadata.user 보존	진입 장벽 0
Q10	workspace	single root + XDG layout	personal v1 적정
–	asset 보존	content-addressable copy, copy_threshold_mb=100 초과 시 reference + checksum	reproducibility + 디스크 절감
–	wire 버전	additive within vN, breaking → vN+1	외부 깨짐 방지
–	ignore	gitignore 문법 + .kbignore	익숙함
–	에러	thiserror per crate, anyhow at boundary	추적성 + UX
–	sync	watch=false default	v1 명시 ingest

---

1. Headline UX scenes

1.1 kb ask (default)

$ kb ask "Markdown chunking 규칙은?"

heading boundary 우선 [1]. code block 중간 분할 금지 [2]. table 가능한 한
단일 chunk 유지 [2]. 긴 section 은 paragraph 단위로 분할 [1]. chunk 마다
heading_path 와 source_span 보존 [1].

─────────────────────────────────────────────────────────
[1] notes/rust/kb-architecture.md#L661-L672
    §14 Chunking 정책
[2] notes/rust/kb-architecture.md#L665-L668
    §14 Chunking 정책

grounded ✓  qwen2.5:14b-instruct  rag-v1  3 chunks

1.2 kb ask --explain

$ kb ask --explain "Markdown chunking 규칙은?"

▎ heading boundary 우선
  └ notes/rust/kb-architecture.md#L662
    「heading boundary를 우선한다」

▎ code block 중간 분할 금지
  └ notes/rust/kb-architecture.md#L663
    「code block은 중간에서 자르지 않는다」

▎ table 단일 chunk 유지
  └ notes/rust/kb-architecture.md#L664
    「table은 가능한 한 하나의 chunk로」

▎ heading_path / source_span 보존
  └ notes/rust/kb-architecture.md#L668-L670

retrieval trace
  query           "Markdown chunking 규칙은?"
  mode            hybrid
  k               8
  threshold (gate) 0.30  → top-1 0.82  pass
  fusion          rrf (k=60)
  chunks (used)   3 / 8 returned
    #1 0.82  notes/rust/kb-architecture.md#L661-L672  bm25=12.4 vec=0.78
    #2 0.78  notes/rust/kb-architecture.md#L692-L713  bm25=10.1 vec=0.74
    #3 0.55  guides/markdown-style.md#L4-L18           bm25=8.2  vec=0.61

grounded ✓  qwen2.5:14b-instruct  rag-v1  3 chunks
prompt   1184 tokens  completion 312 tokens  latency 1842 ms
embedding multilingual-e5-small  index v1.0

1.3 kb ask (refusal — score gate)

$ kb ask "당신의 회사 매출은?"

근거 부족. KB 에 해당 내용 없음.
가까운 후보 (모두 임계 0.30 미만):
  · ~/notes/finance/personal-budget.md#L1-L8  (score 0.21)

grounded ✗  qwen2.5:14b-instruct  rag-v1  0 chunks used

1.4 kb ask (refusal — LLM self-judge)

$ kb ask "이 책의 23쪽 결론은?"

근거 부족. 제공된 chunk 중 결론 내용 없음.
검색은 됨, LLM 이 결론 부재 판단:
  · papers/book.pdf#p=23  (score 0.61)
  · papers/book.pdf#p=24  (score 0.58)

grounded ✗  qwen2.5:14b-instruct  rag-v1  3 chunks searched, 0 grounded

1.5 kb search (dense)

$ kb search "Markdown chunking 규칙"

1. 0.82  notes/rust/kb-architecture.md#L661-L672
   §14 Chunking 정책
   heading boundary 우선. code block 중간 분할 금지.
   table 가능한 한 단일 chunk…

2. 0.71  notes/rust/kb-architecture.md#L692-L713
   §15 검색과 RAG 정책
   검색은 처음부터 hybrid 로 설계하되 구현은 단계적…

3. 0.55  guides/markdown-style.md#L4-L18
   §1 Heading 규약
   문서는 항상 H1 으로 시작한다. H2 부터는…

3 hits  hybrid  index v1.0  bm25+e5-small/RRF

1.6 kb search --explain

각 hit 아래 추가:

   ├ lexical (bm25)   rank 1   score 12.4
   ├ vector (e5-s)    rank 2   score 0.78
   └ rrf fusion       rank 1   score 0.82
   chunker md-heading-v1   chunk_id 9b4a8c…

1.7 exit codes

code	의미
0	hit / grounded answer / success
1	no-hit / refusal (정상 거절)
2	error (parser fail, IO, network, model 미기동)
3	doctor unhealthy

---

2. Wire schema v1

docs/wire-schema/v1/*.schema.json 으로 동결. internal Rust struct ↔ wire 변환은 From/TryFrom. 모든 wire 객체는 schema_version 필드 필수.

2.1 Citation (5 variants — discriminated by kind)

{
  "schema_version": "citation.v1",
  "kind": "line|page|region|caption|time",
  "path": "notes/rust/kb.md",
  "uri":  "notes/rust/kb.md#L12-L34",

  "line":    { "start": 12, "end": 34, "section": "§14 Chunking 정책" },
  "page":    { "page": 13, "section": "Experiment Setup" },
  "region":  { "x": 120, "y": 40, "w": 520, "h": 180 },
  "caption": { "model": "qwen2.5-vl:7b" },
  "time":    { "start_ms": 822000, "end_ms": 850000, "speaker": "S1" }
}

variant 별 해당 키만 채움. path 와 uri 는 항상 채움 (uri 는 path + W3C Media Fragments 합본).

2.2 SearchHit

{
  "schema_version": "search_hit.v1",
  "rank": 1,
  "score": 0.82,
  "chunk_id": "9b4a8c1e7d3f2a05",
  "doc_id":   "3f9a2c10ee4d6b78",
  "doc_path": "notes/rust/kb-architecture.md",
  "heading_path": ["아키텍처", "Chunking 정책"],
  "section_label": "§14 Chunking 정책",
  "snippet": "heading boundary 우선. code block 중간 분할 금지…",
  "snippet_full_text": false,
  "citation": { "...": "citation.v1" },
  "retrieval": {
    "method": "hybrid",
    "lexical_score": 12.4,
    "vector_score": 0.78,
    "fusion_score": 0.82,
    "lexical_rank": 1,
    "vector_rank": 2
  },
  "index_version": "v1.0",
  "embedding_model": "multilingual-e5-small",
  "chunker_version": "md-heading-v1"
}

retrieval.method ∈ {lexical, vector, hybrid}. 단독 모드 시 다른 score/rank 는 null.

