CHANGELOG.md

Changelog

v1.6.0 — Onboarding + Identity (2026-04-10)

Added

• Interactive onboarding (engraph init) — polished CLI with welcome banner, vault scan checkmarks, identity prompts via dialoguer, progress bars, actionable next steps
• Agent onboarding — engraph init --detect --json for vault inspection, --json for non-interactive apply. Two-phase detect → apply flow for AI agents.
• identity MCP tool + CLI + HTTP — returns compact L0/L1 identity block (~170 tokens) for AI session context
• setup MCP tool + HTTP — first-time setup from inside an MCP session (detect/apply modes)
• identity_facts table — SQLite storage for L0 (static identity) and L1 (dynamic context) facts
• L1 auto-extraction — active projects, key people, current focus, OOO status, blocking items extracted during engraph index
• engraph identity --refresh — re-extract L1 facts without full reindex
• [identity] config section — name, role, vault_purpose in config.toml
• [memory] config section — feature flags for identity/timeline/mining

Changed

• MCP tools: 23 → 25
• HTTP endpoints: 24 → 26
• Dependencies: +dialoguer 0.12, +console 0.16, +regex 1

v1.5.5 — Housekeeping (2026-04-10)

Added

• auto_link parameter on create — set to false to skip automatic wikilink resolution. Applies to MCP, HTTP, and CLI. Discovered links still appear as suggestions in the response.
• reindex_file MCP tool + HTTP endpoint — re-indexes a single file after external edits. Reads from disk, re-embeds chunks, rebuilds edges. Available as MCP tool, POST /api/reindex-file, and OpenAPI operation.

Changed

• rmcp bumped from 1.2.0 to 1.4.0 — host validation, non-Send handler support, transport fixes. Does not yet fix #20 (protocol 2025-11-25 needed for Claude Desktop Cowork/Code modes — blocked upstream on modelcontextprotocol/rust-sdk#800).
• MCP tools: 22 → 23
• HTTP endpoints: 23 → 24
• OpenAPI version: 1.5.0 → 1.5.5

v1.5.0 — ChatGPT Actions (2026-03-26)

Added

• OpenAPI 3.1.0 spec (openapi.rs) — hand-written spec for all 23 endpoints, served at GET /openapi.json
• ChatGPT plugin manifest — served at GET /.well-known/ai-plugin.json
• --setup-chatgpt CLI helper — interactive setup: enables HTTP, creates API key, configures CORS, prompts for public URL
• Plugin config — [http.plugin] section for name, description, contact_email, public_url

Changed

• Module count: 25 → 26
• Test count: 417 → 426
• /openapi.json and /.well-known/ai-plugin.json routes require no authentication

v1.4.0 — PARA Migration (2026-03-26)

Added

• PARA migration engine (migrate.rs) — AI-assisted vault restructuring into Projects/Areas/Resources/Archive
• Heuristic classification — priority-ordered rules detect Projects (tasks, active status), Areas (recurring topics), Resources (people, reference), Archive (done, inactive)
• Preview-then-apply workflow — generates markdown + JSON preview for review before moving files
• Migration rollback — engraph migrate para --undo reverses the last migration
• 3 new MCP tools — migrate_preview, migrate_apply, migrate_undo
• 3 new HTTP endpoints — POST /api/migrate/preview, /apply, /undo
• Migration log — SQLite table tracks all moves for rollback support

Changed

• Module count: 24 → 25
• MCP tools: 19 → 22
• HTTP endpoints: 20 → 23
• Test count: 385 → 417

v1.3.0 — HTTP/REST Transport (2026-03-26)

Added

• HTTP REST API (http.rs) — axum-based HTTP server alongside MCP, enabled via engraph serve --http
• 20 REST endpoints mirroring all 19 MCP tools + update-metadata
• API key authentication — eg_ prefixed keys with read/write permission levels
• Rate limiting — configurable per-key token bucket (requests/minute)
• CORS — configurable allowed origins for web-based agents
• Graceful shutdown — CancellationToken coordinates MCP + HTTP + watcher exit
• API key management CLI — engraph configure --add-api-key/--list-api-keys/--revoke-api-key
• --no-auth mode — local development without API keys (127.0.0.1 only)

Changed

• engraph serve gains --http, --port, --host, --no-auth flags
• Module count: 23 → 24
• Test count: 361 → 385
• New dependencies: axum, tower-http, tower, rand, tokio-util

v1.2.0 — Temporal Search (2026-03-26)

