Skip to content

Weekly research scan 2026-04-05 #13

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
max4c wants to merge 1 commit into from
Closed

Weekly research scan 2026-04-05 #13

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
246 changes: 246 additions & 0 deletions research/weekly-scan-2026-04-05.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,246 @@
# Weekly Research Scan — 2026-04-05

Repos monitored: **OpenOats**, **QMD**, **AppFlowy**, **Exo (mail-app)**
Period: March 29 – April 5, 2026

---

## 1. OpenOats (yazinsai/OpenOats)

> Native macOS (Swift/SwiftUI) meeting note-taker with on-device WhisperKit transcription, local knowledge-base search via embeddings, and LLM-powered suggestions. ~2.1k stars, MIT.

### Activity: Moderate (6 merged PRs, 3 open issues)

| PR / Commit | What shipped | Why it matters |
|-------------|-------------|----------------|
| #261 | **User-editable note templates** (`TemplateStore.swift`) — create, edit, and apply custom note templates | Extensible template system for structured output |
| #255 | **EventKit calendar integration** — auto-titles sessions from calendar events, pulls attendee lists | Contextual auto-metadata from system sources |
| #257 | **Move KB file I/O off `@MainActor`** — Knowledge Base reads now async | Critical perf fix for file-heavy architectures |
| #258 | **AirPods mic bleed fix** with output device picker | Audio UX polish |
| #260 | **Template picker in session view** — apply template before/after meeting | Template UX integration |
| #259 | Minor: settings label cleanup | — |

#### Active Issues

- #288: Transcription missing other person's audio — real user pain point
- #279: Speaker attribution issues with MS Teams
- #304 (Closed): Windows support request — shows demand beyond macOS

### Patterns Worth Adapting for Bugbook

| Pattern | Detail | Effort | Impact |
|---------|--------|--------|--------|
| **User-editable templates** | `TemplateStore.swift` — file-based template CRUD. Bugbook could offer templates for database schemas and page layouts. Users define a template as a markdown file with YAML frontmatter; `create` command accepts `--template <name>`. | **Low** | **High** |
| **Calendar-aware auto-metadata** | EventKit integration auto-populates title, attendees, date from calendar. Bugbook pages could auto-tag with calendar context (meeting name, project) when created during a calendar event. | Med | Med |
| **Async file I/O off MainActor** | PR #257 moved all Knowledge Base file reads off `@MainActor`. Bugbook's `RowStore` and `IndexManager` must follow the same pattern — any file I/O on the main actor will freeze the UI with large workspaces. | **Low** | **High** |
| **Meeting format spec** | Formal spec (`docs/meeting-format-spec.md`) with schema versioning (`schema: openoats/v1`), `x_` extension fields, and "parsers must ignore unknown fields" rule. Bugbook should formalize its row file format similarly. | Low | Med |
| **Progressive file enrichment** | Files go through raw → cleaned → AI-enriched stages in-place. Bugbook CLI agents could progressively add AI-generated summaries, tags, or backlinks to existing note bodies. | Low | Med |
| **Dataview-compatible inline fields** | Action items use `[owner:: Name] [due:: YYYY-MM-DD]` syntax for Obsidian Dataview compatibility. If Bugbook supports this convention, files become portable to/from Obsidian. | Low | Med |

---

## 2. QMD (tobi/qmd)

> On-device hybrid search engine for markdown files. BM25 (SQLite FTS5) + vector similarity (sqlite-vec) + LLM reranking — all local. ~17.2k stars, MIT. By Tobi Lütke.

### Activity: High (5 merged PRs, 7 open PRs, 6 open issues)

| PR | What shipped | Why it matters |
|----|-------------|----------------|
| #478 | **`rerank` parameter on MCP query tool** | Bugbook can now skip LLM reranking for fast keyword searches |
| #475 | **`handelize()` filename fix** — fixes mangled filenames with dots/timestamps | Fixes upstream filename resolution bugs |
| #453 | **Configurable `RERANK_CONTEXT_SIZE`** (default 2048→4096) | Better reranking for long documents |
| #458 | **Embedding reliability fixes** | Fewer silent failures during indexing |
| #456 | **`vec0 OR REPLACE` workaround** | sqlite-vec compatibility fix |

