Quill v0.9.1 — Research Platform Export + Reproducibility

What's New

Algorithm Arena Enhancements

• Benchmark Engine: Real model evaluation with pairwise comparisons and statistical significance testing
• Paper Export: Generate publication-ready LaTeX papers, Markdown reports, and CSV data
• Reproducibility: Experiment versioning with full configuration snapshots (git commit, platform, env)
• Model Registry: Register and manage benchmark models with cost tracking

Frontend

• New Export tab with LaTeX/Markdown/CSV download
• Reproducibility information panel
• Enhanced model comparison dashboard

Technical

• All type checks pass (backend + frontend)
• Desktop build verified (macOS aarch64)
• 3 new backend modules, 1120 lines of new code

Installation

Download the DMG below and install Quill Desktop, or clone the repo and run make dev.

Quill v0.9.0 — Algorithm Arena + Research Platform

Algorithm Arena & Research Platform

New Features

• Algorithm Arena: Compare models, configurations, and thinking levels side-by-side
• Model Selection: Select multiple models to benchmark against each other
• Configuration Management: Adjust temperature, thinking level, and max tokens per model
• Benchmark Tasks: Run standardized tasks across all selected models
• Statistical Comparison: View pass rates, latency, token usage, and cost comparisons
• Visualization: Bar charts, line charts, stacked bars, and confidence intervals
• Welcome Page: New default landing page with quick actions and feature overview
• Redesigned UI: Fresh color scheme (deep navy + amber gold), improved sidebar, smooth animations

Backend

• Experiment registry with 8 swappable algorithm modules
• Dataset management with versioning and integrity checks
• Experiment persistence and run tracking
• Bridge between experiments and evaluation framework
• Statistical significance testing (z-test, Cohen's d)

Frontend

• Algorithm Arena dashboard with model comparison
• Real-time benchmark execution with progress tracking
• Interactive charts for result visualization
• Responsive design with dark/light mode support

Desktop

• Updated Tauri configuration with optimized resource bundling
• Fixed symlink issues in frontend bundle preparation

Quill Desktop v0.8.0 — ZCode Sync Round 5

Quill v0.8.0 — ZCode Sync Round 5

Systematic evaluation of 11 leading harness frameworks (ZCode v3.14.3 full-source — no newer release as of 2026-09-28; OpenWork v0.18.54, DeepSeek Harness dsh-v0.1.7, OpenClaw 2026.9.6, Hermes Agent v0.21.5, Kimi Code 2.1.1, OpenAI Codex 0.159-alpha, awesome-harness-engineering, DeerFlow 2.1.0, Agent Framework 1.19, CrewAI 1.15.22). Nine new feature systems — six from ZCode's still-unported feature set, three from the reference frameworks:

Feature	Ported from	Module
Dynamic Workflows — author orchestration as plain TypeScript (top-level await, agent()/ask()/run()/log()/sleep()/escalate()); compiler validation with a string-literal gate on run(); sandboxed execution; append-only journal; amend/resume with warm cache so finished steps are never re-paid	ZCode (flagship)	workflows/dynamic/
Protocol Versioning — /api/protocol/hello handshake with capability negotiation; X-Quill-Protocol-Version (older = deprecation header, newer = fail fast); compat-field tolerance, method-not-found degradation, payload schema versioning	ZCode Protocol V4	server/protocol_version.ts
Conversation Sharing — sanitized (secrets redacted, tool internals stripped), integrity-hashed read-only share links at /share/<id>; three access modes; import shared conversations as new threads — self-hosted, no cloud dependency	ZCode	runtime/sharing.ts
Automation → IM Delivery — scheduled-task outcomes pushed to Slack / Feishu / DingTalk / Telegram / generic webhooks (implements what ZCode's bot_delivery_target column promises but never delivers)	ZCode + Quill	scheduling/automation_delivery.ts
Claude Code Migration — import Claude Code conversations (~/.claude/projects JSONL) as Quill threads with tolerant parsing and origin tagging	ZCode	migrations/claude_code.ts
@-Mention Pickers — typing @ in the input opens a grouped Agents + Skills autocomplete with full keyboard navigation	ZCode	frontend/mention-picker.tsx
LLM Overlay — scoped role→model routing with CrewAI's exact settings-merge semantics (generation params copy, credentials only same-provider)	CrewAI 1.15.22	models/llm_overlay.ts
Delegation Acceptance Criteria — deterministic harness-verified checks on subagent results (file: exists, file_written:, tests_passed: anchored to recorded runs — never self-reported)	DeerFlow 2.1.0	multi_agent/acceptance_checks.ts
Goal Judge — four-verdict continuation state machine (done / blocked / wait / continue) with fail-open on judge errors and pre-judge quality gates	Hermes Agent	agents/goal/judge.ts

✅ Verified end-to-end in the packaged app

Every feature above was exercised through the actual desktop build before release: dynamic workflow submit → completed → event journal → amend chain (predecessor superseded); Claude Code scan of real sessions → import → share → read-only viewer → re-import as a new thread; protocol handshake with capability echo; clean quit with zero leftover processes. Desktop verification itself caught and fixed two real bugs (GUI-cwd journal path, share role mapping).