Added

• Temporal search lane (temporal.rs) — 5th RRF lane for time-aware queries
• Date extraction — from frontmatter date: field or YYYY-MM-DD filename pattern
• Heuristic date parsing — "today", "yesterday", "last week", "this month", "recent", month names, ISO dates, date ranges
• LLM date extraction — orchestrator detects temporal intent and extracts date ranges from natural language
• Temporal scoring — smooth decay function for files near but outside the target date range
• Temporal candidate injection — date-matched files enter candidate pool as graph seeds
• Confidence % display — search results show normalized confidence (0-100%) instead of raw RRF scores
• Date coverage stats — engraph status shows how many files have extractable dates

Changed

• QueryIntent gains Temporal variant with custom lane weights (temporal: 1.5)
• OrchestrationResult gains date_range field (backward-compatible serde)
• LaneWeights gains temporal field (0.0 for non-temporal intents)
• insert_file signature extended with note_date parameter
• Module count: 22 → 23
• Test count: 318 → 361

[1.1.0] - 2026-03-26 — Complete Vault Gateway

Added

• Section parser (markdown.rs) — heading detection, section extraction, frontmatter splitting
• Obsidian CLI wrapper (obsidian.rs) — process detection, circuit breaker (Closed/Degraded/Open), async CLI delegation
• Vault health (health.rs) — orphan detection, broken link detection, stale notes, tag hygiene
• Section-level editing — edit_note() with replace/prepend/append modes targeting specific headings
• Note rewriting — rewrite_note() with frontmatter preservation
• Frontmatter mutations — edit_frontmatter() with granular set/remove/add_tag/remove_tag/add_alias/remove_alias ops
• Hard delete — delete_note() with soft (archive) and hard (permanent) modes
• Section reading — read_section() in context engine for targeted note section access
• Enhanced file resolution — fuzzy Levenshtein matching as final fallback in resolve_file()
• 6 new MCP tools — read_section, health, edit, rewrite, edit_frontmatter, delete
• CLI events table — audit log for CLI operations
• Watcher coordination — recent_writes map prevents double re-indexing of MCP-written files
• Content-based role detection — detect people/daily/archive folders by content patterns, not just names
• Enhanced onboarding — engraph init detects Obsidian CLI + AI agents, engraph configure has new flags
• Config sections — [obsidian] and [agents] in config.toml

Changed

• Module count: 19 → 22
• MCP tools: 13 → 19
• Test count: 270 → 318

[1.0.2] - 2026-03-26

Fixed

• Person search uses FTS — context who now finds person notes via full-text search instead of exact filename matching. Handles hyphens, underscores, any vault structure. Prefers People folder → person tag → fuzzy filename.
• llama.cpp logs suppressed — backend.void_logs() silences Metal/model loading output. Clean terminal output by default.
• Basename resolution — find_file_by_basename normalizes hyphens/underscores/spaces for cross-format matching.

Changed

• Re-recorded demo GIF with v1.0.2 brew binary (clean output, no 2>/dev/null workarounds)

[1.0.1] - 2026-03-26

Changed

• Inference backend switched from candle to llama.cpp — via llama-cpp-2 Rust bindings. Gets full Metal GPU acceleration on macOS (88 files indexed in 70s vs 37+ minutes on CPU with candle). Same backend as qmd.
• Default embedding model produces 256-dim vectors via embeddinggemma-300M (Matryoshka truncation)
• BERT GGUF architecture support added alongside Gemma (future model flexibility)
• Progress bar during indexing via indicatif (was silent for minutes)
• CI workflow installs CMake on Ubuntu (required for llama.cpp build)

Fixed

• Prompt format applied during embedding — embed_one uses search_query prefix, embed_batch uses search_document prefix. Without this, embeddinggemma operated in wrong symmetric mode.
• GGUF tokenizer fallback — added shimmytok crate to extract tokenizer from GGUF metadata when tokenizer.json is unavailable (Google Gemma repos are gated)
• LlamaBackend singleton — global OnceLock prevents double-initialization crash when loading multiple models
• Orchestrator/reranker use built-in tokenizer — llama.cpp reads tokenizer from GGUF metadata, no external tokenizer.json needed
• Dimension migration clears FTS — reset_for_reindex now also clears chunks_fts to prevent duplicate entries
• LLM cache wired into search — search_with_intelligence checks/populates llm_cache table
• MCP server wires intelligence — search handler passes orchestrator + reranker via SearchConfig
• CLI search wires intelligence — run_search loads models when intelligence enabled
• Qwen3 GGUF filename — fixed case sensitivity (was 404)
• Embedding batch params — n_ubatch >= n_tokens assertion, use encode() not decode(), AddBos::Never (PromptFormat adds <bos>)

Removed

• candle-core, candle-nn, candle-transformers dependencies (replaced by llama-cpp-2)

[1.0.0] - 2026-03-25

Intelligence release. Replaced ONNX with GGUF model inference, added LLM-powered search intelligence. Immediately followed by v1.0.1 which switched the inference backend from candle to llama.cpp for Metal GPU support.