2.3 Answer

{
  "schema_version": "answer.v1",
  "answer": "heading boundary 우선 [1]. code block 중간 분할 금지 [2]…",
  "citations": [
    { "marker": "[1]", "citation": { "...": "citation.v1" } },
    { "marker": "[2]", "citation": { "...": "citation.v1" } }
  ],
  "grounded": true,
  "refusal_reason": null,
  "model": { "id": "qwen2.5:14b-instruct", "provider": "ollama" },
  "embedding": { "id": "multilingual-e5-small", "provider": "fastembed", "dimensions": 384 },
  "prompt_template_version": "rag-v1",
  "retrieval": {
    "trace_id": "ret_4a8b2c1e",
    "mode": "hybrid",
    "k": 8,
    "score_gate": 0.30,
    "top_score": 0.82,
    "chunks_returned": 8,
    "chunks_used": 3
  },
  "usage": { "prompt_tokens": 1184, "completion_tokens": 312, "latency_ms": 1842 },
  "created_at": "2026-04-27T15:42:11+09:00"
}

거절 시 grounded=false, answer 는 사람 친화 거절 문장, refusal_reason ∈ {"score_gate","llm_self_judge","no_index","no_chunks"}. citations 는 빈 배열 또는 가까운 후보 (marker null).

2.4 IngestReport

{
  "schema_version": "ingest_report.v1",
  "scope": { "root": "/home/altair/KnowledgeBase", "include": ["**/*.md"], "exclude": [".git/**"] },
  "scanned": 142, "new": 12, "updated": 3, "skipped": 127, "errors": 0,
  "duration_ms": 4231,
  "items": [
    {
      "kind": "new|updated|skipped|error",
      "doc_id": "3f9a2c10ee4d6b78",
      "doc_path": "notes/rust/kb-architecture.md",
      "asset_id": "8c1e7d3f2a05",
      "byte_len": 41822,
      "block_count": 184,
      "chunk_count": 38,
      "parser_version": "pulldown-cmark-0.x",
      "chunker_version": "md-heading-v1",
      "warnings": [],
      "error": null
    }
  ]
}

--summary-only 시 items: null.

2.5 DocSummary (kb list docs)

{
  "schema_version": "doc_summary.v1",
  "doc_id": "3f9a2c10ee4d6b78",
  "doc_path": "notes/rust/kb-architecture.md",
  "title": "Rust 로컬 Knowledge Base 설계",
  "lang": "ko",
  "tags": ["knowledge-base", "rust", "rag"],
  "trust_level": "primary",
  "source_type": "markdown",
  "byte_len": 41822,
  "chunk_count": 38,
  "created_at": "2026-04-27T00:00:00+09:00",
  "updated_at": "2026-04-27T15:42:11+09:00",
  "parser_version": "pulldown-cmark-0.x",
  "chunker_version": "md-heading-v1"
}

2.6 ChunkInspection

{
  "schema_version": "chunk_inspection.v1",
  "chunk_id": "9b4a8c1e7d3f2a05",
  "doc_id": "3f9a2c10ee4d6b78",
  "doc_path": "notes/rust/kb-architecture.md",
  "heading_path": ["아키텍처", "Chunking 정책"],
  "text": "heading boundary 우선…",
  "source_spans": [{ "kind": "line", "start": 661, "end": 672 }],
  "block_ids": ["b_0a", "b_0b"],
  "token_estimate": 480,
  "chunker_version": "md-heading-v1",
  "embeddings": [
    { "model": "multilingual-e5-small", "dimensions": 384, "embedding_id": "e_2f1a" }
  ]
}

2.7 DoctorReport

{
  "schema_version": "doctor.v1",
  "ok": true,
  "checks": [
    { "name": "config_loaded", "ok": true,  "detail": "~/.config/kb/config.toml" },
    { "name": "data_dir_writable", "ok": true, "detail": "~/.local/share/kb" },
    { "name": "sqlite_open", "ok": true, "detail": "kb.sqlite (schema v1)" },
    { "name": "lancedb_open", "ok": true, "detail": "lancedb/" },
    { "name": "embedding_model", "ok": true, "detail": "multilingual-e5-small (384d)" },
    { "name": "ollama_reachable", "ok": true, "detail": "http://127.0.0.1:11434" },
    { "name": "ollama_model_pulled", "ok": false, "detail": "qwen2.5:14b-instruct missing", "hint": "ollama pull qwen2.5:14b-instruct" }
  ]
}

ok=false 가 1개 이상이면 root ok=false, exit 3.

2.8 Versioning 규칙

• 한 schema 안: 새 optional 필드 추가만 OK. 기존 필드 제거/타입변경/enum 값 제거 금지.
• 그 이상의 변경 → *.v2.schema.json 신설. CLI --schema-version v1|v2. default 최신.
• enum 값 추가 시 클라이언트는 unknown 무시 (forward compat).

---

3. 도메인 모델 (kb-core)

3.1 Newtype IDs

#[derive(Clone, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)] pub struct AssetId(pub String);
#[derive(Clone, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)] pub struct DocumentId(pub String);
#[derive(Clone, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)] pub struct BlockId(pub String);
#[derive(Clone, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)] pub struct ChunkId(pub String);
#[derive(Clone, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)] pub struct EmbeddingId(pub String);
#[derive(Clone, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)] pub struct IndexId(pub String);

Display, FromStr 구현. 32-char hex.