🐛 Fixes

• CI: Unit Tests workflow green again — it had been failing since before v0.7.0 (the gateway e2e test called server.listen on the handle and CI lacked a build step before the dist-importing test)
• Dynamic-workflow journal directory resolves against the project root in GUI context (was mkdir /.scitops)
• Share creation maps LangChain message types to share roles and falls back to thread-state history for imported threads

📦 Install

Platform	Asset
macOS Apple Silicon	Quill.Desktop_0.8.0_aarch64.dmg
macOS Intel	Quill.Desktop_0.8.0_x64.dmg
Windows	Quill.Desktop_0.8.0_x64-setup.exe (or .msi)
Linux (deb)	Quill.Desktop_0.8.0_amd64.deb
Linux (rpm)	Quill.Desktop-0.8.0-1.x86_64.rpm

Requirement: Node.js 18+ on the system (the app searches PATH, Homebrew, Volta, nvm, fnm). Everything else — backend, UI server, default config — ships inside the app and starts automatically: install, open, chat.

Tests: backend 583 → 694 (+111), frontend 407 → 411. Architecture policy: zero new violations.

Full changelog: https://github.com/CreamyLong/quill/blob/main/CHANGELOG.md

Quill v0.7.0 — ZCode Sync Round 4 + 8 New Framework Features + Out-of-the-Box Desktop

Quill v0.7.0 — ZCode Sync Round 4

Systematic evaluation of 11 leading harness frameworks (ZCode v3.14.3 full-source, OpenWork, DeepSeek Harness, OpenClaw, Hermes Agent, Kimi Code, OpenAI Codex, awesome-harness-engineering, DeerFlow 2.1, Agent Framework 1.19, CrewAI). Eight new feature systems ported:

Feature	Ported from	Module
Off-Peak Tasks (闲时任务) — ticket-admitted deferred execution with expiry detection + auto re-ticket	ZCode	scheduling/offpeak.ts
CommandInbox — serial message queueing while the agent is busy	ZCode	agents/inbox/
Git Checkpoints & Rewind — hidden-repo workspace snapshots, per-file restore	ZCode	runtime/checkpoints.ts
Bash Read-Only Policy — argv analysis auto-approves read-only commands (git/gh/npm semantics)	ZCode	guardrails/bash_readonly_policy.ts
Auto Review — 3-tier risk classification; exfiltration across trust boundaries hard-denied	DeepSeek Harness	guardrails/auto_review.ts
Durable Kanban Board — lease-based claims, stale reclaim, dispatcher	Hermes Agent	multi_agent/kanban.ts
ObservationPack — content-addressed obs:// handles for large tool results with exact paged recall	SoL-Pi (NVIDIA)	agents/middlewares/observation_pack.ts
Skill Attribution — "this answer used skill X"	OpenWork	skills/attribution.ts

🖥️ Desktop app now works out of the box

Previous packaged builds could not self-start their backend or render the UI. This release ships the full stack inside the app:

• Backend (gateway + compiled code + production dependencies) and the Next.js UI server are bundled as app resources
• The app auto-starts the gateway (:8200) and UI server (:3100) on launch and opens the UI — no make dev needed
• First run seeds a default config.yaml into the app data directory
• Child-process logs are written to ~/Library/Application Support/com.quill.desktop/logs/ (or the platform equivalent)
• Clean shutdown: no orphan processes after quit
• Requirement: Node.js 18+ on the system (the app searches PATH, Homebrew, Volta, nvm, fnm)

Verified end-to-end on macOS Apple Silicon: launch → gateway healthy → UI renders → API calls succeed → clean quit with zero leftover processes.

