Release v0.73.0

[wladimiiir]

Highlights

⚡ Virtualized Message Rendering as Default

We've made message virtualization the default rendering mode, bringing a significant performance improvement - especially for long-running tasks with hundreds of messages. This has been a long time coming, but previous solutions had too many quirks to become the default. The current implementation is a solid replacement that dramatically improves task loading times and scrolling performance. If you do experience any issues, you can still enable the legacy mode, but we're confident this will be a smooth upgrade. If you run into any problems, please report them on GitHub so we can address them.

🔌 Agent Client Protocol (ACP) Support

The aiderdesk CLI binary now supports the Agent Client Protocol (ACP) - an open standard for communication between AI coding agents and editor/IDE clients. What does this mean for you? You can now use AiderDesk as a backend agent directly within ACP-compatible editors like IntelliJ IDEA, Zed, and others, while keeping all of AiderDesk's powerful features: subagents, extensions, tool approvals, worktrees, task management, and more. No need to leave your IDE or switch contexts - your agent works right where you are, with full control over tool execution and approval flows. This brings the best of both worlds: the rich editing experience of your favorite IDE combined with AiderDesk's flexible agent capabilities.

Other Changes

• OpenAI-Compatible Providers: Added support for extra body parameters, giving you fine-grained control over provider-specific options that aren't part of the standard API.
• Anthropic-Compatible Providers: Ensured the correct /v1 suffix in the base URL to prevent configuration issues.
• Model Defaults: Updated default models across providers to use the latest versions.
• Bug Fixes:
  • Showing correct compact threshold values from agent profile overrides.
  • Fixed shell PATH extraction when login messages were present in output.
  • Excluded worktree symlink folders from git tracking and diffs during rebase.
  • Fixed usage report not being attached to tool-only responses, which previously prevented context compaction from triggering.

🧩 Extensions

• Cursor: Added health checks and automatic restart to the proxy for better reliability, plus a build phase during installation.
• Cursor SDK: Added fast and not-fast model variants, and removed the cursor-sdk/ prefix from model names for cleaner references.
• Auggie SDK: Removed a duplicated claude-opus-4.7 model entry.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

---

This discussion was created from the release 0.73.0.

[kuptom]

Hi, thank you very much for your work. I would like to ask how to set the System prompt for the default model (llama.cpp + Agents-A1) in Agent mode so that it correctly calls sub-agents.

In the settings / AGENTS / myProject (not Global), I have configured X sub-agents and 1 orchestrator. For the sub-agents, I can set the System prompt and Description, but for the Orchestrator, which is not a sub-agent, this is not possible. I can only set Rules. However, that is not enough to make it call the sub-agents.

I also tried adding an AGENTS.md file to the project, where I specified that sub-agents should be called as a priority, but so far without success.

Could you or someone else please advise me? Maybe with a simple example setup containing 1 orchestrator and 2 sub-agents?

Thank you very much.

[wladimiiir]

Hi, I’m actually introducing this feature in the next release - the ability to override the default system prompt for a specific agent via the Agent settings.
Currently, it is only possible to override the system prompt globally by creating custom prompt files (system-prompt.hbs in your case).

[kuptom]

Thank you very much! Your advice helped me a lot. I created a custom system prompt based on the documentation you pointed me to, and it started working really well. I really appreciate your help and look forward to the next release!

[richardchen874-sys]

The virtualized message rendering work is a good signal that Aider Desk is hitting real long-running agent sessions. Once sessions reach hundreds of messages, provider choice and context growth become user-facing cost issues.

I would consider showing provider/model, context size, token usage, and retry count at the session level. That would help users decide when to stay local, when to use a hosted OpenAI-compatible endpoint, and when to trim context.

I am testing an OpenAI-compatible multi-model API layer, and coding-agent UX is a strong fit for usage visibility. Are you planning a session-level usage or cost panel for agent mode?

[wladimiiir]

Thanks for the feedback. I think the important thing is to define what a "session" is, as that concept does not currently exist in AiderDesk. For example, you could consider a task and all its subtasks as a single session. In that case, it should be fairly easy to create an extension that adds information to the usage panel in the bottom-right corner - task-usage-info-bottom placement - to calculate and display the total cost for that task and its subtasks.