kebab/CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project

Single-user local-first knowledge base + RAG. Rust 2024 workspace, ~21 crates, single binary (kebab). All inference is local (Ollama + fastembed + whisper.cpp).

The repo's documentation is split by audience — don't duplicate across them:

• README.md — first stop for an end user. Quick start, command table, one Mermaid logical-architecture diagram, configuration pointers, license. Stays narrow.
• HANDOFF.md — phase-level progress dashboard for someone picking the project up. Phase status table, component count, "next task candidates", short summary of post-merge deviations. The README never duplicates this.
• docs/ARCHITECTURE.md — internal structure: crate dependency graph, directory tree, locked-in technical decisions. The README links here from the Mermaid diagram.
• docs/superpowers/specs/2026-04-27-kebab-final-form-design.md — frozen design contract.
• tasks/INDEX.md — per-component task tree.
• tasks/HOTFIXES.md — dated post-merge deviation log; live source of truth where behavior and the frozen spec disagree.

Build / test / lint

cargo test -p <crate>                          # preferred — workspace has 18 crates
cargo test -p <crate> <test_name>              # single test (substring match)
cargo test --workspace --no-fail-fast -j 1     # full suite — see -j 1 below
cargo clippy --workspace --all-targets -- -D warnings   # CI gate
cargo build --release                          # produces target/release/kebab

-j 1 for the full workspace test isn't optional: 18 integration-test binaries each link lance + datafusion + arrow + tantivy and the parallel link step exhausts memory (linker gets SIGKILL'd, build silently fails partway). Per-crate runs are fine in parallel.

target/ is 6–10 GB after a fresh build but balloons to 90+ GB after a few task cycles (each fb-* batch adds incremental compile artifacts on top of the existing 18 × test-binary debug info). The dev/test profile is already trimmed (debug = "line-tables-only", split-debuginfo = "unpacked" — see workspace Cargo.toml). Run cargo clean routinely after each merged PR, not just "if pressure shows up" — disk space is tight and recovery via cargo clean is cheap (one re-link per crate on next build). Verified pattern: 92 GB → 0 GB in seconds, backtraces still resolve to function + line.

The facade rule

kebab-app is the only crate UI binaries (kebab-cli, future kebab-tui, kebab-desktop) may touch. Every user-facing entry has a *_with_config(cfg, …) companion that takes an explicit Config:

• kebab-cli calls the *_with_config form so --config <path> is honored.
• The bare kebab_app::ingest(...) / search(...) / ask(...) form re-loads Config::load(None) (XDG default) and silently bypasses any explicit path. Two regressions of exactly this shape are recorded in tasks/HOTFIXES.md (P3-5 + P4-3 follow-ups). When wiring a new CLI subcommand, always thread the Config through.

*_with_config is #[doc(hidden)] pub fn but it's the official config-explicit API, not a test seam.

Spec contract

docs/superpowers/specs/2026-04-27-kebab-final-form-design.md (12 sections) is the single contract for the whole workspace. Every component task spec under tasks/p<N>/ lists which contract_sections it implements.

• Changing the design doc requires updating every referencing task spec in the same PR.
• Task specs themselves stay frozen as the historical contract once the task is merged. Don't edit them retroactively to match what shipped.
• Live deviations from the original contract go in tasks/HOTFIXES.md as dated entries, plus a one-line cross-link in the original spec's Risks / notes. Treat HOTFIXES.md as the live source of truth when behavior and spec disagree.

tasks/INDEX.md is the dashboard for which phases / components are done; update its phase status when a phase epic completes.

Allowed / forbidden deps

Each task spec lists Allowed dependencies and Forbidden dependencies per design §8. The most load-bearing ones:

• kebab-core MUST NOT depend on any other kebab-* crate. Domain types only.
• kebab-eval's metrics and compare modules MUST NOT import retrieval / embedding / LLM crates directly. The runner is allowed to use kebab-app's facade (P5-1 inheritance — see deviations in that task spec).
• UI crates (kebab-cli, kebab-mcp, kebab-tui, future kebab-desktop) MUST NOT import kebab-store-* / kebab-llm-* / kebab-parse-* directly — only kebab-app.

Read the relevant task spec's deps section before adding an import. New crates inherit the same boundary rules.

