Alpha 6: Many providers, per-agent everything, real task workflow

Alpha 5 put agents at the center. Alpha 6 makes each one yours to shape. Every important behavior (which model talks, when it replies, how it summarizes, what it can read, what extra instructions it follows) is now configured per agent rather than globally. On top of that, the provider list grew from two to nine, and the task tracker finally works like a tracker.

What's new

🔌 Many more model providers
OpenAI, DeepSeek, GLM, Kimi, Mistral, Ollama, and Qwen join Claude and LM Studio. A new OpenAI-compatible base lets the same plumbing serve any OpenAI-style API, so future providers are a config away rather than a new integration.

🧰 Per-agent customization, end to end
Each agent now owns its own:

• Auto-reply policy: always, mentions-only, or never. Stop noisy back-and-forth in group chats without muting the whole agent.
• Summarizer: pick the model and prompt your agent uses to compact long sessions, separate from the model it actually thinks with.
• Append-to-system-prompt: layer custom instructions on top of the base prompt without forking it. Updates to the base template still flow through.
• Skills: provider-agnostic SKILL.md files that get translated into Claude skills or system-prompt text depending on the provider.
• Store: namespaced key-value scratch space with optional TTL, separate from curated memories. Good for "URLs already scanned", "deduped IDs", "cooldown timers".
• Volumes: host directories now mount under the agent's home as ~/{name} with read-only or read-write modes, so cd ~/projects just works.

✅ Tasks that behave like tasks

• Five-state model: pending → todo → in-progress → done / cancelled. The old "completed" / "failed" wording is gone.
• Approval gate: turn it on, and tasks an agent creates for itself wait for your nod before the auto-processor picks them up.
• Comments thread with edit and delete history, separate from the status timeline.
• Kanban order for drag-and-drop boards.
• Archive: done and cancelled tasks fade out of the default list after N days, plus a manual archive button. Archived tasks live in a paginated archive view.

💬 Calmer chat

• Messages queue up so an agent finishes one reply before starting the next.
• A new chat command palette and inline chat panel.
• Chat list resorts by activity and updates the sidebar live as messages arrive.
• Permission prompts now open inside the relevant chat instead of a floating modal.
• Message history view with its own route so you can jump back to old conversations.

📝 Better writing surface

• Upgraded Markdown editor and command input.
• MarkdownTextarea everywhere a plain textarea used to live.
• Global search reaches further across the app.
• Agent tabs in the navigation, with proper deep-linkable routes per agent view.

🐳 Container and browser polish

• A Global Container section shows Docker container status and lets you restart it.
• Volume edits trigger an automatic, coordinated container restart so mounts stay in sync.
• Each agent run gets its own browser profile, cleaned up after.

Known limitations

Still alpha. Expect rough edges, breaking changes before v1.0, and occasional bugs. The provider expansion is wide rather than deep: some providers ship with the same defaults as their OpenAI-compatible peer; tune temperatures and tool prompts to taste. Report issues via GitHub Issues.