3.2 Versions / labels

pub struct ParserVersion(pub String);
pub struct ChunkerVersion(pub String);
pub struct EmbeddingModelId(pub String);
pub struct EmbeddingVersion(pub String);
pub struct IndexVersion(pub String);
pub struct PromptTemplateVersion(pub String);
pub struct SchemaVersion(pub &'static str);

3.3 RawAsset

pub struct RawAsset {
    pub asset_id: AssetId,
    pub source_uri: SourceUri,
    pub workspace_path: WorkspacePath,
    pub media_type: MediaType,
    pub byte_len: u64,
    pub checksum: Checksum,
    pub discovered_at: OffsetDateTime,
    pub stored: AssetStorage,
}

pub enum SourceUri { File(PathBuf), Kb(String) }
pub struct WorkspacePath(pub String);

pub enum MediaType {
    Markdown,
    Pdf,
    Image(ImageType),
    Audio(AudioType),
    Other(String),
}

pub enum AssetStorage {
    Copied   { path: PathBuf },
    Reference{ path: PathBuf, sha: Checksum },
}

3.4 CanonicalDocument / Block / SourceSpan

pub struct CanonicalDocument {
    pub doc_id: DocumentId,
    pub source_asset_id: AssetId,
    pub workspace_path: WorkspacePath,
    pub title: String,
    pub lang: Lang,
    pub blocks: Vec<Block>,
    pub metadata: Metadata,
    pub provenance: Provenance,
    pub parser_version: ParserVersion,
    pub schema_version: u32,
    pub doc_version: u32,
}

pub enum Block {
    Heading(HeadingBlock),
    Paragraph(TextBlock),
    List(ListBlock),
    Code(CodeBlock),
    Table(TableBlock),
    Quote(TextBlock),
    ImageRef(ImageRefBlock),
    AudioRef(AudioRefBlock),
}

pub struct CommonBlock {
    pub block_id: BlockId,
    pub heading_path: Vec<String>,
    pub source_span: SourceSpan,
}

pub struct HeadingBlock { pub common: CommonBlock, pub level: u8, pub text: String }
pub struct TextBlock    { pub common: CommonBlock, pub text: String, pub inlines: Vec<Inline> }
pub struct ListBlock    { pub common: CommonBlock, pub ordered: bool, pub items: Vec<TextBlock> }
pub struct CodeBlock    { pub common: CommonBlock, pub lang: Option<String>, pub code: String }
pub struct TableBlock   { pub common: CommonBlock, pub headers: Vec<String>, pub rows: Vec<Vec<String>> }
pub struct ImageRefBlock{
    pub common: CommonBlock,
    pub asset_id: Option<AssetId>,
    pub src: String,
    pub alt: String,
    pub ocr: Option<OcrText>,
    pub caption: Option<ModelCaption>,
}
pub struct AudioRefBlock{
    pub common: CommonBlock,
    pub asset_id: AssetId,
    pub duration_ms: u64,
    pub transcript: Option<Transcript>,
}

pub enum Inline {
    Text(String),
    Code(String),
    Link { text: String, href: String },
    Strong(Vec<Inline>),
    Emph(Vec<Inline>),
}

pub enum SourceSpan {
    Line   { start: u32, end: u32 },
    Byte   { start: u64, end: u64 },
    Page   { page: u32, char_start: Option<u32>, char_end: Option<u32> },
    Region { x: u32, y: u32, w: u32, h: u32 },
    Time   { start_ms: u64, end_ms: u64 },
}

3.5 Chunk / Citation

pub struct Chunk {
    pub chunk_id: ChunkId,
    pub doc_id: DocumentId,
    pub block_ids: Vec<BlockId>,
    pub text: String,
    pub heading_path: Vec<String>,
    pub source_spans: Vec<SourceSpan>,
    pub token_estimate: usize,
    pub chunker_version: ChunkerVersion,
}

pub enum Citation {
    Line   { path: WorkspacePath, start: u32, end: u32, section: Option<String> },
    Page   { path: WorkspacePath, page: u32, section: Option<String> },
    Region { path: WorkspacePath, x: u32, y: u32, w: u32, h: u32 },
    Caption{ path: WorkspacePath, model: String },
    Time   { path: WorkspacePath, start_ms: u64, end_ms: u64, speaker: Option<String> },
}

impl Citation {
    pub fn path(&self) -> &WorkspacePath;
    pub fn to_uri(&self) -> String;
    pub fn parse(s: &str) -> Result<Self>;
}

3.6 Metadata / Provenance

pub struct Metadata {
    pub aliases: Vec<String>,
    pub tags: Vec<String>,
    pub created_at: OffsetDateTime,
    pub updated_at: OffsetDateTime,
    pub source_type: SourceType,
    pub trust_level: TrustLevel,
    pub user_id_alias: Option<String>,
    pub user: serde_json::Map<String, serde_json::Value>,
}

pub enum SourceType { Markdown, Note, Paper, Reference, Inbox }
pub enum TrustLevel { Primary, Secondary, Generated }

pub struct Provenance { pub events: Vec<ProvenanceEvent> }
pub struct ProvenanceEvent {
    pub at: OffsetDateTime,
    pub agent: String,
    pub kind: ProvenanceKind,
    pub note: Option<String>,
}
pub enum ProvenanceKind {
    Discovered, Parsed, Normalized, Chunked,
    OcrApplied, CaptionApplied, Transcribed,
    Embedded, Indexed, Warning, Error,
}

3.7 SearchQuery / SearchHit

pub enum SearchMode { Lexical, Vector, Hybrid }

pub struct SearchQuery {
    pub text: String,
    pub mode: SearchMode,
    pub k: usize,
    pub filters: SearchFilters,
}

pub struct SearchFilters {
    pub tags_any: Vec<String>,
    pub lang: Option<Lang>,
    pub path_glob: Option<String>,
    pub trust_min: Option<TrustLevel>,
}

pub struct SearchHit {
    pub rank: u32,
    pub chunk_id: ChunkId,
    pub doc_id: DocumentId,
    pub doc_path: WorkspacePath,
    pub heading_path: Vec<String>,
    pub section_label: Option<String>,
    pub snippet: String,
    pub citation: Citation,
    pub retrieval: RetrievalDetail,
    pub index_version: IndexVersion,
    pub embedding_model: Option<EmbeddingModelId>,
    pub chunker_version: ChunkerVersion,
}

pub struct RetrievalDetail {
    pub method: SearchMode,
    pub fusion_score: f32,
    pub lexical_score: Option<f32>,
    pub vector_score: Option<f32>,
    pub lexical_rank: Option<u32>,
    pub vector_rank: Option<u32>,
}