Wire schema v1

All --json output carries a schema_version field. Current schemas: ingest_report.v1, ingest_progress.v1, search_hit.v1, answer.v1, doctor.v1, reset_report.v1, schema.v1, error.v1, chunk_inspection.v1, citation.v1, doc_summary.v1. Schemas live in docs/wire-schema/v1/. The wire shape is the contract for external integrations (Claude Code skills, MCP, etc.); breaking it requires a *.v2 major bump and parallel-running both for one phase. In --json mode, fatal errors emit error.v1 to stderr as ndjson (non---json mode keeps plain stderr text); exit codes 0/1/2/3 are unchanged — error.v1.code provides fine-grained agent branching.

In-tree integration packages live under integrations/<host>/ — currently integrations/claude-code/kebab/ (a Claude Code skill that calls kebab search --json / kebab ask --json). Any wire schema major bump (v1→v2) MUST update each shipped integration in the same PR, same as the version-cascade rule below. Per-user trigger keywords (team / system / acronym) belong in the user's local copy of the skill, not in the repo-shipped frontmatter — keep integrations/claude-code/kebab/SKILL.md's description generic.

Versioning cascade

parser_version / chunker_version / embedding_version / prompt_template_version / index_version follow the cascade rule in design §9. Changing any of these invalidates downstream records (chunks, embeddings, eval runs, …). When changing a version: either ship a re-process job or treat it as a breaking schema bump. The eval runner snapshots all five into eval_runs.config_snapshot_json.

Release / binary version bump

Workspace Cargo.toml 의 version 은 binary release 의 정체성. 다음 트리거 중 하나 발생 시 bump + 새 release 컷:

• 사용자가 새 바이너리로 도그푸딩 또는 실사용 을 할 필요가 있다고 명시.
• breaking schema change (V00X migration / wire schema major bump v1→v2 등) 가 머지된 후 — 이전 릴리즈 binary 가 새 DB / 새 wire 와 호환 안 됨. wire 의 additive minor 변경 (예: IngestReport.unchanged 같은 필드 추가) 은 backward-compat 이라 본 트리거에 해당 안 됨.
• frozen design contract 변경 (design §X 갱신) 이 머지된 후.

Bump 자체는 단순 minor / patch 한 줄 수정 (Cargo.toml workspace version) — 이미 모든 kebab-* crate 가 version = { workspace = true } 라 자동 cascade. 동시에 Cargo.lock 자동 갱신.

Release 절차:

1. gitea-release v<X.Y.Z> (gitea-ops skill) 으로 tag + push + release notes.
2. release notes 는 사용자 도그푸딩에 영향이 가는 surface 변경을 위주로 — wire schema 추가, CLI flag 신규, TUI 키 변경, V00X migration 등 — 다룬다. 이때 추가된 기능과 변경사항은 유저가 이해할 수 있도록 친절하고 자세하게 풀어서 설명해야 하며, 단순히 commit subject 를 나열하는 형태로 끝내면 안 된다. 필요하다면 도그푸딩이나 테스트 결과도 함께 적어 둔다.
3. 프리-1.0 (0.x.y) 단계: minor bump 시 wire schema additive / surface 변경 누적, patch bump 시 bug fix only.

bump 시점 = release 시점 같은 commit. 즉 commit chore: bump version 0.x → 0.y 직후 같은 commit 에 tag. v0.1.0 (2319206) 처럼 bump 없이 tag 만 찍는 패턴은 후속 release 가 대상 commit 을 헷갈리게 함 — pre-release snapshot 은 SHA reference 로 충분.

Dogfood trigger

도그푸딩 = 새 binary 를 실제 KB / 실제 query 로 돌려보고 user-visible 동작이 spec 의 의도와 일치하는지 확인하는 종단 검증. unit / integration test 가 못 잡는 회귀 (UX 어색함, performance regression, 의외의 token 처리, embedding drift, RAG hallucination) 를 catch 함. PR 머지 전 또는 머지 직후 release notes 작성 전에 실시.

도그푸딩이 필요한 시점

다음 트리거 중 하나라도 hit 시 도그푸딩 필수. 모두 release-level 또는 user-visible behavior 변경 임.