#### Notable Open PRs

| PR | Title | Relevance to Bugbook |
|----|-------|---------------------|
| **#503** | Cloud LLM support | Users without powerful local hardware could use cloud models for reranking |
| **#502** | Per-collection model config in `index.yml` | Different embedding models per Bugbook workspace |
| **#492** | Ebbinghaus memory decay with recall scoring | Spaced repetition — surface notes needing review. Extremely relevant for PKM |
| **#496** | Binary search for `findBestCutoff` and `isInsideCodeFence` | Chunking performance improvement |
| **#490** | Remote Ollama embeddings via `OLLAMA_EMBED_URL` | Alternative to local node-llama-cpp |

#### Notable Open Issues

| Issue | Title | Relevance |
|-------|-------|-----------|
| **#499** | Index should be part of `qmd://` link | ⚠️ Would change virtual path URI format — Bugbook's `normalizedRelativePath()` parses these |
| **#497** | `QMD_EMBED_MODEL` env var ignored → dimension mismatch | Users setting custom models may hit this |
| **#498** | sqlite-vec extension loading support | Platform compatibility |

### What Bugbook Should Do

**Upstream changes (update dependency, no Bugbook code changes):**
- FTS5 ranking fixes (#463, #462) — automatic improvement
- `qmd embed` reliability (#458, #456) — upstream fix
- `handelize()` filename fix (#475) — just update qmd

**Bugbook code changes needed:**

| Action | Detail | Effort | Impact |
|--------|--------|--------|--------|
| **Expose `--no-rerank` flag** | Pass `rerank: false` in `QmdBackend` for BM25-only searches via the `query` command. Skips expensive LLM reranking when users just want keyword search. | **Low** | **High** |
| **Pass `--chunk-strategy auto`** | Enable AST-aware chunking (merged last week in #449) for code files in workspaces. Add to `ensureCollection()` calls. | **Low** | Med |
| **Watch issue #499** | If `qmd://` URI format changes to include index, `normalizedRelativePath()` in `SearchCommand.swift` will break. No action yet, but monitor. | — | — |
| **Surface `rerank_context_size` as setting** | Expose `RERANK_CONTEXT_SIZE` env var in Bugbook settings for users with long documents. | Low | Low |

---

## 3. AppFlowy (AppFlowy-IO/AppFlowy)

> Open-source AI-powered collaborative workspace (Flutter + Rust), ~68.9k stars. Notion alternative emphasizing data privacy and local-first architecture.

### Activity: Moderate (v0.11.5 + v0.11.6 released)

**v0.11.6 (April 2, 2026):**
- **Document data corruption fix** — critical stability patch
- Block selection and select-all behavior improvements

**v0.11.5 (March 26, 2026):**
- AI Meeting Transcript — paste YouTube link → transcript
- **Database bulk edit optimization** — multi-row property editing
- Mobile: database embedded views support Feed and List views
- Fixes: shortcut conflicts, plus-menu insertion, filter evaluation

#### Tab Management — Major 2026 Focus (8+ issues updated this week)

| Issue | What | Timeline |
|-------|------|----------|
| #3101 | Navigation history (back/forward) | 2026 |
| #8350 | Support opening multiple windows | Q2 26 |
| #6279 | Change order of opened tabs | Q2 26 |
| #8545 | Right-click "open in new tab" | Q2 26 |
| #8561 | Missing keyboard shortcut for new tab | Q2 26 |
| #8477 | Go to doc section when switching to active tab | Q2 26 |

#### Database Issues Worth Watching

| Issue | What | Relevance |
|-------|------|-----------|
| **#5899** | Row records mentionable as pages | Bugbook's file-per-row architecture does this naturally |
| **#8623** | Kanban "No Status" column annoyance | UX problem Bugbook should solve preemptively |
| **#2418** | Deleted rows should go to Trash | Bugbook could implement soft-delete via a `_trash/` directory |
| **#8624** | Nested templates broken (databases inside templates shared, not copied) | Bugbook's directory-copy approach solves this |