3.7a Forward-declared types

Block::ImageRef / AudioRef variant 은 v1 부터 존재하나, 그 안의 ocr / caption / transcript 필드는 P1 에선 항상 None. 다음 타입은 kb-core 에 stub 으로 둠 (최종 도메인 모델 슬롯):

pub struct OcrText      { pub joined: String, pub regions: Vec<OcrRegion>, pub engine: String, pub engine_version: String }
pub struct OcrRegion    { pub bbox: (u32, u32, u32, u32), pub text: String, pub confidence: f32 }
pub struct ModelCaption { pub text: String, pub model: String, pub model_version: String }
pub struct Transcript   { pub segments: Vec<TranscriptSegment>, pub engine: String, pub engine_version: String, pub language: Lang }
pub struct TranscriptSegment { pub start_ms: u64, pub end_ms: u64, pub text: String, pub speaker: Option<String>, pub confidence: Option<f32> }

pub struct Checksum(pub String);    // full blake3 hex (64 chars)
pub struct Lang(pub String);
pub enum   ImageType { Png, Jpeg, Webp, Gif, Tiff, Other(String) }
pub enum   AudioType { M4a, Mp3, Wav, Flac, Ogg, Other(String) }

ExtractConfig, DocFilter, JobKind, JobStatus, JobFilter, JobRow, JobId, VectorRecord, VectorHit, RefusalSignal, NoHitSignal, DoctorUnhealthy 도 kb-core 에 정의 (자세한 필드는 사용 시 결정, 이 spec 에서 forward-ref 만 보장).

OffsetDateTime 는 time::OffsetDateTime, Result 는 crate-local alias.

3.7b Parser intermediate types — kb-parse-types

Parser 의 중간 표현 (ParsedBlock 류) 은 kb-core 가 아니라 별도의 thin crate kb-parse-types 에 둔다. 이유: kb-normalize 는 medium-agnostic 한 ID/Provenance lift 를 책임지고 어떤 parser 도 직접 import 하면 안 된다. 그러나 normalize 에 들어오는 입력 타입이 어딘가에 정의되어야 하는데, 그것을 kb-core 에 박으면 (a) parser-별 ParsedBlock 변종 (ParsedImageRegion, ParsedPdfPage, ParsedAudioSegment) 이 향후 합류할 때 core 의 namespace 가 폭발하고, (b) parser 의 의미 변경이 core 변경이 되어 모든 의존자가 영향을 받는다.

kb-parse-types 는 이 둘 사이의 유일한 layer 다. 의존 그래프:

kb-core (도메인 모델 — Block, Chunk, SourceSpan, IDs, …)
   ▲
   │
kb-parse-types (parser 중간 표현 — ParsedBlock, ParsedImageRegion[P+], ParsedPdfPage[P+], ParsedAudioSegment[P+], Inline)
   ▲                            ▲
   │                            │
kb-parse-md, kb-parse-pdf,      kb-normalize
kb-parse-image, kb-parse-audio

kb-parse-types 는:

• kb-core 에만 의존 (Block, SourceSpan, Lang 등 도메인 타입 사용).
• 다른 어떤 kb-* 에도 의존하지 않는다.
• 어떤 parser 의 구체 라이브러리 (pulldown-cmark, pdf-extract, image, whisper-rs) 에도 의존하지 않는다.
• serde + thiserror 정도의 외부 의존만 가진다.

P1 에서 정의되는 타입:

// kb-parse-types — depends on kb-core only.
pub struct ParsedBlock {
    pub kind: ParsedBlockKind,
    pub heading_path: Vec<String>,
    pub source_span: kb_core::SourceSpan,
    pub payload: ParsedPayload,
}

pub enum ParsedBlockKind { Heading, Paragraph, List, Code, Table, Quote, ImageRef, AudioRef }

pub enum ParsedPayload {
    Heading   { level: u8, text: String },
    Paragraph { text: String, inlines: Vec<kb_core::Inline> },
    List      { ordered: bool, items: Vec<Vec<kb_core::Inline>> },
    Code      { lang: Option<String>, code: String },
    Table     { headers: Vec<String>, rows: Vec<Vec<String>> },
    Quote     { text: String, inlines: Vec<kb_core::Inline> },
    ImageRef  { src: String, alt: String },
    AudioRef  { src: String },                        // duration_ms filled by extractor before chunking
}

pub struct Warning { pub kind: WarningKind, pub note: String }
pub enum WarningKind { MalformedFrontmatter, MalformedTable, EncodingFallback, ExtractFailed }

Inline 은 kb-core (§3.4) 에 있는 도메인 타입. kb-parse-types 는 그것을 참조 만 한다 — 같은 의미를 두 crate 에 중복 정의하지 않는다 (그러면 normalize 가 identity-conversion 을 해야 해서 무의미).

P6/P7/P8 에서 추가될 타입 (forward-ref):

pub struct ParsedImageRegion { /* OCR/EXIF 추출 전 단계 */ }
pub struct ParsedPdfPage     { pub page: u32, pub text: String }
pub struct ParsedAudioSegment { pub start_ms: u64, pub end_ms: u64, pub text: String }

→ 새 medium 추가 시 kb-core::Block 변종은 변하지 않고, kb-parse-types 만 확장된다.

3.8 Answer / RAG types

pub struct Answer {
    pub answer: String,
    pub citations: Vec<AnswerCitation>,
    pub grounded: bool,
    pub refusal_reason: Option<RefusalReason>,
    pub model: ModelRef,
    pub embedding: Option<ModelRef>,
    pub prompt_template_version: PromptTemplateVersion,
    pub retrieval: AnswerRetrievalSummary,
    pub usage: TokenUsage,
    pub created_at: OffsetDateTime,
}

pub struct AnswerCitation { pub marker: Option<String>, pub citation: Citation }
pub enum RefusalReason { ScoreGate, LlmSelfJudge, NoIndex, NoChunks }

pub struct ModelRef {
    pub id: String,
    pub provider: String,
    pub dimensions: Option<usize>,
}

pub struct AnswerRetrievalSummary {
    pub trace_id: TraceId,
    pub mode: SearchMode,
    pub k: usize,
    pub score_gate: f32,
    pub top_score: f32,
    pub chunks_returned: u32,
    pub chunks_used: u32,
}

pub struct TokenUsage {
    pub prompt_tokens: u32,
    pub completion_tokens: u32,
    pub latency_ms: u32,
}

pub struct TraceId(pub String);

---

4. ID 생성 recipe

규칙: 모든 ID = blake3(canonical_json(tuple)) 의 hex prefix 32 chars.