🐛 Fixes

• Evaluation pass^k aliasing bug (trial-0 score mutated mid-aggregation, misreporting failures)
• GUI-context backend crash: .scitops state dir now resolves against the project root instead of the process cwd (was mkdir /.scitops)
• Node discovery in GUI context (launchd PATH lacks Homebrew/Volta)

📦 Install

Platform	Asset
macOS Apple Silicon	Quill.Desktop_0.7.0_aarch64.dmg
macOS Intel	Quill.Desktop_0.7.0_x64.dmg
Windows	Quill.Desktop_0.7.0_x64-setup.exe (or .msi)
Linux (Debian/Ubuntu)	Quill.Desktop_0.7.0_amd64.deb
Linux (Fedora/RHEL)	Quill.Desktop-0.7.0-1.x86_64.rpm

Tests: 476 → 583 (107 new), all passing. Full changelog: CHANGELOG.md

Quill v0.6.0 — ZCode Sync Round 3 + 7 New Framework Features

Quill v0.6.0 — ZCode Sync Round 3 & Framework Competitiveness Update

What's New

Systematic evaluation of 11 leading harness frameworks/products. Seven new feature systems ported:

New Features

1. Elicitation System — Proactive ambiguity detection that generates structured clarifying questions before the agent commits to expensive runs. Scores 5 dimensions (missing context, vagueness, underspecified goal, contradiction, insufficient detail).
2. Context Rot Detection — Real-time context health monitoring middleware. Detects token bloat, tool result staleness, repetition loops, and topic fragmentation. Auto-triggers compaction/forking.
3. Workflow Concurrency Control — Adjust running workflow concurrency without stopping. Event-driven queue management with real-time UI status.
4. Two-Stage Classifier — Fast heuristic gate before expensive chain-of-thought reasoning. Classifies messages as minimal/standard/deep.
5. ACP Adapter — Agent Client Protocol JSON-RPC server for IDE integration (Zed, JetBrains, VS Code).
6. Funnel Telemetry — Drop-off analysis for user journeys across interaction stages. Identifies UX friction points.
7. Computer Use Agent (CUA) — Desktop automation broker with fail-closed privacy design. PrivacyGuard with configurable rules + immutable audit trail.

Desktop

• Quill Desktop v0.6.0 (macOS aarch64 DMG)
• Bundle ID: com.quill.desktop
• Fixed Info.plist metadata (proper name + version + executable)

Frameworks Analyzed

ZCode v3.14.x, OpenWork, DeepSeek Harness, awesome-harness-engineering, DeerFlow, CrewAI, AutoGen, Kimi Code CLI, OpenAI Codex, OpenClaw, Hermes Agent

Quill v0.4.0 — Framework Competitiveness Update

Quill v0.4.0 — Framework Competitiveness Update

Systematic analysis of 10 leading harness frameworks (OpenWork, DeepSeek Harness, OpenClaw, Hermes Agent, Kimi Code CLI, OpenAI Codex, DeerFlow, AutoGen, CrewAI, awesome-harness-engineering) informed the porting of 7 high-impact feature systems.

New Features

1. Progressive Skill Loading — 3-level disclosure (catalog → summary → deep dive) with intent-term ranking. Reduces token usage by only loading full skill content when needed. (Hermes Agent + DeerFlow 2.0)

2. Context Modes for Subagents — Isolated vs. snapshot context delivery with auto-suggestion based on task characteristics. Subagents with isolated context use ~67% fewer tokens. (DeerFlow 2.0 + awesome-harness-engineering)

3. State Machine Guardrails — Phase-dependent tool access control (planning → research → implementation → review → cleanup) with per-phase tool group policies. (awesome-harness-engineering + DeerFlow RBAC)

4. Adaptive Orchestration Topology — Dynamic coordination pattern selection (parallel/sequential/hierarchical/hybrid) based on task dependency graph analysis. (awesome-harness-engineering AdaptOrch + CrewAI)

5. Extension Manager — Five contribution kinds (middleware, lifecycle hooks, observers, gateway services, HTTP routers) with enable/disable lifecycle and dependency resolution. (DeerFlow 2.0 + OpenClaw plugin SDK)

6. Compute Worker Pool — CPU admission-controlled worker pool with task priority, bounded queue, overload handling, and graceful shutdown. (OpenClaw WorkerTaskPool)

7. Two-Phase Memory Pipeline — Parallel rollout extraction + serialized global consolidation. (OpenAI Codex CLI)

Testing

• 65 new tests across 5 test files, all passing
• Backend builds cleanly (tsc)
• Rust desktop shell compiles (cargo check)
• 473 existing tests continue to pass (3 pre-existing failures unrelated to changes)

Documentation

• Updated harness-framework-comparison.md with full analysis of all 10 frameworks

Competitive Position

Quill now matches or exceeds all 10 frameworks on 22 of 26 capability dimensions.

Quill v0.3.0 — Session Forking, FTS5 Search, Marketplace Install

What's New in v0.3.0

🔀 Session Forking

Deep fork with checkpoint copy — branched threads carry full conversation history. Powered by the new runtime/fork.ts service that copies all checkpoints and pending writes.

