# Orchestrator

Ivy Tendril, 2025 to 2026. The chat surface of Ivy Tendril, where a task becomes a plan and a coding agent runs it.

Role: Founding Design Engineer at Ivy Interactive. Stack: C# / .NET 10, Ivy Framework, React 19, TypeScript, Tailwind CSS, Vite.

Ivy Tendril is a desktop app that turns a task into a plan, runs a coding agent on it in an isolated git worktree, and opens a pull request after review. I designed and built the chat surface: the empty state, the composer, the agent picker, structured questions, events, the message queue, and the plan side panel.

![Ivy Tendril chat home with a greeting, four suggested actions and the composer](https://www.joelbystedt.com/work/orchestrator/full-view.png)

*The chat home: a greeting, four suggested actions, and the composer.*

## Suggested actions on the empty page

**Discovery.** New users opened Tendril, saw an empty chat, and asked in Discord what they were supposed to type.

**Problem.** An empty text field gives no idea of what the product can do, so the first minutes were spent guessing.

**Solution.** The empty state shows the four most useful next actions as buttons: review the waiting plans, ask why a job failed, add a project, edit verifications. Clicking one sends a ready-made prompt.

## Agent picker

**Discovery.** A customer used Claude Code for planning and Codex for execution. Switching meant editing the config file and restarting the app.

**Problem.** The agent was a global setting, so changing it mid-conversation was slow and easy to get wrong.

**Solution.** The picker lives in the composer. The left column lists the agents (Antigravity, Claude Code, Codex, Copilot, Gemini, OpenCode, Ivy Agent, OpenAI Proxy), the right column sets the model and effort for the selected one. The badge next to the send button shows which agent will answer.

![Agent picker listing eight agents, with a Claude Code submenu for model and effort](https://www.joelbystedt.com/work/orchestrator/agent-agnostic.png)

*Agent on the left, model and effort on the right. Claude Code is set to Claude Opus 5 at high effort.*

## Structured questions

**Discovery.** Session logs showed agents asking long open questions. Users answered with one word or stopped responding.

**Problem.** A question buried in prose is easy to miss and hard to answer precisely, which stalls the job.

**Solution.** When an agent needs a decision, the question renders as a form. Each option has a title, a one-line reason, and a recommended default, plus an Other field. A footer shows the time, tokens, and cost of the turn.

![Question form with four options, one marked as recommended, and a Submit response button](https://www.joelbystedt.com/work/orchestrator/chat-questions.png)

*The agent asks how to proceed. The recommended option is marked, and the footer reports 24.7 seconds and 65 cents for the turn.*

## Events in the chat

**Discovery.** Users missed that a plan had been stopped because the notice appeared in a separate tray and disappeared after a few seconds.

**Problem.** State changes happened outside the conversation, so the chat and the real status drifted apart.

**Solution.** Events such as a stopped plan appear inline in the transcript with an icon, followed by the agent's explanation. Tool calls fold into one line that can be expanded.

![Chat transcript with an inline event Stopped plan 485, a folded row reading 2 tool calls, and an answer with headings](https://www.joelbystedt.com/work/orchestrator/events.png)

*A stopped plan appears as an event row, followed by what happened, what is unknown, and what has not changed.*

## Message queue

**Discovery.** People typed follow-ups while the agent was working. The message was either lost or interrupted the run.

**Problem.** The composer had no defined behavior during a run.

**Solution.** Messages sent during a run are queued and listed above the composer, each with send now, edit, and delete. The send button becomes a stop button until the run ends.

![Queued Messages panel above the composer with one queued message and send, edit and delete actions](https://www.joelbystedt.com/work/orchestrator/message-queue.png)

*One queued message waiting for the agent to finish.*

## Plan side panel

**Discovery.** To change a plan, users copied text from the plan page into the chat and pasted the answer back.

**Problem.** The plan and the conversation about it lived on different screens.

**Solution.** Opening a plan splits the view: the document on the left with Plan and Details tabs, the chat on the right with prompts such as tighten the scope, explain the solution, and what could go wrong. Execute is in the header.

![Split view with the plan 483 document on the left and a chat panel on the right](https://www.joelbystedt.com/work/orchestrator/side-panel.png)

*Plan 483 open next to its chat, with Execute one click away.*

## Result

Chat is the entry point, and structure takes over where a decision, an event, or a document needs it. The components are built with Ivy Framework widgets in C# with React underneath, and the same primitives are now used across the app.

Links:
- [Ivy Tendril on GitHub](https://github.com/Ivy-Interactive/Ivy-Tendril)
- [tendril.ivy.app](https://tendril.ivy.app)