### Patterns Worth Adapting for Bugbook

| Pattern | Detail | Effort | Impact |
|---------|--------|--------|--------|
| **Database bulk editing** | Select multiple rows → edit shared property (status, tags). In Bugbook, pipe selection through `bugbook batch` with a shared `--set` value. SwiftUI: multi-select + toolbar action. | **Low** | **High** |
| **Tab management / multi-window** | AppFlowy struggling with this in Flutter (10+ open issues). Bugbook on macOS can use native `WindowGroup` + `openWindow` for free. Implement now while it's a differentiator. | Med | High |
| **Soft-delete to Trash** | Issue #2418. Move deleted row files to `_trash/` with timestamp suffix. Recoverable via CLI (`bugbook restore <db> <row_id>`) or UI. | **Low** | Med |
| **Kanban empty-status handling** | Issue #8623. Make the "No Status" column hideable/collapsible. Validate required fields on card creation to prevent orphaned cards. | **Low** | Med |
| **Database row as mentionable page** | Issue #5899. Bugbook's file-per-row architecture means `[[row filename]]` wikilinks work naturally. Just need a link resolver in the markdown editor. | Med | High |
| **Feed/timeline view** | v0.11.2 added a Feed view type. For Bugbook: a chronological view sorted by `created_at` or `updated_at`, showing title + preview. Good for journals and activity streams. | Med | Med |

---

## 4. Exo (ankitvgupta/mail-app)

> AI-native email client built with Electron + TypeScript. Local-first with SQLite, deep Claude integration, agent system with tool use. Active development, pre-launch.

### Activity: Very High (20+ commits, 8 merged PRs)

| PR | What shipped | Why it matters |
|----|-------------|----------------|
| #49 | **Prompt injection prevention** — `<untrusted_email>` safety tags with loop-based stripping | Security pattern for AI processing of user content |
| #47 | **Prefetch pipeline concurrency fix** — reduced from 10→3 parallel tasks, added event-loop yields | Critical perf lesson for background AI processing |
| #45 | **Split view with custom filters** — configurable email splits with user-defined filter rules | Virtual views over a single data source |
| #43 | **Agent permission gate + audit log** — `permission-gate.ts`, `audit-log.ts` for tool access control | Safe agent automation pattern |
| #41 | **LLM cost tracking** — per-call cost tracking in `llm_calls` table with model-aware pricing | Essential for AI-powered apps |
| #39 | **Memory/learning system** — `draft-edit-learner.ts` extracts preferences from how users edit AI drafts | AI that improves from user behavior |
| #37 | **Dual command palette** — Cmd+K for commands, Cmd+J for agent | Power-user UX pattern |

### Architecture Deep Dive (Most Relevant Patterns)

#### Agent System
- `agent-coordinator.ts` → provider dispatch → `orchestrator.ts` lifecycle management
- `permission-gate.ts` gates tool access; `audit-log.ts` records all calls
- `AgentCommandPalette.tsx` (Cmd+J) for natural-language interaction
- Agent-to-agent delegation and external provider registration

#### Memory/Learning System
- `memory-context.ts` injects relevant memories into AI prompts
- `draft-edit-learner.ts` learns writing preferences from user edits
- `analysis-edit-learner.ts` learns priority preferences from overrides
- Configurable scoping (per-sender in email → per-project/tag in PKM)

#### Background Processing Pipeline
Multi-stage pipeline with lessons learned:
1. Parse/analyze new items
2. Enrich high-priority items
3. Auto-generate drafts
- Concurrency limits matter (10→3 was the fix)
- Event-loop yields prevent UI freezing
- Caching avoids duplicate expensive DB queries

### Patterns Worth Adapting for Bugbook

| Pattern | Detail | Effort | Impact |
|---------|--------|--------|--------|
| **Agent audit log** | Log all CLI agent actions (create/update/delete) with timestamp, caller, and parameters. Write to `.bugbook/agents/events.jsonl`. Bugbook already has the file — just ensure every mutation logs an event. | **Low** | **High** |
| **Dual command palette** | Cmd+K for navigation/commands, Cmd+J for agent chat. SwiftUI `CommandGroup` + sheet overlay. Separate entry points for structured commands vs. freeform AI. | Med | High |
| **Prompt injection prevention** | Wrap user-authored content in `<untrusted_content>` tags before passing to AI. Strip nested bypass attempts. Critical for AI features that process web clippings or imported content. | **Low** | Med |
| **Memory/learning from edits** | Track how users modify AI-generated content (summaries, tags). Over time, improve prompts based on patterns. Store in `.bugbook/agents/memories.json`. | High | High |
| **LLM cost tracking** | Centralized service logging model, tokens, cost per AI call. Store in `agents/events.jsonl` with `type: "llm_call"`. Essential for transparency. | **Low** | Med |
| **Background processing with concurrency control** | Limit parallel AI operations (3, not 10). Yield to event loop between batches. Directly applicable to Bugbook's index rebuilds and future AI enrichment pipeline. | Low | Med |
| **Virtual views / split filters** | User-defined filter presets over the same dataset — Bugbook's `ViewConfig` already supports this. The "split config editor" UX pattern is worth studying for a view builder. | Low | Med |

---

## Top 3 This Week

The three highest-impact, lowest-effort items to act on first. All are changes to Bugbook's own codebase.

