Skip to content

Roadmap

Complete

Storage & Data Model

  • Entry data model with structured metadata and type-specific extensions
  • DistilleryStore protocol — async storage abstraction enabling backend migration
  • DuckDB backend with VSS extension and HNSW index (cosine similarity)
  • Configurable embedding providers (Jina v3 default, OpenAI adapter)
  • Embedding model lock via _meta table — prevents mixed-model corruption
  • MCP server with 16 tools over stdio and streamable-HTTP (consolidated from 20 in staging/api-hardening)
  • distillery.yaml config system with validation

Core Skills

  • /distill — session knowledge capture with duplicate detection
  • /recall — semantic search with provenance display
  • /pour — multi-pass retrieval + structured synthesis with citations
  • /bookmark — URL fetch, auto-summarize, store with dedup check
  • /minutes — meeting notes with --update (append) and --list modes
  • Shared CONVENTIONS.md — author/project identification, error handling patterns

Classification Pipeline

  • ClassificationEngine — LLM prompt-based type assignment with confidence scoring
  • DeduplicationChecker — skip/merge/link/create at configurable thresholds
  • /classify skill — classify by ID, batch inbox, review queue triage

Quality & Observability

  • Implicit retrieval feedback + quality metrics (folded into distillery_status)
  • Stale entry detection — distillery_list(stale_days=…) (formerly the distillery_stale tool)
  • Conflict detection — distillery_find_similar(conflict_check=true)
  • Usage metrics dashboard — distillery_status tool (replaces former distillery_metrics)

Infrastructure

  • FastMCP 2.x/3.x with @server.tool decorators
  • Hierarchical tag namespace with validation; tag inventory via distillery_list(group_by="tag") (formerly distillery_tag_tree)
  • 12 entry types including person, project, digest, github, feed
  • Entry-type schema discovery via MCP resource distillery://schemas/entry-types (replaces distillery_type_schemas)

Team Access

  • HTTP transport — distillery-mcp --transport http
  • GitHub OAuth — team authentication via FastMCP GitHubProvider
  • Prefect Horizon deployment (MotherDuck)
  • Fly.io deployment with persistent DuckDB on volume
  • Namespace taxonomy — hierarchical, validated tag system

Ambient Intelligence

  • /radar — interest-driven feed digest with AI source suggestions
  • /watch — add/remove/list monitored feed sources
  • /tune — adjust relevance thresholds and trust weights
  • Feed polling architecture — FeedPoller with configurable intervals
  • Source adapters — GitHub events (REST API) and RSS/Atom
  • Relevance scoring pipeline — embedding-based cosine similarity
  • Interest extractor — mines entries for tags, domains, repos, expertise
  • Auto-tagging — source tags (source/github/owner/repo, source/reddit/sub) and topic tags from KB vocabulary
  • distillery retag CLI — backfill tags on existing feed entries
  • Hybrid BM25 + vector search — DuckDB FTS extension with Reciprocal Rank Fusion (RRF)
  • Recency decay — configurable time-weighted scoring (90-day window, 0.5 min weight)
  • Graceful degradation — falls back to vector-only if FTS extension unavailable

Team Skills

  • /digest — team activity summaries over configurable time windows
  • /gh-sync — sync GitHub issues/PRs into the knowledge base as searchable entries
  • /investigate — deep context builder with 4-phase retrieval and relationship traversal
  • /briefing — knowledge dashboard with solo mode (5 sections) and team mode (8 sections)

Entry Relations & Corrections

  • entry_relations table with backfill migration
  • distillery_correct tool for structured corrections
  • distillery_relations tool for managing entry links

New Entry Fields

  • expires_at — time-limited entries with UTC normalization
  • verification — orthogonal quality tracking (Unverified, Testing, Verified)
  • session_id — first-class field for session-scoped entries
  • Extended EntrySource — added inference, documentation, external provenance values

Session Hooks

  • Hook dispatcher script (distillery-hooks.sh) — routes UserPromptSubmit, SessionStart, PreCompact
  • Memory nudge — periodic reminder to /distill every 30 prompts
  • SessionStart briefing — automatic context injection via HTTP MCP
  • Scope-aware /setup hook configuration — detects plugin install scope (user/project)

Onboarding

  • /setup skill — MCP connectivity wizard, auto-poll configuration, session hook setup
  • uvx-first setup — uvx distillery-mcp as recommended first-time path

API Hardening (staging/api-hardening → released)

  • API consolidation: 20 → 16 tools. Removed distillery_aggregate, distillery_stale, distillery_tag_tree, distillery_metrics, distillery_interests, distillery_type_schemas, distillery_poll, distillery_rescore. Functionality folded into distillery_list, distillery_status, distillery_configure, REST /api/maintenance, and the distillery://schemas/entry-types resource.
  • #244 — Bulk ingest pipeline: new distillery_store_batch tool; /gh-sync runs as a server-side background job tracked by distillery_sync_status
  • #245 — Hardened MCP interface: tool descriptions, structured error codes, schema validation, INVALID_PARAMS suggestions
  • #232 — distillery_store enum includes github
  • #238 — distillery_store accepts output_mode="summary" to skip dedup/conflict checks
  • #241 — label→tag sanitiser handles underscored labels
  • #240 — /gh-sync passes valid output_mode
  • #317 — distillery_list / distillery_search exclude archived entries by default
  • #311 — distillery_list default output_mode="summary"
  • #346, #347, #349 — DuckDB WAL recovery and FTS replay hardening
  • #351 — Embedding budget default raised to unlimited; provider 429s surface to caller

Planned

P0 — Follow-up

  • #112 — Security Review Follow-Up

P0 — Quality & Bugfixing

PRs go directly to main.

  • #230 — DuckDB WAL corruption on unclean shutdown
  • #236 — RateLimitMiddleware defaults starve local-client bursts
  • #221 — FeedPoller poll cycle exceeds 5 minutes
  • #169 — distillery retag produces no output
  • #235 — Plugin auto-registers hosted demo MCP

P0 — Memory Benchmarking

  • #233 — LongMemEval retrieval benchmark

P1 — Near-term Features

  • #199 — distillery_extract for PreCompact summarisation
  • #237 — Retrieval-hygiene conventions docs
  • #212 — Slim down container image
  • #163 — Relevance-sorted feed queries for /radar
  • #152 — /whois skill
  • #151 — /process skill
  • #149 — Access control (visibility flag)

Deferred

  • #147, #142, #141, #140, #138, #158 — Graph analysis arc (NetworkX, hidden connections, epiphany generation)
  • #167 — Slack conversation adapter
  • #101 — Browser extension
  • #93 — Public knowledge spaces for OSS projects
  • #81 — Tauri desktop frontend
  • LangGraph evaluation for complex skill orchestration
  • Multi-team support and cross-team knowledge sharing
  • Re-embedding migration tooling

Technology Stack

Layer Current Planned
Interface Claude Code skills Same
Transport stdio + streamable-HTTP Same
Auth GitHub OAuth (FastMCP) + multi-team RBAC
Storage DuckDB + VSS + FTS / MotherDuck Same
Search Hybrid BM25 + vector (RRF) + score normalization
Embeddings Jina v3 / OpenAI Same
Language Python 3.11+ Same
Hosting Local / Fly.io / Prefect Horizon Same