Schema / migration:

• 신규 V00X migration (예: V007 trigram, V008 OCR mirror, V009 morphological) — corpus_revision cascade + auto-backfill 정책의 사용자 경험 확인.
• frozen design contract 변경 (docs/superpowers/specs/2026-04-27-kebab-final-form-design.md §X 갱신) — verbatim CI diff-check 외의 user-visible side effect 확인.

Wire schema / CLI surface:

• 신규 --json 필드, exit code 변경, 또는 schema major bump (v1 → v2) — agent / external integration 의 호환성 검증.
• kebab 의 subcommand 또는 flag 추가/삭제/rename — agent skill / muscle memory 영향.

Search / RAG behavior:

• FTS5 tokenizer / chunker / embedder 모델 / RAG prompt template 변경 — 같은 query 의 hit ordering, snippet, RAG citation 패턴이 자연스럽게 변화하는지.
• score gate, RRF fusion ratio, NLI threshold 같은 ranking 파라미터 default 변경.

Performance:

• ingest / search / ask latency 의 의도된 변화 (예: lindera tokenize, OCR 추가, multi-hop RAG) — actual wall-clock 측정 + release notes 에 명시.
• 대용량 KB (수천 doc / 만 chunk) 의 first-boot eager backfill 시간이 사용자 hang 인지에 영향 안 가는지.

Language / locale:

• 한국어 / 일본어 / 중국어 lexical 동작 변경 (V007 trigram, V009 morphological, future N-gram).
• 영어 substring 매칭 같은 ad-hoc 부산물의 회귀.

File / asset surface:

• 신규 source 형식 (PDF OCR, audio, video) — extractor / chunker 의 실제 corpus 동작.
• .kebabignore / _external/ 같은 workspace 정책 변경.

Release-level: 위 트리거 중 하나가 hit 되어 Cargo.toml workspace version bump 가 필요하면, bump commit 이전에 도그푸딩 evidence 가 HOTFIXES + release notes 에 명시 되어 있어야 함. evidence 없는 release 는 사용자가 "왜 bump 했는지" 추적 불가.

도그푸딩 데이터 보관소

모든 도그푸딩 source 문서 + KB state + 로그는 /build/dogfood/ 한 디렉토리에 누적 보관한다. 분류는 문서 의미 / 종류 / 형식 기준만 — kebab version, 생성 시점, scenario name 같은 prefix 금지 (v0.20.1-dogfood/, dogfood-v018/ 같은 디렉토리 신설 X). 자세한 layout 은 /build/dogfood/README.md 참조.

• /build/dogfood/corpus/ — source 문서 (read-only). format 별 분류 (markdown/, code/, html/, images/, pdf/, manifest/, resources/) + 각 format 내 category 별 (예: markdown/{korean,english,bilingual,tech-docs,coding-md-corpus,topics,notes,edge-cases}, code/{rust,python,...}). 새 fixture 는 적절한 category subdir 에 추가.
• /build/dogfood/kb/ — 도그푸딩 run 의 KB 출력 (SQLite + LanceDB + assets + models). 매 run 마다 reset 가능. 별 KB 디렉토리 신설 X.
• /build/dogfood/logs/ — 누적 실행 로그 (ndjson + stderr + summary).
• /build/dogfood/config.toml — canonical 도그푸딩 config (없으면 kebab init 후 path override).
• /build/dogfood/_archive/ — regeneratable stale state (이전 run 의 sqlite/lancedb, XDG snapshot). 디스크 압박 시 wipe 가능.

/tmp/kebab-smoke/, /tmp/kebab-*, /build/cache/dogfood*, /home/altair823/KnowledgeBase, ~/.config/kebab/, ~/.local/share/kebab/, ~/.local/state/kebab/ 같은 위치 신규 사용 금지 — 모두 /build/dogfood/ 로 일관. ad-hoc fixture 가 필요하면 corpus/<format>/<category>/ 에 추가.

도그푸딩 결과 기록

도그푸딩 evidence 는 두 곳에 cascade:

1. tasks/HOTFIXES.md 의 dated entry — 시나리오 별 hit count 표 + snippet evidence + known limitation. 미래에 spec drift 의심 시 git history 외 immediate reference 가 됨.
2. docs/release-notes/v<X.Y.Z>-draft.md (또는 gitea release body) — 사용자 도그푸딩 영향에 영향이 가는 surface 변경을 4 단락 (변경 사실 / trade-off / mitigation / upgrade 절차) 으로 풀어서 설명. evidence link.