### 1. Expose `--no-rerank` for Fast Keyword Search (from QMD #478)
**Effort: Low | Impact: High**
QMD's MCP query tool now accepts a `rerank` parameter. Bugbook's `QmdBackend` in `SearchCommand.swift` should pass `rerank: false` for BM25-only search modes. This skips the expensive LLM reranking step entirely, making keyword searches near-instant. One-line change in the query command builder.

### 2. Database Bulk Editing in SwiftUI (from AppFlowy v0.11.5)
**Effort: Low | Impact: High**
AppFlowy shipped database bulk edit optimization in v0.11.5 — select multiple rows, change a shared property. Bugbook already has `bugbook batch` on the CLI side. The SwiftUI gap is: multi-select in table/kanban views → toolbar action → `MutationEngine.execute()` with batch operations. The engine supports it; just wire up the UI. Multi-select is native in SwiftUI `List` and `Table`.

### 3. Agent Audit Log for All Mutations (from Exo #43)
**Effort: Low | Impact: High**
Exo's `audit-log.ts` records every agent tool call with timestamp, caller, and parameters. Bugbook already has `.bugbook/agents/events.jsonl` — but mutations via the CLI don't always log to it. Ensure every `MutationEngine.execute()` call from `BugbookCLI` writes a structured event to `events.jsonl` with the operation details. This gives full auditability for agent-driven changes and powers the Agent Hub dashboard.

---

## Proposed Tickets

### Ticket 1: Add `--no-rerank` flag to search commands

**Title:** Skip LLM reranking for keyword-only searches

**Description:** Bugbook's `QmdBackend` in `SearchCommand.swift` always delegates to qmd without specifying the `rerank` parameter. Since qmd #478 merged, the `query` MCP tool accepts `rerank: false` to skip LLM reranking entirely. Add a `--no-rerank` flag to `bugbook search` that passes this through. Default behavior: rerank enabled for hybrid/vector searches, disabled for BM25-only (`--mode keyword`). This makes keyword searches near-instant by avoiding the LLM reranking step.

**Effort:** Low (one flag + conditional parameter in `QmdBackend.buildQuery()`)

**Source:** [tobi/qmd#478](https://github.com/tobi/qmd/pull/478) — Add `rerank` parameter to MCP query tool

---

### Ticket 2: Wire up multi-select bulk editing in database views

**Title:** Add bulk property editing to table and kanban views

**Description:** AppFlowy v0.11.5 shipped database bulk editing — selecting multiple rows and changing a shared property in one action. Bugbook's `MutationEngine` already supports batch operations, and the CLI has `bugbook batch`. The missing piece is the SwiftUI UI: enable multi-select in `DatabaseTableView` and `DatabaseKanbanView`, show a floating toolbar when >1 row is selected, and wire the "Set Property" action to `MutationEngine.execute()` with batch `updateRow` operations. Use SwiftUI's native `Table` selection and `List` selection APIs.

**Effort:** Low (engine already supports batch; this is UI wiring)

**Source:** [AppFlowy v0.11.5 release notes](https://github.com/AppFlowy-IO/AppFlowy/releases/tag/0.11.5) — Database bulk edit optimization

---

### Ticket 3: Ensure all CLI mutations emit agent audit events

**Title:** Log all CLI mutations to `events.jsonl` for agent auditability

**Description:** Exo's agent system logs every tool call with timestamp, caller, and parameters to an audit log. Bugbook has `.bugbook/agents/events.jsonl` but CLI mutations (create, update, delete, batch) don't consistently write events to it. Add an `AuditLogger` that `MutationEngine` calls after every successful mutation, writing a JSON line with `{type: "mutation", timestamp, command, database, operations, caller}`. The `caller` field distinguishes "cli", "app", or a specific agent run ID. This powers the Agent Hub dashboard and gives users a complete audit trail of all automated changes.

**Effort:** Low (structured append to an existing JSONL file after each mutation)

**Source:** [ankitvgupta/mail-app#43](https://github.com/ankitvgupta/mail-app/pull/43) — Agent permission gate + audit log
Loading