Let MCP client tabs wrap to their full height and give each client a 44px touch target, keeping the Codex setup panel readable on mobile.
Clarified project verification through the active Codex /mcp view; reserve codex mcp list for checking user-wide configurations.
Keep the setup table of contents focused on visible shared sections when switching client tabs.
v0.15.2
Architecture Diagrams and Codex
Added system, sequence and user diagrams to both websites, generated from a shared Mermaid source in docs/architecture.md and checked in CI.
Added project-scoped Codex MCP setup via hnindex init --mcp codex: merge valid TOML while retaining other server settings and binding the workspace root.
Corrected the Codex skill installation path to .agents/skills/use-vibe-hnindex and documented project trust, verification and TOML setup.
Added Codex to the homepage installer, integrations, installation guide and MCP setup guide.
Lazy-load diagram rendering with accessible titles, keyboard scrolling, loading feedback and a source fallback.
v0.15.1
Website and Documentation Refresh
Redesigned the main website and docs with a shared visual identity, responsive navigation, keyboard focus states and English/Vietnamese interface controls.
Generate both sites’ version and full release history from package.json and docs/changelog.md; check committed data in CI and regenerate before site builds.
Added a docs changelog and code graph guide; refreshed installation and quick start around offline workspace context and source locations.
Synchronized server, CLI, MCP and Claude marketplace metadata for the documentation patch release.
v0.15.0
Agent Workspace and Code Locations
Added local workspace_context: explicit project/path, configured root, MCP client roots or indexed cwd; require selection for multiple roots/project names. Read live bounded purpose documents, Git state and index readiness.
Persist declared tasks in SQLite per project/session with clear and deletion cleanup. Git changes are evidence, never inferred intent.
Added locate_code: file/symbol/keyword locations without embeddings, explicit hybrid mode, ambiguity, excerpts and per-file freshness including relationship endpoints.
Added MCP startup instructions and knowledge://workspace; updated prompts/rules/skills with filesystem fallback for missing/stale/excluded code.
CLI binds project installs to their root and adds --project-root for global configurations; fixed absolute --cwd resolution.
Requires Node 22+; upgraded better-sqlite3 to 13.0.3 with bundled N-API binaries to avoid native statement cleanup crashes on Node 24.
Added offline index_code_graph, find_references, callers and bounded graph_context tools using SQLite and TypeScript/JavaScript AST analysis.
Resolve indexed-file imports, alias/re-export chains, lexical symbols and typed methods using project compiler options; record source evidence and unresolved dynamic/external calls.
Update graph snapshots during full/file/watch indexing, re-resolve consumers after edits/deletions, hide stale graphs after analysis failure, and enrich smart_context with symbol relationships.
Added AST declaration chunks and embedding-only file/line context; versioned chunk profiles require one full re-index after upgrading.
Added native Voyage/custom HTTP reranking, strict score validation and full-body timeouts. Preserve RRF on fallback, rerank a wider pool before trim and cache final ranking with query options.
Added document/query retrieval semantics, ordered responses, strict vector validation, full-request timeouts, bounded 429/5xx retries and cached cloud health probes.
Isolated vector collections by provider/model/dimensions/endpoint; tracked embedding profiles and full re-indexing when switching providers or recovering an incomplete index.
Unified file hashes across indexing paths to avoid paying for unchanged files; automatic non-Ollama worker concurrency is capped at four.
CLI supports provider/model/base URL/API key options and clears stale generic settings when switching provider.
Updated configuration docs, website changelog and package/marketplace metadata to v0.13.0.
v0.12.1
🚫 Tool Priority Enforcement
🚫 Hardened agent_rules_stub — Ultra-compact BANNED/INSTEAD table at top of generated rules. Shorter, punchier, harder to ignore.
📢 priority_prompt tool — One-shot tool to dump enforcement rules mid-conversation. Call when AI uses grep instead of MCP.
💬 MCP prompts capability — tool-priority prompt for AI clients to auto-fetch enforcement rules.
📝 SKILL.md + CLAUDE.md — Stronger BANNED language at top, compact 4-line table.
🧠 Auto-Track — Tự động log mọi tool call (search, smart_context, code_session) vào SQLite, không cần gọi tool thủ công. AI restart session vẫn có ngay context cũ.
💬 chat_context tool — Save/load/clear/ingest conversation messages. Load hỗ trợ semantic search qua Qdrant (semantic_query param) — chỉ trả context thực sự liên quan, tiết kiệm token.
👁️ Watch Auto-Resume — File watchers tự động resume khi MCP server khởi động lại (IDE restart, AI agent restart, connection drop). Không còn lo mất watch mode. Env WATCH_AUTO_RESUME (default true).
🔴 MCP-First Rules — agent_rules_stub giờ sinh rules aggressive: cấm grep/cat/read/Glob, bắt buộc dùng MCP tools. Có bảng so sánh "Instead of → Use MCP".
📝 CLAUDE.md — Thêm MANDATORY section ở đầu với bảng thay thế tool rõ ràng.
📝 SKILL.md + init-skill — Sync strong language về tool priority, đảm bảo AI agent nào cũng thấy.
📝 Init-Skill Sync — Bản inline skill trong CLI hnindex init-skill giờ có đầy đủ hướng dẫn Code Agent (code_session + code_apply). Trước đây bản này thiếu → người dùng init skill không thấy Code Agent.
📄 Smart Context content limit — Tăng từ 2500 → 25000 ký tự (gấp 10×). Thêm env SMART_CONTEXT_MAX_FILE_CHARS, set 0 = unlimited. File 500+ dòng giờ đọc gần như toàn bộ.
🔧 CLI defaults — hnindex init giờ include CODE_AGENT_ENABLED, CODE_AGENT_SCOPE, SMART_CONTEXT_MAX_FILE_CHARS trong config mặc định.
📝 SKILL.md — Hướng dẫn Code Agent chi tiết cho AI agents: params, output format, when to use.