Added

• GGUF model inference — replaced ONNX (ort) with GGUF quantized models for all ML inference
• Research orchestrator — LLM-based query classification (exact/conceptual/relationship/exploratory) with adaptive lane weights. Single LLM call returns intent + 2-4 query expansions.
• Cross-encoder reranker — 4th RRF lane using Qwen3-Reranker for relevance scoring. Two-pass fusion: 3-lane retrieval → reranker scores top 30 → 4-lane RRF.
• Query expansion — each search runs multiple expanded queries through all retrieval lanes, merged via deduplication.
• Heuristic orchestrator — fast-path intent classification via pattern matching (docids, ticket IDs, "who" queries) when intelligence is disabled. Zero latency.
• Intelligence onboarding — opt-in prompt during engraph init and first engraph index. Downloads ~1.3GB of optional models.
• engraph configure command — --enable-intelligence, --disable-intelligence, --model embed|rerank|expand <uri> for model overrides.
• Dimension migration — auto-detects embedding dimension changes and triggers re-index.
• LLM result cache — SQLite cache for orchestrator results (keyed by query SHA256).
• Model override support — configurable embedding, reranker, and expansion model URIs for multilingual support.

Changed

• Embedding model: all-MiniLM-L6-v2 (ONNX, 384-dim, 23MB) → embeddinggemma-300M (GGUF, 256-dim, ~300MB)
• Search pipeline: hardcoded 3-lane weights → adaptive per-query-intent weights
• --explain output now shows query intent and 4-lane breakdown (semantic, FTS, graph, rerank)
• status command shows intelligence enabled/disabled state

Removed

• ort (ONNX Runtime) dependency
• ndarray dependency
• src/embedder.rs and src/model.rs (replaced by src/llm.rs)
• ModelBackend trait (replaced by EmbedModel)

[0.7.0] - 2026-03-25

Added

• File watcher — engraph serve now watches the vault for changes and re-indexes automatically (2s debounce)
• Placement correction learning — detects when users move notes from suggested folders, updates centroids
• Fuzzy link matching — sliding window Levenshtein matching (0.92 threshold) during note creation
• First-name matching — matches "Steve" to [[Steve Barbera]] for People folder notes (suggestion-only)
• created_by column and filter — track note origin, filter with engraph context list --created-by
• placement_corrections table for observability
• link_skiplist table schema (reserved for future use)

Changed

• Centroid updates use true online mean (was EMA 0.9/0.1)
• Indexer refactored: index_file, remove_file, rename_file extracted as public functions
• Bulk indexing uses batched transactions for performance
• run_index_shared variant accepts external store/embedder references

Fixed

• Content hash consistency between diff_vault and index_file (BOM handling)

[0.6.0] - 2026-03-25

Added

• Write pipeline — create, append, update_metadata, move, archive, unarchive notes
• sqlite-vec replaces HNSW for vector search (single SQLite database)
• Tag registry with fuzzy Levenshtein resolution
• Link discovery — exact basename and alias matching during note creation
• Folder placement — type rules, semantic centroids, inbox fallback
• Archive/unarchive — soft delete with metadata preservation
• 6 new MCP write tools (13 total)

Changed

• All vectors stored in SQLite vec0 virtual table (was HNSW + separate files)
• Atomic writes via temp file + rename for crash safety
• Mtime-based conflict detection for concurrent edits

[0.5.0] - 2026-03-24

Added

• MCP server — engraph serve starts stdio MCP server via rmcp SDK
• 7 read-only MCP tools: search, read, list, vault_map, who, project, context

[0.4.0] - 2026-03-24

Added

• Context engine — 6 functions: read, list, vault_map, who, project, topic
• Token-budgeted context bundles for AI agents
• Person and project context assembly from graph + search

[0.3.0] - 2026-03-24

Added

• Vault graph — bidirectional wikilink + mention edges built during indexing
• Graph search agent — 3rd RRF lane with 1-2 hop expansion
• People detection from configured People folder

[0.2.0] - 2026-03-24

Added

• Hybrid search — semantic (embeddings) + keyword (FTS5 BM25) fused via RRF
• Smart chunking with break-point scoring algorithm
• Docid system (6-char hex file IDs)
• Vault profiles with auto-detection (engraph init)
• Pluggable model layer (ModelBackend trait)
• --explain flag for per-lane score breakdown

[0.1.0] - 2026-03-19

Added

• Initial release
• ONNX embedding model (all-MiniLM-L6-v2, 384-dim)
• SQLite metadata storage
• Incremental indexing
• .gitignore-aware vault walking