4.1 canonical_json

• key 정렬 (BTreeMap / serde-json-canonicalizer)
• ASCII whitespace 없음
• UTF-8 NFC 정규화
• 숫자: integer/float 표준 표현
• 배열 순서 보존

4.2 Recipe

fn id_from<T: Serialize>(tuple: T) -> String {
    let bytes = canonical_json::to_vec(&tuple).unwrap();
    let hex = blake3::hash(&bytes).to_hex().to_string();
    hex[..32].to_string()
}

asset_id     = id_from({ kind: "asset", asset_blake3: <full hex of raw bytes> })
doc_id       = id_from({ kind: "doc",   workspace_path, asset_id, parser_version })
block_id     = id_from({ kind: "block", doc_id, block_kind, heading_path, ordinal, source_span })
chunk_id     = id_from({ kind: "chunk", doc_id, chunker_version, block_ids, policy_hash })
embedding_id = id_from({ kind: "embedding", chunk_id, model_id, model_version, dimensions })
index_id     = id_from({ kind: "index", collection, embedding_model, dimensions, index_version, index_kind, index_params_hash })

workspace_path 정규화: workspace root 기준 POSIX 슬래시, NFC, leading ./ 제거, 중복 슬래시 제거.

4.3 변경 영향 행렬

변경	영향 받는 ID
파일 내용 변경	asset_id → doc_id → block_id → chunk_id → embedding_id
파일 이동 (workspace 안)	doc_id → …
parser_version bump	doc_id → block_id → chunk_id → embedding_id
chunker_version 또는 policy 변경	chunk_id → embedding_id
embedding model/version/dim 변경	embedding_id
index 형상 변경	index_id

4.4 Tests

• 동일 입력 → 동일 ID (회귀 1000회).
• 입력 순서 미세 차이 → ID 변화 없음 (key 정렬).
• POSIX path 케이스 (./a/b.md vs a/b.md) → 동일.
• NFC 차이 한국어 글자 → 동일.

---

5. SQLite 스키마

PRAGMA foreign_keys = ON; journal_mode = WAL; synchronous = NORMAL;. UTF-8. timestamps RFC3339 TEXT.

5.1 Migrations meta

CREATE TABLE schema_meta (
  key   TEXT PRIMARY KEY,
  value TEXT NOT NULL
);
CREATE TABLE migrations (
  id          INTEGER PRIMARY KEY,
  applied_at  TEXT NOT NULL,
  description TEXT NOT NULL
);

5.2 Assets

CREATE TABLE assets (
  asset_id        TEXT PRIMARY KEY,
  source_uri      TEXT NOT NULL,
  workspace_path  TEXT NOT NULL,
  media_type      TEXT NOT NULL,
  byte_len        INTEGER NOT NULL,
  checksum        TEXT NOT NULL,
  storage_kind    TEXT NOT NULL CHECK (storage_kind IN ('copied','reference')),
  storage_path    TEXT NOT NULL,
  discovered_at   TEXT NOT NULL
);
CREATE UNIQUE INDEX idx_assets_workspace_path  ON assets(workspace_path);
CREATE INDEX        idx_assets_media_type      ON assets(media_type);

5.3 Documents

CREATE TABLE documents (
  doc_id          TEXT PRIMARY KEY,
  asset_id        TEXT NOT NULL REFERENCES assets(asset_id) ON DELETE RESTRICT,
  workspace_path  TEXT NOT NULL,
  title           TEXT,
  lang            TEXT,
  source_type     TEXT NOT NULL,
  trust_level     TEXT NOT NULL,
  parser_version  TEXT NOT NULL,
  doc_version     INTEGER NOT NULL,
  schema_version  INTEGER NOT NULL,
  metadata_json   TEXT NOT NULL,
  provenance_json TEXT NOT NULL,
  created_at      TEXT NOT NULL,
  updated_at      TEXT NOT NULL
);
CREATE UNIQUE INDEX idx_docs_workspace_path ON documents(workspace_path);
CREATE INDEX        idx_docs_lang           ON documents(lang);
CREATE INDEX        idx_docs_source_type    ON documents(source_type);

CREATE TABLE document_tags (
  doc_id TEXT NOT NULL REFERENCES documents(doc_id) ON DELETE CASCADE,
  tag    TEXT NOT NULL,
  PRIMARY KEY (doc_id, tag)
);
CREATE INDEX idx_document_tags_tag ON document_tags(tag);

5.4 Blocks

CREATE TABLE blocks (
  block_id          TEXT PRIMARY KEY,
  doc_id            TEXT NOT NULL REFERENCES documents(doc_id) ON DELETE CASCADE,
  kind              TEXT NOT NULL,
  heading_path_json TEXT NOT NULL,
  ordinal           INTEGER NOT NULL,
  source_span_json  TEXT NOT NULL,
  payload_json      TEXT NOT NULL
);
CREATE INDEX idx_blocks_doc_id ON blocks(doc_id);

5.5 Chunks + FTS5

CREATE TABLE chunks (
  chunk_id          TEXT PRIMARY KEY,
  doc_id            TEXT NOT NULL REFERENCES documents(doc_id) ON DELETE CASCADE,
  text              TEXT NOT NULL,
  heading_path_json TEXT NOT NULL,
  section_label     TEXT,
  source_spans_json TEXT NOT NULL,
  token_estimate    INTEGER NOT NULL,
  chunker_version   TEXT NOT NULL,
  policy_hash       TEXT NOT NULL,
  block_ids_json    TEXT NOT NULL,
  created_at        TEXT NOT NULL
);
CREATE INDEX idx_chunks_doc_id          ON chunks(doc_id);
CREATE INDEX idx_chunks_chunker_version ON chunks(chunker_version);

