v0.6.5 — Queen Phase Separation 👑

Highlights

Queen Phase Separation

The queen agent's responsibilities are now cleanly separated into three phases, each with its own scoped system prompt and toolset.

• Building phase — Full coding tools for writing and editing agent code, designing graphs, and constructing worker agents
• Staging phase — Agent is loaded but not running; queen can inspect, configure, and launch the worker
• Running phase — Worker is executing; queen monitors, controls, and can stop or restart
• Dynamic prompt swapping — System prompt and available tools change on each phase transition, reducing prompt size and improving instruction adherence
• QueenPhaseState — New state machine managing phase lifecycle, tool availability, and phase-change event notifications
• Smaller model support — Phase-scoped prompts significantly improve instruction following on smaller models

Queen Thinking Hook

• Thinking visibility — New thinking hook surfaces queen reasoning during agent construction
• Max token control — Configurable token limits for thinking output

Telegram Tool Expansion

Expanded Telegram integration with message management, media, and chat info operations. (#5403)

---

What's New

Core Framework

• Queen prompt optimization — Condensed building phase prompts, removed unused prompts, reorganized tool documentation
• Escalation system — New escalation tool for worker-to-queen communication during execution
• Progressive disclosure — Runtime data revealed progressively to avoid information overload
• Dynamic system prompts — System prompts loaded dynamically into context
• Validation improvements — GCU and validation added to initialize_agent_package; agent.json validated before parsing (#5846)
• Judge improvements — Judge can now wait for queen input; skip-judge logic improved
• BOM-safe JSON loading — Handles byte-order marks in JSON files (#5901)
• LLM logger on by default — LLM debug logging enabled by default
• Config error logging — Parse errors now logged for better debugging (#4955)
• Provider key detection — Improved indirect variable expansion for API key detection (#5504)
• Minimax provider — Added Minimax provider mapping and stream fallback

Frontend & Server

• Duplicate session fix — Prevent duplicate session creation when starting from home
• CLI validation — Validate --output path before agent execution (#5838)

Refactoring

• Remove old TUI CodingAgent — Deprecated TUI-specific coding agent removed
• Remove unused builder functions — Cleaned up legacy builder code and output cleaner
• Remove old skills — All unused agent skills removed
• Rename queen mode → phase — Consistent "phase" terminology throughout codebase
• Reorganize tools — Coder tools and agent initialization tools restructured

---

Bug Fixes

• Fix back-to-back compaction edge case
• Fix turn signal edge case
• Fix output key that terminated the queen
• Fix duplicated session calls
• Fix Cannot write to closing transport error
• Fix missing GCU prompts and instructions
• Fix thinking hook max token limits
• Fix MCP tools and templates loading
• Fix logger schema mismatch
• Fix legacy agent.json loading error handling
• Fix Aden client import duplication after rebase
• Skip POSIX permission tests on Windows (#5842, #5847)

Documentation

• Add README for brevo, csv, runtime_logs, account_info tools (#5602)
• Remove old agent skill references
• Restructure docs layout

---

Community Contributors

• Jack (@jackthepunished) — Telegram tool expansion with message management, media, and chat info (#5403)
• VasuBansal7576 — Minimax provider integration

---

Upgrading

git pull origin main
uv sync

# For the web frontend
cd core/frontend
npm install
npm run build

or, simply run

./quickstart.sh (linux/mac)
.\quickstart.ps1 (windows)