Work

You have a feature to build, a bug to fix, or a question about your code. Work (Mod+2) is where you talk to agents and get things done.

Press Mod+2 to open it.

Overview

Work is the main chat surface in kill9. It's where you describe what you want and agents respond with code, answers, test results, or questions. Each initiative and story has its own chat session with full history. You can run multiple sessions in parallel using track tabs and chat splits.

The chat interface has three parts: the conversation history at the top, the input area at the bottom, and contextual information along the sides. Everything you type and everything agents do is captured in the session.

What you can do

In Work, you can:

• Chat with any agent by mentioning them by name
• Let k9 coordinate a full pipeline from planning through review
• Switch between multiple sessions using track tabs
• Run two conversations side by side with chat splits
• Attach files, code snippets, and references as context
• Review agent activity including tool usage, diffs, and delegation chains
• Approve or reject agent actions through permission dialogs

How it works

Starting a chat

Type your request in the input box and press Enter. Mention an agent by name to direct the work:

Use kdev to add email validation to the signup form.

Or let k9 coordinate the whole pipeline:

Use k9 to implement a password reset flow with tests and review.

When you open a new chat, kill9 suggests starter prompts based on context. For a feature initiative, it might suggest "Break this into stories" or "Write acceptance criteria." These are suggestions — type anything you want.

Agent selector

The agent selector lets you pick which agent handles your message. You'll find it in the input area. Choose a specialist for focused work, or pick k9 for multi-step work that needs coordination.

Use the agent selector when you know exactly who should handle the task. Skip it and let k9 decide when the work spans multiple phases.

Model selector

The model selector lets you choose which AI model powers the agent. Different models have different strengths — some are faster, some reason more deeply, some handle code better. Pick the model that matches your task and budget.

If you're unsure which model to use, the default works for most tasks. Switch to a more capable model for complex architecture decisions or tricky bugs.

Reasoning level

Reasoning level controls how much thinking the agent does before responding. A higher reasoning level means the agent spends more time analyzing your request and considering alternatives — useful for complex problems. A lower level means faster responses for straightforward tasks.

Use high reasoning for architecture decisions, tricky bugs, and complex refactors. Use low reasoning for simple edits, quick questions, and routine changes.

File attachments

You can attach files, code snippets, and references to your message. The agent reads everything you attach and uses it as context for its work. This is how you share specific code, design mockups, or external documents without copy-pasting.

You can also send elements from other panels — select something in Browser or Docs and send it to the chat as context.

Vim mode input

If you're a Vim user, enable Vim mode in the input area. You get modal editing — normal mode for navigation, insert mode for typing. All your muscle memory works: dd to delete a line, cw to change a word, Esc to exit insert mode.

Toggle Vim mode in the input area controls.

Track tabs

Track tabs appear at the top of the Work panel. Each open session gets its own tab, so you can switch between conversations without losing context. Click a tab to jump to that session, or close tabs you're done with.

Use track tabs when you're working on multiple things in parallel — one tab for a feature, another for a bug fix, a third for a quick question.

Chat splits

Chat splits let you view two conversations side by side in the same panel. This is useful when you need to reference one agent's output while talking to another, or when you're comparing approaches from two sessions.

Split the chat using the split controls in the panel header. Each side runs independently with its own agent, model, and session.

Delegation breadcrumbs

When an agent delegates work to a subagent, you see delegation breadcrumbs at the top of the message. These show the chain: which agent delegated to which, and what the subagent is working on.

Click a breadcrumb to jump to that subagent's session and see what it's doing. This keeps you oriented when k9 coordinates multiple specialists in sequence.

Permission dialogs

Agents ask for permission before taking sensitive actions — editing files, running commands, or accessing external services. A permission dialog appears showing exactly what the agent wants to do.

Approve the action to let it proceed, or reject it to stop. You can also grant blanket permission for specific actions if you trust the agent's judgment for that category.

Message parts

Agent messages aren't just plain text. Each message can include multiple parts:

• Text — the agent's explanation, reasoning, or questions
• Tool usage — shows what tool the agent used, what it sent, and what came back
• Code blocks — formatted code with syntax highlighting and file references
• Diffs — file changes shown as before/after comparisons
• Agent activity — status updates when agents are working on your behalf

Each message also has controls: copy the text, rollback the changes, retry with different parameters, or delete it from history.

Some messages include entity links — clickable references to initiatives, epics, or stories. Clicking one opens it in Plan (Mod+3).

Tips

• If you're not sure which agent to use, just type your request and let k9 figure it out.
• Attach relevant files before asking a question. The more context the agent has, the better the answer.
• Use chat splits when you need to compare two approaches or reference one session while working in another.
• Check delegation breadcrumbs when k9 is coordinating — they tell you exactly who's doing what.
• Higher reasoning levels cost more tokens but produce better results for hard problems. Don't overthink simple tasks, but don't underthink complex ones.
• You can reference documents, mockups, and entities by name in your messages. The agent understands these references.
• Sessions persist across restarts. Close kill9, come back tomorrow, and your conversation is still there.

Related

• Agents — which agent does what
• Inbox — where work arrives before it becomes a session
• Plan a feature — creating initiatives that become work sessions
• Review changes — accepting or rejecting agent work
• Tabs and splits — managing multiple panels and conversations