CREATE VIRTUAL TABLE chunks_fts USING fts5(
  chunk_id     UNINDEXED,
  doc_id       UNINDEXED,
  heading_path,
  text,
  tokenize = 'unicode61 remove_diacritics 2'
);

CREATE TRIGGER chunks_ai AFTER INSERT ON chunks BEGIN
  INSERT INTO chunks_fts(chunk_id, doc_id, heading_path, text)
  VALUES (new.chunk_id, new.doc_id, new.heading_path_json, new.text);
END;
CREATE TRIGGER chunks_ad AFTER DELETE ON chunks BEGIN
  DELETE FROM chunks_fts WHERE chunk_id = old.chunk_id;
END;
CREATE TRIGGER chunks_au AFTER UPDATE ON chunks BEGIN
  DELETE FROM chunks_fts WHERE chunk_id = old.chunk_id;
  INSERT INTO chunks_fts(chunk_id, doc_id, heading_path, text)
  VALUES (new.chunk_id, new.doc_id, new.heading_path_json, new.text);
END;

5.6 Embedding records (P3)

CREATE TABLE embedding_records (
  embedding_id   TEXT PRIMARY KEY,
  chunk_id       TEXT NOT NULL REFERENCES chunks(chunk_id) ON DELETE CASCADE,
  model_id       TEXT NOT NULL,
  model_version  TEXT NOT NULL,
  dimensions     INTEGER NOT NULL,
  lance_table    TEXT NOT NULL,
  created_at     TEXT NOT NULL,
  UNIQUE(chunk_id, model_id, model_version, dimensions)
);
CREATE INDEX idx_embed_chunk ON embedding_records(chunk_id);
CREATE INDEX idx_embed_model ON embedding_records(model_id, model_version, dimensions);

5.7 Jobs / IngestRuns / Answers / EvalRuns

CREATE TABLE jobs (
  job_id        TEXT PRIMARY KEY,
  kind          TEXT NOT NULL,
  status        TEXT NOT NULL CHECK (status IN ('pending','running','succeeded','failed','canceled')),
  payload_json  TEXT NOT NULL,
  progress_json TEXT,
  error_json    TEXT,
  created_at    TEXT NOT NULL,
  updated_at    TEXT NOT NULL,
  finished_at   TEXT
);
CREATE INDEX idx_jobs_status ON jobs(status);
CREATE INDEX idx_jobs_kind   ON jobs(kind);

CREATE TABLE ingest_runs (
  run_id        TEXT PRIMARY KEY,
  scope_json    TEXT NOT NULL,
  scanned       INTEGER NOT NULL,
  new_count     INTEGER NOT NULL,
  updated_count INTEGER NOT NULL,
  skipped_count INTEGER NOT NULL,
  error_count   INTEGER NOT NULL,
  duration_ms   INTEGER NOT NULL,
  started_at    TEXT NOT NULL,
  finished_at   TEXT NOT NULL,
  items_json    TEXT
);

CREATE TABLE answers (
  trace_id                TEXT PRIMARY KEY,
  query                   TEXT NOT NULL,
  answer                  TEXT NOT NULL,
  grounded                INTEGER NOT NULL,
  refusal_reason          TEXT,
  model_id                TEXT NOT NULL,
  model_provider          TEXT NOT NULL,
  embedding_model_id      TEXT,
  embedding_dimensions    INTEGER,
  prompt_template_version TEXT NOT NULL,
  retrieval_mode          TEXT NOT NULL,
  retrieval_k             INTEGER NOT NULL,
  score_gate              REAL NOT NULL,
  top_score               REAL NOT NULL,
  chunks_returned         INTEGER NOT NULL,
  chunks_used             INTEGER NOT NULL,
  citations_json          TEXT NOT NULL,
  packed_chunks_json      TEXT,
  prompt_tokens           INTEGER,
  completion_tokens       INTEGER,
  latency_ms              INTEGER,
  created_at              TEXT NOT NULL
);
CREATE INDEX idx_answers_created_at ON answers(created_at);
CREATE INDEX idx_answers_grounded   ON answers(grounded);

CREATE TABLE eval_runs (
  run_id              TEXT PRIMARY KEY,
  suite               TEXT NOT NULL,
  config_snapshot_json TEXT NOT NULL,
  aggregate_json      TEXT NOT NULL,
  commit_hash         TEXT,
  created_at          TEXT NOT NULL
);
CREATE TABLE eval_query_results (
  run_id   TEXT NOT NULL REFERENCES eval_runs(run_id) ON DELETE CASCADE,
  query_id TEXT NOT NULL,
  result_json TEXT NOT NULL,
  PRIMARY KEY (run_id, query_id)
);

5.8 트랜잭션 정책

• ingest 1 doc = 1 트랜잭션.
• bulk ingest 는 doc 단위 커밋.
• chunker/embedding 재처리 = 별도 job + per-chunk 트랜잭션.

5.9 마이그레이션

migrations/V001__init.sql, V002__*.sql 형식. 시작 시 schema_meta.schema_version 확인 → 누락된 마이그레이션 적용. 다운그레이드 미지원.

---

6. Filesystem + config layout

6.1 Path resolution (XDG)

종류	기본 위치
워크스페이스	~/KnowledgeBase/
config	~/.config/kb/config.toml
data	~/.local/share/kb/
cache	~/.cache/kb/
state (logs)	~/.local/state/kb/

~, $HOME, ${KB_*} expand. 절대 path 정규화 후 사용.

6.2 Workspace 구조

~/KnowledgeBase/
├── inbox/   notes/   papers/   photos/   recordings/
└── .kbignore

.kbignore 와 config.workspace.exclude 합집합.

