Chat & Composer
The main window is where most of your time is spent. This page describes the composer (the input area at the bottom) and how a turn unfolds from typing to receiving a streamed response.
Anatomy of the composer
The composer is split into a text area at the top (placeholder: Ask, build, or run…) and a control bar underneath. The bar has a left cluster of input-affordance buttons and a right cluster of status / send controls.
Left cluster (input affordances)
| Control | Purpose |
|---|---|
| Attach file | Upload a file (PDF, DOCX, ZIP, …). Attached files are rendered with the right viewer and supplied to the model. |
| Commands | Open the Commands menu — the unified entry point for tools, skills, MCP servers, and saved prompts. Selections apply to this turn. |
| Attach image | Upload an image. Vision-capable models will receive it as an image input. |
| Voice input | Press to record. Whisper transcribes locally; you see a live audio-level meter. Pressing again stops the recording. |
Right cluster (status and send)
| Control | Purpose |
|---|---|
| Active tools | A small badge with a count (e.g. "3") that shows how many tools are currently active for this turn. Click to inspect / adjust. |
| Confirm before tools | A toggle. When on, the agent pauses for approval before running a tool. When off, tools run inline. |
| Model picker | Shows the active model (e.g. Ghast 2 Pro). Click to switch. Lists every configured provider's models. |
| Context ready | A small indicator that turns active once memory / project / workspace context has been assembled for the turn. |
| Send | Submit the prompt. While the response is streaming, the same button becomes Stop. |
If no model is configured yet, the model picker shows the placeholder Select model… and the Send button shows the message Choose or configure a model before sending. Drafts stay saved in this chat. Configure a Router or Advanced provider first (Model Selection).
How a turn unfolds
- Type or paste a prompt. Use Attach file / image / voice as needed.
- (Optional) open Commands to enable a specific tool, skill, MCP server, or saved prompt for this turn.
- (Optional) toggle Confirm before tools if you want approval gates on every tool call.
- Press Enter (Shift-Enter for newline) or click Send.
- The composer locks; the agent's response card appears above it.
- The response streams in chunk-by-chunk. Tool calls appear as collapsible cards inside the response — expand to see the call, arguments, and result.
- When the agent finishes, the composer unlocks.
You can press Stop any time during streaming. The partial response stays in the conversation; you can either send a follow-up or Retry the last turn.
Memory retrieval
When you start a new conversation (or send the first message in a thread fork), Ghast:
- Looks up your vector memory store with the embedding of your message.
- Picks the top matches above a cosine-similarity threshold (default 0.75).
- Surfaces them as a small "memory retrieval" stage card that appears briefly above the response.
- Injects them as context so the agent's reply is anchored in what it already knows about you.
See Memory for what's stored and how to manage it.
Context usage display
The composer surfaces a Context usage breakdown for the current turn:
- Input — what you've sent so far this turn (after attachment expansion).
- Output — running count from the model.
- Reasoning — if applicable for the chosen model.
- Cached — how much of the prefix was served from the model's prompt cache.
used / totaltokens — the running total against the model's context window.
When you approach the window, Ghast runs context compaction: older turns are summarised into a compact representation and replaced in the prompt. The agent doesn't forget; the summary is preserved and shown as a folded "earlier" card you can inspect.
Threads, forks, and side conversations
- New thread — clicking "New" in the sidebar starts a clean conversation with empty context.
- Thread fork — from any message, you can fork into a new conversation that inherits the context up to that point but evolves independently. Useful when you want to try a different approach without losing the current branch.
- Side conversation — some skills open a side conversation with the agent in a focused subtask. The main thread continues; the side conversation closes when its goal is met.
Artefacts
When the agent generates substantial output (a long file, a code block, a diagram), it can break that out into an artefact panel alongside the chat. Artefacts are persisted alongside the conversation and can be opened in the Lightbox viewer for full-screen review.