🔍 FTS5 Session Search

BM25-ranked full-text search across all thread messages with snippet extraction and prefix matching. Uses SQLite FTS5 when available, falls back to in-memory scan.

🛒 Marketplace Install

Install skills directly from GitHub repos or arbitrary URLs. Supports owner/repo shorthand, branch/tag selection, and auto-detection of SKILL.md files.

📊 Competitive Position

Quill now matches or exceeds all 10 leading harness frameworks on 18 of 22 capability dimensions (up from 11 previously).

Framework Analysis Updated

Updated harness-framework-comparison.md with the latest competitive analysis (2026-09-11) covering OpenWork, DeepSeek Harness, OpenClaw, Hermes Agent, Kimi Code, Codex, DeerFlow 2.0, CrewAI, AutoGen, and awesome-harness-engineering.

Verification

• ✅ Backend type-check: clean (0 errors)
• ✅ Backend tests: 408 passed (3 pre-existing failures unrelated to this release)
• ✅ Pushed to main: commit e0b732b

Assets

This release includes the full source. Desktop binaries require Rust/Tauri toolchain to build locally:

cargo install tauri-cli --version "^2.0.0-beta"
cd desktop && cargo tauri build

Full Changelog: v0.2.0...v0.3.0

Quill Desktop v0.2.0

ci: add GitHub Actions workflow for desktop build & release

Automated multi-platform builds triggered on version tags (v*):

- macOS: aarch64 (Apple Silicon) + x64 (Intel) → .dmg
- Windows: x64 → .exe (NSIS installer)
- Linux: amd64 → .AppImage + .deb

Uses tauri-action for builds, creates draft release, then publishes
after all platforms succeed. Supports manual dispatch with tag input.

Requires secrets: TAURI_SIGNING_PRIVATE_KEY (optional, for signed builds)

Co-Authored-By: Claude Opus 4.6 <<EMAIL>>

Quill Desktop v0.1.0

🪶 Quill Desktop v0.1.0

Open-source AI super-agent desktop app built with Tauri 2 (Rust + WebView).

What's Included

• Native folder picker — bypass browser security restrictions, select any local folder
• Auto-start servers — launches Gateway + frontend automatically, no manual setup
• System tray — runs in background, quick access via menu
• Workspace sync — incremental sync between local workspace and Gateway
• System integration — clipboard, notifications, window management
• Auto-update — built-in updater via Tauri v2
• 8 languages — English, 中文, 한국어, 日本語, Français, Русский, Español, العربية

New: Enhanced Folder Discovery

• Quick access — Home, Documents, Desktop, Downloads, Projects
• Recent folders — automatically tracks last 20 opened folders with metadata
• Bookmarks — pin frequently used folders (⭐)
• Folder metadata — file count, total size, last modified
• Subdirectory browsing — navigate folders without reopening picker
• Persistent history — recent/bookmarks saved between sessions

Downloads

Platform	File	Size
macOS (Apple Silicon)	Quill.Desktop_0.1.0_aarch64.dmg	5.7 MB

System Requirements

• macOS 11+ (Apple Silicon)
• Node.js 22+ (for Gateway and frontend)
• 4 GB RAM recommended
• Ollama (optional, for local models)

Quick Start

1. Download the .dmg file
2. Open it and drag Quill Desktop to Applications
3. Launch Quill Desktop — it auto-starts Gateway + frontend
4. Click "Open Workspace" to select a local folder

Architecture

┌─────────────────────────────────────────────┐
│  Tauri 2 Desktop App                         │
│  ┌───────────────────────────────────────┐  │
│  │  WebView (Next.js frontend)            │  │
│  └───────────────────┬───────────────────┘  │
│                      │                       │
│  ┌───────────────────▼───────────────────┐  │
│  │  Rust Backend                          │  │
│  │  • fs_bridge (file CRUD + discovery)   │  │
│  │  • system_bridge (clipboard/notify)    │  │
│  │  • sync_bridge (workspace sync)        │  │
│  │  • gateway (auto-start child servers)  │  │
│  │  • tray (system tray + menu)           │  │
│  └───────────────────┬───────────────────┘  │
│                      │ spawns                │
│  ┌───────────────────▼───────────────────┐  │
│  │  Gateway (port 8200) + Frontend (3100) │  │
│  └───────────────────────────────────────┘  │
└─────────────────────────────────────────────┘

Known Issues

• Windows and Linux builds coming soon
• First launch may take a few seconds to start both servers
• Requires the Quill repo to be cloned for the Gateway/frontend code