6.3 Data dir 구조

~/.local/share/kb/
├── kb.sqlite (+ -wal, -shm)
├── lancedb/
│   └── chunk_embeddings_<model>_<dim>.lance/
├── assets/<aa>/<asset_id>     # shard
├── artifacts/<doc_id>/        # ocr.json / caption.json / transcript.json / pdf-text.json
├── models/                    # fastembed/  ollama 캐시 위임
└── runs/<run_id>/             # eval per_query.jsonl + report.md

6.4 Config (~/.config/kb/config.toml) — frozen schema

schema_version = 1

[workspace]
root    = "~/KnowledgeBase"
include = ["**/*.md"]
exclude = [".git/**", "node_modules/**", ".obsidian/**"]

[storage]
data_dir          = "${XDG_DATA_HOME:-~/.local/share}/kb"
sqlite            = "{data_dir}/kb.sqlite"
vector_dir        = "{data_dir}/lancedb"
asset_dir         = "{data_dir}/assets"
artifact_dir      = "{data_dir}/artifacts"
model_dir         = "{data_dir}/models"
runs_dir          = "{data_dir}/runs"
copy_threshold_mb = 100

[indexing]
max_parallel_extractors = 2
max_parallel_embeddings = 1
watch_filesystem        = false

[chunking]
target_tokens             = 500
overlap_tokens            = 80
respect_markdown_headings = true
chunker_version           = "md-heading-v1"

[models.embedding]
provider   = "fastembed"
model      = "multilingual-e5-small"
version    = "v1"
dimensions = 384
batch_size = 64

[models.llm]
provider       = "ollama"
model          = "qwen2.5:14b-instruct"
context_tokens = 32768
endpoint       = "http://127.0.0.1:11434"
temperature    = 0.0
seed           = 0

[search]
default_k     = 10
hybrid_fusion = "rrf"
rrf_k         = 60
snippet_chars = 220

[rag]
prompt_template_version = "rag-v1"
score_gate              = 0.30
explain_default         = false
max_context_tokens      = 8000

config 우선순위: default → file → env (KB_<SECTION>_<KEY>) → CLI flag.

6.5 kb init 출력

$ kb init
created  ~/.config/kb/config.toml
created  ~/.local/share/kb/
created  ~/KnowledgeBase/
opened   ~/.local/share/kb/kb.sqlite (schema v1)
hint     edit ~/.config/kb/config.toml then `kb ingest ~/KnowledgeBase`

기존 파일 보존, --force 명시 필요.

6.6 Permissions / portability

• 디렉토리 0o755, 파일 0o644.
• 항상 POSIX path 정규화 후 DB 저장. to_posix 단일 함수.
• 심볼릭 링크: 1차 follow + 무한루프 detect (canonicalize 후 set 추적).

---

7. Trait contracts (kb-core)

7.1 입출력 보조

pub struct SourceScope { pub root: PathBuf, pub include: Vec<String>, pub exclude: Vec<String> }
pub struct ExtractContext<'a> { pub asset: &'a RawAsset, pub workspace_root: &'a Path, pub config: &'a ExtractConfig }

pub struct ChunkPolicy {
    pub target_tokens: usize,
    pub overlap_tokens: usize,
    pub respect_markdown_headings: bool,
    pub chunker_version: ChunkerVersion,
}

pub enum EmbeddingKind { Document, Query }
pub struct EmbeddingInput<'a> { pub text: &'a str, pub kind: EmbeddingKind }

pub struct GenerateRequest {
    pub system: String,
    pub user: String,
    pub stop: Vec<String>,
    pub max_tokens: usize,
    pub temperature: f32,
    pub seed: Option<u64>,
}

pub enum TokenChunk {
    Token(String),
    Done { finish_reason: FinishReason, usage: TokenUsage },
}
pub enum FinishReason { Stop, Length, Aborted, Error(String) }

7.2 트레잇

pub trait SourceConnector {
    fn scan(&self, scope: &SourceScope) -> Result<Vec<RawAsset>>;
}

pub trait Extractor: Send + Sync {
    fn supports(&self, media_type: &MediaType) -> bool;
    fn parser_version(&self) -> ParserVersion;
    fn extract(&self, ctx: &ExtractContext, bytes: &[u8]) -> Result<CanonicalDocument>;
}

pub trait Chunker: Send + Sync {
    fn chunker_version(&self) -> ChunkerVersion;
    fn policy_hash(&self, policy: &ChunkPolicy) -> String;
    fn chunk(&self, doc: &CanonicalDocument, policy: &ChunkPolicy) -> Result<Vec<Chunk>>;
}

pub trait Embedder: Send + Sync {
    fn model_id(&self) -> EmbeddingModelId;
    fn model_version(&self) -> EmbeddingVersion;
    fn dimensions(&self) -> usize;
    fn embed(&self, inputs: &[EmbeddingInput]) -> Result<Vec<Vec<f32>>>;
}

pub trait Retriever: Send + Sync {
    fn search(&self, query: &SearchQuery) -> Result<Vec<SearchHit>>;
    fn index_version(&self) -> IndexVersion;
}

pub trait LanguageModel: Send + Sync {
    fn model_ref(&self) -> ModelRef;
    fn context_tokens(&self) -> usize;
    fn generate_stream(
        &self,
        req: GenerateRequest,
    ) -> Result<Box<dyn Iterator<Item = Result<TokenChunk>> + Send>>;
}

pub trait DocumentStore {
    fn put_asset(&self, a: &RawAsset) -> Result<()>;
    fn put_document(&self, d: &CanonicalDocument) -> Result<()>;
    fn put_blocks(&self, doc: &DocumentId, blocks: &[Block]) -> Result<()>;
    fn put_chunks(&self, doc: &DocumentId, chunks: &[Chunk]) -> Result<()>;
    fn get_document(&self, id: &DocumentId) -> Result<Option<CanonicalDocument>>;
    fn get_chunk(&self, id: &ChunkId) -> Result<Option<Chunk>>;
    fn list_documents(&self, filter: &DocFilter) -> Result<Vec<DocSummary>>;
}

pub trait VectorStore {
    fn ensure_table(&self, model: &EmbeddingModelId, dim: usize) -> Result<IndexId>;
    fn upsert(&self, recs: &[VectorRecord]) -> Result<()>;
    fn search(&self, query_vec: &[f32], k: usize, filters: &SearchFilters) -> Result<Vec<VectorHit>>;
}

pub trait JobRepo {
    fn create(&self, kind: JobKind, payload: serde_json::Value) -> Result<JobId>;
    fn update_progress(&self, id: &JobId, progress: serde_json::Value) -> Result<()>;
    fn finish(&self, id: &JobId, status: JobStatus, error: Option<&str>) -> Result<()>;
    fn list(&self, filter: &JobFilter) -> Result<Vec<JobRow>>;
}

