Sundial Agents

The Sundial Agents extension provides an alternative, more code-focused way of working with AI agents. This doesn't replace chats or spec-driven development. However, it does offer a supplementary workflow intended to help you achieve a flow state while still using LLMs to complete work.

For spec-driven development and storing durable decisions and research in code, see the Sundial extension.

In this workflow, you and multiple agents work on the same code branch at the same time. Those agents perform asynchronous work or research for you, with the results either in the code or embedded in the code as annotations. Agent tasks are small, self-contained, and fire-and-forget (come back later). Agents are prompted to understand that they are working on a changing codebase with you and other agents and to mitigate accordingly.

How it works

Each Sundial agent has a memory, model, and reasoning effort that you control.

As you work in code, you can type % to enter Sundial command mode on any empty source line and direct that message to any Sundial agent. For example, typing %q>1 will send a question to agent 1.

This brings up a form with the message text box in focus, so you can type the message to the agent and press Enter to submit (or Escape to cancel). This returns your cursor directly to its previous location in the source for a seamless, hands-on-keyboard experience.

If you have the Vim plugin installed, it brings you back to Normal mode.

Annotated source shows a marker to the left of the source line (kind of like a breakpoint). As your cursor moves between lines, the annotation window (in the lower-right sidebar) switches to that annotation.

The agent will then work independently (and asynchronously), periodically sending status updates that you can see in the Sundial agent sidebar. Agents can also annotate other source locations as part of their work, and you can follow these links from your original code annotation. This helps tremendously with code-explanation tasks. When the agent finishes a task, they will post an official response on the annotation.

Annotations are tagged to the most recent Git commit. As you navigate the source, you see all annotations alongside the agents' official responses, optionally filtered to the most recent commit. Alternatively, select the Annotations tab in the sidebar to see the entire list of annotations.

Sundial Task Types

• %q — answer a question
• %d — deep research with references to primary sources
• %f — fix a bug
• %w — write new code
• %r — refactor code
• %c — cleanup partially implemented code
• %t — write tests

How it works behind the scenes

Each source file has a companion annotation file in the .sundial directory (gitignored by default).

As source changes, the annotations are continuously reanchored to maintain their logical anchor point.

A companion CLI (installed via npm) manages all annotations and work items and connects submitted messages to local coding agents. Currently, the only implemented provider is Codex.

Experimental Features

Diff View Commands

As you work, type %d to bring up the diff commands. You can switch between diff and normal views and between inline and side-by-side layouts without taking your hands off the keyboard.

• %dd toggles the global diff view.
• %di toggles VS Code's inline and side-by-side diff layout.
• %d+ moves the baseline one first-parent commit back; %d- moves it one commit toward HEAD.
• %d0 resets the baseline to HEAD; %dp selects the last permanent commit.

Temp Commit Commands

As you work, use %c on a blank line to bring up a commit menu. You can create temporary commits of work that reset the diff view. These are automatically squashed if you use the built-in commit options to create the "real" commit.

• %cf creates a temporary checkpoint for the current file.
• %ca creates a temporary checkpoint containing all dirty, staged, and non-ignored untracked files.
• %cm asks for a message, then consolidates the temporary checkpoint suffix and any remaining dirty work into one permanent commit. A prior %ca is not required.

VS Code task commands

Sundial also contributes no-argument VS Code commands that open the same composer without inserting or deleting a % line. The active source line becomes the interaction anchor, and the document is saved before the composer opens. These commands are available from the Command Palette, keybindings, and VSCodeVim's :vsc[ode] <command-id> bridge.

For example, :vsc sundialEditor.task.fix opens a line-scoped Fix composer on the line under the cursor. Choose the target agent and send the message normally.

Sundial Agents panel

Sundial Editor contributes the Sundial Agents panel to VS Code's right-hand Secondary Side Bar, beside other collaboration surfaces such as Codex and Claude Code.

By default, five agents are initialized, but there is no limit to the number of agents you can have active.

Each named agent has its own persistent queue. Submitted interactions move independently through waiting, working, and completed, so a busy agent can finish older work in order while other agents continue. A collapsed work card shows its latest status; expand its history to see ordered progress updates. These tend to be less verbose than a typical chat session.

You can open the agent session in Codex using a link that opens it automatically in the terminal.

Agent controls let you add more named agents, rename them, choose each agent's model and reasoning effort from the provider's current catalog, read current-session status history, open a provider conversation, interrupt current work, reset one provider session, or reset all managed sessions and runtime work together.

Agent queues, session records, status histories, and source-annotation companions survive local VS Code restarts under .sundial/. Project bootstrap adds that entire directory to .gitignore, so this state remains ephemeral.

Autosave

The extension contributes defaults for VS Code's built-in delayed autosave: files.autoSave is afterDelay and files.autoSaveDelay is 1000 milliseconds. User, workspace, folder, and language-specific settings retain their normal precedence. It is highly recommended that you work with autosave installed if editing code concurrently with agents.