도그푸딩 단계에서 발견된 bug (spec 과 실제 동작의 mismatch, performance regression, UX 어색함) 는 즉시 fix → re-dogfood. fix 가 별 PR 으로 빠지면 머지 후 HOTFIXES 에 dated entry.

DOGFOOD scenario catalog (§1~§13) 는 docs/DOGFOOD.md. 신규 release 마다 §관련 section 의 scenario list 갱신 + 신규 scenario 추가.

Naming + paths

• Crate prefix: kebab- (kebab-case package, kebab_ snake_case in Rust modules).
• Binary: kebab.
• Env var prefix: KEBAB_* (e.g. KEBAB_RAG_SCORE_GATE, KEBAB_EVAL_GOLDEN, KEBAB_COMMIT_HASH).
• XDG paths: ~/.config/kebab/, ~/.local/share/kebab/, ~/.cache/kebab/, ~/.local/state/kebab/.
• SQLite filename: kebab.sqlite (under data_dir).
• Workspace ignore: .kebabignore (per directory).
• _external/ (under workspace.root): single-file / stdin ingest 가 외부 file 을 deterministic 명명 (<blake3-12>.<ext>) 으로 copy. 첫 생성 시 .kebabignore 자동 append.

The migration from the old kb name lives in commits 911fb49 / f1a448d / f9714aa. If you spot a leftover kb reference, treat it as a leftover and fix it (the rename PR sweep covered crates/, docs/, tasks/, README, design doc, fixtures — but workspace root Cargo.toml comments needed a follow-up; assume similar misses are possible).

Smoke + integration

docs/SMOKE.md walks through running the full pipeline against an isolated TempDir KB via --config /tmp/kebab-smoke/config.toml. Use this instead of touching ~/.local/share/kebab/ when verifying a fresh clone or a CLI flag change. Most CLI regressions surface here, not in unit tests (see HOTFIXES.md).

User-facing docs (README + HANDOFF + ARCHITECTURE)

Three sibling docs split the audience. Every implementation PR (feat/*) keeps them in sync; spec PRs (spec/*) don't touch any of the three.

README.md — end user. Stays narrow. The three surfaces a user touches:

• CLI — new kebab <subcommand>, flag, --json field, or exit-code change. Update the 명령 table and the Quick start block if the new flow needs a different invocation.
• TUI — new pane, key binding, or run-time behavior visible to a kebab tui user. Update the row in the 명령 table and the Mermaid diagram if a new external surface lands.
• Configuration — new config.toml field, KEBAB_* env, default change, or XDG path. Update the Configuration section AND the config example block in docs/SMOKE.md.

The Mermaid logical-architecture diagram stays the only diagram in the README. If a new media type / external service / store crosses the diagram boundary, update it; otherwise leave it alone.

The README does NOT carry: phase status, component count, post-merge deviations, crate dependency graph, directory tree, locked-in technical decisions. Those live in HANDOFF or ARCHITECTURE.

HANDOFF.md — handing off. Phase-level progress + next-task candidates. Flip the relevant phase row from ⏳ to ✅ when a phase epic completes. Add a one-line entry under "머지 후 발견된 버그 / 결정 (요약)" when a HOTFIXES entry lands that's load-bearing for someone picking up the project. Per-component progress lives in tasks/INDEX.md, not here.

docs/ARCHITECTURE.md — implementation detail. Crate dependency graph, directory tree, locked-in technical decisions. Update when:

• A new crate is added — extend the graph + directory tree.
• A locked-in decision flips (e.g. OCR engine default changes per a HOTFIXES entry) — update the table and link the HOTFIXES entry.
• A directory moves — update the tree.

Out of scope for all three: HOTFIXES detail (tasks/HOTFIXES.md), version cascade mechanics (CLAUDE.md §Versioning cascade), per-task spec rationale (tasks/p<N>/).

If a feature ships behind a flag that's off-by-default, mention the flag explicitly in the README so a user reading only the README knows the surface exists but is gated.

Remote

Git remote is Gitea: https://gitea.altair823.xyz/altair823-org/kebab.git. PRs are created via the Gitea REST API (POST /repos/altair823-org/kebab/pulls) — gh CLI does not work against this host. Auth uses ~/.netrc (populated via git credential fill).