DeepWiki App Space

AI code intelligence for any local codebase — a unified app combining a tree-sitter symbol/call graph (à la colbymchenry/codegraph) with an AI-generated, source-grounded wiki (à la Cognition/Devin's DeepWiki). One index powers both: the graph gives agents surgical structural context, the wiki gives humans browsable docs + conversational Q&A.

CodeGraph and DeepWiki were merged into this single self-contained app. The former codegraph app and the shared codeindex-core crate were both folded in and removed — the code-intelligence core now lives in this crate (src/{db,index,lang,model,parse,query}.rs).

• Parsing/index: tree-sitter — 17 languages: Rust, Python, JavaScript, TypeScript/TSX, Go, Java, C, C++, C#, Ruby, PHP, Scala, Bash, Julia, Haskell, OCaml → SQLite + FTS5 graph of symbols/calls/imports. Wiki pages live in the same DB at ~/.senclaw/space-apps-data/deepwiki/index.db.
• Live: a file watcher incrementally re-indexes on change (debounced).
• Grounding: every wiki page / answer is built from retrieved evidence with path:line citations — no hallucinated APIs.

MCP tools (deepwiki-mcp)

Index

Code graph

Wiki

Skills

• deepwiki-generate — plan from the outline, write each page from grounded context, save.
• deepwiki-ask — answer codebase questions, grounded in real path:line sources.
• deepwiki-explore — understand structure via the call graph (who calls what, outlines).
• deepwiki-review — impact/risk review of a change (callers, blast radius, tests at risk).

Sub-agents (personas)

Installed into the SenClaw virtual-agents dir on app install; dispatchable via run_persona / the dispatch DAG:

• wiki-author — plans and writes the wiki from grounded evidence.
• codebase-guide — answers questions about the codebase, grounded in source.
• code-explorer — navigates and explains a codebase via the call graph.
• impact-analyst — assesses the blast radius and risk of a change end-to-end.

REST API (used by the Web UI)

GET /api/status · GET /api/recents (previously-indexed roots) · POST /api/index {path} · GET /api/outline · GET /api/context?q=&depth= (alias /api/ask) · GET /api/pages · GET/POST/DELETE /api/page?slug= · GET /api/search?q= · GET /api/symbol?name= · GET /api/explore?q=&depth= · GET /api/file?path= · GET /api/files · GET /api/snippet?name=|path=&start=&end= · MCP at /api/mcp/sse + /api/mcp/message

Build & run

# Web UI — React + TypeScript + Ant Design + Vite (same stack as the email app)
cd web && npm install && npm run build   # → web/dist (served statically by the binary)
cd ..

cargo build -p deepwiki --release        # binary at target/release/deepwiki
PORT=4330 ./target/release/deepwiki       # serves Web UI + REST + MCP on :4330
# dev server with HMR: cd web && npm run dev

The header's repo field is an AutoComplete: once you've indexed folders, clicking it lists your previously-indexed roots (with file/symbol counts + relative time) for one-click re-indexing, and typing filters that list. History is kept in the indexed_roots table and served by GET /api/recents.

The Web UI (web/, React 19 + AntD 6, theme-synced with the SenClaw host via postMessage) has three tabs:

• Wiki (WikiView) — page tree + react-markdown, a STRUCTURE file tree (/api/files, expandable folders → files; click a file → outline + syntax-highlighted source), and a dual-mode query box:
• Tìm kiếm — returns grounded symbol evidence (/api/context).
• Hỏi AI (Devin-style) — POST /api/ask investigates the question multi-hop through the call graph (callers + callees, both directions, query::investigate), sends that subgraph + source excerpts to SenClaw's configured LLM (Space-App bridge llm.request), and renders: a cited Markdown answer, the model, a "Graph tổng quan — luồng điều tra" (OverviewGraph — the investigation subgraph laid out by depth), a "Xem luồng" button (full Graph tab on the focus symbol), and the evidence. Every Q&A is saved to history (ask_history table) — the Lịch sử panel lists past questions and reopens them (answer + graph) or deletes them (/api/ask-history, /api/ask-history/:id).

Hỏi AI requires: (1) the SenClaw daemon built with this repo (the bridge llm.request handler lives in the main crate — see src/gateway/ui_server/{space.rs,llm_config.rs}), and (2) an active LLM configured in SenClaw (Settings → Models). The app calls the daemon at SENCLAW_BASE_URL (injected when the daemon launches the app; defaults to http://127.0.0.1:18788). FTS uses Porter stemming so NL queries match identifiers (indexing → index_repo).

• Code (CodeView) — symbol search + call-graph/blast-radius explorer, with an inline syntax-highlighted source viewer (CodeBlock, via /api/snippet) and a "Graph" button.
• Graph (GraphView + OverviewGraph) — an interactive multi-hop call-graph (/api/investigate): columns by relative depth (callers N … focus … callees N). Each node shows name + kind badge + file:line so a coder reads real logic, not bare names. Controls: Cả hai/Callers/Callees and Sâu 1/2/3 (depth). Click a node to re-centre.

Path filtering + Settings. Indexing skips build artifacts by default (node_modules, target, dist, build, release, web_dist, *.min.js, *.map, …) and auto-detects minified/generated files, so the graph reflects real source (no flood of 1-char symbols from vendored bundles). The header "Loại trừ path" field adds custom globs; the ⚙️ Settings drawer (SettingsDrawer) edits the full default-excludes list (with "Khôi phục mặc định"), custom excludes, and the minified line-length threshold — all persisted in the settings meta JSON and applied on re-index (/api/settings, index::Settings/load_settings/save_settings). IndexReport.excluded counts skipped files.

LLM (Hỏi AI) uses SenClaw's Main model. The bridge chat_completion reads SenClaw's active (Main) LLM config — same model the daemon uses. The Settings drawer shows it live via GET /api/llm-info (e.g. deepseek-v4-flash), proxied from the daemon's /api/llm-config, so you can confirm a real model (not a mock) is wired. Requires the daemon built from this repo (bridge enabled) + an active model in Settings → Models.

The typed API client is web/src/api.ts.