Shard public beta Get Shard
All docs pages Chatting, 5 of 13

Docs / Using Shard

Chatting

This page covers the Chat window: sending messages, the queue, Stop, your chat history, context compaction and the usage details.

The Chat window#

  • Sidebar (left): New chat, your list of Chats, and buttons at the bottom for the changelog, hiding the sidebar, and opening Settings. It also shows your message limit, for example “Message limit: 12/10000”.
  • Header (top): the chat’s name, the provider, the model, your reasoning choice, and a request count. A button opens Chat details.
  • Messages (middle): your messages, Shard’s answers, and a row for every action Shard takes.
  • Message box (bottom): type here and press Enter or the send button.

If you scroll up while Shard is working, a Jump to latest button appears with a count of new items.

What you see while Shard works#

  • A status line shows what Shard is doing, like “Thinking” or “Compacting context”.
  • Every tool Shard uses gets its own row, grouped by type (for example “Read 2 scripts”). Click a row to open it. You see the Arguments Shard sent, the Output, any Changes and any Error. Use Select to select the text so you can copy it, and Show more for long content.
  • Code blocks in answers have a Select code button.
  • When Shard finishes, the turn shows “Worked for” and the time it took.
  • If Shard changed scripts, a Review section lists each changed script with lines added and removed. Open View recorded script changes to see them. To take changes back, use Studio’s Undo (see Undo).

Sending while Shard is working: the queue#

You don’t have to wait. While Shard is working on a message, the send button is replaced by Stop, and pressing Enter puts your new message in a queue instead of interrupting.

Queued messages appear above the message box. Each one has buttons:

  • Steer adds the message to the current task, right before Shard’s next step. Use it to correct or guide Shard mid-task (“use a ModuleScript instead”). The row shows “Steering into the run…” until Shard picks it up. If Shard is idle and only waiting for its agents, the message is delivered right away.
  • Send appears when nothing is running. It sends the message as a normal new message.
  • Remove deletes it from the queue.

When a task finishes normally, Shard sends the next queued message by itself, then the next, one at a time. If you stop a task, or it fails, the queue waits for you.

The queue lives only in memory. It is lost when Studio closes, when Shard restarts, and when you start a playtest. Deleting a chat also clears its queue.

Stop#

Press Stop to end the current task. Shard shows “Stopped by user.” Any agents working for that task stop too.

Shard can’t take back a change that was already made. If you stop in the middle of an edit or a RunCode call, part of it may already be in your place. Check, and use Undo if needed.

Chat history#

Every chat is saved automatically in Studio’s plugin settings on this computer.

  • The list is in the sidebar, most recently used first. Pinned chats stay on top, with a 📌 before the name.
  • Open a chat by clicking it.
  • Pin or unpin by hovering over the chat and clicking the pin icon.
  • Delete by hovering and clicking the delete icon, then clicking the confirm icon that replaces it within 3 seconds. Deleting can’t be undone.
  • Rename: a new chat is named after the first 60 characters of your first message. TODO (owner) the plugin has no rename control yet (the code supports renaming, but no button triggers it).

Good to know:

  • Your chats are stored on your computer, not in your place file and not on Shard’s server. They don’t follow you to another computer. TODO (owner) confirm whether you want to say the list is shared across all places on this computer (the storage keys don’t include a place ID).
  • When Shard reads a script, the script’s source is kept only while Studio is open. After you reopen Studio, that step shows “Source not saved with the chat”, and Shard reads the script again if it needs it.
  • If Studio closed while Shard was answering, that answer shows “This response was interrupted. Send a new message to continue.”
  • A chat that couldn’t be loaded shows as “Unavailable conversation”. Start a new chat to keep going.

Provider and model per chat#

Each chat remembers its own provider, model and reasoning choice. Changing them in Settings → AI while a chat is open changes that chat only. With New chat open, you set the default for new chats. You can switch models in the middle of a chat: Shard keeps the conversation.

Context compaction and /compact#

Every model can only read a certain amount of text at once (its context). Long chats eventually get too big.

Automatic compaction. Before each request, Shard estimates the size of the conversation. When it gets close to the model’s limit, Shard asks the model to summarize the older part. You see “Compacting context...”, then “Context compacted. Earlier messages remain in this chat.”

  • Your messages stay on screen. Only what the model reads is shortened.
  • Shard keeps the latest part of the conversation word for word.
  • If it can’t make the chat small enough, you see “Choose a model with more context.” Switch to a model with a bigger context, or start a new chat.

Compact by hand with /compact. Type /compact and press Enter.

  • If nothing is running, Shard compacts right away.
  • If a task is running, you see “Compaction queued for the next completed tool boundary.” Shard compacts at the next safe point.
  • “No additional completed turns or tool rounds to compact.” means there isn’t enough finished conversation to summarize yet.

A compaction is a normal request to your model, so it counts toward your message limit and costs tokens.

Chat details and usage#

Click the details button in the chat header to open Chat details. It has two tabs.

Usage shows numbers for the open chat:

  • Input tokens and Output tokens: totals for the chat, with how many requests reported them.
  • Estimated cost: a USD estimate based on Shard’s catalog prices. It’s an estimate, not your bill. For Codex and Claude Code it says API equivalent instead (see Connected accounts).
  • Prompt cache: how often cached input was used.
  • Latest request: provider, model, result, duration and time.
  • Last request context: how full the model’s context was, with a bar and an estimated breakdown.
  • Coverage: how many requests reported each number. Some providers don’t report everything.

The header also shows the number of requests in the chat, and lines added and removed, for example “14 requests · +120 −8”.

Shard’s server keeps the usage report for signed-in accounts only. As a guest, the Usage tab may show less or say the report is unavailable.

Subagents lists the agents that worked in this chat. See Agents.

Announcements and polls#

Shard sometimes shows a notice or a short poll in the chat. You can close a notice, answer a poll (with an optional comment), or press Skip. Your answer, and your display name, are sent to Shard. See Privacy and data.