---

8. 모듈 경계 (Allowed / Forbidden)

kb-cli, kb-tui, kb-desktop
   └─> kb-app
         ├─> kb-source-fs
         ├─> kb-parse-md / kb-parse-pdf / kb-parse-image / kb-parse-audio
         │     └─> kb-parse-types (parser intermediate)
         ├─> kb-normalize
         │     └─> kb-parse-types
         ├─> kb-chunk
         ├─> kb-store-sqlite (DocumentStore, JobRepo, Retriever[lexical])
         ├─> kb-store-vector (VectorStore)
         ├─> kb-embed-local
         ├─> kb-search (Retriever[hybrid])
         ├─> kb-llm-local
         ├─> kb-rag
         ├─> kb-eval
         └─> kb-config
              └─> kb-core (모두 의존)

kb-parse-types 는 kb-core 와 parsers/normalize 사이의 thin layer (§3.7b 참조). parser-별 중간 표현 (ParsedBlock, ParsedImageRegion, ParsedPdfPage, ParsedAudioSegment, Inline) 을 한 곳에 모아 (a) kb-core 의 namespace 폭발을 막고 (b) kb-normalize 가 parser 를 직접 import 하지 않게 한다.

핵심 금지:

• UI → store/llm/parse 직접 의존 ✗
• parse-* → store/llm/embed ✗
• parse-* → kb-normalize ✗ (단방향: parsers → kb-parse-types ← normalize)
• chunk → llm/embed ✗
• normalize → store / parse-* ✗
• kb-parse-types → 어떤 parser/normalize/store/llm/embed/search/rag/ui ✗ (kb-core 만 의존)
• 다른 store 와 cross-write ✗

cargo deny + workspace deny.toml + CI 체크로 강제.

---

9. Versioning rules

식별자	변경 시	bump 규칙
parser_version	파서 의미 변화	semver-suffix string 상수
chunker_version	chunk boundary/policy 변화	라벨 (md-heading-v2)
policy_hash	policy 값만 변경	자동 (config 해시)
embedding_model.id	모델 교체	새 lance 테이블
embedding_model.version	같은 모델 가중치/토크나이저 변경	bump
embedding.dimensions	차원 변경	새 lance 테이블 강제
index_version	retrieval 형상 변화	bump
prompt_template_version	template 변경	코드 상수 (rag-v2)
DB schema_version	DDL 변경	마이그레이션 정수 증가
wire schema (*.v1)	깨는 변경 시	*.v2 신설, v1 additive only
internal Rust struct	자유 진화	wire 분리되어 외부 영향 0

CI:

• 코드 변경 PR 에서 parser_version / chunker_version 동일하게 유지됐는데 동작 테스트 결과 다르면 fail.
• DDL 변경 있는데 마이그레이션 정수 미증가 fail.
• v1 JSON schema 파일 변경 시 additive 검증.

---

10. 에러 모델 + exit codes

// kb-core
pub enum CoreError { InvalidId, InvalidCitation, InvalidSpan, Malformed }
// crate-local examples
pub enum ParseMdError { Yaml(String), Encoding, Pulldown(String), Span }
pub enum StoreError { Sqlx(rusqlite::Error), Migration(String), Conflict(String) }
pub enum LlmError { Unreachable, ModelNotPulled(String), Timeout, Stream(String) }

Boundary (kb-app, kb-cli) 에서 anyhow::Error 합침. exit code 매핑:

fn exit_code(err: &anyhow::Error) -> i32 {
    if err.downcast_ref::<RefusalSignal>().is_some()    { return 1; }
    if err.downcast_ref::<NoHitSignal>().is_some()      { return 1; }
    if err.downcast_ref::<DoctorUnhealthy>().is_some()  { return 3; }
    2
}

레벨	메시지
default	error: <한 줄>\n hint: <조치>
--verbose	+ anyhow chain
--debug 또는 RUST_LOG=debug	+ tracing target/level/span

Refusal 은 에러 아님. kb ask 거절은 정상 stdout (Answer with grounded=false) + exit 1.

Logging: tracing + tracing-subscriber + tracing-appender daily roll, ~/.local/state/kb/logs/. structured (trace_id, doc_id, chunk_id).

kb doctor 출력 (사람):

$ kb doctor
✓ config_loaded         ~/.config/kb/config.toml
✓ data_dir_writable     ~/.local/share/kb
✓ sqlite_open           kb.sqlite (schema v1)
✓ lancedb_open          lancedb/
✓ embedding_model       multilingual-e5-small (384d)
✓ ollama_reachable      http://127.0.0.1:11434
✗ ollama_model_pulled   qwen2.5:14b-instruct missing
                        hint: ollama pull qwen2.5:14b-instruct

1 check failed.

---

11. 동결 범위 / 변경 정책

이 문서가 동결 ↔ 다음 컴포넌트 분해 작업이 안전:

• 모든 wire schema (docs/wire-schema/v1/*.schema.json)
• 모든 trait 시그니처 (kb-core)
• 모든 ID recipe (4.2)
• SQLite DDL (5장)
• Filesystem + config schema (6장)
• 모듈 경계 (8장)
• exit codes / refusal 정책

변경하려면: 이 문서에 다이어그램이나 이슈 포인트를 명기 → 영향 범위 (파급 task 목록) 적시 → 그 후에만 task 분해 수정.

의도적으로 빠진 것 (out of scope, P+):

• multi-workspace
• watch mode
• desktop app kb:// protocol handler
• LLM-as-judge eval
• visual embedding (CLIP)
• real-time collab
• enterprise auth

---

12. 다음 단계

1. 이 문서 검토.
2. 검토 통과 시 tasks/_template.md (작업 단위 spec 템플릿) 작성.
3. P1 (Markdown ingestion) 6 component task 로 분해 — 템플릿 적합성 검증.
4. 나머지 phase 일괄 분해 (~30 component task).

각 task 는 이 문서의 trait 시그니처 + wire schema + DDL 만 인용. 새 도메인 타입 / 새 trait 도입 금지 (이 문서 수정 절차 거쳐야 함).